content
stringlengths 1
15.9M
|
|---|
\section{METHOD}
\label{method}
The proposed methods rely on the the geometry induced by the models~\eqref{eq:staticWrench} and~\eqref{wrenchInSensorCoordinates}.
\subsection{The geometry of the raw measurements}
\label{subsec:geometry}
First, observe that
\[ \text{rank}(M) = 3 .\]
As a consequence, all wrenches $w$ belong to the three-dimensional subspace given by the $\operatorname{span}{(M)} \subset \mathbb{R}^6$. In view of this,
we can state the following lemma.
\lemma{1}{All raw measurements $r$ belong to a three dimensional affine space, i.e. there exist a point $r_m \in \mathbb{R}^6$, an orthonormal basis
$U_1 \in \mathbb{R}^{6 \times 3}$, and for each $r \in \mathbb{R}^6$ a vector $\lambda \in \mathbb{R}^3$ such that
\begin{IEEEeqnarray}{RCL}
\label{decompositionMeasurements}
r &=& r_m + U_1 \lambda.
\end{IEEEeqnarray}
Also, the vector $\lambda$ belongs to a three-dimensional ellipsoid.
}
\begin{proof}
From \eqref{wrenchInSensorCoordinates} and \eqref{eq:staticWrench}, one has:
\begin{equation}
\label{rFromModel}
r = |g|C^{-1} M \hat{g} + o,
\end{equation}
where $\hat{g} := g/|g|$.
The matrix $C^{-1}M \in \mathbb{R}^{6\times3}$ is of rank~$3$.
Consequently, all raw measurements $r$ belong to an affine three-dimensional space
defined by the point $o$ and the basis of $\operatorname{span}{(C^{-1}M)}$.
Now, define $P \in \mathbb{R}^{3 \times 6}$ as the projector of $r$ onto $\operatorname{span}{(C^{-1}M)}$.
Then, the projection $p \in \mathbb{R}^3$ of $r$ onto this span is given by
\begin{IEEEeqnarray}{RCL}
p &=& Pr = |g|PC^{-1} M \hat{g} + Po.
\end{IEEEeqnarray}
By considering all possible orientations of the sensor's frame~$\mathcal{S}$, then the gravity direction $\hat{g}$ spans the unit sphere.
Consequently, the vector $p$ belongs to the \textit{span of the unit sphere applied to the linear transformation $|g|PC^{-1} M$},
i.e. an \textit{ellipsoid centered at the point $Po$}. This in turn implies that when decomposing the vector $r$ as in~\eqref{decompositionMeasurements},
the vector~$\lambda$ necessarily belongs to a three-dimensional ellipsoid.
\end{proof}
To provide the reader with a better comprehension of the above lemma, assume that $r \in \mathbb{R}^{3}$ and that the affine subspace is
a plane, i.e. a two-dimensional space.
As a consequence, all measurements belong to an ellipse lying on this plane -- see Figure~\ref{imageWithPlane}. Observe also that given a point
$\lambda \in \mathbb{R}^2$ on the plane and expressed w.r.t. the basis $U_1$, the relationship~\eqref{decompositionMeasurements} provides with the components of this point in the space
$\mathbb{R}^3$.
\begin{figure}
\vspace{2em}
\centering
\includegraphics[width=0.45\textwidth]{images/imageWithPlane.pdf}
\caption{Example when $r \in \mathbb{R}^3$ and $U_1 = (U_1^1,U_1^2) \in \mathbb{R}^{3 \times 2}$. }
\label{imageWithPlane}
\end{figure}
By leveraging on the above lemma, the next two sections propose a method to estimate the sensor's offset and calibration matrix.
\subsection{Method for estimating the sensor's offset}
\label{offsetEstimationTechnique}
Assume that one is given with a set of measurements $(r_i,g_i)$ with $i \in \{1 \cdots N \}$ corresponding to several body's static orientations.
Then, let
us show how we can obtain the basis $(r_m,U_1)$ in~\eqref{decompositionMeasurements} associated with all measurements $r_i$, and how this basis
can be used for estimating the offset $o$.
Observe that the point $r_m$ can be chosen as any point that belongs to the affine space. Then,
a noise-robust choice for this point is given by the mean value of the measurements~$r_i$,
\begin{IEEEeqnarray}{RCL}
r_m &=& \frac{1}{N} \sum\limits_{i=1}^N r_i.
\end{IEEEeqnarray}
An orthonormal basis $U_1$ can be then obtained by applying the singular value decomposition on the matrix resulting from the difference between
all measurements and $r_m$, i.e.
\begin{IEEEeqnarray}{RCL}
(\tilde{r}_1, \cdots, \tilde{r}_n) = USV^\top,
\end{IEEEeqnarray}
where
\[\tilde{r}_i := r_i - r_m,\] and $U \in \mathbb{R}^{6\times6}$, $S \in \mathbb{R}^{ 6 \times N}$, $V \in \mathbb{R}^{N\times N}$ are
the (classical) matrices obtained from the singular value decomposition.
Note that
only the first three elements on the diagonal of$~S$ are (significantly) different from zero
since all measurements~$r_i$ belong to a three dimensional subspace. Consequently, (an estimate of) the orthonormal basis $U_1$ is given by the first
three columns of the matrix~$U$.
With $(r_m,U_1)$ in hand, the offset $o$ can be easily estimated.
First, note that equation \eqref{decompositionMeasurements} holds for all points belonging to the three dimensional space.
Hence, it holds also for the offset $o$ being the center of the ellipsoid (see Figure~\ref{imageWithPlane}), i.e.
\begin{equation}
\label{eq:offsetInPlane}
o = r_m + U_1 \lambda_o .
\end{equation}
Then to estimate the offset $o$ belonging to $\mathbb{R}^6$, we can estimate the coordinates $\lambda_o$ in the subspace $\mathbb{R}^3$. In view of $U_1^\top U_1 = I_3$, multiplying
the above equation times $U_1^\top$ yields
\begin{equation}
\lambda_o := U^{\top}_1(o-r_m).
\end{equation}
Now, by subtracting $r_m$ from \eqref{rFromModel} and multiplying the resulting equation by $U_1^\top$, one has
\begin{equation}
\label{rFromModel1}
U^{\top}_1\tilde{r}_i = Kg_i + \lambda_o,
\end{equation}
where
\[K := U^\top_1C^{-1}M \quad \text{and} \quad \lambda_o \] are the unknowns
in the above equation. In view of~\eqref{eq:kroneckerVec} the equation~\eqref{rFromModel1} can be written by stacking the obtained vectors for all measurements as
\begin{IEEEeqnarray}{RCLRCL}
\bar{r} &=& \Gamma x, \IEEEyessubnumber \label{eqForOffset} \\
\bar{r} &:=&
\begin{pmatrix}
\tilde{r}^\top_1U_1, && \cdots, && \tilde{r}^\top_NU_1
\end{pmatrix}^\top &\in& \mathbb{R}^{3N\times 1}, \IEEEyessubnumber \\
\Gamma &:=&
\begin{pmatrix}
g^\top_1 \otimes I_3, I_3 \\
. \\
. \\
g^\top_N \otimes I_3, I_3
\end{pmatrix}&\in& \mathbb{R}^{3N\times 12}, \IEEEyessubnumber \\
x &:=&
\begin{pmatrix}
\text{vec}(K) \\
\lambda_o
\end{pmatrix}&\in& \mathbb{R}^{12\times 1}.\IEEEyessubnumber
\end{IEEEeqnarray}
Solving the equation~\eqref{eqForOffset} for the unknown $x$ in the least-square sense provides with an estimate
$\hat{\lambda}_o \in \mathbb{R}^3$ of $\lambda_o$. To obtain the coordinates of this point w.r.t. the six-dimensional space, i.e. the raw measurements space,
we apply the transformation~\eqref{eq:offsetInPlane} as follows:
\[
\hat{o} = r_m + U_1 \hat{\lambda}_o .
\]
\subsection{Method for estimating the sensor's calibration matrix}
\label{calibrationMatrixEstimation}
In this section, we assume no offset, i.e. $o = 0$, which means that this offset has already been estimated
by using one of the existing methods in the literature
or by using the method described in the previous section. Consequently, the relationship between the set of measurements
$(r_i,g_i)$
and the body's inertial characteristics, i.e. mass and center of mass, is given by
\begin{equation}
Cr_i = Mg_i \nonumber.
\end{equation}
In addition, we also assume that the body's inertial characteristics can be modified by adding sample masses at specific relative positions w.r.t. the
sensor frame $\mathcal{S}$. As a consequence, the matrix $M$ in the above equation is replaced by $M_j$, i.e.
\begin{equation}
\label{rawMeasurementsNoOffsetSeveralDataSets}
Cr^j_i = M_jg^j_i ,
\end{equation}
where $j$ indicates that new inertial
characteristics have been obtained by adding sample masses. Observe that $M_j$ can then be decomposed as follows
\begin{IEEEeqnarray}{RCCRCL}
\label{MknownUnkwn}
M_j &:= &M_b& + &M^j_a& \nonumber \\
&=& m
\begin{pmatrix}
I_3 \\
c \times
\end{pmatrix}&
+ &m^j_a
\begin{pmatrix}
I_3 \\
c^j_a \times
\end{pmatrix}&, \nonumber
\end{IEEEeqnarray}
where $(m^j_a,c^j_a)$ are the mass and the vector of the center of mass, expressed w.r.t. the sensor frame $\mathcal{S}$, of the added mass.
In the above equation, $M_b$ is unknown but $M^j_a$ is assumed to be known.
In light of the above, we assume to be given with several data sets
\begin{IEEEeqnarray}{RCL}
\label{dataSets}
R_j := (r^j_1, \cdots,r^j_{N_j}) \in \mathbb{R}^{6\times N_j}, \IEEEyessubnumber \\
G_j := (g^j_1, \cdots,g^j_{N_j}) \in \mathbb{R}^{3\times N_j} ,\IEEEyessubnumber
\end{IEEEeqnarray}
associated with $N_D$ different $(m^j_a,c^j_a)$. Given~\eqref{rawMeasurementsNoOffsetSeveralDataSets} and~\eqref{dataSets},
the measurements associated with the $jth$ data set can be compactly written as
\begin{IEEEeqnarray}{RCL}
CR_j - M_bG_j&=& M^j_aG_j. \nonumber
\end{IEEEeqnarray}
The matrices $C$ and $M_b$ are unknown. Then, in view of~\eqref{eq:kroneckerVec} the above equation can be written for all data sets as follows
\begin{IEEEeqnarray}{RCLRLL}
\Theta x &=& \beta, \IEEEyessubnumber \label{equationForEstimatingC} \\
x &:=&
\begin{pmatrix}
\text{vec}(C) \\
m \\
mc
\end{pmatrix},
&\in& \mathbb{R}^{40\times 1},
\IEEEyessubnumber \\
\Theta &:=&
\begin{pmatrix}
R^\top_1 \otimes I_6, \ -(G^\top_1 \otimes I_6)H \\
. \\
. \\
R^\top_{N_D} \otimes I_6, \ -(G^\top_{N_j} \otimes I_6)H
\end{pmatrix},
&\in& \mathbb{R}^{6N_T\times 40},
\IEEEyessubnumber \IEEEeqnarraynumspace \\
\beta &:=&
\begin{pmatrix}
\text{vec}(M^1_a G_1) \\
. \\
. \\
\text{vec}(M^{N_D}_a G_1)
\end{pmatrix},
&\in& \mathbb{R}^{40\times 1}.
\IEEEyessubnumber
\end{IEEEeqnarray}
with \[N_T = \sum\limits_{j=1}^{N_D} N_j,\] i.e. the number of all measurements, and the matrix $H~\in~\mathbb{R}^{18\times4}$ a properly chosen permutator such that
\[\text{vec}(M_b) = H
\begin{pmatrix}
m \\
mc
\end{pmatrix}.\]
To find the calibration matrix $C$, we have to find the solution~$x$ to the equation~\eqref{equationForEstimatingC}.
The uniqueness of this solution is characterized by the following lemma.
\lemma{2}{
A necessary condition for the uniqueness of
the solution $x$ to the equation~\eqref{equationForEstimatingC} is that the number of data sets
must be greater
than two, i.e.
\begin{IEEEeqnarray}{rcl}
\label{necessityForWellPosedenessC}
N_D \geq 3.
\end{IEEEeqnarray}
}
\begin{figure}[t]
\centering
\subfloat[Dataset 1: no added mass]{\includegraphics[width=0.22\textwidth]{images/dataset0107.pdf}
\label{fig:dataset1}}
\subfloat[Dataset 2]{\includegraphics[width=0.22\textwidth]{images/dataset02.pdf}
\label{fig:dataset2}}
\newline
\subfloat[Dataset 3]{\includegraphics[width=0.22\textwidth]{images/dataset03.pdf}
\label{fig:dataset3}}
\subfloat[Dataset 4]{\includegraphics[width=0.22\textwidth]{images/dataset04.pdf}
\label{fig:dataset4}}
\caption{Added mass configurations for calibration datasets.}
\label{fig:calibration}
\end{figure}
\begin{proof}
This is a proof by contradiction. Assume $N_D~=~2$. In addition, assume, without loss of generality, also that the matrix $M_j$ in
equation~\eqref{rawMeasurementsNoOffsetSeveralDataSets}
is perfectly known (adding unknowns to the considered problem would only require a larger number of data sets).
Then, in view of~\eqref{rawMeasurementsNoOffsetSeveralDataSets} and~\eqref{dataSets}, one has
\begin{IEEEeqnarray}{rcl}
C(R_1,R_2) = (M_1G_1,M_2,G_2). \nonumber
\end{IEEEeqnarray}
The matrix $C$ is the unknown of the above equation. By applying \eqref{eq:kroneckerVec} one easily finds out that there exists a unique~$C$ only
if the rank of the matrix $(R_1,R_2)$ is equal to six, i.e.
\[\text{rank}\Big( (R_1,R_2) \Big) = 6. \]
Recall that the matrix $C$ is invertible by assumption, and thus with rank equal to six.
Consequently
\begin{IEEEeqnarray}{RCL}
\text{rank}\Big( (R_1,R_2) \Big) &=& \text{rank}\Big( (M_1G_1,M_2,G_2)\Big) \nonumber \\
&=& \text{rank}\left( (M_1,M_2)
\begin{pmatrix}
G_1 && 0 \\
0 && G_2
\end{pmatrix} \right) \nonumber \\
&\leq& \min\Big(\text{rank}(M_1,M_2),6\Big).
\nonumber
\end{IEEEeqnarray}
In view of~\eqref{matrixM}, one easily verifies that $\det(M_1,M_2) \equiv 0$, which implies that \[\text{rank}\Big( (R_1,R_2) \Big) \leq 5.\]
\end{proof}
Establishing a sufficient condition for the uniqueness of the solution~$x$ to the equation~\eqref{equationForEstimatingC}
is not as straightforward as proving the above necessary condition, and is beyond the scope of the present paper. Clearly, this uniqueness is related to
the rank of the matrix $\Theta$, and this rank condition can be verified numerically on real data.
Then, the solution~$~x$ can be found by applying least-square techniques, thus yielding estimates of the calibration matrix~$C$ and of the inertial
characteristics of the rigid body.
\section{INTRODUCTION}
The importance of sensors in a control loop goes without saying.
Measurement devices, however, can seldom be used \emph{sine die} without being subject to periodic calibration procedures.
Pressure transducers, gas sensors, and thermistors are only a few examples of measuring devices that need to be calibrated periodically for providing
the user with precise and robust measurements.
Of most importance, calibration procedures may require to move the sensor from the hosting system to specialized laboratories, which are equipped with
the tools for performing the calibration of the measuring device. This paper presents techniques to calibrate
strain gauges six-axis force-torque sensors \emph{in situ}, i.e. without the need of removing the sensor from the hosting system.
The proposed calibration method is particularly suited for robots with embedded six-axis force-torque sensors installed within limbs~\cite{Fumagalli2012}.
Calibration of six-axis force-torque sensors has long attracted the attention of the robotic community~\cite{braun2011}.
The commonly used model for predicting the force-torque
from the raw measurements of the sensor is an affine model.
This model is sufficiently accurate
since these sensors are mechanically designed and mounted so that the strain deformation is (locally) linear with respect to the applied forces and torques.
Then, the calibration of the sensor aims at determining the two components of this model, i.e. a six-by-six matrix and a six element vector.
These two components are usually referred to as the sensor's \emph{calibration matrix} and \emph{offset}, respectively.
In standard operating conditions, relevant changes in the calibration matrix may occur in months.
As a matter of fact, manufacturers such as ATI~\cite{atimanual} and Weiss Robotics~\cite{kms40manual}
recommend to calibrate force-torque sensors at least once a year.
Preponderant changes in the sensor's offset can occur in hours, however, and this in general requires to estimate the
offset before using the sensor.
The offset models the sensitivity of the semiconductor strain gauges with respect to temperature.
\begin{figure}
\vspace{0em}
\centering
\includegraphics[width=0.225\textwidth]{images/leg.pdf};
\caption{iCub's leg with the two force/torque sensors and an additional accelerometer.}
\label{fig:iCubLeg}
\end{figure}
Classical techniques for determining the offset of a force-torque sensor exploit the aforementioned affine model between
the raw measurements and the load attached to the sensor. In particular, if no load is applied to the measuring device, the output of the sensor corresponds to
the sensor's offset. This offset identification procedure, however, cannot be always performed since it may require to take the hosting system apart
in order to unload the force-torque sensor. Another widely used technique for offset identification is to find two sensor's orientations that induce
equal and opposite loads with respect to the sensor. Then, by summing up the raw measurements associated with these two orientations, one can estimate
the sensor's offset. The main drawback of this technique is that the positioning of the sensor at these opposite configurations may require to move
the hosting system beyond its operating domain.
Assuming a preidentified offset, non-in situ identification of the calibration matrix is classically performed by exerting on
the sensor a set of force-torques known \emph{a priori}. This usually requires to place
sample masses at precise relative positions with respect to the sensor. Then, by comparing
the known gravitational force-torque with that measured by the sensor,
one can apply linear least square techniques to identify the sensor's calibration matrix.
For accurate positioning of the sample masses, the use of robotic positioning devices
has also been proposed in the specialized
literature~\cite{uchiyama1991systematic}~\cite{watson1975pedestal}.
To reduce the number of sample masses,
one can apply constrained forces, e.g. constant norm forces, to the measuring device.
Then these constrains can be exploited during the computations for identifying the calibration matrix~\cite{voyles1997shape}.
To avoid the use of added masses, one can use a supplementary already-calibrated measuring device that measures
the force-torque exerted on the sensors~\cite{faber2012force}~\cite{oddo2007}.
On one hand, this calibration technique avoids the problem of precise positioning of the added sample masses.
On the other hand,
the supplementary sensor may not always be available.
All above techniques, however, cannot be performed in situ, thus being usually time consuming and expensive.
To the best of our knowledge, the first \emph{in situ} calibration method for force-torque sensors was proposed
in \cite{shimanoroth}. But this method exploits the topology of a specific kind of manipulators, which are equipped with
joint torque sensors then leveraged during the estimation.
A recent result~\cite{Gong2013} proposes another \emph{in situ} calibration technique for six-axis force-torque sensors.
The technical soundness of this work, however, is not clear to us. In fact, we show that a necessary condition for identifying the calibration matrix
is to take measurements for at least three different added masses, and this requirement was not met by the algorithm~\cite{Gong2013}.
Another in situ calibration technique for force-torque sensors can be found in \cite{roozbahani2013novel}.
But the use of supplementary already-calibrated force-torque/pressure sensors impairs this technique for the reasons we have discussed before.
This paper presents in situ calibration techniques for six-axis force-torque sensors using accelerometer measurements.
The proposed method exploits the geometry induced by the affine model between the raw measurements and the gravitational force-torque applied to the sensor.
In particular, it relies upon the properties that all gravitational raw measurements belong to a three-dimensional space, and that in this space they form an ellipsoid.
Then, the contribution of this paper is twofold. We first propose a method for estimating the sensor's offset, and then a procedure for identifying
the calibration matrix. The latter is independent from the former, but requires to add sample masses to the rigid body attached to the sensor. Both methods are independent from the inertial characteristics
of the rigid body attached to the sensor.
The proposed algorithms are validated on the iCub platform by calibrating
two force-torque sensors embedded in the robot~leg.
The paper is organized as follows. Section~\ref{sec:background} presents the notation used in the paper and the problem statement with the assumptions.
Section~\ref{method} details the proposed method for the calibration of six-axis force-torque sensors.
Validations of the approach are presented in Section~\ref{experiments}.
Remarks and perspectives conclude the paper.
\section{EXPERIMENTS}
\label{experiments}
To test the proposed method, we calibrated the two force-torque sensors embedded in the leg of the iCub humanoid robot
-- see Figure~\ref{fig:iCubLeg}. The mass and the center of mass of this leg are unknown.
To apply the method described in section~\ref{method}, we need to add sample masses to the iCub's leg.
For this purpose, we installed on the robot's foot a beam to which samples masses can be easily attached. This beam also houses a XSens MTx IMU.
The supplementary accelerometer will no longer be required when using
the iCub version 2, which will be equipped with 50 accelerometers
distributed on the whole body.
The iCub's knee is kept fixed with a position controller, so we can consider the robot's leg as a unique rigid body.
Consequently, the accelerometer measures the gravity force w.r.t. both sensors. Recall also that
to apply the proposed methods, we need to modify the orientations of the sensors' frames. To do so, we modified the \emph{front-back} and
\emph{lateral} angles associated with the robot's hip.
\begin{figure}[t]
\centering
\subfloat[Dataset 5]{\includegraphics[width=0.22\textwidth]{images/dataset05.pdf}
\label{fig:dataset5}}
\subfloat[Dataset 6]{\includegraphics[width=0.22\textwidth]{images/dataset06.pdf}
\label{fig:dataset6}}
\newline
\subfloat[Dataset 7: no added mass]{\includegraphics[width=0.22\textwidth]{images/dataset0107.pdf}
\label{fig:dataset7}}
\subfloat[Dataset 8]{\includegraphics[width=0.22\textwidth]{images/dataset08.pdf}
\label{fig:dataset8}}
\caption{Added mass configurations for validation datasets.}
\label{fig:validation_masses}
\end{figure}
We collected data associated with eight different added mass configurations, each of which is characterized by a mass placed at a different
location with respect to the beam. In other words,
we collected eight different data sets. Figures~\ref{fig:calibration} and~\ref{fig:validation_masses} show the configurations of these data sets.
For each of these data sets, we slowly\footnote{The iCub front-back and lateral angles were moved with peak velocities of $2 \deg/s$.}
moved the \emph{front-back} and
\emph{lateral} angles of the robot hip, spanning a range of $70 \deg$ for the \emph{front-back} angle, and a range of $90 \deg$ for the lateral angle.
We sampled the two F/T sensors and the accelerometer at 100 Hz, and we filtered the obtained signals with a Savitzky-Golay filter of third order
with a windows size of 301 samples.
\begin{figure}
\vspace{0.5em}
\centering
\includegraphics[width=0.52\textwidth]{images/raw_ellipsoid.pdf}
\caption{Dark blue: raw measurements of the sensor embedded in the leg for dataset 1 projected in the 3D subspace through $U_1$.
In light blue an ellipsoid fitted to the measured points is added, to highlight the fact that the measured data lie on an ellipsoid. The $o'$ point estimated
with the method described in \ref{offsetEstimationTechnique} is the center of this ellipsoid.
}
\label{ellipsoidWithRawData}
\end{figure}
We estimated the sensors' offsets by applying the method described in section~\ref{offsetEstimationTechnique} on all eight data sets.
Figure~\ref{ellipsoidWithRawData} verifies the statement of Lemma~1, i.e. the raw measurements belong to a three dimensional ellipsoid.
In particular, this figure shows the measurements
$r_i$ projected onto the three dimensional space where the ellipsoid occurs, i.e. the left hand side of the equation~\eqref{rFromModel1}.
Then, we removed the offset from the raw measurements to apply the estimation method for the calibration matrix described in
section~\ref{calibrationMatrixEstimation}.
The two sensors' calibration matrices were identified by using only four data sets (see Figure~\ref{fig:calibration}).
The other four were used to validate the obtained calibration results (see Figure~\ref{fig:validation_masses}).
The qualitative validation of the calibration procedure is based on the fact that the weight of the leg is constant for each data sets.
Consequently, if we plot
the force measured by the sensors, i.e. the first three rows of left hand side of the sensor's equation
\[ w = C(r - o), \]
these forces must belong to a sphere, since they represent the (constant norm) gravity force applied to the sensors.
As for elements of comparisons, we can also plot the first three rows of the above equation when evaluated with the calibration matrix that was
originally provided by the
manufacturer of the sensors.
\begin{figure}[h]
\vspace{0.5em}
\centering
\subfloat[Validation results for leg F/T sensor.]{\includegraphics[width=0.45\textwidth]{images/leg_validation.pdf}}
\newline
\subfloat[Validation results for foot F/T sensor.]{\includegraphics[width=0.45\textwidth]{images/foot_validation.pdf}}
\caption{Dark green: force measurements obtained through the calibration matrix estimated using the proposed technique. Dark red:
force measurements obtained through the calibration matrix provided with the sensor. Light red and light green surfaces: ellipsoids fitted to the
measured forces.}
\label{fig:validation}
\end{figure}
Figure~\ref{fig:validation} depicts the force measured by the sensor with the estimated calibration matrix (in green)
and with the calibration matrix provided by the manufacturer (in red). It is clear to see that the green surfaces are much more spherical than the red ones.
As a matter of fact, Table \ref{table:soa} lists the semi axes of the ellipsoids plotted in Figures~\ref{fig:validation}, and clearly shows
that the green surfaces represent spheres much better than the red ones.
Interestingly enough, the force-torque sensor embedded in the iCub leg is much more mis-calibrated than that embedded in
the foot. In fact, by looking at the data sheets describing the technological lives of these sensors, we found out that the force-torque sensor embedded in
the leg is much older than that in the foot, which means that the calibration matrix of the leg's sensor is much older than that of the foot's sensor.
The quantitative validation of the proposed calibration procedure is performed by comparing the known weights of the added masses with the weights estimated by the sensors.
Table \ref{table:mass} shows that the estimated weights obtained after performing the proposed calibration are better than those estimated by using the calibration matrix
provided by the sensor manufacturer. A similar comparison was conducted on the estimation of the sample mass positions (Table \ref{table:cmass}). It has to be noticed that the errors in estimating the mass position are relatively high, but this is due to the choice of using relatively small masses with respect to the sensor range and signal to noise ratio.
\begin{table*}[ht]
\caption{Qualitative calibration evaluation on validation dataset: ellipse semiaxes after calibration}
\centering
\begin{tabular}{p{0.8cm} | p{1.2cm} | p{1.2cm} | p{1.3cm} p{1.3cm} p{1.3cm} | p{1.3cm} p{1.3cm} p{1.3cm}}
\emph{Sensor} & \emph{Dataset} & \emph{Added mass} (Kg) & \multicolumn{3}{p{3.9cm} |}{\emph{Semiaxes length [N] with proposed calibration}} & \multicolumn{3}{p{3.9cm}}{\emph{Semiaxes length [N] with manufacturer calibration}}\\
\hline \rowcolor[gray]{.9}
\rowcolor[gray]{.9} Foot & Dataset 5 & 0.51 & 13.6 & 13.1 & 12.9 & 13.5 & 10.2 & 4.6 \\
\rowcolor[gray]{.9} Foot & Dataset 6 & 0.51 & 13.5 & 12.9 & 12.7 & 13.6 & 10.5 & 9.4 \\
\rowcolor[gray]{.9} Foot & Dataset 7 & 0 & 8.4 & 7.9 & 7.4 & 8.5 & 6.9 & 6.2 \\
\rowcolor[gray]{.9} Foot & Dataset 8 & 0.51 & 13.7 & 12.5 & 12.0 & 15.7 & 13.6 & 10.4 \\
\hline
Leg & Dataset 5 & 0.51 & 34.4 & 33.3 & 32.5 & 76.5 & 49.4 & 45.6 \\
Leg & Dataset 6 & 0.51 & 34.8 & 33.5 & 32.8 & 82.4 & 49.3 & 47.3 \\
Leg & Dataset 7 & 0 & 30.9 & 28.2 & 26.9 & 77.0 & 44.5 & 40.0 \\
Leg & Dataset 8 & 0.51 & 35.52 & 33.30 & 32.2 & 88.9 & 49.5 & 48.3 \\
\hline
\end{tabular}
\label{table:soa}
\end{table*}
\begin{table*}[ht]
\caption{Qualitative calibration evaluation on validation dataset: sample mass estimations}
\centering
\begin{tabular}{p{0.8cm} | p{1.2cm} | p{3cm} | p{3cm} | p{4cm} }
\emph{Sensor} & \emph{Dataset} & \multicolumn{3}{c}{\emph{Added mass} (Kg)} \\
\hline \rowcolor[gray]{.9}
\hline \rowcolor[gray]{.9} & & Ground truth & With proposed calibration & With manufacturer calibration \\
\rowcolor[gray]{.9} Foot & Dataset 5 & 0.51 & 0.53 & 0.06 \\
\rowcolor[gray]{.9} Foot & Dataset 6 & 0.51 & 0.52 & 0.27 \\
\rowcolor[gray]{.9} Foot & Dataset 7 & 0 & -0.03 & -0.13 \\
\rowcolor[gray]{.9} Foot & Dataset 8 & 0.51 & 0.46 & 0.45 \\
\hline
Leg & Dataset 5 & 0.51 & 0.51 & 2.77 \\
Leg & Dataset 6 & 0.51 & 0.54 & 3.04 \\
Leg & Dataset 7 & 0 & -0.04 & 2.39 \\
Leg & Dataset 8 & 0.51 & 0.51 & 3.25 \\
\hline
\end{tabular}
\label{table:mass}
\end{table*}
\begin{table*}[ht]
\caption{Qualitative calibration evaluation on validation dataset: center of mass estimations}
\centering
\begin{tabular}{p{1.8cm} | p{2.2cm} | p{1.4cm} | p{0.7cm} | p{0.7cm} || p{0.7cm} | p{0.7cm} | p{0.7cm}|| p{0.7cm} | p{0.7cm} | p{1.2cm} }
\emph{Sensor} & \emph{Dataset} & \multicolumn{9}{c}{\emph{Center of mass position for the added mass} [cm]} \\
\hline \rowcolor[gray]{.9} - & - & \multicolumn{3}{c || }{Ground truth} & \multicolumn{3}{c ||}{With proposed calibration} & \multicolumn{3}{c}{With manufacturer calibration} \\
\rowcolor[gray]{.9} Foot & Dataset 5 & 39 & {-}3.5 & 2.9 & 31 & 8.8 & $-$8.3 & 273 & $-$83 & 81 \\
\rowcolor[gray]{.9} Foot & Dataset 6 & 21 & 0 & 6.3 & 19 & 9.9 & $-$5 & 30 & $-$18 & 18 \\
\rowcolor[gray]{.9} Foot & Dataset 7 & - & - & - & - & - & - & - & - & - \\
\rowcolor[gray]{.9} Foot & Dataset 8 & {-}4 & 0 & 6.3 & 5.5 & 9 & $-$3.2 & $-$8.6 & $-$12 & 10 \\
\hline
Leg & Dataset 5 & 39 & $-$3.5 & 39 & 28 & 11 & 29 & 16 & 6.5 & 29 \\
Leg & Dataset 6 & 20 & 0 & 43 & 17 & 9.7 & 30 & 12 & 5.1 & $-$27 \\
Leg & Dataset 7 & - & - & - & - & - & - & - & - & - \\
Leg & Dataset 8 & $-$4.3 & 0 & 43 & 3.8 & 8.8 & 32 & 7.5 & 4.9 & $-$26 \\
\hline
\end{tabular}
\label{table:cmass}
\end{table*}
\section{CONCLUSIONS}
\label{conclusions}
In this paper, we addressed the problem of calibrating six-axis force-torque sensors in situ by using the accelerometer measurements.
The main point was to highlight the geometry behind the gravitational
raw measurements of the sensor, which can be shown to belong to a three-dimensional affine space, and more precisely to a three-dimensional ellipsoid.
Then, we propose a method to identify first the sensor's offset, and then the sensor's calibration matrix. The latter method requires to add sample masses
to the body attached to the sensor, but is independent from the mass and the center of mass of this body. We show that a necessary condition to identify the sensor's calibration matrix is to collect data for more than two
sample masses. The validation of the method was performed by calibrating the two force-torque sensors embedded in the iCub leg.
The main assumption under the proposed algorithm is that the measurements were taken for static configurations of the rigid body attached to the sensor.
Then, future work consists in weakening this assumption, and developing calibration procedures that hold even for dynamic motions of the rigid body. This
extension requires to use the gyros measurements. In addition, comparisons of the proposed method versus existing calibration
techniques is currently being investigated, and will be presented in a forthcoming publication.
\section{BACKGROUND}
\label{sec:background}
\subsection{Notation}
The following notation is used throughout the paper.
\begin{itemize}
\item The set of real numbers is denoted by $\mathbb{R}$. Let $u$ and $v$ be two $n$-dimensional column vectors of real numbers, i.e. $u,v \in \mathbb{R}^n$,
their inner product is denoted as $u^\top v$, with ``$\top$'' the transpose operator.
\item Given $u \in \mathbb{R}^3$, $u \times$ denotes the skew-symmetric matrix-valued operator associated with the cross product in
$\mathbb{R}^3$.
\item The Euclidean norm of either a vector or a matrix of real numbers is denoted by $|\cdot |$.
\item $I_n \in \mathbb{R}^{n \times n}$ denotes the identity matrix of dimension~$n$;
$0_n \in \mathbb{R}^n$ denotes the zero column vector of dimension~$n$; $0_{n \times m} \in \mathbb{R}^{n \times m}$ denotes the zero matrix of dimension~$n \times m$.
\item The vectors $e_1,e_2,e_3$ denote the canonical basis of $\mathbb{R}^3$.
\item Let $\mathcal{I} = \{O;e_1,e_2,e_3\}$ denote a
fixed inertial frame with respect to (w.r.t.) which the sensor's absolute orientation is measured.
Let $\mathcal{S} = \{O';i,j,k\}$ denote a frame attached to the sensor, where the matrix $T := (i,j,k)$ is a rotation matrix
whose column vectors are the vectors
of coordinates of $i,j,k$ expressed in the basis of $\mathcal{I}$.
\item The sensor's orientation w.r.t. $\mathcal{I}$ is characterized by the rotation matrix $T$. Given a vector of coordinates $\bar{x} \in \mathbb{R}^3$
expressed w.r.t. $\mathcal{I}$, we denote with $x$ the same vector expressed w.r.t. $\mathcal{S}$, i.e. $\bar{x} = Tx$.
\item Given $A \in \mathbb{R}^{n \times m}$ and $B \in \mathbb{R}^{p \times q}$, we denote with $\otimes$ the Kronecker product $A \otimes B \in \mathbb{R}^{np \times mq}$.
\item Given $X \in \mathbb{R}^{m \times p}$, $\text{vec}(X) \in \mathbb{R}^{nm}$ denotes the column vector obtained by stacking the columns of the matrix~$X$.
In view of the definition of $\text{vec}(\cdot)$, it follows that \begin{equation}\label{eq:kroneckerVec} \text{vec}(AXB) = \left( B^{\top} \otimes A \right) \text{vec}(X).\end{equation}
\end{itemize}
\subsection{Problem statement and assumptions}
We assume that the model for predicting the force-torque (also called wrench)
from the raw measurements is an affine model, i.e.
\begin{equation}
\label{wrenchInSensorCoordinates}
w = C ( r - o),
\end{equation}
where
$ {w} \in \mathbb{R}^{6}$ is the wrench exerted on the sensor expressed in the sensor's frame,
$r \in \mathbb{R}^{6}$ is the raw output of the sensor,
$ {C} \in \mathbb{R}^{6 \times 6}$ is the
invertible
calibration matrix, and
${o} \in \mathbb{R}^6$ is the sensor's offset.
The calibration matrix and the offset are assumed to be constant.
We assume that the sensor is attached to a rigid body
of
(constant) mass $m~\in~\mathbb{R}^+$ and with a center of mass whose position
w.r.t. the sensor frame $\mathcal{S}$ is characterized the vector $c \in \mathbb{R}^3$.
The gravity force applied to the
body
is given by
\begin{IEEEeqnarray}{RCL}
\label{eq:g}
m\bar{g} = mTg ,
\end{IEEEeqnarray}
with $\bar{g},g \in \mathbb{R}^3$ the gravity acceleration expressed w.r.t. the inertial and sensor frame, respectively. The gravity acceleration $\bar{g}$ is constant, so the vector $g$ does not have a constant direction,
but has a constant norm.
Finally, we make the following main assumption.
\hp{}{
The raw measurements $r$ are taken for static configurations of the
rigid body
attached to the sensor, i.e. the angular velocity of the frame $\mathcal{S}$ is always zero.
Also, the gravity acceleration $g$ is measured by an accelerometer installed on the rigid body.
Furthermore, no external force-torque, but the gravity force, acts on the rigid body. Hence
}
\begin{IEEEeqnarray}{RCL}
\label{eq:staticWrench}
w &=&
M(m,c) g, \IEEEyessubnumber \\
M(m,c) &:=& m
\begin{pmatrix}
I_3 \\
c \times
\end{pmatrix}. \label{matrixM} \IEEEyessubnumber
\end{IEEEeqnarray}
\remark{}{ We implicitly assume that the accelerometer frame is aligned with the force-torque sensor frame.
This is a convenient, but non necessary, assumption.
In fact, if the relative rotation between the sensor frame $\mathcal{S}$ and the accelerometer frame is unknown,
it suffices to consider the accelerometer frame as the sensor frame $\mathcal{S}$.
}
Under the above assumptions, what follows proposes a new method for estimating the sensor's offset $o$
and for identifying the sensor's calibration matrix $C$ without the need of removing the sensor from the
hosting system.
|
\section{Introduction}
Graphene continues to receive significant attention for its numerous intriguing interaction effects and transport properties~\cite{katsnelson06,shytov07,gu11,beenakker06}. Recent experiments with high--quality samples suggest the existence of non--trivial correlated phases in graphene, such as excitonic condensates in non--zero magnetic fields~\cite{gorbachev12}. However, the zero--field condensate predicted theoretically \cite{min08,zhang08,abergel2013} has not been observed~\cite{gorbachev12,kim11}. This has provoked significant interest in the archetypical two--body problem~\cite{sabio10,lee12,mahmoodian13}.
Most interesting properties of graphene are due to the existence of Dirac cones~\cite{wallace47}
located near two points in the Brillouin zone, $\vec{K}^+$ and $\vec{K}^-$. Each cone hosts positive and negative
energy states, each with linear dispersion $\epsilon=\pm v_Fp$, akin to electron and positron states
in quantum electrodynamics. (Here $v_F=10^6ms^{-1}$ is the Fermi velocity\cite{deacon07} and $p$ is the magnitude
of the momentum $\vec{p}=(p_x,p_y)$.) The symmetry between positive and negative cones results in the compensation
of total kinetic energy for two particles with opposite momenta. Thus, the two--particle states can be divided into
\textit{dispersing} (spanned by states where $E\neq 0$) and \textit{non--dispersing} ($E=0$) sectors.
States in the non--dispersing sector have momentum--independent eigenvalues, and are therefore infinitely degenerate.
However the linear dispersion is only accurate at low energies; higher order terms can at times reveal important physics hidden by the conical approximation.
The effect of dispersion on the excitonic physics can be seen if one considers Dirac particles interacting via the Coulomb potential,
which scales as $U(r)\sim\frac{Ze^2}{\epsilon r}$. In the conical approximation, the kinetic energy scales as $v_Fp\sim \frac{\hbar v_F}{r}$ due to the uncertainty principle.
In the case of a single particle interacting with a static charge, Dirac vacuum reconstruction occurs when the potential energy dominates
over kinetic energy\cite{shytov07}: $\frac{Ze^2}{\epsilon}>\frac{\hbar v_F}{2}$. In the case of two carriers, however, a doubling of the
relative velocity effectively doubles the critical charge for which collapse is possible\cite{sabio10}: $Z_C\gtrsim 2$. This renders
the strongly--interacting regime irrelevant for electron--hole physics. For a weaker, quadratic dispersion ($p^2\sim\frac{\hbar^2}{r^2}$),
however, bound states can form for arbitrarily small interaction strength. For this reason, the previously neglected contribution of non--dispersing states \cite{sabio10,lee12} can be
important for understanding the two--particle physics of graphene. To this end, the two--body problem
was re--analyzed in Ref.~[\onlinecite{mahmoodian13}] with the inclusion of trigonal--warping terms which preserve the symmetry between the two cones
but do not lead to non--zero kinetic energy when the total momentum of the pair is zero. In this paper we introduce quadratic momentum terms due to next--nearest--neighbour hopping which were ignored by Ref.~[\onlinecite{mahmoodian13}]
and show that this leads to the finite kinetic energy necessary for bound state formation. In particular, we will show that this leads to a new class of states which exist regardless of the orientation in momentum space (and cannot arise due to trigonal terms alone due to sign--indefinite kinetic energy). The formation of pairs of particles in the same valley is allowed for the model we consider here, which was not the case for the electron--hole case in Ref.~[\onlinecite{mahmoodian13}].\\
\indent In this work, we show that Cooper--pair--like states can be formed in the subspace of non--dispersing two--particle states. The dynamics in this sector is governed by quadratic terms in the single--particle dispersion. Two such contributions are possible: an isotropic term due to next--nearest--neighbour hopping, $\epsilon_I\propto p^2$, and an anisotropic term due to trigonal--warping, $\epsilon_A\propto p^2\sin(3\phi_{\vec{p}})$, where $\phi_{\vec{p}}$ is the polar angle in momentum space defined by $\tan(\phi_{\vec{p}})=\frac{p_y}{p_x}$. We show that, depending upon the relative magnitudes of these two terms, two regimes are possible. When the isotropic contribution dominates, bound states can be formed; otherwise it is possible to form non--dispersing quasibound states (which can leak into the continuum.) We calculate the binding energies of such states numerically, for a double--layer configuration, and discuss the decay rate due to coupling to the continuum of dispersing states.\\
\indent The rest of this work shall be structured as follows. In Sec. II we construct the effective Hamiltonian of a pair of interacting electrons in double--layer graphene. In particular, we discuss the inclusion of a finite band curvature into this Hamiltonian, before projecting out the high energy states and focusing on the non--dispersing sector discussed above. In Sec. III we will calculate the binding energies of these pairs, approaching the problem from two directions. Firstly, we approximate the energies by treating the potential within the harmonic approximation. Secondly, we find the direct--pair energies by calculating the local density of states numerically. We find that these approaches give an order--of--magnitude agreement, with our numerical method yielding $E=45$meV for the bound state of highest energy. In Sec. IV we present an analysis of the semiclassical trajectories of the pair, which gives an intuitive view of how a pair can become bound in the presence of a repulsive interaction. In Sec. V we discuss the coupling of these states to the dispersing sector. Although the potential can destroy the states in principle, the decay rate vanishes by symmetry for the highest energy level of the pair. The kinetic energy also leads to a decay, but we argue that the decay rate is small enough to validate our consideration of the non--dispersing sector in isolation. Finally, in Sec. VI we summarize our results. Discussions of the interlayer electron--electron interaction and our numerical approach to calculating the local density of states can be found in the appendices.
\section{Effective Hamiltonian}
\indent We begin by analyzing the kinetic energy of two Dirac quasiparticles in graphene. Since electrons in graphene can reside on two sublattices, $A$ and $B$, they are to be described by a two--component Dirac spinor. The internal degree of freedom arising due to the presence of the sublattices is known as pseudospin (for a discussion, see Ref. [\onlinecite{katsnelson12}]). In the low--energy approximation, the Dirac spinors for the two valleys, $\vec{K}^+$ and $\vec{K}^-$, can be treated as fully independent. We define these spinors as $\psi_{\vec{K}^+}=\left[\psi_{\vec{K}^+}^A,\psi_{\vec{K}^+}^B\right]^T$ and $\psi_{\vec{K}^-}=\left[\psi_{\vec{K}^-}^B,\psi_{\vec{K}^-}^A\right]^T$, where $A$ and $B$ label the probability amplitudes for the two sublattices. In the conical approximation, the dynamics of the pair is governed by the dispersion arising from the relative motion of its constituent particles, $\widehat{H_L}=v_F\vec{\sigma}_1\cdot\hat{\boldsymbol{p}}_1+v_F\vec{\sigma}_2\cdot\hat{\boldsymbol{p}}_2$, where $\boldsymbol{\sigma}_i$ is the pseudospin operator, subscripts denote the particle number and $\hat{\boldsymbol{p}}_i$ is a small momentum measured with respect to the $\vec{K}^+$-- or $\vec{K}^-$--point. We focus on states with zero total momentum, such that $\vec{p}_1=-\vec{p}_2$. The eigenstates of $\widehat{H_L}$ are given by:
\begin{gather}\label{wf1}
\begin{aligned}
\ket{1,\phi_{\vec{p}}}&=\frac{1}{\sqrt{2}}\left[e^{-i\phi_{\vec{p}}}\ket{\uparrow\uparrow}+e^{i\phi_{\vec{p}}}\ket{\downarrow\downarrow}\right],\\
\ket{2,\phi_{\vec{p}}}&=\frac{1}{\sqrt{2}}\left[\ket{\uparrow\downarrow}+\ket{\downarrow\uparrow}\right],\\
\ket{3,\phi_{\vec{p}}}&=\frac{1}{2}e^{i\phi_{\vec{p}}}\ket{\downarrow\downarrow}-\frac{1}{2}e^{-i\phi_{\vec{p}}}\ket{\uparrow\uparrow}+\frac{1}{2}\left[\ket{\uparrow\downarrow}-\ket{\downarrow\uparrow}\right],\\
\ket{4,\phi_{\vec{p}}}&=\frac{1}{2}e^{-i\phi_{\vec{p}}}\ket{\uparrow\uparrow}-\frac{1}{2}e^{i\phi_{\vec{p}}}\ket{\downarrow\downarrow}+\frac{1}{2}\left[\ket{\uparrow\downarrow}-\ket{\downarrow\uparrow}\right].
\end{aligned}\raisetag{4\baselineskip}
\end{gather}
\noindent The vectors $\ket{\uparrow}$ and $\ket{\downarrow}$ here represent the up-- and down-- pseudospin states for a single particle, so that the two--particle eigenstates are, e.g., $\ket{\uparrow\downarrow}=\ket{\uparrow}\bigotimes\ket{\downarrow}$. The states in Eq. (\ref{wf1}) have corresponding eigenvalues $E_{1,2}=0$ and $E_{3,4}=\pm 2v_Fp$. The subspace spanned by $\ket{1,\phi_{\vec{p}}}$ and $\ket{2,\phi_{\vec{p}}}$ from Eq. (\ref{wf1}) is the \textit{non--dispersing sector}. Such states are formed by electron quasiparticles in opposite cones, with the same magnitude of momentum, so that the relative velocity of the pair vanishes. Similarly, the subspace spanned by $\ket{3,\phi_{\vec{p}}}$ and $\ket{4,\phi_{\vec{p}}}$ forms the \textit{dispersing sector}, in which the velocities are opposite. In the absence of interactions, all states in the non--dispersing sector are infinitely degenerate.\\
\indent We note that this degeneracy is lifted if the symmetry between the upper and lower cones is broken, e.g., by a small band curvature. We extend the kinetic energy by quadratic terms compatible with the symmetries of the honeycomb lattice (see Refs.~[\onlinecite{katsnelson12}] and~[\onlinecite{,mccann06c}]). We write the single--particle kinetic energy in the form:
\begin{gather}
\begin{aligned}
\widehat{H}_{j}&=v_F\vec{\sigma}_j\cdot\vec{p}_j-\frac{p^2_j}{4m^*}+\tau_j\mu (p_{x,j}+ip_{y,j})^2\sigma_{+,j}+\text{H.c.},\label{single}
\end{aligned}\raisetag{0.5\baselineskip}
\end{gather}
\noindent where $j$ is the particle number, $p_j^2=p_{x,j}^2+p_{y,j}^2$, $\sigma_{+,j}=\frac{1}{2}\left(\sigma_{x,j}+ i\sigma_{y,j}\right)$, $\tau_j=\pm 1$ for an electron in the $\vec{K}^{\pm}$ valley (determining the sign of the trigonal--warping) and H.c. denotes the Hermitian conjugate. The second term in Eq. (\ref{single}) is invariant under all two--dimensional rotations, and arises microscopically from contributions due to the hopping of electrons from one atom to its next--nearest--neighbour, giving $m^*=\frac{\hbar^2}{9a^2t'}$, where $t'$ is the next--nearest--neighbour hopping parameter \cite{castro09}. The third term (including H.c.) is invariant under rotations by $120^\circ$. This term represents trigonal--warping, and originates from nearest--neighbour hopping, expanded to second order in momentum\cite{castro09,katsnelson12}, so that $\mu=\frac{3a^2t}{8\hbar^2}$.\\
\indent To examine the dynamics in the non--dispersing sector, we restrict the two--particle Hamiltonian to this subspace. We explicitly treat two distinct cases: \textit{direct pairs} (when both particles are in the same valley) and \textit{indirect pairs} (opposite valleys). All states $\ket{1,\phi_{\vec{p}}}$ and $\ket{2,\phi_{\vec{p}}}$ are annihilated by the operator $(\vec{\sigma}_1-\vec{\sigma}_2)\cdot\vec{p}$. Calculating the matrix elements of the kinetic energy we find the effective Hamiltonian:
\begin{align}
\widehat{H}_{1,2}^\text{eff}=\left[\begin{array}{cc}
-\frac{p^2}{2m^*} & \tau_{1,2}\mu p^2\sin(3\phi_{\vec{p}}) \\
\tau_{1,2}\mu p^2\sin(3\phi_{\vec{p}}) & -\frac{p^2}{2m^*}
\end{array}\right]\label{direct},
\end{align}
\noindent where the rows and columns correspond to states $\ket{1,\phi_{\vec{p}}}$ and $\ket{2,\phi_{\vec{p}}}$ and $\tau_{1,2}=\tau_1+\tau_2$.\\
\indent We will show that some of the features in the dynamics of two--particle states crucially depend upon the signs and relative magnitudes of the quadratic terms, i.e., on the values of $m^*$ and $\mu$. It has been shown by a variety of different approaches \cite{deacon07,castro09,reich02,kretinin13,kuhne12} that $t$ and $t'$ have the same sign, however there is a disagreement on the precise value of $t'$. \textit{Ab initio} calculations\cite{castro09, reich02} give the range $0.02t\leq t'\leq 0.2t$, while cyclotron resonance\cite{deacon07}, quantum capacitance \cite{kretinin13} and polarization--resolved magnetospectroscopy \cite{kuhne12} measurements have produced $t'=0.04t,0.11t\text{ and }0.14t$ respectively. The full two--particle kinetic energy is $\widehat{H}_{1,2}=\widehat{H}_{1}+\widehat{H}_{2}$, and so $-m^*$ plays the role of a two--particle reduced mass due to the $-\frac{p^2}{2m^*}$ term which arises when $\widehat{H}_{1,2}$ is written explicitly, with the corresponding range of values $0.7\leq\frac{m^*}{m_e}\leq 7.5$. This implies that the isotropic kinetic energy term is negative definite, which will be shown to be of crucial importance to the spectrum of two--particle states.\\
\section{Bound states of electron pairs}
\indent To understand the dynamics of pairs described by the kinetic energy terms in $\widehat{H}_{1,2}$, let us first consider the simplest case of indirect pairs, where the electrons are in opposite valleys. In this configuration $\tau_{1,2}=0$, so that the contribution of trigonal--warping vanishes and the only remaining kinetic term is $-\frac{p^2}{2m^*}$. The dynamics of the interacting pair is therefore described by the Hamiltonian $H_I=-\frac{p^2}{2m^*}+U(r)$ for states with configuration $\ket{2,\phi_{\vec{p}}}=\frac{1}{\sqrt{2}}[\ket{\uparrow\downarrow}+\ket{\downarrow\uparrow}]$, where $U(r)$ is the potential energy. The Hamiltonian $H_I$ describes the motion of a particle with \textit{negative} effective mass $-m^*$ in the external potential $U(r)$. We note that $-H_I$ describes the motion of a particle with a \textit{positive} mass $m^*$ in an attractive potential. In two dimensions, an arbitrarily weak attractive potential exhibits at least one bound state at negative energies for massive particles\cite{landau77}. It follows, therefore, that $H_I$ will exhibit positive energy bound states if $U(r)$ is repulsive. This property is a direct consequence of the negative definite kinetic energy of the pair, the dynamics of which is akin to the motion of a hole--like state near the top of the valence band in a semiconductor: the repulsive potential due to a negatively charged impurity is perceived as an attraction due to negative band curvature. In the real space picture, two electrons with opposite momenta reside in different cones and have nearly the same velocities. The repulsive force tends to increase the momentum of one electron, and decrease the momentum of the other. Due to the negative dispersion term, this decreases the velocity of the first particle, and increases the velocity of the other, reducing the distance between them. Unlike conventional bound states, these positive eigenstates are metastable. Formation of regular (electron--hole) excitons is prohibited in this regime.\\
\indent Metastable states similar to the ones described by $H_I$ were previously discovered in connection with inverse hydrogen absoption spectra \cite{gross71}. Simple models of bound state formation due to a negative single--particle energy dispersion near the top band boundary were considered in Refs.~[\onlinecite{mahajan06}] and~[\onlinecite{souza10}]. We note, however, that the origin of the negative dispersion in graphene is different: the leading term in the single particle energy is linear, and the Hamiltonian $H_I$ arises via the compensation between the two sub--bands. This means that the dynamics of these pairs can be represented as a slow relative motion $v\propto p/m^*$ equation superimposed with the fast motion of the pair, $v\propto v_F$.\\
\indent Since the effective mass ($-m^*$) is only about five times larger than the free electron mass, the binding can be quite strong. For the example of a repulsive Coulomb interaction, $U(r)=\frac{e^2}{\epsilon_s\epsilon_N r}$, the problem reduces to the two--dimensional hydrogen atom\cite{zaslow67}. (Here $\epsilon_N=1+\frac{N\pi e^2}{8\epsilon_s\hbar v_F}$ is the intrinsic dielectric constant\cite{hwang07} of graphene embedded in a material with dielectric constant $\epsilon_s$. For single (double) layer graphene, the number of fermion species is $N=4$ ($N=8$).) The highest energy level is given by
$E_1=\frac{2m^*e^4}{\epsilon_s^2\epsilon_4^2\hbar^2}$. We note that the hydrogen--like Hamiltonian $H_I$ results in binding energy $E_1\sim 1.5$eV and Bohr radius $a_B\sim 2.5\text{\AA}$ for $t'=0.1t$. At such short distances, the low--energy approximation to the graphene band structure is not valid \cite{geim07}, rendering the solution inconsistent. More importantly, the dynamics of particles at such high energies is affected by Pauli blocking due to the Dirac sea. For the bound state to be observable, the relevant phase space domain must be free from other particles. This can be achieved by, e.g., gating, if the bound state energy is well below $1$eV. Two--particle states with smaller binding energies can be realized in double--layered structures where the electrons in opposite layers are separated vertically by a dielectric spacer of thickness $d$. Hexagonal boron nitride ($\epsilon_{s}=3-4$) spacers have been experimentally shown to electrically isolate parallel graphene layers at a thickness of 4 atomic layers ($d=1.3$nm)\cite{britnell12}. This suppresses the $\frac{1}{r}$ singularity, yielding smaller binding energies. For a rough estimate of the binding energy we approximate the potential as $V(r)\sim e^2/\epsilon_s\epsilon_4^2\sqrt{r^2+d^2}$ (see Appendix A). The spectrum of the resulting shallow well can be found in the harmonic approximation, assuming $r\ll d$. The energy of the highest bound state is $E_0=-\hbar\omega+\frac{e^2}{\epsilon_s\epsilon_4^2d}$, where $\omega=\sqrt{\frac{e^2}{\epsilon_s\epsilon_4^2m^{*}d^3}}$ is the oscillation's angular frequency near the potential maximum. For example, $t'=0.1t$ gives binding energy $E_0=31$meV. For the case of direct pairs the trigonal warping is not compensated ($\tau_{1,2}\neq 0$), and the situation becomes more complicated. We note, however, that if the mass $m^*$ is small enough the trigonal warping terms cannot change the sign of the kinetic energy. Further, there are several momentum space orientations for which $\sin(3\phi_{\vec{p}})=0$. We therefore take $E_0$ as a first approximation of the binding energy for direct \textit{and} indirect pairs.\\
\indent To analyze the case of the direct pair with anisotropic dispersion, we derive its effective Hamiltonian in momentum space. The potential energy $V(r)$ is represented by a non--local operator proportional to its Fourier transform $\tilde{V}_{\vec{p},\pvec{p}'}=\tilde{V}(|\vec{p}-\pvec{p}'|)$, while the kinetic energy terms are given by Eq. (\ref{direct}) for $\tau_{1,2}=\pm 2$. Restricting the potential energy to the non--dispersing sector requires some care due to a non--trivial overlap between non--dispersing states with different momenta: $\braket{1,\phi_{\pvec{p}'}|1,\phi_{\vec{p}}}=\cos(\phi_{\vec{p}}-\phi_{\pvec{p}'})$. For the case of direct interactions, the $(\vec{p},\pvec{p}')$ block of the Hamiltonian matrix takes the form:
\begin{align}
\widehat{H}_{\vec{p},\pvec{p}'}=\delta_{\vec{p},\pvec{p}'}\widehat{H}_{1,2}^\text{eff}
+\tilde{V}_{\vec{p},\pvec{p}'}&\left[\begin{array}{cc}
\cos(\phi_{\vec{p}}-\phi_{\pvec{p}'}) & 0 \\
0 & 1\label{hamiltonian}
\end{array}\right],
\end{align}
\noindent where $\widehat{H}_{1,2}^\text{eff}$ is given by Eq. (\ref{direct}). In the absence of inter--particle interaction the eigenvalues of this matrix are given by the kinetic energy terms: $\epsilon_2^{(2)}=-2\mu p^2\left[\eta+\sin(3\phi_{\vec{p}})\right]$, where we have introduced the \textit{anisotropy parameter} $\eta=\frac{6t'}{t}$ which is not physically tunable (uncertainty in the value of $t'$ gives a range of possible values $0.12\leq\eta\leq 1.2$). Depending on the value of $\eta$, the kinetic energy is either negative--definite ($\eta>1$) or sign--indefinite ($\eta<1$). \\
\indent We proceed by numerically diagonalizing the Hamiltonian given by Eq. (\ref{hamiltonian}) using the interlayer interaction from Ref.~[\onlinecite{abergel2013}] (see Appendix A). Although it is assumed that the relevant phase space domain is free of other particles to avoid the effect of Pauli blocking, we will treat the case of screening at half--filling ($p_f=0$) as a first approximation. Indeed, the dielectric contribution to the screening giving rise to $\epsilon_N$ occurs at scales smaller than the Fermi wavelength ($\lambda_f$), and is most important in the realistic limit of $\lambda_f\ll d$. To visualize the resulting wavefunctions, we calculate the local density of states (LDOS), $\nu(\epsilon,x,y) =\sum_{n}\delta(\epsilon-\epsilon_{n})|\psi_{n}(x,y)|^2,$ where $x$ and $y$ are the components of the in--plane separation and $n$ labels the eigenstates. In the isotropic regime, $\eta>1$, as is evidenced by Fig. \ref{LDOS_fig}a for $\eta=1.1$, there is a formation of distinct, highly localized, bound states at $\epsilon_1=45$meV, $\epsilon_2=30$meV and $\epsilon_3=27$meV. We note that the energy of the highest bound state, $\epsilon_1$, is very similar to the value predicted in the harmonic approximation above. This validates the further use of such an approximation in the calculations of the transition rates that are to follow. At negative energies there is a low intensity continuum of unbound states, which are only weakly coupled to the bound states due to the symmetry of the Hamiltonian, resulting in their large lifetimes.\\
\begin{figure}[t]
\includegraphics[width=\linewidth]{figure_12.png}
\caption{Two--body LDOS in a graphene hybrid structure as a function of binding energy $E$ and interparticle distance $r$. The energy dependence shows distinct bound states in the (a) isotropic regime, which peel off into the free particle continuum in the (b) anisotropic regime (Coulomb potential in white). In real space, the wavefunctions are those of highly localized bound states at (c) $\epsilon_1=45$meV, and have higher anisotropy for (d) less bound states.}
\label{LDOS_fig}
\end{figure}
\indent In the anisotropic regime ($\eta< 1$), there are six \textit{easy axis} angles, defined by the relation $\sin(3\phi_{0})=-\eta$, along which the dispersion is effectively suppressed despite the broken conduction--valence symmetry. By concentrating the wavefunction along these axes, one constructs a state qualitatively similar to the non--dispersing solutions in which the interparticle distance takes a constant value $r_0$: $\psi(r)\propto\delta(r-r_0)$. The energies of these states (Fig. \ref{LDOS_fig}b) follow the profile of the interaction potential, $\epsilon\approx U(r_0)$. Further, a negative energy state dragged into the positive continuum by the interaction potential can decay by changing its pseudospin configuration rather than by tunnelling through a barrier. This is wholly due to the sign--indefinite kinetic energy.\\
\section{Semiclassical trajectories}
In order to understand the dynamics of the electron pair in real space, it is instructive to analyze their semiclassical trajectories. Treating the momenta as classical variables, the time evolution of the two--particle system is then governed by Hamilton's equations:
\begin{align}
\frac{d\vec{p}_{i}}{dt}&=-\nabla_{\vec{x}_i}H,\nonumber\\
\frac{d\vec{x}_i}{dt}&=\nabla_{\vec{p}_i}H,\label{hamiltons}
\end{align}
where the subscript $i=1,2$ denotes the particle number.\\
\indent We will restrict our discussion to the case of indirect pairs in the subspace of states with configuration $\ket{2,\phi_{\vec{p}}}=\frac{1}{\sqrt{2}}[\ket{\uparrow\downarrow}+\ket{\downarrow\uparrow}]$. In this case the kinetic energy is isotropic and there is a trivial overlap of states with different momenta. The kinetic energy can be taken as the relevant eigenvalue of Eq. (\ref{single}) for each particle. Thus, we write:
\begin{align}
H=-\frac{p_1^2+p_2^2}{4m^*}+v_F(p_1-p_2)+\frac{V_0}{\sqrt{|\vec{x}_1-\vec{x}_2|^2+d^2}},\label{semi}
\end{align}
where $p_i=|\vec{p}_i|$ is the magnitude of the momentum for particle $i$, $\vec{x}_i=(x_i,y_i)$ is its position and $V_0=\frac{e^2}{\epsilon_s\epsilon_N^2}$ determines the strength of the potential. Then, Eqns. (\ref{hamiltons}) take the form:
\begin{align}
\frac{d\vec{p}_{1,2}}{dt}&=\pm\frac{V_0\vec{r}_{12}}{[|\vec{r}_{12}|^2+d^2]^{\frac{3}{2}}}\nonumber,\\
\frac{d\vec{x}_{1,2}}{dt}&=-\frac{\vec{p}_{1,2}}{2m^*}\pm v_F\frac{\vec{p}_{1,2}}{p_{1,2}}\label{systemeqns},
\end{align}
where $\vec{r}_{12}=\vec{x}_1-\vec{x}_2$. We solved Eqns. (\ref{systemeqns}) numerically using a $4^{\text{th}}$--order Runge--Kutta procedure, and a typical result is given in Fig. (\ref{trajectories_fig}).\\
\begin{figure}[h]
\includegraphics[scale=0.35]{trajectories.eps}
\centering
\caption{(Color online) Typical trajectories of an electron pair with zero total momentum $\vec{p}_1=-\vec{p}_2$ in separated graphene layers (separation $d=1.3$nm) with a hBN dielectric spacer ($\epsilon_s=3.9$). The constituent particles are in different colors, and the ticks on the axes correspond to steps of $20$nm.}
\label{trajectories_fig}
\end{figure}
\indent For a pair with vanishing total momentum, $\vec{p}_1=-\vec{p}_2\equiv\vec{p}$, Eqns. (\ref{systemeqns}) define the band velocities of the individual particles, which are given by
\begin{align}
\vec{v}_{1,2}=v_F\hat{\vec{p}}\mp\frac{\vec{p}}{2m^*},
\end{align}
where $\hat{\vec{p}}$ is a unit vector in the direction of $\vec{p}$. It immediately follows that the trajectories of the electrons can be represented in terms of a superposition of two motions: a fast center--of--mass motion, characterized by velocity $\vec{v}_{cm}=\frac{1}{2}\left(\vec{v}_1+\vec{v}_2\right)=v_F\hat{\vec{p}}$, and a much slower relative motion, with velocity $\vec{v}_{r}=\vec{v}_2-\vec{v}_1=\vec{p}/m^*$. In the absence of interactions, and for small momenta, the relative velocity is sufficiently small that the pair behaves as if they were a single particle, moving with velocity $v_F$ in the direction of the momentum.\\
\indent If we now switch on the inter--particle interaction, the momenta of the particles become time--dependent, in accordance with Eqns. (\ref{systemeqns}). The sign of the interaction implies that whenever $\vec{p}_1$ is increasing, $\vec{p}_2$ must be decreasing, and vice versa. Due to the sign of the parabolic energy term, this will cause the particle with the smaller of the two velocities to speed up slightly, and the other particle to slow down, closing the distance between the two particles. Therefore, the finite relative velocity implies that a repulsive force between the two particles will increase the velocity of one particle while decreasing the velocity of the other, causing them each to change direction slightly in such a way that their seperation is almost constant. This causes the two particles to behave as if they were ``stuck together" despite the repulsive force. Note, however, that the momentum $\vec{p}$ is changed by the interaction. Hence, as the particles are orbiting around their mass center, the direction of $\vec{v}_{cm}$ is also changing, so that the average velocity of the pair over a long time is zero, as can be seen from Fig. (\ref{trajectories_fig}). The large value of $v_F$, compared to $\vec{v}_r$, results in non--propagating orbits of relatively large diameter $d\sim v_F/\omega$ where $\omega$ is the angular frequency of the orbit, which can be estimated as $\omega=E_0/\hbar$ ($d\sim 100$nm in Fig. (\ref{trajectories_fig})).\\
\section{Decay into the dispersing sector}
\indent So far we have considered only the non--dispersing sector. Coupling to the dispersing sector could lead to the decay of the metastable states found above, however these transitions are suppressed by momentum mismatch between the sectors. It is instructive to analyze the decay via Fermi's golden rule. Due to energy conservation the decay is only allowed into states with positive energy $E=2v_fp$, i.e. $\ket{3,\phi_{\vec{p}}}$ from Eq. (\ref{wf1}). The coupling between these two sectors occurs via trigonal--warping and potential energy terms, due to the non--trivial overlap between states with different momenta. The relevant matrix elements of the Hamiltonian are
$H_{1,3}=\frac{i}{\sqrt{2}}\tilde{V}_{\vec{p},\vec{p}'}\sin(\phi_{\vec{p}}-\phi_{\pvec{p}'})$ and $H_{2,3}=i\sqrt{2}\mu p^2\sin(3\phi_{\vec{p}})\delta_{\vec{p},\vec{p}'}$, where $H_{i,j}=\bra{i,\phi_{\pvec{p}'}}\widehat{H}\ket{j,\phi_{\vec{p}}}$.
\noindent The matrix element due to the interaction potential, $H_{1,3}$, vanishes by symmetry if $\psi_i(\vec{p})$ is an $s$--state. The kinetic energy term, $H_{2,3}$, conserves the momentum. Therefore the decay occurs when the initial ($p_i$) and final ($p_f$) momenta satisfy conservation laws: $E_0=2v_Fp_{i,f}$. This gives $p_i\ll p_0$, where $p_0=\sqrt{2\hbar m^{*}\omega}$ is the zero point momentum in the initial state. The smallness of $p_i$ results in a small matrix element which is proportional to $p^2$. The decay rate is given by Fermi's golden rule, $\Gamma=\frac{2\pi}{\hbar}|M_{if}|^2\rho_f$, where $M_{if}=\bra{\psi_f}H\ket{\psi_i}$ is the transition matrix element and $\rho_f=\frac{|E|}{8\pi\hbar^2v_F^2}$ is the density of final states\cite{wallace47} in the absence of spin--flipping and inter--valley scattering for particles with velocity $2v_F$. To calculate the kinetic contribution, we approximate the final state wavefunctions by plane waves, $\psi_f\sim\delta(\pvec{p}'-\vec{p}_f)$. The exact initial wavefunction for a direct pair state is not known, but for an order of magnitude estimate we shall employ the harmonic approximation as discussed above: $\psi_i(\vec{p})\sim \frac{1}{p_0}\exp(-\frac{p^2}{p_0^2})$. The transition rate due to $H_{2,3}$ is therefore $\Gamma=\frac{\mu^2p_f^4E_0}{\pi\hbar v_f^2p_0^2}\sim 10^{-10}\frac{E_0}{\hbar}$. This suggests only weak coupling to the continuum, justifying our consideration of the non--dispersing sector independently. Note that $H_{2,3}$ vanishes for indirect pairs, which do not decay by this mechanism.\\
\indent These small values imply that the lifetime of the pair is likely to be limited by other, non--universal mechanisms, such as impurity or electron--electron scattering. Also, one has to bear in mind that higher--order virtual transitions could lift the restriction $p_f=p_i\ll p_0$. The detailed analysis of this strongly depends on the properties of $\tilde{V}_{\vec{p},\pvec{p}'}$.\\
\section{Summary and conclusions}
\indent In this paper we have analyzed the problem of an isolated electron--electron bound state, however further work remains to be done in considering the many--body effects, the simplest example of which is interaction screening. Recall that in order to prevent Pauli blocking the system must be gated so that the Fermi energy exceeds the binding energy. This would introduce a non--zero density of states and metallic screening of the interaction. If the inter--particle separation is larger than the screening radius, the binding energy would be renormalized, but the bound state would not be destroyed: for a massive particle in an arbitrarily weak potential, at least one bound state exists in two dimensions\cite{landau77}. The states we considered here have positive energy and therefore do not represent energy minima, however once created they have long lifetimes. One way to create them is by coupling the aforementioned graphene structure to a superconductor. As the metastable states are akin to Cooper pairs, this would lead to a giant enhancement of the proximity effect, which has been observed in graphene recently \cite{ojeda09,du08,komatsu12}.\\
\indent In conclusion, we have studied the problem of interacting electron--electron pairs in hybrid graphene--dielectric--graphene structures. We have shown that, in the isotropic regime, the conduction--valence band asymmetry allows the formation of a new kind of Cooper--pair--like bound state in the sector spanned by eigenfunctions which are dispersionless in the conical approximation.\\
\section{Acknowledgements}
\indent The authors wish to thank V.I. Fal'ko and M.V. Berry for insightful discussions. A.V.S. is supported by EPSRC/HEFCE Grant No. EP/G036101/1.
|
\section*{Kitagawa: ``Instrumental Variables Before
and Later''}
I am grateful for the kind words by Kitagawa. He has been doing very
interesting work on
testing for validity of instrumental variables in recent years (e.g., Kitagawa, \citeyear{Kit10}, \citeyear{Kit13}) that will undoubtedly be influential in
the literature.
I am also glad that Kitagawa likes my summary of the differences
between econometric and statistical approaches to causality as ``choice
versus chance.''
Kitagawa's comments on the impact of the local average treatment effect
literature on economic practice agree with my views. As emphasized in
the paper, the LATE concept was never intended to change the question
of interest, but to clarify what we could learn from the data.
Nevertheless, in some cases the LATE may well be representative of a
subpopulation that is of substantial interest on its own.
Consider the draft lottery example (Angrist, 1990; Hearst, Newman and
Hully, 1986) where the compliers are the men who served, or would have
served, in the military, because of their draft lottery number.
Arguably, this is the group on the margin for whom the effect of
military service is most interesting.
Similarly, in the Angrist and Krueger (1991) study of the returns to
education using compulsory schooling laws as an instrument, the
compliers are the individuals for whom schooling decisions are affected
by compulsory schooling laws, again arguably an interesting
subpopulation for educational policies that are often targeted at those
receiving lower levels of education. Nonetheless, in general the
subpopulation of compliers is not chosen for its interest, but because
we can hope to learn something about them. It is about the primacy of
internal validity over external validity (Shadish, Cook and Campbell, \citeyear{ShaCamCoo02}).
Kitagawa discusses instrumental variables in the context of another
example that, like the supply-and-demand example I discuss in the
paper, is a classic one, that of the estimation of returns on inputs in
a production function. Specifically, he focuses on the causal effect of
labor inputs on output. The starting point for an economist is exactly
as Kitagawa describes: firms do not choose input levels randomly, but
choose them optimally, for example, to maximize profits.
This leads quickly to settings where we cannot simply regress output on
inputs if we are interested in the causal effect of input on output.
Moreover, the context in combination with economic theory on firm
behavior suggests where a researcher might look for instruments that
satisfy the exclusion restriction, namely cost variables that affect
the choice of input levels but that affect output only through their
effect on input levels.
In his comments, Kitagawa also distinguishes between various objectives
for the researcher. If the goal of the researcher is what he calls
``scientific reporting,'' Kitagawa agrees with my recommendation to
report both estimates of the local average treatment effect and bounds
on the overall average treatment effect. If, on the other hand, the
goal is directly to make a decision, say, on whether to extend the
treatment to the entire population or not, he advocates a decision
theoretic approach, either Bayesian along the lines of Chamberlain
(\citeyear{C11}), or the type of Manski ``data-alone'' frequentist approach. I
agree with that, and I think the distinction between scientific
reporting and decision making is a useful one to bear in mind.
\section*{Richardson and Robins: ``ACE Bounds; SEMs with
Equilibrium Conditions''}
Richardson and Robins make two sets of comments, one about bounds on
the average causal effect (ACE), and one about simultaneous equations
models (SEMs).
In the discussion on bounds, they formulate four sets of assumptions,
captured by different graphical models that allow for construction of
the same set of bounds. They relate these assumptions to their novel
Single World Intervention Graphs (SWIGs).
I find the SWIGs an intriguing approach, and one that might help make
the graphical approach more relevant for researchers interested in
causal effects.
One concern I have with the discussion of the four sets of assumptions
is that it is not clear when there is a substantively important
difference between the assumptions. For example, I find it difficult to
think of substantive applications where the independencies hold one
pair at a time [Assumption (iii)], but not joint independence
[Assumption (i)].
The discussion on market equilibrium and bicausal models is very
interesting and stimulating.
I am happy to see Richardson and Robins endorse my interpretation of
structural equations in terms of potential outcomes. Although, as the
authors point out, this interpretation of structural equations is not
universal, in my view, partly based on conversations with other
economists, it is the leading one in economics. The discussions of
normalization issues that the authors refer to generally arise in the
context of estimation in settings where there are multiple instruments.
In that case, the difference between estimation methods such as limited
information maximum likelihood (LIML, going back to Anderson and Rubin,
1948), and two stage least squares (TSLS) matter. In the recent
literature on weak instruments, these differences have been shown to
potentially matter a great deal. Staiger and Stock (1997) is a key
paper, and Stock and Andrews (2005) provide an overview.
Although economists routinely use the supply-and-demand example in
textbooks and teaching, most discussions no longer explicitly discuss
where the equilibrium that is assumed arises from, making the work in
this area more difficult to access for researchers from other areas
than it need to be. The model used by Richardson (\citeyear{R96}) where the data
come from a discrete approximation to a finer recursive model appears
to capture well the mechanisms researchers implicitly have in mind. See
\citet{Ber66} for a related discussion in the older economics
literature discussing the relationship between nonrecursive (bicausal)
models in discrete time and recursive continuous time models.
\section*{Shpitser:
``Causal Graphs: Addressing the
Confounding Problem Without
Instruments or Ignorability''}
Shpitser is concerned that I did not discuss the growing literature on
causal graphical models.
This is a very interesting and rapidly expanding literature that has
important antecedents (Wright, \citeyear{W21}) that were\linebreak[4] influential in the
economics literature, and where Richardson and Sphitser have made major
contributions.
However, I saw the focus of my paper on an econometrics perspective on
instrumental variables, and there graphical models do not currently
play a major role.
It is an interesting question why economists have not felt that
graphical models have much to offer them.
\citet{Pea13} has also raised this question, and concludes somewhat
dismissively that:
``economists
are still scared of graphs.'' He sees this as an ``educational
deficiency,'' and writes that
``This
educational impairment is the main factor that prevents economists from
appreciating
much of the recent progress in causal inference'' (Pearl, \citeyear{Pea13}, page 8).
My view on the lack of use in the econometrics literature on the
graphical models is more sanguine. I~see substantial evidence that as a
group economists are willing to adopt new methods from other
disciplines that are viewed as useful in practice. There are many
examples of this even within the area of causal inference. The rapid
adoption of the Rubin potential outcome approach starting in the early
1990s with Heckman (1990) and Manski (1990) is one, as is the by now
widespread use of matching and propensity score methods, and the
current boom in studies using methods associated with regression
discontinuity designs that were originally developed in the psychology
literature (see Cook, 2008, for a historical overview). In contrast,
the causal graphs have not caught on in economics. In my view a major
reason is that there have been few compelling applications of causal
graphs to social science questions where the causal-graph approach has
generated novel analyses or prevented researchers from making mistakes
that other frameworks might have encouraged them to make.
A second reason may be that some assumptions are not easy to
incorporate in the graphical approach. Monotonicity, which Swanson and
Hern\'an are particularly concerned with in their comments, and which
plays a key role in instrumental variables analyses, is difficult to
capture in a causal graph. See the discussion in Imbens and Rubin (\citeyear{ImbRub95}).
Let me flesh out the first part of this argument.
There are thousands of empirical studies in economics where researchers
use instrumental variables methods. Implicitly, they may have a causal
graph like Figure~1 in the main paper, or Figure 1(c) in the Shpitser comment,
in mind. Often there is considerable discussion in a particular
application whether the two key assumptions that there is no direct
effect of $Z_i$ on $Y_i^{\mathrm{obs}}$ (no arrow from $Z_i$ to $Y_i$,
and no confounding of the effect of $Z_i$ on $Y^{\mathrm{obs}}_i$ (no
unobserved common cause of $Z_i$ and $Y_i$) are plausible. In
observational studies in social science, both these assumptions tend to
be controversial. In this relatively simple setting, I do not see the
causal graphs as adding much to either the under\-standing of the
problem, or to the analyses.
Similarly, there are thousands of empirical studies in economics where
researchers use matching type methods based on the assumption of no
unmeasured confounders, and where implicitly they may have a causal
graph like Figure~1(b) in mind. Again, the assumptions underlying such
a graph are typically controversial and researchers often put in
substantial effort in arguing for the absence of unobserved
confounders. In this case, again I fail to see what using a
causal-graph approach would add in practice.
Now consider a more complicated setting such as the
``hypothetical longitudinal study represented by the causal graph
shown in Figure~2,'' in the comment by Shpitser, or Figure~1 in Pearl
(1995). Here, identification questions are substantially more complex,
and there is a strong case that the graph-based analyses have more to
contribute. However, I am concerned about the relevance of such
examples in social science settings. I~would like to see more
substantive, rather than hypothetical, applications where a graph such
as that in Figure~2 could be argued to capture the causal structure.
There are a large number of assumptions coded into such graphs, and
given the difficulty in practice to argue for the absences of one or
two arrows in instrumental-variables or no-unobserved-confounders
applications in social sciences, I worry that in practice it is
difficult to convince readers that such a causal graph fully captures
all important dependencies.
In other words, in social sciences applications a graph with many
excluded links may not be an attractive way of modeling dependence
structures. As Andrew Gelman writes on his blog in a discussion of
graphical models and potential outcomes, ``Nothing is zero, everything
matters to some extent'' (\cite{Gel}). Of course, instrumental
variables methods do also critically rely on the absence of particular
dependencies, but my point is that the larger graphical models such as
those in Figure~2 of the Shpitser comment or Figure~1 in Pearl (1995)
with many variables and many excluded links require researchers to
evaluate critically many more of those assumptions. The causal graph
methods appear to be more suited to answering the question whether
given a complex set of conditional independencies particular causal
effects are identified, whereas in my experience in many social science
applications researchers proceed by assessing a few conditional
independencies given which it is known particular effects are identified.
\section*{Swanson and Hern\'an:
``Think Globally, Act Globally: An
Epidemiologist'S Perspective on
Instrumental Variable Estimation''}
First of all, I want to commend Swanson and Hern\'an for their work on
improving the reporting the results of instrumental variables analyses
(Swanson and Hern\'an, \citeyear{S13}). Although many of their recommendations
such as the reporting of estimates of the proportion of compliers are
routinely followed in the economics literature (these estimates are
there often referred to as the first stage coefficients in the
two-stage-least-squares terminology), these practices had not made it
to the epidemiology literature, and their work will likely improve
practice there.
I am also glad to see that they do not attempt to defend the
homogeneity assumptions that would allow for point identification of
the ATE: it appears that there is growing consensus that such
assumptions are not realistic.
There are other areas where there is less agreement.
Swanson and Hern\'an take issue with the focus in the paper on the
local average treatment effect (LATE). Whereas Kitagawa felt LATEs were
``valuable pieces of information about causal effects'' (Kitagawa, page
359), Swanson and Hern\'an take the view that ``the LATE is not generally
relevant to epidemiological questions'' and propose to ``refocus on the
global ATE in the population of interest'' (Swanson and Hern\'an, page 371).
In my response to Swanson and Hern\'an, I want to make three points.
First, I want to correct the record concerning my position on
presenting estimates based on IV assumptions. Swanson and Hern\'an
summarize my position in terms of ``two options \ldots (1) present bounds
for the ATE, \ldots, or (2) present point estimates'' (pages 372--373) and then add
that ``of course \ldots we can always do both.'' Swanson and Hern\'an
appear to have missed that presenting both
the bounds and the point estimate for the LATE (which is the same as
the point estimate for the ATE under homogeneity) was what I in fact
proposed (see also the comments by Kitagawa). One concern with the sole
focus on the ATE that Swanson and Hern\'an appear to favor, either
directly, or in combination with tighter bounds on outcomes, is that
one may discard relevant information.\vadjust{\goodbreak} Let me expand on comments in the
main paper in this regard. Consider the following two versions of an
artificial example with a dichotomous instrument, treatment and
outcome. Let $p_{zxy}$ be the population fraction of units with
$Z_i=z$, $X_i=x$, and $Y_i=y$, for $z,x,y\in\{0,1\}$. In the first
example, suppose $p^1_{000}=1/4$, $p^1_{001}=1/12$, $p^1_{010}=0$,
$p^1_{011}=0$, $p^1_{100}=1/24$, $p^1_{101}=7/24$, $p^1_{110}=7/24$,
$p^1_{111}=1/24$, and suppose these fractions are estimated precisely.
In this case, the fractions of compliers, nevertakers and alwaystakers
are 1$/$2, 1$/$2 and 0, the bounds on the ATE are $[-3/16,5/16]$, and the
point estimate of the LATE is $-$1$/$4. In the second example,
$p^2_{000}=1/6$, $p^2_{001}=1/6$, $p^2_{010}=0$, $p^2_{011}=0$,
$p^2_{100}=1/8$, $p^2_{101}=5/24$, $p^2_{110}=1/8$, $p^2_{111}=5/24$.
In this case,
the fractions of compliers, nevertakers and alwaystakers are again 1$/$2,
1$/$2 and 0, the bounds on the ATE are the same, $[-3/16,5/16]$, and the
point estimate of the LATE is now positive 1$/$4. Under the instrumental
variables assumptions, the bounds for the ATE are identical in the two
examples, but the LATEs are very different. In the first case, there is
evidence of a substantial negative effect for a subpopulation, whereas
in the second example one knows there is a subpopulation for which the
effect is substantial and positive. That would appear to potentially
lead to very different substantive conclusions. Simply reporting bounds
would miss these results.
In the second part of my response to Swanson and Hern\'an, I will
discuss more explicitly the concerns about external validity that are
implicit in the discussions of the relative merits of the overall
average effect (ATE) and the LATE.
Swanson and Hern\'an are interested in the ATE in the population of
interest, and then without explicitly saying so, assume that the study
population is representative for this population of interest.
Matters are rarely so clear cut in practice. The study sample need not
be a random sample from the population of interest because of
nonresponse, or the policy maker may be interested in the average
effect if the treatment were to be extended to a larger population at a
future date, or were to be offered on a voluntary basis to the general
population. What the population of future volunteers looks like may
well depend on the efficacy of the treatment according to the
statistical analysis. There are many examples where even in randomized
experiments the causal effects found for the study population did not
generalize to the population subsequently subject to the treatment.
Once one recognizes that even the study population may differ from the
population of interest much of the concern with the LATE that Swanson
and Hern\'an raise loses its force.
My position here is again essentially similar to the \citet{ShaCamCoo02} view on the primacy of internal validity over external validity.
In the third part of my response, I will make some comments on the
monotonicity assumption.
Swanson and Hern\'an present a generic example where the monotonicity
condition is likely to be violated, and argue that the instrument in
this example is one of the ``most commonly proposed instruments in
epidemiology.'' In fact, the example demonstrates how much there is to
be gained from a closer study of the earlier econometric literature, as it was discussed in the original paper on the LATE
(Example 2, page 472, Imbens and Angrist, 1994); see also Section 5.3 in the current paper.
The generic example is as follows. The assignment of individuals to the
treatment is partly based on preferences of an administrator (physician
in the epidemiological version of the experiment). The assignment of
administrators to individuals is as good as random. Different
administrators may have different preferences on average, but it need
not be the case that the resulting instrument is monotone because the
set of individuals who would be assigned to the treatment by one
administrator need not be a proper subset of the set of individuals who
would be assigned to the treatment by a second administrator. That
setting also arises in applications of instrumental variables in legal
settings where the administrator may be a randomly assigned judge: see Aizer and Doyle (2013) with an application in the criminal justice system, and
Dobbie and Song (2013) with an application to bankruptcy proceedings. It is important to distinguish such
settings from those where the instrument corresponds to an increase in
the incentive to participate, in which case the monotonicity assumption
is plausible. It is precisely by articulating explicitly these
assumptions and describing the role they play that we may be able to
avoid misleading decision-making efforts.
\renewcommand{\bibname}{Additional References}
|
\subsection{Preliminaries}
\smallskip\noindent\emph{ Parity games.}
A \emph{parity game} $P = (G, \alpha)$ consists of a \emph{game graph}
$G = ((V, E),(V_\po, V_\pe))$ and a \emph{parity function} $\alpha: V \rightarrow
\mathbb{Z}$ that assigns an integer value to each vertex. We denote the two
players with ${\mathcal{O}}$ (for odd) and ${\mathcal{E}}$ (for even).
Player~${\mathcal{O}}$ (resp.\ player~${\mathcal{E}}$) wins a play if the \emph{lowest} priority occurring in the play
infinitely often is odd (resp.\ even).
We say that the vertices in $V_\po$ are \emph{${\mathcal{O}}{}$}-vertices
and the vertices in $V_\pe$ are \emph{${\mathcal{E}}{}$}-vertices. We use ${p}$ to denote
one of the players $\set{{\mathcal{O}},{\mathcal{E}}}$ and ${\bar{p}}$ to denote his opponent.
We will specifically consider \emph{parity-3 games} with $\alpha: V
\rightarrow \set{-1,0,1}$ and \emph{B\"uchi games} with $\alpha: V \rightarrow \set{0,1}$,
where the vertices in the set $B = \set{v \mid \alpha(v) = 0}$ are called \emph{B\"uchi vertices}.
B\"uchi games are denoted as $(G, B)$.
\smallskip\noindent\emph{ One-pair Streett and parity-3 games.}
A one-pair Streett objective with pair $(L_1, U_1)$ is equivalent to a parity
game with three priorities.
Let the vertices in $U_1$ have priority~$-1$, let the vertices
in $L_1 \setminus U_1$ have priority~$0$ and let the remaining vertices
have priority~$1$. Then player~1 wins the game with the one-pair Streett
objective if and only if player~${\mathcal{O}}$ wins the parity-3 game.
As the known algorithms for parity-3 games are special cases of algorithms for
general parity games, we will use the notion of parity games (i.e.,
player~${\mathcal{O}}$ and player~${\mathcal{E}}$ instead of player~1 and player~2).
\smallskip\noindent\emph{ Plays.}
For technical convenience we consider that every vertex in the game graph
$G$ has at least one outgoing edge.
A game is initialized by placing a token on a vertex. Then the two
players form an infinite path called \emph{play} in the game graph by
moving the token along the edges. Whenever the token is on a vertex in~$V_{p}$,
player~${p}$ moves the token along one of the outgoing edges of the vertex.
Formally, a \emph{play} is an infinite sequence $\langle v_0,v_1,v_2,\ldots\rangle$ of vertices such that
$(v_j,v_{j+1}) \in E$ for all $j \geq 0$.
For a vertex $u\in V$, we write $\Out(u)=\set{v\in V \mid (u,v) \in E}$
for the set of successor vertices of~$u$ and $\In(u)=\set{v \in V \mid (v,u)
\in E}$ for the set of predecessor vertices of $u$.
We denote by $\OutDeg(u)=|\Out(u)|$ the number of outgoing edges from $u$,
and by $\InDeg(u)=|\In(u)|$ the number of incoming edges.
\smallskip\noindent\emph{ Strategies.}
A \emph{strategy} of a player~${p} \in \set{{\mathcal{O}},{\mathcal{E}}}$ is a
function that, given a finite prefix of a play ending at $v \in V_{p}$,
selects a vertex from $\Out(v)$ to extend the play.
\emph{Memoryless strategies} do not depend on the history of a play but only on the current vertex. That is, a memoryless
strategy of player~${p}$ is a function $\tau: V_{p} \rightarrow V$ such that
$\tau(v) \in \Out(v)$. It is well-known that for parity games it is sufficient
to consider memoryless strategies (see Theorem~\ref{thrm:determinacy} below).
Therefore we will only consider memoryless strategies from now on.
A start vertex $v$ together with a strategy~$\sigma$ for ${\mathcal{E}}{}$ and a
strategy~$\pi$ for ${\mathcal{O}}{}$ defines a unique play $\omega(v,\sigma,
\pi)=\langle v_0,v_1,v_2,\ldots\rangle$,
which is defined as follows:
$v_0=v$ and for all $j \geq 0$, if $v_j \in V_\pe$, then $\sigma(v_j)=v_{j+1}$,
and if $v_j \in V_\po$, then $\pi(v_j)=v_{j+1}$.
\smallskip\noindent\emph{ Winning strategies and sets.}
A strategy $\tau$ is winning for player~${p}$ at start vertex $v$ if the
resulting play is winning for player~${p}$ irrespective of the strategy of
player~${\bar{p}}$. A vertex $v$ belongs to
the \emph{winning set} $W_{p}$ of player~${p}$ if player~${p}$ has a winning
strategy from $v$. By the following theorem every vertex is winning for exactly one of the two players.
When required for explicit reference of a specific game graph $G$ or
specific parity game $P$ we use $W_{p}(G)$ and $W_{p}(P)$ to refer to the
winning sets.
\begin{theorem}[\cite{EJ91,McNaughton93}]\label{thrm:determinacy}
For every parity game the vertices $V$ can be partitioned into the winning
set~$W_\pe$ of~${\mathcal{E}}{}$ and the winning set~$W_\po$ of~${\mathcal{O}}{}$.
There exists a memoryless winning strategy for~${\mathcal{E}}{}$ (resp.\ ${\mathcal{O}}$)
for all vertices in~$W_\pe$ (resp.\ $W_\po$).
\end{theorem}
The algorithmic question for parity games is to compute the set~$W_\pe$. We will
use the following algorithm for B\"uchi games as a subroutine in our algorithm.
\begin{theorem}[\cite{CH12,CH14}]\label{thrm:buchialg}
Let $(G, B)$ be a B\"uchi game with game graph $G$ and B\"uchi vertices $B$.
There is an algorithm $\buchialg(G, B)$ that computes~$W_\pe$ in time~$O(n^2)$.
\end{theorem}
For the analysis of our algorithm we further introduce the notions of \emph{closed sets}, \emph{attractors}, and \emph{dominions}.
\smallskip\noindent\emph{ Closed sets and attractors.}
A set $U \subseteq V$ is \emph{${p}$-closed} if for all ${p}$-vertices
$u$ in $U$ we have $\Out(u) \subseteq U$ and for all ${\bar{p}}$-vertices
$v$ in $U$ there exists a vertex $w \in \Out(v) \cap U$.
Note that player~${\bar{p}}$ can ensure that a play that currently ends in a ${p}$-closed set never leaves the ${p}$-closed set against any strategy of player~${p}$ by choosing an edge $(v,w)$ with $w \in \Out(v) \cap U$ whenever the current
vertex~$v$ is in $U \cap V_{\bar{p}}$ (see also~\cite[Proposition~2.2]{CH14}).
Given a game graph $G$ and a $p$-closed set $X$, we will denote by $G[X]$
the game graph induced by the set $X$ of vertices.
In a game graph $G$ a ${p}$-\emph{attractor} $\mathit{Attr}_{p}(U,G)$ of a set
$U \subseteq V$ is the set of vertices from which player~${p}$ has a strategy
to reach $U$ against all strategies of player~${\bar{p}}$. We have that $U \subseteq \mathit{Attr}_{p}(U,G)$.
A ${p}$-attractor can be constructed inductively as follows: Let $R_0=U$; and for all $i \ge 0$ let
\[
R_{i+1}= R_i \cup \set{v \in V_{p} \mid \Out(v) \cap R_i \neq \emptyset}
\cup \set{v \in V_{\bar{p}} \mid \Out(v) \subseteq R_i}. \quad (\ddag)
\]
Then $\mathit{Attr}_{p}(U,G)= \bigcup_{i\ge 0} R_i$.
The lemma below summarizes some well-known facts about closed sets, attractors, and winning sets.
\begin{lemma}\label{lemm:attr}
Let ${p} \in \set{{\mathcal{O}},{\mathcal{E}}}$.
The following assertions hold for parity games.
\begin{compactenum}
\item The set $V \setminus \mathit{Attr}_{p}(U,G)$ is ${p}$-closed in $G$~\cite[Lemma~4]{Zielonka98}.\label{sublemm:complattr}
\item Let $U \subseteq V$ be ${p}$-closed. Then $\mathit{Attr}_{\bar{p}}(U,G)$ is ${p}$-closed~\cite[Lemma~5]{Zielonka98}.\label{sublemm:attrclosed}
\item The attractor $\mathit{Attr}_{p}(U,G)$ can be computed in
$O(\sum_{v \in \mathit{Attr}_{p}(U,G)} \lvert\In(v)\rvert)$
time~\cite{Beeri80,Immerman81}.\label{sublemm:attrtime}
\item Let $U \subseteq W_{p}(G)$ and let $A = \mathit{Attr}_{p}(U,G)$. Then $W_{p}(G) = A \cup W_{p}(G[V \setminus A])$ and $W_{\bar{p}}(G) = W_{\bar{p}}(G[V \setminus A])$~\cite[Lemma~4.5]{JPZ08}.\label{sublemm:subgraph}
\end{compactenum}
\end{lemma}
\smallskip\noindent\emph{ Dominions.}
A set of vertices $D \subseteq V$ is a ${p}$-\emph{dominion} if $D \ne
\emptyset$, player~${p}$ has a winning strategy from every vertex in $D$ that
also ensures only vertices in $D$ are visited, and $D$ is a ${\bar{p}}$-closed set.
We will only consider ${\mathcal{E}}$-dominions in this paper and therefore usually omit
the reference to the player. Dominions of size $\lvert D
\rvert \le h$ can be computed by running the small-progress measure algorithm
of Jurdzi\'nski~\cite{Jurdzinski00} with a reduced codomain~\cite{Schewe07}.
A description of the small-progress measure algorithm for B\"uchi games is given
in~\cite[Section~2.4.1]{CH14}. We will use the following algorithm as a
subroutine.
\begin{theorem}[\cite{Jurdzinski00,Schewe07,CH14}]\label{thrm:progress}
Let $(G, B)$ be a B\"uchi game with game graph $G$ and B\"uchi vertices $B$.
There is an algorithm $\progress(G, B, h)$ that
returns the set of all
dominions of size at most $h$ in time $O(mh)$.
\end{theorem}
\section{Introduction}
\smallskip\noindent{\em Game graphs and graphs.}
Consider a directed graph $(V,E)$ with a partition $(V_1, V_2)$ of $V$, which is
called a {\em game graph}. Let $n = |V|$ and $m = |E|$.
Two players play the following {\em alternating game} on the graph
that forms an infinite path. They start by placing a token on an initial
vertex and then take turns indefinitely in moving the token: At a vertex
$v \in V_1$ player 1 moves the token along one of the outedges of $v$, at a
vertex $u \in V_2$ player 2 moves the token along one of the outedges of $u$.
If $V_2=\emptyset$, then we simply have a standard graph.
\smallskip\noindent{\em Objectives and winning sets.} Objectives are subsets of infinite paths
that specify the desired set of paths for player~1, and the objective for player~2 is the
complement of player-1 objective (i.e., we consider \emph{zero-sum} games). Given an objective $\Phi$,
an infinite path {\em satisfies the objective} if it belongs to $\Phi$.
Given a starting vertex $x \in V$ and an objective $\Phi$, if player~1 can guarantee that
the infinite path starting at $x$ satisfies $\Phi$, {\em no matter what choices player~2 makes,}
then player~1 can {\em win from $x$} and $x$ belongs to the {\em winning set of player~1},
and since the winning sets partition the game graph~\cite{Mar75}, the complement of the winning
set for player~1 is the winning set for player~2.
In case the game graph is a standard graph (i.e., $V_2=\emptyset$), the \emph{winning set}
consists of those vertices $x$ such that there exists an infinite path starting at $x$
that satisfies $\Phi$.
The winning set computation
for game graphs is more involved than for standard graphs due to
the presence of the adversarial player~2.
\smallskip\noindent{\em Relevant objectives.}
The most basic objective is \emph{reachability} where, given a set $U \subseteq V$
of vertices, an infinite path satisfies the objective if the path visits a vertex
in $U$ {\em at least once}.
The next interesting objective is the \emph{B\"uchi} objective that requires an
infinite path to visit a vertex in $U$ \emph{infinitely often}.
The next and a very central objective in formal verification and automata theory is the
\emph{one-pair Streett objective} that consists of a pair $(L_1,U_1)$ of sets of vertices
(i.e., $L_1 \subseteq V$ and $U_1 \subseteq V$), and an infinite path satisfies the objective
iff the following condition holds:
if some vertex in $L_1$ is visited infinitely often, then some vertex in $U_1$ is visited
infinitely often (intuitively the objective specifies that if one B\"uchi objective holds,
then another B\"uchi objective must also hold).
A generalization of one-pair Streett objectives is the {\em $k$-pair Streett objective}
(aka {\em general Streett objective})
that consists of $k$-Streett pairs $(L_1,U_1), (L_2,U_2), \ldots, (L_k,U_k)$, and
an infinite path satisfies the objective iff the condition for every Streett pair is
satisfied (in other words the objective is the conjunction of $k$ one-pair Streett objectives).
We study
(1)~game graphs with one-pair Streett objectives and
(2)~graphs with general Streett objectives.
\smallskip\noindent{\em Significance in verification.}
Two-player games on graphs
are useful in many problems in computer science,
specially in verification and synthesis of systems such as the synthesis of
systems from specifications and synthesis of reactive
systems~\cite{Church62,PnueliRosner89,RamadgeWonham87},
verification of open systems~\cite{AHK02}, checking interface
compatibility~\cite{InterfaceAutomata}, well-formedness of
specifications~\cite{Dill89book}, and many others.
General and one-pair Streett objectives are central in verification as most
commonly used specifications can be expressed as Streett
automata~\cite{Safra88,Thomas97}.
\emph{Game graphs with one-pair Streett objectives} arise in many applications in
verification. We sketch two of them.
(A)~Timed automaton games are a model for real-time systems.
The analysis of such games
with reachability objectives and safety objectives (which are the dual of reachability objectives) reduces to
game graphs with one-pair Streett objectives~\cite{dAFHMS03,dAF07,CHP11,CP13}.
(B)~The synthesis of Generalized Reactivity(1) (aka GR(1)) specifications exactly
require the solution of game graphs with one-pair Streett objectives;
GR(1) specifications are standard for hardware synthesis~\cite{PPS06} and even used in
synthesis of industrial protocols~\cite{GCH13,BJPPS12}\footnote{A GR(1) specification
expresses that if a conjunction of B\"uchi objectives holds, then another
conjunction of B\"uchi objectives must also hold, and since conjunction of
B\"uchi objectives can be reduced in linear time to a single B\"uchi
objective, a GR(1) specification reduces to implication between two
B\"uchi objectives, which is an one-pair Streett objective.}.
{\em General Streett objectives in standard graphs} arise, for example, in the
verification of closed systems with strong fairness
conditions~\cite{LH2000,DPC2009,Fra86}.
In program verification,
a scheduler is {\em strongly fair} if every event that is enabled infinitely often
is scheduled infinitely often. Thus,
verification of systems with strong fairness conditions directly corresponds to checking
the non-emptiness of Streett automata, which in turn corresponds to determining the winning
set in standard graphs with Streett objectives. Note, however, that
a Streett objective can either specify desired behaviors of the system or
erroneous ones, and for erroneous specifications, it is useful to have
a \emph{certificate} (as defined below) to identify an error trace of the system~\cite{Ehlers2010,LH2000}.
Note that {\em standard graphs} are relevant for testing the non-emptiness of Streett
automata and the verification of {\em closed} systems, while {\em game graphs} are relevant
for the synthesis and verification of {\em open} systems.
\smallskip\noindent{\em Previous results.}
We summarize the previous results for game graphs and graphs with Streett objectives.
\emph{Game graphs.} We consider the computation of the winning set for player~1 in game graphs.
For {\em reachability} objectives, the problem is PTIME-complete, and the computation can be
achieved in time linear in the size of the graph~\cite{Beeri80,Immerman81}.
For {\em B\"uchi} objectives, the current best known algorithm requires $O(n^2)$ time~\cite{CH12,CH14}.
For {\em general Streett} objectives, the problem is coNP-complete~\cite{EJ88}, and for {\em one-pair
Streett} objectives the current best known algorithm requires $O(n m)$ time~\cite{Jurdzinski00}.
One-pair Streett objectives also corresponds to the well-known parity games problem
with three priorities (the parity games problem in general is in UP $\cap$ coUP~\cite{Jur98}; it is one of the rare and intriguing combinatorial problems that lie in UP $\cap$ coUP,
but not known to be in PTIME).
Despite the importance of game graphs with one-pair Streett objectives in numerous
applications and several algorithmic ideas to improve the running time
for general parity games~\cite{VJ00,JPZ08,Schewe07} or B\"uchi games~\cite{CJH03,CH12,CH14},
there has been no algorithmic improvement since 2000~\cite{Jurdzinski00} for one-pair Streett games.
\emph{Graphs.} In standard graphs we study the computation of the winning
set for general Streett objectives. If $x$ belongs to the winning set, it is often useful to output a
\emph{certificate} for $x$. Let $S$ be a (not necessarily maximal)
strongly connected component (\textsc{scc}{}) that is reachable
from $x$ such that for all $1 \leq j \leq k$ we have either $S \cap L_j=\emptyset$ or $S \cap U_j \neq \emptyset$
(i.e., if $S$ contains a vertex from $L_j$ then it also contains a vertex from $U_j$).
A certificate is a ``lasso-shaped'' path that reaches $S$ and then visits all
vertices in $S$ infinitely often to satisfy the general Streett objective.
The basic algorithm~\cite{EL87,LichtensteinP85} for the winning set problem
has an asymptotic running time of $O((m+b)\min(n,k))$ with
$b = \sum_{j=1}^k (\lvert L_j \rvert + \lvert U_j \rvert) \le 2nk$.
Within the same time bound Latvala and Heljanko~\cite{LH2000} additionally
compute a certificate of size at most $n\min(n,2k)$.
Duret-Lutz~et~al.~\cite{DPC2009} presented a space-saving ``on-the-fly''
algorithm with the same time complexity for the slightly different transition-based
Streett automata.
The current fastest algorithm for the problem by Henzinger and Telle~\cite{HT1996} from 1996
has a running time of $O(m \min(\sqrt{m \log n}, k, n) + b \min(\log n, k))$,
however, given a start vertex $x$, to report the certificate for $x$ it
takes time $O(m \min(\sqrt{m \log n}, k, n) + b \min(\log n, k) + n\min(n,k))$.
\smallskip\noindent{\em Our contributions.} In this work our contributions are two-fold.
\emph{Game graphs.} We show that the winning set computation for game graphs with
one-pair Streett objectives can be achieved in $O(n^{2.5})$ time.
Our algorithm is faster for $m > n^{1.5}$, and breaks the long-standing
$O(n m)$ barrier for dense graphs.
We also discuss the implications of our algorithm for general parity games in Remark~\ref{rem:genpar}.
\emph{Graphs.}
We present an algorithm with $O((n^2 + b) \log n)$ running time for the
winning set computation in graphs with general Streett objectives,
which is faster for $m \ge \max(n^{4/3}, b^{2/3})\log^{1/3} n$ and $k \ge n^2 m^{-1} \log n$.
We additionally give an algorithm that computes a certificate
for a vertex $x$ in the winning set in time $O(m + n \min(n, k))$.
We also provide an example where the certificate
has size $\Theta(n \min(n, k))$, showing that no algorithm can
compute and \emph{output} a certificate faster.
In contrast to~\cite{HT1996} the running time of our algorithm for the winning
set computation does not change with certificate reporting.
Thus when certificates need to be reported and $k=\Omega(n)$, our algorithm is optimal up to a factor of
$\log n$ as the size of the input is at least $b$ and the size of the output is $\Omega(n^2)$.
\smallskip\noindent{\em Technical contributions.}
Both of our algorithms use a \emph{hierarchical (game) graph
decomposition} technique that
was developed
by Henzinger et al.~\cite{HKW99}
to handle \emph{edge deletions} in
\emph{undirected} graphs. In~\cite{CH12,CH14} it was extended
to deal with \emph{vertex deletions} in \emph{directed}
and \emph{game} graphs. We combine and extend this technique in two ways.
\emph{Game graphs.} The classical algorithm for one-pair Streett objectives repeatedly solves B\"uchi games such that the union of the winning sets of player~1 in the B\"uchi games is exactly the winning set for the one-pair Streett objective.
Schewe~\cite{Schewe07} showed that an algorithm for parity games by
Jurdzi\'nski~\cite{Jurdzinski00} can be used to compute small subsets of the
winning set of player~1, called \emph{dominions}, and thereby improved the
running time for general parity games. However his ideas
do not improve the running time for one-pair Streett (aka {\em parity-3}) games.
With this algorithm dominions of size $h$ in B\"uchi games can be found in time
$O(m h)$.
We extend this approach by using the hierarchical game graph decomposition
technique to find small dominions quickly and call the $O(n^2)$ B\"uchi game
algorithm of~\cite{CH12,CH14} for large dominions.
This extension is possible as we are able to show that, rather surprisingly,
it is sufficient to consider game graphs with $O(n h)$ edges to detect
dominions of size $h$ (see Lemma~\ref{lemm:level}).
\emph{Graphs.} In prior work that used the hierarchical graph
decomposition technique the runtime analysis relied heavily on the fact that
identified vertex sets that fulfilled a certain desired condition
were \emph{removed} from
the (game) graph after their detection. The work for identifying the vertex set was then charged in an amortization argument to the
removed vertex set. This is not possible for general Streett objectives on graphs, where \textsc{scc}{s} are identified and some but not all of its vertices might be removed. As a consequence a vertex might belong to
an identified \textsc{scc}{} multiple times. We
show how to overcome this difficulty by identifying, when an \textsc{scc}{} $S$ splits into multiple \textsc{scc}{}s,
an \textsc{scc}{} $X \subset S$ whose size is at most half
of the size of $S$. We identify $X$ by using Tarjan's \textsc{scc}{} algorithm~\cite{Tarjan1972}
on the graph and its \emph{reverse graph} in lock-step, thereby finding the smallest {\em top} (i.e. with no incoming edges) or
{\em bottom} (i.e. with no outgoing edges) \textsc{scc}{} contained in $S$. The smallest such \textsc{scc}{} $X$ has size at most $|S|/2$ and the algorithm takes time $O(|X| n)$ to find it, which we charge to the vertices in $X$.
In this way, every time a vertex is ``charged'', the size of the identified vertex set to which it belongs is halved, guaranteeing that each
vertex is charged only $O(\log n)$ times.
In Section~\ref{sec:parity} we present our algorithm for one-pair Streett
objectives in game graphs, in Section~\ref{sec:streett} the algorithm
for general Streett objectives in graphs.
\inmain{A full version with all proofs is provided in the appendix.}
\section{One-Pair Streett Objectives in Game Graphs}\label{sec:parity}
\input{definitions}
\input{newalgo}
\section{K-Pair Streett Objectives in Graphs}\label{sec:streett}
\input{streettdef}
\input{streettalgo}
\section*{Acknowledgements}
\noindent K.~C.\ is supported by the Austrian Science Fund (FWF): P23499-N23 and S11407-N23 (RiSE),
an ERC Start Grant (279307: Graph Games), and a Microsoft Faculty Fellows Award.
M.~H.\ is supported by the Austrian Science Fund (FWF): P23499-N23 and the Vienna Science and Technology Fund (WWTF) grant ICT10-002.
V.~L.\ is supported by the Vienna Science and Technology Fund (WWTF) grant ICT10-002.
The research leading to these results has received funding from the European
Research Council under the European Union's Seventh Framework Programme
(FP/2007-2013) / ERC Grant Agreement no. 340506.
\bibliographystyle{plain}
\subsection{Algorithm}
In this section we present our new algorithm to compute the winning set of
player~${\mathcal{E}}{}$ in a parity-3 game $P = (G, \alpha)$ in time $O(n^{2.5})$.
Its complement is the winning set of player~${\mathcal{O}}$.
\inmain{The correctness of the algorithm is proven in the appendix.}
\begin{algorithm2e}
\SetAlgoRefName{Parity-3}
\caption{New Algorithm for Parity-3 aka One-Pair Streett Objective}
\label{alg:parity}
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\BlankLine
\Input{a game graph $G = ((V, E),(V_\po, V_\pe))$ and a priority function
$\alpha: V \rightarrow \{-1, 0, 1\}$}
\Output{the winning set $W_\pe$ of player~${\mathcal{E}}$}
\BlankLine
$Z = \set{v \in V \mid \alpha(v) = -1}$; $E' = \set{(u, u) \mid u \in Z} \cup \set{(u,v) \in E \mid u \in V \setminus Z}$\;
$G' = (V, E')$ \tcc*{vertices with $\alpha = -1$ are absorbing in $G'$}
$W \leftarrow \emptyset$; $B \leftarrow \set{v \in V \mid \alpha(v) = 0}$\;
\Repeat{$D = \emptyset$}{
$D \leftarrow \domalg(G'[V], B, \sqrt{n})$\;
\lIf{$D = \emptyset$}{$D \leftarrow \buchialg(G'[V], B)$}
$A \leftarrow \mathit{Attr_\pe}(D, G[V])$; $W \leftarrow W \cup A$\;
$V \leftarrow V \setminus A$; $B \leftarrow B \setminus A$\;
}
\Return{W}\;
\BlankLine
\SetKwProg{myproc}{Procedure}{}{}
\myproc{$\domalg(G' = ((V, E'),(V_\po, V_\pe)), B, \domsize_{\text{max}})$}{
\For{$i \leftarrow 1$ \KwTo $\lceil\log(2\domsize_{\text{max}})\rceil$}{
construct $G'_i$; $\mathit{Bl}_i \leftarrow \{v \in V_\po \mid \OutDeg(v) > 2^i\}$\;
$Y_i \leftarrow \mathit{Attr_\po}(\mathit{Bl}_i, G'_i)$\;
$\mathcal{D}_i \leftarrow \progress(G'_i[V \setminus Y_i], B \setminus Y_i, 2^i)$\;
\lIf{$\mathcal{D}_i \ne \emptyset$}{\Return{union of dominions in $\mathcal{D}_i$}}
}
\Return{$\emptyset$}\;
}
\end{algorithm2e}
\smallskip\noindent{\em Initialization (Steps~1--3 of Algorithm~\ref{alg:parity}).}
First the algorithm constructs the modified game graph
$G' = ((V, E'), (V_\po, V_\pe))$ from $G$. Let $Z$ be the vertices in $V$
with priority $-1$. In $G'$ the vertices in $Z$ are made \emph{absorbing},
that is, the outgoing edges of the vertices in $Z$ are replaced with
self-loops. Otherwise $G'$ contains the same edges as $G$.
We will consider a B\"uchi game on $G'$ where the vertices in $Z$ have priority
$1$, and thus in the B\"uchi game there are only two priorities (priority~0
and~1). The construction of $G'$ ensures that dominions in the B\"uchi game
are also dominions in the parity-3 game $P$\inappendix{ (see Lemma~\ref{lemm:domG})}.
\smallskip\noindent{\em Iterated vertex deletions (Steps~4--9 of Algorithm~\ref{alg:parity}).}
The algorithm will repeatedly remove vertices from the graphs $G$ and $G'$.
Initially the set $V$ is the set of vertices in the input game graph $G$.
During the algorithm, we denote with $V$ the set of remaining vertices after
vertex deletions and we denote with $G[V]$ and $G'[V]$ the subgraphs induced
by the vertices remaining in $V$. The set of B\"uchi vertices $B$ maintains
the set of priority-0 vertices in $V$.
The vertex set removal is achieved by identifying dominions
and removing their attractors.
\smallskip\noindent{\em Dominion find and attractor removal.}
The algorithm repeatedly finds dominions in the B\"uchi game $(G'[V], B)$.
After a dominion in the B\"uchi game $G'[V]$ is found, its ${\mathcal{E}}$-attractor in
$G[V]$ is removed from $V$ and $B$.
Then the search for dominions is continued on the
remaining vertices. If all vertices in the B\"uchi game are winning for ${\mathcal{O}}$,
i.e., no dominion exists in the B\"uchi game, then Algorithm~\ref{alg:parity} terminates.
The winning set of player~${\mathcal{E}}$ is the union of the ${\mathcal{E}}$-attractors of all
found dominions. The remaining vertices are winning for player~${\mathcal{O}}$.
We now describe the steps to find dominions.
\smallskip\noindent{\em Steps of dominion find.}
For the search for dominions in the B\"uchi game $(G'[V], B)$ we use two
different algorithms, $\buchialg$ and $\domalg$. We first search for ``small''
dominions with up to $O(\domsize_{\text{max}})$ vertices with $\domsize_{\text{max}} = \sqrt{n}$ with
Procedure~$\domalg$. If no dominion is found, we can conclude that either
all dominions contain more than $\sqrt{n}$ vertices or the winning set of ${\mathcal{E}}{}$ on
the current game graph is empty (in this case the algorithm terminates).
The former case occurs at most $\sqrt{n}$ times and in such a case we use the $O(n^2)$
algorithm $\buchialg$ (Theorem~\ref{thrm:buchialg}) to obtain a dominion.
Below we describe the details of $\domalg$.
\smallskip\noindent{\em Graph decomposition for $\domalg$.}
In the Procedure~$\domalg$ we use the following graph decomposition.
For a game graph $G' = ((V, E'), (V_\po, V_\pe))$ we denote its decomposition with
$\set{G'_i}$. We consider the incoming edges of each vertex in $E'$ in a fixed
order: First the edges from vertices in $V_\pe$, then the remaining edges.
We construct $\log n$ graphs $G'_i = (V, E'_i)$, $1 \le i \le \log n$, where the
set of edges $E'_i$ contains for each vertex $v \in V$ with
$\OutDeg(v) \le 2^i$ all its outgoing edges in $E'$ and in addition for each
vertex $v \in V$ its first $2^i$ incoming edges in $E'$.
Note that (1) $E'_i \subseteq E'_{i+1}$, (2) $\lvert E'_i\rvert \le 2^{i+1} n$, and (3) $G'_{\log{n}} = G'$.
We color ${\mathcal{O}}{}$-vertices $v$ with $\OutDeg(v) > 2^i$ \emph{blue} in $G'_i$ and
denote the set of blue vertices with $\mathit{Bl}_i$. We call vertices with
$\OutDeg(v) \le 2^i$ \emph{white}.
\smallskip\noindent{\em Procedure $\domalg$ (Steps~11--17 of Algorithm~\ref{alg:parity}).}
The Procedure~$\domalg$ searches for dominions in the subgraphs $G'_i$,
starting at $i=1$. The index $i$ is increased one by one up to at most
$i = \lceil\log(2\domsize_{\text{max}})\rceil$ (with $\domsize_{\text{max}} = \sqrt{n}$) as long as no dominion
was found. Let $Y_i$ be the ${\mathcal{O}}{}$-attractor of blue vertices in $G'_i$, i.e.,
of ${\mathcal{O}}{}$-vertices that are missing outgoing edges in $G'_i$.
To ensure that dominions found in the subgraph $G'_i$ are also dominions in $G'$,
only the vertices in $V \setminus Y_i$ are considered. The $\progress$ algorithm
(Theorem~\ref{thrm:progress}) is used to find dominions of size at most $O(2^i)$ in $G'_i[V \setminus Y_i]$.
The following key lemma describes the central connection between dominions of a
certain size and our graph decomposition. Namely, if a dominion $D$ is found
in $G'_i$ but not in $G'_{i-1}$, then $\mathit{Attr_\pe}(D,G')$ contains more than $2^{i-1}$
vertices. This has the remarkable consequence, detailed in
Corollary~\ref{coro:maxit}, that every dominion of size $h$ can be found by
searching for a dominion in $G'_i$ with $i = \lceil \log(2h) \rceil$. This will be crucial for our runtime analysis.
\begin{lemma}\label{lemm:level}
Let $G' = ((V, E'),(V_\po, V_\pe))$ be a game graph and $\set{G'_i}$ its graph decomposition.
For $1 \le i \le \log n$ we define the following sets: the set of blue vertices
$\mathit{Bl}_i = \{v \in V_\po \mid \OutDeg(v) > 2^i\}$, the attractor of blue vertices
$Y_i = \mathit{Attr_\po}(\mathit{Bl}_i, G'_i)$, and the set of dominions $\mathcal{D}_i =
\progress(G'_i[V \setminus Y_i], B \setminus Y_i, 2^i)$. If a dominion $D$ is
contained in $\mathcal{D}_i$ but not in $\mathcal{D}_{i-1}$, then $\mathit{Attr_\pe}(D,G')$
contains more than $2^{i-1}$ vertices.
\end{lemma}
\begin{proof}
We distinguish three cases:
\emph{Case 1}: The dominion $D$ contains more than $2^{i-1}$ vertices. This
situation might arise as Procedure~$\progress(G'_{i-1}[V \setminus Y_{i-1}], B \setminus
Y_{i-1}, 2^{i-1})$ only guarantees to detect dominions of size at most
$2^{i-1}$. In this case the lemma is satisfied trivially.
\emph{Case 2}:
The dominion $D$ contains a vertex $v \in V_\po$ that is blue in $G'_{i-1}$, i.e.,
an ${\mathcal{O}}$-vertex with more than $2^{i-1}$ outgoing edges. Since
$D$ is ${\mathcal{O}}{}$-closed, we have $\Out(v) \subseteq D$. Thus $\lvert
\mathit{Attr_\pe}(D,G') \rvert \ge \lvert D \rvert > 2^{i-1}$ in this case.
\emph{Case 3}: All vertices $v \in V_\po$ in $D$ are white in $G'_{i-1}$ and thus
the outgoing edges of the ${\mathcal{O}}{}$-vertices in $D$ are the same in $G'_{i-1}$ and $G'_i$. There are two subcases.
\emph{Case 3a}:
All edges $(u,v)$ from vertices $u \in V_\pe \cap D$ to vertices $v \in D$
that are present in $G'_i$ are also present in $G'_{i-1}$. Let $\sigma$ be
the winning strategy of~${\mathcal{E}}{}$ for the vertices in $D$ found in $G'_i$.
This implies that (i)
$D$ is ${\mathcal{O}}$-closed in $G'_{i-1}$ and (ii) all edges $(u, v)$ with $u \in D \cap
V_\pe$ and $v = \sigma(u)$ are contained in $G'_{i-1}$. Thus $\sigma$
is also a winning strategy of~${\mathcal{E}}{}$ for the vertices in $D$ in $G'_{i-1}$.
Hence the set $D$ is a dominion in $G'_{i-1}$.
Thus either Case~1 applies or the dominion would already have been detected in
iteration $i-1$, a contradiction.
\emph{Case 3b}: There exists a vertex $u \in V_\pe \cap D$ that has an outgoing
edge $(u,v)$ to a vertex $v \in D$ in $G'_{i}$ but not in $G'_{i-1}$.
This implies $\InDeg(v) > 2^{i-1}$. By the ordering of the incoming edges and
the fact that $u \in V_\pe$, at least $2^{i-1}$ edges in $\In(v)$ emanate from
vertices in $V_\pe$. By the definition of an attractor, all these vertices are
contained in $\mathit{Attr_\pe}(D,G')$. Thus we have $\lvert \mathit{Attr_\pe}(D,G') \rvert > 2^{i-1}$ as required.
\end{proof}
\begin{corollary}\label{coro:maxit}
Let $G'$, $\set{G'_i}$, $\mathit{Bl}_i$, $Y_i$, and $\mathcal{D}_i$ be defined as in
Lemma~\ref{lemm:level}. Let $D$ be a dominion in $G'$ with
$D = \mathit{Attr_\pe}(D, G')$ and $h = \lvert D \rvert$. Then for $i
= \lceil \log(2h) \rceil$ the set of dominions $\mathcal{D}_i$ contains $D$.
\end{corollary}
\inappendix{
\begin{proof}
By the definition of $i$ we have $2^{i-2} < h \le 2^{i-1}$.
Assume by contradiction that $\mathcal{D}_i$ does not contain $D$. Since $G' =
G'_{\log n}$ and $Y_{\log n} = \emptyset$, by Theorem~\ref{thrm:progress} there
exists some $i'$ with $i < i' \le \log n$, such that $\mathcal{D}_{i'}$ contains $D$.
Let $i^*$ be the smallest $i'$ such that $\mathcal{D}_{i'}$ contains $D$. Note
that $i^* > i$. We have that $D \notin \mathcal{D}_{i^*-1}$. By
Lemma~\ref{lemm:level} this implies $\lvert \mathit{Attr_\pe}(D,G') \rvert > 2^{i^*-1}$, a contradiction to $h \le 2^{i-1}$.
\end{proof}
}
\begin{corollary}\label{coro:minsize}
Either the Procedure $\domalg(G'[V], B, \domsize_{\text{max}})$ returns a dominion
or every dominion $D$ in $G'[V]$ with $D = \mathit{Attr_\pe}(D, G'[V])$ has size greater than $\domsize_{\text{max}}$.
\end{corollary}
In the runtime analysis we will additionally use the following lemma, which
follows from the inductive construction of attractors.
\begin{lemma}\label{lemm:attrsubset}
Let the game graphs $G$ and $G'$ and the vertex set $V$ be defined as in
Algorithm~\ref{alg:parity}. Then for a player ${p} \in \set{{\mathcal{E}}, {\mathcal{O}}}$ and
every set $U \subseteq V$ it holds that $\mathit{Attr}_{p}(U, G'[V]) \subseteq \mathit{Attr}_{p}(U, G[V])$.
\end{lemma}
\inappendix{
\begin{proof}
Let us consider the attractor computation $\mathit{Attr}_{p}(U, G'[V])$ and $\mathit{Attr}_{p}(U, G[V])$
as defined in ($\ddag$), and let us call the respective sequences as $R'_i$ and $R_i$
respectively.
By the definition of $G'$ for every vertex $v$ in $V$ either (1) $\Out'(v) = \Out(v)$
or (2) $\Out'(v) = \{v\}$.
It is straightforward to prove by induction that $R'_i \subseteq R_i$ and the desired
result follows.
\end{proof}
}
\begin{lemma}[Runtime]\label{lemm:time}
Algorithm~\ref{alg:parity} can be implemented in $O(n^{2.5})$ time.
\end{lemma}
\begin{proof}
Algorithm~\ref{alg:parity} can be initialized in $O(m)$ time as the graph $G'$
and the set $B$ can be constructed from $G$ in linear time.
Note that the number of edges $m'$ in $G'$ is at most the number of edges $m$ in $G$.
For the operations in the repeat-until loop we analyze the \emph{total}
running time over all iterations of the loop. The runtime analysis relies
heavily on the fact that when a dominion $D$ is identified, the vertices in
$\mathit{Attr_\pe}(D, G)$ and their incident edges are removed from $G$ and $G'$.
In combination with Corollary~\ref{coro:minsize}, this ensures that $\buchialg$
is called at most $O(n / \domsize_{\text{max}})$ times. By Theorem~\ref{thrm:buchialg} one
call to $\buchialg$ takes time $O(n^2)$. With $\domsize_{\text{max}} = \sqrt{n}$ we obtain a
total time spent in $\buchialg$ of $O(n^{2.5})$.
To analyze the total time spent in $\domalg$, we first show how to efficiently
construct the graph decomposition $\set{G'_i}$ of $G'$.
We maintain the following data structure for $G'$ over all iterations of Algorithm~\ref{alg:parity}. At each
vertex~$v$ of $G'$ we maintain (a) a sorted list of inedges $\In(v)$, and (b) a list of
outedges $\Out(v)$. Additionally we maintain for each edge $(u,v)$ a pointer to its position in the
inlist of $v$ and the outlist of $v$. This allows us to update the data structure in time
proportional to the degree of $v$ when a vertex $v$ is removed.
As each vertex can be deleted at most once, the
total time to update this data structure is bounded by $O(n + m)$.
We next analyze the time needed per iteration $i$ of the for-loop
in $\domalg$. Given the above data structure, the graph $G'_i$, the set of blue
vertices~$\mathit{Bl}_i$, and the attractor~$Y_i = \mathit{Attr_\po}(\mathit{Bl}_i,
G'_i)$
can be constructed in time $O(n \cdot 2^i)$.
By Theorem~\ref{thrm:progress} the time for one call of
the subroutine $\progress(.,.,2^i)$ on graph $G'_i$,
is $O(n \cdot 2^i \cdot 2^i) = O(n \cdot 2^{2i})$.
Let $i^*$ be the iteration at which Procedure~$\domalg$ stops after it is
called by Algorithm~\ref{alg:parity}. The runtime
for this call to Procedure~$\domalg$ from $i = 1$ to $i^*$ forms a geometric series that
is bound by $O(n \cdot 2^{2i^*})$.
By Lemmata~\ref{lemm:level} and~\ref{lemm:attrsubset} and Corollary~\ref{coro:minsize}
either (1) a dominion $D$ with $\lvert
\mathit{Attr_\pe}(D,G) \rvert > 2^{i^*-1}$ vertices was found by $\domalg$ or (2) all
dominions in $G'$ have more than $\domsize_{\text{max}}$ vertices or there are no more dominions in $G'$. Thus either (2a) a dominion~$D$
with more than $\domsize_{\text{max}}$ vertices is detected in the subsequent call to $\buchialg$
or (2b) there is no dominion in $G'$ and this is the
last iteration of Algorithm~\ref{alg:parity}. Case (2b) can happen at most
once and its runtime is bounded by $O(n \cdot 2^{2\log(2 \domsize_{\text{max}})}) =
O(n^2)$. In the cases (1) and (2a) more than $2^{i^*-2}$ vertices are removed from the graph
in this
iteration, as $\domsize_{\text{max}} > 2^{i^*-2}$. We charge each such vertex $O(n \cdot 2^{i^*}) = O(n \cdot \domsize_{\text{max}})$ time.
Hence the total runtime for these cases is $O(n^2 \cdot \domsize_{\text{max}}) = O(n^{2.5})$.
It remains to consider the total time needed to compute $A = \mathit{Attr_\pe}(D, G[V])$.
By Lemma~\ref{lemm:attr}.~(\ref{sublemm:attrtime}) the attractor $A$ can be
computed in time $O(\sum_{v \in A} \lvert\In(v)\rvert)$. Since the edges adjacent
to vertices in $A$ are removed from $G$ after the iteration in which $D$ was
found, this attractor computation can be done in total time $O(m)$.
We conclude that the runtime of Algorithm~\ref{alg:parity} is $O(n^{2.5})$.
\end{proof}
\inappendix{
We will show the correctness of Algorithm~\ref{alg:parity} by first proving
that every dominion found in the B\"uchi game on $G'$ is indeed a dominion in
the parity-3 game on $G$. Together with
Lemma~\ref{lemm:attr}.~(\ref{sublemm:subgraph}) this implies that the computed set $W$
is indeed a part of the winning set of player~${\mathcal{E}}$ in the parity-3 game. We
then provide a winning strategy for player~${\mathcal{O}}$ for all remaining vertices.
\begin{lemma}\label{lemm:domG}
Let the game graphs $G$ and $G'$ and the vertex sets $V$ and $B$ be defined as
in Algorithm~\ref{alg:parity}. If $D$ is a dominion in the B\"uchi game
$(G'[V], B)$, then $D$ is a dominion in the parity-3 game $P=(G[V],\alpha)$.
\end{lemma}
\begin{proof}
Let $Z$ be the vertices in $V$ with priority $\alpha$ equal to $-1$. The vertices
in $Z$ have priority~1 in the B\"uchi game, i.e., $Z \cap B = \emptyset$.
Whenever a play in $G'[V]$
reaches a vertex $u$ in $Z$, only $u$ will be visited in the subsequent play
since $\Out(u) = \{u\}$. Thus no vertex in $Z$ is winning for ${\mathcal{E}}$ in $(G'[V], B)$,
i.e., $D \cap Z = \emptyset$. Hence for all vertices in $D$ the outgoing edges
are the same in $G[V]$ and $G'[V]$. Thus $D$ is ${\mathcal{O}}$-closed in $G[V]$ and the
winning strategy of player~${\mathcal{E}}$ for $D$ in the B\"uchi game $(G'[V], B)$ is also
winning for player~${\mathcal{E}}$ for all vertices in $D$ in the parity-3 game $P$.
\end{proof}
\begin{lemma}[Correctness]\label{lemm:correctness}
Given a parity-3 game $P$, let $W$ be the output of Algorithm~\ref{alg:parity}.
We have: (1)~\emph{(Soundness).} $W \subseteq W_\pe(P)$; and
(2)~\emph{(Completeness).} $W_\pe(P) \subseteq W$.
\end{lemma}
\begin{proof}
The first part on soundness follows from Lemmata~\ref{lemm:domG}
and~\ref{lemm:attr}.~(\ref{sublemm:subgraph}).
We now prove the completeness result.
Given the output $W$, let $\overline{W}$ denote the complement set.
When Algorithm~\ref{alg:parity} terminates, the winning set of player~${\mathcal{E}}$
in the B\"uchi game $(G'[\overline{W}],B)$ is empty (otherwise
the algorithm would not have terminated).
Also note that since the algorithm removes attractors for ${\mathcal{E}}$, the set $\overline{W}$
is closed for ${\mathcal{E}}$ (by Lemma~\ref{lemm:attr}.~(1)).
Consider the set $Z = \set{v \in \overline{W} \mid \alpha(v) = -1}$, its
attractor $X = \mathit{Attr_\po}(Z,G[\overline{W}])$, and the subgame induced by
$U = \overline{W} \setminus X$.
Note that in $U$ the game graphs $G$ and $G'$ coincide.
Thus all vertices in $U$ must be winning for player~${\mathcal{O}}$ in the B\"uchi game $(G[U],B)$
as otherwise $W_\pe$ would have been non-empty for $(G'[\overline{W}],B)$.
We prove the lemma by describing a winning strategy for player~${\mathcal{O}}$ in $P$
for all vertices in $\overline{W}$.
Since $\overline{W}$ is ${\mathcal{E}}$-closed, for vertices in $Z \cap V_{{\mathcal{O}}}$, the winning
strategy chooses an edge in $\overline{W}$.
For vertices in $X$ player~${\mathcal{O}}$ follows his attractor strategy to $Z$.
In the subgame induced by $U = \overline{W} \setminus X$ player~${\mathcal{O}}$ follows his winning strategy
in the B\"uchi game $(G[U],B)$. Then in a play either (i)~$X$ is visited infinitely often;
or (ii)~from some point on only vertices in $U$ are visited.
In the former case, the attractor strategy ensures that then some vertex in $Z$ with priority~$-1$
is visited infinitely often; and in the later case, the subgame winning strategy ensures that
only vertices with priority~$1$ and no vertices with priority~$0$
are visited infinitely often.
It follows that $\overline{W} \subseteq W_\po(P)$, i.e., $W_\pe(P) \subseteq W$, and the desired
result follows.
\end{proof}
Lemmata~\ref{lemm:time} and~\ref{lemm:correctness} yield the following result.
}
\begin{theorem}
Algorithm~\ref{alg:parity} correctly computes the winning sets in parity-3 games in $O(n^{2.5})$ time.
\end{theorem}
\smallskip\noindent{\em Computation of winning strategies.}
In parity-3 games the previous results for computing winning strategies for the
players in their respective winning sets are as follows:
The small-progress measure algorithm of~\cite{Jurdzinski00} requires $O(n m)$~time
to compute the winning strategy of the player whose parity is equal to
the parity of the lowest priority and $O(n^2 m)$~time to
compute the respective winning strategies for both players; Schewe~\cite{Schewe08}
shows how to modify the small-progress measure algorithm to compute the
respective winning strategies of both players in
$O(n m)$~time. We show that our algorithm also computes the respective
winning strategies in $O(n^{2.5})$~time.
We first observe that the algorithm of~\cite{CH14} that solves B\"uchi games
in $O(n^2)$~time also computes the respective winning strategies of both players
(the algorithm is based on identifying traps and attractors, and the corresponding
winning strategies are identified immediately with the computation).
In Lemma~\ref{lemm:correctness} we describe the strategy computation for a winning
strategy for player~${\mathcal{O}}$ which involves an attractor strategy and the sub-game
strategy for B\"uchi games, each of which can be computed in $O(n^2)$ time.
A winning strategy for player~${\mathcal{E}}$ is obtained in the iterations of the algorithm,
i.e., whenever we obtain a dominion by solving B\"uchi games we also obtain a
corresponding winning strategy, and similarly for the attractor computation.
Thus the winning strategy for player~${\mathcal{E}}$ can be computed in $O(n^{2.5})$~time.
\begin{corollary}
Winning strategies for player~${\mathcal{E}}$ and player~${\mathcal{O}}$ in parity-3 games in their respective winning sets
can be computed in $O(n^{2.5})$ time.
\end{corollary}
\begin{remark}{{\sc (Discussion on general parity games).}}\label{rem:genpar}
We now discuss the implication of our result for general parity games (we do not
discuss general Streett games where the problem is coNP-complete~\cite{EJ88}).
The current best known algorithm for parity games with dependence on the number
of priorities $d$ is from~\cite{Schewe07}, and the (simplified) running time for $d =
o(\sqrt{n})$ is $O(n^{\gamma(d)} \cdot m)$, where $\gamma(d)$ is approximately
$d/3$ for large $d$.
More precisely,
$\gamma(d) = d/3 + 1/2 - 1/(\lceil0.5 d\rceil \lfloor 0.5 d \rfloor)$ for odd $d$,
and $\gamma(d) = d/3 + 1/2 - 1/(3 d) - 1/(\lceil0.5 d\rceil \lfloor 0.5 d \rfloor)$
for even $d$.
Our algorithm for parity-3 games also extends to parity games as a recursive
algorithm as follows:
we apply our initialization step and iterated vertex deletions, and
to find dominions we replace $\buchialg$ by our recursive algorithm that handles
games that have one less priority and replace $\domalg$ by a procedure to find
dominions with small-progress measure of~\cite{Jurdzinski00} with our graph
decomposition and codomain bounded by $\domsize_{\text{max}}$ (where $\domsize_{\text{max}}$ is chosen to balance the
running time of the two dominion find procedures).
For the sake of simplicity of presentation we consider the case of constantly many
priorities and refer for the analysis of the general case to~\cite{Schewe07}.
Using the notation of~\cite{Schewe07} with $\beta(d) =
\gamma(d)/(\lfloor0.5d\rfloor + 1)$, we obtain with $\domsize_{\text{max}} = n^{\beta(d)}$
a running time of our algorithm of $O(n^{1+\gamma(d+1)})
= O(n^{2 + \gamma(d) - \beta(d)})$ for parity
games with $d$ priorities, i.e., it replaces $m$ of~\cite{Schewe07} by
$n^{2-\beta(d)}$.
We present the details of the calculation.
We show by induction that our algorithm solves parity games with $d-1$ priorities
in $O(n^{1+\gamma(d)})$ time.
The base case of $d-1=3$ follows from our algorithm for parity-3 games.
The inductive case is as follows:
To solve a parity game with $d$ priorities, our algorithm calls the progress
measure algorithm (on the graph decomposition) for $d-1$ priorities with
$\domsize_{\text{max}} = n^{\beta(d)}$ and recursively
calls the algorithm for $d-1$ priorities at most $O(n^{1-\beta(d)})$ times.
The total time for the progress measure algorithm is bounded by
$O(n^2 n^{\beta(d)\lfloor0.5d\rfloor})$ and the total time for all calls to
the algorithm for $d-1$ priorities is bounded by $O(n^{1-\beta(d)}n^{1+\gamma(d)})$.
We obtain the recurrence $\gamma(d+1) = 1 + \gamma(d) - \beta(d)$,
which yields $\gamma(d)$ as defined above and a running time of
$O(n^{1+\gamma(d+1)})$ for $d$ priorities.
In the limit $\beta(d)$ approaches $2/3$. For small $d$ we compare our running
times with Schewe's in Table~\ref{tab:comparison}.
We have presented the details for parity-3 games for the following reasons:
(1)~All the key ideas and conceptual details are easily demonstrated for the
simpler case of parity-3 games and
(2)~while all previous ideas for general parity games (such
as~\cite{JPZ08,Schewe07}) and for B\"uchi games (such
as~\cite{CJH03,CH12,CH14}) fail to improve the running time for parity-3 games,
our approach succeeds to break the long-standing $O(n m)$ barrier for dense
graphs.
\end{remark}
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\# priorities & 3 & 4 & 5 & 6 & 7 \\
\hline
\hline
Schewe~\cite{Schewe07} &
$O(m n)$ & $O(m n^{3/2})$ & $O(m n^{2})$ & $O(m n^{7/3})$ & $O(m n^{11/4})$\\
\hline
\cite{Schewe07} if $m = \Theta(n^2)$ &
$O(n^3)$ & $O(n^{7/2})$ & $O(n^{4})$ & $O(n^{13/3})$ & $O(n^{19/4})$\\
\hline
this paper &
$O(n^{2.5})$ & $O(n^{3})$ & $O(n^{10/3})$ & $O(n^{15/4})$ & $O(n^{65/16})$\\
\hline
\end{tabular}
\caption{Comparison of running times of~\cite{Schewe07} and our algorithm for small priorities.}\label{tab:comparison}
\end{center}
\end{table}
\subsection{Algorithm}\label{sec:alg}
By abuse of notation we denote by $G$ the \emph{current} graph maintained by the algorithm
where some edges and vertices might have been deleted and use \emph{input graph} to denote the
unmodified, strongly connected graph for which a good component is searched.
Our algorithm for good component detection is given in Algorithm~\ref{alg:streett}.
It maintains in a list $Q{}$ a partition of the vertices in $G$ into sets
such that every \textsc{scc}{} of $G$ is contained in the subgraph induced by one of the vertex sets. The list is
initialized with the set of all vertices in the strongly connected input graph.
We will show that if a good component exists, it must be fully contained in one of
the vertex sets in the partition. The algorithm repeatedly removes a set $S$ from
$Q{}$ and identifies and deletes bad vertices from $G[S]$.
If no edge is contained in $G[S]$, the set $S$ is removed as
it can only induce trivial components.
Otherwise the subgraph $G[S]$ is either
determined to be strongly connected and output as a good component or
a ``small'' maximal \textsc{scc}{} in $G[S]$ is identified.
To find a small maximal \textsc{scc}{} the algorithm
searches in lock-step in $G[S]$ and in $\mathit{RevG}[S]$ for a bottom \textsc{scc}{} and stops as soon as one of the searches stops.
(A bottom \textsc{scc}{} in $\mathit{RevG}[S]$ is a top \textsc{scc}{} in $G[S]$.) We only describe
the search in $G[S]$ here, the search in $\mathit{RevG}[S]$ is analogous.
The algorithm uses the hierarchical
graph decomposition in $G[S]$.
The subgraph $G_i[S]$ for any $i$ contains only the outedges
of vertices with an outdegree of at most $2^i$. The search for a
bottom \textsc{scc}{} is started at $i = 1$, then $i$ is increased one by one
if necessary, up to at most $\log n$.
If for some $i$ we can identify a bottom \textsc{scc}{} that
does not contain any blue vertex (i.e. a vertex for which some edges are
missing in $G_i$), then the found \textsc{scc}{} in $G_i[S]$ must also be a bottom \textsc{scc}{} in $G[S]$. If multiple bottom \textsc{scc} s (without blue vertices) are found in $G_i[S]$,
we only consider the smallest one.
We then put the newly detected \textsc{scc}{} and the ``rest'' of $S$ back into $Q{}$.
\begin{algorithm2e}
\SetAlgoRefName{GoodComponent}
\caption{Detection of good components for the winning set computation in graphs with $k$-pair Streett objectives}
\label{alg:streett}
\SetKwInOut{Input}{input}
\SetKwInOut{Output}{output}
\SetKw{continue}{continue}
\SetKw{Break}{break}
\SetKwBlock{parallel}{in lock-step for $H \in \set{G,\mathit{RevG}}$}{end}
\BlankLine
\Input{strongly connected graph $G = (V, E)$, Streett pairs $( L_j, U_j )$ for $j = 1, \dotsc, k$}
\Output{a good component in $G$ if one exists}
\BlankLine
add $\construct(V)$ to $Q{}$\;
\While{$Q{} \ne \emptyset$}{ \label{l:while}
pull $\mathit{D}(S)$ from $Q{}$\label{l:pull}\;
\lWhile{$\bad(\mathit{D}(S)) \ne \emptyset$}{\label{l:bad}
$\mathit{D}(S) \leftarrow \remove(S, \mathit{D}(S), \bad(\mathit{D}(S)))
}
\If{$G[S]$ contains at least one edge}{ \label{l:le2}
\parallel{
\For{$i \leftarrow 1$ \KwTo $\log n$}{\label{l:for}
construct $H_i[S]$;\label{l:forl1}
$\mathit{Bl}_i \leftarrow \{v \in S \mid \OutDeg(v) > 2^i\}$
\tcc*{$\OutDeg$ in $H$
$Z \leftarrow S \setminus \reach(\mathit{Bl}_i, H_i[S])$
\tcc*{$Z$ cannot reach $\mathit{Bl}_i$}
\If{$Z \ne \emptyset$}{
$H[X] \leftarrow \sccalg(H_i[Z])$\;
\lIf{$X = S$}{\KwRet{$H[S]$}}\label{l:stillconnected}
\If{$\lvert X \rvert \le \lvert S \rvert / 2$}{\label{l:smallerhalf}
\Break
}
}
}
}
add $\remove(S, \mathit{D}(S), X)$ and $\construct(X)$ to $Q{}$\label{l:addRest}\;
}
}
\KwRet{no good component exists}
\end{algorithm2e}
\inappendix{
The idea of the running time analysis is as follows.
We can show that a bottom \textsc{scc}{} of $G[S]$ identified in iteration~$i$ of the for-loop
must contain $\Omega(2^i)$ vertices. In
time $O(n2^i)$ a standard \textsc{scc}{} algorithm can compute all \textsc{scc} s of $G_i[S]$
and thus also the smallest bottom \textsc{scc}{}. The time needed for
the search in all graphs $G_{{i}'}[S]$ up to $i$ can be bounded with an additional factor
of two. Thus the work for the search is $O(n)$ per vertex in the identified \textsc{scc}{}.
Given that the subgraph $G[S]$ was split into at least one top and one bottom \textsc{scc}{},
the smallest top or bottom \textsc{scc}{} contains at most half of the vertices of the
subgraph. By searching for a smallest bottom \textsc{scc}{} (without blue vertices)
in $G_i[S]$ and $\mathit{RevG}_i[S]$ we find one top or bottom \textsc{scc}{} with at most
half of the vertices of the subgraph. We charge the work for finding
such an \textsc{scc}{} to the vertices in this \textsc{scc}{}. This guarantees that each
vertex will be charged at most $O(\log n)$ times over the whole running time
of the algorithm. Thus we can bound the total running time for computing \textsc{scc} s
by $O(n^2 \log n)$.
We additionally have to take the time for the maintenance of the data structures
into account. Here we use the properties of the data structure $\mathit{D}(S)$ described
in Lemma~\ref{lemma:ds} to obtain a running time of $O((n + b) \log n)$ for the
maintenance of the data structures and the identification of bad vertices over the
whole algorithm.
Combined these ideas lead to a total running time of $O((n^2 + b) \log n)$.
}
\inmain{In the analysis of the running time of Algorithm~\ref{alg:streett} the
crucial observation is that if we detect a maximal \textsc{scc}{} not containing
any blue vertices in $G_{{i}^*}$, then this \textsc{scc}{} contains at least
$2^{{i}^*-1}$ vertices. Thus we can charge all the work of the search up to level $i^*$ to the vertices in the \textsc{scc}{}.
}
\begin{lemma}\label{lemma:size}
Whenever in Algorithm~\ref{alg:streett} the for-loop stops for
$H \in \set{G,\mathit{RevG}}$ and some $i = {i}^*$
with a nonempty vertex set $Z = S \setminus \reach(\mathit{Bl}_{{i}^*}, H_{{i}^*}[S])$
and the smallest bottom \textsc{scc}{} $H[X]$ in $H_{{i}^*}[Z]$ returned by $\sccalg(H_{{i}^*}[Z])$
with $ \lvert X \rvert \le \lvert S \rvert /2$, then $H[X]$ contains at least $2^{{i}^*-1}$ vertices.
\end{lemma}
\inappendix{
\begin{proof}
As $\mathit{Bl}_{{i}^*-1}$ is the set of vertices in $H_{{i}^*-1}[S]$ with
outdegree larger than $2^{{i}^*-1}$, any bottom $\textsc{scc}{}$ $H[Y]$ that contains a vertex of $\mathit{Bl}_{{i}^*-1}$,
has $\lvert Y \rvert \ge 2^{{i}^*-1}$.
Hence it suffices to show that $X \cap \mathit{Bl}_{{i}^*-1} \ne \emptyset$.
Assume by contradiction that $X \cap \mathit{Bl}_{{i}^*-1} = \emptyset$.
Since $H[X]$ is a bottom \textsc{scc}{},
no vertex
in $X$ can reach any vertex in $\mathit{Bl}_{{i}^*-1}$, i.e.,
$X \subseteq S \setminus \reach(\mathit{Bl}_{{i}^*-1}, H_{{i}^*}[S])$. As all edges in $H_{{i}^*-1}[S]$
are contained in $H_{{i}^*}[S]$, this implies
$X \subseteq S \setminus \reach(\mathit{Bl}_{{i}^*-1}, H_{{i}^*-1}[S])$.
Since $\sccalg$ finds the smallest bottom \textsc{scc}{} in graph $H_i$ for each $i$, the for-loop would thus
have terminated in an iteration $i \le {i}^*-1$. Contradiction.
\end{proof}
}
\begin{lemma}[Runtime]\label{lemm:timeStreett}
Algorithm~\ref{alg:streett} can be implemented in time $O((n^2 + b) \log n)$.
\end{lemma}
\begin{proof}
The preprocessing and initialization of the data structure and the removal
of bad vertices in the whole algorithm take time $O(m + k + b)$ using
Lemma~\ref{lemma:ds}.
Additionally we maintain
at each vertex a list of its incoming and a list of its outgoing edges
including pointers to the lists of its neighbors, which we
use to update the lists of its neighbors. Since each vertex is
deleted at most once, this data structure can be constructed and maintained in
total time $O(n^2)$.
Consider the while loop where a set $S$ is removed from $Q{}$.
The lock-step search for $G[S]$ and $\mathit{RevG}[S]$ only increases the running time
by a factor of two, thus we restrict the analysis of the running time to $G[S]$.
The construction of $ G_i[S]$, $Z$, and $G[X]$ can all be done in time $O(n 2^i)$ for each $i$,
i.e., in total time $O(n2^{{i}^*})$ up to level $i^*$.
If $X=S$, then the algorithm terminates
and the time for processing $S$
can be bounded by $O(n2^{\log n}) = O(n^2)$.
If the processing of $S$ ends when some bottom \textsc{scc}{} $G[X] \subset G[S]$ is found,
let $i^*$ be the value of $i$ when $G[X]$ is detected and inserted into $Q{}$.
By Lemma~\ref{lemma:size} the set $X$ contains at least $2^{{i}^*-1}$ vertices.
We charge $O(n)$ to each vertex in $X$.
Since $\lvert X \rvert \le \lvert S \rvert / 2$, a vertex $v$ is only charged when the
size of the set in $Q{}$
containing $v$ is halved, which can happen at most $\lceil \log n \rceil$ times.
Thus the total running time for processing all sets $S$, except for the work in $\remove$ and $\construct$,
can be bounded by $O(n^2 \log n)$.
The $\remove$ and $\construct$ are called once per found bottom \textsc{scc}{}
$G[X]$ with $X \ne S$ and take by Lemma~\ref{lemma:ds} time $O(\lvert X \rvert + \mathit{bits}(X))$ time. Hence,
by charging $O(1)$ to the vertices in $X$ and, respectively, to $\mathit{bits}(X)$, the
total running time for this part can be bounded by $O((n + b) \log n)$ as each vertex
and bit will only be charged $O(\log n)$ times.
Combining all parts yields the claimed running time bound of $O((n^2 + b) \log n)$.
\end{proof}
To prove the correctness of Algorithm~\ref{alg:streett}
we first show that all candidates for good components are in $Q{}$ before
each iteration of the algorithm.
\begin{lemma}\label{lemma:allcandidates}
Before each iteration of the outer while-loop every good component of the input
graph is contained in one of the subgraphs $G[S]$ for which the data structure
$\mathit{D}(S)$ is maintained in the list $Q{}$.
\end{lemma}
\inappendix{
\begin{proof}
We will show that the algorithm never removes edges or vertices that belong to
a good component, which together with a correct initialization of the list $Q{}$
will imply the lemma.
At the beginning of the algorithm one data structure for the whole strongly connected
input graph is added to $Q{}$. Thus every good component is contained in
this data structure in $Q{}$ after the initialization.
At the beginning of each iteration of
the outer while-loop the data structure of one of the subgraphs $G[S]$
is pulled from the list $Q{}$. In Line~\ref{l:bad} we remove
vertices from the subgraph that are in some set $L_j$ but not
strongly connected to any vertex in $U_j$, i.e., bad vertices.
In Line~\ref{l:le2} we remove trivial \textsc{scc}{s}.
Observe that a good component is non-trivial and does not
contain any bad vertices. Thus the removal of bad vertices
and trivial \textsc{scc}{s} does not remove any vertices of a good component,
i.e., after the removal of these vertices
the updated subgraph $G[S]$ still contains the good components it contained
before. If no good component is identified in this iteration, i.e., the algorithm
does not terminate, we find a bottom or top \textsc{scc}{} $G[X]$, which is by definition
a maximal \textsc{scc}{}. Since a good component has to be strongly connected,
every good component in $G[S]$ must either be a subgraph of the newly identified \textsc{scc}{} $G[X]$
or does not contain \emph{any} vertex in $X$. Thus the removed edges between $G[X]$
and the remaining subgraph cannot belong to a good component. Finally, we add the
data structures for $G[X]$ as well as for $G[S \setminus X]$ to $Q{}$.
Thus no vertex or edge of a good component was removed and every good component
continues to be completely contained in a subgraph in $Q{}$.
\end{proof}
}
As all candidates for good components are maintained in the list $Q{}$,
it remains to show that the algorithm makes progress in each iteration and
correctly outputs a good component if and only if one exists.
\begin{lemma}[Correctness] \label{lemm:correctStreett}
Algorithm~\ref{alg:streett} outputs a good component if one exists,
otherwise the algorithm reports that no such component exists.
\end{lemma}
\begin{proof}
First we show that whenever Algorithm~\ref{alg:streett} outputs a subgraph $G[S]$,
then $G[S]$ is a good component.
Line~\ref{l:le2} ensures only non-trivial \textsc{scc}{s} are considered.
After the removal of bad vertices from $S$ in Line~\ref{l:bad}
we know that for
all $1 \le j \le k$ and all vertices in $S \cap L_j$ there exists a vertex in
$S \cap U_j$. Thus if $G[S]$ is strongly connected, then $G[S]$ is a good \textsc{scc}{}.
The algorithm computes a maximal \textsc{scc}{} in $G[S]$. If $G[S]$ is equal
to the found maximal \textsc{scc}{}, i.e., $G[S]$ remains strongly connected, then
$G[S]$ is a good component and is output in Line~\ref{l:stillconnected}.
This is the only case when Algorithm~\ref{alg:streett} outputs a subgraph
$G[S]$; thus if the algorithm outputs a component, it is a good component.
Algorithm~\ref{alg:streett} terminates if a good component is identified or
$Q{}$ is empty. Lemma~\ref{lemma:allcandidates} shows that before every
iteration of the outer while-loop \emph{every} good component is contained in
one of the subgraphs $G[S]$ in $Q{}$. That is, if a good component exists
in $G$, the algorithm will not terminate until a good component is identified.
Whenever the algorithm does not terminate in an iteration
of the outer while-loop, either (a) a trivial \textsc{scc}{} is removed from
$Q{}$~(Line~\ref{l:le2}) or (b) one of the subgraphs from $Q{}$ is
split into two smaller subgraphs~(Line~\ref{l:addRest}).
Each case can happen at most $n$ times.
This implies that the algorithm terminates
after a finite number of steps if no good component exists.
Next we show that if there exists a good component in $G$, then
the algorithm will output a good component.
Let~$Y$ be a {\em maximal} good component in~$G$
and let~$S_Y$ be the vertex set maintained in $Q{}$
that currently contains the vertices in~$Y$. By the arguments above after a finite number of
steps either (1) another good component is detected or (2) $\mathit{D}(S_Y)$ is pulled
from $Q{}$. By Lemma~\ref{lemma:allcandidates} $Y$ is never split
by the algorithm thus after Case~(2) happened at most $n$ times, one of
the following two cases occurs: either (2a) $\mathit{D}(S_Y)$ is pulled from $Q{}$
with $G[S_Y] \supset Y$ and after the removal of
bad vertices from $S_Y$, $G[S_Y]$ without the bad vertices is equal to $Y$ or (2b) $G[S_Y] = Y$ is pulled
from $Q{}$. In both cases the good component $Y$ is output and the algorithm
terminates: Since $Y$ is non-trivial, the condition in Line~\ref{l:le2} is satisfied.
The algorithm searches for a top or bottom \textsc{scc}{} in~$Y$. Since~$Y$ is
strongly connected, the only top or bottom \textsc{scc}{} in~$Y$ is~$Y$ itself. Hence
the algorithm outputs~$Y$ in Line~\ref{l:stillconnected}.
\end{proof}
Recall Algorithm~{\AlCapFnt{Streett}}{} that calls Algorithm~\ref{alg:streett} for each maximal \textsc{scc}{} in the input graph and then computes reachability to
the union of the identified good components. Lemmata~\ref{lemm:timeStreett}
and~\ref{lemm:correctStreett} yield the following result.
\begin{theorem}
Algorithm~{\AlCapFnt{Streett}}{} correctly computes the winning set in graphs with $k$-pair Streett
objectives in $O((n^2+b)\log n)$ time. Given a vertex $x$ in the winning set,
a certificate for $x$ can be output in time $O(m+ n \min(n,k))$.
\end{theorem}
\subsection{Preliminaries}\label{sec:pre}
Let $G[S]$ denote the subgraph of a graph $G = (V,E)$ induced
by the set of vertices $S \subseteq V$.
$\mathit{RevG}$ denotes the graph with vertices $V$ and all edges of $G$ reversed.
Let $\reach(S,G)$ be the
set of vertices in $G$ that \emph{can reach} a vertex in $S \subseteq V$.
A strongly connected component (\textsc{scc}{}) of a directed graph $G=(V,E)$ is a subgraph
$G[S]$ induced by a subset of vertices $S \subseteq V$ such that there
is a path in $G[S]$ between every pair of vertices in $S$.
We call an \textsc{scc}{} \emph{trivial} if it only contains a single vertex and no edges.
All other \textsc{scc}{s} are \emph{non-trivial}.
The set $\reach(S,G)$ and the maximal \textsc{scc} s of a graph $G$ can be found in
linear time~\cite{Beeri80,Immerman81,Tarjan1972}.
\smallskip\noindent{\em Algorithm~{\AlCapFnt{Streett}}{} and good component detection.}
The input is a directed graph $G=(V, E)$ and $k$ Streett
pairs $( L_j, U_j )$, $j = 1, \dotsc, k$.
The size of the input is measured in terms of $m=\lvert E \rvert$,
$n=\lvert V \rvert$, $k$, and $b = \sum_{j=1}^k (\lvert L_j \rvert + \lvert U_j \rvert)$.
Consider a maximal \textsc{scc}{} $C$; the \emph{good component detection
problem} asks to (a) output a non-trivial \textsc{scc}{} $G[X{}] \subseteq C$
such that for all $1 \le j \le k$ either
no vertex in $L_j$ or at least one vertex in $U_j$ is contained in the \textsc{scc}{}
(i.e., $L_j \cap X{}=\emptyset$ or $U_j \cap X{}\neq \emptyset$),
or (b) detect that no such \textsc{scc}{} exists.
In the former case, there exists an infinite path
that visits $X{}$ infinitely often and satisfies the Streett objective,
while in the later case there exists no infinite path
that visits vertices in $C$ infinitely often and satisfies the Streett objective.
It follows from the results of~\cite{AlurHenzingerBook} that the following algorithm,
called Algorithm~{\AlCapFnt{Streett}}{}, suffices for
the winning set computation: (1) Compute the maximal \textsc{scc}{} decomposition
of the graph; (2) for each maximal \textsc{scc}{} $C$ for which the good component detection
returns an \textsc{scc}{}, label the maximal \textsc{scc}{} $C$ as \emph{satisfying}; (3) output
the set of nodes that can reach a satisfying maximal \textsc{scc}{s} as the winning set. Since the first and last step are linear time, the runtime of
Algorithm~{\AlCapFnt{Streett}}{} is dominated by the detection of good components in maximal \textsc{scc}{s}.
In the following we assume that the input graph is strongly connected and focus
on good component detection.
\inappendix{
\smallskip\noindent\emph{Certificate computation.}
Given a start vertex $x$ that belongs to the winning set,
a certificate is an example of an
{\em accepting run}, i.e., an infinite path from $x$ that satisfies the objective.
The output of Algorithm~{\AlCapFnt{Streett}}{} can be used to construct such an accepting run.
Given a start vertex $x$ and a good component $G[X{}]$ reachable from $x$, we
generate the accepting run as follows. A path from $x$ to $X{}$ can be found
in linear time by a depth-first search.
Let $v$ be the vertex in $X{}$ where this path ends.
We call $v$ the \emph{root} of the \textsc{scc}{} $G[X{}]$. We show next how to obtain, in $O(m + n \min(n, k))$ time,
from the \textsc{scc}{} $G[X{}]$ a cycle starting and ending at the root $v$ such that the resulting
certificate is indeed an accepting run. For this it is sufficient that the cycle in
$G[X{}]$ contains for each $L_j$ with $L_j \cap X{} \ne \emptyset$ a vertex in
$U_j \cap X{}$, i.e., we do not have to include {\em all} vertices in $X{}$.
We can use Tarjan's depth-first search based \textsc{scc}{}
algorithm~\cite{Tarjan1972} to traverse the subgraph $G[X{}]$ in linear $O(m)$ time,
starting from root $v$. Tarjan's algorithm constructs a graph called
\emph{jungle} with $O(\lvert X{} \rvert)$ edges that for an
\textsc{scc}{} $G[X{}]$ consists of a spanning tree and at most one \emph{backedge} per
vertex in $X{}$. The vertices are assigned pre-order numbers in the order they are traversed.
We say an edge of $G[X{}]$ is a \emph{backedge} if it leads from a vertex with
a higher number to a vertex with a lower number. Spanning tree edges always lead
from lower numbered vertices to higher numbered vertices.
In Tarjan's algorithm a \emph{lowlink} is determined for each vertex $u$ which refers
to the lowest numbered vertex $w$ that $u$ can reach by a sequence of tree edges
followed by at most one backedge. We additionally store at each vertex $u \ne v$
a \emph{backlink} that is the first edge on the path from $u$ to its lowlink.
The backlinks can be determined and stored during the depth-first search
without increasing its running time.
With this data structure we can find within $G[X{}]$ a path from root $v$ to a vertex
$u \in X{}$, $u \ne v$, and back by first searching for $u$ in
the spanning tree and then following the backlinks back to $v$.
Since no vertex will appear more than twice on this path, its size and the
time to compute it is $O(\lvert X{} \rvert)$.
As it suffices to find such paths for one vertex per nonempty set $U_j \cap X{}$, we
can generate a certificate from $G[X{}]$
in $O(m + \lvert X{} \rvert \min(\lvert X{} \rvert, \lvert\{j \mid
U_j \cap X{} \ne \emptyset \}\rvert))$ time, which can be bounded with
$O(m + n \min(n, k))$. This certificate has a size of $O(n \min(n, k))$.
As shown in Figure~\ref{fig:certificate},
the smallest existing certificate can be as large as $\Theta(n \min(n, k))$.
\begin{figure}
\centering
\input{exampletarjan}
\caption{An example for a ``jungle'' constructed by Tarjan's \textsc{scc}{} algorithm
for an \textsc{scc}{}.
Backedges are dotted, spanning tree edges are solid. Backlinks are marked with a dot.
The numbers of the vertices represent the order in which the vertices are visited,
the numbers in brackets are the lowlinks.}\label{fig:tarjan}
\end{figure}
\begin{figure}
\centering
\input{certificateworstcase}
\caption{Let the only path between $s$ and $t$ be of length $\Theta(n/2) = \Theta(n)$,
not containing any of the vertices $v_j$ for $1 \le j \le k$.
Let the Streett pairs $( L_j, U_j )$ be given by $L_j = \{s\}$ and
$U_j = \{v_j\}$ for $1 \le j \le k$.
For this example the size of the smallest certificate is $\Theta(nk)$, where
$k$ can be of order $\Theta(n)$.}\label{fig:certificate}
\end{figure}
}
\inmain{For a vertex $x$ in the winning set a \emph{certificate} is
an example of an infinite path from $x$ that satisfies the objective.
In the appendix we provide an algorithm that computes a certificate for $x$ in
time $O(m + n \min(n,k))$ (given a good component reachable from $x$)
and show an example where the certificate has size $\Theta(n \min(n,k))$.}
Next, we introduce the different concepts used in the algorithm for good component detection.
\inappendix{
First we describe the hierarchical graph decomposition
technique for this setting. This decomposition will be crucial for the runtime
analysis.
}
\smallskip\noindent\emph{Graph decomposition.}
In our algorithm we decompose a graph $G$ in the following way.
Let $G_i = (V, E_i)$ be a subgraph of $G$ with $E_i = \{(u,v) \mid \OutDeg(u)
\le 2^i \}$, i.e., the edges of $G_i$ are the outedges of the vertices with
outdegree at most $2^i $.
Note that for $i = \log n$ we have that $G_i = G$.
We say vertices in $G$ with $\OutDeg(v) > 2^i $ are \emph{colored blue} in $G_i$
and denote the set of blue vertices in $G_i$ by $\mathit{Bl}_i$.
All other vertices are \emph{white}.
Note that all vertices in $G=G_{\log n}$ are white and that all vertices in
$\mathit{Bl}_i$ have outdegree zero in $G_i$.
\smallskip\noindent\emph{Top and bottom strongly connected components.}
The algorithm will repeatedly find a top or a bottom \textsc{scc}{} in the remaining
graph $G$. A bottom \textsc{scc}{} $G[S]$ in a directed graph $G$ is an
\textsc{scc}{} with no edges from vertices in $S$ to vertices
in $V \setminus S$, i.e., no \emph{outgoing} edges.
A top \textsc{scc}{} is a bottom \textsc{scc}{} of $\mathit{RevG}$, i.e., an \textsc{scc}{} without \emph{incoming}
edges. Top and bottom \textsc{scc} s are by definition maximal \textsc{scc} s.
Note that every graph has at least one bottom and at least one top \textsc{scc}{}.
If they are not the same, then they are disjoint and thus one of them contains at
most half of the vertices of $G$.
\smallskip\noindent\emph{Lock-step search.}
The lock-step search of Even and Shiloach~\cite{ES1981} was
already applied (in a different way) to the detection of good components in~\cite{HT1996}.
Lock-step search simulates parallel execution by alternatingly making one step
in each parallel instance. The execution finishes as soon as the first instance
terminates.
In our algorithm we use this technique to search in parallel
for the smallest top or bottom \textsc{scc}{}, where the search for a top
\textsc{scc}{} is performed by searching for a bottom \textsc{scc}{} in $\mathit{RevG}$.
\smallskip\noindent\emph{Bad vertices.}
In contrast to \emph{good components} we also define \emph{bad vertices}.
The basic idea behind the algorithms for good component detection, described for example
in~\cite{HT1996}, is to repeatedly delete \textit{bad} vertices until either a
good component is found or it can be concluded that no such component exists.
A vertex is \emph{bad} if for some index~$j$ with
$1 \le j \le k$ the vertex is in~$L_j$ but it is not strongly connected
to any vertex in~$U_j$. All other vertices are \emph{good}.
Note that good vertices can become bad if some vertex deletion disconnects
an \textsc{scc}{} or a vertex in a set $U_j$ is deleted. A good component is then
a non-trivial \textsc{scc}{} that only contains good vertices.
\smallskip\noindent\emph{Data structure.}
The algorithm maintains for the current graph $G=(V,E)$ (some vertices of the input
graph might have been deleted) a decomposition into vertex sets $S\subseteq V$ such
that every \textsc{scc}{} of $G$
is completely contained in $G[S]$ for one of the sets $S$. For all the sets $S$ a
data structure $\mathit{D}(S)$ is saved in a list $Q{}$.
The data structure $\mathit{D}(S)$ supports the following operations:
(1) $\construct(S)$ initializes the data structure for the set $S$,
(2) $\remove(S, \mathit{D}(S), B)$ removes a set $B \subseteq V$ from $S$ and updates
the data structure of $S$ accordingly, and
(3) $\bad(\mathit{D}(S))$ returns the set $\{v \in S \mid
\exists j \text{ with } v \in L_j \text{ and } U_j \cap S = \emptyset \}$.
In~\cite{HT1996} an implementation of this data structure was given that achieves
the following running times. For a set of vertices $S \subseteq V$ let $\mathit{bits}(S)$
be defined as $\sum_{j=1}^{k} \left( \lvert S \cap L_j \rvert + \lvert S \cap U_j \rvert\right)$.
\begin{lemma}[Lemma 2.1 in \cite{HT1996}]\label{lemma:ds}
After a one-time preprocessing of time $O(k)$, the data structure $\mathit{D}(S)$ can
be implemented in time $O(\mathit{bits}(S)+\lvert S \rvert)$
for $\construct(S)$, time $O(\mathit{bits}(B)+\lvert B \rvert)$ for
$\remove(S, \mathit{D}(S), B)$, and constant running time for $\bad(\mathit{D}(S))$.
\end{lemma}
|
\section{Introduction}
Chain conditions appear frequently in the study of countable groups. These are finiteness conditions that forbid certain infinite sequences of subgroups. An elementary but interesting example of such a condition is the property of being polycyclic. From a geometric group theory perspective, these finiteness conditions ought to restrict the complexity of the groups, as in the case of polycyclic groups. From a descriptive set theory perspective, however, the chain conditions are non-Borel co-analytic statements and, therefore, either admit ``nice" non-chain-condition characterizations - e.g.\ polycyclic groups are soluble with each term of the derived series finitely generated - or describe large and wild classes. In this work, we explore this tension in four chain conditions in the space of marked groups.\par
\indent In the space of marked groups, denoted $\mathscr{G}$, we first consider three well-known chain conditions: the minimal condition on centralizers, the maximal condition on subgroups, and the maximal condition on normal subgroups. We characterize each of these in terms of well-founded descriptive-set-theoretic trees. This characterization implies the classes in question are large and wild, whereby they do not admit ``nice" characterizations.
\begin{thm}
Each of the subsets of $\mathscr{G}$ defined by the minimal condition on centralizers, the maximal condition on subgroups, and the maximal condition on normal subgroups are co-analytic and not Borel. This remains true when restricting to finitely generated groups.
\end{thm}
Our techniques additionally give new ordinal-valued isomorphism invariants unbounded below the first uncountable ordinal in the cases of the minimal condition on centralizers and the maximal condition on subgroups. The ordinal-valued isomorphism invariant we obtain in the case of the maximal condition on normal subgroups is not new and has been considered in the literature; cf. \cite{C11}. However, our approach is new, and we show that this invariant is unbounded below the first uncountable ordinal.\par
\indent We next consider the set of elementary amenable marked groups. We likewise characterize these in terms of descriptive-set-theoretic trees. It follows that elementary amenability is indeed a chain condition.
\begin{thm}
A countable group $G$ is elementary amenable if and only if there is no infinite descending sequence of the form
$$G=G_0\geq G_1\geq\ldots\geq G_n \geq \ldots $$
such that for all $n\geq 0$, $G_n\neq\{e\}$ and there is a finitely generated subgroup $K_n\leq G_n$ with $G_{n+1}= [K_n,K_n]\cap H_n$, where $H_n$ is the intersection of the index-$(\leq(n+1))$ normal subgroups of $K_n$.
\end{thm}
Our characterization gives two new invariants of elementary amenable groups: the decomposition rank and decomposition degree. We further obtain
\begin{thm}
The sets of elementary amenable groups and finitely generated elementary amenable groups are co-analytic and non-Borel in the space of marked groups.
\end{thm}
It is well-known that the set of amenable groups is Borel in the space of marked groups. Our theorem thus gives a non-constructive answer to an old question of M. Day \cite{D57}, which was open until R. I. Grigorchuk \cite{G84} constructed groups of intermediate growth: \textit{Are all finitely generated amenable groups elementary amenable?}
\begin{cor}
There is a finitely generated amenable group that is not elementary amenable.
\end{cor}
The paper is organized as follows. In Section \ref{sec:Prelim}, we discuss the basic properties of $\mathscr{G}$ and introduce concepts from descriptive set theory. In Sections \ref{sec:MinCent},\ref{sec:Max}, and \ref{sec:MaxN}, we analyze sets of groups satisfying various chain conditions. This introduces our use of descriptive-set-theoretic trees to study the structure of groups as well as the ordinal-valued invariants arising from those trees. In Section \ref{sec:EAGroups}, we use those same techniques to analyze elementary amenable groups. In Section \ref{sec:Borel}, we prove the maps used throughout the paper are indeed Borel. Those who are content to believe that our constructions are Borel can safely skip this section without missing any group-theoretic content. Finally, Section \ref{sec:Remarks} discusses some questions arising from this paper not touched upon in earlier sections.
\section{Preliminaries}\label{sec:Prelim}
\subsection{The space of marked groups}
In order to apply the techniques of descriptive set theory to groups, we need an appropriate space of groups. Let $\mathbb{F}_{\omega}$ be the free group on the letters $\{a_i\}_{i\in\mathbb{N}}$; so $\mathbb{F}_{\omega}$ is a free group on countably many generators with a distinguished set of generators. The power set of $\mathbb{F}_{\omega}$ may be naturally identified with the Cantor space $\{0,1\}^{\mathbb{F}_{\omega}}=:2^{\mathbb{F}_{\omega}}$. It is easy to check the collection of normal subgroups of $\mathbb{F}_{\omega}$, denoted $\mathscr{G}$, is a closed subset of $2^{\mathbb{F}_{\omega}}$ and, hence, a compact Polish space. Each $N\in\mathscr{G}$ is identified with a \textbf{marked group}. That is the group $G=\mathbb{F}_{\omega}/N$ along with a distinguished generating set $\{f_N(a_i)\}_{i\in \mathbb{N}}$ where $f_N:\mathbb{F}_{\omega}\rightarrow G$ is the usual projection; we always denote this projection by $f_N$. For a marked group $G$, we abuse notation and say $G\in \mathscr{G}$; of course, we formally mean $G=\mathbb{F}_{\omega}/N$ for some $N\in \mathscr{G}$. Since every countable group is a quotient of $\mathbb{F}_{\omega}$, $\mathscr{G}$ gives a compact Polish space of all countable groups. A sub-basis for this topology is given by sets of the form
$$ O_{\gamma} := \left\{N\in\mathscr{G} \mid \gamma\in N\right\}, $$
where $\gamma\in\mathbb{F}_{\omega}$ along with their complements.
\indent Similar reasoning leads us to define the space of \textbf{$m$-generated marked groups} as
\[
\mathscr{G}_m := \bigcap_{i\geq m}\{N \trianglelefteq \mathbb{F}_{\omega} \mid a_i\in N\}.
\]
This is a closed subset of $\mathscr{G}$ and so is a compact Polish space in its own right. We further let $\mathscr{G}_{fg}:=\cup_{m\geq 1} \mathscr{G}_m$ be the space of finitely generated marked groups. As this is an $F_\sigma$ subset of $\mathscr{G}$, it is a standard Borel space, with Borel sets precisely those sets of the form $\mathscr{G}_{fg}\cap B$ with $B$ Borel in $\mathscr{G}$; a \textbf{standard Borel space} is a Borel space which admits a Polish topology that induces the Borel structure. We can thus also talk about Borel functions with domain $\mathscr{G}_{fg}$.\par
\indent It is convenient to give the marked groups $G=\mathbb{F}_{\omega}/N$ a preferred enumeration. To this end, we fix an enumeration $\bfs{\gamma}:=(\gamma_i)_{i\in \mathbb{N}}$ of $\mathbb{F}_{\omega}$. Each $G$ is thus taken to come with an enumeration $f_N(\bfs{\gamma}):=(f_N(\gamma_i))_{i\in \mathbb{N}}$; note the enumeration of $G$ may have many repetitions. When we write $G$ as $G=\{g_0,g_1,\ldots\}$, we will always mean this enumeration. Later in the paper we will work with $\mathbb{N}^{<\mathbb{N}}$, i.e. the set of finite sequences of natural numbers. If $(s_0,\ldots,s_n)=:s\in\mathbb{N}^{<\mathbb{N}}$, we will write $\{g_s\}$ for the set $\{g_{s_0},\ldots,g_{s_n}\}$. Note that this set may have fewer than $n+1$ elements, e.g.\ if $s_0=s_1=\ldots=s_n$, or even if the $s_i$ are distinct but enumerate the same element.\par
\indent We will often discuss quotients of groups or particular subgroups of groups, and of course we wish to view these as elements of $\mathscr{G}$. A quotient of a marked group is obviously again a marked group. However, subgroups of marked groups do not have an obvious marking. The enumeration gives us a preferred way to select markings for subgroups. If $H\leq \mathbb{F}_{\omega}/N=G\in \mathscr{G}$, let $\pi_H\colon\mathbb{F}_{\omega}\to\mathbb{F}_{\omega}$ be induced by mapping the generators $(a_i)_{i\in \mathbb{N}}$ of $\mathbb{F}_{\omega}$ as follows:
\[
\pi_H(a_j):=
\begin{cases}
\gamma_j, & \text{ if }f_N(\gamma_j)\in H \\
e, & \text{ else.}
\end{cases}
\]
We then identify $H$ with $\mathbb{F}_{\omega}/\ker (f_N\circ\pi_H)$. In the case $H$ has a distinguished finite generating set $\{g_{i_0},\dots,g_{i_n}\}$, we instead define $\pi_H(a_{i_j})=\gamma_{i_j}$ and $\pi_H(a_j)=e$ for $j\neq i_k$; this streamlines our proofs later. We often appeal to this convention implicitly.\par
\indent We will consider maps from and on $\mathscr{G}$. A slogan from descriptive set theory is ``Borel = explicit'' meaning if you describe a map ``explicitly'', i.e. without an appeal to something like the axiom of choice, it should be Borel. All of the maps we discuss in the next few sections will be ``explicit'' in this sense, so we will not prove they are Borel when we define them, in order to keep the focus on the group-theoretic aspects of our constructions. We will often use enumerations of groups in our constructions, but this will not require choice since every marked group comes with a preferred enumeration. For those who are interested in the details, we discuss the descriptive-set-theoretic aspects of our constructions in Section \ref{sec:Borel}.
\subsection{Descriptive set theory}
\indent We are interested in certain types of non-Borel subsets of $\mathscr{G}$. The following definitions and theorems are all fundamental in descriptive set theory; a standard reference is \cite{K95}.
\begin{defn}
Let $X,Y$ be uncountable standard Borel spaces. Then $A\subseteq Y$ is \textbf{analytic} (denoted $\Sigma^1_1$) if there is a Borel set $B \subseteq X\times Y$ such that $\operatorname{proj}_Y(B)=A$. A set $C\subseteq Y$ is \textbf{co-analytic} (denoted $\Pi^1_1$) if $Y\setminus C$ is analytic.
\end{defn}
Every Borel set is analytic, but any uncountable standard Borel space contains non-Borel analytic sets. It follows that there are non-Borel co-analytic sets. The collection of analytic sets is closed under countable unions, countable intersections, and Borel preimages. It follows the collection of co-analytic sets is closed under countable unions, countable intersections, and Borel preimages. We remark that sets defined using a single existential quantifier which ranges over an uncountable standard Borel space are often analytic as such quantification can typically be rewritten as a projection of a Borel set. Thus sets defined by using a universal quantifier over an uncountable set are often co-analytic.
\begin{defn}
Let $X,Y$ be standard Borel spaces, and $A\subseteq X$, $B\subseteq Y$. We say that $A$ \textbf{Borel reduces} to $B$ if there is a Borel map $f\colon X\to Y$ such that $f^{-1}(B)=A$.
\end{defn}
If $A$ Borel reduces to $B$ and $B$ is Borel, analytic, or co-analytic, then so is $A$. This gives us a method for proving that sets are, for example, co-analytic simply by showing they Borel reduce to a co-analytic set. One important example comes from the space of (descriptive-set-theoretic) trees.
\begin{defn}
A set $T\subseteq\mathbb{N}^{<\mathbb{N}}$ of finite sequences of natural numbers is a \textbf{tree} if it is closed under initial segments. A sequence $x\in\mathbb{N}^\mathbb{N}$ is a branch of $T$ if for all $n\in\mathbb{N}$, $x\restriction n\in T$. For $s\in T$, $T_s:=\{r\in \mathbb{N}^{<\mathbb{N}}\mid s^{\smallfrown} r\in T\}$ where ``$^{\smallfrown}$" indicates concatenation of finite sequences.
\end{defn}
As with groups, we may identify $X\subseteq\mathbb{N}^{<\mathbb{N}}$ with an element $f_X\in 2^{\mathbb{N}^{<\mathbb{N}}}$. We define
\[
Tr := \{ x\in 2^{\mathbb{N}^{<\mathbb{N}}} \mid x \text{ is a tree }\}.
\]
The set $Tr$ is a closed subset of $2^{\mathbb{N}^{<\mathbb{N}}}$ and so is a compact Polish space. A sub-basis for the topology on $Tr$ is given by sets of the form
$$ O_t := \left\{T\in Tr \mid t\in T\right\}, $$
where $t\in\mathbb{N}^{<\mathbb{N}}$ along with their complements.
\indent There are two subsets of $Tr$ of particular interest to us:
\[
IF := \{ T\in Tr \mid T \text{ has a branch } \}
\]
and $WF := Tr\setminus IF$. We call $WF$ the set of \textbf{well-founded} trees and $IF$ the set of \textbf{ill-founded} trees. One can check that $IF$ is analytic, so $WF$ is co-analytic. The importance of these sets comes from the following fact.
\begin{thm}\cite[Theorem 27.1]{K95}\label{thm:WFComplete}
Every analytic set Borel reduces to $IF$. Therefore, every co-analytic set Borel reduces to $WF$.
\end{thm}
\noindent Thus a set $A$ is co-analytic if and only if it Borel reduces to $WF$.\par
\indent We are interested in $WF$ for a second reason. Let $ORD$ denote the class of ordinals. For any $T\in WF$, we can define a function $\rho_T\colon T\to ORD$ inductively as follows: If $t\in T$ has no extensions in $T$, let $\rho_T(t)=0$. Otherwise let $\rho_T(t) = \sup\{\rho_T(s)+1 \mid t\subsetneq s \}$. We may then define a rank function $\rho\colon Tr\to ORD$ by
\[
\rho(T) =
\begin{cases}
\rho_T(\emptyset)+1, & \text{if }T\in WF\\
\omega_1, & \text{else.}
\end{cases}
\]
For $T=\emptyset$, we define $\rho(T)=0$. The function $\rho$ is bounded above by $\omega_1$, the first uncountable ordinal. Furthermore, this rank function has a special property:
\begin{defn}\label{def:PiRank}
Let $X$ be a standard Borel space and $A\subseteq X$. A function $\phi\colon A\to ORD$ is a \textbf{$\Pi^1_1$-rank} if there are relations $\leq_\phi^\Pi$, $\leq_\phi^\Sigma\subseteq X\times X$ such that $\leq_\phi^\Pi$ is co-analytic, $\leq_\phi^\Sigma$ is analytic, and for all $y\in A$,
\begin{align*}
x\in A \wedge \phi(x)\leq \phi(y) &\Leftrightarrow x \leq_\phi^\Sigma y \\
&\Leftrightarrow x \leq_\phi^\Pi y.
\end{align*}
\end{defn}
Given any rank function on $A$, one may use it to define an order $\leq_\phi$ on $A$. The idea of the above definition is that if $\phi$ is a $\Pi^1_1$-rank, then the initial segments of $\leq_\phi$ are Borel, and this is witnessed in a uniform way.
\begin{thm}\cite[Exercise 34.6]{K95}
The function $\rho\colon WF\to ORD$ is a $\Pi_1^1$-rank.
\end{thm}
\noindent We may use this fact to create other $\Pi_1^1$-ranks in an easy way: Let $X$ be a standard Borel space. If $A\subseteq X$ Borel reduces to $WF$ via $f$, then the map $x \mapsto \rho(f(x))$ is a $\Pi_1^1$-rank. \par
\indent The most important fact about $\Pi_1^1$-ranks for this paper is the following (\cite[Theorem 35.23]{K95}):
\begin{thm}[The Boundedness Theorem for $\Pi_1^1$-ranks]\label{thm:BddnessThm}
Let $X$ be a standard Borel space, $A\subseteq X$ co-analytic, and $\phi\colon A\to\omega_1$ a $\Pi_1^1$-rank. Then
$$A \text{ is Borel} \;\Longleftrightarrow\; \sup \{ \phi(x) \mid x\in A \} < \omega_1.$$
\end{thm}
We will use the Boundedness Theorem to show that certain $\Pi_1^1$ sets are not Borel, by showing that they come with $\Pi^1_1$-ranks with images unbounded below $\omega_1$. To this end, we will often use the following fact about the ranks of trees, which follows immediately from the definition.
\begin{lem}\label{lem:TrRkMonotone}
Suppose $S,T$ are trees and $\phi\colon S\to T$ is a map such that $s\subsetneq t \Rightarrow \phi(s) \subsetneq \phi(t)$. (We call such a map \textbf{monotone}.) Then $\rho_S(s)\leq \rho_T(\phi(s))$ for all $s\in S$. In particular $\rho(S)\leq\rho(T)$.
\end{lem}
\section{The minimal condition on centralizers}\label{sec:MinCent}
We wish to show certain chain conditions give rise to sets of marked groups which are $\Pi^1_1$ and not Borel in $\mathscr{G}$. We begin by looking at the following chain condition.
\begin{defn}
A subgroup $H$ of $G$ is a \textbf{centralizer} in $G$ if $H=C_G(A)$ for some $A\subseteq G$.
\end{defn}
\begin{defn}
A group $G$ satisfies the \textbf{minimal condition on centralizers} if there is no strictly decreasing infinite chain $C_0 > C_1 > \ldots$ of centralizers in $G$. We denote the class of countable groups satisfying the minimal condition on centralizers by $\mc M_C$.
\end{defn}
The class $\mc M_C$ is large, containing abelian groups, linear groups, and finitely generated abelian-by-nilpotent groups; see \cite{Br79} for further discussion. It is not hard to check that a group $G$ satisfies the minimal condition on centralizers if and only if it satisfies the maximal condition on centralizers, but our analysis is easier if we think about the minimal version of the chain condition.
Given a group $G\in\mathscr{G}$, we construct a tree $T_G\subseteq\mathbb{N}^{<\mathbb{N}}$ and associated groups $G_s\in \mathscr{G}$ for each $s\in T_G$. Each $G_s$ will be a centralizer in $G$.
\begin{enumerate}[$\bullet$]
\item Put $\emptyset\in T_G$ and let $G_\emptyset:=G=C_G(\emptyset)$.
\item Suppose that $s\in T_G$ and $G_s=C_G(\{g_s\})$ has already been defined. If $C_G(\{g_s\}\cup\{g_i\})\neq C_G(\{g_s\})$, then put $s^{\smallfrown} i\in T_G$ and $G_{s^{\smallfrown} i}:=C_G(\{g_s\}\cup\{g_i\})$.
\end{enumerate}
\begin{lem}\label{lem:CentMapBorel}
The map $\Phi_C\colon\mathscr{G}\to Tr$ given by $G\mapsto T_G$ is Borel.
\end{lem}
\noindent Intuitively, Lemma~\rm\ref{lem:CentMapBorel} holds since our construction is explicit; we delay a rigorous proof until Section \ref{sec:Borel}.
\begin{lem}\label{lem:CentMapReduction}
$T_G$ is well-founded if and only if $G\in\mc M_C$.
\end{lem}
\begin{proof}
If $G\in\mc M_C$, then $T_G$ contains no infinite branches by definition. If $G\notin\mc M_C$, then there is some infinite $A\subseteq G$ such that for all finite $B\subseteq A$, $C_G(A)\neq C_G(B)$. Let $a_0<a_1<a_2<\ldots$ be such that $A=\{g_{a_0},g_{a_1},\ldots\}$. By moving to a subsequence if necessary, we may assume that $C_G(\{g_{a_0},\ldots,g_{a_n}\}) \gneq C_G(\{g_{a_0},\ldots, g_{a_{n+1}}\})$ for all $n\in\mathbb{N}$. Then $(a_0,\ldots,a_n)\in T_G$ for all $n\in\mathbb{N}$, so $T_G$ has an infinite branch.
\end{proof}
\begin{lem}\label{lem:CentSubRank}
Let $H,G\in\mathscr{G}$. If $H\hookrightarrow G$, then $\rho(T_H)\leq\rho(T_G)$.
\end{lem}
\begin{proof}
Let $\alpha\colon H\hookrightarrow G$ and let $\psi\colon\mathbb{N}\to\mathbb{N}$ be such that $\alpha(h_k)=g_{\psi(k)}$. We now define a map $\phi:T_H\rightarrow \mathbb{N}^{<\mathbb{N}}$: Let $\phi(\emptyset)=\emptyset$. If $s\in T_H$ and $s=(s_0,\ldots,s_n)$, let $\phi(s)=(\psi(s_0),\ldots,\psi(s_n))$. Clearly $\phi$ is monotone. Further, if $s\in T_H$, then $H_s = C_H(\{h_s\}) \hookrightarrow C_G(\{g_{\phi(s)}\})$. Since $C_G(\{g_{\phi(s)}\})\cap \alpha(H) \cong C_H(\{h_s\})$, we have that $C_G(\{g_{\phi(s\restriction k)}\}) \neq C_G(\{g_{\phi(s\restriction (k+1))}\})$ for all $k<|s|$. Thus $\phi(s)\in T_G$. It follows $\phi(T_H)\subseteq T_G$, and by Lemma \ref{lem:TrRkMonotone}, $\rho(T_H)\leq\rho(T_G)$.
\end{proof}
\begin{cor}\label{cor:CentIsoInv}
If $G,G'\in\mathscr{G}$ and $G\cong G'$, then $\rho(T_G)=\rho(T_{G'})$.
\end{cor}
We thus see that $\rho(T_G)$ is an isomorphism invariant, so it makes sense to talk about the rank of a group $G$ with the minimal condition on centralizers, even when not considering a specific marking.
\begin{defn}
If $G$ has the minimal condition on centralizers, then $\rho(T_G)$ for some (any) marking of $G$ is called the \textbf{centralizer rank} of $G$.
\end{defn}
We also mention that the above results, except for Lemma \ref{lem:CentMapBorel}, work with arbitrary enumerations of the group $G$, not just those that can arise from viewing $G$ as a marked group. Certain enumerations may be easier to use to calculate $\rho(T_G)$, and Corollary \ref{cor:CentIsoInv} assures us that using these enumerations will not affect the answer. The same will be true of our later constructions. Of course, in this paper Lemma \ref{lem:CentMapBorel} and analogous results are of central importance, so we will continue to work with groups as elements of $\mathscr{G}$.
We now argue the centralizer rank is unbounded below $\omega_1$.
\begin{lem}\label{lem:CentRankSucc}
For $A,B\in \mc{M}_C$ with $A$ nonabelian, $A\times B\in \mc{M}_C$ and $\rho(T_B)<\rho(T_{A\times B})$.
\end{lem}
\begin{proof}
It is easy to see $A\times B\in \mc{M}_C$. Let $a\in A$ be noncentral. Then
\[
C_{A\times B}(\{(a,e)\})=(A\times B)_i
\]
for some $i\in T_{A\times B}$ since the centralizer is not all of $A\times B$. Further,
$$B\cong\{e\}\times B\leq C_{A\times B}(\{(a,e)\}),$$
so by Lemma \ref{lem:CentSubRank}, $\rho((T_{A\times B})_i)=\rho(T_{(A\times B)_i})\geq\rho(T_B)$. The result now follows.
\end{proof}
\begin{lem}\label{lem:CentRankLim}
Let $\{A_i\}_{i\in\mathbb{N}}$ be countable groups. If $A_i\in\mc M_C$ for all $i\in\mathbb{N}$, then there is a group $A\in \mc M_C$ such that $\rho(T_A)\geq\rho(T_{A_i})$ for all $i\in\mathbb{N}$.
\end{lem}
\begin{proof}
Let $A=\ast_{i\in\mathbb{N}} A_i$. By \cite[Corollary 4.1.6]{MKS66}, which says that centralizers in free products are cyclic or centralizers of a conjugate of a free factor, we infer that $A\in\mc M_C$. Lemma \ref{lem:CentSubRank} now implies that $\rho(T_A)\geq\rho(T_{A_i})$ for all $i\in\mathbb{N}$, as desired.
\end{proof}
\begin{lem}\label{lem:CentRankUnbdd}
For all $\alpha<\omega_1$, there is $G\in\mc M_C$ such that $\rho(T_G)\geq\alpha$.
\end{lem}
\begin{proof}
We prove this inductively. Clearly the lemma holds for $\alpha=0$. Suppose $\alpha=\beta+1$ and the lemma holds for $\beta$. Let $G\in\mc M_C$ be such that $\rho(T_G)\geq\beta$ and $A\in\mc M_C$ be nonabelian. Applying Lemma~\rm\ref{lem:CentRankSucc}, we see $\rho(T_{A\times G})\geq\beta+1$.
Suppose $\alpha$ is a limit ordinal. Since $\alpha$ is countable, there is a countable increasing sequence of $\alpha_i<\alpha$ such that $\sup_{i\in\mathbb{N}} \alpha_i =\alpha$. Let $G_i\in\mc M_C$ be such that $\rho(T_{G_i})>\alpha_i$. Applying Lemma~\rm\ref{lem:CentRankLim}, there is some $G\in\mc M_C$ such that $\rho(T_G)>\alpha_i$ for all $i\in\mathbb{N}$. It now follows that $\rho(T_G)\geq\alpha$.
\end{proof}
\begin{lem}\label{lem:CentFG}
For all $\alpha<\omega_1$, there is a finitely generated $G\in\mc M_C$ such that $\rho(T_G)\geq\alpha$.
\end{lem}
\begin{proof}
Let $H\in\mc M_C$ be a group such that $\rho(T_H)\geq\alpha$. Then \cite[Corollary on pg. 949]{KS71} implies that $H$ embeds into a 3-generated group $G\in\mc M_C$. By Lemma \ref{lem:CentSubRank}, $\rho(T_G)\geq\rho(T_H)\geq\alpha$ verifying the lemma.
\end{proof}
We remark that the proof of the result cited in the previous uses nothing more complicated than free products with amalgamation and is similar to the classical Higman-Neumann-Neumann embedding result \cite{HNN49}.
\begin{thm}\label{thm:MCNotBorel}
$\mc M_C$ is $\Pi^1_1$ and not Borel in $\mathscr{G}$, and $\mc M_C\cap\mathscr{G}_{fg}$ is $\Pi^1_1$ and not Borel in $\mathscr{G}_{fg}$.
\end{thm}
\begin{proof}
Let $\Phi_C$ be the Borel map from Lemma~\rm\ref{lem:CentMapBorel}. By Lemma \ref{lem:CentMapReduction}, $\Phi_C^{-1}(WF)=\mc M_C$, and since $\Phi_C$ is Borel, $\mc M_C$ is $\Pi^1_1$. Lemma \ref{lem:CentRankUnbdd} implies the ranks of the trees in $\Phi_C(\mc M_C)$ are unbounded below $\omega_1$, so the $\Pi_1^1$-rank on $\mc M_C$ given by $G\mapsto \rho(\Phi_C(G))$ is unbounded below $\omega_1$. By Theorem \ref{thm:BddnessThm}, we conclude that $\mc M_C$ is not Borel. Lemma \ref{lem:CentFG} implies the ranks of the trees in $\Phi_C(\mc M_C\cap\mathscr{G}_{fg})$ are also unbounded below $\omega_1$, and by Theorem \ref{thm:BddnessThm}, we conclude that $\mc M_C\cap\mathscr{G}_{fg}$ is also not Borel.
\end{proof}
\section{The maximal condition on subgroups}\label{sec:Max}
We next consider a more basic chain condition. Proving the analogue of Lemma \ref{lem:CentRankLim} in this context is more complicated, which is why we present it after the previous section.
\begin{defn}
A group $G$ satisfies the \textbf{maximal condition on subgroups}, abbreviated by saying a group satisfies max, if there is no strictly increasing chain $H_0 < H_1 < H_2 < \ldots$ of subgroups of $G$. Equivalently, a group $G$ satisfies max if all of its subgroups are finitely generated. We denote the class of groups satisfying max as $\mc M_{\max{}}$.
\end{defn}
Given a group $G\in \mathscr{G}$, we construct a tree $T_G\subseteq\mathbb{N}^{<\mathbb{N}}$ and associated groups $G_s\in \mathscr{G}$ for each $s\in T_G$.
\begin{enumerate}[$\bullet$]
\item Put $\emptyset\in T_G$ and let $G_\emptyset:=\{e\}$.
\item Suppose that $s\in T_G$ and $G_s=\<\{g_s\}\>$ has already been defined. If $\<\{g_s\}\cup\{g_i\}\>\neq\<\{g_s\}\>$, then put $s^{\smallfrown} i\in T_G$ and $G_{s^{\smallfrown} i}:=\<\{g_s\}\cup\{g_i\}\>$.
\end{enumerate}
\begin{lem}\label{lem:MaxMapBorel}
The map $\Phi_M\colon\mathscr{G}\to Tr$ given by $G\mapsto T_G$ is Borel.
\end{lem}
We will prove Lemma~\ref{lem:MaxMapBorel} in Section \ref{sec:Borel}.
\begin{lem}\label{lem:MaxMapReduction}
$T_G$ is well-founded if and only if $G\in \mc M_{\max{}}$.
\end{lem}
\begin{proof}
If $G\in \mc M_{\max{}}$, then $T_G$ contains no infinite branches by definition. If $G\notin \mc M_{\max{}}$, then there is some infinitely generated subgroup $H\leq G$. There is some increasing sequence $a_0<a_1<\ldots$ of natural numbers such that $H=\<g_{a_0},g_{a_1},\ldots\>$. We may assume that $\<g_{a_0},\ldots,g_{a_n}\> \lneq \<g_{a_0},\ldots,g_{a_{n+1}}\>$ for all $n\in\mathbb{N}$. Then $(a_0,\ldots,a_n)\in T_G$ for all $n\in\mathbb{N}$, so $T_G$ has an infinite branch.
\end{proof}
\begin{lem}\label{lem:MaxSubRank}
Let $H,G\in\mathscr{G}$. If $H\hookrightarrow G$, then $\rho(T_H)\leq\rho(T_G)$.
\end{lem}
\begin{proof}
Let $\psi\colon\mathbb{N}\to\mathbb{N}$ be such that $h_k=g_{\psi(k)}$. We now define a map $\phi:T_H\rightarrow \mathbb{N}^{<\mathbb{N}}$: Let $\phi(\emptyset)=\emptyset$. If $s\in\mathbb{N}^{<\mathbb{N}}$ and $s=(s_0,\ldots,s_n)$, let $\phi(s)=(\psi(s_0),\ldots,\psi(s_n))$. Clearly $\phi$ is monotone. Furthermore, if $s\in T_H$, then $H_s \cong G_{\phi(s)}$, hence $\phi(T_H)\subseteq T_G$. Lemma \ref{lem:TrRkMonotone} now implies $\rho(T_H)\leq\rho(T_G)$.
\end{proof}
The previous lemma implies $\rho(T_G)$ is a group invariant.
\begin{cor}
If $G,G'\in\mathscr{G}$ and $G\cong G'$, then $\rho(T_G)=\rho(T_{G'})$.
\end{cor}
\begin{defn}
If $G$ has the maximal condition on subgroups, then $\rho(T_G)$ for some (any) marking of $G$ is called the \textbf{subgroup rank} of $G$.
\end{defn}
\begin{lem}\label{lem:MaxRankSucc}
For all groups $G\in \mc M_{\max{}}$, $G\times\mathbb{Z}\in \mc M_{\max{}}$ and $\rho(T_G)<\rho(T_{G\times\mathbb{Z}})$.
\end{lem}
\begin{proof}
It is easy to see $G\times \mathbb{Z}$ satisfies max. For the latter condition, let $G=\{g_0,g_1,\ldots\}$ and $G\times\mathbb{Z}=\{a_0,a_1,\ldots\}$. There is some $k\in\mathbb{N}$ such that $a_k=(e_G,z)$ where $\mathbb{Z}=\<z\>$. Let $\psi\colon\mathbb{N}\to\mathbb{N}$ be defined such that $a_{\psi(m)}=(g_m,e_\mathbb{Z})$. The map $\phi:T_G\rightarrow \mathbb{N}^{<\mathbb{N}}$ given by $(s_0,\ldots,s_n)\mapsto(\psi(s_0),\ldots,\psi(s_n))$ is clearly monotone, and further, $\phi(T_G)\subseteq (T_{G\times\mathbb{Z}})_k$. By Lemma \ref{lem:TrRkMonotone}, $\rho(T_G) \leq \rho((T_{G\times\mathbb{Z}})_k)<\rho(T_{G\times\mathbb{Z}})$.
\end{proof}
\begin{lem}\label{lem:MaxRankLim}
Let $\{A_i\}_{i\in\mathbb{N}}$ be countable groups. If $A_i\in \mc M_{\max{}}$ for each $i\in\mathbb{N}$, then there is a group $A\in \mc M_{\max{}}$ such that $\rho(T_A)\geq\rho(T_{A_i})$ for all $i\in\mathbb{N}$.
\end{lem}
\begin{proof}
This is a consequence of \cite[Theorem 2]{Ol89} due to A. Y. Olshanskii. This result gives a 2-generated group $A$ containing each of the $A_i$ such that every proper subgroup of $A$ is either contained in a conjugate of some $A_i$, is infinite cyclic, or is infinite dihedral. Thus if every subgroup of each $A_i$ is finitely generated, then every subgroup of $A$ is finitely generated, and so $A\in\mc M_{\max{}}$. Since each $A_i$ is a subgroup of $A$, Lemma \ref{lem:MaxSubRank} implies that $\rho(T_A)\geq\rho(T_{A_i})$ for all $i\in\mathbb{N}$ as desired.
\end{proof}
\begin{lem}\label{lem:MaxRankUnbdd}
For all $\alpha<\omega_1$, there is $G\in \mc M_{\max{}}$ such that $\rho(T_G)\geq\alpha$.
\end{lem}
\begin{proof}
The proof is the same as that of Lemma \ref{lem:CentRankUnbdd}, with Lemmas \ref{lem:MaxRankSucc} and \ref{lem:MaxRankLim} referenced at the appropriate places.
\end{proof}
\begin{thm}
$\mc M_{\max{}}$ is $\Pi^1_1$ and not Borel in $\mathscr{G}$ and $\mathscr{G}_{fg}$.
\end{thm}
\begin{proof}
The proof is the same as that of Theorem \ref{thm:MCNotBorel} using Lemmas \ref{lem:MaxMapBorel} and \ref{lem:MaxRankUnbdd} where appropriate. The statement is true for $\mc G_{fg}$ simply because $\mc M_{\max{}}\subseteq\mc G_{fg}$.
\end{proof}
\section{The maximal condition on normal subgroups}\label{sec:MaxN}
Given a group $G$ and a set $S\subseteq G$, we write $\<\<S\>\>_G$ to denote the normal closure of $S$ in $G$. We suppress the subscript $G$ when the group is clear from context.
\begin{defn}
A group $G$ satisfies the \textbf{maximal condition on normal subgroups}, abbreviated by saying a group satisfies max-n, if there is no infinite strictly increasing chain of normal subgroups of $G$. Equivalently, a group $G$ satisfies max-n if each of its normal subgroups is the normal closure of finitely many elements of $G$. We denote the class of groups satisfying max-n as $\mc M_n$.
\end{defn}
Given $G\in \mathscr{G}$, we construct a tree $T_G\subseteq\mathbb{N}^{<\mathbb{N}}$ and associated groups $G_s\in \mathscr{G}$ for each $s\in T_G$.
\begin{enumerate}[$\bullet$]
\item Put $\emptyset\in T_G$ and let $G_\emptyset:=G$.
\item Suppose that $s\in T_G$ and $G_s=G/\ngrp{\{g_s\}}$ has already been defined. If $\ngrp{\{g_s\}\cup\{g_i\}}\neq \ngrp{\{g_s\}}$, then put $s^{\smallfrown} i\in T_G$ and $G_{s^{\smallfrown} i}:=G/\ngrp{\{g_s\}\cup\{g_i\}}$.
\end{enumerate}
\begin{lem}\label{lem:MaxNMapBorel}
The map $\Phi_{M_n}\colon\mathscr{G}\to Tr$ given by $G\mapsto T_G$ is Borel.
\end{lem}
We prove Lemma~\ref{lem:MaxNMapBorel} in Section \ref{sec:Borel}.
\begin{lem}\label{lem:MaxNMapReduction}
$T_G$ is well-founded if and only if $G\in\mc M_n$.
\end{lem}
\begin{proof}
If $G\in\mc M_n$, then $T_G$ contains no infinite branches by definition. If $G\notin\mc M_n$, then there is a normal subgroup $N\trianglelefteq G$ such that $N=\ngrp{g_{a_0},g_{a_1},\ldots}$ and
$$\ngrp{g_{a_0},\ldots,g_{a_n}} \lneq \ngrp{g_{a_0},\ldots,g_{a_{n+1}}}$$
for all $n\in\mathbb{N}$. Thus the sequence $(a_0,\ldots, a_n)$ is an element of $T_G$ for any $n\in\mathbb{N}$, so $T_G$ has an infinite branch.
\end{proof}
\begin{lem}\label{lem:MaxNImRank}
If $G\in\mc M_n$ and $f:G\twoheadrightarrow G'$, then $\rho(T_{G})\geq \rho(T_{G'})$ with equality if and only if $f$ is injective.
\end{lem}
\begin{proof}
Since $G$ is max-n, $\ker(f)=\ngrp{S}$ for some finite $S=\{g_{s_0},\ldots,g_{s_n}\}$. We may assume that $n$ is minimal, so no element of $S$ is in the normal closure of the others. Setting $(s_0,\dots,s_n)=:s\in\mathbb{N}^{<\mathbb{N}}$, the minimality of $S$ implies $s\in T_G$; in the case $\ker(f)=\{1\}$, we take $s=\emptyset$. Let $\psi\colon\mathbb{N}\to\mathbb{N}$ be a map such that $g'_k=f(g_{\psi(k)})$. Then for all $i_0,\ldots,i_k\in\mathbb{N}$,
\[
G'/\<\<g'_{i_0},\ldots,g'_{i_k}\>\>_{G'} \cong G/\<\<g_{\psi(i_0)},\ldots,g_{\psi(i_k)},S\>\>_G,
\]
and the monotone map $\phi:T_{G'}\rightarrow \mathbb{N}^{<\mathbb{N}}$ given by $(r_0,\ldots,r_n)\mapsto(\psi(r_0),\ldots,\psi(r_n))$ sends $T_{G'}$ into $(T_G)_s$. By Lemma \ref{lem:TrRkMonotone}, $\rho(T_{G'}) \leq \rho((T_G)_s)\leq\rho(T_G)$, and the rightmost inequality is strict if and only if $s\neq \emptyset$.
\end{proof}
We conclude this rank is also isomorphism invariant.
\begin{cor}\label{cor:MaxNRankIsoInv}
If $G,G'\in\mathscr{G}$ and $G\cong G'$, then $\rho(T_G)=\rho(T_{G'})$.
\end{cor}
Recall that a group is \textbf{hopfian} if it is not isomorphic to any of its proper quotients. The following corollary is easy enough to prove directly, but it follows immediately from Lemma \ref{lem:MaxNImRank} and Corollary~\ref{cor:MaxNRankIsoInv}.
\begin{cor}
If $G\in\mc M_n$, then $G$ is hopfian.
\end{cor}
Unlike the previous invariants, this rank has appeared before in the literature; cf.\ \cite{C11}.
\begin{defn}
If $G$ has the maximal condition on normal subgroups, then $\rho(T_G)$ for some (any) marking of $G$ is called the \textbf{length} of $G$.
\end{defn}
If we were to follow our template from previous sections, we would move on to analogues of Lemmas \ref{lem:CentRankSucc} and \ref{lem:CentRankLim}. However, we were unable to prove an analogue of Lemma \ref{lem:CentRankLim} which would take advantage of Lemma \ref{lem:MaxNImRank}. Such a result would be a sort of dual version of the result of Olshanskii cited in the proof of Lemma \ref{lem:MaxRankLim}. Specifically, the following question is open to the best of the authors' knowledge:
\begin{quest}
Suppose $\{A_i\}_{i\in\mathbb{N}}$ is a set of normally $k$-generated max-n groups. Is there a max-n group $A$ such that $A\twoheadrightarrow A_i$ for all $i\in\mathbb{N}$?
\end{quest}
A positive answer to this question would give us exactly the right analogue of Lemma \ref{lem:CentRankLim}. Lacking this, we will use a construction involving (restricted) wreath products. Recall the wreath product of $H$ and $G$ is $H\wr G:=H^{<G}\rtimes G$ where $H^{<G}$ denotes the direct sum and $G\curvearrowright H^{<G}$ by shift; in the case $G\curvearrowright X$ for some set $X$, we write $H\wr_X G:=H^{<X}\rtimes G$. We will see that we can relate $\rho(T_{H\wr G})$ to both $\rho(T_H)$ and $\rho(T_G)$, while Lemma \ref{lem:MaxNImRank} alone only gives us information about how $\rho(T_{H\wr G})$ and $\rho(T_G)$ relate.
We will focus on perfect max-n groups with no central factors; let us call the set of such groups $\mc M'_n$. A group $G$ is said to have a \textbf{central factor} if there are normal subgroups $L\trianglelefteq M $ in $G$ such that $M/L$ is nontrivial and central in $G/L$. Since $\mc M'_n\subseteq \mc M_n$, it is enough for our purposes to show that $\rho$ is unbounded below $\omega_1$ on $\mc M'_n$ and $\mc M'_n\cap \mathscr{G}_{fg}$.
\begin{lem}\label{lem:MaxNRankSucc}
Let $S$ be an infinite simple group. For all groups $G\in\mc M_n$, $G\times S\in\mc M_n$ and $\rho(T_G)<\rho(T_{G\times S})$. If $G\in\mc M'_n$, then so is $G\times S$.
\end{lem}
\begin{proof}
It is easy to see that $G\times S\in\mc M_n$, and since $G$ is a quotient of $G\times S$, Lemma \ref{lem:MaxNImRank} implies $\rho(T_G)<\rho(T_{G\times S})$. If $G$ is perfect, then $G\times S$ is perfect, so for the last statement we need only to check that if $G$ has no central factors, then $G\times S$ has no central factors. Suppose that $L, M\trianglelefteq G\times S$ give a central factor. Let $\pi\colon G\times S \to G$ be the usual projection. Since $G$ has no central factors, $\pi(M)=\pi(L)$. Thus $MS=LS$, so $M=L(S\cap M)$. Since $S$ has no central factors, $S\cap M=S\cap L$. We conclude that $M=L$, whereby $G\times S$ has no central factors, a contradiction.
\end{proof}
Lemma~\rm\ref{lem:MaxNRankSucc} allows us to find a group in $\mc M'_n$ with rank greater than a given group in $\mc M'_n$. However, we also need to be able to find a group in $\mc M'_n$ with rank greater than a countable family of groups from $\mc M'_n$. We begin by looking at properties of the ranks of wreath products.
\begin{lem}\label{lem:max-n}
Suppose $H$ and $G$ are groups satisfying max-n. Then $\rho(T_{H\wr G})\geq \rho(T_G)+\rho(T_H)$.
\end{lem}
\begin{proof}
For each $h\in H$ define $f_h\in H^{<G}$ by
\[
f_h(g)=
\begin{cases}
h, & \text{ if } g=e\\
e, & \text { else.}
\end{cases}
\]
Let $H=\{h_0,h_1,\ldots\}$ and let $\psi\colon\mathbb{N}\to\mathbb{N}$ be a map such that $f_{h_i}=g_{\psi(i)}$. We now define a monotone $\phi:T_H\rightarrow \mathbb{N}^{<\mathbb{N}}$: Put $\phi(\emptyset)=\emptyset$. For non-empty $s\in T_H$, define $\phi$ by
$$(s_0,\dots,s_k)\mapsto \left(\psi(s_0),\dots, \psi(s_k)\right).$$
\indent We argue $\phi$ maps $T_H$ into $T_{H\wr G}$ by induction on the length of $s\in T_H$. As the base case is immediate, say $s\in T_H$ and $s^{\smallfrown} k\in T_H$. By construction, it is the case that $\ngrp{\{h_s\}\cup \{h_k\}}_H\neq\ngrp{\{h_s\}}_H$. For all $t\in T_H$, $\ngrp{\{g_{\phi(t)}\}}_{H\wr G}=\ngrp{\{h_t\}}_H^{<G}$, hence
\[
\ngrp{\{g_{\phi(s)}\}\cup\{g_{\psi(k)}\}}_{H\wr G}\neq \ngrp{\{g_{\phi(s)}\}}_{H\wr G}.
\]
We conclude that $\phi(s^{\smallfrown} k)\in T_{H\wr G}$, so $\phi$ maps $T_H$ into $T_{H\wr G}$.\par
Now if $s=(s_0,\ldots,s_n)\in T_H$ is a terminal node, then $\ngrp{\{h_s\}}_H = H$. In this case $(H\wr G)/\ngrp{\{g_{\phi(s)}\}}_{H\wr G} \cong G$, so $\rho(T_G)=\rho((T_{H\wr G})_{\phi(s)})$ by Corollary \ref{cor:MaxNRankIsoInv}. The desired result now follows.
\end{proof}
In general, $H\wr G$ need not be max-n. A theorem of P. Hall provides a sufficient condition for this.
\begin{thm}[Hall, {\cite[Theorem 4]{H54}}]\label{thm:max-n_wreath}
Let $H$ and $G$ be groups satisfying max-n. If $H$ has no central factors, then $H\wr G$ satisfies max-n.
\end{thm}
Our next lemma allows us to iterate wreath products and remain in $\mc M'_n$.
\begin{lem}\label{lem:central_factors}
If $G$ and $H$ have no central factors, then $H\wr G$ has no central factors.
\end{lem}
\begin{proof}
Suppose $L\trianglelefteq M$ gives a central factor of $H\wr G$. Let $\pi:H\wr G\rightarrow G$ be the usual projection. Since $G$ has no central factors, it must be the case that $\pi(L)=\pi(M)$, so $LH^{<G}=MH^{<G}$. Thus, $M=L(H^{<G}\cap M)$, and it suffices to show $H^{<G}\cap M\leq L$. Since $H$ has no central factors, it follows similarly to the proof of Lemma~\rm\ref{lem:MaxNRankSucc} that $H^{F}\cap M=H^{F}\cap L$ for all finite $F\subseteq G$. We conclude that $H^{<G}\cap M\leq L$ verifying the lemma.
\end{proof}
It is easy to see the wreath product of two perfect groups is perfect, so using Theorem \ref{thm:max-n_wreath} and Lemma \ref{lem:central_factors}, the class $\mc M'_n$ is closed under wreath products. With the following fact from the literature, we are equipped to prove the desired lemma.
\begin{lem}[{\cite[Lemma 3.6]{GKO14}}]\label{lem:wreath}
Suppose $A,B$ are countable groups and form $G=A\wr B$. If $N\trianglelefteq G$ meets $B$ non-trivially, then $[A,A]^{<B}\leq N$.
\end{lem}
\begin{lem}\label{lem:MaxNRankLim}
Let $\{A_i\}_{i\in\mathbb{N}}$ be countable groups. If $A_i\in\mc M'_n$ for all $i\in\mathbb{N}$, then there is a group $A\in\mc M'_n$ such that $\rho(T_{A_i})\leq\rho(T_A)$ for all $i\in\mathbb{N}$.
\end{lem}
\begin{proof}
For each $n$, put $G_n:=A_n\wr\left(\dots \wr A_0\right)$. By making the natural identification, we may assume $G_{n}\leq G_{n+1}$ for all $n$ and form $A:=\bigcup_{n\in \mathbb{N}} G_n$. (Alternatively, one may take the direct limit.) \par
\indent Consider $N\trianglelefteq A$. Certainly, $N\cap G_n$ is non-trivial for some $n$. Fix such an $n$ and take $k>n$. We now see $N\cap G_{k}\trianglelefteq G_{k}=A_{k}\wr G_{k-1}$ is a normal subgroup that meets $G_{k-1}$ non-trivially. Applying Lemma~\rm\ref{lem:wreath}, $[A_{k},A_{k}]^{<G_{k-1}}\leq N\cap G_{k}$. Since $A_{k}$ is perfect, we have that $A_{k}^{<G_{k-1}}\leq N$. It now follows that $A/N$ is isomorphic to a quotient of $G_n$.
\indent Suppose $(N_i)_{i\in \mathbb{N}}$ is an increasing sequence of normal subgroups of $A$. By the previous paragraph, $A/N_0$ is a quotient of $G_n$ for some $n$, and Theorem~\rm\ref{thm:max-n_wreath} implies that each $G_n$ is a max-n group. Letting $\pi:A\rightarrow A/N_0$ be the usual projection, it is thus the case that $\pi(N_i)=\pi(N_j)$ for all sufficiently large $i$ and $j$. Therefore, $N_i=N_j$ for all sufficiently large $i$ and $j$, and $A$ satisfies max-$n$. \par
\indent For each $n$ and $k>n$, define
\[
L_n^k:=A_k\wr_{G_{k-1}}\left( \dots\wr_{G_{n+2}}\left(A_{n+2}\wr_{G_{n+1}}A_{n+1}^{<G_n}\right)\right)
\]
and put $L_n:=\bigcup_{k>n}L_n^k$. We see $L_n\trianglelefteq A$ and $A/L_n\cong G_n$. By Lemmas \ref{lem:max-n} and \ref{lem:MaxNImRank}, $\rho(T_{A_n})\leq\rho(T_{G_n})\leq\rho(T_A)$ for all $n$.
\indent We finally verify $A$ is perfect and has no central factors. That $A$ is perfect is immediate. It follows from Lemma~\rm\ref{lem:central_factors} and induction that each $G_n$ has no central factors. Since any factor of $A$ is a factor of $G_n$ for some $n$, $A$ has no central factors.
\end{proof}
\begin{lem}\label{lem:MaxNRankUnbdd}
For all $\alpha<\omega_1$, there is $G\in\mc M'_n$ such that $\rho(T_G)\geq\alpha$.
\end{lem}
\begin{proof}
The proof is the same as that of Lemma \ref{lem:CentRankUnbdd}, with Lemmas \ref{lem:MaxNRankSucc} and \ref{lem:MaxNRankLim} referenced at the appropriate places.
\end{proof}
The groups given by Lemma~\rm\ref{lem:MaxNRankUnbdd} are not, in general, finitely generated. For finding finitely generated examples another result of Hall is needed.
\begin{lem}[Hall, {\cite[cf. Theorem 4]{H61}}]\label{lem:Hall}
Let $H$ be a countable group. Then there exists a short exact sequence
\[
\{e\}\rightarrow M\rightarrow G\rightarrow \mathbb{Z}\rightarrow \{e\}
\]
where $G$ is 2-generated, $[M,M]=[H,H]^{<\mathbb{Z}}$, and there is $t\in G$ so that the conjugation action of $t$ on $[M,M]$ is by unit shift.
\end{lem}
It is useful to sketch Hall's construction of $G$. Let $\{h_i\}_{i\in \mathbb{N}}$ list $H$ and form the unrestricted wreath product $H^{\mathbb{Z}}\rtimes \mathbb{Z}$. Define $\sigma\in H^{\mathbb{Z}}$ by
\[
\sigma(i):=
\begin{cases}
h_n, & \text{ if }i=2^n\\
e, & \text{ else}
\end{cases}
\]
and let $t$ be a generator for $\mathbb{Z}$ in $H^{\mathbb{Z}}\rtimes \mathbb{Z}$. The desired group is then $G:=\grp{t,\sigma}$. The subgroup $M$ equals $\grp{g\sigma g^{-1}\mid g\in G}$.\par
\indent We point out a consequence of the construction for later use: Suppose $H$ is perfect and $h\in H$. Taking $f_h\in [H,H]^{<\mathbb{Z}}=H^{<\mathbb{Z}}$ as defined in Lemma \ref{lem:max-n}, the construction of $G$ implies $\ngrp{f_h}_G=\ngrp{h}_H^{<\mathbb{Z}}$.
\begin{cor}\label{cor:MaxNFG}
For each $\alpha<\omega_1$, there is a finitely generated group $G\in\mc M_n$ with \mbox{$\rho(T_G)\geq \alpha$}.
\end{cor}
\begin{proof}
Fix $\alpha<\omega_1$ and apply Lemma~\rm\ref{lem:MaxNRankUnbdd} to find a group $H\in\mc M'_n$ with $\rho(T_H)\geq \alpha$. We now apply Lemma~\rm\ref{lem:Hall} to find a 2-generated group $G$ with a short exact sequence
\[
\{e\}\rightarrow M\rightarrow G\rightarrow \mathbb{Z}\rightarrow \{e\}
\]
where $[M,M]=[H,H]^{<\mathbb{Z}}=H^{<\mathbb{Z}}$. \par
\indent The group $G/[M,M]$ is a finitely generated metabelian group, hence it satisfies max-n by \cite[Theorem 3]{H54}. On the other hand, any normal subgroup of $G$ that lies in $[M,M]=H^{<\mathbb{Z}}$ is shift-invariant because $G$ contains an element that acts by shift on $[M,M]$. Since $H\wr\mathbb{Z}$ is max-n, it follows that $H^{<\mathbb{Z}}$ is max-$G$; that is to say $H^{<\mathbb{Z}}$ has the maximal condition on subgroups invariant under the conjugation action by $G$. We conclude the group $G$ is max-n.\par
\indent It remains to compute a lower bound for $\rho(T_G)$. Using again the notation from Lemma \ref{lem:max-n}, find $\psi:\mathbb{N}\rightarrow \mathbb{N}$ such that for each $k\in \mathbb{N}$ we have $f_{h_k}=g_{\psi(k)}$. Since $\ngrp{f_h}_G=\ngrp{h}_H^{<\mathbb{Z}}$, we may argue as in Lemma~\ref{lem:max-n} to conclude that $\alpha\leq \rho(T_H)\leq \rho(T_G)$. That is to say, we can define a monotone $\phi\colon T_H\to T_G$ and by Lemma~\rm\ref{lem:TrRkMonotone} conclude that $\alpha\leq \rho(T_H)\leq \rho(T_G)$.
\end{proof}
\begin{thm}
$\mc M_n$ is $\Pi^1_1$ and not Borel in $\mathscr{G}$, and $\mc M_n\cap G_{fg}$ is $\Pi^1_1$ and not Borel in $\mc G_{fg}$.
\end{thm}
\begin{proof}
This follows from Theorem \ref{thm:BddnessThm}, Lemma \ref{lem:MaxNMapBorel}, and Corollary \ref{cor:MaxNFG}.
\end{proof}
\section{Elementary amenable groups}\label{sec:EAGroups}
Perhaps surprisingly, the property of being elementary amenable may also be characterized by well-founded trees. This in turn gives a chain condition equivalent to elementary amenability.
\subsection{Preliminaries} We study the collection of elementary amenable groups. This class is typically defined as follows:
\begin{defn}\label{def:EA}
The collection of \textbf{elementary amenable groups}, denoted $\mathop{\rm EG} \nolimits$, is the smallest collection of countable groups such that
\begin{enumerate}[(i)]
\item $\mathop{\rm EG} \nolimits$ contains all finite groups and abelian groups.
\item $\mathop{\rm EG} \nolimits$ is closed under group extensions.
\item $\mathop{\rm EG} \nolimits$ is closed under countable increasing unions.
\item $\mathop{\rm EG} \nolimits$ is closed under taking subgroups.
\item $\mathop{\rm EG} \nolimits$ is closed under taking quotients.
\end{enumerate}
\end{defn}
Our results here require a fairly well-known embedding result, which is based on a generalization of Lemma~\ref{lem:Hall}.
\begin{prop}[Hall; Neumann, Neumann {\cite[Theorem 5.1]{NN59}}]\label{prop:EAembedding}
Suppose $K\in \mathop{\rm EG} \nolimits$. Then there exists $H\in \mathop{\rm EG} \nolimits$ and a short exact sequence
\[
\{e\}\rightarrow M\rightarrow G\rightarrow \mathbb{Z}\rightarrow \{e\}
\]
where $G$ is 2-generated, $G\in \mathop{\rm EG} \nolimits$, $[M,M]=[H,H]^{<\mathbb{Z}}$, and $K$ embeds into $[H,H]$.
\end{prop}
\subsection{Decomposition trees}
We now define a tree associated to a marked group $G$. Just as in the previous sections, we will see that this tree being well-founded or not gives group-theoretic information about $G$, in this case characterizing being an elementary amenable group.\par
Let $G\in\mathscr{G}$. For $n\geq 0$, put $R_n(G):=\<g_0,\ldots,g_n\>$ and for $k\geq 1$, define
\[
S_k(G):=[G,G]\cap\bigcap\mc{N}_{k}(G)
\]
where $\mc{N}_k(G):=\{N\trianglelefteq G\;|\;|G:N|\leq k+1\}$. For each $l\geq 1$, we now define a tree $T^l(G)\subseteq\mathbb{N}^{<\mathbb{N}}$ and associated groups $G_s\in\mathscr{G}$ as follows:
\begin{enumerate}[$\bullet$]
\item Put $\emptyset\in T^l(G)$ and let $G_{\emptyset}:=G$.
\item Suppose we have $s\in T^l(G)$ and $G_s$. If $G_s\neq\{e\}$, put $s^{\smallfrown} n\in T^l(G)$ and $G_{s^{\smallfrown} n}:=S_{|s|+l}\left(R_n\left(G_s\right)\right)$.
\end{enumerate}
We call $T^l(G)$ the \textbf{decomposition tree} of $G$ with offset $l$. This tree is always non-empty, and if $s\in T^l(G)$ is terminal, then $G_s=\{e\}$. Since the composition of the functions $R_n$ and $S_k$ is associative, we obtain a useful observation:
\begin{obs}\label{prop:child_rank}
For $s\in T^l(G)$, $T^l(G)_s=T^{|s|+l}(G_s)$, and for each $r\in T^{|s|+l}(G_s)$, $(G_{s})_r=G_{s^{\smallfrown} r}$ as marked groups. This implies, in particular, that if $T^l(G)$ is well-founded, then so is $T^{|s|+l}(G_s)$.
\end{obs}
\begin{lem}\label{lem:Phi_borel}
For each $l\geq 1$, the map $\Phi^l\colon\mathscr{G}\to Tr$ given by $G\mapsto T^l(G)$ is Borel.
\end{lem}
As usual, we postpone the proof of this lemma to Section \ref{sec:Borel}.
\begin{lem}\label{lem:xi_indp}
Let $G,H\in\mathscr{G}$ and $H\hookrightarrow G$. Then for all $l\geq k \geq 1$,
\[
\rho\left(T^l(H)\right)\leq\rho\left(T^k(G)\right).
\]
In particular, for $G,G'\in\mathscr{G}$, if $G\cong G'$, then
\[
\rho\left(T^l(G)\right)=\rho\left(T^l(G')\right)
\]
for all $l\geq 1$.
\end{lem}
\begin{proof}
We induct on $\rho(T^k(G))$ simultaneously for all $k$. If $\rho(T^k(G))=1$, then $G=\{e\}$, so $H=\{e\}$. Suppose the lemma holds for all $G$ and $k$ with $\rho(T^k(G))\leq\beta$. Suppose that $f\colon H\to G$ is an embedding and $\rho(T^k(G))=\beta+1$. For all $n\geq 0$, there is some $k(n)$ so that $f(R_n(H))\leq R_{k(n)}(G)$. It follows that $f(H_n)\leq G_{k(n)}$ for all $n\geq 0$ since $S_{l+1}(G_{k(n)})\leq S_{k+1}(G_{k(n)})$. By the inductive hypothesis and Observation~\ref{prop:child_rank},
\[
\rho\left(T^l(H)\right)=\sup_{n\in\mathbb{N}}\left\{\rho\left(T^{l+1}(H_n)\right)\right\}+1 \leq \sup_{n\in\mathbb{N}}\left\{\rho\left(T^{k+1}(G_{k(n)})\right)\right\}+1 \leq \rho\left(T^k(G)\right)
\]
completing the induction.
\end{proof}
\begin{cor}\label{cor:some_all}
For $G\in \mathscr{G}$, $T^l(G)$ is well-founded for some $l\geq 1$ if and only if $T^l(G)$ is well-founded for all $l\geq 1$.
\end{cor}
\begin{proof} Suppose $G\in \mathscr{G}$ is so that $T^l(G)$ is well-founded. In view of Lemma~\ref{lem:xi_indp}, $T^k(G)$ is well-founded for all $k\geq l$. For $n\leq l$, take $s\in T^n(G)$ with $|s|=l$; if no such $s$ exists then $T^n(G)$ is plainly well-founded. There is an injection $G_s\hookrightarrow G$, so applying Lemma~\ref{lem:xi_indp} once again,
\[
\rho\left(T^{n+|s|}(G_s)\right)\leq \rho\left(T^{n+|s|}(G)\right).
\]
By choice of $s$, $n+|s|\geq l$, hence $\rho(T^{n+|s|}(G))<\omega_1$. Since $T^n(G)_s=T^{n+|s|}(G_s)$, we conclude $T^{n}(G)_s$ is well-founded for each $s$ of length $l$. The tree $T^n(G)$ is therefore well-founded, and the corollary follows.
\end{proof}
Define $\mathrm{W}:=\cup_{l=1}^{\infty} (\Phi^l)^{-1}(WF)$; that is, $\mathrm{W}$ is the collection of marked groups so that some decomposition tree is well-founded. By Corollary~\ref{cor:some_all}, every decomposition tree of a group in $\mathrm{W}$ is well-founded; that is to say, $W=\cap_{l=1}^{\infty} (\Phi^l)^{-1}(WF)$.\par
\indent Lemma~\ref{lem:xi_indp} shows the rank of a decomposition tree is independent of the marking. We thus define
\begin{defn} The \textbf{decomposition rank} of $G\in\mathrm{W}$ is defined to be
\[
\xi(G):=\min_{k\in \omega}\rho\left(T^k(G)\right)
\]
for some (any) marking of $G$. The \textbf{decomposition degree} is defined to be
\[
\deg(G):=\min\left\{k \mid \xi(G)=\rho\left(T^k(G)\right)\right\}
\]
for some (any) marking of $G$.
\end{defn}
\begin{cor}\label{lem:sgrp_xi}
If $G,H\in\mathscr{G}$ and $H\hookrightarrow G$, then $\xi(H)\leq \xi(G)$.
\end{cor}
\begin{rmk}
The decomposition rank in a fairly straightforward manner tracks the number of extensions and unions applied to produce the group. The decomposition degree, on the other hand, is currently mysterious. It somehow tracks the size of the finite groups ``appearing" in the construction of an elementary amenable group. We do not consider the decomposition degree further as it is tangential to our goal. We do study the decomposition rank in detail.
\end{rmk}
We now show that $\mathrm{W} \subseteq \mathop{\rm EG} \nolimits$ and $\mathrm{W}$ enjoys the same closure properties as $\mathop{\rm EG} \nolimits$, so that in fact $\mathop{\rm EG} \nolimits=\mathrm{W}$.
\begin{thm}
If $G\in\mathrm{W}$, then $G\in\mathop{\rm EG} \nolimits$.
\end{thm}
\begin{proof}
We induct on $\xi(G)$. For the base case, if $\xi(G)=1$, then $G=\{e\}$ and $G\in\mathop{\rm EG} \nolimits$. Suppose the theorem holds for all $\alpha<\beta$ and $\xi(G)=\rho(T^l(G))=\beta$. Consider $R_i(G)$. Since $R_i(G)$ is finitely generated, $\mc{N}_{l+1}(R_i(G))$ is finite, so
\[
\left|[R_i(G),R_i(G)]:G_i\right| <\infty.
\]
We infer $R_i(G)/G_i$ is finite-by-abelian and, therefore, elementary amenable.\par
\indent On the other hand, Observation~\rm\ref{prop:child_rank} gives $\rho(T^{1+l}(G_i))=\rho(T^l(G)_i)$. Hence, $\rho(T^{1+l}(G_i))<\beta$, and we conclude that $G_i\in \mathop{\rm EG} \nolimits$ from the inductive hypothesis. As $\mathop{\rm EG} \nolimits$ is closed under group extensions and countable increasing unions, $R_i(G)\in\mathop{\rm EG} \nolimits$ for all $i\in\omega$, whereby $G\in\mathop{\rm EG} \nolimits$.
\end{proof}
The family $\mathrm{W}$ also has the same closure properties as $\mathop{\rm EG} \nolimits$. Lemma~\ref{lem:xi_indp} already shows $\mathrm{W}$ is closed under taking subgroups. For the other closure properties, we require several lemmas.
\begin{lem}\label{lem:finandab}
$\mathrm{W}$ contains all finite groups and all abelian groups.
\end{lem}
\begin{proof}
If $G$ is abelian, then $\rho(T^1(G))\leq 2$. If $G$ is finite with size $m$, then $\rho(T^m(G))\leq 2$.
\end{proof}
We next consider increasing unions.
\begin{lem}\label{lem:union_xi}
If $G=\cup_{i\in\mathbb{N}} H_i$ and each $H_i\in\mathrm{W}$, then $G\in\mathrm{W}$.
\end{lem}
\begin{proof}
For each $i\in\mathbb{N}$, let $\alpha_i:=\rho(T^{1}(H_i))<\omega_1$. Since each $R_n(G)$ is finitely generated, there is some $m_n\in\mathbb{N}$ such that $R_n(G)\leq H_{m_n}$. By Lemma \ref{lem:xi_indp}, $\rho(T^{1}(G_n))\leq\rho(T^{1}(H_{m_n}))=\alpha_{m_n}$. We conclude $\rho(T^1(G))\leq \sup_{i\in\mathbb{N}} (\alpha_{m_i}) + 1 < \omega_1$, and thereby, $G\in\mathrm{W}$.
\end{proof}
In our construction, given $G$ and $k\geq 1$, we are particularly interested in the $G_i$ associated with $i\in T^k(G)$. We will see their decomposition rank is related to that of $G$ in a simple way; this observation is necessary for showing $W$ is closed under taking extensions and quotients.
\begin{lem}\label{lem:rk_xi}
Suppose $G\in\mathrm{W}$ is non-trivial and $\deg(G)=k$. Then
\[
\sup_{i\in \omega}\xi(G_i)+1\leq \xi(G)
\]
where $G_i$ is the subgroup of $G$ associated to $i\in T^k(G)$. In particular, $\xi(G_i)<\xi(G)$ for all $i\in\mathbb{N}$.
\end{lem}
\begin{proof}
By construction, for all $i\in\mathbb{N}$,
\begin{align*}
\rho\left(T^{k+1}(G_i)\right)+1 &=\rho\left(T^k(G)_i\right)+1 \\
&\leq \rho\left(T^k(G)\right).
\end{align*}
Hence,
\begin{align*}
\sup_{i\in \mathbb{N}}\xi(G_i)+1 &= \sup_{i\in \mathbb{N}} \left\{ \min_{l\in\mathbb{N}} \rho\left(T^l(G_i)\right)\right\}+1 \\
&\leq \sup_{i\in\mathbb{N}} \left\{ \rho\left(T^{k+1}(G_i)\right) \right\} +1\\
&= \rho\left(T^k(G)\right) \\
&= \xi(G)
\end{align*}
as desired.
\end{proof}
\noindent The inequality in Lemma~\rm\ref{lem:rk_xi} may be strict; for example, consider $\operatorname{Sym}_{fin}(\mathbb{N})$, the group of finitely supported permutations of $\mathbb{N}$. We also point out that Lemma~\rm\ref{lem:rk_xi} \emph{does not} hold for choices of $k$ such that $\rho(T^k(G))\neq\xi(G)$.\par
We next show that $\mathrm{W}$ is closed under extensions. We will first prove a weaker statement. This approach is inspired by \cite{Os02}.
\begin{lem}\label{lem:finorab_xi}
Suppose that $N\in\mathrm{W}$, $B$ is finite or abelian, and there is a short exact sequence
\[
1 \rightarrow N \rightarrow G \rightarrow B \rightarrow 1 .
\]
Then $G\in\mathrm{W}$.
\end{lem}
\begin{proof}
Suppose first that $B$ is abelian. Thus, $[G,G] \leq N$, so for any $l\geq 1$ and all $n\in T^l(G)$, $G_n\leq N$. It follows from Lemma \ref{lem:xi_indp} that for all $n\in\mathbb{N}$,
\[
\rho(T^{l+1}(G_n))\leq\rho(T^{l+1}(N))<\omega_1.
\]
Appealing to Observation \ref{prop:child_rank}, we infer $\rho(T^l(G))<\omega_1$, so $G\in \mathrm{W}$.
Suppose that $B$ is finite and $|G\colon N|=k$. For all $n\in T^k(G)$, $G_n\leq N$, so as above, $T^k(G)$ is well-founded. Hence, $G\in W$.
\end{proof}
\begin{lem}\label{lem:ext_xi}
Suppose the group $G$ is the extension of a group $B\in\mathrm{W}$ by a group $N\in\mathrm{W}$. Then $G\in\mathrm{W}$. The family $\mathrm{W}$ is thus closed under group extensions.
\end{lem}
\begin{proof}
We first establish the following claim.
\begin{claim*}
If $N\in\mathrm{W}$ and $B$ is finite-by-abelian, then the extension of $B$ by $N$ is in $\mathrm{W}$.
\end{claim*}
\begin{proof}[Proof of claim.]
Suppose that $B$ is the extension of an abelian group $A$ by a finite group $F$. Let $F_0$ be the preimage of $F$ in $G$. Then $G/F_0 \cong B/F \cong A$, so $G/F_0$ is abelian. Since $F_0$ is the extension of the finite group $F$ by $N$, Lemma \ref{lem:finorab_xi} implies that $F_0\in\mathrm{W}$. Applying Lemma \ref{lem:finorab_xi} a second time, $G\in\mathrm{W}$.
\end{proof}
We now prove the lemma by induction on $\beta=\xi(B)$. If $\beta=1$, then $B=\{e\}$ and the induction claim holds trivially. Suppose the result holds for all $\delta<\beta$. First, assume that $B$ is finitely generated, let $\deg(B)=l$, and form the decomposition tree $T^l(B)$. By finite generation, there is some $m\in\mathbb{N}$ such that for all $k\geq m$, $R_k(B)=B$, so $B_k=[B,B]\cap\bigcap\mc N_l(B)$. We now consider $K\trianglelefteq G$ the preimage of $B_k$ under the projection map. The group $K$ is the extension of $B_k$ by $N$, and $\xi(B_k)<\xi(B)$ by Lemma~\ref{lem:rk_xi}. The inductive hypothesis therefore implies $K\in\mathrm{W}$. On the other hand, $G/K$ is finite-by-abelian, so $G\in\mathrm{W}$ by our claim.\par
\indent If $B$ is not finitely generated, then $B=\cup_{n\in\mathbb{N}} R_n(B)$, and $\xi(R_n(B))\leq\xi(B)$ for all $n\in\mathbb{N}$. Letting $C_n$ be the preimage in $G$ of $R_n(B)$, the previous paragraph implies $C_n\in\mathrm{W}$. Since $G=\cup_{n\in\mathbb{N}} C_n$, Lemma \ref{lem:union_xi} ensures that $G\in\mathrm{W}$.
\end{proof}
Finally, we show that $W$ is closed under quotients.
\begin{lem}\label{lem:xi_quot}
If $G\in\mathrm{W}$ and $L\trianglelefteq G$, then $G/L\in \mathrm{W}$.
\end{lem}
\begin{proof} We argue by induction on $\xi(G)$. As the base case is immediate, suppose the lemma holds up to $\beta$ and let $G$ be such that $\xi(G)=\beta+1$. In view of Lemma~\ref{lem:union_xi}, we may assume $G$ is finitely generated, so $R_n(G)=G$ for all suitably large $n$. Say $k=\deg(G)$ and let $G_n$ be the subgroup corresponding to $n\in T^k(G)$.\par
\indent By the inductive hypothesis and Lemma~\ref{lem:rk_xi}, $G_nL/L\cong G_n/G_n\cap L\in \mathrm{W}$ for each $n$. On the other hand, $G/G_n\twoheadrightarrow(G/L)/(G_nL/L)$. Therefore, $(G/L)/(G_nL/L)$ is finite-by-abelian and so is in $W$. It now follows from Lemma \ref{lem:ext_xi} that $G/L\in W$.
\end{proof}
Combining Lemmas \ref{lem:xi_indp}, \ref{lem:finandab}, \ref{lem:union_xi}, \ref{lem:ext_xi}, and \ref{lem:xi_quot}, we obtain the following corollary.
\begin{cor}
If $G\in \mathop{\rm EG} \nolimits$, then $G\in\mathrm{W}$.
\end{cor}
We thus produce a characterization of elementary amenable groups.
\begin{thm}\label{thm:EA_char}
Let $G$ be a marked group. Then the following are equivalent:
\begin{enumerate}[(1)]
\item $G\in\mathop{\rm EG} \nolimits$ .
\item $T^l(G)$ is well-founded for all $l\geq 1$.
\item $T^l(G)$ is well-founded for some $l\geq 1$.
\end{enumerate}
\end{thm}
We can rephrase this to have the form of a chain condition independent of the marking. This corollary may thus be taken to be a definition of elementary amenability.
\begin{cor}
A countable group $G$ is elementary amenable if and only if there is no infinite descending sequence of the form
\[
G=G_0\geq G_1\geq\ldots
\]
such that for all $n\geq 0$, $G_n\neq\{e\}$ and there is a finitely generated subgroup $K_n\leq G_n$ with $G_{n+1}= [K_n,K_n]\cap H_n$ where $H_n$ is the intersection of the index-$(\leq(n+1))$ normal subgroups of $K_n$.
\end{cor}
\begin{proof}
Suppose $G\in\mathscr{G}$ and there is an infinite descending sequence
\[
G=G_0\geq G_1\geq\ldots
\]
as in the statement. Form $T^1(G)$, the decomposition tree of $G$ with offset $1$. We now proceed by induction to build $s_0\subsetneq s_1\subsetneq \dots$ with $s_i\in T^1(G)$ and $|s_i|=i$ such that $G_i\hookrightarrow G_{s_i}$. The base case is immediate: set $s_0=\emptyset$. Suppose we have defined $s_n$, so $G_n\hookrightarrow G_{s_n}$. Let $K_n\leq G_{n}$ be such that $G_{n+1}= [K_n,K_n]\cap H_n$ where $H_n$ is the intersection of the index-$(\leq n+1)$ normal subgroups of $K_n$. Since $K_n$ is finitely generated, there is $R_{m}(G_{s_n})$ such that $K_n\hookrightarrow R_{m}(G_{s_n})$. It follows that $G_{n+1}\hookrightarrow G_{s_n^{\smallfrown} m}$. Setting $s_{n+1}=s_n^{\smallfrown} m$, we have verified the inductive claim. The tree $T^1(G)$ thus has an infinite branch, so by Theorem~\rm\ref{thm:EA_char}, $G\notin\mathop{\rm EG} \nolimits$.
\medskip
Suppose there are no infinite descending sequences as in the statement and form $T^1(G)$. Let $s_0\subsetneq s_1\subsetneq \dots$ with $s_i\in T^1(G)$ and $|s_i|=i$. It suffices to show $s_0\subsetneq s_1\subsetneq \dots$ terminates, and so $T^1(G)$ is well-founded. This is indeed obvious since by construction the sequence of subgroups $G_{s_0}\geq G_{s_1}\geq\dots$ is a sequence of subgroups as in the chain condition.
\end{proof}
There are two main differences between this chain condition and the chain conditions explored in the earlier sections of this paper. First of all, $G_{n+1}$ is not related to $G_n$ only by being a subgroup. This is not unheard of; for example when looking at weak chain conditions one requires that $G_{n+1}$ be an \emph{infinite index} subgroup of $G_n$. The second difference is that the definition of $H_n$ changes with $n$. As far as we are aware, there are no widely-studied chain conditions defined in this way. That elementary amenability can be recast this way suggests that perhaps there are other interesting chain conditions with this property.
\subsection{$\mathop{\rm EG} \nolimits$ is not Borel}
We now study the descriptive-set-theoretic properties of $\mathop{\rm EG} \nolimits$. We show that on $\mathop{\rm EG} \nolimits$ the decomposition rank is unbounded below $\omega_1$.
\begin{lem}\label{lem:xiRankSucc}
For every $K\in \mathop{\rm EG} \nolimits$, there is $L\in \mathop{\rm EG} \nolimits$ with $\xi(K)<\xi(L)$.
\end{lem}
\begin{proof}
Let $G\in \mathop{\rm EG} \nolimits$ be as given by Proposition~\rm\ref{prop:EAembedding} for $K$ and form $L:=G\wr \mathbb{Z}$. Let $k=\deg(L)$, and take $L_i$ to be the subgroup of $L$ corresponding to $i\in T^k(L)$. Since $L$ is finitely generated, we may find $n$ such that $L=R_n(L)$. \par
\indent We now consider $L_n$. The group $[L,L]=[R_n(L),R_n(L)]$ certainly contains $[M,M]=[H,H]^{<\mathbb{Z}}$. On the other hand, if $N\trianglelefteq L$ has index $k+1$, there is $n\in \mathbb{Z}\setminus \{0\}$ so that $n\in N$. Applying Lemma~\ref{lem:wreath}, $[G,G]\leq N$, so $[H,H]^{<\mathbb{Z}}\leq L_n$. The group $K$ thus embeds into $[H,H]$, and Lemma~\rm\ref{lem:sgrp_xi} implies $\xi(K)\leq \xi(L_n)$. Appealing to Lemma~\rm\ref{lem:rk_xi}, we conclude $\xi(K)< \xi(L)$ proving the lemma.
\end{proof}
Our next lemma follows immediately from Corollary~\rm\ref{lem:sgrp_xi} by taking the direct sum.
\begin{lem}\label{lem:xiRankLimit}
Let $\{A_i\}_{i\in \mathbb{N}}$ be countable groups. If $A_i\in \mathop{\rm EG} \nolimits$ for all $i\in \mathbb{N}$, then there is $A\in \mathop{\rm EG} \nolimits$ with $\xi(A)\geq \xi(A_i)$ for all $i\in \mathbb{N}$.
\end{lem}
\begin{lem}
For all $\beta<\omega_1$, there is $G\in \mathop{\rm EG} \nolimits$ such that $\xi(G)\geq \beta$.
\end{lem}
\begin{proof}
The proof is the same as that of Lemma \ref{lem:CentRankUnbdd}, with Lemmas \ref{lem:xiRankSucc} and \ref{lem:xiRankLimit} referenced at the appropriate places.
\end{proof}
\begin{lem}\label{lem:xi_unbounded}
For each $\beta<\omega_1$, there is a finitely generated $G\in \mathop{\rm EG} \nolimits$ such that $\xi(G)\geq \beta$.
\end{lem}
\begin{proof}
Let $H\in\mc \mathop{\rm EG} \nolimits$ be a group such that $\xi(H)\geq\beta$. Proposition~\rm\ref{prop:EAembedding} implies that $H$ embeds into a 2-generated group $G\in\mc \mathop{\rm EG} \nolimits$. By Corollary~\ref{lem:sgrp_xi}, $\xi(G)\geq\xi(H)\geq\beta$.
\end{proof}
\begin{thm}\label{thm:EA_nonborel}
$\mathop{\rm EG} \nolimits$ is a non-Borel $\Pi^1_1$ set in $\ms{G}$, and $\mathop{\rm EG} \nolimits \cap \mathscr{G}_{fg}$ is a non-Borel $\Pi^1_1$ set in $\ms{G}_{fg}$.
\end{thm}
\begin{proof}
This follows from Theorem \ref{thm:BddnessThm}, Lemma \ref{lem:Phi_borel}, and Lemma~\rm\ref{lem:xi_unbounded} along with the facts that $\xi(G)\leq \rho(T^1(G))$ and that $\rho\circ \Phi^1$ is a $\Pi^1_1$-rank on $\mathop{\rm EG} \nolimits$.
\end{proof}
Let $\mathop{\rm AG} \nolimits\subseteq\mathscr{G}$ denote the class of countable amenable groups. Via Theorem~\rm \ref{thm:EA_nonborel}, we now may give a non-constructive answer to an old question of Day \cite{D57}, which was open until Grigorchuk \cite{G84} constructed groups of intermediate growth: \textit{Is it the case that every amenable group is elementary amenable}?
\begin{cor}
There is a finitely generated amenable group that is not elementary amenable.
\end{cor}
\begin{proof}
It is well-known that $\mathop{\rm AG} \nolimits$ is Borel; see Lemma~\rm\ref{lem:AG_borel} for a proof. The set $\mathop{\rm AG} \nolimits\cap\mathscr{G}_{fg}$ is thus Borel. On the other hand, Theorem~\rm \ref{thm:EA_nonborel} gives that $\mathop{\rm EG} \nolimits\cap \mathscr{G}_{fg}$ is not Borel. We conclude that $\mathop{\rm EG} \nolimits\cap \mathscr{G}_{fg} \subsetneq \mathop{\rm AG} \nolimits\cap \mathscr{G}_{fg}$.
\end{proof}
\subsection{Further observations}
By a result of C. Chou \cite[Proposition 2.2.]{Ch80}, the class of elementary amenable groups is the smallest class of countable discrete groups that satisfies (i),(ii), and (iii) of Definition~\rm\ref{def:EA}. Chou's theorem suggests a natural ranking of elementary amenable groups different than our decomposition rank. Indeed, after \cite{Os02}, define
\begin{enumerate}[$\bullet$]
\item $G\in \mathop{\rm EG} \nolimits_0$ if and only if $G$ is finite or abelian.
\item Suppose $\mathop{\rm EG} \nolimits_{\alpha}$ is defined. Put $G\in \mathop{\rm EG} \nolimits_{\alpha}^e$ if and only if there exists $N\trianglelefteq G$ such that $N\in \mathop{\rm EG} \nolimits_{\alpha}$ and $G/N\in \mathop{\rm EG} \nolimits_{0}$. Put $G\in \mathop{\rm EG} \nolimits_{\alpha}^l$ if and only if $G=\bigcup_{i\in \mathbb{N}}H_i$ where $(H_i)_{i\in \mathbb{N}}$ is an $\subseteq$-increasing sequence of subgroups of $G$ with $H_i\in \mathop{\rm EG} \nolimits_{\alpha}$ for each $i\in \mathbb{N}$. Set $\mathop{\rm EG} \nolimits_{\alpha+1}:=\mathop{\rm EG} \nolimits_{\alpha}^e\cup \mathop{\rm EG} \nolimits_{\alpha}^l$.
\item For $\lambda$ a limit ordinal, $\mathop{\rm EG} \nolimits_{\lambda}:=\bigcup_{\beta<\lambda}\mathop{\rm EG} \nolimits_{\beta}$.
\end{enumerate}
By a result of D. Osin \cite[Lemma 3.2]{Os02}, $\bigcup_{\alpha <\omega_1}\mathop{\rm EG} \nolimits_{\alpha}$ is closed under group extension. It now follows from Chou's theorem that $\mathop{\rm EG} \nolimits=\bigcup_{\alpha <\omega_1}\mathop{\rm EG} \nolimits_{\alpha}$. One may then define for $G\in \mathop{\rm EG} \nolimits$
\[
\mathop{\rm rk}(G):=\min\{\alpha\;|\;G\in \mathop{\rm EG} \nolimits_{\alpha}\}.
\]
\noindent We call $\mathop{\rm rk}(G)$ the \textbf{construction rank} of $G$.
We now compare $\xi$ and $\mathop{\rm rk}$ and in the process mostly recover a theorem of Olshanskii and Osin.
\begin{prop}\label{prop:rk_xi}
For $G\in \mathop{\rm EG} \nolimits$, $\mathop{\rm rk}(G)\leq 3\xi(G)$.
\end{prop}
\begin{proof}
We induct on $\xi(G)$ for the proposition. For the base case, if $\xi(G)=1$, then $G=\{e\}$, and the inductive claim obviously holds. Suppose the proposition holds up to $\beta$. Say $\xi(G)=\beta+1$ and $\deg(G)=k$. Then $\xi(G_i)\leq \beta$ for each $G_i$ associated to $i\in T^k(G)$, and applying the inductive hypothesis, $\mathop{\rm rk}(G_i)\leq 3\xi(G_i)$.\par
\indent On the other hand, $R_i(G)/G_i$ is finite-by-abelian, say an extension of the group $A$ by the group $F$. Letting $F_0$ be the inverse image of $F$ in $R_i(G)$ under the usual projection, $\mathop{\rm rk}(F_0)\leq\mathop{\rm rk}(G_i)+1$, and $R_i(G)/F_0\cong A$. Hence,
\[
\mathop{\rm rk}(R_i(G))\leq (\mathop{\rm rk}(G_i)+1)+1 \leq 3\xi(G_i)+2.
\]
We conclude
\[
\mathop{\rm rk}(G)\leq \sup_{i\in \mathbb{N}}\left(3\xi(G_i)+2 \right)+1\leq 3(\beta+1)=3\xi(G).
\]
This finishes the induction.
\end{proof}
Bounding $\xi$ from above by $\mathop{\rm rk}$ involves a bit more work. We begin with a general lemma for well-founded trees.
\begin{lem}\label{lem:wf_tree}
Suppose $T$ is a well-founded tree. Then
\[
\rho(T)\leq \sup_{|s|=k}\rho(T_s) + k.
\]
\end{lem}
\begin{proof}
We argue by induction on $|s|$. For the base case, $|s|=1$,
\[
\rho(T)=\rho_T(\emptyset)+1=\sup_{i\in T}\left(\rho_T(i)+1\right)+1=\sup_{i\in \mathbb{N}}\rho(T_i)+1.
\]
\indent Supposing the lemma holds up to length $k$,
\[
\rho(T)\leq \sup_{|s|=k}\rho(T_s)+k\leq \sup_{|s|=k}\left(\sup_{s^{\smallfrown} i\in T}\rho(T_{s^{\smallfrown} i})+1\right)+k\leq\sup_{|s|=k+1}\rho(T_s)+k+1
\]
completing the induction.
\end{proof}
\begin{prop}\label{prop:xi_rk_ub}
For $G\in \mathop{\rm EG} \nolimits$,
\[
\rho\left(T^1(G)\right)\leq \omega(\mathop{\rm rk}(G)+1).
\]
In particular, $\xi(G)\leq \omega(\mathop{\rm rk}(G)+1)$.
\end{prop}
\begin{proof}
We argue by induction on $\mathop{\rm rk}(G)$. For the base case, $\mathop{\rm rk}(G)=0 $, $G$ is either finite or abelian. There is thus $m\geq 1$ such that every element of $T^1(G)$ has length at most $m$. It follows that $\rho(T^1(G))$ is finite, which proves the base case.\par
\indent Suppose the lemma holds up to $\alpha$ and $\mathop{\rm rk}(G)=\alpha+1$. Let us consider first the case that the construction rank is given by a countable increasing union; say $G=\bigcup_{n\in \omega}H_n$ with $\mathop{\rm rk}(H_n)\leq \alpha$ for each $n$. Since $R_i(G)$ is finitely generated, there is $n(i)$ for which $G_i\leq H_{n(i)}$. We apply the inductive hypothesis and Lemma~\rm\ref{lem:xi_indp} to conclude
\[
\rho\left(T^2(G_i)\right)\leq \rho(T^1(G_i))\leq \omega(\alpha+1).
\]
Hence,
\[
\rho\left(T^1(G)\right)=\sup_{i\in \omega}\rho\left(T^2(G_i)\right) +1 \leq \omega\cdot \alpha +\omega+1\leq \omega(\alpha+2),
\]
verifying the hypothesis in this case. \par
\indent We now consider the case $\mathop{\rm rk}(G)$ is given by a group extension. Suppose $H\trianglelefteq G$ is such that $\mathop{\rm rk}(H)=\alpha$ and $\mathop{\rm rk}(G/H)=0$. If $G/H$ is abelian, $G_i\leq H$ for each $i$. Hence, $\mathop{\rm rk}(G_i)\leq \alpha$, and the desired result follows just as in the increasing union case. Suppose $G/H$ is finite. We may find $k$ such that for all $s\in T^1(G)$ with $|s|=k$, $G_s\leq H$. Applying the inductive hypothesis and Lemma~\rm\ref{lem:xi_indp},
\[
\rho\left(T^{k+1}(G_s)\right)\leq \rho\left(T^{1}(G_s)\right)\leq \omega(\alpha+1).
\]
Lemma~\rm\ref{lem:wf_tree} now implies
\[
\rho\left(T^1(G)\right)\leq \sup_{|s|=k}\rho\left(T^1(G)_s\right)+k\leq \omega(\alpha+1)+k\leq \omega(\alpha+2).
\]
This completes the induction, and we conclude the proposition.
\end{proof}
As a corollary to Lemma~\rm\ref{lem:xi_unbounded} and Proposition~\rm\ref{prop:xi_rk_ub}, we obtain a less detailed version of a theorem from the literature.
\begin{cor}[Olshanskii, Osin {\cite[Corollary 1.6]{OO13}}]\label{cor:OlOs}
For every ordinal $\alpha<\omega_1$, there is $G\in \mathop{\rm EG} \nolimits\cap \mathscr{G}_{fg}$ such that $\alpha\leq\mathop{\rm rk}(G)$. The function $\mathop{\rm rk}\colon\mathop{\rm EG} \nolimits\cap \mathscr{G}_{fg}\to ORD$ is thus unbounded below $\omega_1$.
\end{cor}
\begin{proof} Suppose for contradiction $\alpha <\omega_1$ is such that $\mathop{\rm rk}(G)< \alpha$ for all $G\in \mathop{\rm EG} \nolimits$. By Proposition~\rm\ref{prop:xi_rk_ub}, $\xi(G)\leq \omega(\alpha+1)<\omega_1$ for all $G\in\mathop{\rm EG} \nolimits$, contradicting Lemma~\rm\ref{lem:xi_unbounded}.
\end{proof}
In our proof of Theorem~\rm\ref{thm:EA_nonborel}, we use that $\rho\circ\Phi^1$ is a $\Pi^1_1$-rank. It is natural to ask if $\xi$ itself is a $\Pi^1_1$-rank. This is indeed the case.
\begin{thm}
The decomposition rank is a $\Pi^1_1$-rank on $\mathop{\rm EG} \nolimits$.
\end{thm}
\begin{proof}
Each of the ranks $\phi_l:=\rho\circ \Phi^l$ is a $\Pi^1_1$-rank on $\mathop{\rm EG} \nolimits$ where $\Phi^l$ is as defined in Lemma~\rm\ref{lem:Phi_borel}. Let $\leq_l^{\Pi}\subseteq \mathscr{G}\times \mathscr{G}$ and $\leq_l^{\Sigma} \subseteq \mathscr{G}\times \mathscr{G}$ be the relations given by $\phi_l$ as a $\Pi^1_1$-rank. We now consider the following relations:
\[
\leq_{\xi}^{\Pi}:=\bigcup_{N\in \mathbb{N}}\bigcap _{l\geq N}\leq_l^{\Pi}\text{ and } \leq_{\xi}^{\Sigma}:=\bigcup_{N\in \mathbb{N}}\bigcap _{l\geq N}\leq_l^{\Sigma}
\]
Since co-analytic and analytic sets are closed under countable unions and intersections, $\leq_{\xi}^{\Pi}$ is co-analytic and $\leq_{\xi}^{\Sigma}$ is analytic. To conclude $\xi$ is a $\Pi^1_1$-rank, it thus remains to show for $H\in \mathop{\rm EG} \nolimits$,
\begin{align*}
G\in \mathop{\rm EG} \nolimits \,\wedge\, \xi(G)\leq \xi(H) &\Leftrightarrow G \leq_{\xi}^\Sigma H \\
&\Leftrightarrow G \leq_\xi^\Pi H.
\end{align*}
Suppose $G\in \mathop{\rm EG} \nolimits$ and $\xi(G)\leq \xi(H)$. Letting $M:=\max\{\deg(G),\deg(H)\}$, we see that $\rho\left(T^k(G) \right)\leq \rho \left(T^k(H) \right)$ for all $k\geq M$ via Lemma~\rm\ref{lem:xi_indp}, hence $\phi_k(G)\leq \phi_k(H)$ for $k\geq M$. We conclude that $G \leq_{\xi}^{\Pi} H$ and $G\leq_{\xi}^{\Sigma} H$.\par
\indent Conversely, suppose $G \leq_{\xi}^{\Pi} H$ and $G\leq_{\xi}^{\Sigma} H$ and let $M\geq 0$ be such that $G \leq_{k}^{\Pi} H$ and $G \leq_{k}^{\Sigma} H$ for all $k\geq M$. Immediately, $G\in \mathop{\rm EG} \nolimits$. For each $k\geq M$, we further see $\phi_k(G)\leq \phi_k(H)$, and taking $k=\max\{\deg(G),\deg(H),M\}$,
\[
\xi(G)=\phi_k(G)\leq \phi_k(H)=\xi(H).
\]
Therefore, $\xi$ is a $\Pi^1_1$-rank.
\end{proof}
Propositions \ref{prop:rk_xi} and \ref{prop:xi_rk_ub} combine to give us
$$ \xi(G) \leq \omega(\mathop{\rm rk}(G)+1) \leq \omega(3\xi(G)+1), $$
so $\mathop{\rm rk}$ is closely related to a $\Pi^1_1$-rank. Given this close relationship, it is natural to ask whether or not $\mathop{\rm rk}$ is a $\Pi^1_1$-rank. We suspect, however, that $\mathop{\rm rk}$ is \textit{not} a $\Pi_1^1$-rank as the sets $\mathop{\rm rk}^{-1}(\alpha)$ are likely analytic and non-Borel for suitably large $\alpha$; in fact, we believe $\mathop{\rm rk}^{-1}(2)$ is analytic and non-Borel. Indeed, if $\mathop{\rm EG} \nolimits_\alpha$ is Borel and uncountable, then $\mathop{\rm EG} \nolimits^e_{\alpha}$ is defined by quantifying over $\mathop{\rm EG} \nolimits_\alpha$. We thus expect $\mathop{\rm EG} \nolimits_{\alpha}^e$ to be analytic and, barring some clever argument, non-Borel. (We remark that one can make such a clever argument in the case of $\mathop{\rm EG} \nolimits^e_0$, but it does not seem to work beyond that.) We do not pursue this question further as it is tangential to the aim of this work and somewhat technical.
\section{Borel functions and sets}\label{sec:Borel}
In previous sections we made claims that certain maps and sets were Borel, and from this and the Boundedness Theorem \ref{thm:BddnessThm}, we concluded that certain subsets of $\mathscr{G}$ were not Borel. A slogan from descriptive set theory is ``Borel = explicit'' meaning if you describe a map or set without an appeal to something like the axiom of choice or quantifying over an uncountable space, it should be Borel. As the maps and sets from previous sections are ``explicit'' in this sense, we were content to state they were Borel without further proof. To those not as familiar with descriptive set theory, we offer this section to verify our previous claims.
Recall that $\mathscr{G}=\{ N\trianglelefteq\mathbb{F}_{\omega} \}$ and that we identify $N$ with the group $\mathbb{F}_{\omega}/N$. We make frequent use of the usual projection from $\mathbb{F}_{\omega}$ to $\mathbb{F}_{\omega}/N$ and always denote this projection by $f_N$. Every countable group is identified with an element of $\mathscr{G}$; in fact, a given group $G$ corresponds to many distinct elements of $\mathscr{G}$ as there are many different surjections of $\mathbb{F}_{\omega}$ onto $G$. We fix an enumeration $(\gamma_i)_{i\in \mathbb{N}}$ for $\mathbb{F}_{\omega}$, and this gives rise to an enumeration of $G$ in the obvious way. Let us also enumerate the generators for $\mathbb{F}_{\omega}$ as $(a_i)_{i\in \mathbb{N}}$. Recall finally that $\mc G_{fg} = \cup_{n\in\mathbb{N}} \{ N \trianglelefteq \mathbb{F}_{\omega} \mid \forall k\geq n \;a_k\in N \}$. This is an $F_\sigma$ subset of $\mathscr{G}$. In particular, its Borel sets as a Borel space are precisely those sets of the form $B\cap\mathscr{G}_{fg}$ where $B\subseteq\mathscr{G}$ is Borel.
\subsection{Borel functions}
The sub-basic open sets of $\mathscr{G}$ are those of the form $O_{\gamma}=\{ N \mid \gamma \in N\}$ and their complements. The Borel $\sigma$-algebra on $\mathscr{G}$ is thus generated by the $O_\gamma$, so in order to show a map $\psi \colon\mathscr{G}\to\mathscr{G}$ is Borel, we need only to check that $\psi^{-1}(O_\gamma)$ is Borel for all $\gamma\in\mathbb{F}_{\omega}$.\par
\indent We begin with the easier examples of Borel maps.
\begin{lem}\label{lem:QuotientBorel}
For each $\delta\in\mathbb{F}_{\omega}$, there is a Borel map $Q_\delta\colon\mathscr{G}\to\mathscr{G}$ such that if $N\in\mathscr{G}$ with $\mathbb{F}_{\omega}/N\cong G$, then $\mathbb{F}_{\omega}/Q_\delta(N)\cong G/\ngrp{f_N(\delta)}$.
\end{lem}
\begin{proof}
Since $G/\ngrp{f_N(\delta)}\cong \mathbb{F}_{\omega}/\ngrp{N,\delta}$, the map $Q_\delta(N):=\ngrp{\delta}N$ meets our requirements. We need only check that it is Borel. For this,
\begin{align*}
Q_\delta^{-1}(O_\gamma) &= \{ N\in\mathscr{G} \mid \gamma\in\ngrp{\delta}N \} \\
&= \{ N\in\mathscr{G} \mid \exists g\in\ngrp{\delta} \; g^{-1}\gamma\in N \} \\
&= \bigcup_{g\in\ngrp{\delta}} \{ N\in\mathscr{G} \mid g^{-1}\gamma\in N\}
\end{align*}
which is open, so we have verified the lemma.
\end{proof}
We can now easily prove Lemma \ref{lem:MaxNMapBorel}.
\begin{proof}[Proof of Lemma \ref{lem:MaxNMapBorel}]
By repeated composition, we may define $Q_s\colon\mathscr{G}\to\mathscr{G}$ for all $s\in\mathbb{N}^{<\mathbb{N}}\setminus \{\emptyset\}$ so that
\[
Q_s(N)=\ngrp{\gamma_s}N;
\]
we define $Q_{\emptyset}:=id$. The previous lemma ensures these maps are Borel.\par
\indent Now suppose $t\in\mathbb{N}^{<\mathbb{N}}$ is of the form $v^{\smallfrown} i$ with $v\in\mathbb{N}^{<\mathbb{N}}$ and $i\in\mathbb{N}$ and consider the basic open set $O_t:=\{T\in Tr\mid t\in T\}$ of $Tr$. We see that
\[
\Phi_{M_n}^{-1}(O_t)=\{ N\in\mathscr{G} \mid Q_v(N)\neq Q_t(N)\},
\]
which is Borel. The map $\Phi_{M_n}$ is thus Borel.
\end{proof}
\begin{lem}\label{lem:Rn_borel}
For each $n\geq 0$, there is a Borel map $R_n\colon\mathscr{G}\to\mathscr{G}$ such that if $N\in\mathscr{G}$ with $\mathbb{F}_{\omega}/N\cong G$, then $\mathbb{F}_{\omega}/R_n(N)\cong\<g_0,\ldots,g_n\>$.
\end{lem}
\begin{proof}
Let $\pi_n\colon\mathbb{F}_{\omega}\to\mathbb{F}_{\omega}$ be induced by mapping the generators $(a_i)_{i\in \mathbb{N}}$ as follows:
\[\pi_n(a_i) =
\begin{cases}
\gamma_i, & 0\leq i\leq n \\
e, & \text{otherwise.}
\end{cases}
\]
Suppose that $N\in\mathscr{G}$ with $\mathbb{F}_{\omega}/N=G$. The function $f_N\circ\pi_n\colon\mathbb{F}_{\omega}\to \grp{g_0,\dots,g_n}$ is then a surjection. We thus define $R_n$ to be the map sending $N$ to $\ker (f_N\circ\pi_n)$. Since $\mathbb{F}_{\omega}/\ker (f_N\circ\pi_n)\cong \grp{g_0,\dots,g_n}$, this works as intended. \par
\indent As $\gamma\in\ker (f_N\circ\pi_n)$ iff $\pi_n(\gamma)\in\ker (f_N)$, we conclude that
\[
R_n^{-1}(O_\gamma) = \{ M \in\mathscr{G} \mid \pi_n(\gamma)\in M \},
\]
which is the open set $O_{\pi_n(\gamma)}$. The map $R_n$ is thus Borel.
\end{proof}
The above proof works for subgroups generated by any fixed collection of elements of $G$; that is to say, the same proof shows the maps $G\mapsto G_s$ defined in Section \ref{sec:Max} are Borel, so as before, we get Lemma \ref{lem:MaxMapBorel} as a corollary.
We now move on to proving Lemma \ref{lem:CentMapBorel}; this follows from the next lemma. Its proof is more involved than the previous two.
\begin{lem}
For each $s\in\mathbb{N}^{<\mathbb{N}}\setminus \{\emptyset\}$, there is a Borel map $C_s\colon\mathscr{G}\to\mathscr{G}$ such that if $N\in\mathscr{G}$ with $\mathbb{F}_{\omega}/N\cong G$, then $\mathbb{F}_{\omega}/C_s(N)\cong C_G(\{g_s\})$.
\end{lem}
\begin{proof}
Suppose that $N\in\mathscr{G}$ and $\mathbb{F}_{\omega}/N\cong G$. Define $\pi_N\colon\mathbb{F}_{\omega}\to\mathbb{F}_{\omega}$ by
\[
\pi_N(a_j):=
\begin{cases}
\gamma_j, & \text{ if }f_N(\gamma_j)\in C_G\left(\{f_N(\gamma_s)\}\right) \\
e, & \text{ else.}
\end{cases}
\]
The map $f_N\circ \pi_N\colon\mathbb{F}_{\omega}\to C_G(\{g_s\})$ is then a surjection, so the map $N\mapsto\ker(f_N\circ\pi_N)$ works as intended. In order to check it is Borel, we introduce the set
\[
S_j:=\{N\in\mathscr{G} \mid \pi_N(a_j)=\gamma_j\}.
\]
Since $f_N(\gamma_j)\in C_G(\{f_N(\gamma_s)\})$ iff $[\gamma_j,\gamma_{s_i}]\in N$ for each $0\leq i \leq |s|-1$,
\[
S_j=\{N\in\mathscr{G} \mid [\gamma_j,\gamma_{s_i}]\in N \text{ for each }0\leq i \leq |s|\},
\]
which is an open set.
We now fix a word $\delta=\delta(a_0,\dots,a_m)\in \mathbb{F}_{\omega}$ and consider the pre-image of the basic open set $O_{\delta}$. Our notation $\delta(a_0,\dots,a_m)$ indicates the word $\delta$ only uses the letters appearing in the parentheses. We may evaluate $\pi_N(\delta)$ by substituting in the images of $a_0,\dots,a_m$, so $\pi_{N}(\delta)=\delta(x_0,\dots,x_m)$ for some $\ol{x}:=(x_0,\dots,x_m)\in \Omega:=\prod_{i=0}^m\{\gamma_i,e\}$ that depends on $N$. The set of $N\in \mathscr{G}$ such that $\pi_N(\delta(a_0,\ldots,a_m))=\delta(\ol{x})$ for some fixed $\ol{x}$ is the Borel set
\[
S_{\ol{x}}:=\bigcap_{x_j=\gamma_j} S_j \cap \bigcap_{x_k=e} S_k^c.
\]
Since $\delta\in\ker(f_N\circ \pi_N)$ iff $\pi_N(\delta)\in\ker f_N=N$, we now see that
\begin{align*}
C_s^{-1}(O_\delta) &= \{N\in\mathscr{G} \mid \pi_N(\delta)\in N\} \\
&= \bigcup_{\ol{x}\in\Omega} \left( \{N\in\mathscr{G} \mid \delta(\ol{x})\in N\} \cap S_{\ol{x}} \right)
\end{align*}
which is Borel.
\end{proof}
We next show the maps $S_k$ from Section \ref{sec:EAGroups} are Borel. The main idea is the same as in previous lemma.
\begin{lem}\label{lem:Sk_Borel}
For each $k\geq 1$, there is a Borel map $S_k:\mathscr{G}_{fg}\rightarrow \mathscr{G}$ such that if $\mathbb{F}_{\omega}/N=G$, then
\[
\mathbb{F}_{\omega}/S_k(N)\cong [G,G]\cap \bigcap \mc{N}_k(G)
\]
where $\mc{N}_k(G):=\{M\trianglelefteq G\mid |G:M|\leq k+1\}$.
\end{lem}
\begin{proof}
Suppose $N\in \mathscr{G}_{fg}$ and $G\cong\mathbb{F}_{\omega}/N$. Similarly to the previous lemma, we define $\pi_N\colon\mathbb{F}_{\omega}\to\mathbb{F}_{\omega}$ by
\[
\pi_N(a_i):=
\begin{cases}
\gamma_i, & \text{ if }f_N(\gamma_i)\in [G,G]\cap\bigcap\mc{N}_k(G)\\
e, & \text{ else.}
\end{cases}
\]
Define $S_k:\mathscr{G}_{fg}\rightarrow \mathscr{G}$ by $N\mapsto \ker(f_N\circ\pi_{N})$; this map behaves as desired. We claim this map is also Borel. \par
\indent Define
\[
\mc N_k := \left\{ M\in\mathscr{G}_{fg} \mid |\mathbb{F}_{\omega} : M|\leq k+1 \right\}.
\]
If $N\in\mathscr{G}_{fg}$, then the collection of index-$\leq k+1$ subgroups of $\mathbb{F}_{\omega}/N$ is precisely $\{ MN/N \mid M\in\mc N_k \}$. Therefore, $f_N(\gamma_i)\in [G,G]\cap\bigcap\mc N_k(G)$ iff $\gamma_i\in [\mathbb{F}_{\omega},\mathbb{F}_{\omega}]N\cap \bigcap_{M\in\mc N_k} MN $. As in the previous lemma, we may define
\begin{align*}
S_i &:= \left\{N\in \mathscr{G}_{fg} \mid \pi_N(a_i)=\gamma_i \right\}\\
&= \left\{N\in \mathscr{G}_{fg} \mid \gamma_i \in [\mathbb{F}_{\omega},\mathbb{F}_{\omega}]N\cap \bigcap_{M\in\mc N_k} MN \right\}\\
&= \bigcup_{\delta\in[\mathbb{F}_{\omega},\mathbb{F}_{\omega}]} \left\{N\in\mathscr{G}_{fg} \mid \delta^{-1}\gamma_i\in N \right\} \cap \bigcap_{M\in \mc{N}_k}\bigcup_{\delta\in M}\left\{N\in\mathscr{G}_{fg} \mid \delta^{-1}\gamma_i\in N\right\}.
\end{align*}
The last set is Borel since $\mc{N}_k$ is countable. Given $\ol{x}:=(x_0,\dots,x_m)\in \Omega:=\prod_{i=0}^m\{\gamma_i,e\}$, we define as before $S_{\ol{x}}$. \par
\indent Fixing a word $\delta=\delta(a_0,\dots,a_m)\in \mathbb{F}_{\omega}$, we now consider the pre-image of the basic open set $O_{\delta}$. We see
\begin{align*}
S_k^{-1}(O_\delta) &= \{N\in\mathscr{G} \mid \pi_N(\delta)\in N\} \\
&= \bigcup_{\ol{x}\in\Omega} \big( \{N\in\mathscr{G} \mid \delta(\ol{x})\in N\} \cap S_{\ol{x}} \big)
\end{align*}
which is Borel.
\end{proof}
\indent Using Lemma~\ref{lem:Rn_borel} and \ref{lem:Sk_Borel}, we build Borel maps $\Psi^l_s:\ms{G}\rightarrow \ms{G}$ for each $l\in\mathbb{N}$ and $s\in \mathbb{N}^{<\mathbb{N}}$. For $s=\emptyset$, put $\Psi^l_{\emptyset}=id$. Supposing we have defined $\Psi^l_s$, define $\Psi^l_{s^{\smallfrown} n}$ by
\[
\Psi^l_{s^{\smallfrown} n}(N):=S_{|s|+l}\circ R_n(\Psi^l_s(N)).
\]
It follows that if $s\in T^l(G)$ with $G=\mathbb{F}_{\omega}/N$, then $\mathbb{F}_{\omega}/\Psi^l_s(N)= G_s$. If $s\notin T^l(G)$, then $\mathbb{F}_{\omega}/\Psi^l_s(N)=\{e\}$.
\begin{proof}[Proof of Lemma \ref{lem:Phi_borel}]
\indent Fixing $s\in \mathbb{N}^{<\mathbb{N}}$ and $l\in\mathbb{N}$,
\[
(\Phi^l)^{-1}(O_s)=\left\{N\in \ms{G} \mid s\in T^l(\mathbb{F}_{\omega}/N)\right\}.
\]
If $s=\emptyset$, then $(\Phi^l)^{-1}(O_s)=\ms{G}$ which is plainly Borel. Else, say $s=r^{\smallfrown} n$, so
\[
\begin{array}{ccl}
(\Phi^l)^{-1}(O_s) & = & \left\{N\in \ms{G} \mid r^{\smallfrown} n \in T^l(\mathbb{F}_{\omega}/N)\right\}\\
& = & \left\{N\in \ms{G} \mid (\mathbb{F}_{\omega}/N)_r\neq \{e\}\right\}\\
& = & (\Psi^l_r)^{-1}(\ms{G}\setminus \{e\}),
\end{array}
\]
which is Borel.
\end{proof}
\subsection{Borel sets}
Recall that $\mathop{\rm AG} \nolimits$ denotes the class of countable amenable groups.
\begin{lem}[Folklore]\label{lem:AG_borel}
The set $\mathop{\rm AG} \nolimits$ is Borel in $\ms{G}$, and therefore, $\mathop{\rm AG} \nolimits\cap \mathscr{G}_{fg}$ is Borel.
\end{lem}
\begin{proof}
Amenable groups are characterized by F\o lner's property: A countable group $G$ is amenable if and only if for every finite $F\subseteq G$ and every $n\geq 1$, there is a finite non-empty subset $K\subseteq G$ such that
\[
\frac{|xK\Delta K|}{|K|}\leq \frac{1}{n}
\]
for all $x\in F$ where $\Delta$ denotes the symmetric difference.\par
\indent Letting $P_f(\mathbb{F}_{\omega})$ be the collection of finite subsets of $\mathbb{F}_{\omega}$, we infer
\[
\mathop{\rm AG} \nolimits=\bigcap_{F\in P_f(\mathbb{F}_{\omega})}\bigcap_{n\geq 1}\bigcup_{K\in P_f(\mathbb{F}_{\omega})}\bigcap_{x\in F}\left\{N\in \mathscr{G} \mid \frac{|f_N(x)f_N(K)\Delta f_N(K)|}{|f_N(K)|}\leq \frac{1}{n}\right\}.
\]
It thus suffices to show
\[
\Omega:=\left\{N\in \mathscr{G} \mid \frac{|f_N(x)f_N(K)\Delta f_N(K)|}{|f_N(K)|}\leq \frac{1}{n}\right\}
\]
is Borel. It is easy to see requiring $|f_N(K)|=m$ and $|f_N(x)f_N(K)\Delta f_N(K)|=l$ is Borel, hence
\[
\Omega=\bigcup_{\frac{l}{m}\leq \frac{1}{n}}\left\{N\mid |f_N(x)f_N(K)\Delta f_N(K)|=l\text{ and } |f_N(K)|=m \right\}
\]
is Borel. The set $\mathop{\rm AG} \nolimits$ is thus Borel.
\end{proof}
\section{Further remarks}\label{sec:Remarks}
Our results give tools to study groups enjoying any of the other chain conditions in the literature. Perhaps more interestingly, our results suggest new questions concerning elementary amenable groups and groups with the minimal condition on centralizers, maximal condition on subgroups, and maximal condition on normal subgroups. \par
\indent Most immediately, one desires a better understanding of the various rank functions. In the case of max groups, there are no infinite subgroup rank two groups, the infinite groups with subgroup rank 3 are Tarski monsters, and $\mathbb{Z}$ has rank $\omega+1$. In the case of max-n, examples of finite rank groups are easy to produce and understand; however, transfinite rank examples are somewhat mysterious. Following Olshanskii and Osin, cf. \cite[Corollary 1.6]{OO13}, we ask
\begin{quest}
For which ordinals $\alpha$ is there an infinite group in $\mc M_C$ ($\mc M_{\max{}}, \mc M_n$) such that the centralizer rank (subgroup rank, length) is $\alpha$?
\end{quest}
\indent In a different direction, showing a set is non-Borel in $\mathscr{G}$ demonstrates there is no ``simple" definition of the class. Our techniques give a way to determine if a subset of $\mathscr{G}$ (or of a Borel subset of $\mathscr{G}$) given by a chain condition is not Borel and hence to determine if it does not admit a ``simple" characterization. In the setting of max-n groups, there is a particularly intriguing question along these lines. By an old result of Hall, a two-step solvable group is max-n if and only if it is finitely generated; this is certainly a Borel condition. On the other hand, no such nice characterization of three-step solvable groups with max-n is known. We thus ask
\begin{quest}
Is the set of max-n three-step solvable marked groups Borel?
\end{quest}
\indent In a similar vein, our results on elementary amenable groups, in a sense, show elementary amenable groups are not ``elementary". One naturally asks
\begin{quest}[Hume] Is there an intermediate ``elementary" Borel set between $\mathop{\rm EG} \nolimits\cap \mathscr{G}_{fg}$ and $\mathop{\rm AG} \nolimits\cap \mathscr{G}_{fg}$? More precisely, is there an elementary class $\ms{E}(B)$ in the sense of Osin \cite{Os02} with $B$ ``small" such that $\mathop{\rm EG} \nolimits\cap \mathscr{G}_{fg}\subseteq \ms{E}(B)\subsetneq \mathop{\rm AG} \nolimits\cap\mathscr{G}_{fg}$ and $\ms{E}(B)$ is Borel?
\end{quest}
We also arrive at new questions with a descriptive-set-theoretic flavor.
\begin{defn}
Let $Y$ be a uncountable Polish space. A set $A\subseteq Y$ is \textbf{$\Pi^1_1$-complete} if $A$ is $\Pi^1_1$ and for all $B\subseteq X$ with $X$ an uncountable Polish space and $B$ co-analytic, $B$ Borel reduces to $A$.
\end{defn}
The idea is that $\Pi^1_1$-complete sets are as complicated as they possibly could be; Theorem \ref{thm:WFComplete} says that $WF\subseteq Tr$ is $\Pi^1_1$-complete.
\begin{quest}
Are any of $\mc M_C,\mc M_{\max{}},\mc M_n,$ or $\mathop{\rm EG} \nolimits$ $\Pi^1_1$-complete?
\end{quest}
Note that for a positive answer it suffices to show that $WF$ (or some other $\Pi^1_1$-complete set) Borel reduces to these sets. Under an extra set-theoretic assumption known as $\Sigma^1_1$-Determinacy, every $\Pi^1_1$ set which is not Borel is in fact $\Pi^1_1$-complete. We do not expect that extra set-theoretic assumptions should be necessary to prove any of the sets are $\Pi^1_1$-complete; we mention this as evidence that the positive answer is indeed the correct one. It is worth noting the question is a problem in group theory. For example, in the case of $\mathop{\rm EG} \nolimits$ one must devise a method of building a group from a tree so that well-founded trees give rise to elementary amenable groups and ill-founded trees give rise to non-elementary-amenable groups.
\subsection*{Acknowledgments}
The authors would like to thank Alexander Kechris and Andrew Marks for helpful mathematical discussions.
J. Williams was partially supported by NSF Grant 1044448, Collaborative Research: EMSW21-RTG: Logic in Southern California.
\bibliographystyle{bibgen}
|
\section{BRITE}
The BRITE (BRIght Target Explorer) constellation of nano-satellites monitors
photometrically, in 2 colours, the brightness and temperature variations of
stars with V$\le$4, with high precision and cadence, in order to perform
asteroseismology \citep{weiss2014}. The mission consists of 3 pairs of
nano-satellites, built by Austria, Canada and Poland, carrying 3-cm aperture
telescopes. One instrument per pair is equipped with a blue filter; the other
with a red filter. Each BRITE instrument has a wide field of view
($\sim$24$^\circ$), so up to 25 bright stars can be observed simultaneously, as
well as additional fainter targets with reduced precision. Each field will be
observed during several months. As of September 2014, 6 nano-satellites are
already flying and 5 are observing. Each pair of nano-satellites can (but does
not have to) observe the same field and thus increase the duty cycle of
observations.
BRITE primarily measures pressure and gravity modes of pulsations to probe
the interiors and evolution of stars through asteroseismology. Since the BRITE
sample consists of the brightest stars, it is dominated by the most
intrinsically luminous stars: massive stars at all evolutionary stages, and
evolved cooler stars at the very end of their nuclear burning phases (cool
giants and AGB stars). Analysis of OB star variability will help solve two
outstanding problems: the sizes of convective (mixed) cores in massive stars and
the influence of rapid rotation on their structure and evolution. In addition,
measurements of the timescales involved in surface granulation and differential
rotation in AGB stars, cool giants and cool supergiants will constrain turbulent
convection models.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.36\textwidth,clip]{Neiner2_fig1a
\includegraphics[width=0.36\textwidth,clip]{Neiner2_fig1b}
\caption{Examples of magnetic field detections in the single cool stars
$\delta$\,Oph and $\beta$\,Vir. Shown are the LSD Stokes V magnetic signatures
(top panels) and LSD intensity profiles (bottom panels).}
\label{singlecool}
\end{figure}
\section{Combining asteroseismology and spectropolarimetry}
The study of the magnetic properties of pulsating stars is particularly
interesting since, when combined with the study of their pulsational properties,
it provides (1) a unique way to probe the impact of magnetism on the physics of
non-standard mixing processes inside these stars and (2) strong constraints on
seismic models thanks to the impact of the field on mode splittings and
amplitudes.
The combination of an asteroseismic study with a spectropolarimetric study has
been accomplished for only a couple of massive stars so far, e.g. for the
$\beta$ Cep star V2052\,Oph \citep{briquet2012}. This star presents a magnetic
field with a strength at the poles of about 400 G that has been modelled thanks
to Narval spectropolarimetry \citep{neiner2012}. Moreover our asteroseismic
investigations of this object showed that the stellar models explaining the
observed pulsational behaviour do not have any convective core overshooting
\citep{briquet2012}. This outcome is opposite to other results of dedicated
asteroseismic studies of non-magnetic $\beta$~Cep stars
\citep[e.g.][]{briquet2007}. Indeed, it is usually found that convective core
overshooting needs to be included in the stellar models in order to account for
the observations \citep{aerts2010}. The most plausible explanation is that the
magnetic field inhibits non-standard mixing processes inside V2052\,Oph. Indeed
the field strength observed in V2052\,Oph is above the critical field limit
needed to inhibit mixing determined from theory \citep[e.g.][]{zahn2011}. These
findings opened the way to a reliable exploration of the effects of magnetism on
the physics of mixing inside stellar interiors of main-sequence B-type
pulsators.
Conversely, the deformation of line profiles by pulsations is usually neglected
when modelling the magnetic field present in pulsating stars. However, these
deformations directly impact the shape of the Stokes V signatures and thus our
ability to derive correct magnetic parameters. We recently developed a version
of the Phoebe 2.0 code that allows us to model both the line and Stokes V
profiles at the same time, taking pulsations into account, thus presenting for
the first time coherent spectropolarimetric models including magnetism and
pulsations \citep[see][]{neiner2014}. Thanks to this work, and the combination
of seismic and spectropolarimetric data, much more reliable magnetic parameters
can be derived for pulsators.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.4\textwidth,clip]{Neiner2_fig2a
\includegraphics[width=0.4\textwidth,clip]{Neiner2_fig2b}
\caption{Examples of magnetic field detections in the cool stars $\iota$\,Peg
and $\lambda$\,And, that have been follow-up with several observations.
$\iota$\,Peg is a binary star: the magnetic signature follows the radial
velocity of the magnetic component. The panels are the same as in Figs.~\ref{singlecool}.}
\label{bincool}
\end{figure}
\section{The BRITE spectropolarimetric survey}
There are $\sim$600 stars brighter than V=4, which are the prime targets of
BRITE. We started a systematic survey of all these BRITE targets with
spectropolarimetry. Narval at TBL is used for all targets with declination above
-20$^\circ$, ESPaDOnS at CFHT for stars with declination between -45 and
-20$^\circ$, and HarpsPol at ESO for stars with declination below -45$^\circ$.
From the results of the MiMeS project \citep{wade2014}, we know that $\sim$10\%
of all O and B stars have detectable magnetic fields. A similar occurrence is
found for A stars and down to F5. The magnetic fields observed in these stars
are stable oblique dipoles of fossil origin, with surface strength at the poles
from $B_{\rm pol} \simeq 100$ G to several kG. Therefore we aim at detecting all
fields above 50 G. For stars cooler than F5, the magnetic fields have a dynamo
origin and $\sim$50\% of them are found to be magnetic on average
\citep[see][]{konstantinova2014}. The cool giants and supergiants, however, have
very weak fields with $B_{\rm pol}$ of the order of a few to 10 G. Therefore for
these stars, we aim at detecting all fields above $B_{\rm pol}$ = 5 G. For each
star, we thus acquire one observation with a very high signal-to-noise, to reach
the desired detection level.
Thanks to this very high signal-to-noise spectropolarimetric observation of
each target, we will:
(1) discover new magnetic stars. This is particularly crucial for massive stars,
since only $\sim$65 magnetic OB stars are known as of today, including only a
handful of pulsating massive stars \citep[see][]{petit2013}. Note that one
measurement is enough to detect a field as magnetic signatures appear in Stokes
V profiles even for cross-over phases (i.e. when the longitudinal field is
null).
(2) help select the best high priority targets for BRITE, i.e. the magnetic
massive ones and the most interesting cool ones. BRITE can observe all $\sim$600
stars in 6 years if each field is observed 3 months on average, but it is useful
to oberve the most interesting targets first or longer. In particular the BRITE
sample includes 11 O stars, 160 B stars (including 29 known $\beta$\,Cep stars,
20 known classical Be stars, and 22 chemically peculiar B stars), 106 A stars
(including 6 known Ap stars), 12 eclipsing binaries, 7 known $\delta$\,Scuti
stars, 7 HgMn stars, 3 RR Lyrae stars, 1 known roAp, 22 cool sub-giant stars,
several dozens red giants,... Magnetic stars among them will be prime targets
for asteroseismology.
(3) determine the fundamental parameters of all targets for the BRITE seismic
modelling: effective temperature, gravity, projected rotation velocity (vsini),
as well as abundances in particular for magnetic and chemically peculiar stars
(HgMn, Ap, Am...). See e.g. \cite{fossati2014}.
(4) provide a complete spectropolarimetric census of bright (V$\le$4) stars, by
combining the Narval, ESPaDOnS and HarpsPol data, as well as archival data.
\section{First results}
The first Narval and ESPaDOnS observations already led to the discovery of 14
new magnetic stars. All of them are cool stars and several are binary objects.
Examples of new detections obtained with Narval are shown in Figs.~\ref{singlecool}, \ref{bincool} and \ref{onetimebin}.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.36\textwidth,clip]{Neiner2_fig3}
\caption{Example of a magnetic field detection in the cool binary $\xi$\,UMa
observed only one time. The field can still be attributed to the red-shifted
component. The panels are the same as in Figs.~\ref{singlecool}.}
\label{onetimebin}
\end{figure}
\section{Conclusions}
By combining the data acquired with Narval, ESPaDOnS and HarpsPol, a complete
spectropolarimetric census of bright (V$\le$4) stars will be available. We will
use this database to perform detailed unbiased statistics on the presence of
detectable magnetic fields in stars. The data will also be made available to the
community as a legacy, through the PolarBase database \citep{petit2014}.
All the spectra (whether the star is magnetic or not) will also serve to
determine the fundamental parameters of the BRITE stars, needed for seismic
modelling. For magnetic stars, chemical peculiarities may appear in the spectra
and will be studied as well. The best targets will be followed-up to
characterise their magnetic fields in details and provide crucial inputs for
seismic modelling.
\begin{acknowledgements}
This project is based on observations obtained at the Telescope Bernard Lyot
(USR5026) operated by the Observatoire Midi-Pyr\'en\'ees, Universit\'e de
Toulouse (Paul Sabatier), Centre National de la Recherche Scientifique (CNRS) of
France, at the Canada-France-Hawaii Telescope (CFHT) operated by the National
Research Council of Canada, the Institut National des Sciences de l'Univers of
the CNRS of France, and the University of Hawaii, and at the European Southern
Observatory (ESO), Chile.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{\label{}}
\section{Bayesian physical inference of the initial conditions}
We apply~\cite{JLW2014} the full-scale Bayesian inference code \textsc{borg} (Bayesian Origin Reconstruction from Galaxies,~\cite{JascheWandelt2013}) to the galaxies of the \texttt{Sample dr72} of the New York University Value Added Catalogue (NYU-VAGC~\footnote{\href{http://sdss.physics.nyu.edu/vagc/}{http://sdss.physics.nyu.edu/vagc/}}), based on the final data release (DR7) of the Sloan Digital Sky Survey (SDSS). The physical model for gravitational dynamics is second-order Lagrangian perturbation theory (2LPT), linking initial density fields ($a~=~10^{-3}$) to the presently observed large-scale structure, in the linear and mildly non-linear regime. The algorithm explores numerically the posterior distribution by sampling the joint distribution of all parameters involved, via efficient Hamiltonian Markov Chain Monte Carlo (HMC) dynamics.
Each sample (Fig. \ref{fig:slices}, upper panel) is a ``possible version of the truth'' in the form of a full physical realization of dark matter particles, tracing both the density and the velocity fields. The variation between samples (Fig. \ref{fig:slices}, lower panel) quantifies joint and correlated uncertainties (survey geometry, selection effects, biases, noise) inherent to any cosmological observation.
\section{Data-contrained realizations of the Universe}
We generate~\cite{LJSHW2014} a set of data-constrained realizations of the present large-scale structure (Fig. \ref{fig:slices_filter}): some samples of inferred primordial conditions are evolved with 2LPT to $z=69$, then with a fully non-linear cosmological simulation (using \textsc{gadget-2},~\cite{Springel2005}) from $z=69$ to $z=0$.
A dynamic, non-linear physical model naturally introduces some correlations between the constrained and unconstrained parts, which yields reliable extrapolations for certain aspects of the model that have not yet been constrained by the data (e.g. near the survey boundaries or at high redshift).
\section{Dark matter voids in the SDSS galaxy survey}
We apply~\cite{LJSHW2014} \textsc{vide} (the Void IDentification and Examination toolkit,~ \cite{VIDE}~\footnote{\href{http://www.cosmicvoids.net/}{http://www.cosmicvoids.net/}}) to the constrained parts of these realizations. The void finder is a modified version of \textsc{zobov}~\cite{Neyrinck2008} that uses Voronoi tessellations of the tracer particles to estimate the density field and a watershed algorithm to group Voronoi cells into voids.
We find physical cosmic voids in the field traced by the dark matter particles, probing a level deeper in the mass distribution hierarchy than galaxies, and greatly alleviating the bias problem for cosmological interpretation of final results. Due to the high density of tracers, we find about an order of magnitude more voids at all scales than the voids directly traced by the SDSS galaxies (Fig. \ref{fig:voids}, left panel), which sample the underlying mass distribution only sparsely~\cite{Sutter2014}. Our inference framework therefore yields a drastic reduction of statistical uncertainty in voids catalogs. For usual voids statistics such as radial density profiles of stacked voids (observed in simulations to be of universal character, e.g.~\cite{Hamaus2014}), the results we obtain are consistent with $N$-body simulations prepared with the same setup (Fig. \ref{fig:voids}, right panel).
\section*{References}
|
\section{Introduction}
The slowdown of a rotating neutron star is usually understood to
arise from the loss of kinetic rotational energy in the form of
electromagnetic radiation from a rotating dipolar magnetic field
attached to the star. In this simple model, the slowdown rate of a
pulsar in vacuo, rotating at frequency $\nu$, having a magnetic dipole
moment $M$, moment of inertia $I$ and inclination angle $\alpha$
between the magnetic and rotation axes is given by
\begin{equation}
\label{cubelaw}
\dot\nu = -\frac{8\pi^2}{3c^3}\frac{M^2 \sin^2\alpha}{I}\nu^3 .
\end{equation}
If $M$, $I$ and $\alpha$ are constant, we would expect to observe a
dependency of $\dot\nu \propto \nu^3$ as the pulsar slows down.
However, other braking mechanisms are possible with different
dependence upon $\nu$ and, in a more general power-law slowdown of the
form
\begin{equation}
\label{kappa}
\dot\nu=-\kappa\nu^n ,
\end{equation}
\noindent
the value of the exponent of $\nu$ is known as the braking index $n$,
being 3 for dipolar magnetic torque. Other mechanisms have been
proposed which would give lower values of $n$; for example,
\citet{mt69} showed that the torque from a simple outflow of particles
would lead to $\dot\nu \propto \nu^1$, i.e. $n=1$. The value of $n$ may also
differ from 3 if, for instance, the values of $M$, $\alpha$ or $I$ in
equation~(\ref{cubelaw}) vary with time \citep{br88}. In order to
explore the physics of this slowdown, the value of $n$ can in
principle be determined from observation of higher-order frequency
derivatives; differentiation of equation~(\ref{kappa}) shows that the
second and third derivatives should be related to $n$ by
\begin{eqnarray}
\label{ndef}
n&=& {\ddot\nu\nu\over \dot\nu^2} \;\;\;\; {\rm and}\\
\label{F3}
n(2n-1)&=&{\stackrel{...}{\nu}\nu^2\over \dot\nu^3}.
\end{eqnarray}
For most pulsars, evaluation of the rotational slowdown law is
confined to measurement of the rotational frequency $\nu$ and its
first derivative $\dot\nu$. For old pulsars, the second derivative
$\ddot\nu$ expected from equation~(\ref{ndef}) is usually unmeasurably small,
while in many younger pulsars any secular behaviour is often confused
by unpredictable changes in rotation rate, in the form of either
timing noise or glitches. Timing noise is seen as slow quasi-random
changes in rotation rate and arises from magnetospheric instabilities
\citep{lhk+10}, while glitches are almost instantaneous increases in
rotation rate, often followed by some associated transient behaviour
and have their origin in the neutron star interior \citep{elsk11}.
Because of these effects, values of braking index have been reliably
established for only eight pulsars. For the Crab pulsar, $\ddot\nu$
has been measured between glitches \citep{lps93}, leading to an
observed value $n_{\rm obs}=2.51(1)$, significantly less than the
value of $n=3$ expected for braking by magnetic dipole radiation. The
same is true for all the other seven pulsars (Table \ref{indices}).
These results indicate that the physical process causing the slowdown is not
just simple dipolar electromagnetic radiation.
\begin{table}
\caption{Measured braking indices for young pulsars}
\label{indices}
\begin{tabular}{llll}
\hline\hline
PSR & $n$ & Reference \\
\hline
B0531+21(Crab)& 2.51(1) & \citet{lps93} \\
B0540$-$69 & 2.14(1) & \citet{lkg+07} \\
B0833$-$45(Vela)& 1.4(2) & \citet{lpgc96} \\
J1119$-$6127 & 2.684(2)& \citet{wje11} \\
B1509$-$58 & 2.839(1)& \citet{lkg+07} \\
J1734$-$3333 & 0.9(2) & \citet{elk+11}\\
J1833$-$1034 & 1.857(1)& \citet{rgl12}\\
J1846$-$0258 & 2.65(1) & \citet{lkg+07} \\
\hline
\end{tabular}
\end{table}
\begin{figure}
\includegraphics[width=8.5cm]{spin.eps}
\caption{The spin-frequency history of the Crab pulsar over 45 years.
(a) The observed spin-frequency $\nu$ determined from fits to
100-day data sets every 50 days, showing the monotonic slow-down
of the pulsar.
(b), (c) and (d) The frequency residuals $\delta\nu$ after fitting
to the values in (a) simple slow-down models involving frequency and
respectively one, two and three spin-frequency derivatives in the
Taylor Series of equation~(\ref{taylor}). The fitted values of
$\nu_0$, $\dot{\nu}_0$, $\ddot{\nu}_0$ and $\stackrel{...}{\nu}_0$
for (d) are given in Table~\ref{spinpars}.}
\label{nu}
\end{figure}
\begin{figure}
\includegraphics[width=8.5cm]{spindown.eps}
\caption{The slowdown rate $\dot\nu$ of the Crab pulsar over 45 years.
(a) Observed values of the first derivative $\dot\nu$ determined from
fits to 100-day data sets every 50 days. The magnitude of the
slowdown rate $\vert\dot\nu\vert$ is decreasing overall, with small
step increases at the glitches.
(b) Frequency derivative residuals $\delta\dot\nu$ on an expanded
scale, obtained by subtracting from (a) a linear model using the values of
the first two frequency derivatives given in Table~\ref{spinpars}.}
\label{nudot}
\end{figure}
In this paper we report on the measurement and analysis of the
rotation rate of the Crab pulsar from 1968 to 2013. This 45-year
time-baseline amounts to about 5\% of the pulsar lifetime and allows
the spin-down of the Crab pulsar to be described over a period which
includes many glitches and provides more details of the cumulative
effect that they have on the long-term spin-down \citep{lps93,sj03}.
Elsewhere, the same data have been used to examine the statistics and
physical details of the glitches \citep{eas+14} and to study the
evolution of the radio pulse emission over this time \citep{lgw+13}
to enable a comprehensive picture of the evolution of the pulsar.
\section{Observations and Basic Analysis}
The rotation of the Crab pulsar has been monitored by daily
observations at Jodrell Bank Observatory since 1984, mainly using the
13-m radio telescope at 610 MHz \citep{lps88,lps93}. Regular
observations with the 76-m Lovell telescope at around 1400-1700 MHz,
designed to monitor any changes in dispersion measure, also contribute
to the dataset.
These data have been supplemented with earlier observations taken at
Arecibo \citep{girp77} and in the optical at Princeton
\citep{gro75a} and Hamburg \citep{loh81}.
There are no observations available between February 1979 and
February 1982, this being the only significant gap with no data.
There are in total appproximately 11,000 times of arrival (TOAs) and
together the measurements comprise a record of the rotation of the
pulsar over a total of 45 years from November 1968 to December 2013.
In order to study the long-term rotational history of the pulsar, we
have used standard procedures to reduce the TOAs to the barycentre of
the Solar System. We have then fitted values of the rotation
frequency $\nu$ and its first two derivatives $\dot\nu$ and $\ddot\nu$
over time spans of approximately 100 days. Such analyses were
repeated with the central reference time advancing by typically 50
days between analyses. Close to glitches, the time spans were
adjusted in such a way that no analysis was performed over a glitch,
so that one analysis ended and another started close to the epoch of
the glitch.
These time sequences of rotational frequencies and first derivatives
provide the main forms of the data that we use to study the long-term
behaviour of the pulsar in this paper. Figs.~\ref{nu}a and
\ref{nudot}a illustrate the evolution of the rotational frequency
$\nu({\rm t})$ and slowdown rate $\dot\nu({\rm t})$ over the 45 years.
The rotational slowdown of the pulsar is evident in Fig.~\ref{nu}a,
falling by about 0.5~Hz during this time.
The slowdown rate (Fig.~\ref{nudot}a) also shows a general reduction in
magnitude with time, but there are also considerable long-term effects
resulting from glitches, which we investigate further in a later
section.
Following convention, it is instructive to characterise the variation
in rotation frequency with time as a Taylor series of derivatives of
the form:
\begin{equation}
\nu(t)= \nu_0 +\dot\nu_0(t-t_0)+{1\over 2}\ddot\nu_0(t-t_0)^2
+ {1\over 6}\stackrel{...}{\nu}_0(t-t_0)t^3+\delta\nu(t).
\label{taylor}
\end{equation}
By fitting such a model to the frequency data, we evaluate $\nu_0$ and
the first three derivatives, leaving a residual $\delta\nu(t)$. In
Figs.~\ref{nu}b-\ref{nu}d we see the residuals resulting from fitting
and subtracting Taylor series consisting of 1, 2 and 3 derivative
terms, respectively. In each case, the residuals are dominated by the
next, unfitted derivative term. The remaining residuals shown in
Fig.~\ref{nu}d, which might in previous analyses have been designated
as timing noise, are related to glitch activity, which we discuss
later. The best fit values obtained from the 3-derivative fit are
presented in Table~\ref{spinpars}. These parameters represent the
long-term behaviour of the pulsar, including the effects of glitches.
Using equation~(\ref{ndef}) and the values in Table~\ref{spinpars}, a
braking index $n=2.349(1)$ is obtained. Using this braking index in
equation~(\ref{F3}) it is found that the expected value of
$\stackrel{...}{\nu}$ is $-0.52\times10^{-30}$~Hz\,s$^{-3}$, which is
about one fifth of the measured value. This shows that the general
long-term slowdown of the Crab pulsar is not a simple power-law. We
show later that removal of the effects of glitches brings the slowdown
close to a simple power-law.
Extrapolating the Taylor series back in time to investigate the
behaviour in earlier years should evidently be approached with
caution, but we remark that the series indicates an original rotation
rate of 58~Hz (a period of 17 milliseconds) at birth in AD~1054.
\begin{table}
\caption{Observed and derived parameters for the Crab pulsar. The
spin parameters were obtained by fitting a 3-derivative Taylor series
to the frequency data presented in Fig.~\ref{nu}a. The standard
errors determined in the analysis are given in parentheses after the
values and are in units of the least significant digit.}
\label{spinpars}
\begin{tabular}{ll}
\hline \hline
\multicolumn{1}{c}{Parameter} & \multicolumn{1}{c}{Value} \\
\hline
Data span (MJD) & 40175$-$56665 \\
Epoch t$_0$(MJD) & 48442.5 \\
$\nu_0$ (Hz) & 29.946923(1) \\
$\dot{\nu}_0$ ($10^{-10}$~Hz\,s$^{-1}$) & $-$3.77535(2) \\
$\ddot{\nu}_0$ ($10^{-20}$~Hz\,s$^{-2}$) & 1.1147(5) \\
$\stackrel{...}{\nu}_0$ ($10^{-30}$~Hz\,s$^{-3}$) & $-$2.73(4) \\
RMS residual (Hz) & 0.0000058 \\
Maximum frequency residual (Hz) & 0.000013 \\
Mean braking index $n$ & 2.342(1) \\
\hline
\end{tabular}
\end{table}
\section{The 24 glitches}
The monotonic rotational slowdown of the Crab pulsar seen in
Fig.~\ref{nu} is occasionally reversed discontinuously at a glitch, in
which $\nu$ increases by a small step $\Delta\nu$ of order
$10^{-9}\nu$ to $10^{-7}\nu$, followed by a nearly complete recovery
within about 20 days of the event. The magnitude of the slowdown rate
$\vert\dot\nu\vert$ (Fig.~\ref{nudot}) also presents an initial step
increase at a glitch, again partially reversing the long-term trend
and showing the corresponding short-term recovery \citep{lps93,wbl01}.
However, in contrast with the relaxation in $\nu$, the recovery in
slowdown rate is not complete and a persistent step
($\Delta\dot\nu_{\rm p}$) is commonly observed after glitches. First
remarked upon by \citet{girp77} and \citet{dp83} in relation to the
glitch of 1975, these persistent steps are a general feature of
glitches in this pulsar and have an appreciable effect on the overall
slowdown behaviour.
Unfortunately, the study of the long-term impact of glitches is
often contaminated by the occurrence of other, nearby glitches.
However, there are three large glitches which are ``isolated'', each
having no other detectable glitches within 800 days before or 1200
days after the epoch of the glitch. These three glitches, in 1975,
1989 and 2011 allow the nature of these persistent steps to be
demonstrated in Fig.~\ref{3glits}. As pointed out by \citet{lps93} in
relation to the glitch of 1989, the persistent step increase in
slowdown rate consists of an instantaneous component followed by a
further quasi-exponential asymptotic increase on a timescale of about
265 days. We now see in Fig.~\ref{3glits} that the same description
applies closely to all three large, isolated glitches: all three step
changes in $\dot\nu$ have different amplitudes and all three show a
further increase in $\dot\nu$ which grows exponentially for at least a
year after the initial step. Additionally, the relative amplitudes of
the step and the exponential and the time scales of the exponential
are identical to within the fitted errors for all three. As a result,
we conclude that the long-term recovery can be represented by a
single-parameter function of the form:
\begin{equation}
\delta\dot{\nu} = \left\{ \begin{array}{ll}
0 & \mbox{if $t<0$} \\
\Delta\dot{\nu}_{\rm p} \times (0.46\times {\rm exp}(-t/320)-1.0) & \mbox{if $t>0$,}
\end{array}
\right.
\label{vdot_pers}
\end{equation}
where $t$ is the time from the glitch epoch in days and
$\Delta\dot{\nu}_{\rm p}$ is the total persistent step in slowdown
rate. Note that the asymptotic exponential is in the same sense as
the initial step in $\dot\nu$, unlike the short-term recovery at a
glitch.
We now regard this as an intrinsic component of all glitches, although
for many it is often obscured by the recovery from previous glitches
or the occurrence of later glitches. The asymptotic exponential
component has a timescale of 320$\pm$20 days and comprises 46\%, or
almost half, of the total increase of slowdown rate. In what follows,
we assume that these values apply to all glitches in the Crab pulsar.
Table~\ref{glitch_list} lists 24 glitches, giving the date and MJD of
occurrence, the step change $\Delta\nu$, the fractional value $\Delta
\nu/\nu$ and the persistent change of slowdown rate $\Delta \dot
\nu_{\rm p}$. This is a robust list which includes all glitches for
which $\Delta\nu> 0.01 \mu$Hz. Any glitches smaller than this are of
similar size to the period variations produced by the noise present in
the data. Any glitch which may have occurred during the gap in
observations in 1981 left no obvious change in slowdown rate, and
would presumably have been small and would not affect the results of
our analysis. Glitch epochs and frequency step sizes were taken from
\citet{elsk11} and \citet{ejb+11}. The sequence of glitches is
illustrated in Fig.~\ref{glitches}a, which shows the date of
occurrence and the size $\Delta\nu$ of each glitch (on a logarithmic
scale). Note particularly the pattern of substantially increased
activity from MJD $\sim 50000$ to MJD $\sim 54000$ (years 1995-2006).
A histogram of glitch sizes $\Delta\nu$ is shown in
Fig.~\ref{dfdf1p}a. Espinoza et al. (2014)\nocite{eas+14} have
analysed the Jodrell Bank data on the Crab pulsar and conclude
that the decrease in numbers of glitches towards smaller sizes is
intrinsic and is not related to the detection capabilities.
\begin{figure}
\includegraphics[width=8.5cm]{3glits.eps}
\caption{The variation in slowdown rate $\dot\nu$ of the Crab pulsar
near to the three large isolated glitches which occurred in 1975, 1989
and 2011. In each case, the frequency derivative residuals
$\delta\dot\nu$ were obtained relative to a linear model involving the
first two frequency derivatives fitted to $\dot\nu$ over the 800 days
preceding the glitch, which occurred at day 0. Each glitch is
followed instantly by a large negative transient (increase in
magnitude) in $\dot\nu$. The main transient behaviour ceases after
about 100 days, revealing a persistent negative offset which continues to
increase in a quasi-exponential manner on a
timescale of around 320 days. The smooth lines are the fits to the
post-glitch data described by equation \ref{vdot_pers}.
}
\label{3glits}
\end{figure}
The persistent steps in frequency derivative were estimated by fitting
a function of the form given by equation~(\ref{vdot_pers}) to the observed
$\dot{\nu}$ data centred on the glitch epoch. In order to avoid
contamination by short-term glitch transient recoveries, data taken
within 90 days following each glitch were not used. In some cases the
step is very small and no significant measurement was possible. In
other cases, because of the presence of other nearby glitches, the
available data points were insufficient to perform the measurement,
although in a few cases it was possible to measure the combined effect of
two or three closely-separated glitches (Table \ref{glitch_list}).
Without being extremely precise, this method effectively measures the
basic trends we are studying and it is not affected by the necessary
assumptions of more complicated models. Our values are roughly
consistent with existent measurements using such models
\citep[e.g.][]{wbl01,wwty12}.
The step in $\Delta\dot{\nu}_{\rm p}$ at a glitch is loosely related to
the step-change in rotation frequency $\Delta\nu$. The relationship
is shown in Fig.~\ref{dfdf1p}b, which includes data from 15 glitches
(some points combine data from two or three adjacent glitches). The
approximately linear relationship is fitted by the diagonal line
of unity slope, given by
\begin{equation}
|\Delta\dot{\nu}_{\rm p}| = 7\times10^{-8}\times\Delta\nu \;\;\rm{ Hz\; s^{-1}},
\label{dfdf1pe}
\end{equation}
where $\Delta\nu$ is in Hz.
The persistent slowdown increase is usually attributed to a reduction
in effective moment of inertia due to pinning of neutron superfluid
vortices to other internal components. Because a rotating superfluid
slows down only if its vortices are allowed to move apart, the pinning
of some vortices decouples a fraction of the superfluid from the rest
of the star and reduces the effective moment of inertia of the
rotating star. Equation~(\ref{dfdf1pe}) indicates that the re-pinning
of vortices is more effective after larger glitches.
We offer no explanation for the surprising phenomenon of the long-term
asymptotic increase in $\dot\nu$ and continue to view all increases of
slowdown rate at glitches as decreases in effective moment of inertia,
presumably due to vortex pinning. We note that if there is no
relaxation in this cumulative pinning, the proportional change in
moment of inertia has amounted to $0.3\%$ in 45 years. We remark that
such a large rate of change cannot persist for more than a few
thousand years, by which time a large proportion of the effective
moment of inertia of the neutron star would have disappeared.
\begin{table}
\caption{The steps in rotation rate and slowdown rate for the 24
glitches observed between 1968 and 2013. This table is derived from
\citet{eas+14}}.
\label{glitch_list}
\begin{tabular}{llclc}
\hline\hline
\multicolumn{1}{c}{Date} &
\multicolumn{1}{c}{MJD} &
\multicolumn{1}{c}{$\frac{\Delta\nu}{\nu}$} &
\multicolumn{1}{c}{$\Delta\nu$} &
\multicolumn{1}{c}{$\Delta\dot{\nu}_p$} \\
\multicolumn{1}{c}{ } &
\multicolumn{1}{c}{days} &
\multicolumn{1}{c}{$10^{-9}$} &
\multicolumn{1}{c}{$\mu$Hz} &
\multicolumn{1}{c}{$10^{-15} \textrm{s}^{-2}$} \\
\hline
1969 Sep & 40491.84(3) & 7.2(4) & 0.22(1) & $-$ \\
1971 Jul & 41161.98(4) & 1.9(1) & 0.057(4) & $-$ \\
1971 Oct & 41250.32(1) & 2.1(1) & 0.062(3) & $-$ \\
1975 Feb & 42447.26(4) & 35.7(3) & 1.08(1) & $-$112(2) \\
1986 Aug & 46663.69(3) & 6(1) & 0.18(2) & $-$ \\
\\
1989 Aug & 47767.504(3) & 81.0(4) & 2.43(1) & $-$150(5) \\
1992 Nov & 48945.6(1) & 4.2(2) & 0.13(1) & $-$ \\
1995 Oct & 50020.04(2) & 2.1(1) & 0.063(2) & $-$ \\
1996 Jun & 50260.031(4) & 31.9(1) & 0.953(4) & $^*$ \\
1997 Jan & 50458.94(3) & 6.1(4) & 0.18(1) & $-116(5)^\dagger$ \\
\\
1997 Dec & 50812.59(1) & 6.2(2) & 0.19(1) & $-$ \\
1999 Oct & 51452.02(1) & 6.8(2) & 0.20(1) & $-$25(3) \\
2000 Jul & 51740.656(2) & 25.1(3) & 0.75(1) & $^*$ \\
2000 Sep & 51804.75(2) & 3.5(1) & 0.105(3) & $-53(3)^\dagger$ \\
2001 Jun & 52084.072(1) & 22.6(1) & 0.675(3) & $^*$ \\%$-$82(21) \\
\\
2001 Oct & 52146.7580(3)& 8.87(5) & 0.265(1) & $-70(10)^\dagger$ \\
2002 Aug & 52498.257(2) & 3.4(1) & 0.101(2) & $^*$ \\
2002 Sep & 52587.20(1) & 1.7(1) & 0.050(3) & $-8(2)^\dagger$ \\
2004 Mar & 53067.0780(2)& 214(1) & 6.37(2) & $^{*}$ \\
2004 Sep & 53254.109(2) & 4.9(1) & 0.145(3) & $^{*}$ \\
\\
2004 Nov & 53331.17(1) & 2.8(2) & 0.08(1) & $-250(20)^{\dagger\dagger}$ \\
2006 Aug & 53970.1900(3)& 21.8(2) & 0.65(1) & $-$30(5) \\%-17(5) \\
2008 Apr & 54580.38(1) & 4.7(1) & 0.140(4) & $-$ \\%-12(25) \\
2011 Nov & 55875.5(1) & 49.2(3) & 1.46(1) & $-$132(5) \\%-82(8) \\
\hline
\end{tabular}
$^\dagger$incorporates the persistent step of the previous glitch.
$^{\dagger\dagger}$incorporates the persistent steps of the previous two glitches.
\end{table}
\begin{figure}
\includegraphics[width=8cm]{glits-cumuls.eps}
\caption{Glitches in the Crab pulsar.
(a) The distribution in time and magnitude of the steps $\Delta\nu$ in
spin-frequency of the 24 glitches in Table \ref{glitch_list}.
(b) The cumulative effect of the persistent steps in slowdown rate
$\Delta{\dot\nu}_{\rm p}$ at the glitches.}
\label{glitches}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{df1p-df0_H.eps}
\caption{Glitch sizes.
(a) Histogram of frequency steps $\Delta\nu$ on a logarithmic scale
for the 24 glitches in Table \ref{glitch_list}.
(b) The inter-dependence of the persistent change in slow-down rate
$|\Delta\dot{\nu}_{\rm p}|$ and the step-change in rotation frequency
$\Delta\nu$ for 10 glitches. The diagonal line has a slope of unity.
Only events for which $\Delta\dot{\nu}_{\rm p}<0$ are shown.}
\label{dfdf1p}
\end{figure}
\citet{mpw08} include the Crab pulsar glitches in a comprehensive
analysis of the statistics of the distributions of size and intervals
between glitches, testing the hypothesis that these are determined by
a random process of self-organised criticality. The incidence of
glitches in most pulsars appears to be random, and \citet{wwty12} show
that the distribution of interval times in the Crab pulsar is
Poissonian, although some pulsars show a quasi-periodicity
\citep{lel99,mpw08,wwty12}.
\begin{figure}
\includegraphics[width=8cm]{compareCumuls-grey.eps}
\caption{The cumulative distribution of intervals between all 20
glitches occurring between MJD 46000 and MJD 56600, as observed (bold
line) and in 100 random simulations. }
\label{P_cumul}
\end{figure}
We nevertheless draw attention to the cluster of 15 glitches between
MJD 50000 and 54000 seen in Fig.~4, which suggests that there may be
an extra complexity in the system. We require a statistical test for
the hypothesis that this cluster was a chance concentration of
glitches which occurred at random. The \citet{wwty12} test was not
sensitive to clustering, which is better revealed by the distribution
of intervals between all glitches, rather than only consecutive ones. We
therefore compare the observed distribution of intervals between all
glitches with those from simulated glitch sequences and show that the
probability that the observed distribution is random is low.
We confined our analysis to the 20 glitches which occurred in the last
30 years since high-cadence monitoring commenced in January 1984. As
explained in \citet{eas+14}, the data in this period constitute a
uniform set and are complete, indicating a mean rate of 0.67
glitches/yr. The bold line in Figure \ref{P_cumul} shows the
cumulative distribution of all 190 glitch separation times in this set
of data. We then generated 10,000 simulated distributions of 20
events, each occurring at a random time within a 30-year period;
in Fig.~\ref{P_cumul}, we show 100 of these realisations (randomly
picked) for comparison with the observed distribution.
The curve derived from the observed data lies on the outside of the
spread of curves from the simulations, indicating a low probability
that the observed cluster was a random occurrence. In particular,
there is an excess of small separation times in the observed
cumulative distribution, which is therefore steeper than most of the
simulated distributions. This indicates that the glitches are more
clustered than can be expected from a random occurrence of
glitches. We also compared the means of the separation times in the
distributions; the mean of the 190 observed glitch separation times was
less than the mean in all except 52 (~0.5\%) of the 10,000 simulated
distributions. We therefore conclude that it is unlikely that the
cluster of glitches is a statistical accident. We note that the
11-year period of the cluster coincides with a period of an anomaly in
the underlying slowdown rate, which is seen in Fig.~\ref{underlying_n}
and will be discussed in the next Section.
\section{The effect of glitches on the slowdown}
Fig.~\ref{glitches}b shows the cumulative effect of the steps
$\Delta\dot{\nu}_{\rm p}$ at the 17 glitches for which this quantity
was measured. The total cumulative effect over 45 years is an
increase in slowdown rate $\vert\Sigma\Delta\dot\nu_{\rm p}\vert$ of
$946(25)\times10^{-15}$~Hz\,\,s$^{-1}$. The small but appreciable
overall negative contribution made by the glitches to the second
derivative $\ddot\nu$ may be evaluated by dividing the total sum by
the time interval, giving a value of
$-0.066(4)\times10^{-20}$~Hz\,\,s$^{-2}$, which is approximately 6\%
of the second derivative $\ddot\nu$ evaluated from the long-term
analysis (Table~\ref{spinpars}).
We now address the notion that the slowdown is an underlying steady
phenomenon with superimposed cumulative steps in the slowdown rate
caused by glitches. If the persistent steps $\Delta\dot{\nu}_{\rm
p}$, as tabulated in Table~\ref{glitch_list} and with the form
described by equation \ref{vdot_pers}, are removed from the
$\dot{\nu}$ data we obtain a smoother, ``corrected'' $\dot{\nu}$
evolution which shows the underlying slowdown.
Fig.~\ref{underlying_n}a shows frequency derivative residuals obtained
by removing from the corrected $\dot{\nu}$ data a new linear 2-term
Taylor series fitted to the restricted 6-year span up to MJD 42447,
this being the epoch of the 1975 glitch. This fit yields values of
$\dot{\nu}=-3.84796(1)\times10^{-10}$~Hz\,\,s$^{-1}$ and
$\ddot\nu=1.236(1)\times10^{-20}$~Hz\,\,s$^{-2}$ at MJD 41338.67. We
now find from equation~(\ref{ndef}) that during this span, the braking
index was $n=2.519(2)$, and from equation~(\ref{F3}) the expected
value of $\stackrel{...}{\nu} = -0.636(1)\times 10^{-30}$
Hz\,\,s$^{-3}$. The smooth curve in Fig.~\ref{underlying_n}a is the
calculated variation in $\dot{\nu}$ including the curvature due to
this term, and in general is seen to track the data well.
However, at around MJD 50000, the slope in Fig.~\ref{underlying_n}a
changes abruptly, recovering at around MJD 54000, after an accumulated
offset in $\dot\nu$, to approximately the expected slope. This is the
same 11-year span that contains the large concentration of glitches.
The accumulated deviation in this 11-year span amounts to
$\sim-200\times 10^{-15}$ Hz\,\,s$^{-1}$, and is in addition to the
$\sim-522\times 10^{-15}$ Hz\,\,s$^{-1}$ already accounted for by the
observed persistent glitch contributions $\Delta\dot{\nu}_{\rm p}$ in
this period, and may be compared with the total change during this
time of $\sim3500\times 10^{-15}$ Hz\,\,s$^{-1}$ due to the underlying
slowdown. It seems unlikely that the deviation can be due to glitches
which are below the threshold of our measurements, since
Fig.~\ref{dfdf1p}a shows a deficit of small glitches
\citep{eas+14}. More likely, the accumulated deviation is due to a
phenomenon related to glitches which is not reflected in our present
model.
The variations in $\delta\dot\nu$ may also be interpreted as changes
in the underlying value of the braking index $n$ with time. These
changes are summarised in Fig.~\ref{underlying_n}b, in which the value
of the braking index is close to 2.5 throughout most of the 45 years,
except for the period of high glitch activity from MJD 50000-54000,
when it takes a value of about 2.3.
\begin{figure}
\includegraphics[width=8.5cm]{brakings_2.eps}
\caption{The underlying slowdown of the Crab pulsar after removal of
the effects of the glitches. a) The frequency derivative residuals
$\delta\dot\nu$ obtained by subtracting from $\dot\nu$ data the
persistent steps $\Delta\dot\nu_{\rm p}$ at the glitches given in
Table~\ref{glitch_list}, removing short-term glitch transients by
excluding data within 90 days following each glitch, and by removing
the main linear trend as measured from the start to MJD~50458. The
smooth curve is the expected behaviour during this era for a constant
value of $n=2.519$, showing a curvature corresponding to a value of
$\stackrel{...}{\nu} = -0.636\times 10^{-30}$ Hz\,s$^{-3}$ as
calculated from equation~(\ref{F3}). b) The underlying braking index
of the Crab pulsar using equation~\ref{ndef}, evaluated over sections
of 1000 days at 500-day intervals. The values of $\nu$ and
$\dot{\nu}$ are the mean values of the data in each interval in
Figs.~\ref{nu}a and \ref{nudot}a, and $\ddot\nu$ is taken from the
slope of the corresponding data in (a). The horizontal line at
$n=2.519$ corresponds to the curve in (a).}
\label{underlying_n}
\end{figure}
We remark that the process of separating the effects of glitches from
an underlying steady rotational evolution provides a good description
of the overall behaviour, with braking index $n = 2.5$. The small
effect on the index seen in Fig.~\ref{underlying_n}b remains
unexplained.
\section{The slowdown power law}
\label{powerlaw}
Why is the braking index different from 3, as expected from
equation~(\ref{cubelaw}) for energy loss through electromagnetic
dipole radiation? Either this equation is inadequate because angular
momentum is also lost through a mechanical process such as an
outflowing jet or interaction with an external fall-back disk of
supernova remains \citep{mph01a}, or one or more of the parameters
$I$, $M$, or $\alpha$ of the equation is changing with time. These two
categories might be called the unipolar and dipolar approaches in
which the loss of angular momentum may lead to very different slowdown
laws $\dot\nu\ \propto\nu$ and $\dot\nu\propto\nu^3$ respectively. As
pointed out by \citet{hck99}, \citet{xq01}, and \citet{cl06}, a
combination of both processes could account for the observed low
braking indices of young pulsars.
\subsection{The wind component}
The Crab pulsar, along with other young pulsars, generates a powerful
particle stream whose effect is observed as a wind nebula
(e.g. \citet{gs06}). The possible effect of angular momentum carried
away by this particle stream has been much debated, following
\citet{mt69} who pointed out that if the slowdown were predominantly
due to wind, the torque would follow the first rather than the third
power of the rotation rate. Other values of braking index of less
than three due to particle flows have also been suggested
(e.g. \citet{wxg03}). An alternative approach by \citet{mph01a}
suggests that interaction between the rotating pulsar and an accretion
disc of infalling supernova remains, forming outside the magnetosphere
but coupled by a propeller torque, would lead to a similar slowdown
law. \citet{yps12} consider the effect of this process on the timing
properties of old as well as young pulsars.
For instance, if the departure of the braking index from 3 is due
to such a dynamic process with braking index $n=1$, the
slowdown can be represented by a combination of two power laws:
\begin{equation}
\dot\nu= -(A\nu^3 + B\nu ).
\end{equation}
If at the present time the dynamic component increases the slowdown
rate by a factor $(1+\epsilon)=(1+B/A\nu^2)$, it is easy to show that
the observed braking index is
\begin{equation}
n={\nu\ddot\nu \over\dot\nu^2}=\frac{\epsilon+3}{\epsilon+1}
\label{epsilon1}
\end{equation}
and
\begin{equation}
\epsilon=\frac{3-n}{n-1} .
\label{epsilon2}
\end{equation}
If the wind (or the disk) was to be responsible for the whole of the
difference between $n_{\rm dip}=3$, the expected value from
magnetic-dipole braking, and $n_{\rm obs}$, the observed value, then
equation~(\ref{epsilon2}) shows that for $n_{\rm obs}=2.5$, the wind
accounts for approximately one third of the slowdown torque. A more
precise modelling of the effect of the wind or disk might lead to a more
accurate assessment of this proportion. For the Vela pulsar, the very
low value of $n_{\rm obs}=1.4$ suggests that the wind is responsible
for ${4\over 5}$ of the slowdown torque.
\subsection{A changing magnetic dipole}
If alternatively the low braking index is predominantly due to
magnetic dipole radiation in which the coefficient $\kappa$ in
equation~(\ref{kappa}) is a function of time $\kappa(t)$, then the observed
braking index will be
\begin{equation}
n_{\rm obs}={\ddot\nu\nu \over \dot\nu^2}=n_{\rm dip}+{\dot\kappa(t) \nu \over
\kappa(t)\dot\nu} .
\label{kappadot2}
\end{equation}
The observed interglitch value of braking index for the Crab pulsar,
$n_{\rm obs}= 2.5$, would require
$\dot\kappa/\kappa\simeq2.0\times10^{-4}$ yr$^{-1}$. Since
\begin{equation}
\kappa \,\, {\propto} \,\, M^2 \sin^2\alpha I^{-1},
\label{coefficient}
\end{equation}
we evaluate the changes in these three parameters which
might account for the low value of the braking index.
From equation~(\ref{kappadot2}) the effects are given by
\begin{equation}
n_{\rm obs}= n_{\rm dip}+{\nu\over\dot\nu} \left(-{{\dot I} \over
I}+2{\dot\alpha \over \tan\alpha}+2{\dot M \over M}\right).
\label {bdot}
\end{equation}
The overall changes in the Crab slowdown are too large to be explained
by simple changes in ellipticity and the consequent changes in the
moment of inertia $I$ of a rotating ellipsoid. An apparent reduction
in $I$ might however be due to a continuous decoupling of the interior
produced by an accumulation of pinned vortices in reservoirs (Alpar et
al. 1996; Ho \& Andersson 2012)\nocite{accp96,ha12}, eventually
locking up a substantial fraction of the superfluid. Such a process
would eventually be limited by the available superfluid, which is
usually considered to be the superfluid in the neutron-rich inner
crust. It seems unlikely that the process could persist long
enough to explain the low braking index of older pulsars such as the
Vela pulsar.
Several authors, notably \citet{br88,crz98,ah97,lyn04,elk+11}, have
remarked that a low braking index might be due to an increasing
dipolar magnetic field, or to an increasing inclination angle
$\alpha$. Evidence for a secular increase in $\alpha$ has been
presented by \citet{lgw+13}. From observations of the profile of the
radio pulse over 22 years, they find an increase in the component
separation amounting to $6\times10^{-3}$ degrees per year. They remark
that this can be attributed to a similar rate of increase in $\alpha$,
which for $\alpha$ = 45 deg would account for the low value of the
braking index. Using the long-term value of $n_{obs}=2.35$, they
found that the changing $\alpha$ gives a value of
$\dot\kappa/\kappa=2.6\times10^{-4}$ yr$^{-1}$, which could account
for the observed rate of change.
However, although the secular change in $\alpha$ may well be
sufficient to account for the whole of the reduction in the value of
the braking index below 3, it should be noted that that the
determination of $\dot\alpha$ from the observation of the separation
of the components of the pulse profile is model dependent. Hence the
actual contribution is uncertain and it is also likely that the
relative contributions of the different processes discussed above may
be evolving with time. Further possibilities are models of
pulsar magnetospheres which recognise an inner region corotating with
the star and which can lead to $n<3$ when the evolution of this
region's size changes at a different rate to the constantly-growing
light cylinder region \citep{mel97,cs06b,bta+06}.
\section{Long-term evolution and the $P-\dot{P}$ diagram}
\begin{figure}
\includegraphics[width=8.5cm]{ppdotSection.eps}
\caption{The progress of the Crab Pulsar across the $P-\dot{P}$
diagram. a) The upper part of the standard $P-\dot{P}$ diagram and
b) an expanded view of the region containing the Crab Pulsar, showing
its motion during the past 45 years. The dot-dash line represents the
path that would be followed by a pulsar having braking index $n=3$.}
\label{ppdot}
\end{figure}
The departure from simple magnetic dipole slowdown is illustrated
dramatically in Fig.~\ref{ppdot}. This shows a section of the
familiar $P-\dot{P}$ diagram, a log-log plot on which it is customary
to plot all known pulsars. As a pulsar ages, it moves from left to
right across the diagram, following a path whose slope $2-n$ depends on the
braking index. For $n=3$, the slope is $-$1. For the Crab pulsar, for
which the mean value of $n=2.349$, the slope is $\sim-0.35$,
representing the general movement of the Crab pulsar on the diagram.
However, between glitches the evolution is with a slope of $-0.5$,
corresponding to $n=2.5$. If the wind model is correct, the monopole
term will eventually dominate and the path across the $P-\dot{P}$
diagram will turn upwards towards a slope of +1. As pointed out by
\citet{ac04b}, this leads into a region of the diagram where no pulsars
have been detected; however, it does lead towards the magnetars
\citep{elk+11}. \citet{hck99} remark on the possibility that the high
slowdown rate of the magnetars may be accounted for largely or
completely by monopolar wind torque, in contrast to the current
interpretation in terms of a very high dipole magnetic field.
\section{Conclusions}
The slowdown of the rotation rate of the Crab pulsar, including the
effect of the glitches, could be described by a power law with braking
index of around $2.35$. The spin evolution is, however, affected by
the steps in slowdown rate at glitches; removing these reveals an
underlying simple slowdown with braking index 2.519(2). The average
effect of the glitches is to increase the rate of slowdown by about
6\%. This description of the slowdown fails during a 11-year period
which coincides with a period of increased glitch activity.
Including the exponential component of the persistent step in
slowdown rate at glitches allows an almost complete separation of the effects
of glitches from the underlying slowdown. This component is in the
form of an exponential with time constant 320 days, asymptotically
reaching a value which nearly doubles the previously known effect.
The $n=2.5$ braking index, lower than the conventional value $n=3$
for a rotating magnetic dipole, may be attributed to a combination of
a secular increase in magnetic inclination angle and a monopolar wind torque.
If the rate of mass loss in the wind persists, the braking index is
expected to reduce towards $n=1$, possibly accounting for the low
index observed in the Vela pulsar.
The observed pattern of glitch activity, including the 11-year period
of increased glitch activity, does not agree well with the random
behaviour expected from self-organised criticality.
\section*{Acknowledgements}
Pulsar research at JBCA is supported by a Consolidated Grant from the UK
Science and Technology Facilities Council (STFC).
C.M.E. acknowledges the support from STFC and FONDECYT (postdoctorado 3130512).
|
\section{Omitted proofs}
\label{proofatend:proofs}%
\count@=\z@
\loop
\the\toks\numexpr\prooftoks+\count@\relax
\ifnum\count@<\value{proofcount}%
\advance\count@\@ne
\repeat
\fi
}
\makeatother
\fixstatement{theorem}
\fixstatement{proposition}
\fixstatement{corollary}
\fixstatement{lemma}
\makeatletter
\newcommand{\superimpose}[2]{%
{\ooalign{$#1\@firstoftwo#2$\cr\hfil$#1\@secondoftwo#2$\hfil\cr}}}
\makeatother
\makeatletter
\newcommand{\footnoteref}[1]{%
\protected@xdef\@thefnmark{\ref{#1}}\@footnotemark%
}
\makeatother
\makeatletter
\newcommand{\customlabel}[2]{%
\protected@write \@auxout {}{\string \newlabel {#1}{{#2}{\thepage}{#2}{#1}{}} }%
\hypertarget{#1}{#2}
}
\makeatother
\newcommand{\oversetlabel}[3][\theequation]{
\stepcounter{equation}
\phantomsection
\overset{({\customlabel{#2}{#1}})}{#3}
}
\newcommand{\eqlabel}[2][\theequation]{
\ensuremath{\oversetlabel[#1]{#2}{=}}
}
\newcommand{\mathop{}\!\mathrm{d}}{\mathop{}\!\mathrm{d}}
\newcommand{\texttt{t\!t}}{\texttt{t\!t}}
\newcommand{\texttt{f\!f}}{\texttt{f\!f}}
\newcommand{\triangleq}{\triangleq}
\newcommand{\stackrel{\triangle}{\iff}}{\stackrel{\triangle}{\iff}}
\DeclareMathOperator{\dom}{dom}
\DeclareMathOperator{\cod}{cod}
\DeclareMathOperator{\obj}{obj}
\DeclareMathOperator{\arr}{mrp}
\DeclareMathOperator{\rng}{img}
\newcommand{\cat}[1]{\textnormal{\textsf{#1}}\xspace}
\newcommand{\cat{Set}}{\cat{Set}}
\newcommand{\coalg}[1]{{#1\cat{-CoAlg}}}
\newcommand{^{op}}{^{op}}
\newcommand{ {\mbox{\large\bf-}} }{ {\mbox{\large\bf-}} }
\newcommand{\mathrm{Id}}{\mathrm{Id}}
\newcommand{\f}[1]{\mathcal{F}_{#1}}
\newcommand{\ff}[1]{\f{\mathfrak{#1}}}
\newcommand{\ff{W}}{\ff{W}}
\newcommand{\p}[1]{\mathcal{P}_{#1}}
\newcommand{\p{\!\!f}}{\p{\!\!f}}
\newcommand{\perp}{\perp}
\newcommand{\support}[1]{\lfloor #1 \rfloor}
\newcommand{\totalweight}[1]{\llfloor #1 \rrfloor}
\newcommand{\supportg}[1]{\left\lfloor #1 \right\rfloor}
\newcommand{\totalweightg}[1]{\left\lfloor\kern-4pt\left\lfloor #1 \right\rfloor\kern-4pt\right\rfloor}
\newcommand\restr[2]{{\left.\kern-\nulldelimiterspace#1\vphantom{\big|}\right|_{#2}}}
\newcommand\corestr[2]{{\left.\kern-\nulldelimiterspace#1\vphantom{\big|}\right|^{#2}}}
\newcommand{\lbrace\kern-2.2pt\vert}{\lbrace\kern-2.2pt\vert}
\newcommand{\vert\kern-2.2pt\rbrace}{\vert\kern-2.2pt\rbrace}
\newcommand{\pepasync}[1]{\raisebox{-1.0ex}{$\;\stackrel{\mbox{\large $\rhd\hspace{-1.2ex}\lhd$}}{\scriptscriptstyle #1}\,$}}
\newsavebox{\xtaTempBox}
\newlength{\xtaMinLen}
\newcommand{\rightarrowh}[1]{%
\mathbin{
\tikz[baseline=-.75ex]{
\setlength{\xtaMinLen}{1.2em}
\draw[#1] (0,0) -- (\xtaMinLen,0);}}}
\newcommand{\xrightarrowh}[4]{%
\sbox{\xtaTempBox}{\hbox{\( \scriptstyle\mkern#3#2\mkern#4 \)}}
\mathbin{
\tikz[baseline=-0.6ex]{
\setlength{\xtaMinLen}{1.2em}
\setlength{\xtaMinLen}{\maxof{\wd\xtaTempBox}{\xtaMinLen}}
\draw[#1] (0,0) --
node[midway,above=-0.4ex]{\usebox\xtaTempBox} (\xtaMinLen,0);}}}
\newcommand{\rightarrowh{-open triangle 60}}{\rightarrowh{-open triangle 60}}
\newcommand{\xrightarrowu}[1]{\xrightarrowh{-open triangle 60}{#1}{7mu}{17mu}}
\newcommand{\rightarrowh{-triangle 60}}{\rightarrowh{-triangle 60}}
\newcommand{\xrightarroww}[1]{\xrightarrowh{-triangle 60}{#1}{7mu}{17mu}}
\newcommand{\rightarrowh{-angle 90}}{\rightarrowh{-angle 90}}
\newcommand{\xrightarrowg}[1]{\xrightarrowh{-angle 90}{#1}{7mu}{9mu}}
\newcommand{\rightarrowh{->,double equal sign distance,-implies}}{\rightarrowh{->,double equal sign distance,-implies}}
\newcommand{\rightarrowh{-open triangle 60 reversed}}{\rightarrowh{-open triangle 60 reversed}}
\newcommand{\xrightarrowt}[1]{\xrightarrowh{-open triangle 60 reversed}{#1}{7mu}{17mu}}
\hypersetup{
colorlinks=true,
linkcolor=blue!50!black,
citecolor=blue!50!black,
filecolor=blue!50!black,
urlcolor=blue!50!black,
pdfpagelayout=SinglePage,
pdfpagemode=UseOutlines,
pdftitle={Structural operational semantics for non-deterministic processes with quantitative aspects},
pdfauthor={Marino Miculan and Marco Peressotti},
pdfsubject={Concurrency Theory},
pdfkeywords={rule formats, behavioural congruences, concurrent systems, quantitative models}
}
\title{
Structural operational semantics for\\
non-deterministic processes with quantitative aspects\thanks{
This work is partially supported by MIUR PRIN project 2010LHT4KM, \emph{CINA}.}
}
\author{
\begin{tabular}{ccc}
Marino Miculan&\qquad& Marco Peressotti\\[-.4ex]
\small\href{mailto:<EMAIL>}{\tt <EMAIL>}
&\qquad&
\small\href{mailto:<EMAIL>}{\tt <EMAIL>}
\end{tabular}\\[-.2ex]
\small Laboratory of Models and Applications of Distributed Systems \\[-.4ex]
\small Department of Mathematics and Computer Science\\[-.4ex]
\small University of Udine, Italy\\
}
\date{}
\begin{document}
\maketitle
\begin{abstract}
Recently, unifying theories for processes combining non-determinism
with quantitative aspects (such as probabilistic or stochastically
timed executions) have been proposed with the aim of providing
general results and tools. This paper provides two contributions in
this respect. First, we present a general GSOS specification format
and a corresponding notion of bisimulation for non-deterministic
processes with quantitative aspects. These specifications define
labelled transition systems according to the ULTraS model, an
extension of the usual LTSs where the transition relation associates
any source state and transition label with \emph{state reachability
weight functions} (like, e.g., probability distributions). This
format, hence called \emph{Weight Function GSOS} (WF-GSOS), covers
many known systems and their bisimulations (e.g.~PEPA, TIPP, PCSP)
and GSOS formats (e.g.~GSOS, Weighted GSOS, Segala-GSOS).
The second contribution is a characterization of these systems
as coalgebras of a class of functors, parametric in the weight
structure. This result allows us to prove \emph{soundness} and
\emph{completeness} of the WF-GSOS specification format, and
that bisimilarities induced by these specifications are always
congruences.
\end{abstract}
\section{Introduction}
Process calculi and labelled transition systems have proved very
successful for modelling and analysing concurrent, non-deterministic
systems.
This success has led to many extensions dealing with quantitative
aspects, whose transition relations are endowed with further
information like probability rates or stochastic rates; see
\cite{bg98:empa,denicola13:ultras,hillston:pepabook,ks2013:w-s-gsos,pc95:cj} among others. These calculi are very effective in modelling and
analysing quantitative aspects, like performance analysis of computer
networks, model checking of time-critical systems, simulation of
biological systems, probabilistic analysis of security and safety
properties, etc.
Each of these calculi is tailored to a specific quantitative aspect
and for each of them we have to develop a quite complex theory almost
from scratch. This is a daunting and error-prone task, as it embraces
the definition of syntax, semantics, transition rules,
various behavioural equivalences, logics, proof systems; the proof of
important properties like congruence of behavioural equivalences; the
development of algorithms and tools for simulations, model checking,
etc.
This situation would naturally benefit from general
\emph{frameworks} for LTS with quantitative aspects, i.e.,
mathematical \emph{meta}models offering general methodologies, results, and
tools, which can be uniformly instantiated to a wide range of specific
calculi and models.
In recent years, some of these theories have been proposed; we mention
\emph{Segala systems} \cite{sl:njc95},
\emph{Functional Transition Systems} (FuTS)
\cite{latella:qapl2015},
\emph{weighted labelled transition systems} (WLTSs)
\cite{handbook:weighted2009,ks2013:w-s-gsos},
and \emph{Uniform Labelled Transition Systems} (ULTraS), introduced by
Bernardo, De Nicola and Loreti specifically as ``a uniform setting for
modelling non-deterministic, probabilistic, stochastic or mixed
processes and their behavioural equivalences''
\cite{denicola13:ultras}.
A common feature of most of these meta-models is that their labelled
transition relations do not yield simple states (e.g., processes), but
some mathematical object representing quantitative information about
``how'' each state can be reached. In particular, transitions in
ULTraS systems have the form $P \xrightarrowu{a} \rho$ where $\rho$ is
a \emph{state reachability weight function}, i.e., a function
assigning a \emph{weight} to each possible state.\footnote{The reader
aware of advanced process calculi will be not baffled by the fact
that targets are not processes. Well known previous examples are the
LTS abstractions/concretions for $\pi$-calculus, for the applied
$\pi$-calculus, for the ambient calculus, etc.} By suitably
choosing the set of weights, and how these functions can be combined,
we can recover ordinary non-deterministic LTSs, probabilistic
transition systems, stochastic transition systems, etc. As
convincingly argued in \cite{denicola13:ultras}, the use of weight
functions in place of plain processes simplifies the combination of
non-determinism with quantitative aspects, like in the case of EMPA or
PEPA. Moreover, it paves the way for general definitions and results,
an important example being the notion of $\mathcal M$-bisimulation
\cite{denicola13:ultras}.
Albeit quite effective, these meta-models are at their dawn, with many
results and techniques still to be developed. An important example of
these missing notions is a \emph{specification format}, like the
well-known GSOS, ntyft/ntyxt and ntree formats for non-deterministic
labelled transition systems. These formats are very useful in
practice, because they can be used for ensuring important properties
of the system; in particular, the bisimulations induced by systems in
these formats is guaranteed to be a congruence (which is crucial for
compositional reasoning). From a more foundational point of view,
these frameworks would benefit from a categorical characterization in
the theory of coalgebras and bialgebras: this would allow a
cross-fertilizing exchange of definitions, notions and techniques with
similar contexts and theories.
In this paper, we provide two main contributions in this respect.
First, we present a GSOS-style format, called \emph{Weight Function
GSOS} (WF-GSOS), for the specifications of non-deterministic systems
with quantitative aspects. The judgement derived by rules in this
style is of the form $P \xrightarrowu{a} \psi$, where $P$ is a process
and $\psi$ is a \emph{weight function term}. These terms describe
weight functions by means of an \emph{interpretation}; hence, a
specification given in this format defines a ULTraS. By choosing the
set of weights, the language of weight function terms and their
interpretation, we can readily capture many quantitative notions
(probabilistic, stochastic, etc.), and different kinds of
non-deterministic interactions, covering models like PEPA, TIPP, PCSP,
EMPA, among others. Moreover, the WF-GSOS format supports a general
definition of \emph{(strong) bisimulation}, which can be readily
instantiated to the various specific systems.
The second contribution is more fundamental. We provide a general
categorical presentation of these non-deterministic systems with
quantitative aspects. Namely, we prove that ULTraS systems are in
one-to-one correspondence with coalgebras of a precise class of
functors, parametric on the underlying weight structure. Using this
characterization we define the abstract notion of \emph{WF-GSOS
distributive law} (i.e.~a natural transformation of a specific
shape) for these functors. We show that each WF-GSOS specification
yields such a distributive law (i.e., the format is sound); taking
advantage of Turi-Plotkin's bialgebraic framework, this implies that
the bisimulation induced by a WF-GSOS is always a congruence, thus
allowing for compositional reasoning in quantitative settings.
Additionally, we extend the results we presented in \cite{mp:qapl14}
proving that the WF-GSOS format is also \emph{complete}: every
abstract WF-GSOS distributive law for ULTraSs can be described by
means of some WF-GSOS specification.
The rest of the paper is structured as follows. In
Section~\ref{sec:ultras} we recall Uniform Labelled Transition
Systems, and their bisimulation. In Section~\ref{sec:WF-GSOS} we
introduce the \emph{Weight Function SOS} specification format for the
syntactic presentation of ULTraSs.
In Section~\ref{sec:examples} we provide some application examples,
such as a WF-GSOS specification for PEPA and the translations of
Segala-GSOS and WGSOS specifications in the WF-GSOS format. The
categorical presentation of ULTraS and WF-GSOS, with the results that
the format is sound and complete and bisimilarity is a congruence, are
in Section~\ref{sec:coalg}. Final remarks, comparison with related
work and directions for future work are in Section~\ref{sec:concl}.
\section{Uniform Labelled Transition Systems and their bisimulation}\label{sec:ultras}
In this section we recall and elaborate the definition of ULTraSs, and
define the corresponding notion of (coalgebraically derived)
bisimulation; finally we compare it with the notion of $\mathcal
M$-bisimulation presented in \cite{denicola13:ultras}.
\assumeapx{Additional examples are provided in the Appendix. }%
Although we focus on the ULTraS framework, the results and
methodologies described in this paper can be ported to similar formats
(like FuTS \cite{latella:qapl2015}),
and more generally to a wide range of systems combining
computational aspects in different ways.
\subsection{Uniform Labelled Transition Systems}
ULTraS are
(non-deterministic) labelled transition systems whose transitions lead
to \emph{state reachability weight functions}, i.e.~functions
representing quantitative information about ``how'' each state can be
reached. Examples of weight functions include probability
distributions, resource consumption levels, or stochastic rates. In
this light, ULTraS can be thought of as a generalization of Segala systems
\cite{sl:njc95}, which stratify non-determinism over probability.
Following the parallel with Segala systems, ULTraS transitions can be
pictured as being composed by two steps:
\[
x \xrightarrowu{a} \rho \xrightarroww{w} y
\]
where the first is a labelled non-deterministic (sub)transition
and the second is a weighted one; from this perspective the weight function
plays the r\^ole of the ``hidden intermediate state''.
Akin to Weighted Labelled Transition Systems (WLTS)
\cite{ks2013:w-s-gsos,handbook:weighted2009},
weights are drawn from a fixed
set endowed with a commutative monoid structure, where the unit is
meant to be assigned to disabled transitions (i.e.~those yielding
unreachable states) and the monoidal addition is used to
compositionally weigh sets of transitions given by non-determinism.
\begin{definition}[$\mathfrak W$-ULTraS]
\label{def:ultras}
Given a commutative monoid $\mathfrak W = (W,+,0)$, a
\emph{($\mathfrak W$-weighted) Uniform Labelled Transition System}
is a triple $(X,A,\rightarrowh{-open triangle 60})$ where:
\begin{itemize}
\item
$X$ is a set of \emph{states} (processes) called \emph{state space} or \emph{carrier};
\item
$A$ is a set of \emph{labels} (actions);
\item
${\rightarrowh{-open triangle 60}} \subseteq X\times A \times [X \to W]$
is a \emph{transition relation} where $[X \to W]$
denotes the set of all \emph{weight functions} from $X$ to
the carrier of $\mathfrak W$.
\end{itemize}
\end{definition}
Monoidal addition does not play any r\^ole in the above
definition\footnote{%
Originally, in \cite{denicola13:ultras} $W$ is a partial order
with bottom. Actually, the order is not crucial to the basic definition
of ULTraS as it is only used by some equivalences considered in that
paper.
} but it is crucial to define the notion of bisimulation
and in general how the ``merging'' of two states (e.g.~induced by functions between carriers) affects the transition relation. In fact, bisimulations
can be thought as inducing ``state space refinements that are well-behaved
w.r.t.~the transition relation''. From this perspective, monoidal addition
provides an \emph{abstract, uniform and compositional} way to ``merge'' the
outgoing transitions into one: adding their weight; likewise probabilities
or stochastic rates are added in probabilistic or stochastic systems.
Because the monoidal structure supports finite addition
only\footnote{%
Indeed it is possible to assume sums for any family indexed by some
set; however, in Section~\ref{sec:coalg} we assume image-finiteness
to guarantee the existence of a final coalgebra. } we can only
merge finitely many transitions. Assuming ULTraSs to have a finite
carrier or maps between carriers to define finite pre-images
(i.e.~$|f^{-1}(y)| \in \mathbb N$) is preposterous: since we aim to
provide syntactic description of ULTraSs, state spaces may be infinite
(cf.~initial semantics) and functions may map arbitrary many states to
the same image, e.g., their behaviour (cf.~bisimulations, final
semantics). Therefore, in this paper we shall consider \emph{image
finite} ULTraSs only. This is a mild and common assumption
(e.g.~\cite{ks2013:w-s-gsos,bartels04thesis,bloomIM:95}) and our
results readily generalise to transfinite bounds (e.g.~to deal with
countably-branching systems).
\begin{definition}[Image finiteness]
Let $\mathfrak W = (W,+,0)$ be a commutative monoid.
For a function $\rho : X \to W$ the set $\support{\rho} \triangleq \{ x \mid \rho(x) \neq 0\}$
is called \emph{support of $\rho$} and whenever it is finite
$\rho$ is said to be \emph{finitely supported}.
The set of finitely supported functions with domain $X$
is denoted by $\ff{W} X$.
A $\mathfrak W$-ULTraS $(X,A,\rightarrowh{-open triangle 60})$ is said to be \emph{image finite}
iff for any state $x \in X$ and label $a \in A$ the set
$\{\rho \mid x\xrightarrowu{a}\rho\}$ is finite and contains
only finitely supported weight functions.
\end{definition}
\begin{example}
A weight function $\rho \in \ff{\!\!2}X$
(for $\mathfrak 2 = (\{\texttt{t\!t},\texttt{f\!f}\},\lor,\texttt{f\!f})$)
is a predicate describing a finite subset of $X$.
Thus $\p{\!\!f} X \cong \ff{\!\!2}X$.
Likewise, a function $\rho \in \ff{N}X$
(for $\mathfrak N = (\mathbb N,+,0)$) assigns to
each element of $X$ a multiplicity and hence
describes a finite multiset.
\end{example}
Intuitively, elements of $\ff{W} X$ can be seen as ``generalised
multisets''. Therefore, it is natural to extend a function $f : X \to Y$
to a function $\ff{W}(f) : \ff{W} X \to \ff{W} Y$ mapping (finitely supported)
weight functions over $X$ to (finitely supported) weight functions over
$Y$ as follows:
\begin{equation}\label{eq:fw-action}
\textstyle
\ff{W}(f)(\rho) \triangleq \lambda y:Y.
\sum_{x \in f^{-1}(y)}\rho(x)\text{.}
\end{equation}
This definition generalises the extension of a function to the powerset;
in fact, $\ff{\!\!2}(f)(\rho) = \lambda y : Y.\bigvee_{x \in f^{-1}(y)}\rho(x)$
describes the subset of $Y$ whose elements are image of
some element in the subset of $X$ described by $\rho$.
Henceforth, we shall refer to $\ff{W}(f)(\rho)$ as the
\emph{action of $f$ on $\rho$} and denote it by $\rho[f]$,
when confusion seems unlikely.
We can now make the idea of
``state space maps being well-behaved w.r.t.~the transition relation''
formal:
\begin{definition}[ULTraS homomorphism]
Let $(X,A,\rightarrowh{-open triangle 60}_X)$ and $(Y,A,\rightarrowh{-open triangle 60}_Y)$ be two image-finite
$\mathfrak W$-ULTraS. A \emph{homomorphism} $f : (\rightarrowh{-open triangle 60}_X) \to (\rightarrowh{-open triangle 60}_Y)$
is a function $f : X \to Y$ between their state spaces
such that for any $x \in X$ and $a \in A$:
\[
x \xrightarrowu{a}_{\!\!X} \rho \iff f(x) \xrightarrowu{a}_{\!\!Y} \rho[f]
\text{.}
\]
\end{definition}
Given two homomorphisms
$f : (\rightarrowh{-open triangle 60}_X) \to (\rightarrowh{-open triangle 60}_Y)$ and
$g : (\rightarrowh{-open triangle 60}_Y) \to (\rightarrowh{-open triangle 60}_Z)$,
the function $g \circ f : X \to Z$ is a homomorphism
$g \circ f : (\rightarrowh{-open triangle 60}_X) \to (\rightarrowh{-open triangle 60}_Z)$.
Homomorphism composition is always defined, it is associative
and has identities.
In Section~\ref{sec:coalg} we will show that ULTraSs homomorphisms
indeed form categories equivalent to categories of coalgebras for a suitable
functor. For the time being, consider the degenerate monoid $\mathfrak 1$
containing exactly its unit and let $A$ be a singleton;
then a $\mathfrak 1$-ULTraS $(X,A,\rightarrowh{-open triangle 60}_X)$ is just a relation
$\rightarrowh{-open triangle 60}_X \cong R_X$ on $X$ and
any homomorphism is exactly a relation homomorphism. In fact,
$f : X \to Y$ is a $\mathfrak 1$-ULTraS homomorphism
$f : (\rightarrowh{-open triangle 60}_X) \to (\rightarrowh{-open triangle 60}_Y)$ iff
$(x,x') \in R_x \iff (f(x),f(x')) \in R_Y$. For $A$ with more
than one label we get exactly homomorphisms of labelled relations
i.e.~LTSs.
\subsection{Bisimulation}
We present now the definition of bisimulation for ULTraS
based on the notion of \emph{kernel bisimulation}
(a.k.a.~behavioural equivalence) i.e.~``a relation which is
the kernel of a common compatible refinement of the two\footnote{%
We present bisimulations as relations between two
state spaces instead of considering one system in isolation;
we are aware that in the case of ULTraS two systems can
be ``run in parallel'' still the notion of having a common
refinement allows for different homomorphisms even when
considering a single system and therefore offers greater
generality.
} systems''
\cite{staton11}. This notion naturally stems from the final
semantics approach and, under mild assumptions, coincides
with Aczel-Medler's coalgebraic bisimulation, as we will see in Section~\ref{sec:coalg}.
\begin{definition}[Refinement]
Given $(X,A,\rightarrowh{-open triangle 60}_X)$ a \emph{refinement} for it
is any $(Y,A,\rightarrowh{-open triangle 60}_Y)$ such that there exists
an homomorphism $f : (\rightarrowh{-open triangle 60}_X) \to (\rightarrowh{-open triangle 60}_Y)$.
\end{definition}
Homomorphisms provide the right notion of refinement.
Consider an equivalence relation $R \subseteq X \times X$,
$R$ is \emph{stable w.r.t.~$\rightarrowh{-open triangle 60}_X$}
if, and only if, its equivalence classes are not split by
the transition relation $\rightarrowh{-open triangle 60}_X$, i.e.,
iff there is a refinement whose carrier is $Y = X/R$.
Hence, stability of an equivalence relation corresponds to the
canonical projection $\kappa : X \to X/R$ being a ULTraS homomorphism.
This observation contains all the ingredients needed
to define bisimulations for ULTraSs. Before we formalise this
notion let us introduce some accessory notation.
In the following, we will denote the \emph{total weight} of
$\rho \in \ff{W} X$ by
$\totalweight{\rho} \triangleq \sum_{x \in X} \rho(x)$.
The weight $\rho$ assigned to $C \subseteq X$ is
the total weight of the restriction $\restr{\rho}{C}$
i.e.~$\totalweight{\restr{\rho}{C}} = \sum_{x \in C} \rho(x)$.
Any relation $R$ between two sets $X$ and $Y$ defines a relation
$R_\mathfrak{W}$
between finitely supported weight functions for $X$ and $Y$ as:
\[
(\phi,\psi) \in R_\mathfrak{W} \stackrel{\triangle}{\iff}
\forall (C,D) \in R^\star\,
\totalweight{\restr{\phi}{C}} =
\totalweight{\restr{\psi}{D}}
\]
where $R^\star\subseteq \p{}X\times\p{}Y$ is the
\emph{subset closure} of $R$
i.e.~smallest relation s.t., for $C\subseteq X$, $D\subseteq Y$:
\begin{align*}
(C,D)\in R^\star \iff & (\forall x\in C,\forall y\in Y: (x,y)\in R \Rightarrow
y\in D) \wedge \\&
(\forall x\in X,\forall y\in D: (x,y)\in R \Rightarrow x\in C)
\end{align*}
\begin{definition}[Bisimulation]
\label{def:bisim}
Let $(X,A,\rightarrowh{-open triangle 60}_X)$ and $(Y,A,\rightarrowh{-open triangle 60}_Y)$ be two image-finite
$\mathfrak W$-ULTraS. A relation $R$ between $X$ and $Y$
is a \emph{bisimulation} if, and only if,
for each pair of states $x \in X$ and $y \in Y$,
$(x,y) \in R$ implies that for each label $a \in A$
the following hold:
\begin{itemize}
\item
if ${x\xrightarrowu{a}_{\!\!X}\phi}$ then there exists
${y\xrightarrowu{a}_{\!\!Y}\psi}$ s.t.~$(\phi,\psi) \in R_\mathfrak{W}$.
\item
if ${y\xrightarrowu{a}_{\!\!Y}\psi}$ then there exists
${x\xrightarrowu{a}_{\!\!X}\phi}$ s.t.~$(\phi,\psi) \in R_\mathfrak{W}$.
\end{itemize}
Processes $x$ and $y$ are said to be \emph{bisimilar}
if there exists a bisimulation relation $R$ such that $(x,y) \in R$.
\end{definition}
\looseness=-1
As ULTraSs can be seen as stacking non-determinism over other
computational behaviour, Definition~\ref{def:bisim} stratifies
bisimulation for non-deterministic labelled transition system over
bisimulation for systems expressible as labelled transition systems
weighted over commutative monoids. In fact, two processes $x$ and $y$ are related by some
bisimulation if, and only if, whether one reaches a weight function
via a non-deterministic labelled transition, the other can reach
another function via a transition with the same label, where the
two functions are equivalent in the sense that they assign the
same total weight to the classes of states in the relation. For
instance, in the case of weights being probabilities, functions are
considered equivalent only when they agree on the probabilities
assigned to each class of states which is precisely the intuition
behind probabilistic bisimulation \cite{ls:probbisim}.
More examples will be discussed below\assumeapx{ and in the Appendix}.
\paragraph{Constrained ULTraS}
Sometimes, the ULTraSs induced by a given monoid are too many, and we
have to restrict to a subclass. For instance, fully-stochastic
systems such as (labelled) CTMCs are a strict subclass of ULTraSs
weighted over the monoid of non-negative real numbers $(\mathbb
R_0^+,+,0)$, where weights express rates of exponentially distributed
continuous time transitions. In the case of fully-stochastic systems,
for each label, each state is associated with precisely one weight
function. This kind of ``deterministic'' ULTraSs are called
\emph{functional} in \cite{denicola13:ultras}, because the transition
relation is functional, and correspond precisely to
WLTSs~\cite{ks2013:w-s-gsos,handbook:weighted2009}.
These are a well-known family of systems (especially their automata
counterpart) and have an established coalgebraic understanding as long
as a (coalgebraically derived) notion of \emph{weighted bisimulation}
which are shown to subsume several known kinds of systems such as
non-deterministic, (fully) stochastic, generative and reactive
probabilistic \cite{ks2013:w-s-gsos}. Moreover,
Definition~\ref{def:bisim} coincides with weighted bisimulation on
functional ULTraSs/WLTSs over the same monoid
\cite[Def.~4]{ks2013:w-s-gsos}; hence Definition~\ref{def:bisim}
covers every system expressible in the framework of WLTS.
\assumeapx{(cf.~Appendix~\ref{apx:vs-wlts}).}
\begin{proposition}
\label{prop:w-bisim}
Let $\mathfrak W$ be a commutative monoid and $(X,A,\rightarrowh{-open triangle 60}_X)$,
$(Y,A,\rightarrowh{-open triangle 60}_Y)$
be $\mathfrak W$-LTSs seen as a functional $\mathfrak W$-ULTraSs.
Every bisimulation relation between them is a $\mathfrak W$-weighted bisimulation
and vice versa.
\end{proposition}
\ifappendix
\begin{proof}[Proof (Omitted)]
See Appendix~\hyperref[proof:w-bisim]{\ref*{apx:vs-wlts}}.
\end{proof}
\fi
Another constraint arises in the case of probabilistic
systems, i.e., weight functions are probability distributions. Since
addition is not a closed operation in the unit interval $[0,1]$, there
is no monoid $\mathfrak W$ such that every weight function on it is also a
probability distribution. Altough we could relax
Definition~\ref{def:ultras}
to allow commutative \emph{partial} monoids\footnote{A commutative
partial monoid is a set endowed with a unit and a partial binary
operation which is associative and commutative, where it is defined,
and always defined on its unit. } such as the weight structure of
probabilities $([0,1],+,0)$, not every weight function
on $[0,1]$ is a probability distribution. In fact, probabilistic
systems (among others) can be recovered as ULTraSs over the $(\mathbb
R^+_0,+,0)$ (i.e.~the free completion of $([0,1],+,0)$) and subject to
suitable constraints. For instance, Segala systems \cite{sl:njc95}
are precisely the strict subclass of $\mathbb R^+_0$-ULTraS such that
every weight function $\rho$ in their transition relation is a
probability distribution i.e.~$\totalweight\rho = 1$. Moreover,
bisimulation is preserved by constraints; e.g., bisimulations
on the above class of (constrained) ULTraS corresponds to Segala's (strong)
bisimulations \cite[Def.~13]{sl:njc95}.
\begin{proposition}\label{prop:segala-bisim}
Let $(X,A,\rightarrowh{-open triangle 60}_X)$ and $(Y,A,\rightarrowh{-open triangle 60}_Y)$
be image-finite Segala-systems seen as
ULTraSs on $(\mathbb R^+_0,+,0)$.
Every bisimulation relation between them is a strong bisimulation
in the sense of \cite[Def.~13]{sl:njc95} and vice versa.
\end{proposition}
\ifappendix
\begin{proof}[Proof (Omitted)]
See Appendix~\hyperref[proof:segala-bisim]{\ref*{apx:vs-segala}}.
\end{proof}
\fi
A similar result holds for generative (or fully) or reactive
probabilistic systems and their bisimulations.
In fact, these are functional ${\mathbb R^0_+}$-ULTraS
s.t.~for all $x \in X$
${x \xrightarrowu{a} \rho} \implies \totalweight{\rho} \in \{0,1\}$
and
$\sum_{\{\rho \mid x \xrightarrowu{a} \rho\}}\totalweight{\rho} \in \{0,1\}$
respectively.
\subsection{Comparison with $\mathcal M$-bisimulation}
Bernardo et al.~defined a notion of bisimulation for ULTraS
parametrized by a function $\cal M$ which is used to weight sets of
(sequences of) transitions \cite[Def.~3.3]{denicola13:ultras}.
Notably, $\cal M$'s
codomain may be not the same of that used for weight functions in the
transition relation. This offers an extra degree of freedom with
respect to Definition~\ref{def:bisim}.
We recall the relevant definitions with minor modifications since
the original ones have to consistently weight also sequences of transitions
in order to account also for trace equivalences which are not in the scope of this paper.
\begin{definition}[$M$-function]
\label{def:mfun}
Let $(M,\bot)$ be a pointed\footnote{%
A pointed set (sometimes called based set or rooted set)
is a set equipped with a distinguished element
called (base) point; homomorphisms are
point preserving functions.
%
} set and $(X,A,\rightarrowh{-open triangle 60})$ be a
$\mathfrak W$-ULTraS.
A function $\mathcal M : X \times A\times \mathcal P X \to M$ is an
\emph{$M$-function for $(X,A,\rightarrowh{-open triangle 60})$} if, and only if,
it agrees with termination and class union, i.e.:
\begin{itemize}
\item
for all $x \in X$, $a\in A$ and $C \in \mathcal PX$,
$\mathcal M(x,a,C) = \bot$ whenever
${x \centernot{\xrightarrowu{a}}}$ or
$\totalweight{\restr{\rho}{C}} = 0$ for every $x \xrightarrowu{a}\rho$;
\item
for all $x,y \in X$, $a \in A$ and $C_1,C_2 \in \mathcal P X$, if
$\mathcal{M}(x,a,C_1) = \mathcal{M}(y,a,C_1)$ and $\mathcal{M}(x,a,C_2) = \mathcal{M}(y,a,C_2)$ then
$\mathcal{M}(x,a,C_1\cup C_2) = \mathcal{M}(y,a,C_1 \cup C_2)$.
\end{itemize}
\end{definition}
\begin{definition}[$\mathcal M$-bisimulation {\cite{denicola13:ultras}}]
\label{def:mbisim}
Let $\mathcal M$ be an $M$-function for $(X,A,\rightarrowh{-open triangle 60})$.
An equivalence relation $R \subseteq X\times X$ is a \emph{$\cal M$-bisimulation for $\rightarrowh{-open triangle 60}$}
iff for each pair $(x,y)\in R$, label $a \in A$, and class $C \in X/R$,
$\mathcal{M}(x,a,C) = \mathcal{M}(y,a,C)$.
\end{definition}
\looseness=-1
Differently from Definition~\ref{def:bisim}, $M$ may be not $W$
allowing one to, for instance, consider stochastic rates up-to a
suitable tolerance as a way to account for experimental measurement
errors in the model.
A further distinction between bisimulation and $\mathcal M$-bisimulation
arises from the fact that ULTraSs come with two distinct ways of
\emph{terminating}. A state can be seen as ``terminated'' either when
its outgoing transitions are always the constantly zero function, or
when it has no transitions at all. In the first case, the state has
still associated an outcome, saying that no further state is
reachable; we call these states \emph{terminal}. In the second case,
the LTS does not even tell us that the state cannot reach any further
state; in fact, there is no ``meaning'' associated to the state. In
this case, we say that the state is \emph{stuck}.\footnote{This is
akin to sequential programs: a terminal state is when we reach the
end of the program; a stuck state is when we are executing an
instruction whose meaning is undefined.} The bisimulation given in
Definition~\ref{def:bisim} keeps these two terminations as different
(i.e., they are not bisimilar), whereas $\mathcal M$-bisimulation does not
make this distinction (cf.~\cite[Def.~3.2]{denicola13:ultras} or, for
a concrete example based on Segala systems,
\cite[Def.~7.2]{denicola13:ultras}).
Finally, the two notions differ on the quantification over equivalence
classes: in the case of Definition~\ref{def:bisim} quantification
depends on the non-deterministic step whereas in the case of $\mathcal
M$-bisimulation it does not.
Under some mild assumptions, the two notions agree. In
particular, let us restrict to systems with just one of the two
terminations for each action $a$---i.e.~if for some $x$, $\{\rho\mid
x\xrightarrowu{a}\rho\} = \emptyset$ then for all $y$, $\lambda z. 0
\notin \{\rho\mid y\xrightarrowu{a}\rho\}$, and, symmetrically, if for
some $x$, $\lambda z. 0 \in \{\rho\mid x\xrightarrowu{a}\rho\}$ then
for all $y$, $\{\rho\mid y\xrightarrowu{a}\rho\} \neq \emptyset$.
Then, the bisimulation given in Definition~\ref{def:bisim} corresponds
to a $\mathcal M$-bisimulation for a suitable choice of $\mathcal M$.
\begin{proposition}\label{prop:vs-m-bisim}
Let $(X,A,\rightarrowh{-open triangle 60})$ be a $\mathfrak W$-ULTraS with at most one kind
of termination, for each label.
Every bisimulation $R$ is also an $\mathcal M$-bisimulation for
\[
\mathcal{M}(x,a,C) \triangleq
\{[\rho]_{R_{\mathfrak{W}}} \mid {x\xrightarrowu{a}\rho}
\text{ and }
\totalweight{\restr{\rho}{C}} \neq 0\}
\cup
\{[\lambda z. 0]_{R_{\mathfrak{W}}}\}
\]
where
$(M,\perp) = (\p{\!\!f}(\ff{W} X/{R_{\mathfrak{W}}}), \{[\lambda z. 0]_{R_{\mathfrak{W}}}\})$.
\end{proposition}
\begin{proofatend}
The function $\mathcal M$ is well-given because
$\mathcal M(x,a,C) = \bot = \perp$ whenever
$x \centernot{\xrightarrowu{a}}$ or, for each
$x \xrightarrowu{a} \rho$, $\rho(C) = 0$, and
$\mathcal{M}(x,a,C_1) = \mathcal{M}(y,a,C_1)$ and $\mathcal{M}(x,a,C_2) = \mathcal{M}(y,a,C_2)$
implies $\mathcal{M}(x,a,C_1\cup C_2) = \mathcal{M}(y,a,C_1 \cup C_2)$
by definition of $R_{\mathfrak{W}}$.
By Definition~\ref{def:bisim}, whenever $x \xrightarrowu{a} \phi$
then $y \xrightarrowu{a} \psi$ s.t.~$\phi(C) = \psi(C)$
for each $C\in X/R$ i.e.~$\phi R_{\mathfrak{W}} \psi$ and the symmetric case for $y$.
Therefore $(x,y) \in R$ implies that
$\Phi_{x,a} \triangleq \{[\phi]_{R_{\mathfrak{W}}} \mid x \xrightarrowu{a} \phi\}$
and
$\Phi_{y,a} \triangleq \{[\psi]_{R_{\mathfrak{W}}} \mid y \xrightarrowu{a} \psi\}$
are equal for each $a \in A$.
We can safely add $\perp$ to both $\Phi_{x,a}$
and $\Phi_{y,a}$ since, whenever both $x$ and $y$ terminate,
they are either both stuck or both terminal.
In fact, equality and inequality are preserved
while adding $\perp$ since $\Phi_{x,a} = \emptyset \implies
\perp \notin \Phi_{y,a}$ (and vice versa) by hypothesis.
For each $C\in X/R$ ($x,y \in X$ and $a \in A$) let $\Psi_{x,a,C} \triangleq
(\Phi_{x,a}\setminus\{[\rho]_{R_{\mathfrak{W}}}\mid\rho(C) = 0\})\cup\{\perp\}$.
Clearly $\Phi_{x,a} \cup = \bigcup_{C \in X/R} \Psi_{x,a,C}$
and if $(x,y) \in R$ then $\Psi_{x,a,C} = \Psi_{y,a,C}$.
Complementarly, if $(x,y) \notin R$
then there exists some $\phi \in \Phi_{x,a}$ s.t.~for no $\psi \in \Phi_{y,a}$
$\phi R_{\mathfrak{W}} \psi$ or vice versa; w.l.o.g.~assume the former.
Hence there exists $C \in X/R$ such that $\phi(C) \neq \psi(C)$
whence $\Psi_{x,a,C} \neq \Psi_{y,a,C}$. Finally, we conclude by
$\mathcal M(x,a,C) = \Psi_{x,a,C}$ for each $x \in X$, $a\in A$ and $C \in X/R$.
\end{proofatend}
Intuitively, Definition~\ref{def:bisim} generalises strong bisimulation
for Segala systems (Segala and Lynch's probabilistic bisimilarity \cite{sl:njc95})
and $\mathcal{M}$-bisimulation generalises convex bisimulation
\cite{denicola13:ultras}.
\section{WF-GSOS: A complete GSOS format for ULTraSs}\label{sec:WF-GSOS}
In this section we introduce the \emph{Weight Function SOS}
specification format for the syntactic presentation of ULTraSs. As it
will be proven in Section~\ref{sec:cong-proof}, bisimilarity for
systems given in this format is guaranteed to be a congruence with
respect to the signature used for representing processes.
The format is parametric in the weight monoid $\mathfrak W$ and, as usual,
in the \emph{process signature} $\Sigma$ defining the syntax of system
processes. In contrast with ``classic'' GSOS formats
\cite{klin:tcs11}, targets of rules are not processes but terms whose
syntax is given by a different signature, called the \emph{weight
signature}. This syntax can be thought of as an ``intermediate
language'' for representing weight functions along the line of viewing
ULTraSs as stratified (or staged) systems. An early example of this
approach can be found in \cite{bm:2015stocsos}, where targets are terms
representing measures over the continuous state space.
Earlier steps in this direction can be found e.g.~in Bartels' GSOS
format for Segala systems (cf.~\cite[§5.3]{bartels04thesis} and \cite[§4.2]{mp:qapl14}) or in
\cite{cm:quest10,denicola13:ultras} where targets are described by
meta-expressions.
\begin{definition}[WF-GSOS Rule]\label{def:wf-gsos-rule}
Let $\mathfrak W$ be a commutative monoid and $A$ a set of labels.
Let $\Sigma$ and $\Theta$ be the \emph{process signature} and
the \emph{weight signature}, respectively.
A WF-GSOS rule over them is a rule of the form:
\[\frac{
\begin{array}{c}
\Big\{
x_i \xrightarrowu{a} \phi^a_{ij}
\Big\}
\hspace{-1.2ex}\begin{array}{l}
\scriptstyle 1 \leq i \leq n,\\[-4pt]
\scriptstyle a \in A_i,\\[-4pt]
\scriptstyle 1\leq j \leq m^a_i
\end{array}
\quad
\Big\{
x_i \centernot{\xrightarrowu{b}}
\Big\}
\hspace{-1.2ex}\begin{array}{l}
\scriptstyle \\[-4pt]
\scriptstyle 1 \leq i \leq n,\\[-4pt]
\scriptstyle b \in B_i
\end{array}
\quad
\Big\{
\totalweight{\phi^{a_k}_{i_kj_k}} = w_k
\Big\}
_{1 \leq k \leq p}
\quad
\Big\{
\corestr{\phi^{a_k}_{i_kj_k}}{\mathfrak C_k} \ni y_k
\Big\}
_{1 \leq k \leq q}
\end{array}
}{
\mathtt{f}(x_1,\dots,x_n) \xrightarrowu{c} \psi
}\]
where:
\begin{itemize}\itemsep=0pt
\item
$\mathtt f$ is an $n$-ary symbol from $\Sigma$;
\item
$X = \{x_i\mid 1 \leq i \leq n\}$,
$Y = \{y_k\mid 1 \leq k \leq q\}$
are sets of pairwise distinct \emph{process} variables;
\item
$\Phi = \{\phi^a_{ij}\mid 1 \leq i \leq n,\ a \in A_i,\ 1\leq j \leq m^a_i\}$
is a set of pairwise distinct \emph{weight function} variables;
\item
$\{w_k \in \mathfrak W \mid 1 \leq k \leq p\}$ are
\emph{weight constants};
\item
$\{\mathfrak C_k \mid 1 \leq k \leq q,\, w_k \in \mathfrak C_k\}$ is a set of \emph{clubs} of $\mathfrak W$,
i.e.~subsets of $W$ being monoid ideals whose complements are sub-monoids of $\mathfrak W$;
\item
$a,b,c \in A$ are labels and $A_i \cap B_i = \emptyset$ for
$1 \leq i \leq n$;
\item
$\psi$ is a \emph{weight term} for the signature $\Theta$
such that $\mathrm{var}(\psi) \subseteq X \cup Y \cup \Phi$.
\end{itemize}
A rule like above is \emph{triggered} by
a tuple $\langle C_1,\dots,C_n\rangle$ of \emph{enabled labels} and
by a tuple $\langle v_1,\dots,v_p \rangle$ of weights
if, and only if,
$A_i \subseteq C_i$, $B_i \cap C_i = \emptyset$, and
$w_j = v_j$
for
$1 \leq i \leq n$ and
$1 \leq j \leq p$.
\end{definition}
Intuitively, the four families of premises can be grouped in two
kinds: the first two families correspond to the non-deterministic (and
labelled) behaviour, whereas the other two correspond to the weighting
behaviour of quantitative aspects. The former are precisely the
premises of GSOS rules for LTSs (up-to targets being functions), and
describe the possibility to perform some labelled transitions. The
latter are inspired by Bartels' \emph{Segala-GSOS}
\cite[§5.3]{bartels04thesis} and Klin's WGSOS \cite{ks2013:w-s-gsos} formats;
a premise like $\totalweight{\phi} = w$ constrains the variable
$\phi$ to those functions whose total weight is exactly the constant $w$;
a premise like $\corestr{\phi}{\mathfrak C}\ni y$ binds the process variable
$y$ to those elements being assigned a weight in $\mathfrak C$.
This kind of premises are meant to single out elements from weight
functions domain in a way that is coherent w.r.t.~function actions (hence
independent from carrier maps and variable substitutions).
To this end, selection may depend on weights only and has to be
unaffected by sums, i.e., $z = f(x) = f(x')$ is selected if and only if at
least $x$ or $x'$ is.
Clubs are the finest substructures of commutative monoids that are
``isolated'' w.r.t. the monoidal operation in the sense that:
\begin{itemize}
\item
are commutative monoid ideals, i.e.~subsets $\mathfrak C$ with a module structure;
\item
their complement $\overline{\mathfrak C}$ in $\mathfrak W$ is a sub-monoid of $\mathfrak W$.
\end{itemize}
Because of the first assumption
$v+w \in \mathfrak C \implies v \in \mathfrak C \lor w \in \mathfrak C$
and because of the second
$v,w \notin \mathfrak C \implies v+w \notin \mathfrak C$
In other words, if something is selected depending on its weight, no matter what is added to, it will remain selected and vice versa:
$v+w \in \mathfrak C \iff v \in \mathfrak C \lor w \in \mathfrak C$.
Note that no club can contain the unit $0$ (otherwise $\overline{\mathfrak C} = \emptyset$) and this ensures selections
to be confined within the weight function supports (hence to be finite).
\begin{remark}
The empty set trivially is a club.
Not all complements of submonoids are clubs, for instance
even natural numbers under addition are a submonoid of
$(\mathbb{N},+,0)$ but odd numbers are not a club;
the only non-empty club in $(\mathbb{N},+,0)$ is $\mathbb{N}\setminus\{0\}$.
Elements with an opposite cannot be part of a club: if
$x \in \mathfrak C$ then $x+(-x) = 0$ is in $\mathfrak C$
and hence $\overline{\mathfrak C}$ cannot be a submonoid of $\mathfrak W$.
\end{remark}
Like Segala-GSOS (but unlike WGSOS), there are no
variables denoting the weight of each $y_k$ since this information can be
readily extracted from $\phi^{a_k}_{i_k j_k}$, e.g.~by some
operator from $\Theta$ that ``evaluates'' $\phi^{a_k}_{i_k j_k}$
on $y_k$.
Targets of transitions defined by these rules are terms generated from
the signature $\Theta$. In order to characterize transition relations
for ULTraSs, we need to \emph{evaluate} these terms to weight
functions. This is obtained by adding an \emph{interpretation for
weight terms}, besides a set of rules in the above format.
Before defining interpretations and specifications, we need to
introduce some notation. For a signature $S$ and a set $X$ of
variable symbols, let $T^S X$ denote the set of terms freely
generated by $S$ over the variables $X$ (in the
following, $S$ will be either $\Sigma$ or $\Theta$). A substitution
for symbols in $X$ is any function $\sigma:X \to Y$;
its action extends to terms defining the function $T^S(\sigma) :
T^S X \to T^S Y$ (i.e.~simultaneous substitution). When confusion
seems unlikely we use the more evocative notation $\mathtt t[\sigma]$ instead of
$T^S(\sigma)(\mathtt t)$.
\begin{definition}[Interpretation]\label{def:wf-gsos-eval}
Let $\mathfrak W$ be a commutative monoid, $\Sigma$ and $\Theta$ be
the process and the weight signature respectively. A
\emph{weight term interpretation}
for them is a family of functions
\[
\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace_X : T^\Theta(X + \ff{W}(X)) \to \p{\!\!f}\ff{W}
T^\Sigma(X)
\]
indexed over sets of variable symbols, and respecting substitutions, i.e.:
\[
\forall \sigma : X \to Y, \psi\in T^\Theta(X):
\lbrace\kern-2.2pt\vert\psi\vert\kern-2.2pt\rbrace_X[\sigma] =
\lbrace\kern-2.2pt\vert \psi[\sigma] \vert\kern-2.2pt\rbrace_Y
\text{.}\]
\end{definition}
Different from \cite{mp:qapl14} interpretations allow one term to
represent finitely many weight functions. This generalization offers
more freedom in the use of the format by reducing the
constrains on what can be encoded in weight function terms
and simplifies the proof for completeness.
We are ready to introduce the WF-GSOS specification format.
Basically, this is a set of WF-GSOS rules,
subject to some finiteness conditions to ensure image-finiteness,
together with an interpretation.
\begin{definition}[WF-GSOS specification]\label{def:wf-gsos-spec}
Let $\mathfrak W$ be a commutative monoid, $A$ the set of labels, $\Sigma$
and $\Theta$ the process and the weight signature respectively.
An \emph{image-finite WF-GSOS specification over
$\mathfrak W, A, \Sigma$ and $\Theta$} is a pair $\langle\mathcal R,
\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace\rangle$ where
$\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace$ is a weight term interpretation
and $\mathcal R$ is a set of rules compliant with
Definition~\ref{def:wf-gsos-rule} and such that only finitely many
rules share the same operator in the source ($\mathtt f$), the same
label in the conclusion ($c$), and the same trigger $\langle
A_1,\dots,A_n\rangle$, $\langle w_1,\dots,w_p\rangle$.
\end{definition}
Every WF-GSOS specification induces an
ULTraS over ground process terms.
\begin{definition}[Induced ULTraS]\label{def:induced-ultras}
The ULTraS induced by an image-finite WF-GSOS specification
$\langle\mathcal R,\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace\rangle$ over $\mathfrak
W, \Sigma, \Theta$ is the $\mathfrak W$-ULTraS $(T^\Sigma\emptyset, A,
\rightarrowh{-open triangle 60})$ where $\rightarrowh{-open triangle 60}$ is defined as the smallest subset
of $T^\Sigma\emptyset
\times A \times \ff{W} T^\Sigma\emptyset$ being closed under the following condition.
Let $p = \mathtt{f}(p_1, \dots, p_n) \in T^\Sigma\emptyset$. Since the ground
$\Sigma$-terms $p_i$ are structurally smaller than $p$ assume
(by structural recursion)
that the set $\{\rho \mid p_i \xrightarrowu{a} \rho\}$ --
and hence the trigger $\vec{A} = \langle A_1,\dots,A_n\rangle$,
$\vec{w} = \langle w_1,\dots,w_q\rangle$ --
is determined for every $i \in \{1,\dots,n\}$ and $a\in A$.
For any rule $R \in \mathcal R$ whose conclusion is of the form
$\mathtt{f}(x_1,\dots,x_n) \xrightarrowu{c} \psi$ and triggered by
$\vec A$ and $\vec w$
let $X$, $Y$, $\Phi$ be the set of process and weight function
variables involved in $R$ as per Definition~\ref{def:wf-gsos-rule}.
Then, for any substitution $\sigma:X\cup Y \to T^\Sigma\emptyset$
and map $\theta : \Phi \to \ff{W} T^\Sigma\emptyset$ such that:
\begin{enumerate}
\item
$\sigma(x_i) = p_i$ for $x_i \in X$;
\item
$\theta(\phi^a_{ij}) = \rho$
for each premise $x_i\xrightarrowu{a}\phi^a_{ij}$
and $\totalweight{\phi^{a}_{ij}} = w_k$ of $R$,
and for any $\rho$ such that $p_i\xrightarrowu{a} \rho$
and $\totalweight{\rho} = w_k$;
\item
$\sigma(y_k) = q_k$ for each premise
$\corestr{\phi^{a_k}_{i_k j_k}}{\mathfrak C_k}\ni y_k$
of $R$ and for any $q_k\in T^\Sigma\emptyset$ s.t.~ $\theta(\phi^{a_k}_{i_k j_k})(q_k) \in \mathfrak C_k$;
\end{enumerate}
there is $p \xrightarrowu{c} \rho$ where
$\rho \in \lbrace\kern-2.2pt\vert \psi[\theta]\vert\kern-2.2pt\rbrace_{X\cup Y}[\sigma]$
is an instantiated interpretation of the target $\Theta$-term $\psi$.
\end{definition}
The above definition is well-defined since it is based on structural recursion
over ground $\Sigma$-terms (i.e.~the process $p$ in each triple $(p,a,\rho)$);
in particular, terms have finite depth and only structurally smaller terms
are used by the recursion (i.e.~the assumption of $p_i \xrightarrowu{a} \rho$
being defined for each $p_i$ in $p = \mathtt{f}(p_1, \dots, p_n)$).
Moreover, for any trigger, operator, and conclusion label
only finitely many rules have to be considered.
Finally we can state the main result for the proposed format.
\begin{theorem}[Congruence]\label{th:congruence-1}
The bisimulation on the ULTraS induced by a
WF-GSOS specification is a congruence with respect
to the process signature.
\end{theorem}
The proof is postponed to Section~\ref{sec:cong-proof},
where we will take advantage of the bialgebraic framework.
\begin{remark}[Expressing interpretations]\rm
Weight term interpretation can be defined in many
ways, e.g.~by structural recursion on $\Theta$-terms.
For instance, every substitution-respecting family of maps:
\[
h_X : \Theta\ff{W} T^\Sigma(X) \to \p{\!\!f}\ff{W} T^\Sigma(X)
\qquad
b_X : X \to \p{\!\!f}\ff{W} T^\Sigma(X)
\]
uniquely extends to an interpretation by structural recursion on
$\Theta$-terms where $h_X$ and $b_X$ define the inductive and base
cases respectively. These maps can be easily given by means of a set
of equations, as in~\cite[§4.1]{mp:qapl14}.
\end{remark}
\section{Examples and applications of WF-GSOS specifications}
\label{sec:examples}
In this section we provide some examples of applications of the WF-GSOS format.
First, we show how a process calculus can be given a WF-GSOS
specification; in particular, we consider PEPA, a well known process
algebra with quantitative features. Then we show that Klin's Weighted
GSOS format for weighted systems \cite{ks2013:w-s-gsos} and Bartels'
Segala-GSOS format for Segala systems \cite{bartels04thesis} are
subsumed by our WF-GSOS format; this corresponds to the fact that
ULTraSs subsume both weighted and Segala systems.
\subsection{WF-GSOS for PEPA}\label{sec:pepa-WF-GSOS}
In PEPA \cite{hillston:pepabook,hillston05}, processes are terms over the grammar:
\begin{equation}
\label{eq:pepa-grammar}
P ::= (a,r).P \mid P + P \mid P \pepasync{L} P
\mid P\setminus L
\end{equation}
where $a$ ranges over a fixed set of labels $A$, $L$ over subsets of $A$ and
$r$ over $\mathbb R^+$.
The semantics of process terms is usually defined by the inference rules
in Figure~\ref{fig:pepa-classic-sos}
\begin{figure}
\[\begin{array}{c}
\frac{}{(a,r).P \xrightarrowg{a,r} P}
\quad
\frac{P_1 \xrightarrowg{a,r} Q}{P_1 + P_2 \xrightarrowg{a,r} Q}
\quad
\frac{P_2 \xrightarrowg{a,r} Q}{P_1 + P_2 \xrightarrowg{a,r} Q}
\quad
\frac{P \xrightarrowg{a,r} Q}{P \setminus L \xrightarrowg{a,r} Q
}\ a \notin L
\quad
\frac{P \xrightarrowg{a,r} Q}{
P \setminus L \xrightarrowg{\tau,r} Q
}\ a \in L
\\
\frac{
P_1 \xrightarrowg{a,r_1}Q_1 \quad P_2 \xrightarrowg{a,r_2}Q_2}{
P_1 \pepasync{L}P_2 \xrightarrowg{a,R} Q_1 \pepasync{L} Q_2
}\ a \in L
\quad
\frac{P_1 \xrightarrowg{a,r}Q}{
P_1 \pepasync{L}P_2 \xrightarrowg{a,r} Q \pepasync{L} P_2
}\ a \notin L
\quad
\frac{P_2 \xrightarrowg{a,r}Q}{
P_1 \pepasync{L}P_2 \xrightarrowg{a,r} P_1 \pepasync{L} Q
}\ a \notin L
\end{array}\]\vspace{-1ex}
\caption{Structural operational semantics for PEPA.}
\label{fig:pepa-classic-sos}
\end{figure}
where $a \in A$, $r,r_1,r_2,R \in \mathbb R^+$
(passive rates are omitted for simplicity) and $R$ depends only on
$r_1$, $r_2$ and the intended meaning of synchronisation.
For instance, in applications to performance evaluation
\cite{hillston:pepabook}, rates model time and $R$ is
defined by the \emph{minimal rate law}:
\begin{equation}
\label{eq:minimal-rate-law}
R = \frac{r_1}{r_a(P_1)}\cdot
\frac{r_2}{r_a(P_2)}\cdot
\min(r_a(P_1),r_a(P_2))
\end{equation}
where $r_a$ denotes the apparent rate of $a$ \cite{hillston:pepabook}.
PEPA can be characterized by a specification in the WF-GSOS format where
the process signature $\Sigma$ is the same as \eqref{eq:pepa-grammar}
and weights are drawn from the monoid of positive real numbers under
addition extended with the $+\infty$ element (only for technical
reasons connected with the ${\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace}$ and
process variables---differently from other stochastic process algebras
like EMPA \cite{bg98:empa}, PEPA does not allow instantaneous actions,
i.e.~with rate $+\infty$). The intermediate language of weight terms
is expressed by the grammar:
\[
\theta ::= \perp \mid \diamondsuit_r(\theta) \mid \theta_1\oplus \theta_2 \mid
\theta_1\parallel_L \theta_2 \mid \xi \mid P
\]
where $r\in \mathbb R^+_0$, $L \subseteq A$, $\xi$ range over weight
functions $\ff{W} X$, and $P$ over processes in $T^\Sigma X$ for some set $X$.
Note that the grammar is untyped
since it describes the terms freely generated by the signature $\Theta
= \{\perp:0,\diamondsuit_r:1,\oplus:2,\parallel_L:2\}$, over
weight function variables and processes. Intuitively $\perp$ is
the constantly $0$ function, $\diamondsuit_r$ reshapes its argument to
have total weight $r$, $\oplus$ is the point-wise sum and
$\parallel_L$ parallel composition e.g.~by
\eqref{eq:minimal-rate-law}. The formal meaning of these operators is
given below by the definition (by structural recursion on
$\Theta$-terms) of the interpretation
${\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace}$ which is introduced alongside
WF-GSOS rules for presentation convenience. Each operator is
interpreted as a singleton (PEPA describes functional ULTraSs) and
hence we will describe ${\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace}$ as if a
weight function is returned.
For each action $a\in A$ and rate $r \in \mathbb R^+$, a process
$(a,r).P$ presents exactly one $a$-labelled transition ending in the
weight function assigning $r$ to the (sub)process denoted by the variable $P$
and $0$ to everything else. Hence, the \emph{action axiom} is expressed as
follows:
\[
\frac{}{(a,r).P \xrightarrowu{a} \diamondsuit_r(P)}
\qquad
\lbrace\kern-2.2pt\vert\diamondsuit_r(\psi)\vert\kern-2.2pt\rbrace_{X}(t) =
\begin{cases}
\frac{r}{
\raisebox{-1pt}{$\scriptstyle\left|
\raisebox{-1pt}{$\scriptstyle
\support{\lbrace\kern-2.2pt\vert\psi\vert\kern-2.2pt\rbrace_{X}\!}
$}\right|$}} & \text{if } \lbrace\kern-2.2pt\vert\psi\vert\kern-2.2pt\rbrace_{X}(t)\neq 0\\
0 & \text{otherwise}
\end{cases}
\]
where $\diamondsuit_r$ normalises\footnote{Since the interpretation
$\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace$ is being defined by structural
recursion and has to cover all the language freely generated from
$\Theta$, we can not use the (slightly more intuitive) ``Dirac''
operator $\delta_r(P)$ where $P$ is restricted to be a process
variable instead of a $\Theta$-term. Likewise, indexing
$\delta_{r,P}$ also over processes would break substitution
independence i.e.~naturality.} $\lbrace\kern-2.2pt\vert P \vert\kern-2.2pt\rbrace_{X}$
to equally distribute the weight $r$ over its support; in particular,
since process variables will be interpreted as ``Dirac-like''
functions $\diamondsuit_r(P)$ corresponds to the weight function
assigning $r$ to $\Sigma$-term denoted by $P$.
Conversely to the action axiom, $(a,r).P$ can not perform any action but $a$:
\[
\frac{}{(a,r).P \xrightarrowu{b} \perp}\ a \neq b
\qquad
\lbrace\kern-2.2pt\vert \perp \vert\kern-2.2pt\rbrace_{X}(t) = 0
\]
This rule is required to obtain a functional ULTraS and
is implicit in Figure~\ref{fig:pepa-classic-sos}
where disabled transitions are assumed with rate $0$
as in any specification in the Stochastic GSOS or Weighted
GSOS formats. Without this rule, transitions would have
been disabled in the non-deterministic layer
i.e.~$(a,r).P\centernot{\xrightarrowu{b}}$.
Stochastic choice is resolved by the stochastic race, hence the rate
of each competing transition is added point-wise as in
Figure~\ref{fig:pepa-classic-sos} (and in the SGSOS and WGSOS
formats). This passage belongs to the stochastic layer of the
behaviour (hence to the interpretation, in our setting) whereas the
selection of which weight functions to combine is in the
non-deterministic behaviour represented by the rules and, in
particular, to the labelling. Therefore, the \emph{choice rules}
become:
\[
\frac{P_1 \xrightarrowu{a} \phi_1 \quad P_2 \xrightarrowu{a} \phi_2}{
P_1 + P_2 \xrightarrowu{a} \phi_1 \oplus \phi_2
}\qquad
\lbrace\kern-2.2pt\vert \psi \oplus \phi \vert\kern-2.2pt\rbrace_{X}(t) =
\lbrace\kern-2.2pt\vert \psi \vert\kern-2.2pt\rbrace_{X}(t) +
\lbrace\kern-2.2pt\vert \phi \vert\kern-2.2pt\rbrace_{X}(t)
\]
Likewise, process cooperation depends on the labels to select the
weight function to be combined. This is done in the next two rules:
one when the two processes cooperate, and the other when one process
does not interact on the channel:
\begin{gather*}
\frac{P_1 \xrightarrowu{a} \phi_1 \quad P_2 \xrightarrowu{a} \phi_2}{
P_1 \pepasync{L} P_2 \xrightarrowu{a} \phi_1 \parallel_{L} \phi_2
} a \in L
\quad
\frac{P_1 \xrightarrowu{a} \phi_1 \quad P_2 \xrightarrowu{a} \phi_2}{
P_1 \pepasync{L}P_2 \xrightarrowu{a}
(\phi_1 \parallel_{L} P_2) \oplus
(P_1 \parallel_{L} \phi_2)
} a \notin L
\end{gather*}
The combination step depends on the minimal rate law \eqref{eq:minimal-rate-law}:
\[
\lbrace\kern-2.2pt\vert \psi \parallel_{L} \phi \vert\kern-2.2pt\rbrace_{X}(t) =
\begin{cases}
\frac{\lbrace\kern-2.2pt\vert \psi \vert\kern-2.2pt\rbrace_{X}(t_1)}{
\totalweight{\lbrace\kern-2.2pt\vert \psi \vert\kern-2.2pt\rbrace_{X}}}\cdot
\frac{\lbrace\kern-2.2pt\vert \phi \vert\kern-2.2pt\rbrace_{X}(t_2)}{
\totalweight{\lbrace\kern-2.2pt\vert \phi \vert\kern-2.2pt\rbrace_{X}}}\cdot
\min(\totalweight{\lbrace\kern-2.2pt\vert \psi \vert\kern-2.2pt\rbrace_{X}},\totalweight{\lbrace\kern-2.2pt\vert
\phi \vert\kern-2.2pt\rbrace_{X}}) & \text{if } t = t_1 \pepasync{L} t_2\\
0 & \text{otherwise}
\end{cases}
\]
Each process is interpreted as a weight function over process terms.
This is achieved by a Dirac-like function assigning $+\infty$ to the
$\Sigma$-term composed by the aforementioned variable: $\lbrace\kern-2.2pt\vert
P \vert\kern-2.2pt\rbrace_{X}(t) = +\infty$ if $P=t$, 0 otherwise. The
infinite rate characterizes instantaneous actions as if all the mass
is concentrated in the variable; e.g., in interactions based on the
minimal rate law, processes are not consumed. For the same reason, if
we were dealing with concentration rates and the multiplicative law, we
would assign $1$ to $P$.
The remaining rules for hiding
are straightforward:
\[
\frac{P \xrightarrowu{a} \phi}
{P \setminus L \xrightarrowu{a} \phi}
\ a \notin L
\qquad
\frac{P \xrightarrowu{a} \phi}
{P \setminus L \xrightarrowu{\tau} \phi}
\ a \in L
\]
This completes the definition of $\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace$
by structural recursion and hence the WF-GSOS specification of PEPA. It
is easy to check that the induced ULTraS is functional and correspond
to the stochastic system of PEPA processes, that bisimulations on it
are stochastic bisimulations (and vice versa) and that bisimilarity is
a congruence with respect to the process signature.
\subsection{Segala-GSOS}
In \cite{bartels04thesis}, Bartels proposed a GSOS specification
format\footnote{Segala-GSOS specifications yield distributive
laws for Segala systems but
it still is an open problem whether every such distributive law
arises from some Segala-GSOS specification.} for Segala systems
(hence Segala-GSOS), i.e.~ULTraS where weight functions are exactly
probability distributions. We recall Bartels' definition, with minor
notational differences.%
\begin{definition}[{\cite[§5.3]{bartels04thesis}}
A \emph{GSOS rule for Segala systems} is a rule of the form
\vspace{-1ex}\[
\vspace{-1ex}
\frac{\left\{x_i \xrightarrow{a} \phi^a_{ij}\right\}_{
1 \leq i \leq n,\
a \in A_i,\
1\leq j \leq m^a_i
}
\quad
\left\{x_i \centernot{\xrightarrow{b}}\right\}_{
1 \leq i \leq n,\
b \in B_i
}
\quad
\left\{\phi^a_{ij} \rightarrowh{->,double equal sign distance,-implies} y_k\right\}_{
1 \leq k \leq q
}}
{\mathtt{f}(x_1,\dots,x_n) \xrightarrow{c} w_1\cdot t_1 + \dots + w_m \cdot t_m}
\]
\vspace{-.5ex}where:
\begin{itemize}\itemsep=-1pt
\item
$\mathtt f$ is an $n$-ary symbol from $\Sigma$;
\item
$X = \{x_i\mid 1 \leq i \leq n\}$, $Y = \{y_k\mid 1 \leq k \leq q\}$, and
$V = \{\phi^a_{ij}\mid 1 \leq i \leq n,\ a \in A_i,\ 1\leq j \leq m^a_i\}$
are pairwise distinct \emph{process} and \emph{probability distribution} variables
respectively;
\item
$a,b,c \in A$ are labels and $A_i \cap B_i = \emptyset$ for any
$i \in \{1,\dots ,n\}$;
\item
$t_1,\dots, t_m$ are target terms on variables
$X$, $Y$ and $V$; the latter are associated with colours from
a finite palette to indicate different instances;
\item
$\{w_i \in (0,1] \mid 1 \leq i \leq m\}$ describe a linear composition of the
targets terms i.e.~are weights associated
to the target terms and such that $w_1 + \dots + w_m = 1$.
\end{itemize}
A rule like above is \emph{triggered} by a tuple $\langle
C_1,\dots,C_n\rangle$ of \emph{enabled labels} if, and only if, $A_i \subseteq C_i$ and $B_i \cap C_i = \emptyset$ for each $i \in \{1,\dots ,n\}$. A \emph{GSOS
specification for Segala systems} is a set of rules in the above
format containing finitely many rules for any source symbol $\mathtt f$,
conclusion label $c$ and trigger $\vec C$.
\end{definition}
Segala-GSOS specifications can be easily turned into WF-GSOS ones. The
first two families of premises are translated straightforwardly to the
corresponding ones in our format; the third can be turned into those
of the form $\support{\phi} \ni y$. Targets of transitions describe
finite probability distributions and are evaluated to actual
probability distributions by a fixed interpretation of a form similar
to Definition~\ref{def:wf-gsos-eval}. Some care is needed to
handle copies of probability variables.
In practice, duplicated variables are expressed by adding
``colouring'' operators to $\Theta$; their number is finite
and depends only on the set of rules since multiplicities
are fixed and finite for rules in the above format.
Let $\tilde V$ be the set of ``coloured'' variables from $V$ where the colouring is
used to distinguish duplicated variables (cf. \cite[§5.3]{bartels04thesis}).
Given a substitution $\nu$ from $\tilde V$ to (finite) probability distributions
over $T^\Sigma(X + Y)$, each $t_i$ is interpreted as the probability distribution:
\vspace{-.5ex}\[
\vspace{-.5ex}
\tilde t_i(t) \triangleq
\begin{cases}
\prod^{|\tilde V \cap var(t_i)|}_{k = 1} \nu(\phi_{k})(t_k) &
\text{if $t = t_i[\phi_{k}/t_k]$ for $t_k \in T^\Sigma(X+Y)$} \\
0 & \text{otherwise}
\end{cases}
\]
and each target term $w_1\cdot t_1 + \dots + w_m \cdot t_m$ is
interpreted as the convex combination of $\tilde t_1,\dots,\tilde t_m$.
\subsection{Weighted GSOS}
In \cite{ks2013:w-s-gsos}, Klin and Sassone proposed a GSOS
format\footnote{Weighted GSOS specifications are proved to yield GSOS
distributive laws for Weighted LTSs but it is currently an open
question whether the format is also complete. } for Weighted LTSs
that is parametric in the commutative monoid $\mathfrak W$ and hence called
$\mathfrak W$-GSOS. The format subsumes many known formats for systems
expressible as $WLTS$: for instance, Stochastic GSOS specifications
are in the $\mathbb R^+_0$-GSOS format and GSOS for LTS are in the $\mathbb
B$-GSOS format where $\mathfrak 2 = (\{\texttt{t\!t},\texttt{f\!f}\},\lor,\texttt{f\!f})$.
\begin{definition}[{\cite[Def.~13]{ks2013:w-s-gsos}}]\label{def:wgsos-rule}
A $\mathfrak W$-GSOS rule is an expression of the form:
\vspace{-1ex}\[
\vspace{-1ex}
\frac{\left\{x_i \xrightarrowt{a} w_{ai}\right\}_{
1 \leq i \leq n,\
a \in A_i
}
\quad
\left\{x_{i_k} \xrightarrowu{b_k,u_k} y_{k}\right\}_{
1 \leq k \leq m
}}
{\mathtt{f}(x_1,\dots,x_n) \xrightarrowu{c,\beta(u_1,\dots,u_m)} t}
\]
where:
\vspace{-.5ex}
\begin{itemize}\itemsep=-1pt
\item
$\mathtt f$ is an $n$-ary symbol from $\Sigma$;
\item
$X = \{x_i\mid 1 \leq i \leq n\}$,
$Y = \{y_k\mid 1 \leq k \leq m\}$ and
$\{u_k\mid 1 \leq k \leq m\}$
are pairwise distinct \emph{process} and
\emph{weight} variables;
\item
$\{w_{ai} \in \mathfrak W \mid 1 \leq i \leq n,\ a\in A_i\}$ are \emph{weight constants}
such that $w_{i_k} \neq 0$ for $1 \leq k \leq m$;
\item
$\beta : W^m \to W$ is a multiadditive function on $\mathfrak W$;
\item
$a,b,c \in A$ are labels and $A_i \subseteq A$ for
$1 \leq i \leq n$;
\item
$t$ is a $\Sigma$-term such that $Y \subseteq \mathrm{var}(\mathtt t) \subseteq X \cup Y$;
\end{itemize}
A rule is \emph{triggered} by a $n$-tuple $\vec C$ of \emph{enabled
labels} s.t.~$A_i \subseteq C_i$ and by a family of weights
$\{v_{ai}\mid 1\leq i \leq n,\ a \in A_i\}$ s.t.~$w_{ai} = v_{ai}$. A
$\mathfrak W$-GSOS specification is a set of rules in the above format such
that there are only finitely many rules for the same source symbol,
conclusion label and trigger.
\end{definition}
Each rule describes the weight of $\mathtt t$
in terms of weights assigned to each $y_k$ (i.e.~$u_k$)
occurring in it; if two rules share the same symbol, label, trigger and
target then their contribute for $\mathtt t$ is added.
To turn a $\mathfrak W$-GSOS specification into WF-GSOS ones, the first
step is to make weight function explicit, by means of premises like
\vspace{-2pt}$x_i \xrightarrowu{a} \phi^a_i$ (since WLTS are
functional ULTraS, i.e.~$m^a_i = 1$). Then, each premise $x_i
\xrightarrowt{a} w_{ai}$ is translated into $\totalweight{\phi^a_i}
= w_{ai}$. If $\mathfrak W$ is \emph{positive} (i.e., whenever $a+b=0$
then $a=b=0$) then $W \setminus\{0\}$ is a club and the
translation of a $\mathfrak W$-GSOS into a WF-GSOS is straightforward.
More generally, it suffices to combine rules sharing the same source,
label and trigger into a single WF-GSOS rule with the same source,
label and trigger. Its target is a suitable weight term containing the
functions $\beta$ and targets $\mathtt t$ of the original rules; every
occurrence of variables $y_k$ and $u_k$ is replaced with the
corresponding function variable (i.e.~$\phi^{b_k}_{i_k}$). In
order to deal with multiple copies of the same weight variable, we
wrap each occurrence in a different ``colouring'' operator, like in
the case of Segala-GSOS.
\section{A coalgebraic presentation of ULTraS and WF-GSOS}\label{sec:coalg}
The aim of this section is to prove some important results about
WF-GSOS specifications. We first provide a characterization of
ULTraSs as coalgebras for a specific behavioural functor
(Section~\ref{sec:ultras-as-coalgebras}), and their bisimulations as
\emph{cocongruences}. Then, leveraging this characterization in
Section~\ref{sec:soundness} we apply Turi and Plotkin's bialgebraic
theory \cite{tp97:tmos}, which allows us to define the categorical
notion of \emph{WF-GSOS distributive laws}; these laws describe the
interplay between syntax and behaviour in any GSOS presentation of
ULTraS. We will prove that every WF-GSOS specification yields a
WF-GSOS distributive law, i.e., the format is \emph{sound}. As a
consequence, we obtain that the bisimilarities induced by these
specifications are always congruence relations. Finally, in
Section~\ref{sec:completeness} we prove that WF-GSOS specification are
also \emph{complete}: every abstract WF-GSOS distributive law can be
described by means of a WF-GSOS specification.
\subsection{Abstract GSOS}\label{sec:abstract-gsos}
In \cite{tp97:tmos}, Turi and Plotkin detailed an abstract
presentation of well-behaved structural operational
semantics for systems of various kinds. There syntax and
behaviour of transition systems are modelled by algebras
and coalgebras respectively. For instance, an (image-finite)
LTS with labels in $A$ and states in $X$ is seen as
a (successor) function $h : X \to (\p{\!\!f} X)^A$ mapping
each state $x$ to a function yielding, for each label $a$,
the (finite) set of states reachable from $x$
via $a$-labelled transitions i.e.~$\{ y \mid x \xrightarrow{a} y\}$:
\vspace{-0.5ex}
\[
y \in h(x)(a) \iff x \xrightarrow{a} y\text.
\vspace{-0.5ex}
\]
Functions like $h$ are \emph{coalgebras} for the
(finite) \emph{labelled powerset functor} $(\p{\!\!f})^A$
over the category of sets and functions $\cat{Set}$.
In general, state based transition systems can be viewed
as \emph{$B$-coalgebra} i.e.~sets (\emph{carriers})
enriched by functions (\emph{structures}) like $h : X\to BX$
for some suitable covariant functor $B : \cat{Set} \to \cat{Set}$.
The $\cat{Set}$-endofunctor $B$ is often called \emph{behavioural}
since it encodes the computational behaviour characterizing the
given kind of systems. A \emph{morphism} from a $B$-coalgebra $h : X \to BX$ to
$g : Y \to BY$ is a function $f : X \to Y$ such that
the coalgebra structure $h$ on $X$ is consistently
mapped to the coalgebra structure $g$ on $Y$
i.e.~$g\circ f = Bf \circ h$.
Therefore, $B$-coalgebras and their homomorphisms form the category $\coalg{B}$.
Two states $x,y \in X$ are said to be \emph{behaviourally equivalent}
with respect to the coalgebraic structure $h : X \to BX$ if
they are equated by some coalgebraic morphism from $h$. Behavioural
equivalences are generalised to two (or more) systems in the form
of kernel bisimulations \cite{staton11}
i.e.~as the pullbacks of morphisms extending
to a cospan for the $B$-coalgebas structures associated with the
given systems as pictured below.
\vspace{-.5ex}\[\vspace{-.5ex}
\begin{tikzpicture}[auto,font=\small,yscale=1.2,xscale=1.5,
baseline=(current bounding box.center)]
\node (n0) at (0,1) {\(X_1\)};
\node (n1) at (2,1) {\(X_2\)};
\node (n2) at (1,.5) {\(Y\)};
\node (n3) at (0,0) {\(B X_1\)};
\node (n4) at (2,0) {\(B X_2\)};
\node (n5) at (1,-.5) {\(B Y\)};
\node (n6) at (1,1.5) {\(R\)};
\draw[->] (n0) to node [swap] {\(f_1\)} (n2);
\draw[->] (n1) to node [] {\(f_2\)} (n2);
\draw[->] (n0) to node [swap] {\(h_1\)} (n3);
\draw[->] (n1) to node [] {\(h_2\)} (n4);
\draw[->] (n2) to node [swap] {\(g\)} (n5);
\draw[->] (n3) to node [swap] {\(Bf_1\)} (n5);
\draw[->] (n4) to node [] {\(Bf_2\)} (n5);
\draw[->] (n6) to node [swap] {\(p_1\)} (n0);
\draw[->] (n6) to node [] {\(p_2\)} (n1);
\begin{scope}[shift=($(n6)!0.28!(n2)$),scale=0.4]
\draw +(-.5,.25) -- +(0,0) -- +(.5,.25);
\end{scope}
\end{tikzpicture}
\]
If the cospan $f_1,f_2$ is jointly epic,
i.e.~$j\circ f_1 = k \circ f_2 \implies j = k$ for any $j,k : C \to Z$,
(in general if $\{f_i\}$ is an epic sink, hence $\{p_i\}$ is a monic source)
then the set $Y$ is isomorphic to the equivalence classes induced by $R$.
We refer the interested reader to \cite{rutten:universal}
for more information on the coalgebraic approach to process theory.
Dually, process syntax is modelled via algebras for endofunctors.
Every algebraic signature $\Sigma$ defines an endofunctor
$\Sigma X = \coprod_{\mathtt f\in \Sigma}
X^{ar(\mathtt f)}$ on $\cat{Set}$ such that every model for
the signature is an algebra for the functor i.e.~a set $X$ (carrier) together
with a function $g : \Sigma X \to X$ (structure). A morphism
from a $\Sigma$-algebras $g : \Sigma X \to X$ to
$h : \Sigma Y \to Y$ is a function $f : X \to Y$ such that $f \circ g = h \circ \Sigma f$.
The set of $\Sigma$-terms with variables from a set $X$ is denoted
by $T^\Sigma X$ and the set of ground ones admits an obvious
$\Sigma$-algebra $a : \Sigma T^\Sigma\emptyset \to T^\Sigma\emptyset$
which is the \emph{initial $\Sigma$-algebra} in the sense that
for every other $\Sigma$-algebra $g$, there exists a unique morphism
from $a$ to $g$ i.e.~the \emph{inductive extension} of the underlying
function $f : T^\Sigma\emptyset \to X$. The construction $T^\Sigma$
is a functor, moreover, it is the monad freely generated by $\Sigma$.
In \cite{tp97:tmos}, Turi and Plotkin showed that structural
operational specifications for LTSs in the well-known image
finite GSOS format \cite{bloomIM:95} correspond to
natural transformations of the following form:
\[\lambda : \Sigma(\mathrm{Id} \times B) \Longrightarrow BT^\Sigma\text.\]
\looseness=-1
These transformations, hence called \emph{GSOS distributive laws}, contain the information
needed to connect $\Sigma$-algebra and $B$-coalgebra
structures over the same carrier set
and capture the interplay between syntax and dynamics
at the core of the SOS approach.
These structures are called \emph{$\lambda$-bialgebras} and are
formed by a carrier $X$ endowed with a $\Sigma$-algebra $g$ and a $B$-coalgebra
$h$ structure s.t.:
\[\vspace{-.5ex}
\begin{tikzpicture}[auto,font=\small,yscale=1.2,xscale=1.5,
baseline=(current bounding box.center)]
\node (n0) at (0,1) {\(\Sigma X\)};
\node (n1) at (0,0) {\(\Sigma(X \times BX)\)};
\node (n2) at (2,0) {\(BT^\Sigma X\)};
\node (n3) at (1,1) {\(X\)};
\node (n4) at (2,1) {\(BX\)};
\draw[->] (n0) to node [] {\(g\)} (n3);
\draw[->] (n0) to node [swap] {\(\Sigma \langle id_X, h\rangle\)} (n1);
\draw[->] (n1) to node [] {\(\lambda_X\)} (n2);
\draw[->] (n2) to node [swap] {\(Bg^\flat\)} (n4);
\draw[->] (n3) to node [] {\(h\)} (n4);
\end{tikzpicture}
\vspace{-.5ex}\]
where $g^\flat: T^\Sigma X \to X$ is the canonical extension of $g$
by structural recursion.
In particular, every $\lambda$-distributive law gives rise to a $B$-coalgebra structure
over the set of ground $\Sigma$-terms $T^\Sigma\emptyset$ and to a
$\Sigma$-algebra structure on the carrier of the final $B$-coalgebra.
These two structures are part of the initial and final
$\lambda$-bialgebra respectively and therefore, because the unique
morphism from the former to the latter is both a $\Sigma$-algebra
and a $B$-coalgebra morphism, observational
equivalence on the system induced over $T^\Sigma\emptyset$
is a congruence with respect to the syntax $\Sigma$.
\subsection{ULTraSs as coalgebras}\label{sec:ultras-as-coalgebras}
Since ULTraSs alternate non-deterministic steps with quantitative
steps, the corresponding behavioural functor can be obtained by
composing the usual functor $(\p{\!\!f})^A: \cat{Set} \to \cat{Set}$ of non-deterministic labelled
transition systems with the functors capturing the quantitative
computational aspects: $\ff{W}$.
It is easy to see that the action of a set function on a weight function
\eqref{eq:fw-action} preserves identities and composition rendering
$\ff{W}$ an endofunctor over \cat{Set}.
For any $\mathfrak W$ and any $A$, $A$-labelled image-finite $\mathfrak W$-ULTraSs
and their homomorphisms clearly form a category: $\cat{ULTS}_{\mathfrak W,A}$.
Objects and morphisms of this category are in 1-1 correspondence
with $(\p{\!\!f}\ff{W})^A$-coalgebras and their homomorphisms respectively.
\begin{proposition}\label{prop:ultras-as-coalgebras}
$\cat{ULTS}_{\mathfrak W,A} \cong \coalg{(\p{\!\!f}\ff{W})^A}$.
\end{proposition}
\begin{proof}
Any image-finite $\mathfrak W$-ULTraS $(X,A,\rightarrowh{-open triangle 60})$ determines a
coalgebra $(X,h)$ where, for any $x \in X$ and $a \in A$: $h(x)(a)
\triangleq \{\rho \mid x\xrightarrowu{a} \rho\}$. Image-finiteness
guarantees that these sets are finite and that their elements are
finitely supported weight functions from $X$ to the carrier of $\mathfrak
W$. Then, it is easy to check that the correspondence is bijective.
\end{proof}
A similar result holds for the bisimulation given in
Definition~\ref{def:bisim}. Categorically, a relation between $X$ and
$Y$ is a (jointly monic) span $X\leftarrow R \rightarrow Y$. In our
case, this span has to be subject to some conditions, as shown
next.
\begin{proposition}\label{prop:bisim-correspondence}
Let $(X_1,A,\rightarrowh{-open triangle 60}_1)$ and $(X_2,A,\rightarrowh{-open triangle 60}_2)$ be
two image-finite
$\mathfrak W$-ULTraSs; let $(X_1,h_1)$, $(X_2,h_2)$ be the corresponding
coalgebras according Proposition~\ref{prop:ultras-as-coalgebras}.
A relation between $X_1$ and $X_2$ is a bisimulation iff there exists a coalgebra $(Y,g)$ and
two coalgebra morphisms $f_1:(X_1,h_1)\to (Y,g)$ and
$f_2:(X_2,h_2)\to (Y,g)$ such that $f_1,f_2$ are jointly epic and
$R$ is their pullback, i.e.~the
diagram below commutes.
\[\begin{tikzpicture}[auto,font=\footnotesize,yscale=1.5,xscale=1.96,
baseline=(current bounding box.center)]
\node (n0) at (0,1) {\(X_1\)};
\node (n1) at (2,1) {\(X_2\)};
\node (n2) at (1,.5) {\(Y\)};
\node (n3) at (0,0) {\((\p{\!\!f}\ff{W} X_1)^A\)};
\node (n4) at (2,0) {\((\p{\!\!f}\ff{W} X_2)^A\)};
\node (n5) at (1,-.5) {\((\p{\!\!f}\ff{W} Y)^A\)};
\node (n6) at (1,1.5) {\(R\)};
\draw[->] (n0) to node [swap] {\(f_1\)} (n2);
\draw[->] (n1) to node [] {\(f_2\)} (n2);
\draw[->] (n0) to node [swap] {\(h_1\)} (n3);
\draw[->] (n1) to node [] {\(h_2\)} (n4);
\draw[->] (n2) to node [swap] {\(g\)} (n5);
\draw[->] (n3) to node [swap] {\((\p{\!\!f}\ff{W} f_1)^A\)} (n5);
\draw[->] (n4) to node [] {\((\p{\!\!f}\ff{W} f_2)^A\)} (n5);
\draw[->] (n6) to node [swap] {\(p_1\)} (n0);
\draw[->] (n6) to node [] {\(p_2\)} (n1);
\begin{scope}[shift=($(n6)!0.28!(n2)$),scale=0.4]
\draw +(-.5,.25) -- +(0,0) -- +(.5,.25);
\end{scope}
\end{tikzpicture}\]
\end{proposition}
\begin{proofatend}
Let $\mathfrak W = (W,+,0)$ be a commutative monoid and let $(X,A,\rightarrowh{-open triangle 60}_X)$,
$(Y,A,\rightarrowh{-open triangle 60}_Y)$, $(X,\alpha)$ and $(Y,\beta)$ be two ULTraS
over $\mathfrak W$ and their corresponding coalgebras
(Proposition~\ref{prop:ultras-as-coalgebras}).
Recall that a function $f : X \to Y$ is a also
coalgebra morphism $f : \alpha \to \beta$ iff,
for each $x \in X$, and $a \in A$:
\[ f(x) \xrightarrowu{a}_Y \psi \iff x \xrightarrowu{a}_X \phi
\land \psi = \phi[f] \]
where $\phi[f]$ denotes the action of $f$ on $\phi$ (i.e.~the
function $\lambda y : Y . \sum_{x\in f^{-1}(y)}\phi(x)$)
and function equality is defined point-wise as usual.
Firstly, we prove that if $R$ is a kernel relation
of some jointly epic cospan of coalgebra mophism from $\alpha$ and $\beta$
then it is a bisimulation.
Let the aforementioned cospan be
$(X,\alpha) \xleftarrow{f} (Z,\gamma) \xrightarrow{g} (Y,\beta)$,
$(Z,A,\rightarrowh{-open triangle 60}_Z)$ the ULTraS for $\gamma$
and assume $x$ and $y$ such that $f(x) = g(y)$.
By definition of coalgebra morphism, $f(x) = z$ implies:
\[x \xrightarrowu{a}_X \phi \iff
z \xrightarrowu{a}_Z \rho = \phi[f] = \lambda c:Z\sum_{x\in f^{-1}(c)}\phi(x)
\text.\]
Likewise $g(y) = z$ implies:
\[
y \xrightarrowu{a}_Y \psi \iff
z \xrightarrowu{a}_Z \rho = \psi[g] = \lambda c:Z\sum_{y\in g^{-1}(c)}\psi(y)
\text.\]
Therefore $f(x) = g(y)$ implies:
\begin{gather*}
x \xrightarrowu{a}_X \phi \implies y \xrightarrowu{a}_Y \psi \land
\forall C \in Z.Z\sum_{x\in f^{-1}(C)}\phi(x) = Z\sum_{y\in g^{-1}(C)}\psi(y)\\
y \xrightarrowu{a}_Y \psi \implies x \xrightarrowu{a}_X \phi \land
\forall C \in Z.Z\sum_{x\in f^{-1}(C)}\phi(x) = Z\sum_{y\in g^{-1}(C)}\psi(y)
\end{gather*}
Then, we conclude by noting that if $R$ is the kernel of $f,g$
there is a bijective correspondence between its equivalence classes
and elements in $Z$ since every class is in the image of $f$ or $g$
by the jointly epic assumption.
For the converse, given a bisimulation $R$ for $(X,A,\rightarrowh{-open triangle 60}_X)$
$(Y,A,\rightarrowh{-open triangle 60}_Y)$ let $Z$ be the set of the equivalence classes in $R$ and
consider the ULTraS $(Z,A,\rightarrowh{-open triangle 60}_Z)$ defined as follows:
\begin{gather*}
C\xrightarrowu{a}_Z \lambda D:Z.\sum_{x' \in D} \phi(x') \iff x\xrightarrow{a}_X \phi \land x \in C \\
C\xrightarrowu{a}_Z \lambda D:Z.\sum_{y' \in D} \psi(y') \iff y\xrightarrow{a}_Y \psi\land y \in C
\end{gather*}
The two statements are redundant since $x,y \in C \iff x R y$ and hence
iff for every $x\xrightarrow{a}_X\phi$ there is
$y \xrightarrow{a}_Y \psi$ s.t.~$\phi \equiv_R\psi$ and vice versa.
Finally, class membership defines a jointly epic coalgebra cospan
from the coalgebras associated to $(X,A,\rightarrowh{-open triangle 60}_X)$ and
$(Y,A,\rightarrowh{-open triangle 60}_Y)$ to the one associated to $(Z,A,\rightarrowh{-open triangle 60}_Z)$
by simply mapping each $x \in X$ and each $y \in Y$ to its class.
\end{proofatend}
Intuitively, the system $(Y,g)$ ``subsumes'' both $(X_1,h_1)$ and
$(X_2,h_2)$ via $f_1$, $f_2$; then, $R$ relates the states which are
mapped to the same behaviour in $Y$ ($(x_1,x_2)\in R$ iff
$f_1(x_1)=f_2(x_2)$).
\paragraph{Coalgebraic bisimulation} In Concurrency Theory also
Aczel-Medler's \emph{coalgebraic bisimulation} \cite{am89:final} is
widely used.
In fact, it is known that kernel bisimulations and coalgebraic
bisimulations coincide if the behavioural functor is \emph{weak
pullback preserving} (wpp). This is the case for many behavioural
functors, but not for $\ff{W}$ in general
\cite{ks2013:w-s-gsos}. Actually, the fact that $\ff{W}$ (and
$(\p{\!\!f}\ff{W})^A$) preserves weak pullbacks depends on the underlying monoid only.
\begin{definition}
A commutative monoid is called \emph{positive} (sometimes zerosumfree, positively ordered) whenever $x + y = 0 \implies x = y = 0$ holds true.
It is called \emph{refinement} if for each $r_1 + r_2 = c_1 + c_2$
there is a $2\times 2$ matrix $(m_{i,j})$ s.t.~
$r_i = m_{i,1} + m_{i,2}$ and $c_j = m_{1,j} + m_{2,j}$.
\end{definition}
\begin{lemma}
\label{lem:wpp}
Coalgebraic bisimulation and behavioural equivalence on ULTraSs
coincides if $\mathfrak W$ is a positive refinement monoid.
\end{lemma}
\begin{proof}
$(\p{\!\!f})^A$ is wpp, and under the lemma hypothesis also $\ff{W}$ is wpp,
by \cite{gummS:01monlbl}. Therefore, $(\p{\!\!f}\ff{W})^A$ is wpp, hence
every behavioural equivalence is a coalgebraic bisimulation on
$(\p{\!\!f}\ff{W})^A$-coalgebras. We conclude by
Proposition~\ref{prop:ultras-as-coalgebras}.
\end{proof}
This condition, can be easily verified and in fact holds for several
monoids of interest, e.g.: $(\{\texttt{t\!t},\texttt{f\!f}\},\lor,\texttt{t\!t})$, $(\mathbb N,+,0)$,
$(\mathbb R^+_0,+,0)$, $(\mathbb N,\max,0)$, and $(A^*,\cdot,\varepsilon)$.
A simple counter example is $(\{0,a,b,1\}, +, 0)$ where
$x + y \triangleq 1$ whenever $x \neq 0 \neq y$
for it is positive but not refinement
(cf.~$\vec r = \langle a, a\rangle$ and $\vec c = \langle b, b\rangle$).
\vspace{-1ex}
\subsection{WF-GSOS specifications are WF-GSOS distributive laws}
\label{sec:soundness}\label{sec:cong-proof}
In this subsection we put the WF-GSOS format within the bialgebraic
framework \cite{tp97:tmos}. As a consequence, we obtain that the
bisimilarity induced by the ULTraS defined by this specification is a
congruence.
In particular, we prove that every WF-GSOS specification represents a
distributive law of the signature over the ULTRaS
behavioural functor, i.e., a natural transformation of the form
\vspace{-.5ex}\begin{equation}\label{eq:wf-gsos-nat}\vspace{-.5ex}
\lambda :
\Sigma(\mathrm{Id} \times (\p{\!\!f}\ff{W})^A)
\Longrightarrow
(\p{\!\!f}\ff{W} T^\Sigma)^A
\end{equation}
where $A$ is the set of labels, $\mathfrak W$ is the commutative monoid
of weights,
$\Sigma = \coprod_{\mathtt f\in \Sigma}\mathrm{Id}^{\mathrm{ar}(\mathtt f)}$ is
the syntactic endofunctor induced by the process signature $\Sigma$,
and $T^\Sigma$ is the free monad for $\Sigma$. We will call natural
transformations of this type \emph{WF-GSOS distributive laws}.
Before stating the soundness theorem, we note that every natural
transformation $\lambda$ as above induces a
$(\p{\!\!f}\ff{W})^A$-coalgebra structure over ground
$\Sigma$-terms. Namely, this is the only function $h_\lambda :
T^\Sigma\emptyset \to (\p{\!\!f}\ff{W}(T^\Sigma\emptyset))^A$ such that:\vspace{-.5ex}
\begin{equation}\label{eq:induced-coalg}\vspace{-.5ex}
h_\lambda \circ a =
(\p{\!\!f}\ff{W} T^\Sigma(a^\#))^A\circ\lambda_X \circ \Sigma\langle id,h_\lambda\rangle
\end{equation}
where $a^\# : T^\Sigma T^\Sigma\emptyset \to T^\Sigma\emptyset$ is the
inductive extension of $a$.
We can now provide the soundness result for WF-GSOS specifications
with respect to WF-GSOS distributive laws, and between
systems and coalgebras they induce over ground $\Sigma$-terms.
\begin{theorem}[Soundness]\label{thm:wf-gsos-soundness}
A specification $\langle\mathcal R,\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace\rangle$
yields a natural transformation $\lambda$ as in \eqref{eq:wf-gsos-nat}
such that $h_\lambda$ and the ULTraS induced by $\langle\mathcal R,
\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace\rangle$ coincide.\par
\end{theorem}
\begin{proof}
For any set $X$, define the function $\lambda_X$ as the composite:
\vspace{-.5ex}\[\vspace{-.5ex}
\Sigma(X \times (\p{\!\!f}\ff{W} X)^A)
\xrightarrow{\llbracket\mathcal R\rrbracket_X}
(\p{\!\!f} T^\Theta(X + \ff{W} X))^A
\xrightarrow{(\mu\circ\p{\!\!f}\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace_X)^A\!\!}
(\p{\!\!f}\ff{W} T^\Sigma X)^A
\]
where $\mu : \p{\!\!f}\pf \Rightarrow \p{\!\!f}$ and $\llbracket\mathcal R\rrbracket_X$ is defined as follows: for all
$\psi' \in T^\Theta(X + \ff{W} X)$, $\mathtt f \in \Sigma$, $c \in A$,
trigger $\vec A = \langle A_1,\dots A_n\rangle$,
$\vec w = \langle w_1,\dots w_p\rangle$,
$y'_k\in X$ and
$\Phi_i(a) = \{\phi^a_{ij} \in \ff{W} X \mid 1\leq j \leq m^a_i\}$ for
$n = \mathrm{ar}(\mathtt f)$ and $i \in \{1,\dots,n\}$,
let
\vspace{-.5ex}\[\vspace{-.5ex}
\psi' \in \llbracket\mathcal R\rrbracket_X
(\mathtt f((x'_1,\Phi_1),\dots,(x'_n,\Phi_n)))\]
if, and only if, there exists in $\mathcal R$ a (possibly renamed) rule
\vspace{-1ex}\[\vspace{-1ex}\frac{
\begin{array}{c}
\Big\{
x_i \xrightarrowu{a} \phi^a_{ij}
\Big\}
\hspace{-1.2ex}\begin{array}{l}
\scriptstyle 1 \leq i \leq n,\\[-4pt]
\scriptstyle a \in A_i,\\[-4pt]
\scriptstyle 1\leq j \leq m^a_i
\end{array}
\quad
\Big\{
x_i \centernot{\xrightarrowu{b}}
\Big\}
\hspace{-1.2ex}\begin{array}{l}
\scriptstyle \\[-4pt]
\scriptstyle 1 \leq i \leq n,\\[-4pt]
\scriptstyle b \in B_i
\end{array}
\quad
\Big\{
\totalweight{\phi^{a_k}_{i_kj_k}} = w_k
\Big\}
_{1 \leq k \leq p}
\quad
\Big\{
\corestr{\phi^{a_k}_{i_kj_k}}{\mathfrak C_k}\ni y_k
\Big\}
_{1 \leq k \leq q}
\end{array}
}{
\mathtt{f}(x_1,\dots,x_n) \xrightarrowu{c} \psi
}\]
such that $m^a_i \neq 0$ iff $a \in A_i$ and there exists
a substitution $\sigma$ such that $\psi' = \sigma[\psi]$,
$\sigma x_i = x'_i$, $\sigma y_k = y'_k$, $\sigma\phi^a_{ij} = \phi^a_{ij}$,
$\totalweight{\phi^{a_k}_{i_kj_k}} = w_k$ and
$\phi^{a_k}_{i_kj_k}(\sigma y_k) \in \mathfrak C_k$.
Then, naturality can be proved separately for the two components:
the former can be tackled as in \cite[Th.~1.1]{tp97:tmos}
and the latter readily follows from Definition~\ref{def:wf-gsos-eval}.
Correspondence of $h_\lambda$ with the induced ULTraS follows by noting that
the latter is given by structural recursion on $\Sigma$-terms
by applying precisely $\lambda$ as given above
(cf.~\eqref{eq:induced-coalg} and
Definition~\ref{def:induced-ultras}).
\end{proof}
Now, by general results from the bialgebraic framework, every
behavioural equivalence on $h_\lambda$ is also a congruence on
$T^\Sigma\emptyset$. In order to obtain this result we need the
following (simple yet important) property.
\begin{proposition}\label{prop:final-coalg}
The category of $(\p{\!\!f}\ff{W})^A$-coalgebras has a final object.
\end{proposition}
\begin{proof}
By \cite{barr93} every finitary $\cat{Set}$ endofunctor admits a
final coalgebra. By definition $\ff{W}$ is finitary. The thesis follows from $\p{\!\!f} \cong \ff{2}$ and
from finitarity being preserved by functor composition.
\end{proof}
\begin{corollary}[Congruence]
Behavioural equivalence on the coalgebra over $T^\Sigma\emptyset$
induced by $\langle\mathcal R,\lbrace\kern-2.2pt\vert{ {\mbox{\large\bf-}} }\vert\kern-2.2pt\rbrace\rangle$ is a
congruence with respect to the signature $\Sigma$.
\end{corollary}
\begin{proof}
The syntactic endofunctor $\Sigma$ admits an initial algebra and, by
Proposition~\ref{prop:final-coalg}, the behavioural endofunctor
$(\p{\!\!f}\ff{W})^A$ admits a final coalgebra. The same holds for
their free monad and cofree copointed functor respectively.
The specification $\langle\mathcal R,\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace\rangle$
defines, by Theorem~\ref{thm:wf-gsos-soundness}, a distributive law
which uniquely extends to a distributive law distributing the
free monad over the cofree copointed functor; then the thesis
follows from \cite[Cor.~7.3]{tp97:tmos}.
\end{proof}
\vspace{-1ex}
\subsection{WF-GSOS distributive laws are WF-GSOS specifications}\label{sec:completeness}
In this subsection we give the important result that the WF-GSOS
format is also \emph{complete} with respect to distributive
laws of the form \eqref{eq:wf-gsos-nat}.
\begin{theorem}[Completeness]\label{thm:wf-gsos-completeness}
Every WF-GSOS distributive law $\lambda$ arises from some WF-GSOS
specification $\langle\mathcal R,\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace\rangle$.
\end{theorem}
The proof of this Theorem follows the methodology introduced by
Bartels for proving adequacy of Bloom's GSOS specification format
\cite[§3.3.1]{bartels04thesis}. The (rather technical) proof will
take the rest of this subsection, so for sake of conciseness we omit
to recall some results which can be found in \emph{loc.~cit.}.
The thesis follows from proving that, for every $\lambda$,
there exists an image-finite set of WF-SOS rules $\mathcal R$
(and suitable interpretations $\theta$ and $\xi$)
making the diagram in Figure~\ref{fig:diag-comp} commute.
\begin{figure}
\centering
\begin{tikzpicture}[auto,xscale=2.4,yscale=1.2, font=\footnotesize,
baseline=(current bounding box.center)]
\node (n0) at (0,3) {$\Sigma(\mathrm{Id} \times (\p{\!\!f}\ff{W})^A)$};
\node (n1) at (4,3) {$(\p{\!\!f}\ff{W} T^\Sigma)^A$};
\node (n2) at (0,0) {$(\p{\!\!f} T^\Xi(\mathrm{Id} + \ff{W}))^A$};
\node (n3) at (4,0) {$(\p{\!\!f}\ff{W} T^\Sigma)^A$};
\node (n4) at (1.2,2) {$(\p{\!\!f} T^\Theta(\mathrm{Id} + \ff{W}))^A$};
\node (n5) at (2.8,2) {$(\p{\!\!f}^2\ff{W} T^\Sigma)^A$};
\node (n6) at (1.2,1) {$(\p{\!\!f}^2 T^\Theta(\mathrm{Id} + \ff{W}))^A$};
\node (n7) at (2.8,1) {$(\p{\!\!f}^3\ff{W} T^\Sigma)^A$};
\draw[->] (n0) to node {$\lambda$} (n1);
\draw[->] (n0) to node[swap] {$\llbracket\mathcal R\rrbracket$} (n2);
\draw[->] (n0) to node {$\rho$} (n4);
\draw[->] (n2) to node[swap] {$(\p{\!\!f}\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace)^A$} (n3);
\draw[->] (n2) to node[swap] {$(\p{\!\!f}\xi)^A$} (n6);
\draw[->] (n3) to node[] {$(\mu_{\scriptscriptstyle \ff{W} T^\Sigma})^A$} (n1);
\draw[->] (n4) to node {$(\p{\!\!f}\theta)^A$} (n5);
\draw[->] (n5) to node[pos=.2] {$(\mu_{\scriptscriptstyle \ff{W} T^\Sigma})^A$} (n1);
\draw[->] (n6) to node[] {$(\mu_{\scriptscriptstyle T^\Theta(\mathrm{Id} + \ff{W})})^A$} (n4);
\draw[->] (n6) to node[swap] {$(\p{\!\!f}^2\theta)^A$} (n7);
\draw[->] (n7) to node[swap] {$(\p{\!\!f}\mu_{\scriptscriptstyle \ff{W} T^\Sigma})^A$} (n3);
\draw[->] (n7) to node {$(\p{\!\!f}\mu_{\scriptscriptstyle \ff{W} T^\Sigma})^A$} (n5);
\node[red, font=\small] at (2,2.6) {(Lem.~\ref{lem:wf-gsos-factorization})};
\node[red, font=\small] at (2,.3) {(Def.~$\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace$)};
\node[red, font=\small] at (3.1,1.5) {($=$)};
\node[red, font=\small] at (1.7,1.5) {(Nat.)};
\end{tikzpicture}
\caption{Factorization for $\lambda$-distributive laws as WF-GSOS specifications.}
\label{fig:diag-comp}
\end{figure}
The lower part of the diagram defines the interpretation
$\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace$ out of $\xi$ and $\theta$ completing
the WF-GSOS specification for $\lambda$.
The middle and right parts of the diagram trivially commute.
The upper part of the diagram commutes because of the following
lemma which states that every WF-GSOS distributive law
arises from an interpretation and a natural transformation
having the same type of those defined by image-finite
sets of WF-GSOS rules.
\begin{lemma}\label{lem:wf-gsos-factorization}
Let $\Sigma$, $A$ and $\mathfrak W$ be a signature, a set of labels and
a commutative monoid, respectively.
Let $\lambda$ be a WF-GSOS distributive law as in \eqref{eq:wf-gsos-nat}.
There exist $\Theta$ and an interpretation
factorizing $\lambda$ i.e.~there exists
$\rho : \Sigma(\mathrm{Id} \times (\p{\!\!f}\ff{W})^A)\Rightarrow (\p{\!\!f} T^\Theta(\mathrm{Id} + \ff{W}))^A$
such that $\lambda = (\mu \circ \p{\!\!f} \theta )^A \circ \rho$.
\end{lemma}
\begin{proof}[Proof (sketch)]
In $\cat{Set}$ it is easy to encode finitely supported functions
as terms. For instance let $\Theta$ extend $\Sigma$ with
operators for describing collections and weight assignments (e.g.~$( {\mbox{\large\bf-}} \mapsto w)$ where $w \in \mathfrak W \setminus \{0\}$).
Then, we can turn $\lambda$ into $\rho$ by simply encoding its
codomain. Then $\theta$ simply evaluates these terms
back to weight functions everything else to the $\emptyset$.
\end{proof}
Following Bartels' methodology, the left part of the diagram commutes
by reducing $\rho$
to simpler, but equivalent, families of natural transformations and
eventually deriving a syntactical specification which is then shown to
be equivalent to an image-finite set of WF-GSOS rules and an
intermediate interpretation $\xi$.
The use of another signature $\Xi$
besides $\Theta$
gives us an extra degree of freedom and simplifies the proof. In
particular, it allows us to encode natural transformations of type
$\ff{W}
\Rightarrow \p{\!\!f}\ff{W}$ (yielded by the aforementioned reduction) in
$\xi$
and handle them downstream to the interpretation
$\lbrace\kern-2.2pt\vert {\mbox{\large\bf-}} \vert\kern-2.2pt\rbrace$.
This expressiveness gain is one of the reasons for the introduction of
non-determinism in Definition~\ref{def:wf-gsos-eval}.
First, note that, by \cite[Lem.~A.1.1]{bartels04thesis}, $\rho$ as above is equivalent to:
\vspace{-.5ex}\[\vspace{-.5ex}
\textstyle
\bar\rho:\Sigma(\mathrm{Id} \times (\p{\!\!f}\ff{W})^A)\times A
\Longrightarrow
\p{\!\!f} T^\Theta(\mathrm{Id} +\ff{W})
\]
which is equivalent to a family of natural transformations
\vspace{-.5ex}\begin{equation}\label{eq:alpha-nat}\vspace{-.5ex}
\textstyle
\alpha_{\mathtt f,c} : (\mathrm{Id} \times (\p{\!\!f}\ff{W})^A)^N
\Longrightarrow
\p{\!\!f} T^\Theta(\mathrm{Id} + \ff{W})
\end{equation}
indexed by $\mathtt f \in \Sigma$ and $c \in A$ and
where $N = \{1,\dots,\mathrm{ar}(\mathtt f)\}$.
In fact, $\Sigma$ is a polynomial functor and $\mathrm{Id} \times A \cong A \cdot \mathrm{Id}$
is an $|A|$-fold coproduct.
By \cite[Lem.~A.1.7]{bartels04thesis}, each $\alpha_{\mathtt f,c}$ is equivalent to a natural transformation
\vspace{-.5ex}\begin{equation}\label{eq:alpha-bar-nat}\vspace{-.5ex}
\textstyle
\bar\alpha_{\mathtt f,c} :
(\p{\!\!f}\ff{W})^{A\times N}
\Longrightarrow \p{\!\!f} T^\Theta(N + \mathrm{Id} + \ff{W})
\end{equation}
and, by the natural isomorphism
\vspace{-.5ex}\[\vspace{-.5ex}
\textstyle
(\p{\!\!f})^{A\times N} \cong (\p{\!\!f}^+ + 1)^{A\times N} \cong
\coprod_{E \subseteq A \times N} (\p{\!\!f}^+)^E
\]
each $\bar \alpha_{\mathtt f,c}$ is equivalent to a family of
natural transformations
\vspace{-.5ex}\begin{equation}\label{eq:beta-nat}\vspace{-.5ex}
\textstyle
\beta_{\mathtt f,c,E} : (\p{\!\!f}^+\ff{W})^{E}
\Longrightarrow
\p{\!\!f} T^\Theta(N + \mathrm{Id} + \ff{W})
\end{equation}
where the added index corresponds to the
vector of sets of labels $\langle E_1,\dots,E_{\mathrm{ar}(\mathtt f)}\rangle$
composing the trigger of a WF-GSOS rule.
By the natural isomorphism
\vspace{-.5ex}\[\vspace{-.5ex}
\textstyle
\p{\!\!f}^+\ff{W} \cong \p{\!\!f}^+\coprod_{v \in \mathfrak W}\ff{W}^v \cong
\coprod_{V \in \p{\!\!f}^+\mathfrak W}
\prod_{v \in V}\p{\!\!f}^+\ff{W}^v
\]
where $\ff{W}^v X \triangleq \{ \phi \in \ff{W} X \mid \totalweight{\phi} = v\}$,
each $\beta_{\mathtt f,c,E}$ is equivalent to a family of
natural transformations
\vspace{-.5ex}\begin{equation}\label{eq:gamma-nat}\vspace{-.5ex}
\textstyle
\gamma_{\mathtt f,c,E,w} :
\coprod_{e \in E}\prod_{v \in w(e)}\p{\!\!f}^+\ff{W}^v
\Longrightarrow
\p{\!\!f} T^\Theta(N + \mathrm{Id} + \ff{W})
\end{equation}
where $w : E \to \p{\!\!f}^+\mathfrak W$.
Since total weight premises associate pairs from $E$
to weights, maps like $w$ can be seen as families
of triggering weights.
By \cite[Lem.~A.1.3]{bartels04thesis} and by the natural isomorphism
\vspace{-.5ex}\[\vspace{-.5ex}
\textstyle
T^\Theta \cong \coprod_{\psi \in T^\Theta 1} \mathrm{Id}^{|\psi|_*}
\]
where $|\psi|_*$ denotes the number of occurrences of $* \in 1$
in the $\Theta$-term $\psi$ (cf.~\cite[Lem.~A.1.5]{bartels04thesis})
each $\gamma_{\mathtt f,c,E,w}$ corresponds to a family
of natural transformations
\vspace{-.5ex}\begin{equation}\label{eq:delta-nat}\vspace{-.5ex}
\textstyle
\delta_{\mathtt f,c,E,w,\psi} :
\coprod_{e \in E}\prod_{v \in w(e)}\p{\!\!f}^+\ff{W}^v
\Longrightarrow
\p{\!\!f}^+((\mathrm{Id} + \ff{W})^{|\psi|_*})
\end{equation}
where the added index $\psi$ ranges over some subset of
$T^\Theta(1+N)$ (cf.~target terms of WF-GSOS rules).
Then, following \cite[§3.3.1, Cor.~A.2.8]{bartels04thesis} it is
easy to check that each $\delta_{\mathtt f,c,E,w,\psi}$ describes a
non-empty, finite set of derivation rules as
\vspace{-.5ex}\[\vspace{-.5ex}
\frac{
\phi_{j,{v_j}} \in \pi_{v_j}(\Phi_{e_j}) \quad y_i \in \epsilon_{j,{v_j}}(\phi_{j,{v_j}})
}{
\langle z_{1},\dots,z_{|\psi|_*}\rangle \in
\delta_{\mathtt f,c,E,w,\psi}((\Phi_e)_{e \in E})
}
\]
where $p,q\in \mathbb N$, $e_j \in E$, $1\leq j \leq p$,
$1\leq i\leq q$, $v_j \in w(e_j)$, each $z_k \in \{y_i \mid 1\leq i\leq q \}$
for $1 \leq k \leq |\psi|_*$ and each $\epsilon_{j,v_j}$ is a natural transformation:
\vspace{-.5ex}\[\vspace{-.5ex}\epsilon_{j,{v_j}}:\ff{W}^{v_j} \Longrightarrow \p{\!\!f}^+(\mathrm{Id} + \ff{W})\text{.}\]
Natural transformations of this type can be easily encoded in the
term $\psi$ by suitable extensions of $\Theta$ and therefore
each $\delta_{\mathtt f,c,E,w,\psi}$ can be shown to be equivalent
to a $\delta$-specification i.e.~a non-empty, finite set of derivation rules as above
except for each $z_k$ being a term wrapping $\phi_{j,v}$ with
the symbol denoting $\epsilon_{j,v}$. These terms are then evaluated
by the interpretation $\xi^\delta$ as expected.
This proof points out the trade-off that has to be made in presence of
specifications with interpretation such as WF-GSOS or MGSOS
\cite{bm:2015stocsos}. In fact, clubs were not mentioned in the above
reduction of $\rho$ since each $\epsilon_{j,v_j}$ was handled by the
interpretation $\xi$. However, the following result shows that clubs
(hence, premises like $\corestr{\phi}{\mathfrak C}\ni y$), characterize
natural transformations of type $\ff{W}^v \Rightarrow \p{\!\!f}$.
\begin{lemma}\label{lem:clubs-nat}
For any natural transformation $\upsilon : \ff{W}^w \Rightarrow \p{\!\!f}$
there exists a club $\mathfrak C_\upsilon$ characterizing it:
$x \in \upsilon_X(\phi) \iff \phi(x) \in \mathfrak C_\upsilon$.
\end{lemma}
\begin{proof}[Proof (sketch)]
Intuitively, natural transformations of this type are ``selecting
a finite subset from each weight function domain'' and it is
easy to check that elements can be only singled out by their
weight. Likewise, finiteness and naturality prevent the selection of
anything outside function supports. Then, the problem readily
translates into finding the finest topology on the weight monoid
that ``plays well'' with $\ff{W}$ i.e.~such that monoidal addition,
seen as a continuous map from the product topology,
preserves opens (i.e.~any admissible selection). Clubs are a base for this
topology since, by definition, these are the only substructures
isolated w.r.t.~$\ff{W}$-action. Hence selections made by $\upsilon$
are completely characterized by a single club $\mathfrak C_\upsilon$.
\end{proof}
Finally, we have to translate the set of rules we got so far into the
WF-GSOS format; we do it by reversing the chain that led us from
$\rho$ to $\delta$ and $\delta$-specification. By
Lemma~\ref{lem:clubs-nat} every $\delta$-specification
is equivalent to a $\gamma$-specification
\vspace{-.5ex}\[\vspace{-.5ex}
\left\{
\frac{
\phi_{j} \in \Phi_{e_j} \quad
\totalweight{\phi_j} = v_j \quad
\corestr{\phi_j}{\mathfrak C_i} \ni y_i
}{
\psi(z_1,\dots,z_{|\psi|_*}) \in
\gamma_{\mathtt f,c,E,w}((\Phi_e)_{e \in E})
}\,\middle|
\begin{array}{l}\scriptstyle
v_j \in w(e_j),\\\scriptstyle
z_k \in \{\phi_j[\zeta_j], y_i\}+N,\\\scriptstyle
\psi
\end{array}\!\!\!
\right\}_\text{fininte}
\]
where $\phi_j[\zeta_j]$ is a term build with the $\Xi$-operator
denoting the natural transformation
$\zeta_j : \ff{W}^{v_j} \Rightarrow \p{\!\!f}\ff{W}$ and $\xi^\gamma$ acts as
$\xi^\delta$ on these terms, as the identity on those generated from $\Theta$
(distributing the powerset as expected) and maps everything else to $\emptyset$.
A $\gamma$-specification defines a natural transformation as in \eqref{eq:gamma-nat}
and every family of $\gamma$-specifications characterizing
a natural transformation as in \eqref{eq:beta-nat} is equivalent to
a $\beta$-specification i.e.~a set of derivation rules
\vspace{-.5ex}\[\vspace{-.5ex}
\left\{
\frac{
\phi_{j} \in \Phi_{e_j} \quad
\totalweight{\phi_j} = v_j \quad
\corestr{\phi_j}{\mathfrak C_i} \ni y_i
}{
\psi(z_1,\dots,z_{|\psi|_*}) \in
\beta_{\mathtt f,c,E}((\Phi_e)_{e \in E})
}\,\middle|
\begin{array}{l}\scriptstyle
z_k \in \{\phi_j[\zeta_j], y_i\}+N,\\\scriptstyle
\psi
\end{array}\!\!\!
\right\}_\text{image finite}
\]
finite up to vectors of
total weights $\vec v = \langle v_0,\dots,v_p\rangle$.
Since $E \subseteq A \times N$, every family of
$\beta$-specifications describing a natural transformation
as in \eqref{eq:alpha-bar-nat} is equivalent to a set
\vspace{-.5ex}\[\vspace{-.5ex}
\left\{
\frac{
\Phi_{m_n,b_n} = \emptyset \quad
\phi_{j} \in \Phi_{l_j,a_j} \quad
\totalweight{\phi_j} = v_j \quad
\corestr{\phi_j}{\mathfrak C_i} \ni y_i
}{
\psi(z_1,\dots,z_{|\psi|_*}) \in
\bar\alpha_{\mathtt f,c}(\langle \Phi_1,\dots,\Phi_{|\mathtt f|}\rangle)
}\,\middle|
\begin{array}{l}\scriptstyle
\langle m_n,b_n\rangle \neq \langle l_j,a_j\rangle,\\\scriptstyle
z_k \in \{\phi_j[\zeta_j], y_i\}+N,\\
\psi
\end{array}\!\!\!
\right\}_\text{im.fin.}
\]
containing finitely many rules for every $E$ and $\vec v$.
This set corresponds to an $\alpha$-specification i.e.~an image-finite
set like the following:
\vspace{-.5ex}\[\vspace{-.5ex}
\left\{
\frac{
\Phi_{m_n}(b_n) = \emptyset \quad
\phi_{j} \in \Phi_{l_j}(a_j) \quad
\totalweight{\phi_j} = v_j \quad
\corestr{\phi_j}{\mathfrak C_i} \ni y_i
}{
\psi(z_1,\dots,z_{|\psi|_*}) \in
\alpha_{\mathtt f,c}(\langle \langle x_1,\Phi_1\rangle,\dots,\langle x_{|\mathtt f|},\Phi_{|\mathtt f|}\rangle\rangle)
}\,\middle|
\begin{array}{l}\scriptstyle
\langle m_n,b_n\rangle \neq \langle l_j,a_j\rangle,\\\scriptstyle
z_k \in \{\phi_j[\zeta_j], y_i, x_h\},\\\scriptstyle
\psi
\end{array}\!\!\!
\right\}_\text{im.fin.}
\]
Finally, every family of $\alpha$-specifications equivalent to a natural
transformation as $\rho$ corresponds to an image-finite set of WF-GSOS rules
and an interpretation. Therefore we conclude that for any $\rho$
there exist $\mathcal R$ and $\xi$ as in Figure~\ref{fig:diag-comp}.
\section{Conclusions and future work}\label{sec:concl}
\looseness=-1
In this paper we have presented WF-GSOS, a GSOS-style format for
specifying non-deterministic systems with quantitative aspects. A
WF-GSOS specification is composed by a set of rules for the derivation
of judgements of the form $P \xrightarrowu{a} \psi$, where $\psi$ is a
term of a specific signature, together with an \emph{interpretation} for these
terms as weight functions. We have shown that a specification in this
format defines an ULTraS, and it is expressive enough to subsume other
more specific formats such as Klin's \emph{Weighted GSOS} for WLTS
\cite{ks2013:w-s-gsos}, and Bartel's \emph{Segala-GSOS} for Segala
systems \cite[§5.3]{bartels04thesis}, and those subsumed by them
e.g.~Klin and Sassone's Stochastic GSOS \cite{ks2013:w-s-gsos} and
Bloom's GSOS \cite{bloomIM:95}. WF-GSOS induces naturally a notion of
(strong) bisimulation, which we have compared with
$\mathcal{M}$-bisimulation used in ULTraS. We have also provided a
general categorical presentation of ULTraSs as coalgebras of a precise
class of functors, parametric on the underlying weight structure.
This presentation allows us to define categorically the notion of
\emph{abstract GSOS} for ULTraS, i.e., natural transformations of a
precise type. We have proved that WF-GSOS specification format is
\emph{adequate} (i.e., sound and complete) with respect to this
notion. Taking advantage of Turi-Plotkin's bialgebraic framework, we
have proved that the bisimulation induced by a WF-GSOS is always a
congruence; hence our specifications can be used for compositional and
modular reasoning in quantitative settings (e.g., for ensuring
performance properties). Moreover, the format is at least as
expressive as every GSOS specification format for systems subsumed by
ULTraS.
\paragraph{Related works}
In this paper we have shown that commutative monoids are enough to
define ULTraSs, their homomorphisms and bisimulations. The original
work \cite{denicola13:ultras} assumed weights to be organised into a
partial order with bottom $(W,\leq,\perp)$, but the order plays no
r\^ole in the definition besides distinguishing the point $\perp$
used to express unreachability. A monoidal sum is eventually and
implicitly assumed by the notion of $\mathcal{M}$-bisimulation and,
because of the definition of $M$-function, this operation is assumed
to be be monotone in both its components and to have $\perp$ as
unit. In other words, $\mathcal{M}$-bisimulation implicitly assumes
weights to form a commutative positively ordered monoid
$(W,+,\leq,0)$. Any such a monoid is positive and hence it has a
natural order $a \trianglelefteq b \iff \exists c .\, a+ c = b$; this
order is the weakest one rendering the monoid $W$ positively ordered,
in the sense that for any such ordering $\leq$, it is
${\trianglelefteq} \subseteq {\leq}$.
\looseness=-1
We note that in \cite{denicola13:ultras}, weights used to define
ULTraSs are decoupled from those of $M$-functions; e.g., the formers
can be in $([0,1],\leq,0)$ and the latters in $(\mathbb R^+_0,+,0)$.
However, the notion of constrained ULTraS is sill needed to precisely
capture probabilistic systems or, in other words, the use of partial
orders may still require to embed the systems under study into a
larger class of ULTraSs. We remark that $(\mathbb R^+_0,+,0)$ is the
smaller completion of $([0,1],\leq,0)$ under $+$ and in this sense the
embedding can be seen as canonical. Therefore, defining ULTraSs in
terms of commutative monoids is a conservative generalisation that
additionally provides a natural notion of homomorphisms and hence
bisimulations. As a side note, existence of bottoms does not allow
weights to have opposites, e.g., to model opposite transitions like in
calculi for reversible computations.
Although in this paper we have taken ULTraSs as a reference,
WF-GSOS can be interpreted in other meta-models, such as FuTSs
\cite{latella:qapl2015}. Like ULTraSs, FuTSs have state-to-function
transitions, but admit several distinct domains for weight functions
and more free structure besides the strict alternation between
non-deterministic and quantitative steps. In their more general
form, they can be understood as coalgebras for functors of shape:
\begin{equation}
\label{eq:futs-fun}
F_{\vec{A},\vec{\mathfrak W}} =
(\f{\mathfrak W_{0,k_0}}\dots\ff{\mathfrak W_{1,0}})^{A_0}
\times\dots
(\f{\mathfrak W_{n,k_n}}\dots\ff{\mathfrak W_{1,0}})^{A_n}
\vspace{-1ex}
\end{equation}
where each $\mathfrak W_{i,j}$ in $\vec{\mathfrak W}$ is a commutative monoid and each $A_i$ in $\vec{A}$ is a set.
We remark that, although in \cite{latella:qapl2015} weights are drawn
from semirings, commutative monoids are sufficient to define $\ff{W}$ and hence define FuTS, homomorphisms and eventually bisimulations. Moreover, Lemma~\ref{lem:wpp}
readily generalises to \eqref{eq:futs-fun}: if weights are drawn only
from positive refinement monoids then any such functor is wpp.
No rule format for FuTSs has been published yet;
we believe the WF-GSOS specification format to be a step in this direction
because of the similarities between the behavioural functors involved.
This would allow us to formulate compositionality results for
(meta)calculi defining FuTSs, e.g., the framework for stochastic
calculi proposed in \cite{denicola13:ustoc}. Indeed
since ULTraSs can be viewed as FuTSs
(assuming commutative monoids as a common ground)
any specification format for the latter that is both correct and complete
w.r.t.~the suitable abstract GSOS law will necessarily subsume WF-GSOS.
The systems considered in this paper can be seen as generalised Segala
systems. We showed how the proposed format subsumes Bartels'
Segala-GSOS; however, this is not the only specification format for
this kind of systems. In \cite{gdl2012:treerules} Gebler et
al.~proposed a $nt\mu f\nu/nt\mu x\nu$ rule format for describing
Segala systems. Since Turi-Plotkin seminal paper \cite{tp97:tmos} it
is well known that GSOS and coGSOS (i.e~tree-rule formats such as that
in \cite{gdl2012:treerules}) correspond to distributive laws of
completely different shapes: the former distribute monads over
copointed endofunctors whereas the latter distribute pointed
endofunctors over comonads. These different shapes have obvious
implications on the data available to the derivation rules: monads
provide views ``inside terms'' whereas comonads provide views ``inside
executions''. Their common generalisation are laws distributing
monads over comonads but has limited practical benefits because it
does not translate to any concrete rule format that would be complete
for any specification containing both GSOS and coGSOS
\cite{klin:sos2014}.
\paragraph{Future work}
The categorical characterization of ULTraS systems paves the way for
further interesting lines of research. One is to develop
Hennessy-Milner style modal logics for quantitative systems at the
generality level of the ULTraS framework. In fact, Klin has shown in
\cite{klin09:sosmlogic} that HML and CCS are connected by a
(contravariant) adjunction. A promising direction is to follow this
connection taking advantage of the bialgebraic presentation of ULTraSs
provided in this paper. Another is to explore the implications of the
recent developments in the coalgebraic understanding of unobservable
moves \cite{bmp:arxiv14-unobs,bonchi2015killing} in the context of this work.
An intermediate step in this direction is to develop a suitable monad structure
for $\p{\!\!f}\ff{W}$ which is, in general, not a monad (cf.~$\p{\!\!f}\mathcal{D}$ where $\mathcal{D}$ is the probability distribution monad). This alone will
allow us to define e.g.~trace and testing equivalences in a principled coalgebraic way.
\vspace{-1ex}
\paragraph{Acknowledgements} We thank Rocco De Nicola, Daniel Gebler,
the anonymous reviewers and the QAPL'14 participants for useful discussions on the conference
version of this paper.
This work is partially supported by MIUR PRIN project 2010LHT4KM, \emph{CINA}.
{\footnotesize
|
\section{Introduction}
Diffuse optical tomography (DOT) is an imaging technique that uses near infrared light to image highly scattering media. A good review has been provided in~\cite{arridge1999optical} and an updated version is provided in~\cite{arridge2009optical}. The imaging modality has shown great promise as a low-cost alternative or complement to existing medical imaging technology particularly in brain imaging and breast cancer detection. The region of interest is illuminated with near infrared light over a collection of wavelengths and the data are comprised of observations of the resulting scattered diffuse fields at a number of locations surrounding the medium. Given these measurements as well as the partial differential equation governing the interaction of light and tissue (typically, the diffusion equation), we seek to recover space and time-varying maps (i.e. images) of concentrations of physiologically relevant chromophores such as oxygenated and deoxygenated hemoglobin (HbO$_2$ and HbR respectively), lipid, and water (H$_2$O) as well as properties governing the scattering of light within the medium.
The recovery of images of chromophore concentrations can be mathematically posed as a nonlinear inverse problem. However, due to the diffusive physics associated with this problem as well as limitations concerning the geometric distribution of sources and detectors, image recovery is an ill-posed inverse problem. New technology developed in our research group allows for the collection of hyperspectral data (over 100 bands in the near infrared portion of the spectrum). Although we have demonstrated~\cite{larusson2013parametric,larusson2011hyperspectral,larusson2012parametric} that the availability of more information using multiple wavelengths increases the accuracy of the reconstruction, the use of hyperspectral data poses a significant computational burden in the context of image recovery. We are interested in developing computationally efficient methods for hyperspectral diffuse optical tomography (HyDOT) with specific application towards breast imaging in which the breast is placed in between two parallel plates.
To motivate the need for fast algorithms for hyperspectral DOT, we outline here the expected costs in terms of storage and computation. To make ideas concrete, we consider an experimental setup for detecting tumors in breast tissue (see Figure~\ref{fig:phantom}). We use $N_s$ near-infrared sources to illuminate the medium of interest. The sources are constrained to lie on the top plane and detectors are constrained to be on a different plane so that for a given source, we have $N_{ds}$ detectors measuring photon fluence at $N_\lambda$ wavelengths. This results in $M = N_sN_{ds}N_\lambda$ measurements. We also assume that the domain has been discretized into a grid with $N$ vertices.
\begin{table}[!ht]
\centering
\begin{tabular}{|l|c|c|}\hline
Number of & Symbol & Typical number \\ \hline
Sources & $N_s$ & $10-100$ \\
Detectors / source & $N_{ds}$ & $3-10$ \\
Wavelengths & $N_\lambda$ & $10-200$ \\
Grid size & $N$ & $32^3-100^3$ \\ \hline
\end{tabular}
\caption{Typical range of parameters for the hyperspectral DOT problem}
\label{tab:range}
\end{table}
In Table~\ref{tab:range}, we have listed the range of various parameters that one might encounter in practice. Since the most accurate forward model in terms of the unknown voxel values is nonlinear, standard numerical/optimization approaches to solve the inverse problem repeatedly linearize the forward problem about a current estimate~\cite{arridge1999optical,de2011regularized}. Each optimization step then requires the solution of the forward and the adjoint PDE for each source-detector set~\cite{arridge1999optical}. In all, we need to solve $N_s(N_{ds} + 1)N_\lambda$ systems of equations \textit{at each optimization step}, which amounts to about $2\times 10^5$ systems of equations for the range of parameters described in Table~\ref{tab:range}. Thus, for finely discretized fields, even with a solver of optimal complexity $\bigO(N)$, use of the nonlinear forward model poses a significant computational challenge because the resulting cost is $\bigO(N_s(N_{ds} + 1)N_\lambda N$) flops.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.35]{figs/phantom}
\caption{The breast shaped phantom used as the imaging medium. At its widest, the phantom is $12$ cm long and $5$ cm thick.}
\label{fig:phantom}
\end{figure}
To mitigate the computational burden, in the present work we restrict ourselves to the case where the absorption can be represented as a small perturbation about the background medium. Then, we can use the Born approximation to linearize the forward model. Although this approximation introduces a modeling error and has known limitations~\cite{boas1997fundamental}, studies with experimental data have validated the utility of this model for hyperspectral DOT~\cite{larusson2013parametric,larusson2011hyperspectral,larusson2012parametric}. However, since the domain of interest has a complex shape, the Green's function required for the construction of the linearized operator is not known in closed form. Therefore, we need to use a PDE formulation to compute the incident and adjoint fields. An advantage of the PDE formulation is that we can handle known, non-constant background which may be obtained by imaging using a high resolution modality (such as magnetic resonance imaging, MRI) along with image segmentation~\cite{boverman2005quantitative}. Although the breast is a highly heterogenous medium, based on previous work we model it as a piecewise constant medium with homogenous background. For example, the authors in~\cite{schweiger1999optical} use prior anatomical knowledge to derive a piecewise constant medium. Although the Born approximation reduces the computational burden associated with the imaging problem tremendously, computing the Born operator $\bH$ (which is a linear mapping between the perturbation of interest and the measurements) still requires the solution of many large-scale discretized PDEs for the incident and adjoint field; in sum, about $\bigO(N_s(N_{ds} + 1)N_\lambda)$ large scale, sparse linear systems corresponding to the discrete diffusion equation.
Furthermore, the cost to store $\bH$ and form matrix-vector products (matvecs) with $\bH$ during the optimization that are needed to invert for the desired parameters scales as $\bigO(MN)$, where $M$ is the number of measurements and $N$ is the number of unknowns. Therefore, we develop a recursive algorithm to generate a low-rank approximation to $\bH$ and use this approximation in the optimization. If the numerical rank of the low-rank representation is $R$ then the cost of storing factors and forming matvecs are $\bigO(R(M + N))$.
Of course, the optimal rank-R approximation could in theory be computed from the SVD of $\bH$ if we had $\bH$. This is inefficient on two fronts: a) it requires the full computation and storage of as well as multiple accesses to the very large, dense matrix $\bH$ b) the cost of an SVD on $\bH$, if we had it, is still prohibitively expensive at $\bigO \left( \min \{M,N\}^2, \max\{M,N\} \right)$ flops. On the other hand, our algorithm, because it operates in a spatially recursive manner requiring, at the finest scale, only local information, and using rank revealing factorizations to aggregate information globally, does not require that $\bH$ be fully stored.
We have three main goals in this paper. The first is to design an algorithm to overcome the challenge of computing the solution of $\bigO(10^5)$ number of large-scale parametric linear systems needed to obtain certain information necessary to compute our approximate Born matrix. The second is to design an efficient algorithm to compute and store a meaningful low-rank representation of the measurement operator $\bH$. The third
is to use this low-rank representation for recovering the parameters that define our chromophore image. We now summarize the key components of our fast algorithm and contributions in this paper:
\begin{itemize}
\item For each source-detector set, we need to compute the incident field and the adjoint field for hundreds of wavelengths. We derive a novel Krylov recycling subspace approach to solve the corresponding systems of large-scale parametric linear systems which takes advantage of similarities in the systems across wavelengths. This is described in detail in Section~\ref{sec:krylov}.
\item For the problem at hand, the sources, detectors and the perturbation are well-separated from each other. Under these circumstances there is strong numerical evidence that the Born operator $\bH$ is low-rank, a feature that has been exploited to achieve computational savings~\cite{chaillat2012faims}. In this work, we give a new approach for computing a low rank approximation, and use this in our numerical results on the parametric inverse problem.
The storage of the matrix $\bH$ in its entirety is avoided; instead, the rows of $\bH$ corresponding to a single source-detectors set are constructed, compressed and then recursively compressed across multiple source-detectors pairs to obtain a low-rank factorization $\bH \approx \hat{\bH} = \mathbf{U} \mathbf{V}^T$. We provide a detailed outline of the algorithm and analysis of the computational costs in Section~\ref{sec:compress}. Because we fully compute the entries of $\bH$, we need the fast Krylov solvers because we need to solve $\bigO(10^5)$ systems.
\item Based on recent success in the context of DOT, we employ the parametric level set (PaLS) approach to represent the chromophore image we want to recover. The PaLS approach, developed for inverse problems in~\cite{aghasi2011parametric} and subsequently applied to diffuse optical tomography~\cite{larusson2012parametric,larusson2013parametric}, has the advantage of explicitly describing the geometry of the anomaly.
As a result, the recovery of the chromophore image is obtained by solving a nonlinear least-squares solver on a problem in which we have replaced the Born operator by the approximation described above. We show how to accelerate the reconstruction of the chromophore concentration and shape parameters by using the compressed measurement operator $\hat{\bH}$ is used in place of the full measurement operator. Error bounds are provided based on the error in the low-rank representation and are useful for the optimization routine. This is outlined in Section~\ref{sec:recon}.
\end{itemize}
\textbf{Related work}: We briefly review other approaches to solve the parametric system of equations that represent the propagation of light in diffuse medium. One approach to deal with parametric system of equations is the use of spectral methods by expanding the matrix coefficients in a series of orthogonal polynomials (such as Chebyshev, Legendre, etc.) and solving for the coefficients of the orthogonal polynomials using a Collocation method or a residual minimizing Galerkin projection method (see~\cite{constantine2010spectral} and references therein). The low-rank property of the solutions arising from system of parametric coefficients with coefficients smoothly dependent on the parameters was demonstrated by Kressner and Tobler~\cite{kressner2011low} and they developed global Krylov subspace algorithms that exploited the low-rank nature to minimize computational and storage costs. In both approaches, one has to solve a coupled system of equations. This can be computationally expensive and, in order to ensure rapid convergence, a preconditioner that is effective across all the shifts is necessary but difficult to choose in practice. Other approaches to deal with the expensive cost of solving parametric linear systems is to use parametric model reduction which is reviewed in~\cite{benner2013survey}.
Other works have also considered the compression of the measurement operator $\bH$ that maps the perturbation to the measurements. In~\cite{markel2003inverse}, the authors develop analytical formulas for inversion based on Fourier analysis when the sources and detectors are distributed uniformly on the boundary of a regular geometry such as a plane, cylinder, or sphere. In our previous work, we have exploited the structure of the Green's function in regular geometries to decompose the Born operator into a number of sparse easily computed matrices~\cite{hyde2007analysis}. The approach of compressing the operator $\bH$ is similar to that derived in~\cite{chaillat2012faims}. Here, the authors consider compression across multiple sources and detectors for a given frequency using randomized SVD and then recursively compressing the low-rank factors across multiple frequencies. However, we cannot adopt their framework directly for the following reasons. Firstly, in our system, the detectors are not shared across all of the sources. Consequently, the pre-processing step that compresses the incoming field and the data, as described in~\cite{chaillat2012faims}, cannot be used directly. Secondly, the authors in~\cite{chaillat2012faims} advocate compression of the measurement operator across multiple sources and detectors for a given frequency and then recursively combining the low-rank factorizations across different wavelengths. Memory limitations prevent computations of the entire measurement operator in its entirety. Because of the way our computations are organized, we choose to first compress across multiple wavelengths and detectors for a given source and then combine the low rank factorizations across multiple sources.
\section{Forward Problem}
In this section, we give the specifics of the forward problem and discretization used for the associated PDEs.
\subsection{Governing equations}
The radiative transport physics associated with the propagation of light through a medium can be approximated by the diffusion model of the form in the domain $\Omega$~\cite{arridge1999optical}
\begin{align}\label{eqn:diffusion}
-\nabla \cdot D^\text{tot}(\br,\lambda) \nabla \phi(\br,\lambda) + \nu \mu_a^\text{tot}(\br,\lambda) \phi(\br,\lambda) = & \quad S(\br,\lambda) & \quad \br \in& \Omega \\ \label{eqn:dirichlet}
\phi(\br, \lambda) = & \quad 0 & \quad \br \in& \partial\Omega_D\\ \label{eqn:robin}
\phi(\br,\lambda) + 2AD^\text{tot}(\br,\lambda)\frac{\partial \phi(\br,\lambda)}{\partial n} = & \quad 0 & \quad \br \in& \partial\Omega_R
\end{align}
where $D^\text{tot}(\br,\lambda)$ is the diffusion coefficient and is related to the reduced scattering coefficient $\mu_s'(\lambda)$ as $D(\br,\lambda) = \nu/3\mu_s' (\br,\lambda) $. We also denote by $\partial\Omega_D$ the portion of the boundary over which zero Dirichlet boundary conditions are imposed (curved boundaries and chest wall) and $\partial \Omega_R$ corresponds to the boundary portion over which Robin boundary conditions are imposed corresponding to a refractive index mismatch (top and bottom flat regions). The coefficient $A$ is a function of the refractive index of the medium. We denote by $\phi(\br,\lambda)$ the photon fluence at a position $\br$ due to a source of wavelength $\lambda$ injected into the medium, and $\nu$ is the electromagnetic propagation velocity within the medium. Further, $\mu_a^\text{tot}(\br,\lambda)$ is the absorption coefficient. The quantity $S(\br,\lambda)$ is the photon source with units of optical energy per unit time per unit volume and typically written in terms of a delta function; that is, $S(\br,\lambda) = S_0(\lambda) \delta(\br-\br_s)$, with $S_0(\lambda)$ the source power at wavelength $\lambda$.
We decompose the absorption, $\mu_a^\text{tot}(\br,\lambda)$, into a constant background absorption $\mu_a(\lambda)$ and a spatially varying perturbation $\Delta \mu_a(\br,\lambda)$. The total fluence, $\phi$, is decomposed into an incident field $\phi_i$ and a scattered field $\phi_s$. Likewise, we can expand the diffusion $D^\text{tot}(\br,\lambda)$ into the sum of a homogenous background term $D(\lambda)$ and a perturbation $\Delta D(\br,\lambda)$. However, the spatial dependence of diffusion is minimal in healthy breasts~\cite{shah2004spatial} and spatial contrast in breast tumors is either non-existent or small~\cite{grosenick2004concentration,grosenick2005timea}. To simplify matters we assume that $D^{tot}(\br,\lambda)$ is independent of $\br$ and $\Delta D(\br,\lambda) =0$. Therefore $D^\text{tot}(\br,\lambda)$ can be represented entirely as $D(\lambda)$ and we can then divide throughout by $D(\lambda)$ (see for example~\cite{larusson2011hyperspectral}).
The equation for the incident field $\phi_i$ and the scattered field $\phi_s$ can be obtained by substituting $\phi(\br,\lambda) = \phi_i(\br,\lambda) + \phi_s(\br,\lambda)$ and collecting the appropriate terms, and is therefore,
\begin{align}
\label{eqn:incident}
-\nabla^2 \phi_i(\br,\lambda) + \frac{\nu \mu_a(\lambda)}{D(\lambda)} \phi_i(\br,\lambda) \quad = & \quad \frac{S_0(\lambda)}{D(\lambda)}\delta(\br-\br_s) & \quad \br \in \Omega \\
\label{eqn:born}
-\nabla^2 \phi_s(\br,\lambda) + \frac{\nu \mu_a(\lambda)}{D(\lambda)} \phi_s(\br,\lambda) \quad = & \quad - \frac{\nu\Delta\mu_a(r,\lambda)}{D(\lambda) }\left( \phi_i(\br,\lambda) + \phi_s(\br,\lambda) \right) & \quad \br \in \Omega
\end{align}
Under the Born approximation, the scattered field is assumed to be much smaller than the incident field, i.e. $\phi_s \ll \phi_i$ and therefore, the total fluence $\phi(\br,\lambda) = \phi_i(\br,\lambda) + \phi_s(\br,\lambda)$ in the right hand side of equation~\eqref{eqn:born}
can be replaced by $\phi_i(\br,\lambda)$.
As a result of this approximation, there is a linear relation between the scattered fluence rate $\phi_s(\br,\lambda)$ and the perturbation of absorption $\Delta\mu_a(\br,\lambda)$.
It should also be noted that both the scattered field $\phi_s(\br\,\lambda)$ and $\phi_i(\br,\lambda)$ satisfy the same boundary conditions in the equations~\eqref{eqn:dirichlet}-\eqref{eqn:robin}. Furthermore, if additional information such as spatial variability is known about the background properties of diffusion and absorption (currently assumed to be homogenous) they can be incorporated into this model~\cite{boverman2005quantitative}. The solution to the photon fluence $\phi_s$ computed at the measurement location $\br_d$ for a particular wavelength $\lambda$ can be written using the following integral equation
\begin{equation}
\phi(\br_d, \lambda) = \phi_i(\br_d, \lambda) + \phi_s(\br_d,\lambda) \quad \approx \quad \phi_i(\br_d,\lambda) - \int_\Omega \phi_d(\br,\lambda) \nu\Delta\mu_a(\br,\lambda) \phi_i(\br,\lambda)d\br \label{eqn:lippschwin}
\end{equation}
where $\phi_d (\br,\lambda)$, which we call the adjoint field, can be derived using the reciprocity property of the Greens function and satisfies the system of equations along with the same boundary conditions in equations~\eqref{eqn:dirichlet}-\eqref{eqn:robin}
\begin{equation}
-\nabla^2 \phi_d(\br,\lambda) + \frac{\nu \mu_a(\lambda)}{D(\lambda)} \phi_d(\br,\lambda) \quad = \quad \frac{1}{D(\lambda)}\delta(\br-\br_d) \qquad \br \in \Omega
\end{equation}
and $\br_d$ corresponds to the detector location.
To relate the scattered fluence to the concentrations of chromophores, the perturbation $\Delta\mu_a(\br,\lambda)$ is decomposed in terms of piecewise constant functions as
\begin{equation} \Delta\mu_a(\br,\lambda) \define \sum_{l=1}^{N_{sp}} \varepsilon_l(\lambda)c_l\chi(\br)\qquad \chi(\br) = \left\{ \begin{array}{ll} 1 & \br \in \mathcal{D} \\ 0 & \br \in \Omega\backslash \mathcal{D}\end{array} \right. \label{eqn:perturbation}
\end{equation}
where $N_{sp}$ is the number of species, $\varepsilon_l$ is the extinction coefficient of species $l$ at wavelength $\lambda$, $c_l$ is the concentration of species $l$ and $\chi$ is an indicator function which depends on $\mathcal{D}$, the domain of support for the perturbation we wish to image. For the purpose of this paper, we will consider that the chromophore concentrations are co-located. This choice was also considered in~\cite{larusson2012parametric}. Further details regarding the governing partial differential equations can be obtained from the following references~\cite{larusson2013parametric,larusson2011hyperspectral,larusson2012parametric}.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.29]{figs/mua2}
\includegraphics[scale=0.29]{figs/dmua}
\caption{The absorption coefficient $\mu_a$ as a function of $\lambda$ in the range $\lambda \in [600,1000]$ [nm]. The chromophore concentrations of the background and the perturbation used to generate this plot are provided in Table~\ref{tab:conc}.}
\label{fig:muaD}
\end{figure}
We also model the diffusion coefficient $D(\lambda)$ using Mie scattering theory~\cite{larusson2011hyperspectral} as
\begin{equation}
D(\lambda) =\frac{\nu \Psi}{3} \left(\frac{\lambda}{\lambda_0}\right)^b
\label{eqn:difflambda}
\end{equation}
The reference wavelength $\lambda_0$ is chosen as $600$ nm and $\Psi$ has units of cm$^{-1}$. The scattering pre-factor $\Psi$ depends primarily on the number and size of scatterers, and a scattering exponent $b$ depends on the size of scatterers in the medium~\cite{grosenick2005timeb}.
\subsection{Discretization using finite elements}
To solve the systems of equations~\eqref{eqn:born} and~\eqref{eqn:incident} we use the standard linear Galerkin finite element approach. Expanding the solution field in an appropriately chosen finite dimensional basis $u_m(\br)$, i.e., \[ \phi_i(\br,\lambda) \approx \sum_{m=1}^N \hat{\phi}_{i,m}(\lambda) u_m(\br)\quad \text{and}\quad \phi_d(\br,\lambda) \approx \sum_{m=1}^N \hat{\phi}_{d,m}(\lambda) u_m(\br)\] Let us denote the discretized incident photon fluence field as $\bphii(\lambda) = [\hat{\phi}_{i,1}(\lambda),\dots,\hat{\phi}_{i,N}(\lambda)]^T$ and the scattered photon fluence $\bphi_d(\lambda) = [\hat{\phi}_{d,1}(\lambda),\dots,\hat{\phi}_{d,N}(\lambda)]^T$. The resulting system of equations can be summarized as
\begin{equation} \label{eqn:discincident}
\left(\bK + \frac{\nu\mu_a}{D}(\lambda)\bM + \frac{1}{2AD(\lambda)} \bR\right) \bphi_i(\lambda) = \frac{1}{D(\lambda)}\bb_{i}
\end{equation}
where the matrices $\bK$, $\bM$ and $\bR$ have entries given by
\begin{align} \label{eqn:matrices}
\bK_{jk} \quad & = \quad \int_\Omega \nabla u_k(\br) \cdot \nabla u_j(\br) d\br \qquad \bR_{jk} = \int_{\partial\Omega_R} u_j(\br)u_k(\br)d\br \\ \nonumber
\bM_{jk} \quad & = \quad \int_\Omega u_k(\br) u_j(\br) d\br
\end{align}
for $j,k = 1,\dots,N$ and the vector $\bb_i$ has entries $\bb_{i,j} = \int_\Omega \delta (\br-\br_s) u_j(\br) d\br$. A similar equation can be derived for the adjoint field $\bphid$ with the same matrices and a different right hand side $\bb_d$ which has entries $\bb_{d,j} = \int_\Omega \delta (\br-\br_d) u_j(\br) d\br$.
The measurements that are collected at the detector are the values of the photon fluence produced as a result of different sources excited at various wavelengths. Let us denote by $\by$ the vector of measurements obtained as
\begin{equation} \by = \sum_{k=1}^{\nsp} c_i\bE_i\bH\bmu + \bEta \qquad \bEta \sim \mathcal{N} (\bzero,\bW^{-2})
\label{eqn:measurement}
\end{equation}
where the rows of $\bH$ are constructed by the discretized representation of the integral equation~\eqref{eqn:lippschwin} and $\bmu$ is a discrete representation of the shape of the absorption perturbation $\chi(\br)$. The matrices $\bE_i = \bI_{N_s}\otimes \bI_{N_{ds}} \otimes \text{diag}(\varepsilon_k(\lambda))$ for $k=1,\dots,\nsp$ and $\varepsilon_k$ are the extinction coefficients for species $k$ that is a function of wavelength. Furthermore, $c_k$ are the concentration of the $k$-th chromophore species. The measurements are typically corrupted by noise which we model as a Gaussian distribution $\mathcal{N}(\bzero,\bW^{-2}$). The noise covariance $\bW$ is modeled as a diagonal matrix with diagonal entries $1/\sigma_m$. The relationship between the standard deviation $\sigma_m$ to the signal to noise ratio (SNR) is described in~\cite{larusson2011hyperspectral}.
\section{Fast solvers for the Born approximation}\label{sec:krylov}
In order to construct the measurement operator $\bH$ and thereby solving the inverse problem, we need the solution of the incident field $\bphii$ and $\bphid$ corresponding to multiple source and detector locations and multiple wavelengths. In this Section, we will discuss an efficient solver for the computing the sequence of parametrized systems using a novel recycling approach based on Krylov subspaces. Krylov subspace methods are a popular class of algorithms for iterative solution to linear systems. Recycling Krylov subspaces in the context of parametric systems with smoothly varying coefficients has been previously considered in~\cite{chan1999galerkin,kilmer2006recycling,parks2006recycling}.
\subsection{Recycling across wavelengths}
For simplicity of notation, we denote by \[ \sigma_j \define \frac{\nu\mu_a(\lambda_j)}{D(\lambda_j)} \qquad \text{and} \qquad \sigma_j' \define \frac{1}{2AD(\lambda_j)}\] the shifts corresponding to wavelengths $j=1,\dots,N_\lambda$, and by $\bx_j \define \bphii D(\lambda_j)$ so that equation~\eqref{eqn:discincident}
\begin{equation}\label{eqn:multipleshifted}
(\bK + \sigma_j \bM + \sigma_j'\bR) \bx_j = \bb \qquad j = 1,\dots, N_\lambda \end{equation}
where matrices $\bK,\bM$ and $\bR$ and vector $\bb$ are independent of the shift $\sigma$ and $\sigma'$. The solution to the adjoint field $\bphid$ can be written in a similar fashion with a different right hand side $\bb_d$. We first make the following change of variables $\bK \leftarrow \bM^{-1/2}\bK\bM^{-1/2}$, $\bR \leftarrow \bM^{-1/2}\bR\bM^{-1/2}$ and transform the vectors as $\bx_j \leftarrow \bM^{-1/2}\bx$ and $\bb \leftarrow \bM^{-1/2}\bb$. This can be done efficiently by using a lumped mass matrix~\cite{hughes2012finite}. Denoted by $\krylov{\bK}{\bb}{n}$, the Krylov subspace of the matrix $\bK$ with starting vector $\bb$, is defined as
\[ \krylov{\bK}{\bb}{n} \define \Span{\bb,\bK\bb,\dots,\bK^{n-1}\bb}\]
Krylov subspaces enjoy an interesting property called shift invariance~\cite{simoncini2007recent}, i.e. \[ \krylov{\bK}{\bb}{n} = \krylov{\bK+ \sigma\bI}{\bb}{n}\]
Several efficient methods exist for solving the system of equations~\eqref{eqn:multipleshifted} (if $\bR = 0$ corresponding to Neumann b.c.s), which solve for multiple shifts roughly at the cost of solving a single system. This is accomplished by generating a subspace that is independent of the shift and use the shift-invariant property of Krylov subspaces (for a detailed review, see~\cite[Section 14.1]{simoncini2007recent} and references therein). However, the presence of a third matrix $\bR$ destroys the shift-invariant property of the Krylov subspace methods, unless $\krylov{\bK}{\bb}{n}$ is an invariant subspace of $\bR$. Even though the shift-invariant property does not hold, \textit{we can utilize information from the solution of the shifted system of equations $(\bK+\sigma_j\bI)\bx_j = \bb$ if the perturbation $\bR$ is not too large in magnitude}. This is the main idea behind the recycling approach to Krylov subspaces that we are proposing.
We construct a shift-invariant basis $\krylov{\bK}{\bb}{n}$ by running $n$ steps of the Arnoldi algorithm and we extract the $k$ smallest Harmonic Ritz eigenvalues and eigenvectors to construct $\mathbf{U}$ and $\bC$ both in $\mathbb{R}^{N\times k}$ such that
\begin{equation}\label{eqn:gcr} \bK\mathbf{U} = \bC \qquad \bC^T\bC = \bI
\end{equation}
By using the shift-invariant property of Krylov subspaces, we know that $\mathbf{U}$ is also an approximately invariant subspace of $\bK+\sigma_j\bI$. We now consider the extension of the solution to the parametric system $\bA_j \define \bK + \sigma_j \bI + \sigma_j'\bR$. When $\normtwo{\bR}$ is small relative to $\normtwo{\bK}$, this can be considered a perturbation to the system $\bK + \sigma_j\bI$ for which we have already generated an approximately invariant subspace. To compute a relation of the kind in Equation~\eqref{eqn:gcr} for the matrix $\bA_j$, we proceed as
\[\bA_j \mathbf{U} = \bC + \sigma_j \mathbf{U} + \sigma_j'\bR \mathbf{U} \define \bC_j'\]
In the above equation, the most expensive step is computing $\bR_\mathbf{U} \define \bR \mathbf{U}$. The matrix $\bR$ only has non-zero entries corresponding to boundaries at which there is refraction index mismatch, in our application it is limited to the top and the bottom boundaries. As a consequence, $\bR$ is even more sparse compared to $\bK$. Moreover, the matrix $\bR\mathbf{U}$ can be precomputed since it will be used across each shift. Compute the thin QR decomposition which using MATLAB notation we represent as $[\bQ_j,\bY_j] = \text{qr}(\bA_j \mathbf{U}, 0)$. The updated updated matrices can now be computed as $\mathbf{U}_j \define \mathbf{U}\bY^{-1}_j$ and $\bC_j = \bQ_j$ and satisfy the relation
\[\bA_j \mathbf{U}_j = \bC_j \qquad \bC_j^T\bC_j = \bI. \]
Here and henceforth, $\mathbf{U}_j$ is not computed explicitly, rather a solve using the upper triangular matrix $\bY_j$ is performed when it is necessary to form products with $\mathbf{U}_j$.
\begin{algorithm}[!ht]
\begin{algorithmic}[1]
\STATE\label{shiftinvariant} Generate a basis for Krylov subspace $\krylov{\bK}{\bb}{n}$
\[ \bK \mathbf{V}_n = \mathbf{V}_{n+1}\bar{\bT}_n \qquad \by_j \define \argmin \left\Vert \mathbf{V}_{n+1}^T\bb - \left(\bar{\bT}_{n+1} + \sigma_j \begin{bmatrix}\bI \\ \bzero \end{bmatrix}\right) \by \right\Vert\]
\STATE Solve the eigenvalue problem $\bar{\bT}_n^T\bar{\bT}_n \bz = \theta \bT_n^T\bz $ and retain $k$ smallest eigenvalues in magnitude. Collect the eigenvectors into a matrix $\bZ_k$ and the corresponding eigenvalues as $\boldsymbol\Theta_k$.
\STATE Compute $\bC = \mathbf{V}_{n+1}\bar{\bT}_n\bZ_k\boldsymbol\Theta_k$ and $\mathbf{U}_k = \mathbf{V}_{n+1}\bar{\bT}_n\bZ_k$
\STATE Compute the thin QR $[\bC,\bY] = \text{qr}(\bC,0)$ and set $\mathbf{U} = \mathbf{U}\bY^{-1}$
\COMMENT {Shift-invariant deflation subspace}
\STATE Compute the initial solution $\tilde{\bx}_{0,j} = \mathbf{V}_n\by_j$ for $j=1,\dots, N_\lambda$
\FOR {$j = 1,\dots,N_\lambda$}
\STATE $\bC_j' = \bC + \sigma_j \mathbf{U} + \sigma_j'\bR\mathbf{U} $ \COMMENT{Compute new deflation subspace}
\STATE $[\bC_j, \bY_j] = \text{qr}(\bC_j',0)$ and set $\mathbf{U}_j = \mathbf{U}\bY_j^{-1}$.
\STATE Compute $\br_j = \bb - \bA_j\bx_{0,j}$ and $\bx_{-1,j} \define \tilde{\bx}_{0,j} + \mathbf{U}_j\bC_j^T\br_j$ \COMMENT {Initial deflation}
\STATE Generate $\mathbf{V}_{m-k+1}^{(j)}$ and $\bar{\bT}_{m-k}^{(j)}$ by applying $m-k$ steps of the Arnoldi algorithm using matrix $(\bI - \bC_j\bC_j^T)\bA_j$ applied to the initial deflated vector $(\bI - \bC_j\bC_j^T)\br_j$
\STATE Solve the least-squares system of equations~\eqref{eqn:minres} with $\beta_j = \normtwo{(\bI - \bC_j\bC_j^T)\br_j}$
\[ [\bQ,\bZ] = \text{qr}(\bar{\bT}_{m-k}^{(j)},0) \quad \by_{2,j} = \bZ^{-1}(\bQ^T\beta_j \be_1) \quad \by_{1,j} = - \bF_k^{(j)}\by_{2,j} \]
to generate approximate solution $\bx_{m,j} = \bx_{-1,j} + \mathbf{U}_j\by_{1,j} + \mathbf{V}_{m-k}^{(j)}\by_{2,j}$
\ENDFOR
\end{algorithmic}
\caption{Augmented GMRES for parametric systems}
\label{alg:auggmres}
\end{algorithm}
We adopt the same recycling strategy as~\cite{parks2006recycling}. However, where we differ from this approach is the way we construct and update the approximate invariant subspace, as we now explain. The strategy in~\cite{parks2006recycling} is to first generate an approximate a recycling basis for the matrix $\bA_1$ and then update the recycling subspace of perturbed matrices $\bA_j = \bA_{j-1} + \Delta \bA_j$ after the convergence of the augmented Krylov solver for each of the system. In the context of our problem, note that
\[\Delta \bA_j \define (\sigma_j-\sigma_{j-1})\bI + (\sigma_j' -\sigma_{j-1}')\bR .\]
Since the matrices are known a priori and only the shifts are varying, we are able to do something different: once the basis $\mathbf{U}$ and $\bC$ are known, our approach to update $\mathbf{U}_j$ and $\bC_j$ can be easily performed independently and can be parallelized in a straightforward manner. We will discuss the steps and cost for obtaining
$\bC_j$ at the end of this section. First, let us assume that $\mathbf{U}_j, \bC_j$ are available, and discuss our solution technique.
To begin, we assume we have computed
$\tilde{\bx}_{0,j}$ iteratively and simultaneously for all $j$ using the shift-invariant property as the approximate
solution to
$(\bK + \sigma_j \bI) \bx_j= \bb$. We use $\tilde{\bx}_{0,j}$ as a first estimate of the solution to $(\bK + \sigma_j \bI + \sigma_j' \bR) \bx = \bb$ obtained by exploiting the shift-invariance of the Krylov subspaces (see Step~\ref{shiftinvariant} in Algorithm~\ref{alg:auggmres}). The corresponding initial residual is denoted $\br_{0,j} := \bb - \bA_j \tilde{\bx}_{0,j}$. Because we expect this solution to need augmentation, we next search for a
better estimate of the form $\bx_{-1,j} = \tilde{\bx}_{0,j} + \mathbf{U}_j \bz$. We choose $\bz$ such that
\[ \bz = \argmin \| \bb - \bA_j(\tilde{\bx}_{0,j} + \mathbf{U}_j \bz ) \| = \argmin \| \br_{0,j} - \bC_j \bz \| \]
For this choice, we get a new solution estimate $\bx_{-1,j} = \tilde{\bx}_{0,j} + \mathbf{U}_j \bC_j^T (\bb - \bA_j\tilde{\bx}_{0,j} )$ that gives the residual $\br_{-1,j} = (\bI - \bC_j \bC_j^T) \br_{0,j}$.
The approximate solution $\bx_{m,j}$ for the parametric system $j$ is obtained by searching in the augmented affine subspace
\begin{equation}
\bx_{m,j} \in \bx_{-1,j} + \Span{\mathbf{U}_j} \oplus \krylov{(\bI-\bC_j\bC^T_j)\bA_j}{(\bI-\bC_j\bC^T_j)\br_{0,j}}{m-k}
\end{equation}
that is, by searching for solutions in the deflated subspace $\krylov{(\bI-\bC_j\bC^T_j)\bA_j}{(\bI-\bC_j\bC^T_j)\br_{0,j}}{m-k}$ obtained by applying $m-k$ steps of the Arnoldi algorithm with the deflated matrix $(\bI-\bC_j\bC^T_j)\bA_j$ to the starting vector $(\bI-\bC_j\bC^T_j)\bb$ and augmented with the subspace $\mathbf{U}_j$. The initial guess $\bx_{-1,j}$ is obtained as $ \mathbf{U}_j\bC^T_j(\bb - \bA\tilde{\bx}_{0,j})$ and the initial residual is $\br_{0,j} = \bC_j\bC^T_j\bb$, (where $\tilde{\bx}_{0,j}$ is the approximation solution obtained by using the shift-invariant property to solve $(\bK + \sigma_j\bI)\bx_j = \bb$).
The following discussion closely mirrors the presentation of recycling in \cite{parks2006recycling,wang2007large,kilmer2006recycling,mello2010recycling}.
The Arnoldi algorithm on the deflated problem yields the matrix relationship
\[ (\bI - \bC_j\bC^T_j)\bA_j \mathbf{V}_{m-k}^{(j)} = \mathbf{V}_{m-k+1}^{(j)}\bar{\bT}_{m-k}^{(j)}\]
where the superscripts indicate the system index $j$. The above equation can be rewritten as a modified Arnoldi relationship by defining $\bF_k^{(j)} \define \bC^T_j\bA_j\mathbf{V}_{m-k}^{(j)}$ and reorganizing as
\begin{equation}
\bA [\mathbf{U}_j, \mathbf{V}_{m-k}^{(j)}] = [\bC_j, \mathbf{V}_{m-k+1}^{(j)}] \begin{bmatrix}\bI_k & \bF_k^{(j)} \\ \bzero & \bar{\bT}_{m-k}^{(j)} \end{bmatrix}
\end{equation}
Now, $\mathbf{V}_{m-k}^{(j)}$ forms a basis for the subspace $\krylov{(\bI-\bC_j\bC^T_j)\bA_j}{(\bI-\bC_j\bC^T_j)\br_{0,j}}{m-k}$ and now we search for solutions of the form $\bx_{m,j} \in \bx_{-1,j} + \Span{\mathbf{U}_j} \oplus \Span{\mathbf{V}_{m-k}^{(j)}}$ and can be written as
\[ \bx_{m,j} = \bx_{-1,j} + \mathbf{U}_j\by_{1,j} + \mathbf{V}_{m-k}^{(j)} \by_{2,j}\]
The solution to the coefficients $\by_{1,j}$ and $\by_{2,j}$ are obtained by minimizing the residual which results in the following least squares problem
\begin{equation}\label{eqn:minres} \min_{\by_{1,j},\by_{2,j}} \left\lVert \begin{bmatrix} \bzero \\ \beta_j \be_1 \end{bmatrix} - \begin{bmatrix}\bI_k & \bF_k^{(j)} \\ \bzero & \bar{\bT}_{m-k}^{(j)} \end{bmatrix} \begin{bmatrix} \by_{1,j} \\ \by_{2,j} \end{bmatrix}\right\rVert \end{equation}
where $\beta_j = \normtwo{(\bI-\bC_j\bC^T_j)\br_{0,j}}$. Since the number of iterations are expected to be small, we store the vectors $\mathbf{V}_{m-k}^{(j)}$ and solve the least squares problem in equation~\eqref{eqn:minres} directly. If the number of iterations are expected to be large, we do not need the solution to $\by_{1,j}$ and $\by_{2,j}$ explicitly, only the products $\mathbf{V}_{m-k}^{(j)}\by_{2,j}$ and $\bF_k^{(j)} \by_{2,j}$ which can be obtained using short-term recurrence relation similar to recycled MINRES~\cite{wang2007large} (assuming the matrices $\bK$, $\bM$ and $\bR$ are symmetric).
\subsection{Computational and storage costs}
We now discuss the computational and storage costs involving Algorithm~\ref{alg:auggmres} and the overhead induced by the augmented approach for constructing the augmented basis $\mathbf{U}$. We note that the loop can be executed in parallel because no information is shared across the solves except for the initial choice of $\mathbf{U}$. We focus only on costs that are linear in the size of the matrix, i.e. costs of the form $\bigO(N)$ since the costs involving smaller matrices are negligible. The augmented approach in Algorithm~\ref{alg:auggmres} requires additional storage of $2Nk$ for the matrices $\mathbf{U}$ and $\bC$. For the pre-computation of the new bases $\mathbf{U}_j$ and $\bC_j$ the major cost is the QR factorization which is $\bigO(Nk^2)$ for each wavelength. To accelerate this computation, we use a more efficient approach at the possible expense of some accuracy. We first compute
\begin{align*}(\bC_j')^T\bC_j' = & \quad \bI + \sigma_j ( \bC^T\mathbf{U} + \mathbf{U}^T\bC)+\sigma_j'( \bC^T\bR_\mathbf{U}+ \bR_\mathbf{U}^T\bC) \\
& \quad + \sigma_j\sigma_j' (\bR_\mathbf{U}^T\mathbf{U} + \mathbf{U}^T\bR_\mathbf{U}) + \sigma_j^2 \mathbf{U}^T\mathbf{U}+ (\sigma_j')^2 \bR_\mathbf{U}^T\bR_\mathbf{U} .
\end{align*}
Next, the small $k\times k$ matrices such as $\mathbf{U}^T\bC$ are precomputed and stored since they are independent of the shifts $\sigma_j$ and $\sigma_j'$. Then the Cholesky factorization of $(\bC_j')^T\bC_j'$ is computed and $\bC_j$ is obtained as $\bC_j = \bC_j'\bF^{-1}$. Note as before that the inverse $\bF^{-1}$ is not computed explicitly. Because of this pre-computation, the additional cost per wavelength is now only $\bigO(k^3)$. Suppose the algorithm converges in $m$ iterations, then the algorithm requires $m-k$ additional matrix-vector products and $\bigO(Nm^2 + mkN )$ other floating point operations. For an efficient algorithm, the overhead costs must be offset by the gains obtained by decreasing the number of iterations. Numerical experiments performed in Section~\ref{sec:results} demonstrate that the reduction in the number of matrix-vector products due to deflation offsets the additional computational cost due to pre-computation and re-orthogonalization.
\section{Fast compression of $\bH$}\label{sec:compress}
After computing the incident field $\bphii$ and the adjoint field $\bphid$ the next step is to compute the measurement operator $\bH$. The construction of $\bH$ is described in Equation~\eqref{eqn:measurement}. However, recall that the operator $\bH$ is expensive to store and compute, and therefore our goal is to produce an approximate factorization of $\bH \approx \mathbf{U}\mathbf{V}^T$. Furthermore, an optimal low-rank compression using SVD scales as $\bigO \left( \min \{M,N\}^2, \max\{M,N\} \right)$ which is prohibitively expensive. Here we present an algorithm that avoids computing $\bH$ in its entirety, but computes and compresses sub-blocks and then combines the compressed sub-blocks in a recursive manner.
\begin{figure}
\centering
\begin{subfigure}{0.4\textwidth}\centering
\begin{tikzpicture}[level/.style={sibling distance=40mm/#1},scale=0.55]
\tikzstyle{ann} = [draw=none,fill=none]
\node [ann] (z){$\{1,2,3,4,5 \}$}
child {node [ann] (a) {$\{1,2\}$}
child {node [ann] {\small{$\{ 1\}$}}
}
child {node [ann] {\small{$ \{ 2\} $}}
}
}
child {node [ann] (j) { \small{ $\{ 3,4,5\}$}}
child {
node [ann] {\small{$\{3,4\}$}}
child {
node [ann] {\small{$\{3\}$} }
}
child{
node [ann] {\small{$\{4\}$} }
}
}
child {
node [ann] (l) {\small{$\{5\}$}}
}
};
\end{tikzpicture}
\end{subfigure}
\begin{subfigure}{0.4\textwidth}
\begin{tikzpicture}[level/.style={sibling distance=65mm/#1},scale=0.65]
\tikzstyle{ann} = [draw=none,fill=none]
\node [ann] (z){$ \mathbf{U}\mathbf{V}^T$}
child {node [ann] (a) {$\mathbf{U}_1^{(1)}\mathbf{V}_1^{(1)}{}^T$}
child {node [ann] {\small{$\mathbf{U}_1^{(2)}\mathbf{V}_1^{(2)}{}^T$}}
}
child {node [ann] {\small{$ \mathbf{U}_2^{(1)}\mathbf{V}_2^{(1)}{}^T $}}
}
}
child {node [ann] (j) { \small{ $\mathbf{U}_2^{(2)}\mathbf{V}_2^{(2)}{}^T$}}
child {
node [ann] {\small{$\mathbf{U}_3^{(2)}\mathbf{V}_3^{(2)}{}^T$}}
child {
node [ann] {\footnotesize{$\mathbf{U}_1^{(3)}\mathbf{V}_1^{(3)}{}^T$} }
}
child{
node [ann] {\footnotesize{$\mathbf{U}_2^{(3)}\mathbf{V}_2^{(3)}{}^T$} }
}
}
child {
node [ann] (l) {\footnotesize{$\mathbf{U}_4^{(2)}\mathbf{V}_4^{(2)}{}^T$}}
}
};
\end{tikzpicture}
\end{subfigure}
\caption{The size of the matrices at the leaf level are $ N_{ds}N_\lambda \times N$ and $N_s = 5$. The superscripts denote the level of the tree, where as the subscripts denote the order at a given level (which can be different than the global ordering of the sources). At the leaves the matrices are approximated by low-rank factors and then agglomerated recursively based on the tree. Here we consider a tree corresponding to $5$ sources. A possible source configuration that produces the tree can be obtained in Figure~\ref{fig:domain}. }
\label{fig:part}
\end{figure}
\subsection{Outline} The rows of $\bH$ can be partitioned as
\[ \bH^T = [\bH_1^T,\dots,\bH_{N_s}^T] \]
where each block $\bH_i$ is of size $N_{ds}N_\lambda \times N$. Each row of $\bH$ represents a discretized version of the integral described in Equation~\eqref{eqn:lippschwin} that combines the incident field and the adjoint field. Instead of compressing $\bH$ by using low-rank factorization techniques such as truncated SVD, the idea is to compress each block $\bH_i \approx \mathbf{U}_i\mathbf{V}_i^T$ locally and then combine the factorizations in a recursive fashion. The full algorithm is provided in Algorithm~\ref{alg:recursive} and is illustrated for $N_s = 4$ in Figure~\ref{fig:part}. The low-rank approximation of the blocks $\bH_i$ can be accomplished either using Randomized SVD or partially pivoted ACA algorithms described in Subsection~\ref{sec:compression}. Then a scheme for ordering the blocks $\bH_i$ is presented in Subsection~\ref{sec:part} that uses a spatial bisection tree to order the source locations by their spatial proximity. The agglomeration of the low-rank factors is accomplished by recursion using this tree structure. In Subsection~\ref{sec:costs} we analyze the computational costs of this recursive compression scheme and conclude with an error analysis in Subsection~\ref{sec:acc}.
\subsection{Low Rank representation}\label{sec:compression}
As was mentioned earlier, we require a strategy to compute a low-rank factorization of the sub-blocks $\bH_j^T$ for $j=1,\dots,N_s$. Consider the blocks $ \bH_j $ of size $m \times n$ where $m = N_{ds}N_\lambda$ and $n = N$. Considering only one source, we seek a low-rank approximation of the form $\bH \approx \mathbf{U} \mathbf{V}^T$ and the number of columns of $\mathbf{U}$ and $\mathbf{V}$ is denoted by $r$. It is well known that the best low-rank approximation to rank $r$ is obtained by truncating the SVD to rank $r$. In this case, we have $\| \bH - \mathbf{U}_r\Sigma_r\mathbf{V}_r^T \|_2 = \sigma_{r+1} $. However, computing the SVD is expensive since it requires $\bigO(nm^2)$ operations assuming $n \geq m$. The advantage of the low-rank representation is that the cost of storing the decomposition and computing matrix-vector products are both given by $\bigO(r(m+n))$ instead of $\bigO(mn)$. When $r \ll \min\{m,n\}$, this reduction can represent significant savings.
In the Appendix~\ref{sec:lowrank} we describe two approximate methods that compute a low-rank representation but have a lower computational cost asymptotically than the SVD, namely randomized SVD (RandSVD) and partially pivoted Adaptive Cross Approximation (ppACA). Here, we only summarize the resulting computational costs. For a matrix of dimensions $m\times n$ that has a rank $r$ the cost can be expressed as
\begin{equation}
\mu_\text{Comp}(m,n;r) \quad = \quad \left\{ \begin{array}{cc} C_1rmn + C_2 r^2(m+n) & \text{RandSVD}\\ C_3r^2(m+n)& \text{ppACA}\end{array}\right.
\label{eqn:svdcosts}
\end{equation}
where constants $C_1$, $C_2$ and $C_3$ are assumed to be known and provided in the literature (see Appendix~\ref{sec:lowrank}). Note that the costs of the low-rank approximation both using RandSVD and ppACA are asymptotically smaller than the cost of an SVD which scales as $\bigO(nm^2)$ assuming $n > m$.
\subsection{Agglomerating low ranks}\label{sec:agglomeration}
Having produced low-rank approximations to the sub-blocks $\bH_j$ for $j = 1,\dots,N_s$ we now consider the problem of agglomerating low-rank factors to produce a global low-rank factorization. Here we consider only two sources, i.e. $N_s = 2$ and as before, $\bH^T = [\bH_1^T, \bH_2^T]$ of size $m \times n$ where $m = N_{ds}N_\lambda$ and $n = N$. Suppose we have the low-rank factors $\bH_1 \approx \mathbf{U}_1 \mathbf{V}_1^T$ and $\bH_2 \approx \mathbf{U}_2 \mathbf{V}_2^T$ each with rank $r$ which have been compressed according to some predetermined tolerance $\varepsilon$. The low-rank factors can be then combined as
\[\bH = \begin{bmatrix} - \bH_1 - \\ - \bH_2 - \end{bmatrix} \approx \begin{bmatrix} \mathbf{U}_1\mathbf{V}_1^T \\ \mathbf{U}_2\mathbf{V}_2^T \end{bmatrix} = \begin{bmatrix}\mathbf{U}_1 & \\ & \mathbf{U}_2 \end{bmatrix} \begin{bmatrix} - \mathbf{V}_1^T - \\ - \mathbf{V}_2^T - \end{bmatrix} .\]
Now the leftmost matrix is $m \times r$ and has independent columns by construction.
The rightmost $r \times n$ matrix, however, may have a rank smaller yet than $\min{r,n}$. So we compute
$\begin{bmatrix} - \mathbf{V}_1^T- \\ - \mathbf{V}_2^T - \end{bmatrix} = \mathbf{U}_\mathbf{V} \mathbf{V}^T$ where $\mathbf{U}_\mathbf{V}, \mathbf{V}$ have
$r' < r$ columns, and we set
\[ \begin{bmatrix}\mathbf{U}_1 & \\ & \mathbf{U}_2 \end{bmatrix} \mathbf{U}_\mathbf{V} \mathbf{V}^T = \mathbf{U}\mathbf{V}^T, \qquad \mathbf{U} \in \mathbb{R}^{2N_{ds}N_{\lambda} \times r'} , \mathbf{V} \in \mathbb{R}^{N \times r'} .\]
We will need to form $\mathbf{U}$ explicitly, and this requires computing a rank-r' approximation to the stacked $\mathbf{V}_i$ matrix. The cost must therefore include the matrix-matrix product (using the block diagonal structure of the leftmost matrix in the product) and the cost of this additional factorization, using the same algorithm as was used
to approximately factor the $\bH_i$.
The estimate of the cost of agglomeration of sub-blocks is therefore,
\begin{equation}
\mu_\text{Agg}(m,n;r',r) \quad = \quad
\left\{ \begin{array}{cc} C'rr'(m+n) & \text{RandSVD}\\ C''(r')^2(m+n)& \text{ppACA}\end{array}\right.
\end{equation}
where the constant $C'$ and $C''$ are assumed to be known (see Appendix~\ref{sec:lowrank}). For simplicity of analysis we will only use RandSVD for agglomerating low-rank representations.
\subsection{Recursive SVD}\label{sec:part}
We have, thus far, described a procedure to compute low-rank factors $\bH_j \approx \mathbf{U}_j\mathbf{V}_j^T$ corresponding to source indexes $j=1,\dots,N_s$ and shown how to agglomerate low-rank factors when the number of sources are $2$. However, as we shall show, this procedure can be implemented recursively. The complete algorithm for computing and compressing $\bH$ is summarized in Algorithm~\ref{alg:recursive}.
Let $I = \{1,\dots,N_s\}$ denote an index set and let $|I|$ denote the cardinality of the set $I$. Let the index sets $I$ be partitioned into binary trees denoted by $T_I$ respectively. For each $\tau \in T_I$, we denote the set of its sons by $S(\tau) \subset T_I$ and the leaves of the tree are denoted by $\mathcal{L}(T_I)$. The tree $T_I$ can be used to recursively spatially partition the domain so that the sources are ordered such that recursively combining low-rank factors from sub-blocks corresponds to combining nearby sources since the corresponding Green's functions are known to be highly compressible~\cite{chaillat2012faims}. As a result, by this ordering, we hope to gain a larger compression at each level in the tree. Given the tree $T_I$ we can recursively compress the sub-blocks using Algorithm~\ref{alg:recursive} which is initialized using the root of the tree, which corresponds to the index set $I$. The partitioning of the matrix $\bH$ into sub-blocks and their recursive compression for $N_s = 4$ using the Figure~\ref{fig:part}. The construction of the tree is described in Algorithm~\ref{alg:recursive}.
\begin{algorithm}[!ht] \label{alg:rr}
\begin{algorithmic}[1]
\IF {$S(\tau) \neq \emptyset$}
\STATE $\mathcal{U} := \emptyset$, $\mathcal{V} := \emptyset$
\FORALL {$\tau' \in S(\tau ) $}
\STATE $\mathbf{U}_{\tau'},\mathbf{V}_{\tau'} = \text{RecursiveLowRank}(\tau', \varepsilon)$
\STATE $\mathcal{U}.\text{append} ( \mathbf{U})$, $\mathcal{V}.\text{append} ( \mathbf{V})$
\ENDFOR
\ELSE
\FORALL {$\tau' \in \mathcal{L}(T_I ) $}
\STATE Compute the sub-block $\bH_\tau'$ using Equation~\eqref{eqn:lippschwin} \\
\COMMENT {// Use fast solvers developed in Section~\ref{sec:krylov} for incident $\bphii$ and adjoint field $\bphid$. }
\STATE $[\mathbf{U},\mathbf{V}] = \text{LowRank}(\bH|_{\tau'}, \varepsilon)$ \\
\COMMENT {//Compute low-rank at leaf level such that $\norm{\bH_{\tau'}-\mathbf{U}\mathbf{V}^T} \leq
\varepsilon \norm{\bH_{\tau'}}$. see Section~\ref{sec:compression}}
\STATE $\mathcal{U}.\text{append} ( \mathbf{U})$, $\mathcal{V}.\text{append} ( \mathbf{V})$
\ENDFOR
\ENDIF
\STATE $[ \mathbf{U},\mathbf{V}]=$ Agglomerate($\mathcal{U},\mathcal{V}$, $\varepsilon$)
\COMMENT {// Agglomerate low-rank factors, see Section~\ref{sec:agglomeration}}
\RETURN $\mathbf{U}, \mathbf{V}$ such that $\bH \approx\mathbf{U}\mathbf{V}^T$
\end{algorithmic}
\caption{RecursiveLowRank$(\tau, \varepsilon)$}
\label{alg:recursive}
\end{algorithm}
\begin{figure}
\centering
\begin{subfigure}{0.75\textwidth}\centering
\begin{tikzpicture}[scale = 0.2]
\def10{10}
\def 5{ 5}
\draw (0,0) rectangle (10,10);
\foreach \Point [count = \xi] in {(0.4*10,0.3*10), (0.18*10,0.65*10), (0.6*10,0.2*10), (0.95*10,0.4*10), (0.85*10, 0.85*10)}{
\ifnum\xi=3
\node [label = {[xshift=0.025*10 cm,yshift=-0.05*10 cm]$\xi$}] at \Point {\textbullet};
\else
\node [label = {[xshift=-0.025*10 cm,yshift=-0.05*10 cm]$\xi$}] at \Point {\textbullet};
\fi
}
\draw [->, thick] (10 +0.2* 5, .5*10) -- (10 +0.8* 5,.5*10);
\draw (10+ 5,0) rectangle (2*10 + 5 , 10);
\draw (1.5*10 + 5,0) -- (1.5*10 + 5,10);
\foreach \Point [count = \xi] in {(10+ 5 + 0.4*10,0.3*10), (10+ 5 + 0.18*10,0.65*10), (10+ 5 + 0.6*10,0.2*10), (10+ 5 + 0.95*10,0.4*10), (10+ 5 + .85*10, 0.85*10)} {
\ifnum\xi=3
\node [label = {[xshift=0.025*10 cm,yshift=-0.05*10 cm]$\xi$}] at \Point {\textbullet};
\else
\node [label = {[xshift=-0.025*10 cm,yshift=-0.05*10 cm]$\xi$}] at \Point {\textbullet};
\fi
}
\draw [->, thick] (2*10 + 5 +0.2* 5, .5*10) -- (2*10 + 5 +0.8* 5,.5*10);
\draw (2*10 + 2* 5,0) rectangle (3*10 + 2* 5 ,10);
\draw (2.5*10 + 2* 5,0) -- (2.5*10 + 2* 5,10);
\draw (2.5*10 + 2* 5, .5*10) -- (3*10 + 2* 5,0.5*10);
\foreach \Point [count = \xi] in {(2*10+2* 5 + 0.4*10,0.3*10), (2*10+2* 5 + 0.18*10,0.65*10), (2*10+2* 5 + 0.6*10,0.2*10), (2*10+2* 5 + 0.95*10,0.4*10), (2*10+2* 5 + .85*10, 0.85*10)}{
\ifnum\xi=3
\node [label = {[xshift=0.025*10 cm,yshift=-0.05*10 cm]$\xi$}] at \Point {\textbullet};
\else
\node [label = {[xshift=-0.025*10 cm,yshift=-0.05*10 cm]$\xi$}] at \Point {\textbullet};
\fi
}
\end{tikzpicture}
\end{subfigure}
\caption{ Top view of the bounding box for a possible source configuration. Locations of $5$ different sources that can be subdivided recursively using Algorithm~\ref{alg:geom}. The resulting tree $T_J$ that is returned by the algorithm, when initialized by the index set $J = \{1,\dots,5\}$ is provided in Figure~\ref{fig:part}. }
\label{fig:domain}
\end{figure}
We make the assumption that the locations of the source, detectors and the support of the perturbation are non-overlapping and well separated. For concreteness, let us assume that the breast is placed between parallel plates and is enclosed by a cuboid of dimensions $[-L_x,L_x] \times [-L_y,L_y] \times [0,L_z]$. Furthermore, let us assume that the sources are located on the top plate $z=L_z$ and the detectors are located at $z=0$ and the detectors for any given source are roughly co-axial. The construction of the tree is performed as follows, and a simple example illustrating this construction is provided in Figure~\ref{fig:domain}. Let $J$ be the index set corresponding to source locations. The tree $T_J$ is constructed recursively by geometric bisection applied on a 2D domain. Given the initial bounding box containing all the source (in this case, a box of size $[-L_x,L_x] \times [L_y, L_y]$), the domain is split in a direction perpendicular to the direction of maximum expanse and the sources are split between the newly created split domains. If the number of points in each domain are greater than $2$, the procedure is computed recursively until the number of points in each leaf is no greater than $2$. If the points are uniformly distributed, the number of levels $L = \log_2(N_s)$, where $N_s$ is the number of sources. The algorithm is summarized in Algorithm~\ref{alg:geom}. The algorithm is initialized using $\tau = J$, $\alpha = [-L_x,L_x]$ and $[-L_y,L_y]$.
\begin{algorithm}[!h]
\begin{algorithmic}[1]
\STATE $j_\text{max} :=\arg\max\{ \beta_j- \alpha_j : j \in \{ 1,\dots,d\} \}$
\STATE $\gamma := (\alpha_{j_\text{max}} + \beta_{j_\text{max}})/2$ \COMMENT{// Split cluster in chosen direction }
\STATE $\tau_1 := \emptyset$, $\tau_2 := \emptyset$
\FOR {$i \in \tau$}
\IF {$e_{j_\text{max}}^T\bx_i \leq \gamma $ }
\STATE $\tau_1 = \tau_1 \bigcup \{i\} $
\ELSE
\STATE $\tau_2 = \tau_2 \bigcup \{i\} $
\ENDIF
\ENDFOR
\STATE Define $\beta' := \beta$, $\beta_{j_\text{max}}' := \gamma$, and $\alpha' := \alpha$, $\alpha_{j_\text{max}}' := \gamma$
\STATE Split$(\tau_1,\alpha,\beta') $ and Split$(\tau_2,\alpha',\beta)$ \COMMENT {// Split only if $|\tau_i| > 2$ for $i=1,2$.}
\end{algorithmic}
\caption{Split$(\tau, \alpha, \beta)$ // Geometric bisection to construct tree}
\label{alg:geom}
\end{algorithm}
\subsection{Computational cost}\label{sec:costs}
In this section we present a detailed analysis of the computational cost of the recursive SVD described in Algorithm~\ref{alg:recursive}. Our analysis is in the same spirit as the one provided in~\cite{hackbusch2014new}. However, while they used the full SVD algorithm to produce the optimal truncation at each level, we will consider the use of approximate low-rank factorizations that have better computational costs. As we shall see, this analysis and results are more sharp than~\cite{hackbusch2014new}
since we account for the cost depending on rank of the sub-blocks at each level of the tree.
The starting point of our analysis is splitting the costs into a contribution that comes from the preparation at the leaf level that is computed using the low-rank factorization techniques listed in Section~\ref{sec:compression} and the contribution that comes from agglomerating low-rank sub-blocks. The number of sub-blocks to agglomerate become fewer the higher up we are in the tree, but the size of the matrices to be agglomerated increases. The total cost is therefore equal to
\begin{equation}\label{eqn:genericcosts}
\text{Cost } = \sum_{b \in \mathcal{L}(T_{I})} \mu_{Comp}(m^{(L)},n^{(L)};r_L) + \sum_{\ell=0}^{L-1}\sum_{b \in T^{(l)}_{I} } \mu_\text{Agg}\left(m^{(\ell)},n^{(\ell)};r_{L-\ell-1},r_{L-\ell}\right) \text{ flops}.
\end{equation}
We assume that $N_s$ is a power of $2$ and each partition has the same size, so that the size of the partition at the leaf level is $N_\lambda N_{ds} \times N$, and $N_s = 2^L$ where $L$ is the depth of the tree $T_{I}$. Furthermore, we denote by $r_{\ell}$, the maximum block rank of any partition at level $\ell$. The cost of recursive SVD is then
\begin{align*}
\text{Cost } = & \quad 2^L\mu_\text{Comp}\left(\frac{M}{2^L},N; r_L\right) & +& \quad \sum_{\ell=0}^{L-1}2^{L-\ell-1}\mu_\text{Agg}\left(2^{\ell -L}M,N;r_{L-\ell - 1},r_{L-\ell}\right) \\
= & \quad C_12^Lr_L\frac{M}{2^L}N + 2^LC_2r_L^2\left(\frac{M}{2^L} + N\right) & +& \quad \sum_{\ell=0}^{L-1}C'2^{L-\ell-1}r_{L-\ell-1}r_{L-\ell}(2^{\ell-L} M + N) \\
= & \quad C_1r_LM{N} + C_2r_L^2\left(M+2^LN\right) & + &\quad \sum_{\ell=0}^{L-1}C'2^{L-\ell-1} 2\delta^{L-\ell-1}_{L-\ell}r_{L-\ell}^2(2^{\ell-L} M + N) \\
= & \quad C_1r_LM{N} + C_2r_L^2\left(M+2^LN\right) &+ & \quad 2^{L-1}C'r_{L}^2\sum_{\ell=0}^{L-1}(M/2^L + 2^{-\ell}N)2\delta^{L-\ell-1}_{L-\ell} \prod_{k=L-\ell}^{L-1}\left(2\delta^{k}_{k+1}\right)^2
\end{align*}
Here, we assume that $r_{\ell} = 2\delta^{\ell}_{\ell+1}r_{\ell+1}$, that is, the ratio of the ranks at level $\ell$ that is obtained by agglomeration of two sub-blocks each of rank at most $r_{\ell+1}$ is a factor of $\delta^{\ell}_{\ell+1}$, which is dependent on the particular level under consideration. It is easy to see that $0 < \delta^{\ell}_{\ell+1} \leq 1$ for all $\ell=0,\dots,L-1$. If we further make the assumption that $ \max_{0\leq \ell < L} \delta^{\ell}_{\ell+1} = \delta$ is independent of the level $\ell$, then we have following result that the total computational cost of recursive low-rank truncation can be further simplified to
\begin{equation}
C_1r_LM{N} + C_2r_L^2\left(M+2^LN\right) + 2^{L}C'r_{L}^2\delta\left(\frac{1-(\sqrt{2}\delta)^{2L-1}}{1-\sqrt{2}\delta}\frac{M}{2^L} + \frac{1-(2\delta)^{2L-1}}{1-2\delta}{N}\right) .
\end{equation}
We define the quantities $f_1$ and $f_2$ as follows:
\begin{align*}
f_1 \define \quad \sum_{\ell=0}^{L-1}2^{-\ell}\delta^{L-\ell-1}_{L-\ell}\prod_{k=L-\ell}^{L-1} \left(2\delta^{k}_{k+1}\right) \quad & \leq \quad \delta \frac{1-(\sqrt{2}\delta)^{2L-1}}{1-\sqrt{2}\delta} \\
f_2 \define \quad \sum_{\ell=0}^{L-1}\delta^{L-\ell-1}_{L-\ell}\prod_{k=L-\ell}^{L-1} \left(2\delta^{k}_{k+1}\right) \quad & \leq \quad \delta \frac{1-(2\delta)^{2L-1}}{1-2\delta}.
\end{align*}
Here, we assume that $\delta \neq 1/\sqrt{2}, 1/2$. If this were the case, the appropriate sums would simplify to equal $L$. In the worst case, when there is no compression at higher levels $\delta = 1$ and then $f_1 \sim 2^L$ and $f_2 \sim 4^L$.
\begin{table}[!ht]\centering
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|c|c|} \hline
\multirow{2}{*} {Method} & \multicolumn{2}{|c|}{ Leaf computation} & \multicolumn{2}{|c|}{Tree computation} \\ \cline{2-5}
&RandSVD & ppACA& Average & Worst \\ \hline
Direct & $C_1RMN + C_2 R^2 (M+N)$ & $C_3R^2(M+N)$ & - & - \\ \hline
Recursive & $ C_1r_LMN + C_2r_L^2 (M+ 2^LN)$ & $C_3r_L^2(M+2^LN)$& $Cr_L^2(Mf_2 +2^LNf_1) $ & $r_L^2C4^L(M +N) $ \\ \hline
\end{tabular}
\caption{Summary of computational costs of the recursive SVD algorithm. Here $M=N_sN_{ds}N_\lambda$ is the number of measurements and $N$ is the grid size. The storage costs are $\bigO(R(M+N))$ where $R$ is the global rank of the low-rank factorization. }
\label{tab:compcosts}
\end{table}
We now compare the computational costs between different methods that we have outlined in this section. It can be readily seen that the cost of the recursive factorizations scale asymptotically better than the SVD which scales as $\bigO(NM^2)$ assuming $M \leq N$. We denote by `Direct', the low-rank algorithms described in Section~\ref{sec:compression}. Considering only the costs that are of the order $\bigO(MN)$, it can be readily observed that simply the recursive SVD methods have better scaling than the `Direct' method which scales as $\bigO(NM^2)$, if the rank at the leaves are smaller than the overall rank of the matrix $r_L < R$. By comparing the worst case costs for the recursive SVD methods, we can see that that the terms linear in $M$ and $N$ are comparable if $r_L \sim R/2^L$. If no compression is observed at any levels, including the leaves, then our algorithm performs poorly since we are needlessly computing a large number of ``low-rank'' factorizations at all the levels in the tree. However, there are still a couple of benefits of using the recursive SVD approach. First, for the range of parameters we are interested in exploring, storing the entire matrix $\bH$ could cost $\sim 200 $ GB which may be completely infeasible to store and later compress. By contrast the strategy in Algorithm~\ref{alg:recursive} does not require storage of $\bH$ in its entirety but computes and compresses sub-blocks of $\bH$ on-the-fly and therefore, has favorable storage costs. Second, our algorithm provides more locality in the calculations and therefore, the algorithm is more amenable to parallelization and distributed computing setting. This has also been noted by~\cite{chaillat2012faims}.
To summarize the asymptotic cost of factorization using RandSVD at the leaf level is $\bigO(r_LMN + r_L^2(N +M))$ and using ppACA it is $\bigO( r_L^2(N + M))$. Numerical evidence suggests that there is compression at every level (and therefore $r_L \ll R$) and this justifies the use of this hierarchical approach.
\subsection{Accuracy}\label{sec:acc} We now discuss the accuracy of the recursive low-rank approximation. In the algorithms described above, there are two sources of error - due to the low-rank truncation at the leaf level and the error accumulated due to the agglomeration process at all other levels in the tree. In order to analyze the accuracy of the recursive SVD computation, we first consider the accuracy of the agglomeration step at one level. We consider the matrix $\bH$ which has a partitioned as $\bH = [\bH_1^T, \bH_2^T]^T $. Suppose we compute a low-rank approximation to $\bH_i \approx \hat{\bH}_i\define \mathbf{U}_i\mathbf{V}_i^T$ for $i=1,2$ using the techniques described in Section~\ref{sec:compression}. We assume that the low-rank matrices satisfy the bounds $\| \bH_i - \hat{\bH}_i\| \leq \varepsilon \| \bH_i\| $ for $i=1,2$. We can then bound the error in the approximation $\bH \approx \hat{\bH}\define \mathbf{U}\mathbf{V}^T$.
\begin{align*}
\norm{ \bH - \hat{\bH} } \quad \leq & \quad \norm{ \hat{\bH} - [\hat{\bH}_1^T,\hat{\bH}_2^T]} + \norm{\bH_1-\hat{\bH}_1} + \norm{\bH_2 - \hat{\bH}_2} \\
\leq & \quad \varepsilon \left(\norm{ [\hat{\bH}_1^T,\hat{\bH}_2^T]^T} \right) + \varepsilon\left(\norm{\bH_1}+\norm{\bH_2}\right) \\
\leq & \quad \varepsilon \left(\norm{ \hat{\bH}_1} + \norm{\hat{\bH}_2} \right) + \varepsilon\left(\norm{\bH_1}+\norm{\bH_2}\right) \\
\leq & \quad (2\varepsilon + \varepsilon^2) \left(\norm{\bH_1}+\norm{\bH_2}\right) = 2\varepsilon\left(\norm{\bH_1}+\norm{\bH_2}\right) + \bigO(\varepsilon^2)
\end{align*}•
We have used the fact that the strategy that is used to truncate the rank for the agglomeration is the same as one to compute the low-rank compression. Furthermore, we have also have used the inequality that
\[ \norm{ \hat{\bH}_i} = \norm{ \hat{\bH}_i - \bH_i + \bH_i } \leq \norm{ \hat{\bH}_i - \bH_i} + \norm{ \bH_i} \leq (1+ \varepsilon )\norm{\bH_i} \]
We now extend it to the case where $N_s > 2$ by recursively applying the error bound that was derived above
\begin{align*}
\norm{ \bH - \hat{\bH} } \quad \leq & \quad \sum_{b \in \mathcal{L}(T_{I})} \norm{\bH_b - \hat{\bH}_b} + \sum_{\ell=0}^{L-1}\sum_{b \in T^{(l)}_{I} } \norm{\hat{\bH}_b -\text{Agg}\{ \hat{\bH}_{b'} : b' \in S(b))\} } \\
\leq & \quad (L+1) \varepsilon \left( \sum_{b \in \mathcal{L}(T_{I})} \norm{\bH_b} \right)+ \bigO(\varepsilon^2)
\end{align*}•
If the computations were performed in the Frobenius norm, then using the Cauchy-Schwarz inequality we can conclude that $ \norm{\bH - \hat{\bH} }_{F} \lesssim 2^{L/2}(L+1)\varepsilon \norm{\bH}_{F}$. In order to derive an equivalent relationship for the $2$-norm, we use the following inequality
\[ \normtwo{\bH - \hat{\bH} } \leq \norm{\bH - \hat{\bH} }_F \lesssim 2^{L/2}(L+1)\varepsilon \norm{\bH}_{F} \leq \sqrt{N_r} 2^{L/2}(L+1)\varepsilon \normtwo{\bH}\]
where $N_r = \min\{M,N\} $. Therefore, in order to achieve a desired relative tolerance $\varepsilon_d $, the tolerance that is used in the low-rank approximation and the agglomeration can be computed as $\varepsilon \sim \varepsilon_d/2^{L/2}(L+1)\sqrt{N_r}$ for the Recursive SVD.
This error bound although locally optimal can result in a low-rank factorization that may be suboptimal in terms of compression. For this reason, we propose an additional step
for compressing the low-rank factors $\hat\bH = \mathbf{U}\mathbf{V}^T$.
\begin{enumerate}
\item Compute thin QR factorizations $\bQ_\mathbf{U}\bR_\mathbf{U} = \mathbf{U}$ and $\bQ_\mathbf{V}\bR_\mathbf{V} = \mathbf{V}$
\item Compute SVD $\bR_\mathbf{U}\bR_\mathbf{V}^T = \mathbf{U}'\boldsymbol\Sigma'(\mathbf{V}')^T$
\item Truncate $R$ singular values and return $\mathbf{U} = \bQ_\mathbf{U}\bU_r'$ and $\mathbf{V} = \bQ_\mathbf{V}\bV_r'\boldsymbol\Sigma_r'$
\end{enumerate}
This additional cost is $\bigO(R^2(M+N) + R^3)$ and may be beneficial when $R \ll \min\{M,N\}$.
The take away is that the cost of storage and matvecs with $\hat\bH$ is $\bigO(R(M+N))$, which is
critical when we need to access and multiply with the estimate repeatedly in the course of the optimization for the image parameters. We now describe that optimization problem.
\section{Reconstruction algorithms}\label{sec:recon}
The recovery of the shape of the tumor and the chromophore concentrations from diffuse optics measurements is an ill-posed inverse problem. The inverse problem can be stated as follows: Given a set of measurements $\by$ that measures the scattered field $\bphi_s$ at multiple detector locations $\br_d$ due to incident field $\bphi_i$ from multiple source locations and illuminated at several different wavelengths, recover the spatially varying perturbation of absorption $\Delta \mu_a(\br,\lambda)$ and the concentration of the chromophore species. Standard approaches to deal with ill-posedness introduce some kind of regularization, such as Tikhonov regularization. Here, we consider the parametric level setup approach proposed in~\cite{aghasi2011parametric} (abbreviated as PaLS) and subsequently applied to the diffuse optical tomography problem in~\cite{larusson2012parametric}.
We briefly review the PaLS approach for parameterizing the shape perturbation. The characteristic function $\chi(\br)$ defined in equation~\eqref{eqn:perturbation} is taken as the $\tau$-level set of a Lipschitz continuous function $\varphi (\br): \mathcal{D} \rightarrow \mathbb{R}$. Using $\varphi(\br)$, the characteristic function $\chi(\br)$ can be expressed as
\begin{equation} \chi(\br) = H\left(\varphi(\br) -\tau \right) \qquad \varphi(\br) = \sum_{k=1}^{n_p} \alpha_k\psi\left(\beta_k\normr{\br-\bchi_k}\right) \label{eqn:pals}
\end{equation}
where $H(\cdot)$ is the Heaviside function. In practice, we use smooth approximations $H_\varepsilon$ of the Heaviside function $H$, and its derivative denoted by $\delta_\varepsilon$. We represent the function $\varphi(\br)$ parametrically as weighted combinations of basis functions $\psi(\cdot)$ and we have $\normr{\br} = \sqrt{\normtwo{\br}^2 + \nu^2}$ and $\nu > 0$ is a small parameter to ensure that $\varphi$ is differentiable. Several choices are available for $\psi$ such as polynomials and radial basis functions. Here we choose the compactly supported radial basis functions that were previously used in~\cite{aghasi2011parametric}. The coefficients $\alpha_k$ control the magnitude of the radial basis functions, $\beta_k$ control the width and $\bchi_k$ control the centers. The basis functions and their number control how fine or coarse the representation will be. On the one hand, having a large number of basis functions will be beneficial in reconstructing fine scale features, however, it has additional associated computational cost and further exacerbates the non-convexity.
The parameters that need to be estimated are collected in a vector $\bp = [\balpha^T,\bbeta^T,\bchi_x^T,\bchi_y^T,\bchi_z^T]^T \in \mathbb{R}^{N_p}$, where $N_p = (d+2)n_p$ and $d$ is the dimension of the problem. The reconstruction problem can now be stated as the minimization of the following functional
\begin{equation}\label{eqn:minimization}
\hat{\bc}, \hat{\bp} \quad \define \quad \argmin_{\bc, \bp} \text{ } \normtwo{\beps}^2 \text{ }= \text{ }\normtwo{\bW(\by - \bD(\bp) \bc)}^2
\end{equation}
where the columns of $\bD(\bp)$ are given by $\bE_i\bH\bmu(\bp)$ and $\bc = [c_1,\dots,c_{\nsp}]^T$ represent the concentration of the chromophores.
\begin{algorithm}[!ht]
\begin{algorithmic}[1]
\STATE Given tolerances $\tau_1$ and $\tau_2$ and initial guess for PaLS parameters $\bp$
\WHILE {$\normtwo{\beps} \leq \tau_1$}
\STATE $\bc = \left(\bW\bD(\bp) \right)^{+} (\bW \by)$
\WHILE {$\normtwo{\beps} \leq \tau_2$}
\STATE\label{step:lm} $(\bJ^T\bJ + \nu \bI) \delta \bp = - \bJ^T \beps $
\\ \COMMENT{//The parameter $\nu$ is chosen by a Levenberg-Marquardt procedure. }
\STATE $\bp \leftarrow \bp + \delta \bp$
\ENDWHILE
\ENDWHILE
\RETURN Shape parameters $\bp$ and chromophore concentration $\bc$
\end{algorithmic}
\caption{Optimization procedure for solving shape parameters $\bp$ and chromophore concentrations $\bc$}
\label{alg:opt}
\end{algorithm}
The resulting optimization problem is solved by alternating between solving for the concentration parameters $\bc$ which is a linear least-squares problem and solving for the PaLS parameters $\bp$ using a Levenberg-Marquardt procedure. The optimization algorithm requires constructing the Jacobian
\[ \bJ = \frac{\partial\beps}{\partial\bp} = - \bW\bar{\bE}\bH\frac{\partial\bmu}{\partial\bp} \]
where $\bar{\bE} = \sum_{i=1}^{\nsp}c_i\bE_i$. Analytical expressions for the derivatives $\frac{\partial\bmu}{\partial\bp}$ are provided in~\cite{aghasi2011parametric}. The stopping criteria for the iterative procedure is chosen according to the discrepancy principle, i.e., the iterations are terminated when the norm of the residuals is less than the noise level up to a user defined constant $\gamma > 1$. In mathematical terms, the stopping criterion becomes $\normtwo{\beps} \leq \gamma \normtwo{\bEta}$ and $\bEta$ is the noise defined in Equation~\eqref{eqn:measurement}. More efficient algorithms are available for the reconstruction of PaLS parameters, for example, see TREGS~\cite{de2011regularized}. However, we have chosen the Levenberg-Marquardt algorithm for its relative simplicity of implementation.
Recall in Section~\ref{sec:compress}, we used a compressed low-rank representation of the measurement operator $\bH \approx \mathbf{U} \mathbf{V}^T$. Let us denote $\hat{\bH} = \mathbf{U}\mathbf{V}^T$ and the error as $\bE_\bH$ in the low-rank truncation process such that $\bH = \hat\bH +\bE_\bH $. From the results in Section~\ref{sec:compress} we know that $\norm{\bE_\bH} \leq \varepsilon \|\bW\bar\bE\|\norm{\bH} $ and define $\bar\varepsilon \define \varepsilon \|\bW\bar\bE\|$. To simplify the theoretical analysis, we rescale $\by \leftarrow \bW$, $\bH \leftarrow \bW\bar\bE\bH$ and $\hat{\bH} \leftarrow \bW\bar\bE\hat\bH$. The approximate Jacobian $\bar{\bJ}$ is now given by the expression $\bar{\bJ} = - \bar{\bH} \frac{\partial\bmu}{\partial\bp}$, so that we have $\bJ = \bar{\bJ} + \bE_\bJ $, where $\bE_\bJ =\bE_\bH\frac{\partial\bmu}{\partial\bp}$. It can be readily shown that the approximation to the objective function $\bbf(\bp) \define \normtwo{\by-\bH\bmu(\bp)}^2$ and the gradient $\nabla_\bp \bbf (\bp) \define -\frac{\partial\bmu}{\partial\bp}^T\bH^T(\by-\bH\bmu(\bp))$ (and the equivalent quantities $\hat\bbf$ and $\nabla_\bp \hat\bbf$ with the approximation $\bar\bH$ instead of $\bH$) satisfy the following approximation bounds
\begin{align*}
|\bbf-\hat{\bbf}| \quad \leq & \quad 2\bar\varepsilon\normtwo{\by-\hat\bH\bmu} \normtwo{\bH}\normtwo{\bmu}+ \bigO(\bar\varepsilon^2) \\
\norm{\nabla_\bp \bbf- \nabla_\bp \hat{\bbf}}\quad \leq & \quad \bar\varepsilon \normtwo{\frac{\partial\bmu}{\partial\bp}} \normtwo{\bH} \left( \normtwo{\bmu}\normtwo{\bH} + \normtwo{\by-\hat\bH\bmu} \right)
\end{align*}
Furthermore, assume that $\norm{\nabla_\bp \bbf- \nabla_\bp \hat{\bbf}} \leq \tau_g \norm{ \nabla_\bp \hat{\bbf}}$, then the acute angle $\theta$ between the gradient $\nabla_\bp \bbf$ and $\nabla_\bp \hat\bbf$ satisfies the following inequality
\[ \cos\theta \geq \frac{1-\tau_g^2}{\sqrt{1 + \tau_g^2}}\]
The result follows from the result in~\cite[Lemma 3.1]{yue2013accelerating}. We assume that the angle $\theta$ is acute, i.e., $\tau_g < 1$ which is always possible since we can control the error $\bar\varepsilon$ and therefore the tolerance $\tau_g$. We now present a result that bounds the error between the true and the perturbed search directions in step~\ref{step:lm} of Algorithm~\ref{alg:opt}.
\begin{propos}
Assume that the derivative $\frac{\partial\bmu}{\partial\bp}$ is full rank and $N_p \leq R$, where $R$ is the effective rank of the low-rank representation $\hat\bH$ and $N_p$ is the number of PaLS parameters. In Algorithm~\ref{alg:opt} let $\delta \bp$ be the search direction corresponding to the exact Jacobian $\bJ$ and let $\delta\bar{\bp}$ be the search direction corresponding to the approximate Jacobian $\bar{\bJ}$. Then, we can bound the error between the two search directions as
\begin{equation}
\normtwo{\delta \bp - \delta\bar{\bp}} \leq \left[ \eta(\nu) \normtwo{\delta\bar{\bp}} + \frac{\normtwo{\beps}}{\nu + \bar{\sigma}_{N_p}^2}\right] \normtwo{\bE_\bJ}
\end{equation}
where the factor $\eta(\nu)= \max_{\bar{\sigma}_{N_p} \leq \sigma \leq \bar{\sigma}_1} \sigma/(\nu + \sigma^2)$ and $\bar\sigma_1$ and $\bar{\sigma}_{N_p}$ are upper and lower bounds for the singular values of the unperturbed Jacobian $\bJ$.
\end{propos}
The proof is readily obtained by an application of the result~\cite[Theorem 3.1]{ipsen2011rank}. The matrix $\frac{\partial\bmu}{\partial\bp}$ is full-rank and the dimension of $\bp$ denoted by $N_p$ is smaller than the rank of $\hat\bH$, therefore the Jacobian $\bJ$ and perturbed Jacobian $\bar{\bJ}$ are full-rank and satisfy the requirements of the theorem. If $R < N_p$ then the Jacobian is rank-deficient and we could consider a subset selection procedure similar to~\cite{ipsen2011rank}.
\section{Numerical Experiments}~\label{sec:results}
We present some results of the algorithms that we described in Section~\ref{sec:krylov}. For the rest of this section, we consider the following test problem. The geometry under consideration is a breast shaped phantom that is compressed between two flat plates (see Figure~\ref{fig:phantom}). At its widest it is 12 cm long and the maximum thickness is 5 cm. The domain is discretized using gmsh~\cite{geuzaine2009gmsh}, an open source 3D finite element mesh generator. The finite element matrices corresponding to the discretized representations of the partial differential equations given by equation~\eqref{eqn:matrices} are computed using FEniCS~\cite{logg2010dolfin} accessed using its Python interface. The boundary $\partial \Omega_R$ for which refractive index mismatch conditions are applied are assumed to be the flat top portions of the boundary, where as zero Dirichlet boundaries are applied on the rest of the boundary.
We consider the background medium to be composed of $\nsp = 4$ species, oxygenated and de-oxygenated hemoglobin denoted as HbO$_2$ and HbR respectively, water H$_2$O and lipids. These species have been specifically chosen since they are the most optically active chromophores, found in breast tissue, in the wavelength range $[600,1000]$ nm. The concentration of the various species in the background have been summarized in Table~\ref{tab:conc}. The extinction coefficients for the species have been found in the literature~\cite{prahl}. We take the value~\cite{grosenick2005timeb} of $\Psi = 9.4$ based on the wavelength of $600$ nm and the prefactor $b=1.4$ in Equation~\eqref{eqn:difflambda}.
\begin{table}[!ht]\centering
\begin{tabular}{|c|c|c|c|c|} \hline
Species & HbO$_2$ &HbR& H$_2$O & Lipids \\ \hline
Units & $\mu$M & $\mu$M & $\%$ & $\%$ \\ \hline
Background & 17 & 7 & 0.15 & 0.6 \\ \hline
Tumor & 25 & 15 & 0.25 & 0.5 \\ \hline
\end{tabular}
\caption{Concentration of different species in the background and the tumor. }
\label{tab:conc}
\end{table}
\subsection{Forward solver}
For the preconditioner we choose an incomplete LU factorization implemented using SuperLU~\cite{lishao10} and we considered the parameters \verb;fill_factor; $\in \{5,10,15\}$ and \verb;drop_tol; $\in\{10^{-3},10^{-4},10^{-5}\}$. We consider the following transformation of the linear systems~\eqref{eqn:multipleshifted}
\[ \left( \bK + \bar{\sigma}'\bR + \sigma_j\bM +( \sigma_j' -\bar{\sigma}')\bR \right)\bx_j = \bb\]
for $j = 1,\dots,N_\lambda$ and $\bar{\sigma}'$ is the mean of $\sigma_j'$. We then define $\bK \leftarrow \bK + \bar{\sigma}'\bR $ and $\sigma_j' \leftarrow \sigma_j' - \bar{\sigma}'$. This transformation essentially the leaves the solution unchanged but improves the convergence of our solver, since the modified matrix $ \bK + \bar{\sigma}'\bR$ contains average information about the refractive index mismatch boundary conditions. Other transformations involving the minimum or maximum over $\sigma_j'$ may also be considered.
\iffalse
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.29]{figs/itercount}
\includegraphics[scale=0.29]{figs/reshist}
\caption{(left) Comparison of the number of iterations taken by GMRES(50) applied to the individual systems and Aug-GMRES described in Section~\ref{sec:krylov}. (right) Residual history of the parametric systems using GMRES (50) (solid black lines) and Aug-GMRES (dotted red lines). All systems were solved till the relative residual was smaller than $10^{-8}$. Significant reduction is obtained from the initial deflation and this results in faster convergence. Only $30$ wavelengths are used here. }
\label{fig:krylov}
\end{figure}
\fi
We now report the results of our solver on a variety of test problem sizes, and preconditioner parameter such as fill factor and drop tolerance. The column labeled `Iter' reports the total number of iterations across $100$ wavelengths, `MVP [s]' reports the CPU time spent on matrix-vector products as well as application of the preconditioner, and finally `Tot. [s]' reports the total CPU run time of Algorithm~\ref{alg:auggmres} (including the pre-computation time for computing the initial guess $\tilde{\bx}_{0,j}$ and the matrices $\mathbf{U}, \bC$ obtained by solving the shift-invariant system and generating the augmented space for each system $\mathbf{U}_j$ and $\bC_j$). From the table it can be seen that by increasing the dimension of the deflation space $k$, the total number of iterations decrease but the cost per iteration increases as a result of extra orthogonalization w.r.t. $\bC_j$. Therefore, there is a trade-off between the number of iterations and total run time and adding additional vectors in the deflation space is a case of diminishing returns. For the range of parameters we experimented with, typically $k=5,10$ produces the best results in terms of total CPU time. However, for problem sizes larger than we are considering, the cost of matrix-vector products may be the dominant cost so that it might be beneficial to use a larger deflation space.
\begin{table}\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|} \hline
\multicolumn{10}{|c|}{Varying problem size, drop tol $=10^{-4}$, fill factor $=10$}\\\hline
\multirow{2}{*}{Dim k} & \multicolumn{3}{|c|}{$N=16,271$} & \multicolumn{3}{|c|}{$N=52,425$} & \multicolumn{3}{|c|}{$N=87,431$}\\ \cline{2-10}
• & Iter & MVP [s] & Tot. [s] & Iter & MVP [s] & Tot. [s] & Iter & MVP [s] & Tot. [s]
\\ \hline
$0$ & $2927$ & $27.90$ & $41.84$ & $3992$ & $146.39$ & $238.11$ & $5426$ & $388.35$ & $740.34$ \\ \hline
$5$ & $2164$ & $18.09$ & $32.84$ & $3028$ & $124.39$ & $232.86$ & $4922$ & $315.79$ & $495.68$\\ \hline
$10$ & $2093$ & $18.94$ & $34.29$ & $2845$ & $116.44$ & $195.03$ & $4768$ & $300.87$& $469.59$ \\ \hline
$15$ & $2072$ & $19.34$ & $35.21$ & $2850$ & $103.59$ & $180.86$ & $4709$ & $311.00$ & $479.408$ \\ \hline
\multicolumn{10}{|c|}{Varying fill factor, $N=52,425$, drop tol; $= 10^{-4}$ } \\ \hline
& \multicolumn{3}{|c|}{fill factor $= 5$} & \multicolumn{3}{|c|}{fill factor $= 10$} & \multicolumn{3}{|c|}{fill factor $= 15$}\\ \hline
$0$ & $5080$ & $129.91$ & $297.57$ & $3992$ & $146.39$ & $238.11$ & $3688$ & $195.37$ & $262.99$ \\ \hline
$5$ & $3710$ & $105.21$ & $222.64$ & $3028$ & $124.39$ & $232.86$ & $2809$ & $165.09$ & $248.91$\\ \hline
$10$ & $3614$ & $98.20$ & $210.50$ & $2845$ & $116.44$ & $195.03$ & $2664$ & $145.67$ & $218.17$ \\ \hline
$15$ & $3729$ & $91.93$ & $197.10$ & $2850$ & $103.59$ & $180.86$ & $2651$ & $131.70$ & $194.64$ \\ \hline
\multicolumn{10}{|c|}{Varying drop tol, fill factor = $10$, $N=52,425$ } \\ \hline
& \multicolumn{3}{|c|}{drop tol $= 10^{-3}$} & \multicolumn{3}{|c|}{drop tol $= 10^{-4}$} & \multicolumn{3}{|c|}{drop tol $= 10^{-5}$}\\ \hline
$0$ & $3485$ & $99.61$ & $179.86$ & $3992$ & $146.39$ & $238.11$ & $5293$ & $193.75$ & $314.37$ \\ \hline
$5$ & $2665$ & $77.21$ & $169.98$ & $3028$ & $124.39$ & $232.86$ & $3859$ & $143.45$ & $248.03$\\ \hline
$10$ & $2541$ & $75.81$ & $161.72$ & $2845$ & $116.44$ & $195.03$ & $3721$ & $155.99$ & $259.73$ \\ \hline
$15$ & $2505$ & $71.76$ & $148.10$ & $2850$ & $103.59$ & $180.86$ & $3744$ & $144.06$ & $228.62$ \\ \hline
\end{tabular}•
\caption{Summary of augmented GMRES solver for different problem sizes, and preconditioner parameters. The number $k$ refers to the dimension of the deflation space. All systems were solved till it converged to a relative tolerance of $10^{-8}$. As can be seen, on average $10-20\%$ improvement was observed for all systems in terms of total computation time, by using deflation. }
\label{tab:timingforward}
\end{table}
We would like to emphasize that care should be taken to interpret the results in Table~\ref{tab:timingforward}. The algorithm has been implemented in Python (which is an interpreted language) and because it uses pre-compiled code for parts of the computation, the timing results may be slightly different if the entire algorithm were implemented in a single programming language. In particular, we expect the overall computation time would be lower if using a compiled language such as C/C++ and the performance gains from our algorithm to be higher.
\subsection{Compression}
We now discuss the results of the compression scheme presented in Section~\ref{sec:compress}. We consider the same geometry that was used in the previous subsection. The sources are placed on the top of the phantom whereas the receivers are placed on the bottom. For each source, there are $9$ detectors constrained to move along with the source that are evenly placed co-axially with the source with $0.5$ cm distance from each other. The number of sources varied from $N_s = 4,\dots,25$ and the number of wavelengths vary between $N_\lambda = 11,\dots,81$. With these parameters, the maximum number of parameters are $18,225$. The discretized grid has $N = 52,425$ degrees of freedom.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.33]{figs/breast_H_source}
\includegraphics[scale=0.33]{figs/breast_H_delta}
\caption{(left) Compression defined as Rank/Full rank of the recursive SVD computed as a function of the level of the tree, with increasing number of sources. Here Rank is defined as the number of columns of $\mathbf{U}$ and Full Rank $=\min\{M,N\} $. (right) The compression ratio $\delta^{\ell}_{\ell+1} \define r_\ell/(2r_{\ell+1})$ as function of the level. As can be seen, on average, there is higher levels of compression at higher levels in the tree. Here, level $0$ is the root of the tree. A tolerance of $10^{-6}$ was used for the truncation. Furthermore, $N_{ds} =9$ and $N_\lambda = 41$.}
\label{fig:source}
\end{figure}
In the examples, we will describe, we have used the randomized SVD for both computing the low-rank factorization at the leaf level and to compute the agglomeration of the low-rank factors as we go up the tree. Similar results are obtained using the partially pivoted Adaptive Cross Approximation and will not be displayed here. We first consider the compression by varying the number of sources and keeping all other parameters fixed. We assume that $N_{ds} =9$ and $N_\lambda = 41$ and a tolerance of $10^{-6}$ was used for truncating the rank of the sub-blocks and the agglomeration. As can be seen, at higher levels in the tree (closer to the root) we observe a higher level of compression because there is a greater redundancy of information globally as opposed to locally. Furthermore, with increasing number of sources we observe a higher level of compression which implies that there is redundancy both in terms of wavelengths and the source-detector positions. The results are displayed in Figure~\ref{fig:source}. We also plot the maximum compression ratio $\delta^{\ell}_{\ell+1} \define r_\ell/(2r_{\ell+1})$ (computed across all the nodes at level $\ell$) which is the ratio of the ranks at level $\ell$ obtained by agglomerating 2 sub-blocks at level $\ell +1$ with ranks $\ell+1$. As can be seen, $\delta^{\ell}_{\ell+1} < 1$ at all levels indicating that there is compression, not only at the leaves, but compression at every level in the tree. This justifies using a hierarchical compression scheme and the cost analysis performed in Section~\ref{sec:costs}.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.33]{figs/breast_H_tol}
\includegraphics[scale=0.33]{figs/breast_H_lambda}
\caption{(left) Compression defined as Rank/Full rank of the recursive SVD computed as a function of the level of the tree, with varying tolerance used to compress the low-rank factors. Here $N_{ds}=9$, $N_\lambda = 41$ and $N_s = 9$. (right) The compression of the recursive SVD computed as a function of the level of the tree, with varying $N_\lambda$. Here $N_{ds}=9$, tol = $10^{-6}$ and $N_s = 9$. }
\label{fig:tollambda}
\end{figure}
Next we compute the compression as a function of tree level with varying tolerance used to truncate the ranks of the sub-blocks. The results are presented in Figure~\ref{fig:tollambda}. As can be seen with a higher tolerance the ranks at each level decreases dramatically. Here, we have fixed $N_{ds}=9$, $N_\lambda = 41$ and $N_s = 9$. We also present results of computing the compression as a function of tree level with varying number of wavelengths used to illuminate the object. All other parameters are fixed as $N_{ds}=9$, tol = $10^{-6}$ and $N_s = 9$. The results are also presented in Figure~\ref{fig:tollambda}. We can see that with increasing number of wavelengths there is a higher level of compression at higher levels in the tree.
Finally, we compare the run time of the recursive SVD algorithm proposed in Section~\ref{sec:compress} with RandSVD applied to the entire measurement operator. As can be seen, the computational time for the recursive algorithm is far lower and is therefore, more efficient.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.3]{figs/breast_recur_timing_coarse}
\caption{Comparison of time taken to factorize the Born operator $\bH$ to a tolerance of $10^{-6}$ with $N_s $ ranging from $4$ to $25$ and $N_{ds}$ fixed at 9. We compare the CPU timing of RandSVD in Algorithm~\ref{alg:randsvd} applied to the entire matrix directly with the Recursive SVD algorithm proposed in Section~\ref{sec:compress}. The time for constructing $\bH$ is not included. The grid size was $16,721$.}
\label{fig:comptime}
\end{figure}
\subsection{Reconstruction results}
In the following experiments we will use the following metrics to measure the error in the shape perturbation. In Experiment $1$ and $2$ we use the full Born operator to generate the measurements, however, we use the compressed operator in the reconstructions. In Experiment $3$ we use measurements from the full diffusion equation and use the compressed Born model in the reconstruction. As a result, we avoid Let $\bmu$ denote the true shape perturbation and let $\hat\bmu$ denote the recovered shape perturbation. The first is standard relative L$_2$ error $\normtwo{\bmu-\hat\bmu}/\normtwo{\bmu}$. For piecewise constant medium, a different metric known as the Dice coefficient (see for example,~\cite{larusson2012parametric}) provides more information about localization of the perturbation. It can be defined as
\[ D(\bmu,\hat\bmu) \define 2 \frac{| \bmu \bigcap \hat\bmu|}{|\bmu| + |\hat\bmu|} \]
where $| \bmu \bigcap \hat\bmu|$ corresponds to the number of non-zero pixels that are contained in both the true shape $\bmu$ and the reconstructed shape $\hat\bmu$ whereas $|\bmu|$ and $|\hat\bmu|$ correspond to the number of non-zero pixels in the true and the reconstructed shapes respectively. In order to If the object is recovered fully, then the Dice coefficient would be $1$.
\textbf{Experiment 1}: In this experiment we study how the accuracy of the low-rank representation for $\bH$ affects the error in the reconstruction. We consider $N_s=4$ sources and $N_{ds} = 9$ detectors/source and $N_\lambda=25$ wavelengths totaling $900$ measurements generated using the Born model but with the full measurement operator $\bH$. We only consider reconstruction of the shape perturbation and consider the concentration of the chromophores as known and provided in Table~\ref{tab:conc}. Gaussian noise is added such that the signal-to-noise (SNR) ratio is $33$ dB to simulate observational noise. The SNR is defined as
\[ \text{SNR} = 20\log_{10} \frac{\normtwo{\by}}{\normtwo{\boldsymbol\eta}}\]
where the noise $\bEta$ has been defined in Equation~\eqref{eqn:measurement}. The grid size is chosen to be $52,425$. A ``true'' shape perturbation is obtained by using three randomly generated basis functions which leads to $N_p = 15$. We report the rank of the measurement operator as a function of the global tolerance used for low-rank representation and the error in the reconstruction using the low-rank operator $\hat\bH$. The results are reported in Table~\ref{tab:exp1}. As can be seen that for very small tolerance, the error in the reconstruction is hardly noticeable. Therefore, a larger rank of the low-rank representation $\hat\bH$ does not affect the reconstruction error below a certain tolerance and therefore the compressed operator $\hat\bH$ can be used as a surrogate for the full matrix $\bH$ with little or no loss in accuracy in the reconstruction. When the number of measurements is large, this can represent significant savings in computational time.
\begin{table}[!ht]\centering
\begin{tabular}{|c|c|c|c|c|} \hline
tol & $10^{-3}$ & $10^{-6}$ & $10^{-9}$\\ \hline
Rank & 48& 120& 288 \\ \hline
L$_2$err. & $40.62 \% $ & $40.52 \%$& $40.52 \%$ \\\hline
Dice &$0.82$ & $0.827$ & $0.827$ \\ \hline
Time [s]& $2.20$ & $2.23$ & $2.32$ \\ \hline
\end{tabular}
\caption{Rank of the compressed operator and the error in the reconstruction of the shape perturbation as a function of tolerance used to compress $\bH$. See experiment 1 for more details. `Time [s]' indicates the CPU time of solving the optimization problem with the compressed operator. See Experiment 1 for more details.}
\label{tab:exp1}
\end{table}
\textbf{Experiment 2}: In this experiment we study the reconstruction of chromophore concentrations as well as the shape perturbation. The number of measurements and the process of generating them is the same as Experiment 1 except with SNR $30$ dB. For the reconstruction, the radial basis functions were randomly initialized and a truncated measurement operator $\hat\bH$ computed using tolerance $10^{-6}$ was used in the reconstruction. The error of the reconstruction of the chromophore concentrations as well the relative L$_2$ error of the shape perturbation are reported in Table~\ref{tab:exp2}. As can be seen from the Table, the added difficulty in recovering the shape perturbation as well as the chromophore concentrations affects the reconstruction error of the shape perturbation slightly. Moreover, the concentrations of the chromophore species are recovered fairly accurately. This is consistent with the observations in~\cite{larusson2013parametric,larusson2011hyperspectral,larusson2012parametric}.
\begin{table}[!ht]\centering
\begin{tabular}{|c|c|c|c|c|c|c|} \hline
Species & HbO$_2$ &HbR& H$_2$O & Lipids & L$_2$ err. & Dice \\ \hline
Recon. & $1.6 \%$ & $0.2 \%$ & $4.9 \%$ & $4.4\%$ & $52.72 \%$ & $0.79$ \\\hline
\end{tabular}
\caption{Error in the reconstruction of different species and the shape perturbation corresponding to Experiment 2. }
\label{tab:exp2}
\end{table}
\begin{figure}
\centering
\includegraphics[scale=0.182]{figs/true_exp}
\includegraphics[scale=0.177]{figs/recon_exp2}
\caption{Reconstruction in Experiment 2 (left) true anomaly and (right) reconstruction. The error is described in Table~\ref{tab:exp2}.}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.3]{figs/meas_comp}
\includegraphics[scale=0.22]{figs/recon_exp3}
\caption{(left) Comparison of the measurements generated using the Born model and that generated from the full diffusion equation. As can be seen, the agreement between the measurements is quite close. (right) Reconstruction of the shape perturbation using measurements from full diffusion equation. The errors are reported in Table~\ref{tab:exp3}.}
\label{fig:exp3}
\end{figure}
\textbf{Experiment 3}: In this example we examine the validity of the Born approximation. As mentioned earlier, experimental studies have validated the utility of the Born approximation (see for e.g.~\cite{larusson2012parametric}). However, in order to prevent committing an ``inverse crime'' we use data generated from the full diffusion equation, Equation~\eqref{eqn:diffusion} solving using finite elements with the same boundary conditions and use it to reconstruct both the chromophore concentrations and the shape parameter. The comparison between the measurements generated using the Born model and those generated from the full diffusion equation is provided in Figure~\ref{fig:exp3}. In addition a noise of $50$ dB was added to simulate observational error in realistic conditions. All other settings were the same as that in Experiment 2. The error of the reconstruction of the chromophore concentrations as well the L$_2$ error of the shape perturbation are reported in Table~\ref{tab:exp3}. The error in the reconstruction is higher than that obtained from Experiment 2. Since the full diffusion equation was used in generating the measurements, there is a modeling error which corresponds to about $20$ dB SNR and because of this, the Born model cannot exactly recover the shape perturbation. However, the reconstruction is still satisfactory as can be seen from the Figure~\ref{fig:exp3}.
\begin{table}[!ht]\centering
\begin{tabular}{|c|c|c|c|c|c|c|} \hline
Species & HbO$_2$ &HbR& H$_2$O & Lipids & L$_2$ err. & Dice \\ \hline
Recon. & $5.1 \%$ & $ 1.2 \%$ & $1.2 \%$ & $9.3\%$ & $71.80 \%$ & 0.65\\\hline
\end{tabular}
\caption{Error in the reconstruction of different species and the shape perturbation corresponding to Experiment 3. Measurements were generated from the full diffusion equation, i.e., Equation~\eqref{eqn:diffusion}. See also Figure~\ref{fig:exp3}.}
\label{tab:exp3}
\end{table}
\section{Conclusions and future work}
We have presented a fast algorithm for recovering shape of the perturbations and chromophore concentrations that is composed of three parts - a fast Krylov subspace approach for accelerating the solution of the incident and the adjoint field for multiple wavelengths, constructing a low-rank approximation to the sensitivity matrix $\bH$ using an approach that accounts for redundancies across wavelengths which is then recursively combined across source-detectors pairs, and finally solving an optimization problem for recovering parameters with the low-rank approximation to $\bH$. The results indicate that significant gains can be obtained both in terms of computational costs and storage requirements. We provide detailed numerical experiments that validates our claims and further provide a detailed analysis of the computational costs and the error. The algorithms were demonstrated on a challenging synthetic inversion case on a complex geometry which highlights the flexibility of our algorithms.
Future work includes extension of our algorithms to the fully nonlinear setting. In this setting, computation of the incident $\phi_i$ and adjoint fields $\phi_d$ and the construction of the measurement operator $\bH$ for the Born approximation, must be performed at every Newton or Gauss-Newton step. We therefore, believe that adopting the ideas proposed in this paper will be even more beneficial in the fully nonlinear case. Other possible extensions include a recycling strategy of the kind described in Section~\ref{sec:krylov} to multiple wavelengths and multiple right hand sides and to the full diffusion equation, Equation~\eqref{eqn:diffusion}. Additional work is currently underway in our lab to compare the reconstructions obtained the synthetic case with real data obtained from experiments. Based on previous work~\cite{larusson2013parametric,larusson2011hyperspectral,larusson2012parametric}, we anticipate that the reconstructions will indeed be excellent even in the hyperspectral case.
\section{Acknowledgements}
We would like to thank Nishanth Krishnamurthy, Pami G. Anderson, Jana Kainerstorfer, and Angelo Sassaroli for useful discussions. The first author would also like to thank Tania Bakhos with her help in generating the meshes. This work was supported by NIH Grant R01-CA154774. Additionally, the second author was supported by NSF Grant DMS 1217161.
|
\section{Introduction}
In this paper, we study the invariance properties of periodic cyclic homology under deformations of the algebra structure. Given a family of algebras $\{A_t\}_{t \in J}$ parametrized by a real number $t$, we would like to identify conditions under which we can conclude
\[ HP_\bullet(A_t) \cong HP_\bullet(A_s) \qquad \forall t,s \in J. \]
The types of algebras we consider will be topological algebras, and the deformations will have a smooth dependence on $t$.
In the world of formal deformations, Getzler constructed a connection on the periodic cyclic complex of a deformation \cite{MR1261901}. His connection, called the \emph{Gauss-Manin connection}, commutes with the boundary map on the periodic cyclic complex and descends to a flat connection on the periodic cyclic homology of the deformation. Our goal is to adapt Getzler's connection to our setting of smooth deformations and investigate its properties.
For a real interval $J \subseteq \R$, we consider a smoothly varying family $\{m_t\}_{t \in J}$ of jointly continuous associative multiplications on a locally convex vector space $X$. For each $t \in J$, we have a locally convex algebra $A_t$ whose underlying space is $X$ and whose multiplication is given by $m_t$. These algebras can be collected to form the algebra $A_J$ of smooth sections of the bundle of algebras over $J$ whose fiber at $t \in J$ is $A_t$, where the multiplication in $A_J$ is defined fiberwise. Then $A_J$ is an algebra over $C^\infty(J)$, the space of smooth complex-valued functions defined on the parameter space $J$, where the module action is given by fiberwise scalar multiplication.
The complex of interest to us is the periodic cyclic complex of $A_J$ over the ground ring $C^\infty(J)$. This can be thought of as the space of smooth sections of the bundle of chain complexes over $J$ whose fiber at $t \in J$ is the periodic cyclic complex of $A_t$. It is on this complex that we shall define Getzler's Gauss-Manin connection $\nabla_{GM}$. The connection $\nabla_{GM}$ commutes with the boundary map and thus descends to a connection on the $C^\infty(J)$-linear periodic cyclic homology $HP_\bullet(A_J)$.
The fundamental issue for us is to determine when we can parallel tranport with respect to $\nabla_{GM}$ at the level of periodic cyclic homology. Indeed, doing so would provide isomorphisms $HP_\bullet(A_t) \cong HP_\bullet(A_s)$ between the periodic cyclic homology groups of any two algebras in the deformation. This can be used as a computational device if one already knows the cyclic homology of one particular algebra $A_{t_0}$ in the deformation. Of course, the striking degree of generality for which $\nabla_{GM}$ exists is an indication that any attempt to integrate $\nabla_{GM}$ will fail generally. The goal then is to identify properties of a deformation that allow for parallel translation.
Our main result is a rigidity result for periodic cyclic cohomology of a certain class of Banach algebras. The weak bidimension $\dbw A$ of a Banach algebra $A$ is the smallest integer $n$ such that the Hochschild cohomology $H^{n+1}(A, M^*)$ vanishes for all Banach $A$-bimodules $M$. A Banach algebra $A$ is called amenable if $\dbw A = 0$. This class was defined and studied by Johnson \cite{MR0374934}. If $\dbw A = n$, then $A$ is also called $(n+1)$-amenable. We prove that the Gauss-Manin connection is integrable for small enough deformations of a Banach algebra of finite weak bidimension. Consequently, periodic cyclic cohomology is preserved under such deformations.
A general feature of $\nabla_{GM}$ is the fact that
\[ \nabla_{GM}[\ch P] = 0, \]
where $[\ch P]$ denotes the class in $HP_0(A_J)$ of the Chern character of an idempotent $P$ in the algebra $M_N(A_J)$ of $N \times N$ matrices over $A_J$.
One can also define a dual Gauss-Manin connection $\nabla^{GM}$ on the periodic cyclic cohomology $HP^\bullet(A_J)$ over $C^\infty(J)$. The connections are compatible in the sense that for $[\phi] \in HP^\bullet(A_J)$ and $[\omega] \in HP_\bullet(A_J)$,
\[ \frac{d}{dt} \langle \phi, \omega \rangle = \langle \nabla^{GM}\phi, \omega \rangle + \langle \phi, \nabla_{GM} \omega \rangle, \] where
\[ \langle \cdot, \cdot \rangle: HP^\bullet(A_J) \times HP_\bullet(A_J) \to C^\infty(J) \]
is the canonical pairing. Combined with the above result, this says
\[ \frac{d}{dt} \langle \phi, \ch P \rangle = \langle \nabla^{GM}\phi, \ch P \rangle \]
for an idempotent $P \in M_N(A_J)$. This offers insight into how the pairing between $K$-theory and cyclic cohomology deforms as the algebra deforms.
In a subsequent paper \cite{Yashinski}, we prove the integrability of the Gauss-Manin connection for the deformation of smooth noncommutative tori. We also use the compatibility of the Gauss-Manin connection with the Chern character to prove differentiation formulas for the pairings of cyclic cocycles with $K$-theory classes. Similar work was carried out independently by Yamashita \cite{Yamashita}.
The outline of the paper is as follows. In \textsection 2, we cover necessary background on locally convex topological vector spaces, Hochschild and cyclic homology, and Getzler's Cartan homotopy formula for the action of Hochschild cochains on the periodic cyclic complex. In \textsection 3, we lay the foundation for our study of deformations by studying properties of modules of the form $C^\infty(J,X)$ for some locally convex vector space $X$. Our main techniques for studying deformations use connections and parallel translation, which are discussed in \textsection 4. In \textsection 5, we define what we mean by a smooth deformation of either algebras or chain complexes, and give a criterion for triviality of these deformations in terms of integrable connections. We use our methods to prove some known rigidity results in \textsection 6. In \textsection 7, we define Getzler's Gauss-Manin connection in our setting of smooth deformations, and prove some of its basic properties. The main theorem on Banach algebras of finite weak bidimension is proved in \textsection 8.
\subsection*{Acknowledgements}
I'd like to thank my thesis advisor, Nigel Higson, for suggesting this line of research and offering useful guidance throughout the project. I have benefitted from discussions with Erik Guentner and Rufus Willett on the material. I'd also like to thank Rufus Willett for reading an earlier draft of this document and suggesting ways to improve it.
\section{Preliminaries}
In this section, we shall first quickly review the relevant concepts from the theory of locally convex topological vector spaces. Then we'll discuss Hochschild and cyclic homology. In particular, we'll review the action of Hochschild cochains on the cyclic chain complex and the corresponding Cartan homotopy formula.
\subsection{Locally convex algebras and modules}
We shall work in the category $\LCTVS$ of complete, Hausdorff locally convex topological vector spaces over $\C$ and continuous linear maps. We shall write $X \in \LCTVS$ to mean that $X$ is a complete, Hausdorff locally convex topological vector space. See \cite{MR0225131} for background in the theory of locally convex topological vector spaces. More details concerning topological tensor products can be found in \cites{MR0075539, MR0225131, MR1093462}. We shall rapidly review the necessary facts below.
Recall that $X \in \LCTVS$ is a Fr\'{e}chet space if the topology on $X$ is metrizable. This is equivalent to the topology on $X$ being defined by a countable family of seminorms. Among the Fr\'{e}chet spaces are the Banach spaces, whose topology is defined by a single norm.
Given $X \in \LCTVS$ and a subspace $Y \subset X$, there is a naturally locally convex topology on the quotient space $X/Y$, which is Hausdorff if and only if $Y$ is closed. Even if $Y$ is closed, the quotient $X/Y$ may not be complete. However if $X$ is a Fr\'{e}chet space and $Y$ is a closed subspace, then $X/Y$ is complete and therefore is also a Fr\'{e}chet space.
Given $X, Y \in \LCTVS$, the space $\Hom(X,Y)$ of continuous linear maps from $X$ to $Y$ is a Hausdorff locally convex topological vector space under the topology of uniform convergence on bounded subsets of $X$. Recall that a subset $A \subset X$ is bounded if and only if $\sup_{x \in A} p(x) < \infty$ for each continuous seminorm $p$ on $X$. If $X$ has the additional property of being bornological, then $\Hom(X,Y)$ is complete. Examples of bornological spaces include Fr\'{e}chet spaces and $LF$-spaces, which are strict, countable inductive limits of Fr\'{e}chet spaces. A special case of interest is the \emph{strong dual} $X^* = \Hom(X, \C).$ We remark that the strong dual of a Banach space is a Banach space, but the strong dual of a Fr\'{e}chet space is never a Fr\'{e}chet space, unless the original space is actually a Banach space. One can also consider $\Hom(X,Y)$ with the weaker topology of pointwise convergence, which we denote $\Hom_\sigma(X,Y)$.
Nuclearity (in the sense of Grothendieck) is a nice technical property that a space $X \in \LCTVS$ can have. We shall occasionally need to reference it, but we shall not work with the concept directly. For more details, see \cite{MR0075539} or \cite{MR0225131}. We remark that this should not be confused with the notion of nuclearity for $C^*$-algebras. For example, a Banach space is nuclear if and only if it is finite-dimensional.
The bilinear maps appearing in structures in this paper will be assumed to be jointly continuous. This naturally leads to the projective tensor product of two spaces in $\LCTVS$. If one wishes to consider separately continuous bilinear maps, one should use the inductive tensor product instead, see \cite{MR0075539}. In the cases of Fr\'{e}chet spaces, the notions of joint and separate continuity coincide, and therefore so do these two tensor products.
Given $X, Y \in \LCTVS$, the \emph{projective topology} on $X \otimes Y$ is the strongest locally convex topology such that the canonical bilinear map $\iota: X \times Y \to X \otimes Y$ is jointly continuous, see \cite[Chapter~43]{MR0225131} for more details and an explicit construction. The \emph{(completed) projective tensor product} $X \potimes Y$ is the completion of $X \otimes Y$ with the projective topology. The completed projective tensor product has the universal property that any jointly continuous bilinear map $B$ from $X \times Y$ into a space $Z \in \LCTVS$ induces a unique continuous linear map $\widehat{B}: X \potimes Y \to Z$ such that the diagram
\[ \xymatrix{
X \times Y \ar[r]^-\iota \ar[rd]^-B & X \potimes Y \ar[d]^-{\widehat{B}}\\
& Z
} \]
commutes. The projective tensor product is functorial in the sense that two continuous linear maps $F: X_1 \to X_2$ and $G: Y_1 \to Y_2$ induce a continuous linear map \[ F \otimes G: X_1 \potimes Y_1 \to X_2 \potimes Y_2 \] given on elementary tensors by
\[ (F \otimes G)(x \otimes y) = F(x) \otimes G(y). \] If $X$ and $Y$ are Banach (resp. Fr\'{e}chet) spaces, then $X \potimes Y$ is a Banach (resp. Fr\'{e}chet) space.
In the language of category theory, the tensor product $\potimes$ makes $\LCTVS$ into a symmetric monoidal category. In any such category, one can define a notion of algebra and module. For $\LCTVS$, these notions agree with the definitions below.
By a \emph{locally convex algebra},
we mean a space $A \in \LCTVS$ equipped with a jointly continuous associative multiplication. Notice we are implicitly assuming $A$ is complete. The multiplication induces a continuous linear map $m: A \potimes A \to A$. Joint continuity implies that for every defining seminorm $p$ on $A$, there is another continuous seminorm $q$ such that
\[ p(ab) \leq q(a)q(b), \qquad \forall a,b \in A. \] There may be no relationship between $p$ and $q$ in general. In the special case where
\[ p(ab) \leq p(a)p(b), \qquad \forall a,b \in A, \] for all seminorms in a family defining the topology, the algebra is called \emph{multiplicatively convex} or \emph{$m$-convex}. An $m$-convex algebra can be expressed as a projective limit of Banach algebras. A \emph{Fr\'{e}chet algebra} is a locally convex algebra whose underlying space is a Fr\'{e}chet space. We do not insist that a Fr\'{e}chet algebra is $m$-convex, as some authors do.
Now suppose $R$ is a unital, commutative locally convex algebra. By a \emph{locally convex $R$-module}, we mean a space $M \in \LCTVS$ equipped with a jointly continuous $R$-module structure. Such a module action induces a continuous linear map $\mu: R \potimes M \to M$. All such modules will be assumed to be unital in the sense that $1 \cdot m = m$ for all $m \in M$.
Given two locally convex $R$-modules $M$ and $N$, we topologize $\Hom_R(M,N)$ as a subspace of $\Hom_\C(M,N)$. When $N = R$, we obtain the topological $R$-linear dual of $M$
\[ M^\dual := \Hom_R(M,R). \]
We shall use the notation $M^\dual$ to distinguish from $M^*$, which will always mean the usual $\C$-linear topological dual space.
We shall need to take topological tensor products over an algebra different from $\C$. The basic facts we need are below, but a more detailed exposition can be found in \cite[Chapter II]{MR1093462}. Suppose $R$ is a unital commutative locally convex algebra and let $M$ and $N$ be locally convex $R$-modules.
The \emph{(completed) projective tensor product over $R$} of $M$ and $N$, denoted $M \potimes_R N$, is a locally convex $R$-module together with a jointly continuous $R$-bilinear map $\iota: M \times N \to M \potimes_R N$ which is universal in the sense that any jointly continuous $R$-bilinear map $B$ from $M \times N$ into a locally convex $R$-module $P$ induces a unique continuous $R$-linear map $\widehat{B}: M \potimes_R N \to P$ such that the diagram
\[ \xymatrix{
M \times N \ar[r]^-\iota \ar[rd]^-B & M \potimes_R N \ar[d]^-{\widehat{B}}\\
& P
} \]
commutes. The module $M \potimes_R N$ can be constructed as the completion
of $(M \potimes_\C N)/K$, where $K$ is the closure of the subspace spanned by elements of the form
\[ (r\cdot m) \otimes n - m \otimes (r\cdot n), \qquad r \in R, m \in M, n \in N. \] Any two continuous $R$-linear maps
$F: M_1 \to N_1$ and $G: M_2 \to N_2$ induce a continuous $R$-linear map
\[ F \otimes G: M_1 \potimes_R N_1 \to M_2 \potimes_R N_2 \] in the usual way. From the construction, we see that if $M$ and $N$ are Banach (resp. Fr\'{e}chet) modules, then $M \potimes_R N$ is a Banach (resp. Fr\'{e}chet) module.
A \emph{locally convex $R$-algebra} is a locally convex $R$-module $A$ equipped with a jointly continuous associative $R$-linear product $m$, so that $m$ induces a continuous $R$-linear map
\[ m: A \potimes_R A \to A. \]
A locally convex $R$-module is \emph{free} if it is isomorphic to $R \potimes_\C X$ for some $X \in \LCTVS$. Here the $R$-module action is induced by
\[ r\cdot(s \otimes x) = rs \otimes x. \]
The free module $R \potimes X$ has the universal property that any continuous $\C$-linear map $\widetilde{F}$ from $X$ into a locally convex $R$-module $M$ induces a unique continuous $R$-linear map $F: R \potimes X \to M$ such that the diagram
\[ \xymatrix{
X \ar[r]^-\iota \ar[rd]^-{\widetilde{F}} & R \potimes X \ar[d]^-F\\
& M
} \]
commutes. Here, $\iota(x) = 1 \otimes x$ and $F$ is given by
\[ F(r \otimes x) = r\cdot \widetilde{F}(x). \]
This establishes a linear isomorphism
\[ \Hom_R(R \potimes X, M) \cong \Hom(X, M). \]
\begin{proposition} \label{Proposition-FreeModuleUMPTopologicalIsomorphism}
If $X$ and $R$ are Fr\'{e}chet spaces, one of which is nuclear, then the linear isomorphism \[ \Hom_R(R \potimes X, M) \cong \Hom(X, M) \] is a topological isomorphism.
\end{proposition}
\begin{proof}
The linear isomorphisms
\[ \Phi: \Hom(X,M) \to \Hom_R(R \potimes X, M), \qquad \Psi: \Hom_R(R \potimes X, M) \to \Hom(X,M) \] are given by
\[ \Phi(F) = \mu (1 \otimes F), \qquad \Psi(G)(x) = G(1 \otimes x), \] where $\mu: R \potimes M \to M$ is the module action. Continuity of $\Psi$ follows from the fact that if $B \subset X$ is bounded, then $1 \otimes B \subset R \potimes X$ is bounded.
The continuity of $\Phi$ is more subtle. The map $\Phi$ factors as
\[ \xymatrix{
\Phi: \Hom(X,M) \ar[r]^-{\Phi_1} &\Hom_R(R \potimes X, R \potimes M) \ar[r]^-{\Phi_2} &\Hom_R(R \potimes X, M),
} \]
where $\Phi_1(F) = 1 \otimes F$ and $\Phi_2$ is composition with the module action $\mu$. Continuity of $\Phi_2$ follows from continuity of $\mu$. To show that $\Phi_1$ is continuous, we need to relate the bounded subsets of $R \potimes X$ to the bounded subsets of $R$ and $X$. This is related to the difficult ``probl\`{e}me des topologies" of Grothendieck \cite{MR0075539}. If either $R$ or $X$ are nuclear, then for every bounded subset $D \subset R \potimes X$, there are bounded subsets $A \subset R$, $B \subset X$ such that $D$ is contained in the closed convex hull of
\[ A \otimes B = \{ r \otimes x ~|~ r \in A, x \in B \}, \] see \cite[Theorem 21.5.8]{MR632257}. Continuity of $\Phi_1$ follows from this fact.
\end{proof}
\begin{proposition} \label{Proposition-TensorProductOfFreeModules}
Given $X, Y \in \LCTVS$,
\[ (R \potimes_\C X) \potimes_R (R \potimes_\C Y) \cong R \potimes_\C (X \potimes_\C Y) \]
as locally convex $R$-modules via the correspondence
\[ (r_1 \otimes x) \otimes (r_2 \otimes y) \longleftrightarrow r_1r_2 \otimes (x \otimes y). \]
\end{proposition}
This can be proved using the universal properties of both modules. This shows that the projective tensor product of free modules is free.
\begin{proposition} \label{Proposition-QuotientOfFreeModules}
Let $R$ be a nuclear Fr\'{e}chet algebra. Given a Fr\'{e}chet space $X$ and a closed subspace $Y \subset X$,
\[ (R \potimes X) / (R \potimes Y) \cong R \potimes (X / Y) \]
as Fr\'{e}chet $R$-modules via the correspondence
\[ [r \otimes x] \longleftrightarrow r \otimes [x]. \]
\end{proposition}
\begin{proof}
Nuclearity of $R$ implies that $R \potimes Y$ is a closed subspace of $R \potimes X$ \cite[Proposition 43.7]{MR0225131}. Since all spaces are Fr\'{e}chet, all quotients appearing are complete. One then induces mutually inverse isomorphisms using the universal properties of completed projective tensor products and quotients.
\end{proof}
By a \emph{locally convex cochain complex}, we mean a collection of spaces $\{ \mathcal{C}^n \}_{n \in \Z}$ in $\LCTVS$ and continuous linear maps $\{ d^n: \mathcal{C}^n \to \mathcal{C}^{n+1}\}_{n \in \Z}$ such that $d^{n+1} \circ d^n = 0$. We'll use the notation $Z^n(\mathcal{C}) = \ker d^n$ for cocycles and $B^n(\mathcal{C}) = \im d^{n-1}$ for coboundaries. The cohomology is $H^n(\mathcal{C}) = Z^n(\mathcal{C}) / B^n(\mathcal{C})$, which may not be Hausdorff or complete. We will often drop the superscript $n$ on the coboundary map. By turning the arrows around, we obtain the definition of a locally convex chain complex.
If each $\mathcal{C}^n$ is a locally convex $R$-module and the coboundary maps are $R$-linear, then $\mathcal{C}^\bullet$ is a \emph{locally convex cochain complex of $R$-modules}. In this case, the cohomology spaces are $R$-modules.
\subsection{Hochschild and cyclic homology for locally convex algebras}
A good reference for Hochschild and cyclic homology is \cite{MR1600246}.
Let $R$ be a unital commutative locally convex algebra and let $A$ be a (possibly nonunital) locally convex $R$-algebra. The main examples for us will be $R = \C$ and $R=C^\infty(J)$, the smooth functions on a real interval $J$. All homology theories that follow are the continuous versions of the usual $R$-linear algebraic theories, in that they take into account the topology of the algebra $A$.
Recall that the \emph{unitization} of the algebra $A$ is the algebra
\[ A_+ = A \oplus R \] with multiplication
\[ (a_1, r_1)(a_2, r_2) = (a_1a_2 + r_2\cdot a_1 + r_1\cdot a_2, r_1r_2). \] Then $A_+$ is a unital locally convex $R$-algebra with unit $(0,1)$, which contains $A$ as a closed ideal. We can, and will, form the unitization in the case where $A$ is already unital. We shall let $e \in A_+$ denote the unit of $A_+$, to avoid possible confusion with the original unit of $A$, if it exists.
\subsubsection{Hochschild cochains}
Let $C^k(A,A)$ denote the space of all jointly continuous $k$-multilinear (over $R$) maps $D: A^{\times k} \to A$.
The coboundary map $\delta: C^k(A,A) \to C^{k+1}(A,A)$ is given by
\begin{align*}
\delta D(a_1, \ldots, a_{k+1}) &= D(a_1, \ldots , a_k)a_{k+1} + (-1)^{k+1} a_1D(a_2, \ldots, a_{k+1})\\ & \qquad + \sum_{j=1}^k (-1)^{k-j+1}D(a_1, \ldots , a_{j-1}, a_ja_{j+1}, a_{j+2}, \ldots , a_k),
\end{align*} and satisfies $\delta^2 = 0$.
The cohomology of $(C^{\bullet}(A,A), \delta)$ is the \emph{Hochschild cohomology of $A$ (with coefficients in $A$)}, and is denoted by $H^\bullet(A,A)$. If we wish to emphasize the ground ring $R$, we shall write $H^\bullet_R(A,A).$
There is much additional structure on $C^\bullet(A,A)$, including a cup product and a Lie bracket, called the \emph{Gerstenhaber bracket}. The shifted complex $\g^\bullet(A) := C^{\bullet+1}(A,A)$ is a differential graded Lie algebra under the Gerstenhaber bracket \cite{MR0161898}. This gives the cohomology $H^{\bullet+1}(A,A)$ the structure of a graded Lie algebra.
\subsubsection{Hochschild homology}
For $n \geq 0$, the space of \emph{Hochschild $n$-chains} is defined to be
\[
C_n(A) = \begin{cases} A, &n=0\\ A_+ \potimes_R A^{\potimes_R n}, &n \geq 1
\end{cases} \]
The boundary map $b: C_n(A) \to C_{n-1}(A)$ is given on elementary tensors by
\begin{align*}
b(a_0 \otimes \ldots \otimes a_n) &= \sum_{j=0}^{n-1} (-1)^j a_0 \otimes \ldots \otimes a_{j-1} \otimes a_ja_{j+1} \otimes a_{j+2} \otimes \ldots \otimes a_n\\
& \qquad + (-1)^n a_na_0 \otimes a_1 \otimes \ldots \otimes a_{n-1}.\\
\end{align*}
More rigorously, $b$ is induced by the functoriality of the projective tensor product $\potimes_R$ using the continuous multiplication map $m: A \potimes_R A \to A$. This shows that $b$ is continuous. Associativity of $m$ implies that $b^2 = 0$. The homology of the complex $(C_\bullet(A), b)$ is called the \emph{Hochschild homology of $A$ (with coefficients in $A_+$)} and shall be denoted $HH_\bullet(A)$ or $HH_\bullet^R(A)$ if we wish to emphasize $R$.
\subsubsection{Cyclic homology}
We only introduce the periodic cyclic theory. Let \[ C_{\even}(A) = \prod_{n=0}^\infty C_{2n}(A), \qquad C_{\odd}(A) = \prod_{n=0}^\infty C_{2n+1}(A), \] with the product topologies. Consider the operator $B: C_n(A) \to C_{n+1}(A)$ given on elementary tensors by \[ B(a_0 \otimes \ldots \otimes a_n) = \sum_{j=0}^n (-1)^{jn} e \otimes a_j \otimes \ldots a_n \otimes a_0 \otimes \ldots \otimes a_{j-1} \] if $a_0 \in A$, and
\[ B(e \otimes a_1 \otimes \ldots \otimes a_n) = 0. \]
Then it is immediate that $B^2 = 0$. Moreover, one can check that \[ bB + Bb = 0.\] Extend the operators $b$ and $B$ to the periodic cyclic complex \[ C_{\per}(A) = C_{\even}(A) \oplus C_{\odd}(A). \] This is a $\Z/2$-graded complex
\[ \xymatrix{
C_{\even}(A) \ar@<.5ex>[r]^-{b+B} &C_{\odd}(A) \ar@<.5ex>[l]^-{b+B}
} \]
with differential $b+B$. The homology groups of this complex are called the even and odd \emph{periodic cyclic homology groups} of $A$, and are denoted $HP_0(A)$ and $HP_1(A)$ respectively. As before, we will write $HP_\bullet^R(A)$ if we wish to emphasize the ground ring $R$.
\subsubsection{Dual cohomology theories}
To obtain periodic cyclic cohomology, we dualize the previous notions. Let \[ C^n(A) = C_n(A)^\dual = \Hom_R(C_n(A),R)\] be the topological $R$-linear dual module of $C_n(A)$ with the topology of uniform convergence on bounded subsets. The maps \[ b:C^n(A) \to C^{n+1}(A), \qquad B: C^n(A) \to C^{n-1}(A) \] are induced by duality, and are given explicitly by
\begin{align*}
b\phi(a_0, \ldots, a_n) &= \sum_{j=0}^{n-1} (-1)^j \phi(a_0, \ldots a_{j-1}, a_ja_{j+1}, a_{j+2}, \ldots, a_n)\\
& \qquad \qquad + (-1)^n\phi(a_na_0, a_1, \ldots, a_{n-1}),
\end{align*} and
\[ B\phi(a_0, \ldots , a_{n-1}) = \sum_{j=0}^{n-1} (-1)^{j(n-1)}\phi(e, a_j, \ldots, a_{n-1}, a_0, \ldots a_{j-1}), \qquad a_0 \in A, \]
\[ B\phi(e, a_1, \ldots , a_{n-1}) = 0. \] The cohomology of $(C^\bullet(A), b)$ is called the \emph{Hochschild cohomology of $A$ (with coefficients in $A^\dual = \Hom_R(A, R)$)} and will be denoted by $HH^\bullet(A)$. The periodic cyclic cochain complex is $C^{\per}(A) = C^{\even}(A) \oplus C^{\odd}(A)$, where
\[C^{\even}(A) = \bigoplus_{n=0}^\infty C^{2n}(A), \qquad C^{\odd}(A) = \bigoplus_{n=0}^\infty C^{2n+1}(A). \] Then $C^{\per}(A)$ is a $\Z/2$-graded complex with differential $b+B$, and its cohomology groups are the even and odd \emph{periodic cyclic cohomology} of $A$, denoted $HP^0(A)$ and $HP^1(A)$ respectively.
Since $C^{\per}(A) \cong C_{\per}(A)^\dual$, there is a canonical pairing \[ \langle \cdot, \cdot \rangle: C^{\per}(A) \times C_{\per}(A) \to R \] which descends to a bilinear map \[ \langle \cdot , \cdot \rangle: HP^{\bullet}(A) \times HP_{\bullet}(A) \to R.\]
\subsubsection{Chern character}
We shall review some basic facts about the Chern character in periodic cyclic homology, see \cite[Ch. 8]{MR1600246} for a more detailed account.
Let $A$ be an arbitrary algebra over the ground ring $R$. Given an idempotent $P \in A$, $P^2 = P$, define the element $\ch P \in C_{\even}(A)$ given by $(\ch P)_0 = P$ and for $n \geq 1$, \[ (\ch P)_{2n} = (-1)^n\frac{(2n)!}{n!}(P^{\otimes (2n+1)} - \frac{1}{2} e \otimes P^{\otimes (2n)}). \] One can verify directly that $b (\ch P_{2(n+1)}) = -B (\ch P_{2n})$, so that $(b+B)\ch P = 0$.
More generally, we can define $\ch P \in C_{\even}(A)$ when $P$ is an idempotent in the matrix algebra $M_N(A) \cong M_N(\C) \otimes A$. Consider the \emph{generalized trace} $T: C_\bullet(M_N(A)) \to C_\bullet(A)$ defined by
\[ T((u_0 \otimes a_0) \otimes \ldots \otimes (u_n \otimes a_n)) = \tr(u_0\ldots u_n)a_0 \otimes \ldots \otimes a_n, \] where $\tr: M_N(\C) \to \C$ is the ordinary trace. As shown in \cite[Ch. 1]{MR1600246}, $T$ is a chain homotopy equivalence, and so induces an isomorphism $HP_\bullet(M_N(A)) \cong HP_\bullet(A)$. So define $\ch P \in C_{\even}(A)$ to be the image of $\ch P \in C_{\even}(M_N(A))$ under the map $T$. In this way, we build a homomorphism
\[ \ch: K_0(A) \to HP_0(A), \qquad \ch [P] = [\ch P], \]
where $K_0(A)$ denotes the algebraic $K$-theory group of $A$, and $[P]$ is the $K$-theory class of an idempotent $P \in M_N(A)$.
Given an invertible $U \in A$, there is a cycle $\ch U \in C_{\odd}(A)$ given by \[ (\ch U)_{2n+1} = (-1)^n n! U^{-1} \otimes U \otimes U^{-1} \otimes \ldots \otimes U^{-1} \otimes U.\] Then, one can check that $(b+B)\ch U = 0$. As in the case of idempotents, define $\ch U \in C_{\odd}(A)$ for any invertible $U \in M_N(A)$ by composing with $T$. In this way, we build a homomorphism
\[ \ch: K_1(A) \to HP_1(A), \qquad \ch [U] = [\ch U], \] where $K_1(A)$ denotes the algebraic $K$-theory group of $A$.
There are pairings
\[ HP^0(A) \times K_0(A) \to R, \qquad HP^1(A) \times K_1(A) \to R \] given by
\[ \langle [\phi], [P] \rangle = \langle [\phi], [\ch P] \rangle, \qquad \langle [\phi], [U] \rangle = \langle [\phi], [\ch U] \rangle \] for an idempotent $P \in M_N(A)$ and an invertible $U \in M_N(A)$.
\subsubsection{Noncommutative geometry dictionary}
In the case where $A = C^\infty(M)$, the algebra of smooth functions on a closed manifold $M$ with its usual Fr\'{e}chet topology, the above homology groups have geometric interpretations. The Hochschild cohomology $H^\bullet(A,A)$ is the graded space of multivector fields on $M$. The cup product corresponds to the wedge product of multivector fields, and the Gerstenhaber bracket corresponds to the Schouten-Nijenhuis bracket. The Hochschild homology $HH_\bullet(A)$ is the space of differential forms on $M$. The differential $B$ descends to a differential on $HH_\bullet(A)$, and this can be identified with the de Rham differential $d$ up to a constant. The even (respectively odd) periodic cyclic homology can be identified with the direct sum of the even (respectively odd) de Rham cohomology groups. In a dual fashion, the Hochschild cohomology $HH^\bullet(A)$ is the space of de Rham currents and the periodic cyclic cohomology can be identified with de Rham homology. For more details, see \cite{MR823176}.
When passing to an arbitrary, not necessarily commutative, algebra $A$, we could view $H^\bullet(A,A)$ and $HH_\bullet(A)$ as spaces of noncommutative multivector fields and differential forms respectively. However, these spaces have a tendency to be badly behaved. For example, they may be too small or non-Hausdorff. Instead, we shall work with the chain complexes $C^\bullet(A,A)$ and $C_\bullet(A)$, and view their elements as (generalized) noncommutative multivector fields and differential forms respectively.
Just as multivector fields act on differential forms by Lie derivative and contraction operations, there are Lie derivative and contraction operations
\[ L, \iota: C^\bullet(A,A) \to \End(C_\bullet(A)) \] for any algebra $A$, which we shall review in the next section.
\subsection{Operations on the cyclic complex}
The Cartan homotopy formula that follows was first observed by Rinehart in \cite{MR0154906} in the case where $D$ is a derivation, and later in full generality by Getzler in \cite{MR1261901}, see also \cite{MR2308582}, \cite{MR1667686}. An elegant and conceptual proof of the Cartan homotopy formula can be found in \cite{MR1468938}. Our conventions vary slightly from \cite{MR1261901}, and are like those in \cite{MR2308582}.
To simplify the notation of what follows, the elementary tensor $a_0 \otimes a_1 \otimes \ldots \otimes a_n \in C_n(A)$ will be written as $(a_0, a_1, \ldots , a_n)$. All operators that are defined in this section are given algebraically on elementary tensors, and extend to continuous linear operators on the corresponding projective tensor products.
All commutators of operators that follow are graded commutators. That is, if $S$ and $T$ are homogenous operators of degree $|S|$ and $|T|$, then
\[ [S, T] = ST - (-1)^{|S||T|}TS. \]
\subsubsection{Lie derivatives, contractions, and the Cartan homotopy formula}
Given a Hochschild cochain $D \in C^k(A,A)$, the \emph{Lie derivative along $D$} is the operator $L_D \in \End(C_{\bullet}(A))$ of degree $1 - k$ given by
\begin{align*}
&L_D(a_0 , \ldots , a_n)\\
&\qquad = \sum_{i=0}^{n-k+1}(-1)^{i(k-1)}(a_0, \ldots , D(a_i, \ldots , a_{i+k-1}), \ldots , a_n)\\
& \qquad \qquad + \sum_{i=1}^{k-1}(-1)^{in}(D(a_{n-i+1}, \ldots , a_n, a_0, \ldots , a_{k-1-i}), a_{k-i}, \ldots , a_{n-i}).
\end{align*}
In the case $D \in C^1(A,A)$, the above formula is just \[ L_D(a_0 , \ldots , a_n) = \sum_{i=0}^n (a_0 , \ldots a_{i-1} , D(a_i) , a_{i+1} , \ldots , a_n). \] To be completely precise in the above formulas, we are identifying $C^k(A,A)$ as a subspace of $\Hom((A_+)^{\potimes_R k}, A)$ by extending by zero, so that
\[ D(a_1, \ldots , a_k) = 0, \qquad \text{ if } a_i = e \text{ for some } i. \]
The one exception, where we do not wish to extend by zero, is for the multiplication map $m$ of the unitization $A_+$. Here, the formula for $L_m$ still gives a well-defined operator on $C_\bullet(A)$, and $L_m = b$.
\begin{proposition} \label{Proposition-LieDerivativeCommutator}
If $D, E \in C^{\bullet}(A,A)$, then
\[ [L_D, L_E] = L_{[D,E]}, \qquad [b, L_D] = L_{\delta D}, \qquad [B, L_D] = 0. \]
\end{proposition}
So $C_{-\bullet}(A)$ and $C_{\per}(A)$ are differential graded modules over the differential graded Lie algebra $\g^\bullet(A)$. In particular, the graded Lie algebra $H^{\bullet+1}(A,A)$ acts via Lie derivatives on both the Hochschild homology $HH_{-\bullet}(A)$ and the periodic cyclic homology $HP_{\bullet}(A)$.
Given a $k$-cochain $D \in C^k(A,A)$, the \emph{contraction with $D$} is the operator $\iota_D \in \End(C_\bullet(A))$ of degree $-k$ given by
\[ \iota_D(a_0 , \ldots , a_n) = (-1)^{k-1}(a_0D(a_1, \ldots , a_k) , a_{k+1} , \ldots , a_n). \]
\begin{proposition} \label{Proposition-ContractionCommutator}
For any $D \in C^\bullet(A,A)$, $[b, \iota_D] = -\iota_{\delta D}$.
\end{proposition}
Although $\iota_D$ interacts well with $b$, it does not with the differential $B$, and needs to be adjusted for the cyclic complex. Given $D \in C^k(A,A)$, let $S_D$ denote the operator on $C_{\bullet}(A)$ of degree $2-k$ given by
\begin{align*}
&S_D(a_0 , \ldots , a_n) = \sum_{i=1}^{n-k+1}\sum_{j=0}^{n-i-k+1} (-1)^{i(k-1) + j(n-k+1)}\\
& \qquad (e , a_{n-j+1} , \ldots , a_n , a_0 , \ldots , a_{i-1} , D(a_i, \ldots , a_{i+k-1}) , a_{i+k} \ldots , a_{n-j}),
\end{align*} if $a_0 \in A$ and \[ S_D(e, a_1, \ldots , a_n) = 0. \] The sum is over all cyclic permutations with $D$ appearing to the right of $a_0$.
Given $D \in C^\bullet(A,A)$, the \emph{cyclic contraction with $D$} is the operator \[ I_D = \iota_D + S_D. \]
\begin{theorem}[Cartan homotopy formula] \label{Theorem-CHF}
For any $D \in C^\bullet(A,A)$, \[ [b+B, I_D] = L_D - I_{\delta D}. \]
\end{theorem}
Theorem \ref{Theorem-CHF} implies that the Lie derivative along a Hochschild cocycle $D \in C^{\bullet}(A,A)$ is continuously chain homotopic to zero in the periodic cyclic complex. Thus, the action of $H^{\bullet+1}(A,A)$ on $HP_{\bullet}(A)$ by Lie derivatives is zero.
The results of this section can be summarized in another way. Consider the endomorphism complex $\End_R(C_{\per}(A))$ whose coboundary map is given by the graded commutator with $b + B$. Let
\[ \Op(A) = \Hom_R\big(\g^\bullet(A), \End_R(C_{\per}(A))\big), \] and let $\partial$ denote the boundary map in $\Op(A)$. Given $\Phi \in \Op(A)$ and $D \in \g^\bullet(A)$, we shall write $\Phi_D := \Phi(D)$. So
\[ (\partial \Phi)_D = [b + B, \Phi_D] - (-1)^{|\Phi|}\Phi_{\delta D}. \] Note that the Lie derivative $L$ and the cyclic contraction $I$ are elements of $\Op(A)$ of even and odd degrees respectively. Theorem~\ref{Theorem-CHF} is exactly the statement
\[ \partial I = L. \] So it follows from this that $\partial L = 0$, i.e.
\[ [b+B, L_D] = L_{\delta D}, \]
as in Proposition~\ref{Proposition-LieDerivativeCommutator}.
\begin{example} \label{Example-GradedAlgebraCyclicHomology}
Consider a nonnegatively graded algebra $A = \bigoplus_{n=0}^\infty A_n$. Let $D: A \to A$ be the algebra derivation defined by $D(a) = n\cdot a$ for all $a \in A_n$. The complex $C_{\per}(A)$ decomposes into eigenspaces for $L_D$, depending on the total degree of a tensor. However, $L_D$ acts by zero on $HP_\bullet(A)$ by Theorem \ref{Theorem-CHF}. Thus the nontrivial part of the homology is contained entirely in the $0$-eigenspace for $L_D$, which coincides with $C_{\per}(A_0)$. In this way, we see the inclusion $A_0 \to A$ induces an isomorphism $HP_\bullet(A_0) \cong HP_\bullet(A)$.
\end{example}
\section{$C^\infty(J)$-modules}
Let $J \subseteq \R$ be a nonempty open interval, which will serve as a parameter space. Loosely speaking, our general approach to deformation theory is as follows: given a family of objects $\{E_t\}_{t \in J}$ that depend smoothly on $t$, we form a bundle $E$ over $J$ whose fiber at $t$ is $E_t$. The object of interest is then the space of smooth sections of the bundle $E$. If each $E_t$ has an underlying vector space structure (for example, if we are dealing with algebras, chain complexes, Lie algebras, etc), then the space of smooth sections is a $C^\infty(J)$-module. In what follows, the vector spaces are in $\LCTVS$, so we will deal with locally convex $C^\infty(J)$-modules. The space of sections will generally inherit new structure by considering any additional structure $\{E_t\}_{t \in J}$ had fiberwise.
Let $X \in \LCTVS$ and consider the space $C^\infty(J,X)$ of infinitely differentiable functions on $J$ with values in $X$. By a differentiable function $f: J \to X$, we mean that the usual limit
\[ \lim_{h \to 0} \frac{f(t+h) - f(t)}{h} \] exists in the topology on $X$ for all $t \in J$.
We equip $C^\infty(J,X)$ with its usual topology of uniform convergence of functions and all their derivatives on compact subsets of $J$. We shall write $C^\infty(J) = C^\infty(J, \C)$, which is a nuclear Fr\'{e}chet algebra under this topology. Notice $C^\infty(J,X)$ is a locally convex module over $C^\infty(J)$ using pointwise scalar multiplication.
Since $X$ is complete,
\[ C^\infty(J,X) \cong C^\infty(J) \potimes X, \] see e.g. \cite[Theorem~44.1]{MR0225131}. In other words, $C^\infty(J,X)$ is a free locally convex $C^\infty(J)$-module. If $X$ is a Fr\'{e}chet space, then $C^\infty(J,X)$ is a Fr\'{e}chet space.
The space $C^\infty(J,X)$ is equipped with continuous linear ``evaluation maps" \[ \epsilon_t: C^\infty(J,X) \to X \] for each $t \in J$ given by $\epsilon_t(x) = x(t)$.
There are several notions for what is meant by saying a collection \[\{F_t: X \to Y\}_{t \in J}\] of continuous linear maps depends smoothly on $t$. We consider one of the strongest.
\begin{definition} \label{Definition-SmoothFamilyOfLinearMaps}
Given $X, Y \in \LCTVS$, a \emph{smooth family of continuous linear maps} from $X$ to $Y$ is a collection of continuous linear maps $\{F_t: X \to Y\}_{t \in J}$ with the property that there exists a continuous linear map $\widetilde{F}: X \to C^\infty(J, Y)$ such that $F_t = \epsilon_t \circ \widetilde{F}$ for all $t \in J$.
\end{definition}
From the universal property of free modules, we see that such a smooth family induces a continuous $C^\infty(J)$-linear map $F: C^\infty(J,X) \to C^\infty(J,Y)$ given by
\[ F(x)(t) = F_t(x(t)), \qquad x \in C^\infty(J,X). \] Conversely, all continuous $C^\infty(J)$-linear maps between free $C^\infty(J)$-modules are induced by a smooth family of continuous linear maps.
Given a smooth family $\{F_t: X \to Y\}_{t \in J}$ of continuous linear maps, it is necessary that the map
\[ t \mapsto F_t(x) \]
is smooth for each $x \in X$. Under the technical assumption that $X$ is barreled (Fr\'{e}chet spaces are examples of barreled spaces), this is also sufficient. The Banach-Steinhaus theorem (uniform boundedness principle) \cite[Theorem 33.1]{MR0225131} is the main advantage of considering barreled spaces.
\begin{proposition} \label{Proposition-SmoothFamilyForBarreledX}
If $X$ is barreled, then $\{F_t: X \to Y\}_{t \in J}$ is a smooth family of continuous linear maps if and only if the map
\[ t \mapsto F_t(x) \] is smooth for every $x \in X$.
\end{proposition}
\begin{proof}
Suppose $t \mapsto F_t(x)$ is smooth for each $x \in X$, and let $F_t^{(n)}(x)$ denote the $n$-th derivative of this map. The linear map $F_t^{(n)}: X \to X$ is in fact continuous for each $t$. Using induction, this follows from the Banach-Steinhaus theorem because $F_t^{(n)}(x)$ is a pointwise limit of continuous linear maps by its very definition.
We must show that the map $\widetilde{F}: X \to C^\infty(J,Y)$ defined by
\[ \widetilde{F}(x)(t) = F_t(x) \] is continuous. Our assumption certainly implies that the map $t \mapsto F_t^{(n)}(x)$ is continuous. Thus, for any compact $K \subset J$, any $n$, and any $x \in X$, the set
\[ \{ F_t^{(n)}(x) ~|~ t \in K\} \] is compact in $Y$, hence bounded. By the Banach-Steinhaus theorem, the set $\{F_t^{(n)}: X \to Y\}_{t \in K}$ is equicontinuous. Thus for any continuous seminorm $q$ on $Y$, there exists a continuous seminorm $p$ on $X$ such that
\[ q(F_t^{(n)}(x)) \leq p(x), \qquad \forall x \in X, \quad \forall t \in K. \] Consequently,
\[ \sup_{t \in K} q(F_t^{(n)}(x)) \leq p(x), \qquad \forall x \in X.\] The expression on the left, which depends on $K, n,$ and $q$, is one of the defining seminorms of $C^\infty(J,Y)$ applied to $\widetilde{F}(x)$. The topology of $C^\infty(J,Y)$ is generated by all such seminorms as $K, n,$ and $q$ vary. This shows that $\widetilde{F}$ is continuous.
\end{proof}
One can also view a collection $\{F_t: X \to Y\}_{t \in J}$ as a map $F: J \to \Hom(X,Y)$.
\begin{proposition} \label{Proposition-SmoothPathsInHom}
Consider a collection $\{F_t: X \to Y\}_{t \in J}$ of continuous linear maps and the corresponding map $F: J \to \Hom(X,Y)$.
\begin{enumerate}
\item If $\{F_t\}_{t \in J}$ is a smooth family, then $F: J \to \Hom(X,Y)$ is a smooth curve.
\item If the curve $F: J \to \Hom(X,Y)$ is smooth, then $F: J \to \Hom_\sigma(X,Y)$ is smooth (with respect to the topology of pointwise convergence).
\item If $X$ is barreled and the curve $F: J \to \Hom_\sigma(X,Y)$ is smooth, then $\{F_t\}_{t \in J}$ is a smooth family.
\end{enumerate}
\end{proposition}
\begin{proof}
The second statement is trivial and the third is a restatement of Proposition \ref{Proposition-SmoothFamilyForBarreledX}. For the first statement, we shall prove $F$ is differentiable, and the proof that $F$ is $n$ times differentiable follows by replacing $F$ with $F^{(n-1)}$. Fix $t \in J$, $\epsilon > 0$, and a continuous seminorm $q$ on $Y$. By continuity of $\widetilde{F}$, there is a seminorm $p$ on $X$ such that
\[ \sup_{u \in [t-\epsilon, t+\epsilon]} q(F_u''(x)) \leq p(x), \qquad \forall x \in X. \]
From Taylor's formula
\[ F_{t+h}(x) = F_t(x) + F_t'(x)h + \int_t^{t+h} F_u''(x)(t+h-u)du, \] we see
\[ q\left( \frac{F_{t+h}(x) - F_t(x)}{h} - F_t'(x) \right) \leq \sup_{u \in [t,t+h]} q(F_u''(x))\frac{h}{2} \leq p(x)\frac{h}{2} \] for $|h| < \epsilon$. Given a bounded subset $A \subset X$, consider the seminorm on $\Hom(X,Y)$
\[ q_A(G) = \sup_{x \in A} q(G(x)), \qquad \forall G \in \Hom(X,Y). \] Then we have shown
\[ q_A\left( \frac{F_{t+h} - F_t}{h} - F_t'\right) \leq C_A\frac{h}{2}, \] where $C_A = \sup_{x \in A}p(x) < \infty$. Since the right side goes to $0$ as $h \to 0$, this proves $\frac{d}{dt} F_t = F_t'$ in the topology of $\Hom(X,Y)$.
\end{proof}
\begin{corollary} \label{Corollary-BanachSpaceInvertibleSmoothFamily}
Let $X$ be a Banach space and let $\{F_t: X \to X\}_{t \in J}$ be a smooth family of continuous linear maps such that each $F_t$ is bijective. Then $\{F_t^{-1}\}_{t \in J}$ is a smooth family as well. Consequently, the map $F: C^\infty(J,X) \to C^\infty(J,X)$ induced by $\{F_t\}_{t \in J}$ is a topological isomorphism of $C^\infty(J)$-modules.
\end{corollary}
\begin{proof}
That each $F_t^{-1}$ is continuous follows from the open mapping theorem. It is well-known that the inversion map on the set of invertibles of the Banach algebra $\Hom(X,X)$ is differentiable. If we view $\{F_t\}_{t \in J}$ as a differentiable path in $\Hom(X,X)$, then it follows from the chain rule that the path corresponding to $\{F_t^{-1}\}_{t \in J}$ is differentiable. From Proposition \ref{Proposition-SmoothPathsInHom}, $\{F_t^{-1}\}_{t \in J}$ is a smooth family. The induced endomorphism of $C^\infty(J,X)$ is clearly inverse to $F$.
\end{proof}
We've described three different meanings for maps $\{F_t: X \to Y\}_{t \in J}$ to depend smoothly on $t$. The corresponding inclusions
\[ \Hom_{C^\infty(J)}\left(C^\infty(J,X), C^\infty(J,Y)\right) \to C^\infty\left(J, \Hom(X,Y)\right) \to C^\infty\left(J,\Hom_\sigma(X,Y)\right) \]
are continuous. They are linear, but not necessarily topological, isomorphisms when $X$ is barreled.
\begin{proposition} \label{Proposition-BanachSpaceSmoothPathsOfHoms}
The canonical map
\[ \Hom_{C^\infty(J)}\left(C^\infty(J,X), C^\infty(J,Y)\right) \to C^\infty\left(J, \Hom(X,Y)\right) \]
is a topological isomorphism if either
\begin{enumerate}
\item $X$ and $Y$ are Banach spaces, or
\item $X$ is a nuclear Fr\'{e}chet space and $Y \in \LCTVS$.
\end{enumerate}
\end{proposition}
\begin{proof}
If $X$ and $Y$ are Banach spaces, then we claim that the domain and codomain are both Fr\'{e}chet spaces. Then the result follows from the open mapping theorem. Since $\Hom(X,Y)$ is a Banach space, $C^\infty\left( J, \Hom(X,Y)\right)$ is a Fr\'{e}chet space. Proposition \ref{Proposition-FreeModuleUMPTopologicalIsomorphism} gives a topological isomorphism
\[ \Hom_{C^\infty(J)}\left(C^\infty(J,X), C^\infty(J,Y)\right) \cong \Hom\left(X, C^\infty(J,Y) \right), \] and the topology of the latter is generated by a countable family of seminorms.
If $X$ is a nuclear Fr\'{e}chet space, then there is a topological isomorphism
\[ X^* \potimes Z \cong \Hom(X, Z) \]
for any $Z \in \LCTVS$, see \cite[Proposition 50.5]{MR0225131}. Using this and Proposition \ref{Proposition-FreeModuleUMPTopologicalIsomorphism}, we have topological isomorphisms
\begin{align*}
\Hom_{C^\infty(J)} \left( C^\infty(J,X), C^\infty(J,Y) \right) &\cong \Hom\left(X, C^\infty(J,Y) \right)\\
&\cong X^* \potimes C^\infty(J) \potimes Y\\
&\cong C^\infty(J) \potimes \Hom(X,Y)\\
&\cong C^\infty\left(J, \Hom(X,Y)\right).
\end{align*}
\end{proof}
An important case to consider is when $Y = \C$. Recall that $M^\dual$ denotes the topological $C^\infty(J)$-linear dual of a $C^\infty(J)$-module $M$.
\begin{corollary} \label{Corollary-DualFreeModule}
If $X$ is either a Banach space or a nuclear Fr\'{e}chet space, then
\[ C^\infty(J,X)^\dual \cong C^\infty(J, X^*). \]
\end{corollary}
\section{Connections and parallel translation}
\subsection{Connections}
Since we are only dealing with one-parameter deformations, we shall only treat connections on $C^\infty(J)$-modules where the interval $J$ represents the parameter space. As there is only one direction to differentiate in, a connection is determined by its covariant derivative. In what follows, we shall identify the two notions, and will commonly refer to covariant differential operators as connections.
\begin{definition}
A \emph{connection} on a locally convex $C^\infty(J)$-module $M$ is a continuous $\C$-linear map $\nabla: M \to M$ such that
\[ \nabla( f\cdot m) = f' \cdot m + f\cdot \nabla m \]
for all $f \in C^\infty(J)$ and $m \in M$.
\end{definition}
It is immediate from this Leibniz rule that the difference of two connections is a continuous $C^{\infty}(J)$-linear map. Further, given any connection $\nabla$ and continuous $C^{\infty}(J)$-linear map $F:M \to M$, the operator $\nabla - F$ is also a connection. So if the space of connections is nonempty, then it is an affine space parametrized by the space $\End_{C^{\infty}(J)}(M)$ of continuous $C^\infty(J)$-linear endomorphisms. Since the operator $\frac{d}{dt}$ is an example of a connection on a free module $C^\infty(J,X)$, we obtain the following classification.
\begin{proposition} \label{Proposition-FormOfConnection}
If $\nabla$ is a connection on $C^\infty(J,X)$, where $X \in \LCTVS$, then
\[ \nabla = \frac{d}{dt} - F, \]
for some continuous $C^\infty(J)$-linear map $F: C^\infty(J,X) \to C^\infty(J,X)$.
\end{proposition}
An element in the kernel of a connection $\nabla$ will be called a \emph{parallel section} for $\nabla$. Suppose $M$ and $N$ are two locally convex $C^\infty(J)$-modules with connections $\nabla_M$ and $\nabla_N$ respectively. We shall say that a continuous $C^\infty(J)$-linear map $F: M \to N$ is \emph{parallel} if $F \circ \nabla_M = \nabla_N \circ F$.
A parallel map sends parallel sections to parallel sections.
\begin{proposition} \label{Proposition-TensorAndDualConnections}
Given locally convex $C^\infty(J)$-modules $M$ and $N$ with connections $\nabla_M$ and $\nabla_N$,
\begin{enumerate}[(i)]
\item \label{Part-TensorConnection}
the operator $\nabla_M \otimes 1 + 1 \otimes \nabla_N$ is a connection on $M \potimes_{C^\infty(J)} N$.
\item \label{Part-DualConnection}
the operator $\nabla_M^\dual$ on $M^\dual = \Hom_{C^\infty(J)}(M, C^\infty(J))$ given by
\[ (\nabla_M^\dual\phi)(m) = \frac{d}{dt} \phi(m) - \phi(\nabla_M m) \]
is a connection.
\end{enumerate}
\end{proposition}
The definition of $\nabla_M^\dual$ ensures that the canonical pairing
\[ \langle \cdot , \cdot \rangle : M^\dual \otimes_{C^\infty(J)} M \to C^\infty(J) \] is a parallel map, where we consider the ground ring $C^\infty(J)$ with the connection $\frac{d}{dt}$. That is,
\[ \frac{d}{dt} \langle \phi, m \rangle = \langle \nabla_{M^*}\phi, m \rangle + \langle \phi, \nabla_M m \rangle. \]
\subsection{Parallel translation in free modules}
Let $X \in \LCTVS$ and let $M = C^\infty(J,X)$ be the corresponding free module.
\begin{definition}
A connection $\nabla$ on $M$ is \emph{integrable} if there is a parallel isomorphism
\[ F: (M, \nabla) \to \left(C^\infty(J,X), \frac{d}{dt}\right) \] of locally convex $C^\infty(J)$-modules.
\end{definition}
We shall express this condition in terms of parallel translation. We will think of $M$ as sections of the trivial bundle whose fiber over $t \in J$ is $M_t \cong X$. Parallel translation relies on the existence and uniqueness of a solution $m \in M$ to the initial value problem
\[ \nabla m = 0, \qquad m(s) = x \] for any given $s \in J$ and $x \in M_s$. In this case, the parallel translation operator
\[ P^\nabla_{s,t}: M_s \to M_t \] is the linear map defined by $P^\nabla_{s,t}(x) = m(t)$, where $m$ is the unique solution to the above initial value problem. Evidently, $P^\nabla_{s,t}$ is a linear isomorphism with inverse $P^\nabla_{t,s}$.
\begin{theorem} \label{Theorem-IntegrableConnections}
A connection $\nabla$ on $M = C^\infty(J,X)$ is integrable if and only if the following two conditions hold:
\begin{enumerate}[(i)]
\item For every $s \in J$ and $x \in M_s$, there is a unique $m \in M$ such that
\[ \nabla m = 0, \qquad m(s) = x. \]
\item The linear map $P^\nabla: X \to C^\infty(J \times J, X)$ given by
\[ P^\nabla(x)(s,t) = P^\nabla_{s,t}(x) \] is well-defined and continuous.
\end{enumerate}
\end{theorem}
\begin{proof}
Notice that the connection $\frac{d}{dt}$ on $C^\infty(J,X)$ satisfies both conditions. Moreover, both conditions are preserved by parallel isomorphism. So an integrable connection $\nabla$ satisfies (i) and (ii).
Conversely, suppose $\nabla$ satisfies (i) and (ii) and fix a value $s \in J$. By condition (ii), the linear maps
\[ \widetilde{F}: X \to C^\infty(J,M_s), \qquad \widetilde{F}(x)(t) = P^\nabla_{t,s}(x) \]
\[ \widetilde{G}: M_s \to M, \qquad \widetilde{G}(x)(t) = P^\nabla_{s,t}(x) \]
are continuous, and induce mutually inverse $C^\infty(J)$-linear isomorphisms
\[ F: M \to C^\infty(J, M_s), \qquad G: C^\infty(J,M_s) \to M \]
by the universal property of free modules. We'll show that
\[ G: \left(C^\infty(J,M_s), \frac{d}{dt}\right) \to \left(M, \nabla\right) \]
is parallel. By $C^\infty(J)$-linearity, the Leibniz rule, and continuity, it suffices to check
\[ G \circ \frac{d}{dt} = \nabla \circ G \] for elements of the form $1 \otimes x \in C^\infty(J) \potimes M_s$. But this follows immediately by definition of parallel translation. That $F = G^{-1}$ is parallel follows automatically.
\end{proof}
The second condition in the theorem can be weakened if $X$ is barreled.
\begin{theorem}
If $X \in \LCTVS$ is barreled, then a connection $\nabla$ on $M = C^\infty(J,X)$ is integrable if and only if the following two conditions hold:
\begin{enumerate}[(i)]
\item For every $s \in J$ and $x \in M_s$, there is a unique $m \in M$ such that
\[ \nabla m = 0, \qquad m(s) = x. \]
\item Each $P^\nabla_{s,t}: M_s \to M_t$ is continuous, and for each fixed $x \in X$, the map $(s,t) \mapsto P^\nabla_{s,t}(x)$ is smooth (i.e. all mixed partial derivatives exist).
\end{enumerate}
\end{theorem}
\begin{proof}
Mimic the proof of Proposition~\ref{Proposition-SmoothFamilyForBarreledX} to show that
\[ P^\nabla: X \to C^\infty(J\times J, X), \qquad P^\nabla(x)(s,t) = P^\nabla_{s,t}(x) \] is continuous.
\end{proof}
Essentially, an integrable connection $\nabla$ is one for which we can parallel translate, and moreover the parallel translation operators $P^\nabla_{s,t}$ are isomorphisms of topological vector spaces that depend smoothly on both parameters $s$ and $t$.
Now suppose $X, Y \in \LCTVS$ and $M = C^\infty(J,X)$ and $N = C^\infty(J,Y)$. Suppose $F: M \to N$ is a continuous $C^\infty(J)$-linear map and $\{F_t: M_t \to N_t\}_{t \in J}$ is the corresponding smooth family of continuous linear maps.
\begin{proposition} \label{Proposition-ParallelFamilyCommutesWithParallelTransport}
In the above situation, if $F: M \to N$ is parallel with respect to integrable connections on $M$ and $N$, then the diagram
\[ \xymatrix{
M_s \ar[r]^-{F_s} \ar[d]_-{P^{\nabla_M}_{s,t}} & N_s \ar[d]^-{P^{\nabla_N}_{s,t}}\\
M_t \ar[r]^-{F_t} & N_t\\
} \]
commutes for all $t,s \in J$.
\end{proposition}
\begin{proof}
Given $x \in M_s$, let $m \in M$ be the unique $\nabla_M$-parallel section through $x$. Then $F(m)$ is the unique $\nabla_N$-parallel section through $F(m)(s) = F_s(x)$. Consequently,
\[ P^{\nabla_N}_{s,t}(F_s(x)) = F(m)(t) = F_t(m(t)) = F_t(P^{\nabla_M}_{s,t}(x)). \]
\end{proof}
If $N$ has the trivial connection $\frac{d}{dt}$, we obtain the following.
\begin{corollary} \label{Corollary-ParallelTranslationFormula}
Given any integrable connection $\nabla$ on $M$ and parallel isomorphism
\[ F: (M, \nabla) \to \left(C^\infty(J,X), \frac{d}{dt}\right), \]
then $P^\nabla_{s,t} = F_t^{-1} \circ F_s: M_s \to M_t$.
\end{corollary}
\begin{proposition} \label{Proposition-IntegrableTensorAndDual}
\begin{enumerate}
\item If $\nabla_M$ and $\nabla_N$ are integrable connections on $M$ and $N$, then $\nabla_{\potimes} := \nabla_M \otimes 1 + 1 \otimes \nabla_n$ is integrable on $M \potimes_{C^\infty(J)} N$, and
\[ P^{\nabla_{\potimes}}_{s,t} = P^{\nabla_M}_{s,t} \otimes P^{\nabla_N}_{s,t}: M_s \potimes N_s \to M_t \potimes N_t. \]
\item If $X$ is either a Banach space or a nuclear Fr\'{e}chet space, and $M = C^\infty(J,X)$ has an integrable connection $\nabla_M$, then the dual connection $\nabla_M^\dual$ is integrable on $M^\dual = C^\infty(J,X^*)$ (see Corollary \ref{Corollary-DualFreeModule} for this identification), and
\[ P^{\nabla_M^\dual}_{s,t} = \left(P^\nabla_{t,s}\right)^*: M_s^* \to M_t^*. \]
\end{enumerate}
\end{proposition}
\begin{proof}
Given parallel isomorphisms
\[ F: (M, \nabla_M) \to \left(C^\infty(J,X), \frac{d}{dt}\right), \qquad G: (N, \nabla_N) \to \left(C^\infty(J,Y), \frac{d}{dt}\right), \] we obtain a parallel isomorphism
\begin{align*}F \otimes G: \left( M \potimes_{C^\infty(J)} N, \nabla_{\potimes} \right) &\to \left( C^\infty(J,X) \potimes_{C^\infty(J)} C^\infty(J,Y), \frac{d}{dt} \otimes 1 + 1 \otimes \frac{d}{dt} \right)\\
&\cong \left( C^\infty(J, X \potimes Y), \frac{d}{dt} \right),
\end{align*} which shows $\nabla_{\potimes}$ is integrable. In a similar way, we obtain a parallel isomorphism
\[ (F^{-1})^\dual: (M^\dual, \nabla_M^\dual) \to \left( C^\infty(J, X^*), \frac{d}{dt} \right) \]
because the dual connection of $\frac{d}{dt}$ on $C^\infty(J,X)$ identifies with $\frac{d}{dt}$ on $C^\infty(J,X^*)$ under the isomorphism of Corollary \ref{Corollary-DualFreeModule}. The parallel translation formulas follow from Corollary \ref{Corollary-ParallelTranslationFormula}.
\end{proof}
Let us consider the problem of parallel translation for a connection $\nabla$ on $M = C^\infty(J,X)$. Recall by Proposition~\ref{Proposition-FormOfConnection} that $\nabla = \frac{d}{dt} - F$ for some continuous $C^\infty(J)$-linear map $F: C^\infty(J,X) \to C^\infty(J,X)$. Let $\{F_t: X \to X\}_{t \in J}$ be the corresponding smooth family of continuous linear maps. To parallel translate, we must solve the first order linear ODE
\[ x'(t) = F_t(x(t)), \qquad x(s) = x_0 \]
given $s \in J$ and $x_0 \in X$. By the fundamental theorem of calculus (which is valid for functions with values in $X \in \LCTVS$), any solution satisfies
\[ x(t) = x(s) + \int_s^tx'(u)du = x_0 + \int_s^t F_u(x(u))du. \] Applying the fundamental theorem inductively, we obtain
\begin{align*}
x(t) &= x_0 + \sum_{n=1}^N \int_{s}^t \int_{s}^{u_1} \ldots \int_{s}^{u_{n-1}} (F_{u_1} \circ \ldots \circ F_{u_n})(x_0) du_n \ldots du_1\\ &\qquad \qquad+ \int_{s}^t \int_{s}^{u_1} \ldots \int_{s}^{u_{N}}(F_{u_1} \circ \ldots \circ F_{u_{N+1}})(x(u_{N+1})) du_{N+1}du_N \ldots du_1. \end{align*} for any $N$. If the last term can be shown to converge to $0$ in $C^\infty(J,X)$ as $N \to \infty$, then any solution $x(t)$ has the form
\[ x(t) = x_0 + \sum_{n=1}^\infty \int_{s}^t \int_{s}^{u_1} \ldots \int_{s}^{u_{n-1}} (F_{u_1} \circ \ldots \circ F_{u_n})(x_0) du_n \ldots du_1. \] This gives uniqueness of solutions. If this series can be shown to converge, we obtain existence of solutions. It is straightforward to show both of these in the case where $X$ is a Banach space. The fundamental theorem of calculus ensures that the solution depends smoothly on both $t$ and $s$. These are well-known results from the theory of first order linear ODE's on a Banach space, which we restate in our language.
\begin{theorem}
If $X$ is a Banach space, then every connection on $C^\infty(J,X)$ is integrable.
\end{theorem}
Notice that if $\nabla$ has constant coefficients, i.e. $F_t$ doesn't depend on $t$, then the solution takes the well-known form
\[ x(t) = \sum_{n=0}^\infty \frac{(t-s)^n}{n!}F^n(x_0) = \exp((t-s)F)(x_0). \]
Once we start considering other classes of locally convex vector spaces, e.g. Fr\'{e}chet spaces, the existence and uniqueness theorem for solutions to linear ODE's is false. One cannot guarantee that the above series defining the solution will converge.
Another situation in which we can get control of this series is when $F$ has nilpotence properties. We'll call $F$ \emph{nilpotent (with respect to $\frac{d}{dt}$)} if there is an integer $n$ such that
\[ F_{u_1} \circ \ldots \circ F_{u_n} = 0, \qquad \forall u_1, \ldots , u_n \in J. \] In this case, the above series becomes a finite sum, and we see that the connection $\nabla = \frac{d}{dt} - F$ is integrable.
Let's generalize the above discussion to perturbations $\nabla - F$ of an integrable connection $\nabla$.
\begin{theorem}[Fundamental theorem of calculus]
If $\nabla$ is an integrable connection on $C^\infty(J,X)$, then
\[ x(t) = P_{s,t}^\nabla(x(s)) + \int_s^t P_{u,t}^\nabla\left( (\nabla x)(u) \right)du \] for any $x \in C^\infty(J,X)$ and $s,t \in J$.
\end{theorem}
\begin{proof}
Fix $s \in J$ and view everything as a function of $t$ (so that $\nabla$ differentiates with respect to $t$). Using the fact that $(\nabla \circ P_{u,t})(y(u)) = 0$ for any $u$ or $y$, we see that applying $\nabla$ to the right hand side gives $(\nabla x)(t)$. Thus the two sides differ by a $\nabla$-parallel section, which must be $0$ because the two sides are equal when $t = s$.
\end{proof}
By repeatedly applying this fundamental theorem of calculus, we see that solutions to
\[ (\nabla - F)x = 0, \qquad x(s) = x_0 \]
take the form
\begin{align*} x(t) &= P_{s,t}^\nabla(x_0) + \sum_{n=1}^\infty \int_s^t \int_s^{u_1} \ldots \int_s^{u_{n-1}}\\ &\qquad(P_{u_1,t}^\nabla \circ F_{u_1} \circ P_{u_2,u_1}^\nabla \circ F_{u_2} \circ \ldots \circ P_{u_n,u_{n-1}}^\nabla \circ F_{u_n} \circ P_{s,u_{n}}^\nabla)(x_0)du_n \ldots du_1,
\end{align*} provided the series converges.
So we shall say that $F$ is \emph{nilpotent (with respect to $\nabla$)} if there is an integer $n$ such that
\[ F_{u_1} \circ P_{u_2,u_1}^\nabla \circ F_{u_2} \circ \ldots \circ P_{u_n,u_{n-1}}^\nabla \circ F_{u_n} = 0, \qquad \forall u_1, \ldots u_n \in J. \] Thus a nilpotent perturbation $\nabla - F$ of an integrable connection is integrable. We are interested in a special case of this. We record it here, though it shall be used in a subsequent paper.
\begin{proposition}\label{Proposition-NilpotentPerturbation}
Suppose $\nabla$ is an integrable connection on $C^\infty(J,X)$, and $F$ is a $C^\infty(J)$-linear endomorphism of $C^\infty(J,X)$ such that $[\nabla, F] = 0$ and $F^N = 0$ for some integer $N$. Then $\nabla - F$ is integrable and
\[ P_{s,t}^{\nabla - F} = \sum_{n=0}^{N-1} \frac{(t-s)^n}{n!}F_t^n \circ P_{s,t}^\nabla. \]
\end{proposition}
\begin{proof}
By assumption, $F$ is parallel with respect to $\nabla$. Using Proposition~\ref{Proposition-ParallelFamilyCommutesWithParallelTransport}, we have
\[ P_{u_1,t}^\nabla \circ F_{u_1} \circ P_{u_2,u_1}^\nabla \circ F_{u_2} \circ \ldots \circ P_{u_n,u_{n-1}}^\nabla \circ F_{u_n} \circ P_{s,u_{n}}^\nabla = F_t^n \circ P_{s,t}^\nabla. \]
It follows that $F$ is nilpotent with respect to $\nabla$, so $\nabla - F$ is integrable. From the explicit series solution, we see
\begin{align*} P_{s,t}^{\nabla-F}(x_0) &= P_{s,t}^\nabla(x_0) + \sum_{n=1}^{N-1} \int_s^t \int_s^{u_1} \ldots \int_s^{u_{n-1}} F_t^n(P_{s,t}^\nabla(x_0))du_n\ldots du_1\\
&= \sum_{n=0}^{N-1} \frac{(t-s)^n}{n!}F_t^n(P_{s,t}^\nabla(x_0)).
\end{align*}
\end{proof}
In other words, if $x$ is the $\nabla$-parallel section through $x_0$ over $s \in J$, then $\exp((t-s)F)(x)$ is the $(\nabla - F)$-parallel section through $x_0$.
\section{Smooth deformations}
\subsection{Deformations of algebras}
Let $X \in \LCTVS$ and let $J$ denote an open interval of real numbers.
\begin{definition}
A \emph{smooth one-parameter deformation of algebras} is a smooth family of continuous linear maps $\{ m_t: X \potimes X \to X \}_{t \in J}$ for which each $m_t$ is associative.
\end{definition}
So for each $t \in J$, we have a locally convex algebra $A_t := (X, m_t)$ whose underlying space is $X$.
Consider the continuous $C^\infty(J)$-linear map
\[ m: C^\infty(J, X \potimes X) \to C^\infty(J,X) \] associated to the smooth family $\{m_t\}_{t \in J}$. Letting $A_J = C^\infty(J,X)$, then $m$ can be viewed as a map
\[ m: A_J \potimes_{C^\infty(J)} A_J \to A_J \]
using Proposition~\ref{Proposition-TensorProductOfFreeModules}. Associativity of $m$ follows from associativity of the family $\{m_t\}$. Thus $A_J$ is a locally convex $C^\infty(J)$-algebra, which we shall refer to as the \emph{algebra of sections} of the deformation $\{A_t\}_{t \in J}$. Explicitly, the multiplication in $A_J$ is given by
\[ (a_1 a_2)(t) = m_t(a_1(t), a_2(t)) \] for all $a_1, a_2 \in A_J$. Note that the evaluation maps $\epsilon_t: A_J \to A_t$ are algebra homomorphisms.
\begin{proposition}
Associating to a deformation its algebra of sections gives a one-to-one correspondence between smooth one-parameter deformations over $J$ with underlying space $X$ and locally convex $C^\infty(J)$-algebra structures on $C^\infty(J,X)$.
\end{proposition}
If $X$ is Fr\'{e}chet, then our definition of a smooth deformation is equivalent to a smooth path in $\Hom(X \potimes X, X)$ whose image lies in the set of associative products, by Proposition \ref{Proposition-SmoothPathsInHom}. The following is a useful criterion for checking that the deformation is smooth in this case.
\begin{proposition} \label{Proposition-DeformationsOnFrechetSpace}
If $X$ is a Fr\'{e}chet space, then a set of continuous associative multiplications $\{m_t: X \potimes X \to X\}_{t \in J}$ is a smooth one-parameter deformation if and only if the map
\[ t \mapsto m_t(x_1, x_2) \] is smooth for each fixed $x_1, x_2 \in X$.
\end{proposition}
\begin{proof}
If $\{m_t\}_{t \in J}$ is a smooth one-parameter deformation, then it is immediate that $t \mapsto m_t(x_1, x_2)$ is smooth for all $x_1, x_2 \in X$.
Conversely, if $t \mapsto m_t(x_1, x_2)$ is smooth for each fixed $x_1, x_2 \in X$, then the map
\[ m: X \times X \to C^\infty(J,X) \] given by
\[ m(x_1, x_2)(t) = m_t(x_1, x_2) \]
is separately continuous by Proposition~\ref{Proposition-SmoothFamilyForBarreledX}. Since $X$ is Fr\'{e}chet, it follows that $m$ is jointly continuous and so induces a continuous linear map
\[ m: X \potimes X \to C^\infty(J,X). \] This shows that $\{m_t\}_{t \in J}$ is a smooth family of continuous linear maps.
\end{proof}
\begin{definition}
A \emph{morphism} between the deformations $\{A_t\}_{t \in J}$ and $\{B_t\}_{t\in J}$ is a continuous $C^\infty(J)$-linear algebra homomorphism $F: A_J \to B_J$.
\end{definition}
Thus a morphism is equivalent to a family $\{F_t: A_t \to B_t\}_{t \in J}$ of continuous algebra homomorphisms which vary smoothly in the sense of Definition~\ref{Definition-SmoothFamilyOfLinearMaps}. When $X$ is Fr\'{e}chet, the smoothness can be checked using Proposition~\ref{Proposition-SmoothFamilyForBarreledX}.
A deformation is called \emph{constant} if the products $\{m_t\}$ do not depend on $t$. A deformation is called \emph{trivial} if it is isomorphic to a constant deformation. Thus $\{A_t\}_{t \in J}$ is trivial if and only if there is a locally convex algebra $B$ such that $A_J \cong C^\infty(J,B)$ as algebras. We can characterize triviality of a smooth deformation of algebras in terms of connections.
\begin{proposition} \label{Proposition-TrivialDeformationOfAlgebras}
The deformation $\{A_t\}_{t \in J}$ is trivial if and only if $A_J$ admits an integrable connection that is a derivation with respect to the algebra structure. In this case, the parallel translation maps $P_{s,t}^\nabla: A_s \to A_t$ are isomorphisms of locally convex algebras.
\end{proposition}
\begin{proof}
Notice that $\frac{d}{dt}$ is an integrable connection and a derivation on a constant deformation. If $\{A_t\}_{t \in J}$ is trivial and $F: A_J \to B_J$ is a $C^\infty(J)$-linear algebra isomorphism with the algebra of sections of a constant deformation, then $\nabla = F^{-1}\frac{d}{dt}F$ is a connection and a derivation on $A_J$, and $\nabla$ is integrable because
\[ F: (A_J, \nabla) \to \left(B_J, \frac{d}{dt}\right) \] is a parallel isomorphism.
Conversely, suppose $A_J$ has an integrable connection $\nabla$ that is a derivation. That $\nabla$ is a derivation is equivalent to the multiplication map
\[ m: ( A_J \potimes_{C^\infty(J)} A_J, \nabla \otimes 1 + 1 \otimes \nabla) \to (A_J, \nabla) \] being a parallel map. Combining Proposition \ref{Proposition-ParallelFamilyCommutesWithParallelTransport} with Corollary \ref{Corollary-ParallelTranslationFormula}, we see that each $P^\nabla_{s,t}: A_s \to A_t$ is an algebra isomorphism.
Fixing an $s \in J$, it follows that
\[ \{P^\nabla_{s,t}: A_s \to A_t\}_{t \in J} \]
is an isomorphism between the constant deformation with fiber $A_s$ and $\{A_t\}_{t \in J}$.
\end{proof}
Thus it is important to determine if a deformation has a connection that is a derivation. In analogy with the work of Gerstenhaber on formal deformations \cite{MR0171807}, the obstruction to this is cohomological.
Given any connection $\nabla$ on the algebra of sections $A_J$, define the bilinear map $E$ by
\[ \nabla(a_1 a_2) = \nabla(a_1)a_2 + a_1\nabla(a_2) - E(a_1, a_2). \]
So $E$ is the defect of $\nabla$ from being a derivation, and in fact $E = \delta \nabla$, where $\delta$ is the Hochschild coboundary. It follows that $\delta E = 0$. Using the Leibniz rule for $\nabla$, one can check that $E$ is a $C^\infty(J)$-bilinear map. So $E$ defines a cohomology class $[E] \in H^2_{C^\infty(J)}(A_J,A_J)$. Notice that $\nabla$ is only $\C$-linear and not $C^\infty(J)$-linear. Thus, we may have $[E] \neq 0$.
\begin{proposition} \label{Proposition-ClassOfE}
The cohomology class $[E] \in H^2_{C^\infty(J)}(A_J,A_J)$ is independent of the choice of connection. Moreover, $[E] = 0$ if and only if $A_J$ possesses a connection that is a derivation.
\end{proposition}
\begin{proof}
Let $\nabla$ and $\nabla'$ be two connections with corresponding cocycles $E$ and $E'$. Since $\nabla' = \nabla - F$ for some $F \in C^1_{C^\infty(J)}(A_J,A_J)$, we have
\[ E' = \delta(\nabla - F) = E - \delta F, \] which shows that $[E] = [E'].$
If $\nabla$ is a connection that is a derivation, then $E = \delta \nabla = 0.$ Conversely, if $\nabla$ is any connection on $A_J$ and $[E] = 0$, then $E = \delta F$ for some $F \in C^1_{C^\infty(J)}(A_J,A_J)$. Hence $\delta (\nabla - F) = 0$, which shows that $\nabla - F$ is a connection that is a derivation.
\end{proof}
From this, we see that the cohomology class $[E]$ provides an obstruction to the triviality of a deformation. Even if this obstruction vanishes, there is still an analytic obstruction in that the corresponding connection may not be integrable. These two issues are common to the smooth deformation theory of other types of structures as well, e.g. cochain complexes (see below) or $A_\infty$-algebras \cite{Yashinski-Thesis}.
\subsection{Deformations of cochain complexes}
By a \emph{smooth one-parameter deformation of cochain complexes}, we mean a collection $\{X^n\}_{n \in \Z}$ of spaces in $\LCTVS$ together with a smooth family of continuous linear maps $\{ d^n_t: X^n \to X^{n+1}\}_{t \in J}$ for each $n$ such that $d^{n+1}_t \circ d^n_t = 0$ for all $t \in J$. (By turning the arrows around, we could just as well talk about deformations of chain complexes.) For each $t \in J$, we have a locally convex cochain complex
\[ (\mathcal{C}^\bullet_t, d_t) := \left(
\xymatrix{
\ldots \ar[r]^-{d_t} & X^{n-1} \ar[r]^-{d_t} & X^n \ar[r]^-{d_t} & X^{n+1} \ar[r]^-{d_t} &\ldots
}
\right) \]
built on the same underlying family of spaces.
Let $\mathcal{C}_J^n = C^\infty(J,X^n)$ and let $d: \mathcal{C}_J^n \to \mathcal{C}_J^{n-1}$ be the continuous $C^\infty(J)$-linear map associated to the smooth family $\{d_t\}_{t \in J}$. We obtain a chain complex
\[ (\mathcal{C}_J^\bullet, d) := \left(
\xymatrix{
\ldots \ar[r]^-d & \mathcal{C}_J^{n-1} \ar[r]^-d & \mathcal{C}_J^n \ar[r]^-d & \mathcal{C}_J^{n+1} \ar[r]^-d &\ldots
}
\right) \]
of locally convex $C^\infty(J)$-modules. We'll call $\mathcal{C}_J^\bullet$ the \emph{complex of sections} of the deformation. The cohomology $H^\bullet(\mathcal{C}_J)$ is a $C^\infty(J)$-module, and the evaluation chain maps $\epsilon_t: \mathcal{C}_J^\bullet \to \mathcal{C}_t^\bullet$ induce maps on cohomology
\[ (\epsilon_t)_*: H^\bullet(\mathcal{C}_J) \to H^\bullet(\mathcal{C}_t). \]
By a morphism of two deformations, we mean a continuous $C^\infty(J)$-linear (degree $0$) chain map between their respective complexes of sections. We'll call a deformation \emph{trivial} if it is isomorphic to a constant deformation.
\begin{proposition}
Suppose $(\mathcal{C}^\bullet, d)$ is a cochain complex of Fr\'{e}chet spaces such that the cohomology $H^\bullet(\mathcal{C})$ is Hausdorff. Let $\mathcal{C}_J^\bullet = C^\infty(J, \mathcal{C}^\bullet)$ be the complex of sections of the constant deformation with fiber $\mathcal{C}^\bullet$. Then
\[ H^\bullet(\mathcal{C}_J) \cong C^\infty\left(J, H^\bullet(\mathcal{C})\right) \] as locally convex $C^\infty(J)$-modules.
\end{proposition}
\begin{proof}
Notice that requiring $H^n(\mathcal{C}) = Z^n(\mathcal{C}) / B^n(\mathcal{C})$ to be Hausdorff is equivalent to requiring the space of coboundaries $B^n(\mathcal{C})$ to be closed. In this case, both $B^n(\mathcal{C})$ and $H^n(\mathcal{C})$ are Fr\'{e}chet spaces for all $n$.
Notice that $Z^n(\mathcal{C}_J) = C^\infty(J, Z^n(\mathcal{C}))$, but a priori we only have $B^n(\mathcal{C}_J) \subseteq C^\infty(J, B^n(\mathcal{C}))$. However, since $d: \mathcal{C}^n \to B^{n+1}(\mathcal{C})$ is a surjection of Fr\'{e}chet spaces, it follows from \cite[Proposition 43.9]{MR0225131} that
\[ 1 \otimes d: C^\infty(J) \potimes \mathcal{C}^n \to C^\infty(J) \potimes B^{n+1}(\mathcal{C}) \] is surjective as well. That is, $B^n(\mathcal{C}_J) = C^\infty(J, B^n(\mathcal{C}))$ for all $n$. Thus,
\[ H^n(\mathcal{C}_J) = Z^n(\mathcal{C}_J)/B^n(\mathcal{C}_J) = C^\infty(J, Z^n(\mathcal{C})) / C^\infty(J, B^n(\mathcal{C})) \cong C^\infty(J, H^n(\mathcal{C})), \] where the last isomorphism is from Proposition \ref{Proposition-QuotientOfFreeModules}.
\end{proof}
\begin{example} \label{Example-CyclicComplexOfDeformation}
If $\{A_t\}_{t \in J}$ is a smooth deformation of algebras, then $\{(C_{\per}(A_t), b_t + B)\}_{t \in J}$ is a smooth deformation of chain complexes. Notice that the Hochschild boundary $b_t$ depends on the multiplication of $A_t$, whereas the operator $B$ does not. Since the completed projective tensor product commutes with direct products \cite[Theorem 15.4.1]{MR632257}, the complex of sections of $\{C_{\per}(A_t)\}_{t \in J}$ is naturally identified with the periodic cyclic complex $C_{\per}^{C^\infty(J)}(A_J)$. One can also consider the complexes associated to the various other homology/cohomology theories discussed above.
\end{example}
As in the algebra case, we can characterize triviality of a deformation of chain complexes in terms of connections. The proofs here are analogous those in the algebra case.
\begin{proposition} \label{Proposition-BundleOfComplexesParallelTransport}
A smooth deformation of cochain complexes $\{\mathcal{C}^\bullet_t\}_{t \in J}$ is trivial if and only if the complex of sections $\mathcal{C}_J^\bullet$ admits an integrable connection that is a chain map. For such a connection $\nabla$, the parallel translation map $P^\nabla_{s,t}: \mathcal{C}^\bullet_s \to \mathcal{C}^\bullet_t$ is an isomorphism of locally convex cochain complexes for all $s,t \in J$. In particular, the parallel translation maps induce isomorphisms
\[ (P^\nabla_{s,t})_*: H^\bullet(\mathcal{C}_s) \to H^\bullet(\mathcal{C}_t). \]
\end{proposition}
The obstruction to the existence of such a connection is again cohomological. Let $\nabla$ be any connection on $\mathcal{C}_J^\bullet$, and consider the map
\[ G = [d, \nabla]: \mathcal{C}_J^\bullet \to \mathcal{C}_J^{\bullet + 1}, \]
which is the defect of $d$ from being a $\nabla$-parallel map (equivalently, the defect of $\nabla$ from being a chain map). It follows that $G$ is $C^\infty(J)$-linear and $[d, G] = 0$, so $G$ is a cocycle in the endomorphism complex $\End_{C^\infty(J)}(\mathcal{C}_J)$.
\begin{proposition} \label{Proposition-ClassOfG}
The cohomology class $[G] \in H^1(\End_{C^\infty(J)}(\mathcal{C}_J))$ is independent of the choice of connection $\nabla$. Moreover, $[G] = 0$ if and only if $\mathcal{C}_J^\bullet$ admits a connection that is a chain map.
\end{proposition}
Suppose $\mathcal{C}_J^\bullet$ is equipped with a connection $\nabla$ that is a chain map. Our main goal is to identify the cohomology groups $H^\bullet(\mathcal{C}_s) \cong H^\bullet(\mathcal{C}_t)$ of different fibers via parallel translation. By Proposition \ref{Proposition-BundleOfComplexesParallelTransport}, this happens when $\nabla$ is integrable. In this case, the cochain complexes themselves are fiberwise isomorphic, and this may be too strong of a condition to be useful in practice.
As $\nabla$ is a chain map, it induces a connection $\nabla_*$ on the $C^\infty(J)$-module $H^\bullet(\mathcal{C}_J)$. The homology module may not be free, complete, or even Hausdorff, so we should be careful about what we mean by integrability of $\nabla_*$. Nonetheless, it makes sense to inquire about the existence and uniqueness of a solution $[c] \in H^\bullet(\mathcal{C}_J)$ to the cohomological differential equation
\[ \nabla_* [c] = 0, \qquad [c(s)] = [c_0] \]
with initial value $[c_0] \in H^\bullet(\mathcal{C}_s)$. Having this is enough to construct parallel translation operators
\[ P^{\nabla_*}_{s,t}: H^\bullet(\mathcal{C}_s) \to H^\bullet(\mathcal{C}_t), \]
which are linear isomorphisms. Additionally, if the map $[c_0] \mapsto [c]$ is continuous, then $P^{\nabla_*}_{s,t}$ is continuous, hence an isomorphism of topological vector spaces.
\section{Some rigidity results}
The results in this section are largely not original. Most are stated, with slight variations, in \cite{Crainic}, where they are proved using the homological perturbation lemma. We give different proofs using our methods in the setting of smooth deformations.
We'll call a locally convex algebra $A$ \emph{(smoothly) rigid} if every smooth deformation $\{A_t\}_{t \in J}$ with $A_0 = A$ is trivial on some interval $J' \subset J$ containing $0$. Similarly, we can define rigidity of a cochain complex. Our main tool for proving rigidity results in the setting of Banach spaces is the following lemma.
\begin{lemma} \label{Lemma-DeformationOfContractibleHasZeroCohomology}
Let $\{(\mathcal{C}_t^\bullet, d_t)\}_{t \in J}$ be a smooth deformation of cochain complexes of Banach spaces, and suppose the complex $\mathcal{C}_0^\bullet$ has a continuous linear contracting homotopy in degree $n$
\[ \xymatrix{
\mathcal{C}_0^{n-1} \ar@<.5ex>[r]^-{d_0} & \mathcal{C}_0^n \ar@<.5ex>[l]^-h \ar@<.5ex>[r]^-{d_0}& \mathcal{C}_0^{n+1}, \ar@<.5ex>[l]^-h
} \qquad d_0 h + h d_0 = 1.\]
Then there is a subinterval $J' = (-\epsilon, \epsilon)$ such that \[ Z^n(\mathcal{C}_{J'}) \cong C^\infty(J', Z^n(\mathcal{C}_0)), \qquad B^n(\mathcal{C}_{J'}) \cong C^\infty(J', B^n(\mathcal{C}_0)), \qquad H^n(\mathcal{C}_{J'}) = 0.\]
\end{lemma}
\begin{proof}
By assumption, there are split short exact sequences
\[ \xymatrix{
0 \ar[r] & Z^{n-1}(\mathcal{C}_0) \ar[r] &\mathcal{C}_0^{n-1} \ar@<.5ex>[r]^-{d_0} & B^n(\mathcal{C}_0) \ar@<.5ex>[l]^-h \ar[r] &0,
} \]
\[ \xymatrix{
0 \ar[r] & Z^n(\mathcal{C}_0) \ar[r] &\mathcal{C}_0^n \ar@<.5ex>[r]^-{d_0} & B^{n+1}(\mathcal{C}_0) \ar@<.5ex>[l]^-h \ar[r] &0,
} \]
and $B^n(\mathcal{C}_0) = Z^n(\mathcal{C}_0)$. So the cocycles are complemented in the space of cochains, that is, there are closed subspaces $W^{n-1} = h(B^n(\mathcal{C}_0))$ and $W^n = h(B^{n+1}(\mathcal{C}_0))$ for which
\[ \mathcal{C}_0^{n-1} = Z^{n-1}(\mathcal{C}_0) \oplus W^{n-1}, \qquad \mathcal{C}_0^n = Z^n(\mathcal{C}_0) \oplus W^n. \]
Let $\pi: \mathcal{C}_0^n \to Z^n(\mathcal{C}_0) = B^n(\mathcal{C}_0)$ be the projection. Then $\pi$ induces a continuous $C^\infty(J)$-linear map $\pi: \mathcal{C}_J^n \to C^\infty(J, Z^n(\mathcal{C}_0))$, which restricts to the maps
\[ \pi: Z^n(\mathcal{C}_J) \to C^\infty(J, Z^n(\mathcal{C}_0)), \qquad \pi: B^n(\mathcal{C}_J) \to C^\infty(J, B^n(\mathcal{C}_0)). \]
We claim these are topological isomorphisms for a small enough interval $J'$ containing $0$. We'll prove this by showing $\pi$ is injective on cocycles and surjective on coboundaries. The results then follow from the commutative diagram
\[ \xymatrix{
B^n(\mathcal{C}_{J'}) \ar@{^{(}->}[r] \ar[d]^-\pi &Z^n(\mathcal{C}_{J'}) \ar[d]^-\pi \\
C^\infty(J', B^n(\mathcal{C}_0)) \ar@{=}[r] & C^\infty(J', Z^n(\mathcal{C}_0))
} \]
and the open mapping theorem.
Consider the family of maps $\pi \circ d_t$ restricted to $W^{n-1}$. When $t = 0$,
\[ \pi \circ d_0: W^{n-1} \to B^n(\mathcal{C}_0) \] is a topological isomorphism of Banach spaces. So there is some $\epsilon > 0$ for which $\pi \circ d_t: W^{n-1} \to B^n(\mathcal{C}_0)$ is a topological isomorphism for all $t \in J' := (-\epsilon, \epsilon)$. From Corollary \ref{Corollary-BanachSpaceInvertibleSmoothFamily}, the induced $C^\infty(J')$-linear map
\[ \pi \circ d: C^\infty(J', W^{n-1}) \to C^\infty(J', B^n(\mathcal{C}_0)) \] is an isomorphism, and in particular it is surjective. It follows that $\pi: B^n(\mathcal{C}_{J'}) \to C^\infty(J', B^n(\mathcal{C}_0))$ is surjective.
Now consider the map $d_0: W^n \to B^{n+1}(\mathcal{C}_0)$, which is a topological isomorphism of Banach spaces. In particular it is bounded below, so that
\[ \norm{d_0(w)} \geq C \norm{w}, \qquad \forall w \in W^n \]
for some constant $C >0$.
Since $t \mapsto d_t$ is norm continuous, the maps $d_t: W^n \to \mathcal{C}_t^{n+1}$ are bounded below for $t$ in a small enough interval $J'$. In particular they are injective. Let's show $\pi: Z^n(\mathcal{C}_t) \to Z^n(\mathcal{C}_0)$ is injective for all $t \in J'$. Consider an element $z \in Z^n(\mathcal{C}_t)$. As vector spaces, $\mathcal{C}_t^n = \mathcal{C}_0^n = Z^n(\mathcal{C}_0) \oplus W^n$, so we can write
\[ z = z_0 + w, \qquad z_0 \in Z^n(\mathcal{C}_0), \quad w \in W^n. \]
If $z \in \ker \pi$, then $z_0 = 0$. Since $z \in Z^n(\mathcal{C}_t)$, we have $0 = d_t(z) = d_t(w)$. Since $d_t$ is injective, $w = 0$ and so $z = 0$. This shows $\pi: Z^n(\mathcal{C}_t) \to Z^n(\mathcal{C}_0)$ is injective for all $t \in J'$, and consequently $\pi: Z^n(\mathcal{C}_{J'}) \to C^\infty(J', Z^n(\mathcal{C}_0))$ is injective.
\end{proof}
A variation of the following theorem was first proved in \cite{MR0634038} using a certain ``inverse function theorem".
\begin{theorem}
Let $A$ be a Banach algebra whose Hochschild cochain complex has a continuous linear contracting homotopy in degree $2$
\[ \xymatrix{
C^1(A, A) \ar@<.5ex>[r]^-{\delta} & C^2(A, A) \ar@<.5ex>[l]^-h \ar@<.5ex>[r]^-{\delta} & C^3(A, A), \ar@<.5ex>[l]^-h
} \qquad \delta h + h \delta = 1,\]
so that $H^2(A,A) = 0$. Then $A$ is rigid.
\end{theorem}
\begin{proof}
Given a smooth deformation $\{A_t\}_{t \in J}$ with $A_0 = A$, consider the deformation of cochain complexes $\{C^\bullet(A_t, A_t)\}_{t \in J}$. Using Proposition \ref{Proposition-BanachSpaceSmoothPathsOfHoms}, its complex of sections naturally identifies with the Hochschild complex $C^\bullet_{C^\infty(J)}(A_J, A_J)$. By Lemma \ref{Lemma-DeformationOfContractibleHasZeroCohomology}, $H^2_{C^\infty(J')}(A_{J'}, A_{J'}) = 0$ for some subinterval $J' \subset J$ containing $0$. So $A_{J'}$ has a connection that is a derivation, and it is integrable because the underlying space is a Banach space. This shows $\{A_t\}_{t \in J'}$ is trivial.
\end{proof}
Of course when $A$ is finite dimensional, then such a contracting homotopy will exist whenever $H^2(A,A) = 0$. As an example, a finite direct sum $A$ of matrix algebras satisfies $H^2(A,A) = 0$. Thus, all finite dimensional $C^*$-algebras are rigid, as associative algebras. For a general Banach algebra with $H^2(A,A) = 0$, we will need to assume the existence of the homotopy $h$. One can possibly circumvent this by considering nonlinear homotopy operators as in \cite{Crainic}.
\begin{theorem} \label{Theorem-ContractibleComplexIsRigid}
Let $(\mathcal{C}^\bullet, d)$ be a contractible cochain complex of Banach spaces that is bounded above and below in degree. Then the complex $(\mathcal{C}^\bullet, d)$ is rigid.
\end{theorem}
\begin{proof}
Let $h: \mathcal{C}^\bullet \to \mathcal{C}^{\bullet - 1}$ be a continuous contracting homotopy. Then the endomorphism complex $\End(\mathcal{C})$ is contractible, with homotopy
\[ H: \End(\mathcal{C})^\bullet \to \End(\mathcal{C})^{\bullet-1}, \qquad H(F) = h \circ F. \] Suppose $\{ (\mathcal{C}_t^\bullet, d_t)\}_{t \in J}$ is a smooth deformation with $\mathcal{C}_0^\bullet = \mathcal{C}^\bullet$. The complexes $\{ \End(\mathcal{C}_t) \}_{t \in J}$ form a smooth deformation of cochain complexes whose complex of sections is identified with $\End_{C^\infty(J)} (\mathcal{C}_J)$ by Proposition \ref{Proposition-BanachSpaceSmoothPathsOfHoms}. Our assumption that the degree is bounded guarantees that $\End(\mathcal{C}_t)$ is a Banach space. By Lemma \ref{Lemma-DeformationOfContractibleHasZeroCohomology}, $H^1(\End_{C^\infty(J')}(\mathcal{C}_{J'})) = 0$ for some subinterval $J'$. From Proposition \ref{Proposition-ClassOfG}, the module $\mathcal{C}_{J'}^\bullet$ admits a connection that is a chain map. Since we are working with Banach spaces, the connection is integrable and so $\{ \mathcal{C}_t^\bullet\}_{t \in J'}$ is trivial by Proposition \ref{Proposition-BundleOfComplexesParallelTransport}.
\end{proof}
Next we consider an application to homological perturbation theory. We recall the construction of the mapping cone. Given a chain map
\[ f: (\mathcal{C}^\bullet, d_{\mathcal{C}}) \to (\mathcal{D}^\bullet, d_{\mathcal{D}}) \] between cochain complexes, the mapping cone complex $(C_f^\bullet, \partial)$ is defined by
\[ C_{f}^\bullet = \mathcal{C}^{\bullet+1} \oplus \mathcal{D}^\bullet, \qquad \partial = \begin{bmatrix} -d_{\mathcal{C}} & 0\\ f & d_{\mathcal{D}} \end{bmatrix}. \] If $C_f^\bullet$ is contractible, then it is easy to see that $f$ is a chain homotopy equivalence. Indeed, one can extract the homotopy inverse as well as the homotopy operators from the contracting homotopy of $C_f^\bullet$. The converse is true as well \cite{MR1909353}.
\begin{theorem}
Suppose $\{\mathcal{C}_t\}_{t \in J}$ and $\{\mathcal{D}_t\}_{t \in J}$ are smooth deformations of bounded cochain complexes of Banach spaces, and $\{f_t: \mathcal{C}_t \to \mathcal{D}_t\}_{t \in J}$ is a smooth family of continuous chain maps. If $f_0$ is a chain homotopy equivalence, then there is a subinterval $J' \subset J$ containing $0$ such that $f_t$ is a chain homotopy equivalence for all $t \in J'$. Moreover, the homotopy inverse and the homotopy operators can be chosen to depend smoothly on $t$.
\end{theorem}
\begin{proof}
By assumption, the mapping cone $C_{f_0}^\bullet$ is contractible, and so it is rigid by Theorem \ref{Theorem-ContractibleComplexIsRigid}. Thus, the deformation $\{ C_{f_t}^\bullet\}_{t \in J'}$ is trivial for some subinterval $J' \subset J$. So its complex of sections is isomorphic to $C^\infty(J', C_{f_0})$, which has a $C^\infty(J')$-linear contracting homotopy. Thus each $C_{f_t}^\bullet$ is contractible in a way that depends smoothly on $t$.
\end{proof}
\section{The Gauss-Manin connection} \label{Section-GaussManinConnection}
\subsection{Gauss-Manin connection in periodic cyclic homology}
In this section, we'll construct Getzler's Gauss-Manin connection in our setting of smooth deformations. Let $A_J$ denote the algebra of sections of a smooth one-parameter deformation of locally convex algebras $\{A_t\}_{t \in J}$. Unless specified otherwise, all chain groups and homology groups associated to $A_J$ that follow are over the ground ring $C^\infty(J)$.
Consider the deformation of chain complexes $\{ (C_{\per}(A_t), b_t +B)\}_{t \in J}$. As in Example \ref{Example-CyclicComplexOfDeformation}, we can identify its complex of sections with $(C_{\per}(A_J), b+B).$ We would like to show, under favorable circumstances, that this deformation of complexes is trivial at the level of homology. To that end, we'd like to construct a connection on $C_{\per}(A_J)$ that is a chain map. As described in Proposition \ref{Proposition-ClassOfG}, this is a problem in cohomology. To start, let $\nabla$ be any connection on $A_J$, and let $E = \delta \nabla$ as in Proposition \ref{Proposition-ClassOfE}. We extend $\nabla$ to the unitization $(A_J)_+$ (over $C^\infty(J)$) by $\nabla e = 0$, and then to $C_n(A_J)$ using Proposition~\ref{Proposition-TensorAndDualConnections}. Then $\nabla$ extends to a connection on the periodic cyclic complex $C_{\per}(A_J)$, which is given by the Lie derivative $L_{\nabla}$. From Proposition \ref{Proposition-ClassOfG}, $C_{\per}(A_J)$ has a connection that is a chain map if and only if the class of
\[ G := [b+B, L_\nabla] = L_E \]
vanishes in $H^1(\End(C_{\per}(A_J)))$, i.e. $L_E$ is chain homotopic to zero via a $C^\infty(J)$-linear homotopy operator.
But the Cartan Homotopy formula
\[ [b+B, I_E] = L_E \]
of Theorem~\ref{Theorem-CHF} implies exactly this. Notice that $E$ is $C^\infty(J)$-linear, so $I_E$ is a $C^\infty(J)$-linear endomorphism of $C_{\per}(A_J)$. We conclude that the \emph{Gauss-Manin connection}
\[ \nabla_{GM} = L_\nabla - I_E \]
is a connection on $C_{\per}(A)$ and a chain map. Amazingly, the cohomological obstruction to the existence of such a connection vanishes for any deformation $\{A_t\}_{t \in J}$.
\begin{proposition}
The Gauss-Manin connection $\nabla_{GM}$ commutes with the differential $b+B$ and hence induces a connection on the $C^{\infty}(J)$-module $HP_{\bullet}(A_J)$. Moreover, the induced connection on $HP_\bullet(A_J)$ is independent of the choice of connection $\nabla$ on $A_J$.
\end{proposition}
\begin{proof}
We have already established the first claim. For another connection $\nabla',$ let \[ \nabla'_{GM} = L_{\nabla'} - I_{E'} \] be the corresponding Gauss-Manin connection. Then \[ \nabla' -\nabla = F, \qquad E' - E = \delta F \]
for some $C^\infty(J)$-linear map $F: A_J \to A_J$. Thus,
\[ \nabla_{GM}' - \nabla_{GM} = L_F - I_{\delta F} = [b+B, I_F], \]
by Theorem~\ref{Theorem-CHF}. We conclude that the Gauss-Manin connection is unique up to continuous $C^\infty(J)$-linear chain homotopy.
\end{proof}
\begin{corollary} \label{Corollary-NablaGMOnTrivialDeformation}
If $A$ admits a connection $\nabla$ which is also a derivation, then the Gauss-Manin connection on $HP_{\bullet}(A)$ is given by \[ \nabla_{GM}[\omega] = [L_\nabla \omega]. \]
\end{corollary}
As a trivial example, we see that the Gauss-Manin connection associated to a constant deformation is just the usual differentiation $\frac{d}{dt}$.
The Gauss-Manin connection is a canonical choice of a connection on $HP_\bullet(A_J)$. It is natural in the sense that morphisms of deformations induce parallel maps at the level of periodic cyclic homology.
\begin{proposition}[Naturality of $\nabla_{GM}$] \label{Proposition-NaturalityOfNablaGM}
Let $A_J$ and $B_J$ denote the algebras of sections of two deformations over the same parameter space $J$, and let $F: A_J \to B_J$ be a morphism of deformations. Then the following diagram commutes.
\[ \xymatrix{
HP_{\bullet}(A_J) \ar[r]^-{F_*} \ar[d]^-{\nabla_{GM}} &HP_{\bullet}(B_J) \ar[d]^-{\nabla_{GM}}\\
HP_{\bullet}(A_J) \ar[r]^-{F_*} &HP_{\bullet}(B_J)\\
} \]
\end{proposition}
\begin{proof}
Let $\nabla^A$ and $\nabla^B$ denote connections on $A_J$ and $B_J$ with respective cocycles $E^A$ and $E^B$, and let $F_*: C_{\per}(A_J) \to C_{\per}(B_J)$ be the induced map of complexes. For
\[ h = F_* I_{\nabla^A} - I_{\nabla^B} F_*, \] we have
\begin{align*}
[b+B, h] & = F_* [b+B, I_{\nabla^A}] - [b+B, I_{\nabla^B}] F_*\\
&= F_* (L_{\nabla^A} - I_{E^A}) - (L_{\nabla^B} - I_{E^B}) F_*\\
&= F_* \nabla_{GM}^A - \nabla_{GM}^B F_*.
\end{align*} This shows that the diagram commutes up to continuous chain homotopy. The problem is that $I_{\nabla^A}$ and $I_{\nabla^B}$ are not well-defined operators on the complexes $C_{\per}(A_J)$ and $C_{\per}(B_J)$ respectively (over $C^\infty(J)$), because $\nabla^A$ and $\nabla^B$ are not $C^\infty(J)$-linear operators. However, one can show that thanks to the Leibniz rule, $h$ descends to a map of quotient complexes such that the following diagram
\[ \xymatrix{
C_{\per}^{\C}(A_J) \ar[r]^-h \ar[d]^-\pi &C_{\per}^{\C}(B_J) \ar[d]^-\pi\\
C_{\per}^{C^{\infty}(J)}(A_J) \ar[r]^-{\bar{h}} &C_{\per}^{C^{\infty}(J)}(B_J)
} \]
commutes, and consequently $[b+B, \bar{h}] = F_* \nabla_{GM}^A - \nabla_{GM}^B F_*$ as desired.
\end{proof}
As a simple application of Proposition~\ref{Proposition-NaturalityOfNablaGM}, we get a proof of the differentiable homotopy invariance property of periodic cyclic homology by considering morphisms between constant deformations.
\begin{corollary}[Homotopy Invariance]
Let $A$ and $B$ be locally convex algebras and let $\{F_t: A \to B\}_{t \in J}$ be a smooth family of algebra maps. Then the induced map
\[ (F_t)_*: HP_{\bullet}(A) \to HP_{\bullet}(B) \] is independent of $t$.
\end{corollary}
\begin{proof}
Let $A_J = C^{\infty}(J, A)$ and $B_J = C^{\infty}(J, B)$ be the algebras of sections corresponding to the constant deformations over $J$ with fiber $A$ and $B$ respectively. Then $\{ F_t:A \to B \}_{t \in J}$ is a morphism between these constant deformations. Let $F: A_J \to B_J$ be the induced $C^\infty(J)$-linear algebra map
\[ F(a)(t) = F_t(a(t)). \] Using the canonical connection $\frac{d}{dt}$ on both $A_J$, we see that $\nabla_{GM}$ is given by $\frac{d}{dt}$ under the identification $C_{\per}(A_J) \cong C^\infty(J, C_{\per}^{\C}(A))$, and similarly for $B$. Given a cycle $\omega \in C_{\per}^{\C}(A)$, we view it as a ``constant" cycle in $C_{\per}(A_J),$ and then Proposition~\ref{Proposition-NaturalityOfNablaGM} implies that
\[ \left[ \frac{d}{dt} F_t (\omega) \right] = \left[ F_t \left(\frac{d\omega}{dt}\right) \right] = 0 \] in $HP(B_J)$. So there is $\eta \in C^\infty(J,C_{\per}^{\C}(B))$ such that
\[ \frac{d}{dt} F_t(\omega) = (b+B)(\eta(t)).\] But, by the fundamental theorem of calculus,
\[ F_t(\omega) - F_s(\omega) = \int_s^t(b+B)(\eta(u)) du = (b+B)\Bigg( \int_{s}^{t} \eta(u) du \Bigg) \] for any $s,t \in J$. Hence $[F_t(\omega)] = [F_s(\omega)]$ in $HP_\bullet(B)$.
\end{proof}
\subsection{Dual Gauss-Manin connection}
We define $\nabla^{GM}$ on $C^{\per}(A_J)$ to be the dual connection of $\nabla_{GM}$ as in Proposition~\ref{Proposition-TensorAndDualConnections}. In terms of the canonical pairing,
\[ \langle \nabla^{GM}\phi, \omega \rangle = \frac{d}{dt} \langle \phi, \omega \rangle - \langle \phi, \nabla_{GM} \omega \rangle. \] It is straightforward to verify that $\nabla^{GM}$ commutes with $b+B$ and therefore induces a connection on $HP^\bullet(A_J)$. The connections $\nabla_{GM}$ and $\nabla^{GM}$ satisfy \[ \frac{d}{dt} \langle [\phi], [\omega] \rangle = \langle \nabla^{GM}[\phi], [\omega] \rangle + \langle [\phi], \nabla_{GM} [\omega] \rangle, \] for all $[\phi] \in HP^\bullet(A_J)$ and $[\omega] \in HP_{\bullet}(A_J).$
\subsection{Interaction with the Chern character}
The algebra $A_J$ can be viewed as an algebra over $\C$ or $C^\infty(J)$, and there is a surjective morphism of complexes
\[ \pi: C_{\per}^\C(A_J) \to C_{\per}^{C^\infty(J)}(A_J). \]
\begin{proposition}
If $\omega \in C_{\per}^{C^{\infty}(J)}(A_J)$ is a cycle that lifts to a cycle $\tilde{\omega} \in C_{\per}^{\C}(A_J)$, then $\nabla_{GM} [\omega] = 0$ in $HP_{\bullet}^{C^{\infty}(J)}(A_J)$.
\end{proposition}
\begin{proof}
Let $\nabla_{GM}^{\C} = L_{\nabla} - I_E$, viewed as a linear operator on $C_{\per}^{\C}(A_J)$. By Theorem~\ref{Theorem-CHF}, \[ \nabla_{GM}^\C = L_{\nabla} - I_{\delta \nabla} = [b+B, I_{\nabla}] \] and so $\nabla_{GM}^\C$ is the zero operator on $HP_{\bullet}^{\C}(A_J)$. Thus, at the level of homology, we have
\[ \nabla_{GM} \circ \pi = \pi \circ \nabla_{GM}^\C = 0 \] where $\pi: HP_\bullet^\C(A_J) \to HP_\bullet^{C^{\infty}(J)}(A_J)$ is the map induced by the quotient map. By hypothesis, $[\omega]$ is in the image of $\pi$.
\end{proof}
Note that the homotopy used in the previous proof does not imply that $\nabla_{GM}$ is zero on $HP_{\bullet}^{C^{\infty}(J)}(A_J)$. The reason is that the operator $I_{\nabla}$ is not a well-defined operator on the quotient complex $C_{\per}^{C^{\infty}(J)}(A_J)$.
\begin{theorem} \label{Theorem-ChPChUAreParallel}
If $P \in M_N(A_J)$ is an idempotent and $U \in M_N(A_J)$ is an invertible, then \[ \nabla_{GM}[\ch P] = 0, \qquad \nabla_{GM}[\ch U] = 0 \] in $HP_\bullet(A_J)$.
\end{theorem}
\begin{proof} This is immediate from the previous proposition because the cycle $\ch P \in C_{\per}^\C(A_J)$ is a lift of the cycle $\ch P \in C_{\per}^{C^\infty(J)}(A_J)$, and similarly for $\ch U$.
\end{proof}
Combining this with the identity
\[ \frac{d}{dt} \langle [\phi], [\omega] \rangle = \langle \nabla^{GM}[\phi], [\omega] \rangle + \langle [\phi], \nabla_{GM}[\omega] \rangle, \] we obtain the following differentiation formula for the pairing between $K$-theory and periodic cyclic cohomology.
\begin{corollary}
If $P \in M_N(A_J)$ is an idempotent and $U \in M_N(A_J)$ is an invertible, then
\[ \frac{d}{dt} \langle [\phi], [P] \rangle = \langle \nabla^{GM}[\phi], [P] \rangle, \qquad \frac{d}{dt} \langle [\phi], [U] \rangle = \langle \nabla^{GM}[\phi], [U] \rangle. \]
\end{corollary}
\begin{remark} Proposition~\ref{Proposition-NaturalityOfNablaGM} can be used to give another proof that
\[ \nabla_{GM}[\ch P] = 0 \] when $P \in A_J$ is an idempotent. Indeed, an idempotent in $A_J$ is equivalent to a morphism of deformations \[ \{ F_t: \C \to A_t\}_{t \in J} \] from the constant deformation with fiber $\C$. The induced algebra map
\[ F: C^{\infty}(J, \C) \to A_J \] sends $1$ to $P$. Applying Proposition~\ref{Proposition-NaturalityOfNablaGM}, we see
\[ \nabla_{GM}[\ch P] = \nabla_{GM} F [\ch 1] = F \frac{d}{dt} [\ch 1] = 0. \]
\end{remark}
\subsection{Integrating $\nabla_{GM}$}
The very fact that $\nabla_{GM}$ exists for all smooth one-parameter deformations implies that the problem of proving $\nabla_{GM}$ is integrable cannot be attacked with methods that are too general. Indeed, one cannot expect periodic cyclic homology to be rigid for all deformations, there are plenty of finite dimensional examples for which it is not.
\begin{example} \label{Example-CyclicHomologyNotPreserved}
For $t \in \R$, let $A_t$ be the two-dimensional algebra generated by an element $x$ and the unit $1$ subject to the relation $x^2 = t\cdot 1.$ Then $A_t \cong \C \oplus \C$ as an algebra when $t \neq 0$, and $A_0$ is the exterior algebra on a one dimensional vector space. Consequently,
\[ HP_0(A_t) \cong \begin{cases} \C \oplus \C, &t \neq 0\\ \C, &t = 0. \end{cases} \] A similar result holds for periodic cyclic cohomology $HP^0(A_t)$.
\end{example}
From the point of view of differential equations, one issue is that the periodic cyclic complex is never a Banach space. Even in the case where $A$ is a Banach algebra, e.g. finite dimensional, the chain groups $C_n(A)$ are also Banach spaces, but the periodic cyclic complex
\[ C_{\per}(A) = \prod_{n=0}^\infty C_n(A) \] is a Fr\'{e}chet space, as it is a countable product of Banach spaces. The operator $\nabla_{GM}$ contains the degree $-2$ term $\iota_E: C_n(A) \to C_{n-2}(A)$. Thus unless $E = 0$, one cannot reduce the problem to the individual Banach space factors, as the differential equations are hopelessly coupled together.
One instance in which $\nabla_{GM}$ is clearly integrable is when the deformation $\{A_t\}_{t \in J}$ is trivial. Using Proposition \ref{Proposition-IntegrableTensorAndDual} and Proposition \ref{Proposition-TrivialDeformationOfAlgebras}, we obtain the following.
\begin{proposition}
If $A_J$ has an integrable connection $\nabla$ that is a derivation, then $\nabla_{GM} = L_\nabla$ is integrable on $C_{\per}(A_J)$, and
$P^{\nabla_{GM}}_{s,t}: C_{\per}(A_s) \to C_{\per}(A_t)$ is the map of complexes induced by the algebra isomorphism $P^\nabla_{s,t}: A_s \to A_t$.
\end{proposition}
While this is not surprising, it is interesting to note is that if we consider another connection $\nabla'$ on $A_J$, the corresponding Gauss-Manin connection $\nabla_{GM}'$ on $C_{\per}(A_J)$ need not be integrable, and in general seems unlikely to be so. However the induced connection $\left(\nabla_{GM}'\right)_*$ on $HP_\bullet(A_J)$ is necessarily integrable by the uniqueness of the Gauss-Manin connection up to chain homotopy.
As proving integrability of $\nabla_{GM}$ at the level of the complex $C_{\per}(A_J)$ is both too difficult and, in some cases, too strong of a result, our general approach will be to find a different complex that computes $HP_\bullet(A_J)$ equipped with a compatible connection.
\section{A rigidity theorem for periodic cyclic cohomology}
In this section, we give our main theorem for rigidity of periodic cyclic cohomology of certain Banach algebras. We first review the necessary concepts from homological algebra.
\subsection{Homological bidimension}
Let $A$ be a (possibly nonunital) Banach algebra, and let $A^e = A_+ \potimes A_+^{\op}$ be its topological enveloping algebra. The algebra $A^e$ is designed so that there is a one-to-one correspondence between locally convex $A$-bimodules and locally convex unital left $A^e$-modules. Here, we shall only discuss modules whose underlying space is a Banach space. The continuous Hochschild cohomology of $A$ with coefficients in a Banach $A$-bimodule $M$ is defined as
\[ H^\bullet(A,M) := \Ext_{A^e}^\bullet(A_+, M). \] See \cite{MR1093462} for a discussion of derived functors in the context of locally convex algebras and modules. When $A$ is unital, we do not have to be careful with unitizations, as
\[ H^\bullet(A,M) = \Ext_{A \potimes A^{\op}}^\bullet(A, M). \] The Hochschild cohomology $H^\bullet(A,A)$ coincides with our previous notation, and $HH^\bullet(A) = H^\bullet(A, A^*)$. The bimodule structure on $A^*$ comes from a general construction: given any $A$-bimodule $M$, the topological dual $M^* = \Hom(M, \C)$ is an $A$-bimodule via
\[ (a \cdot \phi \cdot b)(m) = \phi(bma), \qquad \forall \phi \in M^*. \] By considering the topological bar resolution $B_\bullet(A)$, which is a projective resolution of $A_+$ by $A^e$-modules, we obtain the \emph{standard complex}
\[ C^n(A, M) = \Hom_{A^e}(B_n(A), M) \cong \Hom(A^{\potimes n}, M), \] with differential
\begin{align*}
(\delta D)(a_1, \ldots , a_{n+1}) &= a_1 D(a_2, \ldots , a_{n+1}) + (-1)^{n+1}D(a_1, \ldots , a_n)a_{n+1}\\
& \qquad + \sum_{j=1}^n (-1)^j D(a_1, \ldots , a_ja_{j+1}, \ldots , a_{n+1}),
\end{align*} whose cohomology is $H^\bullet(A,M)$, see \cite[Section III.4.2]{MR1093462}.
The \emph{homological bidimension} of a Banach algebra $A$ is
\[ \db A = \inf \{n ~|~ H^{n+1}(A,M) = 0 \text{ for all Banach }A\text{-bimodules }M\} \] and the \emph{weak homological bidimension} of $A$ is
\[ \dbw A = \inf \{n ~|~ H^{n+1}(A,M^*) = 0 \text{ for all Banach }A\text{-bimodules }M\}. \] Clearly, $\dbw A \leq \db A$. It is a fact that if $H^{n+1}(A,M) = 0$ (resp. $H^{n+1}(A,M^*) = 0$) for all $M$, then $H^{m}(A,M) = 0$ (resp. $H^m(A,M^*) = 0$) for all $M$ and all $m \geq n+1$ \cite[Theorem III.5.4]{MR1093462} (resp. \cite{MR1437458}). A Banach algebra $A$ is called \emph{amenable} if $\dbw A = 0$. As an example, Johnson proved that the convolution algebra $L^1(G)$ of a locally compact group with respect to Haar measure is amenable if and only if the group $G$ is amenable \cite[Theorem 2.5]{MR0374934}. A Banach algebra $A$ for which $\dbw A = n$ is also called $(n+1)$-amenable.
As in \cite{MR823176}, we shall consider the universal differential graded algebra $(\Omega^\bullet A, d)$ associated to $A$. However, we shall use the topological version, constructed using completed projective tensor products. Explicitly, $\Omega^0 A \cong A = C_0(A)$ and
\[ \Omega^n A \cong A_+ \potimes A^{\potimes n} = C_n(A), \] under the identification
\[ a_0da_1 da_2 \ldots da_n \longleftrightarrow (a_0, a_1, a_2, \ldots , a_n). \] Then $\Omega^n A$ is a Banach $A$-bimodule with the left action
\[ a\cdot(a_0da_1 \ldots da_n) = (aa_0)da_1\ldots da_n.\] The right action is determined by the relation
\[ (da_1)a_2 = d(a_1a_2) - a_1(da_2). \] The map
\[ d^{\otimes n}: A^{\potimes n} \to \Omega^n A, \qquad (a_1, \ldots , a_n) \mapsto da_1\cdot \ldots \cdot da_n \] is a Hochschild $n$-cocycle in the standard complex with coefficients in the bimodule $\Omega^nA$. In fact, $d^{\otimes n}$ is the universal Hochschild $n$-cocycle in the sense that any Hochschild cocycle $D: A^{\potimes n} \to M$ into a Banach $A$-bimodule factors through a unique $A$-bimodule map $F: \Omega^nA \to M$, determined by
\[ F(da_1 \ldots da_n) = D(a_1, \ldots , a_n),\] as in \cite{MR1303030}. Thus the cohomology class $[D] \in H^n(A,M)$ is the image of $[d^{\otimes n}]$ under the map \[ H^n(A, \Omega^nA) \to H^n(A,M) \] induced by $F$. It follows that
\[ \db A \leq n \qquad \text{if and only if} \qquad H^{n+1}(A, \Omega^{n+1}A) = 0. \]
Let's now consider cocycles with values in dual Banach modules. Compose the universal $n$-cocycle $d^{\otimes n}$ with the canonical embedding into the double dual to obtain an $n$-cocycle
\[ d^{\otimes n}: A^{\potimes n} \to (\Omega^nA)^{**}. \]
Given any standard $n$-cocycle $D: A^{\potimes n} \to M^*$, consider the bimodule map
\[ F: \Omega^n A \to M^* \]
induced by the universal property of $\Omega^nA$. Define a bimodule map
\[ G: M \to (\Omega^nA)^*, \qquad G(m)(\omega) = F(\omega)(m). \] Then the dual map
\[ G^*: (\Omega^nA)^{**} \to M^* \] is an $A$-bimodule map that satisfies $G^* \circ d^{\otimes n} = D$.
It follows that
\[ \dbw A \leq n \qquad \text{if and only if} \qquad H^{n+1}(A, (\Omega^{n+1}A)^{**}) = 0. \]
Now suppose $\{A_t\}_{t \in J}$ is a smooth deformation of Banach algebras and let $A_J$ be its algebra of sections. One can form the space of abstract $n$-forms $\Omega^n A_J$ over the ground ring $C^\infty(J)$ by taking the projective tensor products over $C^\infty(J)$. Notice that the spaces $\{\Omega^nA_t\}_{t \in J}$ are all canonically isomorphic as Banach spaces, and $\Omega^nA_J$ is isomorphic, as a $C^\infty(J)$-module, to the space of smooth functions from $J$ into the underlying Banach space of $\Omega^nA_t$. There is a universal $C^\infty(J)$-linear cocycle
\[ d^{\otimes n}: A_J^{\potimes_{C^\infty(J)} n} \to \Omega^nA_J, \qquad d^{\otimes n}(a_1, \ldots , a_n) = da_1\ldots da_n. \] By Propositions \ref{Proposition-TensorProductOfFreeModules} and \ref{Proposition-BanachSpaceSmoothPathsOfHoms}, the complex
\[ C^\bullet_{C^\infty(J)}(A_J, \Omega^nA_J) = \Hom_{C^\infty(J)}\left( A_J^{\potimes_{C^\infty(J)} \bullet}, \Omega^nA_J \right) \]
is isomorphic to the complex of sections of the deformation $\{C^\bullet(A_t, \Omega^nA_t)\}_{t \in J}$. Moreover, evaluation at $t \in J$ induces a chain map
\[ \epsilon_t: C^\bullet_{C^\infty(J)}(A_J, \Omega^nA_J) \to C^\bullet(A_t, \Omega^nA_t) \] which maps the universal cocycle for $A_J$ to the universal cocycle for $A_t$. Thus if $H^{n+1}_{C^\infty(J)}(A_J, \Omega^{n+1}A_J) = 0$, we have $H^{n+1}(A_t, \Omega^{n+1}A_t) = 0$ for all $t \in J$, and consequently $\db A_t \leq n$ for all $t \in J$.
We can also consider the $C^\infty(J)$-linear double dual module $(\Omega^nA_J)^{\dual\dual}$ and the cocycle
\[ d^{\otimes n}: A_J^{\otimes_{C^\infty(J)}n} \to (\Omega^nA_J)^{\dual\dual} \] obtained by composing the universal cocycle with the canonical embedding into the double dual. Using Proposition \ref{Proposition-BanachSpaceSmoothPathsOfHoms} and Corollary \ref{Corollary-DualFreeModule}, the Hochschild complex $C^\bullet_{C^\infty(J)}(A_J, (\Omega^nA_J)^{\dual\dual})$ identifies with the complex of sections of the deformation $\{C^\bullet(A_t, (\Omega^nA_t)^{**})\}_{t \in J}$. By considering evaluation at $t \in J$, we see that if $[d^{\otimes n}] = 0$ in $H^n_{C^\infty(J)}(A_J, (\Omega^nA_J)^{\dual\dual})$, then $[d^{\otimes n}] = 0$ in $H^n(A_t, (\Omega^nA_t)^**)$ for all $t \in J$. We have proved the following proposition.
\begin{proposition}
Let $\{A_t\}_{t \in J}$ be a deformation of Banach algebras.
\begin{enumerate}
\item If $H^{n+1}_{C^\infty(J)}(A_J, \Omega^nA_J) = 0$, then $\db A_t \leq n$ for all $t \in J$.
\item If $H^{n+1}_{C^\infty(J)}(A_J, (\Omega^nA_J)^{\dual\dual}) = 0$, then $\dbw A_t \leq n$ for all $t \in J$.
\end{enumerate}
\end{proposition}
\begin{lemma} \label{Lemma-FiniteHochschildDimensionContraction}
Let $A$ be a Banach algebra and $M$ be a Banach $A$-bimodule.
\begin{enumerate}
\item If $\db A \leq n$, then the standard complex $C^\bullet(A,M)$ has a contracting homotopy in degree $n+1$
\[ \xymatrix{
C^n(A, M) \ar@<.5ex>[r]^-{\delta} & C^{n+1}(A, M) \ar@<.5ex>[l]^-h \ar@<.5ex>[r]^-{\delta} & C^{n+2}(A, M), \ar@<.5ex>[l]^-h
} \qquad \delta h + h \delta = 1.\]
\item If $\dbw A \leq n$, then the standard complex $C^\bullet(A,M^*)$ has a contracting homotopy in degree $n+1$
\[ \xymatrix{
C^n(A, M^*) \ar@<.5ex>[r]^-{\delta} & C^{n+1}(A, M^*) \ar@<.5ex>[l]^-h \ar@<.5ex>[r]^-{\delta} & C^{n+2}(A, M^*), \ar@<.5ex>[l]^-h
} \qquad \delta h + h \delta = 1.\]
\end{enumerate}
\end{lemma}
\begin{proof}
If $\db A \leq n$, then $A_+$ has a projective resolution of length $n$ \cite[Theorem III.5.4]{MR1093462}. By the ``Comparison theorem" \cite[Theorem III.2.3]{MR1093462}, a projective resolution is unique up to chain homotopy equivalence in the category of complexes of Banach $A$-bimodules. By applying the functor $\Hom_{A^e}(\cdot, M)$, it follows that the standard complex $C^\bullet(A,M)$ has the required homotopy.
If $\dbw A \leq n$, then $A_+$ has a flat resolution of length $n$ \cite[Theorem 1]{MR1437458}, that is, a resolution by Banach $A$-bimodules which are flat as left $A^e$-modules. Since the dual of a flat module is injective \cite[Theorem VII.1.14]{MR1093462}, it follows that $A^*$ has an injective resolution of length $n$. Using the Comparison theorem for injective resolutions, the dual $B_\bullet(A)^*$ of the bar resolution has a contracting homotopy in degree $n+1$ consisting of $A$-bimodule maps. After applying the functor $\Hom_{A^e}(M, \cdot)$, we see the complex $\Hom_{A^e}(M, B_\bullet(A)^*)$ has a contracting homotopy in degree $n+1$. However there is a natural isomorphism of complexes
\[ \Hom_{A^e}(M, B_\bullet(A)^*) \cong \Hom_{A^e}(B_\bullet(A), M^*) = C^\bullet(A, M^*),\] which gives the result, see \cite[Proposition III.4.13]{MR1093462}.
\end{proof}
\begin{corollary} \label{Corollary-UpperSemicontinuity}
Let $\{A_t\}_{t \in J}$ be a smooth deformation of Banach algebras. Then the functions
\[ t \mapsto \db A_t \qquad \text{and} \qquad t \mapsto \dbw A_t \] are upper semi-continuous.
\end{corollary}
\begin{proof}
Simply combine the previous two results with Lemma \ref{Lemma-DeformationOfContractibleHasZeroCohomology}.
\end{proof}
\begin{example}
Let $G$ be a connected semisimple Lie group with maximal compact subgroup $K$. Let $\mathfrak{g}$ and $\mathfrak{k}$ denote their respective Lie algebras. In \cite{MR2391803}, a smooth deformation $\{G_t\}$ of Lie groups is constructed in such a way that $G_0 = \mathfrak{g}/\mathfrak{k} \rtimes K$ and $G_t \cong G$ for all $t \neq 0$. The group $G_0$ is amenable, but $G$ may not be, e.g. $G = SL(2,\R)$. So Corollary \ref{Corollary-UpperSemicontinuity} and Johnson's theorem imply that there is no corresponding smooth deformation $\{L^1(G_t)\}$ of Banach algebras.
\end{example}
\subsection{Contractions and retractions}
When the universal $(n+1)$-cocycle is a coboundary, one can construct a contracting homotopy in the Hochschild complex in a uniform way. As described in \cite{MR1269386}, if $\phi: A^{\potimes n} \to \Omega^{n+1}A$ satisfies $\delta \phi = d^{\otimes (n+1)}$, then
\[ \alpha: \Omega^k A \to \Omega^{k+1}A, \qquad \alpha(a_0 da_1 \ldots da_k) = a_0 \phi(a_1, \ldots , a_n) da_{n+1} \ldots da_k \]
defines a contracting homotopy of the Hochschild chain complex $(C_\bullet(A), b)$ in degrees $k \geq n+1$. The transpose of $\alpha$ gives a contracting homotopy in degree $k \geq n+1$ for the Hochschild cochain complex $(C^\bullet(A), b)$.
Khalkhali showed in \cite{MR1269386} that if the cocycle $d^{\otimes (n+1)}: A^{\potimes (n+1)} \to (\Omega^{n+1}A)^{**}$ is a coboundary, then one can construct a contracting homotopy of $(C^\bullet(A), b)$ in degrees $k \geq n+1$ in a similar way. Given a cochain $\phi: A^{\potimes n} \to (\Omega^{n+1}A)^{**}$ such that $\delta \phi = d^{\otimes (n+1)}$, define
\[ \alpha: (\Omega^{k+1}A)^* \to (\Omega^kA)^*, \qquad k \geq n \] by
\[ (\alpha f)(a_0 da_1 \ldots da_k) = \left[ a_0 \cdot \phi(a_1, \ldots , a_n) \right](f_{da_{n+1} \ldots da_k}), \]
where $f \in (\Omega^{k+1}A)^*$ and $f_{da_{n+1} \ldots da_k} \in (\Omega^{n+1}A)^*$ is given by
\[ f_{da_{n+1} \ldots da_k}(\omega) = f(\omega da_{n+1} \ldots da_k). \]
Then $b\alpha + \alpha b = 1$ in $C^k(A)$ when $k \geq n+1$.
Given a contracting homotopy $\alpha: C^{k+1}(A) \to C^k(A)$, Khalkhali constructed a retract of the periodic cyclic cochain complex with only finitely many degrees \cite{MR1269386}, which we now describe. Let $N$ be such that $\alpha$ is a contracting homotopy in all degrees above $2N$. Let
\[ C_0^{\even}(A) = \left(\bigoplus_{k=0}^{N-1} C^{2k}(A)\right) \bigoplus \ker \left\{ b: C^{2N}(A) \to C^{2N+1}(A) \right\} \] and
\[ C_0^{\odd}(A) = \bigoplus_{k=0}^{N-1} C^{2k+1}(A). \]
The $\Z/2$-graded complex $C_0^{\per}(A) = C_0^{\even}(A) \oplus C_0^{\odd}(A)$ has differential $b+B$. Then $C_0^{\per}(A)$ is a subcomplex of $C^{\per}(A)$, and in fact is a deformation retract. That is, there is a chain map $R: C^{\per}(A) \to C_0^{\per}(A)$ such that $RI = \id$ and $IR$ is chain homotopic to $\id$, where $I: C_0^{\per}(A) \to C^{\per}(A)$ is the inclusion. Thus, the cohomology of $C_0^{\per}(A)$ is $HP^\bullet(A)$. The key feature is that $C_0^{\per}(A)$ is a complex of Banach spaces. We won't need the explicit form of the retraction $R$, but we remark that it depends heavily on the contracting homotopy $\alpha$.
All of the above can be carried out for the algebra of sections $A_J$ of a smooth deformation $\{A_t\}_{t \in J}$ of Banach algebras, where everything is considered over the ground ring $C^\infty(J)$. If $\phi: A_J^{\potimes n} \to \Omega^{n+1}A_J$ satisfies $\delta \phi = d^{\otimes (n+1)}$, then
\[ \alpha: \Omega^k A_J \to \Omega^{k+1}A_J, \qquad \alpha(a_0 da_1 \ldots da_k) = a_0 \phi(a_1, \ldots , a_n) da_{n+1} \ldots da_k \]
defines a contracting homotopy of the Hochschild chain complex $(C_\bullet(A_J), b)$ in degrees $k \geq n+1$. Its dual
\[ \alpha^\dual: C^{k+1}(A_J) \to C^k(A_J) \] is a contraction for the Hochschild cochain complex.
If $\phi: A_J^{\otimes n} \to (\Omega^{n+1}A_J)^{\dual \dual}$, then Khalkhali's contracting homotopy
\[ \alpha: (\Omega^{k+1}A_J)^\dual \to (\Omega^kA_J)^\dual, \qquad k \geq n\] can be defined by the same formula as above
\[ (\alpha f)(a_0 da_1 \ldots da_k) = \left[ a_0 \cdot \phi(a_1, \ldots , a_n) \right](f_{da_{n+1} \ldots da_k}), \] where $f \in (\Omega^{k+1}A_J)^\dual$ and $f_{da_{n+1} \ldots da_k} \in (\Omega^{n+1}A_J)^\dual$ is given by
\[ f_{da_{n+1} \ldots da_k}(\omega) = f(\omega da_{n+1} \ldots da_k). \] So $b\alpha + \alpha b = 1$ in $C^k(A_J)$ for $k \geq n+1$. Moreover, given an $\alpha$ which is a contracting homotopy in degrees above $2N$, we can define
\[ C_0^{\even}(A_J) = \left(\bigoplus_{k=0}^{N-1} C^{2k}(A_J)\right) \bigoplus \ker \left\{ b: C^{2N}(A_J) \to C^{2N+1}(A_J) \right\} \] and
\[ C_0^{\odd}(A_J) = \bigoplus_{k=0}^{N-1} C^{2k+1}(A_J). \]
As in the $\C$-linear case, there is an inclusion $I: C_0^{\per}(A_J) \to C^{\per}(A_J)$ and a retraction $R: C^{\per}(A_J) \to C_0^{\per}(A_J)$ which are $C^\infty(J)$-linear chain maps such that $RI = \id$ and $IR$ is chain homotopic to $\id$. The retraction $R$ is built in the same way as the $\C$-linear case using the homotopy $\alpha$.
\begin{definition}
We'll say that a locally convex algebra $A$ is $HP^\bullet$-rigid if whenever $\{A_t\}_{t \in J}$ is a smooth deformation with $A_0 = A$, then there some subinterval $J' \subseteq J$ containing $0$ for which $HP^\bullet(A_t) \cong HP^\bullet(A_0)$ for all $t \in J'$.
\end{definition}
We now give our main application of the Gauss-Manin connection.
\begin{theorem} \label{Theorem-HPRigidity}
Let $A$ be a Banach algebra such that $\dbw A < \infty$. Then $A$ is $HP^\bullet$-rigid.
\end{theorem}
\begin{proof}
Let $\{A_t\}_{t \in J}$ be a smooth deformation with $A_0 = A$. Suppose $\dbw A_0 = n$. As described above, the Hochschild complex $C^{\bullet}_{C^\infty(J)}(A_J, (\Omega^{n+1}A_J)^{\dual \dual})$ identifies with the complex of sections of the deformation $\{C^\bullet(A_t, (\Omega^{n+1}A_t)^{**})\}_{t\in J}$. From Lemmas \ref{Lemma-FiniteHochschildDimensionContraction} and \ref{Lemma-DeformationOfContractibleHasZeroCohomology}, $H^{n+1}_{C^\infty(J')}(A_{J'}, (\Omega^{n+1}A_{J'})^{\dual \dual}) = 0$ for a subinterval $J' \subseteq J$ containing $0$. So there is a $\phi: A_{J'}^{\potimes n} \to (\Omega^{n+1}A_{J'})^{\dual \dual}$ with $\delta \phi = d^{\otimes (n+1)}$. As described above, we can construct from this the deformation retract $C_0^{\per}(A_{J'})$ of $C^{\per}(A_{J'})$ for a suitable $N$. A priori, the space of cocycles $\ker \{ b: C^{2N}(A_{J'}) \to C^{2N+1}(A_{J'}) \}$ may not be a free $C^\infty(J')$-module. However, the conclusion of Lemma \ref{Lemma-DeformationOfContractibleHasZeroCohomology} guarantees that it is, and moreover $C_0^{\per}(A_{J'})$ is the complex of sections of $\{C_0^{\per}(A_t)\}_{t \in J'}$.
We can now transfer the Gauss-Manin connection to $C_0^{\per}(A_{J'})$. Let
\[ I: C_0^{\per}(A_{J'}) \to C^{\per}(A_{J'}), \qquad R: C^{\per}(A_{J'}) \to C_0^{\per}(A_{J'}) \] be the inclusion and retraction, which are continuous $C^\infty(J')$-linear chain maps. Define $\widetilde{\nabla} = R \circ \nabla_{GM} \circ I$ on $C_0^{\per}(A_{J'})$. Then $\widetilde{\nabla}$ is a chain map and it is a connection because $RI = \id$. Since the underlying space $C_0^{\per}(A_t)$ is a Banach space, the connection $\widetilde{\nabla}$ is integrable, and the result follows from Proposition \ref{Proposition-BundleOfComplexesParallelTransport}.
\end{proof}
Let $HE^\bullet(A)$ denote the entire cyclic cohomology of $A$, see \cite{MR953915}. As Khalkhali showed, the canonical inclusion $HP^\bullet(A) \to HE^\bullet(A)$ is an isomorphism for Banach algebras of finite weak bidimension \cite{MR1269386}. We immediately obtain the following.
\begin{corollary}
Let $A$ be a Banach algebra such that $\dbw A < \infty$. Then $A$ is $HE^\bullet$-rigid.
\end{corollary}
\begin{example}
We'll show how our theorem can be used to give a proof of an instance of a theorem of Block on the cyclic homology of filtered algebras \cite{MR934456}. In Block's setting we have an increasing filtration of an algebra $A$,
\[ F_0 \subset F_1 \subset F_2 \subset \ldots \]
where $A = \bigcup_n F_n$ and $F_n\cdot F_m \subset F_{n+m}$. Letting $B = \gr(A)$ be the associated graded algebra, his result is that if $HH_n(A) = 0$ for all large enough $n$, then the inclusion $F_0 \to A$ induces an isomorphism $HP_\bullet(F_0) \cong HP_\bullet(A)$.
Let's consider the situation of a finite filtration of a Banach algebra
\[ F_0 \subset F_1 \subset \ldots \subset F_N = A, \] and suppose there exist closed subspaces $B_k \subset A$ for which each $F_n \cong \bigoplus_{k=0}^n B_k$ as Banach spaces. Then we can identify the associated graded algebra $\gr(A) \cong \bigoplus_{k=0}^N B_k$ with $A$ as Banach spaces. The multiplication in $\gr(A)$ is such that $B_n\cdot B_m \subset B_{n+m}$. Given $a \in B_n$ and $b \in B_m$, the product in the filtered algebra $A$ can be written as
\[ ab = \sum_{k=0}^{n+m} \pi_{n,m}^k(a,b) \] for uniquely defined operators
\[ \pi_{n,m}^k: B_n \potimes B_m \to B_{n+m-k}. \] Given $t \in \R$, we can define a new associative product $m_t$ on $A$ by
\[ m_t(a,b) = \sum_{k=0}^{n+m} t^k \pi_{n,m}^k(a,b), \qquad a \in B_n, b \in B_m. \] This clearly gives a smooth deformation $\{A_t\}_{t \in \R}$ of Banach algebras, as the products depend polynomially on $t$. We have $A_1 = A$, $A_0 = \gr(A)$, and $A_t \cong A$ for all $t \neq 0$. If $\dbw \gr(A) < \infty$, then the Gauss-Manin connection is integrable for this deformation, and $HP^\bullet(A) \cong HP^\bullet(\gr(A))$. View the inclusions $\{F_0 \to A_t\}_{t \in J}$ as a morphism of deformations out of the constant deformation. From Proposition \ref{Proposition-NaturalityOfNablaGM}, this morphism induces a $\nabla^{GM}$-parallel map. Since $HP^\bullet(\gr(A)) \to HP^\bullet(F_0)$ is an isomorphism (Example \ref{Example-GradedAlgebraCyclicHomology}), it follows that $HP^\bullet(A) \to HP^\bullet(F_0)$ is an isomorphism.
\end{example}
Notice that Example \ref{Example-CyclicHomologyNotPreserved} is such a deformation of a filtered algebra $A_1$ into its associated graded algebra $A_0$. However $\dbw A_0 = \infty$, as one can show that $HH^n(A_0) \cong \C$ for all $n > 0$.
\begin{bibdiv}
\begin{biblist}
\bibselect{SmoothDeformationsGMConnection}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
Symmetry-protected topological (SPT) phases have topological order that is not characterized by a local order parameter and their existence requires symmetry to be preserved~\cite{top1d_wenold, top1d_wennew,top1d_lukasz,top1d_pollmanturner,top1d_schuchgarciacirac}. Ground states of topologically non-trivial SPT phases cannot be continuously connected to trivial product states without either closing the gap or breaking the protected symmetry. In one dimension, a particularly useful way to describe ground states is the matrix-product-state (MPS) representation~\cite{fannes1992,Perez-Garcia2007,hastings2007area} and this has led to many interesting results including a complete classification of SPT phases~\cite{top1d_wennew}. In addition to classifying SPT phases, an intriguing connection of SPT phases to quantum computation was identified in Ref.~\cite{bartlett2012} that SPT ground states of $\mathbb{Z}_2 \times \mathbb{Z}_2$ symmetry can serve as resource states for realizing certain gate operations in quantum computation by local measurement.
Measurement-based quantum computation (MBQC)~\cite{raussendorf_2001,raussendorf2003,raussendorf_2009} is a quantum computational scheme that makes use of only local measurements on a suitably entangled resource state. It was originally invented with a specific resource state, i.e., the cluster state~\cite{raussendorf_2001} but was subsequently shown to be supported by a variety of systems~\cite{VandenNest2006,gross2007_2,gross2007,XieChen2009}, in particular, the Affleck-Kennedy-Lieb-Tasaki (AKLT) states~\cite{aklt,WeiRobert_Review} on various one- and two-dimensional systems~\cite{miyake_aklt,Wei2011,Miyake2011,wei_aklt2012, wei_aklt2013,Cai_Miyake_2010}. In Ref.~\cite{bartlett2012} it was observed that both the 1D cluster and AKLT states, which are capable of supporting arbitrary single-qubit gates, belong to a 1D SPT phase protected by $\mathbb{Z}_2 \times \mathbb{Z}_2$ on-site symmetry. Moreover other ground states of this phase also support a protected identity gate operation and can act as perfect wires for transmission of quantum information. The results in Ref.~\cite{bartlett2012} hinge on features of specific Abelian groups, $i.e.$, groups whose projective representation possesses a maximally non-commutative factor system. This brings forth several interesting questions:
\begin{enumerate}
\item Can we extend the results of Ref~\cite{bartlett2012} to get the ground-state form of SPT phases protected by an arbitrary group (both Abelian and non-Abelian)?
\item Are there SPT phases protected by other groups which protect the perfect operation of the identity gate?
\item Are there SPT phases where other non-trivial operations are also allowed? Is it possible to find an entire SPT phase whose ground states support universal one-qubit gates?
\end{enumerate}
Here we develop a formalism that addresses (1) and allows us to treat an arbitrary finite group $G$, either Abelian or non-Abelian, so that we can examine the associated SPT ground states and protected gate operations. The results of Ref.~\cite{bartlett2012} on the spin-1 system with $\mathbb{Z}_2 \times \mathbb{Z}_2$ are reproduced in this formulation. To address (2), we find that in addition to $\mathbb{Z}_2 \times \mathbb{Z}_2$, 1D topologically non-trivial SPT phases associated with the symmetry groups $A_4$ (the alternating group of degree 4) and $S_4$ (the symmetric group of degree 4), see Sec~\ref{sec:examples}, acting on a three-dimensional on-site irreducible representation (i.e., physical spin-1 entities) also protect the identity gate operation. The latter group was also studied in Ref.~\cite{MillerMiyake14}.
We only make partial progress in answering (3). We consider an example Hamiltonian with $A_4$ and parity invariance and study its ground states in various parameter regimes. This Hamiltonian can be regarded as perturbing the AKLT Hamiltonian. We find an extended region in the parameter space where the ground state is exactly the AKLT state and hence can be used as a resource state capable of universal single-qubit gate operations. Whether or not it is generic that the imposition of an appropriate set of symmetries can allow the entire region of an SPT phase to support protected universal single-qubit gates remains an open question. There has however been progress in reducing certain SPT ground states into resource states that support universal single-qubit operations by a `buffering' technique~\cite{MillerMiyake14}, which in some sense gives an affirmative answer to (3).
The rest of the paper is organized as follows. In Sec~\ref{sec:review}, we review the matrix-product state formalism and its connection to quantum computation, and their utility in SPT phases. In Sec~\ref{sec:mainresults}, we present the key results of our formalism that can determine, in terms of MPS, the structure of SPT ground states constrained by symmetry. The method we used was inspired by Refs.~\cite{vidaltensor,vidaltensor_u1,vidaltensor_su2,vidaltensor_bonddimension} where they consider imposing global symmetries such as $SU(2)$ and $U(1)$ for application in numerical simulations. The formalism we develop here might also find its application in numerical simulations with discrete symmetries imposed~\cite{sukhi_injectivity}. In Sec.~\ref{sec:examples}, we use our formalism to examine SPT phases and their non-trivial ground states protected by symmetries such as $\mathbb{Z}_2 \times \mathbb{Z}_2$, $D_4$, $A_4$ and $S_4$. In Sec.~\ref{sec:model_hamiltonian}, we construct a specific Hamiltonian that is $A_4$ symmetric by perturbing the AKLT Hamiltonian and study its ground states. We find an extended region where the ground states are identical to the AKLT state, which allows universal single qubit operations. We conclude in Sec.~\ref{sec:summary}.
\section{Review of relevant definitions and results}\label{sec:review}
\subsection{Definition of a gapped phase of matter}
In Ref.~\cite{top1d_wenold}, it was argued that in order to talk about phases of matter, we need to specify the class of Hamiltonians we are considering. Two gapped Hamiltonians from a given class are in the same phase if we can `connect' them smoothly without closing the spectral gap. Otherwise, there is a boundary in the space of Hamiltonians where the gap closes separating different phases of matter~\cite{top1d_wenold,top1d_wennew}. In 1D, if we consider the class of all gapped local Hamiltonians, it has been shown~\cite{top1d_wenold} that they all belong to the same phase and we can connect any two such Hamiltonians without closing the gap by adding suitable local operators. Thus, there is no \emph{intrinsic} topological order in 1D and all Hamiltonians can be connected to those in the trivial phase with product ground states. In other words, any ground state can be connected to a product state. On the other hand, if we restrict ourselves to a class of Hamiltonians that respect some global symmetry, there are generally phase boundaries which arise. We cannot connect Hamiltonians in different phases through symmetry respecting operators without closing the gap. Different phases are characterized by a combination of \emph{symmetry fractionalization} and \emph{symmetry breaking}~\cite{top1d_wennew}. When symmetry is not broken, the unique ground states of these \emph{Symmetry Protected Topological} (SPT) phases respect the symmetry of the Hamiltonian and allow us to write down their form using tools from the representation theory of groups. Much of this is possible by using the matrix-product-state representation of gapped ground states of 1D spin chains which we shall briefly review below.
\subsection{Matrix product states}\label{sec:MPSreview}
We begin by giving a brief review of the \emph{Matrix Product State} representation of many-body wavefunctions in 1D~\cite{Perez-Garcia2007}. Consider a one-dimensional chain of $N$ spins. If the Hilbert space of each spin is $d$-dimensional, the Hilbert space of the spin chain itself is $d^N$-dimensional. This means that the number of coefficients needed to describe the wavefunction of the spin chain grows exponentially with the length of the chain. However, if the spin chain is in the ground-state configuration of a gapped Hamiltonian, it can be efficiently written as an MPS wavefunction~\cite{vidal03,hastings2007area,arad_itai}. To do this, we need to associate for every spin site (labeled by $m=1 \dots N$), a $D_m \times D_{m+1}$-dimensional matrix $A^{i_m}_m$ for each basis state $\ket{i_m} = \ket{1} \dots \ket{d}$. $D = max_m(D_m)$ is the maximum `virtual' or `bond' dimension and approaches a constant value that is independent of the size of the chain for gapped spin chains~\cite{hastings2007area}. With these matrices (which we shall refer to as MPS matrices), we can write the wavefunction with periodic boundary conditions as:
\begin{equation} \label{eq:mps_periodic}
\ket{\psi} = \sum_{i_1 \dots i_N} Tr[ A^{i_1}_1 A^{i_2}_2 \dots A^{i_N}_N] \ket{i_1} \dots \ket{i_N}.
\end{equation}
We can also write down the wavefunction for a finite chain as
\begin{equation} \label{eq:mps_open}
\ket{\psi} = \sum_{i_1 \dots i_N} \bra{L} A^{i_1}_1 A^{i_2}_2 \dots A^{i_N}_N \ket{R} \ket{i_1} \dots \ket{i_N},
\end{equation}
where, the vectors $\ket{L}$ and $\ket{R}$ live in the virtual space and encode the boundary conditions for the finite chain. If we consider the class of local gapped Hamiltonians without any symmetry constraint, Eqs.~(\ref{eq:mps_periodic},\ref{eq:mps_open}) would represent the general form of ground states. This means we need about $Nd$ matrices to specify the ground state.
\subsection{Matrix product states and measurement-based quantum computation}\label{sec:MPS_MBQC}
To demonstrate the motivation for this work, we first see how we can use MPS wavefunctions for MBQC in the virtual space. Consider encoding quantum information that needs to be processed in one of the virtual boundary vectors of Eq.~(\ref{eq:mps_open}), say $\ket{R}$~\cite{gross2007,gross2007_2,upload1,upload2}. If we perform a projective measurement of the $N$-th spin in some basis $\{ \ket{\phi_N^i} \}$ with the outcome being a projection of the spin onto state $\ket{\phi'_N} \in \{ \ket{\phi^i_N} \}$, we can write the wavefunction of the remaining $N-1$ spins as $\ket{\psi'} = \innerproduct{\phi'_N}{\psi}$ $i.e$
\begin{equation}
\ket{\psi'} = \sum_{i_1 \dots i_{N-1}} \bra{L} A^{i_1}_1 A^{i_2}_2 \dots A^{i_{N-1}}_{N-1} \ket{R'} \ket{i_1} \dots \ket{i_{N-1}},
\end{equation}
where $\ket{R'}=A'_N \ket{R} $ can be regarded as resulting from $\ket{R}$ undergoing a linear transformation $ A'_N = \sum_{i_N} \innerproduct{\phi'_N}{i_N} A^{i_N}_N$.
Thus, if we know all the MPS matrices $A^{i_m}_m$ and if these matrices span the space of relevant operations on the virtual vector, we can hope to induce any transformation on the vector by measurement in an appropriate choice of basis. Usually, there is also an overall residual operator which we can account for by adapting subsequent bases of measurement.
Let us demonstrate this using two translationally invariant canonical resource states. First, the cluster state~\cite{raussendorf2003,WeiRobert_Review} is a $d=2$ spin chain whose wavefunction can be written in terms of $D=2$ MPS matrices:
\begin{eqnarray}
\label{eq:cluster_mps}
A^0 = \begin{pmatrix}
1 & 0 \\
1 & 0
\end{pmatrix},~
A^1 = \begin{pmatrix}
0 & 1 \\
0 & -1
\end{pmatrix}
\end{eqnarray}
Measuring in the $\ket{\pm}=\frac{1}{\sqrt{2}}(\ket{0} \pm \ket{1})$ basis results in the operation $\ket{R}\mapsto H (\sigma_z)^s\ket{R}$ where $s$ labels the measurement outcome and is $0/1$ if the outcome is $\ket{\pm}$ and $H$ is the Hadamard gate $H \equiv \frac{1}{\sqrt{2}}\begin{pmatrix}
1 & 1\\
1 & -1
\end{pmatrix}$. The measurement thus induces the Hadamard operation up to residual operators $(\sigma_x)^s$ as $H (\sigma_z)^s = (-1)^s (\sigma_x)^s H$.
We can induce a different operation, say $R_z(\theta)=e^{-i\theta \sigma_z/2}$ by measuring in the basis $\ket{\phi,\pm}=\frac{1}{\sqrt{2}}(\ket{0}\pm e^{i\phi}\ket{1})$. This results in the operation $\ket{R}\mapsto H (\sigma_z)^s e^{-i\phi \sigma_z/2} \ket{R}$ where $s$ is the measurement outcome which is $0/1$ if the outcome is $\ket{\phi, \pm}$. This is a single-qubit rotation by $\phi$ about the Z axis up to the operator $H (\sigma_z)^s$.
Similarly, we can also perform rotations about the other orthogonal axes and using sequential rotations about different axes by appropriate angles (using, for example, the Euler angle parametrization for rotations), we can perform any arbitrary single-qubit rotation.
The second prominent resource state is the AKLT state~\cite{aklt,miyake_aklt,WeiRobert_Review} which is a spin-1 $(d=3)$ system whose wavefunction can be described by $D=2$ MPS matrices:
\begin{equation}
A^i = \sigma_i~~ (i=x,y,z),
\end{equation}
where the basis of the spins $\{\ket{x}, \ket{y}, \ket{z}\}$ is chosen as
\begin{eqnarray}
\ket{x} \equiv \frac{1}{\sqrt{2}} (\ket{-1} - \ket{1}), \ket{y} \equiv \frac{i}{\sqrt{2}} (\ket{-1} + \ket{1}), \ket{z} \equiv\ket{0}, \nonumber
\end{eqnarray}
with $\ket{\pm1}$ and $\ket{0}$ being eigenstates of the spin-1 $S_z$ operator.
If we measure the spin in \{$\ket{x}$, $\ket{y}$, $\ket{z}$\} basis, we can induce the operation $\ket{R} \mapsto \sigma_s \ket{R}$ which is the identity operation up to the residual operator $\sigma_s$. We can also induce $R_z(\theta) = e^{-i\theta \sigma_z/2}$ by measuring in the basis $\{\ket{\theta,x} = \cos(\frac{\theta}{2}) \ket{x} - \sin(\frac{\theta}{2}) \ket{y}, \ket{\theta,y}= \sin(\frac{\theta}{2}) \ket{x} + \cos(\frac{\theta}{2}) \ket{y}, \ket{z}\}$. If the measurement outcome is $\ket{z}$ then we have the identity operation with residual operator $\sigma_z$. However, if the outcome is $\ket{\theta,x}$ or $\ket{\theta,y}$ then the operation is $\ket{R} \mapsto \sigma_i e^{(-i\theta \sigma_z/2)} \ket{R}$ where $i = x/y$ if the outcome is $\ket{\theta,i}$. Thus, if we keep measuring till we get either $\ket{\theta,x}$ or $\ket{\theta,y}$ as the outcome, we can induce the required operation up to Pauli residual operators. The extension to rotations about other axes and ultimately to a full set of single-qubit rotations is straightforward. An important difference between the AKLT and cluster states is that for the latter, the length of the spin chain needed for computation is fixed while for the former, it is not.
It was noted that both the 1D AKLT and cluster states belong to a non-trivial topological phase protected by $\mathbb{Z}_2 \times \mathbb{Z}_2$ symmetry~\cite{wen_Tensor_Entanglement,top1d_pollmanturner} and there have been investigations to see if the ability to support quantum computation can be a property of the phase~\cite{miyake_MBQC_edge,bartlett_haldane,bartlett2012,bartlett_njp2012,bartlett_correlation}. In particular, the authors of~\cite{bartlett2012,bartlett_njp2012} deduce that any non-trivial MPS ground state in the non-trivial $\mathbb{Z}_2 \times \mathbb{Z}_2$ invariant spin-1 Hamiltonians (Haldane phase) must have the form $A^i = B_i \otimes \sigma_i$ $(i=x,y,z)$. Thus, there always exists a `protected' two-dimensional virtual subspace in the ground states of the Haldane phase on which the Pauli matrices act and in which quantum information can in principle be encoded and processed. While the ground states of the Haldane phase in general do not support non-trivial gate operations, they do allow a protected identity gate operation by measurements in the $\{\ket{x},\ket{y},\ket{z} \}$ basis that only induces Pauli operation on the boundary vectors.
\section{Main result: Tensor decomposed ground-state form in the presence of a global symmetry}\label{sec:mainresults}
\subsection{SPT phases with an on-site internal symmetry }\label{sec:on-site}
Let us now consider symmetric phases of Hamiltonians that are invariant under the action of a certain symmetry group $G$ on each spin according to some representation $u(g)$. i.e. $[H,\hat{U}(g)]=0$ where $\hat{U}(g) = u_1(g) \otimes \cdots \otimes u_N(g) $. We consider ground states that do not break the symmetry of the Hamiltonian and are hence left invariant under the transformation $\hat{U}(g)$ up to a complex phase
\begin{equation} \label{eq:wavefunction_invariance}
\hat{U}(g) \ket{\psi} = \chi(g)^N \ket{\psi}.
\end{equation}
Eq.~(\ref{eq:wavefunction_invariance}) can be imposed as a condition on the MPS matrix level (Suppressing the site labels for brevity) as~\cite{top1d_wenold, top1d_wennew, top1d_lukasz,top1d_pollmanturner}
\begin{equation}
\label{eq:mps_invariance}
u(g)_{ij} A^j = \chi(g)V^{-1}(g) A^i V(g).
\end{equation}
Note that here and henceforth, when no confusion will arise, we use the Einstein summation convention wherein repeated indices are summed over. Because $u$ is a group representation, group properties impose $\chi$ to be a 1D representation and $V$ to be a \emph{projective representation} of $G$. A projective representation respects group multiplication up to an overall complex phase.
\begin{equation}\label{eq:projective_defn}
V(g_1) V(g_2) = \omega(g_1,g_2) V(g_1 g_2).
\end{equation}
The complex phases $\omega(g_1,g_2)$ are constrained by associativity of group action and fall into classes labelled by the elements of the second cohomology group of $G$ over complex numbers $H^2(G,\mathbb{C})$ (See Appendix~\ref{app:proj} for some comments on projective representations). In other words, the different elements of $H^2(G,\mathbb{C})$ label different classes of projective representations. It was also shown in~\cite{top1d_wenold,top1d_wennew,top1d_lukasz,top1d_pollmanturner,top1d_schuchgarciacirac} that the different elements of $H^2(G,\mathbb{C})$ represent different SPT phases of matter. In particular, the identity element labels the set of \emph{linear representations} of $G$ (which respect group multiplication exactly) and the corresponding phase of matter is trivial, containing product ground states. We now use the symmetry constraint of Eq.~(\ref{eq:mps_invariance}) to deduce the form of the MPS matrices for a given phase labelled by $\omega \in H^2(G,\mathbb{C})$ using a technique similar to the one presented in~\cite{vidaltensor}.
With only on-site symmetry, the different 1D representations $\chi$ all correspond to the same SPT phase~\cite{top1d_wenold,top1d_wennew}. Hence, we just consider the case when $\chi(g) = 1$ $i.e.$ the trivial 1D irreducible representation (irrep) of $G$. With this, we can rewrite Eq.~(\ref{eq:mps_invariance}) in a more illuminating form:
\begin{equation}
\label{eq:invariant_tensor}
u(g)_{ii'} V(g)_{\alpha \alpha'}V^{-1}(g)_{\beta' \beta} A^{i'}_{\alpha' \beta'} = A^{i}_{\alpha \beta}.
\end{equation}
Eq.~(\ref{eq:invariant_tensor}) shows that the matrices $A^i$ are invariant 3 index tensors. We now organize the vector space of each index as a reduced representation constructed out of copies of linear or projective irreps of $G$.
\begin{equation} \label{eq:vectorspace}
\mathbb{V} \cong~ \bigoplus_a n_a \mathbb{V}_a \cong ~\bigoplus_a \mathbb{D}_a \otimes \mathbb{V}_a.
\end{equation}
If $\mathbb{V}$ is the vector space of any index, $a$ runs over the irreps, $n_a$ is the degeneracy (number of copies) of the irrep $a$ and $\mathbb{D}_a$ is the corresponding degeneracy vector space of $a$. Any basis element in the vector space $ \mathbb{V}$ can be labelled by three numbers as $\ket{a_i, m_i, d_i}$ where $a_i$ labels the irreducible representation and is analogous to the angular momentum label in $SU(2)$, $m_i$ labels the state in $a_i$ and is analogous to the azimuthal quantum number $m_i$ and $d_i$ labels which copy of the irreducible representation $a_i$ is being considered. Symmetry transformations are block-diagonal and act on the $m_i$ labels of each sector $a_i$ but leave the $d_i$ labels alone. So if $U(g)$ is a symmetry that acts on the vector space Eq.~(\ref{eq:vectorspace}) and if $U^a(g)$ is the representation of the $a$-th irrep then
\begin{equation}
U(g) \cong~ \bigoplus_{a}\mathbb{1}^a \otimes U^a(g).
\end{equation}
Note that for a given physical system, we assume that the vector space of the physical index is known in terms of which irreps and how many copies are contained. However, for a given $\omega \in H^2(G,\mathbb{C})$ which labels the phase we are trying to study the ground-state form of, we have to allow an arbitrary number of copies of each projective irrep from the class $\omega$ to appear in the virtual space indices. Using this organization, Eq.~(\ref{eq:invariant_tensor}) and an application of Schur's lemma after decomposing the fusion of the irreps $a_i$ and $a_\alpha$ determined by the Clebsch-Gordan (CG) series $i \otimes \alpha=\oplus_\gamma n_{i \alpha}^\gamma \gamma$ (see Appendix~\ref{app:cg} for more details), we can write down the MPS matrices for the SPT phase labelled by $\omega$ using a generalized Wigner-Eckart theorem as follows
\begin{multline}
\label{eq:ACB}
A[\omega]^{a_i m_i d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta)} = \\ \sum_{n=1}^{n^\beta_{i \alpha}} B^{a_i d_i}_{(a_\alpha d_\alpha) (a_\beta d_\beta;n)} C[\omega]^{a_\beta m_\beta;n}_{a_i m_i, a_\alpha m_\alpha},
\end{multline}
where $C[\omega]^{a_\beta m_\beta;n}_{a_i m_i, a_\alpha m_\alpha}$ denotes the CG coefficients associated with the change of basis of the direct product of linear irrep $i$ and the irrep $\alpha$ of projective class $\omega$, to the $n$-th copy of irrep $\beta$ of the same projective class $\omega$ (See Appendix~\ref{app:cg} for more details)
\begin{multline}
\label{eq:CG_definition}
\ket{a_\beta, m_\beta ; n} = \\ \sum_{a_i,m_i,a_\alpha, m_\alpha} C[\omega]^{a_\beta m_\beta;n}_{a_i m_i, a_\alpha m_\alpha} \ket{a_i, m_i} \ket{a_\alpha, m_\alpha}.
\end{multline}
The entries $B^{a_i d_i}_{(a_\alpha d_\alpha) (a_\beta d_\beta;n)}$ of the MPS matrices are not determined by on-site symmetry considerations alone and depend on the parameters of the Hamiltonian amongst other things. Finally, putting back the site dependence, $\mathfrak{m}=1 \cdots N$ in the MPS matrices, we have
\begin{multline}\label{eq:mps_onsite}
A[\omega]^{a_i m_i d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta);\mathfrak{m}} = \\ \sum_{n=1}^{n^\beta_{i \alpha}} B^{a_i d_i}_{(a_\alpha d_\alpha) (a_\beta d_\beta;n); \mathfrak{m}} C[\omega]^{a_\beta m_\beta;n}_{a_i m_i, a_\alpha m_\alpha} ,
\end{multline}
We see that to construct the ground-state form of an SPT phase labelled by $\omega$, we need the CG coefficients for the direct product of the linear representation of the physical spins and the projective irreps of class $\omega$: $\ket{i}$ and $\ket{\alpha}$. To make sense this, we use the result that every finite group $G$ has associated to it at least one other finite group $\tilde{G}$, called a Schur cover, with the property that every projective representation of $G$ can be lifted to a linear representation of $\tilde{G}$~\cite{karpilovsky}. So we can reinterpret the CG coefficients of a linear and projective representation of $G$ simply as the CG coefficients of two linear representations of $\tilde{G}$. For example, half odd integer $j$ representations are projective representations of $SO(3)$ while integer $j$ are linear representations. However, if we consider the group $SU(2)$ which is the cover of $SO(3)$, both half odd integer and integer $j$ are linear representations and we know that we can find CG coefficients for decompositions of the kind $1 \otimes \frac{1}{2} = \frac{1}{2} \oplus \frac{3}{2}$.
To summarize, in order to find the ground-state forms of different SPT phases of a spin chain that transforms under a certain representation $u(g)$ of $G$, we need to follow the following steps:
\begin{enumerate}
\item Obtain the second cohomology group of $G$, $H^2(G,\mathbb{C})$ whose elements $\omega$ will label the different SPT phases.
\item Obtain the covering group $\tilde{G}$
\item Identify the irreps `$i$' of the physical spin among the irreps of $\tilde{G}$.
\item Identify the irreps `$\alpha$' that correspond to the projective class $\omega$.
\item Obtain CG coefficients corresponding to the fusion of the irreps of the physical spin with each irrep of the projective class $\omega$. (Ref.~\cite{sakata} and Appendix~\ref{app:sakata} gives a technique to calculate the CG coefficients for certain types of decompositions of finite group irreps)
\item Use the CG coefficients in Eq.~(\ref{eq:mps_onsite}) allowing $\alpha$ and $\beta$ to run over all the irreps of class $\omega$ and $i$ to run over the irreps of the physical spin. Each block of the MPS matrices split into a part that is calculated purely from the group $G$ for each phase $\omega$ and a part that is undetermined.
\end{enumerate}
\subsection{Obtaining the tensor decomposition of Eq.~(\ref{eq:ACB})}
\label{app:cg}
For what follows, it is useful to employ a basis independent representation of the tensor $A$,
\begin{equation}
\label{eq:basisindeptensor}
\hat{A} = \sum_{i \alpha \beta } A^i_{\alpha \beta} \outerproduct{i \alpha} {\beta}
\end{equation}
We organize the vector space of each index and label it by three quantum numbers--the irrep $a_j$ (analogous to the spin label $j$), the irrep multiplicity $m_j$ (analogous to the azimuthal quantum number $m_j$) and the irrep degeneracy (the number of copies of the irrep, $d_j$), i.e., $\ket{i} = \ket{a_i,m_i,d_i}$, $\ket{\alpha} = \ket{a_\alpha,m_\alpha,d_\alpha}$ and so on.
\begin{multline}
\!\!\!\!\!\! \hat{A} = A^{a_i m_i d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta)}
\outerproduct{a_i, m_i,d_i; a_\alpha,m_\alpha,d_\alpha} {a_\beta,m_\beta,d_\beta}. \nonumber
\end{multline}
The invariance condition is
\begin{equation} \label{eq:invariancetensor}
\hat{U}(g) \hat{A} = \hat{A},
\end{equation}
where $\hat{U}(g)$ effects a symmetry transformation on the basis bras and kets of each irrep as
\begin{multline} \label{eq:transformtensor}
\hat{U}(g) \hat{A} \equiv \\ A^{a_i m_i d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta)} U(g)^i_{m_i m'_i} V(g)^\alpha_{m_\alpha m'_\alpha} V(g)^{-1\beta}_{m'_\beta m_\beta} \\ \outerproduct{a_i,m'_i,d_i; a_\alpha,m'_\alpha,d_\alpha} {a_\beta,m'_\beta,d_\beta} .
\end{multline}
Note that symmetry transformations act on the $m$ indices for each irrep but leave the $d$ indices unchanged. Eqs.~(\ref{eq:invariancetensor}) and~(\ref{eq:transformtensor}) together give us back the tensor invariance condition
\begin{multline}
U(g)^i_{m_i m'_i} V(g)^\alpha_{m_\alpha m'_\alpha} V(g)^{-1\beta}_{m'_\beta m_\beta} A^{a_i m'_i d_i}_{(a_\alpha m'_\alpha d_\alpha) (a_\beta m'_\beta d_\beta)} \\= A^{a_i m_i d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta)}.
\end{multline}
This condition is valid for each set of irreps labelled by $(a_i, d_i, a_\alpha, d_\alpha, a_\beta, d_\beta)$. Now consider the Clebsch-Gordan (CG) series $i \otimes \alpha=\oplus_\beta n_{i \alpha}^\beta \beta$. On the basis level we have,
\begin{multline}
\label{eq:cg}
\ket{a_\beta, m_\beta ; n} = \\ \sum_{a_i,m_i,a_\alpha, m_\alpha} C^{a_\beta m_\beta;n}_{a_i m_i, a_\alpha m_\alpha} \ket{a_i, m_i} \ket{a_\alpha, m_\alpha}.
\end{multline}
$C[\omega]^{a_\beta m_\beta;n}_{a_i m_i a_\alpha m_\alpha}$ denotes the CG coefficients associated with the change of basis of the direct product of irreps $i$ and $\alpha$ to the $n$-th copy of irrep $\beta$. With this, we rewrite Eq.~(\ref{eq:basisindeptensor}) as
\begin{multline}
\hat{A} = A^{a_i m_i d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta)} (C^{-1})^{a_\gamma m_\gamma;n}_{a_i m_i, a_\alpha m_\alpha } \\ \outerproduct{a_\gamma, m_\gamma;n, d_i, d_\alpha} {a_\beta,m_\beta,d_\beta}
\end{multline}
The ket $\ket{a_\gamma, m_\gamma;n, d_i, d_\alpha}$ denotes a basis in the $n$-th copy of $a_\gamma$ irrep obtained from fusing the $d_i$-th copy of irrep $a_i$ and $d_\alpha$-th copy of irrep $a_\alpha$. If we impose invariance Eq.~(\ref{eq:invariancetensor}) in this new form, we get
\begin{multline}
V(g)^{\gamma;n}_{m_\gamma m'_\gamma} (C^{-1})^{a_\gamma m'_\gamma;n }_{a_i m_i, a_\alpha m_\alpha} A^{a_i m_i
d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m'_\beta d_\beta)}
V(g)^{-1\beta}_{m'_\beta m_\beta}\\ = (C^{-1})^{a_\gamma m_\gamma;n }_{a_i m_i ,a_\alpha m_\alpha } A^{a_i m_i
d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta)},
\end{multline}
which is equivalent to
\begin{multline}
V(g)^{\gamma;n}_{m_\gamma m'_\gamma} \left[ (C^{-1})^{a_\gamma m'_\gamma;n }_{a_i m_i, a_\alpha m_\alpha} A^{a_i m_i
d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta)}\right]
=\\ \left[ (C^{-1})^{a_\gamma m_\gamma;n }_{a_i m_i, a_\alpha m_\alpha} A^{a_i m_i
d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m'_\beta d_\beta)}\right] V(g)^{\beta}_{m'_\beta m_\beta}\
\end{multline}
Using Schur's lemmas, we can now determine that
\begin{eqnarray}
\gamma \neq \beta &:& (C^{-1})^{a_\gamma m_\gamma;n }_{(a_i m_i) (a_\alpha m_\alpha) } A^{a_i m_i
d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta)} = 0 \nonumber, \\
\gamma = \beta &:& (C^{-1})^{a_\gamma m_\gamma; n }_{(a_i m_i) (a_\alpha m_\alpha)} A^{a_i m_i
d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta)} \propto \delta_{m_\gamma m_\beta} .\nonumber
\end{eqnarray}
This gives us
\begin{eqnarray*}
&&\left[ (C^{-1})^{a_\beta m_\beta;n }_{a_i m_i, a_\alpha m_\alpha} A^{a_i m_i
d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta n_\beta d_\beta)}\right] \\&& = \delta_{m_{\beta} n_{\beta}}~ B^{a_i d_i}_{(a_\alpha d_\alpha) (a_\beta d_\beta;n)}
\end{eqnarray*}
This is again a condition valid for each set of irreps labelled by $(a_i, d_i, a_\alpha, d_\alpha, a_\beta, d_\beta)$. Finally, moving $C$ to the right hand side, we get
\begin{equation}
\label{eq:decomposed_mps}
A^{a_i m_i d_i}_{(a_\alpha m_\alpha d_\alpha)(a_\beta m_\beta d_\beta)} = \sum_{n=1}^{n_{i \alpha}^\beta} B^{a_i d_i}_{(a_\alpha d_\alpha) (a_\beta d_\beta;n)} C^{a_\beta m_\beta;n }_{a_i m_i, a_\alpha m_\alpha}
\end{equation}
If we restrict $V$ to contain only irreps of a class $\omega$, we get Eq.~(\ref{eq:ACB}).
\subsection{SPT phases with on-site symmetry and lattice translation invariance}\label{sec:on-site+trans}
Gapped Hamiltonians with only lattice translation invariance all belong to the same phase~\cite{top1d_wenold,top1d_wennew}. Ground-states of such Hamiltonians and can be described by MPS matrices $A^{i_m}_m$ that are site independent $i.e.$ $A^{i_m}$~\cite{Perez-Garcia2007}. This means that unlike the case for an arbitrary gapped phase where we needed $N d$ matrices to describe a ground state, we now only need $d$ matrices. Eq.~(\ref{eq:mps_periodic}) is simplified to
\begin{equation} \label{eq:mps_trans}
\ket{\psi} = \sum_{i_1 \dots i_N} Tr[ A^{i_1} A^{i_2} \dots A^{i_N}]\ket{i_1} \dots \ket{i_N}.
\end{equation}
If we consider gapped Hamiltonians invariant under translation and an on-site symmetry transformation $u(g)$, the conditions of Eqs.~(\ref{eq:wavefunction_invariance}, \ref{eq:mps_invariance}) again hold. However, unlike the case for just on-site symmetry, the different 1D irreps, $\chi(g)$ that appear in Eq.~(\ref{eq:mps_invariance}) now label distinct phases of matter~\cite{top1d_wenold,top1d_wennew}. Different SPT phases are now labelled by $\{ \omega, \chi \}$ where, $\omega \in H^2(G,\mathbb{C})$ labels the different projective classes and $\chi$ labels the different 1D irreps of the group $G$. We now see how we can constrain the ground-state form of these SPT phases extending the results of Sec~\ref{sec:on-site}
Let us rewrite Eq.~(\ref{eq:mps_invariance}) by absorbing $\chi(g)$ on the right hand side into $u(g)$ on the left and call $\tilde{u}(g) = \chi^*(g) u(g)$
\begin{equation}
\tilde{u}(g)_{ij} A^j =V^{-1}(g) A^i V(g),
\end{equation}
Since re-phasing a representation with a 1D irrep is still a representation, we can find the new irrep content of $\tilde{u}(g)$. With this, we can repeat the procedure of Sec~\ref{sec:on-site} and obtain the MPS matrices for ground states of a given spin system in any phase labelled by $\{\omega, \chi\}$ as
\begin{multline}\label{eq:mps_onsite+trans}
A[\omega,\chi]^{a_i m_i d_i}_{(a_\alpha m_\alpha d_\alpha) (a_\beta m_\beta d_\beta)} =\\ \sum_{n=1}^{n^\beta_{i \alpha}} B^{a_i d_i}_{(a_\alpha d_\alpha) (a_\beta d_\beta;n)} C[\omega,\chi]^{a_\beta m_\beta;n}_{a_{i'} m_{i'}, a_\alpha m_\alpha}.
\end{multline}
Where $i \otimes \chi \cong i'$ is some linear irrep of $G$ that can easily be identified by calculating the characters of $i'$ and $C[\omega,\chi]^{a_\beta m_\beta;n}_{a_{i'} m_{i'} a_\alpha m_\alpha}$ denote the CG coefficients associated with the change of basis of the direct product of linear irrep $i'$ and the irrep $\alpha$ of projective class $\omega$, to the $n$-th copy of irrep $\beta$ of the same projective class $\omega$.
To summarize, in order to find the ground-state forms of different SPT phases for a spin chain that transforms under a certain representation $u(g)$ of $G$ and that is translationally invariant, we need to follow the steps below:
\begin{enumerate}
\item Obtain $H^2(G,\mathbb{C})$ and the covering group $\tilde{G}$.
\item Identify the irreps `$i$' of the physical spin among the irreps of $\tilde{G}$.
\item Identify the different 1D irreps of $G$, $\chi$ among the 1D irreps of $\tilde{G}$.
\item Identify the irreps `$i'$' corresponding to re-phasing the physical spin irreps `$i$' with $\chi$.
\item Identify which irreps `$\alpha$' correspond to the projective class $\omega$.
\item Obtain CG coefficients corresponding to the fusion of the re-phased irreps of the physical spin with each irrep of the projective class $\omega$.
\item Use the CG coefficients in Eq.~(\ref{eq:mps_onsite+trans}) allowing $\alpha$ and $\beta$ to run over all the irreps of class $\omega$ and $i'$ to run over the re-phased irreps of the physical spin.
\end{enumerate}
We can also consider the ground-state forms constrained by other space-time symmetries like inversion and time-reversal and combinations with on-site symmetry which have also been classified. While there are constraints imposed on the entries of the MPS matrices, we do not immediately see a useful structure like we do with on-site symmetries with or without translation invariance mentioned above. However, for the sake of completeness, we have presented the results in the Appendices.~\ref{app:parity+time},\ref{app:parity+time+onsite}.
\section{Examples of ground-state forms for various on-site symmetries}\label{sec:examples}
In this section, we use the results of the decomposition scheme discussed in the previous section to write down several ground-state forms of SPT phases protected by various on-site symmetries.
We will focus on some subgroups of $SO(3)$ that have a particular non-trivial second cohomology group $H^2(G,\mathbb{C}) = \mathbb{Z}_2$ and hence one class of non-trivial projective representations. (But our formalism can be applied to groups of other second cohomology group as well.) We will also focus on constructing ground states that are topologically non-trivial $i.e.$ states that cannot be connected to the product state and whose virtual space representation corresponds to non-trivial projective representation. This is because these non-trivial states are sufficiently entangled and may offer advantages for information processing. We shall use the following conventions:
\begin{enumerate}
\item Groups are defined by a \emph{presentation} $\innerproduct{S}{R}$ $i.e.$ by listing the set $S$ of generators and the set $R$ of relations between them.
\item Representations are written by listing those of the generating set $S$. Any element in the group can always be written as the product of powers of the subset of $S$.
\item $\tilde{G}$ denotes the Schur cover of $G$ that contains the linear and projective irreps of $G$.
\item We list the irreps of $\tilde{G}$ and label different classes of irreps by elements of $H^2(G,\mathbb{C})$. These correspond to the linear and projective irreps of $G$.
\item $\chi_i$ denotes different 1D irreps of $G$ (and $\tilde{G}$).
\item MPS matrices are constructed up to a similarity transformation for a particular basis of the physical spin that will be mentioned.
\item Pauli matrices are denoted as $\sigma_i = \{\sigma_x, \sigma_y, \sigma_z\}$ or $\sigma_i = \{\sigma_1, \sigma_2, \sigma_3\}$.
\end{enumerate}
\subsection{Haldane phase ($\mathbb{Z}_2 \times \mathbb{Z}_2$)}\label{sec:z2z2}
Consider a chain of three level spins (d=3) that is invariant under a three-dimensional representation of $\mathbb{Z}_2 \times \mathbb{Z}_2$ written as a restricted set of spin-1 $SO(3)$ rotations,
\begin{equation}\label{eq:u_z2z2}
u(g) = \{\mathbb{1},R_x(\pi),R_y(\pi),R_z(\pi)\}.
\end{equation}
$\mathbb{Z}_2 \times \mathbb{Z}_2$, also known as the Klein four-group, is the group of symmetries of a rhombus or a rectangle (which are not squares) generated by $\pi$ flips about perpendicular axes in the plane of the object. Some information about the group are follows:
\begin{itemize}
\item $G=\mathbb{Z}_2 \times \mathbb{Z}_2 = \langle a,x | a^2 = x^2 = (ax)^2 = e\rangle $
\item $H^2(G,\mathbb{C}) = \mathbb{Z}_2 = \{e,a\}$
\item $\tilde{G}= D_8 : \langle a,x|a^4 = x^2 = (ax)^2 = e \rangle$
\item Class $e$ irreps of $\tilde{G}$:\\
$1_{(p,q)}: a \mapsto (-1)^p,~x \mapsto(-1)^q$, $~(p,q)$ $\in$ $\{0,1\}$
\item Class $a$ irreps of $\tilde{G}$:\\
$\tilde{2}: a \mapsto i\sigma_z $,~$x \mapsto\sigma_x$
\end{itemize}
The three-dimensional representation can be shown to be $u(g) \cong 1_{(0,1)} \oplus 1_{(1,0)} \oplus 1_{(1,1)}$. Which means, with an appropriate choice of basis, each basis state of the 3 level spin transforms as one of the non-trivial 1D irreps. We can check that
$\{\ket{x} \equiv \frac{1}{\sqrt{2}} (\ket{-1}-\ket{1}),~\ket{y} \equiv \frac{i}{\sqrt{2}} (\ket{-1}+\ket{1}), ~\ket{z} \equiv \ket{0}\}$ is such an appropriate basis where $u(g)$ is block diagonal. Calculating the CG coefficients, we get the following MPS matrices:
\begin{equation}
A^i = B_i \otimes \sigma_i, \label{eqn:factorize}
\end{equation}
where $B_i$ are undetermined and $\sigma_i$ are the Pauli matrices. We thus have reproduced the result of Ref.~\cite{bartlett2012} using our general framework.
\subsection{$D_4$ invariant SPT phase}\label{sec:d4}
$D_{n}$, the dihedral group is the symmetry group of a planar $n$ sided polygon and has projective representations when $n$ is even. Some information about the group are as follows. We only look at the case of even $n$.
\begin{enumerate}
\item G = $D_n$ = $\innerproduct{a,x}{a^n = x^2 = (ax)^2 = e }$
\item $H^2(G,\mathbb{C}) = \mathbb{Z}_2 = \{e,a\}$
\item $\tilde{G}$ = $Q_{n}$: \\$\innerproduct{a,x}{a^{2n} = x^4 = e, a^n = x^2, x a x^{-1 } = a^{-1}}$.
\item Class $e$ irreps of $\tilde{G}:$
\begin{enumerate}
\item $1_{(p,q)}: a \mapsto (-1)^p,~x \mapsto (-1)^q$, $~(p,q)$ $\in$ $\{0,1\}$
\item $2_{(k)}: a \mapsto \begin{pmatrix}
e^{ {-i k \eta_n}/{2} } & 0\\
0 & e^{ {i k \eta_n}/{2} },
\end{pmatrix} $,
$x \mapsto -i \sigma_y$,\\ $k = 2,4,\ldots n-2$, $\eta_n = 2 \pi/ n$
\end{enumerate}
\item Class $a$ irreps of $\tilde{G}:$
\begin{enumerate}
\item $\tilde{2}_{(k)}: a \mapsto \begin{pmatrix}
e^{ {-i k \eta_n}/{2} } & 0\\
0 & e^{ {i k \eta_n}/{2} }
\end{pmatrix} $,
$x \mapsto \sigma_y~$,\\ $k = 1,3,\ldots n-1$, $\eta_n = 2 \pi/ n$
\end{enumerate}
\end{enumerate}
Let us now consider the group $D_4$. This is the group of symmetries of a square generated by $\frac{\pi}{2}$ rotations about the symmetry axis perpendicular to the plane and reflections about symmetry axes in the plane of the square. We consider the following irreps (using a different choice of basis than the one mentioned above).
\begin{enumerate}
\item Linear irrep $2_{(2)}: a \mapsto i \sigma_y$, $x\mapsto \sigma_z$
\item Projective irreps:\\ $\tilde{2}_{(1/3)}:a \mapsto \frac{1}{\sqrt{2}} (\pm \mathbb{1}-i \sigma_y),~x \mapsto i \sigma_z$
\end{enumerate}
If we consider a $d=2$ physical spin transforming under the 2D irrep $2_{(2)}$ the non-trivial MPS matrices associated with the two basis states $\ket{i} = \ket{0}, \ket{1}$ are obtained by calculating the CG coefficients:
\begin{eqnarray}
A^0 &=& \begin{pmatrix}
B_{11} \otimes \sigma_z & B_{13} \otimes \mathbb{1}\\
B_{31} \otimes \mathbb{1} & B_{33} \otimes \sigma_z\\
\end{pmatrix}, \\
A^1 &=& \begin{pmatrix}
B_{11} \otimes -\sigma_x & B_{13} \otimes -i \sigma_y\\
B_{31} \otimes i \sigma_y & B_{33} \otimes \sigma_x\\
\end{pmatrix}.
\end{eqnarray}
The MPS matrices cannot be further factorized, and thus we do not even have the protected identity gate.
\subsection{$A_4$ invariant SPT phase}\label{subsec:a4}
$A_4$, the alternating group of degree four, is the group of chiral or rotational symmetries of a regular tetrahedron generated by rotations (no reflections) about various symmetry axes. It is also the group of even permutations on four elements, i.e. a subgroup of $S_4$ to be discussed next. Some information about the group are as follows.
\begin{enumerate}
\item G = $A_4$ = $\innerproduct{a,x}{a^3 = x^2 = (ax)^3 = e }$
\item $H^2(G,\mathbb{C}) = \mathbb{Z}_2 = \{e,a\}$
\item $\tilde{G}$ = $\tilde{T}$ : $\innerproduct{a,x}{a^3 = x^2 = v, v^2= (ax)^3 = e}$.
\item Class $e$ irreps of $\tilde{G}:$
\begin{enumerate}
\item $1_{(p)}: a \mapsto e^{2 \pi i p/3},~x \mapsto 1$, $p=0,1,2$
\item $3: a \mapsto \begin{pmatrix}
0 & 1 & 0\\
0 & 0 & 1\\
1 & 0 & 0
\end{pmatrix}$, $x \mapsto \begin{pmatrix}
1 & 0 & 0 \\
0 & -1 & 0\\
0 & 0 & -1
\end{pmatrix}$
\end{enumerate}
\item Class $a$ irreps of $\tilde{G}:$
\begin{enumerate}
\item $\tilde{2}_{(p)}: a \mapsto e^{2 \pi i p/3} \frac{1}{2} [\mathbb{1} + i (\sigma_x + \sigma_y + \sigma_z)] $,\\$ x \mapsto i \sigma_x $, $p = 0,1,2$
\end{enumerate}
\end{enumerate}
If we consider the physical spin transforming under the only 3D linear irrep, the non-trivial MPS matrices associated with the three basis states $\ket{i} = \ket{1}, \ket{2}, \ket{3}$ are obtained by calculating the CG coefficients:
\begin{eqnarray}
A^i &=& B_{i} \otimes \sigma_i\\
B_{i} &=& V^{i-1} B V^{*i-1}\\
V &=& \begin{pmatrix}
\mathbb{1} & 0 & 0\\
0 & \omega~\mathbb{1} & 0\\
0 & 0 & \omega^*~\mathbb{1}
\end{pmatrix},~\omega = e^{2 \pi i /3} \nonumber \\
B &=& \begin{pmatrix}
B_{00} & B_{01} & B_{02}\\
B_{10} & B_{11} & B_{12}\\
B_{20} & B_{21} & B_{22}\\
\end{pmatrix}. \nonumber
\end{eqnarray}
Similar to Eq.~(\ref{eqn:factorize}) the MPS matrices are factorized into to two parts, and the indenity gate is protected by the symmetry. We remark that imposing inversion or time-reversal symmetry does not further simplify the $B$'s structure.
\subsection{$S_4$ invariant SPT phase}\label{sec:s4}
$S_4$, the symmetric group of degree four, is the group of achiral or full symmetries of a tetrahedron generated by rotations and reflections about various symmetry axes. It is also the group of all permutations of four elements. Some information about the group are as follows.
\begin{enumerate}
\item G = $S_4$ = $\innerproduct{a,b,c}{a^2 = b^3 = c^4 = abc = e }$
\item $H^2(G,\mathbb{C}) = \mathbb{Z}_2 = \{e,a\}$
\item $\tilde{G}$ = $O'$ : $\innerproduct{a,b,c}{a^2 = b^3 = c^4 = abc = v, v^2= e }$
\item Class $e$ irreps of $\tilde{G}:$ $(a= tk,~ b = s,~ c = s^2 k t)$
\begin{enumerate}
\item $1_{(p)}:t \mapsto (-1)^p,~ k \mapsto 1,~ s \mapsto 1$, $p= \{0,1\}$
\item $s: t \mapsto \sigma_x,~ k \mapsto \mathbb{1},~ s\mapsto \begin{pmatrix}
e^{2 \pi i/3} & 0\\
0 & e^{-2 \pi i/3}
\end{pmatrix}$
\item $3_{(p)}$: $k \mapsto \begin{pmatrix}
1 & 0 & 0 \\
0 & -1 & 0\\
0 & 0 & -1
\end{pmatrix}$,~$
s \mapsto \begin{pmatrix}
0 & 1 & 0\\
0 & 0 & 1\\
1 & 0 & 0
\end{pmatrix}$,\\ $t \mapsto (-1)^p \begin{pmatrix}
1 & 0 & 0\\
0 & 0 & 1\\
0 & 1 & 0
\end{pmatrix}$, $p = 0,1$
\end{enumerate}
\item Class $a$ irreps of $\tilde{G}:$
\begin{enumerate}
\item $\tilde{2}_{(p)}:t \mapsto (-1)^p \frac{i}{\sqrt{2}}(\sigma_z - \sigma_y),~ k \mapsto i \sigma_x,\\ s \mapsto \frac{1}{2} [\mathbb{1} + i (\sigma_x + \sigma_y + \sigma_z)]$ , $p =0,1$
\item $\tilde{4} = 2 \otimes \tilde{2}_{(0)}$
\end{enumerate}
\end{enumerate}
If we consider the physical spin transforming under one of the 3D linear irreps, $3_{(1)}$ the non-trivial MPS matrices associated with the three basis states are obtained by calculating the CG coefficients:
\begin{eqnarray}
A^i &=& B_{i} \otimes \sigma_i\\
B_{i} &=& \begin{pmatrix}
B_{2_02_0} & 0 & B_{2_0 4} \otimes u^\dagger_{i-1}\\
0 & B_{2_12_1} & B_{2_1 4} \otimes v^\dagger_{i-1}\\
B_{4 2_0} \otimes u_{i-1} &
B_{4 2_1} \otimes v_{i-1} &
B_{4 4} \otimes \mathbb{1}_2 + \tilde{B}_{44} \otimes f_{i-1}
\end{pmatrix} \nonumber\\
u_i &=& \begin{pmatrix}
\omega^{*i}\\
\omega^{i}
\end{pmatrix},~
v_i = \begin{pmatrix}
\omega^{*i}\\
-\omega^{i}
\end{pmatrix} \nonumber\\
f_i &=& \begin{pmatrix}
0 & \omega^i\\
\omega^{*i} & 0
\end{pmatrix},~ \omega = e^{2 \pi i/3} \nonumber
\end{eqnarray}
We observe that if we restrict the $B_i$ matrix to only the bottom right block and set the two matrices to scalars, $B_{4 4} = \cos(\frac{\theta}{2})$ and $\tilde{B}_{4 4} = e^{i\phi} \sin(\frac{\theta}{2})$, then it reduces to the one used for the buffering scheme in Ref.~\cite{MillerMiyake14} up to a change of basis.
\subsection{Summary of new SPT phases with identity gate protection}
We now list, from the examples in the previous section, those SPT ground states which allow the perfect operation of the identity gate according to the scheme reviewed in Sec.~\ref{sec:MPS_MBQC}. We see that the MPS matrices for non-trivial ground states of $d=3$ (i.e. of spin magnitude $S=1$) spin chains protected by $\mathbb{Z}_2 \times \mathbb{Z}_2,~A_4,~S_4$ all have the form
\begin{eqnarray} \label{eq:good_form_for_identity}
A^i = B_i \otimes \sigma_i.
\end{eqnarray}
Following the convention of Ref.~\cite{bartlett2012}, we call $B_i$ the junk part and $\sigma_i$ the protected part. We also note that our convention of placing the protected and junk parts is in reverse order as compared to the convention used in Refs.~\cite{bartlett2012,bartlett_njp2012,MillerMiyake14}. This is for notational consistency in this paper.
Consider encoding qubit information $\ket{\psi}$ in the protected part of the right boundary virtual space with the junk part arbitrarily set to some state $\ket{J}$ in any of these ground states.
\begin{eqnarray}
\ket{R} = \ket{J} \otimes \ket{\psi}
\end{eqnarray}
If we perform a measurement on the rightmost $i.e$ $N$-th spin in the basis $\ket{x}, \ket{y}, \ket{z}$ in which the MPS matrices have the form of Eq.~(\ref{eq:good_form_for_identity}) with an outcome $\ket{k_N}$, we induce a transformation of the boundary vector by~(\ref{sec:MPS_MBQC})
\begin{eqnarray}
\ket{R} &\mapsto& A^{k_N} \ket{R}\\
\implies \ket{J} \otimes \ket{\psi} &\mapsto& B_{k_N} \ket{J}\otimes \sigma_{k_N} \ket{\psi}
\end{eqnarray}
The qubit information $\ket{\psi}$ is unchanged upto an inconsequential Pauli operator $\sigma_{k_N}$ which can be corrected for by a change of readout basis. In fact, we can measure several spins (say $m$ from the right) and we still have the perfect operation of the identity gate upto a residual operator $\sigma_{k_{N-m}} \ldots \sigma_{k_N}$. This means all these ground states allow a protected subspace with perfect identity gate operation, which allows for perfect transmission of quantum information encoded in the projected subspace.
However, note that if we measure in a different basis formed by a linear combination of $\ket{x}, \ket{y}, \ket{z}$, it is easy to check that the boundary vector $\ket{R} = \ket{J} \otimes \ket{\psi}$ no longer remains decomposed into protected and junk parts and, in general, there will be mixing between the two vector spaces. As an illustration, if a measurement outcome of $\frac{1}{\sqrt{2}}(\ket{x} + \ket{y})$ is obtained, the induced transformation on $\ket{R}$ is (up to an overall factor)
\begin{eqnarray}
\ket{J} \otimes \ket{\psi} \mapsto B_x \ket{J} \otimes \sigma_x \ket{\psi} + B_y \ket{J} \otimes \sigma_y \ket{\psi}.
\end{eqnarray}
Thus, in general only the identity gate is protected in the ground states of these phases. However, if it were possible that $B_i$ is independent of physical index $i$, then arbitrary single-qubit gates would be possible, as mixing will not occur. It is worth noting that when $B_i$ is independent of the index $i$, the corresponding wavefunction is identically the AKLT state. We had hoped that imposing additional symmetry like parity and/or time reversal invariance might give further constraints on the matrices $B_i$'s and thereby allow non-trivial gate operations. But we checked (using results of Appendix.~\ref{app:parity+time+onsite}) that imposing these additional symmetries on the $\mathbb{Z}_2 \times \mathbb{Z}_2, A_4$ and $S_4$ SPT ground states listed above does not induce ground states that could provide universal qubit operations.
\section{An $A_4$ symmetric Hamiltonian}
\label{sec:model_hamiltonian}
Here we ask a slightly less general question: can one find a particular Hamiltonian with symmetry such that there is an extended region (not necessarily at all points of a phase) in the phase diagram that the ground states can provide universal qubit operations in the framework of MBQC?
Below we first construct a specific Hamiltonian that possesses $A_4$ and parity symmetry, which can be regarded as perturbing the spin-1 AKLT Hamiltonian.
Then we present a numerical investigation and show that indeed there exists a finite parameter region where the ground states are exactly (here and henceforth, exact is defined up to machine precision) the AKLT state, and can therefore serve as a resource state for implementing universal single-qubit gates. After the numerical investigation, we present analytic understanding why such an extended region of AKLT ground states can exist.
\begin{figure}[ht]
\includegraphics[scale = 0.63]{Fidelity.png}
\caption{(Color online) Fidelity of ground states with the AKLT state. It is seen that there is an extended region such that the ground state is exactly the AKLT state.}
\label{fig:Fidelity}
\end{figure}
\subsection{Construction of $A_4$ symmetric Hamiltonian}
We will now construct the $A_4$ and inversion symmetric Hamiltonian and study its phase diagram. We use group invariant polynomials as building blocks to construct Hermitian operators invariant under group action. A group $G$ invariant $n$-variable polynomial $f(x_1,x_2, \dots x_n)$ is unchanged when the $n$-tuplet of variables $\left(x_1, x_2 \dots x_n \right)$ is transformed under an $n$-dimensional representation of the group $D(g)$.
\begin{eqnarray}
f(x'_1, x'_2 \dots x'_n) &=& f(x_1, x_2 \dots x_n) \\
x'_i &=& D(g)_{i j} x_j.
\end{eqnarray}
If we have $n$ Hermitian operators $X_{i=1 \dots n}$ that are $n$-dimensional and transform covariantly like the $n$ variables of the polynomial $x_{i=1 \dots n}$, $i.e.$ $D(g) X_i \dagr{D}(g) = D(g)_{i j} X_j$, then we can elevate the group invariant polynomials to group invariant operators as $f(x_1,x_2, \dots x_n) \rightarrow f(X_i, X_2 \dots X_n)$ carefully taking into account that unlike the numbers $x_i$, the operators $X_i$ do not commute.
Since we need three-dimensional operators of $A_4$, we consider the set of independent three variable polynomials invariant under the 3D irrep of $A_4$~\cite{ramond_group}:
\begin{eqnarray}
f_1(x,y,z) &=& x^2 + y^2 + z^2,\\
f_2(x,y,z) &=& x^4 + y^4 + z^4,\\
f_3(x,y,z) &=& xyz.
\end{eqnarray}
We know that the spin operators $S^i$ satisfying $ [S^i, S^j] = i \epsilon_{ijk} S^k $ transform covariantly under any $SO(3)$ rotation, in particular for the finite set of rotations that corresponds to the subgroup $A_4 \in SO(3)$. Thus, to find invariant operators for the three-dimensional representation $3$, we need to take the spin operators in the appropriate three-dimensional basis $\ket{x}, \ket{y}, \ket{z}$ as defined in Sec.~\ref{subsec:a4} and elevate the polynomials $f_1, f_2, f_3$ to operators as
\begin{eqnarray}
F_1 &=& S^x_a S^x_b + S^y_a S^y_b + S^z_a S^z_b,\\
F_2 &=& (S^x_a S^x_b)^2 + (S^y_a S^y_b)^2 + (S^z_a S^z_b)^2 ,\\
F_3 &=& S^x_a S^y_b S^z_c + S^z_a S^x_b S^y_c + S^y_a S^z_b S^x_c \nonumber \\
&+& S^y_a S^x_b S^z_c + S^x_a S^z_b S^y_c + S^z_a S^y_b S^x_c,
\end{eqnarray}
where the indices $a,b,c$ label collectively any other quantum numbers like lattice sites and can be chosen as per convenience, say to make the operators local. As a model Hamiltonian, we could use any function of the invariant operators $F_1$, $F_2$ and $F_3$ and ensure that everything is symmetric under the exchange of lattice labels to impose inversion symmetry. In particular, the AKLT state is the unique ground state of a particular combination of the invariant operators but it has a larger symmetry group, $SO(3)$.
\begin{eqnarray}
H_{AKLT} = \sum_i \left[ \vec{S}_i\cdot\vec{S}_{i+1} + \frac{1}{3} (\vec{S}_i\cdot\vec{S}_{i+1})^2 \right], \\
\mbox{where}~\vec{S}_i\cdot\vec{S}_{i+1} \equiv S^x_i S^x_{i+1} + S^y_i S^y_{i+1} + S^z_i S^z_{i+1}. \nonumber
\end{eqnarray}
Thus we can consider adding two other combinations to the AKLT Hamiltonian so as to break the $SO(3)$ symmetry to $A_4$ by using $A_4$ invariant perturbations:
\begin{eqnarray}
H_q &=& \sum_i \left[ (\vec{S}^2_i\cdot\vec{S}^2_{i+1}) - \frac{1}{3} (\vec{S}_i\cdot\vec{S}_{i+1})^2 \right],\\
\mbox{where}&&\vec{S}^2_i\cdot\vec{S}^2_{i+1} \equiv (S^x_i S^x_{i+1})^2 + (S^y_i S^y_{i+1})^2 + (S^z_i S^z_{i+1})^2,\nonumber
\end{eqnarray}
and
\begin{multline}
H_c = \sum_i [ (S^x S^y)_i S^z_{i+1} + (S^z S^x)_i S^y_{i+1} + (S^y S^z)_i S^x_{i+1} \\
+ (S^y S^x)_i S^z_{i+1} + (S^x S^z)_i S^y_{i+1} + (S^z S^y)_i S^x_{i+1} \\
+ S^x_{i} (S^y S^z)_{i+1} + S^z_{i} (S^x S^y)_{i+1} + S^y_{i} (S^z S^x)_{i+1} \\
+ S^x_{i} (S^z S^y)_{i+1} + S^z_{i} (S^y S^x)_{i+1} + S^y_{i} (S^x S^z)_{i+1}].
\end{multline}
With these pieces, we arrive at the total Hamiltonian which is $A_4$ and inversion symmetric,
\begin{equation}
\label{eqn:H}
H = H_{AKLT} + \lambda H_c + \mu H_q.
\end{equation}
\subsection{Checking AKLT as the ground state}
The AKLT state $|\psi_{\rm AKLT}\rangle$ has the MPS representation $A^x=\sigma_x$, $A^y=\sigma_y$, and $A^z=\sigma_z$ in the basis of \{$|x\rangle$, $|y\rangle$, $|z\rangle$\} defined earlier. We know that at $\lambda=\mu=0$ the ground state of the Hamiltonian~(\ref{eqn:H}) is uniquely the AKLT state. We would like to know whether there is an extended region of $(\lambda,\mu)$ around $(0,0)$ such that the ground state is also the AKLT state. We do this numerically by first solving the ground state $|\psi_G\rangle$ of the Hamiltonian~(\ref{eqn:H}) using the infinite time-evolving bond decimation (iTEBD) algorithm invented by Vidal~\cite{vidal2007} and then calculating the fidelity between these two states $f=|\langle \psi_G|\psi_{\rm AKLT}\rangle|^2$. As shown in Fig.~\ref{fig:Fidelity} we indeed see that there is an extended region in this Hamiltonian such that the ground state is exactly the AKLT state and thus a useful resource state for universal single-qubit MBQC. \\
\subsection{Analytic understanding}
We now analyze why such an extended region of AKLT is possible and calculate analytically the boundary of the AKLT region in the $\lambda$-$\mu$ plane, shown in Fig.~\ref{fig:Fidelity}. First we recall that the interaction between sites $i$ and $i+1$ of ${H_{AKLT}}$ is a projection to the joint $S=2$ subspace. More precisely,
\begin{equation}
(H_{AKLT})_{i,i+1}= 2 \sum_{m=-2}^{2} P_{|S=2,m\rangle}- \frac{2}{3}\openone,
\end{equation}
where we have defined the projector $P_{|\psi\rangle}\equiv |\psi\rangle\langle\psi|$ associated with the state $|\psi\rangle$, $|S=2,m\rangle$ denotes the eigenbasis of the joint spin-2 states for neighboring sites $i$ and $i+1$, and $\openone$ is the identity operator in the spin-2 subspace.
For the quartic Hamiltonian, it is seen by straightforward calculation that
\begin{equation}
(H_q)_{i,i+1}=P_{ (|S=2,2\rangle+|S=2,-2\rangle)/\sqrt{2}}\,+P_{|S=2,0\rangle}+\frac{2}{3}\openone.
\end{equation}
For the cubic Hamiltonian, it is seen that
\begin{equation}
(H_c)_{i,i+1}=2\sqrt{3}\Big(P_{|\phi^+\rangle}-P_{|\phi^-\rangle}\Big),
\end{equation}
where
\begin{eqnarray}
|\phi^\pm\rangle&\equiv & (|S=2,m=2\rangle+|S=2,m=-2\rangle \nonumber \\
&&\pm i \sqrt{2}|S=2,m=0\rangle)/2.
\end{eqnarray}
Since the AKLT state is annihilated by any spin-2 projectors, it will remain the ground state if the following operator is positive,
\begin{eqnarray}
h(\lambda,\mu)&\equiv& 2 P_{|S=2,m=2\rangle}+2 P_{|S=2,m=-2\rangle}+2 P_{|S=2,m=0\rangle} \nonumber\\
&&+ 2\sqrt{3}\lambda\Big(P_{|\phi^+\rangle}-P_{|\phi^-\rangle}\Big) + \mu \, P_{|S=2,m=0\rangle} \nonumber\\
&& + \mu\, P_{ (|S=2,m=2\rangle+|S=2,m=-2\rangle)/\sqrt{2}},
\end{eqnarray}
which, in the basis of $|S=2,m=\{\pm 2, 0\}\rangle$ is the following $3\times3$ matrix,
\begin{equation}
h(\lambda,\mu)=\left(\begin{array}{ccc}
2+{\mu}/{2}&{\mu}/{2} & -i\sqrt{6}\lambda\\
{\mu}/{2} & 2 +{\mu}/{2} & -i\sqrt{6}\lambda\\
i\sqrt{6}\lambda & i \sqrt{6}\lambda & 2+\mu \end{array}\right).
\end{equation}
By direct diagonalization, we find that the matrix $h(\lambda,\mu)$ is non-negative when $\mu \pm 2\sqrt{3}\lambda +2 > 0$ which indeed gives the region of the AKLT in Fig.~\ref{fig:Fidelity}.
\section{Summary}\label{sec:summary}
We have presented a straightforward and general formalism for investigating the structure of a wavefunction as constrained (or protected) by a discrete symmetry group. The wavefunction is organized into two parts: (1) a CG part, whose form is inferred from the symmetry group and (2) a part whose form is not constrained by the symmetry. From the viewpoint of measurement-based quantum computation, one can then use this formalism to discuss whether the ground state of an SPT phase protected by a given symmetry group allows protected gate operations. This happens when, for example, the MPS matrices $A^i$ decompose into the form $A^i = B_i \otimes \sigma_i$ $i.e.$ the virtual vector space decomposes into junk and protected parts. Generically speaking, the identity gate is not necessarily protected in an arbitrary SPT phase. With the new formalism, we recovered the results of the $\mathbb{Z}_2 \times \mathbb{Z}_2$ case previously obtained in Ref.~\cite{bartlett2012} and obtain the MPS forms for several other groups. We show that $A_4$ and $S_4$ groups also allow protected identity gate operation. We also constructed a Hamiltonian with $A_4$ and inversion symmetry and found that in an extended region of a two-parameter space, the ground state is exactly the AKLT state. Using the formulation developed here, further exploration of 1D SPT phases and gate protection can be made with arbitrary finite groups. The MPS forms can also allow the study of the properties of 1D SPT phases which would be of interest to the condensed matter community.
Despite the search we still have not identified a 1D SPT phase that generically supports arbitrary universal single-qubit gates, contrary to what we had hoped for. The only gate that can be naturally protected generically in an entire SPT phase is the identity gate. Although not satisfying in terms of quantum computation, it is useful in terms of transmitting quantum information over long distances. The buffering technique recnetly invented in Ref.~\cite{MillerMiyake14} seems necessary to bring forth universal gates, as demonstrated for the $S_4$ symmetry, and it would be interesting to know whether this can be applied generically to all SPT phases with the identity gate known to be protected. Another open question that naturally arises is whether there exists a 2D SPT phase where all ground states in the phase support protected universal quantum computation.
\label{sec:summary}
\medskip \noindent {\bf Acknowledgement.} The authors would like to thank Naveen Prabhakar and Robert Raussendorf for several helpful discussions, and especially Akimasa Miyake for pointing out a mistake in an earlier draft, where the inversion symmetry was incorrectly imposed. This work was supported by the
National Science Foundation under Grants No. PHY 1314748 and No. PHY
1333903.
|
\subsection{The Gamma Background}
All materials contain radioactive contaminants. Thus, although care was
taken to minimize radiocontamination in the construction of the CDMS~II cryostat,
support structures, detectors housings, and the detectors themselves,
each component contains some contamination.
The majority of the gamma background observed in CDMS~II is caused by decays
from radioactive U, Th (and their decay chain daughters) and $^{40}$K
occurring in the surrounding materials. Additionally, Ge has radioactive isotopes that
can be produced by neutrons or cosmogenic radiation ($^{68}$Ge and $^{71}$Ge).
These isotopes decay via electron capture producing characteristic lines
at 10.4~keV (K-shell) and 1.3~keV (L-shell). We chose our analysis energy
range of 3--14~keVnr to avoid these activation lines.
Figure~\ref{bavsdata_largeRange}
shows the gamma background for a recoil energy up to 30~keVnr for both single and
multiple scatters, with the K-shell activation line clearly visible at
$\sim$17~keVnr in the single-scatter data.
Other low-energy electron recoils (or ``gammas") result from a variety of
sources, including cosmogenic activation of non-Ge isotopes and small-angle
Compton scattering.
\subsection{The Gamma Background Model}
\label{Gamma_background}
Bulk gamma events are modeled using $^{133}$Ba calibration data.
Although dominated by a line at 356~keVee, sufficient Compton scattering
occurs throughout the surrounding mechanical structures that a flat recoil spectrum
is observed between 3 and 14~keVnr (see Fig.~\ref{bavsdata}). In order for the Ba calibration
data to be a good proxy for WIMP-search (all data taken without a calibration source) gamma events, the energy spectrum of WIMP-search
and barium calibration events must be the same in the energy region of interest.
Figure~\ref{bavsdata} shows a comparison between the energy spectrum of barium
and WIMP-search events in the ER band.
Kolmogorov--Smirnov (KS) tests \cite{KS} comparing the two distributions between
3 and 14~keVnr indicate that their shapes are compatible. The individual detectors
T1Z2, T1Z5, T2Z5 and T3Z4 have KS p-values of 0.8 (0.5), 0.7 (0.2), 0.07 (0.8),
and 0.013 (0.8) for multiples (singles), respectively.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{Figure3.pdf}
\caption{(Color online) WIMP-search (WS) data and Ba calibration data for events
within the ER band for detector T1Z2, given in NR (ER) equivalent energy along the
bottom (top) axis. The 10.4 keVee (17 keVnr) Ge activation
line is clearly seen in the single-scatter WIMP-search data (left panel) but absent in
the Ba calibration and multiple-scatter WIMP-search data (right panel).}
\label{bavsdata_largeRange}%
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{Figure4.pdf}
\caption{(Color online) Spectral comparison of events in the ER band.
The singles spectrum is shown in the left panel while the
multiples spectrum is shown in the right panel. Ba calibration and WIMP-search
(WS) data are well matched. The p-value from a Kolmogorov–-Smirnov test
for the detector shown (T1Z2) is 0.8 (0.5) for multiples
(singles). The p-values for the other detectors are stated in the text.}
\label{bavsdata}%
\end{figure}%
Differences in ionization energy between Ba and WIMP-search data in the ER
band may result in systematic effects.
Figure~\ref{bavsdata_q} compares the ionization-energy spectra inside the ER band. Again,
KS-test p-values indicate that the Ba and WIMP-search spectral shapes are
compatible (for both singles and multiples). The
individual detectors (in the same order) have KS p-values of 0.2 (0.6),
0.02 (0.2), 0.7 (0.5), 0.8 (0.9) for multiples (singles), respectively. This provides assurance that
any ionization-yield differences between the two data types have minimal
influence on the modeling of the gamma background in the ER band.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{Figure5.pdf}
\caption{(Color online) Comparison of the ionization-energy spectra for events
in the ER band. The singles spectrum is shown in the left
panel while the multiples spectrum is shown in the right panel. Ba calibration
and WIMP-search (WS) data are well matched. The p-value from a Kolmogorov–-Smirnov test
for the detector shown (T1Z2) is 0.2 (0.6) for
multiples (singles). The p-values for the other detectors are stated in the text.}
\label{bavsdata_q}%
\end{figure}%
\subsection{The Surface-Event Background}
Surface events are defined as particle interactions that occur within a few
$\mu$m of the surface of a detector. Such events can have diminished charge
collection and can even result in a complete loss of the ionization signal,
in which case they are called zero-charge (ZC) events.
The majority of surface events come from decays of the $^{210}$Pb decay chain,
a long-lived
product of the ubiquitous $^{222}$Rn whose daughters implant into
surfaces during fabrication
of the detectors and housings~\cite{Pb206_Redl}.
The $^{210}$Pb decay chain produces relatively low-energy decay products
that do not penetrate the detectors deeply enough to have full charge
collection, leading to a significant number of surface or ZC events.
\subsection{The Surface-Event Model}
\label{SE}
We start our GEANT4~\cite{Geant-B} simulation of the surface-event background
by contaminating the
surface of both the Cu detector housings and the Ge detectors with $^{214}$Po
nuclei that are allowed to decay isotropically. In addition to using the standard
GEANT4 physics lists it is imperative that the ``Screened Nuclear Recoil Physics List"
(SNRPL)~\cite{iZipRejection,SNR,Pb206_Redl} is invoked in order to correctly simulate
implantation of heavy, low-energy ($\lesssim$~500~keV) nuclei. The
SNRPL is based on algorithms used in SRIM~\cite{SRIM} and has been confirmed
to produce compatible results~\cite{SNR}. After the initial implantation of the
$^{210}$Pb nuclei, we simulate the full $^{210}$Pb decay chain shown in Fig.~\ref{decayChain}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{Figure6.pdf}
\caption{(Color online) The dominant decay modes of the $^{210}$Pb decay chain. The
alpha decay concluding this decay chain gives the $^{206}$Pb nucleus 103~keV
of recoil energy.}
\label{decayChain}%
\end{figure}%
The initial $^{210}$Pb decay produces a mix of electrons and low-energy photons, most of
which are sufficiently low in energy to be classified as surface events
\cite{iZipRejection,Pb206_Redl}. The $^{210}$Bi beta decay has an endpoint of
$\sim$1.2~MeV. In this analysis we focus on low-energy events (below
$\sim$14~keV) and therefore only a small fraction of $^{210}$Bi decays will
fall into our signal region, making the $^{210}$Bi a sub-dominant component.
The final decay in the $^{210}$Pb decay chain is
another Po-isotope alpha decay; $^{210}$Po decays to $^{206}$Pb, producing a
5.3~MeV alpha particle. The alpha particle is unlikely to
contribute to our background because of its high energy.
The 103~keV recoil energy of the $^{206}$Pb nucleus, however, may be degraded
sufficiently to appear in the low-energy signal region because it may have to
travel some distance through the surface in which the parent $^{210}$Po atom
is implanted~\cite{iZipRejection,Pb206_Redl}. The number of particles produced
in this decay chain is proportional to the number of alpha particles produced in the
$^{210}$Po decay, so the observed $^{210}$Po alpha rate (measured
from a high-energy sideband in the WIMP-search data)
is used to estimate the total number of events expected to be observed from
$^{210}$Pb decay chain products (in the low-energy signal region) for each detector.
As described above, GEANT4~\cite{Geant-B} is utilized to
simulate the particle interactions
in our detectors. However, the standard GEANT4 framework is currently not
capable of simulating the phonons and electron-hole pairs produced
by particle interactions in semiconductor crystals (i.e., the detailed
detector response), and therefore this
estimation must be made after the GEANT4 simulation completes. We
extended the GEANT4 framework to include these processes~\cite{Geant4_sim}, however
we did not use this new software here since it would have gone beyond the scope of this paper.
The amount of phonon and ionization energy collected by the sensors is also not
modeled with GEANT4 and
must be done post-simulation. Consider a particle interaction that produces an
initial combination of phonons and electron-hole pairs. The phonons diffuse
through the crystal, and the electron-hole pairs are drifted to the sensors
using a small drift field ($\sim$~3~V/cm), emitting additional phonons on the
way. The amount of charge collected depends on a few factors. The first
(and most obvious) is the absolute number of electron-hole pairs produced by
an event. For events producing recoiling electrons (from incident photons
and electrons), one electron-hole pair escapes the interaction
region per 3.0~eV of deposited recoil energy (on average)~\cite{eEV_cite}.
Events recoiling
off nuclei produce fewer charge carriers, with the amount given approximately by
standard Lindhard theory \cite{Lindhard,CDMSII}.
A particle hitting a
detector near a surface (within $\sim$~1~$\mu$m) will have suppressed charge
collection as well~\cite{DeadLayer}. In Section~\ref{systematics} we show evidence
that for the detector sidewalls this depth scale is likely a factor of 5 smaller
than for the detector faces. This led us to systematically vary the sidewall surface depth in the
limit calculation presented in Section~\ref{results} to account for
systematic uncertainties. For an event
that occurs further than
$\sim$~1~$\mu$m away
from a surface most of the produced charge is collected. For events on the side
instrumented for phonon readout (henceforth referred to as the ``phonon side'')
or on a sidewall we model the amount of charge collected to exponentially go to
zero at the surface, while for events on the side instrumented for ionization readout
(henceforth referred to as the ``ionization side'') we collect a
minimum of $\sim$~50$\%$ of the produced charge carriers, exponentially increasing
to 100$\%$ with a characteristic length of 1~$\mu$m (see Section~\ref{systematics}).
Considering that the ionization and phonon sides of the detectors have different
charge collection characteristics, it is possible to separate the simulation into the
five components shown in Fig.~\ref{schematicContribution};
\textbf{1:} Events that originate in the Cu housings (Housing),
\textbf{2:} Events that originate on the detector currently being studied, on
either the charge side (Q-same) or \textbf{3:} the phonon side (P-same).
Events can also originate either from \textbf{4:} the detector adjacent
to the charge side (Q-opposite), or \textbf{5:} the detector adjacent to the
phonon side (P-opposite). Figure~\ref{SimContribution} shows how
these components contribute to the overall event distribution in the ionization- versus
recoil-energy plane.
To obtain a more realistic detector response, electronic noise
(as measured with calibration data) was added to the simulated
ionization and phonon energies.
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{Figure7.pdf}
\caption{(Color online, legend applies to all panels) Expected distribution of background
events in the plane of ionization versus~recoil energy (for a generic CDMS~II Ge detector),
from simulations of surface events originating
from the adjacent detector and striking the detector's charge side (far left column),
from the detector's charge side (second column from left), from the adjacent
detector and striking the detector's
phonon side (third column), from the detector's phonon side (fourth column),
and from the housing (last column),
as labeled (see Fig.~\ref{schematicContribution} for a schematic).
The top row shows single-scatter events and the bottom shows multiple-scatter
events. The upper dark (blue) pair of curves represents the electron recoil
band while the nuclear
recoil band is shown by the lower, lighter (teal) pair of curves. Events
from $^{210}$Po decays that produce nuclear recoils are
highlighted in a lighter shade (orange) at Q near zero.
This simulation has $\sim$100$\times$ more events than expected in the WIMP-search data.
Note that the relative numbers of events in each plot are fixed here;
none of the relative normalizations (either of the 5 components or
of singles to multiples) are allowed to float. Furthermore,
note that decays on different surfaces cause quite different spectra,
which need to be considered
for a reliable background model.}
\label{SimContribution}%
\end{figure*}%
\subsection{The Full Background Model}
\label{FullBackgroundModel}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Figure8.pdf}
\caption{Spectra of alpha events observed in (from top to bottom) detector
T1Z2, T1Z5, T2Z5, and T3Z4.
Our analysis assumes that any alpha with
energy $>$~4~MeV is from a $^{210}$Po decay.
Shown are alpha events detected on the inner (``Q-inner", solid) and
outer (``Q-outer", dashed) electrodes. Q-outer events most likely
originate from the copper housings. }
\label{AlphaEventPlot}%
\end{figure}%
\begin{figure*}
\centering
\includegraphics[width=0.49\textwidth]{Figure9_1.pdf}
\includegraphics[width=0.49\textwidth]{Figure9_2.pdf}
\includegraphics[width=0.49\textwidth]{Figure9_3.pdf}
\includegraphics[width=0.49\textwidth]{Figure9_4.pdf}
\caption{(Color online) Stacked histograms of either phonon or ionization energy.
Figures (a) and (c) show the
ML best-fit result to WIMP-search data singles, while figures (b) and (d) show the ML
best fit to WIMP-search data
multiples. Figures (a) and (b) show the projection in ionization energy while (c) and
(d) show the projection in phonon energy. The four canvases in each figure show
the result for each of the four detectors.
The combined components of the surface-event background model are represented
by the solid green histogram (legend title: $^{210}$Pb PDF), while the
gamma-background model is shown in blue (legend title: Ba PDF).
The combined probability density functions from
simulation and calibration data are shown as the thick
line on top of the solid histograms (the thin lines
indicate the statistical uncertainties), while the
WIMP-search data is shown in black error bars. The orange histogram
(legend title: NR PDF) represents the
best-fit nuclear recoil-like component. The agreement is good, with T3Z4 having
the worst fit of the four detectors caused by two histogram bins after the low-energy peak
(figures (a) and (b)) which are not fit well.
The low-energy peak of that detector in the multiples is also not fit well,
which may be a result of mis-calibration due to a lack of penetration-depth
calibration data for this detector.}
\label{bestFit}%
\end{figure*}%
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{Figure10.pdf}
\caption{(Color online) A comparison between WIMP-search data (left column)
and the background model (right column) for one detector (T1Z2), with a
nuclear recoil (NR) component representative of an 8 GeV/c$^2$ WIMP shown
with the background-model singles.
The number of points displayed for the background model
has been reduced to match the number in the WIMP-search data.
The likelihood
fit is performed in this two-dimensional space of ionization versus~phonon energy.}
\label{bestFit_scatter}%
\end{figure}%
The gamma background is straightforwardly modeled with Ba calibration events.
The surface-event background is more difficult due to uncertainty in the locations
of the radiocontaminants, which are only partially constrained by the observed $^{210}$Po
alpha decays.
Figure~\ref{AlphaEventPlot} shows energy histograms of the observed alpha
events that produced most of their ionization signal in the inner sensor
(detector face) and on the outer sensor (housing).
Unfortunately, uncertainties on the observed alpha rates are large because saturation effects make it
difficult to reliably reconstruct events at such high energies, and
because CDMS~II detectors, in contrast to SuperCDMS~iZIPs~\cite{iZipRejection}, cannot
reliably determine whether an event occurred on the top or the bottom of a detector.
We construct the surface-event component of the detector-face background model by
assuming that 1/4 of the contamination is on each of the four flat surfaces: the
detector's top and bottom and the facing sides of the adjacent detectors.
Consequently, the number of simulated primaries from each detector face is equal.
The number of events expected from the detector housings can be constrained by
counting the number of alpha events that are identified as events occurring on
the outer wall of a detector. Hence the 5 components discussed at the end of
Section~\ref{SE} and in Fig.~\ref{SimContribution} are reduced to a ``Housing"
and a ``Detector Face" component. In the ML fit the relative normalization of
the different components is fixed accordingly.
\section{Introduction}
\input{Introduction.tex}
\label{intro}
\section{CDMS~II Detectors}
\label{detectors_CDMSII}
\input{Detectors.tex}
\section{The Background model}
\label{bg}
\input{BackgroundModel.tex}
\section{Maximum Likelihood analysis}
\label{MLA}
\input{MLH.tex}
\section{Systematics}
\label{systematics}
\input{Systematics.tex}
\section{Results}
\label{results}
\input{Results.tex}
\section{Comparisons to the Collar-Fields Style Fits}
\label{comp}
\input{ComparisonToCF.tex}
\section{Conclusion}
\input{Conclusion.tex}
\section{Acknowledgements}
\input{Acknowledgements.tex}
\bibliographystyle{ieeetr}
|
\section{Introduction}
\label{sec1}
Historically, the definition of noise was related to the sound: A noise is an unwanted, unpleasant and confusing type of sound.\footnote{In fact, the word noise is etymologically derived from the Latin word \emph{nausea}, meaning seasickness.} However, such definition is ambiguous. What does it mean unwanted, unpleasant or confusing? An attempt to provide a more academic definition comes from music: Noise is a non-harmonious or discordant group of sounds. Again, however, the definition is not free from ambiguities because one man's noise is another man's music \cite{Landauer1}.
A more scientific definition closer to our interest in electrical devices comes from communications: A noise is an electric disturbance that interferes with or prevents reception from a signal or information. For example, the buzz in a telephone call. Thus, we realize that once we have a precise definition of what is a signal, the meaning of what is noise becomes perfectly clear: It is the difference between the measured value and the signal.
\subsection{Quantum noise in electrical devices from an experimental point of view}
\label{sec1.1}
As discussed above, the answer to what is noise in electrical devices depends on our definitions of the electrical signal. For most DC applications, the signal is just a time average value of the current. For frequency applications, the signal is equivalently defined as a time average value, but using a shorter time interval (related to the inverse of the operating frequency). In other applications, mainly digital applications, the signal is related to a time average value of the voltage in a capacitor. Hereafter, we will assume that the electrical signal is the DC value of the current, referenced by the symbol $\langle I \rangle$. All fundamental and practical issues discussed here for the DC signal (and its noise) can be easily and straightforwardly extended to those other types of electrical signals.
What is measured in a laboratory for the DC signal is the time average value of the instantaneous current $I(t)$ in a unique device during a large period of time $T$:
\begin{equation}
\langle I \rangle = \lim_{T\rightarrow \infty}\frac{1}{T}\int^{T}_0 I(t) dt.
\label{noise1}
\end{equation}
Once we have defined the signal $\langle I \rangle$ as the DC value, in principle, the noise can be quantified by time averaging the difference between the measured value of the current $I(t)$ and the signal in a unique device:
\begin{equation}
\triangle I^2 = \lim_{T\rightarrow \infty}\frac{1}{T}\int^{T}_0 (I(t)-\langle I \rangle)^2dt.
\label{noise2}
\end{equation}
The square of the difference avoids positive and negative cancellations.\\
At this point, it is very important to realize that $I(t)$ presents very rapid fluctuations that cannot be captured by the standard laboratory apparatuses. Any experimental setup that measures the current fluctuations behaves as a low-pass filter (i.e. the current fluctuations at frequencies higher than the apparatus cut-off frequency are not measured). Therefore, the experimentally accessible information about the current fluctuations is not given by \eref{noise2}, but by the power spectral density of the fluctuations $S(w)$ (and its related magnitudes). From the Wiener-Khinchine relation, the power spectral density can be defined as the Fourier transform of the time average definition of this autocorrelation function $\triangle R(\tau)$ :
\begin{equation}
\triangle R(\tau) =\lim_{T\rightarrow \infty}\frac{1}{T}\int^{T}_0 \triangle I(t_1) \triangle I(t_1+\tau) dt_1,
\label{noise3}
\end{equation}
where $\triangle I(t)=I(t)-\langle I \rangle$. A straightforward calculation shows that \eref{noise3} can be rewritten as $\triangle R(\tau) = R(\tau)-\langle I \rangle^2$ with:
\begin{eqnarray}
R(\tau) = \lim_{T\rightarrow \infty}\frac{1}{T}\int^{T}_0 I(t_1) I(t_1+\tau) dt_1.
\label{noise4}
\end{eqnarray}
Then, the Fourier transform of \eref{noise3} gives the noise power spectral density $S(w)$:
\begin{eqnarray}
S(w) = \int_{-\infty}^{\infty} \triangle R(\tau) e^{-jw\tau} d\tau.
\label{noise5}
\end{eqnarray}
It is quite trivial to realize that the definition of the spectral density $S(w)$ in \eref{noise5} and \eref{noise3} is consistent with the definition of the total noise\footnote{Technically, $S(w)$ defined in \eref{noise5} is non-negative and symmetric with respect to $w$. Then, since only positive frequencies $w$ are measured in a laboratory, the measured density includes our $S(w)$ and $S(-w)$, and the integral of the noise spectrum measured in a laboratory runs from $0$ till $\infty$.} in \eref{noise2}:
\begin{eqnarray}
\triangle I^2 = \int_{-\infty}^{\infty} S(w) dw,
\label{noise6}
\end{eqnarray}
where we have used the definition of the delta function $\delta(\tau)=\int_{-\infty}^{\infty} e^{-jw\tau} dw$. \\
It is very relevant for the rest of the paper to realize that the measurement of $S(w)$ through the function $R(\tau)$ defined in \eref{noise4} requires the knowledge of the measured value of the current during all $t$. Thus, we have to make predictions about the evolution of the electronic device while being (continuously) measured. In a classical scenario, such discussion about measurement is generally ignored. On the contrary, for quantum systems, it has very relevant implications because the evolution of a system with or without measurement can be dramatically different.
If the electronic device satisfies the ergodic theorem \cite{Price,ergodic}, a continuous measurement of the system can be avoided in practical computations. Let us see in what sense ergodicity can simplify our noise computations. In general, the \emph{mathematical} concept of a random process is used to deal with noise. A random process requires a sample space. In our case, we can define an ensemble of \emph{identical} electrical devices\footnote{At this point, the reader will wonder that, in typical laboratory experiments, only one electronic device is available (not an ensemble of them). Then, as a practical definition of ensemble, we can define the instantaneous current measured in different time-intervals: $I^{\gamma_1}(t)$ for the instantaneous current measured during the first time interval, $I^{\gamma_2}(t)$ for the second interval, and so on. }, each one labeled by the sample space variable $\gamma$. Then, the (instantaneous) current is labeled by the random process $I^{\gamma}(t)$. For a fixed time, $t_1$, the quantity $I^{\gamma}(t_1)$ is a random variable. For a fixed device $\gamma_1$, the function $I^{\gamma_1}(t)$ is a well-defined non-random function of time. Finally, $I^{\gamma_1}(t_1)$ is just a real number. Often the sample space variable $\gamma$ is omitted in the notation. The DC value of the current in \eref{noise1} can be alternatively defined for an ergodic system as:
\begin{equation}
\langle I \rangle = \sum_{i} I_{i}(t_1) P(I_i(t_1)),
\label{noise7}
\end{equation}
where $P(I_i(t_1))$ is the probability of getting $I_i$ at time $t_1$. These probabilities are defined as the ratio of the number of devices providing $I_i$ divided by the total number of devices. It is important to realize that the experimental evaluation of \eref{noise7} requires only one measurement of the current at $t_1$ in a large number of \emph{identical} $\gamma$-devices. Then, the theoretical predictions of \eref{noise7} do only need to determine the free (without measuring apparatus) evolution of the electronic device from the initial time $t_0$ till $t_1$. See a detailed discussion in appendix \ref{app-noise} about how ergodicity avoids the complications of the measurement in a quantum system. Obviously, we can compute the total noise represented in \eref{noise2} from a unique measurement in ergodic systems:
\begin{equation}
\triangle I^2 = \sum_{i} (I_{i}(t_1)-\langle I \rangle )^2 P(I_i(t_1)),
\label{noise7bis}
\end{equation}
However, the noise measured in a laboratory is not given by $\triangle I^2$, but by $S(w)$ in \eref{noise5}. We repeat the reason explained in \eref{noise2}. The amount of noise generated by an instantaneous current evolving for example from $I(t_1)=5$ mA to $I(t_2)=10$ mA during a time interval of $t_2-t_1=\tau=1$ fs, is not captured from the state-of-the-art laboratory apparatuses (which already have difficulties to capture noise at frequencies higher than a few of Terahertzs). From an experimental point of view, in fact, it is easy to get $S(w\rightarrow 0)$, but impossible to get $S(w\rightarrow \infty)$. We can compute the noise power spectral density $S(w)$ from the ensemble average version of the autocorrelation defined in \eref{noise4} as:
\begin{eqnarray}
R(t_1,t_2)=\sum_i\sum_jI_j(t_2)I_i(t_1)P\big(I_j(t_2),I_i(t_1)\big).
\label{noise8}
\end{eqnarray}
In general, we can assume that the instantaneous current in an electronic device behaves as a wide-sense stationary random process. Then, $\langle I \rangle$ in \eref{noise7} is constant and time-independent. Identically, then, the autocorrelation function in \eref{noise8} depends only on the time difference $ R(t_1,t_1+\tau)= R(\tau)$ with $t_2=t_1+\tau$. Finally, we use \eref{noise5} with $\triangle R(\tau) = R(\tau)-\langle I \rangle^2$ computed from \eref{noise8}, to get the noise power spectral density $S(w)$.
It is important to emphasize (for a posterior discussion) that the probability $P\big(I_j(t_2),I_i(t_1)\big)$ implies a two-measurement process for each electronic device. The system evolves freely (without interaction with the measurement apparatus) from $t_0$ till $t_1$ when the current is measured, giving the value $I_i$. Then, the system evolves freely again until time $t_2$, when the system is measured again giving $I_j$. In summary, even if the ergodicity argument is invoked, the noise computation through the autocorrelation function requires, at least, two measurements at different times in a single device (and the average over all $\gamma$-devices). We anticipate that our computations with Bohmian mechanics will not assume ergodicity (which is not an obvious property for open systems out of equilibrium \cite{Price}), but the prior expressions requiring a continuous measurement of the current.
Let us emphasize that the previous discussion is valid for either classical or quantum devices. The adjective \emph{quantum} emphasizes that the signal and the noise are computed or measured in an electrical device governed by quantum laws \cite{Landauer2,Beenakker1,Beenakker2,Buttiker1}. If the electronic device is not ergodic, expression (\ref{noise4}) requires a continuous measurement of the current $I(t)$. On the contrary, for an ergodic electron device, expression (\ref{noise7}) requires one unique measurement, while expression (\ref{noise8}) requires a two-times measurement when dealing with the power spectral density $S(w)$.
Up to here, we realize that the definition of quantum noise seems very trivial. Then, why does the concept of quantum noise have a halo of mystery around it?
\subsection{Quantum noise in electrical devices from a computational point of view}
\label{sec1.2}
Our previous definition about what is quantum noise does not answer the question of how we compute it. If we want to predict the values $I(t)$ used in \eref{noise2} and \eref{noise4} or the probabilities $P(I)$ and $P\big(I_j(t_2),I_i(t_1)\big)$ for \eref{noise7} and \eref{noise8}, we require a quantum theory.
There are several quantum theories available in the literature that, by construction, are empirically equivalent when explaining all quantum phenomena. Among others, the so-called Copenhagen or orthodox interpretation \cite{Copenhagen1,Copenhagen2}, Bohmian mechanics \cite{Bohm,OriolsBook,NinoBook} or the many-worlds theory \cite{Everett}. Any theory has usually two different planes. First, the formalism, which is a set of mathematical rules (using elements such as wave functions, operators, trajectories) that allow
us to make practical computations that reproduce experimental results. The formalism of a theory provides an answer to the question: How do we compute quantum noise? The second plane of a theory is its interpretation. It tries to provide a deep connection on how the mathematical rules and its elements explain how nature works. The interpretation of the theory provides answers to the question: Which is the physical origin of quantum noise? Each quantum theory will provide its own answers to both questions.
Many people argue that the only important part of a quantum theory (once we know it is empirically valid) is its formalism because it is the only part we need to make computations. Certainly, one can make noise computations using any of the available formalisms without worrying about its interpretation. At the end of the day, by construction, each theory should give the same predictions. Other people argue that even when one is only interested in computations, a correct understanding of the interpretational issues of each theory is fruitful because it provides an enlarged vision about how correctly apply the theory in unsolved problems (abandoning the \emph{shut up and calculate} \cite{SUAC}). We will return to this very point later, at the conclusions in \sref{sec5}.
Now, we want to clarify why quantum noise is specially sensible to fundamental quantum mechanical issues. Any electrical device (or any experiment) is connected to a measuring apparatus. In our case, an ammeter to get the electrical current. Quantum noise is sensible to the (ammeter) measuring process. As stated in \eref{noise8}, in order to obtain the noise, the quantum system has to be measured, at least, twice. This two-time measurement faces directly with one of the most complex issues in quantum mechanics, namely, which is the perturbation of the quantum wave function when a measurement is performed. Historically, this perturbation is known with the somehow scary name of \emph{the collapse of the wave function}. Can we ignore it? Definitively not if temporal correlations need to be correctly predicted. See for example, Ref. \cite{Buttiker3}:\emph{\lq\lq{}The fluctuations ... are a consequence of a probabilistic reflection and transmission probability (a wave phenomena) and are a consequence of the fact that detectors register either a transmitted or a reflected particle (a particle phenomena)\rq\rq{}}. The measurement process is hidden in the word \emph{detectors}.
We also mention that the fundamental understanding/computing of the measurement process can be largely relaxed when dealing with DC predictions. They can be computed from an ensemble of devices with only one measurement in each device, so we can ignore the evolution of the quantum system after the measurement. See \aref{app-noise} to enlarge this point.
In this paper, we will provide an explanation to the role of collapse in quantum noise from a Bohmian perspective. We emphasize that we are not saying that the Bohmian answer is the best one. Answers from other theories are equally satisfactory, and provide the same predictions. We are just defending that it is a consistent answer that in the authors\rq{} opinion provides a quite intuitive and understandable explanation of quantum noise and also a numerically accessible formalism. In \sref{sec2} we explain how the Copenhagen interpretation explains the multi-time measurement process in a experiment with a flux of electrons impinging upon a tunneling barrier, by introducing the notion of operators. In \sref{sec3} we provide an explanation of the same experiment using Bohmian mechanics, without using operators. Then, in \sref{sec4} we illustrate how the formalism of Bohmian mechanics exposed in \sref{sec3} can be applied in practical problems to calculate the quantum noise in electrical devices, including Coulomb and exchange interactions. Finally, we conclude in \sref{sec5} explaining how the different theories explain the origin of quantum noise.
\section{Multi-time measurement with operators}
\label{sec2}
A typical scenario when discussing quantum noise in electrical devices is a flux of electrons impinging upon a partially transparent barrier (located in the middle of the active region). Electron transport through the barrier takes place by tunneling. Electron is either transmitted or reflected, but not both! \cite{Beenakker1,Buttiker2,Landauer3} We get a transmitted electron with a probability $T$, while a reflected one with probability $R=1-T$. To simplify the discussion, we consider a constant injection of electrons (at zero temperature), one by one. Each electron, after measurement at time $t_1$, will appear randomly at the left or at the right of the barrier. The time averaged number of transmitted electrons will be proportional to $T$, but the number of transmitted electrons fluctuates instantaneously because of the randomness of the transmission. These fluctuations of the number of transmitted electrons (when compared with the DC signal) are named partition noise \cite{Beenakker2,Buttiker3,Landauer3}.
There are many other sources of noise in electrical devices, for example, the $1/f$ noise which become very relevant at low frequencies \cite{Beenakker2,Buttiker1}. In this paper, we will only deal with partition noise due to a tunneling barrier. In \sref{sec4}, we will discuss partition noise considering also the injection of electrons at a finite temperature (the so-called thermal noise). The fluctuations due to both processes simultaneously are known in the literature as shot noise \cite{Landauer1,Beenakker1,Buttiker1,Levitov}.
In this section we discuss how the partition noise is understood within the orthodox interpretation of quantum mechanics, also known as Copenhagen interpretation \cite{Copenhagen1,Copenhagen2}. Let us specify that most available formulations of shot noise are developed within this orthodox interpretation \cite{Landauer1,Beenakker1,Beenakker2,Buttiker3,Buttiker2,Landauer3}. We consider a very simple example, but with a detailed discussion of the role played by the measuring apparatus (the ammeter). The Copenhagen interpretation associates a wave function $\Psi(\bar{x}_N,t)$ to a system of $N$ particles. In principle, such wave function \emph{lives} in a $3N + 1$ dimensional configuration space. Within the first non-relativistic quantization language, the evolution of this wave function is defined by two laws \cite{Cohen}. The first law, known as Schr\"odinger equation, states that (when the system is not measured) the wave function evolves unitarily and deterministically according to the following equation
\begin{equation}
i\hbar \frac{\partial \Psi(\bar{x}_N,t)}{\partial t}=H\Psi(\bar{x}_N,t),
\label{TDSE}
\end{equation}
where $H=\big[\sum_i-\frac{\hbar^2}{2m_i}\nabla^2_i+U(\bar{x}_N,t)\big]$. With $U(\bar{x}_N,t)$ we denote a generic interaction potential in the position representation, with $m_i$ the mass of the $i$-th particle and with $\bar{x}_N = (x_1,x_2,..., x_N)$ the multidimensional vector in the configuration space. To provide a simple discussion of the partition noise in a tunneling barrier, let us assume that each electron in our experiment can be described by a single-particle wave function (we neglect the exchange and the Coulomb interaction among electrons). In Fig.~\ref{figure1} we plot the (unitary) evolution of such wave function solution of \eref{TDSE}. Is the (unitary) Schr\"odinger equation alone depicted in Fig.~\ref{figure1} enough to understand quantum noise? No. The orthodox theory has a second law, known as the \emph{collapse} of the wave function, that takes into account the effects of the interaction of a measuring apparatus with the quantum (sub)system. This second law can be found in many textbooks \cite{Cohen}. It requires a new non-unitary operator $A$. This operator is different from the unitary Schr\"odinger evolution, which is generated by the Hamiltonian seen below \eref{TDSE}, and it must be able to encapsulate all the interactions of the quantum systems with the rest of the particles (including the ammeter, the cables, the environment, etc). This new operator $A$ is the only tool provided by the theory to determine the possible results of a measurement. In principle we do not know anything about this operator except that it is a (hermitian $A = A^{\dagger}$) function whose (real) eigenvalues $a_n$ of its spectral decomposition are the possible results of the measurement. Once the system in Fig.~\ref{figure2} is measured (and not before), the wave function is projected to one of the eigenstates of the mentioned operator in a non-unitary evolution.\footnote{The measurement described in most textbooks is called ``projective" (\emph{strong}) measurement. There exists, for example, another type of measurement known as \emph{weak} measurement, which is useful to describe situations where the effects of the apparatus on the measured system is just a small perturbation.} After the collapse, the \textit{new} wave packet evolves again according to the time-dependent Schr\"odinger equation until a new measurement is done.
\begin{figure}
\includegraphics[width=0.97\columnwidth]{figure1.pdf}
\caption{Evolution of the squared modulus of the wave function of an electron impinging on a tunneling barrier (green solid line). We plot four different times corresponding to (a) initial time, (b) the moment when the wave function interacts with the barrier, (c) the time $t_1$ when it occurs the first measurement and (d) time $t_2$ corresponding to the second measurement. At time $t_1$ and $t_2$, because of the unitary evolution, the electron can be detected at both sides of the barrier (Color figure online)}
\label{figure1}
\end{figure}
For simplicity, in our present conceptual discussion let us assume a reasonable (but ad-hoc) operator (why this operator is reasonable will be clarified in \sref{sec3}). Such operator provides the following perturbation of the wave function. If the electron is \emph{randomly} measured as a reflected electron at $t_1$, the transmitted part of the wave function is eliminated. This measuring process corresponds to Figs. \ref{figure2} (c) and (d) where only the reflected wave function survives after $t_1$. Equivalently, the measurement process associated to \emph{randomly} getting a transmitted electron corresponds to eliminating the reflected part, as seen in Figs. \ref{figure2} (g) and (h).
Now, by comparing the evolutions of the wave functions in \fref{figure1} and \fref{figure2}, it is obvious that the former is wrong. By looking at \fref{figure1}, it could be the case that an electron found at time $t_1$ at the right (transmitted) can be found in a later time $t_2$ at the left as a reflected electron (see the evolution of the probability density in \fref{figure1}). This sequence of possibilities is wrong. Experimental results confirm that once, say time $t_1$, the electron is detected at one side, in a later time $t_2$ it is always found at the same side. Then, we get a very valuable lesson from the Copenhagen explanation: the (unitary) Schr\"odinger equation alone is not able to explain completely quantum noise. It is necessary to include the collapse of the wave function to understand properly what is quantum noise (temporal correlations). The popular arguments that \emph{\lq\lq{}Shot noise is a consequence of quantization of charge\rq\rq{}} \cite{Buttiker1} or \emph{``This is the noise that arises from the graininess of the current''} \cite{Landauer3} emphasize exactly this very point.
\begin{figure}
\includegraphics[width=0.97\columnwidth]{figure2.pdf}
\caption{(a), (b), (c) and (d) Non-unitary evolution of the wave function for a reflected electron detected at time $t_1$ at the left side. (e), (f), (g) and (h) Non-unitary evolution of the wave function for a transmitted electron detected at time $t_1$ at the right side. Symbols are the same as in \fref{figure1} (Color figure online)}
\label{figure2}
\end{figure}
All mentioned orthodox formalisms dealing with quantum noise reproduce experimental results successfully because they include the measurement process inside \cite{Landauer1,Beenakker1,Beenakker2,Buttiker3,Buttiker2,Landauer3,Levitov}.
Most of them do not discuss explicitly which is the operator associated with the ammeter.
Over the years, physicists have identified the operators, by developing instincts on which are the effects of measurements in the wave function. There are scenarios (as the one depicted in \fref{figure2}) where it is quite obvious which operator is the \emph{right} one. On the contrary, for example, when measuring the total (conduction plus displacement) current it is not at all obvious which are the relevant operators. Is this measurement process \emph{continuous} or \emph{instantaneous}? Does it provide a \emph{strong} or a \emph{weak} perturbation of the wave function? The answers to these questions are certainly not simple. The Copenhagen theory itself does not answer these \emph{technical} questions about how to find the \emph{right} operator. Can other quantum theories provide additional help?
\section{Multi-time measurement without operators}
\label{sec3}
In the previous section, we discussed how the Copenhagen formalisms can be successfully used to understand quantum noise in electrical devices. One technical difficulty with this formalism is the proper definition of the \emph{right} operator that determines the collapse of the wave function when, for example, the total (conduction plus displacement) current is measured.
There are alternative theories which account for the perturbation of the wave function during a measurement process in a different way, without operators. The one that we will develop here is Bohmian mechanics. Let us emphasize again that both (Copenhagen and Bohmian) theories are empirically equivalent so that the preferences of one in front of the other are related to computational abilities, intuitive results, etc. \cite{OriolsBook}.
In the Bohmian theory, the complete description of a quantum system of $N$ particles is given by the (same) wave function, $\Psi(\bar{x}_N,t)$ mentioned in \sref{sec2}, and by the actual positions of the point-like particles, $\bar{X}_N(t) = (X_1(t), X_2(t), ..., X_N(t)) $\footnote{We denote with capital letter $X$ the actual position of the particle, while the lower case letter $x$ is used to indicate generic positions. With the barred letter we refer to a multidimensional vector in the configuration space, while a letter without bar denotes the 3-dimensional vector in physical space.} (see \aref{app-bohm} and Refs. \cite{OriolsBook,NinoBook} for a more detailed discussion on this theory).
We emphasize that the evolution law for the wave function $\Psi(\bar{x}_N,t)$ is the same as in standard quantum mechanics: the Schr\"odinger equation (\eref{TDSE}). The wave function \emph{guides} the movement of the actual positions of the particles in time, according to the so-called \emph{guidance equation}, which defines the velocity of the $i$-th particle as
\begin{eqnarray}
v_i(t) &=& \frac{J_i(\bar{x}_N,t)}{|\Psi(\bar{x}_N,t)|^2}\Big|_{\bar{x}_N=\bar{X}_N(t)} = \nonumber \\
&=& \frac{\hbar}{m_i} \text{Im} \frac{\nabla_i \Psi(\bar{x}_N,t)}{\Psi(\bar{x}_N,t)}\Big|_{\bar{x}_N=\bar{X}_N(t)},
\label{eq-guidance}
\end{eqnarray}
where $J_i(\bar{x}_N,t)$ is the usual probability current density, defined as $J_i(\bar{x}_N,t) = \frac{i\hbar}{2m_i}\left(\Psi \nabla_i \Psi^*-\Psi^* \nabla_i \Psi \right)$, associated to the $i$-particle, $|\Psi(\bar{x}_N,t)|^2$ is the usual probability distribution and Im denote the imaginary part. We note that \eref{eq-guidance} describes the \emph{velocity field} for the $i$-particle and depends on the actual position of all particles of the system $\bar{X}_N(t)$. Each particle follows a definite trajectory which can be obtained integrating in time the velocity field
\begin{equation}
X_i(t) = X_i(0) +\int_0^t v_i(t') dt',
\label{eq-position}
\end{equation}
where $X_i(0)$ is the initial position of particle $i$.\\
A proper ensemble of these trajectories (proper means that the initial position of each trajectory of the ensemble is selected according to the initial squared modulus of the wave function, see \eref{QEH} in \aref{app-bohm-3}) reproduces the time-evolution of the many-particle wave function at any later time. \\
In \sref{sec2}, we saw that in order to reproduce the experimental results, we have used the notion of operators to describe how the wave function of a measured system is modified under a measurement process. In the Bohmian theory, we simply consider the apparatus as another (big and complex) quantum system interacting with our measured system. The interaction among them is then included in the Hamiltonian of \eref{TDSE} as any other interaction. Then from the unitary evolution of the many-particle wave function (system plus apparatus) we can look at the behavior of the wave function of the measured system. The latter is called \emph{conditional wave function} (an exclusive concept belonging to Bohmian mechanics) and it is defined from the many-particle wave function in the configuration space, fixing all the actual particles positions excluding that of our subsystem (see \aref{app-bohm-2} for more details).
Let us provide a quite realistic (in particular, non-instantaneous, but in some ways schematic) example in which we can numerically track the behavior of the conditional wave function during the measurement process of the partition noise discussed in \sref{sec2}. The quantum system is an electron labelled as $X_1$ impinging on an
external tunneling barrier. Behind the barrier there is a measuring device, that we call ``transmitted charge detector" modeled as a single degree of freedom $X_2$ (thought as the center of mass of a complex system), which can detect the successful transmission of an electron. First, we have an interaction of the electron with the potential barrier and, subsequently, an interaction with the transmitted charge detector. It is important to stress that both interactions are regarded at the very same level within Bohmian mechanics. The measurement interaction introduces a channelling of the wave function in the configuration space such that the desired property of the ``quantum system'' (here, whether the electron is reflected or transmitted) can be read off from the final position $X_2$ of a particle, thought of as the pointer of the apparatus.
The interaction between the electron and the pointer can be modeled as:
\begin{eqnarray}
H_{int}= \lambda Q(x_1) P_{x_2} = -i \hbar\lambda Q(x_1) \frac {\partial } {\partial x_2},
\label{eq-interaction2}
\end{eqnarray}
where $P_{x_2} = -i\hbar \partial /\partial x_2 $ is the momentum operator of the detector and $\lambda =50\; nm/ps $ is the interaction constant. $Q(x_1)$ is a function that is equal to zero when the electron is outside the detector, ($x_1 < 75 \; nm$ in \fref{conditional wave function-figure1}), and is equal to one when the particle is inside the detector ($x_1 > 75 \; nm$).\footnote{The transition of $Q(x_1)$, from zero to one, is done softly in order to minimize the perturbation of the ``quantum system'' as explained in \cite{albareda}.} In \fref{conditional wave function-figure1} the region in the configuration space in which this function is different from zero is represented by a rectangle.
The many-particle Schr\"odinger equation reads
\begin{eqnarray}
&&i \hbar \frac{\partial \Psi(x_1,x_2,t)}{\partial t}=
\Big(- \frac{\hbar^2}{2m}\frac {\partial^2} {\partial x_1^2} - \frac{\hbar^2}{2 M}\frac {\partial^2} {\partial x_2^2} + \nonumber \\
&+& U(x_1) - i \hbar \lambda Q(x_1) \frac {\partial} {\partial x_2}\Big) \Psi(x_1,x_2,t), \; \; \;
\label{eq-schr2D}
\end{eqnarray}
where $m$ is the effective mass of the electron, $M$ is the mass of
the apparatus pointer and $U(x_1)$ is the external potential energy barrier.
\begin{figure}
\centering
\includegraphics[width=0.97\columnwidth]{figure3.pdf}
\caption{Time evolution of the squared modulus of $\Psi(x_1,x_2,t)$ at four different times. The configuration space region where the \emph{transmitted charge detector} is present is indicated by a rectangle and the barrier by a solid line. The $+$ line indicates the modulus of the conditional wave function $|\psi_R|^2=|\Psi(x_1,X_2^{\alpha=1}(t),t)|^2$, while the $\odot$ line corresponds to $|\psi_T|^2=|\Psi(x_1,X_2^{\alpha=3}(t),t)|^2$. Four trajectories $\{X_1^\alpha(t),X_2^\alpha(t)\}$ with different initial positions are presented with $\square$, $*$, $\times$ and $+$. The actual detector position associated with the reflected trajectory ($+$) with $\alpha=1$ does not move because there is no interaction between this trajectory and the detector (Color figure online)}
\label{conditional wave function-figure1}
\end{figure}
The main feature of a transmitted charge detector is that the center of mass of the wave function in the $x_2$ direction has to move if the
electron is transmitted and it has to be at rest if the electron is reflected. We solve \eref{eq-schr2D} numerically considering as initial wave function the products of two gaussian wave packets, i.e. $\Psi(x_1,x_2,0) = \psi(x_1,0)\phi(x_2,0)$. All details of this simulation can be found in \cite{albareda}. In particular we are considering $M=75000 \; m$. In \fref{conditional wave function-figure1} the numerical solution of the squared modulus of $\Psi(x_1,x_2,t)$ is plotted at four different times. At the initial time $t = 0$, \fref{conditional wave function-figure1}~(a), the entire wave function is at the left of the barrier. At a later time $t_0$ the wave function has split up into reflected and transmitted parts due to the barrier, see \fref{conditional wave function-figure1}~(b). Then, because the electron has not yet arrived at the transmitted charge detector, the wave function has the following form:
\begin{equation}
\Psi(x_1,x_2,t_0) = \left[\psi_{T}(x_1,t_0)+\psi_{R}(x_1,t_0)\right]\phi(x_2,t_0).
\end{equation}
After that, Figs. \ref{conditional wave function-figure1} (c) and (d), the
interaction of the detector with the transmitted part of the wave function appears. For time $t>t_0$ the transmitted part of the wave
function is shifted up in the $x_2$ direction while the reflected part does not move. The interaction with the apparatus thus produces
two channels in the configuration space, one corresponding to the electron being transmitted and the other corresponding to the electron being
reflected, getting an entangled superposition among the electron and the apparatus.
\begin{figure}
\centering
\includegraphics[width=0.97\columnwidth]{figure4.pdf}
\caption{ The $+$ line in (a), (b), (c) and (d) is the time evolution of the squared modulus of the conditional wave function associated to the trajectory $\alpha = 1$ in \fref{conditional wave function-figure1}, i.e. $\psi_{R} = |\Psi(x_1,X_2^{\alpha=1}(t),t)|$. The $\odot$ line in (e), (f), (g) and (h) is the squared modulus of the conditional wave function associated to the trajectory $\alpha = 3$ in \fref{conditional wave function-figure1}. i.e. $\psi_{T} = |\Psi(x_1,X_2^{\alpha=3}(t),t)|$. The actual detector position $X_2(t)$ is plotted at each time in order to compare these results with those in \fref{conditional wave function-figure1} (Color figure online)}
\label{conditional wave function-figure2}
\end{figure}
In \fref{conditional wave function-figure1} we also plot the actual positions of the system and detector $\{X_1(t),X_2(t)\}$ for four different possible initial positions $\{X_1(0), X_2(0)\}$, corresponding (say) to four distinct runs of the experiment (labelled by $\alpha=1,...,4$). Of the four possible evolutions shown, three show the electron being transmitting ($\alpha=2,3,4$) and one being reflecting ($\alpha=1$). While the pointer position $X_2(t)$ does not move for the reflected particle, its evolution for the transmitted ones clearly shows a movement. In conclusion, looking at the \emph{detector} position we can perfectly certify if the particle has been reflected ($X_1(t)< -50\;nm$ and $X_2(t)=0\;nm $) or transmitted ($X_1(t)> -50\;nm$ and $X_2(t) \approx 15\;nm $). We hope the reader will realize how trivially we have been
able to explain the measurement, using only a \emph{channelized} (unitary) time-evolution of 2D wave function plus two Bohmian trajectories, one for the system and another for the measuring apparatus.
Once we have solved the complete problem of the measurement in (2D) configuration space, we can describe the same measurement in (1D) physical space with the help of the conditional wave function. The key point illustrated here is that the collapse of the one-particle wave function for the electron, which collapse is of course postulated through the second law seen in \sref{sec2} in ordinary quantum theory, instead arises naturally and automatically in Bohmian mechanics. It is simply a consequence of slicing the unitary-evolving (2D) wave function $\Psi$ along the (moving) line $x_2 = X_2(t)$, resulting $\psi_1(x_1,t)=\Psi(x_1,X_2(t),t)$. In \fref{conditional wave function-figure1} we have plotted two solid horizontal lines corresponding to a slice of the wave function at two different values of $X_2(t)$. In \fref{conditional wave function-figure2} we report the evolution of these (time-dependent) slices of the many-particle wave function, the \emph{conditional wave function} for the electron, for the trajectories $\alpha = 1$ and $\alpha = 3$ from \fref{conditional wave function-figure1}. We clearly see that if the particle is reflected, as it is the case for $\alpha = 1$, the position of the pointer does not change with time and, after the interaction with the detector has been performed, the electron's \emph{conditional wave function} includes only a reflected part. See Figs. \ref{conditional wave function-figure2} (c) and (d). On the other hand, when the particle is transmitted (e.g., $\alpha =3$), it is the reflected part of the \emph{conditional wave function} which collapses away, leaving only the transmitted packet. See Figs. \ref{conditional wave function-figure2} (g) and (h). Note in particular that the 1D evolution of $\psi_1(x_1,t)$ (the electron's \emph{conditional wave function}) is not unitary, even though the 2D evolution of $\Psi$ is.
While a wave function formulation of quantum mechanics provides only statistical information about the experimental results, with the help of the Bohmian trajectories, we have been able to recover the individual result of each experiment. In fact, for each experiment the pointer of the detector is either moving (corresponding to a transmitted electron) or not (reflected electron), while an ensemble of repeated experiments (where the initial positions of the particles, both the electron $X_1(0)$ and the detector $X_2(0)$, are selected according to the squared modulus of the wave function at the initial time $|\Psi(x_1,x_2,0)|^2$) reproduce the same statistical results.
Thus with the previous numerical example we have reproduced the collapse-behaviour of the wave function of a transmitted (or reflected) electron. This allows us to conclude that the same results of standard formalism (explained in \sref{sec2}) are obtained within Bohmian mechanics (see \cite{OriolsBook,Goldstein1} for a formal derivation of the empirical equivalence of the two theories). Apart from irrelevant technicalities (related to how we define the measuring apparatus) the results in \fref{figure2} and \fref{conditional wave function-figure2} are conceptually identical. We emphasize that, the collapse in Bohmian theory is naturally derived. Such a natural derivation of the collapse behavior demystifies the measurement process (and the quantum noise). We underline that we achieve the non-unitary evolution of the wave function of a measured system simply slicing the enlarged wave function (which includes the apparatus) in the configuration space.
Let us return to the questions posed at the end of \sref{sec2} about the measurement of the total (conduction or displacement) current. Is this measurement process \emph{continuous} or \emph{instantaneous}? Does it provide a \emph{strong} or a \emph{weak} perturbation of the wave function? The Bohmian theory does not provide simple answers to these questions, but it clearly indicates the path. We need to include (somehow) the measuring apparatus in the Hamiltonian. Here, the electrostatic interaction between the electrons in the system and those in the contacts, cables, etc. See a preliminary work in this direction in Ref. \cite{IWCE-Paris}.
A powerful simulator which uses Bohmian mechanics to compute DC and quantum noise is the BITLLES simulator \cite{xavier2,BITLLES}. This simulator allows us to work with a lot of flexibility, being able of introducing any sort of potential, including Coulomb correlations and exchange interaction. The next Section is dedicated to expose the fundamental ideas of the simulator and an example of the calculation of noise with it.
\section{Practical application}
\label{sec4}
We have previously exposed the main features of Bohmian mechanics explaining in a quite trivial way the partition noise in a tunneling barrier. However, as it occurs for all theories, there is a huge step between its general formulation and its practical application. In fact, it happens many times that the practical problem we want to solve is unsolvable both analytically and numerically, and some kind of approximations are required. The paradigmatic example of the need of approximations in quantum theories is the well-known \emph{many-body problem} that reminds us that the celebrated Schr\"odinger equation in \eref{TDSE} (or any alternative formulation) can only be solved exactly for very few (one, two, three,..) degrees of freedom \cite{Manybody,Dirac}.
In principle we have seen in \sref{sec3} that to reproduce the collapse of the wave function in Bohmian mechanics we have to include a suitable interaction with an external apparatus. Then we can write down the Schr\"odinger equation for our complete system including all the electrons in the active region of the device plus all the particles composing the detector. But solving numerically this problem is obviously impossible. Again, the \emph{many-body problem} appears. Then we should look for suitable approximations able to reduce the complexity of our problem. Let us emphasize that the (technical) approximations that we will show, do not alter the general framework we have previously presented.
\subsection{An approximation for the interaction between the electron and the measuring apparatus}
\label{sec4.1}
The first kind of approximation regards the inclusion of the apparatus in our simulations. It seems that its inclusion is unavoidable in order to provide the collapse of the wave function. And this is true, but in the particular case of quantum noise in electrical devices, the fact of \emph{playing} with (Bohmian) trajectories will greatly simplify the problem. In \aref{app-bohm-3}, we have reported how any experimental value is calculated in Bohmian mechanics. The important thing is that any expectation value of a given operator is simply calculated as a function of the actual particles positions over an ensemble of repeated experiments (see \eref{mean-value} in \aref{app-bohm-3}). Thus what really matters in the computation of a property of the quantum system are only the trajectories of the Bohmian particles (not the wave functions). Therefore, if the trajectories without measuring apparatus are enough accurate (this means if the error on these trajectories due to neglecting the apparatus is reasonably small compared to the exact solution) we can get accurate results with a minimal computational effort. In the case of the transmitted charge detector of \sref{sec3}, it has been demonstrated \cite{albareda} that the error due to the exclusion of the apparatus from the simulations is almost negligible for the computation of the trajectories. In this way we can decrease enormously the computational burden, removing all the degrees of freedom related to the apparatus from our computations.
We can provide a more didactic discussion on why the previous technical approximation for the measuring apparatus works quite well when using Bohmian trajectories. In \sref{sec2}, we conclude that the reason why the wave function evolution in \fref{figure1} was wrong is due to the wrong possibility that an electron that is transmitted at time $t_1$ is later reflected at time $t_2$. This unphysical result simply disappears when using Bohmian trajectories: the dynamic of a transmitted electron at time $t_1$ will be determined \emph{locally} by the guidance law \eref{eq-guidance} that only takes into account the transmitted part of the wave function. We can, for all practical purposes, completely ignore the reflected part of the wave function. Therefore, at time $t_2$, this electron will remain as a transmitted electron with full certainty.\\
Finally, let us emphasize that, in principle, the measuring apparatus has also a role in the classical simulation of electronic devices. Such interaction with the apparatus is included at a classical level, at best, by a proper boundary conditions for the scalar potential of the Hamiltonian (i.e. the Poisson equation) ensuring overall charge neutrality. Obviously, this kind of approximation can also be included here.
\subsection{An approximation for the Coulomb and exchange interaction among electrons}
\label{sec4.2}
Once we have \emph{practically} eliminated the apparatus from our computations, a second kind of approximation regards the interactions among the electrons of our device. The active region of the electronic device can contain hundreds of electrons. Also in this case, as we mentioned, the many-particle Schr\"{o}dinger equation can be solved only for very few degrees of freedom. A standard way to proceed consists then on reducing the complexity of the problem by \emph{tracing out} certain degrees of freedom. This process ends up with what is called the \emph{reduced density matrix}. When the reduced density matrix is used, its equation of motion is no longer described by the Schr\"odinger equation but in general by a non-unitary operator. The reduced density matrix is no longer a pure state, but a mixture of states and their evolution is in general irreversible \cite{DiVentra}. Now we discuss how Bohmian mechanics allows us to proceed in a very different way. As it will be seen below, the concept of \emph{conditional wave function} \cite{Goldstein1} provides an original tool to deal with many-body open quantum systems \cite{xavier1,Nino1}. \\
As said, once again the key instrument is the conditional wave function. In order to use the conditional wave function to reduce the degrees of freedom of a system we must know how it evolves in time. It can be demonstrated \cite{xavier1} that the single-particle conditional wave function of particle $1$, $\psi_{1}(x_1, t)$, for a system of $N$ interacting particles, obeys the following wave equation:
\begin{eqnarray}\label{Conditional_eq}
i\hbar\frac{\partial \psi_1(x_1,t)}{\partial t}=\Big\{-\frac{\hbar^2}{2m} \nabla^2_1+U_{1}(x_1,\bar{X}_{N-1}(t),t) \nonumber \\ +G_{1}(x_1,\!\bar{X}_{N\!-\!1}(t),t)\!+\!i J_{1}(x_1,\!\bar{X}_{N-1}(t),t) \Big\} \psi_{1}(x_1,t).
\end{eqnarray}
The explicit expression of the potentials $G_{1}(x_1,\!\bar{X}_{\!N\!-\!1\!}(t),t)$ and $J_{1}(x_1,\bar{X}_{N-1}(t),t)$ that appears in \eref{Conditional_eq} can be found in reference \cite{xavier1}. However, their numerical values are in principle unknown and need some educated guesses.
On the other hand, the total electrostatic potential energy among the $N$ electrons that appears in \eref{TDSE}, has been divided into two parts:
\begin{eqnarray}\label{Potential_energy}
U(x_1,\bar{X}_{N-1}(t),t)&=&U_{1}(x_1,\bar{X}_{N-1}(t),t)+ \nonumber \\
&+&U_{N-1}(\bar{X}_{N-1}(t),t).
\end{eqnarray}
The term $U_{1}(x_1,\bar{X}_{N-1}(t),t)$ can be any type of many-particle potential defined in the position-representation, in particular it can include short-range and long-range Coulomb interactions. The other term $U_{N-1}(\bar{X}_{N-1}(t),t)$ in \eref{Potential_energy} without dependence on $X_1$, is contained in the coupling potential $G_1$ in \eref{Conditional_eq}. The same procedure can be done for the rest of the $N-1$ particles, for example for particle $2$ we fix the positions of particle $1,3,...,N$ obtaining the analogous of \eref{Conditional_eq} for $\psi_2(x_2,t)$. From a practical point of view, all quantum trajectories $\bar{X}_N(t)$ have to be computed simultaneously. In order to gather all the above concepts, let us discuss a practical computation with conditional wave functions by detailing a sequential procedure:
\begin{enumerate}
\item At the initial time $t=0$, we fix the initial position of all $i$-particles, $X_i(0)$, according to the initial probability distribution ($|\Psi(\bar{x}_N,0)|^2$), and their associated single-particle wave function $\psi_{i}(x_i,0)$.
\item From all particle positions, we compute the exact value of the potential $U_{i}(x_i,\bar{X}_{N-1}(0),0)$ for each particle. An approximation for the terms $G_{i}$ and $J_{i}$ is required at this point. We use the simplest one \cite{albareda}.
\item We then solve each single-particle Schr\"{o}dinger-type equation, \eref{Conditional_eq}, from $t=0$ till $t=dt$.
\item From the knowledge of the single-particle wave function $\psi_{i}(x_i, dt)$, we can compute the new velocities $v_{i}(dt)$ for each $\!i$-particle (see \eref{eq-guid-cwf} in \aref{app-bohm-2}).
\item With the previous velocity, we compute the new position of each $i$-particle as $X_{i}(dt)=X_{i}(0)+v_{i}(dt)dt$.
\item Finally, with the set of new positions and wave functions, we repeat the whole procedure (steps 1 till 5) for another infinitesimal time $dt$ till the total simulation time is finished.
\end{enumerate}
The advantage of the above algorithm using \eref{Conditional_eq} instead of the many-particle Schr\"odinger equation (\eref{TDSE}) is that, in order to find approximate trajectories, $X_i(t)$, we do not need to evaluate the wave function and potential energies in the whole configuration space, but only over a smaller number of configuration points, $\{\!x_i,\! \bar{X}_{\!N\!-\!1\!}(t)\!\}$, associated with those trajectories defining the highest probabilities according to $|\psi(\bar{x}_N,t)|^2$ . \\
The exchange interaction is naturally included in \eref{Conditional_eq} through the terms $G_{i}$ and $J_i$. Due to the Pauli exclusion principle, the modulus of the wave function tends to zero, $|\psi(x_i, \bar{X}_{\!N\!-\!1\!}(t),t)|\to 0$, in any neighborhood of $x_{i}$ such that $|x_{i}-\bar{X}_{N-1}(t)|\to 0$. Thus, both terms, $G_{i}(x_i, \bar{X}_{\!N\!-\!1\!}(t),t)$ and $J_{i}(x_i, \bar{X}_{\!N\!-\!1\!}(t),t)$, have asymptotes at ${x_{i}}\to {\bar{X}_{N-1}}(t)$ that \textit{repel} the $i-$ particle from other electrons. However, in order to exactly compute the terms $G_i$ and $J_i$ we must know the total wave function, which is in principle unknown. There are however a few ways to introduce the symmetry of the wave function without dealing directly with these two coupling terms \cite{xavier1,aalarcon09pps,alarconUnpublished}. Clearly, the complexity of the algorithm increases as we go beyond the single-particle quantum transport scenario mentioned in \sref{sec2} and \sref{sec3}.
\subsection{Practical example}
\label{practical}
An electron device is an open system, where many parameters can only be estimated from the knowledge of their statistical (typical) distribution. Apart from the uncertainty in the initial position in the quantum trajectories (the $\alpha$ distribution explained in \aref{app-bohm}), we also take into account the uncertainty on the properties of the injected electrons (initial energies, momentums, etc) which we refer to the parameter $h$. The random process $I^{\gamma}(t)$ mentioned in \sref{sec1.1} is now written as $I^{\alpha,h}(t)$. At finite temperature, the thermal noise introduces fluctuations on the energies of the electrons entering inside the device. As discussed in the introduction of \sref{sec2}, the study of the noise in electrical devices due to the partition noise of the barrier plus the thermal noise of the injection are traditionally known as quantum shot noise \cite{Landauer1,Beenakker1,Beenakker2,Buttiker3,Buttiker2,Landauer3}. This is the noise studied in this \sref{sec4}. In many systems, one obtains the well known Schottky's result \cite{Schottky} or Poissonian shot noise, $S_{II\;shot}(0)=2q\left\langle I\right\rangle$, for the noise power spectral density defined in \eref{noise5} at zero frequency, i.e. $w=0$.\\
We select a particular (large) set of wave packets with values $\alpha$ and $h$ for selecting their initial conditions. We refer to such selection as $\alpha_1$ and $h_1$. We evolve the wave packets and trajectories as explained in previous paragraphs. Within the approximation mentioned in \sref{sec4.1} and \sref{sec4.2}, the total current value can be calculated as the sum of the particle or conduction current plus the displacement current:
\begin{eqnarray}
I^{\alpha,h}(t) & = & I_c^{\alpha,h}(t)+I_d^{\alpha,h}(t)= \nonumber \\
& = & \int_{S} \sum_{i=1}^N q_i v_i(X_i^{\alpha,h}(t))\delta(x_S-X_{i}^{\alpha,h}(t)) \cdot ds + \nonumber \\
&+& \int_{S} \sum_{i=1}^N\epsilon(x_S)\dfrac{dE(x_S;X_i^{\alpha,h}(t),t)}{dt} \cdot ds,
\label{current}
\end{eqnarray}
where $S$ is the surface where we want to calculate the current, $x_S$ are the points of the chosen surface, $\epsilon(x_S)$ is the dielectric constant in the same surface and $E(x_S;X_i^{\alpha,h}(t),t)$ is the electric field in the surface $S$ which depends on the actual position of all the electrons.
Once we know $I^{\alpha_1,h_1}(t)$ for a large interval of time, the algorithm to compute the current fluctuations is quite simple following \eref{noise3} and \eref{noise5}. This discussion can be familiar for those people who works in semi-classical approaches. In fact, the Bohmian procedure explained here for quantum transport is very similar to that of, for instance, the Monte-Carlo simulations of the Boltzmann equation. But instead of being the electric-field the one who \emph{guides} the electrons, it is the wave function, through the guiding velocity field in \eref{eq-guidance}.
\begin{figure}
\centering
\includegraphics[width=0.97\columnwidth]{figure5.pdf}
\caption{Transient current $I_{tran}(t)$ computed analytically and numerically (Color figure online)}
\label{F6_alarcon:Figure6_9a}
\end{figure}
As a practical example of the computation of the fluctuations, we show here the current response to a step input voltage in the Negative Differential Conductance region of a RTD. The input signal is the step voltage $V(t)=V_{1}u(t)+V_{2}\left[ 1-u(t)\right]$ where $u(t)$ is the Heaviside (step) function. The voltages $V_{1}$ and $V_{2}$ are constant. Then the current response can be expressed as $I(t)=I_{tran}(t)+I_{1}u(t)+I_{2}\left[ 1-u(t)\right]$ where $I_{1}$ and $I_{2}$ are the stationary currents corresponding to $V_{1}$ and $V_{2}$ respectively, and $I_{tran}$ is the \textit{intrinsic} transient current.
The results are reported in \fref{F6_alarcon:Figure6_9a} where $I_{tran}(t)$ manifests a delay with respect to the step input voltage, due to the dynamical adjustment of the electric field in the conductors. After this delay, the current response becomes a RLC-like response (dot-dashed line RLC response $2$) i.e. purely exponential. Performing the Fourier transform of $I_{tran}(t)$ in \fref{F6_alarcon:Figure6_9b} and comparing it with the single pole spectra (Fourier transform of the RLC-like responses, dashed and dashed dotted lines), we are able to estimate the cut-off frequency and the frequency offset due to the delay \cite{xavier2}.
\begin{figure}
\centering
\includegraphics[width=0.97\columnwidth]{figure6.pdf}
\caption{Fourier transform of $I_{tran}(t)$ of \fref{F6_alarcon:Figure6_9a}. A logarithmic scale is used to resolve the cut-off frequency offset (Color figure online)}
\label{F6_alarcon:Figure6_9b}
\end{figure}
In order to understand how the many-body Coulomb interaction affects the noise in RTDs, we also investigate the correlation between an electron trapped in the resonant state during a dwell time $\tau_{d}$ and those remaining in the left reservoir.
This correlation occurs essentially because the trapped electron perturbs the potential energy felt by the electrons in the reservoir. In the limit of non-interacting electrons, the Fano factor will be essentially proportional to the partition noise, however, if the dynamical Coulomb correlations are included in the simulations (see \fref{F6_alarcon:Figure6_10}) this result is no longer reached, it shows super-poissonian values. Finally, we are also interested in the high frequency spectrum $S(w)$ given by \eref{noise5} revealing information about the internal energy scales of the RTD that is not available from DC transport (see \fref{F6_alarcon:Figure6_12}).
\begin{figure}
\includegraphics[width=0.97\columnwidth]{figure7.pdf}
\caption{Fano Factor $F$ defined as $F=S(0)/(2q\left\langle I\right\rangle)$, evaluated using the current fluctuations directly available from the BITLLES simulator (Color figure online)}
\label{F6_alarcon:Figure6_10}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.97\columnwidth]{figure8.pdf}
\caption{Current noise power spectrum referred to Poissonian shot noise at different biases (Color figure online)}
\label{F6_alarcon:Figure6_12}
\end{figure}
\section{Conclusions}
\label{sec5}
In the present paper we discuss the understanding and the computation of quantum noise in electrical devices from a Bohmian perspective. Computations of quantum noise are quite complex because noise is generally quantified in terms of temporal correlations. Such correlations must include the time-evolution of a quantum system during and after a measurement. Usually, many other quantum computations do only require a final measurement, so that their time-evolution from the initial until the final time is uniquely determined by the unitary (Schr\"odinger like) evolution. As discussed in \fref{figure2}, this unitary evolution is not enough to compute time correlations which require mixing unitary and non-unitary (the so-called collapse of the wave function) time evolutions.
There are several (empirically equivalent) quantum theories. Each quantum theory has its own formalism that is able to connect the experimental values with some abstract elements such as wave functions, operators, trajectories, etc. that are able to satisfactorily reproduce (or predict) experimental results. We discuss how the Copenhagen (also known as standard or orthodox) interpretation and Bohmian mechanics give explanation to the partition noise. For a flux of electrons impinging upon a tunneling barrier, we analyze how a measurement process affects partition noise in a quantum device. For simplicity, to focus on the importance of the measurement process, we consider spinless electrons without Coulomb and exchange interaction. In \sref{sec2} we explain how standard quantum theory provides an answer for the measurement problem by means of the introduction of the notion of operators. We see that this notion is not always satisfactory even for practical purposes, because the definition of which is the \emph{right} operator is not obvious. Then in \sref{sec3} we discuss an alternative way to deal with the collapse without introducing the idea of operators. In fact within Bohmian mechanics, a theory of wave \emph{and} particles, the collapse is derived trivially by means of the introduction of the \emph{conditional wave function} (the wave function of a subsystem), a tool exclusively belonging to Bohm's theory. Obviously, each theory gives a different formalism to compute quantum noise and different interpretation of its origin. In any case, at the end of the day, the same empirical predictions are achieved by using the orthodox quantum theory or Bohmian mechanics.
In \sref{sec4}, because the Bohmian formulation uses trajectories to compute experimental results, we see that a very reasonable approximation to include collapse can be achieved with a very small computational effort. Finally, details of the simulator named BITLLES based on Bohmian mechanics and numerical results for low and high frequency noise of the current in a resonant tunneling diode are presented. We emphasize that the presented formalism and the procedure for computing the properties of a system (in our case current, noise, etc.) have many similarities with the one used in semi-classical simulations (for example Monte-Carlo of the Boltzmann equation \cite{tomas}). In any case, Bohmian formalism is not at all a semi-classical approach but a complete quantum theory that can be applied to study any non-relativistic quantum phenomena, quantum noise and collapse among them.
Finally, we wish to discuss which is the ultimate origin of the quantum noise according to orthodox and Bohmian interpretations. Before entering into details, let us recall that the definition of noise given in \eref{noise2} in \sref{sec1} is just the difference between the experimental value $I(t)$ and what we define as the signal. Therefore, even a sinusoidal current $I(t)$ provides a value of $\triangle I^2$ different from zero. What we want to discuss hereafter are not all the possible sources of fluctuations in $I(t)$, but only if there is any new type of randomness in quantum devices that is not present in classical ones and what is its origin.
Given this last specification, we can provide an answer to the question: \emph{What is the ultimate origin of quantum noise according to the orthodox interpretation?} As we see in \sref{sec2}, the transmission or reflection of a single electron impinging upon a tunneling barrier becomes unpredictable. This is an example of a new source of randomness present only in quantum devices. According to the orthodox theory, this randomness appears because of the collapse of the wave function due to measurements. Without the collapse (that is put in by hand as an additional postulate in the orthodox theory), the wave function follows a deterministic law dictated by the Schr\"odinger equation. The partition noise in the tunneling barrier discussed in \sref{sec2} is due to the action of the operator which implements the random collapse of the wave function (selecting the final wave function stochastically among the set of available eigenstates).
Alternatively, we can also answer the question: \emph{What is the ultimate origin of quantum noise according to Bohmian mechanics?} The randomness in the values of the current in $I(t)$ provided by the BITLLES simulator comes from the $h$ and $\alpha$ distributions mentioned in \sref{practical}. The $h$ distribution is due to the uncertainty of the initial energies, momentums, entering times, etc. of the electrons. This source of ``extrinsic'' randomness can be minimized imagining technological improvements of the setup (for example, well-controlled single electron sources). On the contrary, the $\alpha$ distribution of the conditional wave function (explained in \aref{app-bohm}) is an unavoidable source of randomness. This randomness of the $\alpha$ distribution cannot be minimized by any technological improvement \cite{wiseman,acin}. Therefore, whether the particle is transmitted or reflected becomes unpredictable in Bohmian theory too. Thus, though Bohmian mechanics is deterministic, an \emph{appearance} of randomness emerges in the subsystems \cite{Goldstein1,Zanghi_chaos}. It is important to notice that the measurement of the system does not introduce any additional randomness. The ultimate origin of the unpredictability is the fact that the uncertainty principle does not allow us to known the (well-defined in the Bohmian theory) initial position of the particles in each experiment.
In summary, according to the orthodox interpretation, the partition noise has its origin in the stochasticity of the orthodox measurement process. On the contrary, Bohmian mechanics says that the origin of noise is the uncertainty of the initial position of the trajectory in each realization of the experiment. Although both theories give the same predictions, in the authors' opinion, the latter has a more natural, common and understandable explanation of the origin of quantum noise. While following deterministic laws, the transmission or reflection of a Bohmian electron is unpredictable in a given experiment. A classical dice is a simpler example of a system following deterministic laws that becomes unpredictable. Collapse in Bohmian mechanics is so naturally derived that the quantum measurement problem, in general, and quantum noise, in particular, are somehow demystified. We underline that Bohmian mechanics achieves the non-unitary evolution of the wave function of a measured system simply slicing the enlarged wave function in the configuration space (without introducing any measurement-associated randomness).
We accept that preferences between the explanation of the origin of quantum noise in terms of the orthodox or Bohmian interpretations are subjective. Therefore, in this paper we have also developed objective arguments about the computational advantages of the Bohmian formalism. The facts that the measuring apparatus, what we call the ammeter, is directly included into the Hamiltonian of the Schr\"odinger equation and that the current values are computed from trajectories (not from the wave functions) allow us to study system plus apparatus scenarios (or look for reasonable approximations). This ability is very relevant, for example, in the computation of high frequency currents where it is difficult to find the \emph{right} operator. For all these reasons, we conclude that quantum noise is easily understood and computed from a Bohmian perspective in many practical scenarios.
\begin{acknowledgements}
We want to acknowledge T. Norsen, N. Zangh\`i, G. Albareda and F.L. Traversa for insightful discussions.This work has been partially supported by the \lq\lq{}Ministerio de Ciencia e Innovaci\'{o}n\rq\rq{} through the Spanish Project TEC2012-31330 and by the Grant agreement no: 604391 of the Flagship initiative \lq\lq{}Graphene-Based Revolutions in ICT and Beyond\rq\rq{}. Z. Z acknowledges financial support from the China Scholarship Council (CSC). D.M. is supported in part by INFN and acknowledges the support of COST action (MP1006) through STSM.
\end{acknowledgements}
\begin{appendices}
\section{The quantum DC current in ergodic systems}
\label{app-noise}
The DC current measured in a laboratory $\langle I \rangle$ can be computed by time-averaging the measured value of the total current $I(t)$ from a \emph{unique device} during a large (ideally infinite) period of time $T$ as mentioned in \eref{noise1}. If we can justify the ergodicity of electronic devices, we can alternatively compute $\langle I \rangle$ from an ensemble-average of all possible values of the current $I_i$ measured, at one particular time $t$, over an \emph{ensemble of (identical) devices} as seen in \eref{noise7}. For DC quantum transport computations, Eq. (\ref{noise7}) is greatly preferred because it deals directly with the probabilistic interpretation of the wave function. It is important to realize that while \eref{noise1} implies measuring the quantum current many times, \eref{noise7} involves only one measurement. We do not need to worry about the evolution of the wave function after the measurement when using \eref{noise7}. Let us discuss this point in more detail. We can define the eigenstates $| \psi_i \rangle$ of a particular operator $I$, as those vectors that satisfy the equation $I |\psi_i \rangle = I_i | \psi_i \rangle $. The eigenvalue $I_i$ is one of the $M$ possible measured values in \eref{noise7}.\footnote{For simplicity we assume that there is no degeneracy. Our qualitative discussion does not change if degeneracy is considered.} Since the entire set of eigenstates form a basis of the Hilbert space, the wave function can be decomposed as $| \psi (t) \rangle = \sum_{i=1}^M c_i(t) |\psi_i \rangle $, with $c_i(t) = \langle \psi_i | \psi(t) \rangle$. Then, we can rediscover \eref{noise7} as follow:
\begin{eqnarray}
\langle I \rangle &=& \langle \psi(t)| I | \psi(t) \rangle = \nonumber \\
&=& \sum_{j=1}^M c^{*}_j(t) \langle \psi_j | \sum_{i=1}^{M} I_ic_i(t) |\psi_i\rangle = \sum_{i=1}^{M} I_i P(I_i),
\end{eqnarray}
where we have used the orthonormal property of the eigenstates $\langle \psi_j | \psi_i \rangle = \delta_{ij}$ and the definition of the (Born) probability $P(I_i) \!=\! |c_i(t)|^2$. We emphasize that $\langle\! \psi(t)| I | \psi(t)\! \rangle$ does not require the explicit knowledge of the eigenstates. Only the free evolution of the state $| \psi(t) \rangle$ and the \emph{measuring} operator $I$ are needed.
At this point, it is mandatory to provide some discussion about the use of the ergodic theorem. Strictly speaking, no ergodic theorem exists for an out of equilibrium system \cite{Price}. Indeed, an out of equilibrium system is represented by a distribution function, or probability function, that is different from that in equilibrium and arises from a balance between the driving forces and the dissipative forces. The applied bias used to measure the DC current of any device implies that the device is quite likely in a far from equilibrium state. Therefore, the ergodic connection between Eq. (\ref{noise1}) and Eq. (\ref{noise7}) has to be considered as only a very reasonable approximation for DC transport, but not as an exact result \cite{Price}.
\section{Bohmian mechanics}
\label{app-bohm}
Bohmian mechanics is a version of quantum theory whose basic elements are waves and point-like particles. The many-particle wave function evolves according to the Schr\"odinger equation (\ref{TDSE}) while particles have definite position at any time with a law given by \eref{eq-guidance}, therefore being a fully deterministic theory. The configuration of the particles, say at time $t=0$, is chosen randomly according to $|\Psi|^2$ at the initial time, known as \emph{quantum equilibrium hypothesis} \cite{Goldstein1}. Thanks to the continuity equation
\begin{eqnarray}
\label{eq-continuity}
\frac{\partial \rho}{\partial t} = - \nabla \left( \rho v \right),
\end{eqnarray}
where $\rho = |\Psi|^2$ and $v$ the Bohmian vector field. An important consequence of the quantum equilibrium hypothesis and equivariance is the empirical equivalence between Bohmian mechanics and orthodox quantum theory for any kind of non-relativistic quantum experiments.
\subsection{The conditional wave function}
\label{app-bohm-2}
Consider a quantum system of $N$ particles and a partition of it in such a way that its spatial coordinates can be split as $\bar{x}_N = \{x_1, \bar{x}_{N-1}\}$. Where we denote with $x_1$ the position in $\mathbb{R}^3$ space of the electron $1$, while with $\bar{x}_{N-1}$ the positions of the rest of the electrons in a $\mathbb{R}^{3 (N-1)}$ space. The actual particle trajectories are accordingly denoted by $\bar{X}_N(t) = \{X_1(t), \bar{X}_{N-1}(t)\}$. \emph{How can one assign a wave function to the electron $1$?} In general this is not possible if the two subsystems are entangled, i.e. the total wave function cannot be written as a product $\Psi(\bar{x}_N) = \psi_1(x_1)\psi_{N-1}(\bar{x}_{N-1})$. However, we can modify our question and ask what is the wave function of the electron $1$ that provides the exact velocity $v_1$ given a particular configuration $\bar{X}_{N-1}(t)$ for the rest of the particles. The answer given by Bohmian mechanics is the so called \emph{conditional wave function} \cite{Goldstein1,Nino1}:
\begin{equation}
\psi_{1}(x_1,t) = \Psi(x_1,\bar{X}_{N-1}(t),t),
\label{Conditional}
\end{equation}
which constitutes a slice of the whole multi-dimensional wave function. The wave function constructed in such a way gives exactly the same Bohmian velocity
\begin{eqnarray}
v_1(t) \!=\! \frac{\hbar}{m_1} \text{Im} \frac{\nabla_1 \Psi}{\Psi}\Big|_{\bar{x}_N\!=\!\bar{X}_N(t)}\! \!\equiv\! \frac{\hbar}{m_1} \text{Im} \frac{\nabla_1 \psi_1}{\psi_1}\Big|_{x_1 \!=\! X_1(t)}.
\label{eq-guid-cwf}
\end{eqnarray}
\subsection{Computation of mean value of an operator}
\label{app-bohm-3}
If needed, Bohmian mechanics can make use of operators, but only as a mathematical trick. Without any physical or fundamental role in the operator. We briefly explain how it is possible to calculate the mean value of a general hermitian operator with Bohmian trajectories. The \emph{quantum equilibrium hypothesis} at the initial time $t=0$ can be expressed in terms of the trajectories as follows
\begin{eqnarray}
\label{QEH}
|\Psi(\bar{x}_N,0)|^2 = \lim_{M_{\alpha} \to \infty} \frac{1}{M_{\alpha}} \sum_{\alpha=1}^{M_\alpha} \prod_{i=1}^{N} \delta(x_i-X_i^{\alpha}(0)),
\end{eqnarray}
where the superindex $\alpha$ takes into account the uncertainty in the initial position of the particles. It can be easily demonstrated \cite{OriolsBook} that the evolution of the above infinite set of quantum trajectories $\alpha=1,2,...,M_{\alpha}$ reproduce at any time $t$ the probability distribution, $|\Psi(\bar{x}_N,t)|^2$. \\
For computing the mean value of an operator $A$ it can be demonstrated \cite{OriolsBook} that
\begin{eqnarray}
\langle A \rangle_{\Psi} = \lim_{M_{\alpha} \to \infty} \frac{1}{M_{\alpha}} \sum_{\alpha=1}^{M_\alpha} A_B(\bar{X}_N^{\alpha}(t)),
\label{mean-value}
\end{eqnarray}
where $A_B(\bar{x}_N)$ is the ``local" mean value of $A$.
\end{appendices}
|
\section{Introduction}
Let $q$ be a non-zero complex number which is not a root of unity.
Let $\Glie := \mathfrak{gl}(M,N)$ be the {\it general linear Lie superalgebra}. Let $U_q(\Gaff)$ be the associated quantum affine superalgebra. (We refer to \S \ref{sec: quantum affine superalgebra} for the precise definition.) This is a Hopf superalgebra neither commutative nor co-commutative, and it can be seen as a deformation of the universal enveloping algebra of the following affine Lie superalgebra:
\begin{align*}
& L \Glie := \Glie \otimes \BC[t,t^{-1}] = \bigoplus_{1 \leq i,j \leq M+N} E_{ij} \otimes \BC[t,t^{-1}].
\end{align*}
Here the $E_{ij}$ for $1 \leq i,j \leq M+N$ are the elementary matrices in $\Glie$.
In this paper, we are concerned with a distinguished family of finite-dimensional $U_q(\Gaff)$-modules, the so-called {\it Kirillov-Reshetikhin modules}. We would like to construct inductive systems of these modules, and realize their inductive limits as modules over the upper Borel subalgebra of $U_q(\Gaff)$, which is in our context the $q$-Yangian $Y_q(\Glie)$. These limits may or may not carry an action of the full superalgebra $U_q(\Gaff)$, and they may also be finite-dimensional.
\subsection{Background.}
Our motivation for studying representations of quantum affine superalgebras comes, on the one hand, from the integrability structure of AdS/CFT correspondence where various quantum superalgebras related to $\mathfrak{sl}(2,2)$ show up (see for example \cite{Beisert3}), and on the other hand, from the generalization of Hernandez-Jimbo's construction of asymptotic representations related to Baxter's $\BQ$-operators to the super case. In is paper, we concentrate on the second point.
In the early seventies, towards the study of the eight-vertex model, Baxter \cite{Baxter} introduced the concept of $\BQ$-operators and $\textbf{T-Q}$ functional relations to determine the eigenvalues of transfer matrices. Ever since, various progress has been made towards understanding and generalizing Baxter's $\BQ$-operators and $\textbf{T-Q}$ relations, notably the approach by using representation theory of quantum affine (super)algebras.
In a series of papers \cite{BLZ1,BLZ2,BLZ}, Bazhanov-Lukyanov-Zamolodchikov (BLZ for short) generalized $\textbf{T},\BQ$-operators from lattice models to integrable quantum field theory. The idea goes roughly as follows. Firstly, one gets a representation $W$ of the lower Borel subalgebra $\mathfrak{B}_-$ of the quantum affine algebra $U_q(\widehat{\mathfrak{sl}_2})$, which is either a finite-dimensional evaluation module for the lattice model, or an infinite-dimensional vertex representation for the quantum field theory. $W$ is referred to as a {\it quantum space}. Secondly, one constructs an $\textbf{L}$-operator, which is an element of the completed tensor product $\mathfrak{B}_+ \otimes \End W$. (In both integrable structures, $\textbf{L}$ is presumed to be the universal $R$-matrix $\mathcal{R} \in \mathfrak{B}_+ \otimes \mathfrak{B}_-$ of $U_q(\widehat{\mathfrak{sl}_2})$ with $\mathfrak{B}_-$ specialized to $\End W$. See \cite[Conjecture]{BLZ} for the statement and \cite[\S 3.3]{BHK} for a proof.) Thirdly, $\textbf{T},\BQ$-operators, as elements of $\End W$, are defined as twisted traces of $\textbf{L}$ over various representations of $\mathfrak{B}_+$: finite-dimensional evaluation representations for $\textbf{T}$ and {\it oscillator representations} for $\BQ$. Baxter's $\textbf{T-Q}$ relations are then deduced from tensor product decompositions of representations of $\mathfrak{B}_+$.
The oscillator representations are constructed quite explicitly in \cite{BLZ} by adapting the so-called oscillator realization of $\mathfrak{B}_+$. There are several generalizations on BLZ's construction of oscillator representations and $\textbf{T-Q}$ relations when replacing $U_q(\widehat{\mathfrak{sl}_2})$ by higher rank quantum affine (super)algebras. In the super case, this has been partly done: by Bazhanov-Tsuboi \cite{BT} for the quantum affine superalgebra $U_q(\widehat{\mathfrak{sl}(2,1)})$, and later generalized by Tsuboi \cite{Tsuboi} for the quantum affine superalgebra $U_q(\widehat{\mathfrak{gl}(M,N)})$; by Kulish-Zeitlin \cite{KZ} and Ip-Zeitlin \cite{IZ} for the (twisted) quantum affine superalgebra $C_q^{(2)}(2)$. In both approaches, evaluation morphisms, oscillator realizations of Borel subalgebras and universal $R$-matrix remain the ingredients indispensable.
Recently, it was expected by Bazhanov-Lukyanov \cite{BL} that the $\textbf{T-Q}$ operators corresponding to the exceptional quantum affine superalgebra $U_q(\widehat{D}(2,1;\alpha))$ should be related to the connection coefficients and the Wilson loop for various (perturbed) hypergeometric equations and there is need to construct both finite-dimensional ({\bf T}-operator) and oscillator ({\bf Q}-operator) representations for $U_q(\widehat{D}(2,1;\alpha))$. However, evaluation morphisms from this quantum affine superalgebra to the finite type quantum superalgebra are not available, and to construct finite-dimensional representations is already an interesting problem. Nevertheless, Drinfeld realizations corresponding to various Dynkin diagrams for this quantum affine superalgebra are known \cite{HSTY}.
We remark also that in the super case, except for some small rank quantum affine superalgebras such as $C_q^{(2)}(2)$ \cite{IZ} and $U_q(\widehat{\mathfrak{gl}(2,2)})$ \cite{Gade}, there is still no explicit formula of Damiani type \cite{Damiani} for the universal $R$-matrix of quantum affine superalgebras.
\subsection{Asymptotic representations.}
Recently, Hernandez-Jimbo \cite{HJ} constructed the analogue of oscillator representations for an arbitrary (non-twisted) quantum affine algebra $U_q(\widehat{\Glie''})$. Their construction is based on a well-studied family of finite-dimensional modules, the so-called Kirillov-Reshetikhin modules. They first constructed inductive systems of Kirillov-Reshetikhin modules, and then endowed their inductive limits with actions of the upper Borel subalgebra in an asymptotic way, resulting in oscillator representations for $U_q(\widehat{\Glie''})$. The asymptotic construction eventually enables Frenkel-Hernandez \cite{FH} to interpret generalized $\textbf{T-Q}$ relations in terms of representations and to prove a conjecture of Frenkel-Reshetikhin on the spectra of quantum integrable systems \cite{FR}. We refer to \cite{Hernandez2} for a general review.
The advantage of Hernandez-Jimbo's construction is that the asymptotic modules have simpler representation structures compared to finite-dimensional modules and they give rise to generalized \textbf{T-Q} relation in \cite{FH} whose proof does not need universal $R$-matrix. (The complete proof of Frenkel-Reshetikhin conjecture however does.)
In the present paper, we would like to apply Hernandez-Jimbo's asymptotic construction to the quantum affine superalgebra $U_q(\Gaff)$. As we shall see, their inductive systems of Kirillov-Reshetikhin modules give rise to the oscillator representations of the upper Borel subalgebra $Y_q(\Glie)$ in \cite{BT,Tsuboi}. We also propose a new application of these inductive systems by realizing their inductive limits as one-parameter families of representations of the full quantum affine superalgebra.
\subsection{Main results.}
To state in a neat way the resulting asymptotic representations in this paper, let us introduce several notations (slightly different from the main text).
Let $I_0 := \{1,2,\cdots,M+N-1\}$ be the set of Dynkin vertices for the Lie superalgebra $\Glie$. For $r$ a Dynkin vertex and $f(z)$ a rational function with $f(0) \in \BC^{\times}$, there is a simple $Y_q(\Glie)$-module, denoted by $S_r(f)$, which is generated by a highest $\ell$-weight vector $v$ satisfying $\phi_r(z) v = f(z) v$ and $\phi_i(z) v = v$ whenever $i \neq r$. Here the $\phi_i(z)$, as formal power series in $Y_q(\Glie)$, are quantum affine analogues of diagonal matrices in $\Glie$.
Notably, for $r$ a Dynkin vertex, $a$ a non-zero complex number (called spectral parameter), and $k$ a positive integer (called lever of representation), the simple module $S_r(q_r^k \frac{1-za}{1-zaq_r^{2k}})$ is called a Kirillov-Reshetikhin module. It is a finite-dimensional evaluation module, and its $Y_q(\Glie)$-module structure can be extended to that of $U_q(\Gaff)$-module.
\subsubsection{Inductive systems of Kirillov-Reshetikhin modules.} From now on, fix the Dynkin vertex $r$ and the spectral parameter $a$. Our first main task in this paper is to construct an inductive system of the Kirillov-Reshetikhin modules $S_r(q_r^k\frac{1-za}{1-zaq_r^{2k}})$ with respect to the level of representation (see \S \ref{sec: positive asymptotic modules}-\ref{sec: odd asymptotic modules})
\begin{equation} \label{introd: inductive system of KR modules}
F_{k,l}: S_r(q_r^l \frac{1-za}{1-zaq_r^{2l}}) \longrightarrow S_r(q_r^k \frac{1-za}{1-zaq_r^{2k}}) \quad \textrm{for}\ l < k.
\end{equation}
The idea of construction is the same as that of Hernandez-Jimbo \cite{HJ}: by using the cyclicity property of some particular tensor products of Kirillov-Reshetikhin modules. In our case, the following guarantees the existence of inductive systems (Theorem \ref{thm: cyclicity of KR modules}):
\begin{itemize}
\item[(A)] Assume $l_1 < l_2 < l_3$. The tensor product $S_r(\frac{1-zaq_r^{2l_2}}{1-zaq_r^{2l_3}}) \otimes S_r(\frac{1-zaq_r^{2l_1}}{1-zaq_r^{2l_2}}) \otimes S_r(\frac{1-za}{1-zaq_r^{2l_1}})$, as a $Y_q(\Glie)$-module, is of highest $\ell$-weight.
\end{itemize}
The proof relies on a more general cyclicity result in our previous paper \cite{Z2} and on some duality argument. (See \S \ref{sec: KR} for details.)
One special feature of the above inductive system is that the structural maps $F_{k,l}$ do not respect the $U_q(\Gaff)$-module structures. That is to say, given a generator $x$ of the algebra, the two maps $F_{k,l} x$ and $x F_{k,l}$ are in general different. Nevertheless, one can establish stability and asymptotic properties of the $F_{k,l}$, which enable us to express asymptotically these maps $x F_{k,l}$, with $l,x$ fixed and with $k$ varying; they turn out to be Laurent polynomials in $q^k$ of a particular form (see Propositions \ref{prop: first properties of inductive system}-\ref{prop: asymptotic properties of inductive system} for precise statements).
In the non-graded case, to prove these properties, Hernandez-Jimbo used some deep theory of $q$-characters of tensor products \cite[Proposition 3.2]{Hernandez}, which is by no means available to us as we do not even have the notion of $q$-character for representations of quantum affine superalgebras. However, when working with the RTT realization of $U_q(\Gaff)$, we are able to prove these two properties in a straightforward manner.
\subsubsection{Asymptotic construction of Hernandez-Jimbo.} \label{par: motivation 1}
The argument of Hernandez-Jimbo, which provides the inductive limit $V_{\infty}$ of the inductive system \eqref{introd: inductive system of KR modules} with a module structure of the upper Borel subalgebra, can be adapted to our situation without difficulty. Take the highest $\ell$-weight vectors $v_k$ in $S_r(q_r^k \frac{1-za}{1-zaq_r^{2k}})$ as an example. They give rise to the same vector $v_{\infty} \in V_{\infty}$. To get $\phi_r(z) v_{\infty}$, note first of all $\phi_r(z) v_k = q_r^{k} \frac{1-za}{1-zaq_r^{2k}} v_k$. By forgetting the term $q_r^k$, and then by taking the limit $\lim_{k\rightarrow\infty}q_r^k = 0$, we obtain $\phi_r(z) v_{\infty} = (1-za) v_{\infty}$. The stability and asymptotic properties of $F_{k,l}$ explains in a conceptual way the validity of this argument for all vectors in $V_{\infty}$. Eventually, we get a representation $(\rho,V_{\infty})$ of $Y_q(\Glie)$ having $S_r(1-za)$ as a simple sub-quotient, which is the desired oscillator representation of $Y_q(\Glie)$ within the framework of Bazhanov-Tsuboi \cite{BT,Tsuboi}.
There are formal series $\phi_i^-(z)$ in $U_q(\Gaff)[[z^{-1}]]$ such that $\phi_r^-(z) v_k = q_r^{-k} \frac{1-z^{-1}a^{-1}}{1-z^{-1}a^{-1}q_r^{-2k}}$, giving rise to the expression $\lim_{k\rightarrow \infty} q_r^{-2k}$, which is nonsense. For this reason, $S_r(1-za)$ does not carry an action of $U_q(\Gaff)$.
\subsubsection{Generic asymptotic construction.} \label{par: motivation 2}
Starting from the same inductive system \eqref{introd: inductive system of KR modules} and from the same stability and asymptotic properties, we propose in this paper a new limit process which endows the inductive limit $V_{\infty}$ with $U_q(\Gaff)$-module structures. Again take the highest $\ell$-weight vectors $v_k$ as an example. We have the asymptotic expressions $\phi_r(z) v_k = q_r^k \frac{1-za}{1-zaq_r^{2k}} v_k$. Now, fix a non-zero complex number $b$. By taking the limit $\lim_{k\rightarrow \infty} q_r^k = b$, we obtain $\phi_r(z) v_{\infty} = b \frac{1-za}{1-zab^2} v_{\infty}$. Note that the expression $\lim_{k \rightarrow \infty} q_r^{-k} = b^{-1}$ makes perfect sense. Again, thanks to the stability and asymptotic properties, we get a representation $(\rho^b,V_{\infty})$ of the full quantum affine superalgebra $U_q(\Gaff)$ on the inductive limit. In particular, it has a simple sub-quotient $S_r(b\frac{1-za}{1-zab^2})$.
Informally, one can think of $(\rho^+, V_{\infty})$ as $(\rho^0,V_{\infty})$.
\subsubsection{Category $\BGG$ and $q$-character for representations of $Y_q(\Glie)$.} Now, following \cite{HJ}, we introduce a category of representations of $Y_q(\Glie)$ including all the Kirillov-Reshetikhin modules, $(\rho^+,V_{\infty})$ and $(\rho^b,V_{\infty})$ constructed above and we define the notions of $q$-character $\chi_q$ and normalized $q$-character $\widetilde{\chi}_q$. Let us put together the main results obtained in this paper on category $\BGG$ and (normalized) $q$-character.
\begin{itemize}
\item[(B)] As in the non-graded case \cite{He,HJ,Mukhin}, there is a classification of simple $Y_q(\Glie)$-modules, simple $U_q(\Gaff)$-modules, and finite-dimensional simple $Y_q(\Glie)$-modules in category $\BGG$ in terms of rational functions (Lemma \ref{lem: simple modules in category O}). Contrary to the non-graded case, there are finite-dimensional simple $Y_q(\Glie)$-modules which cannot be $U_q(\Gaff)$-modules.
\item[(C)] In the case $\Glie = \mathfrak{gl}(1,1)$, the explicit formula $\widetilde{\chi}_q(S_1(f))$ for all rational function $f$ such that $f(0) = 1$ is deduced. See Equation \eqref{equ: q character simple module gl(1,1)}.
\item[(D)] If $S$ is an evaluation module of a polynomial representation, or of the dual of a polynomial representation of $\Uc_q(\Glie)$, then $\widetilde{\chi}_q(S)$ is multiplicity-free, and it is a polynomial in the $[A_{i,x}]^{-1}$ where $i \in I_0$ and $x \in \BC^{\times}$ (Corollaries \ref{cor: normalized q-character for polynomial representations},\ref{cor: normalized q character for dual polynomial representations}). Here as in the non-graded case, the $A_{i,x}$ are the generalized simple roots, and $\Uc_q(\Glie)$ is the finite-type quantum superalgebra.
\item[(E)] Fix a Dynkin vertex $r \in I_0$ and a spectral parameter $a$. Then for all $b \in \BC^{\times}$
\begin{displaymath}
\widetilde{\chi}_q (\rho^+,V_{\infty}) = \widetilde{\chi}_q (\rho^b, V_{\infty}) = \lim_{k \rightarrow \infty} \widetilde{\chi}_q (S_r(q_r^k \frac{1-za}{1-zaq_r^{2k}})).
\end{displaymath}
as formal power series in the $[A_{i,x}]^{-1}$ with coefficients $0$ or $1$. $(\rho^+,V_{\infty})$ is a simple $Y_q(\Glie)$-module, and $(\rho^+,V_{\infty})$ is simple provided some generic condition on $b$. (Corollaries \ref{cor: simplicity of even asymptotic modules},\ref{cor: simplicity of odd asymptotic modules})
\item[(F)] Assume that $b$ is generic. Then related to the following chains of subalgebras:
\begin{eqnarray*}
&\ & Y_q(\mathfrak{gl}(1,0)) \subset Y_q(\mathfrak{gl}(2,0)) \subset \cdots \subset Y_q(\mathfrak{gl}(M-1,0)) \subset Y_q(\mathfrak{gl}(M,0)) \\
&\subset & Y_q(\mathfrak{gl}(M,1)) \subset Y_q(\mathfrak{gl}(M,2)) \subset \cdots \subset Y_q(\mathfrak{gl}(M,N-1)) \subset Y_q(\mathfrak{gl}(M,N)) = Y_q(\Glie)
\end{eqnarray*}
the simple modules in (D) and (E) admit Gelfand-Tsetlin bases (Propositions \ref{prop: q-character and GT basis},\ref{prop: q-character and dual GT basis}, Theorem \ref{thm: main result GT basis of all asymtotic modules}). We refer to \S \ref{sec: definition of GT} for the precise meaning of a Gelfand-Tsetlin basis. Notable properties are semi-simplicity and multiplicity-free when we consider the restrictions of these $Y_q(\Glie)$-modules to the above subalgebras.
\end{itemize}
The proof of (D)-(F) is based on the compatibility of the structural maps $F_{k,l}$ and $q$-characters (Lemmas \ref{lem: GT bases vs structural maps},\ref{lem: dual GT and q character}), and on the explicit computation of $\widetilde{\chi}_q(S)$ when $S$ are the simple modules in (D), by using Gelfand-Tsetlin bases. When $S$ is an evaluation module of a polynomial representation, this is a straightforward generalization of a similar computation carried out by Frenkel-Mukhin \cite[Lemma 4.7]{FM} in the non-graded case, and a closed formula for $\widetilde{\chi}_q(S)$ in terms of Young tableaux is obtained in Equation \eqref{equ: q-character for evaluation modules}. When $S$ is an evaluation module of the dual of a polynomial representation, we do not have similar closed formula for $\widetilde{\chi}_q (S)$.
The asymptotic constructions is quite Lie theoretic. It should eventually be done for more general quantum affine superalgebras like $U_q(\widehat{D}(2,1;\alpha))$, once we know how to define analogues of Kirillov-Reshetikhin modules, their character formulas and some cyclicity result of particular tensor products. (This is the reason why we are restricted to $U_q(\Gaff)$ in the present paper.) Moreover, such constructions should work for non-quantum algebras: semi-simple Lie algebras, current algebras and affine algebras, as we have good candidates for Kirillov-Reshetikhin modules such as Demazure modules and Weyl modules. We hope to return to these issues in future works.
\subsection{Outline.} This paper is organized as follows. In \S \ref{sec: 1}, we recall the definitions concerning quantum (affine) superalgebras and Yangians. Then we study the Gelfand-Tsetlin bases for certain finite-dimensional simple representations for the quantum superalgebra $\Uc_q(\Glie)$. \S \ref{sec: KR} proves a cyclicity result of tensor products of Kirillov-Reshetikhin modules to be needed in the construction of inductive systems of Kirillov-Reshetikhin modules. In \S \ref{sec: asymptotic representations}-\ref{sec: generic asymptotic representations} we carry out in detail the asymptotic constructions for the quantum affine superalgebra $U_q(\Gaff)$, which are illustrated with explicit examples in \S\ref{sec: examples}. In \S \ref{sec: O and q} we introduce category $\BGG$ and $q$-character, discuss their general properties, and study in detail the case $\Glie = \mathfrak{gl}(1,1)$ in \S \ref{sec: example gl(1,1)}. In \S \ref{sec: GT bases} we compute the normalized $q$-character for some evaluation modules. Then in \S \ref{sec: GT for asymptotic modules} we compute normalized $q$-character of asymptotic modules and establish Gelfand-Tsetlin bases for them.
About notations, in the introduction, for the sake of simplicity and for unifying the cases $r \leq M$ and $r > M$, we have used the nonstandard notation $S_r(f)$ for the Kirillov-Reshetikhin modules. In the main text, we adopt the more classical notation $W_{k,a}^{(r)}$. Relevant results and proofs are usually divided into two parts depending on $r$.
\subsection*{Acknowledgments.} The author is grateful to his supervisor David Hernandez and to Vyjayanthi Chari for their interest in the present work and for valuable discussions. Part of it was done while he was visiting Centre de Recherches Math\'{e}matiques in Montr\'{e}al. He would like to thank Masaki Kashiwara for remarks on cyclicity of tensor products, and thank Eugene Mukhin and Weiqiang Wang for remarks on Gelfand-Tsetlin basis and $q$-character.
\section{Preliminaries} \label{sec: 1}
In this section we first collect basic facts about the RTT realizations of the quantum affine superalgebra $U_q(\Gaff)$, the $q$-Yangian $Y_q(\Glie)$, and the quantum superalgebra $\Uc_q(\Glie)$. Next, we review part of the Schur-Weyl duality theory for tensor powers of the natural representation of $\Uc_q(\Glie)$, following Benkart-Kang-Kashiwara \cite{BKK}. Then we show the existence of Gelfand-Tsetlin basis for certain simple representations of $\Uc_q(\Glie)$: polynomial representations and their duals.
\subsection{Conventions.} \label{sec: convention}
Throughout this paper, all the vector superspaces and superalgebras are defined over the base field $\BC$. We fix $q \in \BC$ to be non-zero and not a root of unity. Fix $M,N \in \BZ_{>0}$. Let $I := \{1,2,\cdots,M+N\}$. Define the following maps:
\begin{displaymath}
|\cdot|: I \longrightarrow \super, i \mapsto |i| =: \begin{cases}
\even & (i\leq M), \\
\odd & (i > M);
\end{cases} \quad d_{\cdot}: I \longrightarrow \BZ, i \mapsto d_i := \begin{cases}
1 & (i \leq M),\\
-1 & (i > M).
\end{cases}
\end{displaymath}
Set $q_i := q^{d_i}$. Set $\BP := \oplus_{i \in I} \BZ \epsilon_i$. Let $(,): \BP \times \BP \longrightarrow \BZ$ be the bilinear form defined by $(\epsilon_i,\epsilon_j) = \delta_{ij}d_i$. Let $|\cdot|: \BP \longrightarrow \super$ be the morphism of abelian groups such that $|\epsilon_i| = |i|$. Set $I_0 := I \setminus \{M+N\}$. For $i \in I_0$, let $\alpha_i := \epsilon_i - \epsilon_{i+1}$. Introduce the root lattice $\BQ = \bigoplus_{i\in I_0} \BZ \alpha_i \subset \BP$. Let $\BQ_{\geq 0} := \bigoplus_{i \in I_0} \BZ_{\geq 0} \alpha_i$.
In the following, {\it only} three cases of $|x| \in \super$ are admitted: $x \in I$; $x \in \BP$; $x$ is a $\super$-homogeneous vector of a vector superspace. If a vector space $V = \bigoplus_{\alpha \in P} (V)_{\alpha}$ carries a grading by an abelian group $P$, then we write $|x|_{P} = \alpha$ for $\alpha \in P$ and $x \in (V)_{\alpha}$.
Let $\BV := \bigoplus_{i \in I} \BC v_i$ be the vector superspace with $\super$-grading $|v_i| = |i|$. Then $\End \BV$ is naturally a superalgebra. Let $E_{ij} \in \End \BV$ be the linear endomorphism $v_k \mapsto \delta_{jk} v_i$. In particular, $|E_{ij}| = |i| + |j|$. Let $\Glie$ be the {\it general linear Lie superalgebra} associated with $\BV$, which is, the vector superspace $\End \BV$ endowed with the Lie bracket:
\begin{displaymath}
[f,g] := f g - (-1)^{|f||g|} g f
\end{displaymath}
for $f,g \in \End \BV$ homogeneous. In the following, $\End \BV$ will always be viewed as a superalgebra, and $\Glie$ as a Lie superalgebra, although they are the same as vector superspaces.
\subsection{Quantum superalgebras.} In this subsection, we review the RTT realizations of the quantum affine superalgebra $U_q(\Gaff)$, the $q$-Yangian $Y_q(\Glie)$, and the finite type quantum superalgebra $\Uc_q(\Glie)$, following \cite{Z2}.
\subsubsection{The Perk-Schultz $R$-matrix.} \label{sec: Perk-Schultz matrix}
This is the element $R(z,w) \in (\End \BV \otimes \End \BV)[z,w]$:
\begin{eqnarray} \label{for: Perk-Schultz matrix coefficients}
\begin{array}{rcl}
R(z,w) &:=& \sum\limits_{i\in I}(zq_i - wq_i^{-1}) E_{ii} \otimes E_{ii} + (z-w) \sum\limits_{i \neq j} E_{ii} \otimes E_{jj} \\
&\ & + z \sum\limits_{i<j} (q_i-q_i^{-1}) E_{ji} \otimes E_{ij} + w \sum\limits_{i<j}(q_j-q_j^{-1}) E_{ij} \otimes E_{ji}.
\end{array}
\end{eqnarray}
$R(z,w)$ enjoys many fundamental properties (see \cite[Proposition 3.6]{Z2}), notably the Yang-Baxter equation, and the following {\it ice rule}: for $a,b,c,d \in I$,
\begin{equation} \label{rel: ice rule}
R_{ab,cd}(z,w) \neq 0 \Longrightarrow \epsilon_a + \epsilon_b = \epsilon_c + \epsilon_d \in \BP
\end{equation}
Here the $R_{ab,cd}(z,w) \in \BC[z,w]$ are matrix coefficients defined by:
\begin{displaymath}
R(z,w) (v_c \otimes v_d) = \sum_{a,b \in I} R_{ab,cd}(z,w) (v_a \otimes v_b) \in \BV^{\otimes 2}[z,w].
\end{displaymath}
Let us introduce $R := R(1,0)$ and $R' := -R(0,1)$. Then $R(z,w) = z R - w R'$.
\subsubsection{Quantum affine superalgebra.}\label{sec: quantum affine superalgebra}
This is the superalgebra $U_q(\Gaff)$ defined by
\begin{itemize}
\item[(R1)] generators $s_{ij}^{(n)}, t_{ij}^{(n)}$ for $i,j \in I$ and $n \in \BZ_{\geq 0}$;
\item[(R2)] $\super$-grading $|s_{ij}^{(n)}| = |t_{ij}^{(n)}| = |i| + |j|$;
\item[(R3)] RTT-relations \cite{FRT2}
\begin{eqnarray}
&& R_{23}(z,w) T_{12}(z) T_{13}(w) = T_{13}(w) T_{12}(z) R_{23}(z,w), \label{rel: RTT = TTR} \\
&& R_{23}(z,w) S_{12}(z) S_{13}(w) = S_{13}(w) S_{12}(z) R_{23}(z,w), \label{rel: RSS = SSR} \\
&& R_{23}(z,w) T_{12}(z) S_{13}(w) = S_{13}(w) T_{12}(z) R_{23}(z,w), \label{rel: RTS = STR} \\
&& t_{ij}^{(0)} = s_{ji}^{(0)} = 0 \quad \mathrm{for}\ 1 \leq i < j \leq M+N, \label{rel: FRTS zero condition} \\
&& t_{ii}^{(0)} s_{ii}^{(0)} = 1 = s_{ii}^{(0)} t_{ii}^{(0)} \quad \mathrm{for}\ i \in I. \label{rel: FRTS invertibility condition}
\end{eqnarray}
\end{itemize}
Here $T(z) = \sum_{i,j \in I} t_{ij}(z) \otimes E_{ij} \in (U_q(\Gaff) \otimes \End \BV)[[z^{-1}]]$ and $t_{ij}(z) = \sum_{n \in \BZ_{\geq 0}} t_{ij}^{(n)} z^{-n} \in U_q(\Gaff)[[z^{-1}]]$ (similar convention for $S(z)$ with the $z^{-n}$ replaced by the $z^{n}$).
Remark that \eqref{rel: RTT = TTR}-\eqref{rel: RTS = STR} are operator equations in $(U_q(\Gaff) \otimes \End \BV^{\otimes 2})[[z,z^{-1},w,w^{-1}]]$. Let us express for example Equation \eqref{rel: RSS = SSR} in terms of matrix coefficients:
\begin{eqnarray} \label{equ: definition of q-yangian}
\begin{array}{rcl}
&\ & \sum_{a,b \in I} (-1)^{|a|(|b|+|j|)} R_{kl,ab}(z,w) s_{ai}(z) s_{bj}(w) \\
&=& \sum_{c,d \in I} (-1)^{|c|(|d|+|l|)} R_{cd,ij}(z,w) s_{ld}(w)s_{kc}(z) \quad \textrm{for}\ i,j,k,l \in I.
\end{array}
\end{eqnarray}
$U_q(\Gaff)$ is a Hopf superalgebra with coproduct $\Delta: U_q(\Gaff) \longrightarrow U_q(\Gaff)^{\otimes 2}$
\begin{eqnarray}
&&\Delta (s_{ij}^{(n)}) = \sum_{a=0}^n \sum_{k \in I} (-1)^{(|i|+|k|)(|k|+|j|)} s_{ik}^{(a)} \otimes s_{kj}^{(n-a)}, \label{for: coproduct for quantum affine superalgebra S} \\
&&\Delta (t_{ij}^{(n)}) = \sum_{a=0}^n \sum_{k \in I} (-1)^{(|i|+|k|)(|k|+|j|)} t_{ik}^{(a)} \otimes t_{kj}^{(n-a)}, \label{for: coproduct for quantum affine superalgebra T}
\end{eqnarray}
and antipode $\Sm: U_q(\Gaff) \longrightarrow U_q(\Gaff)$
\begin{eqnarray}
&&\sum_{i,j \in I} \Sm(s_{ij}(z)) \otimes E_{ij} = (\sum_{i,j \in I} s_{ij}(z) \otimes E_{ij})^{-1} \in (U_q(\Gaff)\otimes \End \BV)[[z]], \label{for: antipode for S} \\
&&\sum_{i,j \in I} \Sm(t_{ij}(z)) \otimes E_{ij} = (\sum_{i,j \in I} t_{ij}(z) \otimes E_{ij})^{-1} \in (U_q(\Gaff) \otimes \End \BV)[[z^{-1}]]. \label{for: antipode for T}
\end{eqnarray}
Here the RHS of the above formulas are well defined thanks to Relation \eqref{rel: FRTS zero condition}.
\subsubsection{$q$-Yangian.} \label{sec: Yangian}
The subalgebra of $U_q(\Gaff)$ generated by the $s_{ij}^{(n)}, (s_{ii}^{(0)})^{-1}$ is called the $q$-Yangian, denoted by $Y_q(\Glie)$. As shown in \cite[Proposition 3.10]{Z2}, these generators together with $\super$-grading (R2) in \S \ref{sec: quantum affine superalgebra}, Relation \eqref{rel: FRTS invertibility condition} with the $t_{ii}^{(0)}$ replaced by the $(s_{ii}^{(0)})^{-1}$, Relation \eqref{rel: FRTS zero condition} without the $t_{ij}^{(0)}$, and Relation \eqref{rel: RSS = SSR} (or equivalently Relation \eqref{equ: definition of q-yangian}) give a full presentation for the superalgebra $Y_q(\Glie)$.
According to Formulas \eqref{for: coproduct for quantum affine superalgebra S} and \eqref{for: antipode for S}, $Y_q(\Glie)$ is a sub-Hopf-superalgebra of $U_q(\Gaff)$.
\subsubsection{Quantum superalgebra.} It is the superalgebra $\Uc_q(\Glie)$ generated by $s_{ij},t_{ji}$ for $1 \leq i \leq j \leq M+N$, with $\super$-degrees
\begin{displaymath}
|s_{ij}| = |t_{ji}| = |i|+|j|
\end{displaymath}
and with RTT relations (\cite{FRT})
\begin{eqnarray*}
&& R_{23}T_{12}T_{13} = T_{13}T_{12}R_{23},\quad R_{23}S_{12}S_{13} = S_{13}S_{12}R_{23} \\
&& R_{23}T_{12}S_{13} = S_{13}T_{12}R_{23},\quad s_{ii} t_{ii} = 1 = t_{ii}s_{ii}.
\end{eqnarray*}
Here, as usual, $T = \sum_{i\leq j} t_{ji} \otimes E_{ji},\ S = \sum_{i\leq j}s_{ij} \otimes E_{ij} \in \Uc_q(\Glie) \otimes \End \BV$. $\Uc_q(\Glie)$ is endowed with a Hopf superalgebra structure with similar coproduct as in Formulas \eqref{for: coproduct for quantum affine superalgebra S}-\eqref{for: coproduct for quantum affine superalgebra T}.
The relationship between $U_q(\Gaff)$ and $\Uc_q(\Glie)$ is explained as follows.
\begin{prop} \label{prop: evaluation morphism}
(1) The assignment $s_{ij} \mapsto s_{ij}^{(0)},\ t_{ji} \mapsto t_{ji}^{(0)}$ extends uniquely to a Hopf superalgebra morphism $\iota: \Uc_q(\Glie) \longrightarrow U_q(\Gaff)$.
(2) The assignment $s_{ij}(z) \mapsto s_{ij} - z t_{ij},\ t_{ij}(z) \mapsto t_{ij} - z^{-1}s_{ij}$ extends uniquely to a superalgebra morphism $\ev: U_q(\Gaff) \longrightarrow \Uc_q(\Glie)$.
\end{prop}
We understand that $s_{ji} = t_{ij} = 0$ in the superalgebra $\Uc_q(\Glie)$ for $1 \leq i < j \leq M+N$. The morphism $\ev$ is called an {\it evaluation morphism}. It is clear that $\ev \circ \iota = \Id_{\Uc_q(\Glie)}$.
\subsubsection{Structures of quantum superalgebras.} Let us gather together in this paragraph the main properties of $U_q(\Gaff),Y_q(\Glie), \Uc_q(\Glie)$ which will be used later on.
\begin{itemize}
\item[(a)] For $a \in \BC^{\times}$, there is an automorphism of Hopf superalgebra
\begin{equation} \label{for: automorphisms of Z-graded superalgebras}
\Phi_a: U_q(\Gaff) \longrightarrow U_q(\Gaff),\quad s_{ij}^{(n)} \mapsto a^n s_{ij}^{(n)},\ t_{ij}^{(n)} \mapsto a^{-n} t_{ij}^{(n)}.
\end{equation}
$\Phi_a$ restricts to an automorphism of the $q$-Yangian still denoted by $\Phi_a: Y_q(\Glie) \longrightarrow Y_q(\Glie)$. Let us define the evaluation morphism $\ev_a := \ev \circ \Phi_a$. So $\ev_1 = \ev$.
\item[(b)] The following relations hold in $U_q(\Gaff)$ in view of Equations \eqref{rel: RTT = TTR}-\eqref{rel: FRTS invertibility condition}:
\begin{eqnarray}
&& s_{ii}^{(0)} s_{jk}^{(n)} = q^{(\epsilon_i,\epsilon_j - \epsilon_k)} s_{jk}^{(n)}s_{ii}^{(0)},\quad s_{ii}^{(0)}t_{jk}^{(n)} = q^{(\epsilon_i,\epsilon_j - \epsilon_k)} t_{jk}^{(n)}s_{ii}^{(0)} \quad (i,j,k \in I), \label{rel: Cartan relations for weight grading} \\
&& [s_{i,i+1}^{(0)}, t_{j+1,j}^{(0)}]_{q^{\delta_{i,j+1} - \delta_{i+1,j}}} = \delta_{ij}(q_i-q_i^{-1}) (t_{ii}^{(0)}s_{i+1,i+1}^{(0)} - s_{ii}^{(0)}t_{i+1,i+1}^{(0)}) \quad (i,j \in I_0), \label{rel: relations between e_i and f_j} \\
&& [s_{ji}^{(1)},t_{kj}^{(0)}] = (q_j-q_j^{-1}) t_{jj}^{(0)}s_{ki}^{(1)} \quad (1 \leq i < j < k \leq M+N). \label{rel: relations between e_0 and f_i}
\end{eqnarray}
Relation \eqref{rel: Cartan relations for weight grading} gives rise to the weight grading on $U_q(\Gaff)$: for $\alpha \in \BQ$,
\begin{displaymath}
(U_q(\Gaff))_{\alpha} = \{ x \in U_q(\Gaff) \ |\ s_{ii}^{(0)} x (s_{ii}^{(0)})^{-1} = q^{(\epsilon_i,\alpha)} x \quad \textrm{for}\ i \in I \}.
\end{displaymath}
In particular, for $i,j \in I$ we have $|s_{ij}^{(n)}|_{\BQ} = |t_{ij}^{(n)}|_{\BQ} = \epsilon_i - \epsilon_j$.
\item[(c)] Let $f \in 1 + z^{-1}\BC[[z^{-1}]]$ and $g \in 1 + z\BC[[z]]$. There exists a superalgebra automorphism
\begin{equation} \label{for: automorphism by power series}
\phi_{(f,g)}: U_q(\Gaff) \longrightarrow U_q(\Gaff),\quad t_{ij}(z) \mapsto f t_{ij}(z),\ s_{ij}(z) \mapsto g s_{ij}(z).
\end{equation}
These automorphisms behave well under coproduct in the following way:
\begin{displaymath}
(\phi_{(f_1,g_1)} \otimes \phi_{(f_2,g_2)} ) \circ \Delta = \Delta \circ \phi_{(f_1f_2,g_1g_2)}: U_q(\Gaff) \longrightarrow U_q(\Gaff)^{\otimes 2}
\end{displaymath}
for $f_1,f_2 \in 1 + z^{-1}\BC[[z^{-1}]],\ g_1,g_2 \in 1 + z\BC[[z]]$. Moreover $\phi_{(f,g)}$ restricts to a superalgebra automorphism of $q$-Yangian denoted by $\phi_g: Y_q(\Glie) \longrightarrow Y_q(\Glie)$ as it does not depend on $f$.
\item[(d)] The following defines an isomorphism of Hopf superalgebras
\begin{equation} \label{equ: transposition in quantum affine superalgebra}
\Psi: U_q(\Gaff) \longrightarrow U_q(\Gaff)^{\mathrm{cop}},\quad s_{ij}^{(n)} \mapsto \varepsilon_{ji}t_{ji}^{(n)},\quad t_{ij}^{(n)} \mapsto \varepsilon_{ji}s_{ji}^{(n)}.
\end{equation}
Here $\varepsilon_{ij} := (-1)^{|i|(|i|+|j|)}$ for $i,j \in I$.
\end{itemize}
We remark that the automorphism $\Psi$ and Relations \eqref{rel: Cartan relations for weight grading}-\eqref{rel: relations between e_0 and f_i} degenerate directly to the quantum superalgebra $\Uc_q(\Glie)$ thanks to Proposition \ref{prop: evaluation morphism}. In particular, $\Uc_q(\Glie)$ is $\BQ$-graded in the obvious way.
\subsection{Schur-Weyl duality.} There is a natural representation $\rho_{(1)}$ of $\Uc_q(\Glie)$ on the vector superspace $\BV$ defined by the following matrix equations \cite[\S 4.4]{Z2}:
\begin{displaymath}
(\rho_{(1)} \otimes \Id_{\End \BV}) (T) = (\Id_{\End \BV} \otimes \tau) (R^{-1}),\quad (\rho_{(1)} \otimes \Id_{\End \BV})(S) = (\Id_{\End \BV} \otimes \tau) ((R')^{-1}).
\end{displaymath}
We would like to understand the $\Uc_q(\Glie)$-module structure of the tensor powers $\BV^{\otimes s}$.
\subsubsection{Highest weight representations.} \label{sec: hwt for quantum superalgebra}
Let $\lambda \in \BP$. Up to isomorphism, there exists a unique simple $\Uc_q(\Glie)$-module, denoted by $L(\lambda)$, which is generated by a vector $v_{\lambda}$ satisfying:
\begin{displaymath}
|v_{\lambda}| = |\lambda|,\quad s_{kk} v_{\lambda} v_{\lambda} = q^{(\epsilon_k,\lambda)} v_{\lambda}, \quad t_{kk} v_{\lambda} = q^{-(\epsilon_k,\lambda)} v_{\lambda},\quad s_{ij} v_{\lambda} = 0 \quad (i,j,k \in I,\ i < j).
\end{displaymath}
For example, as $\Uc_q(\Glie)$-modules, $\BV \cong L(\epsilon_1)$ in view of the following equations:
\begin{eqnarray*}
&& \rho_{(1)}(s_{ii}) = q_i E_{ii} + \sum_{j \neq i} E_{jj} = \rho_{(1)} (t_{ii}^{-1}) \quad (\textrm{for}\ i \in I), \\
&& \rho_{(1)}(s_{ij}) = (q_i - q_i^{-1}) E_{ij},\quad \rho_{(1)}(t_{ji}) = (q_i^{-1}-q_i) E_{ji}\quad (\textrm{for}\ 1 \leq i < j \leq M+N).
\end{eqnarray*}
\subsubsection{Characters.} \label{sec: character}
Usually a $\Uc_q(\Glie)$-module $V$ is $\BP$-graded via the action of the $s_{ii}$:
\begin{displaymath}
(V)_{\alpha} := \{ x \in V \ |\ s_{ii} x = q^{(\epsilon_i,\alpha)} x \ \textrm{for}\ i \in I \}.
\end{displaymath}
From the $\BQ$-grading on $\Uc_q(\Glie)$ we see that $(\Uc_q(\Glie))_{\alpha} (V)_{\beta} \subseteq (V)_{\alpha + \beta}$ for $\alpha \in \BQ$ and $\beta \in \BP$. Assume furthermore that all weight spaces $(V)_{\alpha}$ are finite-dimensional. Introduce {\it characters}
\begin{equation}
\chi (V) := \sum_{\alpha \in \BP} \dim (V)_{\alpha} [\alpha] \in \BZ^{\BP}.
\end{equation}
Here $\BZ^{\BP}$ is the abelian group of functions $\BP \longrightarrow \BZ$ and $[\alpha] = \delta_{\alpha,\cdot}$. For example,
\begin{displaymath}
\chi (\BV) = \sum_{i \in I} [\epsilon_i] \in \BZ[\BP] \subset \BZ^{\BP}
\end{displaymath}
where $\BZ[\BP]$ is the group ring of $\BP$ over $\BZ$.
\subsubsection{Young combinatorics.} Let us introduce several combinatorial objects before stating Schur-Weyl duality.
\begin{defi} \label{def: Young combinatorics}
(1) Let $\CP \subseteq \BP$ be the subset consisting of $\lambda = \sum_{i \in I} \lambda_i \epsilon_i$ such that
\begin{itemize}
\item[(BKK1)] $\lambda_i \geq 0$ for $i \in I$;
\item[(BKK2)] for $i \in I_0\setminus \{M\}$, $\lambda_i \geq \lambda_{i+1}$;
\item[(BKK3)] if $\lambda_{M+j} > 0$ for some $1 \leq j \leq N$, then $\lambda_M \geq j$.
\end{itemize}
For such $\lambda \in \CP$, define $\hit(\lambda) := \sum_{i \in I} \lambda_i \in \BZ_{\geq 0}$.
(2) A $\Glie$-Young diagram is a Young diagram $Y$ such that $(M+1,N+1) \notin Y$. In other words, it is a finite subset $Y \subset \BZ_{>0} \times \BZ_{>0}$ satisfying
\begin{align*}
& (i,j+1) \in Y \Longrightarrow (i,j) \in Y \quad \textrm{for}\ i,j \in \BZ_{>0}, \\
& \sharp \{j \in \BZ_{>0}\ |\ (i,j) \in Y \} \geq \sharp \{j \in \BZ_{>0}\ |\ (i+1,j) \in Y \} \quad \textrm{for}\ i \in \BZ_{>0}, \\
& \sharp \{j \in \BZ_{>0}\ |\ (M+1,j) \in Y \} \leq N.
\end{align*}
Let $\YD$ be the set of $\Glie$-Young diagrams.
(3) For $Y \in \YD$, a $\Glie$-Young tableau of shape $Y$ is a map $f: Y \longrightarrow I$ such that
\begin{itemize}
\item[(T1)] if $(i,j),(i',j') \in Y$ and $i \leq i',j \leq j'$, then $f(i,j) \leq f(i',j')$;
\item[(T2)] if $(i,j),(i+1,j) \in Y$ and $f(i,j) \leq M$, then $f(i,j) < f(i+1,j)$;
\item[(T3)] if $(i,j),(i,j+1) \in Y$ and $f(i,j) \geq M+1$, then $f(i,j) < f(i+1,j)$.
\end{itemize}
Let $\CB(Y)$ be the set of $\Glie$-Young tableaux of shape $Y$.
(4) For $\lambda = \sum_i \lambda_i \epsilon_i \in \CP$, define $Y^{\lambda}$ to be the $\Glie$-Young diagram formed of such $(i,j)$ that
\begin{itemize}
\item[(Y1)] if $1 \leq i \leq M$, then $j \leq \lambda_i$;
\item[(Y2)] if $i > M$, then $j \leq N$ and $\lambda_{M+j} \geq i - M$.
\end{itemize}
Let $\Ym: \CP \longrightarrow \YD, \lambda \mapsto Y^{\lambda}$ be the bijective map thus obtained.
\end{defi}
\begin{rem} \label{rem: notations for parabolic subalgebras gl(a,b)}
Later on, we shall consider $\mathfrak{gl}(a,b)$-Young diagrams and Young tableaux either for $(a = M, 1 \leq b \leq N)$ or for $(1 \leq a \leq M, b = 0)$. In this case, we have analogous definitions of $I_{a,b},\BP_{a,b},\CP_{a,b}, \YD_{a,b}, \CB_{a,b}(Y), \Ym_{a,b}$ by replacing $(M,N)$ with $(a,b)$ everywhere. On the other hand, we shall make the obvious inclusions:
\begin{align*}
& \BP_{1,0} \subset \BP_{2,0} \subset \cdots \subset \BP_{M-1,0} \subset \BP_{M,0} \subset \BP_{M,1} \subset \cdots \subset \BP_{M,N-1} \subset \BP_{M,N} = \BP; \\
& \CP_{1,0} \subset \CP_{2,0} \subset \cdots \subset \CP_{M-1,0} \subset \CP_{M,0} \subset \CP_{M,1} \subset \cdots \subset \CP_{M,N-1} \subset \CP_{M,N} = \CP.
\end{align*}
\end{rem}
\begin{theorem} \label{thm: BKK Schur-Weyl duality} \cite{BKK}
For all $s \in \BZ_{\geq 1}$, the $\Uc_q(\Glie)$-module $\BV^{\otimes s}$ is completely reducible. More precisely, we have a decomposition into simple sub-$\Uc_q(\Glie)$-modules as follows:
\begin{displaymath}
\BV^{\otimes s} = \bigoplus_{\lambda \in \CP} L(\lambda)^{\oplus c_{\lambda}}.
\end{displaymath}
Here $c_{\lambda} \in \BZ_{\geq 0}$ and $c_{\lambda} \neq 0$ if and only if $\hit(\lambda) = s$. Furthermore, for $\lambda \in \CP$
\begin{equation} \label{equ: character of simple modules}
\chi (L(\lambda)) = \sum_{f \in \CB(Y^{\lambda})} [\sum_{(i,j) \in Y^{\lambda}} \epsilon_{f(i,j)}] \in \BZ[\BP].
\end{equation}
\end{theorem}
Such representations $L(\lambda)$ with $\lambda \in \CP$ are usually called {\it polynomial representations}, as they appear as simple submodules of tensor powers of the natural representation.
Let us end this paragraph with the following simple application of the character formula \eqref{equ: character of simple modules}. This result, on the asymptotic behaviour of weight spaces, serves as a motivation for our limit construction carried out later.
\begin{defi} \label{def: fondamental weights}
For $r \in I_0$, the $r$th fundamental weight $\varpi_r \in \BP$ is defined by the formula $\varpi_r := \begin{cases}
\sum_{j=1}^r \epsilon_j & (r \leq M), \\
-\sum_{j=r+1}^{M+N} \epsilon_j & (r > M). \end{cases} $
\end{defi}
\begin{cor} \label{cor: asymptotic behaviour of weight spaces}
Suppose that $(L(l\varpi_r))_{l\varpi_r - \beta} \neq 0$ where $1 \leq r \leq M, l > 0$ and $\beta \in \BQ_{\geq 0}$. Then
\begin{displaymath}
\dim (L(k \varpi_r))_{k \varpi_r - \beta} = \dim (L(rl \varpi_r))_{rl \varpi_r - \beta}
\end{displaymath}
for all $k \geq rl$.
\end{cor}
\begin{proof}
Note that $k \varpi_r \in \CP$ for all $k \in \BZ_{>0}$. Furthermore,
\begin{displaymath}
Y^{k\varpi_r} = \{1,2,\cdots, r \} \times \{ 1,2,\cdots,k \} \subset \BZ_{>0} \times \BZ_{>0}.
\end{displaymath}
In view of Formula \eqref{equ: character of simple modules}, we only need to check the following: for all $k > lr$ and for all $f \in \CB^{Y^{k\varpi_r}}$ such that
\begin{displaymath}
\sum_{(i,j) \in Y^{k \varpi_r}} \epsilon_{f(i,j)} = k \varpi_r - \beta,
\end{displaymath}
we must have $f(i,1) = i$ for $1 \leq i \leq r$. Let us fix such an $f$.
The condition $(L(l\varpi_r))_{l\varpi_r - \beta} \neq 0$ says the existence of $g \in \CB(Y^{l\varpi_r})$ such that
\begin{displaymath}
l \varpi_r - \beta = \sum_{(i,j) \in Y^{l\varpi_r}} \epsilon_{g(i,j)}.
\end{displaymath}
This says that
\begin{displaymath}
\sum_{(i,j) \in Y^{k \varpi_r}} \epsilon_{f(i,j)} = (k-l) \sum_{i=1}^r \epsilon_i + \sum_{(i,j) \in Y^{l \varpi_r}} \epsilon_{g(i,j)} =: \sum_{i=1}^{M+N} c_i \epsilon_i.
\end{displaymath}
In particular, $c_i \geq k-l$ for $1 \leq i \leq r$. We prove by induction on $1 \leq i \leq r$ that: $f(i,j) = i$ for all $1 \leq j \leq k - il$. For $i = 1$, this is obvious from the definition of Young tableau. Suppose the assertion for $i-1$ true. If there exists $1 \leq j_0 \leq k-il$ such that $f(i,j_0) > i$, consider the pre-image $f^{-1}(i)$ of $i$. Suppose that $(m,n) \in f^{-1}(i)$. Then we have the following:
\begin{itemize}
\item[(1)] $m \leq i$, from the definition of Young tableau;
\item[(2)] if $m = i$, then $1 \leq n < j_0$. In particular, $n \leq k - il - 1$;
\item[(3)] if $m < i$, then $k - (i-1)l < n \leq k$ as $f(i-1, j) = i- 1$ for $1 \leq j \leq k-(i-1)l$.
\end{itemize}
Note that $(m,n),(m',n) \in f^{-1}(i)$ forces $m = m'$. By counting the number of $n$ in $f^{-1}(i)$ we get
\begin{displaymath}
\sharp f^{-1}(i) = c_i \leq k - il - 1 + (i-1) l = k - l - 1,
\end{displaymath}
which contradicts with the fact that $c_i \geq k-l$. Thus, $f(i,j) = i$ for all $1 \leq j \leq k - il$ and $1 \leq i \leq r$, as desired.
\end{proof}
For example, when $r = 1$, all the weight spaces of $L(k \varpi_1)$ are one-dimensional.
\subsection{Gelfand-Tsetlin basis.} \label{sec: Gelfand-Tsetlin}
Following Remark \ref{rem: notations for parabolic subalgebras gl(a,b)}, let $\Xm$ be the set of pairs $(a,b) \in \BZ^2$ such that: either $(a = M, 1 \leq b \leq N)$; or $(1 \leq a \leq M, b = 0)$. Let $k \mapsto (M_k,N_k)$ be the unique bijective map $I \longrightarrow \Xm$ such that $M_k + N_k = k$ for $k \in I$. For $k \in I$, let $\Uc_q(\Glie_k)$ be the subalgebra of $\Uc_q(\Glie)$ generated by the $s_{ij},t_{ji}$ with $1 \leq i \leq j \leq k$. Then $\Uc_q(\Glie_k)$ is a sub-Hopf-superalgebra canonically isomorphic to $\Uc_q(\mathfrak{gl}(M_k,N_k))$, and we have the following sequence of inclusions of Hopf superalgebras:
\begin{equation*}
\Uc_q(\Glie_1) \subset \Uc_q(\Glie_2) \subset \cdots \subset \Uc_q(\Glie_{M+N-1}) \subset \Uc_q(\Glie_{M+N}) = \Uc_q(\Glie).
\end{equation*}
Identify $\Glie_k$ with $\mathfrak{gl}(M_k,N_k)$ from now on.
Suppose $2\leq k \leq M+N$. Let $\lambda \in \CP_{M_k,N_k}$. Let $Y^{\lambda} = \Ym_{M_k,N_k}(\lambda)$ be the $\Glie_k$-Young diagram. If $f \in \CB_{M_k,N_k}(Y^{\lambda})$ is a $\Glie_k$-Young tableau of shape $Y^{\lambda}$, then the subset $f^{-1}(I_{M_{k-1},N_{k-1}}) \subseteq Y^{\lambda}$ is a $\Glie_{k-1}$-Young diagram. Moreover, the restriction of $f$ to this subset gives us a $\Glie_{k-1}$-Young tableau of corresponding shape.
\begin{defi} \label{def: Gelfand-Tsetlin patterns}
(1) Suppose $2 \leq k \leq M+N$. For $\lambda \in \CP_{M_k,N_k}$, let $S_{k}(\lambda)$ be the set
\begin{equation*}
S_{k}(\lambda) := \{ \Ym_{M_{k-1},N_{k-1}}^{-1}(f^{-1}(I_{M_{k-1},N_{k-1}})) \in \CP_{M_{k-1},N_{k-1}} \ |\ f \in \CB_{M_k,N_k}(\Ym_{M_k,N_k}(\lambda)) \}.
\end{equation*}
(2) A sequence of weights $(\lambda^{(k)}: k \in I) \in \CP^{M+N}$ is called a Gelfand-Tsetlin pattern if $\lambda^{(k-1)} \in S_k(\lambda^{(k)})$ for all $2 \leq k \leq M+N$. For $\lambda \in \CP$, let $\GT(\lambda)$ be the set of Gelfand-Tsetlin patterns $(\lambda^{(k)})$ with $\lambda^{(M+N)} = \lambda$.
\end{defi}
As in the non-graded case, Gelfand-Tsetlin patterns are in one-to-one correspondence with Young tableaux.
\begin{lem} \label{lem: GT patterns vs Young tableaux}
Let $\lambda \in \CP$. The following defines a bijective mapping
\begin{displaymath}
GT_{\lambda}: \CB(Y^{\lambda}) \longrightarrow \GT(\lambda), \quad GT_{\lambda}(f)^{(k)} = \Ym_{M_k,N_k}^{-1}(f^{-1}(I_{M_k,N_k})).
\end{displaymath}
\end{lem}
\begin{example}
If $N = 0$, then $\CP$ is the set of weights $\sum_{i=1}^M \lambda_i \epsilon_i \in \BP$ with $\lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_M$. A Gelfand-Tsetlin pattern is a sequence of weights $(\sum_{i=1}^k \lambda_{i}^{(k)} \epsilon_i: 1 \leq k \leq M)$ with:
\begin{itemize}
\item[(1)] $\lambda_1^{(k)} \geq \lambda_2^{(k)} \geq \cdots \geq \lambda_k^{(k)}$ for $1 \leq k \leq M$;
\item[(2)] $\lambda_i^{(k)} \geq \lambda_{i}^{(k-1)} \geq \lambda_{i+1}^{(k)}$ for $2 \leq k \leq M$ and $1 \leq i \leq k-1$.
\end{itemize}
This is the classical definition of Gelfand-Tsetlin pattern for $\mathfrak{gl}_M$ (see \cite[pp.118]{MolevGT}). When $N > 0$, the situation becomes complicated. For example, take $(M,N) = (2,1)$. Then the set $\GT(2 \epsilon_1)$ has $5$ elements, namely
\begin{eqnarray*}
(2 \epsilon_1, 2 \epsilon_1, 2 \epsilon_1), \quad (\epsilon_1, 2\epsilon_1, 2 \epsilon_1), \quad (\epsilon_1,\epsilon_1,2\epsilon_1), \quad
(0, 2 \epsilon_1, 2 \epsilon_1),\quad (0, \epsilon_1, 2 \epsilon_1).
\end{eqnarray*}
In particular, $(0,0,2\epsilon_1)$ is not permitted.
\end{example}
For $k \in I, \lambda \in \CP_{M_k,N_k}$ and $s \in \super$, let $L(\lambda;\Glie_k)$ be the simple $\Uc_q(\Glie_k)$-module generated by a highest weight vector of highest weight $\lambda$ and of $\super$-degree $|\lambda|$; let $\BC_s$ be the one-dimensional $\Uc_q(\Glie_k)$-module of zero weight and of $\super$-degree $s$. The following is a direct consequence of Theorem \ref{thm: BKK Schur-Weyl duality}.
\begin{cor}
Let $2 \leq k \leq M+N$ and let $\lambda \in \CP_{M_k,N_k}$. Then the $\Uc_q(\Glie_{k-1})$-module $\Res_{\Uc_q(\Glie_{k-1})}^{\Uc_q(\Glie_k)} L(\lambda; \Glie_k)$ is completely reducible with the following decomposition:
\begin{equation*}
\Res_{\Uc_q(\Glie_{k-1})}^{\Uc_q(\Glie_k)} L(\lambda; \Glie) = \bigoplus_{\lambda' \in S_k(\lambda)} L(\lambda';\Glie_{k-1}) \otimes \BC_{(\hit(\lambda)-\hit(\lambda'))|k|}.
\end{equation*}
as $\Uc_q(\Glie_{k-1})$-modules.
\end{cor}
Now we get an explicit description of Gelfand-Tsetlin basis as follows.
\begin{cor} \label{cor: Gelfand-Tsetlin basis for modules in BKK}
For $\lambda \in \CP$, the $\Uc_q(\Glie)$-module $L(\lambda)$ is equipped with a basis $(v_{\underline{\lambda}}: \underline{\lambda} \in \GT(\lambda))$ satisfying: $v_{\underline{\lambda}}$ is contained in a sub-$\Uc_q(\Glie_k)$-module isomorphic to
\begin{equation*}
L(\lambda^{(k)};\Glie_k) \otimes \BC_{|\sum_{l=k+1}^{M+N} \sharp (GT_{\lambda}^{-1}\underline{\lambda})^{-1}(l) l|}
\end{equation*}
for $\underline{\lambda} = (\lambda^{(k)}) \in \GT(\lambda)$ and $k \in I$.
\end{cor}
\subsection{Dual Gelfand-Tsetlin basis.} \label{sec: dual GT basis}
We shall also need Gelfand-Tsetlin bases for $\Uc_q(\Glie)$-modules $L(\lambda)^*$ with $\lambda \in \CP$. In general, if $H$ is a Hopf superalgebra and if $V$ is an $H$-module, then the vector superspace $V^* = \hom(V,\BC)$ is endowed with an $H$-module structure by:
\begin{displaymath}
(a, f) \mapsto a f: v \mapsto (-1)^{|a||f|} f(\Sm(a) v)
\end{displaymath}
for $\super$-homogeneous $a \in H$ and $f \in V^*$. Here $\Sm: H \longrightarrow H$ is the antipode.
Let $(v_{\underline{\lambda}}: \underline{\lambda} \in \GT(\lambda))$ be a Gelfand-Tsetlin basis for $L(\lambda)$. Let $(v_{\underline{\lambda}}^*: \underline{\lambda} \in \GT(\lambda))$ be its dual basis for $L(\lambda)^*$. Since the embeddings $\Uc_q(\Glie_{k-1}) \longrightarrow \Uc_q(\Glie_k)$ respect the Hopf superalgebra structures, and since the restricted modules $\Res_{\Uc_q(\Glie_k)}^{\Uc_q(\Glie)} L(\lambda)$ are always completely reducible, we conclude that: $v_{\underline{\lambda}}^*$ is contained in a simple sub-$\Uc_q(\Glie_k)$-module isomorphic to
\begin{equation*}
L(\lambda^{(k)};\Glie_k)^* \otimes \BC_{|\sum_{l=k+1}^{M+N} \sharp (GT_{\lambda}^{-1}\underline{\lambda})^{-1}(l) l|}
\end{equation*}
for $\underline{\lambda} \in \GT(\lambda)$ and $k \in I$. For this reason, call $(v_{\underline{\lambda}}^*: \underline{\lambda} \in \GT(\lambda))$ a {\it dual Gelfand-Tsetlin basis}.
Let us compute the highest weight of the $\Uc_q(\Glie)$-module $L(\lambda)^*$ when $\lambda \in \CP$. Let $Y^{\lambda} = \Ym(\lambda)$ be the associated $\Glie$-Young diagram. For all $i,j \in \BZ_{>0}$, define
\begin{eqnarray*}
&& r_i := \sharp \{j \in \BZ_{>0}\ |\ (i,j) \in Y^{\lambda} \},\quad c_j := \sharp \{i \in \BZ_{>0}\ |\ (i,j) \in Y^{\lambda} \}; \\
&& r_i' := \max (r_i - N, 0),\quad c_j' := \max (c_j - M, 0).
\end{eqnarray*}
Then the highest and lowest weights $\lambda, \lambda_b$ of $L(\lambda)$ can be written as:
\begin{displaymath}
\lambda = \sum_{i=1}^M r_i \epsilon_i + \sum_{j=1}^N c_j' \epsilon_{M+j}, \quad \lambda_b = \sum_{i=1}^M r_{M+1-i}' \epsilon_i + \sum_{j=1}^N c_{N+1-j} \epsilon_{M+j}.
\end{displaymath}
Moreover, as $\Uc_q(\Glie)$-modules, we have the identification
\begin{equation} \label{equ: dual of a module in BKK}
L(\lambda)^* \cong L(-\lambda_b).
\end{equation}
In the non-graded case, $\lambda_b = w_0 (\lambda)$ where $w_0$ is a longest element of the Weyl group associated with the Lie algebra. In particular, $(\lambda + \mu)_b = \lambda_b + \mu_b$ for $\lambda,\mu$ dominant. Such additive formula is no longer true in our case. For example, take $(M,N) = (2,1)$. Then $(\epsilon_1)_b = \epsilon_3$ and $(2\epsilon_1)_b = \epsilon_2 + \epsilon_3$.
\begin{rem}
(1) For the quantum superalgebra $U_q(\Glie)$, Palev-Stoilova-Van der Jeugt \cite{PSV} established the Gelfand-Tsetlin bases for the so-called {\it essentially typical representations}. In our situation, for $\lambda \in \CP$, $L(\lambda)$ is essentially typical if and only if $(M,N) \in Y^{\lambda}$. When this is the case, a combinatorial description of the set $\GT(\lambda)$ similar to the example above has been given (cf. Equations (9)-(10) in {\it loc. cit}). Moreover, explicit actions of the generators $e_i^{\pm}$ with respect to the Gelfand-Tsetlin bases were given therein.
(2) For the Lie superalgebra $\Glie$, there are analogs of such representations $L_0(\lambda)$ as $\lambda \in \CP$ constructed as simple submodules of tensor powers of the natural representations of $\Glie$ on $\BV$, also called {\it covariant representations}. (One can view $L(\lambda)$ as a deformation of $L_0(\lambda)$.) In \cite{MolevGT2}, Molev constructed Gelfand-Tsetlin bases for all the $L_0(\lambda)$ and deduced explicit formulas of the actions of the generators $E_{ij}$ with $i,j \in I, |i-j| = 1$ with respect to these bases. The main ingredients used by Molev are the Yangian $Y(\mathfrak{gl}_N)$ and the Mickelsson-Zhelobenko algebra $Z(\Glie,\mathfrak{gl}_M)$.
(3) We shall give another characterization of Gelfand-Tsetlin basis for $L(\lambda)$ and dual Gelfand-Tsetlin basis for $L(\lambda)^*$ within the framework of representations of the quantum affine superalgebra $U_q(\Gaff)$. (See Propositions \ref{prop: q-character and GT basis} and \ref{prop: q-character and dual GT basis}.)
\end{rem}
\section{Kirillov-Reshetikhin module and tensor products} \label{sec: KR}
We prove a cyclicity result on tensor products of Kirillov-Reshetikhin modules over $U_q(\Gaff)$, based on our previous result \cite[Theorem 4.4]{Z2} and on duality arguments.
\subsection{Kirillov-Reshetikhin modules.} \label{sec: KR modules}
Let us recall the definition of Kirillov-Reshetikhin modules $W_{k,a}^{(r)}$ with $r \in I_0, a \in \BC^{\times}, k \in \BZ_{>0}$. These are $U_q(\Gaff)$-modules
\begin{equation}
W_{k,a}^{(r)} := \ev_a^* L(k \varpi_r) \otimes \BC_{|k \varpi_r|}
\end{equation}
where as usual $\BC_{s}$ is the one-dimensional $U_q(\Gaff)$-module of $\super$-degree $s \in \super$.
Let $V$ be a $U_q(\Gaff)$-module. A non-zero vector $v \in V \setminus \{0\}$ is called a {\it highest $\ell$-weight vector} if $v$ is $\super$-homogeneous and
\begin{displaymath}
s_{ij}^{(n)} v = 0 = t_{ij}^{(n)}v,\quad s_{kk}^{(n)}v,t_{kk}^{(n)}v \in \BC v \quad (n \in \BZ_{\geq 0},\ i,j,k \in I,\ i < j).
\end{displaymath}
V is called a {\it highest $\ell$-weight module} if $V = U_q(\Gaff)v$ for some highest $\ell$-weight vector $v$. Similarly, the notions of {\it lowest $\ell$-weight vector} and {\it lowest $\ell$-weight module} are defined by replacing $(i< j)$ with $(i > j)$. Let $V,V'$ be two $U_q(\Gaff)$-modules and let $v,v'$ be highest/lowest $\ell$-weight vectors in $V,V'$ respectively. Write
\begin{displaymath}
s_{ii}(z) v = f_i(z) v,\quad s_{ii}(z) v' = f_i'(z) v' \quad \textrm{for}\ i \in I.
\end{displaymath}
Then $v \otimes v'$ is a highest/lowest $\ell$-weight vector such that
\begin{displaymath}
s_{ii}(z) (v \otimes v') = f_i(z) f_i'(z) v \otimes v' \quad \textrm{for}\ i \in I.
\end{displaymath}
For example, let $v_{k\varpi_r} \in L(k\varpi_r)$ be as in \S \ref{sec: hwt for quantum superalgebra}, and let $v_s \in \BC_s$ be a non-zero vector with $s = |k \varpi_r|$. Then $w := \ev_a^* v_{k\varpi_r} \otimes v_s \in W_{k,a}^{(r)}$ is a highest $\ell$-weight vector and
\begin{displaymath}
|w| = \even,\quad s_{ii}(z) w = (q^{(\epsilon_i,k\varpi_r)} - z a q^{-(\epsilon_i,k\varpi_r)}) w,\quad t_{ii}(z) w = (q^{-(\epsilon_i,k\varpi_r)} - z^{-1}a^{-1} q^{(\epsilon_i,k\varpi_r)}) w.
\end{displaymath}
The rest of this section is devoted to proving the following cyclicity result.
\begin{theorem} \label{thm: cyclicity of KR modules}
Let $r \in I_0, a \in \BC^{\times}$ and $l_1,l_2,l_3 \in \BZ_{>0}$ such that $l_1 < l_2 < l_3$.
\begin{itemize}
\item[(1)] If $r \leq M$, then $W_{l_1,a}^{(r)} \otimes W_{l_2-l_1,aq^{-2l_1}}^{(r)} \otimes W_{l_3-l_2,aq^{-2l_2}}^{(r)}$ is of highest $\ell$-weight.
\item[(2)] If $r > M$, then $W_{l_3-l_2,aq^{-2l_2}}^{(r)} \otimes W_{l_2-l_1,aq^{-2l_1}}^{(r)} \otimes W_{l_1,a}^{(r)}$ is of highest $\ell$-weight.
\end{itemize}
\end{theorem}
Let us recall another cyclicity result before turning to the proof.
\begin{theorem} \label{thm: cyclicity of fundamental modules} \cite[Theorem 4.2, Proposition 4.7]{Z2}
Let $r \in I_0, a \in \BC^{\times}$ and $k \in \BZ_{>0}$.
\begin{itemize}
\item[(I)] $\bigotimes_{j=1}^k W_{1,aq_r^{-2j}}^{(r)}$ is of highest $\ell$-weight.
\item[(II)] $\bigotimes_{j=1}^k W_{1,aq^{2j}}^{(1)}$ is of lowest $\ell$-weight.
\end{itemize}
\end{theorem}
\subsubsection{Reduction to the case $r \leq M$.} \label{sec: reduction}
Following \cite[\S 4.2]{Z2}, let $\Glie' := \mathfrak{gl}(N,M)$. Let us define the quantum affine superalgebra $U_q(\Gafft)$ in the same way as in \ref{sec: quantum affine superalgebra}, except that $M,N$ are replaced by each other everywhere. For distinction, let $s_{ij;J}^{(n)},t_{ij;J}^{(n)}$ be the defining generators. For $i \in I$, set $\overline{i} := M+N+1-i$ and $|i|^J := \begin{cases}
\even & (i \leq N) \\
\odd & (i > N)
\end{cases}$. So that $|s_{ij;J}^{(n)}| = |t_{ij;J}^{(n)}| = |i|^J + |j|^J$. Let $\varepsilon_{ij}^J = (-1)^{|i|^J(|i|^J + |j|^J)}$. By \cite[Proposition 4.3]{Z2}, the following
\begin{equation} \label{for: transposition of quantum affine superalgebras}
f_{J,I}: U_q(\Gafft)^{\textrm{cop}} \longrightarrow U_q(\Gaff),\quad s_{ij;J}^{(n)} \mapsto \varepsilon_{ji}^J s_{\overline{j}\overline{i}}^{(n)},\quad t_{ij;J}^{(n)} \mapsto \varepsilon_{ji}^J t_{\overline{j}\overline{i}}^{(n)}
\end{equation}
defines an isomorphism of Hopf superalgebras.
Let $V$ be a $U_q(\Gaff)$-module. If $v \in V$ is a highest $\ell$-weight vector, then so is $f_{J,I}^* v \in f_{J,I}^* V$. Let us $W_{k,a;J}^{(r)}$ be the corresponding Kirillov-Reshetikhin modules over $U_q(\Gafft)$. Then, as $U_q(\Gafft)$-modules, there are isomorphisms $W_{k,a;J}^{(N+M-r)} \cong f_{J,I}^* W_{k,a}^{(r)}$ for $r \neq M$.
Hence, if (1) of Theorem \ref{thm: cyclicity of KR modules} is proved, (2) follows automatically.
\subsubsection{Dualities.} Let $V$ be a $U_q(\Gaff)$-module. On the dual space $V^* = \hom(V,\BC)$, there are two structures of $U_q(\Gaff)$-modules. The first one is afforded by the Hopf superalgebra structure on $U_q(\Gaff)$, the corresponding $U_q(\Gaff)$-module still denoted by $V^{*}$ as in \S \ref{sec: dual GT basis}. The second one is the pull back $\Psi^* V^*$, where $\Psi$ is the superalgebra isomorphism given by \eqref{equ: transposition in quantum affine superalgebra}. Let us write $\Psi^* V^*$ as $V^{\vee}$ for distinction. As usual, when $V,W$ are finite-dimensional $U_q(\Gaff)$-modules, there are canonical isomorphisms of $U_q(\Gaff)$-modules
\begin{displaymath}
(V \otimes W)^* \cong W^* \otimes V^*,\quad (V \otimes W)^{\vee} \cong V^{\vee} \otimes W^{\vee}.
\end{displaymath}
Let $V$ be a $U_q(\Gaff)$-module. Let $v \in V \setminus \{0\}$ be $\super$-homogeneous. We say that $V$ is {\it cogenerated} by $v$ if every non-zero submodule of $V$ contains $v$. In this case, $v$ is also called a {\it cogenerator} and we write
\begin{displaymath}
U_q(\Gaff) v =: \socle (V).
\end{displaymath}
Note that $\socle(V)$ is {\bf the} simple sub-$U_q(\Gaff)$-module of $V$.
\begin{rem} \label{rem: cyclicity vs cocyclicity}
Let $V$ be a $U_q(\Gaff)$-module $\BP$-graded with respect to the action of the $s_{ii}^{(0)}$:
\begin{displaymath}
(V)_{\lambda} = \{x \in V\ |\ s_{ii}^{(0)} x = q^{(\lambda,\epsilon_i)} x\ \textrm{for}\ i \in I \}.
\end{displaymath}
Suppose that there exists $\lambda_0 \in \BP$ such that $(V)_{\lambda_0} = \BC v_0 \neq 0$ and $(V)_{\lambda} \neq 0$ implies $\lambda_0 - \lambda \in \BQ_{\geq 0}$.
Clearly, for such $V$, the vector $v_0$ must be a highest $\ell$-weight vector. Let $v_0^*$ be the linear form on $V$ defined by: $v_0 \mapsto 1$ and $(V)_{\lambda} \mapsto 0$ for $\lambda \neq \lambda_0$. Then by definition of dual modules, $V$ is of highest $\ell$-weight if and only if $v_0^*$ is a cogenerator in $V^{\vee}$.
Analogously, suppose that all the weights of $V$ lie in the cone $\lambda_1 + \BQ_{\geq 0}$ and that $(V)_{\lambda_1} = \BC v_1$ is one-dimensional. Then one can define $v_1^{*} \in V^{\vee}$ similarly, and $V$ is of lowest $\ell$-weight if and only if $v_1^*$ is a cogenerator in $V^{\vee}$.
\end{rem}
\begin{rem} \label{rem: cogenerator and tensor products}
Let $s \in \BZ_{>0}$. For $1 \leq i \leq s$, let $V_i$ be a $U_q(\Gaff)$-module and let $v_i \in V_i$ be a non-zero $\super$-homogeneous vector. Set $S_i := U_q(\Gaff) v_i \subseteq V_i$.
If $\bigotimes_{i=1}^s V_i$ is cogenerated by $\bigotimes_{i=1}^s v_i$, then so is $\bigotimes_{i=1}^s S_i$.
\end{rem}
The proof of Theorem \ref{thm: cyclicity of KR modules} (1) will go as follows: first we compute the duals $(W_{k,a}^{(r)})^{\vee}$ and rephrase (1) as a statement of cogenerators; next we realize the duals $(W_{k,a}^{(r)})^{\vee}$ as simple submodules arising from cogenerators and apply Theorem \ref{thm: cyclicity of fundamental modules} and Remark \ref{rem: cogenerator and tensor products} to conclude.
We close this subsection with the following convention. Let $S_1,S_2$ be $U_q(\Gaff)$-modules. We say that $S_1 \simeq S_2$ if there exists a one-dimensional $U_q(\Gaff)$-module $D$ making the two $U_q(\Gaff)$-modules $S_1$ and $S_2 \otimes D$ isomorphic. Remark that $\simeq$ does not change the property of being of highest $\ell$-weight, of lowest $\ell$-weight, or cogenerated.
\subsection{Duals of Kirillov-Reshetikhin modules.} Throughout this subsection, $1 \leq r \leq M$. We shall determine $(W_{k,a}^{(r)})^{\vee}$ up to $\simeq$.
\subsubsection{Dual of $W_{1,a}^{(1)}$.} \label{sec: duals}
In view of \S \ref{sec: hwt for quantum superalgebra}, $W_{1,a}^{(1)} = \BV$ as vector superspace.
Let $\rho_a$ be the corresponding representation of $U_q(\Gaff)$ on $\BV$. Similarly, the $U_q(\Gaff)$-module $(W_{1,a}^{(1)})^{\vee}$ gives rise to the representation $(\rho_a^{\vee}, \BV^*)$. Let $(v_i^* \in \BV^* : i \in I)$ be the dual basis of $(v_i \in \BV: i \in I)$. Then $v_1^{*}$ is a highest $\ell$-weight vector in $(W_{1,a}^{(1)})^{\vee}$. In view of the $\BP$-grading on the simple module $(W_{1,a}^{(1)})^{\vee}$, it is enough to determine
\begin{align*}
s_{ii}(z) v_1^* = v_1^* (\Sm(t_{ii}(z^{-1})) v_1 ) v_1^{*},\quad t_{ii}(z) v_1^* = v_1^* (\Sm(s_{ii}(z^{-1}))) v_1^{*}.
\end{align*}
For this, let us introduce
\begin{displaymath}
X(z) := (\rho_a \otimes \Id_{\End \BV}) (S(z)),\quad Y(z) := (\rho_a \otimes \Id_{\End \BV})(T(z)).
\end{displaymath}
From Formulas \eqref{for: antipode for S}-\eqref{for: antipode for T} we get
\begin{displaymath}
X(z)^{-1} = (\rho_a \otimes \Id_{\End \BV}) ((\Sm \otimes \Id_{\End \BV}) (S(z))),\quad Y(z)^{-1} = (\rho_a \otimes \Id_{\End \BV})((\Sm \otimes \Id_{\End \BV}) (T(z))).
\end{displaymath}
We are led to determine the inverses of $X(z),Y(z)$. In view of Proposition \ref{prop: evaluation morphism},
\begin{displaymath}
Y(z) = - z^{-1} a^{-1} X(z) \in \End (\BV^{\otimes 2})[[z^{-1}]].
\end{displaymath}
Therefore, it is enough to find the inverse of $X(z) \in (\End \BV)^{\otimes 2}[[z]]$. By \S \ref{sec: hwt for quantum superalgebra}
\begin{eqnarray*}
X(z) &=& \sum_i (q_i - za q_i^{-1})E_{ii} \otimes E_{ii} + (1 - za) \sum_{i\neq j} E_{ii} \otimes E_{jj} \\
&\ & + \sum_{i<j}(q_i - q_i^{-1})E_{ij} \otimes E_{ij} + \sum_{i<j}za(q_i - q_i^{-1}) E_{ji}\otimes E_{ji}.
\end{eqnarray*}
Let $i,j \in I$. When $i \neq j$,
\begin{displaymath}
X(z) (v_i \otimes v_j) = (1 - z a) v_i \otimes v_j,\quad X(z)^{-1} (v_i \otimes v_j) = (1 - z a)^{-1} (v_i \otimes v_j).
\end{displaymath}
When $i = j$,
\begin{displaymath}
X(z) (v_i \otimes v_i) = (q_i - za q_i) v_i \otimes v_i + (q_i - q_i^{-1}) \sum_{s < i} v_s \otimes v_s + za (q_i - q_i^{-1}) \sum_{t>i} v_t \otimes v_t.
\end{displaymath}
Let us introduce the following matrix $\mathcal{M}_{M,N}(z) \in \mathrm{Mat}(M+N, \BC[[z]])$:
\begin{equation} \label{equ: matrix for duals of natural representations}
(\mathcal{M}_{M,N}(z))_{ij} := \begin{cases}
q_j - q_j^{-1} & \textrm{if}\ i < j, \\
q_j - za q_j^{-1} & \textrm{if}\ i = j, \\
za (q_j - q_j^{-1}) & \textrm{if}\ i > j.
\end{cases}
\end{equation}
Let $(x_{ij}(z))_{i,j \in I}$ be the inverse of $\mathcal{M}_{M,N}(z)$. Then
\begin{displaymath}
X(z)^{-1} = \sum_i x_{ii}(z) E_{ii} \otimes E_{ii} + (1-za)^{-1} \sum_{i \neq j} E_{ii} \otimes E_{jj} + \sum_{i \neq j} (-1)^{|j|(|i|+|j|)} x_{ij}(z) E_{ij} \otimes E_{ij}.
\end{displaymath}
As a result, in the $U_q(\Gaff)$-module $(W_{1,a}^{(1)})^{\vee}$, we have
\begin{eqnarray*}
t_{ii}(z) v_1^* = v_1^* \begin{cases} x_{11}(z^{-1}) & \textrm{if}\ i = 1, \\
(1-z^{-1}a)^{-1} & \textrm{if}\ i > 1,
\end{cases} \quad
s_{ii}(z) v_1^* = v_1^* \begin{cases} -z^{-1}a x_{11}(z) & \textrm{if} \ i = 1, \\
-z^{-1}a(1-z^{-1}a)^{-1} &\textrm{if}\ i > 1.
\end{cases}
\end{eqnarray*}
\begin{rem} \label{rem: inverse of matrices and theta_i}
One can write down explicitly the inverse of $\mathcal{M}_{M,N}(z)$:
\begin{equation*}
(\mathcal{M}_{M,N}^{-1}(z))_{ij} = \frac{\theta_i^{-1}\theta_j}{(1-za)(1-zaq^{-2M+2N})} \begin{cases}
q_j^{-1}-q_j &\textrm{if}\ i < j, \\
q_j^{-1}-zaq^{-2M+2N}q_j &\textrm{if}\ i = j, \\
(q_j^{-1}-q_j) za q^{-2M+2N} & \textrm{if}\ i > j
\end{cases}
\end{equation*}
where the $\theta_i$ for $i \in I$ are defined inductively by: $\theta_1 = 1,\ \theta_{i+1} = q_{i+1}^{-1}q_i^{-1} \theta_i$.
\end{rem}
By comparing the highest $\ell$-weight vectors we conclude that
\begin{equation} \label{for: dual of natural representation}
(W_{1,a}^{(1)})^{\vee} \simeq W_{1,a^{-1}q^{2M-2N}}^{(1)}.
\end{equation}
\subsubsection{Dual of $W_{1,a}^{(r)}$.} Let $b \in \BC^{\times}$. It follows from Equation \eqref{for: dual of natural representation}, Theorem \ref{thm: cyclicity of fundamental modules} (II) and Remark \ref{rem: cyclicity vs cocyclicity} that the tensor product $\bigotimes_{j=1}^r W_{1,bq^{-2j}}^{(1)}$ is cogenerated by $v := \bigotimes_{j=1}^r v_j^-$, where $v_j^-$ is a lowest $\ell$-weight vector in $W_{1,bq^{-2j}}^{(1)}$ for $1 \leq j \leq r$. In particular, the submodule $U_q(\Gaff) v$ is simple and it is exactly the socle of $\bigotimes_{j=1}^r W_{1,bq^{-2j}}^{(1)}$. By comparing the lowest $\ell$-weight vectors, we see that
\begin{equation*}
W_{1,bq^{-2r}}^{(r)} \simeq \socle (\bigotimes_{j=1}^{r} W_{1,bq^{-2j}}^{(1)}) = U_q(\Gaff) v.
\end{equation*}
On the other hand, the $U_q(\Gaff)$-module $(\bigotimes_{j=1}^r W^{(1)}_{1,bq^{-2j}})^{\vee}$ is of lowest $\ell$-weight generated by $v^*$ by Remark \ref{rem: cyclicity vs cocyclicity}, and its simple quotient is the twisted dual of $U_q(\Gaff) v$. Making use of Equation \eqref{for: dual of natural representation} we conclude that
\begin{equation} \label{for: dual of fundamental representations}
(W_{1,a}^{(r)})^{\vee} \simeq W_{1,a^{-1}q^{2(M-N-r+1)}}^{(r)}.
\end{equation}
\subsubsection{Dual of $W_{k,a}^{(r)}$.} Let $b \in \BC^{\times}$. By Theorem \ref{thm: cyclicity of fundamental modules}, the tensor product $\bigotimes_{j=1}^k W_{1,bq^{-2j}}^{(r)}$ is of highest $\ell$-weight. We argue in a similar way as in the preceding paragraph (replacing lowest by highest):
\begin{equation} \label{for: KR as socle}
W_{k,bq^{2k}}^{(r)} \simeq \socle (\bigotimes_{j=1}^k W_{1,bq^{2j}}^{(r)}).
\end{equation}
Similar duality argument shows that
\begin{equation} \label{for: dual of KR modules}
(W_{k,a}^{(r)})^{\vee} \simeq W_{k,a^{-1}q^{2(M-N-r+k)}}^{(r)}.
\end{equation}
\subsection{Proof of Theorem \ref{thm: cyclicity of KR modules}.} According to \S \ref{sec: reduction}, it is enough to prove (1). Fix $1 \leq r \leq M$ and $a \in \BC^{\times}$. Let $l_1,l_2,l_3 \in \BZ_{>0}$ be such that $l_1 < l_2 < l_3$. Set $c := q^{2(M-N-r+1)}$. In view of Remark \ref{rem: cyclicity vs cocyclicity} (1) is equivalent to the following assertion:
\begin{itemize}
\item[(2)] $(W_{l_1,a}^{(r)})^{\vee} \otimes (W_{l_2-l_1,aq^{-2l_1}}^{(r)})^{\vee} \otimes (W_{l_3-l_2,aq^{-2l_2}}^{(r)})^{\vee}$ is cogenerated by the tensor product of highest $\ell$-weight vectors.
\end{itemize}
On the other hand, Equations \eqref{for: dual of fundamental representations} and \eqref{for: KR as socle} say that
\begin{eqnarray*}
&\ & (W_{l_1,a}^{(r)})^{\vee} \otimes (W_{l_2-l_1,aq^{-2l_1}}^{(r)})^{\vee} \otimes (W_{l_3-l_2,aq^{-2l_2}}^{(r)})^{\vee} \\
&\simeq & \socle(\bigotimes_{j=0}^{l_1-1} W_{1,ca^{-1}q^{2j}}^{(r)}) \otimes \socle (\bigotimes_{j=l_1}^{l_2-1}W_{1,ca^{-1}q^{2j}}^{(r)} ) \otimes \socle (\bigotimes_{j=l_2}^{l_3-1} W_{1,ca^{-1}q^{2j}}^{(r)}).
\end{eqnarray*}
By Theorem \ref{thm: cyclicity of fundamental modules}, Remark \ref{rem: cyclicity vs cocyclicity} and Equation \eqref{for: dual of fundamental representations}, the big tensor product $\bigotimes_{j=0}^{l_3-1} W_{1,ca^{-1}q^{2j}}^{(r)}$ is cogenerated by the tensor product of highest $\ell$-weight vectors. Hence (2) follows from Remark \ref{rem: cogenerator and tensor products}. \hfill $\Box$
\section{Asymptotic construction of Hernandez-Jimbo} \label{sec: asymptotic representations}
In this section, following the idea of Hernandez-Jimbo in \cite{HJ}, we construct inductive systems of Kirillov-Reshetikhin modules and endow $Y_q(\Glie)$-module structures on their inductive limits with the help of asymptotic algebras.
\subsection{Highest $\ell$-weight modules.} \label{sec: prefundamental modules}
Let $V$ be a $Y_q(\Glie)$-module. Following \S \ref{sec: KR modules}, we say that a non-zero $\super$-homogeneous vector $v \in V$ is a highest $\ell$-weight vector if
\begin{displaymath}
s_{kk}^{(n)} v \in \BC v,\quad s_{ij}^{(n)} v = 0 \quad \textrm{for}\ n \in \BZ_{\geq 0},\ i,j,k \in I,\ i < j.
\end{displaymath}
To the end of this section, we will construct $Y_q(\Glie)$-modules $L_{r,a}^{\pm}$ where $r \in I_0$ and $a \in \BC^{\times}$ such that $L_{r,a}^+$ contains a highest $\ell$-weight vector $v^+$ with
\begin{displaymath}
|v^+| = \even,\quad s_{jj}(z) v^+ = v^+ \begin{cases}
1 - z a & \textrm{if}\ j \leq r, \\
1 & \textrm{if}\ j > r,
\end{cases}
\end{displaymath}
and $L_{r,a}^-$ contains a highest $\ell$-weight vector $v^-$ with
\begin{displaymath}
|v^-| = \even,\quad s_{jj}(z) v^- = v^- \begin{cases}
1 & \textrm{if}\ j \leq r, \\
1-z a & \textrm{if}\ j > r.
\end{cases}
\end{displaymath}
\subsection{Asymptotic algebras.} \label{sec: asymptotic algebras}
We propose in this subsection two versions of asymptotic algebras in the super case. Note that in the non-graded case, asymptotic algebras have been defined for all quantum affine algebras in \cite[\S 2.2]{HJ}.
\begin{defi} \label{def: asymptotic representations}
Define $\tilde{Y}_q(\Glie)$ and $\widehat{Y}_q(\Glie)$ to be the two subalgebras of $Y_q(\Glie)$ generated by $\tilde{s}_{ij}^{(n)}$ and $\widehat{s}_{ij}^{(n)}$ with $i,j \in I, n \in \BZ_{\geq 0}$ respectively. Here
\begin{displaymath}
\tilde{s}_{ij}^{(n)} := s_{ij}^{(n)}(s_{jj}^{(0)})^{-1},\quad \widehat{s}_{ij}^{(n)} := (s_{ii}^{(0)})^{-1}s_{ij} \quad \textrm{for}\ i,j \in I, n \in \BZ_{\geq 0}.
\end{displaymath}
Call $\tilde{Y}_q(\Glie)$ and $\widehat{Y}_q(\Glie)$ asymptotic algebras.
\end{defi}
Indeed, one can write out the full defining relations of asymptotic algebras in terms of the generators $\tilde{s}_{ij}^{(n)},\widehat{s}_{ij}^{(n)}$. Take $\widetilde{Y}_q(\Glie)$ for example. It is the superalgebra defined by
\begin{itemize}
\item[(As1)] generators $\tilde{s}_{ij}^{(n)}$ for $i,j \in I$ and $n \in \BZ_{\geq 0}$;
\item[(As2)] $\super$-grading $|\tilde{s}_{ij}^{(n)}| = |i| + |j|$;
\item[(As3)] defining relations (by taking $\tilde{s}_{ij}(z) := \sum_{n \in \BZ_{\geq 0}} \tilde{s}_{ij}^{(n)} z^n \in \tilde{Y}_q(\Glie)[[z]]$)
\begin{eqnarray*}
&&\begin{array}{rcl}
&& \sum_{a,b \in I} (-1)^{|a|(|b|+|j|)} q^{(\epsilon_i,\epsilon_b - \epsilon_j)}R_{kl,ab}(z,w) \tilde{s}_{ai}(z) \tilde{s}_{bj}(w) \\
&=& \sum_{c,d \in I} (-1)^{|c|(|d|+|l|)} q^{(\epsilon_d,\epsilon_k - \epsilon_c)} R_{cd,ij}(z,w) \tilde{s}_{ld}(w)\tilde{s}_{kc}(z)\ \ \textrm{for}\ i,j,k,l \in I,
\end{array} \\
&&\tilde{s}_{ii}^{(0)} = 1 \ \textrm{for}\ i \in I,\quad \tilde{s}_{ji}^{(0)} = 0 \ \textrm{for}\ 1 \leq i < j \leq M+N.
\end{eqnarray*}
\end{itemize}
Remark also that $\tilde{Y}_q(\Glie)$ and $\widehat{Y}_q(\Glie)$ are $\BQ$-homogeneous subalgebras as their generators are $\BQ$-homogeneous. Endow $\tilde{Y}_q(\Glie)$ and $\widehat{Y}_q(\Glie)$ with the following $\BC[\BP]$-module structures:
\begin{displaymath}
[\alpha] x = q^{(\alpha,\beta)} x \quad \textrm{for}\ \alpha \in \BP, \beta \in \BQ, x \in (\tilde{Y}_q(\Glie))_{\beta} \ \textrm{or}\ x \in (\widehat{Y}_q(\Glie))_{\beta}.
\end{displaymath}
From the Ice Rule \eqref{rel: ice rule} and \S \ref{sec: Yangian} comes the following:
\begin{lem}
As superalgebras, $Y_q(\Glie) \cong \tilde{Y}_q(\Glie) \rtimes \BC[\BP] \cong \widehat{Y}_q(\Glie) \rtimes \BC[\BP]$ where we identify the $[\epsilon_i]$ with the $s_{ii}^{(0)}$ for $i \in I$.
\end{lem}
We adapt the notion of $\BQ$-graded modules over asymptotic algebras in \cite[\S 2.4]{HJ}.
\begin{defi} \label{def: graded modules over asymptotic algebras}
Let $V$ be a module over $\tilde{Y}_q(\Glie)$ (resp. over $\widehat{Y}_q(\Glie)$). We say that $V$ is $\BQ$-graded in the sense of Hernandez-Jimbo if there is a decomposition into a direct sum of sub-vector-superspaces $V = \bigoplus_{\alpha \in \BQ} V^{(\alpha)}$ such that
\begin{displaymath}
\tilde{s}_{ij}^{(n)} V^{(\alpha)} \subseteq V^{(\alpha + \epsilon_i - \epsilon_j)} \quad (\textrm{resp.}\ \widehat{s}_{ij}^{(n)} V^{(\alpha)} \subseteq V^{(\alpha + \epsilon_i - \epsilon_j)})
\end{displaymath}
for $\alpha \in \BQ, i,j \in I$ and $n \in \BZ_{\geq 0}$.
\end{defi}
The following corollary is an application of the lemma above on the semi-direct product constructions of $q$-Yangian. It is parallel to \cite[Proposition 2.4]{HJ}.
\begin{cor} \label{cor: from asymptotic algebras to Yangian}
Let $V = \bigoplus_{\alpha \in \BQ} V^{(\alpha)}$ be a $\BQ$-graded $\tilde{Y}_q(\Glie)$-module (resp. $\widehat{Y}_q(\Glie)$-module) in the sense of Hernandez-Jimbo. Then the module structure over the asymptotic algebra can be extended to that over $Y_q(\Glie)$ by setting
\begin{displaymath}
s_{ii}^{(0)} x = q^{(\epsilon_i,\alpha)} x
\end{displaymath}
for $\alpha \in \BQ, x \in V^{(\alpha)}$ and $i \in I$.
\end{cor}
\subsection{Asymptotic construction of $L_{r,a}^-$ with $1 \leq r \leq M$.} \label{sec: negative asymptotic modules}
In this subsection, we construct the $Y_q(\Glie)$-module $L_{r,a}^{-}$ with $1 \leq r \leq M$ as a limit of the Kirillov-Reshetikhin modules $(W_{k,a}^{(r)}: k \in \BZ_{>0})$ with $\tilde{Y}_q(\Glie)$ the underlying asymptotic algebra. Fix $a \in \BC^{\times}$.
For $k \in \BZ_{>0}$, let $v_k$ be a highest $\ell$-weight vector of the $U_q(\Gaff)$-module $W_{k,a}^{(r)}$. For $k,l \in \BZ_{>0}$ such that $l < k$, let $Z^{(r)}(l<k,a)$ be the $U_q(\Gaff)$-module
\begin{displaymath}
Z^{(r)}(l<k,a) := \phi^*_{((1-z^{-1}a^{-1}q^{2l})^{-1}, (1-zaq^{-2l})^{-1})} W_{k-l,aq^{-2l}}^{(r)}.
\end{displaymath}
Let $v_{lk} \in Z^{(r)}(l<k,a)$ be a highest $\ell$-weight vector.
The following lemma is a direct consequence of Theorem \ref{thm: cyclicity of KR modules} (1).
\begin{lem} \label{lem: cyclicity for inductive system}
Let $l,k \in \BZ_{>0}$ with $l < k$. Then the $U_q(\Gaff)$-module $W_{l,a}^{(r)} \otimes Z^{(r)}(l<k,a)$ is of highest $\ell$-weight. Moreover, its simple quotient is isomorphic to $W_{k,a}^{(r)}$.
\end{lem}
In consequence, for $l < k$, there exists a unique morphism of $U_q(\Gaff)$-modules
\begin{displaymath}
\mathcal{F}_{k,l}: W_{l,a}^{(r)} \otimes Z^{(r)}(l<k,a) \longrightarrow W_{k,a}^{(r)},\quad v_l \otimes v_{lk} \mapsto v_{k}.
\end{displaymath}
Let $F_{k,l}: W_{l,a}^{(r)} \longrightarrow W_{k,a}^{(r)}$ be the restriction: $x \mapsto \mathcal{F}_{k,l}(x \otimes v_{lk})$.
\subsubsection{First properties of the $F_{k,l}$.} As in Remark \ref{rem: cyclicity vs cocyclicity}, the $U_q(\Gaff)$-modules $W_{k,a}^{(r)}$ are $\BP$-graded with respect to the action of the $s_{ii}^{(0)}$. Also, Theorem \ref{thm: BKK Schur-Weyl duality} gives a combinatorial description of dimensions of the weight spaces $(W_{k,a}^{(r)})_{\lambda}$ for $\lambda \in \BP$. Let $(W_{k,a}^{(r)}, \rho^k)$ be the representation associated to the $U_q(\Gaff)$-module $W_{k,a}^{(r)}$.
\begin{prop} \label{prop: first properties of inductive system}
The maps $(F_{kl}: l < k \in \BZ_{>0})$ verify the following properties:
\begin{itemize}
\item[(1)] for $\beta \in \BQ_{\geq 0}$, $F_{k,l}(W_{l,a}^{(r)})_{l\varpi_r - \beta} \subseteq (W_{k,a}^{(r)})_{k\varpi_r - \beta}$;
\item[(2)] $F_{k,l}: W_{l,a}^{(r)} \longrightarrow W_{k,a}^{(r)}$ is injective;
\item[(3)] for $l < k < u \in \BZ_{>0}$, $F_{u,k} F_{k,l} = F_{u,l}$;
\item[(4)] there exists a strictly increasing function $t: \BZ_{>0} \longrightarrow \BZ_{>0}$ such that
\begin{displaymath}
\rho^k(s_{ij}^{(n)}) F_{k,l} (W_{l,a}^{(r)}) \subseteq F_{k,t(l)}(W_{t(l),a}^{(r)})
\end{displaymath}
for all $i,j \in I, n \in \BZ_{\geq 0}$ and $l,k \in \BZ_{>0}$ with $t(l) < k$.
\end{itemize}
\end{prop}
\begin{proof}
(1) comes easily from the action of the $s_{ii}^{(0)}$ on $v_{lk}$.
For (2), consider the $\tilde{s}_{i,i+1}^{(0)}$ with $i \in I_0$. We have
\begin{displaymath}
\Delta (\tilde{s}_{i,i+1}^{(0)}) = \tilde{s}_{i,i+1}^{(0)} \otimes 1 + s_{ii}^{(0)} (s_{i+1,i+1}^{(0)})^{-1} \otimes \tilde{s}_{i,i+1}^{(0)}.
\end{displaymath}
As $\tilde{s}_{i,i+1}^{(0)}v_{lk} = 0$, we see that $F_{k,l}$ respects the action of the $\tilde{s}_{i,i+1}^{(0)}$ with $i \in I_0$. We are left to verify the following: let $\beta \in \BQ_{\geq 0} \setminus \{0\}$ and $x \in (W_{l,a}^{(r)})_{l\varpi_r - \beta}$ be such that $\tilde{s}_{i,i+1}^{(0)} x = 0$ for all $i \in I_0$, then $x = 0$. Note that $W_{l,a}^{(r)} = \ev_a^* L(l\varpi_r)$ with $L(l\varpi_r)$ the simple $\Uc_q(\Glie)$-module of highest weight $l\varpi_r$. Let such $x \in L(l\varpi_r)$ be given (we use the quantum superalgebra $\Uc_q(\Glie)$ and its generators $s_{ij},t_{ij}$). Then $s_{i,i+1} x = 0$ for all $i \in I_0$. It follows from Equation \eqref{rel: relations between e_0 and f_i} and Proposition \ref{prop: evaluation morphism} that $s_{ij} x = 0$ for all $1 \leq i < j \leq M+N$. $x$ is a highest weight vector for the $\Uc_q(\Glie)$-module $L(l\varpi_r)$, which is impossible.
For (3), remark first that the $U_q(\Gaff)$-module $S := W_{l,a}^{(r)} \otimes Z^{(r)}(l<k,a) \otimes Z^{(r)}(k<u,a)$ is of highest $\ell$-weight by Theorem \ref{thm: cyclicity of KR modules} (1). In consequence, $\mathcal{F}_{u,k} \circ (\mathcal{F}_{k,l} \otimes \Id_{Z^{(r)}(k<u,a)})$ is the unique morphism of $U_q(\Gaff)$-modules $F: S \longrightarrow W_{u,a}^{(r)}$ sending $v_l \otimes v_{lk} \otimes v_{ku}$ to $v_u$. On the other hand, that $Z^{(r)}(l<k,a) \otimes Z^{(r)}(k<u,a)$ is of highest $\ell$-weight gives us a surjective morphism of $U_q(\Gaff)$-modules
\begin{displaymath}
G: Z^{(r)}(l<k,a) \otimes Z^{(r)}(k<u,a) \longrightarrow Z^{(r)}(l<u,a),\quad v_{lk} \otimes v_{ku} \mapsto v_{lu}.
\end{displaymath}
It follows from the uniqueness of $F$ that
\begin{displaymath}
\mathcal{F}_{u,l} \circ (\Id_{W_{l,a}^{(r)}}\otimes G) = \mathcal{F}_{u,k} \circ (\mathcal{F}_{k,l} \otimes \Id_{Z^{(r)}(k<u,a)}).
\end{displaymath}
Applying the above equation to $W_{l,a}^{(r)} \otimes v_{lk} \otimes v_{ku}$ we get $F_{u,l} = F_{u,k} \circ F_{k,l}$.
(4) comes from the asymptotic behaviour of dimensions of weight subspaces observed in Corollary \ref{cor: asymptotic behaviour of weight spaces}.
\end{proof}
We remark that in the non-graded case \cite{HJ}, (4) is proved by using a deep property \cite[Proposition 3.2]{Hernandez} of $q$-character concerning tensor product of two vectors (not necessarily two modules). In particular, it was shown \cite[Lemma 4.4]{HJ} that $t(l) = l+1$.
Let us fix $t: \BZ_{>0} \longrightarrow \BZ_{>0}$ in (4). Thanks to (2) the following operators
\begin{displaymath}
F_{k,t(l)}^{-1} \rho^k (\tilde{s}_{ij}^{(n)}) F_{k,l}: W_{l,a}^{(r)} \longrightarrow W_{t(l),a}^{(r)}
\end{displaymath}
are well-defined for $i,j \in I,n \in \BZ_{\geq 0}$ and $k,l \in \BZ_{>0}$ such that $k > t(l)$. Also, $(W_{k,a}^{(r)},F_{kl})$ is an inductive system of vector superspaces. Let $(W_{\infty},F_k)$ be its inductive limit with $F_k: W_{k,a}^{(r)} \longrightarrow W_{\infty}$ the structural maps. As the $F_{k,l}$ are injective, so are the $F_k$.
\subsubsection{Asymptotic properties of the $F_{k,l}$.}
\begin{prop} \label{prop: asymptotic properties of inductive system}
For $i,j \in I, n \in \BZ_{\geq 0}$ and $l \in \BZ_{>0}$, there exist uniquely two linear maps $A_{ij}^{(n)}(l), B_{ij}^{(n)}(l): W_{l,a}^{(r)} \longrightarrow W_{t(l),a}^{(r)}$ such that
\begin{equation} \label{equ: asymptotic equations}
F_{k,t(l)}^{-1} \rho^k (\tilde{s}_{ij}^{(n)}) F_{k,l} = A_{ij}^{(n)}(l) + q^{-2k} B_{ij}^{(n)}(l)
\end{equation}
for all $k > t(l)$.
\end{prop}
\begin{proof}
It is enough to establish the existence of these maps $A_{ij}^{(n)}(l),B_{ij}^{(n)}(l)$, as their uniqueness follows automatically from assumption. For this, we follow essentially the argument of \cite[Proposition 4.5]{HJ}.
By definition (Proposition \ref{prop: evaluation morphism}), $\rho^k(s_{ij}^{(n)}) = 0$ in the following cases: either $(n \geq 2)$, or $(n = 1,i<j)$, or $(n = 0,i > j)$.
\noindent {\textbf{Case A}: $i = j$.} For $\beta \in \BQ_{\geq 0}$ and $x \in (W_{l,a}^{(r)})_{l\varpi_r-\beta}$, we have $F_{k,l}(x) \in (W_{k,a}^{(r)})_{k\varpi_r - \beta}$, meaning
\begin{displaymath}
\rho^k(s_{jj}(z)) F_{k,l}(x) = (q^{(\epsilon_j,k\varpi_r-\beta)} - z a q^{-(\epsilon_j,k\varpi_r-\beta)}) F_{k,l}(x).
\end{displaymath}
It follows that $F_{k,t(l)}^{-1}\rho^k(\tilde{s}_{jj}(z)) F_{k,l}(x) = F_{t(l),l}(x) - q^{-2k(\epsilon_j,\varpi_r)} z a q^{2(\epsilon_j,\beta)} F_{t(l),l}(x)$.
\noindent {\textbf{Case B}: $i < j$.} It is enough to consider $s_{ij}^{(0)}$. Note that
\begin{displaymath}
\Delta (s_{ij}^{(0)}) = \sum_{i'=i}^j s_{ii'}^{(0)} \otimes s_{i'j}^{(0)}.
\end{displaymath}
For $x \in W_{l,a}^{(r)}$, since $s_{i'j}^{(0)} v_{lk} = 0$ for $i' < j$, we get
\begin{eqnarray*}
\rho^k(\tilde{s}_{ij}^{(0)}) F_{k,l}(x) &=& \tilde{s}_{ij}^{(0)}\mathcal{F}_{k,l}(x \otimes v_{lk}) = \mathcal{F}_{k,l}(\tilde{s}_{ij}^{(0)} (x \otimes v_{lk})) \\
&=& \mathcal{F}_{k,l} (\tilde{s}_{ij}^{(0)} x \otimes v_{lk}) = F_{k,l}(\tilde{s}_{ij}^{(0)} x).
\end{eqnarray*}
In other words, $\rho^k (\tilde{s}_{ij}^{(0)})F_{k,l} (W_{l,a}^{(r)}) \subseteq F_{k,l}(W_{l,a}^{(r)})$ and hence $F_{k,t(l)}^{-1}\rho^k(\tilde{s}_{ij}^{(0)}) F_{k,l} = F_{t(l),l} \rho^l (\tilde{s}_{ij}^{(0)})$.
\noindent {\textbf{Case C}: $i > j$.} It is enough to consider $\rho^k(\tilde{s}_{ij}^{(1)}) = -a\rho^k(t_{ij}^{(0)})\rho^k((s_{jj}^{(0)})^{-1})$. Observe furthermore that $s_{j'j}(z) v_{lk} = 0$ if $j > r$ and $j' \neq j$. Hence, when $j > r$, we have
\begin{eqnarray*}
\rho^k(\tilde{s}_{ij}(z)) F_{k,l}(x) &=& \mathcal{F}_{k,l}(\tilde{s}_{ij}(z) (x \otimes v_{lk})) = \mathcal{F}_{k,l}(\tilde{s}_{ij}(z) x \otimes \tilde{s}_{jj}(z) v_{lk})\\
&=& \mathcal{F}_{k,l}(\tilde{s}_{ij}(z)x \otimes v_{lk}) = F_{k,l}(\tilde{s}_{ij}(z) x)
\end{eqnarray*}
for $x \in W_{l,a}^{(r)}$. This says in particular that $F_{k,t(l)}^{-1}\rho^k( \tilde{s}_{ij}^{(1)}) F_{k,l} = F_{t(l),l} \rho^l(\tilde{s}_{ij}^{(1)})$. We are left to consider the case $i > j \leq r$.
{\it Case C.1: $i \leq r$.} Observe that $t_{i'j}^{(0)} v_{lk} = 0$ for $j< i' \leq i$. In view of the coproduct for $t_{ij}^{(0)}$, we find that for $x \in W_{l,a}^{(r)}$:
\begin{eqnarray*}
\rho^k(t_{ij}^{(0)}) F_{k,l}(x) &=& \rho^k(t_{ij}^{(0)})\mathcal{F}_{k,l}(x \otimes v_{lk}) = \mathcal{F}_{k,l}(t_{ij}^{(0)} (x \otimes v_{lk})) \\
&=& \mathcal{F}_{k,l}(t_{ij}^{(0)} x \otimes t_{jj}^{(0)} v_{lk}) = q^{-(k-l)} F_{k,l}(t_{ij}^{(0)} x), \\
\rho^k (s_{jj}^{(0)}) F_{k,l}(x) &=& \mathcal{F}_{k,l}(s_{jj}^{(0)} x \otimes s_{jj}^{(0)} v_{lk}) = q^{k-l} F_{k,l}(s_{jj}^{(0)} x), \\
\end{eqnarray*}
It follows that $F_{k,t(l)}^{-1}\rho^k(s_{ij}^{(1)})F_{k,l} = - a q^{-2(k-l)} F_{t(l),l} \rho^l(t_{ij}^{(0)} (s_{jj}^{(0)})^{-1}) = q^{-2(k-l)}F_{t(l),l} \rho^l(s_{ij}^{(1)})$.
{\it Case C.2: $(i,j) = (r+1,r)$.} Introduce $f_r := (t_{r+1,r+1}^{(0)})^{-1} t_{r+1,r}^{(0)}$. Then in the superalgebra $U_q(\Gaff)$ by Equation \eqref{rel: relations between e_i and f_j}
\begin{displaymath}
[\tilde{s}_{i',i'+1}^{(0)}, f_r] = \delta_{i',r} (q-q^{-1}) (t_{rr}^{(0)}s_{r+1,r+1}^{(0)} - s_{rr}^{(0)} t_{r+1,r+1}^{(0)})
\end{displaymath}
for $i' \in I_0$. In order to establish Equation \eqref{equ: asymptotic equations}, it is enough to prove the following assertions AS($\beta$) for all $\beta \in \BQ_{\geq 0}$.
\begin{itemize}
\item[AS($\beta$)]: for $x \in (W_{l,a}^{(r)})_{l\varpi_r-\beta}$, there exist $x',x'' \in (W_{t(l),a}^{(r)})_{t(l)-\beta}$ such that for all $k > t(l)$, $F_{k,t(l)}^{-1}\rho^k(f_r) F_{k,l}(x) = q^k x' + q^{-k} x''$.
\end{itemize}
For $\beta = \sum_{i'\in I_0} n_{i'} \alpha_{i'} \in \BQ_{\geq 0}$, let $\ell(\beta) := \sum_{i'\in I_0} n_{i'}$ be the height of $\beta$. We argue by induction on $\ell(\beta)$. First suppose that $\beta = 0$. Then up to scalar $x = v_l$. For $k > t(l)$, define $w_k := F_{k,t(l)}^{-1}\rho^k( f_r) F_{k,l} (v_l) \in W_{t(l),a}^{(r)}$. For $i' \in I_0$, we compute
\begin{eqnarray*}
\rho^{t(l)}(\tilde{s}_{i',i'+1}^{(0)}) w_k &=& \rho^{t(l)}(\tilde{s}_{i',i'+1}^{(0)}) F_{k,t(l)}^{-1} \rho^k(f_r) F_{k,l}(v_l) = F_{k,t(l)}^{-1} \rho^k(\tilde{s}_{i',i'+1}^{(0)} f_r) (v_k) \\
&=& \delta_{i',r} (q-q^{-1}) F_{k,t(l)}^{-1} \rho^k (t_{rr}^{(0)}s_{r+1,r+1}^{(0)} - s_{rr}^{(0)}t_{r+1,r+1}^{(0)}) (v_k) \\
&=& \delta_{i',r} (q-q^{-1})(q^{-k}-q^k) v_{t(l)}.
\end{eqnarray*}
Here the second equality comes from Case B and the definition of $F_{k,l}$. We have therefore found $C_{i'},D_{i'} \in (W_{t(l),a}^{(r)})_{t(l)\varpi_r}$ such that $\rho^{t(l)}(\tilde{s}_{i',i'+1}^{(0)}) w_k = q^k C_{i'} + q^{-k} D_{i'}$ for all $k > t(l)$ and $i' \in I_0$. It follows that
\begin{displaymath}
\rho^{t(l)}(\tilde{s}_{i',i'+1}^{(0)}) (q^{k+2}w_k - q^{k+1}w_{k+1}-q^{k+3}w_{k+1}+q^{k+2}w_{k+2}) = 0
\end{displaymath}
for all $i' \in I_0$. As we see in the proof of Proposition \ref{prop: first properties of inductive system}, this says that
\begin{displaymath}
q^{k+2}w_k - q^{k+1}w_{k+1}-q^{k+3}w_{k+1}+q^{k+2}w_{k+2} = 0
\end{displaymath}
for all $k > t(l)$, since $w_k \in (W_{t(l),a}^{(r)})_{t(l)\varpi_r - \alpha_r}$. In matrix form, this reads
\begin{displaymath}
\begin{pmatrix}
w_{k+2} \\
w_{k+1}
\end{pmatrix} = \begin{pmatrix}
q+q^{-1} & -1 \\
1 & 0
\end{pmatrix} \begin{pmatrix}
w_{k+1} \\
w_k
\end{pmatrix}.
\end{displaymath}
The $2\times 2$ matrix above is diagonalizable with eigenvalues $q,q^{-1}$. We can therefore find $x',y' \in (W_{t(l),a}^{(r)})_{t(l)\varpi_r-\alpha_r}$ which are linear combinations of $w_{t(l)+1},w_{t(l)+2}$ such that
\begin{displaymath}
w_k = q^k x' + q^{-k'} x''
\end{displaymath}
for all $k > t(l)$. This proves the assertion AS(0). Suppose $m > 0$ and that the assertions AS($\beta$) for $\ell(\beta) < m$ have been proved. Fix $\beta \in \BQ_{\geq 0}$ with $\ell(\beta) = m$. Let $x \in (W_{l,a}^{(r)})_{l\varpi_r - \beta}$. As before, take $w_k := F_{k,t(l)}^{-1} \rho^k(f_r) F_{k,l}(x) \in (W_{t(l),a}^{(r)})_{t(l)\varpi_r - \beta- \alpha_r}$.
\begin{eqnarray*}
\rho^{t(l)}(\tilde{s}_{i',i'+1}^{(0)}) w_k &=& \rho^{t(l)}(\tilde{s}_{i',i'+1}^{(0)}) F_{k,t(l)}^{-1} \rho^k(f_r) F_{k,l}(x) = F_{k,t(l)}^{-1} \rho^k(\tilde{s}_{i',i'+1}^{(0)} f_r) F_{k,l}(x) \\
&=& \delta_{i',r}(q-q^{-1})(q^{-k+(\beta,\alpha_r)} - q^{k-(\beta,\alpha_r)}) F_{t(l),l}(x) \\
&\ & + (-1)^{(|i'|+|i'+1|)|r+1|} F_{k,t(l)}^{-1} \rho^k(f_r) F_{k,l} \rho^l(s_{i',i'+1}^{(0)})(x).
\end{eqnarray*}
For the last term, note that $\rho^l(s_{i',i'+1}^{(0)})(x) \in (W_{l,a}^{(r)})_{l\varpi_r - \beta + \alpha_{i'}}$. Hence the assertion AS($\beta-\alpha_{i'}$) applies. We can find $C_{i'},D_{i'} \in (W_{t(l),a}^{(r)})_{t(l)\varpi_r - \beta + \alpha_{i'} - \alpha_r}$ such that
\begin{displaymath}
\rho^{t(l)}(\tilde{s}_{i',i'+1}^{(0)}) w_k = q^k C_{i'} + q^{-k} D_{i'}
\end{displaymath}
for all $k > t(l)$. The rest is parallel.
{\it Case C.3: $i > r + 1$ and $j = r$.} This comes from Case C.2 and Equation \eqref{rel: relations between e_0 and f_i}:
\begin{displaymath}
t_{r+1,r}^{(0)} t_{i,r+1}^{(0)} - t_{i,r+1}^{(0)} t_{r+1,r}^{(0)} = (q-q^{-1}) t_{r+1,r+1}^{(0)} t_{ir}^{(0)}.
\end{displaymath}
Remark that $(|r+1|+|r|)(|i|+|r+1|) = \even$, so there is no sign on the left hand side.
{\it Case C.4: $j < r < i$.} This comes from Cases C.1-C.3 and the following relation in $U_q(\Gaff)$:
\begin{displaymath}
t_{rj}^{(0)} t_{ir}^{(0)} - t_{ir}^{(0)} t_{rj}^{(0)} = (q-q^{-1}) t_{rr}^{(0)}t_{ij}^{(0)}.
\end{displaymath}
This concludes the proof.
\end{proof}
From the above proof, we see that: if $i < j$ then $A_{ij}^{(n)}(l) = F_{t(l),l} \rho^l(\tilde{s}_{ij}^{(n)})$ and $A_{ij}^{(n)}(l) = 0$ for $n > 0$; if $j < i \leq r$ then $A_{ij}^{(n)}(l) = 0$; if $j < i$ then $A_{ij}^{(n)}(l) = 0$ for $n \neq 1$.
\subsubsection{$\tilde{Y}_q(\Glie)$-module structure on $W_{\infty}$.} \label{sec: asymptotic construction}
Since $t$ is strictly increasing, we get another inductive system of vector superspaces $(W_{t(k),a}^{(r)}, F_{f(k),t(l)})$ over the linearly ordered set $\BZ_{> 0}$. Furthermore, the following defines a morphism of inductive systems
\begin{displaymath}
F_{t(k),k}: W_{k,a}^{(r)} \longrightarrow W_{t(k),a}^{(r)}
\end{displaymath}
which induces an isomorphism of inductive limits. Identify their inductive limits.
Let us show that $(A_{ij}^{(n)}(k): W_{k,a}^{(r)} \longrightarrow W_{t(k),a}^{(r)})$ gives a morphism of inductive system of vector superspaces. In other words,
\begin{displaymath}
F_{t(l+1),t(l)} A_{ij}^{(n)}(l) = A_{ij}^{(n)}(l+1) F_{l+1,l} : W_{l,a}^{(r)} \longrightarrow W_{t(l+1),a}^{(r)}.
\end{displaymath}
Indeed, for all $k > t(l+1)$, we have
\begin{eqnarray*}
&\ & F_{t(l+1),t(l)}A_{ij}^{(n)}(l) + q^{-2k} F_{t(l+1),t(l)} B_{ij}^{(n)}(l) = F_{t(l+1),t(l)} F_{k,t(l)}^{-1} \rho^k(\tilde{s}_{ij}^{(n)})F_{k,l} \\
&=& F_{k,t(l+1)}^{-1} \rho^k(\tilde{s}_{ij}^{(n)}) F_{k,l+1}F_{l+1,l} = A_{ij}^{(n)}(l+1) F_{l+1,l} + q^{-2k} B_{ij}^{(n)}(l+1)F_{l+1,l}.
\end{eqnarray*}
The desired equations follow. We obtain therefore linear endomorphisms of $W_{\infty}$:
\begin{displaymath}
A_{ij}^{(n)} := \lim_{\rightarrow} A_{ij}^{(n)}(l) : W_{\infty} \longrightarrow W_{\infty}.
\end{displaymath}
Moreover, $A_{ij}^{(n)}$ is $\super$-homogeneous of degree $|i|+|j|$ in view of Equation \eqref{equ: asymptotic equations}.
Now, as in \cite[\S 4]{HJ}, the assignment $\tilde{s}_{ij}^{(n)} \mapsto A_{ij}^{(n)}$ defines a representation of the superalgebra $\tilde{Y}_q(\Glie)$ on the vector superspace $W_{\infty}$. In other words, the $A_{ij}^{(n)}$ respect the condition (As3) in \S \ref{sec: asymptotic algebras}. Clearly, $A_{ij}^{(0)} = 0$ for $i > j$. By considering the operators
\begin{displaymath}
F_{k,t^2(l)}^{-1}\rho^k(\tilde{s}_{ij}^{(n)} \tilde{s}_{i'j'}^{(n')}) F_{k,l} = (F_{k,t^2(l)}^{-1}\rho^k(\tilde{s}_{ij}^{(n)}) F_{t^2(l),t(l)} )(F_{k,t(l)}^{-1}\rho^k(\tilde{s}_{i'j'}^{(n')})F_{k,l})
\end{displaymath}
and by using Equation \eqref{equ: asymptotic equations}, we see that (As3) is indeed true for the $A_{ij}^{(n)}$.
\subsubsection{$Y_q(\Glie)$-module structure on $W_{\infty}$.} Proposition \ref{prop: first properties of inductive system} (1) implies that $W_{\infty}$ is endowed with a $\BQ$-grading: for $\alpha \in \BQ$, $x \in (W_{\infty})^{(\alpha)}$ if there exists $k \in \BZ_{>0}$ such that $x \in F_k(W_{k,a}^{(r)})_{\alpha+k\varpi_r}$. Now Equation \eqref{equ: asymptotic equations} says that this is a $\BQ$-grading in the sense of Hernandez-Jimbo. In conclusion, $W_{\infty}$ becomes a $Y_q(\Glie)$-module thanks to Corollary \ref{cor: from asymptotic algebras to Yangian}.
From the proof of Proposition \ref{prop: asymptotic properties of inductive system}, we see that for $\alpha \in \BQ, x \in (W_{\infty})^{(\alpha)}$ and $j \in I$,
\begin{displaymath}
|x| = |\alpha|,\quad s_{jj}(z) x = x \begin{cases}
q^{(\epsilon_j,\alpha)} & (j \leq r), \\
q^{(\epsilon_j,\alpha)} - z a q^{-(\epsilon_j,\alpha)} & (j > r).
\end{cases}
\end{displaymath}
Moreover, $W_{\infty}$ contains a highest $\ell$-weight vector $v_{\infty} := F_k v_k$ with the same action of the $s_{ii}(z)$ as that of the highest $\ell$-weight vector $v^- \in L_{r,a}^-$ in \S \ref{sec: prefundamental modules}.
For this reason, set $L_{r,a}^-$ to be the $Y_q(\Glie)$-module $W_{\infty}$ thus obtained.
\subsection{Asymptotic construction of $L_{r,a}^+$ with $1 \leq r \leq M$.}\label{sec: positive asymptotic modules} This is parallel to the construction of $L_{r,a}^-$. We start from an inductive system of Kirillov-Reshetikhin modules, establish stability (Proposition \ref{prop: first properties of inductive system}) and asymptotic (Proposition \ref{prop: asymptotic properties of inductive system}) properties for this inductive system, and endow $Y_q(\Glie)$-module structure on the inductive limit through an asymptotic algebra. As the proofs of these properties are identical to the case of $L_{r,a}^-$, we omit them in this subsection.
The Kirillov-Reshetikhin modules involved will be $W_{k,aq^{2k}}^{(r)}$ for $k \in \BZ_{>0}$. For $l < k$, set
\begin{displaymath}
Z_{kl} := \phi^*_{((1-z^{-1}a^{-1}q^{-2l})^{-1},(1-zaq^{2l})^{-1})} W_{k-l,aq^{2k}}^{(r)}.
\end{displaymath}
Let $v_k \in W_{k,aq^{2k}}^{(r)}$ (resp. $v_{kl} \in Z_{kl}$) be a highest $\ell$-weight vector.
For $l<k$, the $U_q(\Gaff)$-module $Z_{kl} \otimes W_{l,aq^{2l}}^{(r)}$ is of highest $\ell$-weight with simple quotient isomorphic to $W_{k,aq^{2k}}^{(r)}$. This affords a unique morphism of $U_q(\Gaff)$-modules
\begin{displaymath}
\mathcal{F}_{k,l}: Z_{kl} \otimes W_{l,aq^{2l}}^{(r)} \longrightarrow W_{k,aq^{2k}}^{(r)},\quad v_{kl} \otimes v_l \mapsto v_k.
\end{displaymath}
Let $F_{k,l}: W_{l,aq^{2l}}^{(r)} \longrightarrow W_{k,aq^{2k}}^{(r)}$ be the restriction: $x \mapsto \mathcal{F}_{k,l}(v_{kl} \otimes x)$. Then the $F_{k,l}$ verify all the properties in Proposition \ref{prop: first properties of inductive system} with $W_{k,a}^{(r)}$ replaced by $W_{k,aq^{2k}}^{(r)}$ everywhere. Furthermore, a detailed analysis as in the proof of Proposition \ref{prop: asymptotic properties of inductive system} shows that: for all $i,j \in I, n \in \BZ_{\geq 0}$ and $l \in \BZ_{>0}$, there exist linear operators $A_{ij}^{(n)}(l),B_{ij}^{(n)}(l): W_{l,aq^{2l}}^{(r)} \longrightarrow W_{t(l),aq^{2t(l)}}^{(r)}$ such that
\begin{equation} \label{equ: asymptotic equation 2}
F_{k,t(l)}^{-1}\rho^k(\widehat{s}_{ij}^{(n)}) F_{k,l} = A_{ij}^{(n)}(l) + q^{2k} B_{ij}^{(n)}(l) \quad \textrm{for\ all}\ k > t(l).
\end{equation}
As before, the $A_{ij}^{(n)}(l)$ give rise to a morphism of inductive systems of vector superspaces $(W_{k,aq^{2k}}^{(r)},F_{k,l})$ and $(W_{t(k),aq^{2t(k)}}^{(r)},F_{t(k),t(l)})$. We obtain therefore a representation of the asymptotic algebra $\widehat{Y}_q(\Glie)$ on the inductive limit:
\begin{displaymath}
(W_{r,a}^+,F_k) := \lim_{\rightarrow} (W_{k,aq^{2k}}^{(r)},F_{k,l}),\quad \widehat{s}_{ij}^{(n)} \mapsto \lim_{\rightarrow} A_{ij}^{(n)}(l).
\end{displaymath}
Also $W_{r,a}^+$ is $\BQ$-graded in the sense of Hernandez-Jimbo: for $\alpha \in \BQ$, $x \in (W_{r,a}^+)^{(\alpha)}$ if $x \in F_k (W_{k,aq^{2k}}^{(r)})_{\alpha+k\varpi_r}$ for some $k \in \BZ_{>0}$. In this way, we get a representation of $Y_q(\Glie)$ on $W_{r,a}^+$ by Corollary \ref{cor: from asymptotic algebras to Yangian}. For $\alpha \in \BQ, x \in (W_{r,a}^+)^{(\alpha)}$ and $j \in I$,
\begin{displaymath}
|x| = |\alpha|,\quad s_{jj}(z) x = x \begin{cases}
q^{(\alpha,\epsilon_j)} - z a q^{-(\alpha,\epsilon_j)} & (j \leq r), \\
q^{(\alpha,\epsilon_j)} & (j > r).
\end{cases}
\end{displaymath}
The vector $v_{\infty} := F_k(v_k) \in W_{r,a}^+$ verifies the same conditions as $v^+ \in L_{r,a}^+$ in \S \ref{sec: prefundamental modules}. Let $L_{r,a}^+$ be the $Y_q(\Glie)$-module $W_{r,a}^+$ thus obtained.
\subsection{Construction of $L_{r,a}^{\pm}$ for $M+1 \leq r \leq M+N-1$.} \label{sec: odd asymptotic modules}
For this purpose, let $\Glie' = \mathfrak{gl}(N,M)$ be as in \S \ref{sec: reduction}. Recall that we have defined the quantum affine superalgebra $U_q(\Gafft)$ by the RTT generators $s_{ij;J}^{(n)}, t_{ij;J}^{(n)}$ as in \S \ref{sec: reduction}. Let $Y_q(\Glie')$ be the subalgebra of $U_q(\Gafft)$ generated by the $s_{ij;J}^{(n)}$. Then the isomorphism $f_{J,I}$ defined by Formula \eqref{for: transposition of quantum affine superalgebras} restricts to an isomorphism of Hopf superalgebras which we write as $f_{J,I}$
\begin{displaymath}
f_{J,I}: Y_q(\Glie')^{\mathrm{cop}} \longrightarrow Y_q(\Glie),\quad s_{ij;J}^{(n)} \mapsto \varepsilon_{ji}^J s_{\overline{j}\overline{i}}^{(n)}.
\end{displaymath}
We construct as in \S \ref{sec: negative asymptotic modules}-\ref{sec: positive asymptotic modules} the $Y_q(\Glie')$-modules $L_{r,a;J}^{\pm}$ as limits of corresponding Kirillov-Reshetikhin modules for $1 \leq r \leq N$ and $a \in \BC^{\times}$. Now the $Y_q(\Glie)$-modules $L_{r,a}^{\pm}$ for $M+1 \leq r \leq M+N-1$ and $a \in \BC^{\times}$ are realized as:
\begin{displaymath}
L_{r,a}^{\pm} \cong (f_{J,I}^{-1})^* L_{M+N-r,a;J}^{\mp}.
\end{displaymath}
Consider $L_{r,a}^+$ with $M+1\leq r \leq M+N-1$ for example. For $l<k$, define
\begin{displaymath}
Z_{kl} := \phi^*_{((1-z^{-1}a^{-1}q^{2l})^{-1}, (1-zaq^{-2l})^{-1})} W_{k-l,aq^{-2l}}^{(r)}.
\end{displaymath}
Then the tensor product $Z_{kl} \otimes W_{l,a}^{(r)}$ is of highest $\ell$-weight. This affords an inductive system $(W_{k,a}^{(r)}, F_{k,l})$ where $F_{k,l}: W_{l,a}^{(r)} \longrightarrow W_{k,a}^{(r)}$ comes from the surjective map
\begin{displaymath}
\mathcal{F}_{k,l}: Z_{kl} \otimes W_{l,a}^{(r)} \longrightarrow W_{k,a}^{(r)}.
\end{displaymath}
The rest is completely parallel to \S \ref{sec: positive asymptotic modules}.
Following \cite{FH}, the $Y_q(\Glie)$-modules $L_{r,a}^{\pm}$ are called {\it positive/negative prefundamental modules}. We shall see that they are always simple. Contrary to the non-graded case where all prefundamental modules are infinite-dimensional, $L_{r,a}^{\pm}$ is finite-dimensional if and only if $r = M$. This says that there are \lq\lq more\rq\rq\ finite-dimensional simple $Y_q(\Glie)$-modules than finite-dimensional simple $U_q(\Gaff)$-modules, as we have seen in \cite[\S 5]{Z2} on representation theory of $Y_q(\mathfrak{gl}(1,1))$. See also Proposition \ref{prop: finite-dimensional and extension}.
\subsection{Relation with Tsuboi's work.} The $Y_q(\Glie)$-modules $L_{r,a}^{\pm}$ have been constructed by Tsuboi in a different way. In \cite{Tsuboi}, Tsuboi proposed the notion of a {\it contracted quantum superalgebra}. This is the superalgebra defined by generators $\dot{s}_{ii},\dot{s}_{ii}^{-1},\dot{t}_{ii},\dot{s}_{jk},\dot{t}_{kj}$ for $i,j,k \in I, j < k $ with $\super$-degrees
\begin{displaymath}
|\dot{s}_{ii}| = |\dot{s}_{ii}^{-1}| = |\dot{t}_{ii}| = \even,\quad |\dot{s}_{jk}| = |\dot{t}_{kj}| = |j| + |k|
\end{displaymath}
subject to the following relations (take $\dot{T} = \sum_{jk} \dot{t}_{jk} \otimes E_{jk}, \dot{S} = \sum_{jk} \dot{s}_{jk} \otimes E_{jk}$)
\begin{align*}
& R_{23} \dot{T}_{12}\dot{T}_{13} = \dot{T}_{13}\dot{T}_{12}R_{23},\quad R_{23}\dot{S}_{12}\dot{S}_{13} = \dot{S}_{13}\dot{S}_{12}R_{23} \\
& R_{23}\dot{T}_{12}\dot{S}_{13} = \dot{S}_{13}\dot{T}_{12}R_{23},\quad \dot{s}_{ii} \dot{s}_{ii}^{-1} = 1 = \dot{s}_{ii}^{-1}\dot{s}_{ii}.
\end{align*}
Let $\dot{\Uc}_q(\Glie)$ be the superalgebra obtained. Then the proof of Proposition \ref{prop: evaluation morphism} implies that
\begin{displaymath}
\dot{\ev}_a: Y_q(\Glie) \longrightarrow \dot{\Uc}_q(\Glie),\quad s_{ij}(z) \mapsto \dot{s}_{ij} - z a \dot{t}_{ij},\quad (s_{ii}^{(0)})^{-1} \longrightarrow \dot{s}_{ii}^{-1}
\end{displaymath}
defines a morphism of superalgebras for $a \in \BC^{\times}$. Let $\dot{L}_r^+$ be the simple $\dot{\Uc}_q(\Glie)$-module generated by a highest weight vector $v^+$:
\begin{displaymath}
|v^+| = \even,\quad \dot{s}_{jk} v^+ = 0,\quad \dot{s}_{ii} v^+ = v^+,\quad \dot{t}_{ii} v^+ = v^+ \begin{cases}
1 & (i \leq r), \\
0 & (i > r).
\end{cases}
\end{displaymath}
Similarly, let $\dot{L}_r^-$ be the simple $\dot{\Uc}_q(\Glie)$-module generated by a highest weight vector $v^-$:
\begin{displaymath}
|v^-| = \even,\quad \dot{s}_{jk} v^- = 0,\quad \dot{s}_{ii}v^- = v^-,\quad \dot{t}_{ii}v^- = v^- \begin{cases}
0 & (i \leq r),\\
1 & (i > r).
\end{cases}
\end{displaymath}
Then, as $Y_q(\Glie)$-modules, $L_{r,a}^{\pm} \cong \dot{\ev}_a^* \dot{L}_r^{\pm}$. Tsuboi constructed the $\dot{L}_r^{\pm}$ via the $q$-oscillator realizations of the contracted quantum superalgebra $\dot{\Uc}_q(\Glie)$, which is a generalization of the construction carried out in \cite{BT} for $U_q(\widehat{\mathfrak{sl}(2,1)})$ to the case $U_q(\widehat{\mathfrak{gl}(M,N)})$.
Compared with Tsuboi's oscillation realizations, our results in \S \ref{sec: negative asymptotic modules}-\ref{sec: positive asymptotic modules} can be rephrased as asymptotic realizations of the $\dot{\Uc}_q(\Glie)$-modules $\dot{L}_r^{\pm}$.
\section{Generic asymptotic construction} \label{sec: generic asymptotic representations}
In this section, based on the inductive systems of Kirillov-Reshetikhin modules and their stability and asymptotic properties in he previous section, we propose a new asymptotic construction which realizes the inductive limits as modules over the full quantum affine superalgebra $U_q(\Gaff)$ instead of the $q$-Yangian $Y_q(\Glie)$.
Let us fix two parameters $a,b \in \BC^{\times}$ and a Dynkin node $r \in I_0$. To the end of this section, we shall construct a $U_q(\Gaff)$-module, written as $L_{r,a}(b)$, which has a non-zero vector $v$ of $\super$-degree $\even$ such that
\begin{displaymath}
s_{ij}^{(n)} v = t_{ij}^{(n)} v = 0 \quad \textrm{for}\ n \in \BZ_{\geq 0}, 1 \leq i < j \leq M+N
\end{displaymath}
and in the case $1 \leq r \leq M$
\begin{equation} \label{equ: even generic asymptotic modules hwt vector}
s_{ii}(z) v = v \begin{cases}
b - z a b & (i \leq r), \\
1-zab^2 & (i > r),
\end{cases} \quad t_{ii}(z) v = v \begin{cases}
b^{-1} - z^{-1} a^{-1} b^{-1} & (i \leq r), \\
1 - z^{-1} a^{-1} b^{-2} & (i > r),
\end{cases}
\end{equation}
whereas in the case $M+1 \leq r < M+N$
\begin{equation} \label{equ: odd generic asymptotic modules hwt vector}
s_{ii}(z) v = v \begin{cases}
1 - z a & (i \leq r), \\
b -zab^{-1} & (i > r),
\end{cases} \quad t_{ii}(z) v = v \begin{cases}
1 - z^{-1} a^{-1} & (i \leq r), \\
b^{-1} - z^{-1} a^{-1} b & (i > r).
\end{cases}
\end{equation}
\subsection{Construction of $L_{r,a}(b)$ with $1 \leq r \leq M$.} \label{sec: even generic asymptotic modules}
Let $V_k := W_{k,aq^{2k}}^{(r)}$ and fix $v_k \in V_k$ a vector of highest $\ell$-weight for $k \in \BZ_{>0}$. Let $(F_{k,l}: V_l \longrightarrow V_k)_{l<k}$ be the inductive system of vector superspaces constructed in \S \ref{sec: positive asymptotic modules}. Let $(V_{\infty}, F_l: V_l \longrightarrow V_{\infty})$ be its inductive limit. Define $v_{\infty} := F_1 (v_1) \in V_{\infty}$. Then from $F_{k,l}(v_l) = v_k$ we see that $F_l (v_l) = v_{\infty}$ for all $l \in \BZ_{>0}$.
Let $S$ be the subset of $U_q(\Gaff)$ consisting of the RTT generators $s_{ij}^{(n)}, t_{ij}^{(n)}$ for $i,j \in I, n \in \BZ_{\geq 0}$. Let $\rho^k$ be the representation of $U_q(\Gaff)$ on $V_k$ for $k \in \BZ_{>0}$. Let $t: \BZ_{>0} \longrightarrow \BZ_{>0}$ be a strictly increasing function such that (Proposition \ref{prop: first properties of inductive system})
\begin{displaymath}
\rho^k (s_{ij}^{(n)}) F_{k,l} V_l \subseteq F_{k,t(l)} V_{t(l)} \quad \textrm{for}\ k > t(l).
\end{displaymath}
By definition of $V_k = \ev_{aq^{2k}}^* L(k\varpi_r)$ we see that
\begin{displaymath}
\rho^k(t_{ij}(z)) = - z^{-1}a^{-1} q^{-2k} \rho^k (s_{ij}(z)) \in \End V_k [z^{-1}].
\end{displaymath}
It follows that $\rho^k(s) F_{k,l} V_l \subseteq F_{k,t(l)} V_{t(l)}$ for $k > t(l)$ and $s \in S$. Since the $F_{k,l}$ are injective, the operators
\begin{displaymath}
F_{k,t(l)}^{-1} \rho^k(s) F_{k,l} : V_l \longrightarrow V_{t(l)}
\end{displaymath}
for $s \in S, k > t(l)$ are well-defined.
\begin{lem} \label{lem: Laurent polynomiality}
Let $l \in \BZ_{>0}$ and $s \in S$. Then there exists uniquely a $\textrm{Hom}(V_l,V_{t(l)})$-valued Laurent polynomial $P_{l,s}(z) \in \textrm{Hom}(V_l,V_{t(l)})[z,z^{-1}]$ such that
\begin{displaymath}
F_{k,t(l)}^{-1} \rho^k(s) F_{k,l} = P_{l,s}(z)|_{z = q^k}\quad \textrm{for}\ k > t(l).
\end{displaymath}
Furthermore, the coefficients of $z^n$ in $P_{l,s}(z)$ are non-zero only if $-2 \leq n \leq 3$.
\end{lem}
\begin{proof}
Assume without loss of generality $s = s_{ij}^{(n)}$. Since $\rho^k(s_{ii}^{(0)}) F_{k,t(l)} V_{t(l)} \subseteq F_{k,t(l)}V_{t(l)}$
\begin{displaymath}
F_{k,t(l)}^{-1} \rho^k(s) F_{k,l} = F_{k,t(l)}^{-1}\rho^k(s_{ii}^{(0)} \widehat{s}_{ij}^{(n)}) F_{k,l} = (F_{k,t(l)}^{-1} \rho^k(s_{ii}^{(0)})F_{k,t(l)}) (F_{k,t(l)}^{-1}\rho^k(\widehat{s}_{ij}^{(n)}) F_{k,l}).
\end{displaymath}
On the other hand, by definition of the $F_{k,l}$
\begin{displaymath}
F_{k,t(l)}^{-1} \rho^k(s_{ii}^{(0)}) F_{k,t(l)} = q^{(k-t(l)) (\varpi_r, \epsilon_i)} \rho^{t(l)}(s_{ii}^{(0)}).
\end{displaymath}
Combining with Equation \eqref{equ: asymptotic equation 2}, we find a Laurent polynomial $P_{l,s}(z)$ with the desired property. Clearly such $P_{l,s}$ is unique.
\end{proof}
Now we argue as in \S \ref{sec: asymptotic construction}. By using the defining property of the $P_{l,s}$ we see that for all $s \in S$ and $l \in \BZ_{>0}$:
\begin{displaymath}
F_{t(l+1),t(l)} P_{l,s}(z) = P_{l+1,s}(z) F_{l+1,l} \in \textrm{Hom}(V_l, V_{t(l+1)})[z,z^{-1}].
\end{displaymath}
In other words, if we write
\begin{displaymath}
P_{l,s}(z) := \sum_{n=-2}^3 P_{l,s}[n] z^n,\quad P_{l,s}[n] \in \textrm{Hom}(V_l,V_{t(l)}),
\end{displaymath}
then for all $-2 \leq n \leq 3$, $P_{l,s}[n]$ induces a morphism of inductive systems of vector spaces:
\begin{displaymath}
(P_{l,s}[n])_l: (V_l, F_{k,l}) \longrightarrow (V_{t(l)}, F_{t(k),t(l)}).
\end{displaymath}
Let $P_s[n]$ be the inductive limit of the above morphism. Then $P_s[n] \in \End (V_{\infty})$ as both inductive systems give rise to the same inductive limit. As a result, the following assignment
\begin{displaymath}
s \mapsto \sum_{n=-2}^3 P_s[n] b^n \in \End (V_{\infty})\quad \textrm{for}\ s \in S
\end{displaymath}
defines a representation of the quantum affine superalgebra $U_q(\Gaff)$ on $V_{\infty}$. Let us compute the action of $s_{ij}^{(n)},t_{ij}^{(n)}$ on $v_{\infty}$ for $1 \leq i \leq j \leq n$. First, if $i < j$,
\begin{displaymath}
F_{k,t(l)}^{-1} \rho^k(s^{(n)}_{ij}) F_{k,l} v_l = F_{k,t(l)}^{-1} \rho^k(t_{ij}^{(n)}) F_{k,l} v_l = 0.
\end{displaymath}
In other words, $P_{l,s_{ij}^{(n)}}(z) v_l = 0 = P_{l,t_{ij}^{(n)}}(z) v_l$. Henceforth $s_{ij}^{(n)} v_{\infty} = t_{ij}^{(n)} v_{\infty} = 0$. Next, assume $i = j$, in view of the following equation
\begin{eqnarray*}
F_{k,t(l)}^{-1} \rho^k(s_{ij}(z)) F_{k,l} v_l &=& v_{t(l)} \begin{cases}
q^k - z a q^k & (i \leq r), \\
1 - za q^{2k} & (i > r),
\end{cases}\\
F_{k,t(l)}^{-1} \rho^k(t_{ij}(z)) F_{k,l} v_l &=& v_{t(l)} \begin{cases}
q^{-k} - z^{-1}a^{-1}q^{-k} & (i \leq r), \\
1 - z^{-1}a^{-1}q^{-2k} & (i > r).
\end{cases}
\end{eqnarray*}
In a similar way, we get $s_{ii}(z) v_{\infty}$ and $t_{ii}(z) v_{\infty}$ by regarding $q^k$ in the above equation as $b$. The $U_q(\Gaff)$-module $V_{\infty}$ is the desired $L_{r,a}(b)$, as $v_{\infty}$ verifies Equation \eqref{equ: even generic asymptotic modules hwt vector}.
\subsection{Construction of $L_{r,a}(b)$ with $r > M$.} \label{sec: odd generic asymptotic modules}
In this case, by abuse of language, set $V_{k} := W_{k,a}^{(r)}$. Following \S \ref{sec: odd asymptotic modules}, let $v_k \in V_k$ be a highest $\ell$-weight vector and let $F_{k,l}: V_l \longrightarrow V_k$ be the structural maps of the inductive system of vector superspaces $(V_k, F_{k,l})$ so that $F_{k,l} v_l = v_k$ for $l < k$. Let $(V_{\infty}, F_l: V_l \longrightarrow V_{\infty})$ be the inductive limit. Let $v_{\infty} := F_1(v_1) \in V_{\infty}$.
As before, choose a strictly increasing function $t: \BZ_{>0} \longrightarrow \BZ_{>0}$ so that the following operators ($\rho^k$ denotes the representation of $U_q(\Gaff)$ on $V_k$)
\begin{displaymath}
F_{k,t(l)}^{-1} \rho^k(s) F_{k,l} : V_l \longrightarrow V_{t(l)}
\end{displaymath}
with $s \in S, k > t(l)$ are well-defined. The following lemma is proved in a similar way as Lemma \ref{lem: Laurent polynomiality}.
\begin{lem}
Let $l \in \BZ_{>0}$ and $s \in S$. Then there exists uniquely a $\textrm{Hom}(V_l,V_{t(l)})$-valued Laurent polynomial $P_{l,s}(z) \in \textrm{Hom}(V_l,V_{t(l)})[z,z^{-1}]$ such that
\begin{displaymath}
F_{k,t(l)}^{-1} \rho^k(s) F_{k,l} = P_{l,s}(z)|_{z = q^k}\quad \textrm{for}\ k > t(l).
\end{displaymath}
Furthermore, the coefficients of $z^n$ in $P_{l,s}(z)$ are non-zero only if $-2 \leq n \leq 1$.
\end{lem}
The rest is also parallel to the preceding subsection. We get a representation of $U_q(\Gaff)$ on the vector superspace $V_{\infty}$. Moreover, $v_{\infty}$ is killed by $s_{ij}(z),t_{ij}(z)$ whenever $1 \leq i < j \leq n$. Also by replacing $q^k$ and $v_l,v_{t(l)}$ in the following equation with $b$ and $v_{\infty}$, we conclude that $v_{\infty}$ verifies Equation \eqref{equ: odd generic asymptotic modules hwt vector}:
\begin{eqnarray*}
F_{k,t(l)}^{-1}\rho^k(s_{ii}(z)) F_{k,l} v_l &=& v_{t(l)} \begin{cases}
1 -z a & (i \leq r), \\
q^k - z a q^{-k} & (i > r),
\end{cases} \\
F_{k,t(l)}^{-1} \rho^k(t_{ii}(z)) F_{k,l} v_l &=& v_{t(l)} \begin{cases}
1 -z^{-1}a^{-1} & (i \leq r), \\
q^{-k} - z^{-1}a^{-1}q^k & (i > r).
\end{cases}
\end{eqnarray*}
The $U_q(\Gaff)$-module $V_{\infty}$ is the desired $L_{r,a}(b)$.
\subsection{Examples.} \label{sec: examples}
Let us give three examples to illustrate the general construction. In this subsection $\Glie = \mathfrak{gl}(2,1)$. We construct $L_{2,a}^{\pm},L_{2,a}(b)$ explicitly. Fix $a,b \in \BC^{\times}$.
\subsubsection{Construction of $L_{2,a}^-$.} For $k \in \BZ_{>0}$, let $v_4 \in W_{k,a}^{(2)}$ be a lowest $\ell$-weight vector. Define
\begin{displaymath}
v_3 := \tilde{s}_{23}^{(0)} v_4,\quad v_2 := \tilde{s}_{12}^{(0)} v_3,\quad v_1 := \tilde{s}_{23}^{(0)} v_2.
\end{displaymath}
Then by Theorem \ref{thm: BKK Schur-Weyl duality}, $(v_1,v_2,v_3,v_4)$ constitute a basis for the $U_q(\Gaff)$-module $W_{k,a}^{(2)}$. Moreover, from Relation \ref{rel: RSS = SSR}, we deduce the explicit action of the $s_{ij}(z)$. Let $\rho^k$ be the representation corresponding to the $U_q(\Gaff)$-module $W_{k,a}^{(2)}$. Then
\begin{eqnarray*}
&& \begin{cases}
\rho^k(s_{11}(z)) = (q^k - za q^{-k})(E_{11}+E_{22}) + (q^{k-1}-zaq^{-k+1}) (E_{33}+E_{44}), \\
\rho^k(s_{22}(z)) = (q^k - za q^{-k})(E_{11}+E_{33}) + (q^{k-1}-zaq^{-k+1}) (E_{22}+E_{44}), \\
\rho^k(s_{33}(z)) = (1-za) E_{11} + (q^{-1}-zaq)(E_{22}+E_{33}) + (q^{-2}-zaq^2)E_{44},
\end{cases} \\
&& \begin{cases}
\rho^k(s_{12}(z)) = q^k E_{23}, \\
\rho^k(s_{13}(z)) = q^{-1}(q-q^{-1})^{-1} (E_{24}-E_{13}), \\
\rho^k(s_{23}(z)) = q^{-1} E_{12} + q^{-2}E_{34},
\end{cases} \\
&& \begin{cases}
\rho^k(s_{21}(z)) = za (q-q^{-1})^2 q^{-k} E_{32}, \\
\rho^k(s_{31}(z)) = za q(q-q^{-1})^2(q^{k+1}-q^{-k-1}) E_{42} - zaq(q-q^{-1})^2(q^k-q^{-k}) E_{31}, \\
\rho^k(s_{32}(z)) = za q(q-q^{-1})(q^k - q^{-k})E_{21} + za q^2(q-q^{-1})(q^{k+1}-q^{-k-1}) E_{43}.
\end{cases}
\end{eqnarray*}
Here the $E_{ij}: v_l \mapsto \delta_{jl} v_i$ are the linear transformations on the underlying vector superspace $W = \bigoplus_{j=1}^4 \BC v_j$ of $W_{k,a}^{(2)}$. Observe that for all $1 \leq i,j \leq 3$, there exist $A_{ij}(z),B_{ij}(z) \in (\End W)[[z]]$ such that
\begin{displaymath}
\rho^k(\tilde{s}_{ij}(z)) = A_{ij}(z) + q^{-2k} B_{ij}(z) \quad \textrm{for\ all}\ k \in \BZ_{>0}.
\end{displaymath}
By using these $A_{ij}(z)$, we get the $Y_q(\Glie)$-module $L_{2,a}^- = W$ defined by:
\begin{eqnarray*}
&& \begin{cases}
\rho^-(s_{11}(z)) = E_{11}+E_{22} + q^{-1} (E_{33}+E_{44}), \\
\rho^-(s_{22}(z)) = E_{11}+E_{33} + q^{-1} (E_{22}+E_{44}), \\
\rho^-(s_{33}(z)) = (1-za) E_{11} + (q^{-1}-zaq)(E_{22}+E_{33}) + (q^{-2}-zaq^2)E_{44},
\end{cases} \\
&& \begin{cases}
\rho^-(s_{12}(z)) = E_{23}, \\
\rho^-(s_{13}(z)) = q^{-1}(q-q^{-1})^{-1} (E_{24}-E_{13}), \\
\rho^-(s_{23}(z)) = q^{-1} E_{12} + q^{-2}E_{34},
\end{cases} \\
&& \begin{cases}
\rho^-(s_{21}(z)) = 0, \\
\rho^-(s_{31}(z)) = za q^2(q-q^{-1})^2 E_{42} - zaq(q-q^{-1})^2 E_{31}, \\
\rho^-(s_{32}(z)) = za q(q-q^{-1})E_{21} + za q^3(q-q^{-1}) E_{43}.
\end{cases}
\end{eqnarray*}
\begin{rem} \label{rem: negative prefundamental module gl(2,1)}
It is straightforward to check that the representation $(W,\rho^-)$ is simple. Following \S \ref{sec: Gelfand-Tsetlin}, let $Y_q(\Glie_2)$ be the subalgebra of $Y_q(\Glie)$ generated by the $s_{ij}^{(n)}, (s_{ii}^{(0)})^{-1}$ with $n \in \BZ_{\geq 0}$ and $1 \leq i,j \leq 2$. Then as superalgebras $Y_q(\Glie_2) \cong Y_q(\mathfrak{gl}(2,0))$. Furthermore,
\begin{displaymath}
W = \BC v_1 \oplus \BC v_4 \oplus (\BC v_2 \oplus \BC v_3)
\end{displaymath}
is a Krull-Schmidt decomposition of the $Y_q(\Glie_2)$-module $\Res_{Y_q(\Glie_2)}^{Y_q(\Glie)} W$ into indecomposable submodules. Note that the third factor is not simple. Hence the underlying $Y_q(\Glie_2)$-module structure on $W$ is not semi-simple.
\end{rem}
\subsubsection{Construction of $L_{2,a}(b)$ and $L_{2,a}^+$.} For $k \in \BZ_{>0}$, let $w_4 \in W_{k,aq^{2k}}^{(2)}$ be a lowest $\ell$-weight vector. Define
\begin{displaymath}
w_3 := \widehat{s}_{23}^{(0)} w_4,\quad w_2 := \widehat{s}_{12}^{(0)} w_3,\quad w_1 := \widehat{s}_{23}^{(0)}w_2.
\end{displaymath}
Then $(w_1,w_2,w_3,w_4)$ is a basis for $W_{k,aq^{2k}}^{(2)}$. In the following, we identify the underlying vector superspaces of the $W_{k,aq^{(2k)}}^{(r)}$ for $k \in \BZ_{>0}$ with the vector superspace $W = W_{1,aq^2}^{(2)}$ by using this preferred basis. Let $\rho^k$ be the representation associated to the $U_q(\Gaff)$-module $W_{k,aq^{2k}}^{(2)}$. We have
\begin{eqnarray*}
&& \begin{cases}
\rho^k(s_{11}(z)) = (q^k - za q^{k})(E_{11}+E_{22}) + (q^{k-1}-zaq^{k+1}) (E_{33}+E_{44}), \\
\rho^k(s_{22}(z)) = (q^k - za q^{k})(E_{11}+E_{33}) + (q^{k-1}-zaq^{k+1}) (E_{22}+E_{44}), \\
\rho^k(s_{33}(z)) = (1-zaq^{2k}) E_{11} + (q^{-1}-zaq^{2k+1})(E_{22}+E_{33}) + (q^{-2}-zaq^{2k+2})E_{44},
\end{cases} \\
&& \begin{cases}
\rho^k(s_{12}(z)) = q^k E_{23}, \\
\rho^k(s_{13}(z)) = q^{k}(q-q^{-1})^{-1} (qE_{24}-E_{13}), \\
\rho^k(s_{23}(z)) = q^{k} E_{12} + q^{k}E_{34},
\end{cases} \\
&& \begin{cases}
\rho^k(s_{21}(z)) = za q^{k}(q-q^{-1})^2 E_{32}, \\
\rho^k(s_{31}(z)) = za q^{k-1}(q-q^{-1})^2(q^{k+1}-q^{-k-1}) E_{42} - zaq^k(q-q^{-1})^2(q^k-q^{-k}) E_{31}, \\
\rho^k(s_{32}(z)) = za q^k(q-q^{-1})(q^k - q^{-k})E_{21} + za q^k(q-q^{-1})(q^{k+1}-q^{-k-1}) E_{43}.
\end{cases}
\end{eqnarray*}
Now by replacing $\rho^k, q^k$ in the above formulas with $\rho,b$ respectively, we get a representation $\rho$ of $U_q(\Gaff)$ $W := \bigoplus_{i=1}^4\BC w_i$ for $W_{1,aq^{2}}^{(2)}$. The corresponding $U_q(\Gaff)$-module is $L_{2,a}(b)$.
As before, for $1 \leq i,j \leq 3$, there exist $A_{ij}(z),B_{ij}(z) \in (\End W)[[z]]$
\begin{displaymath}
\rho^k (\widehat{s}_{ij}(z)) = A_{ij}(z) + q^{2k} B_{ij}(z)\quad \textrm{for\ all}\ k \in \BZ_{>0}.
\end{displaymath}
Similarly, we get the $Y_q(\Glie)$-module $L_{2,a}^+ = W$:
\begin{eqnarray*}
&& \begin{cases}
\rho^+(s_{11}(z)) = (1 - za)(E_{11}+E_{22}) + (q^{-1}-zaq) (E_{33}+E_{44}), \\
\rho^+(s_{22}(z)) = (1 - za )(E_{11}+E_{33}) + (q^{-1}-zaq) (E_{22}+E_{44}), \\
\rho^+(s_{33}(z)) = E_{11} + q^{-1}(E_{22}+E_{33}) + q^{-2}E_{44},
\end{cases} \\
&& \begin{cases}
\rho^+(s_{12}(z)) = E_{23}, \\
\rho^+(s_{13}(z)) = (q-q^{-1})^{-1} (qE_{24}-E_{13}), \\
\rho^+(s_{23}(z)) = E_{12} + E_{34},
\end{cases} \\
&& \begin{cases}
\rho^+(s_{21}(z)) = za (q-q^{-1})^2 E_{32}, \\
\rho^+(s_{31}(z)) = - za q^{-2}(q-q^{-1})^2 E_{42} + za(q-q^{-1})^2 E_{31}, \\
\rho^+(s_{32}(z)) = - za (q-q^{-1})E_{21} - za q^{-1}(q-q^{-1}) E_{43}.
\end{cases}
\end{eqnarray*}
\begin{rem} \label{rem: positive prefundamental modules gl(2,1)}
The $Y_q(\Glie)$-module $L_{2,a}^+$ defined above is easily seen to be simple. Furthermore, contrary to Remark \ref{rem: negative prefundamental module gl(2,1)}, it is semi-simple as a $Y_q(\Glie_2)$-module.
\end{rem}
We point out that $L_{2,a}^-$ and $L_{2,a}^+$ have been constructed by Bazhanov-Tsuboi \cite[Appendix B.3]{BT} as representations of the upper Borel subalgebra of $U_q(\Gaff)$ defined by Drinfeld-Jimbo generators, bearing the name $\overline{W}_3^{-+}(x)$ and $W_{3}^{+-}(x)$ respectively after scattering.
\section{Category $\BGG$ and $q$-character} \label{sec: O and q}
We have constructed in \S \ref{sec: asymptotic representations}-\ref{sec: generic asymptotic representations} the asymptotic modules: $L_{r,a}^{\pm}$ as modules over $Y_q(\Glie)$ and $L_{r,a}(b)$ as modules over $U_q(\Gaff)$, as certain limits of Kirillov-Reshetikhin modules $W_{k,a}^{(r)}$. In this section, we introduce a category $\BGG$ of $Y_q(\Glie)$-modules including these three kinds of modules following Hernandez-Jimbo \cite{HJ} and study the Frenkel-Reshetikhin $q$-character for this category.
\subsection{Quantum Berezinian.} \label{sec: quantum Berezinian}
The quantum affine superalgebra admits another system of generators, the so-called {\it Drinfeld generators}, $X_{i,n}^{\pm}, K_{j,\pm s}^{\pm}$ where $i \in I_0, j \in I, n \in \BZ$ and $s \in \BZ_{\geq 0}$, arising from Gauss decomposition:
\begin{equation*}
\begin{cases}
S(z) = (\sum\limits_{i<j} f_{ji}^+(z) \otimes E_{ji} + 1 \otimes \Id_{\BV}) (\sum\limits_l K_l^+(z) \otimes E_{ii} )(\sum\limits_{i<j} e_{ij}^+(z) \otimes E_{ij} + 1 \otimes \Id_{\BV}), \\
T(z) = (\sum\limits_{i<j} f_{ji}^-(z) \otimes E_{ji} + 1 \otimes \Id_{\BV}) (\sum\limits_l K_l^-(z) \otimes E_{ii} )(\sum\limits_{i<j} e_{ij}^-(z) \otimes E_{ij} + 1 \otimes \Id_{\BV}).
\end{cases}
\end{equation*}
For example, $K_1^+(z) = s_{11}(z)$ and $K_1^-(z) = t_{11}(z)$. Now for $i \in I_0 = I$, we have
\begin{displaymath}
X_i^+(z) = e_{i,i+1}^+(z) - e_{i,i+1}^-(z) = \sum_n X_{i,n}^+ z^n,\quad X_i^-(z) = f_{i+1,i}^-(z) - f_{i+1,i}^+(z) = \sum_{n} X_{i,n}^- z^n.
\end{displaymath}
We refer to \cite[\S 3.4]{Z2} for more details on the relations and on the coproduct of these Drinfeld generators. Recall the definition of the $d_i, \theta_i$ in \S \ref{sec: convention} and in Remark \ref{rem: inverse of matrices and theta_i}. For $i \in I$, define the {\it quantum Berezinian}
\begin{equation} \label{equ: quantum berezinian}
C_i(z) := \prod_{j=1}^i (K_j^+(z \theta_j^{-1}))^{d_j} \in Y_q(\Glie)[[z]].
\end{equation}
Here $K_j^+(z) = \sum_{s \in \BZ_{\geq 0}} K_{j,s}^+ z^s \in Y_q(\Glie)[[z]]$.
The following result comes from \cite[Theorem 3.2, Proposition 3.13]{Z2}.
\begin{cor} \label{cor: JM elements}
Let $k \in I$.
\begin{itemize}
\item[(1)] $\Delta (C_k(z)) - C_k(z) \otimes C_k(z) \in \sum\limits_{\alpha \in \BQ_{\geq 0} \setminus \{0\}} (Y_q(\Glie))_{\alpha} \otimes (Y_q(\Glie))_{-\alpha}[[z]]$.
\item[(2)] For all $i,j \in I$ such that $i,j \leq k$, we have $C_k(z) s_{ij}(w) = s_{ij}(w) C_k(z)$ and $C_k(z) t_{ij}(w) = t_{ij}(w) C_k(z)$ as formal power series in $U_q(\Gaff)$.
\end{itemize}
\end{cor}
Let $C_q(\Gaff)$ be the subalgebra of $U_q(\Gaff)$ generated by the $(s_{ii}^{(0)})^{-1}$ and by the coefficients of the $C_i(z)$. Then $C_q(\Gaff)$ is indeed a commutative subalgebra of $Y_q(\Glie)$. Let $C_{i,0} \in Y_q(\Glie)$ be the constant term of $C_i(z)$. Then
\begin{equation} \label{equ: quantum berezinian cartan}
C_{i,0} = \prod_{j=1}^i (s_{jj}^{(0)})^{d_j}.
\end{equation}
\subsection{Weights and $\ell$-weights.} \label{sec: weights and l-weights} By using the commutative subalgebra $C_q(\Gaff)$ of $Y_q(\Glie)$, let us introduce the notion of a {\it weight} and an {\it $\ell$-weight}.
Denote $\DP := (\BC^{\times})^I$. Endow it with an additive abelian group structure:
\begin{displaymath}
(a_i)_{i\in I} + (b_i)_{i \in I} := (a_i b_i)_{i \in I}.
\end{displaymath}
For $i \in I$, define $\epsilon_i' \in \DP$ by $(\epsilon_i')_j = \begin{cases}
1 & (j < i), \\
q & (j \geq i).
\end{cases}$ Then the assignment $\epsilon_i \mapsto \epsilon_i'$ extends uniquely to an injective homomorphism of abelian groups $\BP \longrightarrow \DP$. From now on, we view $\BP$ as a free abelian subgroup of $\DP$ and identify $\epsilon_i = \epsilon_i'$ for $i \in I$. In this way, $\BQ \subset \BP \subset \DP$. Moreover $\alpha_i \in \DP$ for $i \in I_0$ is the $I$-tuple $(q^{\delta_{ij}})_{j \in I}$.
Similarly, denote $\ADP := (\BC[[z]]^{\times})^I$. Endow it with a multiplicative abelian group structure:
\begin{displaymath}
(f_i)_{i\in I} (g_i)_{i \in I} := (f_i g_i)_{i \in I}.
\end{displaymath}
Let $\varpi: \ADP \longrightarrow \DP$ and $\sigma: \DP \longrightarrow \ADP$ be two maps defined as follows:
\begin{displaymath}
\varpi (f_i(z))_{i\in I} = (f_i(0))_{i\in I},\quad \sigma (a_i)_{i\in I} = (a_i)_{i \in I}.
\end{displaymath}
These are homomorphisms of abelian groups. Furthermore $\varpi \circ \sigma = \Id_{\DP}$. For the precise statements of results, we shall not identify $\DP$ as a sub-abelian-group of $\ADP$ by $\sigma$.
Let us introduce the analogues of $\epsilon_i,\alpha_j \in \DP$ in $\ADP$ to be used later.
\begin{defi} \label{def: generalized simple roots}
For $i \in I$ and $a \in \BC$, define $\mathfrak{X}_{i,a} \in \ADP$ by
\begin{displaymath}
(\mathfrak{X}_{i,a})_j(z) = \begin{cases}
1 & (j < i), \\
q \frac{1 - zaq^{-1}}{1 - zaq} & (j \geq i).
\end{cases}
\end{displaymath}
For $i \in I_0$, define $A_{i,a} := \mathfrak{X}_{i,a} \mathfrak{X}_{i+1,a}^{-1} \in \ADP$. Call $A_{i,a}$ a {\it generalized simple root} if $a \neq 0$.
\end{defi}
By definition $\varpi(\mathfrak{X}_{i,a}) = \epsilon_i$ and $\varpi(A_{i,a}) = \alpha_i$.
Let $V$ be a $Y_q(\Glie)$-module. For $\underline{a} = (a_i)_{i\in I} \in \DP$, define the {\it weight space}
\begin{displaymath}
(V)_{\underline{a}} := \{ v \in V \ |\ C_{i,0} v = a_i v \quad \textrm{for}\ i \in I \}.
\end{displaymath}
Note that when $\underline{a} \in \BP$, the definition of weight space $(V)_{\underline{a}}$ is the same as the one given in Remark \ref{rem: cyclicity vs cocyclicity}, if we view $U_q(\Gaff)$-modules therein as $Y_q(\Glie)$-modules. Similarly, for $\underline{f} = (f_i(z))_{i \in I}$, define the {\it $\ell$-weight space}
\begin{displaymath}
(V)_{\underline{f}} := \{ v \in V\ |\ (C_i(z) - f_i(z))^n x = 0 \quad \textrm{for}\ i \in I \ \textrm{and}\ n \gg 0 \}.
\end{displaymath}
\subsection{Category $\BGG$.} \label{sec: BGG}
We say that a $Y_q(\Glie)$-module $V$ is in category $\BGG$ if:
\begin{itemize}
\item[(BGG1)] $V$ has a weight space decomposition $V = \bigoplus_{\alpha \in \DP} (V)_{\alpha}$;
\item[(BGG2)] for all $\alpha \in \DP$ we have $\dim (V)_{\alpha} < \infty$;
\item[(BGG3)] there exist a finite number of weights $\lambda_1,\cdots,\lambda_s \in \DP$ such that $(V)_{\alpha} \neq 0$ implies $\alpha \in \bigcup_{j=1}^s (\lambda_j - \BQ_{\geq 0})$.
\end{itemize}
As usual $\BGG$ is a tensor category (stable under direct sum, sub-quotient and tensor product).
\begin{example} \label{example: KR and asymptotic modules}
For $\lambda \in \BP$ and $a \in \BC^{\times}$, the evaluation modules $\ev_a^* L(\lambda)$, viewed as $Y_q(\Glie)$-modules, are in category $\BGG$. Furthermore, the $Y_q(\Glie)$-modules $L_{r,a}^{\pm}$ for $1 \leq r \leq M+N-1$ are in category $\BGG$. Indeed, the weights of these modules lie in $-\BQ_{\geq 0}$, and for $\alpha \in \BQ_{\geq 0}$,
\begin{displaymath}
\dim (L_{r,a}^{\pm})_{-\alpha} = \lim_{k\rightarrow \infty} \dim (W_{k,a}^{(r)})_{k\varpi_r - \alpha}
\end{displaymath}
where the limit at the RHS exists thanks to Corollary \ref{cor: asymptotic behaviour of weight spaces} in the case $r \leq M$ and thanks to \S \ref{sec: reduction} in the case $r > M$.
\end{example}
\begin{example} \label{example: generic asymptotic modules}
Let $r \in I_0$ and $a, b \in \BC^{\times}$. Then $L_{r,a}(b)$, viewed as a $Y_q(\Glie)$-module, is in category $\BGG$. Consider the case $r \leq M$. Let us come back to the situation of \S \ref{sec: even generic asymptotic modules}. Let $\alpha \in \BQ_{\geq 0}$. For $l \in \BZ_{>0}$ and $x \in (V_l)_{l\varpi_r - \alpha}$, we have
\begin{displaymath}
F_{k,t(l)}^{-1} \rho^k (s_{ii}^{(0)}) F_{k,l} (x) = F_{t(l),l}(x) \begin{cases}
q^{k-(\epsilon_i,\alpha)} & (i \leq r), \\
q^{-(\epsilon_i,\alpha)} & (i > r).
\end{cases}
\end{displaymath}
It follows that $s_{ii}^{(0)} F_l(x) = F_l(x) \begin{cases}
b q^{-(\alpha,\epsilon_i)} & (i \leq r), \\
q^{-(\alpha,\epsilon_i)} & (i > r).
\end{cases}$
In other words, $F_l(x) \in (L_{r,a}(b))_{\lambda_{r,b} - \alpha}$ where $\lambda_{r,b} \in \DP$ is given by $(\lambda_{r,b})_j = b^{\min (r,j)}$. In consequence, $(L_{r,a}(b))_{\lambda} \neq 0$ only if $\lambda = \lambda_{r,b} - \alpha$ for some $\alpha \in \BQ_{\geq 0}$, in which case
\begin{displaymath}
\dim (L_{r,a}(b))_{\lambda_{r,b} - \alpha} = \lim_{k\rightarrow \infty} \dim (W_{k,aq^{2k}}^{(r)})_{k\varpi_r - \alpha} < \infty.
\end{displaymath}
This says that $L_{r,a}(b)$ is in category $\BGG$.
When $r > M$, similarly one can find $\lambda_{r,b} \in \DP$ with $(\lambda_{r,b})_j = b^{\max(j-r,0)}$ such that the above statements on weight spaces of $L_{r,a}(b)$ in the case $r \leq M$ remain true.
\end{example}
\subsection{Character and $q$-character.}
To define classical character and $q$-character, let us first introduce the target rings. Let $\CE \subset \BZ^{\DP}$ be the set of maps $c: \DP \longrightarrow \BZ$ satisfying $c(\alpha) = 0$ for all $\alpha$ outside a finite union of sets of the form $\lambda - \BQ_{\geq 0}$. Endow $\CE$ with a ring structure:
\begin{displaymath}
[\alpha]:= \delta_{\alpha,\cdot},\quad [\alpha] [\beta] := [\alpha + \beta],\quad (c+d)(\alpha) := c(\alpha) + d(\alpha).
\end{displaymath}
In particular, we see that $\CE$ contains the group ring $\BZ[\DP]$. For $V$ a $Y_q(\Glie)$-module in category $\BGG$, define its {\it classical character} as in \S \ref{sec: character}:
\begin{displaymath}
\chi(V) := \sum_{\alpha \in \DP} \dim (V)_{\alpha} [\alpha] \in \CE.
\end{displaymath}
Let $\CEl \subset \BZ^{\ADP}$ be the set of maps $c: \ADP \longrightarrow \BZ$ satisfying $c(f) = 0$ for all $f$ such that $\varpi(f)$ is outside a finite union of sets of the form $\mu - \BQ_{\geq 0}$ and such that for each $\alpha \in \DP$, there are finitely many $f$ verifying $\varpi(f) = \alpha$ and $c(f) \neq 0$. Make $\CEl$ into a ring:
\begin{displaymath}
[f]:= \delta_{f,\cdot},\quad [f] [g] := [f g],\quad (c + d)(f) := c(f) + d(f).
\end{displaymath}
Extend $\varpi: \ADP \longrightarrow \DP$ to a surjective ring morphism $\varpi: \CEl \longrightarrow \CE$ and $\sigma: \DP \longrightarrow \ADP$ to an injective ring morphism $\sigma: \CE \longrightarrow \CEl$.
Let $V$ be a $Y_q(\Glie)$-module in category $\BGG$. For $\alpha \in \DP$, the weight space $(V)_{\alpha}$, being finite-dimensional, is stable by the action of $C_q(\Glie)$. It follows that $(V)_{\alpha}$ (and hence $V$) admits an $\ell$-weight space decomposition $V = \bigoplus_{f \in \ADP} (V)_f$ with finite-dimensional $\ell$-weight spaces. Define the $q$-{\it character}
\begin{displaymath}
\chi_q(V) := \sum_{f \in \ADP} \dim (V)_f [f] \in \BZ^{\ADP}.
\end{displaymath}
Note that for all $f \in \ADP$, $(V)_f \subseteq (V)_{\varpi(f)}$. Hence
\begin{displaymath}
\chi_q(V) \in \CEl,\quad \varpi \chi_q(V) = \chi(V).
\end{displaymath}
In other words, $\chi_q$ is a refinement of $\chi$.
\begin{rem} \label{rem: normalized q character}
We shall also need the notion of a {\it normalized $q$-character}. Let $V$ be in category $\BGG$. Assume that there exists a $\lambda \in \DP$ such that: $(V)_{\lambda} = \BC v$ is one-dimensional; if $(V)_{\mu} \neq 0$ then $\mu \in \lambda \in \lambda - \BQ_{\geq 0}$. (Compare Remark \ref{rem: cyclicity vs cocyclicity}.) Since the one-dimensional space $(V)_{\lambda}$ is stable by $C_q(\Gaff)$, there exists $f = (f_i: i \in I) \in \ADP$ such that
\begin{displaymath}
C_i(z) v = f_i(z) v \quad \textrm{for}\ i \in I.
\end{displaymath}
Clearly, for such a $Y_q(\Glie)$-module $V$, the weight $\lambda$ and the $\ell$-weight $f$ are uniquely determined.
Define the normalized character $\widetilde{\chi}$ and normalized $q$-character $\widetilde{\chi}_q$ as follows:
\begin{equation}
\widetilde{\chi}(V) := [\lambda]^{-1} \chi(V),\quad \widetilde{\chi}_q(V) := [f]^{-1} \chi_q(V).
\end{equation}
In the following, whenever we write $\widetilde{\chi}(V)$ or $\widetilde{\chi}_q(V)$, it should be understood implicitly that $V$ verifies the above weight conditions.
\end{rem}
\begin{example}
Let $r \in I_0$ and $a,b \in \BC^{\times}$. Then from Examples \ref{example: KR and asymptotic modules}-\ref{example: generic asymptotic modules} we see that
\begin{equation} \label{equ: normalized character for asymptotic modules}
\widetilde{\chi} (L_{r,a}^{\pm}) = \widetilde{\chi}(L_{r,a}(b)) = \lim_{k \rightarrow \infty} \widetilde{\chi}(W_{k,a}^{(r)}) \in \CE.
\end{equation}
\end{example}
\begin{lem} \label{lem: multiplicative structure of q character}
Let $V,W$ be in category $\BGG$. Then
\begin{align*}
& \chi (V \oplus W) = \chi(V) + \chi(W),\quad \chi(V \otimes W) = \chi(V) \chi(W), \\
& \chi_q(V \oplus W) = \chi(V) + \chi(W),\quad \chi_q(V \otimes W) = \chi_q(V) \chi_q(W).
\end{align*}
\end{lem}
\begin{proof}
The proof is exactly the same as in the non-graded case (see \cite[Remark 2.6]{FR}) in view of the coproduct formula in Corollary \ref{cor: JM elements} (1), by using the partial order on $\DP$ induced by $\BQ_{\geq 0}$.
\end{proof}
\begin{rem} \label{rem: q character and tensor product}
Let $V,W$ be in category $\BGG$. In general, it is not true that $(V)_f \otimes (W)_g \subseteq (V \otimes W)_{fg}$ for $f,g \in \ADP$. However, if $v \in V$ is a highest $\ell$-weight vector (see \S \ref{sec: prefundamental modules}), then $v \in (V)_f$ for a unique $f \in \ADP$ and for all $g \in \ADP$
\begin{displaymath}
v \otimes (W)_g \subseteq (V \otimes W)_{fg}.
\end{displaymath}
For example, if $V,W$ verify such weight conditions that their normalized characters exist (Remark \ref{rem: normalized q character}), then so does $V \otimes W$ and
\begin{displaymath}
\widetilde{\chi}(V \otimes W) = \widetilde{\chi}(V) \widetilde{\chi}(W),\quad \widetilde{\chi}_q(V \otimes W) = \widetilde{\chi}_q(V) \widetilde{\chi}_q(W).
\end{displaymath}
\end{rem}
\subsection{Simple modules in category $\BGG$.} \label{sec: simple in O}
As in the non-graded case \cite[Theorem 3.11]{HJ}, we also have a classification of simple modules in category $\BGG$ in terms of highest $\ell$-weight.
Following \cite[\S 5.1.3]{Z2}, let $\CR$ be the subset of $(1 + z \BC[[z]])^{I_0}$ consisting of $I_0$-tuples of power series $(f_i(z): i \in I_0)$ such that $f_i(z) \in \BC(z)$ for all $i \in I_0$.
\begin{lem} \label{lem: simple modules in category O}
(1) For all $f = (f_i(z): i \in I_0) \in \CR$, there exists uniquely a simple $Y_q(\Glie)$-module generated by a highest $\ell$-weight vector $v$ with
\begin{displaymath}
|v| = \even,\quad s_{ii}(z) v = (\prod_{j=i}^{M+N-1} f_j(z)) v\ \ \textrm{for}\ i \in I_0, \quad s_{M+N,M+N}(z) v = v.
\end{displaymath}
Moreover, such a module is in category $\BGG$. Let $V(f)$ be the $Y_q(\Glie)$-module thus obtained.
(2) All simple modules in category $\BGG$ can be factorized uniquely into $V(f) \otimes \BC_f$ with $f \in \CR$ and $\BC_f$ one-dimensional.
\end{lem}
\begin{proof}
The proof of Part (2) is standard as in the non-graded case. Let $S$ be a simple module in category $\BGG$. Then $S$ must be a highest $\ell$-weight $Y_q(\Glie)$-module. Let $v \in S$ be a highest $\ell$-weight vector with weight $\lambda \in \DP$. Then it is enough to show that
\begin{displaymath}
K_{i+1}^+(z)K_{i}^+(z)^{-1} v \in \BC(z) v \quad \textrm{for}\ i \in I_0.
\end{displaymath}
This comes essentially from the fact that $(S)_{\lambda-\alpha_i}$ is finite-dimensional combined with the following relations and $\BQ$-grading on $U_q(\Gaff)$:
\begin{align*}
& |K_{i,n}^+|_{\BQ} = 0,\quad |X_{i,n}^{\pm}|_{\BQ} = \pm \alpha_i, \quad (Y_q(\Glie))_{\alpha} (S_{\mu}) \subseteq S_{\alpha + \mu}, \\
& \sum_{n\in \BZ_{>0}} [X_{i,0}^+, X_{i,n}^-] z^n = (q_i - q_i^{-1}) (K_{i+1}^+(z)K_{i}^+(z)^{-1} - K_{i+1,0}^+ (K_{i,0}^+)^{-1} ), \\
& [h_i, X_{i,n}^-] = X_{i,n+1}^-,\quad |h_i|_{\BQ} = 0,
\end{align*}
where $h_i \in Y_q(\Glie)$ is some properly chosen element (see \cite[Appendix A]{Z2}).
The proof of Part (1) is the same as that of \cite[Theorem 3.11]{HJ} or \cite[Lemma 5.1]{Z2} by realizing $V(f)$ as a simple sub-quotient of certain tensor product of the $L_{r,a}^{\pm}$ and one-dimensional $Y_q(\Glie)$-modules.
\end{proof}
As in \cite[\S 5.1.1]{Z2}, one-dimensional $Y_q(\Glie)$-modules are factorized uniquely into the form
\begin{displaymath}
\BC_s \otimes \BC_{\underline{a}} \otimes \BC_f,\quad s \in \super,\ \underline{a} = (a_i)_{i \in I} \in \DP,\ f \in 1 + z\BC[[z]].
\end{displaymath}
Here $\BC_{\underline{a}}$ is the usual sign module, $\BC_{\underline{a}} = \BC v$ is the $Y_q(\Glie)$-module with
\begin{displaymath}
|v| = \even,\quad s_{ii}(z) v = a_i \delta_{ij} v,
\end{displaymath}
and $\BC_f = \BC w$ is the $Y_q(\Glie)$-module such that
\begin{displaymath}
|v| = \even,\quad s_{ii}(z) = f(z) \delta_{ij} v.
\end{displaymath}
\begin{example} \label{example: asymptotic modules simple modules}
For $r \in I_0$ and $a \in \BC^{\times}$, let $\varpi_{r,a}^{\pm} \in \CR$ be such that
\begin{displaymath}
(\varpi_{r,a}^{\pm})_j := (1 - z a \delta_{jr})^{\pm 1} \in 1 + z\BC[[z]].
\end{displaymath}
Then the $Y_q(\Glie)$-module $L_{r,a}^{\pm}$ has a simple sub-quotient isomorphic to $V(\varpi_{r,a}^{\pm}) \otimes \BC_{f_{r,a}^{\pm}}$ where $f_{r,a}^{+} = 1$ and $f_{r,a}^- = 1 - za$.
\end{example}
\begin{example} \label{example: asymptotic modules generic simple modules}
For $r \in I_0$ and $a,b \in \BC^{\times}$, the $Y_q(\Glie)$-module $L_{r,a}(b)$ has a simple sub-quotient isomorphic to $V(f) \otimes D$ with $D$ one dimensional and $f \in \CR$ defined by
\begin{displaymath}
f_j = 1\ \textrm{for}\ j \neq r,\quad f_r = \begin{cases}
\frac{1-za}{1-zab^2} & (r \leq M), \\
\frac{1-za}{1-zab^{-2}} & (r > M).
\end{cases}
\end{displaymath}
\end{example}
We end this subsection with the following observations.
\begin{prop} \label{prop: finite-dimensional and extension}
Let $f = (f_i(z): i \in I_0) \in \CR$. Consider the $Y_q(\Glie)$-module $V(f)$.
\begin{itemize}
\item[(1)] $V(f)$ is finite-dimensional if and only if for all $i \in I_0 \setminus \{M\}$ there exists $P_i(z) \in 1 + z\BC[z]$ such that $f_i(z) = \frac{P_i(z q_i^{-1})}{P_i(z q_i)}$.
\item[(2)] $V(f)$ admits a $U_q(\Gaff)$-module structure extending that of $Y_q(\Glie)$ up to tensor product by one-dimensional modules if and only if for all $i \in I_0$, seen as a meromorphic function, $f_i(z)$ is regular at $z = \infty$ and $f_i(\infty) \in \BC^{\times}$.
\end{itemize}
\end{prop}
The proof of this proposition is again standard as in the non-graded case, by using the Drinfeld generators of $U_q(\Gaff)$. See \cite[\S 3.2]{HJ} for Part (1), \cite[Lemma 4.11]{He} and \cite[Theorem 3.6]{Mukhin} for Part (2). For example, As seen from their constructions in \S \ref{sec: asymptotic representations}-\ref{sec: generic asymptotic representations} the $L_{r,a}(b)$ are $U_q(\Gaff)$-modules, whereas the $L_{r,a}^{\pm}$ are not. Note that this proposition also gives a classification of finite-dimensional simple $U_q(\Gaff)$-modules in terms of highest $\ell$-weight, as done in \cite{Z} with a quite different approach.
\subsection{Category $\BGG$ of the $q$-Yangian $Y_q(\mathfrak{gl}(1,1))$.} \label{sec: example gl(1,1)}
In this subsection, we compute the normalized $q$-character for all simple modules $V(f)$ in category $\BGG$ of the $q$-Yangian $Y_q(\mathfrak{gl}(1,1))$. This serves as the first non-trivial example for the study of normalized $q$-character of the asymptotic modules.
Note that in this case $\CR$ is the set of rational functions of the form $\frac{P(z)}{Q(z)}$ where $P(z),Q(z) \in 1 + z\BC[z]$ are co-prime.
\subsubsection{Prime simple modules.} Let $a,b \in \BC$ be such that $a \neq b$. As we see in \cite[\S 5.2.2]{Z2}, the simple module $V(\frac{1-za}{1-zb})$ is two-dimensional with basis $v_1,v_2$ and with the action of the $s_{ij}(z)$ given as follows:
\begin{eqnarray*}
(s_{ij}(z))_{1\leq i,j \leq 2} &=& \begin{pmatrix}
\frac{1-za}{1-zb}E_{11} + \frac{q^{-1}-zaq}{1-zb} E_{22} & \frac{(q^{-1}-q)(b-a)}{1-zb} E_{12} \\
\frac{-z}{1-zb} E_{21} & E_{11} + \frac{q^{-1}-zbq}{1-zb} E_{22}
\end{pmatrix}
\end{eqnarray*}
where the $E_{ij}: v_k \mapsto \delta_{jk} v_i$ are endomorphisms of $V(\frac{1-za}{1-zb})$. Let us compute the action of $K_i^+(z)$. By the Gauss decomposition,
\begin{eqnarray*}
K_1^+(z) &=& s_{11}(z) = \frac{1-za}{1-zb}E_{11} + \frac{q^{-1}-zaq}{1-zb} E_{22}, \\
K_2^+(z) &=& s_{22}(z) + s_{21}(z) s_{11}(z)^{-1} s_{12}(z) \\
&=& E_{11} + \frac{q^{-1}-zbq}{1-zb} E_{22} + \frac{-z}{1-zb} E_{21} (\frac{1-za}{1-zb})^{-1}E_{11} \frac{(q^{-1}-q)(b-a)}{1-zb} E_{12} \\
&=& E_{11} + \frac{q^{-1}-zaq}{1-za}.
\end{eqnarray*}
It follows that (\S \ref{sec: quantum Berezinian})
\begin{displaymath}
C_1(z) = \frac{1-za}{1-zb}E_{11} + \frac{q^{-1}-zaq}{1-zb} E_{22},\quad C_2(z) = \frac{1-za}{1-zb} \Id.
\end{displaymath}
By definition of normalized $q$-character in Remark \ref{rem: normalized q character},
\begin{equation} \label{equ: q character prime simple module gl(1,1)}
\widetilde{\chi}_q (V(\frac{1-za}{1-zb})) = 1 + [A_{1,aq}]^{-1}.
\end{equation}
In particular, the normalized $q$-character of $V(\frac{1-za}{1-zb})$ depends only on $a \in \BC$.
\subsubsection{General simple modules.} Let $1 \neq f \in \CR$. Then there exists a unique factorization
\begin{displaymath}
f(z) = \frac{\prod_{i=1}^s (1- z a_i)}{\prod_{j=1}^t (1- z b_j) }
\end{displaymath}
where $a_i,b_j \in \BC^{\times}$ and $a_i \neq b_j$ for all $1 \leq i \leq s, 1 \leq j \leq t$. Let us set $a_i = 0$ for $s < i \leq \max(s,t)$ and $b_i = 0$ for $t < i \leq \max(s,t)$. Then $a_i \neq b_j$ for all $1 \leq i,j \leq \max(s,t)$. From \cite[\S 5.2.1]{Z2} we have the factorization
\begin{displaymath}
V(f) \cong \bigotimes_{i=1}^{\max(s,t)} V(\frac{1-z a_i}{1-zb_i})
\end{displaymath}
In view of Remark \ref{rem: q character and tensor product} this amounts to
\begin{equation} \label{equ: q character simple module gl(1,1)}
\widetilde{\chi}_q (V(f)) = \prod_{i=1}^{\max (s,t)} (1 + [A_{1,a_i q}]^{-1}).
\end{equation}
\begin{rem} \label{rem: q character gl(1,1)}
(1) Let $1 \neq f \in \CR$. Then $\widetilde{\chi}_q (V(f))$ is a polynomial in the $[A_{1,a}]^{-1}$ with $a \in \BC^{\times}$ if and only if $f(\infty) \neq 0$. Retain the above factorization for $f \in \CR$. Then $V(f)$ has multiplicity-free $q$-character if and only if $a_i \neq a_j$ for all $1 \leq i < j \leq \max(s,t)$.
(2) $\widetilde{\chi}_q (V(\frac{1}{1-za})) = \sigma \widetilde{\chi} (V(\frac{1}{1-za}))$. In other words, the normalized character and normalized $q$-character for the simple module $V(\frac{1}{1-za})$ coincide. In the non-graded case, as proved by Hernandez-Jimbo \cite[Theorem 7.5]{HJ} in the simply-laced case and later by Frenkel-Hernandez \cite[Theorem 4.1]{FH} in full generality, this is true for the positive prefundamental modules within the framework of Hernandez-Jimbo, and it gives rise to Baxter's \textbf{T-Q} relation for an arbitrary non-twisted quantum affine algebra.
(3) As in \cite[\S 5]{Z2}, let $\CF$ be the category of finite-dimensional representations of $Y_q(\Glie)$. It is not true that $\CF$ must be a subcategory of $\BGG$. Nevertheless, the $q$-character $\chi_q(V) \in \BZ[\ADP] \subset \CEl$ is still well-defined for $V$ in $\CF$ and it induces a ring homomorphism
\begin{displaymath}
\chi_q: K_0(\CF) \longrightarrow \BZ[\ADP]
\end{displaymath}
where $K_0(\CF)$ is the usual Grothendieck ring of the tensor category $\CF$. In order to have an injective $q$-character map as in the non-graded case, we need to take into account the $\super$-grading. This is done by extending $\chi_q$ to the super $q$-character map $s.\chi_q$:
\begin{displaymath}
s.\chi_q (V) := \sum_{f \in \ADP} s.\dim (V)_f [f] \in \BZ[\ADP] \otimes_{\BZ} \BZ[\varepsilon]/(\varepsilon^2)
\end{displaymath}
where $s.\dim W := \dim W_{\even} + \varepsilon \dim W_{\odd}$ is the {\it superdimension} associated with a vector superspace \cite[\S 1]{Serganova}. The resulting map $s.\chi_q: K_0(\CF) \longrightarrow \BZ[\ADP] \otimes_{\BZ} \BZ[\varepsilon]/(\varepsilon^2)$ is an injective ring homomorphism. In a similar way as in Remark \ref{rem: normalized q character}, normalized super $q$-character $s.\widetilde{\chi}_q$ can be defined, and Equation \eqref{equ: q character simple module gl(1,1)} becomes
\begin{equation}
s.\widetilde{\chi}_q (V(f)) = \prod_{i=1}^{\max (s,t)} (1 + \varepsilon [A_{1,a_i q}]^{-1}).
\end{equation}
(4) If we work directly with category $\BGG$, then $s.\chi_q(V)$ is still well-defined and it induces an injective ring homomorphism $s.\chi_q: K_0^c(\BGG) \longrightarrow \CEl \otimes_{\BZ} \BZ[\varepsilon]/(\varepsilon^2)$ where $K_0^c(\BGG)$ is certain completion of the usual Grothendieck ring $K_0(\BGG)$. Already in the non-graded case \cite[Proposition 3.12]{HJ} concerning the injective $q$-character morphism, a completed version $K_0^c(\BGG)$ of Grothendieck ring is implicitly used. We are grateful to David Hernandez for this comment.
\end{rem}
\section{$q$-character of evaluation representations} \label{sec: GT bases}
In this section, we study the (normalized) $q$-character of the evaluation modules $\ev_a^* L(\lambda)$ and $\ev_a^* L(\lambda)^*$ for $\lambda \in \CP$, following the idea of Frenkel-Mukhin \cite[Lemma 4.7]{FM} relating $\ell$-weight spaces to (dual) Gelfand-Tsetlin bases.
\subsection{$q$-character and Gelfand-Tsetlin bases.} For $k \in I$, let $U_q(\Gaff_k)$ be the subalgebra of $U_q(\Gaff)$ generated by the $s_{ij}^{(n)},t_{ij}^{(n)}$ with $i,j \leq k$. Then as seen in Corollary \ref{cor: JM elements}, the coefficients of $C_k(z)$ are central elements in $U_q(\Gaff_k)$. Moreover the following are isomorphic as superalgebras:
\begin{displaymath}
U_q(\Gaff_k) \cong U_q(\widehat{\mathfrak{gl}(M_k,N_k)})
\end{displaymath}
where $M_k,N_k$ are defined at the beginning of \S \ref{sec: Gelfand-Tsetlin}. Also from Proposition \ref{prop: evaluation morphism}
\begin{displaymath}
\ev_a U_q(\Gaff_k) \subseteq \Uc_q(\Glie_k).
\end{displaymath}
Hence, for $V$ a $\Uc_q(\Glie)$-module and for $a \in \BC^{\times}$, as $U_q(\Gaff_k)$-modules
\begin{displaymath}
\Res_{U_q(\Gaff_k)}^{U_q(\Gaff)} \ev_a^* V \cong \ev_a^* \Res_{\Uc_q(\Glie_k)}^{\Uc_q(\Glie)} V.
\end{displaymath}
At the RHS, $\ev_a: U_q(\Gaff_k) \cong U_q(\widehat{\mathfrak{gl}(M_k,N_k)}) \longrightarrow \Uc_q(\mathfrak{gl}(M_k,N_k)) \cong \Uc_q(\Glie_k)$.
\begin{prop} \label{prop: q-character and GT basis}
Let $\lambda \in \CP$. Let $V(\lambda;a) := \phi^*_{((1-z^{-1}a^{-1})^{-1}, (1-za)^{-1})} \ev_a^* L(\lambda)$. Then
\begin{equation} \label{equ: q-character for evaluation modules}
\chi_q (V(\lambda;a)) = \sum_{f \in \CB(Y^{\lambda})} [\prod_{(i,j) \in Y^{\lambda}} \mathfrak{X}_{f(i,j),aq^{2(i-j)-1}}].
\end{equation}
In particular, $V(\lambda;a)$ has multiplicity-free $q$-character.
\end{prop}
\begin{proof}
The idea is the same as that of \cite[Lemma 4.7]{FM}. For completeness, let us explain briefly the main steps.
Let $(v_{\underline{\lambda}}: \underline{\lambda} \in \GT(\lambda))$ be a basis of $L(\lambda)$ verifying Corollary \ref{cor: Gelfand-Tsetlin basis for modules in BKK}. Fix a Gelfand-Tsetlin pattern $\underline{\lambda} = (\lambda^{(k)})_{k \in I} \in \GT(\lambda)$ and let $f := GT_{\lambda}^{-1}(\underline{\lambda})$ be the associated Young tableau. For $k \in I$, there exists a sign module $\BC_s$ such that $v_{\underline{\lambda}}$ is contained in a simple sub-$U_q(\Gaff_k)$-module $S$ isomorphic to
\begin{displaymath}
\phi^*_{((1-z^{-1}a^{-1})^{-1}, (1-za)^{-1})} \ev_a^* L(\lambda^{(k)};\Glie_k) \otimes \BC_s.
\end{displaymath}
As $C_k(z)$ is central in $U_q(\Gaff_k)[[z]]$, there exists $g_k(z) \in \BC[[z]]$ such that $C_k(z) v = g_k(z) v$ for all $v \in S$. In particular, $v_{\underline{\lambda}}$ is stable by $C_k(z)$. Now $g_k(z)$ can be computed via a highest $\ell$-weight vector $v^+$ of $S$. Indeed, if $\lambda^{(k)} = \sum_{i=1}^k x_i \epsilon_i$, then
\begin{eqnarray*}
C_k(z) v^+ &=& \prod_{j=1}^k (K_j^+(z \theta_j^{-1}))^{d_j} v^+ = \prod_{j=1}^k (s_{jj}(z \theta_j^{-1}))^{d_j} v^+ \\
&=& v^+ \prod_{j=1}^k (\frac{q_j^{x_j} - z a \theta_j^{-1}q_j^{-x_j}}{1-za\theta_j^{-1}})^{d_j} = v^+ g_k(z).
\end{eqnarray*}
Here the second equation above comes from Gauss decomposition and from the fact that $v^+$ is a highest $\ell$-weight vector of the $U_q(\Gaff_k)$-module $S$. It is enough to show
\begin{displaymath}
g_k(z) = (\prod_{(i,j) \in Y^{\lambda}} \mathfrak{X}_{f(i,j),aq^{2(i-j)-1}})_k = \prod_{f(i,j) \leq k} q \frac{1-zaq^{2(i-j)-2}}{1-zaq^{2(i-j)}}.
\end{displaymath}
For example, let us assume $k > M$. Then $f(i,j) \leq k$ if and only if $(1 \leq i \leq M, 1 \leq j \leq x_i)$ or $(1 \leq j \leq k-M, M+1 \leq i \leq M+x_{M+j}$. It follows that
\begin{eqnarray*}
RHS &=& (\prod_{i=1}^M \prod_{j=1}^{x_i} q \frac{1-zaq^{2(i-j)-2}}{1-zaq^{2(i-j)}}) (\prod_{j=1}^{k-M} \prod_{i=M+1}^{M+x_{M+j}} q \frac{1-zaq^{2(i-j)-2}}{1-zaq^{2(i-j)}}) \\
&=& (\prod_{i=1}^M \frac{q^{x_i} - z a q^{2(i-1)-x_i}}{1-zaq^{2(i-1)}}) \times (\prod_{j=1}^{k-M} \frac{q^{-x_{M+j}} - z a q^{2(M-j) + x_{M+j}}}{1 - z a q^{2(M-j)}} )^{-1} \\
&=& \prod_{j=1}^k (\frac{q_j^{x_j} - z a \theta_j^{-1} q_j^{-x_j}}{1 - z a \theta_j^{-1}})^{d_j} = g_k(z).
\end{eqnarray*}
That $V(\lambda;a)$ has multiplicity-free $q$-character comes from Definition \ref{def: Young combinatorics} and from the fact that the $\mathfrak{X}_{i,b}$ with $i \in I, b \in \BC^{\times}$ are algebraically independent in $\CEl$.
\end{proof}
As an immediate corollary, we have
\begin{cor} \label{cor: normalized q-character for polynomial representations}
$\tilde{\chi}_q (V(\lambda;a)) \in \BZ[[A_{i,b}]^{-1} : i \in I_0, b \in aq^{2\BZ+1}]$ for $\lambda \in \CP$ and $a \in \BC^{\times}$.
\end{cor}
\begin{proof}
Let $\lambda = \sum_{i \in I1} x_i \epsilon_i \in \CP$. Let $Y^{\lambda}$ be the associated $\Glie$-Young diagram. Let $f^+ \in \CB(Y^{\lambda})$ be the $\Glie$-Young tableau corresponding to the Gelfand-Tsetlin pattern
\begin{displaymath}
GT_{\lambda}(f^+)^{(k)} = \sum_{i=1}^k x_i \epsilon_i \quad \textrm{for}\ k \in I.
\end{displaymath}
Then for all $f \in \CB(Y^{\lambda})$ and $(i,j) \in Y^{\lambda}$, we have $f(i,j) \geq f^+(i,j)$. The rest is clear from Equation \eqref{equ: q-character for evaluation modules} and Definition \ref{def: generalized simple roots}.
\end{proof}
\begin{example} \label{example: normalized q character for gl(2,1)}
Let $(M,N) = (2,1)$ and $\Glie = \mathfrak{gl}(2,1)$. Then
\begin{eqnarray*}
\widetilde{\chi}_q (W_{k,aq^{2k}}^{(2)}) &=& 1 + [A_{2,aq^3}]^{-1} + [A_{1,aq}]^{-1}[A_{2,aq^3}]^{-1} + [A_{1,aq}]^{-1}[A_{2,aq}]^{-1}[A_{2,aq^3}]^{-1}, \\
\widetilde{\chi}_q (W_{k,aq^{2k}}^{(1)}) &=& 1 + (1 + [A_{2,aq}]^{-1}) \sum_{l=1}^k \prod_{j=1}^l [A_{1,aq^{2j-1}}]^{-1}.
\end{eqnarray*}
\end{example}
The following combinatorial property of normalized $q$-character will be used later on.
\begin{lem} \label{lem: factorization of q character}
Let $a,b \in \BC^{\times}$ and $k \in \BZ_{>0}$. Let $1 \leq r \leq M$. Write
\begin{displaymath}
\widetilde{\chi}_q (W_{k,a}^{(r)}) = \sum_{\Phi \in \ADP} c_{\Phi} [\Phi],\quad \widetilde{\chi}_q (W_{k,ab^2}^{(r)}) = \sum_{\Phi \in \ADP} d_{\Phi} [\Phi].
\end{displaymath}
Assume that $b^2 \notin q^{2\BZ_{>-r}}$. Then for all $\Phi,\Phi_1,\Phi_2 \in \ADP$ such that
\begin{displaymath}
\Phi = \Phi_1 \Phi_2,\quad c_{\Phi} = c_{\Phi_1} = d_{\Phi_2} = 1,
\end{displaymath}
we must have $\Phi_2 = 1$.
\end{lem}
\begin{proof}
Let $Y = \{(i,j) \in \BZ^2 | 1 \leq i \leq r, 1 \leq j \leq k \}$ be the $\Glie$-Young diagram associated with $L(k\varpi_r)$. Let $\CB = \CB(Y)$ be the set of $\Glie$-Young tableaux of shape $Y$. In view of Proposition \ref{prop: q-character and GT basis}, it is enough to prove the following:
\begin{itemize}
\item[(1)] Let $f,g,h \in \CB$ be such that in the group $\ADP$:
\begin{displaymath}
\prod_{(i,j) \in Y} \mathfrak{X}_{g(i,j),aq^{2(i-j)}}^{-1} \mathfrak{X}_{f(i,j), aq^{2(i-j)-1}} = \prod_{(i,j) \in Y} \mathfrak{X}_{i,ab^2 q^{2(i-j)}}^{-1} \mathfrak{X}_{h(i,j), ab^2 q^{2(i-j)-1}},
\end{displaymath}
then $h(i,j) = i$ for $(i,j) \in Y$.
\end{itemize}
Remark that the $\mathfrak{X}_{i,x}$ with $i \in I, x \in \BC^{\times}$ generate a free abelian subgroup of $\ADP$. Let us compare the index $x$ of the $X_{i,x}$ on both sides. For $1 \leq i \leq r$ by assumption
\begin{displaymath}
ab^2 q^{2(i-k)-1} = aq^{2(i-j-s)-1} \notin \{ aq^{2(i'-j')-1} \ |\ (i',j') \in Y \}.
\end{displaymath}
Hence the $\mathfrak{X}_{i,ab^2q^{2(i-k)-1}}$ for $1 \leq i \leq r$ on the RHS must disappear. Furthermore, $i \leq h(i,k)$ and $h(i,k) \leq h(i',k)$ for $1 \leq i \leq i' \leq k$. It follows that $h(i,k) = i$ for $1 \leq i \leq r$. By definition of a $\Glie$-Young tableau, this says that $h(i,j) = i$ for all $(i,j) \in Y$.
\end{proof}
As we see in Example \ref{example: normalized q character for gl(2,1)} (by taking $\Glie = \mathfrak{gl}(2,1), r = 2, b = q^{-1}$), The Lemma above is false if we remove the condition on $b$. More generally, let $1 \leq s < r \leq M$ and define three $\Glie$-Young tableaux $f,g,h$ of shape $Y$ as follows:
\begin{align*}
& f(i,j) = \begin{cases}
M+1 & (i \geq r-s, j = k), \\
i & \textrm{otherwise},
\end{cases} \quad g(i,j) = \begin{cases}
M+1 & (i > r-s, j = k), \\
r-s & (i = r-s,j = k), \\
i & \textrm{otherwise},
\end{cases} \\
& h(i,j) = \begin{cases}
M+1 & (i=r,j=k),\\
i & \textrm{otherwise}.
\end{cases}
\end{align*}
Then (1) in the proof of the Lemma is satisfied by taking $b = q^{-s}$.
\subsection{$q$-character and dual Gelfand-Tsetlin bases.} Our next task is to consider such $U_q(\Gaff)$-modules $\ev_a^* L(\lambda)^*$ with $\lambda \in \CP$.
\begin{prop} \label{prop: q-character and dual GT basis}
Let $\lambda \in \CP$. Let $(v_{\underline{\lambda}}^*: \underline{\lambda} \in \GT(\lambda))$ be a dual Gelfand-Tsetlin basis for the $\Uc_q(\Glie)$-module $L(\lambda)^*$ as in \S \ref{sec: dual GT basis}. Then there exist $f_{i,\underline{\lambda}}(z) \in \BC[[z]]$ for $i \in I$ and $\underline{\lambda} \in \GT(\lambda)$ such that in the $U_q(\Gaff)$-module $\ev_a^* L(\lambda)^*$
\begin{displaymath}
C_i(z) v_{\underline{\lambda}}^* = f_{i,\underline{\lambda}}(z) v_{\underline{\lambda}}^* \in \BC v_{\underline{\lambda}}[[z]].
\end{displaymath}
Furthermore, if $f_{i,\underline{\lambda}}(z) = f_{i,\underline{\mu}}(z)$ for $\underline{\lambda},\underline{\mu} \in \GT(\lambda)$, then $\lambda^{(i)} = \mu^{(i)}$. In particular, $\ev_a^*L(\lambda)^*$, viewed as a $Y_q(\Glie)$-module in category $\BGG$, has multiplicity-free $q$-character.
\end{prop}
Contrary to Proposition \ref{prop: q-character and GT basis}, we do not have a closed formula for $\chi_q(\ev_a^* L(\lambda)^*)$.
\begin{proof}
The existence of the $f_{i,\underline{\lambda}}(z)$ comes from the defining properties of the dual Gelfand-Tsetlin basis in \S \ref{sec: dual GT basis}. We only need to show that $f_{i,\underline{\lambda}} = f_{i,\underline{\mu}}$ implies $\lambda^{(i)} = \mu^{(i)}$.
For $\underline{\lambda} \in \GT(\lambda)$, let $\lambda_b^{(i)}$ be the lowest weight of the $\Uc_q(\Glie_k)$-module $L(\lambda^{(i)};\Glie_k)$ defined in \S \ref{sec: Gelfand-Tsetlin}. Then $\lambda^{(i)}_b = \mu^{(i)}_b$ if and only if $\lambda^{(i)} = \mu^{(i)}$. In the following, we show that $\lambda^{(i)}_b = \mu^{(i)}_b$ under the condition $f_{i,\underline{\lambda}} = f_{i,\underline{\mu}}$.
Assume first $1 \leq i \leq M$. Write $\lambda_b^{(i)} = \sum_{s=1}^i a_{i+1-s} \epsilon_s$ and $\mu_b^{(i)} = \sum_{s=1}^i a_{i+1-s}' \epsilon_s$. Then
\begin{displaymath}
a_1 \geq a_2 \geq \cdots \geq a_i,\quad a_1' \geq a_2' \geq \cdots \geq a_i'.
\end{displaymath}
Since $v_{\underline{\lambda}}^*$ is contained in a sub-$\Uc_q(\Glie_i)$-module of $L(\lambda)^*$ isomorphic to $L(-\lambda_b^{(i)};\Glie_i) \otimes \BC_{|\lambda - \lambda^{(i)}|}$, we get an explicit formula for $f_{i,\underline{\lambda}}$:
\begin{displaymath}
f_{i,\underline{\lambda}} = \prod_{s=1}^i (q^{-a_{i+1-s}} - z a q^{2(s-1)+ a_{i+1-s}}),\quad f_{i,\underline{\mu}} = \prod_{s=1}^i (q^{-a_{i+1-s}'} - z a q^{2(s-1) + a_{i+1-s}'}).
\end{displaymath}
Note that $s-1+a_{i+1-s} < t-1+a_{i+1-t}$ for $1 \leq s < t \leq i$. Hence $a_s = a_s'$ for $1 \leq s \leq i$.
Next assume $i = M+k$ with $1 \leq k \leq N$. Let us write
\begin{displaymath}
\lambda^{(M+k)}_b = \sum_{l=1}^M x_{M+1-l} \epsilon_l + \sum_{j=1}^k y_{k+1-j} \epsilon_{M+j},\quad \mu^{(M+k)}_b = \sum_{l=1}^M x_{M+1-l}' \epsilon_l + \sum_{j=1}^k y_{k+1-j}' \epsilon_{M+j}.
\end{displaymath}
Then the $x_l,y_j$ (similar for the $x_l',y_j'$) verify the following conditions:
\begin{itemize}
\item[(A)] $x_1 \geq x_2 \geq \cdots \geq x_M$ and $y_1 \geq y_2 \geq \cdots \geq y_k$;
\item[(B)] if $y_j \leq r$, then $x_{r+1} = 0$ (with the convention $x_{M+1} = x_{M+2} = \cdots = 0$).
\end{itemize}
Our aim is to show that $x_l = x_l'$ and $y_j = y_j'$ for $1 \leq l \leq M$ and $1 \leq j \leq k$. Note that $f_{M+k,\underline{\lambda}} = f_{M+k,\underline{\mu}}$ implies $\frac{N(z)}{D(z)} = \frac{N'(z)}{D'(z)}$
where
\begin{displaymath}
N(z) = \prod_{l=1}^M (1 - za q^{2(l-1) + 2 x_{M+1-l}} ), \quad D(z) = \prod_{j=1}^k (1 - z a q^{2(M-k-1+ j)- 2 y_j } )
\end{displaymath}
and $N'(z),D'(z)$ are similarly defined using $x_l',y_j'$. We shall prove $N(z) = N'(z)$. If one of the fractions $\frac{N(z)}{D(z)}, \frac{N'(z)}{D'(z)}$ is in reduced form, then by counting the zeros and the poles we get $N(z) = N'(z)$. Suppose therefore neither $\frac{N(z)}{D(z)}$ nor $\frac{N'(z)}{D'(z)}$ is reduced.
Let $P$ (resp. $Z$) the set of poles (resp. zeros) not including $\infty$ of the rational function $\frac{N(z^{-1})}{D(z^{-1})}$. Let $P_0$ (resp. $Z_0$) be the set of zeros of $D(z^{-1})$ (resp. $N(z^{-1})$). Define similarly $P',Z',P_0',Z_0'$. Then
\begin{displaymath}
P = P',\quad Z = Z', \quad P \subsetneq P_0,\quad Z \subsetneq Z_0.
\end{displaymath}
Moreover, there exists $1 \leq j_1 < j_2 < \cdots < j_s \leq k$ such that
\begin{displaymath}
P_0 \setminus P = \{aq^{2(M-k-1+j_t - y_{j_t})} \ |\ 1 \leq t \leq s \}.
\end{displaymath}
One can find a unique $1 \leq i_s \leq M$ such that $aq^{2(i_s-1+x_{M+1-i_s})} = aq^{2(M-k-1+j_s - y_{j_s})}$. So
\begin{displaymath}
y_{j_s} = M-k+j_s-i_s-x_{M+1-i_s} \leq M-k+j_s-i_s.
\end{displaymath}
Now Condition (B) above says that $x_{M-k+j_s-i_s+1} = 0$. In view of Condition (A) ($j_s \leq k$)
\begin{displaymath}
x_{M+1-i_s} = 0,\quad y_{j_s} = M-k+j_s-i_s.
\end{displaymath}
Again Condition (A) says that $x_{M+1-l} = 0$ for $1\leq l \leq i_s$. It follows that
\begin{displaymath}
\{aq^{2(l-1)+x_{M+1-l}} | i_s < l \leq M \} \subseteq Z,\quad N(z) = \prod_{l\leq i_s}(1-zaq^{2(l-1)})\prod_{l>i_s}(1-zaq^{2(l-1)+x_{M+1-l}}) .
\end{displaymath}
Similar analysis of the rational function $\frac{N'(z)}{D'(z)}$ leads to $i_s'$ with the above properties. We are therefore led to show that $i_s = i_s'$. Assume the contrary: say $i_s > i_s'$. Then observe first from $Z = Z'$ that
\begin{displaymath}
x_{M+1-l} = x_{M+1-l}' \quad \textrm{for}\ l > i_s.
\end{displaymath}
Next, we have $a q^{2(i_s-1+x_{M+1-i_s}')} \in Z' = Z$. Hence there must be $1 \leq j \leq M$ such that
\begin{displaymath}
\quad j-1+x_{M+1-j} = i_s - 1 + x_{M+1-i_s}'.
\end{displaymath}
Since $aq^{2(i_s-1+x_{M+1-i_s})} \notin Z$ by assumption, $j \neq i_s$.
When $j < i_s$, $x_{M+1-j} = 0$ and the above equation can not hold. So $j > i_s$. It follows that $x_{M+1-j} = x_{M+1-j}'$ and $i_s - 1 + x_{M+1-i_s}' = j-1 + x_{M+1-j}'$, which is impossible as $x_{M+1-i_s}' \leq x_{M+1-j}'$. This completes the proof.
\end{proof}
As in Proposition \ref{prop: q-character and GT basis}, all the $\ell$-weight spaces of $\ev_a^* L(\lambda)^*$ are one-dimensional of the form $\BC v_{\underline{\lambda}}$ with $\underline{\lambda} \in \GT(\lambda)$.
\begin{cor} \label{cor: normalized q character for dual polynomial representations}
$\tilde{\chi}_q (\ev_a^* L(\lambda)^*) \in \BZ[[A_{i,b}]^{-1} : i \in I_0, b \in aq^{2\BZ+1}]$ for $\lambda \in \CP$ and $a \in \BC^{\times}$.
\end{cor}
\begin{proof}
First observe from Definition \ref{def: generalized simple roots} that
\begin{displaymath}
(A_{i,a})_i = q \frac{1-zaq^{-1}}{1-zaq},\quad (A_{i,a})_j = 1 \quad \textrm{for}\ j \neq i.
\end{displaymath}
As in the proof of Corollary \ref{cor: normalized q-character for polynomial representations}, let $\lambda \in \CP$. Let $Y^{\lambda}$ be the associated $\Glie$-Young diagram. Let $f^- \in \CB(Y^{\lambda})$ be the $\Glie$-Young tableau such that the associated Gelfand-Tsetlin vector $v_{\GT(f^-)} \in L(\lambda)$ is a lowest weight vector. For simplicity, let us denote $\underline{\mu} := \GT(f^-)$. In view of Proposition \ref{prop: q-character and dual GT basis}, it is enough to show that for all $i \in I$ and for all $\underline{\lambda} \in \GT(\lambda)$, the power series $f_{i,\underline{\mu}}(z) f_{i,\underline{\lambda}}(z)^{-1}$ can be written as a product of $q \frac{1-zbq}{1-zbq^{-1}}$ with $b \in \BC^{\times}$. We consider the case $i = M+k$ with $1 \leq k \leq N$. (The case $i \leq M$ is similar and simpler.)
Let $g := GT_{\lambda}^{-1}(\underline{\lambda})$. Then as in the proof of Corollary \ref{cor: normalized q-character for polynomial representations} we have
\begin{displaymath}
g(x,y) \leq f^-(x,y) \quad \textrm{for}\ (x,y) \in Y^{\lambda}.
\end{displaymath}
As before, for $\theta \in \CP_{M_i,N_i}$, let $\theta_b$ be the lowest weight of the $\Uc_q(\Glie_i)$-module $L(\theta; \Glie_k)$. Write
\begin{displaymath}
\mu^{(M+k)}_b = \sum_{j=1}^M x_{M+1-j} \epsilon_j + \sum_{l=1}^k y_{k+1-j} \epsilon_{M+j}, \quad \lambda^{(M+k)}_b = \sum_{j=1}^M x_{M+1-j}' \epsilon_j + \sum_{l=1}^k y_{k+1-j}' \epsilon_{M+j}.
\end{displaymath}
Then as in the proof of Proposition \ref{prop: q-character and GT basis} we have
\begin{align*}
& f_{i,\underline{\mu}} = \prod_{j=1}^M (q^{-x_{M+1-j}} - zaq^{2(j-1) + x_{M+1-j}}) \times \prod_{l=1}^k (q^{y_{k+1-l}} - zaq^{2(M-l) - y_{k+1-l}})^{-1}, \\
& f_{i,\underline{\lambda}} = \prod_{j=1}^M (q^{-x'_{M+1-j}} - zaq^{2(j-1) + x'_{M+1-j}}) \times \prod_{l=1}^k (q^{y'_{k+1-l}} - zaq^{2(M-l) - y'_{k+1-l}})^{-1}.
\end{align*}
It is therefore enough to show that $x_j \leq x_j'$ and $y_l \leq y_l'$ for $1 \leq j \leq M$ and $1 \leq l \leq k$.
Let us verify $y_l \leq y_l'$ for $1 \leq l \leq k$. (The case for $x$ is parallel.) For this, we use the description in \S \ref{sec: dual GT basis} on the relationship between Young diagrams and lowest weight vectors. Namely, we have (Definition \ref{def: Gelfand-Tsetlin patterns})
\begin{eqnarray*}
y_{k+1-j} &=& \sharp \{ x \in \BZ_{>0} \ |\ (x, N+1-j) \in Y^{\lambda} \} = \sharp (f^-)^{-1}(M+k), \\
y'_{k+1-j} &=& \sharp \{ x \in \BZ_{>0} \ |\ (x, k+1-j) \in g^{-1}(I_{M_i,N_i}) \}.
\end{eqnarray*}
Suppose that $(x,N+1-j) \in Y^{\lambda}$. Then $f^-(x,N+1-j) = M+k$. It follows that $g(x,N+1-j) \leq M+k = i$. Hence $g(x,k+1-j) \leq g(x,N+1-j) \leq i$. In other words,
\begin{displaymath}
\{ x \in \BZ_{>0} \ |\ (x, N+1-j) \in Y^{\lambda} \} \subseteq \{ x \in \BZ_{>0} \ |\ (x, k+1-j) \in g^{-1}(I_{M_i,N_i}) \}.
\end{displaymath}
This says that $y_l \leq y_l'$, as desired.
\end{proof}
\begin{example} \label{example: q character gl(1,2)}
Let $(M,N) = (1,2)$ and $\Glie = \mathfrak{gl}(1,2)$. Consider the $U_q(\Gaff)$-module $W_{k,a}^{(2)}$. Let us compute its normalized $q$-character by using dual Gelfand-Tsetlin basis. The $\Glie$-Young diagram in this case is
\begin{displaymath}
Y = \{ (i,1) \in \BZ^2\ |\ 1 \leq i \leq k \}.
\end{displaymath}
There are $2k+1$ $\Glie$-Young tableaux of shape $Y$, namely $f_i,g_j$ with $0 \leq i \leq k, 0 \leq j \leq k-1$:
\begin{displaymath}
f_i(l,1) = \begin{cases}
2 & (1 \leq l \leq i), \\
3 & (i < l \leq k),
\end{cases}\quad g_j(l,1) = \begin{cases}
1 & (l = 1), \\
2 & (1<l \leq j+1), \\
3 & (j+1 < l \leq k).
\end{cases}
\end{displaymath}
Let $v_i,w_j \in W_{k,a}^{(2)}$ be dual Gelfand-Tsetlin vectors associated with $f_i,g_j$. Then $v_0$ is a highest $\ell$-weight vector. The action of $C_q(\Glie)$ on these vectors becomes:
\begin{align*}
& C_1(z) v_i = (1-za) v_i, \quad C_1(z) w_j = (q^{-1} - z a q) w_j, \\
& C_2(z) v_i = (1-za) (q^i - z a q^{-i})^{-1} v_i, \quad C_2(z) w_j = (1-za)(q^{j+1}-za q^{-j-1})^{-1} w_j.
\end{align*}
It follows that
\begin{equation*}
\widetilde{\chi}_q (W_{k,a}^{(2)}) = 1 + (1 + [A_{1,aq}]^{-1}) \sum_{l=1}^k \prod_{j=1}^l [A_{2,aq^{-2j+1}}]^{-1}.
\end{equation*}
\end{example}
\section{Gelfand-Tsetlin basis of asymptotic representations} \label{sec: GT for asymptotic modules}
In this section, we study the normalized $q$-character for the asymptotic modules $L_{r,a}^+$ and $L_{r,a}(b)$ constructed in previous sections \S \ref{sec: asymptotic representations}-\ref{sec: generic asymptotic representations} and we establish Gelfand-Tsetlin basis for these modules (provided that $b$ is generic).
\subsection{Gelfand-Tsetlin basis.} \label{sec: definition of GT}
For $k \in I$, let $Y_q(\Glie_k)$ be the subalgebra of $Y_q(\Glie)$ generated by the $s_{ij}^{(n)},(s_{ii}^{(0)})^{-1}$ with $i,j \leq k$. Then there exists canonical isomorphisms of superalgebras
\begin{displaymath}
Y_q(\mathfrak{gl}(M_k,N_k)) \cong Y_q(\Glie_k) \subset U_q(\Gaff_k) \cong U_q(\widehat{\mathfrak{gl}(M_k,N_k)}).
\end{displaymath}
We have therefore a chain of sub-superalgebras of $Y_q(\Glie)$:
\begin{displaymath}
Y_q(\Glie_1) \subset Y_q(\Glie_2) \subset \cdots \subset Y_q(\Glie_{M+N-1}) \subset Y_q(\Glie_{M+N}) = Y_q(\Glie).
\end{displaymath}
Let $V$ be a $Y_q(\Glie)$-module. A basis $(v_j)_{j \in \Lambda}$ of $V$ is called a {\it Gelfand-Tsetlin basis} if
\begin{itemize}
\item[(B1)] $V$ is semi-simple as a $Y_q(\Glie_k)$-module for all $k \in I_0$ and simple as a $Y_q(\Glie)$-module;
\item[(B2)] for all $j \in \Lambda$ there exists a sequence $(S_s)_{s\in I}$ such that: $S_s$ is a simple sub-$Y_q(\Glie_s)$-module; $S_s \subset S_{s+1}$ for $1 \leq k < M+N$; $v_j \in S_1$;
\item[(B3)] for all $1 \leq k \leq M+N$ and for $S$ a simple sub-$Y_q(\Glie_s)$-module, the decomposition of $\Res_{Y_q(\Glie_{s-1})}^{Y_q(\Glie_s)} S$ into simple sub-$Y_q(\Glie_{s-1})$-modules is multiplicity-free.
\end{itemize}
Now we can state the central result of this section.
\begin{theorem} \label{thm: main result GT basis of all asymtotic modules}
Let $r \in I_0, a,b \in \BC^{\times}$ be such that $b \notin \pm q^{\BZ}$. Then the $Y_q(\Glie)$-modules $L_{r,a}^+$ and $L_{r,a}(b)$ admit Gelfand-Tsetlin bases.
\end{theorem}
The idea of proof goes as follows. Fix $r \in I_0$. There exists an inductive system $(V_k, F_{k,l})$ constructed in \S\ref{sec: positive asymptotic modules}-\ref{sec: generic asymptotic representations} where $V_k$ are properly defined Kirillov-Reshetikhin modules. The representations of $Y_q(\Glie)$ on $L_{r,a}^+$ and on $L_{r,a}(b)$ are built upon the same underlying space, namely the inductive limit $(V_{\infty}, F_k)$. For all $k \in \BZ_{>0}$, the $Y_q(\Glie)$-module $V_k$ admits a Gelfand-Tsetlin basis. Furthermore the structural maps $F_{k,l}: V_l \longrightarrow V_k$ respect Gelfand-Tsetlin bases on both sides. From this we construct directly a basis of $V_{\infty}$ which serves at the same time as a Gelfand-Tsetlin basis for $L_{r,a}^+$ and for $L_{r,a}(b)$.
The next two subsections are devoted to the detailed proof of this theorem in cases $r \leq M$ and $r > M$. Along the proof, we shall find the simple sub-$Y_q(\Glie_k)$-modules in (B2) in a combinatorial way (by using semi-infinite Young tableaux).
Before the proof, let us make the following convention.
For $1 \leq s \leq M+N$, related to the $q$-Yangian $Y_q(\Glie_s) \cong Y_q(\mathfrak{gl}(M_s,N_s))$ we can define in exactly the same way the category $\BGG_s$, the group $\ADP_{M_s,N_s}$, the generalized simple roots $A_{i,a;s} \in \ADP_{M_s,N_s}$ for $1 \leq i < s$, the ring $(\CEl)_s$, the $q$-character $\chi_q^s$ and the normalized $q$-character $\tilde{\chi}_q^s$ . We identify $\ADP_{M_s,N_s}$ (resp. $(\CEl)_s$) with a subset of $\ADP$ (resp. $\CEl$) in the following natural way:
\begin{displaymath}
f = (f_i: 1 \leq i \leq k) \mapsto f' = (f_i': 1 \leq i \leq M+N),\quad f_i' = \begin{cases}
f_i & (i \leq k), \\
1 & (i > k).
\end{cases}
\end{displaymath}
Then under this identification, $A_{i,a;s} = A_{i,a}$ for $1 \leq i < k$ as seen from Definition \ref{def: generalized simple roots}.
\subsection{Proof of Theorem \ref{thm: main result GT basis of all asymtotic modules} when $r \leq M$.} \label{sec: part 1 of proof}
Fix $a \in \BC^{\times}$. Let us be in the situation of the first paragraph of \S \ref{sec: even generic asymptotic modules}, so that we have an inductive system of vector superspaces $(V_k, F_{k,l})$ with inductive limit $(V_{\infty}, F_k)$. Note that we have fixed a highest $\ell$-weight vector $v_k \in V_k$ for all $k \in \BZ_{>0}$ such that $F_{k,l} v_l = v_k$ whenever $k > l$. Moreover, we take $v_{\infty} = F_1(v_1)$. Denote $(\rho^k,V_k), (\rho^+,V_{\infty}), (\rho^b, V_{\infty})$ the representations of $Y_q(\Glie)$-modules $W_{k,aq^{2k}}^{(r)}, L_{r,a}^+, L_{r,a}(b)$. Let $t: \BZ_{>0} \longrightarrow \BZ_{>0}$ be a strictly increasing function such that the $F_{k,t(l)} \rho^k(s_{ij}^{(n)}) F_{k,l}$ for $k > t(l)$ are well-defined. (See Proposition \ref{prop: first properties of inductive system}.)
\subsubsection{Compatibility of Gelfand-Tsetlin bases and structural maps.} Fix $l,k \in \BZ_{>0}$ such that $l < k$. Choose for all $l > 0$ a Gelfand-Tsetlin basis $(v_{\underline{\lambda}}: \underline{\lambda} \in \GT(l\varpi_r))$ of $V_l$ satisfying the properties of Corollary \ref{cor: Gelfand-Tsetlin basis for modules in BKK}. (Note that $V_k \cong L(k\varpi_r)$ as $\Uc_q(\Glie)$-modules.)
\begin{lem} \label{lem: GT bases vs structural maps}
In the above situation, we have: for all $\underline{\lambda} = (\lambda^{(i)}) \in \GT(l\varpi_r)$,
\begin{displaymath}
F_{k,l} v_{\underline{\lambda}} \in \BC^{\times} v_{ \underline{\mu}}
\end{displaymath}
with $\underline{\mu} = (\mu^{(i)}) \in \GT(k\varpi_r)$ defined by $\mu^{(i)} = \lambda^{(i)} + (k-l) \begin{cases}
\varpi_i & (i \leq r), \\
\varpi_r & (i > r).
\end{cases}$
\end{lem}
\begin{proof}
This comes from Remark \ref{rem: q character and tensor product} combined with (the proof of) Proposition \ref{prop: q-character and GT basis} on relationship between Gelfand-Tsetlin vectors and $\ell$-weight spaces and on multiplicity-free property of $q$-character, in view of the definition of $F_{k,l}$ in \S\ref{sec: positive asymptotic modules}.
\end{proof}
\subsubsection{Semi-infinite Young tableau and Gelfand-Tsetlin bases.} Based on Lemmas \ref{lem: GT bases vs structural maps} and \ref{lem: GT patterns vs Young tableaux}, we introduce another index set of Gelfand-Tsetlin basis of $V_k$. This will be convenient for the statement of results.
Let $Y^{(r)}$ be the subset of $\BZ^2$ consisting of $(i,j)$ such that $1 \leq i \leq r$ and $j < 0$. Let $\CB^{(r)}$ be the set of functions $f: Y^{(r)} \longrightarrow I$ satisfying (T1)-(T3) in Definition \ref{def: Young combinatorics} and: for all $1 \leq i \leq r$, there exists $j < 0$ such that $f(i,j) = i$. Such an $f$ is also called a {\it semi-infinite Young tableau}. For $k > 0$, let $\CB^{(r)}_k$ be the subset of $\CB^{(r)}$ consisting of such $f$ that $f(i,j) = i$ for $j < -k$. We have therefore a chain of subsets of $\CB^{(r)}$:
\begin{displaymath}
\CB_1^{(r)} \subseteq \CB_2^{(r)} \subseteq \CB_3^{(r)} \subseteq \cdots, \quad \CB^{(r)} = \bigcup_{k=1}^{\infty} \CB_k^{(r)}.
\end{displaymath}
There is a canonical bijective map $\pi_k:\CB(Y^{k\varpi_r}) \longrightarrow \CB_k^{(r)}$ sending $g$, a $\Glie$-Young tableau of shape $Y^{k\varpi_r}$, to a semi-infinite Young tableau $f$ where $f(i,j) = \begin{cases}
i & (j < -k), \\
g(i,j+k+1) & (j \geq -k).
\end{cases}$
Let $f_0 \in \CB^{(r)}$ be such that $f_0(i,j) = i$. By using Lemma \ref{lem: GT patterns vs Young tableaux}, Corollary \ref{cor: Gelfand-Tsetlin basis for modules in BKK} and Lemma \ref{lem: GT bases vs structural maps}, one can construct by induction on $k > 0$, a Gelfand-Tsetlin basis $\{ v[k,f] | f \in \CB_k^{(r)} \}$ of $V_k$ verifying the following properties.
\begin{itemize}
\item[(a)] $v[k,f_0] = v_k$ for all $k > 0$.
\item[(b)] Let $l < k$ and $f \in \CB_l^{(r)}$. Then $F_{k,l} v[l,f] = v[k,f]$.
\item[(c)] $v[k,f] \in \BC^{\times} v_{\underline{\mu}}$ with $\underline{\mu} = GT_{k\varpi_r} (\pi_k^{-1}f) \in \GT(k\varpi_r)$.
\end{itemize}
For $f \in \CB^{(r)}$, define $v[f] := F_k v[k,f]$ whenever $f \in \CB_k^{(r)}$. Note that $v[f] \in V_{\infty}$ is well-defined thanks to (b). Moreover $v[f_0] = v_{\infty}$, and $\{v[f]\ |\ f \in \CB^{(r)} \}$ is a basis of $V_{\infty}$.
For $f \in \CB^{(r)}$ and $s \in I$, let $S_s(f)$ be the sub-vector-superspace of $V_{\infty}$ spanned by the $v[g]$ where $g \in \CB^{(r)}$ satisfies $g(i,j) = f(i,j)$ whenever $f(i,j) > s$ or $g(i,j) > s$. Clearly
\begin{displaymath}
v[f] \in S_1(f) \subseteq S_2(f) \subseteq \cdots \subseteq S_{M+N-1}(f) \subseteq S_{M+N}(f) = V_{\infty}.
\end{displaymath}
For $1\leq i \leq \min(r,s)$, define $l_i := \sharp \{ j<0 \ |\ f(i,j) > s \}$. Set $\lambda_s(l,f) := \sum_{i=1}^{\min(r,s)} (l-l_i) \epsilon_i$. Let $f_{\downarrow s} \in \CB^{(r)}$ be the semi-infinite Young tableau obtained from $f$ by replacing any $f(i,j)\leq s$ with $i$. At last,
let $\Phi^l,\Phi^+,\Phi^b \in \ADP$ such that
\begin{displaymath}
\rho^l(C_i(z)) v_l = (\Phi^l)_i(z) v_l,\quad \rho^+(C_i(z)) v_{\infty} = (\Phi^+)_i(z) v_{\infty},\quad \rho^b(C_i(z)) v_{\infty} = (\Phi^b)_i(z) v_{\infty}.
\end{displaymath}
The precise formulas for $\Phi^l,\Phi^+,\Phi^b$ will be given in the proof of Lemma \ref{lem: auxiliary}.
\begin{lem} \label{lem: inductive system 1}
Let $f \in \CB_l^{(r)}$ and $s \in I$. The sub-vector-superspace $F_l^{-1} S_s(f) \subseteq V_l$ is a simple sub-$Y_q(\Glie_s)$-module isomorphic to $\ev_{aq^{2l}}^* L(\lambda_s(l,f); \Glie_s) \otimes D$ where $D$ is a sign module. Moreover, $v[l,f_{\downarrow s}]$ is a highest $\ell$-weight vector of the $Y_q(\Glie_s)$-module $F_l^{-1}S_s(f)$.
\end{lem}
\begin{proof}
Indeed, $F_{l}^{-1} S_s(f)$ is the sub-vector-superspace of $V_l$ spanned by the $v[l,g]$ where $g \in \CB_l^{(r)}$ and $g(i,j) = f(i,j)$ whenever $g(i,j) > s$ or $f(i,j) > s$. In view of (c), the map $\pi_l$, and the correspondence between Young tableaux and Gelfand-Tsetlin patterns in Lemma \ref{lem: GT patterns vs Young tableaux}, the lemma is evident.
\end{proof}
\begin{lem} \label{lem: inductive system 2}
For $s \in I$ and $f \in \CB^{(r)}$, the sub-vector-superspace $S_s(f)$ of $V_{\infty}$ is stable by $\rho^x(Y_q(\Glie_s))$ whenever $x \in \{+,b\}$.
\end{lem}
\begin{proof}
By definition, $S_s(f)$ is stable by the $\rho^x(s_{ii}^{(0)})$. (See also Lemma \ref{lem: auxiliary} below for more general statements.) It is enough to show that $S_s(f)$ is stable by the $\rho^x(\widehat{s}_{ij}^{(n)})$ where $n \in \BZ_{\geq 0}$ and $1 \leq i,j \leq s$. Let us assume $f \in \CB_{v}^{(r)}$.
For $k > v$, the preceding lemma says that $F_{k}^{-1} S_s(f) \subseteq V_k$ is stable by the $\rho^k(\widehat{s}_{ij}^{(n)})$ with $1 \leq i,j \leq s$. It follows that ($v < k < t(k) < u$)
\begin{eqnarray*}
&\ & F_{u,t(k)}^{-1} \rho^u(\widehat{s}_{ij}^{(n)}) \underline{F_{u,k} F_k^{-1}} S_s(f) \subseteq F_{u,t(k)}^{-1} \underline{ \rho^u(\widehat{s}_{ij}^{(n)}) F_u^{-1} S_s(f)} \\
&\subseteq & \underline{F_{u,t(k)}^{-1} F_{u}^{-1}} S_s(f) \subseteq F_{t(k)}^{-1} S_s(f).
\end{eqnarray*}
On the other hand, we know from \S \ref{sec: positive asymptotic modules} the following asymptotic formula
\begin{displaymath}
F_{u,t(k)}^{-1} \rho^u(\widehat{s}_{ij}^{(n)}) F_{u,k} = A_k + q^{2u} B_k \quad \textrm{for}\ u > t(k)
\end{displaymath}
where $A_k,B_k: V_k \longrightarrow V_{t(k)}$ are linear operators depending only on $i,j,n$ and inductively on $k$. As $u$ can be arbitrarily large, we must have
\begin{displaymath}
A_k F_k^{-1} S_s(f) \subseteq F_{t(k)}^{-1} S_s(f),\quad B_k F_k^{-1}S_s(f) \subseteq F_{t(k)}^{-1} S_s(f)
\end{displaymath}
for all $k > v$. But this says exactly that $\rho^+(\widehat{s}_{ij}^{(n)})$ (resp. $\rho^b(\widehat{s}_{ij}^{(n)})$), being defined as inductive limit of the $A_k$ (resp. $A_k + b^2 B_k$), stabilizes $S_s(f)$, as desired.
\end{proof}
Based on the proof of the above lemma, the following lemma is clear.
\begin{lem} \label{lem: hwv of general asymptotic modules}
Let $f \in \CB^{(r)}$ and let $s \in I$. Then $v[f_{\downarrow s}] \in S_s(f)$ is a highest $\ell$-weight vector for the $Y_q(\Glie_s)$-modules $(\rho^x,S_{f})$ whenever $x \in \{+,b\}$. Moreover
\begin{align*}
& \rho^b(s_{ii}(z)) v[f_{\downarrow s}] = v[f_{\downarrow s}] \begin{cases}
b(q^{-l_i} - zaq^{l_i}) & (1 \leq i \leq \min(r,s)), \\
1 - z a b^2 & (\min(r,s) < i \leq s),
\end{cases} \\
& \rho^+(s_{ii}(z)) v[f_{\downarrow s}] = v[f_{\downarrow s}] \begin{cases}
q^{-l_i} - z a q^{l_i} & (1 \leq i \leq \min(r,s)), \\
1 & (\min(r,s) < i \leq s).
\end{cases}
\end{align*}
Here as usual $l_i := \sharp \{j< 0 \ |\ f(i,j) > s \}$ for $1 \leq i \leq \min(r,s)$.
\end{lem}
\begin{rem} \label{rem: more general inductive system even}
Lemmas \ref{lem: inductive system 1}-\ref{lem: hwv of general asymptotic modules} can be rephrased as follows. The original inductive system $(V_k, F_{k,l})$ admits a sub-inductive-system of vector superspaces of the form:
\begin{displaymath}
( \ev_{aq^{2k}}^* L(\lambda_s(k,f);\Glie_s), F_{k,l} ),\quad \lambda_s(k,f) = \sum_{i=1}^{\min(r,s)} (k - l_i) \epsilon_i.
\end{displaymath}
The two asymptotic constructions in \S \ref{sec: positive asymptotic modules} and in \S \ref{sec: even generic asymptotic modules} can be restricted to this sub-inductive-system, resulting in asymptotic modules $(\rho^+,S_s(f))$ and $(\rho^b,S_s(f))$ over $Y_q(\Glie_s)$.
\end{rem}
\begin{lem} \label{lem: auxiliary}
For $f \in \CB^{(r)}$, there exists uniquely $\Phi_f \in \ADP$ such that whenever $f \in \CB^{(r)}_l$:
\begin{displaymath}
\rho^l(C_i(z)) v[l,f] = (\Phi^l\Phi_f)_i(z) v[l,f],\quad \rho^x(C_i(z)) v[f] = (\Phi^x\Phi_f)_i(z) v[f],\quad x \in \{+,b\}, i \in I.
\end{displaymath}
Moreover, $\Phi_f$ is a product of the $A_{i,x}^{-1}$ where $i \in I_0$ and $x \in a q^{2\BZ + 1}$.
\end{lem}
\begin{proof}
For such $l$ and $f$, $v[l,f]$ is a Gelfand-Tsetlin vector. In view of Proposition \ref{prop: q-character and GT basis}, $\BC v[l,f]$ is an $\ell$-weight space of $(\rho^l,V_l)$. Hence by Corollary \ref{cor: normalized q-character for polynomial representations} there exists a unique $\Phi_f$ as a product of the $A_{i,x}^{-1}$ with $i \in I_0, x \in a q^{2\BZ +1}$ such that
\begin{displaymath}
\rho^l (C_i(z)) v[l,f] = (\Phi^l\Phi_f)_i(z) v[l,f].
\end{displaymath}
Now Lemma \ref{lem: GT bases vs structural maps} says that $\rho^k(C_i(z)) v[k,f] = (\Phi^k\Phi_f)_i(z) v[k,f]$ for all $k > l$.
To compute $\rho^b(C_i(z)) v[f]$, remark that for all $k > l$, by assumption $F_{k,l} v[l,f] = v[k,f]$. Hence for all $k > t(l)$
\begin{displaymath}
F_{k,t(l)}^{-1} \rho^k(C_i(z))F_{k,l} v[l,f] = (\Phi^k\Phi_f)_i(z) v[t(l),f]
\end{displaymath}
On the other hand, in view of the following explicit formulas for $\Phi^k,\Phi^+,\Phi^b$:
\begin{eqnarray*}
(\Phi^k)_i(z) &=& \begin{cases}
q^{ki} \prod_{j=1}^i (1-za\theta_j^{-1}) & (i \leq r), \\
q^{kr} \prod_{j=1}^r (1-za\theta_j^{-1}) \times \prod_{j=r+1}^i(1-za\theta_j^{-1}q^{2k})^{d_j} & (i > r),
\end{cases} \\
(\Phi^+)_i(z) &=& \begin{cases}
\prod_{j=1}^i (1-za\theta_j^{-1}) & (i \leq r), \\
\prod_{j=1}^r (1-za\theta_j^{-1}) & (i > r),
\end{cases} \\
(\Phi^b)_i(z) &=& \begin{cases}
b^i \prod_{j=1}^i (1-za\theta_j^{-1}) & (i \leq r), \\
b^r \prod_{j=1}^r (1-za\theta_j^{-1}) \times \prod_{j=r+1}^i(1-za\theta_j^{-1}b^2)^{d_j} & (i > r),
\end{cases}
\end{eqnarray*}
we see from the construction of $\rho^b$ in \S \ref{sec: even generic asymptotic modules} that (by replacing $q^k$ with $b$ everywhere)
\begin{displaymath}
\rho^b (C_i(z)) v[f] = (\Phi^k \Phi_f)_i(z)|_{q^k \rightarrow b} v[f] = (\Phi^b\Phi_f)_i(z) v[f].
\end{displaymath}
Finally, for $\rho^+(C_i(z)) v[f]$, from the above formulas of $\Phi^k,\Phi^+$ we deduce that there exists a formal power series $h_i(w) \in 1 + w\BC[[w]]$ for $i \in I$ such that for all $k > t(l)$,
\begin{eqnarray*}
&&(\Phi^k)_i(z) (\Phi^k)_i(0)^{-1} = (\Phi^+)_i(z) h_i(zq^k), \quad (\Phi^+)_i(0) = 1 \\
&&F_{k,t(l)}^{-1} \rho^k(C_i(z)C_i(0)^{-1}) F_{k,l} v[l,f] = h_i(zq^k) (\Phi^+ \Phi_f)_i(z) (\Phi^+\Phi_f)_i(0)^{-1} v[t(l),f].
\end{eqnarray*}
From Gauss decomposition observe that $C_i(z)C_i(0)^{-1} \in \widehat{Y}_q(\Glie)[[z]]$. Now from the construction of $\rho^+$ in \S \ref{sec: positive asymptotic modules} we obtain (by taking $q^k = 0$)
\begin{displaymath}
\rho^+(C_i(z)C_i(0)^{-1}) v[f] = (\Phi^+\Phi_f)_i(z) (\Phi^+\Phi_f)_i(0)^{-1} v[f].
\end{displaymath}
It is therefore enough to compute $\rho^+(C_i(0)) v[f]$. For this, note that in $(\rho^l,V_l)$,
\begin{displaymath}
|v[l,f]|_{\BP} = l\varpi_r + \sum_{i=1}^r\sum_{j=1}^l (\epsilon_{f(i,j-l-1)} - \epsilon_i).
\end{displaymath}
Hence in $(\rho^+,V_{\infty})$, the $\BQ$-degree of $v[f]$ is given by the second term of the RHS. From the proof of Proposition \ref{prop: q-character and GT basis} we get an explicit expression of the $(\Phi_f)_i(0)$:
\begin{displaymath}
\rho^+(C_i(0)) v[f] = (\Phi_f)_i(0) v[f] = (\Phi^+\Phi_f)_i(0) v[f].
\end{displaymath}
In consequence, $\rho^+(C_i(z)) v[f] = (\Phi^+\Phi_f)_i(z) v[f]$, as desired.
\end{proof}
We arrive at the following important consequence of the preceding lemmas: simplicity of $S_{M+N}(f)$ for $\rho^+$ and $\rho^b$.
\begin{cor} \label{cor: simplicity of even asymptotic modules}
Let $1 \leq r \leq M$ and $a,b \in \BC^{\times}$. Then
\begin{displaymath}
\widetilde{\chi}_q (L_{r,a}^+) = \widetilde{\chi}_q (L_{r,a}(b)) = \lim_{k \rightarrow \infty} \widetilde{\chi}_q (W_{k,aq^{2k}}^{(r)})
\end{displaymath}
as formal power series in the $[A_{i,x}]^{-1}$ where $i \in I_0$ and $x \in a q^{2\BZ+1}$. Furthermore, $L_{r,a}^+$ is a simple $Y_q(\Glie)$-module. $L_{r,a}(b)$ is simple if $b \notin \pm q^{\BZ_{>-r}}$.
\end{cor}
\begin{proof}
Let us explain firstly that the limit of normal $q$-characters above makes sense. For $V$ a module in category $\BGG$ with well-defined normalized character (Remark \ref{rem: normalized q character}), set
\begin{displaymath}
\widetilde{\chi}_q (V) = \sum_{\Phi \in \ADP} d(\Phi, V) [\Phi] \in \CEl.
\end{displaymath}
Then according to Proposition \ref{prop: q-character and GT basis}, for $k \in \BZ_{>0}$ and $\Phi \in \ADP$, either $d(\Phi, V_k) = 0$, or $d(\Phi, V_k) = 1$, in which case $\Phi$ is a product of the $A_{i,x}^{-1}$ with $i \in I_0, x \in a q^{2\BZ+1}$. Furthermore, Lemma \ref{lem: GT bases vs structural maps} indicates that for $l < k$
\begin{displaymath}
d(\Phi, V_l) \leq d(\Phi, V_k).
\end{displaymath}
Combining with the facts that $\widetilde{\chi}_q(V)$ is a refinement of $\widetilde{\chi}(V)$ and that $\lim\limits_{k\rightarrow \infty} \widetilde{\chi}(V_k)$ exists (Equation \eqref{equ: normalized character for asymptotic modules}), we conclude that so does $\lim\limits_{k\rightarrow\infty} \widetilde{\chi}_q (V_k)$, and it is a formal power series in the $[A_{i,x}]^{-1}$ (where $i \in I_0, x \in a q^{2\BZ+1}$) with coefficients $0$ or $1$.
Next the character formulas are immediate consequences of Lemma \ref{lem: auxiliary}:
\begin{displaymath}
\widetilde{\chi}_q (L_{r,a}^+) = \widetilde{\chi}_q (L_{r,a}(b)) = \sum_{f \in \CB^{(r)}} [\Phi_f] = \lim_{k \rightarrow \infty} \widetilde{\chi}_q (W_{k,aq^{2k}}^{(r)}).
\end{displaymath}
Thirdly we show that $L_{r,a}^+$ is a simple $Y_q(\Glie)$-module, imitating the proof of \cite[Theorem 6.1]{HJ}. According to Proposition \ref{prop: finite-dimensional and extension}, $L_{r,a}^+$ has a simple sub-quotient, denoted by $S_{r,a}^+$, which contains a highest $\ell$-weight vector $w$ such that $s_{ii}(z) w = w \begin{cases}
1 - z a & (i \leq r), \\
1 & (i > r).
\end{cases}$ Furthermore, for $x \in \BC^{\times}$, there exists a simple $Y_q(\Glie)$-module $S_{r,x}^-$ in category $\BGG$ containing a highest $\ell$-weight vector $u_x$ such that $s_{ii}(z) u_x = u_x \begin{cases}
1 & (i \leq r), \\
1 - z x & (i > r).
\end{cases}$
Let us show that $\widetilde{\chi}_q (L_{r,a}^+) = \widetilde{\chi}_q(S_{r,a}^+)$. In other words, if $l > 0$ and $d(\Phi, V_l) = 1$, then $\Phi$ must appear in $\widetilde{\chi}_q (S_{r,a}^+)$. Indeed, for all $k > l$, we can find a one-dimensional $Y_q(\Glie)$-module $D$ such that:
\begin{itemize}
\item[(F)] $D \otimes V_k$ is a simple sub-quotient of $S_{r,a}^+ \otimes S_{r,aq^{2k}}^-$.
\end{itemize}
It follows from Remark \ref{rem: q character and tensor product} that $\Phi$ must appear in $\widetilde{\chi}_q (S_{r,a}^+) \widetilde{\chi}_q (S_{r,aq^{2k}}^-)$. In other words, there exists $\Phi_k^+$ (resp. $\Phi_k^-$) in $\widetilde{\chi}_q (S_{r,a}^+)$ (resp. in $\widetilde{\chi}_q (S_{r,aq^{2k}}^-)$) such that
\begin{displaymath}
\Phi = \Phi_k^+ \Phi_k^-.
\end{displaymath}
Since $\varpi(\Phi),\varpi(\Phi_{\pm}^k) \in - \BQ_{\geq 0}$, the set $\{\Phi_k^+ : k > l \}$ is finite. The pigeonhole principle indicates that there exists an infinite sequence $(k_n)_{n > 0}$ of integers such that
\begin{displaymath}
l < k_1 < k_2 < k_3 < \cdots, \quad \Phi_{k_1}^- = \Phi_{k_2}^- = \Phi_{k_3}^- = \cdots.
\end{displaymath}
Lemma \ref{lem: stability lemma} below says that $\Phi^{-}_{k_1} \in \sigma \DP$, so does $\Phi_{k_1}^+ \Phi^{-1}$. The later, as a product of the $A_{i,x}^{\pm 1}$, must be $1$. So $\Phi = \Phi^+_{k_1}$ appears in $\widetilde{\chi}_q (S_{r,a}^+)$, as desired.
At last we show that $L_{r,a}(b)$ is simple provided that $b \notin \pm q^{\BZ_{-r}}$. For this, let $S(r,a,b)$ be a simple sub-quotient of $L_{r,a}(b)$ containing a highest $\ell$-weight vector $u$ such that $s_{ii}(z) u = u \begin{cases}
b - z a b & (i \leq r), \\
1 - z a b^2 & (i > r).
\end{cases}$ Let $l > 0$. There exists a one-dimensional $Y_q(\Glie)$-module $D$ such that:
\begin{itemize}
\item[(F1)] $D \otimes V_l$ is a simple sub-quotient of $S(r,a,b) \otimes S(r,ab^2,b^{-1}q^l)$.
\end{itemize}
Assume $d(\Phi, V_l) = 1$. There exist a monomial $\Phi_1$ in $\widetilde{\chi}_q (S(r,a,b))$ and a monomial $\Phi_2$ in $\widetilde{\chi}_q (S(r,ab^2, b^{-1}q^l))$ such that $\Phi = \Phi_1 \Phi_2$. On the other hand, from the definition of $S(r,a,b)$ and from the normalized $q$-character of $L_{r,a}(b)$ we deduce that there exists $k > l$ such that
\begin{displaymath}
d(\Phi_1, W_{k,aq^{2k}}^{(r)}) = 1 = d(\Phi_2, W_{k,ab^2q^{2k}}^{(r)}).
\end{displaymath}
Note that $d(\Phi, W_{k,aq^{2k}}^{(r)}) = 1$ from Lemma \ref{lem: GT bases vs structural maps}. Now Lemma \ref{lem: factorization of q character} forces $\Phi_2 = 1$. Hence $\Phi = \Phi_1$ appears in $\widetilde{\chi}_q (S(r,a,b))$. It follows that $L_{r,a}(b) \cong S(r,a,b)$ is simple.
\end{proof}
\begin{lem} \label{lem: stability lemma}
Let $W$ be a $Y_q(\Glie)$-module in category $\BGG$ admitting normalized character. Let $(k_n)_{n \in \BZ_{>0}}$ be a strictly increasing sequence of integers. For $n \in \BZ_{>0}$ take $W_n := \Phi_{q^{2k_n}}^* W$. If $\Phi \in \ADP$ is such that $d(\Phi, W_n) > 0$ for all $n$, then $\Phi \in \sigma \DP$.
\end{lem}
\begin{proof}
Let $\theta: \ADP \longrightarrow \ADP$ be the group isomorphism $(f_i(z)) \mapsto (f_i(zq))$. Then for $n \in \BZ_{>0}$,
\begin{displaymath}
\widetilde{\chi}_q (W_n) = \sum_{\Psi \in \ADP} d(\Psi, W) [\theta^{2k_n} \Psi].
\end{displaymath}
One can therefore find $\Phi_n \in \ADP$ such that $\Phi = \theta^{2k_n} \Phi_n$ and $d(\Phi_n, W) > 0$. As $W$ is in category $\BGG$ and the $\varpi(\Phi_n)$ are the same, $\{\Phi_n\ | \ n > 0\}$ is a finite set. By pigeonhole principle, there exists a sub-sequence $(l_n)_{n\in \BZ_{>0}}$ of strictly increasing integers such that $\Phi_{l_1} = \Phi_{l_n}$ for all $n > 0$. It follows that $\Phi_{l_1} = \theta^{2l_2-2l_1} \Phi_{l_1}$. This forces $\Phi_{l_1} \in \sigma (\DP)$, so does $\Phi$.
\end{proof}
\begin{rem} \label{rem: simplicity of generic asymptotic modules}
Corollary \ref{cor: simplicity of even asymptotic modules} says that $L_{r,a}(b)$ is simple if $b \in \pm q^{\BZ_{>-r}}$. In the case $r = M$, it is possible to determine the $b \in \pm q^{\BZ_{-r}}$ making $L_{M,a}(b)$ non-simple as $L_{M,a}(b)$ is of dimension $2^{MN}$. In the case $r < M$, if $b \in \pm q^k$ with $k \geq 0$, then $L_{r,a}(b)$ has a finite-dimensional simple sub-quotient isomorphic to $W_{k,aq^{2k}}^{(r)}$ (up to tensor product by one-dimensional modules, with the convention that $W_{0,a}^{(r)}$ is the trivial module). Since $L_{r,a}(b)$ is infinite-dimensional, it is not simple. It remains to find integers $1 \leq s < r$ such that $L_{r,a}(\pm q^{-s})$ is not simple. Unfortunately, limited by Lemma \ref{lem: factorization of q character}, we still do not know the answer.
\end{rem}
The proof of Corollary \ref{cor: simplicity of even asymptotic modules} applied perfectly if we replace $S_{M+N}(f)$ with more general $S_s(f)$, we obtain the following corollary, whose proof is omitted, on normalized $q$-character formulas and on simplicity of asymptotic modules.
\begin{cor} \label{cor: simplicity and GT}
Let $f \in \CB_l^{(r)}, s \in I$ and let $b \in \CB^{\times}$.
\begin{itemize}
\item[(1)] The $Y_q(\Glie_s)$-modules $(\rho^+, S_s(f))$ and $(\rho^b, S_s(f))$ are in category $\BGG_s$ and
\begin{displaymath}
\widetilde{\chi}_q^s (\rho^+, S_s(f)) = \widetilde{\chi}_q^s (\rho^b, S_s(f)) = \lim_{k\rightarrow \infty} \widetilde{\chi}_q^s (\ev_{aq^{2k}}^* L(\lambda_s(k,f); \Glie_s))
\end{displaymath}
as formal power series in the $[A_{i,x}]^{-1}$ with $1 \leq i < s$ and $x \in a q^{2\BZ+1}$.
\item[(2)] The $Y_q(\Glie_s)$-module $(\rho^+, S_s(f))$ is simple, while the $Y_q(\Glie_s)$-module $(\rho^b,S_s(f))$ is simple if $b \notin \pm q^{\BZ}$.
\end{itemize}
\end{cor}
\subsubsection{End of proof of Theorem \ref{thm: main result GT basis of all asymtotic modules} when $r \leq M$.} \label{sec: end of proof}
Assume $b \notin \pm q^{\BZ}$. Let $x \in \{+,b\}$. Consider the $Y_q(\Glie)$-module $(\rho^x, V_{\infty})$. According to Corollary \ref{cor: simplicity and GT}, its basis $\{v[f]\ |\ f \in \CB^{(r)} \}$ together with the chains (one for each $f \in \CB^{(r)}$)
\begin{displaymath}
v[f] \in S_1(f) \subseteq S_2(f) \subseteq S_3(f) \subseteq \cdots \subseteq S_{M+N-1}(f) \subseteq S_{M+N}(f) = V_{\infty}
\end{displaymath}
satisfies (B1)-(B2) in the definition of a Gelfand-Tsetlin basis in \S \ref{sec: definition of GT}. Observe that for $f,g \in \CB^{(r)}$ and $s \in I$, by definition of $S_s(f)$, Lemma \ref{lem: hwv of general asymptotic modules} and Corollary \ref{cor: simplicity and GT} the following statements are equivalent:
\begin{itemize}
\item[(1)] $v[f] \in S_s(g)$;
\item[(2)] $\{(i,j) \in Y^{(r)}\ |\ f(i,j) > s \} = \{(i,j) \in Y^{(r)}\ |\ g(i,j) > s \}$;
\item[(3)] $(\rho^x,S_s(f))$ and $(\rho^x,S_s(g))$ are isomorphic as $Y_q(\Glie_s)$-modules;
\item[(4)] $S_s(f) = S_s(g)$.
\end{itemize}
Now (B3) on multiplicity-free property of the decomposition of $\Res_{Y_q(\Glie_{s-1})}^{Y_q(\Glie_s)}(\rho^x, S_s(f))$ into simple sub-$Y_q(\Glie_{s-1})$-modules, which are the $S_{s-1}(g)$, is obvious. \hfill $\Box$
\subsection{Proof of Theorem \ref{thm: main result GT basis of all asymtotic modules} when $r > M$.} The proof is almost identical to the case $r \leq M$. We explain the main steps.
Take $V_l := W_{l,a}^{(r)}$ which is a tensor product of $\ev_a^*L(l\varpi_r)$ with a sign module. We borrow the notations from \S \ref{sec: part 1 of proof}, so that $v_k \in V_l, v_{\infty} \in V_{\infty}, F_{k,l}, F_k, \rho^k, \rho^+, \rho^b, \Phi^k,\Phi^+,\Phi^b$ and the strictly increasing function $t$ will be used freely. In the present case,
\begin{eqnarray*}
(\Phi^k)_i(z) &=& \begin{cases}
\prod_{j=1}^i (1-za\theta_j^{-1})^{d_j} & (i \leq r), \\
\prod_{j=1}^r(1-za\theta_j^{-1})^{d_j} \times \prod_{j=r+1}^i (q^k - z a \theta_j^{-1} q^{-k})^{-1} & (i > r),
\end{cases} \\
(\Phi^b)_i(z) &=& \begin{cases}
\prod_{j=1}^i (1-za\theta_j^{-1})^{d_j} & (i \leq r), \\
\prod_{j=1}^r(1-za\theta_j^{-1})^{d_j} \times \prod_{j=r+1}^i (b - z a \theta_j^{-1} b^{-1})^{-1} & (i > r),
\end{cases} \\
(\Phi^+)_i(z) &=& \begin{cases}
\prod_{j=1}^i (1-za\theta_j^{-1})^{d_j} & (i \leq r), \\
\prod_{j=1}^r(1-za\theta_j^{-1})^{d_j} & (i > r).
\end{cases}
\end{eqnarray*}
Set $Y^{(r)}$ to be the subset of $\BZ^2$ consisting of $(i,j)$ such that $1 \leq j \leq M+N-r$ and $i > 0$. Define $\CB^{(r)}$ to be the set of functions $f: Y^{(r)} \longrightarrow I$ satisfying (T1)-(T3) in Definition \ref{def: Young combinatorics} and: for all $1\leq j \leq M+N-r$, there exists $i>0$ such that $f(i,j) = r+j$. Let $\CB^{(r)}_k$ be the subset of $\CB^{(r)}$ consisting of $f$ such that $f(k+1,j) = r+j$ for all $1 \leq j \leq M+N-r$.
For $s \in I$ and $f \in \CB^{(r)}_l$, the subset $Y_s(l,f)$ of $Y^{(r)}$ consisting of $(i,j)$ such that $i\leq l$ and $f(i,j) \leq s$ is easily seen to be a $\Glie_s$-Young diagram. Set $\mu_s(l,f) := \Ym_{M_s,N_s}(Y_s(l,f))$. (See Remark \ref{rem: notations for parabolic subalgebras gl(a,b)}.) Let $\lambda_s(l,f) \in \BP_{M_s,N_s}$ be the highest weight of the simple $\Uc_q(\Glie_s)$-module $L(\mu_s(f);\Glie_s)^*$. Then we have the following observation.
\begin{itemize}
\item[(1)] If $s \leq r$, then $\lambda_s(l,f) = \lambda_s(k,f)$ whenever $l < k$ and $f \in \CB^{(r)}_{l}$.
\item[(2)] If $s > r$, then $\lambda_s(k,f) = \lambda_s(l,f) - (k-l)\sum_{j=r+1}^s \epsilon_j$ whenever $l < k$ and $f \in \CB^{(r)}_{l}$.
\end{itemize}
Set $f_0 \in \CB^{(r)}$ to be such that $f(i,j) = j+r$ for all $1 \leq j \leq M+N-r$.
According to \S \ref{sec: dual GT basis}, for all $l > 0$, there is a dual Gelfand-Tsetlin basis $\{ v_{l,f}\ |\ f \in \CB^{(r)}_l \}$ of $V_l$ such that for all $s \in I$ and $f \in \CB_l^{(r)}$ the vector $v_{l,f}$ is contained in a simple sub-$Y_q(\Glie_s)$-module of $V_l$ of the form:
\begin{displaymath}
\ev_a^* L(\lambda_s(l,f);\Glie_s) \otimes D \cong \ev_a^* L(\mu_s(l,f);\Glie_s)^* \otimes D
\end{displaymath}
where $D$ is a sign module (irrelevant to us). Now from Proposition \ref{prop: q-character and dual GT basis}, Corollary \ref{cor: normalized q character for dual polynomial representations}, their proofs, and the construction of $F_{k,l}$ in \S \ref{sec: odd asymptotic modules} we conclude that
\begin{lem} \label{lem: dual GT and q character}
For $l < k, f \in \CB_l^{(r)}$ we have $F_{k,l} v_{l,f} \in \BC^{\times} v_{k,f}$.
\end{lem}
Now we can choose by induction on $k$ a basis $\{v[k,f]\ |\ f \in \CB_k^{(r)} \}$ of $V_k$ such that
\begin{itemize}
\item[(a)] $F_{k,l} v[l,f] = v[k,f]$ whenever $k > l$ and $f \in \CB_l^{(r)}$;
\item[(b)] $v[l,f_0] = v_l$ for all $l > 0$;
\item[(c)] $v[l,f] \in \BC^{\times} v_{l,f}$ for $f \in \CB_l^{(r)}$.
\end{itemize}
Define $v[f] := F_k v[k,f] \in V_{\infty}$ whenever $f \in \CB^{(r)}_k$. As before, for $f \in \CB^{(r)}$ and $s \in I$, let $S_s(f)$ be the sub-vector-superspace of $V_{\infty}$ spanned by the $v[g]$ where $g \in \CB^{(r)}$ verifies $g(i,j) = f(i,j)$ whenever one of them is bigger than $s$.
Now Corollary \ref{cor: simplicity of even asymptotic modules} stays the same in our situation, as the proof of which requires only the multiplicity-free property of normalized $q$-character of evaluation modules. As in \S \ref{sec: end of proof}, we conclude the proof of Theorem \ref{thm: main result GT basis of all asymtotic modules} in the case $r > M$. \hfill $\Box$
\begin{rem} \label{rem: more general inductive system odd}
Remark \ref{rem: more general inductive system even} remains true in our situation. Let $s \in I$ and let $f \in \CB^{(r)}_l$. Write
\begin{displaymath}
\lambda_s(l,f) = \sum_{i=1}^s x_i \epsilon_i.
\end{displaymath}
Let $v[f_{\downarrow s}] \in S_s(f)$ be a highest $\ell$-weight vector for both $(\rho^+,S_s(f))$ and $(\rho^b,S_s(f))$. If $s \leq r$, then $S_s(f)$ is finite-dimensional and for $1 \leq i \leq s$
\begin{displaymath}
\rho^+(s_{ii}(z)) v[f_{\downarrow s}] = \rho^b(s_{ii}(z)) v[f_{\downarrow s}] = (q_i^{x_i} - za q_i^{-x_i}) v[f_{\downarrow s}].
\end{displaymath}
If $s > r$, then $S_s(f)$ is infinite-dimensional and
\begin{eqnarray*}
\rho^+(s_{ii}(z)) v[f_{\downarrow s}] &=& v[f_{\downarrow s}] \begin{cases}
q_i^{x_i} - z a q_i^{-x_i} & (1 \leq i \leq r), \\
q^{-x_i-l} & (r < i \leq s),
\end{cases} \\
\rho^b(s_{ii}(z)) v[f_{\downarrow s}] &=& v[f_{\downarrow s}] \begin{cases}
q_i^{x_i} - z a q_i^{-x_i} & (1 \leq i \leq r), \\
bq^{-x_i-l} - z a b^{-1} q^{l+x_i} & (r < i \leq s).
\end{cases}
\end{eqnarray*}
In this case, $(\rho^+,S_s(f))$ and $(\rho^b,S_s(f))$ can be seen as the asymptotic modules obtained from the following sub-inductive-system
\begin{displaymath}
( \ev_{a}^* L(\lambda_s(k,f);\Glie_s) \otimes D_k, F_{k,l} )
\end{displaymath}
of the original one $(V_k,F_{k,l})$ by carrying out the asymptotic constructions in \S\ref{sec: odd asymptotic modules} and \S \ref{sec: odd generic asymptotic modules}.
Here the $D_k$ are some properly defined sign modules.
\end{rem}
\subsection{Consequences on normalized $q$-character.} Let us go back to \S \ref{sec: simple in O}. The following corollary is now direct in view of Corollary \ref{cor: simplicity of even asymptotic modules} (which holds for all $r \in I_0$) applied to $S_{M+N}(f)$ and Examples \ref{example: asymptotic modules simple modules}-\ref{example: asymptotic modules generic simple modules}.
\begin{cor}
Let $f = (f_i: i \in I_0) \in \CR$. Assume that $f_i(\infty) \neq 0$ for all $i \in I_0$. Then $\widetilde{\chi}_q (V(f))$ is a formal power series in the $[A_{i,x}]^{-1}$ where $i \in I_0$ and $x \in \BC^{\times}$. In particular, if $S$ is a finite-dimensional simple $U_q(\Gaff)$-module, then the normalized $q$-character $\widetilde{\chi}_q (S)$ is a polynomial in the $[A_{i,x}]^{-1}$.
\end{cor}
Let $r \in I_0$ be such that $M< r < M+N$. Consider the $U_q(\Gaff)$-module $L_{r,a}(b)$. By pulling it back with respect to $f_{J,I}: U_q(\Gafft) \longrightarrow U_q(\Gaff)$, we get a $U_q(\Gafft)$-module $f_{J,I}^* L_{r,a}(b)$ such that
\begin{displaymath}
s_{ii;J}(z) f_{J,I}^* v = f_{J,I}^* v \begin{cases}
b-zab^{-1} & (i \leq M+N-r), \\
1-za & (i > M+N-r)
\end{cases}
\end{displaymath}
where $v \in L_{r,a}(b)$ is a highest $\ell$-weight vector of $L_{r,a}(b)$. By comparing the normalized characters and highest $\ell$-weights of $f_{J,I}^* L_{r,a}(b)$ and $L_{M+N-r,ab^{-2};J}(b)$ we conclude that
\begin{cor} \label{cor: simplicity of odd asymptotic modules}
Let $M< r < M+N$ and $a,b \in \BC^{\times}$. Assume that $b \notin \pm q^{\BZ_{>-M-N+r}}$. Then the $U_q(\Gaff)$-module $L_{r,a}(b)$ is simple.
\end{cor}
Let us rephrase Lemma \ref{cor: simplicity of even asymptotic modules} and Corollary \ref{cor: simplicity of odd asymptotic modules} in another way. Fix $i \in I_0$ and $a \in \BC^{\times}$. For $b \in \BC \setminus \{a\}$, set $\theta_{i,a,b} \in \CR$ to be
\begin{displaymath}
(\theta_{i,a,b})_j = 1\ (j \neq i),\quad (\theta_{i,a,b})_i = \frac{1-za}{1-zb}.
\end{displaymath}
Then whenever $b$ is generic (for example $b \notin q^{2\BZ}$), we have
\begin{displaymath}
\widetilde{\chi}_q (V(\theta_{i,a,b})) = \widetilde{\chi}_q (V(\theta_{i,a,0})).
\end{displaymath}
In other words, when $b$ is generic, $\widetilde{\chi}_q (V(\theta_{i,a,b}))$ is independent of $b$. Note that a similar phenomenon happens for more general rational functions than the $\frac{1-za}{1-zb}$, as we have observed in Equation \eqref{equ: q character simple module gl(1,1)} concerning the $q$-Yangian $Y_q(\mathfrak{gl}(1,1))$.
|
\section{Introduction}
Since Sedl\'{a}$\breve{\mbox{c}}$ek~\cite{se} introduced the notion of magic labeling of a graph, a variety of magic labelings of a graph have been defined and studied (see Gallian~\cite{Gal}). Kotzig and Rosa~\cite{KR} introduced the notion of a magic valuation of a graph in 1970. A {\em magic valuation} of a graph $G=(V,E)$ is a bijection $f$ from $V \cup E$ to $\{1,2,\ldots, |V \cup E|\}$ such that for all edge $xy$, $f(x)+f(y)+f(xy)$ is a constant called the {\em magic constant} of $f$. Given a graph $G$, a positive integer $k$ is said to be a magic constant of $G$ if $k$ is the magic constant of a magic valuation of $G$. Later, Ringel and Llad\'{o}~\cite{RL} rediscovered this notion and called it {\em edge-magic labeling}. More recently, Wallis~\cite{Wal} used the term {\em edge-magic total} labeling to distinguish it from other kinds of labeings that use the word magic.
An edge-magic total labeling is called a {\em $b$-edge consecutive magic labeling} if the edge labels are $\{b+1, b+2, \ldots, b+|E|\}$ where $0 \le b \le |V|$. Sugeng and Miller~\cite{SM} claimed that if a connected graph $G$ with $n$ vertices has a $b$-edge consecutive magic labeling with $1 \le b \le n-2$, then $G$ is a tree; the union of $r$ stars and a set of $r-1$ isolated vertices has an $r$-edge consecutive magic labeling.
In this paper, we extend these existing results about consecutive edge magic labelings of graphs in the following way. We find all the values of $b$ for which a connected bipartite graph has a $b$-edge consecutive magic labeling; We show that a connected bipartite graph $G=(X,Y)$ having an $|X|$-edge consecutive magic labeling is a tree having a graceful labeling and a super edge-magic labeling; We give a necessary and sufficient condition for a caterpillar having a $b$-edge consecutive magic labeling, which actually answers an open problem posed in Wallis~\cite{Wal}; We also obtain an interesting result on consecutive edge magic labelings for lobsters.
For any undefined term, the reader may refer to \cite{k10}.
\section{Properties of edge consecutive magic labelings of bipartite graphs}
We first present the following proposition which is simple but very useful.
\begin{Prop}
Suppose that $G$ has a $b$-edge consecutive magic labeling $\lambda$
for $1 \le b \le |V(G)|$. Then, if $y$ and $z$ are neighbors of $x$,
then either $\{\lambda(y),\lambda(z)\} \subset \{1,\ldots, b\}$ or $\{\lambda(y),\lambda(z)\} \subset \{b+|E(G)|+1, \ldots, |V(G)|+|E(G)|\}$.
\label{prop:neighbor}
\end{Prop}
\begin{proof}
Without loss of generality, we may assume that $\lambda(y) < \lambda (z)$. Suppose to the contrary that $\lambda(z) > b$ and $\lambda(y) < b+|E(G)|+1$. Then, by the definition of $b$-edge consecutive magic labeling, $\lambda(z) \ge b+|E(G)|+1$ and $\lambda(y) \le b$. Since $\lambda$ is an edge-magic total labeling,
\[\lambda(x)+\lambda(y)+\lambda(xy)=\lambda(x)+\lambda(z)+\lambda(xz),\]
which implies
\begin{equation}
\lambda(y)+\lambda(xy)=\lambda(z)+\lambda(xz)
\label{eq1}
\end{equation}
Since $\lambda$ is a $b$-edge consecutive magic labeling,
$\lambda(xz) \ge b+1$. From (\ref{eq1}) and the assumption that $\lambda(z) \ge b+|E(G)|+1$, we obtain
\[\lambda(y)+\lambda(xy) \ge (b+|E(G)|+1)+(b+1)=2b+|E(G)|+2.\]
Again, by the definition of $b$-edge consecutive magic labeling,
$\lambda(xy) \le b+|E(G)|$. Thus we have
\[\lambda(y) \ge 2b+|E(G)|+2-(b+|E(G)|)=b+2,\]
which contradicts the assumption that $\lambda(y) \le b$.
\end{proof}
The following theorem shows that there are only four possible values of $b$ for
which a connected bipartite graph has a $b$-edge consecutive magic labeling.
\begin{Thm}
If a connected bipartite graph $G=(X,Y)$ has a $b$-edge consecutive magic labeling, then $b \in \{0,|X|,|Y|,|X|+|Y|\}$.
\label{thm:bipartite}
\end{Thm}
\begin{proof}
For simplicity, we denote the set $\{1, 2, \ldots, b\}$ by $[b]$. By the definition of a $b$-edge consecutive magic labeling, it is sufficient to show that $b=|X|$ or $|Y|$ if $1 \le b < |X|+|Y|$. Suppose that $1 \le b < |X|+|Y|$. We claim that either $\lambda(X) \subset [b]$ or $\lambda(X) \subset \{b+|E(G)|+1,\ldots, |X|+|Y|+|E(G)|\}$. Take two vertices $x$ and $x'$ in $X$.
Since $G$ is connected, there is an $(x,x')$-path $P$ in $G$. Since $G$ is bipartite,
\[P=x_1w_1x_2w_2x_3\cdots x_lw_lx'\]
where $x=x_1$ and $x_i \in X$ and $w_i \in Y$ for $i=1$, $\ldots$, $l$. Then, by Proposition~\ref{prop:neighbor}, if $\lambda(x) \le b$, then $\lambda(x_i) \le b$ for $i=2$, $\ldots$, $l$ and so $\lambda(x') \le b$, and if $\lambda(x) \ge b+|E(G)|+1$, then $\lambda(x_i) \ge b+|E(G)|+1$ for $i=2$, $\ldots$, $l$ and so $\lambda(x') \ge b+|E(G)|+1$. Since $x$ and $x'$ were arbitrarily chosen, $\lambda(X) \subset [b]$ or $\lambda(X) \subset \{b+|E(G)|+1,\ldots, |X|+|Y|+|E(G)|\}$. Similarly we may show that $\lambda(Y) \subset [b]$ or $\lambda(Y) \subset \{b+|E(G)|+1,\ldots, |X|+|Y|+|E(G)|\}$.
If $\lambda(X) \subset [b]$ and $\lambda(Y) \subset [b]$, then $b \ge |X|+|Y|$, which contradicts the assumption. If $\lambda(X) \subset \{b+|E(G)|+1,\ldots, |X|+|Y|+|E(G)|\}$ and $\lambda(Y) \subset \{b+|E(G)|+1,\ldots, |X|+|Y|+|E(G)|\}$, then $b=0$, which contradicts the assumption again. Therefore if $\lambda(X) \subset \{b+|E(G)|+1,\ldots, |X|+|Y|+|E(G)|\}$, then $\lambda(Y) \subset [b]$, or $\lambda(Y) \subset \{b+|E(G)|+1,\ldots, |X|+|Y|+|E(G)|\}$, then $\lambda(X) \subset [b]$.
If $\lambda(X) \subset [b]$ and $\lambda(Y) \subset \{b+|E(G)|+1,\ldots, |X|+|Y|+|E(G)|\}$, then $|X| \le b$ and $|Y| \le |X|+|Y|-b$, which implies $b=|X|$. If $\lambda(X) \subset \{b+|E(G)|+1,\ldots, |X|+|Y|+|E(G)|\}$ and $\lambda(Y) \subset [b]$, then $|X| \le |X|+|Y|-b$ and $|Y| \le b$, which implies $b=|Y|$.
\end{proof}
Suppose that a graph $G$ with $m$ edges has a labeling of its vertices with some subset of $\{0,1,\ldots,m\}$ such that no two vertices share a label and the edge labels are the set $\{1,2,\ldots,m\}$ where an edge label is the difference of the values assigned to its end vertices. Then $G$ is said to be {\em graceful} and such a labeling is called a {\em graceful labeling} of $G$.
A $b$-edge consecutive magic labeling of a graph is called a {\it super edge-magic labeling} if $b=|V(G)|$. A graph $G$ is said to be {\it super edge-magic} if it has a super edge-magic labeling. Enomoto~{\it et al.}~\cite{ELNR} conjectured that every tree is super-edge magic, which still remains open.
We obtain an interesting result related to graceful labeling and super edge-magic labeling. As a matter of fact, it gives a stronger version of the result by Sugeng and Miller~\cite{SM} stating that if a connected graph $G$ with $n$ vertices has a $b$-edge consecutive magic labeling with $1 \le b \le n-2$, then $G$ is a tree. Furthermore, the conjecture given by Enomoto~{\it et al.}~\cite{ELNR} is true for a tree $T$ that has an $|X|$-edge consecutive magic labeling where $X$ is one of bipartitions of $T$ when it is considered as a bipartite graph.
We first show the following.
\begin{Thm} \label{nonbipartite} If a connected non-bipartite graph $G$ has a
$b$-edge consecutive magic labeling, then $b= 0$ or $|V(G)|$.
\end{Thm}
\begin{proof}
It suffices to show that if $b \ge 1$, then $b =|V(G)|$. Since $b \ge 1$, there is a vertex $x$ in $G$ such that $\lambda(x) \le b$. Take a vertex $y$ in $G$. Since $G$ is not a bipartite graph, there is an odd cycle $C$ in $G$. Let $w$ be a vertex on $C$. Since $G$ is connected, there is an $(x,w)$-path $P_1$ and a $(w,y)$-path $P_2$ in $G$.
Let $W$ be a walk which is obtained by concatenating $P_1$ and $P_2$ if the sum of lengths of $P_1$ and $P_2$ is even, and by concatenating $P_1$, $C$, and $P_2$ if the sum of lengths of $P_1$ and $P_2$ is odd. In both cases, the walk $W$ is an $(x,y)$-walk of even length. Thus, by Proposition~\ref{prop:neighbor}, $\lambda(y) \le b$ since $\lambda(x) \le b$. Since $y$ was arbitrarily chosen, $\lambda(v) \le b$ for any $v \in V(G)$. Thus $|V(G)| \le b$, and therefore $|V(G)| = b$. \end{proof}
\noindent Let $G$ be a connected graph $G$ with $n$ vertices having a $b$-edge consecutive magic labeling for some $b \in [n-1]$. Then, by the above theorem, $G$ is bipartite. If $b=n-1$, then one of the partite sets of $G$ has size $n-1$ by Theorem~\ref{thm:bipartite} and so $G$ is a tree. Hence $G$ is a tree even for $b=n-1$. The following theorem shows that $G$ should be a special tree that has both a graceful labeling and a super edge-magic labeling.
\begin{Thm}
If a connected bipartite graph $G=(X,Y)$ has an $|X|$-edge consecutive magic labeling, then $G$ is a tree having both a graceful labeling and a super edge-magic labeling.
\label{dual}
\end{Thm}
\begin{proof}
By the above observation, $G$ is a tree. We first show that $G$ has a graceful labeling.
Let $\lambda$ be an $|X|$-edge consecutive magic labeling of $G$. Then, by Proposition~\ref{prop:neighbor}, $\lambda(X)=\{1,\ldots, |X|\}$, or $|X|=|Y|$ and $\lambda(Y)=\{1,\ldots, |X|\}$. By symmetry, we may assume that $\lambda(X)=\{1,\ldots, |X|\}$. Then, by the definition of $|X|$-edge consecutive labeling, $\lambda(Y)=\{|X|+|E(G)|+1, \ldots, |X|+|Y|+|E(G)|\}$. Therefore
\begin{equation}
\lambda(X)=\{1,\ldots,|X|\}, \quad \lambda(Y)=\{|X|+|E(G)|+1,\ldots,|X|+|E(G)|+|Y|\},
\label{all}
\end{equation}
and
\[\lambda(E(G))=\{|X|+1, \ldots,|X|+|E(G)|\}.\]
Since $|\lambda(E(G))|=|E(G)|$ and $\lambda(x)+\lambda(y)+\lambda(xy)$ is constant for each edge $xy$ of $G$,
\begin{equation}
|\{\lambda(x)+\lambda(y) \mid x \in X, y \in Y, xy \in E(G)\}|=|E(G)|.
\label{edge}
\end{equation}
We define a labeling $\varphi$ from $X \cup Y$ to $\{0, 1, \ldots, |X|+|Y|-1\}$ as follows:\[\varphi(z)=
\begin{cases}
\lambda(z) -1 & \mbox{if } z \in X;
\\ |E(G)|+2|X|+|Y| - \lambda(z ) & \mbox{if } z \in Y.
\end{cases}
\]
\noindent Then, by (\ref{all}),
\begin{equation}
\varphi(X)=\{0,1,\ldots,|X|-1\} \quad\mbox{ and } \varphi(Y)=\{|X|, \ldots, |X|+|Y|-1\}
\label{sets}
\end{equation}
and
\[1 \le \varphi(y)-\varphi(x) \le |X|+|Y|-1.\]
Since $G$ is connected, $|X|+|Y|-1 \le |E(G)|$ and so
\begin{equation}
1 \le \varphi(y)-\varphi(x) \le |X|+|Y|-1 \le |E(G)|.
\label{absolute}
\end{equation}
Now, for any pair of adjacent vertices $x \in X$ and $y \in Y$, $\varphi(y) > \varphi(x)$ and
\begin{eqnarray}
\varphi(y) - \varphi(x)
&=& (|E(G)|+2|X|+|Y| - \lambda(y)) - (\lambda(x) -1) \notag \\
&=& |E(G)|+2|X|+|Y|+1 -(\lambda(x) + \lambda(y)). \label{eqs}
\end{eqnarray}
Thus, by (\ref{edge}) and (\ref{eqs}),
$S:= \{\varphi(y)-\varphi(x) \mid x \in X, y \in Y, xy \in E(G)\}$ has $|E(G)|$ elements.
Therefore, by (\ref{absolute}),
\[S=\{1, 2, \ldots, |E(G)|\}\]
which, together with (\ref{sets}), implies that $\varphi$ is a graceful labeling of $G$.
Now we show that $G$ has a super edge-magic labeling. Let $\varphi$ be an $|X|$-edge consecutive magic labeling. Then $\varphi(X)=\{1, \ldots, |X|\}$; $\varphi(E(G)) = \{|X|+1, \ldots, |X|+|E(G)|\}$; $\varphi(Y)= \{|X|+|E(G)|+1,\ldots, |X|+|Y|+|E(G)|\}$.
Now we define $\varphi^*:X \cup Y \cup E(G) \to \{1, \ldots, |X|+|Y|+|E(G)|\}$ by
\[\varphi^*(v)=\begin{cases} \varphi(v) & \mbox{ if } v \in X, \\
\varphi(v)-|E(G)| & \mbox{ if } v \in Y,\end{cases}\]
and, for each pair of adjacent vertices $x$ and $y$,
\[\varphi^*(xy)=\varphi(xy)+|Y|.\]
For an edge $xy$ of $G$ for $x \in X$, $y \in Y$,
\[\varphi^*(x)+\varphi^*(y)+\varphi^*(xy)=\varphi(x)+(\varphi(y)-|E(G)|)+(\varphi(xy)+|Y|),\]
which is a constant number. Thus $\varphi^*$ is an edge-magic labeling of $G$. By the definition of $\varphi^*$, $\varphi^*$ is an $(|X|+|Y|)$-edge consecutive magic labeling, that is, a super edge-magic labeling.
\end{proof}
Given a graph $G$, let $\gamma:V(G) \cup E(G) \to \{1,2,\ldots,|V(G)|+|E(G)|\}$ be an edge-magic labeling for a graph $G$. Define the labeling $\gamma':V(G) \cup E(G) \to \{1,2,\ldots,|V(G)|+|E(G)|\}$ as follows:
For a vertex $x$,
\[\gamma'(x) = |V(G)|+|E(G)|+1-\gamma(x),\]
for an edge $xy$,
\[\gamma'(xy)=|V(G)|+|E(G)|+1-\gamma(xy).\]
Then $\gamma'$ is called the {\em dual} of $\gamma$. From Wallis~\cite{Wal}, we know that the dual of an edge-magic labeling of a graph $G$ is also an edge-magic labeling of $G$. Moreover, if $k$ is the magic constant corresponding to $\gamma$, then for any adjacent vertices $x$ and $y$ of $G$,
\begin{align*}
&\gamma'(x)+\gamma'(y)+\gamma'(xy) \\ & =(|V(G)|+|E(G)|+1-\gamma(x))+(|V(G)|+|E(G)|+1-\gamma(y))+
(|V(G)|+|E(G)|+1-\gamma(xy)) \\
& =3(|V(G)|+|E(G)|+1)-(\gamma(x)+\gamma(y)+\gamma(xy))=3(|V(G)|+|E(G)|+1)-k,
\end{align*}
that is, $3(|V(G)|+|E(G)|+1)-k$ is the magic constant corresponding to $\gamma'$.
From the fact that the dual of an edge-magic labeling of a graph $G$ is also an edge-magic labeling of $G$, the following theorem is immediately true.
\begin{Thm}
For a connected bipartite graph $G=(X,Y)$, exactly one of the following is true:
\begin{itemize}
\item[(i)] $G$ does not have a $b$-edge consecutive magic labeling for any $b$;
\item[(ii)] $G$ has only $0$-edge consecutive magic labeling and super edge-magic labeling;
\item[(iii)] $G$ is a tree having a $b$-edge consecutive magic labeling for each $b=0$, $|X|$, $|Y|$, $|X|+|Y|$.
\label{dual1}
\end{itemize}
\label{eachcase}
\end{Thm}
\begin{proof}
We suppose that $G$ has a $b$-edge consecutive magic labeling $\lambda$ and denote the dual of $\lambda$ by $\lambda'$. By Theorem~\ref{thm:bipartite}, $b=0$, $|X|$, $|Y|$, or $|X|+|Y|$.
Since $|V(G)|=|X|+|Y|$, it is true that
$\lambda(X \cup Y)=\{1,\ldots,|X|+|Y|\}$ and $\lambda(E(G))=\{|X|+|Y|+1, \ldots, |X|+|Y|+|E(G)|\}$
if and only if $\lambda'(X \cup Y)=\{|E(G)|+1,\ldots,|X|+|Y|+|E(G)|\}$ and $\lambda'(E(G))=\{1, \ldots, |E(G)|\}$.
Similarly,
$\lambda(X)=\{1,\ldots,|X|\}$, $\lambda(E(G))=\{|X|+1, \ldots, |X|+|E(G)|\}$, and $\lambda(Y)=\{|X|+1+|E(G)|,\ldots,|X|+|Y|+E(G)\}$
if and only if $\lambda'(Y)=\{1,\ldots,|Y|\}$, $\lambda'(E(G))=\{|Y|+1, \ldots, |Y|+|E(G)|\}$, and $\lambda'(X)=\{|Y|+1+|E(G)|,\ldots,|X|+|Y|+|E(G)|\}$.
Thus, for $b=|X|+|Y|$ or $|X|$, then the dual of a $b$-edge consecutive magic labeling of $G$ is a $(|V(G)|-b)$-edge consecutive magic labeling of $G$ and vice versa.
Since $\lambda$ is the dual of $\lambda'$, this statement is true even for $b=0$ or $|Y|$. Therefore, if $b=|X|$ or $|Y|$, then, by the observation by Wallis~\cite{Wal} together with Theorems~\ref{dual}, the statement (iii) is true. If neither (i) nor (iii) is true, then $b=0$ or $|X|+|Y|$. By the above argument again, the statement (ii) is immediately true.
\end{proof}
Given a $b$-edge consecutive magic labeling $\lambda$ of a connected bipartite graph, there is a way of deducing a new $b$-edge consecutive magic labeling and a magic constant from $\lambda$ other than using the dual of edge magic labeling.
\begin{Prop}
If $\lambda$ is a $b$-edge consecutive magic labeling of a connected bipartite graph $G=(X,Y)$ with a magic constant $k$, then the mapping $\lambda^*:V(G) \cup E(G) \to \{1,\ldots,|V(G)|+|E(G)|\}$ defined by
\begin{itemize}
\item[(i)] $\lambda^*(x)=|V(G)|+2|E(G)|+1-\lambda(x)$ for a vertex $x$ and $\lambda^*(xy)=|E(G)|+1-\lambda(xy)$ for two adjacent vertex $x$ and $y$ is also a $b$-edge consecutive magic labeling of $G$ with magic constant $2|V(G)|+5|E(G)|+3-k$ if $b=0$;
\item[(ii)] $\lambda^*(x)=|X|+|Y|+1-\lambda(x)$ for a vertex $x$ and $\lambda^*(xy)=2|V(G)|+|E(G)|+1-\lambda(xy)$ for two adjacent vertex $x$ and $y$ is also a $b$-edge consecutive magic labeling of $G$ with magic constant $4|V(G)|+|E(G)|+3-k$ if $b=|V(G)|$;
\item[(iii)] $\lambda^*(x)=|X|+1-\lambda(x)$ for a vertex $x \in X$, $\lambda^*(y)=2|X|+|Y|+2|E(G)|+1-\lambda(y)$ for a vertex $y \in Y$, and $\lambda^*(xy)=2|X|+|E(G)|+1-\lambda(xy)$ for two adjacent vertex $x$ and $y$ is also a $b$-edge consecutive magic labeling of $G$ with magic constant $5|X|+|Y|+3|E(G)|+3-k$ if $b=|X|$;
\item[(iv)] $\lambda^*(x)=|X|+2|Y|+2|E(G)|+1-\lambda(x)$ for a vertex $x \in X$, $\lambda^*(y)=|Y|+1-\lambda(y)$ for a vertex $y \in Y$, and $\lambda^*(xy)=2|Y|+|E(G)|+1-\lambda(xy)$ for two adjacent vertex $x$ and $y$ is also a $b$-edge consecutive magic labeling of $G$ with magic constant $|X|+5|Y|+3|E(G)|+3-k$ if $b=|Y|$.
\end{itemize}
\label{lem:another}
\end{Prop}
\begin{proof} Suppose that $b=0$. Then
\begin{align*}
& \lambda^*(x)+\lambda^*(y)+\lambda^*(xy) \\ & \ \ =(|V(G)|+2|E(G)|+1-\lambda(x))+(|V(G)|+2|E(G)|+1-\lambda(y))+(|E(G)|+1-\lambda(xy))
\\ & \ \ =2|V(G)|+5|E(G)|+3-k,
\end{align*}
which is a constant number. For the remaining cases, it can similarly be checked.
\end{proof}
\section{Edge consecutive magic labelings for trees}
In the previous section, we have shown that if a connected bipartite graph $G$ has a $b$-consecutive labeling for some $b \in \{1,\ldots,|V(G)|-1\}$, then $G$ is a tree.
In this section, we study edge consecutive magic labelings of interesting families of trees.
A {\em caterpillar} is a tree derived from a path by joining leaves to the vertices of the path. We denote by $S_{n_1,n_2,\ldots,n_r}$ the caterpillar derived from a path $P_r=c_1c_2\cdots c_r$ for a positive integer $r$ by joining $n_i$ leaves to $c_i$, where $n_i$ is a nonnegative integer, for each $i=1$, $\ldots$, $r$. We denote the neighbors of $c_i$ by $c_{i,1}$, $\ldots$, $c_{i,n_i}$ for $i=1$, $\ldots$, $r$.
\begin{Thm}
The caterpillar $S_{n_1,n_2,\ldots,n_r}$ has a $b$-edge consecutive magic labeling if and only if \[t \in \{0,\sum_{i=1}^{\left\lceil\frac{r}{2}\right\rceil}n_{2i-1}+\left\lfloor\frac{r}{2}\right\rfloor,\sum_{i=1}^{\left\lfloor\frac{r}{2}\right\rfloor}n_{2i}+\left\lceil\frac{r}{2}\right\rceil,\sum_{i=1}^r n_i+r\}.\]
\label{thm:ctplr}
\end{Thm}
\begin{proof}
The caterpillar $S_{n_1,n_2,\ldots,n_r}$ is a connected bipartite graph. Let $(X,Y)$ be a bipartition of $S_{n_1,n_2,\ldots,n_r}$ such that
\begin{align*}
X&=\{c_1,c_{21},\ldots,c_{2n_2},c_3, \ldots, c_{r1},\ldots,c_{rn_r}\}; \\
Y&=\{c_{11}, \ldots, c_{1n_1},c_2,c_{31},\ldots,c_{3n_3} \ldots, c_{r-1,1},\ldots,c_{r-1,n_{r-1}},c_r\}
\end{align*}
if $r$ is even, and
\begin{align*} X&=\{c_1,c_{21},\ldots,c_{2n_2},c_3, \ldots, c_{r-1,1},\ldots,c_{r-1,n_{r-1}},c_r\}; \\
Y&=\{c_{11}, \ldots, c_{1n_1},c_2,c_{31},\ldots,c_{3n_3} \ldots, c_{r1},\ldots,c_{rn_r}\}
\end{align*}
if $r$ is odd. Then $|X|=\sum_{i=1}^{\lfloor\frac{r}{2}\rfloor}n_{2i}+\lceil \frac{r}{2} \rceil$ and $|Y|=\sum_{i=1}^{\lceil\frac{r}{2}\rceil}n_{2i-1}+\lfloor\frac{r}{2} \rfloor$.
Now the `only if' part immediately follows from Theorem~\ref{thm:bipartite}.
For a notational convenience, we let $\alpha=|X|=\sum_{i=1}^{\lfloor\frac{r}{2}\rfloor}n_{2i}+\lceil \frac{r}{2} \rceil$ and $\beta=|Y|=\sum_{i=1}^{\lceil\frac{r}{2}\rceil}n_{2i-1}+\lfloor\frac{r}{2} \rfloor$.
To show the `if' part, we give a a $\beta$-edge consecutive magic labeling first.
For each edge of $S_{n_1,n_2,\ldots,n_r}$, the labels of its end vertices are $c_i$ and $c_{i+1}$ for some $i \in \{1, \ldots, r-1\}$, or $c_{2i-1}$ and $c_{2i-1,j}$ for some $i \in \{1, \ldots, \lceil r/2 \rceil\}$ and $j \in \{1,\ldots, n_{2i-1}\}$, or $c_{2i}$ and $c_{2i,j}$ for some $i \in \{1, \ldots, \lceil r/2 \rceil\}$ and $j \in \{1,\ldots, n_{2i}\}$.
We define $\lambda:V(S_{n_1,n_2,\ldots,n_r})\cup E(S_{n_1,n_2,\ldots,n_r})\to\{1,2,\cdots,2\sum_{i=1}^rn_i + 2r-1\}$ by
\[\lambda(c_{2i-1}) =\alpha+2\beta-1+\sum_{l=1}^{i-1} n_{2l}+i; \quad
\lambda(c_{2i}) =\sum_{l=1}^{i} n_{2l-1}+i\] \[ \lambda(c_{2i-1,j})=\sum_{l=1}^{i-1} n_{2l-1}+i+j-1; \quad \lambda(c_{2i,j})=\alpha+2\beta-1+\sum_{l=1}^{i-1} n_{2l}+i+j\]
\[\lambda(c_{i}c_{i+1}) =\alpha+2\beta-i-\sum_{l=1}^{i}n_l \mbox{ for any } 1 \le i \le r-1;\]
\[\lambda(c_{2i-1}c_{2i-1,j}) =\alpha+2\beta-2i+2-\sum_{l=1}^{2i-2}n_l-j;\]
\[\lambda(c_{2i}c_{2i,j}) =\alpha+2\beta-2i+1-\sum_{l=1}^{2i-1}n_l-j\]
for $j=1$, $\ldots$, $n_{2i}$ and $i=1$, $\ldots$, $\left\lceil r/2 \right\rceil$. Then
\begin{align*} \lambda(c_{2i-1})+\lambda(c_{2i})+\lambda(c_{2i-1}c_{2i})&=\left(\alpha+2\beta-1+\sum_{l=1}^{i-1} n_{2l}+i\right)+\left(\sum_{l=1}^{i} n_{2l-1}+i\right)\\ & \ \ \ \ +\left(\alpha+2\beta-2i+1-\sum_{l=1}^{2i-1}n_l\right) =2\alpha+4\beta.
\end{align*} Similarly, one can check that
\[\lambda(c_{2i-1})+\lambda(c_{2i-1,j})+\lambda(c_{2i-1}c_{2i-1,j})=\lambda(c_{2i})+\lambda(c_{2i,j})+\lambda(c_{2i}c_{2i,j})=2\alpha+4\beta.\]
Thus $\lambda$ is a $\beta$-edge consecutive magic labeling. Hence the statement is true by Theorem~\ref{eachcase}.
\end{proof}
Since the double star $S_{m,n}$ is a special case of the caterpillar $S_{n_1,n_2,\ldots,n_r}$ for $r=2$, the following corollary is immediately true.
\begin{Cor}
The double star $S_{m,n}$ has a $b$-edge consecutive magic labeling if and only if $b \in \{0,m+1,n+1, m+n+2\}$.
\label{cor:doublestar}
\end{Cor}
\noindent Wallis~\cite{Wal} asked whether or not double stars are edge-magic. Since a consecutive edge magic labeling is an edge-magic labeling, the above proposition answers his question.
By Corollary~\ref{cor:doublestar}, the double star $S_{m,n}$ has an $(m+1)$-edge consecutive magic labeling and an $(n+1)$-edge consecutive magic labeling. We show that there are only two such labelings for each of $m$, $n$.
\begin{Prop}
For some positive integers $m$ and $n$, the double star $S_{m,n}$ has only two $(m+1)$-edge consecutive magic labelings (resp.\ $(n+1)$-edge consecutive magic labelings) both of which have magic constant $ 4m+2n+6$ (resp.\ $4n+2m+6$).
\label{thm:doublestarmagicnumber}
\end{Prop}
\begin{proof} We may regard $G:=S_{m,n}$ as a bipartite graph with bipartition $(X,Y)$ with $|X|=m+1$ and $|Y|=n+1$.
Let $\lambda$ be an $(m+1)$-edge consecutive magic labeling of $G$ and $k$ be the magic constant of $\lambda$. Then, by the definition of $(m+1)$-labeling and Proposition~\ref{prop:neighbor},
\begin{equation}
\lambda(X) =[m+1] \quad \mbox{ and } \quad \lambda(Y)=\{2m+n+3,\ldots,2m+2n+3\}
\label{eq:prop1}
\end{equation}
for all $x \in X$ and $y \in Y$.
Let $u$ and $v$ be the central vertices of $G$ and let $\lambda(u)=\alpha$ and $\lambda(v)=\beta$. Without loss of generality, we may assume that $u \in X$ and $v \in Y$. As $\alpha$ and $\beta$ belong to different partite sets and the vertices other than the central vertices form an independent set, we know from (\ref{eq:prop1}) that the labeling $\lambda$ is completely determined by $\alpha$ and $\beta$. Therefore it suffices to show that there are only two possible pairs of integers for $(\alpha,\beta)$.
By (\ref{eq:prop1}) and the assumption that $u \in X$ and $v \in Y$,
\[A:=\{\lambda(u)+\lambda(y) \mid y \in Y \}=\{\alpha+2m+n+3, \cdots, \alpha+2m+2n+3\}\]
and\[
B: = \{\lambda(v)+\lambda(x) \mid x \in X \} = \{\beta+1, \cdots, \beta+m+1\}.\]
Since $X \cap Y =\emptyset$, it is true that $A \cap B=\{\lambda(u)+\lambda(v)\}$, so $|A \cap B|=1$. Moreover, $X \cup Y=V(G)$ and each edge is incident to $u$ or $v$, so $A \cup B=\{ \lambda(x) + \lambda(y) \mid xy \in E(G)\}$.
Since $\lambda$ is an edge consecutive magic labeling, $\{k-\lambda(xy) \mid xy \in E(G)\}$ is a set of $m+n+1$ consecutive integers and therefore $\{ \lambda(x) + \lambda(y) \mid xy \in E(G)\}$ is a set of $m+n+1$ consecutive integers. Since $A \cup B=\{ \lambda(x) + \lambda(y) \mid xy \in E(G)\}$, $A \cup B$ is a set of $m+n+1$ consecutive integers. This together with the fact $|A \cap B|=1$ imply that there are only two possible cases:
\[\alpha+2m+n+3 = \beta+m+1 \quad \mbox{or} \quad \alpha+2m+2n+3 = \beta+1.\]
Assume the former. Since $2m+n+3 \le \beta$ and $\alpha \le m+1$, \[
(2m+n+3)+m+1 \le \beta+m+1 =\alpha+2m+n+3 \le (m+1)+2m+n+3.\]
Since the left hand side of the first inequality and the right hand side of the second inequality both equal $3m+n+4$, we have $\beta+m+1=3m+n+4$ and $\alpha+2m+n+3=3m+n+4$. Hence $\alpha=m+1$ and $\beta=2m+n+3$. Now assume the latter. Since $\beta \le 2m+2n+3$ and $1 \le \alpha$, \[
1+2m+2n+3 \le \alpha +2m+2n+3 = \beta+1 \le 2m+2n+3+1.\]
Thus $\beta=2m+2n+3$ and $\alpha=1$. We can easily check that the magic constant is $4m+2n+6$ in both cases.
By symmetry, the double star $S_{m,n}$ has only two $(n+1)$-edge consecutive magic labelings and their magic constant is $2m+4n+6$. \end{proof}
Proposition~\ref{thm:doublestarmagicnumber} tells us that magic constants of $(m+1)$-edge consecutive magic labelings and $(n+1)$-edge consecutive magic labelings for a double star are unique. As a matter of fact, the magic constants of a double star are of specific form.
\begin{Thm}
The magic constants of the double star $S_{m,n}$ are in the form of $dt+6$ for some nonnegative integer $t$ where $d$ is the greatest common divisor of $m$ and $n$.
\label{thm:magicconstant}
\end{Thm}
\begin{proof}
Suppose that the double star $G:=S_{m,n}$ has a $b$-edge consecutive magic labeling $\lambda$ and $k$ is a magic constant of $\lambda$. Let $x$ and $y$ be the central vertices of the double star $G$ and $\lambda(x)=i$ and $\lambda(y)=j$. Then $1 \le i, j \le 2m+2n+3$.
\begin{align}
k(m+n+1)&= \sum_{uv\in
E(G)}\left[\lambda(u)+\lambda(uv)+\lambda(v)\right] \notag
\\ &= \frac{ [1+(2m+2n+3)](2m+2n+3)}{2}+m\lambda(x) +n\lambda(y) \notag
\\ & = (m+n+2)(2m+2n+3)+mi +nj \notag
\\ & = (m+n+1)(2m+2n+5)+mi +nj+1. \label{eq2}
\end{align}
Since magic constant $k$ is a positive integer, $mi+nj+1$ is a multiple of $m+n+1$ by the equality (\ref{eq2}). That is, $mi+nj+1=l(m+n+1)$ for some positive integer $l$. Then, by (\ref{eq2}),
\begin{equation}
k=2m+2n+5+l
\label{sep}
\end{equation}
Since it is impossible for both $i$ and $j$ to equal $1$, we have $l \ge 2$. Let $d$ be the greatest common divisor of $m$ and $n$. Then $m=dm'$ and $n=dn'$ for relatively prime positive integers $m'$ and $n'$. Suppose that $d=1$, that is, $m$ and $n$ are relatively prime. Then, since $l \ge 2$, by the B\'{e}zout's identity, $l-1=\mu_1m+\nu_1n$ or $l=\mu_1m+\nu_1n+1$ for some integers $\mu_1$ and $\nu_1$. Then, by (\ref{sep}),
$k=2m+2n+\mu_1m+\nu_1n+6=d(2m+2n+\mu_1m+\nu_1n)+6$.
Now suppose $d \ge 2$. By the division algorithm, $l=dq+r$ for some integers $q$ and $r$ with $0 \le r \le d-1$. Then
\[mi+nj+1=(m+n+1)(dq+r)\]
or
\[m(i-dq-r)+(j-dq-r)n-dq=r-1.\]
Since the left hand side is divisible by $d$, $r-1$ is a multiple of $d$. Since $r \le d-1$, $r=1$ and so $l=dq+1$. Hence, by (\ref{sep}),
\[k=2m+2n+5+(dq+1)=d(2m'+2n'+q)+6\]
and we complete the proof.
\end{proof}
We may regard $S_{m,n}$ as a bipartite graph with bipartition $(X,Y)$ with $|X|=n+1$ and $|Y|=m+1$. In the proof of Theorem~\ref{thm:ctplr}, we have shown that $2(n+1)+4(m+1)=4m+2n+6$ is a magic constant for $b=n+1$. For the $\beta$-edge consecutive magic labeling given in the proof of Theorem~\ref{thm:ctplr} where $\beta=|Y|$, we define $\varphi: V(S_{n_1,n_2,\ldots,n_r})\cup E(S_{n_1,n_2,\ldots,n_r})\to\{1,2,\cdots,2\sum_{i=1}^rn_i + 2r-1\}$ by
$\varphi(y)=\lambda(y)$ for each $y \in Y$; $\varphi(x)=\lambda(x)-\alpha-\beta+1$ for each $x \in X$;
$\varphi(xy)=\lambda(xy)+\alpha$
for each pair of adjacent vertices $x$ and $y$ of $S_{n_1,n_2,\ldots,n_r}$. It can easily be checked that $\varphi$ is super edge-magic labeling by the fact that $\lambda$ is a $\beta$-edge consecutive magic labeling. Now we take two adjacent vertices $x$ and $y$ of the caterpillar. Then we may assume that $x \in X$ and $y \in Y$. By the definition, \[\varphi(x)+\varphi(y)+\varphi(xy)=(\lambda(x)-\alpha-\beta+1)+\lambda(y)+(\lambda(xy)+\alpha)=2\alpha+3\beta+1=3m+2n+6.\] Then, by Lemma~\ref{lem:another}(ii),
\[4|V(G)|+|E(G)|+3-(3m+2n+6)=4(m+n+2)+(m+n+1)+3-(3m+2n+6)=2m+3n+6\]
for $b=m+n+2$.
Furthermore, recalling that the dual $\gamma'$ of a $b$-edge consecutive magic labeling $\gamma$ with a magic constant $k$ is a $(|V(G)|-b)$-edge consecutive magic labeling with the magic constant $3(|V(G)|+|E(G)|+1)-k$, we obtain \[3(|V(G)|+|E(G)|+1)-(2m+3n+6)=4m+3n+6\] is a magic constant of $G$ for $b=0$. Then, by Lemma~\ref{lem:another}(i),
\[2|V(G)|+5|E(G)|+3-(4m+3n+6)=2(m+n+2)+5(m+n+1)+3-(4m+3n+6)=3m+4n+6\]
is another magic constant for $b=0$.
By the symmetry, $4(n+1)+2(m+1)=2m+4n+6$ is a magic constant for $b=m+1$.
In the rest of paper, we take a look at a special type of a lobster which is obtained from a star graph $G$ by attaching a leaf to each leaf of $G$. For a positive integer $p$, we denote by $L_p$ the lobster obtained from a star with $p$ leaves in such a way. In addition, we denote the center of $L_p$ by $x$, the vertices at distance $1$ from $x$ by $y_1$, $\ldots$, $y_p$, the vertex adjacent to $y_i$ by $x_i$ for each $i=1$, $\ldots$, $p$. Now the following is true for $L_p$.
\begin{Thm}
For $p \ge 3$, $L_p$ has a $b$-edge consecutive magic labeling if and only if $b \in \{0,2p+1\}$.
\label{thm:lobster}
\end{Thm}
\begin{proof}
Kim and Park~\cite{KP} showed that $L_p$ has an $(2p+1)$-edge consecutive magic labeling. Thus, by Theorem~\ref{dual1}, the `only if' part is true.
Now we show the `if' part. By Theorem~\ref{thm:bipartite}, $b \in \{0,p,p+1,2p+1\}$. For notational convenience, we denote $\{1,\ldots,p+1\}$ by $[p+1]$. To reach a contradiction, suppose that $b=p+1$. Then there is a $(p+1)$-edge consecutive magic labeling $\lambda$ of $L_p$ such that
\[\lambda(E(G))=\{p+2,\ldots, 3p+1\}.\]
By Proposition~\ref{prop:neighbor},
\[\lambda(\{x,x_1,\ldots,x_p\})=[p+1] \quad \mbox{and} \quad \lambda(\{y_1,\ldots,y_p\})=\{3p+2, \ldots,4p+1\}.\]
Set $\lambda(x)=i$. Then $i \in [p+1]$.
Let $k$ be the magic constant corresponding to $\lambda$. Then the values of $\lambda$ for the edges joining $x$ and the vertices $y_1$, $\ldots$, $y_p$ are
\[k-4p-i-1, \ldots, k-3p-i-2.\]
Suppose that $i+j \in [p+1]$ for some integer $j$. Then $i+j$ is assigned to a vertex in $\{x,x_1,\ldots,x_p\}$. The set of possible values of $\lambda$ for the edge joining a vertex in $\{y_1, \ldots, y_p\}$ and the vertex labeled with $i+j$ is
\[A(i,j):=\{k-[(i+j)+(4p+1)], \ldots, k-[(i+j)+(3p+2)]\}.\]
Then
\[A(i,1)=\{k-4p-i-2, \ldots, k-3p-i-3\}\]
and
\[A(i,-1):=\{k-4p-i, \ldots, k-3p-i-1\}.\]
We first note that all the elements of $A(i,1)$ except $k-4p-i-2=k-[(i+1)+(4p+1)]$ have been already assigned to edges joining $x$ and $y_1$, $\ldots$, $y_p$, and so the vertex labeled with $4p+1$ must be joined with the one labeled with $i+1$. We also note that all the elements of $A(i,-1)$ except $k-3p-i-1=k-[(i-1)+(3p+2)]$ are occupied by edges joining $x$ and $y_1$, $\ldots$, $y_p$, and so the vertex labeled with $3p+2$ must be joined with the one labeled with $i-1$
Now suppose that $i \ge 3$. Then
$i-2 \in [p+1]$ and $k-[(i-2)+(3p+2)]=k-3p-i$ is the only element in $A(i,-2)$ that was not assigned to edges joining $x$ and $y_1$, $\ldots$, $y_p$. Thus the vertex labeled with $i-2$ and the one labeled with $3p+2$ should be joined. However, $i-1 \in [p+1]$, and the vertex labeled with $3p+2$ must be joined to the one labeled with $i-1$ by the above argument and we reach a contradiction.
Now suppose that $i = 2$. Then $i+2=4 \in [p+1]$ since $p \ge 3$ and
\[A(i,2)=\{k-[(i+2)+(4p+1)], \ldots, k-[(i+2)+(3p+2)]\},\]
in which $k-[(i+2)+(4p+1)]=k-4p-i-3$ is the only available label. However, in that case, the vertex labeled with $i+2$ and the one labeled with $4p+1$ should be joined, which is a contradiction as the vertex labeled with $4p+1$ is already joined to the one labeled with $i+1$.
Now suppose that $i=1$. Then the values of $\lambda$ for the edges joining $x$ and vertices $y_1$, $\ldots$, $y_p$ are $k-4p-2$, $\ldots$, $k-3p-3$. Moreover, $2=i+1 \in [p+1]$ since $p \ge 3$, and
\[A(1,1)=\{k-4p-3, \ldots, k-3p-4\},\]
in which $k-4p-3=k-[(1+1)+(4p+1)]$ is the only available label. Thus the vertex labeled $2$ and the one labeled with $4p+1$ are joined. Now $3=i+2 \in [p+1]$ since $p \ge 3$, and
\[A(1,2)=\{k-4p-4, \ldots, k-3p-5\}\] in which $k-4p-4=k-[(1+2)+(4p+1)]$ is the only available label for the edge incident to the vertex labeled with $3$. Then, however, the other end vertex of the edge must be labeled with $4p+1$, which is impossible as the vertex labeled with $4p+1$ is adjacent to the one labeled with $2$.
Suppose that $i=p+1$. Then the values of $\lambda$ for the edges joining $x$ and vertices $y_1$, $\ldots$, $y_p$ are $k-5p-2$, $\ldots$, $k-4p-3$. Now for $p=i-1$,
\[A(p+1,-1)=\{k-5p-1, \ldots, k-4p-2\}\]
in which $k-4p-2=k-[(p+1)+(-1)+(3p+2)]$ is the only available label for the edge incident to the vertex labeled with $p$. Thus the vertex labeled $p$ and the one labeled with $3p+2$ are joined. On the other hand, for $p-1=i-2$,
\[A(p+1,-2)=\{k-5p, \ldots, k-4p-1\}\]
in which $k-5p=k-[(p+1)+(-2)+(4p+1)]$ is the only available label for the edge incident to the vertex labeled with $p-1$. Then, however, the vertex labeled with $p-1$ and the one labeled $3p+2$ should be joined, which is impossible as the vertex labeled with $3p+2$ also must be adjacent to the one labeled with $p$.
Thus there is no $(p+1)$-edge consecutive magic labeling for $L_p$. Then, by Theorem~\ref{dual1}, the statement is true.
\end{proof}
\begin{Rem} For $p=1$ or $2$, the above theorem is false. We may assign $(p+1)$-edge consecutive magic labelings to $L_1$ and $L_2$, which are the paths $P_3$ and $P_5$ of lengths $2$ and $4$, respectively (see Figure~\ref{fig:counterexample}).
\end{Rem}
\begin{figure}
\centering
\psfrag{1}{$1$}\psfrag{1}{$1$}\psfrag{2}{$2$}\psfrag{3}{$3$}\psfrag{4}{$4$}
\psfrag{5}{$5$}\psfrag{6}{$6$}\psfrag{7}{$7$}\psfrag{8}{$8$}
\psfrag{9}{$9$}\psfrag{A}{$L_1$}\psfrag{B}{$L_2$}
\includegraphics[width=5cm]{lobster.eps}\\
\caption{A $2$-edge consecutive magic labeling for $L_1$ and a $3$-edge consecutive magic labeling for $L_2$. }\label{fig:counterexample}
\end{figure}
\begin{Rem} By Theorem~\ref{thm:lobster}, we know that the converse of Theorem~\ref{dual} is false as the lobster $L_4$ is super edge-magic but it does not have a $b$-edge consecutive magic labeling for $b=4$ or $5$. Furthermore, $L_4$ is graceful as shown in Figure~\ref{fig:graceful}.
\end{Rem}
\begin{figure}
\centering
\psfrag{0}{\small \bf 0}\psfrag{1}{\small \bf 1}\psfrag{2}{\small \bf 2}\psfrag{3}{\small \bf 3}\psfrag{4}{\small \bf 4}
\psfrag{5}{\small \bf 5}\psfrag{6}{\small \bf 6}\psfrag{7}{\small \bf 7}\psfrag{8}{\small \bf 8}\psfrag{a}{\small\texttt 1}
\psfrag{b}{\small\texttt 2}\psfrag{c}{\small\texttt 3}\psfrag{d}{\small\texttt 4}
\psfrag{e}{\small\texttt 5}\psfrag{f}{\small\texttt 6}\psfrag{g}{\small\texttt 7}\psfrag{h}{\small\texttt 8}
\includegraphics[width=5cm]{l4graceful.eps}\\
\caption{A graceful labeling of $L_4$. }
\label{fig:graceful}
\end{figure}
\section{Closing Remarks}
We have shown that there are only two possible values of $b$ for which a connected non-bipartite graph has a $b$-edge consecutive magic labeling.
As Enomoto {\it et al.}~\cite{ELNR} showed that an odd cycle of length at least $3$ is super edge-magic and an even cycle does not have a super edge-magic labeling. Thus, by Theorem~\ref{nonbipartite}, an odd cycle of length $l$ ($l \ge 3$) has a $b$-edge consecutive labeling if and only if $b \in \{0,l\}$. Furthermore, by Theorems~\ref{dual} and \ref{eachcase}, an even cycle does have a consecutive edge magic labeling.
On the other hand, in the same paper, they proved that a complete bipartite graph $K_{m,n}$ is super edge-magic if and only if $m=1$ or $n=1$ or equivalently $K_{m,n}$ is not super edge-magic if and only if $m \ge 2$ and $n\ge 2$. This result together with Theorem~\ref{dual1} imply that $K_{m,n}$ has a total edge consecutive magic labeling if and only if $m=1$ or $n=1$.
|
\section{Introduction}
\label{sec:intro}
The Erd\H os-Szekeres theorem \cite{erdos} states that for every positive integer {\it m}, there exists a smallest integer $ES(m)$, such that any set of at least $ES(m)$ points in the plane, no three on a line, contains $m$ points which lie on the vertices of a convex polygon. The best known bounds on $ES(m)$ are:
\begin{equation}
2^{m-2} +1\leq ES(m) \leq {{2m-5}\choose {m-3}}+1.
\label{eq:es_bound}
\end{equation}
The lower bound is due to Erd\H os \cite{erdosz} and the upper bound is due to T\'oth and Valtr \cite{tothvaltr}.
It is known that $ES(4)=5$ and $ES(5)=9$ \cite{kalb}. Following a long computer search, Szekeres and Peters
\cite{szekeres} proved that $ES(6)=17$. The exact value of $ES(m)$ is unknown for all $m> 6$, but Erd\H os conjectured that the lower bound in (\ref{eq:es_bound}) is, in fact, sharp.
In 1978 Erd\H os \cite{erdosempty} asked whether for every positive integer $k$,
there exists a smallest integer $H(k)$, such that any set of at least
$H(k)$ points in the plane, no three on a line, contains $k$ points which lie on the vertices of a
convex polygon whose interior contains no points of the set. Such a subset is called an {\it empty
convex $k$-gon} or a {\it convex k-hole}. Esther Klein showed that {$ H(4)=5$} and Harborth \cite{harborth}
proved that {$ H(5)=10$}. Horton \cite{horton} showed
that it is possible to construct an arbitrarily large set of points
without a 7-hole, thereby proving that {$ H(k) $} does not exist
for {$ k \geq 7$}. After a long wait, the existence of $ H(6)$ was proved
by Gerken \cite{gerken} and independently by Nicol\'as
\cite{nicolas}. Later, Valtr \cite{valtrhexagon} gave a simpler version of Gerken's proof. Recently, Koshelev \cite{koshelev} has proved that $H(6) \leq \max\{ES(8), 400\}\leq 463$ (the full version of the paper appeared in Russian \cite{koshelev_russian}). The counting version of this problem, that is, the minimum number $N_n(k)$ of
$k$-holes in a set of $n$ points in general position in the plane, is also widely studied \cite{barany_valtr_empty_convex_polygons,dumitrescu} (for the best known bounds refer to the recent papers by Aichholzer et al. \cite{aichholzerI,aichholzerII} and Valtr \cite{valtr_count} and the references therein).
The notion of {\it almost empty} convex polygons, that is, convex polygons with few interior points, was introduced by Nyklov\'a \cite{nyklova}. For integers $k\geq 3$ and $s\geq 0$, let $H(k, s)$ be the minimum positive integer such that any set $S$ of at least $H(k, s)$ points in the plane, no three on a line, contains a subset $Z$ of cardinality $k$ whose elements are on the vertices of a convex $k$-gon and there are at most $s$ points of $S$ in the interior of the convex hull of $Z$. It is clear that $ES(k) \leq H(k, s) \leq H(k, 0)= H(k)$. Nyklov\'a \cite{nyklova} showed that $H(6, 6) = ES(6)=17$ and $H(6, 5) = 19$. It was also shown that $H(k, s)$ does not exist for every $s \leq 2^{(k+6)/4}-k-3$ \cite{nyklova}. Recently, Koshelev \cite{koshelev_almost_empty_hexagon} proved that $H(6, 1)\leq ES(7)$. Koshelev \cite{koshelev_computer_solution,koshelev_almost_empty_hexagon} also showed that Nyklov\'a's proof is incorrect, and using a computer search based on the Szekeres-McKay-Peters algorithm \cite{szekeres} proved that $H(6, k)=17$, for all $2\leq k\leq 6$ and $H(6, 1)=18$.\footnote{Koshelev's papers were originally published in Russian. Here we refer to their English translations \cite{koshelev_computer_solution,koshelev_almost_empty_hexagon}.}
Colored variants of the Erd\H os-Szekeres problem were considered by Devillers et al. \cite{erdosszekereschromatic}.
In such problems, the set of points is partitioned into $r\geq 2$ color classes, and a convex polygon is said to be monochromatic if all its vertices belong to the same color class. It is easy to see that any 2-colored point set of size 10 has an empty monochromatic triangle. Grima et al. \cite{disjoint_monochromatic_triangles} showed that 9 points are necessary and sufficient for a bi-colored point set to have a monochromatic empty triangle. Aichholzer et al. \cite{aichholzer_triangle_number} proved that any set of $n$ bi-colored points in the plane contains $\mathrm\Omega(n^{5/4})$ empty monochromatic triangles, which was later improved to $\mathrm\Omega(n^{4/3})$ by Pach and T\'oth \cite{pach_toth_monochromatic_empty_triangles}. Devillers et al. \cite{erdosszekereschromatic} also constructed a specific coloring for the Horton sets to prove that there exists arbitrarily large 3-colored point sets with no monochromatic empty triangles. Using this they showed that there exists arbitrarily large 2-colored point sets with no monochromatic 5-hole.
It was conjectured by Devillers et al.\cite{erdosszekereschromatic} that any sufficiently large set of bi-colored points in the plane, no three on a line, contains a monochromatic 4-hole. However, in spite of substantial attempts over the last few years, the problem remains open.
Devillers et al.\cite{erdosszekereschromatic} showed that for $n\geq 64$ any bi-chromatic Horton set contains a monochromatic 4-hole. The best known lower bound is a two-colored set of 36 points that contains no monochromatic 4-hole by Huemer and Seara \cite{huemer_geombinatorics}.
As no real progress for the monochromatic empty 4-hole problem was being made, a weaker version arose by relaxing the convexity condition \cite{pach_non_convex}. Aichholzer et al. \cite{aichholzer_siam_dm} showed that this relaxed conjecture is indeed true. They proved that if the cardinality of the bi-chromatic point set $S$ is at least 5044, there always exists an empty monochromatic 4-gon spanned by $S$, that need not be convex. Using observations on vertex degree parity constraints for triangulations of $S$, this bound was lowered to 2760 points by Aichholzer et al. \cite{aiccholzer_parity_wads}.
To the best of our knowledge, the analogous version of the almost empty convex polygon problem for colored point sets has never been considered. In this paper, we initiate the study of the combinatorial quantities associated with the existence of monochromatic empty polygons with few interior points in colored point sets.
For positive integers $c, s \geq 1$ and $r\geq 3$ define $M_r(c, s)$ to be the least integer such that any set of at least $M_r(c, s)$ points in the plane, no three on a line, and colored with $c$ colors contains a monochromatic convex polygon of $r$ vertices and with at most $s$ interior points. In this paper, we primarily concentrate on almost empty monochromatic triangles, that is, the case $r=3$. The case $s=0$, which corresponds to empty monochromatic triangles, has been studied extensively over the last few years. We already mentioned that $M_3(1, 0)=3$, $M_3(2, 0)=9$ \cite{disjoint_monochromatic_triangles} and $M_3(c, 0)=\infty$, for $c\geq 3$ \cite{erdosszekereschromatic}. In this paper we extend these results when $c \geq 3$ and $s \geq 1$. We begin by showing that
\begin{thm}
$M_3(3, 1)=13$.
\label{th:31}
\end{thm}
Using this result and induction on $c$, and a Horton set construction, we obtain the main result of the paper:
\begin{thm}
For $c\geq 2$, the least integer $\lambda_3(c)$ such that $M_3(c, \lambda_3(c))< \infty$ satisfies:
$$\left\lfloor\frac{c-1}{2}\right\rfloor\leq\lambda_3(c)\leq c-2.$$
\label{th:lambda}
\end{thm}
To prove the lower bound in Theorem \ref{th:lambda} we use colored Horton sets. The bounds obtained in this way are equal for $c\in \{2, 3\}$. For $c=4$, the first value for which the bounds in Theorem \ref{th:lambda} do not match, we show that
\begin{thm}
Any Horton set with at least $26$ points colored arbitrarily with $4$ colors contains a monochromatic triangle with at most $1$ interior point.
\label{th:41_restriction}
\end{thm}
Due to the above theorem and several other observations we conjecture that the lower bound in Theorem \ref{th:lambda} is also tight for $c=4$, that is, every sufficiently large 4-colored point set contains a monochromatic triangle with at most 1 interior point.
\section{Preliminaries}
We begin by introducing the definitions and notation required for the remainder of
the paper. Let $S$ be a finite set of points in the plane in general position,
that is, no three on a line. Denote the convex hull of $S$ by $CH(S)$. The
boundary vertices of $CH(S)$, and the interior of $CH(S)$
are denoted by ${\mathcal V}(CH(S))$ and ${\mathcal I} (CH(S))$, respectively. Finally, for any finite set $Z\subseteq S$, $|Z|$ denotes the cardinality of $Z$.
A set of points $S$ in general position is said to be $c$-{\it colored} if $S$ can be partitioned into $c$ non-empty sets $S = S_1\bigsqcup S_2 \bigsqcup \cdots \bigsqcup S_c$, where each set $S_i$ will be referred to as the set of points of color $i$, and $\bigsqcup$ denotes the union of disjoint sets. A subset $T\subset S$ is called {\it monochromatic} if all its points have the same color.
A triangulation of a set of points $S$ in the plane in general position is a triangulation of the convex hull of $S$, with all points from $S$ being among the vertices of the triangulation. We now recall a few standard facts from triangulations of point-sets that we shall use repeatedly in the subsequent sections.
\begin{fact}
For any set $S$ of points in the plane in general position, the number of triangles in any triangulation of $S$ is $2|S|-|{\mathcal V}(CH(S))|-2$.
\label{fact1}
\end{fact}
Since for point sets in general position $|{\mathcal V}(CH(S))|\leq |S|$, the following is an immediate consequence of the above fact:
\begin{fact}
For any set $S$ of points in the plane in general position, the number of triangles in any triangulation of $S$ is at least $|S|-2$.
\label{fact2}
\end{fact}
Grima et al. \cite{disjoint_monochromatic_triangles} proved that every 2-colored 9-point set contains a monochromatic empty triangle. Moreover, there exists a set of 8 points with 4 points colored red and 4 points colored blue and no empty monochromatic triangle. In the following lemma we show that this is the only possible distribution of 2 colors among 8 points that can lead to a point set with no monochromatic empty triangles.
\begin{lem}\label{lm2}
Let $S=R\bigsqcup B$ be a 2-colored point set with color class $R$ and $B$, with $|S|=8$ and $|R|\neq |B|$. Then $S$ always contains an empty monochromatic triangle.
\end{lem}
\begin{proof} Without loss of generality we assume that $|R|>|B|$, so $|R|\geq 5$. If $|R|\ge 6$, then by Fact \ref{fact2} we get at least four triangles by triangulating $R$. Since $|B|\le 2$ we obtain an empty triangle in $R$. If $|R|=5$ and $|\mathcal V(CH(R))|\le 4$, then we get at least four triangles in $R$ by Fact \ref{fact1}. In this case we have $|B|=3$, so we again obtain a monochromatic empty triangle in $R$. If $|\mathcal V(CH(R))|=5$, then there is an empty triangle in $R$ unless ${\mathcal I}(CH(R))\cap S=B$. In that case, the 3 points in $B$ form an empty triangle.
\end{proof}
\section{Monochromatic Triangles with At Most 1 Interior Point: Few Exact Values}
\label{sec:31}
In this section we obtain the exact values of $M_3(c, 1)$ for $c \leq 3$. These results will be used later on in the proofs of the main theorem, and are also interesting in their own right.
\begin{prop}
$M_3(2,1)=6$.
\label{lm1}
\end{prop}
\begin{proof} We observe that $M_3(2,1)>5$, by arbitrarily placing 2 blue-colored points inside a red-colored triangle.
Consider a set of 6 points colored arbitrarily with 2 colors, red and blue. Let $R$ and $B$ denote the set of red and blue points respectively. Without loss of generality assume that $|R|\ge |B|$. If $|R|\geq 4$, by Fact \ref{fact2} we can partition $R$ in at least two interior disjoint triangles. In this case, $|B|\leq 2$ and we have a monochromatic triangle in $R$ with at most one interior point. If $|R|=|B|=3$, then there is a monochromatic triangle with at most one interior point whenever $|{\mathcal I}(CH(R))\cap S|\ne 2$. Otherwise, $|{\mathcal I}(CH(R))\cap S|=2$ and the monochromatic triangle spanned by $B$ has at most one interior point.
\end{proof}
\begin{figure}[!hbt]
\centering
\centering
\includegraphics[width=0.45\columnwidth]
{pic.pdf}\\
\caption{A set of 3-colored 12 points that does not contain any monochromatic triangle with at most 1 interior point.}
\label{fig1}
\end{figure}
\subsection{Proof of Theorem \ref{th:31}}
Figure \ref{fig1} shows that $M_3(3,1)\ge 13$. Therefore, to prove the theorem it suffices to show that $M_3 (3,1)\le 13$.
Let $S=R\bigsqcup B \bigsqcup G$ be a any set of 13 points colored by 3 colors. The three color classes $R, B, G$, will be referred to as red, blue, and green, respectively. Assume, without loss of generality, that $|R|\ge |B| \ge |G|$. As $|R|+|B|+|G|=13$, we have $|R|\ge 5$. We now have the following cases:
\begin{description}
\item[$|R|\ge 6$] In this case any triangulation of $R$ has at least 4 red triangles. If all of these 4 triangles have more than one interior point then $|S\backslash R| \geq 4 \times 2 =8$, which contradicts the fact $|S\backslash R|=13-6=7$. Hence we get a red triangle with at most one interior point.
\item[$|R|=5$] We consider three cases based on the size of ${\mathcal V}(CH(R))$:
\begin{description}
\item[$|{\mathcal V}(CH(R))|=5$] Fact \ref{fact1} implies that that any triangulation of $R$ has three triangles. Assume that all these three triangles have more than one interior point. Then $Z={\mathcal I}(CH(R))\cap S$ is a 2-colored point set with $|Z|\geq 3\times 2=6$, and we get a monochromatic empty triangle with at most one interior point by Proposition \ref{lm1}.
\item[$|{\mathcal V}(CH(R))|=4$] Fact \ref{fact1} implies that that any triangulation of $R$ has four triangles. As $|S\backslash R|=8$, we have a monochromatic red triangle with at most one interior point unless ${\mathcal I}(CH(R))\cap S=S\backslash R=B\bigsqcup G$. Now, If $|B|\ne |G|$, we then get a monochromatic empty triangle by Lemma \ref{lm2}. Note that this triangle is empty in $B\bigsqcup G$. But, it can contain at most 1 red point in the interior, since $|{\mathcal I}(CH(R))\cap R|=1$, and we are done. Finally, if $|B|=|G|=4$, then we look at $CH(B)$. If $|{\mathcal V}(CH(B))|=3$, we have 3 interior disjoint blue triangles by Fact \ref{fact1}. In this case $|S\backslash ({\mathcal V}(CH(R))\cup B)|= |{\mathcal I}(CH(R))\cap R|+|G|=5 < 3\times 2 =6$, and so we get a blue triangle with at most one interior point. Otherwise, $|{\mathcal V}(CH(B))|=4$. In this case, we are done unless $|{\mathcal I}(CH(B))\cap S| \geq 4$. Since $|{\mathcal I}(CH(R))\cap R|=1$, three of the 4 points in ${\mathcal I}(CH(B))\cap S$ must be green, and we get a green triangle with at most one red point inside.
\item[$|{\mathcal V}(CH(R))|=3$] In this case any triangulation of $R$ has 5 triangles. As $|S\backslash R| =8 < 5 \times 2 =10$, we always have a red triangle with at most one interior point. \vspace{0.05in}
\end{description}
\end{description}
This completes the proof of Theorem \ref{th:31}, where the exact value of $M_3(3,1)$ is obtained. In the following section we show that $M_3(5, 1)=\infty$. It remains open to determine whether $M_3(4, 1)$ is finite. More discussions about $M_3(4, 1)$ are in Section \ref{sec:41}.
\section{Proof of Theorem \ref{th:lambda}}
\label{sec:lambda}
In this section bounds on $\lambda_3(c)$ as stated in Theorem \ref{th:lambda} will be proved. Note that $\lambda_3(c)$ is the least integer such that $M_3(c, \lambda_3(c)) < \infty$. Proofs of the upper and lower bounds are presented separately in two sections. First it is shown that $M_3(c, c-2)\leq c^2+c+1$ for $c\geq 3$, that is, any $c$-colored point set with cardinality at least $c^2+c+1$ contains a monochromatic triangle with at most $c-2$ interior points. This proves that $\lambda_3(c) \leq c-2$. In the next section, constructing colorings for the Horton set, the lower bound on $\lambda_3(c)$ is proved.
\subsection{Proof of Upper Bound}
We begin with the following lemma where an upper bound on $M_3(c, c-1)$ is obtained. This bound will then be used to bound $M_3(c, c-2)$.
\begin{lem}
For $c\geq 2$, $M_3(c,c-1) \leq \max\{c^2 + 1, 6\}$
\label{lm:2}
\end{lem}
\begin{proof} Let the $c$ colors be indexed by $\{1, 2, \ldots, c\}$ and $S$ be a any set of $c^2+1$ points in the plane colored by $c$ colors. The result will be proved by induction on $c$. The base case $c=2$ is true as $M_3(2,1)=6$ by Proposition \ref{lm1}.
Now, assume that $c\geq 3$ and $M_3(c-1, c-2)\leq \max\{(c-1)^2+1, 6\}$. Let $S_1 \subseteq S$ be the set of points with color 1. As $|S|=c^2+1$, by the pigeon-hole principle, it can be assumed that $|S_1|\geq c+1$.
If $|S_1|\geq c+2$, by Fact \ref{fact2} any triangulation of $S_1$ has at least $c$ triangles. There exists a monochromatic triangle of color 1 with at most $c-1$ interior points unless each of the triangles in any triangulation of $S_1$ contains at least $c$ interior points. This implies that $|S|\geq c^2+|S_1|\geq c^2+c+2$, which is impossible.
Therefore, $|S_1|=c+1$. In this case any triangulation of $S_1$ has at least $c-1$ triangles. If any triangulation of $S_1$ has $c$ triangles, then by similar arguments as before there exists a monochromatic triangle with at most $c-1$ interior points. This implies that any triangulation of $S_1$ contains exactly $c-1$ triangles, that is, $|{\mathcal V}(CH(S_1))|=c+1$. Let $S_{-1}=(S\backslash S_1)\cap CH(S_1)$. Note that if $|S_{-1}| \leq c(c-1)-1$, then there exists a monochromatic triangle of color 1, with at most $c-1$ interior points. Hence, it suffices to assume that $|S_{-1}|\geq c(c-1)\geq M_3(c-1, c-2)$. As $|S_{-1}|$ is a set of points colored with $c-1$ colors, by induction hypothesis there exists monochromatic triangle in $|S_{-1}|$ with at most $c-2$ interior points.
\end{proof}
\subsubsection{Completing the Proof of the Upper Bound}
For $c=2$, it is known that $M_3(2, 0)=9$ \cite{disjoint_monochromatic_triangles}, which implies that $\lambda_3(2)=0$. For $c\geq 3$, it will be shown by induction on $c$ that $M_3(c, c-2)< c^2 + c + 1$. The base case $c=3$ is true as $M_3(3,1) = 13$ by Theorem \ref{th:31}.
Let $c\geq 4$ and $S$ be any set of $c^2+c+1$ points in the plane colored by $c$ colors, and assume that the theorem is true for $c-1$, that is, $M_3(c-1,c-3) \leq c^2 - c + 1$. Let $S_1 \subseteq S$ be the set of points with color 1. As $|S|=c^2+c+1$, by the pigeon-hole principle, it can be assumed that $|S_1|\geq c+2$.
If $|S_1|\geq c+3$, by Fact \ref{fact2} any triangulation of $S_1$ has at least $c+1$ triangles. There exists a monochromatic triangle with at most $c-2$ interior points unless each of the triangles in any triangulation of $S_1$ contains at least $c-1$ interior points. This implies that $|S|\geq (c-1)(c+1)+|S_1|\geq c^2+c+2$, which is impossible.
Therefore, $|S_1|=c+2$. If $|{\mathcal V}(CH(S_1))|\leq c$, then any triangulation of $S_1$ has at least $c+2$ triangles. By similar arguments as before, there exists a monochromatic triangle with at most $c-2$ interior points. Therefore, it suffices to assume that $c+1\leq |{\mathcal V}(CH(S_1))|\leq c+2$. These two cases are considered separately as follows:
\begin{description}
\item[$|{\mathcal V}(CH(S_1))|=c+2$] Any triangulation of $S_1$ contains $c$ triangles. Let $S_{-1}=(S\backslash S_1)\cap CH(S_1)$. Note that if $|S_{-1}|<c(c-1)$, then there exists a monochromatic triangle of color 1, with at most $c-2$ interior points. Hence, it suffices to assume that $|S_{-1}|\geq c(c-1)$. As $S_{-1}$ is a set of $c(c-1)$ points colored with $c-1$ colors, and $M_3(c-1, c-2)\leq (c-1)^2+1$ by Lemma \ref{lm:2} there exists a monochromatic triangle in $S_{-1}$ with at most $c-2$ interior points.
\item[$|{\mathcal V}(CH(S_1))|=c+1$] Any triangulation of $S_1$ contains $c+1$ triangles by Fact \ref{fact1}. As before, let $S_{-1}=(S\backslash S_1)\cap CH(S_1)$. Note that if $|S_{-1}|<(c+1)(c-1)$, then there exists a monochromatic triangle of color 1, with at most $c-2$ interior points. Hence, it suffices to assume that $|S_{-1}|= c^2-1$. As $S_{-1}$ has $c^2 -1$ points colored with $c-1$ colors, and by induction hypothesis $M_3(c-1, c-3)\leq c^2-c+1$, there exists a monochromatic triangle $\Delta$ in $S_{-1}$ with at most $c-3$ interior points of $S_{-1}$. As the convex hull of $S_1$ contains only one point of color 1, $\Delta$ can contain at most one point of color 1. This implies that $\Delta$ is monochromatic triangle with at most $c-2$ interior points.
\end{description}
\subsection{ Proof of Lower Bound}
In this section we prove the lower bound on $\lambda_3(c)$ by appropriately coloring a Horton set with $c$ colors. A {\it Horton set} is a set $H$ of $n$ points sorted by $x$-coordinates: $h_1 <_x h_2 <_x h_3 <_x \ldots <_x h_n$, such that the set
$H^+:= \{h_2,h_4, \ldots\}$ of the even points and the set $H^- := \{h_1,h_3, \ldots\}$ of the odd points are Horton sets and any line through two even points (the {\it upper set}) leaves all odd points below and any line through two odd points (the {\it lower set}) leaves all even points above. A Horton set of size $n$ can be recursively obtained by adding a large vertical separation after intertwining in the $x$-direction an upper Horton set $H^+$ of size $\lfloor n/2 \rfloor $ and a lower set $H^-$ of size $\lceil n /2\rceil$ \cite{horton}.
With these definitions the following observation can be proved easily.
\begin{prop}
Let $H=\{h_1, h_2, \ldots, h_n\}$ be a Horton set, sorted in the increasing order of the $x$-coordinates. For any line joining $h_{i}$ and $h_{j}$, define
$ H_{ij} $ as all points of $H$ with $x$-coordinates between that of $h_{i}$ and $h_{j}$. Also we define $H_{ij}^+:=H_{ij} \cap H^+$ and $H_{ij}^-:=H_{ij} \cap H^-$. If $i \in H^+$ and $j \in H^-$ (or vice versa), then the line joining $h_{i} \in H^+$ and $h_{j}$ has all points of $H^+_{ij}$ above it and all points of $H^-_{ij}$ below it.
\label{lm6}
\end{prop}
\begin{proof}
Let $h_k \in H_{ij}$ be a point which is below the line joining $h_{i} $ and $h_{j}$. If $h_k \in H^+_{ij}$, then the line through $h_{i} \in H^+$ and $h_k \in H^+$ leaves $h_{j} \in H^-$ above, which is not the case for the Horton set since any $2$ points of $H^+$, when joined by a line must have all points of $H^-$ below it. Hence, $h_k \in H^-_{ij}$.
Similarly, it can be shown that all points which are above the line joining $h_{i} $ and $h_{j}$ are in $H^+_{ij}$.
\end{proof}
\begin{figure}[!hbt]
\centering
\centering
\includegraphics[width=0.48\columnwidth]
{horton5.pdf}\\
\caption{A Horton set colored with 5 colors. The number next to a point denotes the color of the point. Observe that the coloring splits in a cyclical manner in $H^+$ and $H^-$.}
\label{fig2}
\end{figure}
Using the above lemma we now complete the proof of the lower bound. Let $c = 2q + 1$ be an odd number. The $c$ colors are indexed by $\{1, 2, \ldots, c\}$. Consider a Horton set $H$ of size $n$ and arrange the points $h_1,h_2, \ldots h_{n}$ of $H$ in increasing order of $x$-coordinates. Color $H$ with these $c$ colors as follows:
The points $h_1, h_{2q+2}, h_{4q+3} \ldots $ are colored by color 1. $h_2, h_{2q+3}, h_{4q+4} \ldots$ are colored by color 2, and in general, the points $h_i, h_{2q+i+1}, h_{4q+i+2}, \ldots$ by color $i$, for $i \in \{1, 2, \ldots, c\}$. As $c=2q+1$ is odd, this coloring splits in a similar pattern in $H^+$ and $H^-$, that is, $H^-$ is colored in cyclical manner as $1,3,5,\ldots ,2q+1,2,4, \ldots 2q, \ldots$ and so on, and $H^+$ is colored in cyclical manner as $2,4, \ldots 2q,1,3,5,\ldots,2q+1, \ldots$ and so on. The coloring scheme with $5$ colors is illustrated in Figure \ref{fig2}.
If $c=2q+2$ is an even number, and a Horton set $H$ of size $n$ is given, then color the rightmost point of $H$ with color $2q+2$. The remaining $n-1$ points are colored with $c-1$ colors as before.
Hereafter, we assume that $c$ is odd. The proof for the case $c$ is even can be done similarly. Let $\Delta=\{h_a,h_b,h_c\}$ be a monochromatic triangle in a Horton set $H$, colored with $2q+1$ colors as above. From the inductive definition of a Horton set, w.l.o.g. it suffices to assume $h_b, h_c\in H^-$ and $h_a \in H^+$. The coloring is done in such a way that the points in $H^+$ have all the $2q+1$ colors. Let $H_{bc},H_{bc}^-$ and $H_{bc}^+$ be defined as before.
By our coloring scheme and using the fact that $h_b, h_c\in H^-$, we get $|H_{bc}|=4\gamma q+2\gamma-1$ and $|H_{bc}^-|=2\gamma q+\gamma-1$ for some $\gamma\in \{1, 2, \ldots, \}$. By recursively breaking $H^-$ into smaller Horton sets, after some steps $h_b$ and $h_c$ must belong to the upper set and lower set, respectively, of some subset of $H^-$ (say $H_0$) which is itself another Horton set.
So by Proposition \ref{lm6} if $h_b \in {H_0}^+ $ and $h_c \in {H_0}^-$ the line $h_b h_c$ will have at least $q$ points of ${H_0}^+ \subset H^-$ above it. Moreover, these $q$ points, since they are in $H^-$, must either lie beneath the line $h_a h_b$ or beneath the line $h_a h_c$. Therefore, the monochromatic triangle $h_a h_b h_c$ must have at least $q$ interior points. This proves that any monochromatic triangle in $H$ must have at least $q$ interior points.
Therefore, we have shown that it is possible to construct arbitrarily large sets of points which when colored with $c$-colors, has no monochromatic triangle with less than $q$ interior points. This implies that $M_3(2q+2, q-1)=M_3(2q+1, q-1)= \infty$ and hence $\lambda_3(c) \geq \lfloor \frac{c-1}{2} \rfloor $. This completes the proof of Theorem \ref{th:lambda}.
\section{Proof of Theorem \ref{th:41_restriction}}
\label{sec:41}
In this section we prove that any sufficiently large Horton set colored arbitrarily with 4 colors contains a monochromatic triangle with at most 1 interior point. To this end, let $H=\{h_1, h_2, \ldots, h_{26}\}$ be a $4$ colored Horton set of size $26$. If all the 4 colors are not present in $H^+$ (respectively $H^-$), then by Theorem \ref{th:31} there exists a monochromatic triangle with at most 1 interior point in $H^+$ (respectively $H^-$). Therefore, assume that all the 4-colors are present in both $H^+$ and $H^-$. W.l.o.g. let $h_a<_x h_b$ be two points in $H^-$ of same color, such that the colors of all points of $H^-$ in between $h_a$ and $h_b$ are different from that color. We can choose that color such that $b-a \in \{2,4,6\}$.
\begin{description}
\item[$b-a=2$] In this case the points $h_a$ and $h_b$ together with any point $h_c\in H^+$ of the same color (say color 1) is an empty monochromatic triangle because, from the definition of a Horton set, no point in $H^+$ or $H^-$ can be inside the triangle $h_a h_b h_c$.
\item[$b-a=4$] By Proposition \ref{lm6} there is at most 1 point of $H^-$ above the line joining $h_a$ and $h_b$. Therefore, triangle $h_a h_b h_c$, where $h_c \in H^+$ has color 1, has at most 1 interior point.
\item[$b-a=6$] W.l.o.g assume that $h_a \in (H^-)^-, h_b \in (H^-)^+$. Let the $2$ points in $H^-$ between $h_a$ and $h_b$ be $h_u <_x h_v$. Naturally, $h_u \in (H^-)^+$ and $h_v \in (H^-)^-$. By Proposition \ref{lm6} the only point over the line joining $h_a$ and $h_b$ is $h_u$, and $h_a h_b h_c$, where $h_c \in H^+$ has color 1, has at most 1 interior point.
\end{description}
This completes the proof of Theorem \ref{th:41_restriction} and strengthens evidence for the following conjecture:
\begin{conj}
Every sufficiently large set of points in the plane, in general position, colored arbitrarily with $4$ colors contains a monochromatic triangle with at most $1$ interior point.
\label{conjecture1}
\end{conj}
\section{Conclusions}
In this paper, we study the existence of monochromatic empty triangles with few interior points. We prove that any large enough $c$-colored point set contains a monochromatic triangle with at most $c-2$ interior points. Using a Horton set construction, we also show that for any $c\geq 2$, there exist arbitrarily large $c$-colored point sets in which every monochromatic triangle contains at least $\left\lfloor \frac{c-1}{2} \right\rfloor $ interior points.
This paper just scratches the surface of the almost empty monochromatic polygon problem. Several interesting questions are left open. Improving the bounds on $\lambda_3(c)$ and proving Conjecture \ref{conjecture1} are the main challenges. Generalizing these results to monochromatic convex $r$-gons would also be interesting.\\
\small{\noindent{\it Acknowledgement}: The authors are indebted to the anonymous referees for carefully reading the manuscript and providing many valuable comments which improved the presentation of the paper.}
|
\section{Introduction}\label{sc:intro} The Hall-Littlewood symmetric polynomials are very well studied objects that arise naturally in a variety of group theoretic, representation theoretic, and combinatorial contexts; chapters II-V of Macdonald's book \cite{Macdonald1995} contain a detailed description of their origins as well as a rich structural theory. In the simplest instance, the Hall-Littlewood polynomials have the form
\begin{equation}\label{eq:hl}
P_\lambda(x_1,\dots,x_n)=\mathrm{const}(\lambda)\cdot \sum_{\sigma\in S_n}\sigma\left(\prod_{1\le i<j\le n} \frac{x_i-q x_j}{x_{i}-x_{j}}\prod_{i=1}^n x_i^{\lambda_i}\right),
\end{equation}
where the index $\lambda$ is a finite string of integers $\lambda_1\ge\lambda_2\ge\dots\ge 0$, $S_n$ is the symmetric group on $n$ symbols, permutations $\sigma\in S_n$ act on functions in $n$ variables by permuting the variables, and $q\in \C$ is a parameter.\footnote{This parameter is traditionally denoted by $t$, but in the context of the present paper, $q$ happens to be a much more natural notation.}
In recent years, a rational deformation of the Hall-Littlewood polynomials turned to be extremely useful in probability, more exactly, in large time analysis of certain interacting particle systems in (1+1)-dimensions. This deformation is obtained through replacing $x_i^{\lambda_i}$ in the above formula by $((\alpha+\beta x_i)/(\gamma+\delta x_i))^{\lambda_i}$ with $\alpha,\beta,\gamma,\delta\in \C$.
In a pioneering work, Tracy and Widom \cite{TW_ASEP_Fredholm2008}-\cite{TW_total_current2009} considered the case of $\alpha=\gamma=1$, $\beta\delta=q^{-1}$, and showed that the corresponding functions are eigenfunctions for the generator of the \emph{asymmetric simple exclusion process}, or ASEP for short. In the equivalent context of the infinite volume, finite magnon sector XXZ model, same functions were considered in a much earlier work of Babbitt and Gutkin \cite{BabbittGutkin1990}, \cite{Gutkin2000} (that followed similar but more extensive work of Babbitt and Thomas for the (less general) XXX model \cite{BabbittThomas1977}), but those papers did not contain complete proofs and remained essentially unnoticed. In a very recent work of Borodin-Corwin-Gorin \cite{BCG}, these functions were also utilized for asymptotic analysis of stochastic (spin $\frac 12$) six vertex model in a quadrant.
A few years after the work of Tracy and Widom, the case of $\alpha=\beta=\gamma=1$, $\delta=0$, was considered by Borodin-Corwin-Petrov-Sasamoto \cite{BCPS1} in connection with the so-called $q$-TASEP and a $q$-Boson particle system (here `T' in `TASEP' stands for `totally', and total asymmetry means that particles in the system are allowed to move in only one direction).
The fully general case was introduced by Povolotsky \cite{Povolotsky} and developed by Borodin-Corwin-Petrov-Sasamoto \cite{BCPS2} in connection with the so-called $q$-Hahn TASEP and a corresponding zero range process.
Let us also remark that the quantum integrable systems perspective on the Hall-Littlewood polynomials themselves was developed earlier by Van Diejen \cite{vanDiejen2004HL}, and degenerating Hall-Littlewood polynomials further leads to classical works on the quantum delta Bose gas, see the introduction to \cite{BCPS2} and references therein.
The essential property of the deformed functions that made them useful for probabilistic analysis (in addition to them being eigenfunctions for generators of interesting interacting particle systems) consisted in completeness and (bi)orthogonality of them viewed as functions of the index $\lambda$. This made it possible to explicitly construct and in some cases analyze at large times the transition matrices (or Green's functions) for the corresponding Markov chains. However, from a structural viewpoint, these two properties (orthogonality and being eigenfunctions of a nice difference operator) are merely a tip of the iceberg of a wealth of algebraic and combinatorial facts that are available for Hall-Littlewood polynomials.
The principal goal of the present work is to develop further the structural properties of the rational deformations of the Hall-Littlewood polynomials.
If one takes into account harmless renormalizations of functions and variables, the four deformation parameters $\alpha,\beta,\gamma,\delta$ yield a single independent one, and we shall choose it in a specific way and denote it by $s$. The analog of \eqref{eq:hl} then reads
\begin{equation}\label{eq:F}
F_\mu(u_1,\dots,u_M)=\frac{(1-q)^M}{\prod_{i=1}^M (1-su_i)}\,\sum_{\sigma\in S_M}\sigma\left(\prod_{1\le i<j\le M}\frac{u_i-qu_j}{u_i-u_j}\cdot\prod_{i=1}^M \left(\frac{u_i-s}{1-su_i}\right)^{\mu_i}\right),
\end{equation}
where $\mu=(\mu_1\ge \dots\ge \mu_M)\in\mathbb{Z}_{\ge 0}^M$. Up to simple prefactors, these are the functions that have been previously considered, for different values of $s$, in the above referenced papers.\footnote{The prefactor in the right-hand side of \eqref{eq:F} was chosen so that the whole expression can be viewed as a certain partition function, see below. For other purposes, e.g., for Cauchy type identities like \eqref{eq:cauchy-intr} below, it might be more natural to consider normalized functions $F_\mu/F_{(0,\dots,0)}$.}
We introduce several new objects that are closely related to these functions.
$\bullet$\quad We define a `dual' set of functions $G_\nu$ defined as follows:
For $\nu=(\nu_1\ge\dots\ge \nu_n)\in\mathbb{Z}_{\ge 0}^n$ with last $k\ge 0$ coordinates equal to 0, and any $N\ge 0$, we set
\begin{multline}\label{eq:G}
G_\nu(v_1,\dots,v_N)=\frac{(1-q)^N(s^2;q)_n}{(q;q)_{N-n+k}(s^2;q)_k}\\ \times \sum_{\sigma\in S_N}\sigma\left(\prod_{1\le i<j\le N} \frac{v_i-qv_j}{v_i-v_j}\cdot \prod_{i=1}^{n-k} \frac{v_i}{(1-sv_i)(v_i-s)}\left(\frac{v_i-s}{1-sv_i}\right)^{\nu_i}\cdot \prod_{j=n-k+1}^N\frac{1-q^ksv_j}{1-sv_j}\right).
\end{multline}
$\bullet$\quad We introduce skew functions $F_{\mu/\lambda}$, $G_{\nu/\lambda}$, whose special cases with $\lambda=\varnothing$ or $\lambda=(0,\dots,0)$ coincide with $F_\mu$ and $G_\nu$, respectively, and show that all these functions can be defined combinatorially, as partition functions for ensembles of paths in the square grid with specific boundary conditions pictured in Figure \ref{fg:paths}. In the Hall-Littlewood limit $s=0$, this turns into the standard combinatorial formula involving summation over semi-standard Young tableaux, cf. \cite[Section III.5]{Macdonald1995}.
In this combinatorial definition of the $F$- and $G$-functions, the weight of an ensemble of paths is given by the product of vertex weights over all vertices of the grid, and our vertex weights are close relatives of the matrix elements of the higher spin $R$-matrix for the XXZ (or six-vertex) integrable lattice model, with one representation of $U_q(\widehat{sl_2})$ being two-dimensional (spin $\frac 12$) and the other one being a generic Verma module (whose highest weight is related to the parameter $s$). In fact, the $F$-function \eqref{eq:F} can be viewed as the (half)infinite volume, finite magnon sector limit of the eigenfunctions of the higher spin XXZ model with periodic boundary conditions, and the summation over permutations in \eqref{eq:F} is related to the (coordinate or algebraic) Bethe ansatz for this model.
The connection to the integrable lattice models was essential for us; it provided motivation as well as a broader viewpoint. However, familiarity with such models is not necessary for most statements and proofs of this work, with the exception of Theorem \ref{th:skew-R}.
$\bullet$\quad We employ a version of the Yang-Baxter equation (that is central to the theory of integrable lattice models) to prove several (skew) Cauchy and Pieri type identities involving our $F$- and $G$-functions. The fact that $F_\mu$'s are eigenfunctions of simple difference operators in $\mu$, which made them useful in probabilistic models, may be viewed as a corollary of one of the Pieri type identities.
To give an example of our identities, the analog of the Cauchy identity has the form (Corollary \ref{cr:cauchy} below):
\begin{equation}\label{eq:cauchy-intr}
\frac{\prod_{i=1}^M(1-su_i)}{(s^2;q)_M}\sum_{\substack{\nu_1\ge\dots\ge \nu_M\ge 0\\ \nu=0^{n_0}1^{n_1}2^{n_2}\cdots}}\prod_{k\ge 0} \frac{(s^2;q)_{n_k}}{(q;q)_{n_k}} \,F_\nu(u_1,\dots,u_M)G_\nu(v_1,\dots,v_N)=\prod_{\substack{1\le i\le M\\1\le j\le N}} \frac{1-qu_iv_j}{1-u_iv_j}\,.
\end{equation}
$\bullet$\quad We also show that the so-called principal specialization into a geometric progression with ratio $q$ of the skew $G$-functions can be viewed as the (half)infinite volume, finite magnon sector transfer-matrix of the higher spin XXZ model with both representations of $U_q(\widehat{sl_2})$ being arbitrary; this is related to the well-known fusion procedure of Kirillov-Reshetikhin \cite{Kirillov-Reshetikhin} for the higher spin XXZ models.
Thus, our focus in this paper is a one-parameter generalization of the Hall-Littlewood theory. There is another one-parameter generalization of Hall-Littlewood polynomials known as Macdonald polynomials, cf. \cite[Chapter VI]{Macdonald1995}. These two generalizations
seems to be completely different at the moment, and it is natural to conjecture that there should be a two-parameter lift of the Hall-Littlewood theory that would unite the two.
One possible direction that could help in finding such a connection is the theory of Hecke algebras. In a recent work, Takeyama \cite{Takeyama1}, \cite{Takeyama2} showed that the $F$-functions \eqref{eq:F} are closely related to certain rational deformations of the affine Hecke algebras. Turning `affine' to `double affine' may lead to a common generalization of the functions in this paper and Macdonald polynomials, but for now this remains out of our reach.
Our $F$- and $G$-functions have a few degenerations as the parameters $q$ and $s$ tend to certain special values, and some of those appear to be new, cf. Section \ref{sc:degeneration}. The case of inhomogeneous Schur polynomials discussed in Section \ref{ss:ischur} bears a certain similarity to a recent work of Motegi and Sakai \cite{MotegiSakai1}, \cite{MotegiSakai2} on the so-called Grothendieck polynomials, see also Lascoux and Schützenberger \cite{LS}, Lenart \cite{Lenart} for much earlier works on those polynomials, but we were not able to establish a direct connection yet.
Another recent work that seems to be related to the present one on the level of ideas, but not yet directly, is that of Betea, Wheeler, and Zinn-Justin \cite{BW}, \cite{BWZ}.
The paper is organized as follows.
In Section \ref{sc:weights} we introduce vertex weights and establish their connection with $R$-matrices for the higher spin XXZ model.
In Section \ref{sc:functions} we define the $F$- and $G$-functions as partition functions of certain collections of paths in the square grid, and show that these functions are symmetric in their parameters.
Section \ref{sc:cauchy} contains (skew) Cauchy and Pieri type identities.
In Section \ref{sc:symmetrization} we prove symmetrization formulas \eqref{eq:F} and \eqref{eq:G}. Section \ref{sc:principal} deals with principal specializations of $F$- and $G$-functions and their connection to fully general higher spin XXZ $R$-matrices and fusion.
In Section \ref{sc:orthogonality} we discuss orthogonality relations for the $F$-functions \eqref{eq:F} proved earlier in \cite{BCPS2} and their connection to the present work.
Section \ref{sc:degeneration} contains a brief description of degenerations of $F$- and $G$-functions as parameters $q$ and $s$ tend to special values.
\subsection*{Acknowledgments} I am very grateful to Ivan Corwin, Vadim Gorin, and Leonid Petrov for numerous discussions that were extremely helpful. I am also very grateful to Ole Warnaar for a number of very valuable remarks. The research was partially supported by NSF grant DMS-1056390.
\section{Vertex weights}\label{sc:weights}
We start by fixing two parameters that we denote by $q$ and $s$. They should be viewed as complex numbers with the condition of being generic --- vanishing of certain algebraic expressions in $q$ and $s$ may make some of our statements below meaningless. As a rule, we will not focus on such degenerations.
\begin{definition}\label{df:weights} For any four-tuple $(i_1,j_1;i_2,j_2)$ of nonnegative integers, define the corresponding \emph{vertex weight} depending on a (generic) complex parameter $u$ as follows: For any $m\ge 0$,
\begin{align}
w_u(m,0,m,0)&=\frac{1-sq^m u}{1-su}, \\
w_u(m,1,m,1)&=\frac{u-sq^m}{1-su}, \\
w_u(m+1,0,m,1)&=\frac{(1-s^2q^m)u}{1-su},\\
w_u(m,1,m+1,0)&=\frac{1-q^{m+1}}{1-su},\label{eq:weight4}
\end{align}
and $w_u(i_1,j_1;i_2,j_2)=0$ for any other values of $i_1,j_1,i_2,j_2\ge 0$.
\end{definition}
We shall also represent vertices of type $(i_1,j_1;i_2,j_2)$ pictorially as in Figure \ref{fg:vertex}, where $i_1,j_1,i_2,j_2$ denote the number of arrows on South, West, North, and East edges, respectively.
\begin{figure}
\includegraphics[scale=1]{vertex.pdf}
\caption{Graphical representation of a vertex of type $(i_1,j_1;i_2,j_2)=(3,3;5,1)$.}\label{fg:vertex}
\end{figure}
\begin{remark}\label{rm:arrow_preserve} \textbf{(i)} The set of four-tuples $(i_1,j_1;i_2,j_2)\in\mathbb{Z}_{\ge 0}^4$ whose weights are (generically) nonzero are described by two conditions: $i_1+j_1=i_2+j_2$, and $j_1,j_2\le 1$. The first condition is the `arrow preservation' --- for every vertex with nonzero weight, the number of incoming arrows is equal to the number of outgoing ones. This arrow preservation will be upheld throughout the paper. The second condition says that each horizontal edge carries at most one arrow. This condition will remain relevant until Section \ref{sc:principal}, where it will be lifted, and vertices with arbitrary $(i_1,j_1;i_2,j_2)\in\mathbb{Z}_{\ge 0}^4$ subject to $i_1+j_1=i_2+j_2$ will be allowed to have nonzero weights, cf. Corollary \ref{cr:skew-G}.
\noindent \textbf{(ii)} The normalization (i.~e., the common denominator $1-su$) is chosen so that $w_u(0,0;0,0)=1$, cf. Remark \ref{rm:partition_function} below.
\end{remark}
The above-defined vertex weights are closely related to matrix elements of the higher spin $R$-matrix associated with $U_q(\widehat{sl_2})$. To make the connection exact, we need to fix a normalization of the $R$-matrix, and we do so by utilizing the $R$-matrices of \cite{Manga}. There $R_{I,J}$ denotes the image of the universal $R$-matrix in the tensor product of two highest weight representations with arbitrary weights $I$ and $J$ with a particular choice of bases.
\begin{remark} As was noted in the introduction, familiarity with the theory of integrable lattice models ($R$-matrices etc.) is not really needed for almost all statements and proofs of this paper (with the exception of Theorem \ref{th:skew-R}). Proposition \ref{pr:w-to-R} below details the connection of vertex weights of Definition \ref{df:weights} and $R$-matrices. It is used in our proofs of Proposition \ref{pr:YB} and Theorem \ref{th:skew-R}, but one can also verify the claim of Proposition \ref{pr:YB} in a completely elementary fashion by multiplying $4\times 4$ matrices, as indicated in the beginning of its proof. Thus, a reader could safely omit Proposition \ref{pr:w-to-R} as well as all other mentions of the $R$-matrices without much damage to the content of this work.
\end{remark}
An explicit formula for $R_{I,J}$ can be seen in (1.1)-(1.3) of \cite{Manga}, where it is assumed that $I\le J$ are nonnegative integers; a more general formula is in \cite[(5.8)-(5.9)]{Manga}. We shall always assume the `field parameter' $\phi$ of \cite{Manga} to be equal to 1, and we shall also re-denote the parameter $q$ from \cite{Manga} by $Q$; it is related to our $q$ above via $Q^2=q$.
With these conventions we have the following
\begin{proposition}\label{pr:w-to-R} Let $R_{I,J}$ be as in \cite{Manga} with $q$ replaced by $Q$. Then for any $i_1,j_2,i_2,j_2\ge 0$,
\begin{equation}\label{eq:w-to-R}
w_u(i_1,j_1;i_2,j_2)=const\cdot (-1)^{j_1}Q^{\frac{i_2(i_2-1)-i_1(i_1-1)+2i_2+j_1-j_2}2} u^{\frac{j_2-j_1}{2}}\cdot
{\left[R_{I,1}(\lambda;1)\right]}_{i_1,j_1}^{i_2,j_2},
\end{equation}
where $q=Q^2$, the `spectral' parameter $\lambda$ is chosen so that $\lambda^2=(uQ)^{-1}$, and $$const=\frac 1{(1-su)\lambda\, Q^{\frac{1+I}2}}\,.$$
\end{proposition}
\begin{proof} A direct comparison of Definition \ref{df:weights} above and \cite[(5.2) and (5.10)]{Manga}.
\end{proof}
It is natural to ask why we need a different set of weights here rather than using the $R$-matrix itself. The answer lies in simpler explicit formulas for the symmetric functions involved that are also easier to relate to the Hall-Littlewood polynomials, see Sections \ref{sc:symmetrization} and \ref{sc:degeneration} below.
Our next goal is to describe the Yang-Baxter equation\footnote{The exact meaning of the term `Yang-Baxter equation' may depend on the context in which it is used. Our usage is close to what is called the \emph{star-triangle transformation} in the context of the six vertex model in \cite{Baxter1986}.} in terms of our vertex weights.
To that end, define the \emph{two-vertex weights} by
\begin{equation}\label{eq:two-vertex}
w_{u_1,u_2}^{(m,n)}(k_1,k_2;k_1',k_2')=\sum_{l\ge 0} w_{u_1}(m,k_1;l,k_1')w_{u_2}(l,k_2;n,k_2'),
\end{equation}
where $u_1,u_2\in \C$, $l,m,n\in \mathbb{Z}_{\ge 0}$, $k_1,k_2,k_1',k_2'\in\{0,1\}$. This is the weight of two vertices $(m,k_1;l,k_1')$ and $(l,k_2;n,k_2')$ attached along the $l$-edges with $l\ge 0$ being arbitrary, cf. Figure \ref{fg:two-vertices}. Note that the sum over $l\ge 0$ contains at most one nonzero term, because for both factors to be nonzero we must have
\begin{equation}\label{eq:arrow_preserve}
l=m+k_1-k_1'=n+k_2'-k_2,
\end{equation}
cf. Remark \ref{rm:arrow_preserve}(i).
\begin{figure}
\includegraphics[scale=1.2]{two-vertex.pdf}
\caption{Graphical representation of two vertices as in \eqref{eq:two-vertex}.}\label{fg:two-vertices}
\end{figure}
We also set
$$
\widetilde w_{u_1,u_2}^{(m,n)}(k_1,k_2;k_1',k_2')=w_{u_1,u_2}^{(m,n)}(k_2,k_1;k_2',k_1').
$$
As $k_j,k_j'$ vary over $\{0,1\}$, let us organize these weights into $4\times 4$ matrices
$$
w_{u_1,u_2}^{(m,n)}=\begin{bmatrix}
w_{u_1,u_2}^{(m,n)}(0,0;0,0)&w_{u_1,u_2}^{(m,n)}(0,0;0,1)&w_{u_1,u_2}^{(m,n)}(0,0;1,0)&w_{u_1,u_2}^{(m,n)}(0,0;1,1)\\
w_{u_1,u_2}^{(m,n)}(0,1;0,0)&w_{u_1,u_2}^{(m,n)}(0,1;0,1)&w_{u_1,u_2}^{(m,n)}(0,1;1,0)&w_{u_1,u_2}^{(m,n)}(0,1;1,1)\\
w_{u_1,u_2}^{(m,n)}(1,0;0,0)&w_{u_1,u_2}^{(m,n)}(1,0;0,1)&w_{u_1,u_2}^{(m,n)}(1,0;1,0)&w_{u_1,u_2}^{(m,n)}(1,0;1,1)\\
w_{u_1,u_2}^{(m,n)}(1,1;0,0)&w_{u_1,u_2}^{(m,n)}(1,1;0,1)&w_{u_1,u_2}^{(m,n)}(1,1;1,0)&w_{u_1,u_2}^{(m,n)}(1,1;1,1)
\end{bmatrix},
$$
and similarly for $\widetilde w_{u_1,u_2}^{(m,n)}$.
\begin{proposition}[Yang-Baxter equation]\label{pr:YB} For any $m,n\ge 0$, $u_1,u_2\in \C$,
\begin{equation}\label{eq:YB-w}
\widetilde w_{u_2,u_1}^{(m,n)}=X\, w_{u_1,u_2}^{(m,n)}\, X^{-1},
\end{equation}
where
\begin{equation}\label{eq:matrix_X}
X=\begin{bmatrix}
1&0&0&0\\ \\
0&\dfrac{q(u_1-u_2)}{u_1-qu_2}& \dfrac{(1-q)u_1}{u_1-qu_2}&0\\ \\
0&\dfrac{(1-q)u_2}{u_1-qu_2}&\dfrac{u_1-u_2}{u_1-qu_2}&0\\ \\
0&0&0&1
\end{bmatrix}.
\end{equation}
\end{proposition}
\begin{proof} This relation is easy to check directly, although it may be a bit tedious as one needs to go over the cases $m=n,n\pm 1, n\pm 2$. Let us instead derive it from the Yang-Baxter equation for the $R$-matrices. Its special case that is relevant to us now reads, cf. \cite[(4.8)]{Manga},
\begin{equation}\label{eq:YB-R}
R_{1,I}^{(21)}(\lambda_1;1)R_{1,I}^{(31)}(\lambda_2;1)R_{1,1}^{(23)}\left(\frac{\lambda_2}{\lambda_1};1\right)=R_{1,1}^{(23)}\left(\frac{\lambda_2}{\lambda_1};1\right)R_{1,I}^{(31)}(\lambda_2;1)R_{1,I}^{(21)}({\lambda_1};1),
\end{equation}
where the equality takes place in the tensor product of the highest weight representation with weight $I$ that we denote as $V_I$ and $\C^2 \otimes \C^2$, and (omitting zeroes)
\begin{equation}\label{eq:R11}
R_{1,1}(\mu)=\begin{bmatrix}
\mu Q-(\mu Q)^{-1}&&&\\
&\mu-\mu^{-1}&Q-Q^{-1}&\\
&Q-Q^{-1}&\mu-\mu^{-1}&\\
&&&\mu Q-(\mu Q)^{-1}\\
\end{bmatrix}.
\end{equation}
The conjugating matrix $X$ in \eqref{eq:YB-w} is the inverse of $R_{1,1}$ modified according to Proposition \ref{pr:w-to-R}. More exactly, set $u_1=(\lambda_1^2 Q)^{-1}$, $u_2=(\lambda_2^2 Q)^{-1}$, and $\mu=\lambda_2/\lambda_1$ in \eqref{eq:R11}. Then, using \eqref{eq:arrow_preserve} and its analog for $\widetilde w$, we obtain
\begin{multline*}
w_{u_1,u_2}^{(m,n)}(k_1,k_2;k_1',k_2')=const^2 (-1)^{k_1+k_2} Q^{\frac{n(n-1)-m(m-1)}{2}+2n}
Q^{\frac{(k_1-k_2)-(k_1'-k_2')}{2}}u_1^{\frac{k_1'-k_1}{2}}u_2^\frac{k_2'-k_2}{2}\\ \times \left[R_{1,I}^{(21)}(\lambda_1;1)R_{1,I}^{(31)}(\lambda_2;1)\right]_{k_1,k_2}^{k_1',k_2'},
\end{multline*}
\begin{multline*}
\widetilde w_{u_2,u_1}^{(m,n)}(k_1,k_2;k_1',k_2')=const^2 (-1)^{k_1+k_2} Q^{\frac{n(n-1)-m(m-1)}{2}+2n}
Q^{\frac{(k_2-k_1)-(k_2'-k_1')}{2}}u_1^{\frac{k_1'-k_1}{2}}u_2^\frac{k_2'-k_2}{2}\\ \times \left[R_{1,I}^{(31)}(\lambda_2;1)R_{1,I}^{(21)}({\lambda_1};1)\right]_{k_1,k_2}^{k_1',k_2'}.
\end{multline*}
The sub- and super-indices $(k_1,k_2)$ and $(k_1',k_2')$ refer to the basis labels in $\C^2\otimes \C^2$, which are ordered as $(0,0),(0,1),(1,0),(1,1)$. We thus obtain
\begin{multline*}
\begin{bmatrix}1&&&\\
&-\left(\frac 1{Qu_2}\right)^\frac12&&\\
&&-\left(\frac Q{u_1}\right)^\frac12&\\
&&&\frac 1{(u_1u_2)^{\frac12}}
\end{bmatrix}^{-1}
w_{u_1,u_2}^{(m,n)}
\begin{bmatrix}
1&&&\\
&-\left( {u_2}Q\right)^\frac12&&\\
&&-\left(\frac{u_1}Q\right)^\frac12&\\
&&&(u_1u_2)^\frac12
\end{bmatrix}^{-1}R_{1,1}\left(\frac{\lambda_2}{\lambda_1};1\right)\\
=R_{1,1}\left(\frac{\lambda_2}{\lambda_1};1\right)\begin{bmatrix}1&&&\\
&-\left(\frac Q{u_2}\right)^\frac12&&\\
&&-\left(\frac 1{Qu_1}\right)^\frac12&\\
&&&\frac 1{(u_1u_2)^{1/2}}
\end{bmatrix}^{-1}
\widetilde w_{u_2,u_1}^{(m,n)} \begin{bmatrix}
1&&&\\
&-\left(\frac{u_2}Q\right)^{\frac12}&&\\
&&-\left( {Q u_1}\right)^\frac12&\\
&&&(u_1u_2)^{\frac 12}
\end{bmatrix}^{-1}.
\end{multline*}
Simplifying yields \eqref{eq:YB-w}.
\end{proof}
For a future use we record the following computation, cf. \eqref{eq:YB-w}.
\begin{lemma}\label{lm:} Let $A=[A_{ij}]_{i,j=1}^4$ be a $4\times 4$ matrix and $X$ be as in \eqref{eq:matrix_X}. Then
\begin{gather}
(XAX^{-1})_{11}=A_{11},\qquad (XAX^{-1})_{41}=A_{41},\label{eq:lemma1}\\
(XAX^{-1})_{42}=\frac{u_2-u_1}{u_2-qu_1}A_{42}+\frac{(1-q)u_2}{u_2-qu_1}A_{43}.\label{eq:lemma2}
\end{gather}
\end{lemma}
The proof is straightforward.
\section{Symmetric rational functions}\label{sc:functions}
The goal of this section is to define certain rational functions in finitely many variables and to show that these functions are symmetric with respect to permutations of the variables.
We shall define two families of functions, with functions in each family parametrized by pairs of nonnegative signatures. A \emph{signature} is a finite string of ordered integers $\lambda=(\lambda_1\ge \lambda_2\ge \dots \ge \lambda_L)$.\footnote{I apologize for the use of ``$\lambda$'' here, given that it was utilized in the previous section to denote the spectral parameter of the $R$-matrix. The notation is traditional in both places, and I find it hard to avoid.} The length $L$ of the string is called the length of the signature. The set of all signatures of a given length $L$ will be denoted as $\mathrm{Sign}_L$,
and the set of all \emph{nonnegative} signatures (i.e. the signatures that consist of nonnegative integers) of length $L$ will be denoted as $\mathrm{Sign}_L^+$. We will also occasionally write nonnegative signatures in the form $\lambda=0^{m_0}1^{m_1}2^{m_2}\cdots$, where $m_j$'s are the multiplicities: $m_j=|\{i:\lambda_i=j\}|$.
We agree that the set $\mathrm{Sign}_0=\mathrm{Sign}_0^+$ consists of the single empty signature $\varnothing=0^01^02^0\cdots$. Whenever we speak of the set of all (nonnegative) signatures below, we mean $\bigsqcup_{L\ge 0} \mathrm{Sign}_L$ or $\bigsqcup_{L\ge 0} \mathrm{Sign}^+_L$, including the possibility that the length may be zero.
Let us now consider (a part of) the standard square grid, and let us assign to each vertex of the grid a complex variable in such a way that all the vertices in the same row are assigned the same variable.
For a finite up-right path in the square grid, we define its weight as the product of weights of its interior vertices\footnote{We call a grid vertex \emph{interior} to a path if it lies on the path and does not coincide with its beginning and ending vertices.}, where the weight of each vertex is determined via Definition \ref{df:weights} with $i_1,j_1,i_2,j_2$ being $0$ or $1$ depending on whether the corresponding (South, West, North, or East) edge adjacent to the vertex is a part of the path or not (0 if it is not, and 1 if it is). To apply the formulas of Definition \ref{df:weights} we assume that $q$ and $s$ are universal parameters fixed once and for all, and for $u$ we utilize the variables that we have just assigned to the vertices of the grid (recall that this variable does not change if we move along any row).
Similarly, for a collection of finitely many up-right paths in the square grid, we define the corresponding weight as the product of weights of interior vertices of all the paths in the collection, where the weight of each vertex is determined via Definition \ref{df:weights} with $i_1,j_1,i_2,j_2$ equal to the number of paths of the collection that contain the corresponding (South, West, North, or East) edge adjacent to the vertex, with the same convention about $q,s,u$ as in the previous paragraph. Note that the weight of a collection of paths is, generally speaking, not equal to the product of weights of its members (but it will be if the paths do not have common edges).
We view two collections of up-right paths as identical if their sets of interior vertices coincide and the numbers $(i_1,j_1,i_2,j_2)$ for each of such vertices in the two collections are equal.
If a collection of paths has no interior vertices we assign to it the weight 1.
\begin{definition}\label{df:F}
Fix $L\ge M\ge 0$, $\lambda\in\mathrm{Sign}^+_L$, $\mu\in\mathrm{Sign}^+_M$, and indeterminates $u_1,\dots,u_{L-M}$.
Assign to each vertex $(i,j)\in \mathbb{Z}\times \{1,2,\dots,L-M\}$ the variable $u_j$.
Define a rational function $F_{\lambda/\mu}(u_1,\dots, u_{L-M})$ as the sum of weights of all possible collections of $L$ up-right paths that (cf. Figure \ref{fg:paths}, left panel)
\begin{enumerate}
\item start with the (vertical) edges $\{(\mu_m,0)\to (\mu_m,1),\ 1\le m\le M\}$, and with the (horizontal) edges $\{(-1,j)\to (0,j),\ 1\le j\le L-M\}$;
\item end with the (vertical) edges $\{(\lambda_l,L-M)\to (\lambda_l,L-M+1),\ 1\le l\le L\}$.
\end{enumerate}
We shall also use the abbreviated notation $F_{\lambda/\varnothing}=F_\lambda$.
\end{definition}
\begin{definition}\label{df:G}
Fix $L,k\ge 0$, $\lambda\in\mathrm{Sign}_L$, $\nu\in\mathrm{Sign}_L$, and indeterminates $u_1,\dots,u_k$.
Assign to each vertex $(i,j)\in \mathbb{Z}\times \{1,2,\dots,k\}$ the variable $u_j$.
Define the rational function $G_{\lambda/\nu}(u_1,\dots, u_k)$ as the sum of weights of all possible collections of $L$ up-right paths that (cf. Figure \ref{fg:paths}, right panel)
\begin{enumerate}
\item start with the (vertical) edges $\{(\nu_n,0)\to (\nu_n,1),\ 1\le n\le L\}$;
\item end with the (vertical) edges $\{(\lambda_l,k)\to (\lambda_l,k+1),\ 1\le l\le L\}$.
\end{enumerate}
We shall also use the abbreviated notation $G_{\lambda/(0,\dots, 0)}=G_\lambda$.
\end{definition}
\begin{figure}
\includegraphics[scale=1.3]{paths_F.pdf}\qquad \includegraphics[scale=1.3]{paths_G.pdf}
\caption{Paths for $F_{\lambda/\mu}$ (left) and $G_{\lambda/\mu}$ (right).}
\label{fg:paths}
\end{figure}
Note that in the second definition the signatures are not required to be nonnegative.
\begin{remark}\label{rm:partition_function}
Because of our normalization $w(0,0;0,0)=1$, we could have equivalently defined the weight of the collection of paths in Definition \ref{df:F} as the product of weights of \emph{all} vertices of the half-strip $\mathbb{Z}_{\ge 0}\times \{1,\dots,L-M\}$. Furthermore, instead of restricting our attention to up-right paths, we could have taken the sum over all possible assignments of nonnegative integers to all the edges adjacent to the vertices of the half-strip that agree with our boundary conditions on the bottom, left, and top boundaries, and such that only finitely many edges carry nonzero numbers (which can be thought of as a boundary condition at the infinite right edge of the strip). Namely to a vertical edge with coordinate $x\in\mathbb{Z}_{\ge 0}$ at the bottom boundary we assign $\mathbf{1}_{x\in \{\mu_m\}}$, to a vertical edge with coordinate $x\in\mathbb{Z}_{\ge 0}$ at the top boundary we assign $\mathbf{1}_{x\in\{\lambda_l\}}$, and to all horizontal edges at the left boundary we assign 1's.
It is not hard to see that only the assignments that correspond to the collections of up-right paths would give nonzero contributions.
A similar statement applies for Definition \ref{df:G} as well, with the half-strip replaced by the full strip $\mathbb{Z}\times \{1,\dots,k\}$, and the boundary conditions enforced on top and bottom boundaries (left and right infinities are both taken care of by the finiteness condition).
\end{remark}
The definitions immediately imply, by splitting the (half)-strip into two narrower (half)-strips, the following \emph{branching rules}.
\begin{proposition}\label{pr:branching} \textbf{(i)} For any $L\ge K\ge M\ge 0$, $\lambda\in\mathrm{Sign}^+_L$, $\mu\in \mathrm{Sign}^+_M$,
\begin{equation}\label{eq:branching-F}
F_{\lambda/\mu}(u_1,\dots,u_{L-M})=\sum_{\kappa\in\mathrm{Sign}^+_K} F_{\lambda/\kappa}(u_{K-M+1},\dots,u_{L-M})F_{\kappa/\mu}(u_1,\dots,u_{K-M}).
\end{equation}
\textbf{(ii)} For any $L,k_1,k_2\ge 0$, $\lambda,\nu\in\mathrm{Sign}_L$,
\begin{equation}\label{eq:branching-G}
G_{\lambda/\mu}(u_1,\dots,u_{k_1+k_2})=\sum_{\kappa\in\mathrm{Sign}_L} G_{\lambda/\kappa}(u_{k_1+1},\dots,u_{k_1+k_2})G_{\kappa/\nu}(u_1,\dots,u_{k_1}).
\end{equation}
\end{proposition}
The following result is less obvious. It is essentially equivalent to the statement about commutation of transfer matrices with different spectral parameters for the higher spin XXZ model in infinite volume and finite-magnon sector\footnote{The words `finite-magnon sector' refer to the situation when the total number of up-spins in the system remains finite. In our situation this corresponds to finitely many vertical arrows in any row of vertical edges.}.
\begin{theorem}\label{th:symmetry} The functions $F_{\lambda/\mu}(u_1,\dots,u_{L-M})$ and $G_{\lambda/\nu}(u_1,\dots,u_k)$ of Definitions \ref{df:F}, \ref{df:G} are symmetric with respect to permutations of their $u$-variables.
\end{theorem}
\begin{proof} Due to the branching relations above, it suffices to consider the case of two variables. (In other words, it suffices to show that swapping the variables corresponding to two neighboring rows of the grid does not affect the partition function.)
Also, due to translation invariance of the path collections of Definition \ref{df:G}, for $G_{\lambda/\nu}$ we may assume that $\lambda$ and $\nu$ are nonnegative without loss of generality.
Let us recall the $4\times 4$ matrices $w_{u_1,u_2}^{(m,n)}$ of the two-vertex weights \eqref{eq:two-vertex}. If we consider the product of a few such matrices
$$
w_{u_1,u_2}^{(m_0,\dots,m_S;n_0,\dots,n_S)}=w_{u_1,u_2}^{(m_0,n_0)}w_{u_1,u_2}^{(m_1,n_1)}\cdots w_{u_1,u_2}^{(m_S,n_S)},\qquad S\ge 0,
$$
then its matrix elements $w_{u_1,u_2}^{(m_0,\dots,m_S;n_0,\dots,n_S)}(k_1,k_2;k_1',k_2')$ can be viewed as sums of products of weights of all vertices in the rectangle $\{0,\dots,S\}\times\{1,2\}$; the summation goes over all possible assignments of nonnegative numbers to the grid edges adjacent to the vertices of the rectangle, subject to boundary conditions given by
$(m_0,\dots,m_S)$ at the row of vertical edges on the bottom, by $(n_0,\dots,n_S)$ at the row of vertical edges at the top, by $(k_1,k_2)$ at the two horizontal edges on the left boundary, and by $(k_1',k_2')$ at the two horizontal edges on the right boundary, cf. Remark \ref{rm:partition_function}.
Given $\lambda=0^{n_0}1^{n_1}2^{n_2}\cdots$, $\mu=0^{m_0}1^{m_1}2^{m_2}\cdots$ and taking $S\ge \lambda_1$, we have (by Definition \ref{df:F})
$$
F_{\lambda/\mu}(u_1,u_2)=w_{u_1,u_2}^{(m_0,\dots,m_S;n_0,\dots,n_S)}(1,1;0,0),
$$
or with $\lambda=0^{n_0}1^{n_1}2^{n_2}\cdots$, $\nu=0^{m_0}1^{m_1}2^{m_2}\cdots$ and $S\ge \lambda_1$ we we have (by Definition \ref{df:G})
$$
G_{\lambda/\mu}(u_1,u_2)=w_{u_1,u_2}^{(m_0,\dots,m_S;n_0,\dots,n_S)}(0,0;0,0).
$$
Further, Proposition \ref{pr:YB} implies that
\begin{equation}\label{eq:conjugated-products}
\widetilde w_{u_2,u_1}^{(m_0,n_0)}\widetilde w_{u_2,u_1}^{(m_1,n_1)}\cdots \widetilde w_{u_2,u_1}^{(m_S,n_S)}= X\, w_{u_1,u_2}^{(m_0,n_0)}w_{u_1,u_2}^{(m_1,n_1)}\cdots w_{u_1,u_2}^{(m_S,n_S)}\, X^{-1}
\end{equation}
with $X$ as in \eqref{eq:matrix_X}, and the two relations of \eqref{eq:lemma1} yield
$G_{\lambda/\nu}(u_1,u_2)=G_{\lambda/\nu}(u_2,u_1)$ and $F_{\lambda/\mu}(u_1,u_2)=F_{\lambda/\mu}(u_2,u_1)$, respectively.
\end{proof}
\begin{remark}\label{rm:general-symmetry} We used the left boundary conditions $(k_1,k_2)=(1,1)$ and $(0,0)$ to define $F_*(u_1,u_2)$ and $G_*(u_1,u_2)$. For the symmetry $u_1\leftrightarrow u_2$ it is essential that $k_1$ and $k_2$ are equal. One could define similar rational functions with $k_j$'s being different (either two or more of them for a larger number of variables), but then the symmetry relations would be replaced by more complicated ones; those could be extracted from relations on other matrix elements in the setting of Lemma \ref{lm:}.
\end{remark}
\section{Identities of Cauchy and Pieri type}\label{sc:cauchy}
In this section we prove several identities involving $F$- and $G$-functions defined above. The terminology we use for these identities (as well as for a few results in further sections as well) originate from the theory of symmetric functions, where it is traditionally used for similar results involving classical Schur symmetric functions and their generalizations. Exact references to analogs of our results for the Hall-Littlewood symmetric functions (which form a one-parameter generalization of the Schur functions, and the $s=0$ specialization of our $F$- and $G$-functions) are collected in Section \ref{ss:HL} below.
Our first goal is to derive the simplest skew-Cauchy type identity using Proposition \ref{pr:YB} and \eqref{eq:lemma2}. We need more notation to state it.
\begin{definition}\label{df:conjugated} For any $(i_1,j_1;i_2,j_2)\in \mathbb{Z}_{\ge 0}^4$ we define the \emph{conjugated} {vertex weight} (depending on a complex parameter $u$)
by
$$
w_u^c(i_1,j_1;i_2,j_2)=\frac{(q;q)_{i_1}(s^2;q)_{i_2}}{(q;q)_{i_2}(s^2;q)_{i_1}}\,w_u(i_1,j_1;i_2,j_2),
$$
with $w_u$ as in Definition \ref{df:weights} and with the standard $q$-Pochhammer notation
$$
(a;q)_n=\begin{cases}
(1-a)(1-qa)\cdots(1-aq^{n-1}),&n\ge 1,\\
1,&n=0.
\end{cases}
$$
Utilizing such conjugated weights instead of the usual ones in Definitions \ref{df:F} and \ref{df:G} leads to the \emph{conjugated} $F$ and $G$ functions
$$
F^c_{\lambda/\mu}:=\frac{c(\lambda)}{c(\mu)}\, F_{\lambda/\mu},\qquad G^c_{\lambda/\mu}:=\frac{c(\lambda)}{c(\mu)}\,G_{\lambda/\mu},
$$
where for a signature $\nu=0^{n_0}1^{n_1}2^{n_2}\cdots$ we define
\begin{equation}\label{eq:def-c}
c(\nu)=\prod_{k\ge 0} \frac{(s^2;q)_{n_k}}{(q;q)_{n_k}}\,.
\end{equation}
\end{definition}
\begin{theorem}[skew-Cauchy identity with single variables]\label{th:skew-cauchy-single} Let $u,v\in\C$ be such that
\begin{equation}\label{eq:for-convergence-single}
\left|\frac{u-s}{1-su}\cdot\frac{v-s}{1-sv}\right|<1.
\end{equation}
Then for any nonnegative signatures $\lambda$ and $\mu$ we have
\begin{equation}\label{eq:skew-Cauchy-single}
\sum_\nu F_{\nu/\lambda}(u)G^c_{\nu/\mu}(v)=\frac{1-quv}{1-uv}\sum_\kappa G^c_{\lambda/\kappa}(v) F_{\mu/\kappa}(u),
\end{equation}
where both summations are taken over the set of all nonnegative signatures.
\end{theorem}
\noindent\textbf{Comments.} \textbf{(i)} The summation over $\kappa$ always has finitely many nonzero terms, while the summation over $\nu$ may have infinitely many ones. Condition \eqref{eq:for-convergence-single} is needed to insure the convergence of the series.
\noindent\textbf{(ii)} For the statement to be nontrivial one must take the length of $\mu$ to be one more than the length of $\lambda$. Then the only nonzero contributions to the left-hand side of \eqref{eq:skew-Cauchy-single} will come from $\nu$ that are of the same length of $\mu$, and nonzero contributions to the right-hand side will come from $\kappa$ of the same length as $\lambda$.
\noindent \textbf{(iii)} The conjugation in \eqref{eq:skew-Cauchy-single} can be placed on the $F$-factors instead of the $G$-factors; the statement obviously does not change.
\noindent \textbf{(iv)} The formulation and the proof of Theorem \ref{th:skew-cauchy-single} assumes that the `set of all nonnegative signatures' includes the empty signature $\varnothing\in \mathrm{Sign}_0^+$.
\begin{proof} The argument is similar to the proof of Theorem \ref{th:symmetry}. Namely, we begin with \eqref{eq:conjugated-products} with $\lambda=0^{n_0}1^{n_1}2^{n_2}\cdots$, $\mu=0^{m_0}1^{m_1}2^{m_2}\cdots$, and we also take $u_1=u$, $u_2=v^{-1}$. Further, we look at the matrix element $(4,2)$ of both sides (that corresponds to $(k_1,k_2;k_1',k_2')=(1,1;0,1)$). The type of paths that contribute to the left-hand side can be seen on Figure \ref{fg:skew-cauchy-proof}, left panel. Here for the bottom row of vertices we use parameter $v^{-1}$, and for the top row of vertices we use parameter $u$.
\begin{figure}
\includegraphics[scale=1.2]{skew-cauchy-2.pdf}\qquad
\includegraphics[scale=1.2]{skew-cauchy-1.pdf}
\caption{Two types of paths in the proof of Theorem \ref{th:skew-cauchy-single}.}
\label{fg:skew-cauchy-proof}
\end{figure}
According to \eqref{eq:lemma2}, on the right-hand side we obtain a linear combination of $(4,2)$ and $(4,3)$ matrix elements of $w_{u,v^{-1}}^{(m_0,n_0)}w_{u,v^{-1}}^{(m_1,n_1)}\cdots w_{u,v^{-1}}^{(m_S,n_S)}$. The $(4,3)$-matrix element is again the sum of weights of paths of the same type as before, but with $u$ and $v^{-1}$ interchanged. On the other hand, the $(4,2)$-matrix element is the sum of weights corresponding to paths of the type pictured in Figure \ref{fg:skew-cauchy-proof}, right panel. The bottom row of vertices uses $u$, and the top row of vertices uses $v^{-1}$.
As $S$ --- the horizontal size of our rectangles --- tends to infinity, all three terms collect a growing number of factors, which are the weights of vertices $(i_1,j_1;i_2,j_2)=(0,1;0,1)$ lying on the long horizontal parts of the paths that exit through the right boundary.
Let us divide all three terms by
$$
(w_{v^{-1}}(0,1;0,1))^S=\left(\frac{1-sv}{v-s}\right)^S.
$$
This will remove most factors from left-hand side and from the (4,2)-matrix element on the right hand-side, with both tending to a finite limit as $S\to\infty$ (we will identify these limits shortly). On the other hand, the (4,3)-element on the right-hand side will equal to a finite expression times
$$
\left(\frac{w_u(0,1;0,1)}{w_{v^{-1}}(0,1;0,1)}\right)^S=\left(\frac{(u-s)(v-s)}{(1-us)(1-vs)}\right)^S,
$$
which will tend to zero because of our hypothesis \eqref{eq:for-convergence-single}. Hence, from \eqref{eq:lemma2} in the limit $S\to\infty$ we read
\begin{equation}\label{eq:(4,2)-elts}
\text{(4,2)-element of the LHS} = \frac{1-uv}{1-quv}\cdot \left(\text{(4,2)-element of } w_{u,v^{-1}}^{(m_0,n_0)}w_{u,v^{-1}}^{(m_1,n_1)}\cdots w_{u,v^{-1}}^{(m_S,n_S)}\right),
\end{equation}
with the two (4,2)-elements represented as sum of weights of paths on the left and right panels of Figure \ref{fg:skew-cauchy-proof}, respectively, where for the top row of vertices on the left figure and for the bottom row of vertices on the right figure we use the weights $w_u$ of Definition \ref{df:weights}, while for the bottom row of vertices on the left figure and for the top row of vertices on the right figure we use the weights
$$
\text{weight}(i_1,j_1;i_2,j_2)=\frac{v-s}{1-sv}\cdot w_{v^{-1}}(i_1,j_1;i_2,j_2).
$$
Observe that we re-packaged the renormalization as the extra pre-factor $(v-s)/(1-vs)$ on the right, which turns the weight of each of the infinitely many $(0,1;0,1)$ vertices into 1.
It remains to identify \eqref{eq:(4,2)-elts} with \eqref{eq:skew-Cauchy-single}. This readily follows from the pictorial interpretation and the identity
$$
\frac{v-s}{1-sv}\cdot w_{v^{-1}}(i_1,j_1;i_2,j_2)=w_v^c(i_2,1-j_1;i_1,1-j_2).
$$
Note that pictorially, the change $(j_1,j_2)\mapsto (1-j_1,1-j_2)$ in the above relation correspond to swapping filled and unfilled horizontal edges on the top row of the left panel and on the bottom row of the right panel of Figure \ref{fg:skew-cauchy-proof}.
\end{proof}
\begin{remark} Similarly to Remark \ref{rm:general-symmetry}, we could have used other matrix elements in the above argument. This leads to different identities.
More exactly, in the setting of Lemma \ref{lm:} we have
$$
(XAX^{-1})_{22}=\frac{q(u_1-u_2)^2}{(u_1-qu_2)(qu_1-u_2)}\,A_{22}+\frac{(1-q)u_1(u_1-u_2)}{(u_1-qu_2)(qu_1-u_2)}\,A_{32}+\text{ lin. comb. of }(A_{23},A_{33}),
$$
which translates into (under the same assumption \eqref{eq:for-convergence-single})
\begin{multline*}
(1-quv)(q-uv)\sum_\kappa G_{\mu/\kappa}(u)G^c_{\lambda/\kappa}(v)\\={q(1-uv)^2}\sum_{\nu} G_{\nu/\lambda}(u)G^c_{\nu/\mu}(v)-{(1-q)uv(1-uv)}\sum_\nu F_{\nu/\lambda}(u)F_{\nu/\mu}^c(v).
\end{multline*}
Similarly,
$$
(XAX^{-1})_{32}=\frac{(u_1-u_2)^2}{(u_1-qu_2)(qu_1-u_2)}\,A_{32}+\frac{(1-q)u_2(u_1-u_2)}{(u_1-qu_2)(qu_1-u_2)}\,A_{22}+\text{ lin. comb. of }(A_{23},A_{33})
$$
translates into
\begin{multline*}
(1-quv)(q-uv)\sum_\kappa F_{\mu/\kappa}(u)F^c_{\lambda/\kappa}(v)\\={(1-uv)^2}\sum_{\nu} F_{\nu/\lambda}(u)F^c_{\nu/\mu}(v)-{(1-q)(1-uv)}\sum_\nu G_{\nu/\lambda}(u)G_{\nu/\mu}^c(v).
\end{multline*}
\end{remark}
We now draw a few corollaries of Theorem \ref{th:skew-cauchy-single}.
\begin{corollary}[skew-Cauchy identity]\label{cr:skew-cauchy}
Let $u_1,\dots,u_M;v_1,\dots,v_N\in\C$ be such that
\begin{equation}\label{eq:for-cauchy-convergence}
\left|\frac{u_i-s}{1-su_i}\cdot\frac{v_j-s}{1-sv_j}\right|<1,\qquad 1\le i\le M, \quad 1\le j\le N.
\end{equation}
Then for any nonnegative signatures $\lambda$ and $\mu$ we have
\begin{equation}\label{eq:skew-cauchy}
\sum_\nu F_{\nu/\lambda}(u_1,\dots,u_M)G^c_{\nu/\mu}(v_1,\dots,v_N)=\prod_{\substack{1\le i\le M\\1\le j\le N}}\frac{1-qu_iv_j}{1-u_iv_j}\sum_\kappa G^c_{\lambda/\kappa}(v_1,\dots,v_N) F_{\mu/\kappa}(u_1,\dots,u_M),
\end{equation}
where both summations are over the set of all nonnegative signatures.
\end{corollary}
\begin{proof} One first uses branching rules of Proposition \ref{pr:branching} (they hold for conjugated $F$ and $G$ functions too), and then applies Theorem \ref{th:skew-cauchy-single} a total of $MN$ times.
\end{proof}
\begin{corollary}[Pieri type rules]\label{cr:pieri} \textbf{(i)} For any $M\ge 0$, $\mu\in\mathrm{Sign}_M^+$, $u_1,\dots,u_M,v\in \C$ such that
$$
\left|\frac{u_i-s}{1-su_i}\cdot\frac{v-s}{1-sv}\right|<1,\qquad 1\le i\le M,
$$
we have
\begin{equation}\label{eq:pieri-F}
\prod_{i=1}^M\frac{1-qu_iv}{1-u_iv}\, F_\mu(u_1,\dots,u_M)=\sum_{\nu\in \mathrm{Sign}_{M}^+}G^c_{\nu/\mu}(v)F_\nu(u_1,\dots,u_M).
\end{equation}
\textbf{(ii)} For any $l,N\ge 0$, $\lambda\in \mathrm{Sign}_l^+$, $u,v_1,\dots,v_N\in \C$ such that
$$
\left|\frac{u-s}{1-su}\cdot\frac{v_j-s}{1-sv_j}\right|<1,\qquad 1\le j\le N,
$$
\begin{equation}\label{eq:pieri-G}
\prod_{j=1}^N\frac{1-quv_j}{1-uv_j}\, G_\lambda^c(v_1,\dots,v_N)=\frac{1-su}{1-q^{l+1}}\sum_{\nu\in \mathrm{Sign}_{l+1}^+}F_{\nu/\lambda}(u)G_\nu^c(v_1,\dots,v_N).
\end{equation}
\end{corollary}
\begin{proof} For (i) we set $\lambda=\varnothing$ in \eqref{eq:skew-cauchy}. For (ii) we set $\mu=0^{l+1}$ in \eqref{eq:skew-cauchy} and note that $F_{0^{l+1}/\kappa}(u)$ may only be nonzero if $\kappa=0^l$, in which case $$F_{0^{l+1}/0^l}(u)=w_u(l,1;l+1,0)=\frac{1-q^{l+1}}{1-su}$$ according to Definition \ref{df:weights}.
\end{proof}
\begin{remark}\label{rm:bethe} Corollary \ref{cr:pieri}(i) can be viewed as the statement that the vector $$
\{c(\mu)F_\mu(u_1,\dots,u_M)\mid \mu\in\mathrm{Sign}_M^+\}
$$
is an eigenvector of the `transfer-matrix'
$$
\{G_{\nu/\mu}(v) \mid \mu,\nu\in \mathrm{Sign}_M^+\}.
$$
This matrix can indeed be seen as the infinite volume, finite-magnon sector limit of the transfer-matrix of the higher spin XXZ model with periodic boundary conditions (modulo some modifications, cf. Proposition \ref{pr:w-to-R}). As eigenvectors of such transfer-matrices are computable by (coordinate or algebraic) Bethe ansatz, one might expect that there should be a symmetrization formula for $F_\mu$. We shall derive such a formula (and another one for $G_\nu$) in the next section.
\end{remark}
\begin{corollary}[Cauchy identity]\label{cr:cauchy} For any $M,N\ge 0$, $u_1,\dots,u_M;v_1,\dots,v_N\in \C$ such that \eqref{eq:for-cauchy-convergence} holds, we have
\begin{equation}\label{eq:cauchy}
\frac{\prod_{i=1}^M(1-su_i)}{(q;q)_M}\sum_{\nu\in \mathrm{Sign}_M^+}F_\nu(u_1,\dots,u_M)G_\nu^c(v_1,\dots,v_N)=\prod_{\substack{1\le i\le M\\1\le j\le N}} \frac{1-qu_iv_j}{1-u_iv_j}\,.
\end{equation}
\end{corollary}
\begin{proof} Substitute $\lambda=\varnothing$, $\mu=0^M$ into \eqref{eq:skew-cauchy}, and using \eqref{eq:weight4} evaluate
$$
F_{0^M}(u_1,\dots,u_M)=(q;q)_M{\prod_{i=1}^M(1-su_i)^{-1}}\,.\qquad\qquad \qedhere
$$
\end{proof}
\section{Symmetrization formulas for $F_\lambda$ and $G_\lambda$}\label{sc:symmetrization} The goal of this section is to prove the following statement, cf. Remark \ref{rm:bethe}. In what follows we denote the symmetric group on $n$ symbols by $S_n$, and for $\sigma\in S_n$ and a function $f$ in $n$ variables we also use the notation $\sigma(f)(x_1,\dots,x_n)=f(x_{\sigma(1)},\dots,x_{\sigma(n)}).$
\begin{theorem}\label{th:symmetrization} \textbf{(i)} For any $M\ge 0$, $\mu\in \mathrm{Sign}_M^+$, and $u_1,\dots,u_M\in \C$, we have
\begin{equation}\label{eq:symmetrization-F}
F_\mu(u_1,\dots,u_M)=\frac{(1-q)^M}{\prod_{i=1}^M (1-su_i)}\,\sum_{\sigma\in S_M}\sigma\left(\prod_{1\le i<j\le M}\frac{u_i-qu_j}{u_i-u_j}\cdot\prod_{i=1}^M \left(\frac{u_i-s}{1-su_i}\right)^{\mu_i}\right).
\end{equation}
\textbf{(ii)} Fix $n\ge 0$, $\nu\in\mathrm{Sign}_n^+$, and assume $k\ge 0$ last coordinates of $\nu$ are zero: $\nu_{n-k+1}=\dots=\nu_n=0$. Then for any $N\ge n-k$ we have
\begin{multline}\label{eq:symmetrization-G}
G_\nu(v_1,\dots,v_N)=\frac{(1-q)^N(s^2;q)_n}{(q;q)_{N-n+k}(s^2;q)_k}\\ \times \sum_{\sigma\in S_N}\sigma\left(\prod_{1\le i<j\le N} \frac{v_i-qv_j}{v_i-v_j}\cdot \prod_{i=1}^{n-k} \frac{v_i}{(1-sv_i)(v_i-s)}\left(\frac{v_i-s}{1-sv_i}\right)^{\nu_i}\cdot \prod_{j=n-k+1}^N\frac{1-q^ksv_j}{1-sv_j}\right).
\end{multline}
\end{theorem}
\begin{remark}\label{rm:symmetrization} \textbf{(i)} One sees directly from Definition \ref{df:F} that increasing all coordinates of $\mu\in\mathrm{Sign}_M^+$ by the same integer $a\ge 0$, $\mu\mapsto \mu+a^M$, is equivalent to adding $a$ vertices of type $(i_1,j_1;i_2,j_2)=(0,1;0,1)$ to each row of the path collections for $F_\mu$. This yields an extra weight factor:
$$
F_{\mu+a^M}(u_1,\dots,u_M)=\prod_{i=1}^M (w_{u_i}(0,1;0,1))^a F_\mu(u_1,\dots,u_M)=\prod_{i=1}^M\left(\frac{u_i-s}{1-su_i}\right)^a F_\mu(u_1,\dots,u_M).
$$
This is obviously in agreement with \eqref{eq:symmetrization-F}.
\smallskip
\noindent\textbf{(ii)} Definition \ref{df:F} implies that the number of variables of $F_\mu=F_{\mu/\varnothing}$ must be equal to the length of $\mu$, and this is what we have in \eqref{eq:symmetrization-F}. On the other hand, the number of variables $N$ of $G_\nu(v_1,\dots,u_N)$ can be arbitrary. But if $N$ is smaller than the number of nonzero coordinates of $\nu$, then one easily sees that collections of paths of Definition \ref{df:G} with nonzero weight do not exist, and thus $G_\nu(v_1,\dots,v_N)\equiv 0$. The case of the number of variables being at least as large as the number of nonzero coordinates of $\nu$ is covered by \eqref{eq:symmetrization-G}.
\smallskip
\noindent\textbf{(iii)} If in \eqref{eq:symmetrization-G} we have $N-n+k>0$ then the summation over $\sigma\in S_N$ can be partially performed explicitly by symmetrizing over indices $(n-k+1,\dots,N)$ first. The resulting formula looks as follows:
\begin{multline}\label{eq:symmetrization-G'}
G_\nu(v_1,\dots,v_N)=\frac{(1-q)^{n-k}(s^2;q)_n}{(s^2;q)_k}
\sum_{\substack{I\subset\{1,\dots,N\}\\|I|=n-k}}
\prod_{i\in I} \frac{v_i}{(1-sv_i)(v_i-s)}
\cdot\prod_{j\notin I} \frac{1-q^ksv_j}{1-sv_j}
\cdot \prod_{\substack{i\in I\\ j\notin I}}
\frac{v_i-qv_j}{v_i-v_j}
\\
\times
\sum_{\substack{\sigma:\{1,\dots,n-k\}\to I\\ \sigma\text{ is a bijection}}}\sigma\left(\prod_{1\le i<j\le n-k} \frac{v_i-qv_j}{v_i-v_j}\cdot \prod_{i=1}^{n-k} \left(\frac{v_i-s}{1-sv_i}\right)^{\nu_i}\right),
\end{multline}
and we used the symmetrization identity (see \cite[(1.4) in Chapter III]{Macdonald1995})
\begin{equation}\label{eq:symm-identity}
\sum_{\sigma\in S_p}\sigma\left(\prod_{1\le i<j\le p} \frac{z_i-qz_j}{z_i-z_j}\right)=\frac{(q;q)_p}{(1-q)^p}
\end{equation}
along the way.
\smallskip
\noindent\textbf{(iv)} Formula \eqref{eq:symmetrization-G'} immediately implies that the functions $G_\nu$ are \emph{stable} in the sense that adding 0's to the string of their variables does not change them (indeed, the factor $v_i$ forces indices of the zero variables not to be included in the set $I$ in the summation). On the other hand, this fact is also easy to see from Definition \ref{df:G}, as having a zero variable forces the absence of occupied horizontal edges in the corresponding row, and $w_{u=0}(m,0;m,0)=1$ for any $m\ge 0$. This actually proves a more general stability relation: For any signatures $\lambda,\nu$ we have
\begin{equation}
\label{eq:stable}
G_{\nu/\lambda}(v_1,\dots,v_N)=G_{\nu/\lambda}(v_1,\dots,v_N,0).
\end{equation}
\smallskip
\noindent\textbf{(v)} If $\mu\in\mathrm{Sign}_M^+$ has no zero coordinates then the collections of paths in Definitions \ref{df:F} and \ref{df:G} for $F_\mu(u_1,\dots,u_M)$ and $G_\mu(u_1,\dots,u_M)$ are almost identical apart from the left-most column. More exactly, one has (using (i) above)
\begin{equation}\label{eq:G-via-F}
\begin{aligned}
G_\mu(v_1,\dots,v_M)&=\prod_{i=1}^M w_{v_i}(i,0;i-1,1)\cdot F_{(\mu-1^M)}(v_1,\dots,v_M)\\ &= (s^2;q)_M
\prod_{i=1}^M \frac{v_i}{v_i-s}\cdot F_{\mu}(v_1,\dots,v_M),
\end{aligned}
\end{equation}
where $(\mu-1^M)=(\mu-1,\dots,\mu_M-1)\in \mathrm{Sign}_M^+$. This relation also immediately follows from \eqref{eq:symmetrization-F} and \eqref{eq:symmetrization-G'}.
\smallskip
\noindent\textbf{(vi)} The proof of Theorem \ref{th:symmetrization} we give below is a verification rather than a derivation argument, and one might wonder where \eqref{eq:symmetrization-F} and \eqref{eq:symmetrization-G} came from. The symmetrization formula \eqref{eq:symmetrization-F} for $F_\mu$ can be derived with standard (coordinate or algebraic) Bethe ansatz techniques, cf. Remark \ref{rm:bethe}. As for the symmetrization formula \eqref{eq:symmetrization-G} for $G_\nu$, its derivation is given in Proposition \ref{pr:spatial-check} below, and it is based on \eqref{eq:symmetrization-F}, the Cauchy identity \eqref{eq:cauchy}, and the spatial orthogonality of Theorem \ref{th:spatial}.
\end{remark}
\begin{proof}[Proof of Theorem \ref{th:symmetrization}] We shall use the branching relations of Proposition \ref{pr:branching} and induction on the number of variables. For a single variable, Definitions \ref{df:F} and \ref{df:G} imply
$$
F_{(\mu_1)}(u)=(w_{u}(0,1;0,1))^{\mu_1}w_u(0,1;1,0)=\frac{1-q}{1-su_1}\left(\frac{u-s}{1-su}\right)^{\mu_1},
$$
\begin{multline*}
G_{(\nu_1,0^{n-1})}(v)\\=\begin{cases} w_v(n,0;n-1,1)(w_{v}(0,1;0,1))^{\nu_1-1}w_v(0,1;1,0)=\dfrac{(1-q)(1-s^2q^{n-1})v}{(v-s)(1-sv)}\left(\dfrac{v-s}{1-sv}\right)^{\nu_1},&\nu_1>0,\\
w_{v}(n,0;n,0)=\dfrac{1-sq^nv}{1-sv},&\nu_1=0,
\end{cases}
\end{multline*}
and all three expressions are in agreement with \eqref{eq:symmetrization-F}, \eqref{eq:symmetrization-G}.
Let us first prove the inductive step for $F_\mu$. The instance of the branching relation \eqref{eq:branching-F} that we need is
\begin{equation}\label{eq:branching-F-single}
F_\mu(u_1,\dots,u_M)=\sum_{\lambda\in\mathrm{Sign}_{M-1}^+} F_{\mu/\lambda}(u_M)F_\lambda(u_1,\dots,u_{M-1}).
\end{equation}
Split $\mu$ into (nonempty) clusters of equal coordinates
\begin{equation}\label{eq:clusters}
\mu_1=\dots=\mu_{c_1},\quad \mu_{c_1+1}=\dots=\mu_{c_1+c_2},\quad \dots,\quad \mu_{c_1+\dots+c_{m-1}+1}=\dots=\mu_{M},
\end{equation}
where $\{c_j\}_{j=1}^m$ are the cluster sizes. One easily sees from Definition \ref{df:F} that nonzero contributions to the right-hand side of \eqref{eq:branching-F-single} come only from $\lambda$'s such that
\begin{multline}\label{eq:interlacing-F}
\lambda_1=\dots=\lambda_{c_1-1}=\mu_{c_1},\quad \mu_{c_1}\le\lambda_{c_1}\le \mu_{c_1+1},\\
\lambda_{c_1+1}=\dots=\lambda_{c_1+c_2-1}=\mu_{c_1+c_2},\quad \mu_{c_1+c_2}\le \lambda_{c_1+c_2}\le \mu_{c_1+c_2+1},\ \dots,\\
\mu_{c_1+\dots+c_{m-1}}\le\lambda_{c_1+\dots+c_{m-1}}\le \mu_{c_1+\dots+c_{m-1}+1},\quad \lambda_{c_1+\dots+c_{m-1}+1}=\dots=\lambda_{M-1}=\mu_{M}.
\end{multline}
It is convenient to switch from variables $\{u_i\}$ to their fractional-linear images
\begin{equation}\label{eq:xi}
\xi_i:=\frac{u_i-s}{1-su_i}\,,\qquad i\ge 1.
\end{equation}
By the induction hypothesis, we know that $F_\lambda(u_1,\dots,u_{M-1})$ is a linear combination of monomials $\prod_{i}\xi_i^{\lambda_{\sigma(i)}}$, $\sigma\in S_{M-1}$, with coefficients in $\C(\xi_1,\dots,\xi_{M_1})$ --- the field of rational functions in $\xi_1,\dots,\xi_{M-1}$. As a first step, we want to prove a similar statement for $F_\mu(u_1,\dots,u_M)$. Before doing that, let us make sure that such a representation is unique.
\begin{lemma}\label{lm:uniqueness} For any $\alpha,\beta\ge 1$, the functions of the form
$$
f_A:\mathbb{Z}_{\ge 0}^\beta\to \C(\xi_1,\dots,\xi_\alpha),\qquad f_A:(p_1,\dots,p_\beta)\mapsto \prod_{i=1}^\alpha \xi_i^{\sum_{j=1}^\beta A_{ij}p_j},
$$
with $A\in \mathrm{Mat}(\alpha\times\beta,\mathbb{Z})$, are linearly independent over $\C(\xi_1,\dots,\xi_\alpha)$. In other words, if for $\phi_1,\dots,\phi_R\in \C(\xi_1,\dots,\xi_\alpha)$ and pairwise distinct $A^{(1)},\dots,A^{(R)}\in\mathrm{Mat}(\alpha\times\beta,\mathbb{Z})$ we have
$$
\phi_1f_{A^{(1)}}(p_1,\dots,p_\beta)+\dots+\phi_R f_{A^{(R)}}(p_1,\dots,p_\beta)=0\quad \textrm{for any }\ p_1,\dots,p_\beta\in \mathbb{Z}_{\ge 0},
$$
then $\phi_1=\dots=\phi_R=0$.
\end{lemma}
The linear transformation $(p_1,\dots,p_\beta)\mapsto (p_1+\dots+p_\beta,p_2+\dots+p_\beta,\dots,p_\beta)$ allows one to replace $\mathbb{Z}_{\ge 0}^\beta$ by $\{p_1\ge p_2\ge \dots\ge p_\beta\}\subset\mathbb{Z}^\beta_{\ge 0}$ in the statement of the lemma. It is this version that we actually need.
Although Lemma \ref{lm:uniqueness} is not far from being a triviality, we supply a proof at the end of this section.
Let us return to showing that $F_\mu(u_1,\dots,u_M)$ is a linear combination of monomials $\prod_i\xi_i^{\mu_{\tau(i)}}$, $\tau\in S_M$, with coefficients in $\C(\xi_1,\dots,\xi_M)$. As $F_\lambda(u_1,\dots,u_{M-1})$ is a linear combination of monomials of the form $\prod_i\xi_i^{\lambda_{\sigma(i)}}$, let us start with one such monomial in \eqref{eq:branching-F-single} and see what the summation over $\lambda$'s as in \eqref{eq:interlacing-F} gives. We shall take $\sigma=\mathrm{id}$, for other $\sigma$'s the argument and the conclusion are similar.
We want to trace what happens to powers of $\xi_i$'s as we do the summation over $\lambda$'s. For now we shall ignore rational coefficients that are independent of the values $\lambda_i$'s and $\mu_i$'s (but they may depend on multiplicities (cluster sizes) in $\lambda$ and $\mu$). For each $\lambda_i$ that is free to move in the corresponding interval, cf. \eqref{eq:interlacing-F}, we split its range into three parts --- the left end, the right end, and strictly between the two ends. We then observe that
\begin{itemize}
\item If $\lambda_i$ is locked then $\xi_i^{\lambda_i}$ can be simply read as $\xi_i^{\mu_i}$. The same is valid if $\lambda_i$ is at the right edge of its range.
\item If $\lambda_i$ is at the left edge of its (nontrivial) range then $\lambda_i=\mu_{i+1}$ so that $\xi^{\lambda_i}=\xi_i^{\mu_{i+1}}$. But in addition to that, the factor $F_{\mu/\lambda}(u_M)$ in \eqref{eq:branching-F-single} contains $\xi_M^{\mu_{i}-\mu_{i+1}}$ from the vertices of type $(i_1,j_1;i_2,j_2)=(0,1;0,1)$ between $\mu_{i+1}$ and $\mu_i$ in the top ($M$th) row of the corresponding path collection.\footnote{It is actually $\xi_M^{\mu_i-\mu_{i+1}-1}$ as there are $\mu_i-\mu_{i+1}-1$ such vertices. However, since we are ignoring rational coefficients, we can remove the `$-1$' from the exponent.} Thus, together we obtain $\xi_i^{\mu_{i+1}}\xi_M^{\mu_i-\mu_{i+1}}$. Note that the total degree is $\mu_i$.
\item If $\lambda_i$ ranges strictly between $\mu_{i+1}$ and $\mu_i$, we need to do the summation
\begin{equation}\label{eq:sum-inside}
\sum_{\lambda_i:\mu_{i+1}<\lambda_i<\mu_i} \xi_i^{\lambda_i} \xi_M^{\mu_i-\lambda_i-1}=\frac{1}{\xi_i-\xi_M}\,\xi_i^{\mu_i}-\frac{\xi_i}{\xi_M(\xi_i-\xi_M)}\,\xi_i^{\mu_{i+1}}\xi_M^{\mu_i-\mu_{i+1}},
\end{equation}
which is a linear combination of the contributions of the two previous cases. Here in the left-hand side $\xi_M^{\mu_i-\lambda_i-1}$ comes from $F_{\mu/\lambda}(u_M)$, where it corresponds to vertices of type $(0,1;0,1)$ at the top row of the path collection between $\lambda_i$ and $\mu_i$, similarly to the previous case.
\item We also need to add the factor $(w_{u_M}(0,1;0,1))^{\mu_M}=\xi_M^{\mu_M}$ from the vertices between $0$ and $\mu_M$ at the top row of the path collection.
\end{itemize}
We conclude that in all cases any $\xi_i$ with $1\le i\le M-1$ is being raised to the power given by a coordinate of $\mu$, and the total power over all $\xi_i$'s, $1\le i\le M$, is always $\mu_1+\dots+\mu_M$. (Note that we have ignored contributions to $F_{\mu/\lambda}(u_M)$ of vertices of all types different from $(0,1;0,1)$; such vertices simply add rational coefficients.) Relying on the symmetry of $F_\mu(u_1,\dots,u_M)$ in the $u$-variables, cf. Theorem \ref{th:symmetry}, and on Lemma \ref{lm:uniqueness}, we can now conclude that $F_\mu(u_1,\dots,u_M)$ is the sum of monomials of the form $\prod_i\xi_i^{\mu_{\tau(i)}}$ over $\tau\in S_M$. Furthermore, because of the symmetry, to obtain a formula for $F_\mu(u_1,\dots,u_M)$ it suffices to find the (rational) coefficient of only one such monomial, for example, of $\prod_i \xi_i^{\mu_i}$ that corresponds to $\tau=\mathrm{id}$. This is what we do next.
Let us focus on a monomial $\prod_i\xi_{\sigma(i)}^{\lambda_i}$, $\sigma\in S_{M-1}$, with its coefficient in $F_{\lambda}(u_1,\dots,u_{M-1})$, substitute it into \eqref{eq:branching-F-single} instead of $F_{\lambda}(u_1,\dots,u_{M-1})$, do the summation over $\lambda$ subject to \eqref{eq:interlacing-F}, and read off the resulting coefficient of $\prod_i{\xi}_i^{\mu_i}$.
By going through the same three cases as above for each $\lambda_i$, we see that there may be a nontrivial contribution to the coefficient of $\prod_i\xi_i^{\mu_i}$ only if $\sigma$ preserves the subsets
$$
\{1,\dots,c_1\},\quad \{c_1+1,\dots,c_1+c_2\}, \quad \dots,\quad \{c_1+\dots+c_{m-1}+1,\dots,M-1\}
$$
of $\{1,\dots,M-1\}$, where $c_j$'s are the cluster sizes of $\mu$, cf. \eqref{eq:clusters}, and, furthermore, no $\lambda_i$ can assume the lowest possible value of its range as long as this range is nontrivial. This means that the relevant ranges of different $\lambda_i$'s do not intersect, and we can perform the summations over each of them independently.
Let us start with summation over the interval between the first two clusters. From \eqref{eq:symmetrization-F}, the part of $F_\lambda(u_1,\dots,u_{M-1})$ to be summed has the form, apart from the easy factor $(1-q)^{M-1}/\prod_{i=1}^{M-1} (1-su_i)$,
\begin{multline}\label{eq:to-be-summed}
\prod_{1\le i<j\le M-1}\frac{u_{\sigma(i)}-qu_{\sigma(j)}}{u_{\sigma(i)}-u_{\sigma(j)}}\prod_{i=1}^{M-1} \xi_{\sigma(i)}^{\lambda_i}=\prod_{c_1+1\le i<j\le M-1}\frac{u_{\sigma(i)}-qu_{\sigma(j)}}{u_{\sigma(i)}-u_{\sigma(j)}}\prod_{\substack{1\le i\le c_1\\ c_1+1\le j\le M-1}}\frac{u_{i}-qu_{j}}{u_{i}-u_{j}}\prod_{i=c_1+1}^{M-1} \xi_{\sigma(i)}^{\lambda_i}\\ \times
\prod_{1\le i<j\le c_1}\frac{u_{\sigma(i)}-qu_{\sigma(j)}}{u_{\sigma(i)}-u_{\sigma(j)}}\prod_{i=1}^{c_1} \xi_{\sigma(i)}^{\lambda_i},
\end{multline}
where we used the fact that $\sigma$ preserves $\{1,\dots,c_1\}$. Let us denote the restriction of $\sigma$ to $\{1,\dots,c_1\}$ by $\sigma_1$; we want to sum over $\sigma_1\in S_{c_1}$ as well. Before doing the summation, the above expression needs to be multiplied by the corresponding part of $F_{\mu/\lambda}(u_M)$ in \eqref{eq:branching-F-single}, which is
$$
\begin{cases}
w_{u_M}(1,0;0,1)(w_{u_M}(0,1;0,1))^{\mu_{c_1}-\lambda_{c_1}-1}w_{u_M}(c_1-1,1;c_1,0),&\mu_{c_1+1}<\lambda_{c_1}<\mu_{c_1},\\
w_{u_M}(c_1,0;c_1;0),&\lambda_{c_1}=\mu_{c_1}.
\end{cases}
$$
In view of \eqref{eq:interlacing-F}, we have
$$
\prod_{i=1}^{c_1} \xi_{\sigma(i)}^{\lambda_i}=\left(\prod_{i=1}^{c_1-1}\xi_{\sigma_1(i)}\right)^{\mu_{c_1}}
\xi_{\sigma_1(c_1)}^{\lambda_{c_1}}.
$$
Using \eqref{eq:symm-identity} to sum over $\sigma_1$'s with fixed $\sigma_1(c_1)=l$ and substituting explicit weights from Definition \ref{df:weights}, we rewrite the sum over $\sigma_1$ and $\lambda_{c_1}$ as
\begin{multline}
\prod_{c_1+1\le i<j\le M-1}\frac{u_{\sigma(i)}-qu_{\sigma(j)}}{u_{\sigma(i)}-u_{\sigma(j)}}\prod_{\substack{1\le i\le c_1\\ c_1+1\le j\le M-1}}\frac{u_{i}-qu_{j}}{u_{i}-u_{j}}\prod_{i=c_1+1}^{M-1} \xi_{\sigma(i)}^{\lambda_i}\cdot \frac{(q;q)_{c_1-1}}{(1-q)^{c_1-1}}\prod_{1\le i\le c_1}\xi_{i}^{\mu_{c_1}}\\ \times
\sum_{l=1}^{c_1}\prod_{\substack{1\le i\le c_1\\ i\ne l}}\frac{u_{i}-qu_{l}}{u_{i}-u_{l}}\left(\frac{(1-s^2)(1-q^{c_1})u_M}{(1-su_M)^2\xi_l^{\mu_{c_1}}}\sum_{\mu_{c_1+1}<\lambda_{c_1}<\mu_{c_1}}\xi_M^{\mu_{c_1}-\lambda_{c_1}-1}\xi_l^{\lambda_{c_1}}+\frac{1-sq^{c_1}u_M}{1-su_M}\right).
\end{multline}
The sum over $\lambda_1$ is like in \eqref{eq:sum-inside}, and we can omit the
term that is similar to the second term of the right-hand side of
\eqref{eq:sum-inside}, because it has $\xi_l^{\mu_{c_1}-\mu_{c_1+1}}$, while we
need $\xi_l^{\mu_{c_1}}$ to contribute to the coefficient of
$\prod_i\xi_i^{\mu_i}$. Further,
$$
\frac{(1-s^2)(1-q^{c_1})u_M}{
(1-su_M)^2(\xi_l-\xi_M)}+\frac{1-sq^{c_1}u_M}{1-su_M}=\frac{u_l-q^{c_1}u_M}{u_l-
u_M}\,,
$$
and the final simplification is achieved with the identity
\begin{equation}\label{eq:old-identity}
\sum_{l=1}^{c_1}\prod_{\substack{1\le i\le c_1\\ i\ne
l}}\frac{u_{i}-qu_{l}}{u_{i}-u_{l}}\cdot \frac{u_l-q^{c_1}u_M}{u_l-
u_M}=\frac{1-q^{c_1}}{1-q}\prod_{i=1}^{c_1} \frac{u_i-qu_M}{u_i-u_M}\,,
\end{equation}
which follows from evaluating $\oint_{\text{around poles at }
z=u_1,\dots,u_{c_1}} f(z)dz$ with
$$
f(z)=\frac
1{(q-1)z}\,\frac{z-q^{c_1} u_M}{z-u_M}\prod_{i=1}^{c_1} \frac{qz-u_i}{z-u_i}
$$
in two different ways --- as the sum of residues at $z=u_i$, $1\le i\le c_1$,
and as the negative sum of residues at $z=0,u_M,\infty$ (the contributions of
$z=0$ and $z=\infty$ cancel out).\footnote{Identities of this type are rather old. For example, an elliptic generalization of \eqref{eq:old-identity} can be extracted from \cite[No. 400]{TM}. I am very grateful to Ole Warnaar for pointing this out.}
Hence, the summation over $\lambda_{c_1}$ and $\sigma_1$ yields the expression
$$
\prod_{c_1+1\le i<j\le
M-1}\frac{u_{\sigma(i)}-qu_{\sigma(j)}}{u_{\sigma(i)}-u_{\sigma(j)}}\prod_{
\substack{1\le i\le c_1\\ c_1+1\le j\le
M-1}}\frac{u_{i}-qu_{j}}{u_{i}-u_{j}}\prod_{i=c_1+1}^{M-1}
\xi_{\sigma(i)}^{\lambda_i}\cdot \frac{(q;q)_{c_1}}{(1-q)^{c_1}}\prod_{1\le i\le
c_1}\xi_{i}^{\mu_{i}}\prod_{i=1}^{c_1} \frac{u_i-qu_M}{u_i-u_M}\,.
$$
We continue summing in this fashion (the next step is to sum over $\lambda_{c_2}$
and $\sigma_2=\sigma|_{c_1+1,\dots,c_1+c_2}$), and in the end, having summed
over the whole of $\lambda$ and $\sigma$ and multiplied by the weight
$$
(w_{u_M}(0,1;0,1))^{\mu_M}w_{u_M}(c_m-1,1;c_m,0)=\frac{1-q^{c_m}}{1-su_M}\,\xi_M^{\mu_M}
$$
of the vertices between $0$ and $\mu_M$ in the $M$th row,
we obtain
$$
\frac{1-q}{1-su_M}\prod_{i=1}^m \frac{(q;q)_{c_i}}{(1-q)^{c_i}}
\prod_{1\le a<b\le m}\prod_{\substack{i\in\,\text{$a$th cluster of $\mu$}\\
j\in\, \text{$b$th cluster of $\mu$}}}\frac{u_{i}-qu_{j}}{u_{i}-u_{j}}
\prod_{1\le i\le
M}\xi_{i}^{\mu_{i}}
$$
Together with the previously omitted factor $(1-q)^{M-1}/\prod_{i=1}^{M-1}
(1-su_i)$, this gives the correct coefficient of $\prod_{i}\xi_{i}^{\mu_{i}}$
in the right-hand side of \eqref{eq:symmetrization-F} (here we need
\eqref{eq:symm-identity} again). This completes the inductive step for $F_\mu$
and the proof of part (i) of Theorem \ref{th:symmetrization}.
Let us proceed to path (ii) --- the inductive step for $G_\nu$. Recall that the base of the induction, the case of a single variable, was discussed in the beginning of the proof.
The proof of the inductive step for $G_\nu$ is largely similar to that for $F_\mu$ above.
Let us comment on the differences.
One starts with a branching relation, cf. \eqref{eq:branching-G}, \eqref{eq:branching-G}, which says that for $\nu\in\mathrm{Sign}_n^+$,
$$
G_\nu(v_1,\dots,v_N)=\sum_{\lambda\in\mathrm{Sign}_n^+}G_{\nu/\lambda}(v_N)G_\lambda(v_1,\dots,v_{N-1}).
$$
It implies that the $\lambda$-coordinates must interlace with the $\nu$-coordinates. The interlacing is similar to \eqref{eq:interlacing-F}, except in addition the smallest coordinate $\lambda_n$ may vary between $\nu_n$ and $0$.
Same inductive arguments as for $F_\mu$ above show that $G_\nu(v_1,\dots,v_N)$ must be a linear combination of monomials $\prod_i \xi_{\tau(i)}^{\nu_i}$, $\tau\in S_N$, with coefficients in $\C (v_1,\dots,v_N)$, where we take, cf. \eqref{eq:xi},
$$
\xi_i=\frac{v_i-s}{1-sv_i},\qquad i\ge 1.
$$
Hence, by the symmetry in $v$-variables of Theorem \ref{th:symmetry}, it suffices to evaluate the coefficient of $\prod_i \xi_i^{\nu_i}$. Then one needs to consider two cases: (a) The number of nonzero coordinates in $\nu$ is strictly smaller than the number of $v$-variables, i.e. $n-k<N$; and (b) The number of nonzero coordinates in $\nu$ is equal to the number of $v$-variables , i.e. $n-k=N$.
For case (a) the computation literally repeats the one we did for the coefficient of $\prod_i\xi_i^{\mu_i}$ in $F_\mu$ except for the very last factor, where one needs to multiply by the weight of vertex 0 in the top row, which for $G_\nu$ is $w_{v_N}(k,0;k;0)=(1-sq^ku_M)/(1-su_M)$, with $k$ being the multiplicity of $0$ in $\nu$, instead of $w_{u_M}(c_m-1,1;c_m,0)$ that would have been there for $F_\mu$. One easily checks that this exactly gives the coefficient of $\prod_i \xi_i^{\nu_i}$ in the right-hand side of \eqref{eq:symmetrization-G} or \eqref{eq:symmetrization-G'}.
A more substantial difference comes up in case (b). If $n-k=N$, then even though interlacing of $\lambda$ and $\nu$ (that comes from non-vanishing of $G_{\nu/\lambda}(v_N)$) allows $\lambda_N=\lambda_{n-k}$ to vary between $\nu_{N}$ --- the last nonzero coordinate of $\nu$ ---
and 0, the fact that for $G_\lambda(v_1,\dots,v_{N-1})$ to be nonzero the number of nonzero coordinates in $\lambda$ cannot be greater than $N-1$ forces $\lambda_{N}$ to be zero. If we denote by $c$ the size of the smallest nonzero cluster of $\nu$ (located at $\nu_{N}$), then the total contribution of vertices between $0$ and $\nu_{N}$ to $G_{\nu/\lambda}(v_N)$
is
\begin{multline*}
w_{v_N}(k+1,0;k,1)(w_{v_N}(0,1;0,1))^{\nu_{N}-1}w_{v_N}(c-1,1;c,0)=\frac{(1-s^2q^k)v_N}{1-sv_N}\,\xi_N^{\nu_N-1}\,\frac{1-q^c}{1-sv_N}\\
=(1-s^2q^k)(1-q^c)\,\frac{v_N}{(1-sv_N)(v_N-s)}\,\xi_N^{\nu_N}.
\end{multline*}
This allows us to verify that the induction step yields the correct coefficient of $\prod_i \xi_i^{\nu_i}$ in the right-hand side \eqref{eq:symmetrization-G} and concludes the proof of Theorem \ref{th:symmetrization}.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lm:uniqueness}] Assume that we found nonzero $\phi_1,\dots,\phi_R\in\C(\xi_1,\dots,\xi_\alpha)$ and pairwise distinct $A^{(1)},\dots,A^{(R)}\in\mathrm{Mat}(\alpha\times\beta,\mathbb{Z})$ such that $\sum_{r}\phi_r f_{A^{(r)}}\equiv 0$. Pick an $\alpha$-tuple of positive reals $(\zeta_1,\dots,\zeta_\alpha)\in \mathbb{R}^\alpha_{>0}$ and a direction $(\omega_1,\dots,\omega_\beta)\in\mathbb{Z}^\beta_{>0}$ so that the numbers
$$
\Omega_r =\sum_{i,j} A_{ij}^{(r)}\zeta_i \omega_j,\qquad 1\le r\le R,
$$
are pairwise distinct (this is always possible as an equality of two such numbers is a nontrivial quadratic equation on $\zeta$'s and $\omega$'s, and solutions to finitely many such equations cannot exhaust
$\mathbb{R}_{>0}^\alpha\times \mathbb{Z}_{>0}^\beta$). Assign weights to variables $\xi_i$ via $\mathrm{wt}(\xi)=\zeta_i$, $1\le i\le \alpha$, and single out top homogeneous components of the polynomial numerators and denominators of the rational functions $\phi_1,\dots,\phi_R$ with respect to this weighting. Pick a point $(c_1,\dots,c_\alpha)\in (\C\setminus\{0\})^\alpha$ so that none of these top homogeneous components vanishes at this point.
Let us now take a fixed large integer $L$ and look at the behavior of $\sum_{r=1}^R \phi_r f_{A^{(r)}}$ as we substitute
$$
(\xi_1,\dots,\xi_\alpha)=(c_1\xi^{\zeta_1},\dots,c_\alpha\xi^{\zeta_\alpha}),\qquad
(p_1,\dots,p_\beta)=(\omega_1L,\dots,\omega_\beta L)
$$
and take $\xi\to\infty$. We observe that each $|\phi_r|$ behaves as a nonzero constant times $|\xi|^{const_r}$. Each $|f_{A^{(r)}}|$ equals a nonzero constant times $|\xi|^{\Omega_rL}$. As long as $L\cdot \min_{1\le r\le R}(\max_{1\le r\le R}\Omega_r-\Omega_i)$ is greater than all $const_r$ coming from $\phi_r$ (which we can guarantee by taking $L$ large enough), the term corresponding to the maximal $\Omega_r$ will dominate all the other ones, and hence $\sum_r \phi_r f_{A^{(r)}}$ cannot vanish. The contradiction completes the proof of Lemma \ref{lm:uniqueness}.
\end{proof}
\section{Principal specializations}\label{sc:principal}
In this section we provide explicit formulas for $F_\mu$, $G_\nu$, and $G_{\nu/\lambda}$ specialized at geometric progressions with ratio $q$. While the first two results are elementary corollaries of the symmetrization formulas in Theorem \ref{th:symmetrization}, the third one is less obvious as it relies on fusion rules for transfer matrices of the higher spin XXZ model.
\begin{proposition}\label{pr:principal} \textbf{(i)} For any $M\ge 0$, $\mu\in\mathrm{Sign}_M^+$, and $u\in \C$, we have
\begin{equation}\label{eq:principal-F}
F_\mu(u,qu,\dots,q^{M-1}u)=\frac{(q;q)_M}{(su;q)_M}\prod_{i=1}^{M}\left(\frac{q^{i-1}u-s}{1-sq^{i-1}u}\right)^{\mu_{i}}.
\end{equation}
\textbf{(ii)} Fix $n\ge 0$, $\nu\in\mathrm{Sign}_n^+$, and assume $k\ge 0$ coordinates of $\nu$ are zero. Then for any $N\ge n-k$ and $v\in\C$ we have
\begin{equation}
\label{eq:principal-G}
G_\nu(v,qv,\dots,q^{N-1}v)=\frac{(q;q)_N(s^2;q)_n(sv;q)_{N+k}}{(q;q)_{N-n+k}(s^2;q)_k(sv;q)_n(sv;q)_N(sv^{-1};q^{-1})_{n-k}}\,\prod_{i=1}^{n-k}\left(\frac{q^{i-1}v-s}{1-sq^{i-1}v}\right)^{\nu_{i}}.
\end{equation}
\end{proposition}
\begin{proof} For (i) we use \eqref{eq:symmetrization-F}. Observe that substituting $\{u_i=uq^{i-1}\}_{i=1}^M$ into
$$
\sigma\left(\prod_{1\le i<j\le M} \frac{u_i-qu_j}{u_i-u_j}\right),\qquad \sigma\in S_M,
$$
gives 0 unless $\sigma=\mathrm{id}$, in which case we get $(q;q)_M/(1-q)^M$. This implies \eqref{eq:principal-F}. In the same way \eqref{eq:symmetrization-G} implies \eqref{eq:principal-G}.
\end{proof}
Let us proceed to the skew functions $G_{\lambda/\mu}$.
\begin{definition} \label{df:corresponds}
We say that a function $H(\lambda,\mu)$ of two signatures $\lambda$ and $\mu$ of the same length corresponds to vertex weights $\{w^{(H)}(i_1,j_1;i_2,j_2)\mid i_1,j_1,i_2,j_2\ge 0\}$ if each value $H(\lambda,\mu)$ is given by the sum of products of these vertex weights over all possible collections of upright paths as in Definition \ref{df:G} and Figure \ref{fg:paths}, right panel, with a single horizontal row of vertices (i.e. $k=1$ in the notation of Definition \ref{df:G}).
\end{definition}
Clearly, with the terminology of Definition \ref{df:corresponds}, the single variable specialization $G_{\lambda/\mu}(v)$ corresponds to the weights $w_v$ of Definition \ref{df:weights}. This is the case when Definition \ref{df:G} and \ref{df:corresponds} simply coincide.
Because of the unfortunate overload of the letter $\lambda$, cf. the footnote in the beginning of Section \ref{sc:functions}, in what follows we speak about functions of $\nu$ and $\mu$, where $\nu$ plays the role of $\lambda$ in Definition \ref{df:corresponds}.
\begin{theorem}\label{th:skew-R} For any $J\ge 1 $, $G_{\nu/\mu}(v,qv,\dots,q^{J-1}v)$ corresponds (as a function of $\nu$ and $\mu$, in the sense of Definition \ref{df:corresponds}) to the vertex weights
\begin{equation}\label{eq:w-via-R}
w_v^{(J)}(i_1,j_1;i_2,j_2)=\frac{(-1)^{j_1}Q^{\frac{(2i_2-I-1)J+i_2^2-i_1^2}{2}}}{\lambda^J(sv;q)_J}\cdot {\left[R_{I,J}(\lambda;1)\right]}_{i_1,j_1}^{i_2,j_2},
\end{equation}
where $R_{I,J}$ is the (fully general) higher spin $R$-matrix of the XXZ model as in \cite[(5.8)-(5.9)]{Manga} with the parameter $q$ of \cite{Manga} re-denoted by $Q$ and related to our $q$ through $Q^2=q$, the spectral parameter $\lambda$ given by $\lambda^2=(Q^{J}v)^{-1}$, parameter $I$ given by $Q^{-I}=s$, and $m(I,J)$ in \cite[(5.9)]{Manga} (used to denoted the minimum of $I$ and $J$) set to $J$.
\end{theorem}
\begin{remark}\label{rm:skew-R} \textbf{(i)} For $J=1$, Theorem \ref{th:skew-R} immediately follows from Proposition \ref{pr:w-to-R}. Comparing the two formulas one may notice discrepancy in certain factors of the form $f(j_1)/f(j_2)$; this is explained by the fact that such factors cancel out in products of weights over vertices of up-right paths, and are thus irrelevant (for the purpose of Theorem \ref{th:skew-R}).
\smallskip
\noindent\textbf{(ii)} As we know from Definition \ref{df:weights}, for $J=1$ the vertex weights vanish as long as $\max(j_1,j_2)>1$. Similarly, $w_v^{(J)}(i_1,j_1;i_2,j_2)$ from \eqref{eq:w-via-R} vanishes if $\max(j_1,j_2)>J$; this corresponds to the highest weight representation of $U_q(\widehat{sl_2})$ with weight $J$ having dimension $J+1$.
Also, if $I\in \{1,2,\dots\}$, i.e. $s^2\in\{q^{-1}, q^{-2},\dots\}$, then $w_v^{(J)}(i_1,j_1;i_2,j_2)=0$ as long as $\max(i_1,i_2)>I$. In particular, $s^2=q^{-1}$ corresponds to the spin-$\frac 12$ situation with no more than one particle (vertical arrow) at each location.
\end{remark}
\begin{proof}[Proof of Theorem \ref{th:skew-R}] The argument combines Proposition \ref{pr:w-to-R} and an infinite volume limit of the fusion relation \cite[(7.13)]{Manga} (the fusion relations in this context were first derived in \cite{Kirillov-Reshetikhin}, but it is convenient for us to use the notation of \cite{Manga}).
More exactly, follow \cite{Manga} in combining the $R$-matrices into transfer-matrices via
\begin{equation}\label{eq:trace}
T_{J,I}(\lambda)=\mathrm{Trace}_{V_J} \left[R^{(01)}_{J,I}(\lambda)\otimes \cdots\otimes R^{(0S)}_{J,I}(\lambda)\right],
\end{equation}
where $V_J$ is the 0th `auxiliary' highest weight representation of $U_q(\widehat{sl_2})$ with weight $J$, and the tensor product is taken in the `quantum space' $V_I^{\otimes M}$, $M\ge 1$. The fusion relation we are interested in reads, see \cite[(7.13)-(7.14)]{Manga},
\begin{equation}\label{eq:fusion}
T_{1,I}(\lambda)T_{J,I}(\lambda Q^{-\frac{J+1}{2}})=
\left( (Q\lambda-(Q\lambda)^{-1})(Q^{-I}\lambda-Q^I\lambda^{-1}) \right)^M
T_{J-1,I}(\lambda Q^{-\frac{J}2 -1})+
T_{J+1,I}(\lambda Q^{-\frac{J}2}),
\end{equation}
where we take $J<I$. We want to take the limit $M\to\infty$ while keeping the number of quantum particles finite (i.e., looking at matrix elements of transfer-matrices with finitely many indices that are different from 0). To do that, we first normalize $R_{J,I}$ so that $[R_{J,I}(\lambda,1)]_{0,0}^{0,0}$ turns into 1; this is achieved by
\begin{equation}\label{eq:normalize-R}
R_{J,I}(\lambda,1)\mapsto \widetilde R_{J,I}(\lambda,1):=\frac{Q^{-\frac{I(J+1)}{2}}}{\lambda^J(\lambda^{-2}Q^{-I-J};Q^2)_J} \,R_{J,I}(\lambda,1).
\end{equation}
Note that this does not destroy the second coefficient `1' in the right-hand side of \eqref{eq:fusion}. Further, let us assume that $Q$ is sufficiently close to 1, and $\lambda$ is sufficiently close to $s=Q^{-I}$; as the final formula \eqref{eq:w-via-R} easily admits analytic continuation from such a domain (it is even an identity of rational functions for fixed $\nu$ and $\mu$), this is not a serious restriction. This leads to the first coefficient in the right-hand side of \eqref{eq:fusion} being a small constant raised to the power $M$, and the fusion relation now makes sense in the $M\to \infty$ limit; it reads
$$
\widetilde T_{1,I}(\lambda)\widetilde T_{J,I}(\lambda Q^{-\frac{J+1}{2}})=
\widetilde T_{J+1,I}(\lambda Q^{-\frac{J}2}),
$$
or, iterating,
\begin{equation}\label{eq:transfer}
\widetilde T_{1,I}(Q^{\frac{-J+1}{2}}\lambda)\widetilde T_{1,I}(Q^{\frac{-J+3}{2}}\lambda)\cdots \widetilde T_{1,I}(Q^{\frac{J+1}{2}}\lambda)=\widetilde T_{J,I}(\lambda),
\end{equation}
where we use the tilde in $\widetilde T$ to signify both the $M\to \infty$ limit and the normalization \eqref{eq:normalize-R}.
From the point of view of path interpretation of matrix elements of the transfer matrix, we started with the periodic boundary conditions in \eqref{eq:trace}, and by making $\lambda$ close to $s$ we made $[R_{1,I}(\lambda,1)]_{j_1=1,i_1=0}^{j_2=1,i_2=0}$ (which corresponds to $w_{v}(0,1;0,1)$) small, which in the limit $M\to\infty$ prevents paths from going around the infinite loop.
It only remains to utilize Proposition \ref{pr:w-to-R} in the left-hand side, keeping in mind the normalization \eqref{eq:normalize-R} and the symmetry relation \cite[(5.10)]{Manga}.
\end{proof}
Theorem \ref{th:skew-R} and results of \cite{Manga} allow us to write down an explicit formula for the principal specialization of the skew $G$-functions. For that we need additional notation.
Following \cite{Manga}, we shall use the following extended notation for $q$-Pochhammer symbols:
$$
(x;q)_n=\begin{cases} (1-x)(1-qx)\cdots(1-q^{n-1}x),&n>0,\\
1,&n=0,\\
\left((1-q^nx)(1-q^{n+1}x)\cdots (1-q^{-1}x)\right)^{-1},&n<0,
\end{cases}
$$
and also regularized terminating basic hypergeometric series
\begin{multline*}
{}_{r+1}\bar{\phi}_r\left(\begin{matrix} q^{-n};a_1,\dots,a_r\\b_1,\dots,b_r\end{matrix}
\Bigl| q,z\right)=\sum_{k=0}^n z^k\,\frac{(q^{-n};q)_k}{(q;q)_k} \prod_{i=1}^r (a_i;q)_k(b_iq^k;q)_{n-k}\\
=\prod_{i=1}^r (b_i;q)_n\cdot {}_{r+1}{\phi}_r\left(\begin{matrix} q^{-n};a_1,\dots,a_r\\b_1,\dots,b_r\end{matrix}
\Bigl| q,z\right) .
\end{multline*}
\begin{corollary}\label{cr:skew-G} For any $J\ge 1 $, $G_{\nu/\mu}(v,qv,\dots,q^{J-1}v)$ corresponds (as a function of $\nu$ and $\mu$, in the sense of Definition \ref{df:corresponds}) to the vertex weights given by
\begin{multline}\label{eq:w-via-phi}
\widetilde w_v^{(J)}(i_1,j_1;i_2,j_2)\\=\frac{(-1)^{i_1+j_1} q^{\frac{i_1^2+i_2^2}4+\frac{i_1(j_1-1)+i_2j_2}{2}}s^{j_1-i_1}v^{i_1}(vs^{-1};q)_{j_1-i_2}}{(q;q)_{i_1} (vs;q)_{i_1+j_1}}\, {}_{4}\bar{\phi}_3\left(\begin{matrix} q^{-i_1};q^{-i_2},q^J sv,qsv^{-1}\\mathrm{Sign}^2,q^{1+j_1-i_2},q^{1+J-i_1-j_1}\end{matrix}
\Bigl|\, q,q\right)
\end{multline}
if $i_1+j_1=i_2+j_2$, and by 0 otherwise.
\end{corollary}
\begin{proof} This statement is a mere substitution of \cite[(5.8)-(5.9)]{Manga} into Theorem \ref{th:skew-R} and subsequent removal of certain factors of the form $f(j_1)/f(j_2)$, cf. Remark \ref{rm:skew-R}(i).
\end{proof}
Proposition \ref{pr:principal}(ii) (and hence (i) too via \eqref{eq:G-via-F}) can, in principle, be derived from Corollary \ref{cr:skew-G}, because for $G_{\nu/0^M}(v,qv,\dots,q^{J-1}v)$ only vertices with either $i_1=0$ or $j_1=0$ participate. For $i_1=0$, the ${}_4\bar{\phi}_3$ in \eqref{eq:w-via-phi} is simply equal to 1, and for $j_1=0$ only one term in the series expansion of ${}_4\bar{\phi}_3$ contributes and gives an elementary expression. Multiplying the weights over the row of vertices should yield \eqref{eq:principal-G}, but the computation is rather tedious.
Observe that the right-hand side of \eqref{eq:w-via-phi} is manifestly a polynomial in $q^J$, while the definition of $G_{\nu/\mu}(v,qv,\dots,q^{J-1}v)$ requires $J$ to be a positive integer. One might wonder if $G_{\nu/\mu}(v,qv,\dots,q^{J-1}v)$ can be analytically continued in $q^J$ in a natural way. One answer is provided by the following statement.
\begin{corollary}\label{cr:cauchy-principal} For any $M\ge 0$, $\mu\in \mathrm{Sign}_M^+$, and $v,q^J\in \C$, \emph{define} $G_{\nu/\mu}(v,qv,\dots,q^{J-1}v)$ via Corollary \ref{cr:skew-G}. Then for any $u_1,\dots,u_M\in \C$ in a small enough neighborhood of $s$, we have
\begin{equation}\label{eq:pieri-principal}
\prod_{i=1}^M\frac{1-q^J u_iv}{1-u_iv}\, F_\mu(u_1,\dots,u_M)=\sum_{\nu\in \mathrm{Sign}_{M}^+}\frac{c(\nu)}{c(\mu)}\, G_{\nu/\mu}(v,qv,\dots,q^{J-1}v)F_\nu(u_1,\dots,u_M),
\end{equation}
where the function $c(\,\cdot\,)$ is defined in \eqref{eq:def-c}.
\end{corollary}
\noindent\textbf{Comments.} \textbf{(i)} One way to think about \eqref{eq:pieri-principal} is as of a decomposition of the left-hand side, viewed as a function of $\mu\in \mathrm{Sign}_M$, in the basis of functions $\{F_{\nu}(u_1,\dots,u_M)\}_{\nu\in \mathrm{Sign}_M}$ on $\mathrm{Sign}_M$. Because of certain orthogonality relations for $F_\nu$ that we describe in the next section, the coefficients in such an expansion can be effectively extracted via contour integrals. This provides an alternative expression for $G_{\nu/\mu}(v,qv,\dots,q^{J-1}v)$ that is manifestly polynomial in $q^J$, cf. Remark \ref{rm:spatial-check} below. Proving the equality between this expression and that of Corollary \ref{cr:cauchy-principal} directly seems to be challenging though.
\smallskip
\noindent\textbf{(ii)} One can also view \eqref{eq:pieri-principal} as an eigenrelation for the `fused transfer-matrix' $$[G_{\nu/\mu}(v,qv,\dots,q^{J-1}v)]_{\nu,\mu\in\mathrm{Sign}_M}$$ and its eigenvector $\{c(\nu)F_{\nu}(u_1,\dots,u_M)\}_{\nu\in \mathrm{Sign}_M}$, cf. Remark \ref{rm:bethe}.
The product in the left-hand side of \eqref{eq:pieri-principal} is the corresponding eigenvalue, and it is indeed an infinite volume limit of an eigenvalue of the fused transfer matrix of the higher spin XXZ model.
\noindent\textbf{(iii)} For $\mu=0^M$, $q^J=(vs)^{-1}$, Corollary \ref{cr:cauchy-principal} yields \cite[Proposition 5.18]{BCPS2}. Note that for such value of $q^J$, $\nu$ must have no nonzero coordinates because otherwise the factor $(sv;q)_{N+k}=(q^{-J};q)_{J+k}$ in the right-hand side of \eqref{eq:principal-G} vanishes, and then for $G_\nu(v,qv,\dots,q^{J-1}v)$ we can use \eqref{eq:G-via-F} and \eqref{eq:principal-F}. I am very grateful to Leonid Petrov for pointing this connection out.
\begin{proof}[Proof of Corollary \ref{cr:cauchy-principal}] We argue by analytic continuation of the skew Cauchy identity \eqref{eq:skew-cauchy} with $N=J$, $(v_1,\dots,v_N)=(v,qv,\dots,q^{J-1}v)$, $\lambda=\varnothing$, viewed as an identity between polynomials in $q^J$ for $q^J\in \{q,q^2,q^3,\dots\}$.
Indeed, one readily sees from Corollary \ref{cr:skew-G} that $G_{\nu/\mu}(v,qv,\dots,q^{J-1}v)$ is a polynomial in $q^J$ of degree at most $M$, and that its absolute value grows at most as $const^{\nu_1}$ and $\nu_1\to\infty$ as long as $q^J$ stays in a compact subset of $\C$. On the other hand, one sees from \eqref{eq:symmetrization-F} that by taking $u_1,\dots,u_M$ into a small enough neighborhood of $s$ one can achieve that $|F_\nu(u_1,\dots,u_M)|$ decays, as $\nu_1\to\infty$, faster than any (small) positive constant to the power $\nu_1$. Hence, the series in the right-hand side of \eqref{eq:pieri-principal} is uniformly convergent for bounded $q^J$ and $u_i$'s close to $s$, the sum remains a polynomial in $q^J$ of degree at most $M$, and it can be analytically continued off any $M+1$ distinct points. Since the skew Cauchy identity implies \eqref{eq:pieri-principal} for $q^J$ in the infinite set $\{q,q^2,\dots,\}$, the proof is complete.
\end{proof}
While the expression in the right-hand side of \eqref{eq:w-via-phi} does not look too appetizing, it may simplify for special values of parameters. Here is an example of such a simplification at $v=s$.
\begin{proposition}\label{pr:q-hahn}
For any $J\ge 1$, and by analytic continuation of Corollaries \ref{cr:skew-G}-\ref{cr:cauchy-principal}, for any $q^J\in\C$, $G_{\nu/\mu}(s,qs,\dots,q^{J-1}s)$ corresponds (as a function of $\nu$ and $\mu$, in the sense of Definition \ref{df:corresponds}) to the vertex weights given by
\begin{equation}\label{eq:q-hahn}
{\widehat {w}}^{(J)}_s(s,qs,\dots,q^{J-1}s)=(-s)^{-j_1}(s^2q^J)^{j_1}\,\frac{(q^{-J};q)_{j_1}(s^2q^J;q)_{i_2-j_1}}{(s^2;q)_{j_2}}\,\frac{(q,q)_{i_2}}{(q,q)_{j_1}(q;q)_{i_2-j_1}}
\end{equation}
for $i_2\ge j_1$ and $i_1+j_1=i_2+j_2$, and by 0 otherwise.
\end{proposition}
\begin{remark}\label{rm:q-hahn} \textbf{(i)} The tilde and hat over `$w$' in the left-hand sides of \eqref{eq:w-via-phi} and \eqref{eq:q-hahn} symbolize that the expressions on the right-hand sides are different from \eqref{eq:w-via-R} by (irrelevant) factors of the form $f(j_1)/f(j_2)$, cf. Remark \ref{rm:skew-R}(i), that we remove to make the resulting expressions simpler.
\smallskip
\noindent\textbf{(ii)} The condition $i_2\ge j_1$ or, equivalently (modulo the default condition $i_1+j_1=i_2+j_2$), $i_1\ge j_2$, pictorially means that the number of paths that exit any vertex on the top is at least as large as the number of paths that enter the vertex from the left. This could be thought of as the condition that the up-right paths are not allowed to move horizontally by more than one unit. This restriction can also be seen in a different way: At $v=s$, the eigenvalue in the left-hand side of \eqref{eq:pieri-principal} has the form
$$
\prod_{i=1}^M \frac{1-q^Ju_is}{1-u_is}=\prod_{i=1}^M\left(\frac{1-q^Js^2}{1-s^2}+\frac{(q^J-1)s}{1-s^2}\cdot\frac{s-u_i}{1-u_is}\right),
$$
and the important part is that the factors are linear functions in $\xi_i=(s-u_i)/(1-u_is)$. When one multiplies such an expression by \eqref{eq:symmetrization-F}, it is natural to expect that powers $\nu_i$ of $\xi_j$'s increase by no more than one.
\smallskip
\noindent\textbf{(iii)} The right-hand side of \eqref{eq:q-hahn} coincides, up to the factor $(-s)^{j_1}$, with the jumping probabilities \cite[(8)]{Povolotsky} with the parameters $(\mu,\nu)$ of \cite{Povolotsky} related to ours via $\mu=s^2 q^J$, $\nu=s^2$ (I apologize again for the overloading of Greek letters). Further, Corollary \ref{cr:cauchy-principal} coincides with the eigenrelation first proved in \cite{Povolotsky}, and also re-stated and re-proved as \cite[Proposition 5.13]{BCPS2}.
The extra factor $(-s)^{j_1}$ is explained by a change of variables in the eigenfunctions: If we replace our variables $u_i$ in $F_\nu(u_1,\dots,u_M)$ by $sz_i$, $1\le i\le M$, then
$$
\frac{u_i-s}{1-su_i}=(-s)\frac{1-z_i}{1-s^2z_i}, \qquad 1\le z_i\le M,
$$
and when we raise this expression to power $\nu_{\sigma(i)}$ and take the product over $i$ as in \eqref{eq:symmetrization-F}, we obtain the extra factor of $(-s)^{\sum_i \nu_i}$. On the other hand, extra $(-s)^{j_1}$ in the weight of any vertex of type $(i_1,j_1;i_2,j_2)$
leads to the multiplication of $G_{\nu/\mu}$ corresponding to these vertex weights by $(-s)^{\sum_i(\mu_i-\nu_i)}$. We thus see that all these powers of $(-s)$ cancel out in the eigenrelation \eqref{eq:pieri-principal}.
\end{remark}
\begin{proof}[Proof of Proposition \ref{pr:q-hahn}] In principle, we simply need to substitute $v=s$ into the right-hand side of \eqref{eq:w-via-phi}, but what we literally read is not very illuminating. One way to proceed is to apply the transformation formula \cite[(B.3)]{Manga} that reads
$$
{}_{4}\bar{\phi}_3\left(\begin{matrix} q^{-m};a,b,c\\q^{1-m+n},e,f\end{matrix}
\Bigl|\, q,q\right)=\frac{(-1)^{m+n}(ab)^n(a,b,c;q)_{m-n}}{q^{n+\frac{(m-n)(m-n-1)}{2}}}
{}_{4}\bar{\phi}_3\left(\begin{matrix} q^{-n};\frac qa,\frac qb,cq^{m-n}\\q^{1-n+m},\frac{qe}{ab},\frac{qf}{ab}\end{matrix}\Bigl|\, q,q\right)
$$
with $m,n\in\mathbb{Z}_{\ge 0}$ and $abc=ef q^n$ (we used the abbreviated notation $(a,b,c;q)_l=(a;q)_l(b;q)_l(c;q)_l$ here). Setting $v=s$, choosing
$$
m=i_1,\ n=j_2,\ a=s^2q^J,\ b=q,\ c=q^{-i_2},\ e=s^2,\ f=q^{1+J-i_2-j_2},
$$
noting that $q/b=1$ and
$$
{}_{4}\bar{\phi}_3\left(\begin{matrix} q^{-N};A,1,C\\D,E,F\end{matrix}
\Bigl|\, q,z\right)=(D,E,F;q)_N,
$$
we rewrite the right-hand side of \eqref{eq:w-via-phi} with $v=s$, remembering that $i_1+j_1=i_2+j_2$, as
\begin{gather*}
\frac{(-1)^{i_1+j_1} q^{\frac{i_1^2+i_2^2}4+\frac{i_1(j_1-1)+i_2j_2}{2}}s^{j_1}(1;q)_{j_1-i_2}}{(q;q)_{i_1} (s^2;q)_{i_2+j_2}}\\
\times
\frac{(-1)^{i_1+j_2}(s^2q^{J+1})^{j_2}(s^2q^J,q,q^{-i_2};q)_{i_2-j_1}}{q^{j_2+\frac{(i_1-j_2)(i_1-j_2-1)}{2}}}\cdot(q^{1+i_1-j_2},q^{-J},q^{1-i_2-j_2}s^{-2};q)_{j_2}.
\end{gather*}
A few factors can now be simplified:
\begin{gather*}
(1;q)_{j_1-i_2}=\mathbf{1}_{i_2\ge j_1} \cdot (-1)^{i_2+j_1}\, \frac{q^\frac{(i_2-j_1)(i_2-j_1+1)}{2}}{(q;q)_{i_2-j_1}}\,;\\
\frac{(q^{1-i_2-j_2}s^{-2};q)_{j_2}}{(s^2;q)_{i_2+j_2}}=\frac{(-1)^{j_2}s^{-2j_2}q^{-\frac{(2i_2+j_2-1)j_2}{2}}}{(s^2;q)_{i_2}}\,;\\
\frac{(q^{-i_2};q)_{i_2-j_1}(q^{1+i_1-j_2};q)_{j_2}}{(q;q)_{i_1}}=\frac{(-1)^{i_2+j_1}q^{-\frac{(i_2+j_1+1)(i_2-j_1)}{2}}(q;q)_{i_2}}{(q;q)_{j_1}(q;q)_{i_2-j_1}}\,.
\end{gather*}
Collecting powers of $(-1)$ gives $(-1)^{j_1}$. Remembering that we can multiply vertex weights by $f(j_1)/f(j_2)$, cf. Remarks \ref{rm:skew-R}(i) and \ref{rm:q-hahn}(i), we can replace
$$
(s^2q^J)^{j_2}(q^{-J};q)_{j_2}\ \text{ by }\ (s^2q^J)^{j_1} (q^{-J};q)_{j_1} \qquad \text{and}\qquad s^{j_1}s^{-2j_2}\text{ by } s^{-j_1}.
$$
It remains to collect the powers of $q$. We read
$$
\frac{i_1^2+i_2^2}{4}+\frac{i_1(j_1-1)+i_2j_2}{2}+\frac{(i_2-j_1)(i_2-j_1+1)}{2}-\frac{(2i_2+j_2-1)j_2}{2}-\frac{(i_2+j_1+1)(i_2-j_1)}{2}.
$$
Substituting $i_1=i_2+j_2-j_1$ into this expression yields $\frac14(j_1^2-j_2^2)$, which can be removed from the vertex weight, because $q$ raised to that power has the form $f(j_1)/f(j_2)$. Gathering remaining factors leads to the desired expression.
\end{proof}
\section{Orthogonality relations}\label{sc:orthogonality} The functions $F_\nu(u_1,\dots,u_n)$ satisfy two types of (bi)orthogonality relations, both were proved in \cite{BCPS2} (one of the relations had been previously conjectured in \cite{Povolotsky}), and we shall restate them below. The goal of the section is to explain how results of the previous sections connect to these orthogonality relations.
We shall show that the Cauchy identity of Corollary \ref{cr:cauchy} essentially implies one of them, the so-called \emph{spectral biorthogonality}. We shall also show that the second one, the so-called \emph{spatial biorthogonality}, can be viewed as a link between the two symmetrization formulas of Theorem \ref{th:symmetrization}. This theorem can thus be used to give a proof of the spatial biorthogonality, but we stop short of doing that because the proof given in \cite{BCPS2} is shorter and more direct. On the other hand, one can also say that the spatial biorthogonality provides a route of \emph{deriving} (rather than verifying) the symmetrization formula for $G_\nu$ given the simpler, Bethe ansatz type symmetrization formula for $F_\mu$, and this is indeed how the formula for $G_\nu$ was obtained for the first time.
The orthogonality relations can be complemented with \emph{completeness} of the corresponding functional bases, thus providing Plan\-che\-rel type isomorphism theorems between suitable functional spaces. Such theorems were the primary goal of \cite{BCPS2}, where an interested reader can find their statements and proofs.
\begin{theorem}[Spectral orthogonality; Theorems 4.3, 6.7 of \cite{BCPS2}]\label{th:spectral} For any $n\ge 1$, we have
\begin{multline}
\label{eq:spectral-1}\sum_{\nu\in\mathrm{Sign}_n}\oint\cdots\oint\prod_{i=1}^n\frac{du_i}{2\pi\i} \oint\cdots\oint\prod_{i=1}^n\frac{dv_i}{2\pi\i}\sum_{\nu\in\mathrm{Sign}_n} c(\nu)F_{\nu}(u_1,\dots,u_n)F_\nu(v_1^{-1},\dots, v_n^{-1})\\ \times \prod_{1\le i<j\le n}(u_i-u_j)(v_i-v_j)\varphi(u_1,\dots,u_n)\psi(v_1,\dots,v_n)\\
=(-1)^{\frac{n(n-1)}{2}}\oint\cdots\oint\prod_{i=1}^n\frac{du_i}{2\pi\i}\prod_{1\le i, j\le n} (u_i-qu_j)\sum_{\sigma\in S_n}\varphi(u_1,\dots,u_n)\psi(u_{\sigma(1)},\dots,u_{\sigma(n)}),
\end{multline}
where $u_i$'s and $v_j$'s are integrated over positively oriented circles $|u_i-s|=|v_i-s|=\epsilon\ll 1$, $\varphi,\psi$ are suitable test functions, and $c(\,\cdot\,)$ is given by \eqref{eq:def-c}. Less formally, the above relation can be rewritten as
\begin{equation}
\begin{gathered}\label{eq:spectral-2}
\prod_{1\le i<j\le n}(u_i-u_j)(v_i-v_j)\sum_{\nu\in\mathrm{Sign}_n} c(\nu)F_{\nu}(u_1,\dots,u_n)F_\nu(v_1^{-1},\dots, v_n^{-1}) \\
=(-1)^{\frac{n(n-1)}{2}}\prod_{1\le i, j\le n} (u_i-qu_j)\cdot\det\left[\delta(v_i-u_j^{-1})\right]_{i,j=1}^n.
\end{gathered}
\end{equation}
\end{theorem}
The statement of Theorem \ref{th:spectral} is somewhat sloppy with not defining exactly the class of test functions for which \eqref{eq:spectral-1} holds. One possible class is described in \cite{BCPS2}, another one will come out of our proof of Theorem \ref{th:spectral} below. Neither of them is optimal, and since our goals are mostly algebraic here, we do not pursue this issue further.
\begin{theorem}[Spatial orthogonality; Corollary 3.13 of \cite{BCPS2}]\label{th:spatial} For any $n\ge 1$, $\mu,\nu\in\mathrm{Sign}_n$, we have
\begin{equation}\label{eq:spatial}
\frac{c(\nu)}{(q-1)^nn!} \oint\cdots\oint\prod_{i=1}^n\frac{du_i}{2\pi\i u_i}\prod_{1\le i\ne j\le n}\frac{u_i-u_j}{u_i-qu_j}\,F_\nu(u_1,\dots,u_n)F_\mu(u_1^{-1},\dots,u_n^{-1})=\mathbf 1_{\mu=\nu},
\end{equation}
where $u_i$'s are integrated over a positively oriented contour that contains points $\{s,qs,\dots, q^{n-1}s\}$ and its own image under multiplication by $q$, and that does not contain $s^{-1}$.
\end{theorem}
Let us proceed to showing how the Cauchy identity implies spectral orthogonality. The argument is similar to the proof of \cite[Proposition 6.10]{BCPS1}, where it was used for the Hall-Littlewood symmetric polynomials.
\begin{proof}[Proof of Theorem \ref{th:spectral}] We start with the Cauchy identity \eqref{eq:cauchy} with $M=N=n$, substitute $v_i\mapsto v_i^{-1}$, and rewrite it using \eqref{eq:G-via-F}:
\begin{multline}\label{eq:cauchy-rewrite}
\frac{\prod_{i=1}^n(1-su_i)}{(q;q)_n}\sum_{\nu:\,\nu_n=0} F_\nu(u_1,\dots,u_n) G_\nu^c(v_1^{-1},\dots, v_n^{-1})\\+
{\prod_{i=1}^n(1-su_i)(1-sv_i)}{}\sum_{\nu:\nu_n\ge 1}{c(\nu)} F_\nu(u_1,\dots,u_n)F_\nu(v_1^{-1},\dots,v_n^{-1})=\prod_{i,j=1}^n \frac{v_i-qu_j}{v_i-u_j}\,.
\end{multline}
It is convenient to use the notation
$$
\xi_i=\frac{u_i-s}{1-su_i},\quad u_i=\frac{\xi_i+s}{1+s\xi_i},\qquad \zeta_i=\frac{v_i-s}{1-sv_i},\quad v_i=\frac{\zeta_i+s}{1+s\zeta_i},\qquad 1\le i\le n.
$$
Let us multiply both sides of \eqref{eq:cauchy-rewrite} by
\begin{equation}\label{eq:multiply-by}
\prod_{i=1}^n \frac{\zeta_i^M}{\xi_i^M}\prod_{1\le i<j\le n}(u_i-u_j)(v_i-v_j)\cdot \varphi(\xi_1,\dots,\xi_n)\varphi(\zeta_1,\dots,\zeta_n),
\end{equation}
where $M$ is a large positive integer, and $\varphi(\xi_1,\dots,\xi_n),\psi(\zeta_1,\dots,\zeta_n)$ are test functions such that
$$
\prod_{i=1}^n(1+s\xi_i)^{-n-1}\varphi(\xi_1,\dots,\xi_n)\in\C[\xi^{\pm 1},\dots,\xi_n^{\pm 1}],
$$
and $\psi(\zeta_1,\dots,\zeta_n)$ is analytic in a punctured neighborhood of the origin with a possible finite order pole at the origin (as was noted above, these conditions on test functions are not optimal can be substantially relaxed, but we do not pursue that here). Let us further integrate over $|u_i-s|=\epsilon\ll 1$, $|v_i-s|=2\epsilon$.
Rewriting the $u$-integrals in terms of $\xi$-variables, we observe that our assumptions and \eqref{eq:symmetrization-F} guarantee that in each term of the left-hand side of \eqref{eq:cauchy-rewrite} we are integrating a Laurent polynomial in $\xi_i$'s. Furthermore, in the first sum with $\nu_n=0$, each term will have at least one $\xi_i$ raised to only very negative degrees --- this follows from the symmetrization identity \eqref{eq:symmetrization-F} and the presence of $\prod_i \xi_i^{-M}$ with $M\gg 1$. This implies that the sum with $\nu_n=0$ vanishes after the $u$-integration.
To deal with the sum with $\nu_n\ge 1$ we observe that, cf. Remark \ref{rm:symmetrization}(i),
\begin{equation}
\prod_{i=1}^n\frac{\zeta_i^M}{\xi_i^M}\sum_{\nu:\nu_n\ge 1} c(\nu)F_\nu(u_1,\dots,u_n)F_\nu(v_1^{-1},\dots,v_n^{-1})=\sum_{\nu:\nu_n\ge -M+1} c(\nu)F_\nu(u_1,\dots,u_n)F_\nu(v_1^{-1},\dots,v_n^{-1}).
\end{equation}
In the limit $M\to \infty$ we thus obtain the sum over all $\nu\in\mathrm{Sign}_n$. (With our test functions, this limit is actually a stabilization as terms with small or large enough $\nu_j$'s give zero contribution by the same reasoning that we used to remove the part with $\nu_n=0$ above.)
It remains to understand what the integral of the right-hand side of \eqref{eq:cauchy-rewrite} multiplied by \eqref{eq:multiply-by} gives. For that we shrink the $v$-contours to $s$. Due to $\prod_i \zeta_i^M$, $M\gg 1$, there is no singularity at $v_i=s$ for any $i$, $1\le i\le n$. Hence, for a nonzero contribution one needs to pick the residues at the first order poles $v_i=u_j$. Two different $v_i$'s cannot utilize that same $u_j$ because of the $\prod_{i<j}(v_i-v_j)$ factor in \eqref{eq:multiply-by}. Therefore, we end up with the sum
$$
\sum_{\sigma\in S_n} \mathrm{Res}_{v_i=u_{\sigma(i)}}(\cdots),
$$
and this yields the right-hand side of \eqref{eq:spectral-1}.
\end{proof}
Let us proceed to the spatial orthogonality \eqref{eq:spatial}. We are going to test it on the Cauchy identity \eqref{eq:cauchy}. More exactly, this identity provides a decomposition of any function of $u_1,\dots,u_M$ that has the form as in the right-hand side of \eqref{eq:cauchy} divided by the prefactor of the left-hand side, on $\{F_\nu(u_1,\dots,u_M)\}_{\nu\in s_M^+}$. We are going to extract the coefficients in such decomposition using \eqref{eq:spatial} and see that they are given by \eqref{eq:symmetrization-G'}:
\begin{proposition}\label{pr:spatial-check} For any $n,N\ge 0$, $\nu\in\mathrm{Sign}_n^+$, $v_1,\dots, v_N\in \C$ in a sufficiently small neighborhood of $s$, the expression
\begin{equation}\label{eq:spatial-check}
\frac{(q;q)_nc(0^n)}{(q-1)^nn!} \oint\cdots\oint\prod_{i=1}^n\frac{du_i}{2\pi\i u_i}\prod_{1\le i\ne j\le n}\frac{u_i-u_j}{u_i-qu_j}
\prod_{1\le i\le n}\left(\frac{1}{1-su_i}\prod_{\substack{1\le j\le N}}\frac{1-qu_iv_j}{1-u_iv_j}\right) F_\nu(u_1^{-1},\dots,u_n^{-1})
\end{equation}
with integration contours as in Theorem \ref{th:spatial},
coincides with the right-hand side of \eqref{eq:symmetrization-G'} with $k$ being the number of zero coordinates in $\nu$.
\end{proposition}
\begin{remark}\label{rm:spatial-check} \textbf{(i)} The above statement coupled with Theorem \ref{th:symmetrization}(ii) can be used to provide a proof of Theorem \ref{th:spatial}. To do that one needs to argue that functions $\{F_\nu(u_1,\dots,u_n)\}_{\nu\in\mathrm{Sign}_n}$ are linearly independent so that decompositions on them yield well-defined coefficients, and that the set of functions
$$
\prod_{1\le i\le n}\left(\frac{1}{1-su_i}\left(\frac{u_i-s}{1-su_i}\right)^a\prod_{\substack{1\le j\le N}}\frac{1-qu_iv_j}{1-u_iv_j}\right), \qquad
a\in\mathbb{Z},\ N\in \mathbb{Z}_{\ge 0},\ v_1,\dots,v_N\in\C,
$$
is sufficiently rich (the extra parameter $a$ appeared due to Remark \ref{rm:symmetrization}(i)). We do not pursue this direction here as the proof of Theorem \ref{th:spatial} given in \cite{BCPS2} is simpler.
\smallskip
\noindent\textbf{(ii)} One can similarly apply spatial orthogonality \eqref{eq:spatial} to the skew Cauchy identity \eqref{eq:pieri-F} with $\lambda=\varnothing$, thus obtaining a contour integral formula for the skew functions $G_{\nu/\mu}(v_1,\dots,v_N)$ with arbitrary $\mu,\nu\in \mathrm{Sign}_n$ and $v_1,\dots,v_N\in \C$. Via Theorem \ref{th:skew-R}, this can be used to obtain explicit expressions for the fully general higher spin $R$-matrix for the XXZ model.
\end{remark}
\begin{proof}[Proof of Proposition \ref{pr:spatial-check}] All parts of the integrand in \eqref{eq:spatial-check} are symmetric in $u_i$'s. This implies that we can de-symmetrize $F_\mu(u_1^{-1},\dots,u_n^{-1})$ using Theorem \ref{th:symmetrization}(i) and equivalently rewrite \eqref{eq:spatial-check} as (also recalling that $c(0^n)=(s^2;q)_n/(q;q)_n$)
\begin{multline}\label{eq:spatial-desymm}
{(s^2;q)_n} \oint\cdots\oint\prod_{i=1}^n\frac{du_i}{2\pi\i}\prod_{1\le i< j\le n}\frac{u_i-u_j}{u_i-qu_j}\\ \times
\prod_{1\le i\le n}\left(\frac{1}{(u_i-s)(1-su_i)}\left(\frac{1-su_i}{u_i-s}\right)^{\nu_i}\prod_{\substack{1\le j\le N}}\frac{1-qu_iv_j}{1-u_iv_j}\right).
\end{multline}
Let us first handle integration over the last $k$ variables (recall that $k$ is the number
of zero coordinates in $\nu$). If $k\ge 1$, then inside the $u_n$-integration contour, the integrand viewed as a function in $u_n$ has only one simple pole at $u_n=s$. (Indeed, by our assumption on the contours, the points $s^{-1}, v_1^{-1}, \dots, v_N^{-1}$ lie outside the $u_n$-contour.) Evaluating the residue gives, from all the factors that involve $u_n$,
$$
\frac{1}{1-s^2}\prod_{i=1}^{n-1}\frac{u_i-s}{u_i-qs}\prod_{j=1}^N \frac{1-qsv_j}{1-sv_j}\,.
$$
Assuming $k\ge 2$, i.e. $\nu_{n-1}=0$, we see that in $u_{n-1}$ the pole at $u_{n-1}=s$ gets canceled by the last expression, which has also introduced a simple pole at $u_{n-1}=qs$. As for $u_n$, there are no other singularities inside the $u_{n-1}$-integration contour, and we can evaluate the residue at $u_{n-1}=qs$, which gives, writing only factors that depend on $u_{n-1}$,
$$
\frac{1}{1-qs^2}\prod_{i=1}^{n-2}\frac{u_i-qs}{u_i-q^2s}\prod_{j=1}^N \frac{1-q^2sv_j}{1-qsv_j}\,.
$$
As we continue in this fashion, after $k$ steps our integral \eqref{eq:spatial-desymm} turns into
\begin{multline}\label{eq:spatial-without-zeroes}
\frac{(s^2;q)_n}{(s^2;q)_k} \oint\cdots\oint\prod_{i=1}^{n-k}\frac{du_i{(u_i-s)}}{2\pi\i(u_i-q^ks)}\prod_{j=1}^N\frac{1-q^ksv_j}{1-sv_j}\prod_{1\le i< j\le n-k}\frac{u_i-u_j}{u_i-qu_j}\\ \times
\prod_{1\le i\le n-k}\left(\frac{1}{(u_i-s)(1-su_i)}\left(\frac{1-su_i}{u_i-s}\right)^{\nu_i}\prod_{\substack{1\le j\le N}}\frac{1-qu_iv_j}{1-u_iv_j}\right).
\end{multline}
All remaining $\nu_1\ge\dots\ge \nu_{n-k}$ are at least 1, and we cannot proceed in the same way. However, this means that there are no singularities at $u_i=s^{-1}$. Also, there is $\sim |u_i|^{-2}$ decay near $u_i=\infty$, which means that if we peel off the contours and deform them to $\infty$, we only need to take into account the poles at $u_i=v_j^{-1}$. No two $u_i$'s are allowed to share the same pole because of the factor $\prod_{i<j}(u_i-u_j)$ in the integrand. Evaluating (negative) residues at all possible pole choices yields
\begin{multline*}
\frac{(s^2;q)_n}{(s^2;q)_k} \sum_{\substack{\sigma:\{1,\dots,n-k\}\to\{1,\dots,N\}\\ \sigma \text{ is injective}}}\prod_{i=1}^{n-k}\frac{{v_{\sigma(i)}^{-1}-s}}{v_{\sigma(i)}^{-1}-q^ks}\prod_{j=1}^N\frac{1-q^ksv_j}{1-sv_j}\prod_{1\le i< j\le n-k}\frac{v_{\sigma(i)}^{-1}-v_{\sigma(j)}^{-1}}{v_{\sigma(i)}^{-1}-qv_{\sigma(j)}^{-1}}\\ \times
\prod_{1\le i\le n-k}\left(\frac{1}{(v_{\sigma(i)}^{-1}-s)(1-sv_{\sigma(i)}^{-1})}\left(\frac{1-sv_{\sigma(i)}^{-1}}{v_{\sigma(i)}^{-1}-s}\right)^{\nu_i}(1-q)v_{\sigma(i)}^{-1}\prod_{\substack{1\le j\le N\\ j\ne \sigma(i)}}\frac{1-qv_{\sigma(i)}^{-1}v_j}{1-v_{\sigma(i)}^{-1}v_j}\right).
\end{multline*}
Setting $I=\mathrm{Ran}(\sigma)$ and simplifying, we see that the above expression coincides with the right-hand side of \eqref{eq:symmetrization-G'}.
\end{proof}
\section{Degenerations}\label{sc:degeneration}
The aim of this section is to indicate certain degeneration of the above results as parameters $(q,s)$ tend to certain special values. All of these degenerations can and should be approached independently, and we hope to return to them in a future work.
\subsection{Hall-Littlewood symmetric polynomials}\label{ss:HL} The Hall-Littlewood symmetric polynomials are very well studied, and we refer to \cite[Chapter III]{Macdonald1995} for definitions and notations.
In order to reach these polynomials from our definitions, it suffices to set $s=0$. Then the symmetrization formulas of Theorem \ref{th:symmetrization} imply that for any $M,N,n\ge 0$, $\mu=0^{m_0}1^{m_1}2^{m_2}\cdots\in\mathrm{Sign}_M^+$, $\nu=0^{n_0}1^{n_1}2^{n_2}\cdots\in\mathrm{Sign}_n^+$, we have
\begin{gather*}
F_\mu(u_1,\dots,u_M)=\prod_{i\ge 0}{(q;q)_{m_i}}\cdot P_{\mu}(u_1,\dots,u_M)={(q;q)_{m_0}}\cdot Q_{\mu}(u_1,\dots,u_n),\\
G_\nu(v_1,\dots,v_N)=\prod_{i\ge 1}{(q;q)_{n_i}}\cdot P_{\nu}(v_1,\dots,v_N)=Q_{\nu}(v_1,\dots,v_N),
\end{gather*}
with $P_\lambda$'s and $Q_\lambda$'s as in \cite[\S III.2]{Macdonald1995}, and the parameter $t$ of \cite{Macdonald1995} is identified with our $q$.
Recalling Definition \ref{df:conjugated}, we can also compare the skew functions \cite[Ch. III, (5.11'), (5.14)]{Macdonald1995} and Definitions \ref{df:F}, \ref{df:G} to conclude that
\begin{align*}
F_{\lambda/\mu}^c(u)&=P_{\lambda/\mu}(u), \qquad \lambda\in\mathrm{Sign}_L^+,\quad \mu\in\mathrm{Sign}_{L+1}^+,\quad u\in \C,\\
G_{\lambda/\nu}(v)&=Q_{\lambda/\nu}(v), \qquad \lambda\in\mathrm{Sign}_L^+,\quad \mu\in\mathrm{Sign}_{L}^+,\quad v\in \C.
\end{align*}
Since the branching relations of Proposition \ref{pr:branching} are exactly the same as for the Hall-Littlewood polynomials, cf. \cite[Ch. III, (5.5), (5.6)]{Macdonald1995}, the formulas for $F_\mu,G_\nu$ above also follow from the formulas for the skew functions.
The skew Cauchy identity of Theorem \ref{th:skew-cauchy-single} and Corollary \ref{cr:skew-cauchy} now matches the corresponding identity for the Hall-Littlewood polynomials, see \cite[Ch. VI, Ex. 7.6]{Macdonald1995} where this identity is stated in the more general context of Macdonald polynomials. The Cauchy identity of Corollary \ref{cr:cauchy} turns into \cite[Ch. III, (4.4)]{Macdonald1995}.
The Pieri type rules of Corollary \ref{cr:pieri}, after decomposing the product-eigenvalue in the left-hand sides according to the Cauchy identity and comparing the same degree coefficients of both sides, coincide with \cite[Ch. III, (5.7), (5.7')]{Macdonald1995}.
The spatial orthogonality of Theorem \ref{th:spatial} is \cite[Ch. VI, (9.5)-(9.6)]{Macdonald1995}, where again the statement in \cite{Macdonald1995} is in the more general context of Macdonald polynomials.
The principal specializations of $F_\mu$ and $G_\nu$ as in Proposition \ref{pr:principal} correspond to \cite[Ch. III, Ex. 2.1]{Macdonald1995}. I do not know however if analogs
of Theorem \ref{th:skew-R} and Corollary \ref{cr:skew-G} that describe principal specialization of the skew functions $G_{\nu/\mu}$, have been considered in the Hall-Littlewood context.
\subsection{Inhomogeneous Hall-Littlewood polynomials}\label{ss:iHL} Instead of simply setting $s=0$ as we did in Section \ref{ss:HL}, let us send $s\to 0$ but also simultaneously scale the variables $u_i=sz_i$, $v_i=sw_i$. Then the weights $w_u(i_1,j_1;i_2,j_2)$ of Definition \ref{df:weights} divided by $s^{j_1}$ turn into
\begin{align*}
\widetilde w_z(m,0,m,0)&=1, \qquad
\widetilde w_z(m,1,m,1)={z-q^m}, \\
\widetilde w_z(m+1,0,m,1)&={z},\qquad
\widetilde w_z(m,1,m+1,0)={1-q^{m+1}},
\end{align*}
which implies, via Definitions \ref{df:F}, \ref{df:G}, that there exist limits
$$
\widetilde F_{\lambda/\mu}(z_1,\dots,z_m)=\lim_{s\to 0} F_{\lambda/\mu}(sz_1,\dots,sz_m),\qquad
\widetilde G_{\lambda/\nu}(w_1,\dots,w_n)=\lim_{s\to 0} G_{\lambda/\mu}(sw_1,\dots,sw_n),
$$
and they are (inhomogeneous) polynomials whose top homogeneous coefficients coincide with the corresponding Hall-Littlewood versions of $F_{\lambda/\mu}$ and $G_{\lambda/\mu}$ from the previous section.
Taking the same limit in the symmetrization formulas of Theorem \ref{th:symmetrization}, we read (using the notations of Theorem \ref{th:symmetrization})
\begin{align*}
\widetilde F_\mu(z_1,\dots,z_M)&={(1-q)^M}\,\sum_{\sigma\in S_M}\sigma\left(\prod_{1\le i<j\le M}\frac{z_i-qz_j}{z_i-z_j}\cdot\prod_{i=1}^M \left({z_i-1}\right)^{\mu_i}\right),
\\
\widetilde G_\nu(w_1,\dots,w_N)&=\frac{(1-q)^N}{(q;q)_{N-n+k}}\sum_{\sigma\in S_N}\sigma\left(\prod_{1\le i<j\le N} \frac{w_i-qw_j}{w_i-w_j}\cdot \prod_{i=1}^{n-k} {w_i}\left({w_i-1}\right)^{\nu_i-1}\right).
\end{align*}
All the results we proved for $F_{\lambda/\mu}$ and $G_{\lambda/\mu}$ carry over to $\widetilde F_{\lambda/\mu}$ and $\widetilde G_{\lambda/\mu}$, and to my best knowledge, none of them have appeared in the literature before, with the exception of the orthogonality relations and Remark \ref{rm:q-hahn}(iii), whose analogs were proved in \cite{BCPS1}.
\subsection{The Schur like case: $q=0$}\label{ss:schur} The Schur symmetric polynomials, see e.g. \cite[Chapter I]{Macdonald1995} can be thought of as specializations of the Hall-Littlewood symmetric polynomials with the parameter set to 0. Accordingly, we can set $q=0$ in our definitions of $F$- and $G$-functions. The vertex weights of Definition \ref{df:weights} then take the form
\begin{align*}
w_u^{(q=0)}(m,0,m,0)&=\frac{1-\mathbf{1}_{m=0}\cdot su}{1-su},\qquad
w_u^{(q=0)}(m,1,m,1)=\frac{u-\mathbf{1}_{m=0}\cdot s}{1-su}, \\
w_u^{(q=0)}(m+1,0,m,1)&=\frac{\left(1-\mathbf{1}_{m=0}\cdot s^2\right)u}{1-su},\qquad
w_u^{(q=0)}(m,1,m+1,0)=\frac{1}{1-su}.
\end{align*}
The symmetrization formulas of Theorem \ref{th:symmetrization} take the form of ratios of two determinants:
\begin{align*}
F_\mu^{(q=0)}(u_1,\dots,u_M)&=\frac{\det\left[\dfrac{u_i^{M-j}}{1-su_i}\left(\dfrac{u_i-s}{1-su_i}\right)^{\mu_j}\right]_{i,j=1}^M}{\prod_{1\le i<j\le M}(u_i-u_j)}\,,\\
G_\nu^{(q=0)}(v_1,\dots,v_N)
&=(1-s^2)^{\mathbf{1}_{k=0}}\frac{\det\left[\dfrac{v_i^{N-j}}{1-sv_i}\left(\dfrac{v_i}{v_i-s}\right)^{\mathbf{1}_{\nu_j>0}}\left(\dfrac{v_i-s}{1-sv_i}\right)^{\nu_j}\right]_{i,j=1}^N}{\prod_{1\le i<j\le N}(v_i-v_j)}\,.
\end{align*}
At $s=0$ both formulas turn into the celebrated formula for the Schur polynomials as a ratio of two alternants \cite[Ch. I, (3.1)]{Macdonald1995}.
One easily sees that for $q=0$, the conjugation of Definition \ref{df:conjugated} leaves all skew $G$-functions unaffected, which means that it can be removed from all the Cauchy and Pieri type formulas of Section \ref{sc:cauchy}.
As in the previous section, all our results carry over to this degenerate case, and it remains unclear to me whether any of them have been considered before.
\subsection{Inhomogeneous Schur polynomials}\label{ss:ischur} By combining the degeneration procedures of Sections \ref{ss:iHL} and \ref{ss:schur}, i.e. taking $q=0$, $s\to 0$, and scaling the variables $u_i=sz_i$, $v_i=sw_i$, we observe the vertex weights
\begin{align*}
\widetilde w_z^{(q=0)}(m,0,m,0)&=1,\qquad
\widetilde w_z^{(q=0)}(m,1,m,1)={z-\mathbf{1}_{m=0}}, \\
\widetilde w_z^{(q=0)}(m+1,0,m,1)&=z,\qquad
\widetilde w_z^{(q=0)}(m,1,m+1,0)=1,
\end{align*}
and symmetrization formulas
\begin{align*}
\widetilde F_\mu^{(q=0)}(z_1,\dots,z_M)&=\frac{\det\left[{z_i^{M-j}}\left({z_i-1}\right)^{\mu_j}\right]_{i,j=1}^M}{\prod_{1\le i<j\le M}(z_i-z_j)}\,,\\
\widetilde G_\nu^{(q=0)}(w_1,\dots,w_N)
&=\frac{\det\left[{w_i^{N-j}}\left(\dfrac{w_i}{w_i-1}\right)^{\mathbf{1}_{\nu_j>0}}\left({w_i-1}\right)^{\nu_j}\right]_{i,j=1}^N}{\prod_{1\le i<j\le N}(w_i-w_j)}\,.
\end{align*}
These are inhomogeneous polynomials whose top homogeneous components coincide with the Schur polynomials.
The polynomials $\widetilde F_\mu^{(q=0)}(z_1,\dots,z_M)$ bear a certain similarity to the so-called Grothendieck polynomials as presented in \cite{MotegiSakai1}, \cite{MotegiSakai2}, see also references therein, as well as \cite{LS}, \cite{Lenart} for much earlier works on those polynomials. However, on the surface the connection to integrable lattice models, skew functions, and the Cauchy formulas for these two objects appear to be different.
It would be very interesting to establish a direct link.
As in Sections \ref{ss:iHL}, \ref{ss:schur}, our results carry over to this case as well, and we have so far been unable to find them in the literature.
\subsection{Trigonometric to rational limit: $q\to 1$}\label{ss:XXX} The limit we consider here is equivalent to the well-known transition from the XXZ to the XXX model. We take
$$
q=\exp(\epsilon),\quad s=\exp(\epsilon\zeta), \quad u_i=\exp(\epsilon x_i),\quad v_i=\exp(\epsilon y_i), \qquad \epsilon\to 0.
$$
Such limit of the vertex weights of Definition \ref{df:weights} gives
\begin{align*}
w_x^{\mathrm{(rational)}}(m,0,m,0)&=\frac{m+\zeta+x}{\zeta+x}, \qquad
w_x^{\mathrm{(rational)}}(m,1,m,1)=\frac{m+\zeta-x}{\zeta+x}, \\
w_x^{\mathrm{(rational)}}(m+1,0,m,1)&=\frac{m+2\zeta}{\zeta+x},\qquad
w_x^{\mathrm{(rational)}}(m,1,m+1,0)=\frac{m+1}{\zeta+x}.
\end{align*}
Taking the limit of symmetrization formulas of Theorem \ref{th:symmetrization} yields
\begin{align*}
F_\mu^{\mathrm{(rational)}}(x_1,\dots,x_M)&=\frac{1}{\prod_{i=1}^M (\zeta+x_i)}\,\sum_{\sigma\in S_M}\sigma\left(\prod_{1\le i<j\le M}\frac{x_i-x_j-1}{x_i-x_j}\cdot\prod_{i=1}^M \left(\frac{\zeta-x_i}{\zeta+x_i}\right)^{\mu_i}\right)\,,\\
G_\nu^{\mathrm{(rational)}}(y_1,\dots,y_N)&=\frac{(2\zeta)_n}{(N-n+k)!(2\zeta)_k}\\ \times \sum_{\sigma\in S_N}\sigma\Biggl(\prod_{1\le i<j\le N} &\frac{y_i-y_j-1}{y_i-y_j}\cdot \prod_{i=1}^{n-k} \frac{1}{(\zeta+y_i)(\zeta-y_i)}\left(\frac{\zeta-y_i}{\zeta+y_i}\right)^{\nu_i}\cdot \prod_{j=n-k+1}^N\frac{k+\zeta+y_i}{\zeta+y_i}\Biggr),
\end{align*}
with the Pochhammer notation $(a)_m=a(a+1)\cdots(a+m-1)$ for $m\ge 1$, and 1 for $m=0$.
Once again, our results also have such limits, and we have not seen those in the literature.
|
\section{Introduction}
A variety of unconventional superconductors present low density of the charge carriers as a common factor, implying that it could be the basis for
a unifying picture to understand the superconductivity in such exotic systems. Low density of charge carriers is one of the characteristic features which is shared by
cuprates, fullerenes and MgB$_2$ \cite{Fleming1991,Holczer1991,Nagamatsu2001}. This is quite surprising since low carrier density is an unfavourable element for superconductivity within the conventional framework of BCS \cite{Bardeen1957} or Migdal${-}$Eliashberg \cite{Migdal1958,Eliashberg1960} theories. Moreover, a small superfluid density, is unavoidably related to poor screening and strong electronic correlations, ingredients which are expected to be also detrimental for conventional superconductivity. On these grounds it is hard to understand why these low carrier materials are the best superconductors. As far as the superconductivity exhibited by inter-metallic compounds is concerned, the role of electron-phonon interaction cannot be overlooked. However, one may have to look beyond the conventional framework of BCS or Migdal${-}$Eliashberg theories in order to understand the unconventional superconductivity in these compounds. From the experimental side, it is important to look for new superconducting materials with low carrier density.
\section{Experimental Details}
Single crystals of Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$ have been grown using Czochralski crystal pulling method in a tetra-arc furnace under high purity argon atmosphere. Stoichiometric ammount of Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$ (10~g each) was taken and melted 4-5 times in the tetra-arc furnace to make a homogeneous polycrystalline mixture. Single crystals were pulled using a tungsten seed rod at the rate of 10 mm/h for about 6 h to get 5-6~mm long and 3-4~mm thick crystals. The phase purity was characterized by powder X-ray diffraction using PANanalytical X-ray diffractometer. Single crystals were oriented along the crystallographic direction [100] using Laue back reflection using Huber Laue diffractometer and cut to desired shape and dimensions using a spark erosion cutting machine. Resistivity measurements were done in a home made setup using standard four-probe technique. Magnetization measurements were done in commercial SQUID magnetometer (MPMS5, Quantum Design, USA) and heat capacity measurements were done using PPMS.
\section{Results and Discussion}
The crystal structure of Y$_3$Ru$_4$Ge$_{13}$ is shown in Fig.~\ref{fig:fig1}. Both compounds have same crystal structure and cubic symmetry ($\it{Pm3n}$, space group $\#$ 223). Rietveld analysis \cite{Carvajal1993} of the Powder X-ray diffraction (PXRD) of Y$_3$Ru$_4$Ge$_{13}$ is shown in Fig.~\ref{fig:fig2}. The temperature dependence of resistivity $\rho (T)$ from 300 to 2K for Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$ are shown in Fig.~\ref{fig:fig3}. A semi-metallic behaviour $( \frac{d\rho}{dT}<0 )$ can be observed in the normal state resistivity data of both the compounds.
\begin{figure}[h]
\begin{minipage}[t]{0.45\linewidth}
\includegraphics[height=6cm]{1a}
\caption {Crystal structure of Y$_3$Ru$_4$Ge$_{13}$. Yttrium atoms are shown in green in 6d position, Ruthenium are shown in light pink in 8e position and Germanium are shown in
dark blue in 2a and 24k Wyckoff positions.}
\label{fig:fig1}
\end{minipage}
\hfill
\begin{minipage}[t]{0.45\linewidth}
\includegraphics[height=6cm]{1b}
\caption{Rietveld analysis of the powder XRD pattern of Y$_3$Ru$_4$Ge$_{13}$. No impurity peaks are observed indicating single phase nature of the compound. Similar PXRD pattern is also observed for Lu$_3$Os$_4$Ge$_{13}$.}
\label{fig:fig2}
\end{minipage}
\end{figure}
\begin{figure}[ht]
\hfill
\subfigure{\includegraphics[width=7cm]{2a}}
\label{fig:fig2a}
\hfill
\subfigure{\includegraphics[width=7cm]{2b}}
\label{fig:fig1b}
\hfill
\caption{The temperature dependence of the electrical resistivity ($\rho$) along the (100) directions of cubic (Pm3n) Y$3$Ru$4$Ge$13$ and Lu$_3$Os$_4$Ge$_{13}$. Insets show the low temperature data indicating superconducting transition in both compounds. Resistivity data from 2 to 300K clearly shows the semi-metallic nature of both the compounds.}
\label{fig:fig3}
\end{figure}
\begin{figure}[h]
\hfill
\subfigure{\includegraphics[width=7cm]{3a.eps}}
\label{fig:fig3a}
\hfill
\subfigure{\includegraphics[width=7cm]{3b.eps}}
\label{fig:fig3b}
\hfill
\caption{The resistance vs temperature data at different magnetic fields for Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$. With the increase in applied magnetic field the superconducting transition becomes slightly broader for both compounds. The temperature dependence of the upper critical field ($\mu{_0}Hc{_2}(T)$) is extracted from these measurements.}
\label{fig:fig4}
\end{figure}
The magnetoresistance data for Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$ is shown in Fig.~\ref{fig:fig4}. The width of the superconducting transition increases with increasing magnetic field. The transition temperature is taken at the point where resistivity becomes half of its normal state value.
The temperature dependence of the upper critical field for Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$ is shown in Fig.~\ref{fig:fig5}. We estimate the orbital upper critical field, $\mu{_0}Hc{_2}(0)$, for both the compounds using Werthamer-Helfand-Hohenberg (WHH) expression\cite{Werthamer1966}, $\mu{_0}Hc{_2}(0) = -0.693 ~T{_c}\frac{dHc{_2}}{dT}\vert{_{T=T{_c}}}$ in the dirty limit for type-II superconductors. A nearly linear relationship is observed in Fig.~\ref{fig:fig5} between $\mu{_0}Hc{_2}$ and $T{_c}$ in the proximity of the transition temperature ($T{_c}$ at H = 0) for both the compounds but the linear trend is more prominent for Lu$_3$Os$_4$Ge$_{13}$.
\begin{figure}[ht]
\hfill
\subfigure{\includegraphics[width=7cm]{4a.eps}}
\label{fig:fig4a}
\hfill
\subfigure{\includegraphics[width=7cm]{4b.eps}}
\label{fig:fig4a}
\hfill
\caption{$\mu{_0}H{_{c2}}$ as a function of temperature for Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$. The upper critical field $\mu{_0}H{_{c2}}$ increases linearly as the temperature is lowered in the vicinity of the transition temperature $T_{c}$ for both compounds, though the linear dependence is more prominent for Lu$_3$Os$_4$Ge$_{13}$.}
\label{fig:fig5}
\end{figure}
The slope $\frac{dHc{_2}}{dT}\vert{_{T=T{_c}}}$ is used to calculate $\mu{_0}H{_{c2}} = 4.63\pm0.09$~T for Y$_3$Ru$_4$Ge$_{13}$ and $\mu{_0}Hc{_2}= 5.68\pm 0.12$~T for Lu$_3$Os$_4$Ge$_{13}$ using the WHH formula in the dirty limit. The value of $\mu{_0}H{_{c2}}$ is smaller than the weak coupling Pauli paramagnetic limit $\mu{_0}H{^\mathrm{Pauli}}=1.82T_{c}=5.09$~T for Y$_3$Ru$_4$Ge$_{13}$ and $\mu{_0}H{^\mathrm{Pauli}}=5.80$~T for Lu$_3$Os$_4$Ge$_{13}$. The upper critical field value $\mu{_0}H{_{c2}}(0)$ can be used to estimate the Ginzburg-Landau coherence length $\xi(0){_{GL}}=\sqrt{\Phi{_0}/{2\pi{H{_{c2}}(0)}}}= 80.4\pm{0.5}\AA$~for Y$_3$Ru$_4$Ge$_{13}$ and $\xi(0){_{GL}}= 76.1\pm{0.7}\AA$~for Lu$_3$Os$_4$Ge$_{13}$, where $\Phi{_0}={hc}/{2e}$ is the magnetic flux quantum.\\
The DC-magnetisation data of both compounds is shown in Fig.~\ref{fig:fig6} indicating diamagnetic transitions of Y$_3$Ru$_4$Ge$_{13}$ at 2.8~K and Lu$_3$Os$_4$Ge$_{13}$ at 3.1~K. Very similar values of $T_{c}$ from both resistivity and susceptibility data confirm that our single crystals are of very high quality. Large vortex pinning can be observed in the field cooled (FC-Meissner) data in shown in Fig.~\ref{fig:fig6}.
\begin{figure}[h]
\hfill
\subfigure{\includegraphics[width=7cm]{5a.eps}}
\label{fig:fig5a}
\hfill
\subfigure{\includegraphics[width=7cm]{5b.eps}}
\label{fig:fig5b}
\hfill
\caption{DC magnetic susceptibility as a function of temperature for magnetic field $H\parallel[100]$ direction for Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$. The superconducting transition temperatures determined from susceptibility measurements are in excellent agreement with the resistivity data reflecting the high quality of the single crystals. The ZFC and FC susceptibility data indicate significant amount of pinning of vortices in both the compounds.}
\label{fig:fig6}
\end{figure}
\begin{figure}[ht]
\hfill
\subfigure{\includegraphics[width=7cm]{6a.eps}}
\label{fig:fig6a}
\hfill
\subfigure{\includegraphics[width=7cm]{6b.eps}}
\label{fig:fig6b}
\hfill
\caption{{C/T} vs $T{^2}$ data for Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$ respectively. Large jump in heat capacity confirms bulk superconductivity in both the compounds.}
\label{fig:fig7}
\end{figure}
The characterisation of the superconducting transition using heat capacity measurements is shown in Fig.~\ref{fig:fig7}. The specific heat jump at the thermodynamic transition confirm the bulk superconductivity in both the compounds. The low temperature normal state specific heat can be well fitted with $\frac{C}{T} = \gamma + \beta T^2$, where $\gamma T$ represents the electronic contribution and $\beta T^3$ describe the lattice-phonon contributions to the specific heat in the normal state. Fitting the above formula give electronic specific heat coefficient $\gamma= 7.08\frac{mJ}{mol K^2}$ ($\gamma =25.4 \frac{mJ}{mol K^2}$ ) and the phonon/lattice contributions $\beta = 3.52\frac{mJ}{mol K^4}$$(\beta = 2.30\frac{mJ}{mol K^2})$ for Y$_3$Ru$_4$Ge$_{13}$ (Lu$_3$Os$_4$Ge$_{13}$). The ratio $\frac{\Delta C}{\gamma Tc}$ can be used to measure the strength of the electron-phonon coupling. The specific heat jump $\frac{\Delta C}{Tc}$ is $6.07\frac{mJ}{mol K^2}$ ($29\frac{mJ}{mol K^2}$), setting the value of $\frac{\Delta C}{\gamma Tc} = 0.85$ ( $\frac{\Delta C}{\gamma Tc} = 1.15$ ) for Y$_3$Ru$_4$Ge$_{13}$ (Lu$_3$Os$_4$Ge$_{13}$). These values are smaller than the weak-coupling limit of 1.43 for a conventional BCS superconductor, suggesting that these two compounds are moderately electron-phonon coupled superconductor.\\
The comparison among the normal and superconducting state parameters of the both compounds is shown in Table-\ref{table:table1}. We also notice from Table-\ref{table:table1} that value of $\gamma$ is larger in Lu$_3$Os$_4$Ge$_{13}$ suggesting stronger electronic correlations in Lu$_3$Os$_4$Ge$_{13}$ as compared to electronic correlations in Y$_3$Ru$_4$Ge$_{13}$.
\begin{center}
\begin{table}[h]
\centering
\caption{\label{tab:params} Normal and superconducting state parameters of Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$}
\begin{tabular}{lcc}
\br
Parameters &Y$_3$Ru$_4$Ge$_{13}$ & Lu$_3$Os$_4$Ge$_{13}$\\
\mr
$T_{\rm c}$~(K) &2.85 & 3.1 \\
$\gamma$~(mJ/mol\,K$^{2}$) &7.1 & 25.4 \\
$\Theta_{\rm D}$~(K) &223 & 257 \\
$\Delta C_{\rm el}/\gamma T_{\rm c}$ &0.85 & 1.15 \\
$\mu{_0}H{_{c2}}$~(T) &4.63 &5.68\\
$\xi(0)_{GL}$ ($\AA$) &80.4 & 78 \\
\br
\end{tabular}
\label{table:table1}
\end{table}
\end{center}
\section{Conclusion}
We have grown single crystals and characterised the superconducting properties of two semi-metallic compounds Y$_3$Ru$_4$Ge$_{13}$ and Lu$_3$Os$_4$Ge$_{13}$. A bulk superconducting transition is confirmed and characterised through electrical transport, magnetisation and heat capacity measurements on the single crystals. The magnetic susceptibility measurements show large pinning of vortices in both the compounds. The analysis of the low temperature heat capacity data suggests that both these compounds are moderately electron-phonon coupled type-$II$ superconductors.
\section*{References}
\raggedright
|
\section{Introduction}
The Westervelt equation for the acoustic pressure $p$
\begin{equation}\label{Westervelt}
(p-k(p^2))_{tt} -c^2 \Delta p - b\Delta p_t = 0,
\end{equation}
where $k = \beta_a / \lambda$, $\beta_a = 1 + B/(2A)$,
$\lambda=\varrho c^2$ is the bulk modulus,
$\varrho$ is the mass density,
$c$ is the speed of sound,
$b$ is the diffusivity of sound,
$B/A$ is the parameter of nonlinearity,
is a classical model of nonlinear acoustics,
cf.~\cite{HamiltonBlackstock,manfred,Westervelt}.
In particular, it is widely used for the simulation of high intensity focused ultrasound (HIFU) which
has a broad range of technical and medical applications ranging from lithotripsy or thermotherapy
to ultrasound cleaning or welding and sonochemistry, see~\cite{abramov,manfred} and the references therein.
Note that the Westervelt equation can be written alternatively in terms of the acoustic velocity potential $\psi$,
\begin{equation}\label{Westervelt_psi}
(\psi_t-\tilde{k}(\psi_t^2))_t -c^2 \Delta \psi - b\Delta\psi_t
= 0,
\end{equation}
where $\varrho \psi_t = p$ and $\tilde{k}=\varrho k$.
An analysis of the Westervelt equation with homogeneous~\cite{KL08} and inhomogeneous~\cite{KLV10} Dirichlet
and Neumann~\cite{Neumann} boundary conditions as well as with boundary instead of interior damping~\cite{K10}
has yielded well-posedness and exponential decay of small and regular (i.e., $H^2(\Omega)$) solutions.
Here $\Omega\subseteq\mathbb{R}^d$, $d\in\{1,2,3\}$ is the spatial domain on which \eqref{Westervelt} is considered.
An important open question to be addressed in this paper is existence of {\em spatially less regular solutions}
as needed, e.g., for coupling with elastic or acoustic regions exhibiting different material parameters in the
simulation of a focusing silicone lens immersed in an acoustic medium.
An important feature of equation \eqref{Westervelt} (similarly also of \eqref{Westervelt_psi}) is the potential
degeneracy due to the factor $(1-2kp)$ of the second time derivative $p_{tt}$. In order to avoid degeneracy,
it is crucial to obtain an $L_\infty$-estimate of $p$ in order to bound $(1-2kp)$ away from zero.
This has been achieved until now by combining a bound of $\Delta p$ (obtained by energy estimates) with
Sobolev's embedding $H^2(\Omega)\to L_\infty(\Omega)$. In this paper, we will employ the nonlinear damping
for this purpose instead. Note that, indeed, the particular choice of the damping is to some extent left open
from the point of view of the physical modeling. We will use this freedom to devise possible damping terms
leading to existence of $H^1(\Omega)$-solutions.
We consider three modifications of the homogeneous Dirichlet boundary value problem associated with
\eqref{Westervelt} and \eqref{Westervelt_psi}, namely
\begin{equation}\label{Wpress_pLaplace}
\begin{cases}
(1-2ku)u_{tt}
-c^2\,\text{div}\,\Bigl(\nabla u+\varepsilon|\nabla u|^{p-1}\nabla u\Bigr)-b\Delta u_t=2k(u_t)^2,
\\
(u,u_t)|_{t=0}=(u_0, u_1),
\\
u|_{\partial \Omega} =0,
\end{cases}
\end{equation}
\begin{equation}\label{Wpress_viscosity}
\begin{cases}
(1-2ku)u_{tt}-c^2\Delta u-b\,\text{div}\,\Bigl(((1-\delta) +\delta|\nabla u_t|^{q-1})\nabla u_t\Bigr)
=2k(u_t)^2,
\\
(u,u_t)|_{t=0}=(u_0, u_1),
\\
u|_{\partial \Omega} =0,
.\\
\end{cases}
\end{equation}
\begin{equation}\label{Wpot_viscosity}
\begin{cases}
u_{tt}-\frac{c^2}{1-2\tilde{k}u_t}\Delta u
-b\,\text{div}\,\Bigl(((1-\delta) +\delta|\nabla u_t|^{q-1})\nabla u_t\Bigr)
=0,
\\
(u,u_t)|_{t=0}=(u_0, u_1),
\\
u|_{\partial \Omega} =0,
.\\
\end{cases}
\end{equation}
in a smooth domain $\Omega\subseteq\mathbb{R}^d$, $d\in\{1,2,3\}$ with $\varepsilon\geq0$, $\delta\in [0,1]$ and $p,q\geq1$.
The constant parameters $b,c$ will be positive while $k$ will not be assumed to have a particular sign.
Equations \eqref{Wpress_pLaplace} and \eqref{Wpress_viscosity} are motivated by \eqref{Westervelt}
while \eqref{Wpot_viscosity} comes from \eqref{Westervelt_psi}.
In all equations we changed the notation to the typical mathematical one for solutions of PDEs, i.e., $p\to u$, and $\psi\tou$, respectively.
Note that if $\varepsilon=0$ or if $p=1$ in \eqref{Wpress_pLaplace}, one obtains the classical Westervelt equation;
likewise, for $\delta=0$ or $q=1$ in \eqref{Wpress_viscosity} and \eqref{Wpot_viscosity}. Here we will analyze the problem for
\[
\varepsilon,\delta>0\text{ and }p,q>d-1\,,
\]
where $d$ is the space dimension
such that $W^{1,p+1}(\Omega)$ and $W^{1,q+1}(\Omega)$ are continuously embedded in $L_\infty(\Omega)$; precise conditions for these inclusions to hold will be discussed later in the paper. In Sections \ref{secWpot_viscosity} and \ref{secWelastic_viscosity} we will have to to use the stronger assumption $q\geq 3$,
since there we will need $W^{1,q+1}(\Omega)$ to be continuosly embedded in $W^{1,4}(\Omega)$.
In all three cases the damping terms enable us to derive an $L_\infty$-estimate on $u$ (or $u_t$)
and thus avoid degeneracy of the coefficient $1-2ku$ (or $1-2ku_t$). Namely, for the damping term
of \eqref{Wpress_pLaplace} we have, using homogeneity of the Dirichlet boundary values of $u$,
\begin{equation}
\label{estLinftyW1}
\begin{aligned}
&|u(t,x)|\leq C_{W_0^{1,p+1},L_\infty}^\Omega
|\nabla u(t)|_{L_{p+1}}\\
&\qquad =C_{W_0^{1,p+1},L_\infty}^\Omega
\Bigl(|\nabla u_0|_{L_{p+1}}^{p+1}
+\int_0^t\frac{d}{dt}\int_\Omega|\nabla u(s,y)|^{p+1}\,dy\,ds
\Bigr)^{\frac{1}{p+1}}\\
&\qquad =C_{W_0^{1,p+1},L_\infty}^\Omega\Bigl[
|\nabla u_0|_{L_{p+1}}^{p+1}\\
&\qquad \quad +
\int_0^t\int_\Omega (p+1)|\nabla u(s,y)|^{p-1}\nabla u(s,y) \nabla u_t(s,y) \, dy\, ds \Bigr]^{\frac{1}{p+1}} \\
&\qquad =C_{W_0^{1,p+1},L_\infty}^\Omega\Bigl[
|\nabla u_0|_{L_{p+1}}^{p+1}\\
& \qquad \quad
-(p+1)\int_0^t\int_\Omega \text{div}\,\Bigl(|\nabla u(s,y)|^{p-1}\nabla u(s,y)\Bigr) u_t(s,y) \, dy\, ds \Bigr]^{\frac{1}{p+1}}, \\
\end{aligned}
\end{equation}
where $C_{W_0^{1,p+1},L_\infty}^\Omega$ denotes a combination of the constant in the Sobolev embedding
$W_0^{1,p+1}(\Omega)\to L_\infty(\Omega)$ with the one from the Poincar\'{e}-Friedrichs inequality.\\
For the damping term of \eqref{Wpress_viscosity}, we have
\begin{equation}
\label{estLinftyW2}
\begin{aligned}
&|u(t,x)|\leq C_{W_0^{1,q+1},L_\infty}^\Omega
|\nabla u(t)|_{L_{q+1}}\\
&\qquad=C_{W_0^{1,q+1},L_\infty}^\Omega
\Big|\nabla u_0+\int_0^t \nabla u_t(s)\,ds\Big|_{L_{q+1}(\Omega)} \\
&\qquad \leq C_{W_0^{1,q+1},L_\infty}^\Omega
\Bigl[|\nabla u_0|_{L_{q+1}}
+\Big|\int_0^t\nabla u_t(s)\,ds\Big|_{L_{q+1}}\Bigr]\\
&\qquad \leq C_{W_0^{1,q+1},L_\infty}^\Omega
\Bigl[|\nabla u_0|_{L_{q+1}}
+\Bigl(t^{q}\int_0^t\int_\Omega \Big|\nabla u_t(s,y)\Big|^{q+1} \, dy \,ds
\Bigr)^{\frac{1}{q+1}}\Bigr] \\
&\qquad =C_{W_0^{1,q+1},L_\infty}^\Omega
\Bigl[|\nabla u_0|_{L_{q+1}}\\
&\qquad \quad +\Bigl(-t^q\int_0^t\int_\Omega \text{div}\,\Bigl(|\nabla u_t(s,y)|^{q-1}\nabla u_t(s,y)\Bigr) u_t(s,y) \, dy\, ds \Bigr)^{\frac{1}{q+1}}\Bigr].
\end{aligned}
\end{equation}
Finally, for \eqref{Wpot_viscosity}, we get, replacing $u$ by $u_t$ and $p$ by $q$ in \eqref{estLinftyW1}
\begin{equation}
\label{estLinftyW3}
\begin{aligned}
&|u_t(t,x)|\leq
C_{W_0^{1,q+1},L_\infty}^\Omega\Bigl[
|\nabla u_1|_{L_{q+1}}^{q+1}\\
& \qquad -(q+1)\int_0^t\int_\Omega \text{div}\,\Bigl(|\nabla u_t(s,y)|^{q-1}\nabla u_t(s,y)\Bigr)
u_{tt}(s,y) \, dy\, ds\Bigr]^{\frac{1}{q+1}}. \\
\end{aligned}
\end{equation}
Hence, in all three cases, multiplication of the damping term with $u_t$ (or $u_{tt}$) and integration over time and space
will provide us with the desired $L_\infty$-estimate on $u$ (or $u_t$).
In this manner, we avoid estimates on $\Delta u$ in order to conclude $L_\infty$-boundedness of $u$,
a strategy that has been used in previous studies of the Westervelt equation~\cite{KL08,KLV10}.
A major advantage of this relaxed regularity is the fact that it can be expected to enable solutions
of the modified Westervelt equation to be coupled with other equations or with jumping coefficients,
a situation that is also of high practical relevance for HIFU devices based on the use of acoustic lenses
immersed in a fluid medium~\cite{manfred}.
We will first consider an acoustic-acoustic coupling. This can be modeled by the presence of spatially varying,
namely piecewise constant coefficients, in a pressure formulation (cf.~\cite{BambergerGlowinskiTran} for the linear case)
\begin{equation}\label{Wacoustic_viscosity}
\begin{cases}
\frac{1}{\lambda(x)}(1-2k(x)u)u_{tt}-\text{div}\,(\frac{1}{\varrho(x)}\nabla u)\\ -\text{div}\,\Bigl(b(x)((1-\delta(x)) +\delta(x)|\nabla u_t|^{q-1})\nabla u_t\Bigr)
=\frac{2k(x)}{\lambda(x)}(u_t)^2,
\\
(u,u_t)|_{t=0}=(u_0, u_1),
\\
u|_{\partial \Omega} =0
.\\
\end{cases}
\end{equation}
Here we have emphasized space dependence of the coefficients while suppressing space and
time dependence of $u$ in the notation as before.
While in the above equations \eqref{Wpress_pLaplace}, \eqref{Wpress_viscosity}, \eqref{Wpot_viscosity}
and in the regions of nonlinearity (i.e., $k\not=0$) in \eqref{Wacoustic_viscosity}, strong damping $b>0$ is needed for
ensuring well-posedness, we may set $b=0$ in regions where $k$ vanishes.
This corresponds to the physically relevant situation of a linearly acoustic (possibly to be considered as approximation
to linearly elastic) silicone lens immersed in a nonlinear acoustic fluid.
The physically more relevant model requires a linearly elastic model for the lens. Therefore, we consider a velocity based formulation for elastic-acoustic coupling (see also the displacement based formulation in~\cite{BermudezRodriguezSantamarina} and the velocity potential formulation in~\cite{FlemischKaltenbacherWohlmuth}, both for the linear case)
\begin{equation}\label{Welastic_viscosity}
\begin{cases}
\varrho(x) \mathbf{u}_{tt}-\mathcal{B}^T\frac{1}{1-2\tilde{k}(x)\psi_t} [c](x)\mathcal{B}\mathbf{u}\\
\qquad +\mathcal{B}^T\Bigl(((1-\delta(x)) +\delta(x)|\mathcal{B} \mathbf{u}_t|^{q-1})[b](x)\mathcal{B} \mathbf{u}_t\Bigr)
=0,
\\
(\mathbf{u},\mathbf{u}_t)|_{t=0}=(u_0, u_1),
\\
\mathbf{u}|_{\partial \Omega} =0
,\\
\end{cases}
\end{equation}
where $\mathbf{u}$ plays the role of the velocity, $\psi$ determines the gradient part in the Helmholtz decomposition of $\mathbf{u}$,
$$\mathbf{u}=\nabla\psi+\nabla\times\mathbf{A}\,,$$
and the first order differential operator $\mathcal{B}$ is given by
$$ \mathcal{B}=\left( \begin{array}{cccccc}
\partial_{x_1} & 0 & 0 & 0 & \partial_{x_3} & \partial_{x_2}\\
0 & \partial_{x_2} & 0 & \partial_{x_3} & 0 & \partial_{x_1}\\
0 & 0 & \partial_{x_3} & \partial_{x_2} & \partial_{x_1} & 0
\end{array} \right)^T
.$$
Note that $\psi$ can be determined from $\mathbf{u}$ as the solution of
\[
-\Delta \psi=-\text{div}\,\mathbf{u}
\]
which is unique, e.g., if we imposed homogeneous Dirichlet boundary conditions on $\psi$, which we will do below.
Here we think of $\Omega$ being decomposed into an acoustic (fluid) and an elastic (solid) subdomain
$$\Omega=\Omega_f\cup\Omega_s\,, \quad \Omega_f\cap\Omega_s=\emptyset\,.$$
No particular smoothness assumption will have to be imposed on the domains $\Omega_f$ and $\Omega_s$ as these subdomains are just characterized as the sets of points where the $L_\infty$-coefficients $\varrho$, $c$, $\tilde{k}$, and $b$ take on certain values that are typical for fluid and solid, respectively.
The acoustic region $\Omega_f$ is characterized as the region of vanishing shear modulus $\mu=0$ in the tensor
$$[c]=\left(\begin{array}{cccccc}
\lambda+2\mu&\lambda&\lambda&0&0&0\\
\lambda&\lambda+2\mu&\lambda&0&0&0\\
\lambda&\lambda&\lambda+2\mu&0&0&0\\
0&0&0&\mu&0&0\\
0&0&0&0&\mu&0\\
0&0&0&0&0&\mu\end{array}\right),
$$
i.e.,
$$\lambda>0\,, \quad
\mu\left\{\begin{array}{ll}
=0 &\mbox{in }\Omega_f\\
>0 &\mbox{in }\Omega_s\end{array}\right. .
$$
(Note that $[c]$ could be set to any symmetric positive definite $6\times6$ matrix with entries in $L^\infty(\Omega)$ in the elastic region $\Omega_s$, thus allowing for anisotropic elasticity.)
The tensor $[b](x)$ is assumed to be symmetric nonnegative definite and to have the same structure as $[c](x)$ in the fluid region, i.e.,
$$[b]=\left(\begin{array}{cccccc}
\hat{b}&\hat{b}&\hat{b}&0&0&0\\
\hat{b}&\hat{b}&\hat{b}&0&0&0\\
\hat{b}&\hat{b}&\hat{b}&0&0&0\\
0&0&0&0&0&0\\
0&0&0&0&0&0\\
0&0&0&0&0&0\end{array}\right) \mbox{ in }\Omega_f \mbox{ with } \hat{b}(x)\geq\underline{b}>0 \mbox{ in }\Omega_{nl}
\,.
$$
From the Westervelt equation on the subdomain of nonlinearity
$$\Omega_{nl}=\{x\in\Omega \colon \tilde{k}(x)\not=0\}\subseteq\Omega_f$$
we have that
the vector potential $\mathbf{A}$ may be set to zero in the acoustic region ($\psi$ equals the acoustic velocity potential there) and therewith
$$[c]\mathcal{B}\mathbf{u}=\lambda\Delta\psi \mathbf{e}\mbox{ in }\Omega_f, \mbox{ where }
\mathbf{e}=(1,1,1,0,0,0)^T,$$
so that on $\Omega_f$ the PDE in \eqref{Welastic_viscosity} becomes
$$
\varrho \nabla\psi_{tt}-\mathcal{B}^T\frac{\lambda}{1-2\tilde{k}\psi_t}\Delta\psi \mathbf{e}
-\mathcal{B}^T\Bigl(((1-\delta) +\delta|\Delta\psi_t|^{q-1})\hat{b}\Delta\psi_t\mathbf{e}\Bigr)
=0.
$$
Multiplying with an arbitrary vector valued test function $\mathbf{v}=\nabla w+\nabla\times W$ compactly supported in
$\Omega_f$, integrating by parts on $\Omega$, using the fact that
$$\mathbf{e}^T\mathcal{B} \mathbf{v} =\Delta w\,, \quad \text{div}\,\mathbf{v}=\Delta w$$
and assuming that $\varrho$ is constant on $\Omega_f$ such that
$\varrho \nabla\psi_{tt}= \nabla(\varrho\psi_{tt})$,
we arrive at
$$\int_{\Omega_f} \Delta w
\Bigl(\varrho \psi_{tt}-\frac{\lambda}{1-2\tilde{k}\psi_t}\Delta\psi
-((1-\delta) +\delta|\Delta\psi_t|^{q-1})\hat{b}\Delta\psi_t\Bigr)\, dx
=0
$$
for any smooth compactly supported $w$. With
$c^2=\frac{\lambda}{\varrho}$, $b=\frac{\hat{b}}{\varrho}$ we get \eqref{Wpot_viscosity}
which is (up to the damping term) equivalent to \eqref{Westervelt_psi}.
\medskip
In order to be able to make use of the embeddings
\begin{eqnarray}
H_0^1(\Omega)\to L_4(\Omega) \mbox{ with norm }C_{H_0^1,L_4}^\Omega\,,
\label{H01L4}\\
W_0^{1,q+1}(\Omega)\to L_\infty(\Omega) \mbox{ with norm }C_{W_0^{1,q+1},L_\infty}^\Omega\,,
\label{W01qp1linfty}
\end{eqnarray}
in this paper, we will impose zero Dirichlet boundary conditions and assume that $\Omega\subset\mathbb{R}^d$ is an open bounded set with Lipschitz boundary, $d\in\{1,2,3\}$.
Actually, \eqref{H01L4} allows to increase the space dimension even to $d=4$. However, this case is not of practical relevance.
The proofs below show that the embedding $H_0^1(\Omega)\to L_r(\Omega)$ with $r=4$
indeed suffices and we do not need to use the maximal possible exponent $r=6$ in $\mathbb{R}^3$.
Throughout the paper we will use Poincar{\'e}'s inequality; on convex domains the inequality reads (see for example page 364 in \cite{Leoni})
\[
\int_{\Omega} |u(x)-u_{\Omega}|^p dx \leq C(d,p) (\mathrm{diam}(\Omega))^p \int_{\Omega} |\nabla u(x)|^p dx.
\]
where $u_{\Omega}$ is the average of $u$ over the domain $\Omega$ and $\mathrm{diam}(\Omega)$ denotes the diameter of the domain $\Omega$. The Poincar{\'e} constant plays a role in the existence time for the solutions, however, our focus is on establishing local well-posedness of solutions and not on estimating the time of existence.
Several times we will make use of Young's inequality in the form
\begin{equation}\label{abeps}
a b\leq \epsilon a^r+ C(\epsilon,r)b^{\frac{r}{r-1}}
\end{equation}
with
\begin{equation}\label{Cepsr}
C(\epsilon,r)
=(r-1)r^{\frac{r}{r-1}} \epsilon^{-1/(1-r)} \,.
\end{equation}
\medskip
The remainder of this paper consists of five sections, each of them dealing with one of the equations \eqref{Wpress_pLaplace},
\eqref{Wpress_viscosity}, \eqref{Wpot_viscosity}, \eqref{Wacoustic_viscosity} and \eqref{Welastic_viscosity}, respectively.
We prove local in time existence (for sufficiently small times and initial data) for all of them. Global existence and exponential
decay can only be established for the $p$-Laplace damping case \eqref{Wpress_pLaplace}, which lacks uniqueness, though.
\section{The Westervelt equation in acoustic pressure formulation with nonlinear strong damping \eqref{Wpress_viscosity}}
\label{secWpress_viscosity}
First we consider the initial boundary value problem
\begin{equation}\label{W2lin}
\begin{cases}
(1+\alpha) u_{tt}-c^2\Delta u-b\,\text{div}\Bigl(
((1-\delta) +\delta|\nabla u_t|^{q-1})\nabla u_t\Bigr)+fu_t=g,
\\
(u,u_t)|_{t=0}=(u_0, u_1),
\\
u|_{\partial \Omega} =0
.\\
\end{cases}
\end{equation}
\begin{proposition} \label{prop:W2lin}
\begin{enumerate}
\item[(i)]
Let $T>0$, $c^2,b,\delta,1-\delta>0$, $q\geq1$ and assume that
\begin{itemize}
\item
$\alpha\in C(0,T;L_\infty(\Omega))$, $\alpha_t\in L_\infty(0,T;L_2(\Omega))$,
$-1<-\underline{\alpha}\leq \alpha(t,x)\leq \overline{\alpha}$,
\item
$f\in L_\infty(0,T;L_{2}(\Omega))$,
\item
$g\in (L_{q+1}(0,T;W_0^{1,q+1}(\Omega)))^*$,
\item
$u_0\in H_0^1(\Omega)$, $u_1\in L_2(\Omega)$.
\end{itemize}
with
$$\|f-\frac12\alpha_t\|_{L_\infty(0,T;L_{2}(\Omega))}\leq\overline{b}
<\frac{b(1-\delta)}{(C_{H_0^1,L_4}^\Omega)^2}\,. $$
Then \eqref{W2lin} has a weak solution
\begin{align*}
u\in \tilde{X} :=~&C^1(0,T;L_2(\Omega))\cap C(0,T;H_0^1(\Omega))\\
&\cap W^{1,q+1}(0,T;W_0^{1,q+1}(\Omega)).
\end{align*}
Moreover, this solution is unique in $\tilde{X}$ and it satisfies the energy estimate
\begin{eqnarray}
\lefteqn{\frac12\left[\int_\Omega (1+\alpha)(u_t)^2\, dx
+ c^2 |\nabla u|_{L_2(\Omega)}^2 \right]_0^t}
\nonumber\\
&&+ \int_0^t \Bigl( \hat{b}|\nabla u_t|_{L_2(\Omega)}^2
+\tfrac{b\delta}{2}
|\nabla u_t|_{L_{q+1}(\Omega)}^{q+1}\Bigr)\, ds
\nonumber\\
& \leq &
C(\tfrac{b\delta}{2},q+1) \|g\|_{(L_{q+1}(0,T;W_0^{1,q+1}(\Omega)))^*}^{\frac{q+1}{q}}\,.
\label{enest}
\end{eqnarray}
with
\[
\hat{b}=b(1-\delta)- (C_{H_0^1,L_4}^\Omega)^2 \overline{b}
\]
and $C(\epsilon,r)$ as in \eqref{Cepsr}.
\item[(ii)]
If in addition to (i) $\alpha$ is independent of time, we have the estimate
\begin{equation}
\label{utt}
\begin{aligned}
&\|(1+\alpha)u_{tt}\|_{(L_{q+1}(0,T;W_0^{1,q+1}(\Omega)))^*}\\
&\leq\overline{C} \left\{ \|u\|_{\tilde{X}}
+ \|g\|_{(L_{q+1}(0,T;W_0^{1,q+1}(\Omega)))^*}
+ \|u\|_{\tilde{X}} \|f\|_{L_\infty(0,T;L_2(\Omega))} \right\}.
\end{aligned}
\end{equation}
for some constant $\overline{C}>0$.
\item[(iii)]
If in addition to (i)
\begin{itemize}
\item $f\in L_\infty(0,T;L_4(\Omega))$
with $\|f\|_{L_\infty(0,T;L_4(\Omega))}\leq\overline{\overline{b}}$,
\item $g\in L_2(0,T;L_2(\Omega))$,
\item $u_1\in H_0^1(\Omega)\cap W_0^{1,q+1}(\Omega)$,
\end{itemize} then
\begin{align*}
u\in X:=~& H^2(0,T;L_2(\Omega))\cap C^1(0,T;W_0^{1,q+1}(\Omega)) \,,
\end{align*}
and there exist constants $c_1,C_1$ depending only on $\underline{\alpha}$, $b$, $c$, $\delta$, $C_{H_0^1,L_4}^\Omega$, $\overline{\overline{b}}$,
such that
\begin{eqnarray}
E_1[u](t)
+c_1\int_0^t \Bigl(
|u_{tt}|_{L_2(\Omega)}^2
+|\nabla u_t|_{L_2(\Omega)}^2
+|\nabla u_t|_{L_{q+1}(\Omega)}^{q+1}
\Bigr)\, ds
&&\nonumber\\
\leq C_1 (E_1[u](0)+\|g\|_{L_2(0,T;L_2(\Omega))}
+\|f\|_{L_\infty(0,T;L_4(\Omega))}^2
)
&&
\label{enest1}
\end{eqnarray}
for
\begin{equation}\label{defE1}
E_1[u](t)=\left[|u_t|_{L_2(\Omega)}^2
+ |\nabla u|_{L_2(\Omega)}^2
+ |\nabla u_t|_{L_2(\Omega)}^2
+ |\nabla u_t|_{L_{q+1}(\Omega)}^{q+1}
\right](t).
\end{equation}
\end{enumerate}
\end{proposition}
\begin{remark}
Some simplified versions of Proposition \ref{prop:W2lin} appear in the literature; usually,
$\alpha$ is taken to be zero (i.e. constant coefficient for $u_{tt}$), $\delta= 1$.
This version appears in \cite{Biazutti, Clements} where the damping appears under a regular ($p=2$) Laplacian,
and the authors allow a more general divergence operator, instead of $\Delta u$,
to prove global existence of weak solutions.
Nonlinear $p$-Laplace damping was in considered in \cite{RW}, where a nonlinear source term appears too.
However, other features of the equation (such as variable $\alpha$) were not present.
For the sake of self-completeness we include below the proof of Proposition \ref{prop:W2lin}.
\end{remark}
\begin{proof}
The weak form of \eqref{W2lin} reads as
\begin{equation}
\label{W2linweak}
\begin{aligned}
\int_\Omega \Bigl\{(1+\alpha)u_{tt} w + c^2 \nabla u \nabla w
+ b\Bigl((1-\delta) +\delta |\nabla u_t|^{q-1}
\Bigr)\nabla u_t\nabla w \Bigr\} \, dx&\\
=\int_\Omega (g -fu_t) w\, dx, \quad \forall w\in W_0^{1,q+1}(\Omega),&
\end{aligned}
\end{equation}
with initial conditions $(u_{0},u_{1})$.\\
{\it Step 1: Smooth approximation of $\alpha$, $f$, and $g$}:
We consider sequences \\
$(\alpha_k)_{k\in\mathbb{N}}$, $(f_k)_{k\in\mathbb{N}}$ and $(g_k)_{k\in\mathbb{N}}$ such that
\begin{itemize}
\item
$(\alpha_k)_{k\in\mathbb{N}}\subseteq C^\infty([0,T]\times\overline{\Omega})\cap W^{1,\infty}(0,T;L_2(\Omega))$,\\
$\alpha_k\to \alpha$ in $C(0,T;L_\infty(\Omega))\cap W^{1,\infty}(0,T;L_2(\Omega))$,
$-1<-\underline{\alpha}\leq \alpha_k(t,x)\leq \overline{\alpha}$,
\item
$(f_k)_{k\in\mathbb{N}}\subseteq C^\infty((0,T)\times\Omega)$,
$f_k\to f$ in $L_\infty(0,T;L_{2}(\Omega))$,
\item
$(g_k)_{k\in\mathbb{N}}\subseteq C^\infty((0,T)\times\Omega)$,
$g_k\to g$ in
$(L_{q+1}(0,T;W_0^{1,q+1}(\Omega)))^*$,
\item
$\|f_k-\frac12\alpha_{k,t}\|_{L_\infty(0,T;L_{2}(\Omega))}\leq\overline{b}$,
\end{itemize}
and, for fixed $k\in\mathbb{N}$, prove that there exists a solution $u^{(k)}$ of \eqref{W2linweak} with $\alpha$, $f$ and $g$ replaced by $\alpha_k$, $f_k$ and $g_k$, respectively, i.e.,
\begin{equation}
\label{W2linweak_k}
\begin{aligned}
\int_\Omega \Bigl\{(1+\alpha_k)u^{(k)}_{tt} w + c^2 \nabla u^{(k)} \nabla w
+ b\Bigl((1-\delta) +\delta |\nabla u^{(k)}_t|^{q-1}
\Bigr)\nabla u^{(k)}_t\nabla w \Bigr\} \, dx&\\
=\int_\Omega (g_k -f_ku^{(k)}_t) w\, dx\quad \forall w\in W_0^{1,q+1}(\Omega),&
\end{aligned}
\end{equation}
with initial conditions $(u_{0},u_{1})$.
Later we will consider limits as $k\to\infty$ to prove well-posedness of \eqref{W2linweak}.
The existence proof follows the line of the standard approach for linear parabolic or second order hyperbolic PDEs as it can be found, e.g., in \cite{Evans}. The proof is divided into three subparts: (a) Galerkin approximation, (b) energy estimates and (c) weak limit.
\\[1ex]
{\it Step 1 (a): Galerkin approximation.} We will first show existence and uniqueness of solutions for a finite-dimensional approximation of \eqref{W2linweak_k}.\\
Assume $w_m=w_m(x)$, $m\in\mathbb{N}$ are smooth functions such that
\begin{align*}
&\{w_m\}_{m\in\mathbb{N}}\text{ is an orthonormal basis of }L_2^{\tilde{\alpha}_k}(\Omega),\\
&\{w_m\}_{m\in\mathbb{N}}\text{ is a basis of } W_0^{1,q+1}(\Omega),
\end{align*}
where $L_2^{\tilde{\alpha}_k}$ is the weighted $L_2$-space based on the inner product
\begin{equation*}
\langle f,g\rangle_{L_2^{\tilde{\alpha}_k}(\Omega)} := \int_\Omega (1+\tilde{\alpha}_k) fg \, dx\,,
\end{equation*}
with
\[
\tilde{\alpha}_k=\frac{1}{T}\int_0^T\alpha(t)\, dt\,.
\]
Moreover, let $V_n$ be the finite dimensional subspace of $L_2^{\alpha_k} (\Omega) \cap W_0^{1,q+1}(\Omega)$ spanned by $\{w_m\}_{m=1}^n$.
Thus $(V_n)_{n\in\mathbb{N}}$ is a nested sequence ($V_k \subseteq V_l$ for $k \leq l$) of finite dimensional subspaces
$V_n\subseteq L_2^{\alpha_k} (\Omega) \cap W_0^{1,q+1}(\Omega)$ such that $\bigcup_{n\in\mathbb{N}} V_n=W_0^{1,q+1}(\Omega)$.\\
Furthermore, suppose $(u_{0,n})_{n\in\mathbb{N}}$, $(u_{1,n})_{n\in\mathbb{N}}$ are sequences satisfying
\begin{itemize}
\item
$u_{0,n}\in V_n$,
$u_{0,n}\to u_0$ in $H_0^1(\Omega)$,
\item
$u_{1,n}\in V_n$,
$u_{1,n}\to u_1$ in $L_2(\Omega)$.
\end{itemize}
Based on this, we consider a sequence of discretized versions of \eqref{W2linweak_k},
\begin{equation}
\label{W2linweakdis}
\begin{aligned}
&\int_\Omega \Bigl\{(1+\alpha_k)u_{n,tt}^{(k)} w_n + c^2 \nabla u_n^{(k)} \nabla w_n \\
&\qquad + b\Bigl((1-\delta) +\delta|\nabla u_{n,t}^{(k)}|^{q-1}\Bigr)\nabla u_{n,t}^{(k)} \nabla w_n \Bigr\} \, dx\\
&=\int_\Omega (g_k -f_k u_{n,t}^{(k)})w_n \, dx\quad \forall w_n\in V_n,
\end{aligned}
\end{equation}
with $u_n^{(k)}(t)\in V_n$ and initial conditions $(u_{0,n},u_{1,n})$.\\
For each $n\in\mathbb{N}$ the equality in (\ref{W2linweakdis}) together with initial conditions $(u_{0,n},u_{1,n})$ gives an initial value problem for a second order
system of ordinary differential equations which has smooth (with respect to time) coefficients and right-hand side.\\
The standard existence theory for ordinary differential equations (cf. \cite{Teschl}) provides us with a unique solution
$u_n^{(k)}\in C^\infty(0,\tilde{T};V_n)$ of the finite-dimensional approximation \eqref{W2linweakdis} of \eqref{W2linweak} for some $\tilde{T}\leq T$ sufficiently small.
By the uniform energy and norm estimates below,
we obtain $\tilde{T}=T$, i.e. there is a unique solution $u_n^{(k)}\in C^1(0,T;V_n)$ of \eqref{W2linweakdis}.
{\it Step 1 (b): Energy estimate.} Testing \eqref{W2linweakdis} with $w_n=u_{n,t}^{(k)}(t)$, integrating with respect to time and using the identity
$$\frac{d}{dt}\alpha_k \left(u_{n,t}^{(k)}\right)^2 = 2\alpha_k u_{n,t}^{(k)} u_{n,tt}^{(k)}+\alpha_{k,t} \left(u_{n,t}^{(k)}\right)^2$$
as well as Young's inequality in the form \eqref{abeps} with \eqref{Cepsr},
we obtain
\begin{align*}
&\frac12\left[\int_\Omega (1+\alpha_k)\left(u_{n,t}^{(k)}\right)^2\, dx
+ c^2 |\nabla u_n^{(k)}|_{L_2(\Omega)}^2 \right]_0^t\\
&\quad+ b \int_0^t \int_\Omega \Bigl((1-\delta) +\delta|\nabla u_{n,t}^{(k)}|^{q-1}\Bigr)|\nabla u_{n,t}^{(k)}|^2 \, dx \, ds\\
&=
- \int_0^t \int_\Omega (f_k-\frac12\alpha_{k,t}) \left(u_{n,t}^{(k)}\right)^2 \, dx \, ds
+ \int_0^t \int_\Omega g_k u_{n,t}^{(k)} \, dx \, ds\\
&\leq
\overline{b}\int_0^t | u_{n,t}^{(k)}|_{L_4(\Omega)}^2 \, ds
+\int_0^t \Bigl(\tfrac{b\delta}{2} |u_{n,t}^{(k)}|_{W_0^{1,q+1}(\Omega)}^{q+1} \, dx
+C(\tfrac{b\delta}{2},q+1) |g_k|_{(W_0^{1,q+1}(\Omega))^*}^{\frac{q+1}{q}}\Bigr)\, ds\,,
\end{align*}
hence
\begin{equation}\label{enest_kn}
\begin{aligned}
&\frac12\left[\int_\Omega (1+\alpha_k)\left(u_{n,t}^{(k)}\right)^2\, dx
+ c^2 |\nabla u_{n}^{(k)}|_{L_2(\Omega)}^2 \right]_0^t\\
&+ \int_0^t \Bigl( \hat{b}|\nabla u_{n,t}^{(k)}|_{L_2(\Omega)}^2
+b\delta|\nabla u_{n,t}^{(k)}|_{L_{q+1}(\Omega)}^{q+1}\Bigr)\, ds\\
& \leq
C(\tfrac{b\delta}{2},q+1) \|g_k\|_{(L_{q+1}(0,T;W_0^{1,q+1}(\Omega)))^*}^{\frac{q+1}{q}}\, ,
\end{aligned}
\end{equation}
which corresponds to the energy estimate \eqref{enest} upon replacement of $\alpha,f,g$ by $\alpha_k,f_k,g_k$.
As, by assumption,
$g_k \in (L_{q+1}(0,T;W_0^{1,q+1}(\Omega)))^*$,
we have that $\big(u_n^{(k)}\big)_{n\in\mathbb{N}}$ is a bounded sequence in the Banach space
\begin{equation}\label{space:un}
\begin{aligned}
\tilde{X}:=~&C^1(0,T;L_2(\Omega)) \cap C(0,T;H_0^1(\Omega)) \cap W^{1,q+1}(0,T; W_0^{1,q+1}(\Omega)).
\end{aligned}
\end{equation}
Hence
\begin{equation}\label{Xhat}
\begin{aligned}
\big(u_n^{(k)}\big)_{n\in\mathbb{N}}\text{ is bounded in }W^{1,q+1}(0,T; W_0^{1,q+1}(\Omega)),
\end{aligned}
\end{equation}
which is a reflexive Banach space.
Furthermore, we obtain that
\begin{equation}\label{Lqdual}
|\nabla u_{n,t}^{(k)}|^{q-1} \nabla u_{n,t}^{(k)} \text{ is uniformly bounded in } L_{\frac{q+1}{q}}(0,T;L_{\frac{q+1}{q}}(\Omega)).
\end{equation}
{\it Step 1 (b) --- part (ii): Estimate of $u^{(k)}_{n,tt}$}.
To show (ii) in case $\alpha_{k,t}=0$ and therewith $\tilde{\alpha}_k=\alpha_k$, we now prove an analog of \eqref{utt} with $\alpha_k,f_k,g_k$ in place of $\alpha,f,g$. For this purpose, let $v\in W_0^{1,q+1}(\Omega)$. We decompose $v$,
\begin{equation*}
v= v_n+z_n \text{ with } v_n \in V_n \text{ and } z_n \in V_n^{\bot_{L_2^{\alpha_k}}}.
\end{equation*}
Here
\begin{equation*}
|v_n|_{W_0^{1,q+1}(\Omega)}\leq |v|_{W_0^{1,q+1}(\Omega)}
\end{equation*}
holds due to the fact that with a linearly independent set of dual vectors $\{w_m^*\}_{m\in\mathbb{N}} \subseteq (W_0^{1,q+1}(\Omega))^*$ of $V_n^*$, i.e., such that $\langle w_i^*,w_j\rangle_{W_0^{1,q+1}(\Omega)^*,W_0^{1,q+1}(\Omega)}=\delta_{i,j}$ (which exists by the Hahn-Banach Theorem) we have
\begin{align*}
|v_n|_{W_0^{1,q+1}(\Omega)}
=&\sup_{v^*\in V_n^*\,, \ |v^*|_{W_0^{1,q+1}(\Omega)^*}=1}
\langle v^*,v_n\rangle_{W_0^{1,q+1}(\Omega)^*,W_0^{1,q+1}(\Omega)}\\
=&\sup_{v^*\in \text{span}(w_1^*,\ldots,w_n^*)\,, \ |v^*|_{W_0^{1,q+1}(\Omega)^*}=1}
\langle v^*,v_n\rangle_{W_0^{1,q+1}(\Omega)^*,W_0^{1,q+1}(\Omega)}\\
=&\sup_{v^*\in \text{span}(w_1^*,\ldots,w_n^*)\,, \ |v^*|_{W_0^{1,q+1}(\Omega)^*}=1}
\langle v^*,v\rangle_{W_0^{1,q+1}(\Omega)^*,W_0^{1,q+1}(\Omega)}\\
\leq&\sup_{v^*\in W_0^{1,q+1}(\Omega)^*\,, \ |v^*|_{W_0^{1,q+1}(\Omega)^*}=1}
\langle v^*,v\rangle_{W_0^{1,q+1}(\Omega)^*,W_0^{1,q+1}(\Omega)}\\
=& |v|_{W_0^{1,q+1}(\Omega)}\,.
\end{align*}
By using orthogonality we obtain the first equality below; by \eqref{W2linweakdis} we obtain the second equality); finally, by using $|.|_{L^2(\Omega)} \leq {C_{L_{q+1},L_2}^\Omega} |.|_{L^{q+1}(\Omega)}$ we have
\begin{align*}
&\int_\Omega (1+\alpha_k)u_{n,tt}^{(k)} v \, dx = \int_\Omega (1+\alpha_k)u_{n,tt}^{(k)} v_n \, dx \\
&=-c^2 \int_\Omega \nabla u_n^{(k)} \nabla v_n \, dx - b(1-\delta) \int_\Omega \nabla u_{n,t}^{(k)} \nabla v_n \, dx\\
&\quad- b\delta \int_\Omega |\nabla u_{n,t}^{(k)}|^{q-1} \nabla u_{n,t}^{(k)} \nabla v_n \, dx + \int_\Omega (g_k - f_k u_{n,t}^{(k)}) v_n \, dx\\
&\leq \left\{c^2 |\nabla u_n^{(k)}|_{L_2(\Omega)} + b(1-\delta) |\nabla u_{n,t}^{(k)}| \right\} {C_{L_{q+1},L_2}^\Omega} | \nabla v_n |_{L_{q+1}(\Omega)} \\
&\quad + \left\{b\delta |\nabla u_{n,t}^{(k)} |_{L_{q+1}(\Omega)}^q + |g_k - f_k u_{n,t}^{(k)}|_{(W_0^{1,q+1}(\Omega))^*}\right\} |\nabla v_n|_{L_{q+1}(\Omega)}\\
&= \left\{c^2 {C_{L_{q+1},L_2}^\Omega} |u_n^{(k)}|_{H^1(\Omega)} + b(1-\delta) {C_{L_{q+1},L_2}^\Omega} |u_{n,t}^{(k)} |_{H^1(\Omega)} \right. \\
&\quad + \left. b\delta |u_{n,t}^{(k)} |_{W_0^{1,q+1}(\Omega)}^q + |g_k - f_k u_{n,t}^{(k)}|_{(W_0^{1,q+1}(\Omega))^*}\right\} \underbrace{|v_n|_{W_0^{1,q+1}(\Omega)}}_{\leq |v|_{W_0^{1,q+1}(\Omega)}}.
\end{align*}
Multiplying the resulting inequality with a test function $\phi \in C^{\infty}(0,T)$ and integrating with respect to time yields
\begin{align*}
&\int_0^t \int_\Omega (1+\alpha_k)u_{n,tt}^{(k)} v \, dx\, \phi \, ds\\
& \leq \int_0^t \left\{c^2 {C_{L_{q+1},L_2}^\Omega} |u_n^{(k)}|_{H^1(\Omega)} + b(1-\delta) {C_{L_{q+1},L_2}^\Omega} |u_{n,t}^{(k)} |_{H^1(\Omega)} \right. \\
&\quad + \left. b\delta |u_{n,t}^{(k)} |_{W_0^{1,q+1}(\Omega)}^q + |g_k - f_k u_{n,t}^{(k)}|_{(W_0^{1,q+1}(\Omega))^*}\right\}
|v|_{W_0^{1,q+1}(\Omega)} |\phi| \, ds
\end{align*}
\begin{align*}
&\leq \|v \phi\|_{L_{q+1}(0,T;W_0^{1,q+1}(\Omega))} \left\{ c^2 {C_{L_{q+1},L_2}^\Omega} \|u_n^{(k)}\|_{L_{\frac{q+1}{q}}(0,T;H^1(\Omega))} \right.\\
&\quad+b(1-\delta) {C_{L_{q+1},L_2}^\Omega} \|u_n^{(k)}\|_{L_{\frac{q+1}{q}}(0,T;H^1(\Omega))} \\
&\quad \left. +b\delta \|u_n^{(k)} \|_{L_{\frac{q+1}{q}}(0,T;W_0^{1,q+1}(\Omega))}+\|g_k-f_k u_{n,t}^{(k)}\|_{(L_{q+1}(0,T;W_0^{1,q+1}(\Omega)))^*)} \right\},
\end{align*}
which shows that
\begin{equation}
\label{unttk}
\begin{aligned}
&\|(1+\alpha_k)u_{n,tt}^{(k)}\|_{(L_{q+1}(0,T; W_0^{1,q+1}(\Omega)))^*}\\
&\leq\overline{C} \left\{ \|u_n^{(k)}\|_{\tilde{X}}
+ \|g_k\|_{(L_{q+1}(0,T; (W_0^{1,q+1}(\Omega)))^*}
+ \|u_n^{(k)}\|_{\tilde{X}} \|f_k\|_{L_\infty(0,T;L_2(\Omega))} \right\}
\end{aligned}
\end{equation}
for some constant $\overline{C}>0$. Hence
$$(1+\alpha_k)u_{n,tt}^{(k)} \in (L_{q+1}(0,T; (W_0^{1,q+1}(\Omega)))^*$$
with a uniform bound with respect to $n$.
\\[1ex]
{\it Step 1 (c): Weak limit.}
As a consequence of \eqref{Xhat}, \eqref{Lqdual},
there exists a weakly convergent subsequence, which for simplicity we denote by $u_n^{(k)}$ again, and a sequence $u^{(k)} \in W^{1,q+1}(0,T; W_0^{1,q+1}(\Omega))$ such that
\begin{align}\label{weakconv1}
&u_n^{(k)}\rightharpoonup u^{(k)} \mbox{ in }W^{1,q+1}(0,T; W_0^{1,q+1}(\Omega)), \\
& |\nabla u_{n,t}^{(k)}|^{q-1} \nabla u_{n,t}^{(k)}
\rightharpoonup |\nabla u_t^{(k)}|^{q-1} \nabla u_t^{(k)}
\mbox{ in }L_{\frac{q+1}{q}} (0,T;L_{\frac{q+1}{q}}(\Omega)).\label{weakconv3}
\end{align}
This weak limit satisfies the estimates \eqref{enest_kn} and \eqref{unttk} with $u_n^{(k)}$ replaced by $u^{(k)}$.\\
For fixed $k,m\in\mathbb{N}$ and $\phi_m\in C^\infty(0,T,V_m)\subset L_{q+1}(0,T;W_0^{1,q+1}(\Omega))$ with $\phi_m(T)=0$ we have for any $n\geq m$ by $V_m\subseteq V_n$ that
\begin{equation}
\label{convntoinfty_W2}
\begin{aligned}
&\int_0^T \int_\Omega \Bigl\{ (1+\alpha_k)u_{tt}^{(k)} \phi_m
+ c^2 \nabla u^{(k)} \nabla \phi_m\\
&\quad+ b\Bigl((1-\delta) +\delta|\nabla u_t^{(k)}|^{q-1}\Bigr)\nabla u_t^{(k)})\nabla \phi_m
+ f_k u_t^{(k)} \phi_m - g_k \phi_m\Bigr\} \, dx\, ds\\
&=-\int_0^T\int_\Omega [u_{t}^{(k)}-u_{n,t}^{(k)}] \Bigl((1+\alpha_k) \phi_m\Bigr)_t \, dx \, ds\\
&\quad-\int_\Omega [u_1-u_{1,n}] (1+\alpha_k(0)) \phi_m(0) \, dx \, ds + \int_0^T\int_\Omega [u_t^{(k)} -u_{n,t}^{(k)}]f_k \phi_m \, dx\, ds\\
&\quad+ \int_0^T\int_\Omega \left(c^2[\nabla u^{(k)}-\nabla u_n^{(k)}] + b (1-\delta) [\nabla u_t^{(k)}-\nabla u_{n,t}^{(k)}]\right)\nabla \phi_m \, dx\, ds \\
&\quad + b \int_0^T\int_\Omega \delta[|\nabla u_t^{(k)}|^{q-1}\nabla u_t^{(k)}-|\nabla u_{n,t}^{(k)}|^{q-1}\nabla u_{n,t}^{(k)}]\nabla \phi_m\Bigr)
\Bigr\} \, dx\, ds \to 0 \\
\end{aligned}
\end{equation}
$\mbox{ as }n\to \infty$, due to \eqref{weakconv1}
and \eqref{weakconv3}.
Since $\bigcup_{m\in\mathbb{N}} V_m=W_0^{1,q+1}(\Omega)$, the relation \eqref{convntoinfty_W2} proves that $u^{(k)}$ solves \eqref{W2linweak_k}. Moreover $u^{(k)}$ satisfies the energy estimates \eqref{enest} with $\alpha$, $f$, $g$ and $u$ replaced by $\alpha_k$, $f_k$, $g_k$ and $u^{(k)}$, respectively. In case (ii) $\alpha_{k,t}=0$, so additionally, we have that \eqref{utt}, with $\alpha$, $f$, $g$ and $u$ replaced by $\alpha_k$, $f_k$, $g_k$ and $u^{(k)}$, respectively, holds.
\\[1ex]
{\it Step 2: $k\to \infty$.} For all $w\in C^\infty(0,T;W_0^{1,q+1}(\Omega))$ with $w(T)=0$ we have
\begin{align*}\nonumber
&\int_0^t \int_\Omega \Bigl\{ (1+\alpha)u_{tt} w + c^2 \nabla u \nabla w\\
&\quad + b\Bigl((1-\delta) +\delta|\nabla u_t|^{q-1}\Bigr)\nabla u_t\nabla w +f u_t w - gw \Bigr\} \, dx\, ds
\nonumber\\
&=-\int_0^t \int_\Omega [u_{t}-u_{t}^{(k)}] \Bigl((1+\alpha)w\Bigr)_t \, dx \, ds
- \int_0^t \int_\Omega u_{t}^{(k)} \Bigl([\alpha-\alpha_k]w\Bigr)_t \, dx \, ds\\
&\quad- \int_\Omega u_{1} \Bigl([\alpha(0)-\alpha_k(0)]w(0)\Bigr) \, dx \nonumber\\
&\quad+ \int_0^t \int_\Omega c^2 [\nabla u - \nabla u^{(k)})] \nabla w \, dx \, ds + \int_0^t \int_\Omega b(1-\delta) [\nabla u_t - \nabla u_t^{(k)}] \nabla w \, dx \, ds \nonumber\\
&\quad + \int_0^t \int_\Omega b\delta [|\nabla u_t|^{q-1}\nabla u_t - |\nabla u_t^{(k)}|^{q-1}\nabla u_t^{(k)}]\nabla w \, dx \, ds \nonumber\\
&\quad + \int_0^t \int_\Omega [u_t - u_t^{(k)}] f w \, dx \, ds + \int_0^t \int_\Omega [f-f_k] u_t^{(k)} w \, dx\, ds\nonumber\\
&\quad - \int_0^t \int_\Omega [g - g_k] w \, dx \, ds \to 0 \text{ as } k \to \infty, \nonumber
\end{align*}
since we imposed $\alpha_k \to \alpha$ in $C(0,T;L_\infty(\Omega))\cap W^{1,\infty}(0,T;L_2(\Omega))$, $f_k \to f$ in $L_\infty(0,T;L_2(\Omega))$, $g_k \to g$ in
$(L_{q+1}(0,T;W_0^{1,q+1}(\Omega)))^*$.
Due to the fact that the estimate \eqref{enest} remains valid for $u^{(k)}$, analogously to Step 1(c), we find a convergent subsequence (which
we relabel) and a function
$u\in W^{1,q+1}(0,T; W_0^{1,q+1}(\Omega))$
such that
\begin{align}\label{weakconv4}
&u^{(k)}\rightharpoonup u \mbox{ in }
W^{1,q+1}(0,T; W_0^{1,q+1}(\Omega)),
\\
& |\nabla u_{t}^{(k)}|^{q-1} \nabla u_{t}^{(k)}
\rightharpoonup |\nabla u_t|^{q-1} \nabla u_t
\mbox{ in }L_{\frac{q+1}{q}} (0,T;L_{\frac{q+1}{q}}(\Omega)).\label{weakconv6}
\end{align}
This weak limit is, by construction, a weak solution of \eqref{W2lin} and it satisfies the energy estimate
\eqref{enest}; in case (ii) $\alpha_t=0$, it also satisfies \eqref{utt}.
{\it Step 3: Uniqueness of weak solutions.} Uniqueness of weak solutions follows from the fact that the difference $\hat{u}=u^1-u^2$ between any
two weak solutions $u^1,u^2$ of \eqref{W2lin} satisfies
\begin{equation}\label{W2_uniq}
\begin{cases}
(1+\alpha)\hat{u}_{tt}-c^2\Delta \hat{u}-b(1-\delta)\Delta \hat{u}_t
-b\delta \int_0^1\,\text{div}\Bigl( w^\sigma
\Bigl[|\nabla (u^2+\sigma \hat{u})_t|^2 \nabla \hat{u}_t\\
\quad+(q-1)(\nabla (u^2+\sigma \hat{u})_t \nabla \hat{u}_t)
\nabla(u^2+\sigma \hat{u})_t \Bigr]\, \Bigr)d\sigma
+f\hat{u}_t\ = \ 0
\\
(\hat{u},\hat{u}_t)|_{t=0}=(0,0)
\\
\hat{u}|_{\partial \Omega} =0
,\\
\end{cases}
\end{equation}
in a weak sense, with
\begin{equation}\label{wsig}
w^\sigma(x,t)= |\nabla (u^2+\sigma \hat{u})_t(x,t)|^{q-3}\,.
\end{equation}
Above we used the fact that for
\begin{equation*}
F(\lambda)=|\lambda|^{q-1}\lambda, \quad F'(\lambda)=|\lambda|^{q-1}I + (q-1)|\lambda|^{q-3}\lambda\lambda^T
\end{equation*}
we have
\begin{equation*}
F(\nabla u^1_t(x,t))-F(\nabla u^2_t(x,t))=
\int_0^1 F'(\nabla (u^2+\sigma\hat{u})_t)(x,t))\, d\sigma\nabla\hat{u}_t(x,t).
\end{equation*}
Multiplication of \eqref{W2_uniq} by $\hat{u}_t$ yields
\[
\begin{aligned}
&\frac12\left[\int_\Omega (1+\alpha)(\hat{u}_t)^2\, dx
+ c^2 |\nabla \hat{u}|_{L_2(\Omega)}^2 \right]_0^t
+ \int_0^t b(1-\delta) |\nabla \hat{u}_t(s)|^2 \, ds\\
&\hfill+ \int_0^t \int_\Omega (f-\frac12\alpha_t) (\hat{u}_t)^2 \, dx \, ds
\ \leq \ 0
\end{aligned}
\]
since
\begin{equation}\label{wsig_nonneg}
\begin{aligned}
&b\delta\int_0^1\int_0^t\int_\Omega w^\sigma\Bigl(
|\nabla (u^2+\sigma \hat{u})_t|^2 |\nabla \hat{u}_t|^2\\
&\qquad+(q-1)(\nabla (u^2+\sigma \hat{u})_t \cdot\nabla \hat{u}_t)^2
\Bigr) \, dx\, ds\, d\sigma \geq 0.
\end{aligned}
\end{equation}
Therefore $\hat{u}=0$ almost everywhere and the proof of uniqueness is complete.
\\[1ex]
{\it Step 4: Higher energy estimate.} For the proof of (iii), we need a higher energy estimate which can be obtained by testing
\eqref{W2linweak} with $w=u_{tt}(t)$ (strictly speaking, we multiply by a smooth approximation of $u_{tt}$ and take weak limits); then integration with respect to time yields
\begin{equation}
\label{enest01A}\begin{aligned}
&\int_0^t\int_\Omega (1+\alpha)(u_{tt})^2\, dx\, ds
+\left[\frac{b(1-\delta)}{2}|\nabla u_{t}|_{L_2(\Omega)}^2
+\frac{b\delta}{q+1}|\nabla u_{t}|_{L_{q+1}(\Omega)}^{q+1}\right]_0^t\\
&=
c^2\int_0^t |\nabla u_{t}|_{L_2(\Omega)}^2\, ds
-c^2\left[\int_\Omega \nabla u \nabla u_{t}\right]_0^t
-\int_0^t\int_\Omega (f u_t-g)u_{tt}\, dx\, ds\\
& \leq
c^2\int_0^t |\nabla u_{t}|_{L_2(\Omega)}^2\, ds
+\breve{b}\Bigl(|\nabla u_{t}(t)|_{L_2}^2 + |\nabla u_{t}(0)|_{L_2}^2\Bigr)\\
&\quad + \frac{c^4}{4\breve{b}}
\Bigl(|\nabla u(t)|_{L_2}^2 + |\nabla u(0)|_{L_2}^2\Bigr)
+ (\tau+\sigma) \int_0^t|u_{tt}|_{L_2}^2\, ds\\
& \quad+ \frac{\left(C_{H_0^1,L_4}^\Omega\right)^2}{4\tau}
\|f\|_{L_\infty(0,T;L_4(\Omega))}^2
\int_0^t |\nabla u_{t}|_{L_2}^2\, ds
+\frac{1}{4\sigma} \|g\|_{L_2(0,T;L_2(\Omega))},
\end{aligned}
\end{equation}
where we have used integration by parts with respect to time for the $c^2$-term.
Choosing $\breve{b}<\frac{b(1-\delta)}{2}$, then adding \eqref{enest} to \eqref{enest01A} multiplied by $\lambda$
for any $\tau,\sigma>0$ such that
$$\tau+\sigma<1-\underline{\alpha},~\lambda\in \left(0,\min\left\{\tfrac{\hat{b}}{s}\,,\, 2\tfrac{\breve{b}}{c^2}\right\}\right)
\text{ with } s=2c^2+\tfrac{1}{4\tau}\overline{\overline{b}}(C_{H_0^1,L_4}^\Omega)^2$$
yield the energy estimate \eqref{enest1}.
\end{proof}
Note that by the assumptions $\alpha(t,x)\geq-\underline{\alpha}>-1$, $c^2>0$, \eqref{enest} implies an estimate of the form
\begin{eqnarray}
E_0[u](t)
+c_0\int_0^t \Bigl(
|\nabla u_t|_{L_2(\Omega)}^2
+|\nabla u_t|_{L_{q+1}(\Omega)}^{q+1}
\Bigr)\, ds
&&\nonumber\\
\leq C_0 (E_0[u](0)+\|g\|_{(L_{q+1}(0,T;L_{q+1}(\Omega)))^*}) &&
\label{enest0}
\end{eqnarray}
for the usual lower order energy
\begin{equation}\label{defE0}
E_0[u](t)=\left[|u_t|_{L_2(\Omega)}^2 +|\nabla u|_{L_2(\Omega)}^2 \right](t).
\end{equation}
Using Proposition \ref{prop:W2lin} in a fixed point argument, we now show local well-posedness.
\begin{theorem}\label{th:W2locex}
Let $c^2,b,\delta,1-\delta>0$, $k\in\mathbb{R}$, $q>d-1$.\\
For any $T>0$ there is a $\kappa_T>0$ such that for all $u_0,u_1\in W_0^{1,q+1}(\Omega)$ with
\[\begin{aligned}
&E_1[u](0)+|\nabla u_0|_{L_{q+1}}^2 \\
&= |u_1|_{L_2(\Omega)}^2 +|\nabla u_0|_{L_2(\Omega)}^2
+ |\nabla u_1|_{L_2(\Omega)}^2
+ |\nabla u_1|_{L_{q+1}(\Omega)}^{q+1}
+|\nabla u_0|_{L_{q+1}}^2\leq \kappa_T^2
\end{aligned}\]
there exists a weak solution $u\in \mathcal{W} $ of \eqref{Wpress_viscosity} where
\begin{eqnarray}\label{defcW}
\mathcal{W} =\{v\in X
&:& \|v_{tt}\|_{L_2(0,T;L_2(\Omega))}\leq \overline{m}\nonumber\\
&& \wedge \, \|\nabla v_t\|_{L_\infty(0,T;L_2(\Omega))}\leq \overline{m}\nonumber\\
&& \wedge\, \|\nabla v_t\|_{L_{q+1}(0,T;L_{q+1}(\Omega))}\leq \overline{M}\}
\end{eqnarray}
with
\begin{equation}\label{smallnessMbar}
2k C_{W_0^{1,q+1},L_\infty}^\Omega
(\kappa_T+ T^{\frac{q}{q+1}} \overline{M}) <1
\end{equation}
and $\overline{m}$ sufficiently small, and $u$ is unique in $\mathcal{W} $.
\end{theorem}
\begin{proof}
We define the fixed point operator $\mathcal{T} :\mathcal{W} \to X$, $v\mapsto \mathcal{T} v=u$ where $u$ solves \eqref{W2lin} with
\begin{equation}\label{alphafg_W}
\alpha=2k v\,, \quad f=2kv_t\,, \quad g=0\,,
\end{equation}
which is well-defined by Proposition \ref{prop:W2lin}.
Indeed, by $v\in \mathcal{W} $, \eqref{smallnessMbar}, and the penultimate line in \eqref{estLinftyW2} we have that
$\alpha\in C(0,T;L_\infty(\Omega))$, $-1<-\underline{\alpha}\leq \alpha(t,x)\leq \overline{\alpha}$, with $\underline{\alpha}=\overline{\alpha}=
2k C_{W_0^{1,q+1},L_\infty}^\Omega (\kappa_T+ T^{\frac{q}{q+1}} \overline{M}) <1$.
Moreover,
\begin{align*}
\|f-\tfrac{1}{2}\alpha_t\|_{L_\infty(0,T;L_{2}(\Omega)}&\leq k C_{PF} \overline{m},\\
\|f\|_{L_\infty(0,T;L_4(\Omega))}^2&\leq 4k^2 (C_{H_0^1,L_4}^\Omega)^2 \overline{m}^2,
\end{align*}
with $C_{PF}$ the constant in the Poincar\'{e}-Friedrichs inequality
$$ \|w\|_{L_2(\Omega)} \leq C_{PF} \|\nabla w\|_{L_2(\Omega)}\,.$$
Hence, we can make use of the higher energy estimate \eqref{enest1} to conclude that for any $\overline{m},\overline{M}>0$ with
\[
2k C_{W_0^{1,q+1},L_\infty}^\Omega T^{\frac{q}{q+1}} \overline{M} <1,
\quad
k C_{PF} \overline{m}<\frac{1-\delta}{(C_{H_0^1,L_4}^\Omega)^2},
\]
we obtain that under the assumption
\[
E_1[u](0)\leq \kappa_T<
\min\Big\{\frac{1}{2k C_{W_0^{1,q+1},L_\infty}^\Omega}-T^{\frac{q}{q+1}} \overline{M}, \frac{\overline{m}}{\sqrt{C_1}}, \sqrt{\frac{c_1}{C_1}}\overline{m}\Big\},
\]
the operator $\mathcal{T}$ maps into $\mathcal{W} $.
Contractivity is obtained by considering $v^i\in\mathcal{W}$, $u^i=\mathcal{T} v^i\in\mathcal{W}$, $i=1,2$ and subtracting the equations for $u^1$ and $u^2$, which yields
\begin{equation}\label{W2_diff}
\begin{cases}
(1-2kv^1)\hat{u}_{tt}
-b\delta\int_0^1\,\text{div}\Bigl( w^\sigma
\Bigl[|\nabla (u^2+\sigma \hat{u})_t|^2 \nabla \hat{u}_t\\
\quad+(q-1)(\nabla (u^2+\sigma \hat{u})_t \nabla \hat{u}_t)
\nabla(u^2+\sigma \hat{u})_t \Bigr]\, \Bigr)d\sigma \\
\quad -b(1-\delta)\Delta \hat{u}_t -c^2\Delta \hat{u}
-2k v^1_t \hat{u}_t = \ 2k(\hat{v}u^2_{tt}+\hat{v}_tu^2_t),
\\
(\hat{u},\hat{u}_t)|_{t=0}=(0,0),
\\
\hat{u}|_{\partial \Omega} =0,
\\
\end{cases}
\end{equation}
for $\hat{u}=u^1-u^2$, $\hat{v}=v^1-v^2$, with $w^\sigma(x,t)=|\nabla(u^2+\sigma\hat{u})_t(x,t)|^{q-3}$ as in \eqref{wsig}.
Due to the special form of the nonlinear strong damping term here, we cannot apply Proposition \ref{prop:W2lin} directly,
but we can proceed analogously to its proof. By multiplication of \eqref{W2_diff} with $\hat{u}_t$, integration
with respect to space and time, and the fact that the $b\delta$ term yields a nonnegative contribution
\begin{equation}\label{wsig_nonneg1}
\begin{aligned}
&b\delta\int_0^1\int_0^t\int_\Omega w^\sigma\Bigl(
|\nabla (u^2+\sigma \hat{u})_t|^2 |\nabla \hat{u}_t|^2\\
&\qquad+(q-1)(\nabla (u^2+\sigma \hat{u})_t \cdot\nabla \hat{u}_t)^2
\Bigr) \, dx\, ds\, d\sigma \geq 0.
\end{aligned}
\end{equation}
on the left hand side,
we obtain
\begin{eqnarray*}
\lefteqn{\frac12\left[\int_\Omega (1-2kv^1)(\hat{u}_t)^2\, dx
+ c^2 |\nabla \hat{u}|_{L_2(\Omega)}^2 \right]_0^t
+ \hat{b} \int_0^t |\nabla \hat{u}_t|_{L_2(\Omega)}^2
\, ds}
\nonumber\\
& \leq & 2k \int_0^t\int_\Omega \Bigl(
v^1_t (\hat{u}_t)^2 +\hat{v}u^2_{tt}\hat{u}_t+\hat{v}_tu^2_t\hat{u}_t
\Bigr)\, dx\, ds\\
& \leq & 2k (C_{H_0^1,L_4}^\Omega)^2\Bigl(
\|v^1_t\|_{L_\infty(0,T;L_2(\Omega))}
\int_0^t |\nabla \hat{u}_t|_{L_2(\Omega)}^2\, ds\\
&&\qquad+\|u^2_{tt}\|_{L_2(0,T;L_2(\Omega))}
\frac12\Bigl[\|\nabla\hat{v}\|_{L_\infty(0,T;L_2(\Omega))}^2+
\int_0^t |\nabla \hat{u}_t|_{L_2(\Omega)}^2\, ds\Bigr]\\
&&\qquad+\|u^2_t\|_{L_\infty(0,T;L_2(\Omega))}
\frac12\Bigl[\|\nabla\hat{v}_t\|_{L_2(0,T;L_2(\Omega))}^2+
\int_0^t |\nabla \hat{u}_t|_{L_2(\Omega)}^2\, ds\Bigr].
\end{eqnarray*}
This by $v^1,v^2,u^1,u^2\in\mathcal{W}$ yields
\begin{eqnarray*}
\lefteqn{\min\Big\{\tfrac{1-\underline{\alpha}}{2},\tfrac{c^2}{2},
\hat{b}- k (C_{H_0^1,L_4}^\Omega)^2 (3 C_{PF}+1)\overline{m}\Big\}
|||\hat{u}|||}\\
&\leq& k (C_{H_0^1,L_4}^\Omega)^2 (\sqrt{T}+C_{PF})\overline{m}
|||\hat{v}|||\,,
\end{eqnarray*}
where we have used $|\nabla \hat{v}|_{L_\infty(0,T;L_2(\Omega))}^2
\leq T |\nabla \hat{v}_t|_{L_2(0,T;L_2(\Omega))}^2$ since $\nabla \hat{v}(0)=0$,
hence contractivity of $\mathcal{T}$ with respect to the norm
$$|||v|||=\|v_t\|_{L_\infty(0,T;L_2(\Omega))}
+\|\nabla v\|_{L_\infty(0,T;L_2(\Omega))}
+\|\nabla v_t\|_{L_2(0,T;L_2(\Omega))}$$
provided $\overline{m}$ is chosen sufficiently small.
\end{proof}
\begin{remark}\label{rem:W2glob}
To establish a self-mapping property of the fixed point operator used in the local well-posedness proof above, a condition on smallness of the initial data multiplied with the final time is required, cf. \eqref{smallnessMbar}. This results from the appearance of the factor $t^q$ in the $L_\infty$ estimate \eqref{estLinftyW2}, that we have used to exclude the possibility of degeneracy. Thus, whenever \eqref{estLinftyW2} is sharp, degeneracy may occur after finite time even for small initial data. We therefore expect a global in time wellposedness result not to hold for \eqref{Wpress_viscosity}.
\end{remark}
\section{The Westervelt equation in acoustic velocity potential formulation with nonlinear strong damping \eqref{Wpot_viscosity}}
\label{secWpot_viscosity}
Like in Section \ref{secWpress_viscosity} we assume
that $\Omega\subseteq \mathbb{R}^d$, $d \in \{1,2,3\}$, is an open bounded set with Lipschitz boundary in order to make use of \eqref{H01L4}, \eqref{W01qp1linfty}.
\\
Again we first consider an equation in which the nonlinearity occurs only through damping, namely
\begin{equation}\label{Wpot_viscosity_lin}
\begin{cases}
u_{tt}-\alpha\Delta u
-b\,\text{div}\Bigl(((1-\delta) +\delta|\nabla u_t|^{q-1})\nabla u_t\Bigr)
=0
\\
(u,u_t)|_{t=0}=(u_0, u_1)
\\
u|_{\partial \Omega} =0
.\\
\end{cases}
\end{equation}
\begin{proposition} \label{prop:Wpot_viscosity_lin}
\begin{enumerate}
\item[(i)]
Let $T>0$, $b,\delta,1-\delta>0$ and assume
\begin{itemize}
\item $\alpha\in L_2(0,T;L_\infty(\Omega))$, $\nabla\alpha\in L_2(0,T;L_2(\Omega))$
\item $u_0\in H_0^1(\Omega)$.
\item $q\geq1$ if $d=1$, $q>d-1$ if $d\in \{2,3\}$
\end{itemize}
and
\[\begin{cases}
b-(\tfrac12+\sqrt{T})\|\alpha\|_{L_2(0,T;L_\infty(\Omega))}\\
\qquad-(\tfrac12 T+({C_{W_0^{1,q+1},L_\infty}^\Omega})^2)\|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}:=b_1>0 &
\mbox{ if }q=1,\\
b(1-\delta)-(\tfrac12+\sqrt{T})\|\alpha\|_{L_2(0,T;L_\infty(\Omega))}\\ \hfill
-\tfrac12 T\|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}:=b_1>0 & \mbox{ if }q>1.
\end{cases}\]
Then \eqref{Wpot_viscosity_lin} has a weak solution
\begin{align*}
u\in C^1([0,T];L_2(\Omega))\cap W^{1,q+1}(0,T;W_0^{1,q+1}(\Omega))
\end{align*}
and any solution satisifies the energy estimate
\begin{equation}\label{enest_Wpot_viscosity_lin0}
\begin{aligned}
&\tfrac12\left[|u_{t}|_{L_2(\Omega)}^2\right]_0^t
+\min\{b_1,\tfrac{b\delta}{2}\}\int_0^t \Bigl(|\nablau_{t}|_{L_2(\Omega)}^2
+|\nablau_{t}|_{L_{q+1}(\Omega)}^{q+1}\Bigr)\, ds
\\
&\leq
\begin{cases}
\Bigl(\|\alpha\|_{L_2(0,T;L_\infty(\Omega))} +\|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}\Bigr)
\tfrac12 |\nabla u_0|_{L_2(\Omega)} &\mbox{ if }q=1,\\
\Bigl(\|\alpha\|_{L_2(0,T;L_\infty(\Omega))} +\|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}\Bigr)
\tfrac12 |\nabla u_0|_{L_2(\Omega)} \\
\hfill+C(\tfrac{b\delta}{2},\tfrac{q+1}{2})\|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}^{\frac{q+1}{q-1}}
&\mbox{ if }q>1,
\end{cases}
\end{aligned}
\end{equation}
for $C(\tfrac{b\delta}{2},\tfrac{q+1}{2})$ according to \eqref{Cepsr}.
\item[(ii)]
Let $T>0$, $b,\delta,1-\delta>0$, and assume that
\begin{itemize}
\item
$\alpha(t,x)\geq\underline{\alpha}>0$
\item
$\alpha\in C(0,T;L_\infty(\Omega))$, $\alpha_t\in L^{4/3}(0,T;L^2(\Omega))$, $\nabla\alpha\in L^2(0,T;L^4(\Omega))$
\item
$\|\nabla\alpha\|_{L^2(0,T;L^4(\Omega))}< 1$
\item
$\|\alpha_t\|_{L_2(0,T;L_\infty(\Omega))}+\|\nabla\alpha\|_{L^2(0,T;L^4(\Omega))}<\frac{\underline{\alpha}}{2}$
\item
$u_0\in H_0^1(\Omega)$, $u_1\in W_0^{1,q+1}(\Omega)$.
\item
$q\geq 3$
\end{itemize}
Then \eqref{Wpot_viscosity_lin} has a weak solution
\begin{align*}
u\in H^2(0,T;L^2(\Omega))\cap C^1(0,T,W_0^{1,q+1}(\Omega))
\end{align*}
and any solution satisfies the energy estimate
\begin{equation}\label{enest_Wpot_viscosity_lin2}
\begin{aligned}
&\tfrac12\|u_{t}\|_{C(0,t;L_2(\Omega))}^2+\|\sqrt{\alpha}\nablau\|_{C(0,t;L_2(\Omega))}^2
+\|\nablau_{t}\|_{L^2(0,t;L_2(\Omega))}^2 \\
&\quad+\|\nablau_{t}\|_{L_{q+1}(0,t;L_{q+1}(\Omega))}^{q+1}\\
&\quad+\|u_{tt}\|_{L^2(0,t;L_2(\Omega))}^2
+\|\nablau_{t}\|_{C(0,t;L_2(\Omega))}^2
+|\nablau_{t}|_{C(0,t;L_{q+1}(\Omega))}^{q+1}\\
&\leq
\overline{C}\Bigl(
\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))}^{\frac{q+1}{q-1}}
+|u_1|_{L_2(\Omega)}^2+ \|\alpha\|_{C(0,T;L^\infty(\Omega))}|\nabla u_0|_{L_2(\Omega)}^2\\
&\quad+(\|\alpha_t\|_{L^{4/3}(0,T;L_2(\Omega))}
+\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))}) |\nabla u_0|_{L_4(\Omega)}^2\\
&\quad+\|\alpha\|_{C(0,T;L_\infty(\Omega))}
\Bigl(|\nabla u_{0}|_{L_2(\Omega)}^2+|\nabla u_1|_{L^2(\Omega)} |\nabla u_0|_{L_2(\Omega)}\Bigr)\\
&\quad+\|\alpha_t\|_{L^{4/3}(0,T;L_2(\Omega))}^{\frac{q+1}{q-1}}
+(t^{3/2} \|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))})^{\frac{q+1}{q-1}}\\
&\quad+|\nablau_{1}|_{L_2(\Omega)}^2
+|\nablau_{1}|_{L_{q+1}(\Omega)}^{q+1}
\Bigr)
\end{aligned}
\end{equation}
for some constant $\overline{C}>0$.
\end{enumerate}
\end{proposition}
\begin{proof}
Since we deal with an autonomous second order in time PDE, the proof can be done directly via Galerkin discretization, energy estimates and weak limits.
\\
We therefore only provide the crucial energy estimates for (i) and (ii),
which are obtained by multiplying the discretized version of \eqref{Wpot_viscosity_lin} with $u_{n,t}$ and $u_{n,tt}$, respectively and integrating by parts with respect to space and time.
\\
For (i), from multiplication of the discretized version of \eqref{Wpot_viscosity_lin} with $u_{n,t}$ we get
\begin{equation*
\begin{aligned}
&\tfrac12\left[|u_{n,t}|_{L_2(\Omega)}^2\right]_0^t
+\int_0^t \left(b(1-\delta)|\nablau_{n,t}|_{L_2(\Omega)}^2
+b\delta|\nablau_{n,t}|_{L_{q+1}(\Omega)}^{q+1}\right)\, ds
\\
&=\int_0^t\int_\Omega\Bigl( -\alpha \nabla u_{n,t}\nabla u_{n} -u_{n,t}\nabla\alpha \nabla u_{n}
\Bigr) \, dx \, ds\\
&\leq \int_0^t \Bigl(|\alpha(s)|_{L_\infty(\Omega)} |\nabla u_{n,t}(s)|_{L_2(\Omega)}
+|\nabla\alpha(s)|_{L_2(\Omega)}|u_{n,t}(s)|_{L_\infty(\Omega)}\Bigr)\\
&\qquad\left[|\nabla u_{0,n}|_{L_2(\Omega)}+\sqrt{s\int_0^s|\nabla u_{n,t}(\sigma)|_{L_2(\Omega)}^{2}\,d\sigma}\right] \, ds\\
&\leq \left(\|\alpha\|_{L_2(0,T;L_\infty(\Omega))} \sqrt{\int_0^t |\nabla u_{n,t}(s)|_{L_2(\Omega)}^2\, ds}
+\|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}\sqrt{\int_0^t |u_{n,t}(s)|_{L_\infty(\Omega)}^2\, ds}\right)\\
& \quad\cdot\left[|\nabla u_{0,n}|_{L_2(\Omega)}+\sqrt{t\int_0^t|\nabla u_{n,t}(\sigma)|_{L_2(\Omega)}^{2}\,d\sigma}\right]
\end{aligned}
\end{equation*}
which in case $q=1$ yields
\begin{equation}\label{Wpot_visc_lin_enid0}
\begin{aligned}
&\tfrac12\left[|u_{n,t}|_{L_2(\Omega)}^2\right]_0^t
+\int_0^t \Bigl(b(1-\delta)|\nablau_{n,t}|_{L_2(\Omega)}^2
+b\delta|\nablau_{n,t}|_{L_{q+1}(\Omega)}^{q+1}\Bigr)\, ds
\\
&\leq \|\alpha\|_{L_2(0,T;L_\infty(\Omega))} \Bigl( \sqrt{t}\int_0^t |\nabla u_{n,t}(s)|_{L_2(\Omega)}^2\, ds +\tfrac12 \int_0^t |\nabla u_{n,t}(s)|_{L_2(\Omega)}^2\, ds
+\tfrac12 |\nabla u_{0,n}|_{L_2}^2\Bigr)\\
&\quad+\|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}
\Bigl(({C_{W_0^{1,q+1},L_\infty}^\Omega})^2\int_0^t |\nablau_{n,t}(s)|_{L_{q+1}\Omega)}^2\, ds\\
&\qquad+\tfrac12 t\int_0^t |\nabla u_{n,t}(s)|_{L_2(\Omega)}^2\, ds +\tfrac12 |\nabla u_{0,n}|_{L_2}^2\Bigr).
\end{aligned}
\end{equation}
If $q>1$, using \eqref{abeps} and \eqref{Cepsr}, we further estimate
\begin{equation}\label{Wpot_visc_lin_enid0_qgt1}
\begin{aligned}
&\tfrac12\left[|u_{n,t}|_{L_2(\Omega)}^2\right]_0^t
+\int_0^t \Bigl(b(1-\delta)|\nablau_{n,t}|_{L_2(\Omega)}^2
+b\delta|\nablau_{n,t}|_{L_{q+1}(\Omega)}^{q+1}\Bigr)\, ds\\
&\leq \|\alpha\|_{L_2(0,T;L_\infty(\Omega))} \Bigl( \sqrt{t}\int_0^t |\nabla u_{n,t}(s)|_{L_2(\Omega)}^2\, ds \\
&\quad+\tfrac12 \int_0^t |\nabla u_{n,t}(s)|_{L_2(\Omega)}^2\, ds
+\tfrac12 |\nabla u_{0,n}|_{L_2}^2\Bigr)+\tfrac{b\delta}{2}\int_0^t |\nablau_{n,t}(s)|_{L^{q+1}\Omega)}^{q+1}\, ds\\
&\quad+\|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}
\Bigl(\tfrac12 t\int_0^t |\nabla u_{n,t}(s)|_{L_2(\Omega)}^2\, ds +\tfrac12 |\nabla u_{0,n}|_{L_2}^2\Bigr)\\
&
\quad+C(\tfrac{b\delta}{2},\tfrac{q+1}{2}) t (({C_{W_0^{1,q+1},L_\infty}^\Omega})^2 \|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))})^{\frac{q+1}{q-1}}\,.
\end{aligned}
\end{equation}
By our assumptions on smallness of $\alpha$, the estimates \eqref{Wpot_visc_lin_enid0} and \eqref{Wpot_visc_lin_enid0_qgt1} yield
\begin{equation}\label{Wpot_visc_lin_enid0_1}
\begin{aligned}
&\tfrac12\left[|u_{n,t}|_{L_2(\Omega)}^2\right]_0^t
+\min\{b_1,\tfrac{b\delta}{2}\}\int_0^t \Bigl(|\nablau_{n,t}|_{L_2(\Omega)}^2
+|\nablau_{n,t}|_{L_{q+1}(\Omega)}^{q+1}\Bigr)\, ds
\\
\leq&
\begin{cases}
\Bigl(\|\alpha\|_{L_2(0,T;L_\infty(\Omega))} +\|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}\Bigr)\tfrac12 |\nabla u_{0,n}|_{L_2}^2&\mbox{ if }q=1,\\
\Bigl(\|\alpha\|_{L_2(0,T;L_\infty(\Omega))} +\|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}\Bigr)\tfrac12 |\nabla u_{0,n}|_{L_2}^2\\
\hfill+ C(\tfrac{b\delta}{2},\tfrac{q+1}{2}) t \|\nabla\alpha\|_{L_2(0,T;L_2(\Omega))}^{\frac{q+1}{q-1}}&\mbox{ if }q>1.
\end{cases}
\end{aligned}
\end{equation}
For (ii),
we multiply \eqref{Wpot_viscosity_lin} with $u_{n,t}$ (but carry out the estimates after multiplication differently) and with $u_{n,tt}$.
By multiplication with $u_{n,t}$ we get
\begin{equation}\label{Wpot_visc_lin_enid0_iii}
\begin{aligned}
&\tfrac12\left[|u_{n,t}|_{L_2(\Omega)}^2+|\sqrt{\alpha}\nablau_{n}|_{L_2(\Omega)}^2\right]_0^t+\int_0^t \Bigl(b(1-\delta)|\nablau_{n,t}|_{L_2(\Omega)}^2
+b\delta|\nablau_{n,t}|_{L_{q+1}(\Omega)}^{q+1}\Bigr)\, ds\\
&=\int_0^t\int_\Omega\Bigl( \tfrac12\alpha_t |\nabla u_{n}|^2 -u_{n,t}\nabla\alpha \nabla u_{n}
\Bigr) \, dx \, ds\\
&\leq \tfrac12\|\alpha_t\|_{L_2(0,T;L_\infty(\Omega))} \|\nabla u_{n}\|_{C(0,T;L_2(\Omega))}^2\\
&\quad+\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))}\Bigl(\tfrac12 \|u_{n,t}\|_{C(0,T;L_4(\Omega))}^2
+\tfrac12 \|\nabla u_{n}\|_{C(0,T;L_2(\Omega))}^2\Bigr)
\end{aligned}
\end{equation}
which due to our smallness assumption on $\alpha$ implies
\begin{equation}\label{Wpot_visc_lin_enid0_iii_1}
\begin{aligned}
&\tfrac12|u_{n,t}(t)|_{L_2(\Omega)}^2+c_1|\sqrt{\alpha}\nablau_{n}(t)|_{L_2(\Omega)}^2+\int_0^t \Bigl(b(1-\delta)|\nablau_{n,t}|_{L_2(\Omega)}^2
+b\delta|\nablau_{n,t}|_{L_{q+1}(\Omega)}^{q+1}\Bigr)\, ds
\\
&\leq \epsilon_0 \|u_{n,t}\|_{C(0,T;L_4(\Omega))}^{q+1}+
C(\epsilon_0,\tfrac{q+1}{2}) (\tfrac12\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))})^{\frac{q+1}{q-1}} \\
&\quad+\tfrac12|u_{1,n}|_{L_2(\Omega)}^2+\tfrac12 \|\alpha\|_{C(0,T;L^\infty(\Omega))}|\nabla u_{0,n}|_{L_2(\Omega)}^2
\end{aligned}
\end{equation}
for $c_1=\frac{\underline{\alpha}}{2}-\|\alpha_t\|_{L_2(0,T;L_\infty(\Omega))}-\|\nabla\alpha\|_{L^2(0,T;L^4(\Omega))}$ and some $\epsilon_0 >0$.
\\
Multiplication with $u_{n,tt}$ yields
\begin{align*}
&\int_0^t |u_{n,tt}(s)|_{L_2(\Omega)}^2\, ds
+\left[\tfrac{b(1-\delta)}{2}|\nablau_{n,t}|_{L_2}^2
+\tfrac{b\delta}{q+1}|\nablau_{n,t}|_{L_{q+1}(\Omega)}^{q+1}
\right]_0^t
\\
&=\int_0^t\int_\Omega\left( -\alpha \nabla u_{n,tt}\nabla u_{n} -u_{n,tt}\nabla\alpha \nabla u_n
\right) \, dx \, ds\\
&=\int_0^t\int_\Omega\left( \alpha_t \nabla u_{n,t}\nabla u_{n}+\alpha|\nabla u_{n,t}|^2
-u_{n,tt}\nabla\alpha \nabla u_n\right)dx\, ds-\left[\int_\Omega\left( \alpha \nabla u_{n,t}\nabla u_{n} \right)\,dx\right]_0^t\\
& \leq
\int_0^t\left( |\alpha_t(s)|_{L_2(\Omega)} |\nabla u_{n,t}(s)|_{L_4(\Omega)}
+|\nabla\alpha(s)|_{L^4(\Omega)}|u_{n,tt}(s)|_{L^2(\Omega)} \right) \\
&\quad\left[|\nabla u_{0,n}|_{L_4(\Omega)}+\sqrt[4]{s^3\int_0^s|\nabla u_{n,t}(\sigma)|_{L_4(\Omega)}^{4}\,d\sigma}\right] \, ds
\\
&+\|\alpha\|_{C(0,T;L_\infty(\Omega))} \Bigl(|\nabla u_{n,t}(t)|_{L^2(\Omega)}
|\nabla u_{n}(t)|_{L^2(\Omega)}+|\nabla u_{1,n}|_{L^2(\Omega)} |\nabla u_{0,n}|_{L_2(\Omega)}+\int_0^t|\nabla u_{n,t}(s)|_{L^2(\Omega)}^2\, ds\Bigr)\\
&\leq
\|\alpha_t\|_{L^{4/3}(0,T;L_2(\Omega))} \left((1+\tfrac12 t^{\frac{3}{2}})\|\nabla u_{n,t}\|_{L^4(0,t;L_4(\Omega))}^2+\tfrac12|\nabla u_{0,n}|_{L_4(\Omega)}^2\right)\\
&+\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))} \left(\|u_{n,tt}\|_{L^2(0,t;L_2(\Omega))}^2
+\tfrac12 |\nabla u_{0,n}|_{L_4(\Omega)}^2 +\frac{t^{\frac{3}{2}} }{2}\|\nabla u_{n,t}\|_{L^4(0,T;L_4(\Omega))}^2\right)\\
&\quad+\|\alpha\|_{C(0,T;L_\infty(\Omega))} \Bigl(|\nabla u_{1,n}|_{L^2(\Omega)} |\nabla u_{0,n}|_{L_2(\Omega)}
+\|\nabla u_{n,t}\|_{L^2(0,t;L_2(\Omega))}^2\Bigr)\\
&\quad+\tfrac{b(1-\delta)}{4}|\nabla u_{n,t}(t)|_{L^2(\Omega)}^2
+\tfrac{1}{b(1-\delta)}\|\alpha\|_{C(0,T;L_\infty(\Omega))}^2|\nabla u_{n}(t)|_{L^2(\Omega)}^2\\
&\leq
\|\alpha_t\|_{L^{4/3}(0,T;L_2(\Omega))}\tfrac12|\nabla u_{0,n}|_{L_4(\Omega)}^2\\
&\quad+\|\nabla u_{n,t}\|_{L^4(0,T;L_4(\Omega))}^{q+1}
+C(1,\tfrac{q+1}{2})((1+\tfrac12 t^{\frac{3}{2}})\|\alpha_t\|_{L^{4/3}(0,T;L_2(\Omega))})^{\frac{q+1}{q-1}}\\
&\quad+\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))} \Bigl(\|u_{n,tt}(s)|_{L^2(0,t;L_2(\Omega))}^2
+\tfrac12 |\nabla u_{0,n}|_{L_4(\Omega)}^2\Bigr)\\
&\quad+\|\nabla u_{n,t}\|_{L^4(0,T;L_4(\Omega))}^{q+1}
+C(1,\tfrac{q+1}{2})(t^{3/2}\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))})^{\frac{q+1}{q-1}}\\
&\quad+\|\alpha\|_{C(0,T;L_\infty(\Omega))} \Bigl(|\nabla u_{1,n}|_{L^2(\Omega)} |\nabla u_{0,n}|_{L_2(\Omega)}
+\|\nabla u_{n,t}\|_{L^2(0,t;L_2(\Omega))}^2\Bigr)\\
&\quad+\tfrac{b(1-\delta)}{4}|\nabla u_{n,t}(t)|_{L^2(\Omega)}^2
+\tfrac{1}{b(1-\delta)}\|\alpha\|_{C(0,T;L_\infty(\Omega))}^2|\nabla u_{n}(t)|_{L^2(\Omega)}^2\,,
\end{align*}
which, by smallness of $\alpha$, implies
\begin{equation}\label{Wpot_visc_lin_enid1_iii_1}
\begin{aligned}
&\int_0^t c_2|u_{n,tt}(s)|_{L_2(\Omega)}^2\, ds
+\left[\tfrac{b(1-\delta)}{4}|\nablau_{n,t}|_{L_2}^2
+\tfrac{b\delta}{q+1}|\nablau_{n,t}|_{L_{q+1}(\Omega)}^{q+1}
\right]_0^t\\
&\leq
(\|\alpha_t\|_{L^{4/3}(0,T;L_2(\Omega))}
+\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))}) \tfrac12 |\nabla u_{0,n}|_{L_4(\Omega)}^2\\
&\quad+\|\alpha\|_{C(0,T;L_\infty(\Omega))}
\Bigl(|\nabla u_{0,n}|_{L_2(\Omega)}^2+|\nabla u_{1,n}|_{L^2(\Omega)} |\nabla u_{0,n}|_{L_2(\Omega)}\\
&\quad+\|\nabla u_{n,t}\|_{L^2(0,t;L_2(\Omega))}^2\Bigr)+ 2 \|\nabla u_{n,t}\|_{L^4(0,T;L_4(\Omega))})^{q+1}\\
&\quad+C(1,\tfrac{q+1}{2})\|\alpha_t\|_{L^{4/3}(0,T;L_2(\Omega))}^{\frac{q+1}{q-1}}
+C(1,\tfrac{q+1}{2})(t^{3/2}\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))})^{\frac{q+1}{q-1}}\\
&\quad+\tfrac{1}{b(1-\delta)}\|\alpha\|_{C(0,T;L_\infty(\Omega))}^2|\nabla u_{n}(t)|_{L^2(\Omega)}^2
\end{aligned}
\end{equation}
for $c_2=1-\|\nabla\alpha\|_{L^2(0,T;L^4(\Omega))}$.
Adding \eqref{Wpot_visc_lin_enid0_iii_1} and $\epsilon$ times \eqref{Wpot_visc_lin_enid1_iii_1} with $\epsilon_0,\epsilon$ sufficiently small
\begin{align*}
&\epsilon<\min\left\{\frac{b(1-\delta)}{2\|\alpha\|_{C(0,T;L_\infty(\Omega))}},
\frac{c_1\underline{\alpha}b(1-\delta)}{2\|\alpha\|_{C(0,T;L_\infty(\Omega))}^2},
\frac{b\delta}{4({C_{L_{q+1},L_4}^\Omega})^{q+1}}\right\}\,,
\\
&\epsilon_0<\frac{b\delta\epsilon}{(q+1)({C_{L_{q+1},L_4}^\Omega})^{q+1}},
\end{align*}
where we use $q+1\geq 4$, yields the energy estimate
\begin{align*}
&\tfrac12\|u_{n,t}\|_{C(0,t;L_2(\Omega))}^2+c_1\|\sqrt{\alpha}\nablau_{n}\|_{C(0,t;L_2(\Omega))}^2
+\tfrac{b(1-\delta)}{2}\|\nablau_{n,t}\|_{L^2(0,t;L_2(\Omega))}^2 \\
&\quad+c_3\|\nablau_{n,t}\|_{L^{q+1}(0,t;L_{q+1}(\Omega))}^{q+1}
+\epsilon c_2\|u_{n,tt}\|_{L^2(0,t;L_2(\Omega))}^2\\
&\quad+\epsilon \tfrac{b(1-\delta)}{2} \|\nablau_{n,t}\|_{C(0,t;L_2(\Omega))}^2
+\epsilon \tfrac{b\delta}{q+1}|\nablau_{n,t}|_{C(0,t;L_{q+1}(\Omega))}^{q+1}\\
&\leq
C(\epsilon_0,\tfrac{q+1}{2})(\tfrac12\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))})^{\frac{q+1}{q-1}}\\
&\quad+\tfrac12|u_{1,n}|_{L_2(\Omega)}^2+\tfrac12 \|\alpha\|_{C(0,T;L^\infty(\Omega))}|\nabla u_{0,n}|_{L_2(\Omega)}^2\\
&\quad+\epsilon(\|\alpha_t\|_{L^{4/3}(0,T;L_2(\Omega))}
+\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))})\tfrac12 |\nabla u_{0,n}|_{L_4(\Omega)}^2\\
&\quad+\epsilon\|\alpha\|_{C(0,T;L_\infty(\Omega))}
\Bigl(|\nabla u_{0,n}|_{L_2(\Omega)}^2+|\nabla u_{1,n}|_{L^2(\Omega)} |\nabla u_{0,n}|_{L_2(\Omega)}\Bigr)\\
&\quad+\epsilon C(1,\tfrac{q+1}{2})\|\alpha_t\|_{L^{4/3}(0,T;L_2(\Omega))}^{\frac{q+1}{q-1}}
+\epsilon C(1,\tfrac{q+1}{2})(t^{\frac{3}{2}} \|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))})^{\frac{q+1}{q-1}}\\
&\quad+\epsilon \tfrac{b(1-\delta)}{2} |\nablau_{1,n}|_{L_2(\Omega)}^2
+\epsilon \tfrac{b\delta}{q+1}|\nablau_{1,n}|_{L_{q+1}(\Omega)}^{q+1},
\end{align*}
for some $c_3>0,$ which leads to \eqref{enest_Wpot_viscosity_lin2}.
\end{proof}
\begin{remark}
Note that multiplication with $u_t$ (via $q\geq d-1$) according to \eqref{enest_Wpot_viscosity_lin0} only gives an
$L_2(0,T;L_\infty(\Omega))$ bound on $u_{t}$ and in order to obtain a $C(0,T;L_\infty(\Omega))$ bound,
multiplication with $u_{tt}$ is required, cf. \eqref{estLinftyW3}.
Thus part (i) of Proposition \ref{prop:Wpot_viscosity_lin} is only an intermediate result.
\\
In part (ii) we make use of a positive sign of the term pertaining to the potential energy in the equation.
However this leads to an $L_4(\Omega)$ norm term on the right hand side of the energy identity, which we can
only dominate by means of the $L_{q+1}(\Omega)$ norm term on the left hand side. Hence, for this part, $q+1\geq 4$ is needed.
\end{remark}
\begin{theorem}\label{th:W3locex}
Let $c^2,b,\delta,1-\delta>0$, $k\in\mathbb{R}$, $q\geq3$.\\
There exist $\kappa>0$, $\overline{m}>0$, $\overline{M}>0$, $T>0$ such that for all $u_0 \in H_0^1(\Omega)$, $u_1\in H_0^1(\Omega) \cap W^{1,q+1}(\Omega)$ with
\begin{equation}\label{kappaW3}
|u_1|_{L_2(\Omega)}^2 +|\nabla u_0|_{L_2(\Omega)}^2
+ |\nabla u_1|_{L_2(\Omega)}^2
+ |\nabla u_1|_{L_{q+1}(\Omega)}^{q+1}
\leq \kappa^2
\end{equation}
there exists a weak solution $u\in \mathcal{W}\subseteq X=H^2(0,T;L_2(\Omega))\cap C^1(0,T;W^{1,q+1}(\Omega))$ of \eqref{Wpot_viscosity} where
\begin{eqnarray}\label{defcW_W3}
\mathcal{W} =\{v\in X
&:& \|v_{tt}\|_{L_2(0,T;L_2(\Omega))}\leq \overline{m}\nonumber\\
&& \wedge \|\nabla v_t\|_{C(0,T;L_2(\Omega))}\leq \overline{m}\nonumber\\
&& \wedge \|\nabla v_t\|_{C(0,T;L_{q+1}(\Omega))}\leq \overline{M}\}\,.
\end{eqnarray}
\end{theorem}
\begin{proof}
Relying on Proposition \ref{prop:Wpot_viscosity_lin}, we carry out the proof by means of a fixed point argument.
To this end, we define the fixed point operator $\mathcal{T} :\mathcal{W} \to X$, $v\mapsto \mathcal{T} v=u$ where $u$ solves \eqref{Wpot_viscosity_lin} with
\begin{equation}\label{alpha_Wpotviscosity}
\alpha=\frac{c^2}{1-2\tilde{k}v_t}\,,
\end{equation}
which is well-defined by Proposition \ref{prop:Wpot_viscosity_lin} (ii), since we have
\[
\begin{aligned}
&\alpha(t,x)\geq\frac{c^2}{1+2\tilde{k}\|v_t\|_{C(0,T;L_\infty(\Omega))}}
\geq \frac{c^2}{1+2\tilde{k}C_{W_0^{1,q+1},L_\infty}^\Omega\overline{M}}\geq \frac{2c^2}{3}\\
&\|\alpha\|_{C(0,T;L_\infty(\Omega))}\leq\frac{c^2}{1-2\tilde{k}\|v_t\|_{C(0,T;L_\infty(\Omega))}}
\leq \frac{c^2}{1-2\tilde{k}C_{W_0^{1,q+1},L_\infty}^\Omega\overline{M}}\leq 2c^2\\
&\|\nabla\alpha\|_{L_2(0,T;L_4(\Omega))}=\|\frac{2\tilde{k}c^2}{(1-2\tilde{k}v_t)^2}\nabla v_t\|_{L_2(0,T;L_4(\Omega))}\\
&\quad \leq \frac{2\tilde{k}c^2}{(1-2\tilde{k}C_{W_0^{1,q+1},L_\infty}^\Omega\overline{M})^2}
C_{L_{q+1},L_4}^\Omega \sqrt{T}\overline{M}
\leq \tilde{k}c^2 C_{L_{q+1},L_4}^\Omega \sqrt{T}\overline{M}
\\
&\|\alpha_t\|_{L^{4/3}(0,T;L_2(\Omega))}=\|\frac{2\tilde{k}c^2}{(1-2\tilde{k}v_t)^2} v_{tt}\|_{L^{4/3}(0,T;L_2(\Omega))}\\
&\quad \leq \frac{2\tilde{k}c^2}{(1-2\tilde{k}C_{W_0^{1,q+1},L_\infty}^\Omega\overline{M})^2}
\sqrt[4]{T} \overline{m}
\leq 8\tilde{k}c^2 \sqrt[4]{T} \overline{m}
\,,
\end{aligned}
\]
where we have used $v\in\mathcal{W}$, so the positivity and smallness assumptions on $\alpha$ in Proposition \ref{prop:Wpot_viscosity_lin} are satisfied, provided $\overline{m}$, $\overline{M}$ are sufficiently small, in particular
$\overline{M}\leq (4\tilde{k}C_{W_0^{1,q+1},L_\infty}^\Omega)^{-1}$.
\\
Hence, the energy estimate \eqref{enest_Wpot_viscosity_lin2} yields
\begin{equation*}
\begin{aligned}
&\tfrac12\|u_{n,t}\|_{C(0,T;L_2(\Omega))}^2+\|\sqrt{\alpha}\nablau_{n}\|_{C(0,T;L_2(\Omega))}^2
+\|\nablau_{n,t}\|_{L^2(0,T;L_2(\Omega))}^2 \\
&+\|\nablau_{n,t}\|_{L^{q+1}(0,T;L_{q+1}(\Omega))}^{q+1}
\\
&+\|u_{n,tt}\|_{L^2(0,T;L_2(\Omega))}^2
+\|\nablau_{n,t}\|_{C(0,T;L_2(\Omega))}^2
+|\nablau_{n,t}|_{C(0,T;L_{q+1}(\Omega))}^{q+1}\\
\leq &
\overline{C}\Bigl(
(\sqrt{T}\overline{M})^{\frac{q+1}{q-1}}
+|u_1|_{L_2(\Omega)}^2+ |\nabla u_0|_{L_2(\Omega)}^2\\
&+(\sqrt[4]{T}\overline{m}+\sqrt{T}\overline{M}) |\nabla u_0|_{L_4(\Omega)}^2\\
&+|\nabla u_0|_{L_2(\Omega)}^2+|\nabla u_1|_{L^2(\Omega)} |\nabla u_0|_{L_2(\Omega)}\\
&+(\sqrt[4]{T}\overline{m})^{\frac{q+1}{q-1}}
+(T^2\overline{M})^{\frac{q+1}{q-1}}\\
&+|\nablau_1|_{L_2(\Omega)}^2
+|\nablau_1|_{L_{q+1}(\Omega)}^{q+1}
\Bigr)\,.
\end{aligned}
\end{equation*}
Thus by making $T$ and the bound $\kappa$ on the initial data sufficiently small, by an appropriate choice of
$\overline{m}$, $\overline{M}$ we can achieve that $u\in\mathcal{W}$. By closedness of $\mathcal{W}$ we obtain existence of a solution.
\\
\end{proof}
\begin{remark}\label{rem:Wpot_viscosity_uniqueness}
To see that contractivity and therefore also uniqueness fails for \eqref{Wpot_viscosity}, consider two solutions $u^i=\mathcal{T}(v^i)$, $i=1,2$
as well as their difference $\hat{u}=u^1-u^2$, which is a weak solution to
\begin{equation}\label{Wpot_vicosity_lin_diff}
\begin{cases}
\hat{u}_{tt}-\frac{c^2}{1-2\tilde{k}v^1_t}\Delta \hat{u}-b(1-\delta)\Delta \hat{u}_t b\delta\int_0^1\,\text{div}\Bigl( w^\sigma
\Bigl[|\nabla (u^2+\sigma \hat{u})_t|^2 \nabla \hat{u}_t\\
+(q-1)(\nabla (u^2+\sigma \hat{u})_t \nabla \hat{u}_t)
\nabla(u^2+\sigma \hat{u})_t \Bigr]\, \Bigr)d\sigma
=\frac{2\tilde{k}c^2}{(1-2\tilde{k}v^1_t)(1-2\tilde{k}v^2_t)}\Delta v^2\hat{v}_t
\\
(\hat{u},\hat{u}_t)|_{t=0}=(0,0),
\\
\hat{u}|_{\partial \Omega} =0
,\\
\end{cases}
\end{equation}
with $\hat{v}=v^1-v^2$ and $w^\sigma(x,t)=|\nabla(u^2+\sigma\hat{u})_t(x,t)|^{q-3}$ as in \eqref{wsig}.
Upon multiplication with $\hat{u}_t$ and integration with respect to space and time, like in the proof of Proposition \ref{prop:W2lin} and
Theorem \ref{th:W2locex}, the $b\delta$ term yields a nonnegative contribution on the left hand side.
Therewith, similarly to \eqref{Wpot_visc_lin_enid0_iii}, we obtain
\[
\begin{aligned}
&\tfrac12\left[|\hat{u}_t|_{L_2(\Omega)}^2+|\frac{c}{\sqrt{1-2\tilde{k}v^1_t}}\nabla\hat{u}|_{L_2(\Omega)}^2\right]_0^t
+\int_0^t b(1-\delta)|\nabla\hat{u}_t|_{L_2(\Omega)}^2
\\
=&\int_0^t\int_\Omega\Bigl( -\frac{\tilde{k}c^2}{(1-2\tilde{k}v^1_t)^2} v^1_{tt} |\nabla \hat{u}|^2
+\hat{u}_t \frac{2\tilde{k}c^2}{(1-2\tilde{k}v^1_t)^2}\nabla v^1_t \nabla \hat{u}\\
&\quad
-\frac{2\tilde{k}c^2}{(1-2\tilde{k}v^1_t)(1-2\tilde{k}v^2_t)}\nabla v^2 \nabla\hat{v}_t
-\frac{4\tilde{k}^2c^2}{(1-2\tilde{k}v^1_t)^2(1-2\tilde{k}v^2_t)}\nabla v^1_t\nabla v^2 \hat{v}_t\\
&\quad
-\frac{4\tilde{k}^2c^2}{(1-2\tilde{k}v^1_t)(1-2\tilde{k}v^2_t)^2}\nabla v^2_t\nabla v^2 \hat{v}_t
\Bigr) \, dx \, ds
\end{aligned}
\]
We see that estimation of the first term on the right hand side (the one containing $v^1_{tt} |\nabla \hat{u}|^2$) would require higher spatial summability of $v^1_{tt}$ and/or of $\nabla \hat{u}$. The estimates on these two quantities, following from $v^1\in\mathcal{W}$ and from the left hand side of the above estimate on $\hat{u}$ are not strong enough for this purpose. However, multiplication of \eqref{Wpot_viscosity} with higher time or space derivatives of $u$ leads to difficulties involving the strong nonlinear damping term; the same holds for multiplication of \eqref{Wpot_vicosity_lin_diff} with $\hat{u}_{tt}$.
\end{remark}
\section{Linear wave -- nonlinear Westervelt equation coupling via nonlinear strong damping \eqref{Wacoustic_viscosity}}
\label{secWacoustic_viscosity}
We make the following assumptions on the spatially varying coefficients
$\lambda$, $k$, $\varrho$, $b$, $\delta$ in \eqref{Wacoustic_viscosity}:
\begin{equation}\label{condcoeffacoust}
\begin{cases}
k, \lambda, \varrho, b\geq0, \delta \in L_\infty(\Omega), 0 \leq\delta\leq 1,\\
\exists \underline{\lambda},\underline{\varrho}>0: \
\underline{\lambda}\leq \lambda(x) \,, \ \underline{\varrho}\leq\varrho(x)\
\mbox{ in }\Omega,\\
\exists \underline{b},\underline{\delta}>0: \
\underline{b}\leq b(x) \,, \ \underline{\delta}\leq\delta(x)\
\mbox{ in }\Omega_{nl},\\
\mbox{where } \Omega_{nl}:=\{x\in\Omega\, : \, k(x)\not=0\}.
\end{cases}
\end{equation}
Since \eqref{Wacoustic_viscosity} can be viewed as a version of \eqref{Wpress_viscosity} with spatially varying coefficients, inspection of the proof of Theorem \ref{th:W2locex} immediately yields
\begin{corollary}
Let $q>d-1$ and suppose assumptions \eqref{condcoeffacoust} are satisfied.\\
For any $T>0$ there is a $\kappa_T>0$ such that for all $u_0,u_1\in H_0^1(\Omega)\cap W^{1,q+1}(\Omega)$ with
$$E_1[u](0)+|\nabla u_0|_{L_{q+1}}^2 \leq \kappa_T^2\,,$$
there exists a weak solution $u\in \mathcal{W} $ (defined as in \eqref{defcW} with
\eqref{smallnessMbar} and $\overline{m}$ sufficiently small) of \eqref{Wacoustic_viscosity} and $u$ is unique in $\mathcal{W} $.
\end{corollary}
\begin{remark}\label{rem:interfacecond}
Consider the case that the coefficients $\lambda, k, \varrho, b, \delta$ are piecewise constant, for simplicity of exposition, only on two open subdomains $\Omega_I$, $\Omega_{II}$ with $\Omega_I\cap\Omega_{II}=\emptyset$, $\Omega_I\cup\Omega_{II}=\Omega$, with a smooth interface $\Gamma$ defined by $\overline{\Gamma}=\overline{\Omega}_I\cap\overline{\Omega}_{II}$.
Due to the fact that (cf., e.g., \cite{BrezziFortin}, \cite{RaviartThomas})
$$
H^1(\Omega)= \{v\in L^2(\Omega)\, : \, v\vert_{\Omega_i}=v_i\in H^1(\Omega_i)
\,, \, i=I,II\,, \ \mbox{tr}_\Gamma v_I=\mbox{tr}_\Gamma v_{II}\}\,,
$$
multiplying the PDE \eqref{Wacoustic_viscosity} with $v\in H^1(\Omega)$ and integrating by parts separately on each subdomain (cf. equations (2.3), (2.5) in \cite{BambergerGlowinskiTran}), we arrive at the interface conditions
\begin{eqnarray*}
&&\mbox{tr}_\Gamma (u\vert_{\Omega_I})=\mbox{tr}_\Gamma (u\vert_{\Omega_{II}})\\
&&\mbox{ i.e., continuity of pressure in a trace sense}\\
&&\int_\Gamma \Bigl\{\mathbf{v}_u\vert_{\Omega_I}-\mathbf{v}_u\vert_{\Omega_{II}}
\Bigr\} \nu \, v \,ds=0 \quad \forall v\in H^1(\Omega)\\
&& \mbox{where }\mathbf{v}_u:= \frac{1}{\varrho} \nabla u
+ b\,\Bigl((1-\delta) +\delta|\nabla u_t|^{q-1}\Bigr)\nabla u_t\,,\\
&&\mbox{ i.e., continuity of (modified) normal velocity in a variational sense.}
\end{eqnarray*}
Here the notation for velocity is motivated by the linearized Euler equation
$\varrho\mathbf{v}_t=-\nabla p$ relating the velocity $\mathbf{v}$ with the pressure $p$, which is denoted by $u$ here.
\end{remark}
\section{Linear elasticity -- nonlinear Westervelt equation coupling via nonlinear strong damping \eqref{Welastic_viscosity}}
\label{secWelastic_viscosity}
In this section we consider the coupled nonlinear acoustic -- elastic system
\begin{equation}\label{Welastic_viscosity1}
\begin{cases}
\varrho(x) \mathbf{u}_{tt}-\mathcal{B}^T\frac{1}{1-2\tilde{k}(x)\psi_t} [c](x)\mathcal{B}\mathbf{u}\\
\qquad \qquad+\mathcal{B}^T\Bigl(((1-\delta(x)) +\delta(x)|\mathcal{B} \mathbf{u}_t|^{q-1})[b](x)\mathcal{B} \mathbf{u}_t\Bigr)
=0,
\\
(\mathbf{u},\mathbf{u}_t)|_{t=0}=(u_0, u_1),
\\
\mathbf{u}|_{\partial \Omega} =0
,\\
-\Delta \psi=-\text{div}\, \mathbf{u}
\psi|_{\partial \Omega} =0.
\end{cases}
\end{equation}
We again assume that $\Omega\subseteq\mathbb{R}^d$ is an open bounded set with Lipschitz boundary and
impose the following conditions on the coefficients in \eqref{Welastic_viscosity1}
\begin{equation}\label{condcoeffelast}
\begin{cases}
\tilde{k}, \varrho, [b], [c], \delta \in L_\infty(\Omega), ~0\leq\delta\leq 1,\\
\hfill \mbox{ with }[b(x)], [c(x)]\mbox{ symmetric positive semidefinite matrices},\\
\exists \underline{\varrho},\overline{\varrho},\underline{c},\overline{c},\underline{\gamma}>0: \
\quad \underline{\varrho}\leq \varrho(x)\leq \overline{\varrho}, \\
\qquad \qquad \qquad \qquad \quad \underline{c}\leq \lambda_{\min}([c(x)])\leq \lambda_{\max}([c(x)])\leq \overline{c},\\
\qquad \qquad \qquad \qquad \quad \underline{\gamma}\leq (1-\delta(x))\lambda_{\min}([b(x)]),\quad\hfill \mbox{ in }\Omega, \\
\exists \underline{\delta},\underline{b},\overline{b}>0\ \exists \overline{\delta}<1: \
\underline{b}\leq \lambda_{\min}([b(x)])\leq \lambda_{\max}([b(x)])\leq\overline{b},\\
\qquad \qquad \qquad \qquad \quad \underline{\delta}\leq\delta(x)\leq\overline{\delta}, \quad \hfill \mbox{ in }\Omega_{nl},\\
\qquad\qquad\mbox{where } \Omega_{nl}:=\{x\in\Omega\, : \, \tilde{k}(x)\not=0\}\mbox{ is an open domain and}\\
\qquad\qquad \mbox{$\overline{\Omega_{nl}}$ is a compact subset of $\Omega$ or $\Omega$ has $C^2$ boundary}\,.
\end{cases}
\end{equation}
In particular, strong linear damping is required in the whole domain, whereas strong nonlinear damping is only needed in the region $\Omega_{nl}$ of nonlinear acoustics.
\\
Similarly to Theorem \ref{th:W3locex} we obtain the following local existence result
\begin{theorem}\label{th:W5locex}
Let $q\geq3$ and
assumption \eqref{condcoeffelast}
be satisfied.\\
There exist $\kappa>0$, $\overline{m}>0$, $\overline{a}>0$, $\overline{M}>0$, $T>0$ such that for all $\mathbf{u}_0,\mathbf{u}_1\in (H_0^1(\Omega))^d$, $\mathbf{u}_1\vert_{\Omega_{nl}}\in (W^{1,q+1}(\Omega_{nl}))^d$ with
\[
|\mathbf{u}_1|_{L_2(\Omega)}^2 +|\mathcal{B} \mathbf{u}_0|_{L_2(\Omega)}^2
+ |\mathcal{B} \mathbf{u}_1|_{L_2(\Omega)}^2
+ |\mathcal{B} \mathbf{u}_1|_{L_{q+1}(\Omega_{nl})}^{q+1}
\leq \kappa^2
\]
we have existence of a weak solution $\mathbf{u}\in \mathcal{W}\subseteq H^2(0,T;(L_2(\Omega))^d)\cap C^1(0,T;(H_0^1(\Omega))^d)$ of \eqref{Welastic_viscosity1} where
\begin{equation}\label{defcW_W5}
\begin{aligned}
\mathcal{W} =\{\mathbf{v}\in X
:~~& \|\mathbf{v}_{tt}\|_{L_2(0,T;L_2(\Omega))}\leq \overline{m}\\
& \wedge \|\mathcal{B} \mathbf{v}_t\|_{C(0,T;L_2(\Omega))}\leq \overline{a}\\
& \wedge \|\mathcal{B} \mathbf{v}_t\|_{C(0,T;L_{q+1}(\Omega_{nl}))}\leq \overline{M}\}\,.
\end{aligned}
\end{equation}
\end{theorem}
\begin{proof}
Recall that $\psi$ is related to $\mathbf{u}$ via $-\Delta \psi=-\text{div}\,\mathbf{u}$ and that we impose homogeneous Dirichlet boundary conditions on $\psi$.
Like in the proof of Theorem \ref{th:W3locex} we can use a fixed point argument with an operator $\mathcal{T}$ mapping $\mathbf{v}$ to a weak solution of
\begin{equation}\label{Welastic_viscosity_fp}
\begin{cases}
\varrho(x) \mathbf{u}_{tt}-\mathcal{B}^T\alpha(t,x) [c](x)\mathcal{B}\mathbf{u}\\
\qquad \qquad +\mathcal{B}^T\Bigl(((1-\delta(x)) +\delta(x)|\mathcal{B} \mathbf{u}_t|^{q-1})[b](x)\mathcal{B} \mathbf{u}_t\Bigr)
=0,
\\
(\mathbf{u},\mathbf{u}_t)|_{t=0}=(u_0, u_1),
\\
\mathbf{u}|_{\partial \Omega} =0,
\\
\end{cases}
\end{equation}
where
\[
\alpha(t,x)=\frac{1}{1+2\tilde{k}(x)(-\Delta)^{-1}\text{div}\,\mathbf{v}_t(t,x)}\,,
\]
$-\Delta$ denotes the Laplace operator on $\Omega$ with homogeneous Dirichlet boundary conditions on $\partial \Omega$, and
$C_{\Delta,\Omega_{nl}}$ is the constant in the regularity estimate
\[
\|(-\Delta)^{-1} f\|_{H^2(\Omega_{nl})}\leq C_{\Delta,\Omega_{nl}} \|f\|_{L_2(\Omega)}\,.
\]
Existence of a weak solution for fixed $\mathbf{v}\in\mathcal{W}$ can be shown like in Proposition \ref{prop:Wpot_viscosity_lin}.
In particular by choosing $\overline{a}$ sufficiently small so that
\[
2\|\tilde{k}\|_{L_\infty(\Omega_{nl})} C_{\Delta,\Omega_{nl}} C_{H^2(\Omega_{nl}),L_\infty(\Omega_{nl})}\overline{a}
< 1
\]
we have that
\[
\begin{aligned}
\alpha(t,x)
\geq&\frac{1}{1+ 2\|\tilde{k}\|_{L_\infty(\Omega_{nl})} C_{\Delta,\Omega_{nl}} C_{H^2(\Omega_{nl}),L_\infty(\Omega_{nl})}\overline{a}}=:\underline{\alpha}>0\\
\alpha(t,x)
\leq&\frac{1}{1- 2\|\tilde{k}\|_{L_\infty(\Omega_{nl})} C_{\Delta,\Omega_{nl}} C_{H^2(\Omega_{nl}),L_\infty(\Omega_{nl})}\overline{a}}=:\overline{\alpha}<\infty\\
|\alpha_t(t,x)|=&
\left|\frac{2\tilde{k}(x)}{(1+2\tilde{k}(x)(-\Delta)^{-1}\text{div}\,\mathbf{v}_t(t,x))^2}
(-\Delta)^{-1}\text{div}\,\mathbf{v}_{tt}(t,x)\right|\\
\leq& \frac{2\|\tilde{k}\|_{L_\infty(\Omega_{nl})}|(-\Delta)^{-1}\text{div}\,\mathbf{v}_{tt}(t,x)|
}{1- 2\|\tilde{k}\|_{L_\infty(\Omega_{nl})} C_{\Delta,\Omega_{nl}} C_{H^2(\Omega_{nl}),L_\infty(\Omega_{nl})}\overline{a}} \quad \mbox{ on }\Omega_{nl}\\
\alpha_t(t,x)=&\ 0 \mbox{ on }\Omega\setminus\Omega_{nl}\\
\|\alpha_t\|_{L^p(0,T;L_2(\Omega))}\leq&
\frac{2\|\tilde{k}\|_{L_\infty(\Omega_{nl})} \|\nabla (-\Delta)^{-1}\text{div}\,\mathbf{v}_{tt}\|_{L^p(0,T;L_2(\Omega))}
}{1-2\|\tilde{k}\|_{L_\infty(\Omega_{nl})} C_{\Delta,\Omega_{nl}} C_{H^2(\Omega_{nl}),L_\infty(\Omega_{nl})}\overline{a}} \\
\leq&
\frac{2\|\tilde{k}\|_{L_\infty(\Omega_{nl})} T^{\frac{2-p}{2p}}\overline{m}
}{1-2\|\tilde{k}\|_{L_\infty(\Omega_{nl})} C_{\Delta,\Omega_{nl}} C_{H^2(\Omega_{nl}),L_\infty(\Omega_{nl})}\overline{a}} \\
=:&\, \overline{\overline{\alpha}} T^{\frac{2-p}{2p}} \overline{m}\,, \quad p\in[1,2]
\end{aligned}
\]
for $\mathbf{v}\in\mathcal{W}$, provided $\overline{a}$ is chosen sufficiently small.
By multiplying \eqref{Welastic_viscosity_fp} with
$\mathbf{u}_t+\varepsilon\mathbf{u}_{tt}$
we obtain
\begin{align*}
&\tfrac12\left[|\sqrt{\varrho}\mathbf{u}_t|_{L_2(\Omega)}^2\right]_0^t
+\tfrac12\left[|\sqrt{\alpha [c]}\mathcal{B}\mathbf{u}|_{L_2(\Omega)}^2 \right]_0^t +\varepsilon\int_0^t|\sqrt{\varrho}\mathbf{u}_{tt}|_{L_2(\Omega)}^2\, ds\\
&\quad+\int_0^t \Bigl(|\sqrt{(1-\delta)[b]}\mathcal{B}\mathbf{u}_t|_{L_2(\Omega)}^2
+|\sqrt{\delta[b]}\mathcal{B}\mathbf{u}_t|_{L_{q+1}(\Omega_{nl})}^{q+1}\Bigr)\, ds\\
&\quad
+\varepsilon\left[\tfrac12|\sqrt{(1-\delta)[b]}\mathcal{B}\mathbf{u}_t|_{L_2(\Omega)}^2
+\tfrac{1}{q+1}|\sqrt[q+1]{\delta[b]}\mathcal{B}\mathbf{u}_t|_{L_{q+1}(\Omega_{nl})}^{q+1}\right]_0^t\\
&=
\tfrac12\int_0^t\int_{\Omega_{nl}}\Bigl(\alpha_t[c]\mathcal{B}\mathbf{u}\Bigr)^T\mathcal{B}\mathbf{u}\, dx\, ds\\
&\quad+\varepsilon\int_0^t |\sqrt{\alpha [c]}\mathcal{B}\mathbf{u}_t(s)|_{L_2(\Omega)}^2\, ds
+\varepsilon\int_0^t\int_{\Omega_{nl}}\Bigl(\alpha_t[c]\mathcal{B}\mathbf{u}\Bigr)^T\mathcal{B}\mathbf{u}_t\, dx\, .ds\\
&\quad-\varepsilon\left[ \int_\Omega \Bigl(\alpha[c]\mathcal{B}\mathbf{u}\Bigr)^T\mathcal{B}\mathbf{u}_t\, dx\right]_0^t.
\end{align*}
We estimate the right-hand side and arrive at
\begin{align*}
&\tfrac12\left[|\sqrt{\varrho}\mathbf{u}_t|_{L_2(\Omega)}^2\right]_0^t
+\tfrac12\left[|\sqrt{\alpha [c]}\mathcal{B}\mathbf{u}|_{L_2(\Omega)}^2 \right]_0^t+\varepsilon\int_0^t|\sqrt{\varrho}\mathbf{u}_{tt}|_{L_2(\Omega)}^2\, ds \\
&\quad+\int_0^t \Bigl(|\sqrt{(1-\delta)[b]}\mathcal{B}\mathbf{u}_t|_{L_2(\Omega)}^2
+|\sqrt{\delta[b]}\mathcal{B}\mathbf{u}_t|_{L_{q+1}(\Omega_{nl})}^{q+1}\Bigr)\, ds\\
&\quad
+\epsilon\left[\tfrac12|\sqrt{(1-\delta)[b]}\mathcal{B}\mathbf{u}_t|_{L_2(\Omega)}^2
+\tfrac{1}{q+1}|\sqrt[q+1]{\delta[b]}\mathcal{B}\mathbf{u}_t|_{L_{q+1}(\Omega_{nl})}^{q+1}\right]_0^t\\
&\leq
\tfrac12\overline{\overline{\alpha}} t^{\frac12} \overline{m}\overline{c}\int_0^t |\mathcal{B}\mathbf{u}(s)|_{L_4(\Omega_{nl})}^2 ds\\
&\quad+\varepsilon\overline{\alpha}\overline{c}\int_0^t |\mathcal{B}\mathbf{u}_t(s)|_{L_2(\Omega)}^2\, ds
+\varepsilon \overline{\overline{\alpha}} \overline{m}\overline{c}\int_0^t |\mathcal{B}\mathbf{u}(s)|_{L_4(\Omega_{nl})} |\mathcal{B}\mathbf{u}_t(s)|_{L_4(\Omega_{nl})} ds\\
&\quad+\varepsilon\overline{\alpha}\overline{c}\Bigl(|\mathcal{B}\mathbf{u}(t)|_{L_2(\Omega)} |\mathcal{B}\mathbf{u}_t(t)|_{L_2(\Omega)}
+|\mathcal{B}\mathbf{u}_0|_{L_2(\Omega)} |\mathcal{B}\mathbf{u}_1|_{L_2(\Omega)} \Bigr)\\
&\leq
\overline{\overline{\alpha}}\overline{m}\overline{c} \Bigl(t^{\frac32}|\mathcal{B}\mathbf{u}_0|_{L_4(\Omega_{nl})}+t^3\|\mathcal{B}\mathbf{u}_t\|_{L_4(0,T;L_4(\Omega_{nl}))}^2\Bigr)\\
&\quad+\varepsilon\overline{\alpha}\overline{c}\int_0^t |\mathcal{B}\mathbf{u}_t(s)|_{L_2(\Omega)}^2\, ds
+\varepsilon \overline{\overline{\alpha}}\overline{m}\overline{c}
\Bigl( \tfrac12 |\mathcal{B}\mathbf{u}_0|_{L_4(\Omega_{nl})}+\tfrac32 t^{\frac32}\|\mathcal{B}\mathbf{u}_t\|_{L_4(0,T;L_4(\Omega_{nl}))}^2\Bigr)\\
&\quad+\varepsilon\tfrac{\gamma}{4} |\mathcal{B}\mathbf{u}_t(t)|_{L_2(\Omega)}^2
+\varepsilon\tfrac{(\overline{\alpha}\overline{c})^2}{\gamma} |\mathcal{B}\mathbf{u}(t)|_{L_2(\Omega)}^2
+\varepsilon\overline{\alpha}\overline{c}|\mathcal{B}\mathbf{u}_0|_{L_2(\Omega)} |\mathcal{B}\mathbf{u}_1|_{L_2(\Omega)},
\end{align*}
thus by \eqref{condcoeffelast} using \eqref{abeps}, \eqref{Cepsr} in
\begin{align*}
&\overline{\overline{\alpha}}\overline{m}\overline{c} (t^3+\varepsilon \tfrac32 t^{\frac32}) \|\mathcal{B}\mathbf{u}_t\|_{L_4(0,T;L_4(\Omega_{nl}))}^2\\
&\leq \tfrac{\underline{\delta}\underline{b}}{4} \|\mathcal{B}\mathbf{u}_t\|_{L_{q+1}(0,T;L_{q+1}(\Omega_{nl}))}^{q+1}
+C(\tfrac{\underline{\delta}\underline{b}}{4},\tfrac{q+1}{2})
\Bigl(\overline{\overline{\alpha}}\overline{m}\overline{c} (t^3+\varepsilon \tfrac32 t^{\frac32}) t^{\frac{q-3}{2(q+1)}}({C_{L_{q+1},L_4}^\Omega})^2\Bigr)^{\frac{q+1}{q-1}}
\end{align*}
we get
\begin{equation}\label{enest_Welastic_viscosity}
\begin{aligned}
&\tfrac14\underline{\varrho}\|\mathbf{u}_t\|_{C(0,T;L_2(\Omega))}^2
+(\tfrac14\underline{\alpha}\underline{c}-\tfrac{(\varepsilon\overline{\alpha}\overline{c})^2}{\underline{\gamma}})\|\mathcal{B}\mathbf{u}\|_{C(0,T;L_2(\Omega))}^2 +\tfrac{\varepsilon}{2}\underline{\varrho}\|\mathbf{u}_{tt}|_{L_2(0,T;L_2(\Omega))}^2\\
&\quad+(\tfrac12\underline{\gamma}-\varepsilon\overline{\alpha}\overline{c})\|\mathcal{B}\mathbf{u}_t\|_{L^2(0,T;L_2(\Omega))}^2
+\tfrac{\underline{\delta}\underline{b}}{4}\|\mathcal{B}\mathbf{u}_t\|_{L_{q+1}(0,T;L_{q+1}(\Omega_{nl}))}^{q+1}\\
&\quad
+\tfrac{\varepsilon}{2}\Bigl(\tfrac12\underline{\gamma}\|\mathcal{B}\mathbf{u}_t\|_{C(0,T;L_2(\Omega))}^2
+\tfrac{1}{q+1}\underline{\delta}\underline{b}\|\mathcal{B}\mathbf{u}_t\|_{C(0,T;L_{q+1}(\Omega_{nl}))}^{q+1}\Bigr)\\
&\leq\varepsilon\overline{\alpha}\overline{c}|\mathcal{B}\mathbf{u}_0|_{L_2(\Omega)} |\mathcal{B}\mathbf{u}_1|_{L_2(\Omega)}+\varepsilon\overline{\alpha}\overline{c}|\mathcal{B}\mathbf{u}_0|_{L_2(\Omega)} |\mathcal{B}\mathbf{u}_1|_{L_2(\Omega)} \\
&\quad +\overline{\varrho} |\mathbf{u}_1|_{L_2(\Omega)}^2 +\overline{\alpha} |\mathcal{B}\mathbf{u}_0|_{L_2(\Omega)}^2
+\varepsilon\Bigl(\overline{b}|\mathcal{B}\mathbf{u}_1|_{L_2(\Omega)}^2
+\tfrac{2}{q+1}\overline{b}|\mathcal{B}\mathbf{u}_1|_{L_{q+1}(\Omega_{nl})}^{q+1}\Bigr)\\
&\quad+C(\tfrac{\underline{\delta}\underline{b}}{4},\tfrac{q+1}{2})
\Bigl(\overline{\overline{\alpha}}\overline{m}\overline{c} (t^3+\varepsilon \tfrac32 t^{\frac32}) t^{\frac{q-3}{2(q+1)}}({C_{L_{q+1},L_4}^\Omega})^2\Bigr)^{\frac{q+1}{q-1}}
\end{aligned}
\end{equation}
Thus by choosing $\varepsilon,\kappa$ sufficiently small, we obtain $\mathbf{u}\in\mathcal{W}$ .
\end{proof}
Due to the appearance of positive powers of $t$ in the energy estimate \eqref{enest_Welastic_viscosity}, also here only a local in time well-posedness result can be expected. Morever, for similar reasons as in Section \ref{secWpot_viscosity} uniqueness is not likely to hold here, see Remark \ref{rem:Wpot_viscosity_uniqueness}.
\begin{remark}\label{rem:interfacecond_ae}
Again we can derive an interface condition in the case that the coefficients $\lambda, \mu, k, \varrho, b, \delta$ are piecewise
constant, i.e., in a setting similar to the one of Remark \ref{rem:interfacecond}. This yields
\begin{eqnarray*}
&&\mbox{tr}_\Gamma (\mathbf{u}\vert_{\Omega_I})=\mbox{tr}_\Gamma (\mathbf{u}\vert_{\Omega_{II}})\\
&&\mbox{ i.e., continuity of velocity in a trace sense}\\
&&\int_\Gamma \Bigl\{\sigma_u\vert_{\Omega_I}-\sigma_u\vert_{\Omega_{II}}
\Bigr\} \mathbf{\nu} \, \mathbf{v} \,ds=0 \quad \forall \mathbf{v}\in (H^1(\Omega))^d\\
&& \mbox{where }\sigma_u:= [c] \mathcal{B} \mathbf{u}
+ ((1-\delta(x)) +\delta(x)|\mathcal{B} \mathbf{u}_t|^{q-1})[b](x)\mathcal{B} \mathbf{u}_t\,,\\
&&\mbox{ i.e., continuity of (modified) normal stresses in a variational sense.}
\end{eqnarray*}
\end{remark}
\section{The Westervelt equation in acoustic pressure formulation with $p$-Laplace damping \eqref{Wpress_pLaplace}}
\label{secWpress_pLaplace}
Similarly to Section \ref{secWpress_viscosity}, we start with a result on the equation
\begin{equation}\label{W1lin}
\begin{cases}
(1+\alpha) u_{tt}-c^2\,\text{div}\Bigl(\nabla u+\varepsilon|\nabla u|^{p-1}\nabla u\Bigr)-b\Delta u_t+fu_t=g
\\
(u,u_t)|_{t=0}=(u_0, u_1)
\\
u|_{\partial \Omega} =0
.\\
\end{cases}
\end{equation}
\begin{proposition} \label{prop:W1lin}
Let $T>0$, $c^2,\varepsilon>0$, $b\geq0$ and assume that
\begin{itemize}
\item
$\alpha\in C(0,T;L_\infty(\Omega))$, $\alpha_t\in L_\infty(0,T;L_2(\Omega))$,
$-1<-\underline{\alpha}\leq \alpha(t,x)\leq \overline{\alpha}$,
\item
$f\in L_\infty(0,T;L_{2}(\Omega))$,
\item
$g\in L_2(0,T;L_{4/3}(\Omega))$,
\item
$u_0\in H_0^1(\Omega)$, $u_1\in L_2(\Omega)$.
\end{itemize}
with
$$\|f-\frac12\alpha_t\|_{L_\infty(0,T;L_{2}(\Omega))}\leq\overline{b}
<\frac{b}{(C_{H_0^1,L_4}^\Omega)^2}\,. $$
Then \eqref{W1lin} has a weak solution
\begin{align*}
u\in\tilde{X} := ~&C^1(0,T;L_2(\Omega))\, \cap \, C(0,T;H_0^1(\Omega)) \\
&\cap C(0,T;W_0^{1,p+1}(\Omega))\, \cap \, H^1(0,T;H_0^1(\Omega))
\end{align*}
which is unique in $\tilde{X}$ and satisfies the energy estimate
\begin{equation}
\label{W1estimate}
\begin{aligned}
&\left[\frac{1}{2}\int_\Omega (1+\alpha)(u_{t})^2\, dx
+ \frac{c^2}{2} |\nabla u|_{L_2(\Omega)}^2+ \frac{c^2 \varepsilon}{p+1} |\nabla u|_{L_{p+1}(\Omega)}^{p+1} \right]_0^t\\
&\qquad+ \hat{b} \int_0^t |\nabla u_{t}|_{L_2(\Omega)}^2 \, ds \leq \frac{1}{4\check{b}} \|g\|_{L_2(0,T;L_{4/3}(\Omega))}^2.
\end{aligned}
\end{equation}
where $\hat{b}=b-(\overline{b}+\check{b})(C_{H^1,L^4}^\Omega)^2$ and $\check{b}= \frac{b}{(C_{H^1,L^4}^\Omega)^2}- \overline{b}$.\\
In particular, for the energy
\begin{equation}
\label{energy:W1}
\mathcal{E}[u](t):= \left[ |u_t|_{L_2(\Omega)}^2 + |\nabla u|_{L_2(\Omega)}^2 + |\nabla u |_{L_{p+1}(\Omega)}^{p+1}\right](t)
\end{equation}
we have
\begin{equation}\label{enest:W1}
\mathcal{E}[u](t) + c_1 \int_0^t |\nabla u_t|_{L_2(\Omega)}^2 \, ds \leq C_2 \left(\mathcal{E}[u](0) + \|g\|_{L_2(0,T;L_{4/3}(\Omega))}\right)
\end{equation}
for some constants $c_1$, $C_2$ only depending on $\alpha$, $c$, $b$, $\varepsilon$, $p$ and $C_{H^1,L^4}^\Omega$.
\end{proposition}
\begin{proof}
The proof can be done analogously to the one of Proposition \ref{prop:W2lin}
by first showing that \eqref{W1linweak} below holds for smooth approximations $\alpha_k$, $f_k$
and $g_k$ of $\alpha$, $f$ and $g$ by means of Galerkin's method, energy estimates and weak limits
and then letting $\alpha_k\to \alpha$, $f_k \to f$ and $g_k \to g$, respectively. Here, we omit details and only
show the key energy estimates. The weak form of \eqref{W1lin} reads as
\begin{equation}
\label{W1linweak}
\begin{aligned}
\int_\Omega \Bigl\{(1+\alpha)u_{tt} w + c^2 \left( \nabla u + \varepsilon |\nabla u|^{p-1} \nabla u\right) \nabla w
+b \nabla u_t \nabla w \Bigr\} \, dx&
\\
=\int_\Omega (g -fu_t) w\, dx\quad \forall w\in W_0^{1,p+1}(\Omega),&
\end{aligned}
\end{equation}
with initial conditions $(u_{0},u_{1})$. \\
Testing \eqref{W1linweak} with $w=u_t(t)$ and integrating with respect to time yields
\begin{eqnarray*}
&\lefteqn{\left[\frac{1}{2}\int_\Omega (1+\alpha)\left(u_{t}\right)^2\, dx
+ \frac{c^2}{2} |\nabla u|_{L_2(\Omega)}^2\right.}\\
&&\left.+ \frac{c^2 \varepsilon}{p+1} |\nabla u|_{L_{p+1}(\Omega)}^{p+1} \right]_0^t + b \int_0^t |\nabla u_{t}|_{L_2(\Omega)}^2 \, ds\\
&=&
- \int_0^t \int_\Omega \left(f-\frac12\alpha_{t}\right) (u_{t})^2 \, dx \, ds
+ \int_0^t \int_\Omega g u_{t} \, dx \, ds\\
&\leq&
(\overline{b}+ \check{b})\int_0^t | u_{t}|_{L_4(\Omega)}^2 \, ds +\frac{1}{4\check{b}} \|g\|_{L_2(0,T;L_{4/3}(\Omega))}^2,
\end{eqnarray*}
hence, by $\hat{b}=b-(\overline{b}+\check{b})(C_{H^1,L^4}^\Omega)^2>0$, we arrive at \eqref{W1estimate},
which directly implies \eqref{enest:W1}.
\end{proof}
\begin{remark}
Note that in the case of $p$-Laplace damping we do not obtain a higher order energy estimate
like in the proof of Proposition \ref{prop:W2lin}. Testing \eqref{W1linweak}
with $w=u_{tt}$ yields
\begin{align*}
\lefteqn{\int_0^t\int_\Omega \Bigl\{(1+\alpha)(u_{tt})^2 \, dx\, ds + \frac{b}{2} |\nabla u_{n,t}|^2 \big|_0^t}\nonumber\\
&=c^2 \int_0^t |\nabla u_{t}|_{L_2(\Omega)}^2 \, ds - c^2\left[ \int_\Omega \nabla u \nabla u_{t} \, dx\right]_0^t \nonumber\\
& \qquad- c^2 \varepsilon \int_0^t \int_\Omega |\nabla u|^{p-1} \nabla u \nabla u_{tt} \, dx \, ds - \int_0^t \int_\Omega (f u_{t}- g)u_{tt}\, dx \, ds,\nonumber
\end{align*}
but here the first term in the third line cannot be controlled.
\end{remark}
Using Proposition \ref{prop:W1lin}
we now show local existence of solutions.
\begin{theorem}\label{th:W1locex}
Let $c^2,b,\varepsilon>0$, $k\in\mathbb{R}$, $p>d-1$.\\
For any $T>0$ there is a sufficiently small
$\kappa_T>0$ such that for all $u_0 \in W_0^{1,p+1}(\Omega)$, $u_1\in L_2(\Omega)$ with
\begin{equation}\label{W1initialbound}
\mathcal{E}[u](0) = |u_1|_{L_2(\Omega)}^2+ |\nabla u_0|_{L_2(\Omega)}^2 + |\nabla u_0 |_{L_{p+1}(\Omega)}^{p+1} \leq \kappa_T^2
\end{equation}
there exists a weak solution $u\in \mathcal{W} $ of \eqref{Wpress_pLaplace} where
\begin{align}\label{defcWplaplace}
\mathcal{W} =\{v\in \tilde{X}
:~& \|v_{t}\|_{L_\infty(0,T;L_2(\Omega))}\leq \overline{m}\nonumber\\
& \wedge \|\nabla v_t\|_{L_2(0,T;L_2(\Omega))}\leq \overline{m}\nonumber\\
& \wedge \|\nabla v\|_{L_{\infty}(0,T;L_{p+1}(\Omega))}\leq \overline{M}\}
\end{align}
where $\overline{m}$ and $\overline{M}$ are sufficiently small and $2 k C_{W_0^{1,p+1}, L_\infty} \overline{M}<1$ is satisfied.
\end{theorem}
\begin{proof}
We define the fixed point operator $\mathcal{T}: \mathcal{W} \to \tilde{X}$, $v\mapsto \mathcal{T}v =u$ where $u$ is a solution of \eqref{W1lin} with
\begin{equation*}
\alpha=2k v\,, \quad f=2kv_t\,, \quad g=0\,,
\end{equation*}
which is well-defined by Proposition \ref{prop:W1lin}. First, note that
\begin{equation*}
\|f-\frac12\alpha_t\|_{L_\infty(0,T;L_{2}(\Omega)}\leq k \overline{m}
\end{equation*}
and, that by assumption we have $\alpha \in C(0,T; L_\infty(\Omega))$, $-1<-\underline{\alpha} \leq \alpha(t,x) \leq \overline{\alpha}$,
with $\underline{\alpha} = \overline{\alpha} = 2k C_{W_0^{1,p+1}, L_\infty} \overline{M}<1$. \\
Hence, we can make use of the energy estimate \eqref{W1estimate} to conclude
\begin{align*}
&\frac{1-\underline{\alpha}}{2} |u_t(t)|_{L_2(\Omega)}^2 + \frac{c^2}{2} |\nabla u(t)|_{L_2(\Omega)}^2
+ \frac{c^2 \varepsilon}{p+1} |\nabla u(t)|_{L_{p+1}(\Omega)}^{p+1} + \hat{b} \|\nabla u_t\|_{L_2(0,t;L_2(\Omega))} \\
&\leq \frac{1+\overline{\alpha}}{2} |u_1|_{L_2(\Omega)}^2 + \frac{c^2}{2} |\nabla u_0|_{L_2(\Omega)}^2
+ \frac{c^2 \varepsilon}{p+1} |\nabla u_0|_{L_{p+1}(\Omega)}^{p+1}
\end{align*}
which, together with the initial bound \eqref{W1initialbound}, leads to the estimates
\begin{equation*}
\begin{aligned}
\|u_t\|_{L_\infty(0,T;L_2(\Omega))}^2 \leq \frac{2}{1-\underline{\alpha}} L\, \kappa_T^2,\\
\|\nabla u_t\|_{L_2(0,T;L_2(\Omega))}^2 \leq \frac{1}{\hat{b}} L\, \kappa_T^2,\\
\|\nabla u\|_{L_\infty(0,T;L_{p+1}(\Omega))}^{p+1} \leq \frac{p+1}{c^2 \varepsilon} L\, \kappa_T^2,
\end{aligned}
\end{equation*}
where $L:=\frac{1+\overline{\alpha}}{2} + \frac{c^2}{2} + \frac{c^2 \varepsilon}{p+1}$.
Choosing $\kappa_T$ so small that
\begin{equation*}
\frac{2}{1-\underline{\alpha}} L\, \kappa_T^2 \leq \overline{m}^2, \quad
\frac{1}{\hat{b}} L\, \kappa_T^2\leq \overline{m}^2, \quad
\frac{p+1}{c^2 \varepsilon} L\, \kappa_T^2 \leq \overline{M}^{p+1}
\end{equation*}
implies $\|u_t\|_{L_\infty(0,T;L_2(\Omega))} \leq \overline{m}$,
$\|\nabla u_t\|_{L_2(0,T;L_2(\Omega))} \leq \overline{m}$,
$\|\nabla u\|_{L_\infty(0,T;L_{p+1}(\Omega))} \leq \overline{M}$
and thus $u\in\mathcal{W}$ which proves that $\mathcal{T}: \mathcal{W} \to \mathcal{W}$ is a self-mapping. The closedness of $\mathcal{W}$ is trivial.
Existence of solutions can be obtained by a compactness argument: Since $\mathcal{W}$ is bounded in the dual of a
separable Banach space, it is w$\ast$-compact. Hence the sequence of fixed point iterates $u^n$ defined by
$\mathcal{T}u^n = \mathcal{T} u^{n-1}$ where $u_0$ is chosen compatible with initial and boundary conditions has a
w$\ast$-convergent subsequence whose w$\ast$-limit $\overline{u}$ lies in $\mathcal{W}$. Furthermore, as
by
\begin{equation*}
(1-2ku^{n-1})u^n_{tt}
-c^2\,\text{div}\Bigl(\nabla u^n+\varepsilon|\nabla u^n|^{p-1}\nabla u^n\Bigr)-b\Delta u^n_t=2k u^{n-1}_t u^n_t
\end{equation*}
and integration by parts with respect to time we have
\begin{align*}
&\int_0^T \int_\Omega \big\{ -\overline{u}_t (( 1-2k \overline{u}) \phi )_t + [ c^2 ( \nabla \overline{u} + \varepsilon | \nabla \overline{u} |^{p-1} \nabla \overline{u})
+ b \nabla \overline{u}_t ] \nabla \phi - 2k(\overline{u}_t )^2 \phi \big\} dx \, ds \\
&= \int_0^T \int_\Omega \Big\{ - (\overline{u}- u^n)_t ((1-2k\overline{u})\phi)_t + 2k u_t^n ((\overline{u}- u^{n-1} )\phi)_t \\
&- 2k(\overline{u}_t-u_t^n)\overline{u}_t\phi - 2k(\overline{u}_t-u_t^{n-1})u^n_t\phi + [c^2 \nabla(\overline{u}-u^n)+b\nabla (\overline{u}-u^n)_t]\nabla\phi\\
&+c^2 \varepsilon \int_0^1 \tilde{w}^\sigma [ |\nabla(u^n+\sigma \hat{u})|^2 \nabla \hat{u} +(p-1)(\nabla(u^n+\sigma \hat{u}) \nabla\hat{u})\nabla(u^n+\sigma\hat{u}) ] d\sigma \nabla\phi \Big\} \,dx\, ds\\
& \to 0 \text{ as } k \rightarrow \infty
\end{align*}
for any $\phi \in C_0^\infty((0,T)\times \Omega)$ where $\tilde{w}^\sigma(x,t)=|\nabla(u^n+\sigma\hat{u})|^{p-3}$, $\hat{u}=\overline{u}-u^n$, the w$\ast$-limit $\overline{u}$ satisfies the PDE in a weak sense which completes
the proof of existence.
\end{proof}
\begin{remark}
In Theorem \ref{th:W1locex}
but contractivity of $\mathcal{T}$ could not be proved, similarly to Sections \ref{secWpot_viscosity} and \ref{secWelastic_viscosity}; see also Remark \ref{rem:Wpot_viscosity_uniqueness}.
Suppose $v^i \in\mathcal{W}$, $u^i = \mathcal{T}v^i$, $i=1,2$ solve \eqref{Wpress_pLaplace}. Then, similarly to \eqref{W2_diff} we obtain
\begin{equation}\label{W1_diff}
\begin{cases}
(1-2kv^1)\hat{u}_{tt}-c^2\Delta \hat{u}
-c^2 \varepsilon \int_0^1\,\text{div}\Bigl( \tilde{w}^\sigma
\bigl[|\nabla (u^2+\sigma \hat{u})|^2 \nabla \hat{u}\\
\quad+(p-1)(\nabla (u^2+\sigma \hat{u}) \nabla \hat{u})
\nabla(u^2+\sigma \hat{u}) \bigr]\, \Bigr)d\sigma\\
\quad-b\Delta \hat{u}_t -2k v^1_t \hat{u}_t \ = \ 2k(\hat{v}u^2_{tt}+\hat{v}_tu^2_t)
\\
(\hat{u},\hat{u}_t)|_{t=0}=(0,0)
\\
\hat{u}|_{\partial \Omega} =0
.\\
\end{cases}
\end{equation}
for $\hat{u}=u^1-u^2$, $\hat{v}=v^1-v^2$, with $\tilde{w}^\sigma(x,t)= |\nabla(u^2+\sigma \hat{u})(x,t)|^{p-3}$. Multiplication of
\eqref{W1_diff} with $\hat{u}$ as well with $\hat{u}_t$ leads to problems with the first or fourth term, respectively.
Therefore, uniqueness of solutions of solutions remains open.
\end{remark}
For this equation we even have global existence and exponential decay of the energy
\[
E[u](t)=\frac12|\sqrt{1-2ku}u_t|_{L_2(\Omega)}^2+\frac{c^2}{2}|\nabla u|_{L_2(\Omega)}^2+\frac{c^2\varepsilon}{p+1}|\nabla u|_{L_{p+1}(\Omega)}^{p+1}.
\]
\begin{theorem}\label{th:W1globex}
Let $c^2,b,\varepsilon>0$, $k\in\mathbb{R}$, $p>d-1$.\\
There exist $\kappa,\overline{\kappa}>0$ such that for all $(u_0,u_1)\in W_0^{1,p+1}(\Omega)\times L_2(\Omega)$ with
\[
E[u](0)=\frac12|\sqrt{1-2ku_0}u_1|_{L_2(\Omega)}^2+\frac{c^2}{2}|\nabla u_0|_{L_2(\Omega)}^2+\frac{c^2\varepsilon}{p+1}|\nabla u_0|_{L_{p+1}(\Omega)}^{p+1}\leq \kappa^2
\]
there exists a weak solution $u$ of \eqref{Wpress_pLaplace} for all times which satisfies
\[
E[u](t)\leq \overline{\kappa}^2.
\]
Moreover, there exists $\omega>0$ such that the exponential decay estimate
\[
E[u](t)\leq E[u](0)\exp(-\omega t)
\]
holds.
\end{theorem}
\begin{proof}
The crucial energy estimate is obtained by multiplication of \eqref{Wpress_pLaplace} with $u_t$ (using $(1-2ku)u_tu_{tt}=\frac12 \frac{d}{dt}[(1-2ku)(u_t)^2+k(u_t)^3$)
\begin{align*}
&\Bigl[\frac12|\sqrt{1-2ku}u_t|_{L_2(\Omega)}^2
+\frac{c^2}{2}|\nabla u|_{L_2(\Omega)}^2
+\frac{c^2\varepsilon}{p+1}|\nabla u|_{L_{p+1}(\Omega)}^{p+1}
\Bigr]_0^t
+b\int_0^t|\nabla u_t(s)|_{L_2(\Omega)}^2\,ds\\
&=k\int_0^t \int_\Omega (u_t(s))^3\,dx\,ds
\leq k\sup_{s\in(0,t)} |u_t(s)|_{L_2(\Omega)} \ \int_0^t |u_t|_{L_4(\Omega)}^2 \, ds\\
&\leq k {C_{H_0^1,L_4}^\Omega}^2 \sup_{s\in(0,t)} |u_t(s)|_{L_2(\Omega)} \ \int_0^t |\nabla u_t|_{L_2(\Omega)}^2 \, ds
\end{align*}
as long as $u$ exists and is pointwise bounded away from $\frac{1}{2k}$
\begin{equation*}
-1 < -\underline{\alpha}\leq -2k u(s,x)\leq\overline{\alpha} \quad \forall s\in(0,t)\,, \ x\in\Omega
\end{equation*}
which is true for sufficiently short time according to Theorem \ref{th:W1locex}.
Hence, if
\[
|u_t(0)|_{L_2}^2\leq \frac{2}{1-\underline{\alpha}} E[u](0)\leq \frac{2}{1-\underline{\alpha}} \kappa^2< \Big(\frac{b}{k {(C_{H_0^1,L_4}^\Omega)}^2}\Big)^2,
\]
the energy decreases monotonically with time,
\begin{equation}\label{enestW1_1}
E[u](t)+\tilde{b}\int_0^t|\nabla u_t(s)|_{L_2(\Omega)}^2\, ds\leq E[u](0)
\end{equation}
with $\tilde{b}=b-\sqrt{\frac{2}{1-\underline{\alpha}}}\kappa k ({C_{H_0^1,L_4}^\Omega})^2 >0$ and global existence can be concluded.
Equipartition of energy, which we get by multiplication of \eqref{Wpress_pLaplace} with $u$ and
integration over space and time, here reads as
\begin{eqnarray*}
\int_0^t\Bigl\{-|\sqrt{1-4ku}u_t|_{L_2(\Omega)}^2
+c^2|\nabla u|_{L_2(\Omega)}^2
+c^2\varepsilon|\nabla u|_{L_{p+1}(\Omega)}^{p+1}
\Bigr\}\,ds\\
+b\left[|\nabla u|_{L_2(\Omega)}^2\right]_0^t
=-\left[ \int_\Omega (1-2ku)u u_t\, dx\right]_0^t,
\end{eqnarray*}
hence
\begin{equation}
\label{equipartW1}
\begin{aligned}
&\int_0^t|u_t(s)|_{L_2(\Omega)}^2\, ds
\geq \frac{1}{1+2\overline{\alpha}}\int_0^t\Bigl\{
c^2|\nabla u|_{L_2(\Omega)}^2
+c^2\varepsilon|\nabla u|_{L_{p+1}(\Omega)}^{p+1}
\Bigr\}\,ds\\
&+\frac{b}{1+2\overline{\alpha}}\Bigl(|\nabla u(t)|_{L_2(\Omega)}^2-|\nabla u(0)|_{L_2(\Omega)}^2\Bigr)-\frac{1+\overline{\alpha}}{1+2\overline{\alpha}}\Bigl(C_{PF}^2|\nabla u(t)|_{L_2(\Omega)}^2\\
&+ |u_t(t)|_{L_2(\Omega)}^2+C_{PF}^2|\nabla u(0)|_{L_2(\Omega)}^2+ |u_t(0)|_{L_2(\Omega)}^2\Bigr).
\end{aligned}
\end{equation}
Thus we can split the term $b\int_0^t|\nabla u_t(s)|_{L_2(\Omega)}^2\,ds$ in \eqref{enestW1_1} as follows
\begin{align*}
\lefteqn{b\int_0^t|\nabla u_t(s)|_{L_2(\Omega)}^2\,ds}\\
&\geq \frac{b}{C_{PF}^2} \int_0^t\Bigl\{(1-\lambda)|u_t(s)|_{L_2(\Omega)}^2
+\frac{\lambda c^2}{1+2\overline{\alpha}}|\nabla u(s)|_{L_2(\Omega)}^2
+\frac{\lambda c^2\varepsilon}{1+2\overline{\alpha}}|\nabla u(s)|_{L_{p+1}(\Omega)}^{p+1}
\Bigr\}\,ds
\nonumber\\
&+\frac{\lambda b^2}{C_{PF}^2(1+2\overline{\alpha})}
\Bigl(|\nabla u(t)|_{L_2(\Omega)}^2-|\nabla u(0)|_{L_2(\Omega)}^2\Bigr)
\nonumber\\
&-\frac{\lambda b}{2C_{PF}^2}\frac{1+\overline{\alpha}}{1+2\overline{\alpha}}
\Bigl(C_{PF}^2|\nabla u(t)|_{L_2(\Omega)}^2+ |u_t(t)|_{L_2(\Omega)}^2
+C_{PF}^2|\nabla u(0)|_{L_2(\Omega)}^2+ |u_t(0)|_{L_2(\Omega)}^2\Bigr)
\end{align*}
which inserted into \eqref{enestW1_1} with $\lambda\in(0,1)$ sufficiently small implies
\[
E[u](t)+c\int_0^t E[u](s)\, ds\leq C E[u](0)
\]
for some constants $c,C>0$ and all $t>0$.
This relation by a standard argument implies exponential decay of $E[u](t)$.
\end{proof}
\section{Conclusions and Remarks}
The introduction of nonlinear strong damping in equations of nonlinear acoustics allows us to prove existence of solutions with weaker regularity; in particular, this enables us to show well-posedness of solutions to coupled acoustic-acoustic and acoustic-elastic problems.
However, the nonlinear strong damping also introduces additional challenges to the analysis: due to the relatively
high order of differentiation they contain, these terms only allow to derive energy estimates for certain low order multipliers.
For this reason, for some of the equations under consideration, uniqueness of solutions remains open.
The presence of the linear strong damping term $-\Delta u_t$ would seem to imply that the use of a
nonlinear strong damping $-\text{div}\,(|\nabla u_t|^{q-1}\nabla u_t)$ of viscosity type is more natural, it turns out that a $p$-Laplace damping
term $-\text{div}\,(|\nabla u|^{q-1}\nabla u)$ yields some nicer mathematical properties such as global in time existence and
exponential decay; uniqueness, however, remains an open problem for this formulation.
Here we have only investigated the $p$-Laplace damping for the acoustic pressure formulation;
we do expect that global existence and exponential decay will also carry over to the velocity potential formulation
\eqref{Wpot_viscosity} and to the coupled problems \eqref{Wacoustic_viscosity}, \eqref{Welastic_viscosity}
upon replacement of viscosity by $p$-Laplace damping. This setting together with a choice of different boundary conditions
(e.g., practically relevant Neumann as well as absorbing boundary conditions) will be subject of future research.
\bigskip
{\bf Acknowledgments.}
The authors gratefully acknowledge the referee's careful reading of the manuscript and many fruitful comments which led to an improved version of this paper.
R.B. and B.K. gratefully acknowledge financial support of their research by the FWF (Austrian Science Fund): P24970. The work of P.R. was supported by the NSF Grant DMS 0908435.
|
\section{Introduction}
{\bf Local search.} Local search techniques are popular heuristics for hard combinatorial optimization problems. Given a feasible solution, the algorithm repeatedly performs operations from the given class, each improving the cost of the current solution, until a solution is reached for which no operation yields an improvement (a locally optimal solution). Alternatively, we can view this as a neighborhood search process, where each
solution has an associated neighborhood of adjacent solutions, i.e., those that can be reached with a
single operation, and one moves to a better neighbor until none.
Such techniques are easy to implement, easy to parallelize, and fast and give good results.
One advantageous feature of local search algorithms is their flexibility; they can be applied to arbitrary cost functions, even in the presence of
additional constraints.
However, there has long been a gap between worst-case guarantees and real-world experience.
Thus, it is interesting to analyze such algorithms rigorously and, even in settings where alternative, theoretically optimal polynomial-time algorithms are known.
{\bf Problems studied.} We focus on Euclidean problems in the plane (the results extend to small dimensions), and study clustering and network connectivity type problems: the traveling
salesman problem (TSP), Steiner tree, facility location, and k-median.
The \textit{traveling salesman} problem is to connect $n$ input points with a tour of minimum total length. The \textit{Steiner tree} problem, given $n$ terminal points, is to choose additional {\em Steiner} points so as to minimize the length of the minimum tree spanning terminal and Steiner points. The \textit{facility location} problem, given $n$ client points and a facility opening cost $f$, chooses how many facilities to open and where to open them to minimize the combination of the cost of opening facilities and of the total distance from each client to the nearest open facility. The \textit{$k$-median} problem, given $n$ points and an integer $k$, chooses where to open $k$ facilities so as to minimize the total distance from each client to the nearest open facility.
{\bf Algorithms.} Our goal is to prove, under minimal assumptions, that local search finds solutions whose cost is within a $(1+\epsilon)$ factor of optimal. For that goal, local search must do a little more: instead of modifying the current solution by swapping a single point, edge or edge pair (depending on the problem) in and out of the solution, our version of local search swaps up to $1/\epsilon^c$ points, edges or edge pairs. This is a standard variation of local search (particularly for the traveling salesman tour), whereby each iteration is slowed down due to an increase in the size of the neighborhood, but the local optimum tends to be reached after fewer iterations and is of higher quality. Moreover, most implementations of local search do not continue iterating all the way to a local optimum, but stop once the gain obtained by each additional iteration is essentially negligible. Our algorithm thus has a stopping condition, when no local exchange could improve the cost by more than a factor of $1-1/n$. Then, the runtime is polynomial, at most $n^{1/\epsilon^{O(1)}}$.
{\bf Results.} Our results are as follows.
\begin{enumerate}
\item
For TSP, we assume that the input points are random uniform in $[0,1]^2$. Here
local search swaps $O(1/\epsilon^c)$ edges in the tour. Then local
search finds a solution with cost $(1+\mathcal{O}(\epsilon)) OPT$.
The proof is not difficult and serves as a warm-up to the later sections. The random input
assumption is necessary : in the worst-case setting, we give an
example where a locally optimal solution has cost more than
$(2-\epsilon)OPT$.
\item
Similarly, for Steiner tree, assuming random uniform input, again
local search finds a solution with cost $(1+\epsilon) OPT$.
\item
For facility location, we prove the following: consider the
version of local search where local moves consist of adding, deleting
or swapping $O(1/\epsilon^c)$ facilities. Then, even for worst case inputs, local search finds a
solution with cost $(1+\epsilon) OPT$. This is the core result of the paper.
We transform the dissection technique from Kolliopoulos and Rao~\cite{Kolliopoulos07} into a tool for analyzing local search.
\item
For $k$-median, our result is similar, except that local search uses
$(1+\epsilon)k$ medians instead of $k$, so that result is bicriteria. This is a technical, variant of the facility location result.
\end{enumerate}
{\bf Related work.}
\textbf{TSP and Steiner Tree.}
The TSP problem in the Euclidean plane has a long history, including work with local search~\cite{Croes58,Lin65,Lin73}. Most relevant is the work of Karp~\cite{Karp77} giving a simple construction of a near-optimal tour when points are drawn from a random distribution. That work has been subsumed by the approximation schemes of Arora~\cite{Arora96} (and its improvements~\cite{Arora97,Rao98}) and of Mitchell~\cite{Mitchell99}, using a hierarchical dissection technique. Arora noted the relation between that technique and local search, observing:
\begin{quote}
\textit{
Local-exchange algorithms for the TSP work by identifying possible edge exchanges in the current tour that lower the cost
[…]
Our dynamic programming algorithm can be restated as a slightly more inefficient backtracking […].
Thus it resembles $k$-OPT for $k=O(c)$, except that cost-increasing exchanges have to be allowed in order to undo bad guesses. Maybe it is closer in spirit to more ad-hoc heuristics such as genetic algorithms, which do allow cost-increasing exchanges.
}
\end{quote}
In fact, even with neighborhoods of size $f(\epsilon)$, even in the Euclidean plane, local search for TSP can get stuck in a local optimum whose value is far from the global optimum (See Fig. \ref{fig:lb_tsp}). However, in the case of random inputs the intuition is correct.
Local search algorithms have been widely studied for TSP, but mostly for either a local neighborhood limited to size of 2 or 3 (the 2-OPT or 3-OPT algorithms), or for the general metric case. Those studies lead to proofs of constant factor approximations, see~\cite{Chandra94,Johnson97,Mersmann12,Lin73,Rosenkrantz77}. In particular, in~\cite{Chandra94}, it is proved (by example) that for Euclidean TSP 2-OPT cannot be a constant-factor approximation in the worst case.
For the metric Steiner Tree problem, the best approximation algorithm up to 2010 was a constant factor approximation due to Robins and Zelikovsky and was by local search \cite{Robins00}.
\textbf{Facility Location and $k$-Median.}
For clustering problems — facility location and $k$-median — there has also been much prior work. A proof of NP-Hardness of $k$-median even in the Euclidean setting is given in~\cite{Megiddo84}.
The first theoretical guarantees for local search algorithms for
clustering problems are due to Korupolu et al. \cite{Korupolu00}. They show that the local search algorithm which allows swaps of size $p$ is a constant factor approximation for the metric case of the $k$-Median and Facility Location problems. However, for $k$-Median the algorithm requires a constant-factor
blowup in the parameter $k$.
By further refining the analysis, Charikar et al. \cite{Charikar05} improved the
approximation ratio.
More recently, Arya et al. showed in \cite{Arya04} that the local search
algorithm which allows swaps of size $p$ is a $3+2/p$-approximation
without any blowup in the number of medians.
Nevertheless, no better results were known for the Euclidean
case (See the survey paper \cite{Vygen05}).
Kolliopoulos and Rao define in \cite{Kolliopoulos07} a recursive ``adaptive'' dissection of a square enclosing the input points. At each dissection step
\footnote{There is also a ``sub-rectangle'' step not described here.},
they cut the longer side of each rectangle produced by the previous step in such a way that each of the two parts has roughly the same surface area.
Our analysis uses a new version of their dissection algorithm to analyze the local search algoritm.
\textbf{Other related work.}
The question of the efficiency of local search for Euclidean problems was already posed by Mustafa and Ray and Chan and Har-Peled. They proved that local search (with local neighborhood enabling moves of size $\Theta(1/\epsilon)$) gives approximation schemes for hitting circular disks in two dimensions with the fewest points, for several other Euclidean hitting set problems~\cite{Mustafa09}, and for independent sets of pseudo-disks~\cite{Chan09}.
This led to further PTASs by local search for dominating set in disks graph \cite{Gibson10} and for terrain guarding \cite{Krohn14}.
Those papers rely on the combinatorial properties of bipartite planar graphs.
Our analysis technique is different since we rely on dissections.
One problem related to facility location is $k$-means. For $k$-means, Kanungo, Mount, Netanyahu and Piatko~\cite{Kanungo04} proved that local search gives a constant factor approximation. Much remains to be understood.
We also note that there exists proofs of constant factor approximation by local search for the metric capacitated facility location \cite{Chudak05}.
\textbf{Plan.}
The paper is organized as follows: in the next section, as a warm-up we prove the results on TSP and Steiner tree for random inputs. We then analyze local search for facility location, proposing a new recursive dissection. We suitably extend lemmas from~\cite{Kolliopoulos07}. The meat of that section is the proof of Proposition
\ref{thm:new_struct}, which is our main technical contribution.
We end with the $k$-median result, that requires additional ideas to deal with the cardinality constraint.
\section{Polynomial-Time Local Search Algorithms}
Throughout this paper, we denote by $L \bigtriangleup L'$ the symmetric difference of the sets $L$ and $L'$.
We present the local search algorithm that is considered in this paper (see Algorithm \ref{algo:LS} below).
\begin{algorithm}
\caption{Local Search ($\varepsilon$)}
\label{algo:LS}
\begin{algorithmic}[1]
\State \textbf{Input:} A set $\mathcal{C}$ of points in the Euclidean plane
\State $S \gets$ Arbitrary feasible solution (of cost at most $\mathcal{O}(2^{n} \text{OPT})$).
\While{$\exists$ $S'$ s.t. Condition($S',\varepsilon$)
\textbf{and} cost($S'$) $\le$ $(1-1/n)$ cost($S$)\\}
\State $S \gets S'$
\EndWhile
\State \textbf{Output:} $S$
\end{algorithmic}
\end{algorithm}
Note that the type of $S$, Condition, $f(\varepsilon)$ and $\text{Cost}(S)$ are problem dependent. Namely,
\begin{itemize}
\item for Facility Location, $S$ is a set of points, Condition($S',\varepsilon$) is $ |S \bigtriangleup S'| = \mathcal{O}(1/\varepsilon^3)$ and
$\text{Cost}(S) = |S| + \sum\limits_{c \in \mathcal{C}} \min\limits_{s \in S} d(c,s)$;
\item for $k$-Median, $S$ is a set of points, Condition($S',\varepsilon$) is $|S \bigtriangleup S'| = \mathcal{O}(1/\varepsilon^9)$ and $|S'| \le (1+3\varepsilon)k$
and $\text{Cost}(S) = \sum\limits_{c \in \mathcal{C}} \min\limits_{s \in S} d(c,s)$;
\item for TSP $S$ is a set of edges, Condition($S',\varepsilon$) is $|S \bigtriangleup S'| = \mathcal{O}(1/\varepsilon^2)$ and ``$S'$ is a tour and there
is no two edges intersecting'' (if the
initial tour contains intersecting edges we start by modifying the tour so that no two edges intersect)
and $\text{Cost}(S) = \sum\limits_{s \in S} \text{length}(s)$;
\item for Steiner Tree, $S$ is a set of points, Condition($S',\varepsilon$) is $|S \bigtriangleup S'| = \mathcal{O}(1/\varepsilon^2)$ and
$|S'| \le n$ (if the initial set of Steiner vertices is greater than $n$, we greedily remove Steiner vertices until the set has size $n$)
and $\text{Cost}(S) = \text{MST}(S \cup \mathcal{C})$, where $\text{MST}(S \cup \mathcal{C})$ is the length of the minimum spanning tree of the points in
$S \cup \mathcal{C}$.
\end{itemize}
We now focus on the guarantees on the execution time of the algorithms presented in this paper.
The proof of the following Lemma is deferred to Appendix \ref{appx:ptime}.
\begin{lem}
\label{lem:polytime}
The number of iterations of Algorithm \ref{algo:LS} is polynomial for the Facility Location, $k$-Median, Traveling Salesman and Steiner Tree Problems.
\end{lem}
\begin{rem}
Up to discretizing the plane and replacing $(1-1/n)$ by $(1-\Theta(1/n))$, finding $S'$ takes time $\mathcal{O}(n^{\mathcal{O}(1/\varepsilon^c)}\varepsilon^{-1})$, for some constant $c$ which depends on the algorithm.
\end{rem}
\section{Euclidean Traveling Salesman Problem and Steiner Tree}
\begin{thm}\label{thm:ptas_expct}
Consider a set of points chosen independently and uniformly in $[0,1]^2$. Algorithm \ref{algo:LS} produces:
\begin{itemize}
\item In the case of the Traveling Salesman problem, a tour whose length is at most $(1 + \mathcal{O}(\varepsilon)) T_{\text{OPT}} $,
where $T_{\text{OPT}}$ is the length of the optimal solution.
\item In the case of the Steiner Tree problem, a tree whose length is at most $(1 + \mathcal{O}(\varepsilon)) T_{\text{OPT}} $,
where $T_{\text{OPT}}$ is the length of the optimal solution.
\end{itemize}
\end{thm}
To prove Theorem \ref{thm:ptas_expct}, we first prove the following result.
\begin{thm}\label{thm:ub_tsp_st}
Consider an arbitrary set of points in $[0,1]^2$. Algorithm \ref{algo:LS} produces:
\begin{itemize}
\item In the case of the Traveling Salesman problem, a tour whose length is at most $(1 + \mathcal{O}(\varepsilon^2))T_{\text{OPT}} + \mathcal{O}(\varepsilon \sqrt{n})$,
where $T_{\text{OPT}}$ is the length of the optimal solution.
\item In the case of the Steiner Tree problem, a tree whose length is at most $(1 + \mathcal{O}(\varepsilon^2)) T_{\text{OPT}} + \mathcal{O}(\varepsilon \sqrt{ n})$,
where $T_{\text{OPT}}$ is the length of the optimal solution.
\end{itemize}
\end{thm}
We model a random distribution of points in a region $\mathcal{P}$ of the plane by a
two-dimensional Poisson distribution $\Pi_n(\mathcal{P})$. The distribution $\Pi_n(\mathcal{P})$ is determined
by the following assumptions:
\begin{enumerate}
\item the numbers of points occurring in two or more disjoint sub-regions are
distributed independently of each other;
\item the expected number of points in a region $A$ is $n v(A)$ where $v(A)$ is the area
of $A$; and
\item as $v(A)$ tends to zero, the probability of more than one point occurring in $A$
tends to zero faster than $v(A)$.
\end{enumerate}
From these assumptions it follows that
Pr$[A \text{ contains exactly } m \text{ points}] = e^{- \lambda}\lambda^m/m!$, where $\lambda = nv(A)$.
The following result is known.
\begin{thm}
\label{thm:random_tsp}
\cite{Beardwood59}
Let $\mathcal{P}$ be a set of $n$ points distributed according to a two-dimensional Poisson distribution $\Pi_n(\mathcal{P})$ in $[0,1]^2$ and
let $T_n(\mathcal{P})$ be the random variable that denotes the length of the shortest tour
through the points in $\mathcal{P}$.
There exists a positive constant $\beta$ (independent of $\mathcal{P}$) such that
$T_n(\mathcal{P})/\sqrt{n} \rightarrow \beta$ with probability 1.
\end{thm}
Assuming Theorems \ref{thm:ub_tsp_st} and \ref{thm:random_tsp}, we can prove Theorem \ref{thm:ptas_expct}.
\begin{proof}[Proof of Theorem \ref{thm:ptas_expct}]
We focus on the Traveling Salesman case.
Let $L$ be the tour produced by Algorithm \ref{algo:LS} and $T_{\text{OPT}}$ be the optimal tour.
By Theorem \ref{thm:random_tsp}, we have that $\text{Cost}(T_{\text{OPT}}) = \mathcal{O}(\sqrt{n})$ with probability 1.
Hence, Theorem \ref{thm:ub_tsp_st} implies
$$(1-\varepsilon^2)\cdot \text{Cost}(L) \le \text{Cost}(T_{\text{OPT}}) + \mathcal{O}(\varepsilon\sqrt{n}) = (1 + \mathcal{O}(\varepsilon)) \cdot \text{Cost}(T_{\text{OPT}}).$$
We now consider the random variable $ST_n(\mathcal{P})$ that denotes the length of the shortest Steiner Tree through the points in $\mathcal{P}$.
Since the length of the optimal Steiner Tree is at least half the length of the optimal Traveling Salesman Tour,
Theorem \ref{thm:random_tsp} implies that there exists a constant $\delta$ such that $ST_n(\mathcal{P})/\sqrt{n} \ge \delta$ with probability 1.
Then, the exact same reasoning applies to prove the Steiner Tree case.
\end{proof}
The rest of the section is dedicated to the proof of Theorem \ref{thm:ub_tsp_st}.
To this aim,
we define a recursive dissection of the unit square according to a set of points $\mathcal{P}$.
At each step we cut the longer side of each rectangle produced by the previous step in such a way
that each of the two parts contains half the points of $\mathcal{P}$ that lie in the rectangle.
The process stops when each rectangle contains $\Theta (1/\varepsilon^2)$ points of $\mathcal{P}$.
We now consider the final rectangles and we refer to them as \emph{boxes}.
Let $\mathcal{B}$ be the set of boxes.
\begin{lem}
\label{lem:Karp77}
\cite{Karp77} $\sum\limits_{b \in \mathcal{B}} |\partial b| = \mathcal{O}(\varepsilon \sqrt{ |\mathcal{P}|})$, where $|\partial b|$ is the perimeter
of box $b$ and $|\mathcal{P}|$ is the number of points in $\mathcal{P}$.
\end{lem}
For any set of segments $S$ and box $b$ and for each segment $s$, let $s_b$ be the part of $s$ that lies inside $b$.
We define $\text{In}(S, b) := \{ s_b \mid s \in S \text{ and } s \text{ has at least one endpoint in }b\}$ and
$\text{Cross}(S,b) := \{s_b \mid s \in S \text{ and } s \text{ has no endpoint in }b \}$.
Moreover we define $\text{Out}(S,b) := \{ s_{b'} \mid s \in S \text{ and } b \neq b'\}$.
Additionally, let $S(b) = \sum_{s \in S } \text{length}(s_b)$.
We can now prove the two following structural Lemmas. See Fig. \ref{fig:lem_locality_TSP} for an illustration of the proof.
\begin{figure}
\begin{center}
\includegraphics{locality_tsp.pdf}
\caption{The solid black segments form the tour $L_\text{TSP}$ outside $b$. The dotted line segments are the tour $T_\text{TSP}$ inside $b$.
The red segments are the one needed to connect the two tours.}
\label{fig:lem_locality_TSP}
\end{center}
\end{figure}
\begin{lem}\label{lem:local_opt_st}
Let $L_{\text{ST}}$ be a locally optimal solution to the Steiner Tree problem and let $T_{\text{ST}}$ be any Steiner Tree.
Let $\mathcal{B}$ be a set of boxes produced by a dissection of $\mathcal{P} \cup L_\text{ST} \cup T_\text{ST}$.
Using the same notation for a set of segments and their total length, we then have for any box $b \in \mathcal{B}$
$$(1-\mathcal{O}(\varepsilon^2)) L_\text{ST}(b) \le \text{In}(T_\text{ST},b) + |\partial b| + L_\text{ST}/n,$$
where $|\partial b|$ is the perimeter of $b$.
\end{lem}
\begin{proof}
For each box $b$, the segments of $\text{Cross}(L_\text{ST},b)$ can be distributed into
6 different classes according to which side of $b$ they intersect.
We divide further. Since the segments of a class are pairwise disjoint,
there is a natural ordering of the segments inside each class.
For each class that contains more than $1/\varepsilon^2$ segments, we partition them into
subsets that contain $\Theta(1/\varepsilon^2)$ consecutive segments
(in the natural order of the class).
We define a sub-box for each subset of each class as follows. Let $s$ and $s'$ be the two extreme segments of the set in the ordering of the class.
The sides of the sub-box associated to
this subset consists of $s$ and $s'$ and the two shortest paths $p,p'$ along the sides of $b$ that connects the endpoints
of $s$ and $s'$.
Remark that the sum of the lengths of the sides of all the sub-boxes is at most $|\partial b| + \mathcal{O}(\varepsilon^2 L_\text{ST}(b))$.
For each sub-box $b_0$, let $L'$ be the set of vertices of $L_\text{ST}$ that are outside $b_0$, plus the set of vertices of $T_\text{ST}$ that are
inside $b_0$,
plus the set of the intersection points of the edges of $L_\text{ST}$ and $T_\text{ST}$ with the sides of $b_0$.
Thus, $L' \le \text{Out}(L_\text{ST}, b_0) + \text{In}(T_\text{ST},b_0) + |\partial b_0|$.
Moreover, we have $|L_\text{ST} \bigtriangleup L'| = \mathcal{O}(1/\varepsilon^2)$ and the local near-optimality argument
applies.
Namely, we obtain that $(1-1/n) L_\text{ST} \le L'$, and so
$$ - 1/n \cdot L_\text{ST} + \text{In}(L_\text{ST},b_0) + \text{Cross}(L_\text{ST}, b_0) \le \text{In}(T_\text{ST},b_0) + |\partial b_0|.$$
We now sum over all sub-boxes of box $b$ and we obtain
$$ L_\text{ST}(b) = \text{In}(L_\text{ST},b_0) + \text{Cross}(L_\text{ST}, b_0) \le \text{In}(T_\text{ST},b) + |\partial b| + \mathcal{O}(\varepsilon^2 L_\text{ST}(b)) + L_\text{ST}/n.$$
\end{proof}
\begin{lem}\label{lem:local_opt_tsp}
Let $L_{\text{TSP}}$ be a locally optimal solution to the Traveling Salesman problem and let $T_{\text{TSP}}$ be any tour.
Let $\mathcal{B}$ be a set of boxes produced by a dissection of $\mathcal{P}$.
Using the same notation for a set of segments and their total length, we then have for any box $b \in \mathcal{B}$
$$(1-\mathcal{O}(\varepsilon^2)) L_\text{TSP}(b) \le \text{In}(T_\text{TSP},b) + 3|\partial b|/2 + L_\text{TSP}/n,$$
where $|\partial b|$ is the perimeter of $b$.
\end{lem}
\begin{proof}
We again further divide the boxes into sub-boxes as we did for Lemma \ref{lem:local_opt_st}.
For each sub-box $b_0$, we define a tour $L'$ obtained by a traversal of the following Eulerian graph.
The graph vertices are $\mathcal{P}$, plus the corners of $\partial b_0$, plus all points of intersection of $L_\text{TSP}$ and $T_\text{TSP}$
with $\partial b_0$.
The edges are the segments of $\text{Out}(L_\text{TSP},b_0)$, plus the segments of $\text{In}(T_\text{TSP}, b_0)$, plus $\partial b_0$ (so that the result is connected),
plus a minimum length matching of the odd vertices of $\partial b_0$ (so that the result is Eulerian).
Thus, $L' \le \text{Out}(L_\text{TSP},b_0) + \text{In}(T_\text{TSP},b_0) + 3|\partial b_0|/2$.
Since the number of edges of $L$ intersecting $b_0$ is $\mathcal{O}(1/\varepsilon^2)$ and the number of edges in $\text{In}(T_\text{TSP},b_0)$
is $\mathcal{O}(1/\varepsilon^2)$, we have $|L_\text{TSP} \bigtriangleup L'| = \mathcal{O}(1/\varepsilon^2)$ and the local near-optimality argument applies.
Namely, we obtain $(1-1/n) L_\text{TSP} \le L'$, and so
$$ - 1/n \cdot L_\text{TSP} + \text{In}(L_\text{TSP},b_0) + \text{Cross}(L_\text{TSP},b_0) \le \text{In}(T_\text{TSP},b_0) + 3|\partial b_0|/2.$$
We now sum over all sub-boxes of box $b$ and we obtain
$$ L_\text{TSP}(b) = \text{In}(L_\text{TSP},b) + \text{Cross}(L_\text{TSP},b) \le \text{In}(T_\text{TSP},b) + 3|\partial b|/2 + \mathcal{O}(\varepsilon^2 L_\text{TSP}(b)) + L_\text{TSP}/n.$$
\end{proof}
We can now prove Theorem \ref{thm:ub_tsp_st}.
\begin{proof}[Proof of Theorem \ref{thm:ub_tsp_st}]
We first consider the Traveling Salesman case. Let $L_\text{TSP}$ be a tour produced by Algorithm \ref{algo:LS} and $T_\text{TSP}$ be any tour.
Lemma \ref{lem:local_opt_tsp} implies that for any box $b$, we have
$$(1-\mathcal{O}(\varepsilon^2)) L_\text{TSP}(b) \le \text{In}(T_\text{TSP},b) + 3|\partial b|/2 + L_\text{TSP}/n.$$
Since there are $\mathcal{O}(\varepsilon^2 n)$ boxes in total, by summing over all boxes, we obtain
$$-\mathcal{O}(\varepsilon^2 L_\text{TSP}) + \sum\limits_{b \in \mathcal{B}} L_\text{TSP}(b) = (1 - \mathcal{O}(\varepsilon^2)) L_\text{TSP} \le \sum\limits_{b \in \mathcal{B}} (\text{In}(T_\text{TSP},b) + 3|\partial b|/2) \le T_\text{TSP} + \frac{3}{2} \sum_{b \in \mathcal{B}} |\partial b|.$$
By Lemma \ref{lem:Karp77}, $\sum_{b \in \mathcal{B}} |\partial b| = \mathcal{O}(\varepsilon \sqrt{n})$ and so,
$$(1-\mathcal{O}(\varepsilon^2)) \cdot L_\text{TSP} \le T_\text{TSP} + \mathcal{O}(\varepsilon \sqrt{ n}).$$
To prove the Steiner Tree case, it is sufficient to notice that the total number of vertices in $\mathcal{P} \cup L_\text{ST} \cup T_\text{ST}$ is
at most $3n$.
It follows that the total number of boxes is $\mathcal{O}(\varepsilon^2 n)$ and by Lemma \ref{lem:Karp77},
$\sum_{b \in \mathcal{B}} |\partial b| = \mathcal{O}(\varepsilon \sqrt{n})$.
We apply a reasoning similar to the one for the TSP case to conclude the proof.
\end{proof}
Notice that we do not assume that the points are randomly distributed in the $[0,1]^2$ for the proofs of
Lemmas \ref{lem:local_opt_st} and \ref{lem:local_opt_tsp} and Theorem \ref{thm:ub_tsp_st}, thus they hold in the worst-case.
\begin{rem}
One can ask whether it is possible to prove that the local search for TSP is a PTAS without the random input assumption.
However, as shown in Fig. \ref{fig:lb_tsp} there exists a set of points such that there is a local optimum whose length
is at least $(2-o(\varepsilon)) \text{Cost}(\text{OPT})$.
\end{rem}
\section{Clustering Problems}
We now tackle the analysis of the local search algorithm for some Clustering problems.
Recall that $L$ and $G$ denote the local and global optima respectively.
In the following, for each facility $l$ of $L$ (resp. $G$), we denote by
$V_L(l)$ (resp. $V_G(l)$) the Voronoi cell of $l$
in the Voronoi diagram induced by $L$ (resp. $G$).
We extend this notation to any subset $F$ of $L$, namely,
$V_L(F)$ denotes the union of the Voronoi cells of the facilities
of $F$ induced by $L$.
We define a recursive randomized decomposition (Algorithm \ref{algo:dissection}) based on $L $ and $ G$ (and the Voronoi cells induced by $L$).
This decomposition produces a tree encoded by the function Children(), where each node is associated to a region of the Euclidean plane.
In the first step of the dissection, $B$ is the smallest square that contains all the facilities
of $L \cup G$. At every recursive call of the procedure for $(B_r,L_r,G_r)$, the algorithm maintains the following invariants:
\begin{itemize}
\item $B_r$ is a rectangle of bounded aspect ratio;
\item $L_r$ consists of all the facilities of $L$ that are contained in $B_r$;
\item $G_r$ consists of all the facilities of $G$ that are contained in $B_r$, plus some
facilities of $G$ that belong to $V_L(L_r)$.
\end{itemize}
\begin{algorithm}[H]
\caption{Recursive Adaptive Dissection Algorithm}
\label{algo:dissection}
\begin{algorithmic}[1]
\Procedure{Adaptive\_Dissection}{$B, L, G, V_L$}
\If{$|L| + |G| \ge 1/2\varepsilon^2 $ }
\If{$|L| > 1/2\varepsilon$}
\State \underline{Sub-Rectangle Process}:
\State $B'\gets $ minimal rectangle containing all facilities of
$L$ in $B$
\State $b' \gets$ maximum side-length of $B'$
\State $B'_+ \gets$ Rectangle centered on $B'$ and extended by $b'/3$ in all four directions.
\State $B''\gets B'_+\cap B$
\State \underline{Cut-Rectangle Process}:
\State $s''\gets $maximum side-length of $B''$
\State $\ell \gets $line segment that is orthogonal to the side of length $s''$ and intersects it in a random position in the middle $s''/3$.
\State Cut $B''$ into two rectangles $B_1$ and $B_2$ with $\ell$.
\State
\State Children($B$) $\gets \{B_1,B_2\}$
\State $L_1 \gets L \cap B_1$
\State $L_2 \gets L \cap B_2$
\State $G_1 \gets G \cap \{g \mid g \in V_L(L_1) \text{ and } g \notin B_2\}$
\State $G_2 \gets G \setminus G_1$
\State \textsc{Dissection}($B_1$, $L_1$, $G_{1}$, $V_L$)
\State \textsc{Dissection}($B_2$, $L_2$, $G_{2}$, $V_L$)
\Else
\State \underline{Partition Process}:
\State Children($B$) $\gets$ Arbitrary partition of the facilities of $L \cup G$ in parts of size in $[1/2\varepsilon^2, 1/\varepsilon^2]$
\EndIf
\EndIf
\EndProcedure
\end{algorithmic}
\end{algorithm}
\textbf{Regions.} We now introduce the crucial definition of \emph{regions} of a dissection tree $\mathcal{T}$ of solutions $L$ and $G$.
For any node $N$ of the dissection produced by the Partition Process,
we consider that the associated rectangle is the bounding box of the
facilities of $L_N \cup G_N$.
We assign labels to the nodes of the tree. The label of a leaf $B$
is $|L_B| + |G_B|$. Then we proceed bottom-up, for each node of
the tree, the labels of a node is equal to the sum of the labels of its two children.
Once a node has a label greater than $1/2\varepsilon^2$, we say that this node is a \emph{region node} of the tree
and set its label to 0.
We define the regions according to the region nodes. For each region node $R$, the associated region is the rectangle defined by the
node minus the regions of its descendants, namely minus the rectangles
of nodes of label 0 that are descendants of $R$. See Fig. \ref{fig:regions} for an illustration of the regions.
In the following, we denote by $\mathcal{R}$ the set of regions.
\textbf{Portals.} Let $\mathcal{D}$ be a dissection produced by
Algorithm \ref{algo:dissection}.
For any region $R$ of $\mathcal{D}$ not produced by the Partition Process,
we place $p$ equally-spaced \textit{portals} along each boundary of
$R$.
We refer to the dissection $\mathcal{D}$ along with the associated
portals as $\mathcal{D}_p$.
See Fig. \ref{fig:regions} for more details on the regions and portals.
\begin{figure}
\centering
\includegraphics[scale=0.3]{regions.pdf}
\caption{Details of the regions and portals associated to a dissection. The star-shaped points are the portals associated to Region $R_1$.
Regions $R_2,R_3,R_4$ are the only regions sharing portals with region $R_1$. All the regions are disjoint.}
\label{fig:regions}
\end{figure}
\textbf{Definitions and Notations.}
For any clustering problem, we denote by $\mathcal{C}$ the sets of the input points.
We refer to an input point as a \textit{client}.
A solution to a clustering problem is a set of facilities $S \subset \mathds{R}^2$.
For any solution $S$ and any client $c$, we denote by $c_S$ the distance from client $c$
to the closest facility of $S$: $c_S = \min\limits_{s \in S} d(c,s)$.
The service cost of a solution $S$ to a clustering problem is $\sum\limits_{c \in \mathcal{C}} c_S$.
Additionally, for any solution $S$ and client $c$, we define $c(S)$ as the facility of $S$ that
serves $c$ in solution $S$, namely $c(S) := \text{argmin}_{s \in S} d(c,s)$
Let $B$ be the smallest rectangle that contains all the clients.
Let $L$ and $G$ be two sets of facilities.
We now give the definition of an assignment which is crucial for the main proposition.
\begin{defn}
We define an \textit{assignment} as a function that maps the clients to the facility of $L \cup G$.
\end{defn}
Let $E_0$ be the assignment that maps each client $c$ to the facility of $\{c(L),c(G)\}$ that is the farther,
namely, $\forall c \in \mathcal{C}$, $E_0(c)=\text{argmax} (dist(c,c(G)),dist(c,c(L))).$
We show the following proposition which is the technical center of the proof.
\begin{prop}
\label{thm:new_struct}
Let $1/\varepsilon^2 > 0$ be an integer, $G$ and $L$ be two sets of facilities.
Let $\mathcal{D}_{1/\varepsilon^2}$ be a dissection tree with portals.
There exists an assignment $E$ that satisfies the following properties.
Let $R$ be a region not produced by the Partition Process.
If a client $c$ is such that $c(L) \in R$ and $c(G) \notin R$ then $E(c)$ is
either a portal of $R$ or a facility of $L \setminus R$.\\
Moreover,
$$\mathds{E} [\sum_{c \in \mathcal{C}} | dist(c,E(c))-dist(c,E_0(c))|] =\sum\limits_{c \in \mathcal{C}} \mathcal{O}(\varepsilon^2 \log(1/\varepsilon^2) \cdot (c_G + c_L)).$$
\end{prop}
We start by proving some properties of Algorithm \ref{algo:dissection}\footnote{Lemma \ref{lem:sub-rect} is essentially
Lemma 4 from \cite{Kolliopoulos07} but a careful writing of the details of the calculation reveals slightly different constants.}.
The proofs of the following Lemmas are deferred to Appendix \ref{appx:struct_thm}.
\begin{defn}[Aspect Ratio]
We define the aspect ratio of a rectangle $R$ that has sides of lengths $r$ and $r'$ as $\max(\frac{r}{r'},\frac{r'}{r})$.
\end{defn}
\begin{lem}\label{lem:aspect_ratio}
Let $R$ be a rectangle produced by either the Sub-Rectangle or the Cut-Rectangle process of Algorithm \ref{algo:dissection}.
The aspect ratio of $R$ is at most $5$.
\end{lem}
\begin{lem}[\cite{Kolliopoulos07}]\label{lem:sub-rect}
Let $l \in L$ be a facility and $v \in \mathds{R}^2$ be any point. Let $d$ be the distance between $v$ and $l$.
If a cutting line segment $s$ produced by the Sub-Rectangle process during Algorithm \ref{algo:dissection}
separates $v$ and $l$ for the first time, then $\text{length}(s) \le 5d.$
\end{lem}
\begin{lem}\label{lem:4-generalized}
Let $L$ be a set of facilities
Let $v \in \mathds{R}^2$, $l \in L$, $d_0=\text{dist}(v,l)$.
Suppose that, in Algorithm \ref{algo:dissection}, $v$ and $l$ are first separated by a line $s$ that is vertical and that $l$ is to the right of $s$.
Let $d_1$ be the distance from $v$ to the closest open facility located to its left.
Then, the length of $s$ is either:
(i) larger than $d_1/4$ or
(ii) smaller than $12d_0$.
\end{lem}
\begin{lem}[\cite{Kolliopoulos07}]\label{lem:fact1}
Let Event$_0(d, s)$ denote the event that an edge $e$ of length $d$ is separated by a cutting line of
side-length $s$ that is produced by Cut-Rectangle.\\
Then, $Pr[\text{Event}_0(d, s)] \le 3d/s.$
\end{lem}
We now show the proof of the Structure Theorem.
\begin{proof}[Proof of Proposition \ref{thm:new_struct}]
Let $p := 1/\varepsilon^2$.
By linearity of expectation, we only need to show this on a per-client basis.
Let $c$ be a client and $R$ a region containing $l :=c(L)$ but not $g :=c(G) $.
Let $B$ be the first box of the dissection, in top-down order, that contains $l$ but not $g$,
and let $s$ be the side of $B$ that is crossed by $[l,g]$.
We have: $dist(g,l)\leq dist(g,c)+dist(c,l)=c_G+c_L$.
Up to a rotation of center $g$, $l$ is to the north-west of $g$.
Let $u, w$ be the closest facilities of $L$ respectively to the south and to the east of $g$.
To construct $E$, we start with $E := E_0$, and modify $E$ one client at a time so that each client satisfies the first property, and we bound
the corresponding expected cost increase. The initial cost of $E$ is
$\sum\limits_{c \in \mathcal{C}} \max(c_G,c_L)$.
We modify $E(c)$ depending on whether $s$ is vertical or horizontal and according to the length of $s$.
We first provide an upper bound on the expected cost increase induced by $E(c)$ for the case where $s$ is vertical.
It is easy to see that, when $s$ is horizontal, applying the same reasoning on $w$ instead of $u$ leads to
an identical cost increase and thus, the total cost increase is at most twice the cost increase computed
for the case where $s$ is vertical.
By Lemma \ref{lem:4-generalized}, the following cases cover all possibilities for the case where $s$ is vertical.
%
\begin{itemize}
%
\item $s$ is vertical and $s$ was produced by Sub-Rectangle. Then we define $E(c)$ as the portal on $s$ that is closest to $[g,l]$. By Lemma \ref{lem:sub-rect}, the cost increase is at most $\mathcal{O}((c_G+c_L)/p)$.
%
\item $s$ is vertical and $s$ was produced by Cut-Rectangle and its length is at most $12(c_L+c_G)$. Then again we define $E(c)$ as the portal on $s$ that is closest to $[g,l]$. By assumption, again the cost increase is at most $\mathcal{O}((c_G+c_L)/p)$.
%
\item $s$ is vertical and $s$ was produced by Cut-Rectangle and its length is greater that $12(c_L+c_G)$.
Lemma \ref{lem:4-generalized} implies that $s$ has length greater than $d_{u}/4$.
If the length of $s$ is in $[d_u/4,p d_u]$. Then again we define $E(c)$ as the portal on $s$ that is closest to $[g,l]$.
Let $\mathcal{E}_0$ be the event that $d_u/4 \le |s| \le p \cdot d_u$ and $s$ is vertical.
The expected cost increase in this case is, by Lemma \ref{lem:fact1}, at most
$$\sum\limits_{\substack{d_u/4 \le i \le p \cdot d_u \\ \text{s.t } i/d_u \text{ is power of 2}}} pr[|s| = i\text{ and } \mathcal{E}_0] \cdot (i/p) \le
\mathcal{O}(\log(p)/p \cdot (c_G + c_L)).$$
\item We now turn to the last case. Namely, $s$ was produced by Cut-Rectangle and its length is greater than or equal to $p \cdot d_u$.
We define $E(c)$ depending on whether $u$ is in $R$ or not. This leads to two different sub-cases.
1. $u \notin R$.
Then we define $E(c):=u$.
The cost is bounded by the cost to go to $g$ ($\max(c_G,c_L)$) plus
the cost to go from $g$ to $u$, which is $d_u$.
Let $\mathcal{E}_1$ be the event that $u \notin R$ and $p \cdot d_u < |s|$ and $s$ is vertical.
The cost increase is, by Lemma \ref{lem:fact1}, at most,
$$\sum\limits_{\substack{i > p \cdot d_u \\ \text{s.t } i/d_u \text{ is power of 2}}} pr[|s| = i\text{ and } \mathcal{E}_1] \cdot (d_u)
\le \mathcal{O}((c_G + c_L)/p).$$
%
2. $u \in R$.
Let $d$ denotes the first line that separates $u$ from $g$. Since $u$ is to the right of $g$,
$d$ is different from $s$ and has size at least $d_u$. We have two sub-cases.
First, if $d$ was produced before $s$ in the dissection, then we also have $|d| > |s|$.
Let $\mathcal{E}_2$ be the event $|d| > |s| > p \cdot d_u$ and $s$ is vertical.
We now fix $d$. We assign $E(c)$ to be the closest portal on $R$, the expected cost increase conditioned upon $d$ is then at most:
$$\sum\limits_{\substack{p \cdot d_u < i \le |d| \\ \text{s.t } i/d_u \text{ is power of 2}}} pr[|s| = i\text{ and } \mathcal{E}_2] \cdot (i/p)
\le \mathcal{O}(\log(\frac{|d|}{p \cdot d_u}) \cdot (c_G + c_L)/p).$$
We then remove the conditioning on $d$.
If $d$ was produced by the Sub-Rectangle process, then $p \cdot d_u < |d| \le 5d_u$ by Lemma \ref{lem:sub-rect}
and the expected cost increase is at most $\mathcal{O}((c_G+c_L)/p)$.
Otherwise, $d$ was produced by the Cut-Rectangle process, and then the expected cost increase is at most
$$\sum\limits_{\substack{i > p \cdot d_u \\ \text{s.t } i/d_u \text{ is power of 2}}} pr[|d| = i\text{ and } \mathcal{E}_2] \cdot \mathcal{O}(\log(\frac{i}{p\cdot d_u}) \cdot (c_G + c_L)/p)
\le
\mathcal{O}((c_G + c_L)/p).$$\\
Second, if $d$ was produced after $s$ in the dissection, namely $|s| > |d|$.
Let $\mathcal{E}_3$ denote the event that $|s| > |d|$ and $|s| > p \cdot d_u$ and $s$ is vertical.
We assign $c$ to the closest portal located on $d$, which is at distance at most $d_u + |d|/p$ from $g$ (and so at distance at most
$c_G + d_u + |d|/p$ from $c$).
We start by fixing $s$. The expected cost conditioned upon $s$ is then (no matter how $d$ was produced), at most\\
$$\sum\limits_{\substack{d_u < i < |s| \\ \text{s.t } i/d_u \text{ is power of 2}}} pr[|d| = i\text{ and } \mathcal{E}_3] \cdot (d_u + i/p)$$
We then remove the conditioning on $s$, which leads to an expected cost of at most
$$\sum\limits_{\substack{j > p \cdot d_u \\ \text{s.t } i/d_u \text{ is power of 2}}} pr[|s| = j \text{ and } \mathcal{E}_3] \sum\limits_{d_u < i < j} 3(d_u/i) \cdot (d_u + i/p) \le \mathcal{O}((c_G + c_L)/p)$$
Thus, the total expected cost increase for $E$ is at most $\mathcal{O}((\log(p)/p) \cdot (c_G + c_L))$.
\end{itemize}
\end{proof}
\textbf{Partitioning the Clients and the Facilities.}
Before going further, we need to define a partition of the clients and
the facilities
according to the dissection produced by Algorithm \ref{algo:dissection}.
We partition the clients into two sets $C_G$ and $C_L$. $C_G$ contains the clients
that are closer to a facility of $G$ than to a facility of $L$
and $C_L$ contains the rest of the clients, namely $C_G := \{c \mid c_L = \max(c_L,c_G) \}$
and $C_L := \{c \mid c_G \neq \min(c_L,c_G)\}$.
Let $\mathcal{D}$ be a dissection produced by Algorithm
\ref{algo:dissection} and the set of its associated regions $\mathcal{R}$.
For any region $R$, we denote $C_{G}(R)$ the set of clients that are served
by $G_R$ in $G$ and that do not lay on a region not in $P$.
Furthermore, we define $C_L(R)$ as the set of clients that are served by $L_R$ in $L$
and let $C_R := V_{G}(G_R) \setminus \left( C_L \cap (V_L({L \setminus L_R}) \right)$
\footnote{This can be rewritten as $C_R := V_G(G_R) \cap (C_G \cup V_L(L_R))$.}.
This set contains the clients served by $G_R$ in $G$ except those
that belong to $C_L$ and that are served by $L \setminus L_R$ in $L$.
See Fig. \ref{fig:NR} for an illustration.
Additionally, we define $\Delta_R := V_L(L_R) \setminus V_G(G_R)$.
\begin{figure}
\centering
\includegraphics[scale=0.3]{CR.pdf}
\caption{Details of the partitioning of the client.
The star-shaped points are the facilities of $G$ and the square-shaped one are the facilities of $L$.
The blue star-shaped and square-shaped belong to respectively $G_R$ and $L_R$.
Since client $a$ is closer to facility $l$ than to facility $s$, it belongs to the set $C_L$.
Moreover, it is served in $L$ by a facility that does not belong to $V_L(L_{R})$, and so, it is not included in set $C_{R}$.
Client $b$ is closer to facility $s$ than to facility $l$ and so, it is included in set $C_{R}$ albeit it is served by a facility located on
another region in $L$.
Client $c$ is served by a facilities that belongs to $V_L(L_{R})$ (in $L$ and $G$) and so, it belongs to $C_{R}$. Finally, client $d$
does not belong to $V_{G}(G_R)$ and so, is no included in set $C_{R}$.}
\label{fig:NR}
\end{figure}
\subsection{Facility Location}
We now prove the approximation ratio of Algorithm \ref{algo:LS} for
facility location.
\begin{thm}
\label{thm:FL}
For Facility Location, Algorithm \ref{algo:LS} produces a solution $L$ of cost at most $(1 + \mathcal{O}(\varepsilon)) \cdot \text{Cost}(\text{OPT})$.
\end{thm}
\begin{proof}
Let $\text{OPT}$ be a globally optimum solution and $L$ be a locally optimum solution.
By Proposition \ref{thm:new_struct}, for any $p>0$ there exists an assignment $E$ for each random dissection $\mathcal{D}_p$ with portals
of $L \cup \text{OPT}$,
such that for any client $c$ and region $R$, if $c(L) \in R$ and $c(G) \notin R$ then
$c$ is served by a portal of $R$ or a facility of $L \setminus R$ in $E$ and
the expected cost of $E$ is at most $\mathds{E} = \sum\limits_{c \in \mathcal{C}} \max(c_L,c_G) + \mathcal{O}(\log(p)/p \cdot (\sum\limits_{c \in \mathcal{C}} (c_G + c_L)))$.
This implies that there exists a dissection $\mathcal{D}_p$ for which $E$ has value at most $\mathds{E}$.
Throughout the proof, we consider this dissection $\mathcal{D}_p$ and fix $\varepsilon := \log(p)/p$.
Let $\mathcal{R}$ be the set of regions associated to $\mathcal{D}_{p}$.
We start by constructing a solution $G$ based on $\text{OPT}$ and we compare the cost of $L$ to the cost of $G$.
The solution $G$ contains all the facilities of $\text{OPT}$ plus some extra facilities.
First, it has one facility at each portal of $\mathcal{D}_{p}$.
Moreover, for each region $R$ that is produced by the Partition Process,
we open the facilities of $L_R$. Recall that for each of these regions, $|L_R| \le 1/\varepsilon $.
We keep the same assignment for the clients.
Since there are $\mathcal{O}(\varepsilon^2 (|G|+|L|))$ regions and that for each region $G$ uses at most $1/\varepsilon$ extra facilities,
the cost of $G$ is at most Cost($\text{OPT}$)+ $\mathcal{O}(\varepsilon (|\text{OPT}| + |L|)f)$.
We now prove that the cost of $L$ is at most $(1 + \mathcal{O}(\varepsilon))/(1 - \mathcal{O}(\varepsilon))$ times the cost of $G$, namely
$$
|L| \cdot f + \sum\limits_{c \in \mathcal{C}} c_L \le (\frac{1 + \mathcal{O}(\varepsilon)}{1 - \mathcal{O}(\varepsilon)}) (|G|\cdot f
+ \sum\limits_{c \in \mathcal{C}} c_G).$$
We focus on the cost of a region $R$.
We show that, by local optimality, for each region $R$, replacing solution $L$ by solution $G$ does not lead to a much better cost.
We serve the clients of $C_R$ optimally (namely by the facilities that serve them in $G$) and
the clients of $L_{R} \setminus G_R$ by the facilities located on the portals of $R$ or by the facilities
of $L \setminus L_R$, depending on whether they belong to $C_L$ or $C_G$ and according to the assignment $E$.
Since $|L_R \setminus G_R| + |G_R \setminus L_R| = \mathcal{O}(\varepsilon^{-3})$, the locality argument applies.
Namely, we have
$$
(|G_R| - |L_R|)f + \sum\limits_{c \notin C_R \cup \Delta_R}c_L + \sum\limits_{c \in C_R} c_G + \sum\limits_{c \in \Delta_R} c_E
\ge (1-1/n) (|L|f + \sum\limits_{c} c_L).
$$
The rest of the proof is mainly computational and can be found in
the appendix \ref{appx:FL}.
\end{proof}
\subsection{$k$-Median}
Let $L$ and $\text{OPT}$ be respectively local and global optimal solutions to
the $k$-Median problem.
We start with a technical Lemma which allows us to find
``clusters'' of regions of the plane that have roughly the same number of facilities
of $L$ and $G$. See Fig. \ref{fig:balanced_cluster} for an illustration.
The proof of the Lemma is deferred to Appendix \ref{appx:kmed}.
\begin{lem}[Balanced Clustering]
\label{lem:kmed_partition}
Let $\mathcal{R} = \{r_1,...,r_p\}$ be a collection of disjoint sets. Each set contains elements of type either $L$ or $G$
and has size at least $1/2\varepsilon^2$ and at most $1/\varepsilon^2$. The total number of elements of type $L$
is $(1+3\varepsilon)$ times higher than the number of elements of type $G$.
There exists a clustering of $\{r_1,...,r_p\}$ in clusters
satisfying the following two properties. For any cluster $C$,
\begin{itemize}
\item $C$ contains at most $\mathcal{O}(1/\varepsilon^5)$ elements of $\mathcal{R}$, namely $|C| = \mathcal{O}(1/\varepsilon^5)$;
\item the difference between the number of elements of $L$ in the sets contained in $C$
and the number of elements of $G$ in the sets contained in $C$ is at least $|C|/\varepsilon$:
$$\sum_{r_i\in C} |r_i \cap L| - \sum_{r_i\in C} |r_i \cap G| \geq |C|/\varepsilon,$$
for any $1/\varepsilon \in \mathds{N}$.
\end{itemize}
\end{lem}
\begin{figure}
\begin{center}
\includegraphics[scale=0.25]{clusters.pdf}
\caption{The circle-shaped points are the elements of type $L$ and the square-shaped ones the elements of type $G$.
The black circles mark the sets $\{r_1,\ldots,r_p \}$ and the red ones show a clustering of those sets
that satisfy the property of Lemma \ref{lem:kmed_partition}.
}
\label{fig:balanced_cluster}
\end{center}
\end{figure}
\begin{thm}
\label{thm:kmed}
For $k$-Median, Algorithm \ref{algo:LS} for $k$-Median produces a solution $L$ of cost at most $(1 + \mathcal{O}(\varepsilon) \text{Cost}(\text{OPT})$ using at most $1 +
\mathcal{O}(\varepsilon))k$ Medians.
\end{thm}
\begin{proof}
Remark first that solution $L$ uses $(1+\mathcal{O}(\varepsilon))k$ facilities.
We now show that the cost of solution $L$ is at most
$1 + \mathcal{O}(\varepsilon)$ times higher than the cost of the optimal solution.\\
Recall that by Proposition \ref{thm:new_struct}, for any $p>0$ there exists an assignment $E$ for each random dissection $\mathcal{D}_p$
of $L \cup \text{OPT}$ with portals, such that for any client $c$ and region $R$, if $c(L) \in R$ and $c(\text{OPT}) \notin R$ then
$c$ is served by a portal of $R$ or a facility of $L \setminus R$ in $E$ and
the expected cost of $E$ is at most $\mathds{E} = \sum\limits_{c \in \mathcal{C}} \max(c_L,c_\text{OPT}) + \mathcal{O}(\log(p)/p \cdot (\sum\limits_{c \in \mathcal{C}} (c_\text{OPT} + c_L)))$. \\
This implies that there exists a dissection $\mathcal{D}_p$ for which $E$ has value at most $\mathds{E}$.
Throughout the proof, we consider such a dissection $\mathcal{D}_p$ and fix $\varepsilon := \log(p)/p$.
Let $\mathcal{R}$ be the set of regions associated to $\mathcal{D}_{p}$.
We prove that the cost of $L$ is at most $(1 + \mathcal{O}(\varepsilon))/(1 - \mathcal{O}(\varepsilon))$ times the cost of $S$, namely
$$
\sum\limits_{c \in \mathcal{C}} c_L \le \frac{1 + \mathcal{O}(\varepsilon)}{1 - \mathcal{O}(\varepsilon)} \sum\limits_{c \in \mathcal{C}} c_\text{OPT}.
$$
Let $\mathcal{P}$ be a clustering of the regions satisfying the properties of Lemma \ref{lem:kmed_partition}
(depending on $L$ and $\text{OPT}$).
We start by constructing a solution $G$ based on $\text{OPT}$ and we compare the cost of $L$ to the cost of $G$.
We construct $G$ in a similar way to in the proof of Theorem \ref{thm:FL}.
Namely, the solution $G$ contains all the facilities of $\text{OPT}$ plus some extra facilities:
one facility at each portal of $\mathcal{D}_{p}$ and for each region $R$ that is produced by the Partition Process,
we open the facilities of $L_R$. Recall that for each of these regions, $|L_R| \le 1/\varepsilon $.
We keep the same assignment for the clients.
We now compare the costs of $L$ and $G$. To do so, we consider all the regions of each cluster of the clustering $\mathcal{P}$
at the same time. Namely for each cluster $R$, $L$ uses at least as many facilities as $G$.
Therefore $|S_P \setminus L| + |L \setminus S_P| = \mathcal{O}(1/\varepsilon^9)$ and the locality argument applies.
The rest of the proof is similar to the proof of \ref{thm:FL} and is mainly computational and can be found in Appendix \ref{appx:kmed}.
\end{proof}
\textbf{Higher Dimensions.}
Previous results generalize to any dimension $d$. It leads to
algorithms that have exponential dependency in $d$.
More precisely, for any dimension $d$, more portals are needed to maintain the expected cost increase
for the assignment $E$ provided by the Structure Theorem.
Each of the $2d$ faces of each region has to count $p^{d-1}$ portals.
Proposition \ref{thm:new_struct} generalizes to any dimension
$d$ with $\mathcal{O}(d p^{d-1})$ portals instead of $p$.
For Facility Location, Condition($S',\varepsilon$) has to be adapted to $|S \setminus S| + |S \setminus S| = \mathcal{O}(d/ \varepsilon^{d+1})$.
Thus, Theorem \ref{thm:FL} still applies to show that the adapted Algorithm provides a $(1+\mathcal{O}(\varepsilon))$ approximation.
For the $k$-Median problem, Condition($S',\varepsilon$) has to be adapted to $|S'| \le (1+3\varepsilon)k$ and
$|S \setminus S'| + |S' \setminus S| = \mathcal{O}(d/\varepsilon^{7+d})$.
Theorem \ref{thm:kmed} still applies to prove the approximation ratio of the adapted Algorithm.
\bibliographystyle{plain}
|
\section{#1}}
\renewcommand{\thesection.\arabic{equation}}{\arabic{section}.\arabic{equation}}
\def\thesection.\arabic{equation}{\thesection.\arabic{equation}}
\newcommand{\ell}{\ell}
\newcommand{\nonumber}{\nonumber}
\newcommand\para{\paragraph{}}
\newcommand{\ft}[2]{{\textstyle\frac{#1}{#2}}}
\newcommand{\eqn}[1]{(\ref{#1})}
\newcommand\mf{\mathcal{F}}
\newcommand\ma{\mathcal{A}}
\newcommand\Q{\mathcal{Q}}
\newcommand\R{\mathbb{R}}
\newcommand\Z{\mathbb{Z}}
\newcommand\F{\mathbb{F}}
\newcommand\ml{\mathcal{L}}
\newcommand\Pf{\mathrm{Pf}}
\newcommand\D{\mathcal{D}}
\newcommand\cp{\mathbb{CP}}
\newcommand\C{\mathbb{C}}
\newcommand\T{\mathbb{T}}
\newcommand\diff{\mathrm{d}}
\newcommand{\partial}{\partial}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{\mathrm{d}}{\mathrm{d}}
\newcommand{\mathrm{e}}{\mathrm{e}}
\newcommand{\mathrm{i}}{\mathrm{i}}
\newcommand{\overline\nabla}{\overline\nabla}
\newcommand{\mathrm{e}}{\mathrm{e}}
\newcommand{\mathrm{vol}}{\mathrm{vol}}
\newcommand{\mathrm{Vol}}{\mathrm{Vol}}
\newcommand{\mathrm{Im}\, }{\mathrm{Im}\, }
\newcommand{\mathrm{Re}\, }{\mathrm{Re}\, }
\newcommand\g{\gamma}
\newcommand\e{\epsilon}
\newcommand\G{\Gamma}
\newcommand{\zeta}{\zeta}
\newcommand{{\cal P}_-}{{\cal P}_-}
\newcommand{\mg}[1]{\textbf{mg: {#1}}}
\newcommand{\mathrm{spin}^{\mathrm{c}}}{\mathrm{spin}^{\mathrm{c}}}
\newcommand\pamo{{\cal P}_\mathrm{A}(p)}
\newcommand\pb{{\cal P}_\mathrm{B}(p)}
\newcommand\qamo{{\cal Q}_\mathrm{A}(q)}
\newcommand\qb{{\cal Q}_\mathrm{B}(q)}
\newcommand\p{\mathcal{P}(p)}
\newcommand\q{\mathcal{Q}(q)}
\newlength{\sswidth}
\newcommand{\sla}[1]{
\settowidth{\sswidth}{$#1$}
\mbox{$\not{\hspace*{-0.3\sswidth}#1}$}}
\newcommand{\mathrm{bulk}}{\mathrm{bulk}}
\newcommand{\mathrm{grav}}{\mathrm{grav}}
\newcommand{F}{F}
\newcommand{\mathrm{bdry}}{\mathrm{bdry}}
\newcommand{\mathrm{ct}}{\mathrm{ct}}
\newcommand{\mathcal{F}}{\mathcal{F}}
\newcommand{s}{s}
\newcommand{\tau}{\tau}
\newcommand{\dag}{\dag}
\newcommand{\ddag}{\ddag}
\newcommand{\mathtt{h}}{\mathtt{h}}
\newcommand\cN{\mathcal{N}}
\newcommand\tr{\mathrm{Tr}}
\newcommand{\mathcal{F}}{\mathcal{F}}
\newcommand{\mathscr{L}}{\mathscr{L}}
\newcommand{\mathcal{N}}{\mathcal{N}}
\newcommand{\mathcal{M}}{\mathcal{M}}
\newcommand{\gamma}{\gamma}
\newcommand{\tau}{\tau}
\newcommand{\alpha}{\alpha}
\DeclareMathOperator{\image}{\mathrm{Im}\,}
\numberwithin{equation}{section}
\newcommand{\FA}[1]{{\bf [FA: #1]}}
\newcommand{\PR}[1]{{\bf [PR: #1]}}
\newcommand{\JFS}[1]{{\bf [JFS: #1]}}
\begin{document}
\bibliographystyle{utphys}
\begin{titlepage}
\begin{center}
\today
\vskip 2.3 cm
\vskip 5mm
{\Large \bf The holographic supersymmetric R\'enyi entropy }
\vskip 0.5cm
{\Large \bf in five dimensions }
\vskip 15mm
{Luis F. Alday, Paul Richmond and James Sparks}
\vspace{1cm}
\centerline{{\it Mathematical Institute, University of Oxford,}}
\centerline{{\it Andrew Wiles Building, Radcliffe Observatory Quarter,}}
\centerline{{\it Woodstock Road, Oxford, OX2 6GG, UK}}
\end{center}
\vskip 2 cm
\begin{abstract}
\noindent We compute the supersymmetric R\'enyi entropy across an entangling three-sphere
for five-dimensional superconformal field theories using localization.
For a class of $USp(2N)$ gauge theories we construct
a holographic dual 1/2 BPS black hole solution of Euclidean Romans $F(4)$ supergravity.
The large $N$ limit of the gauge theory results agree perfectly with the
supergravity computations.
\end{abstract}
\end{titlepage}
\pagestyle{plain}
\setcounter{page}{1}
\newcounter{bean}
\baselineskip18pt
\tableofcontents
\section{Supersymmetric R\'enyi entropy}\label{sec:fieldtheory}
\subsection{R\'enyi entropy in CFT}
Given a quantum field theory an interesting observable is the R\'enyi entropy.
To define this one divides a spatial slice $\Sigma$ into a region $A$ and its complement
$B=\Sigma\setminus A$. The Hilbert space then factorizes
\begin{eqnarray}
\mathcal{H} &\cong & \mathcal{H}_A\otimes \mathcal{H}_{B}~.
\end{eqnarray}
The reduced density matrix $\rho_A$ is defined as
\begin{eqnarray}
\rho_A & = & \mathrm{Tr}_{B} |\, 0\, \rangle\langle \, 0\, |~,
\end{eqnarray}
where $|\, 0\, \rangle$ is the ground state of the theory.
For any positive integer $n>1$, the R\'enyi entropy $S_n(A)$ associated to $A$ is then defined as
\begin{eqnarray}
S_n(A) &=& \frac{1}{1-n}\log\, \frac{\mathrm{Tr}_A \rho^n_A}{\left(\mathrm{Tr}_A \rho_A\right)^n}~.
\end{eqnarray}
This is a refinement of the entanglement entropy, which arises by analytically continuing in
$n$ and taking the limit
\begin{eqnarray}
S_{\mathrm{EE}}(A) &=& \lim_{n\rightarrow 1} S_n(A)~.
\end{eqnarray}
One can define the R\'enyi entropy using the path integral
formalism as follows.
Consider a Euclidean spacetime with coordinates $(t_E,x,\vec{z})$, where $t_E=\mathrm{i} t$ is the Euclidean time
and the spatial slice $\Sigma=\{t_E=0\}$.
The coordinate $x$ is then defined such that
$x \geq 0$ is the region $A$, and $x<0$ its complement $B$.
The ground state wave function is given by the path integral
\begin{eqnarray}
\Psi[\psi_0(x,\vec{z})] &=& \int_{\displaystyle\left.\psi\right|_{t_E=0}=\psi_0} {\cal D}\psi \, \mathrm{e}^{-I_E(\psi)} \, ,
\end{eqnarray}
where the fields $\psi$ we are integrating over are defined for negative Euclidean time (or positive imaginary Minkowskian time),
and $I_E$ is the Euclidean action. The factorization of the $t_E=0$ slice into
$A\cup B$ leads to a factorization of the boundary data
\begin{eqnarray}
\psi_0(x,\vec{z})&=& \begin{cases}\ \psi_A(x,\vec{z}) & ~~~\textrm{for}~ x\geq 0 \, ,\\
\ \psi_B(x,\vec{z}) & ~~~\textrm{for}~ x<0 \, . \end{cases}
\end{eqnarray}
The reduced density matrix is then
\begin{eqnarray}
\rho_A(\psi_A^{+},\psi_A^{-}) &=& \int {\cal D}\psi_B \, \Psi^\dagger[\psi_A^{+},\psi_B]\Psi[\psi_A^{-},\psi_B] \, .
\end{eqnarray}
If we let the imaginary time in the two path integral definitions of $\Psi$ run from $0$ to $\pm \infty$ respectively, the density matrix becomes the path integral over fields defined on the full Euclidean space, with the $(t_E,x)$ plane cut along the $x>0$ ray and with $\psi_A$ taking values $\psi_A^{\pm}$ above and below the cut, respectively. The trace of the density matrix is obtained by equating the fields across the cut and carrying out the unrestricted Euclidean path integral. More generally this construction shows that
\begin{eqnarray}
\mathrm{Tr}_A \rho^n_A &=& Z_n \, ,
\end{eqnarray}
where $Z_n$ is given by the Euclidean path integral over an $n$-sheeted covering of the cut spacetime. This formulation of the R\'enyi entropy is known as the replica trick \cite{Callan:1994py}, and
leads to the formula
\begin{eqnarray}\label{SnZn}
S_n(A) &=& \frac{1}{1-n}\log\, \frac{Z_n}{(Z_1)^n}~.
\end{eqnarray}
The calculability of $S_n(A)$ depends on the choice of spacetime and region $A$. A natural choice is a spacetime of the form $\R_{t_E} \times \R^{d-1} = \R^d$ and $A$ the unit ball inside $\R^{d-1}$, so that $\partial A = S^{d-2}$. The metric is
\begin{eqnarray}
\mathrm{d} s^2_{\R^{d}} &=&\mathrm{d} t_E^2+\mathrm{d}\rho^2+\rho^2 \mathrm{d} s^2_{S^{d-2}} \, ,
\end{eqnarray}
where $\mathrm{d} s^2_{S^{d-2}}$ denotes the round metric on the unit $(d-2)$-sphere.
The region $A$ is the ball $0\leq\rho\leq 1$. For conformal theories it is convenient
\cite{Calabrese:2004eu}
to perform the computation in the conformally equivalent space $S^d$ with metric
\begin{eqnarray}\label{Sd}
\diff s^2_{S^d} &=& \cos^2 \alpha\, \diff \tau^2 + \diff\alpha^2 + \sin^2\alpha \, \diff s^2_{S^{d-2}}~,
\end{eqnarray}
where the change of coordinates is
\begin{eqnarray}
t_E &=& \frac{\cos\alpha \sin \tau}{1+\cos \alpha \cos \tau} \, , \\
\rho &=& \frac{\sin \alpha}{1+\cos \alpha \cos \tau} \, .
\end{eqnarray}
Here $0 \leq \alpha \leq \pi/2$ and $\tau$ is periodic with period $2\pi$. In these coordinates the branch locus is at $\alpha=\pi/2$ and the cut is at $\tau=0$. In order to compute the R\'enyi entropy we need to evaluate the partition function on the $n$-branched $d$-sphere, in which case the periodicity of $\tau$ is $2\pi n$.
The above replica trick, in which one studies field theory on a singular space, is a convenient method to compute entanglement entropies in conformal field theories. However, if one is interested in constructing holographic duals this singularity persists into the bulk, where gravity becomes dynamical. This raises the issue of how to treat the singularity in gravity
\cite{Headrick:2010zt}. An ingenious way to circumvent this problem is to instead conformally map the space to $S^1 \times \mathbb{H}^{d-1}$ \cite{Casini:2011kv}
\begin{eqnarray}\label{S1H}
\mathrm{d} s^2_{S^1 \times \mathbb{H}^{d-1}} &=& \mathrm{d}\tau^2+\frac{\mathrm{d} q^2}{1+q^2} +q^2 \mathrm{d} s^2_{S^{d-2}} \, ,
\end{eqnarray}
where $q=\tan \alpha$ takes the range $q\in[0,\infty)$. The coordinates in (\ref{S1H}) realize the hyperbolic space $\mathbb{H}^{d-1}$ in a spherical slicing. The branch cut at $\alpha=\pi/2$ has now moved to $q=\infty$. In \cite{Casini:2011kv} it was argued that the entanglement entropy maps to a thermal entropy in this space, where the new Euclidean time $\tau$ has period $\beta=2\pi n$, the inverse temperature. The holographic duals are then naturally black hole solutions with hyperbolic horizons (so called topological black holes).
\subsection{Supersymmetry and localization}\label{sec:SUSY}
In \cite{Nishioka:2013haa} the authors studied a supersymmetric version
of the above R\'enyi entropy for $\mathcal{N}=2$ supersymmetric theories
on the round three-sphere with $d=3$. This is similarly obtained by
computing the partition function on $S^3$ branched $n$ times over
the $S^1$ at $\alpha=\pi/2$, but in addition one needs to turn on an appropriate
background R-symmetry gauge field to preserve supersymmetry. After a lengthy computation using localization
they find that the
partition function $Z_n$ is simply the partition function of
the \emph{squashed} sphere $S^3_b$, with squashing parameter $\sqrt{\frac{b_1}{b_2}}=\sqrt{n}$.
In this section we give a simple explanation for this result,
which works in general dimensions. Since
we will be interested mainly in dimension $d=5$,
we shall present the argument for this case.
We write the metric on $S^5$ in the form (\ref{Sd}), where we choose coordinates
on $S^{d-2}=S^3$ as
\begin{eqnarray}\label{S3coords}
\mathrm{d} s^2_{S^3} &=& \mathrm{d} \theta^2 +\cos^2\theta\mathrm{d}\psi^2 + \sin^2\theta\mathrm{d}\phi^2~.
\end{eqnarray}
Here $\psi$ and $\phi$ both have period $2\pi$, while $0\leq \theta\leq \pi/2$.
In order to define supersymmetric field theories on $S^5$ (or its branching
along $S^3$) one needs to choose a Killing spinor $\epsilon$.
The Killing spinors on $S^5$ have charges
$\pm{1}/{2}$ under the Lie derivatives along
$\partial_\tau$, $\partial_{\psi}$, $\partial_\phi$, which
generate a $U(1)^3\subset SO(6)$ subset of isometries. In particular
our choice of spinor will be such that
\begin{eqnarray}\label{charge12}
\mathcal{L}_{\partial_\tau}\epsilon &=& -\frac{\mathrm{i}}{2}\epsilon~.
\end{eqnarray}
This charge guarantees that the spinor
$\epsilon$ is smooth at $\alpha=\pi/2$, where
$\partial_\tau=0$. Indeed, the normal
space to $\alpha=\pi/2$ is a copy of $\R^2\cong \C$.
One can then introduce a polar radial variable $R=\pi/2-\alpha$,
and corresponding Cartesian coordinates $X=R\cos\tau$,
$Y=R\sin\tau$ on this normal space. The
frame $e_1=\mathrm{d} X$, $e_2=\diff Y$ rotates with charge 1
under $\partial_\tau$, so that positive and negative chirality
spinors in $\R^2$ correspondingly rotate with charge $\pm {1}/{2}$,
respectively. One could instead move to the non-rotating
frame $\hat{e}_1=\diff R$, $\hat{e}_2=R\diff\tau$,
in which the spinor will then have an explicit overall phase
$\mathrm{e}^{-\mathrm{i}\tau/2}$. However, this frame is singular at the origin $R=0$
which is why the spinor looks singular there.
Similarly, we choose conventions so that $\epsilon$
has charge $-{1}/{2}$ under $\partial_\psi$
and $+{1}/{2}$ under $\partial_\phi$. The vector bilinear
$K^\mu=\epsilon^\dagger\gamma^\mu\epsilon$
is then the Killing vector
\begin{eqnarray}\label{Ksphere}
K &=& -\partial_\phi+\partial_\psi+\partial_\tau~,
\end{eqnarray}
which generates a Hopf foliation of $S^5$.
To form the $n$-branched sphere one simply takes $\tau$ to have period $2\pi n$.
We may introduce a local complex coordinate $w=R\mathrm{e}^{\mathrm{i}\tau}$ on
the normal space $\R^2$ to the branch locus $\alpha=\pi/2$.
Then $z=w^{1/n}=R^{1/n}\mathrm{e}^{\mathrm{i}\varphi_3}$
has $\arg z=\varphi_3=\tau/n$, which has the canonical period $2\pi$.
Moreover, a function is \emph{smooth} at the branch point $R=0$
means that it is smooth in the coordinate $z$. For example,
when we come to discuss the computation of one-loop determinants
below, it is convenient to expand in Fourier modes of the
$(S^1)^3=U(1)^3$ isometry, and a complete set of modes in the $\tau$ direction
is then $\mathrm{e}^{\mathrm{i} m_3\varphi_3}$, with $m_3\in\Z$.
The Killing spinor $\epsilon$ on the round
sphere has charge $-{\mathrm{i} n}/{2}$ under $\partial_{\varphi_3}$,
and is thus singular along the branch locus when $n>1$.
We may remedy this, as in \cite{Nishioka:2013haa},
by introducing the background R-symmetry gauge field
\begin{eqnarray}\label{AR}
\mathcal{A} &=& -\frac{n-1}{n}\mathrm{d}\tau \ =\ -(n-1)\diff\varphi_3~.
\end{eqnarray}
In a five-dimensional supersymmetric gauge theory we view this
as embedded in $U(1)_R\subset SU(2)_R$, where the gauge covariant derivative
on $\epsilon$ is\footnote{The second spinor in the $SU(2)_R$ doublet then has the opposite charge under $\mathcal{A}_\mu$.}
\begin{eqnarray}\label{gaugecov}
D_\mu \epsilon &=& \nabla_\mu \epsilon + \frac{\mathrm{i}}{2}\mathcal{A}_\mu\epsilon~.
\end{eqnarray}
The flat gauge field (\ref{AR}) is smooth everywhere on the branched sphere, except at the
branch locus $\alpha=\pi/2$ where $\partial_\tau=0$. This singularity
is designed precisely so as to render the resulting Killing spinor smooth. To see this,
note that we may write $\mathcal{A}=\mathrm{i} g^{-1}(\diff g)=\mathrm{i} \diff \log g$, where
\begin{eqnarray}
g &=& \mathrm{e}^{\mathrm{i} (n-1)\tau /n}~.
\end{eqnarray}
The factor of $1/2$ in (\ref{gaugecov}) is chosen to match our Romans supergravity conventions in the next section,
but in particular this implies that the spinor transforms as
\begin{eqnarray}
\epsilon_{\mathrm{branched}} & = & g^{1/2} \epsilon~.
\end{eqnarray}
Of course then
\begin{eqnarray}
D_\mu \epsilon_{\mathrm{branched}} &=& g^{1/2}\nabla_\mu \epsilon~,
\end{eqnarray}
so that $\epsilon_{\mathrm{branched}}$ satisfies the same Killing spinor
as on the round sphere, but with the Levi-Civita spin
connection replaced by gauge-covariant derivative (\ref{gaugecov}).
Moreover, in a non-rotating frame for $\partial_\tau$, the $\tau$-dependent phase
of the new spinor is
\begin{eqnarray}
\mathrm{e}^{-\mathrm{i}\tau/2}\cdot \mathrm{e}^{\mathrm{i} (n-1)\tau/2n} \ = \ \mathrm{e}^{-\mathrm{i} \tau/2n} \ =\ \mathrm{e}^{-\mathrm{i}\phi_3/2}~.
\end{eqnarray}
This shows that the charged spinor $\epsilon_{\mathrm{branched}}$
is non-singular at the branch locus, and hence non-singular everywhere on the branched five-sphere.
Moreover, since the Killing vector bilinear $K^\mu=\epsilon_{\mathrm{branched}}^\dagger\gamma^\mu\epsilon_{\mathrm{branched}}$
is the same as that for the uncharged spinor $\epsilon$, we have from (\ref{Ksphere})
\begin{eqnarray}
K &=& -\partial_\phi+\partial_\psi+\partial_\tau \ =\ b_1\partial_{\varphi_1}+b_2\partial_{\varphi_2}+b_3\partial_{\varphi_3}~.
\end{eqnarray}
Here we have introduced the angular coordinates
\begin{eqnarray}\label{Fernandoangles}
\varphi_1 &=& -\phi\, ,\qquad \varphi_2 \ = \ \psi \, , \qquad \varphi_3 \ = \ \frac{1}{n}\tau\, ,
\end{eqnarray}
on $U(1)^3$, which all have canonical $2\pi$ periods, and $(b_1,b_2,b_3)=(1,1,\frac{1}{n})$.
Imagine now computing the perturbative partition function of a supersymmetric field theory
on the $n$-branched five-sphere using localization. Locally this computation is the same as that on the round sphere \cite{Kallen:2012va, Kim:2012ava}. What changes are the boundary conditions along the
branch locus at $\alpha=\pi/2$. However, as explained above, these boundary conditions
simply mean that fields are smooth in the $z$ coordinate, rather than the original $w$
coordinate. In particular, one expands in Fourier modes $\exp[\mathrm{i}(m_1\varphi_1+m_2\varphi_2+m_3\varphi_3)]$ where $m_i\in\Z$, $i=1,2,3$.
The charge of such a mode under the supersymmetric Killing vector $K$ is
then $m_1b_1+m_2b_2+m_3b_3$. The Killing spinor $\epsilon_{\mathrm{branched}}$ itself similarly
has charge $-(b_1+b_2+b_3)/2$. Combining these observations with the
structure of the one-loop calculations in three dimensions in \cite{Hama:2011ea, Alday:2013lba}
then leads immediately to the result in \cite{Nishioka:2013haa}:\ the partition function $Z_n$ is simply the partition function
of any three-sphere background with $(b_1,b_2)=(1,\frac{1}{n})$.
In particular most modes in the one-loop determinant
pair under supersymmetry, so that their contributions cancel.
For the remaining unpaired modes, their eigenvalues depend on the
background geometry only via their charge $m_1b_1+m_2b_2$.
The determinant over normalizable modes then leads to a double
sine function $S_2(\cdot \mid (b_1,b_2))$.
In five dimensions, similar reasoning applies to
the
explicit computation
of the perturbative partition function on $S^5$.
We then expect the result
\begin{eqnarray}
{Z}^{\text{pert}}_n \, = \, C(\mathbf{b})\prod_{a =1}^{\mathrm{rank} \ G} \int_{-\infty}^{\infty} \mathrm{d} \sigma_a \, \mathrm{e}^{ - \frac{(2 \pi)^{3}}{b_1 b_2 b_3} \mathscr{F}(\sigma)}\frac{ \prod_\alpha S_3 \left( - \mathrm{i} \alpha (\sigma) \mid \mathbf{b} \right)}{\prod_\rho S_3 \left( - \mathrm{i} \rho (\sigma) + \tfrac{1}{2}(b_1+b_2+b_3)
\mid \mathbf{b} \right)} \, . \label{Z}
\end{eqnarray}
Here the prefactor $C(\mathbf{b})$ depends only on $\mathbf{b}=(b_1,b_2,b_3)=(1,1,\frac{1}{n})$, and
in particular will not contribute to the large $N$ limit of interest in the next subsection.
The product over $\alpha$ in the numerator
is over roots of the gauge group $G$, while the product over $\rho$ in the denominator is over weights
in a weight space decomposition of the matter representation $\mathbf{R}$.
The integral in $\sigma_a$ is over the Cartan of $G$, $\mathscr{F}$ is the prepotential
of the theory, while
$S_3 ( \cdot \mid \mathbf{b} )$ is the triple sine function.
The result (\ref{Z}) also agrees
with the conjecture the authors made in \cite{Alday:2014rxa, Alday:2014bta}:\ that the partition function for any supersymmetric five-sphere background
depends on the background only via the Killing vector $K$.
In particular, (\ref{Z}) equals the squashed five-sphere perturbative partition
function \cite{Imamura:2012bm}.
\subsection{Large $N$ limit of $USp(2N)$ superconformal theories}\label{sec:largeN}
The result for the perturbative partition function (\ref{Z})
is valid for a general supersymmetric gauge theory in five dimensions.
We now focus on a
particular class of theories with gauge group
$G=USp(2N)$ and matter consisting of $N_f$ hypermultiplets in the fundamental and a single hypermultiplet in the anti-symmetric representation of $G$. These theories arise from a system of $N$ D4-branes and some number of D8-branes and orientifold
planes in massive type IIA string theory, and have a large $N$ limit that has a dual description in massive type IIA supergravity
\cite{Ferrara:1998gv, Brandhuber:1999np, Bergman:2012kr}.
For these theories, the large $N$ limit of (\ref{Z}) gives the free energy \cite{Alday:2014bta}
\begin{eqnarray}
{F} \ = \ -\log Z_n^{\mathrm{pert}} & =& \frac{ (b_1+b_2+b_3)^3}{27 b_1 b_2 b_3} {F}_{{S}^{5}_{\mathrm{round}}} \nonumber\\
&=& \frac{(1+2n)^3}{27n^2}{F}_{{S}^{5}_{\mathrm{round}}}~.
\label{freeenergyfinal}
\end{eqnarray}
Here ${F}_{{S}^{5}_{\mathrm{round}} } = \frac{9 \sqrt{2} \pi N^{5/2}}{5 \, \sqrt{8-N_f}} + \mathcal{O}\left( N^{3/2} \right) $ is the large $N$ limit of the free energy on the round five-sphere computed in reference \cite{Jafferis:2012iv}. This results
in the following large $N$ R\'enyi entropy
\begin{eqnarray}
S_n \ = \ S_n(S^3) &=& -\frac{1+7n+19n^2}{27n^2}F_{{S}^{5}_\mathrm{round}}~.
\end{eqnarray}
In the next section we will reproduce this result
from the holographic dual computation.
\section{Holographic dual}\label{sec:gravitydual}
Following \cite{Casini:2011kv}, and similar computations in lower dimensions \cite{Huang:2014gca,Nishioka:2014mwa,Huang:2014pda,Crossley:2014oea}, the holographic supersymmetric
R\'enyi entropy is computed from a 1/2 BPS Euclidean black hole solution.
As explained in \cite{Alday:2014bta}, we may construct this dual
solution in Euclidean Romans $F(4)$ supergravity, and then uplift
this to a solution of massive IIA string theory.
\subsection{Euclidean Romans $F(4)$ supergravity}
The bosonic fields of the six-dimensional Romans supergravity theory \cite{Romans:1985tw} consist of the metric, a scalar field $X$, a two-form potential $B$, and a one-form potential $A$, together with an $SO(3)\sim SU(2)$ gauge field $A^i$ where $i=1,2,3$. For the solution of interest in this paper the two-form potential vanishes, $B=0$, and we work in a gauge in which the Stueckelberg one-form $A$ is zero. Setting also the gauge coupling constant to unity, the Euclidean equations of motion are \cite{Alday:2014rxa,Alday:2014bta}
\begin{eqnarray}
\label{FullEOM}
F^i \wedge F^i & =& 0\, , \nonumber\\
D(X^{-2}*F^i) & = & 0\, ,\nonumber\\
\diff\left(X^{-1}*\mathrm{d} X\right) &=& - \left(\tfrac{1}{6}X^{-6}-\tfrac{2}{3}X^{-2}+\tfrac{1}{2}X^2\right)*1 -\tfrac{1}{8}X^{-2}\left(F^i\wedge *F^i\right) \, .
\end{eqnarray}
The first equation is a remnant of the $B$-field equation of motion, and $D\omega^i=\mathrm{d}\omega^i -\epsilon_{ijk}A^j\wedge \omega^k$ is the
$SO(3)$ covariant derivative.
The Einstein equation is
\begin{eqnarray}
R_{\mu\nu} &=& 4X^{-2}\partial_\mu X\partial_\nu X + \left(\tfrac{1}{18}X^{-6}-\tfrac{2}{3}X^{-2}-\tfrac{1}{2}X^2\right) g_{\mu\nu} \nonumber \\
& &+ \tfrac{1}{2}X^{-2}\left((F^i)^2_{\mu\nu}-\tfrac{1}{8}(F^i)^2g_{\mu\nu}\right)\, ,
\end{eqnarray}
where $(F^i)^2_{\mu\nu} = F^i_{\mu\rho} F^i_\nu{}^\rho$.
The Euclidean action is
\begin{eqnarray}\label{Fullaction}
I_E \ =\ & -\displaystyle\frac{1}{16\pi G_6}\int_{M_6} & R*1-4X^{-2} ( \diff X\wedge *\diff X + \tfrac{1}{8} F^i \wedge * F^i )\nonumber\\
&& -\left(\tfrac{2}{9}X^{-6}-\tfrac{8}{3}X^{-2}-2X^2\right)*1 \, .
\end{eqnarray}
A solution to the above equations of motion is supersymmetric provided there
exist non-trivial Dirac spinors $\epsilon_I$, $I=1,2$, satisfying
the following Killing spinor and dilatino equation
\begin{align}
D_\mu \epsilon_I \ \ =& \ \ \tfrac{\mathrm{i}}{4\sqrt{2}} ( X + \tfrac{1}{3} X^{-3} ) \Gamma_\mu \Gamma_7 \epsilon_I + \tfrac{1}{16\sqrt{2}}X^{-1} F_{\nu\rho}^i ( \Gamma_\mu{}^{\nu\rho} - 6 \delta_\mu{}^\nu \Gamma^\rho ) \Gamma_7 ( \sigma^i )_I{}^J \epsilon_J \, ,\label{KSE}\\[10pt]
0 \ \ = &\ \ - \mathrm{i} X^{-1} \partial_\mu X \Gamma^\mu \epsilon_ I + \tfrac{1}{2\sqrt{2}} \left( X - X^{-3} \right) \Gamma_7 \epsilon_I - \tfrac{\mathrm{i}}{8\sqrt{2}} X^{-1} F^i_{\mu\nu} \Gamma^{\mu\nu} \Gamma_7 ( \sigma^i )_I{}^J \epsilon_J \, .\label{dilatino}
\end{align}
Here $\Gamma_\mu$ generate the Clifford
algebra $\mathrm{Cliff}(6,0)$ in an orthonormal frame, and we have defined the chirality operator
$\Gamma_7 = \mathrm{i} \Gamma_{012345}$, which satisfies $(\Gamma_7)^2=1$.
The covariant derivative acting on the spinor is $D_\mu\epsilon_I=\nabla_\mu\epsilon_I+\frac{\mathrm{i}}{2} A_\mu^i(\sigma^i)_I{}^J\epsilon_J$.
\subsection{$1/2$ BPS black hole solution}
Our starting point is the charged AdS black hole solution of \cite{Cvetic:1999un}. After a Wick rotation and a relabelling of parameters, the solution is
\begin{eqnarray}\label{BH}
\mathrm{d} s^2 & = & \frac{H(r)^{1/2}}{f(r)} \mathrm{d} r^2 + \frac{9f(r)}{2H(r)^{3/2}} \mathrm{d} \tau^2 + r^2 H(r)^{1/2} \mathrm{d} s^2_{ \mathbb{H}^4} \, ,
\end{eqnarray}
where
\begin{eqnarray}
H(r) &=& 1 + \frac{Q}{r^3} \nonumber \, , \\
f(r) &=& -1 -\frac{\gamma}{r^3}+ \frac{2}{9} r^2 H(r)^2 \, .
\end{eqnarray}
The solution depends on the two parameters $Q$ and $\gamma$,
and $\mathrm{d} s^2_{ \mathbb{H}^4}$ is the metric of a unit radius hyperbolic space. As in
section \ref{sec:fieldtheory} we choose coordinates so that
\begin{eqnarray}
\mathrm{d} s^2_{ \mathbb{H}^4} &=& \frac{1}{(1+q^2)} \mathrm{d} q^2 + q^2 \big( \mathrm{d} \theta^2 + \cos^2 \theta \mathrm{d} \psi^2 + \sin^2 \theta \mathrm{d} \phi^2 \big) \, .
\end{eqnarray}
{\it Cf.} equations (\ref{S1H}), (\ref{S3coords}).
The remaining fields are
\begin{eqnarray}\label{XA}
X(r) &=& H(r) ^{-1/4} \nonumber \, , \\
\mathcal{A} &\equiv& A^3 \ = \ 3 \sqrt{1- \frac{\gamma}{Q}} \, \frac{H(r)-1}{H(r)} \mathrm{d} \tau +\mu\mathrm{d}\tau\, .
\end{eqnarray}
Notice that the parameter $Q$ is necessarily non-zero if $\gamma\neq0$. We have also
added a pure gauge term $\mu\mathrm{d}\tau$ to $\mathcal{A}$, which as we shall see is required
in order that the gauge field is non-singular at the horizon.
The metric (\ref{BH}) is asymptotically locally AdS for large $r$. Specifically
\begin{eqnarray}
\mathrm{d} s^2 &\simeq & \frac{9 \mathrm{d} r^2}{2r^2} + r^2\left(\mathrm{d}\tau^2 +\diff s^2_{\mathbb{H}^4}\right) \, ,
\end{eqnarray}
to leading order as $r\rightarrow\infty$. Moreover, the scalar field $X\rightarrow 1$ while
$\mathcal{A}\rightarrow \mu\mathrm{d}\tau$. Since $\tau$ will be periodically identified
in the next subsection, the conformal boundary geometry is
$S^1\times \mathbb{H}^4$.
The solution is supported by a single component of the $SU(2)$ gauge field, and without loss of generality we have chosen this to lie along the $i=3$ direction. For this choice of gauge the Killing spinor equations for $\epsilon_1$ and $\epsilon_2$ decouple. Moreover, if the fields are all real then the Killing spinor equation for $\epsilon_2$ is simply the charge conjugate of that for $\epsilon_1$
\cite{Alday:2014bta}. Hence we can consider only the spinor $\epsilon = \epsilon_1$ which satisfies
\begin{eqnarray}
D_\mu \epsilon & =& \tfrac{\mathrm{i}}{4\sqrt{2}} ( X + \tfrac{1}{3} X^{-3} ) \Gamma_\mu \Gamma_7 \epsilon + \tfrac{1}{16\sqrt{2}}X^{-1} \mathcal{F}_{\nu\rho} ( \Gamma_\mu{}^{\nu\rho} - 6 \delta_\mu{}^\nu \Gamma^\rho ) \Gamma_7 \epsilon \label{KSEdecoupled} \, , \\[10pt]
0 & = & - \mathrm{i} X^{-1} \partial_\mu X \Gamma^\mu \epsilon + \tfrac{1}{2\sqrt{2}} \left( X - X^{-3} \right) \Gamma_7 \epsilon - \tfrac{\mathrm{i}}{8\sqrt{2}} X^{-1} \mathcal{F}_{\mu\nu} \Gamma^{\mu\nu} \Gamma_7 \epsilon \label{Dilatinodecoupled} \, ,
\end{eqnarray}
with $\mathcal{F} = \mathrm{d} \mathcal{A}$.
The above black hole solution is 1/2 BPS for $\gamma=0$. To see this we
introduce the frame
\begin{eqnarray}
e^{0} & =& \frac{H(r)^{1/4}}{f(r)^{1/2}} \mathrm{d} r~,~ \quad \ \, e^{1}\ =\ \frac{3}{\sqrt{2}} \frac{ f(r)^{1/2} }{H(r)^{1/4} } \mathrm{d} \tau ~,\qquad ~ \,e^{2} \ = \ \frac{r H(r)^{1/4} }{(1+q^2)^{1/2}} \mathrm{d} q~,\\[8pt]
e^{3} & =& q r H(r)^{1/4} \mathrm{d} \theta ~,~~~e^{4} \ = \ q r H(r)^{1/4} \cos \theta \mathrm{d} \psi~,~~~e^{5} \ = \ q r H(r)^{1/4} \sin \theta \mathrm{d} \phi~,\nonumber
\end{eqnarray}
and the following basis of six-dimensional gamma matrices
\begin{eqnarray}
\Gamma_0 & = & \left( \begin{array}{cc} 0 & 1_4 \\ 1_4 & 0 \end{array} \right) \, , \quad \Gamma_m \ = \ \left( \begin{array}{cc} 0 & \mathrm{i} \gamma_m \\ - \mathrm{i} \gamma_m & 0 \end{array} \right)\, , \ \ m \, = \, 1, \ldots, 5~, \nonumber\\ \Gamma_7 &=& \left( \begin{array}{cc} - 1_4 & 0 \\ 0 & 1_4 \end{array} \right) \, ,
\end{eqnarray}
where $1_4$ is the $4 \times 4$ unit matrix and $\gamma_m$ are a basis for Cliff$(5,0)$.
In this basis the dilatino condition \eqref{Dilatinodecoupled} can be written as
\begin{eqnarray}
M \epsilon & = & 0 \, ,
\end{eqnarray}
where $M$ is an $8\times8$ matrix. A necessary condition
to have a non-trivial Killing spinor is $\det M=0$. We compute
\begin{eqnarray}
\det M &=& \frac{3^8}{2^{16}} \frac{Q^4 \gamma^4}{ r^{14} \left(r^3+Q \right)^6} \, ,
\end{eqnarray}
from which we conclude that $\gamma=0$ is necessary for supersymmetry.
In order to show that $\gamma=0$ is also sufficient, we next directly
solve the Killing spinor equation (\ref{KSEdecoupled}).
Defining
\begin{eqnarray}
f_1(r) & =& \frac{ r^{1/8} \sqrt{3 \sqrt{2} r^2+2 r^3+2 Q } }{\left(r^3+Q\right)^{3/8}} \, ,\nonumber\\
f_2(r) & =& \frac{ r^{1/8} \sqrt{-3 \sqrt{2} r^2+2 r^3+2 Q} }{\left(r^3+Q\right)^{3/8}} \, ,
\end{eqnarray}
the general solution to the dilatino and Killing spinor equation takes the form
\allowdisplaybreaks{
\begin{align}
\epsilon \ = \ \phantom{+}\sqrt{1+\sqrt{1+q^2}}
&\left(
\begin{array}{c}
\mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau+\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} \phi } \kappa_1+\mathrm{e}^{\mathrm{i} \psi } \kappa_2\right) f_1(r) \\
\mathrm{i}\mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau-\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} \phi } \kappa_1-\mathrm{e}^{\mathrm{i} \psi } \kappa_2\right) f_1(r) \\
\mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau-\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} (\phi +\psi )} \kappa_3+\kappa_4\right) f_2(r) \\
\mathrm{i} \mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau+\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} (\phi +\psi )} \kappa_3-\kappa_4\right) f_2(r) \\
-\mathrm{i} \mathrm{e}^{-\frac{1}{2}\mathrm{i} (\tau+\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} \phi } \kappa_1+\mathrm{e}^{\mathrm{i} \psi } \kappa_2\right) f_2(r) \\
\mathrm{e}^{-\frac{1}{2}\mathrm{i} (\tau-\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} \phi } \kappa_1-\mathrm{e}^{\mathrm{i} \psi } \kappa_2\right) f_2(r)\\
-\mathrm{i} \mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau-\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} (\phi +\psi )} \kappa_3+\kappa_4\right) f_1(r) \\
\mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau+\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} (\phi +\psi )} \kappa_3-\kappa_4\right) f_1(r) \\
\end{array}
\right) \nonumber\\
\ + \ \frac{q}{\sqrt{1+\sqrt{1+q^2}}}
&\left(
\begin{array}{c}
\mathrm{i} \mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau-\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} (\phi +\psi )} \kappa_3+\kappa_4\right) f_1(r) \\
\mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau+\theta +\phi +\psi )} \left(-\mathrm{e}^{\mathrm{i} (\phi +\psi )} \kappa_3+\kappa_4\right) f_1(r) \\
-\mathrm{i} \mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau+\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} \phi } \kappa_1+\mathrm{e}^{\mathrm{i} \psi } \kappa_2\right) f_2(r) \\
\mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau-\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} \phi } \kappa_1-\mathrm{e}^{\mathrm{i} \psi } \kappa_2\right) f_2(r) \\
\mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau-\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} (\phi +\psi )} \kappa_3+\kappa_4\right) f_2(r) \\
\mathrm{i} \mathrm{e}^{-\frac{1}{2}\mathrm{i} (\tau+\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} (\phi +\psi )} \kappa_3-\kappa_4\right) f_2(r) \\
-\mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau+\theta +\phi +\psi )} \left(\mathrm{e}^{\mathrm{i} \phi } \kappa_1+\mathrm{e}^{\mathrm{i} \psi } \kappa_2\right) f_1(r) \\
\mathrm{i} \mathrm{e}^{-\frac{1}{2} \mathrm{i} (\tau-\theta +\phi +\psi )} \left(-\mathrm{e}^{\mathrm{i} \phi } \kappa_1+\mathrm{e}^{\mathrm{i} \psi } \kappa_2\right) f_1(r) \\
\end{array}
\right) \, .
\end{align}}
The four integration constants $\kappa_a$, $a=1,2,3,4$, show that the solution
preserves half of the maximal 8 supercharges.
When the supergravity fields are all real
the vector field
\begin{eqnarray}\label{Killing}
K^\mu &=& \epsilon^\dagger \Gamma^\mu\epsilon
\end{eqnarray}
is Killing
\cite{Alday:2014bta}. In the case at hand we obtain a family of Killing vectors, depending
on the integration constants $\kappa_a$. For generic values of the parameter
$Q$ the black hole solution has symmetry $U(1)_\tau\times SO(4,1)$,
where $SO(4,1)$ is the isometry group of $\mathbb{H}^4$.
In particular this contains the maximal torus $U(1)^3\subset U(1)_\tau\times SO(4,1)$.
By choosing the integration constants $\kappa_a$ as
\begin{eqnarray}
\kappa_1 &=& \frac{1}{2\sqrt{2}}\, , \qquad \kappa_2 \ = \ \kappa_3 \ = \ \kappa_4 \ = \ 0\, ,
\end{eqnarray}
the Killing vector (\ref{Killing}) can be chosen to lie in the Lie algebra of this maximal torus.
Explicitly, we find
\begin{eqnarray}\label{Ksol}
K &=& -\partial_{\phi}+\partial_{\psi}+\partial_{\tau}\, .
\end{eqnarray}
\subsection{Global regularity}\label{sec:regular}
In order to have a globally regular supergravity solution we must in particular
check that the Euclidean black hole (\ref{BH}) smoothly closes
off at the horizon. This occurs at the largest root $r_h>0$ of the function
$f(r)$. Imposing $f(r_h)=0$ leads to the relation
\begin{eqnarray}
Q &=& r_h^2\left(\frac{3}{\sqrt{2}}-r_h\right)\, .
\end{eqnarray}
When $Q=0$ we note that the metric (\ref{BH}) is simply Euclidean AdS$_6$, written
in a hyperbolic slicing, and $r_h=\frac{3}{\sqrt{2}}=\ell$ is the AdS radius.
In general, near to $r=r_h$ the metric is to leading order
\begin{eqnarray}
\mathrm{d} s^2 & \simeq & \diff R^2 + \left(\sqrt{2}r_h-{2}\right)^2 R^2 \diff \tau^2 + H(r_h)^{1/2}r_h^2\diff s^2_{\mathbb{H}^4}\, ,
\end{eqnarray}
where we have defined the new radial coordinate
\begin{eqnarray}
R &=& 2^{3/8}3^{1/4} \frac{r_h^{1/4}}{\left(\sqrt{2}r_h-2\right)^{1/2}}(r-r_h)^{1/2}\, .
\end{eqnarray}
We see that the space smoothly closes off at the horizon $R=0$
provided $\tau$ has period $\beta$, where
\begin{eqnarray}
\beta &=& \frac{2\pi}{\sqrt{2}r_h-2}\, .
\end{eqnarray}
Comparing to section \ref{sec:fieldtheory}, where $\tau$ has
period $2\pi n$ with $n$ the replica index, we see that
$\beta=2\pi n$ and
\begin{eqnarray}\label{rh}
r_h &=& \frac{1+2n}{\sqrt{2}n}\, .
\end{eqnarray}
Notice that $n=1$ gives the Euclidean AdS$_6$ solution with $Q=0$.
Similarly, in order that the gauge field $\mathcal{A}$ in (\ref{XA}) is non-singular at the horizon
we have
\begin{eqnarray}
3\frac{H(r_h)-1}{H(r_h)}+\mu &=& 0\, ,
\end{eqnarray}
which using (\ref{rh}) becomes
\begin{eqnarray}
\mu &=& -\frac{(n-1)}{n}\, .
\end{eqnarray}
Thus the restriction of $\mathcal{A}$ to the conformal boundary gives
\begin{eqnarray}
\left.\mathcal{A}\, \right|_{r=\infty} &=& \mu \mathrm{d}\tau \ = \ -\frac{(n-1)}{n}\mathrm{d}\tau \, .
\end{eqnarray}
Note that this agrees with the R-symmetry gauge field (\ref{AR}) required
for supersymmetry on the $n$-branched sphere.
The resulting supergravity solution is then smooth, with the global topology
being a product of $\R^2$ with $\mathbb{H}^4\cong \R^4$, with the origin of $\R^2$ being
the horizon at $r=r_h$. Thus the solution is defined on $\R^6$,
with the action of the maximal torus $U(1)^3\subset U(1)_\tau\times SO(4,1)$
making this naturally into $\R^6\cong \R^2\oplus\R^2\oplus\R^2$.
Introducing standard $2\pi$ period coordinates (\ref{Fernandoangles}),
the Killing vector bilinear (\ref{Ksol}) becomes
\begin{eqnarray}
K &=& b_1\partial_{\varphi_1}+b_2\partial_{\varphi_2}+b_3\partial_{\varphi_3}\, ,
\end{eqnarray}
where $(b_1,b_2,b_3)=\left(1,1,\tfrac{1}{n}\right)$. Also notice
that the restriction of this vector to the conformal boundary
at $r=\infty$ agrees with the supersymmetric Killing vector
in section~\ref{sec:SUSY}.
\subsection{Free energy}\label{sec:freeenergy}
The holographic free energy is computed by evaluating the renormalized
on-shell action. This takes the form
\begin{eqnarray}
{F} &=& I_{\mathrm{ren}} \ \ = \ \ I_E + I_{\mathrm{GH}} + I_{\mathrm{counterterms}}\, .
\end{eqnarray}
Here $I_E$ is the Euclidean supergravity action \eqref{Fullaction}. The Gibbons-Hawking boundary term
is
\begin{eqnarray}
I_{\mathrm{GH}} &=& -\frac{1}{8\pi G_6}\int_{\partial M_6}\mathcal{K}\sqrt{\det h}\, \diff^5 x\, ,
\end{eqnarray}
where the space $M_6$ has boundary $\partial M_6$, $h_{mn}$ is the induced metric and
$\mathcal{K}$ denotes the trace of the second fundamental form.
The boundary counterterms for the general six-dimensional Euclidean Romans $F(4)$ theory were first given in \cite{Alday:2014rxa, Alday:2014bta}. For the present case
the two-form potential $B=0$, and consequently the general counterterm expression simplifies greatly to
\begin{align}
I_{\mathrm{counterterms}} \ = \ \frac{1}{8\pi G_6}\int_{\partial M_6}\bigg[ &\frac{4\sqrt{2}}{3}+\frac{1}{2\sqrt{2}}R(h) +\frac{3}{4\sqrt{2}}R(h)_{mn}R(h)^{mn}-\frac{15}{64\sqrt{2}}R(h)^2 \nonumber \\
&-\frac{3}{4\sqrt{2}}\| \mathcal{F}\|^2_h+\frac{4\sqrt{2}}{3}(1-X)^2 \bigg]\sqrt{\det h}\,\mathrm{d}^5x \, ,\label{Icounterterms}
\end{align}
where $R(h)_{mn}$, $R(h)$ are respectively the Ricci tensor and scalar of the induced metric.
For the Euclidean black hole solution of interest the restriction of the field strength $\mathcal{F}$ to the conformal boundary is zero.
In order to compute the regularized free energy
we cut off
the radial coordinate
at $r=\Lambda$:
\begin{eqnarray}\label{freeLambda}
{F} & = & \lim_{\Lambda \rightarrow \infty} \left[I_E(\Lambda)+ I_{\mathrm{GH}} ( \Lambda ) + I_{\mathrm{counterterms}} ( \Lambda ) \right] \, ,
\end{eqnarray}
where the relevant integrals are over $M_6(\Lambda)$ and $\partial M_6(\Lambda)$, respectively.
For our black hole solution
the integrals over $\tau$ and the hyperbolic space $\mathbb{H}^4$
factorize, so that the former contributes $2\pi n$ to the integral, while
the latter contributes a factor of $\mathrm{vol}(\mathbb{H}^4)$.
The integral over the radial variable $r$ is then easily evaluated in
(\ref{freeLambda}), and we obtain
\begin{eqnarray}
F &=& -\frac{3n}{ 4\sqrt{2} G_6}\mathrm{vol}(\mathbb{H}^4)\, r_h^3\, .
\end{eqnarray}
The volume $\mathrm{vol}(\mathbb{H}^4)$ is divergent. However, one can also
regularize this using boundary counterterms (notice
that $\mathbb{H}^4$ is Euclidean AdS$_4$). Doing so one obtains
\begin{eqnarray}
\mathrm{vol}(\mathbb{H}^4) &=& \frac{4\pi^2}{3}\, .
\end{eqnarray}
Substituting for the horizon radius $r_h$ in terms of $n$ (\ref{rh}), the final
formula for the free energy is
\begin{eqnarray}\label{Fn}
F &=& F_n \ = \ \frac{(1+2n)^3}{27n^2}F_1\, ,
\end{eqnarray}
where $F_1$ agrees with the free energy of Euclidean AdS$_6$ in a round
$S^5$ slicing. This agrees precisely with (\ref{freeenergyfinal}).
\subsection{Wilson loop}
As explained in \cite{Cvetic:1999un}, solutions of the Euclidean Romans supergravity theory uplift to solutions of massive type IIA supergravity, of the warped product form $M_6\times S^4$. In \cite{Assel:2012nf} the holographic dual of a BPS Wilson loop
in the fundamental representation was argued to be a fundamental string,
sitting at the pole of $S^4$. Here the boundary superconformal field theories are the $USp(2N)$ gauge theories discussed in section \ref{sec:largeN}.
In \cite{Alday:2014bta} the string action for a general background was shown to be
\begin{eqnarray}
S_{\mathrm{string}} &=& \frac{5\pi}{4N^2G_6}\left[\int_{\Sigma_2}\left(X^{-2}\sqrt{\det \gamma}\, \diff^2 x+\mathrm{i} B\right)-\frac{3}{\sqrt{2}}\mathrm{length}(\partial\Sigma_2)\right]\, .
\end{eqnarray}
Here $\Sigma_2$ is the string worldsheet, $\gamma_{ij}$ is the induced metric,
and we have included a boundary counterterm to regularize the string action.
For the black hole background recall that $B=0$. We then consider
a fundamental string $\Sigma_2\cong \R^2$ wrapping the
$\tau$ and $r$ directions, at a point on $\mathbb{H}^4$.
The powers of the harmonic function $H(r)$ cancel in the integrand,
so that
\begin{eqnarray}
\int_{\Sigma_2}X^{-2}\sqrt{\det \gamma}\, \diff^2 x-\frac{3}{\sqrt{2}}\mathrm{length}(\partial\Sigma_2) &=& \lim_{\Lambda\rightarrow\infty}\frac{3}{\sqrt{2}}\left(\int_{r=r_h}^{\Lambda}\diff r -\sqrt{\frac{9f(\Lambda)}{2H(\Lambda)^{3/2}}}\right)2\pi n\nonumber\\
&=& -\frac{6\pi n}{\sqrt{2}}r_h\, .
\end{eqnarray}
Identifying $-S_{\mathrm{string}}$ with $\log \, \langle W\rangle$,
we thus find
\begin{eqnarray}\label{Wn}
\log \, \langle W \rangle_n &=& \frac{1+2n}{3}\log \, \langle W \rangle_{n=1}\, .
\end{eqnarray}
Using the identification of the $n$-branched sphere partition
function with the squashed sphere result explained in section \ref{sec:SUSY},
this result agrees with the large $N$ limit of the field theory computation.
\section{Discussion}\label{sec:conclusions}
In this paper we have
computed the supersymmetric R\'enyi entropy across an entangling three-sphere
for five-dimensional superconformal field theories using localization.
In particular we presented a simple argument for why this
equals the squashed five-sphere partition function,
for appropriate squashing parameters. This argument applies
in general dimensions.
For a class of $USp(2N)$ gauge theories we have constructed
the holographic dual 1/2 BPS black hole solution of Euclidean Romans $F(4)$ supergravity.
The large $N$ limit of the gauge theory result agrees perfectly with the
supergravity computation.
In \cite{Alday:2014rxa, Alday:2014bta} it was conjectured
that for any supersymmetric Romans supergravity solution with
the topology of $\R^6$, with at least $U(1)^3$
isometry, and for which the Killing vector $K$ takes the
form $K=b_1\partial_{\varphi_1}+b_2\partial_{\varphi_2}+b_3\partial_{\varphi_3}$, the holographic free energy is
\begin{eqnarray}
F &=& \frac{(|b_1|+|b_2|+|b_3|)^3}{27|b_1b_2b_3|}F_{\mathrm{AdS}_6}\, .
\end{eqnarray}
For the explicit 1/2 BPS black hole solution we have found in the
present paper, the result (\ref{Fn}) agrees with this conjecture.
Moreover, it was also conjectured that for a BPS Wilson loop
wrapping the $\varphi_i$ circle, at the origin of the perpendicular
$\R^4$, one has $\log \, \langle W \rangle = \frac{|b_1|+|b_2|+|b_3|}{3|b_i|} \log \, \langle
W\rangle_{\mathrm{AdS}_6}$. Again, our result (\ref{Wn}) agrees with this formula.
Notice that the argument in section \ref{sec:SUSY} can be applied
to any squashed sphere background. In three dimensions,
the results of \cite{Alday:2013lba} imply that
the partition function for an $n$-branched \emph{squashed} three-sphere,
with supersymmetric Killing vector $K=b_1\partial_{\varphi_1}+b_2\partial_{\varphi_2}$,
is given by the partition function on a different squashed
sphere with $(b_1,b_2)\rightarrow (b_1,b_2/n)$.
Here the branch locus is the $S^1$ at $\partial_{\varphi_2}=0$.
We expect a similar result to hold also in five dimensions. It would be
interesting to study the implications of this for R\'enyi entropy computations.
\subsection*{Acknowledgments}
\noindent
The work of L.~F.~A. and P.~R. is supported by ERC STG grant 306260. L.~F.~A. is a Wolfson Royal Society Research Merit Award holder. J.~F.~S. is supported by the Royal Society.
|
\section{Relevance of a Local Average Treatment Effect in Epidemiologic
Research}
Some authors claim the LATE is actually what we are primarily
interested in, even if the ``compliers'' are not identifiable. A common
argument is that we care about the treatment effect for the
``compliers'' because this is the only subset of the population whose
treatment behaviors are modifiable. This rationale is problematic,
however, as the definition of ``compliers'' is instrument-dependent
\citep{pearl2011principal}. If multiple instruments were separately
used to estimate the effect of treatment in the ``compliers'' in the
same study, each effect estimate would be pertinent to a different
subset of the population: the ``compliers'' are different for each IV
analysis. It is unclear why the effects in all these various subsets
would be of primary interest. The perception of the ``compliers'' being
the subset whose behaviors are modifiable is overly simplistic because
it ignores this instrument dependence.
Other authors, like Imbens in his review, perceive the LATE as a
``second choice'' estimand, yet advocate it can sometimes be useful. He
argues for reporting subgroup effects even if the subgroup-specific
analysis is not exactly addressing the primary research question. He
proposes an analogy between estimating the effect in the ``compliers''
and estimating an effect in an all-male randomized trial, where males
are, like ``compliers,'' a subset of the general population. This
analogy begs the question: why would we be interested in the effect
estimate from a male-only trial? There are two possible reasons: (1) we
wish to inform clinical or policy decisions for men only, or (2) we
wish to extrapolate from the study to inform decisions for men and
women. If the former, the analogy with the ``compliers'' seems
ill-placed: as we\vadjust{\goodbreak} do not know who is a~``complier,'' we would not know
to whom our new policy should apply. If the latter, then we would need
to assume effect homogeneity between men and women. However, in IV
analyses, the LATE is often chosen over the global ATE precisely
because we expect too much effect heterogeneity for the ATE to be
validly identified. As such, extrapolation of the LATE to the entire
population could be ill-advised.
Finally, the LATE does not naturally translate to time-varying
treatments. Because many if not most exposures studied in epidemiologic
research vary over time, we cannot rely on the LATE to meaningfully
study their effects. If we want to study the effects of time-varying
treatments or exposures within the IV framework, we may instead
consider g-estimation of structural nested models. This approach
requires detailed modeling assumptions about the effect of treatment
\citep{robins2009estimation}.
\section{Plausibility of Monotonicity in Epidemiologic Research}
Part of the argument for favoring the LATE is that the requisite
monotonicity assumption appears more reasonable than the homogeneity
assumptions required to estimate the ``global'' ATE. For dichotomous
treatments and instruments, monotonicity requires no ``defiers'' exist,
while homogeneity requires there is no effect modification by the
instrument among the treated and untreated \citep{robins1989}. However,
while it may be plausible that there are essentially zero ``defiers''
in a randomized trial, the monotonicity condition may not hold for
instruments used in observational studies.
Consider one of the most commonly proposed instruments in epidemiologic
research, physician or facility prescribing preference \citep
{swanson2013}. Suppose we are interested in estimating the effect of a
specific treatment relative to no treatment among patients attending a
clinic where two physicians with different preferences work. The first
physician usually prefers to prescribe the treatment, but she makes
exceptions for her patients with diabetes (because of some known
contraindications). The second usually prefers to not prescribe the
treatment, but he makes exceptions for his more physically active
patients (because of some perceived benefits). Any patient who was both
physically active and diabetic would have been treated contrary to both
of these physicians preferences and, therefore, would be a ``defier.''
Because physicians' preferences represent the weighing of a variety of
risks and benefits, there may be many opportunities for a patient to be
treated contrary to physicians' preferences, and thus exhibit a
violation of monotonicity \citep{swansonmono}.
Moreover, the compliance types (``compliers,'' ``defiers,''
``always-takers,'' ``never-takers'') are not well-defined for such
instruments. Our example above considers a study with only two
physicians that could possibly have seen our patients. In more common
research settings with multiple physicians, for the compliance types to
be well-defined, all physicians with the same level of preference who
could have seen a patient would have to then treat the patient in the
exact same way. Because this is unrealistic, not only is it more likely
that there are monotonicity violations but whoever the ``compliers''
are that our effect pertains to is not just an unidentifiable but an
ill-defined subset of our population \citep{swansonmono}.
Further, most of the commonly proposed instruments in epidemiologic
research use a noncausal proxy instrument in their analyses. This is
done out of necessity, for example, we cannot measure the actual
preference of the physician when using a preference-based instrument,
or we sometimes only have the means to measure approximate locations in
the genome when using a genetic-based instrument. Although the use of
such a noncausal instrument could satisfy the other identifying
assumptions, this measurement error complicates our interpretation of a
LATE-like effect \citep{hr2006}. In particular, if the unmeasured
causal instrument is continuous, then the standard IV estimator using a
dichotomous proxy instrument would not be an effect in a specific
``compliant'' subpopulation but rather identifies a weighted average of
everybody with weights that are not particularly meaningful to any
policy decision. This is assuming that monotonicity held for the
unmeasured causal instrument, which is unlikely for instruments like
preference where the instrument is a summary of multiple dimensions of
encouragement \citep{swansonmono}.
\section{Alternative Approaches: A~Refocus on the Global Average
Treatment Effect}
Because the LATE is not generally relevant to epidemiologic research
questions, and the apparently plausible monotonicity assumption is
actually implausible in many common settings, we suggest shifting focus
back to the effect of primary interest, which is often the global ATE
\citep{rg1996}. Imbens summarized two options for this using IV
methods: (1)~present bounds for the ATE \citep{bp1997}, which are often
too wide to directly inform the particular decision at hand, or (2)
present a point estimate for the ATE assuming effect homogeneity \citep
{robins1994}, even though this assumption often is not palatable. Of
course this dichotomy is somewhat artificial: we can always do both.
Moreover, there are middle grounds.
Consider the canonical flu vaccine trial that\break Imbens described:
physicians were randomized\break to either receive or not receive a letter
encouraging~influenza vaccinations for their patients, and we are
interested in the effect of vaccination on flu-re\-lated hospitalizations
(\citeauthor{mcdonald1992}, \citeyear{mcdonald1992}). Under the instrumental conditions but not
monotonicity, Imbens calculated the Balke--Pearl bounds of [$-$0.24,
0.64] for the global ATE. These bounds do not allow us to conclude
whether vaccines are incredibly helpful, harmful, or somewhere in
between. If we further assume effect homogeneity, the point estimate is
$-$0.12 using the standard IV estimator that assumes additive
homogeneity. However, these homogeneity
assumptions are often perceived as too strong. Next, we propose a
middle ground between the uninformative bounds based on reasonable
assumptions (at least in the flu vaccine trial) and the point estimate
based on the often heroic assumption of homogeneity.
One reason the Balke--Pearl bounds are often wide is because (by
definition) we have no information on what would have happened to the
always-takers had they not been vaccinated and what would have happened
to the never-takers had they been vaccinated. The bounds are estimated
under the most extreme\linebreak[4] scenarios where all or none of these patients
would be hospitalized under these unobserved counterfactual treatments.
However, we could use subject-matter knowledge to assume a more
reasonable range of possibilities. For example, we might propose that
at most 10\% of the never-takers under treatment and 10\% of the
always-takers under no treatment would be hospitalized. We can then use
extensions of the Balke--Pearl bounds to estimate bounds of [$-$0.07,
0.02] that are consistent with this further constraint and monotonicity
\citep{rr2010}. If our narrower bounds are correct, the estimated LATE
using the standard IV estimator under monotonicity ($-$0.12) overstates
the benefit of vaccination that would have occurred had we vaccinated
the whole population. If we assume stricter limits on what would have
happened to the never-takers under treatment (e.g., at most 5\% would
have been hospitalized), we can narrow the bounds and identify the
direction of the effect: [$-$0.07, $-$0.02]. A disadvantage of this
approach is that, like approaches based on estimating the LATE, it
requires well-defined compliance types, an assumption that may be
reasonable for this randomized trial but less appropriate in other
settings as we detailed above. For a review of other approaches to
partial identification of the global ATE under IV-type assumptions, see
\citet{swansonbds}.
Another middle ground approach is to describe the sensitivity of the
point estimate to the suspected effect heterogeneity. A problem with
this approach is that the homogeneity condition is mathematically
stated with respect to the instrument, which is not intuitive, and thus
makes it difficult to apply subject-matter knowledge toward
understanding the validity of the condition. To solve this problem,
\citet{hr2006} proposed a sufficient condition for heterogeneity that
is stated with respect to the confounders. This sufficient condition
allows us to use subject matter knowledge to understand its
plausibility---and, therefore, we can also propose sensitivity
analyses based on plausible violations of this assumption. An advantage
of this approach is we no longer need to assume the compliance types
are well-defined or of a known distribution. Some authors have
previously proposed ways to understand the implications of measured
effect modifiers \citep{brookhart2007preference}, and these ideas could
be extended to consider unmeasured effect modifiers as well.
\section{Conclusion}
Imbens states we are ``limited in the questions we can answer credibly
and precisely.'' We agree, but there are differences between the
questions we can answer and the questions we want answered. Choosing
only answerable questions (e.g., identifying the LATE in some settings)
can mislead decision-making efforts: our estimates may be
misinterpreted as directly relevant to a decision when in fact they are
only tangentially related. On the other hand, exact answers for our
questions (e.g., identifying the global ATE) may be often unattainable,
but a combination of data and assumptions based on subject-matter
knowledge may go a long way towards partly answering them (e.g.,
obtaining narrow bounds for the ATE). At the very least, incomplete
answers should serve as a reminder and encouragement that further
studies---using other data and/or other assumptions---are warranted.
\section*{Acknowledgments}
This work was supported in part by the National Institutes of Health
(R01 AI073127).
|
\section{Introduction}
\label{sec:1}
Neutron star crusts are important laboratories for the study of asymmetric
nuclear matter at subnuclear density~\cite{Cham08,PR00,PR07}.
The crust is divided into an outer crust and an inner crust at the neutron drip
density $n_{\rm{drip}} \sim 4 \times 10^{11}\,\rm{g\,cm^{-3}}$
where neutrons begin to drip out of nuclei~\cite{Stei98}.
It is well known that the outer crust consists of a lattice of nuclei
with a gas of electrons, while the inner crust contains neutron-rich nuclei,
dripped neutrons, and relativistic electrons~\cite{Cham08,PR00,PR07}.
Great efforts have been devoted to the study of neutron star crusts
because of their importance in astrophysical observations and complex phase
structure~\cite{Cham08,PR00,PR07,Stei98,Haen00,Haen01,Pote13,Maru13}.
At low densities around $n_{\rm{drip}}$,
the stable shape of the nucleus is spherical, but it may change from droplet
to rod, slab, tube, and bubble, known as nuclear pasta phases,
at relatively high densities~\cite{Rave83,Wata00,Oyam07,Bao14}.
The neutron drip density is determined by the neutron chemical
potential, which is strongly dependent on the nuclear symmetry energy
and its density dependence.
Therefore, it is interesting to investigate the effects of the symmetry
energy on the neutron drip density and properties of neutron star crusts
around the neutron drip point.
The nuclear symmetry energy and its density dependence play a crucial
role in understanding various phenomena in nuclear physics and
astrophysics~\cite{PR07,LiBA08}.
The symmetry energy $E_{\rm sym}$ at saturation density
can be constrained by experiments to be around $30\pm 4$ MeV,
while the symmetry energy slope $L$ at saturation density
is still very uncertain and may vary from about $20$
to $115$ MeV~\cite{Chen13}. Many properties of neutron stars,
such as the crust structure, the crust-core transition,
and the star radius, are sensitive to the symmetry energy
and its density dependence~\cite{Oyam07,Duco11,Mene11}.
In Ref.~\cite{Oyam07}, the properties of nuclei in the inner
crust were studied using a parametrized Thomas--Fermi approach;
they were found to be sensitive to the density dependence of
the symmetry energy. In Ref.~\cite{Gril12}, a self-consistent Thomas--Fermi
approximation was used to calculate the properties of the inner
crust including pasta phases, and it was found that $L$ could
have dramatic effects on the crust structure.
The equilibrium state of neutron star crusts can be determined by
minimizing the total energy density at a given average baryon density $n_b$
under the conditions of $\beta$ equilibrium and charge neutrality.
The outer crust is well described based on experimental
masses of neutron-rich nuclei, but the inner crust has to be
studied by using phenomenological models due to the presence of
dripped neutrons. In past decades, the structure of the inner crust
has been investigated by using various methods, such as the liquid-drop
model~\cite{Wata00,BBP71,RBP72} and the Thomas--Fermi
approach~\cite{Oyam07,Gril12,Mene08,Mene10}.
Using the Wigner--Seitz approximation, the crust is divided into
spherical cells, in which a nucleus is located in
the center surrounded by a gas of electrons and neutrons.
A simple treatment for the matter inside the Wigner--Seitz
cell is referred to as the coexisting phases (CP) method~\cite{Bao14,Mene08},
in which the matter inside the cell separates into a dense phase
and a dilute phase with a sharp interface. The two coexisting phases
satisfy Gibbs conditions for phase equilibrium, which correspond
to bulk equilibrium without finite-size effects.
The surface and Coulomb energies are perturbatively taken into
account after the coexisting phases are achieved.
Another treatment of the inner crust is based on a compressible
liquid-drop (CLD) model and in this treatment the equilibrium state
is determined by minimization of the total energy density including
the surface and Coulomb energies~\cite{BBP71,Latt85,Latt91}.
Therefore, the finite-size effects due to the surface and Coulomb
energies are properly taken into account in this method.
The Thomas--Fermi (TF) approximation is considered to be self-consistent
in the treatment of finite-size effects and nucleon distributions
and has been widely used in atomic and nuclear physics~\cite{TF07}.
The TF approximation has been used to study neutron star
crusts including pasta phases at zero temperature~\cite{Gril12,Mene09}
and finite temperature~\cite{Mene10}.
It is important to compare and analyze the differences between
these methods and explore their validity at low density.
This paper has two aims. The first one is to analyze the differences
between the methods used for the study of neutron star crusts,
so as to examine the finite-size effects due to the surface
and Coulomb energies in determining the equilibrium
state at low density. In fact, we find that
the energy density obtained in the CP method is generally
larger than that of the corresponding homogeneous phase at low densities
around the neutron drip point. The failure of
the CP method at low densities may be due to the improper
treatment of the surface and Coulomb energies.
The second aim of this paper is to investigate the effects of the symmetry
energy on the neutron drip density and properties of neutron
star crusts. To calculate the properties of neutron star crusts,
we employ the TF approximation, which is considered to be self-consistent
in the treatment of finite-size effects and nucleon distributions.
For the nuclear interaction,
we adopt the relativistic mean-field (RMF) theory,
which has been successfully used to study various phenomena
in nuclear physics~\cite{Sero86,Ring90,Meng06}.
In the RMF approach, nucleons interact via the exchange of
scalar and vector mesons, while the parameters
are fitted to nuclear matter saturation properties or ground-state
properties of finite nuclei.
We consider several different parametrizations of the RMF theory,
so that we can examine the model dependence of the results obtained.
This article is arranged as follows. In Sec.~\ref{sec:2},
we briefly describe the three methods used for the study of
neutron star crusts, namely, the TF approximation, the CP method,
and the CLD model with finite-size effects.
In Sec.~\ref{sec:3}, we discuss the RMF parameters
to be used in this study.
In Sec.~\ref{sec:4}, we show the numerical results and compare
the differences between these methods, as well as discuss
the effects of the symmetry energy on the neutron drip density
and properties of the inner crust.
Section~\ref{sec:5} is devoted to the conclusions.
\section{Model and methods}
\label{sec:2}
We employ the RMF theory to study a system consisting of
protons, neutrons, and electrons. In the RMF approach, nucleons interact via
the exchange of various mesons. The mesons considered are isoscalar scalar
and vector mesons ($\sigma$ and $\omega$) and the isovector vector meson ($\rho$).
Electrons and protons interact through the electromagnetic field $A^{\mu}$.
The Lagrangian density reads
\begin{eqnarray}
\label{eq:LRMF}
\mathcal{L}_{\rm{RMF}} & = & \sum_{i=p,n}\bar{\psi}_i
\left\{i\gamma_{\mu}\partial^{\mu}-\left(M+g_{\sigma}\sigma\right)
-\gamma_{\mu} \left[g_{\omega}\omega^{\mu} +\frac{g_{\rho}}{2}\tau_a\rho^{a\mu}
+\frac{e}{2}\left(1+\tau_3\right)A^{\mu}\right]\right\}\psi_i \notag \\
& & +\bar{\psi}_{e}\left[i\gamma_{\mu}\partial^{\mu} -m_{e} +e \gamma_{\mu}
A^{\mu} \right]\psi_{e} \notag \\
&& +\frac{1}{2}\partial_{\mu}\sigma\partial^{\mu}\sigma -\frac{1}{2
m^2_{\sigma}\sigma^2-\frac{1}{3}g_{2}\sigma^{3} -\frac{1}{4}g_{3}\sigma^{4}
\notag \\
&& -\frac{1}{4}W_{\mu\nu}W^{\mu\nu} +\frac{1}{2}m^2_{\omega}\omega_{\mu
\omega^{\mu} +\frac{1}{4}c_{3}\left(\omega_{\mu}\omega^{\mu}\right)^2 \notag
\\
&& -\frac{1}{4}R^a_{\mu\nu}R^{a\mu\nu} +\frac{1}{2}m^2_{\rho}\rho^a_{\mu
\rho^{a\mu} +\Lambda_{\rm{v}} \left(g_{\omega}^2
\omega_{\mu}\omega^{\mu}\right)
\left(g_{\rho}^2\rho^a_{\mu}\rho^{a\mu}\right) -\frac{1}{4
F_{\mu\nu}F^{\mu\nu},
\end{eqnarray}
where $W^{\mu\nu}$, $R^{a\mu\nu}$, and $F^{\mu\nu}$ are the antisymmetric
field tensors for $\omega^{\mu}$, $\rho^{a\mu}$, and $A^{\mu}$,
respectively. We include the $\omega$-$\rho$ coupling term as described
in~\cite{FSU}, which is essential in modifying the symmetry energy slope.
In the RMF approach, the meson fields are treated as
classical fields and the field operators are replaced by their expectation
values. For a static system, the nonvanishing expectation values are
$\sigma =\left\langle \sigma \right\rangle$, $\omega =\left\langle
\omega^{0}\right\rangle$, $\rho =\left\langle \rho^{30} \right\rangle$,
and $A =\left\langle A^{0}\right\rangle$. From the Lagrangian
density, we can derive the equations of motion for these mean fields
in a uniform or nonuniform system.
We employ the Wigner--Seitz approximation to describe the nonuniform matter
in neutron star crusts. In the present study, we focus on examining the
symmetry energy effects on properties of neutron star crusts around the
neutron drip density, where the inhomogeneous matter is composed of
spherical nuclei arranged in a body-centered-cubic (BCC) lattice.
Generally, nonspherical nuclei (pasta phases) may appear only
at densities higher than 0.05 fm$^{-3}$~\cite{Oyam07,Bao14}.
Therefore, we consider the matter of the crust to be divided
into spherical cells treated in the Wigner--Seitz approximation.
The Wigner--Seitz cell has the same volume as the unit cell in the BCC lattice.
The lattice constant $a$ and the Wigner--Seitz cell radius $r_{\rm{ws}}$ are
related to the cell
volume by $V_{\rm{cell}}=a^3=4 \pi r_{\rm{ws}}^3 / 3=N_b / n_b $, where $N_b$
and $n_{b}$ are the baryon number per cell and the average baryon number
density, respectively. We assume that each spherical nucleus is located in
the center of a charge-neutral cell consisting of a gas of nucleons and
electrons. It is well known that the electron screening effects are
negligible at subnuclear densities~\cite{Maru05}, so we ignore the
electron screening effect caused by the nonuniform charged particle
distributions and assume the electron density to be uniform inside the
Wigner--Seitz cell. At a given average baryon density $n_b$, the equilibrium
state is determined by minimizing the total energy density of the
system. To calculate the total energy per cell, we use the self-consistent
TF approximation with the RMF model, while the CP
method with Gibbs equilibrium conditions and the CLD model including
finite-size effects due to the surface and Coulomb
energies are adopted for comparison.
\subsection{Thomas--Fermi approximation}
\label{sec:2.1}
In the TF approximation, the total energy per cell can be written as
\begin{equation}
E_{\rm{cell}}=\int_{\rm{cell}}{\varepsilon }_{\rm{rmf}}(r)d^{3}r
+{\varepsilon}_e V_{\rm{cell}} +\Delta E_{\rm{bcc}},
\label{eq:TFe}
\end{equation
where ${\varepsilon}_e$ denotes the electron kinetic energy density.
$\Delta E_{\rm{bcc}}$ is a correction term for the BCC
lattice, which is negligible when the nuclear size is much smaller
than the cell size~\cite{Oyam93,Shen11}.
${\varepsilon}_{\rm{rmf}}(r)$ is the local energy density at
radial position $r$, which is calculated in the RMF model as
\begin{eqnarray}
{\varepsilon }_{\rm{rmf}} &=&\displaystyle{\sum_{i=p,n}\frac{1}{\pi ^{2}
\int_{0}^{k_{F}^{i}}dk\,k^{2}\,\sqrt{k^{2}+{M^{\ast }}^{2}}} \notag \\
&&+\frac{1}{2}(\nabla \sigma )^{2}+\frac{1}{2}m_{\sigma }^{2}\sigma ^{2}
\frac{1}{3}g_{2}\sigma ^{3}+\frac{1}{4}g_{3}\sigma ^{4} \notag \\
&&-\frac{1}{2}(\nabla \omega )^{2}-\frac{1}{2}m_{\omega }^{2}\omega ^{2}
\frac{1}{4}c_{3}\omega ^{4}+g_{\omega }\omega \left( n_{p}+n_{n}\right)
\notag \\
&&-\frac{1}{2}(\nabla \rho )^{2}-\frac{1}{2}m_{\rho }^{2}\rho ^{2}
-\Lambda_{\rm{v}}g_{\omega }^{2}g_{\rho }^{2}\omega ^{2}\rho ^{2}+\frac{g_{\rho }}{2
\rho \left( n_{p}-n_{n}\right) \notag \\
&&-\frac{1}{2}(\nabla A)^{2}+eA\left( n_{p}-n_{e}\right) ,
\label{eq:ETF}
\end{eqnarray
where $n_{i}$ is the number density of species $i$ and
$M^{\ast}=M+g_{\sigma}\sigma $ is the effective nucleon mass.
From the Lagrangian density (\ref{eq:LRMF}), we obtain the
equations of motion for the mean fields:
\begin{eqnarray}
&&-\nabla ^{2}\sigma +m_{\sigma }^{2}\sigma +g_{2}\sigma ^{2}+g_{3}\sigma
^{3}=-g_{\sigma }\left( n_{p}^{s}+n_{n}^{s}\right) ,
\label{eq:eqms} \\
&&-\nabla ^{2}\omega +m_{\omega }^{2}\omega +c_{3}\omega^{3}
+2\Lambda_{\rm{v}}g^2_{\omega}g^2_{\rho}{\rho}^2 \omega
=g_{\omega}\left( n_{p}+n_{n}\right) ,
\label{eq:eqmw} \\
&&-\nabla ^{2}\rho +m_{\rho }^{2}{\rho}
+2\Lambda_{\rm{v}}g^2_{\omega}g^2_{\rho}{\omega}^2{\rho}
=\frac{g_{\rho }}{2}\left(n_{p}-n_{n}\right) ,
\label{eq:eqmr} \\
&&-\nabla ^{2}A=e\left( n_{p}-n_{e}\right) ,
\label{eq:eqma}
\end{eqnarray
where $n_{i}^{s}$ is the scalar density of species $i$. The equations of
motion for nucleons give the standard relations between the densities and
chemical potentials,
\begin{eqnarray}
\mu _{p} &=&{\sqrt{\left( k_{F}^{p}\right)^{2}+{M^{\ast }}^{2}}}+g_{\omega
}\omega +\frac{g_{\rho }}{2}\rho +e A,
\label{eq:mup} \\
\mu _{n} &=&{\sqrt{\left( k_{F}^{n}\right)^{2}+{M^{\ast }}^{2}}}+g_{\omega
}\omega -\frac{g_{\rho }}{2}\rho .
\label{eq:mun}
\end{eqnarray
We note that the chemical potential is spatially constant throughout the
Wigner--Seitz cell, while other quantities such as densities and mean fields
depend on the position $r$. In the Wigner--Seitz cell of neutron star crusts,
the conditions of $\beta$ equilibrium and charge neutrality are imposed,
which provide the constraints
\begin{eqnarray}
\mu _{n} &=&\mu _{p}+\mu _{e},
\label{eq:mue} \\
N_{e} &=&N_{p}=\int_{\rm{cell}}n_{p}(r)d^{3}r.
\label{eq:charge}
\end{eqnarray}
At a given average baryon density $n_{b}$, we minimize the total energy density with
respect to the cell radius $r_{\rm{ws}}$. To compute the total energy per cell at
fixed $r_{\rm{ws}}$ and $n_{b}$, we numerically solve the coupled
Eqs.~(\ref{eq:eqms})--(\ref{eq:eqma}) under the
constraints~(\ref{eq:mue}) and (\ref{eq:charge}).
In practice, we start with an initial guess for the mean fields
$\sigma (r)$, $\omega (r)$, $\rho (r)$, and $A(r)$, then determine the
chemical potentials $\mu _{n}$, $\mu _{p}$, and $\mu _{e}$
by the constraints~(\ref{eq:mue}) and (\ref{eq:charge})
and the given average density
$n_{b}=\left(N_{p}+N_{n}\right) /$ $V_{\rm{cell}}$.
Once the chemical potentials are obtained,
we can calculate various densities and solve
Eqs.~(\ref{eq:eqms})--(\ref{eq:eqma}) to get new mean fields.
This procedure is iterated until convergence is achieved.
\subsection{Coexisting phases method}
\label{sec:2.2}
In the CP method~\cite{Bao14,Mene08,Mene09,Maru05},
the matter inside the Wigner--Seitz cell
separates into a dense phase and a dilute phase with a sharp interface.
The coexisting phases satisfy Gibbs conditions for phase equilibrium,
which correspond to bulk equilibrium without finite-size effects.
The surface and Coulomb energies can be perturbatively taken into account
after the coexisting phases are achieved.
We denote the dense liquid phase and dilute gas phase
by $L$ and $G$, respectively. The Gibbs conditions for a nuclear
liquid phase in coexistence with a neutron gas at zero temperature
are written as
\begin{eqnarray}
\label{eq:CP1}
& & P^{L} = P^{G}, \\
\label{eq:CP2}
& & {\mu}^{L}_{n} = {\mu}^{G}_{n}.
\end{eqnarray}
The conditions of $\beta$ equilibrium and charge neutrality with a fixed
average baryon density $n_{b}$ provide the following constraints:
\begin{eqnarray}
\label{eq:CPbeta}
\mu_e &=& \mu_n^{L}-\mu_p^{L}, \\
\label{eq:CPcharge}
n_e &=& n_p=u n_p^{L},\\
\label{eq:CPnb}
n_b &=& u n_b^{L}+ \left(1-u\right) n_b^{G},
\end{eqnarray}
where $u$ denotes the volume fraction of the liquid phase.
We numerically solve Eqs.~(\ref{eq:CP1})--(\ref{eq:CPnb}) within the RMF model
to obtain all properties of the two coexisting phases and the volume fraction $u$
at given density $n_{b}$.
The total energy density of the system is given by
\begin{equation}
\label{eq:ews1}
\varepsilon=u{\varepsilon}_{\rm{bulk}}^{L}
+\left(1-u\right){\varepsilon}_{\rm{bulk}}^{G}+{\varepsilon}_e
+{\varepsilon}_{\rm{surf}}+{\varepsilon}_{\rm{Coul}},
\end{equation}
where ${\varepsilon}_{\rm{bulk}}^{L (G)}$ is the bulk energy density
of phase $L (G)$ obtained in the RMF model.
The surface and Coulomb energy densities for a spherical cell
are given by
\begin{eqnarray}
{\varepsilon}_{\rm{surf}} &=& \frac{3 \tau u}{r_d}, \label{eq:es1} \\
{\varepsilon}_{\rm{Coul}} &=& \frac{e^{2}}{5}
\left(n^L_b Y^L_p\right)^{2}r_d^{2} u D\left( u\right) , \label{eq:ec1}
\end{eqnarray
wit
\begin{eqnarray}
\label{eq:eu1}
D\left( u\right) =1-\frac{3}{2}u^{1/3}+\frac{1}{2}u.
\end{eqnarray
Here $\tau$ is the surface tension, which can be obtained by
a TF calculation for semi-infinite nuclear matter~\cite{Mene10,Bao14,Cent98}.
$e=\sqrt{4\pi/137}$ is the electromagnetic coupling constant.
The radius of the droplet, $r_d$, is determined by minimizing
${\varepsilon}_{\rm{surf}}+{\varepsilon}_{\rm{Coul}}$, which leads to
${\varepsilon}_{\rm{surf}}=2{\varepsilon}_{\rm{Coul}}$.
The radius of the droplet and that of the Wigner--Seitz cell are, respectively,
given by
\begin{eqnarray}
\label{eq:RD}
r_d &=& \left[\frac{15\tau}{2 e^2
\left(n^L_b Y^L_p\right)^2 D(u)}\right]^{1/3}, \\
\label{eq:RW}
r_{\rm{ws}} &=& u^{-1/3} r_d.
\end{eqnarray}
We calculate the energy density of the cell by using Eq.~(\ref{eq:ews1})
at a given average baryon density $n_{b}$ and compare to that of corresponding
homogeneous phase. It is believed that the nonuniform matter in the Wigner--Seitz
approximation should have a smaller energy density than the homogeneous
phase at low density. However, we find that the energy density obtained
in the CP method is generally larger than that of the corresponding homogeneous
phase around the neutron drip density. The failure of the CP method
at low density may be due to the improper treatment of the surface
and Coulomb energies.
\subsection{Compressible liquid-drop model}
\label{sec:2.3}
In the CP method, the equilibrium conditions
are determined by the bulk properties without finite-size effects.
To incorporate the surface and Coulomb energies in determining the
equilibrium conditions, we employ the CLD model
to calculate the energy density of the Wigner--Seitz cell
and derive the equilibrium equations by minimization of the
total energy density including the surface and Coulomb
contributions~\cite{BBP71,Latt85,Latt91}.
The energy density of the cell is generally
expressed as a function of the following six variables:
the volume fraction and radius of the droplet ($u$ and $r_d$),
the baryon density and proton fraction inside the droplet ($n^L_{b}$ and $Y^L_{p}$),
and the number densities of the neutron and electron gases ($n^G_{b}$ and $n_e$).
The total energy density of the cell is given by
\begin{equation}
\label{eq:ews2}
\varepsilon=u {\varepsilon}_{\rm{bulk}}(n^L_{b},Y^L_{p})
+\left({1-u}\right){\varepsilon}_{\rm{bulk}}(n^G_{b},0)
+{\varepsilon}_e(n_e)
+{\varepsilon}_{\rm{surf}}(u,r_d,\tau)
+{\varepsilon}_{\rm{Coul}}(u,r_d,n^L_{b},Y^L_{p}),
\end{equation}
where ${\varepsilon}_{\rm{bulk}} (n^i_b,Y^i_p)$ is the energy density
of homogeneous nuclear matter in phase $i$ ($i=L,G$),
which can be calculated in the RMF model.
The surface and Coulomb terms are given by
Eqs.~(\ref{eq:es1}) and (\ref{eq:ec1}), respectively.
Under the constraints of charge neutrality
and fixed average baryon density given by Eqs.~(\ref{eq:CPcharge}) and~(\ref{eq:CPnb}),
there are only four independent variables and we may choose
$u$, $r_d$, $n^L_{b}$, and $Y^L_{p}$. Therefore, $n_e$ and $n^G_b$
are related to the independent variables by
\begin{eqnarray}
\label{eq:charge2}
n_e &=& u n_b^{L} Y_b^L,\\
\label{eq:nb2}
n_b^{G} &=& \frac{n_b-u n_b^{L}}{1-u}.
\end{eqnarray}
By minimizing the total energy density with respect to the independent
variables~\cite{Latt91}, we obtain the following equilibrium equations:
\begin{eqnarray}
\label{eq:ee21}
0 =\frac{\partial \varepsilon}{\partial r_d} &:& \hspace{0.5cm}
r_d =\left[\frac{15\tau}{2 e^2
\left(n^L_b Y^L_p\right)^2 D(u)}\right]^{1/3}, \\
\label{eq:ee22}
0 =\frac{1}{un^L_b} \frac{\partial \varepsilon}{\partial Y_p^L} &:& \hspace{0.5cm}
\mu_{e}=\mu_{n}^L-\mu_{p}^L-\frac{2 e^{2}}{5} n_b^L Y_p^L r_d^{2}D(u), \\
\label{eq:ee23}
0 =\frac{\partial \varepsilon}{\partial n^L_b}-\frac{Y_p^L}{n^L_b}
\frac{\partial \varepsilon}{\partial Y_p^L} &:& \hspace{0.5cm}
\mu_{n}^L = \mu_{n}^G, \\
\label{eq:ee24}
0 =\frac{n_b^L}{u}\frac{\partial \varepsilon}{\partial n^L_b}
-\frac{\partial \varepsilon}{\partial u} &:& \hspace{0.5cm}
P^L-P^G=\frac{ e^{2}}{5} \left(n^L_b Y^L_p\right)^2 r_d^{2}
\left( 1-2u^{1/3}+u\right).
\end{eqnarray}
We note that the terms involving derivatives of the surface tension are ignored
in deriving these equilibrium equations.
As discussed by Iida and Oyamatsu~\cite{Oyam04},
the surface tension $\tau$ may depend
on the inner density and proton fraction ($n^L_{b}$ and $Y^L_{p}$).
Furthermore, $\tau$ could be affected by the size of the droplet ($r_d$),
which is known as a curvature correction to the surface tension~\cite{Haen00}.
However, the dependence of $\tau$ on these variables is poorly known,
especially in a neutron-rich system. This is because the surface tension
is generally obtained by a TF calculation for semi-infinite nuclear matter.
Due to the equilibrium conditions between the nuclear liquid and gas phases,
the surface tension would be a function of only one of the four variables
$n^L_{b}$, $Y^L_{p}$, $n^G_{b}$, and $Y^G_{p}$. Therefore, it is not
possible to obtain the partial derivatives of $\tau$ with respect to
each independent variable from this calculation.
For simplicity, we neglect contributions from the derivatives of
the surface tension in deriving the above equilibrium equations.
One can see that equilibrium equations~(\ref{eq:ee21})--(\ref{eq:ee24})
of the present paper are equivalent to Eqs.~(43)--(47) of Ref.~\cite{Cham08}.
By comparing Eq.~(\ref{eq:ee22}) with Eq.~(\ref{eq:CPbeta}), we can see that
the $\beta$ equilibrium condition is altered due to the inclusion of
finite-size effects in the minimization procedure.
The last term of Eq.~(\ref{eq:ee22}) comes from the Coulomb energy,
which favors a smaller electron chemical potential.
This leads to the conclusion that the electron fraction (equal to
the average proton fraction) is overestimated in the CP
method with bulk equilibrium.
Also, the inclusion of finite-size effects affects
the mechanical equilibrium as can be seen by comparing
Eq.~(\ref{eq:ee24}) with Eq.~(\ref{eq:CP1}).
The last term of Eq.~(\ref{eq:ee24}) comes from the sum of the surface
and Coulomb energies. Generally, the bulk pressure inside the droplet
is larger than that outside due to the surface and Coulomb contributions,
which leads to a higher density at the center of the droplet.
\section{Parameters}
\label{sec:3}
In this section, we discuss the choice of the RMF parameters
to be used in this study. The parameters of the RMF models are generally
fitted to nuclear matter saturation properties or ground-state
properties of finite nuclei. To study the properties of neutron star
crusts and compare the differences among various methods,
we consider four different RMF parametrizations,
NL3~\cite{NL3}, TM1~\cite{TM1}, FSU~\cite{FSU}, and IUFSU~\cite{IUFSU},
so that we can examine the model dependence of the results obtained.
These RMF models are known to be successful in reproducing
the ground state properties of finite nuclei including unstable ones.
The NL3 parametrization includes nonlinear terms of the $\sigma$ meson only,
while the TM1 parametrization includes nonlinear terms for both $\sigma$
and $\omega$ mesons. An additional $\omega$-$\rho$ coupling term is added
in the FSU and IUFSU parametrizations, and it plays an important role in
modifying the density dependence of the symmetry energy and
affecting the neutron star properties~\cite{Mene11,FSU,IUFSU,Horo01,Horo03,Prov13}.
The IUFSU parametrization was developed from FSU by reducing the neutron skin
thickness of $^{208}$Pb and increasing the maximum neutron star mass in
the parameter fitting~\cite{IUFSU}.
The TM1 model was successfully used to construct the equation of
state for supernova simulations and neutron star calculations~\cite{Shen02,Shen11}.
For completeness, we present the parameters and saturation properties
of these RMF models in Table~\ref{tab:1}.
In order to examine the influence of the symmetry energy slope $L$,
we generate two sets of models based on the TM1 and IUFSU parametrizations.
We determine the model parameters by simultaneously adjusting
$g_{\rho}$ and ${\Lambda}_{\rm{v}}$ so as to achieve a given $L$ at saturation
density and keep $E_{\rm{sym}}$ fixed at a density of 0.11 fm$^{-3}$.
The choice of the fixed density $n_{\rm{fix}}=0.11\, \rm{fm}^{-3}$
is based on the following consideration.
In one set of generated models, the variation of $L$ at
saturation density would not affect the reproduction of
well-known properties of finite nuclei.
It has been pointed out that the binding energy of finite nuclei is essentially
determined by the symmetry energy at a density of $\sim 0.11\, \rm{fm}^{-3}$,
not by the symmetry energy at saturation density~\cite{Chen13,Horo01}.
To examine the sensitivity of the binding energy to the fixed density
$n_{\rm{fix}}$ of the symmetry energy, we perform a standard RMF calculation
as described in Refs.~\cite{Ring90,TM1}
for $^{208}$Pb using the two sets of generated models
with different choices of $n_{\rm{fix}}$.
One can see in Fig.~\ref{fig:1Pb208} that
the binding energy per nucleon of $^{208}$Pb remains almost unchanged
with varying $L$ using $n_{\rm{fix}}=0.11\, \rm{fm}^{-3}$,
whereas it deviates from the experimental value (7.87 MeV)
using $n_{\rm{fix}}=0.10\, \rm{fm}^{-3}$ or $n_{\rm{fix}}=n_0$
(where $n_0$ is the saturation density).
In Tables~\ref{tab:2} and~\ref{tab:3}, we present the parameters
$g_{\rho}$ and ${\Lambda}_{\rm{v}}$
generated based on TM1 and IUFSU by producing
a given $L$ at saturation density and fixed symmetry energy
at $n_{\rm{fix}}=0.11\, \rm{fm}^{-3}$.
We also show in these tables the symmetry energy at saturation density,
$E_{\rm{sym}} (n_0)$, and the neutron-skin thickness
$\Delta r_{np} =\langle r^2_n \rangle^{1/2} - \langle r^2_p \rangle^{1/2}$
of $^{208}$Pb, both of which generally increase with increasing $L$.
We stress that all models in each set have the same isoscalar saturation
properties and fixed symmetry energy at $n_{\rm{fix}}=0.11\, \rm{fm}^{-3}$,
but they have different symmetry energy slope $L$.
By using the set of models with different $L$, it is possible to study
the impact of $L$ on the neutron drip density and properties of
neutron star crusts.
\section{Results and discussion}
\label{sec:4}
In this section, we investigate the effects of the symmetry energy on
the neutron drip density and properties of neutron star crusts.
We first make a detailed comparison among the three methods used
for the study of neutron star crusts,
namely, the TF approximation, the CP method, and the CLD
model with finite-size effects. We analyze the differences among
these methods and explore their validity at low densities near
the neutron drip point. To study the influence of the symmetry
energy slope $L$, we employ the TF approximation, which is
considered to be self-consistent in the treatment of finite-size
effects and nucleon distributions.
\subsection{Comparison between different methods}
\label{sec:4.1}
To describe nonuniform matter in the Wigner--Seitz cell,
we consider three different methods: (1) the simple CP method
with bulk Gibbs equilibrium conditions; (2) the CLD model with
equilibrium conditions determined by including the surface and
Coulomb energies; and (3) the self-consistent TF approximation.
We note that treatments of surface and Coulomb energies
are obviously different among these methods. In the CP method,
Gibbs equilibrium conditions are used which correspond to bulk equilibrium
without finite-size effects, while the surface and Coulomb energies
are perturbatively incorporated after the two coexisting phases
are achieved. In the CLD model, equilibrium conditions are determined
by minimization of the total energy density including the surface
and Coulomb energies; therefore they are incorporated in a consistent
manner. In the TF approximation, the surface effect and nucleon
distributions are treated self-consistently, rather than
a sharp surface being assumed in the CP and CLD methods.
In addition, a neutron skin can be well described within
the TF approximation, but it is not explicitly included
in the CP and CLD methods.
In Fig.~\ref{fig:2ea}, we show the total energy per nucleon,
$E=\varepsilon /n_b -M$, as a function of the average baryon density $n_b$
obtained using the TF, CLD, and CP methods, while that of
homogeneous matter is also displayed.
It is interesting to see that the three methods yield very similar
$E$ at higher densities, but there are significant differences
at lower densities. Moreover, one can see that the simple CP method
fails to describe the nonuniform matter near the neutron drip
density, since $E$ of CP is larger than that of homogeneous matter.
We note that the kinks of CP at $n_b < 10^{-3}$ fm$^{-3}$ correspond
to the neutron drip point.
The failure of the CP method may be due to its improper treatment
of the surface and Coulomb energies.
It implies that the finite-size effect due to the surface and Coulomb
energies is too large to be treated perturbatively at low densities,
so that we have to include contributions from surface and Coulomb
energies in determining the equilibrium state as done in
the CLD and TF methods.
By comparing the results between CLD and CP,
one can see an obvious improvement due to the inclusion of
finite-size effects in the CLD method. Furthermore,
the results of CLD are very close to those obtained in
the self-consistent TF calculation.
In order to analyze the results of Fig.~\ref{fig:2ea}, we plot
various contributions to $E$ in Fig.~\ref{fig:3ebec}.
The Coulomb energy per nucleon,
$E_{\rm{Coul}}={\varepsilon}_{\rm{Coul}}/n_b$,
is calculated by using Eq.~(\ref{eq:ec1}) in the CP and CLD methods,
while it can be easily computed in the TF approximation by using
$E_{\rm{Coul}}=\frac{1}{2 N_b}
\int_{\rm{cell}} e A\left(r\right)\left[n_p\left(r\right)-n_e\right] d^3r$.
However, it is difficult to separate the surface energy from the bulk
energy in the TF approximation, because both are involved
in Eq.~(\ref{eq:ETF}). To estimate the surface energy in the TF approximation,
we use the equilibrium condition
${\varepsilon}_{\rm{surf}}=2\,{\varepsilon}_{\rm{Coul}}$
obtained in the liquid-drop model, which yields the sum
${\varepsilon}_{\rm{surf}}+{\varepsilon}_{\rm{Coul}}=3\,{\varepsilon}_{\rm{Coul}}$.
Therefore, we can define the bulk energy density in the TF approximation by
${\varepsilon}_{\rm{bulk}}=\left( E_{\rm{cell}}-\Delta E_{\rm{bcc}}\right)/V_{\rm{cell}}
-3{\varepsilon}_{\rm{Coul}}-{\varepsilon}_e$ according to Eq.~(\ref{eq:TFe}),
while it is given by ${\varepsilon}_{\rm{bulk}}=u{\varepsilon}_{\rm{bulk}}^{L}
+\left(1-u\right){\varepsilon}_{\rm{bulk}}^{G}$ in the CP and CLD methods.
In Fig.~\ref{fig:3ebec}, from top to bottom, we show, respectively,
the bulk energy per nucleon,
$E_{\rm{bulk}}={\varepsilon}_{\rm{bulk}}/n_b-M$,
the electron kinetic energy per nucleon,
$E_{e}={\varepsilon}_{e}/n_b$,
and the Coulomb energy per nucleon,
$E_{\rm{Coul}}$, obtained in the CP, CLD, and TF methods using the TM1 parametrization.
One can see that $E_e$ and $E_{\rm{Coul}}$ increase
with decreasing $n_b$, and the differences between CP and CLD methods
become very large at low density.
Due to the increasing contributions of $E_e$ and
$3E_{\rm{Coul}}$ (the sum of surface and Coulomb energies per nucleon),
the total energy per nucleon, $E$, obtained in the CP method is even larger
than that of homogeneous matter near the neutron drip density (see Fig.~\ref{fig:2ea}),
which implies that the simple CP method is not applicable to describing
nonuniform matter at low density.
In order to understand the differences in $E_e$ and $E_{\rm{Coul}}$ between
the CP and CLD methods, we display the electron fraction $Y_e=n_e/n_b$
as a function of $n_b$ in Fig.~\ref{fig:4ye}.
At a given $n_b$, a large $Y_e$ corresponds to large $n_e$ and $\mu_e$,
which results in more contributions from $E_e$ and $E_{\rm{Coul}}$.
One can see that $Y_e$ of the CP method is significantly larger than
that of the CLD and TF methods in all cases of Fig.~\ref{fig:4ye}.
This can be understood by comparing
Eqs.~(\ref{eq:CPbeta}) and~(\ref{eq:ee22}).
In the CP method, $\mu_e$ is determined by using Eq.~(\ref{eq:CPbeta}),
while an additional term (the last term) appears in Eq.~(\ref{eq:ee22})
caused by the Coulomb energy in the CLD method.
This term leads to a smaller $\mu_e$ in the CLD method compared to the CP case.
Therefore, we conclude that the inclusion of surface and Coulomb energies
in determining the equilibrium state plays a crucial role
in the description of nonuniform matter at low density.
In Fig.~\ref{fig:5rcd}, we plot the radius of the droplet, $r_d$,
and that of the Wigner--Seitz cell, $r_{\rm{ws}}$,
as a function of $n_b$ obtained by using the TF, CLD, and CP methods.
In the CP and CLD methods, $r_d$ is given by Eq.~(\ref{eq:RD}),
while it is defined by
$r_d=\sqrt{\frac{5}{3}} \langle r^2_p \rangle^{1/2}$ in the TF approximation.
One can see that $r_d$ does not explicitly depend on $n_b$
and there is no significant difference among the three methods.
This is because the equilibrium nuclear size $r_d$ is mainly determined
by a competition between the surface and Coulomb energies,
which is a common feature in these methods.
On the other hand, $r_{\rm{ws}}$ obviously decreases with increasing $n_b$.
Moreover, $r_{\rm{ws}}$ in the CP method is generally smaller than that
of the CLD and TF methods.
This tendency is related to the behavior of $Y_e$ shown in Fig.~\ref{fig:4ye}.
As discussed above, a large $Y_e$ corresponds to large $n_e$ and $\mu_e$,
which results in a large volume fraction $u$ and a small $r_{\rm{ws}}$
according to the relations given in Eqs.~(\ref{eq:CPcharge}) and (\ref{eq:RW}).
In Fig.~\ref{fig:6z}, we present the proton number $Z$ of the droplet
as a function of $n_b$ obtained by using the TF, CLD, and CP methods.
It is well known that $Z$ is sensitive to the surface energy~\cite{Oyam07}.
We can see that the density dependence of $Z$ is relatively weak
at low density for all cases, while it shows a strong density dependence
with increasing $n_b$. The behavior of IUFSU is different from others
due to its relatively low value of $L$.
It has been shown in Refs.~\cite{Oyam07,Bao14} that
a small $L$ favors a large surface tension $\tau$,
which leads to a large $Z$ since $Z$ increases monotonically with $\tau$.
Comparing results among the three methods, we find that $Z$ of the TF method
is generally larger than that of the CP and CLD methods.
This may be due to the different treatment of nucleon distributions.
In the TF approximation, the surface effect and nucleon distributions
are calculated self-consistently and the neutron skin is well described.
\subsection{Neutron drip density}
\label{sec:4.2}
We perform the self-consistent TF calculation to study the effects
of the symmetry energy on the neutron drip density. To examine the
influence of the symmetry energy slope $L$, we use two sets of models
generated from the TM1 and IUFSU parametrizations.
We note that all models in each set have the same isoscalar saturation
properties and fixed symmetry energy at $n_{\rm{fix}}=0.11\, \rm{fm}^{-3}$,
but they have different symmetry energy slope $L$.
The neutron drip point is determined by the condition $\mu_n = Mc^2$.
Beyond this point, neutrons begin to drip out of the nuclei and
form a free neutron gas.
In Fig.~\ref{fig:7ndrip}, we show the neutron drip density $n_{\rm{drip}}$
as a function of $L$ using the two sets of models generated from TM1 and IUFSU,
while the results of NL3 and FSU are also displayed.
It is found that $n_{\rm{drip}}$ increases with $L$
in both TM1 and IUFSU cases.
This tendency can be understood from the following analysis.
The neutron drip density is related to the nucleon number and radius of
the Wigner--Seitz cell as $n_{\rm{drip}}=A/ \frac{4}{3} \pi r_{\rm{ws}}^3 $.
The nucleon number $A$ at $n_{\rm{drip}}$ is not obviously affected
by $L$ [see Fig.~\ref{fig:9dripA}(a)].
However, the cell radius $r_{\rm{ws}}$ at $n_{\rm{drip}}$
decreases significantly with increasing $L$, as shown in Fig.~\ref{fig:8dripR}.
One reason for the decrease of $r_{\rm{ws}}$ is
because the generated models in each set have
fixed symmetry energy at $n_{\rm{fix}}=0.11\, \rm{fm}^{-3}$
with different $L$, and, therefore, a larger $L$ corresponds to
a larger symmetry energy $E_{\rm sym}$ near the saturation
density (see Tables~\ref{tab:2} and~\ref{tab:3}).
Based on the relation derived from the liquid-drop model,
$\mu_e = \mu_n^{L}-\mu_p^{L} \simeq 4 \delta E_{\rm{sym}}$
with $\delta=1-2Y^L_p$ being the neutron excess,
a large $E_{\rm{sym}}$ at the center of the nucleus
(corresponding to a large value of $L$) favors a high $\mu_e$,
although it corresponds to a small $\delta$ and a low nucleon density
in the center region (see Fig.~\ref{fig:10dripD}).
As mentioned above, a high value of $\mu_e$
results in a large volume fraction $u$ and a small $r_{\rm{ws}}$
according to the relations given in Eqs.~(\ref{eq:CPcharge}) and (\ref{eq:RW}).
Therefore, a larger $L$ in one set of generated models leads to
a smaller $r_{\rm{ws}}$ and a larger $n_{\rm{drip}}$, as shown
in Figs.~\ref{fig:7ndrip} and~\ref{fig:8dripR}.
The $L$ dependence of $n_{\rm{drip}}$ can also be explained by
the behavior of the neutron chemical potential $\mu_n$.
At the average baryon density $n_b$, a small $L$ generally
corresponds to a high $\mu_n$ due to the large contribution from
the $\rho$ meson [see Fig.~\ref{fig:13mu}(b)].
Therefore, the model with a smaller $L$ can reach the threshold condition for
the neutron drip $\mu_n = Mc^2$ at a lower density, which implies
an increasing $n_{\rm{drip}}$ with $L$, as shown in Fig.~\ref{fig:7ndrip}.
We display in Fig.~\ref{fig:9dripA} some properties of the nucleus
at the neutron drip density as a function of $L$
obtained in the TF calculation.
As one can see from Fig.~\ref{fig:9dripA}(a), the nucleon number
$A$ of the equilibrium nucleus is almost independent of $L$.
This is because the generated models with different $L$
have fixed symmetry energy at $n_{\rm{fix}}=0.11\, \rm{fm}^{-3}$,
which can produce very similar binding energies for finite nuclei
within one set of generated models (see Fig.~\ref{fig:1Pb208}).
The proton number $Z$ slightly decreases with increasing $L$,
which can be understood from
the $L$ dependence of the surface tension.
As discussed in Refs.~\cite{Oyam07,Bao14,Gril12},
a small $L$ favors a large surface tension $\tau$, which leads to
a large $Z$ since $Z$ increases monotonically with $\tau$.
The average proton fraction $Z/A$ of the nucleus is found to decrease
with increasing $L$ [see Fig.~\ref{fig:9dripA}(b)],
which is caused by the decrease of $Z$ with $L$.
The root-mean-square (rms) radius of the neutron ($R_n$) increases
with $L$, whereas that of the proton ($R_p$)
decreases [see Fig.~\ref{fig:9dripA}(c)].
The difference between $R_n$ and $R_p$,
known as the neutron skin thickness ($\Delta r_{np}$),
is displayed in Fig.~\ref{fig:9dripA}(d).
It is well known that a larger $L$ results in a thicker neutron
skin~\cite{IUFSU,Horo01,Horo03,Gril12,Cent10},
which is also observed in Fig.~\ref{fig:9dripA}(d).
We plot in Fig.~\ref{fig:10dripD} the nucleon density distributions
in the Wigner--Seitz cell at the neutron drip density obtained
with two extreme values of $L$ in the set of TM1.
It is shown that the nucleon distributions, especially the neutron
distributions, can be significantly affected by the value of $L$.
One can see that a large $L$ results in a small nucleon density
at the center of the cell. This may be understood from the
analysis based on a liquid-drop model. As discussed
by Iida and Oyamatsu~\cite{Oyam04}, the equilibrium density
of the nucleon liquid can be estimated by the condition
of zero pressure when there is no neutron gas outside.
The bulk pressure vanishes at
$n^L_b=n_0-\frac{3L n_0}{K} \delta^2$
derived from the liquid-drop model~\cite{Oyam04}.
This implies that the equilibrium density $n^L_b$ decreases
with increasing $L$ for a fixed neutron excess $\delta$.
On the other hand, a larger $L$ in one set of generated models
corresponds to a larger symmetry energy $E_{\rm sym}$ near the
saturation density, as mentioned above.
As a result, a larger $L$ favors fewer neutrons (equivalent
to a smaller $\delta$) in the central region of the nucleus due to
its larger $E_{\rm sym}$. Meanwhile, more neutrons are distributed
in the surface region due to its smaller $E_{\rm sym}$ for a larger $L$
at very low density. Therefore, a large value of $L$ results
in relatively large neutron rms radius and neutron skin thickness
[see Figs.~\ref{fig:9dripA}(c) and~\ref{fig:9dripA}(d)].
\subsection{Properties of nuclei in neutron star crusts}
\label{sec:4.3}
We employ the TF approximation to study the effects of the symmetry
energy on properties of nuclei in the inner crust.
Above the neutron drip density $n_{\rm{drip}}$,
a gas of free neutrons coexists with
a lattice of spherical nuclei, and the equilibrium nuclei
become more and more neutron rich as the density increases.
In Fig.~\ref{fig:11tfdrop}, we display the droplet proton number $Z$,
nucleon number $A_d$, and proton fraction $Z/A_d$
as a function of the average baryon density $n_b$ using the two sets of
generated models. The droplet nucleon number $A_d$ is defined by subtracting
the background neutrons in order to isolate the nucleus from a
surrounding neutron gas~\cite{Gril12,Sub02}.
It is shown that $Z$ and $A_d$ weakly depend on $n_b$ at lower densities,
while they rapidly change at relatively high densities.
For the $L$ dependence of $Z$ and $A_d$, it is found that $Z$
decreases monotonically with increasing $L$, while $A_d$ is almost
independent of $L$ at low densities.
These behaviors are consistent with those shown in Fig.~\ref{fig:9dripA}(a).
Our results are very similar to those reported in Ref.~\cite{Gril12}.
The $L$ dependence of $Z$ may be understood from the behavior of the surface
tension $\tau$. Based on the size equilibrium condition of the liquid-drop model,
${\varepsilon}_{\rm{surf}}=2\,{\varepsilon}_{\rm{Coul}}$,
a large value of $\tau$ leads to large nuclear size $r_d$ and proton number $Z$.
It has been shown in Refs.~\cite{Oyam07,Bao14,Gril12} that
a large $L$ corresponds to a small $\tau$. Therefore,
a small $Z$ is achieved for a large $L$ due to its small $\tau$.
The $L$ dependence of $A_d$ at high densities is mainly because
the nuclear size increases with decreasing $L$ (equivalent to increasing $\tau$).
The proton fraction $Z/A_d$ at low densities is found to decrease
with increasing $L$, which is related to the behaviors of $Z$ and $A_d$,
but the opposite tendency is observed at high densities.
A similar behavior of $Z/A_d$ was also observed in Fig.~4(f) of Ref.~\cite{Gril12}.
The strong $L$ dependence at high densities obtained in the present TF
calculation is consistent with that shown in our previous study
using the CP method~\cite{Bao14}, where a large value of $L$
leads to small $\tau$, $Z$, $A_d$, and $r_d$ values
(see Figs.~4--7 of Ref.~\cite{Bao14}).
It has been shown in Sec.~\ref{sec:4.1} that the difference
between the TF and CP methods is relatively small in the high-density region.
In Figs.~\ref{fig:12ws} and~\ref{fig:13mu}, we present equilibrium
properties of the Wigner--Seitz cell as a function of $n_b$ obtained
in the TF approximation using the two sets of generated models.
For clarity of presentation, we show chemical potentials $\mu_e$,
$\mu_n$, and $\mu_p$ in Fig.~\ref{fig:13mu} with only
the smallest and largest values of $L$ in each set of generated models.
One can see from Fig.~\ref{fig:12ws}(a) that the radius of
the Wigner--Seitz cell, $r_{\rm{ws}}$,
significantly decreases with increasing $n_b$, while the proton
rms radius $R_p$ weakly depends on $n_b$ only at high densities.
These behaviors are consistent with those shown
in Fig.~\ref{fig:5rcd}, where the droplet radius $r_d$ in the TF
approximation is calculated from the proton rms radius $R_p$ as
$r_d=\sqrt{\frac{5}{3}} R_p$.
The decrease of $r_{\rm{ws}}$ is caused by the
increase of nuclear volume fraction $u$ with increasing $n_b$.
On the other hand, the proton density at the center of the cell, $n_p(0)$,
obviously decreases with increasing $n_b$ [see Fig.~\ref{fig:12ws}(c)].
This is because the matter gets more neutron rich and the difference
between the neutron and proton chemical potentials,
which is equivalent to the electron chemical potential as
$\mu_e = \mu_n-\mu_p$, becomes larger as the density
increases [see Fig.~\ref{fig:13mu}(a)].
Moreover, the decrease of $n_p(0)$ at high
densities shows a strong $L$ dependence; namely, a small $L$ leads to
a rapid decrease of $n_p(0)$. This may be understood from
the influence of the $\omega$-$\rho$ coupling term,
which plays an important role in neutron-rich matter.
At the center of the cell, we have the following relation between
densities and chemical potentials according to
Eqs.~(\ref{eq:eqmr})--(\ref{eq:mue}):
\begin{eqnarray}
\mu_e = \mu_n-\mu_p &=& \sqrt{\left( k_{F}^{n}\right)^{2}+{M^{\ast}}^{2}}
-\sqrt{\left( k_{F}^{p}\right)^{2}+{M^{\ast}}^{2}}
-eA -g_{\rho}\rho \nonumber \\
& \simeq & \frac{(3\pi^2)^{2/3}}{2M^{\ast}}\left( n_{n}^{2/3}-n_{p}^{2/3}\right)
-eA +\frac{g_{\rho }^{2}}{2\left(m_{\rho }^{2}
+2\Lambda_{\rm{v}}g_{\omega}^{2}g_{\rho}^{2}{\omega}^{2}\right) }\left( n_{n}-n_{p}\right).
\label{eq:cwr}
\end{eqnarray}
As $n_b$ increases, $\mu_e=\mu_n-\mu_p$ increases monotonically,
as shown in Fig.~\ref{fig:13mu}(a), which yields increasing $n_{n}(0)-n_{p}(0)$
and decreasing $n_{p}(0)$ [see Figs.~\ref{fig:12ws}(b) and~\ref{fig:12ws}(c)].
One can see from Tables~\ref{tab:2} and~\ref{tab:3} that the model with
a small $L$ has relatively large $\Lambda_{\rm{v}}$ and $g_{\rho}$.
Hence, the last term of Eq.~(\ref{eq:cwr}) can make a more significant
contribution in the case of small $L$, which may be the main reason for
the high $\mu_e$ and the rapid decrease of $n_p(0)$,
corresponding to small $L$ at high densities
[see Figs.~\ref{fig:13mu}(a) and \ref{fig:12ws}(c)].
Furthermore, large $\Lambda_{\rm{v}}$ and $g_{\rho}$, corresponding to
small $L$, results in high $\mu_n$ and low $\mu_p$,
as shown in Figs.~\ref{fig:13mu}(b) and \ref{fig:13mu}(c).
For the neutron density at the center, $n_{n}(0)$, and that at the boundary,
$n_{n}(r_{\rm{ws}})$, plotted in Fig.~\ref{fig:12ws}(b), it is seen that
the model with a larger $L$ predicts smaller $n_{n}(0)$ and
larger $n_{n}(r_{\rm{ws}})$, which are more pronounced at high densities.
The behaviors of $n_{n}(0)$ and $n_{n}(r_{\rm{ws}})$ obtained in the present
study are consistent with those reported in Refs.~\cite{Oyam07,Gril12}.
The $L$ dependence of $n_{n}(0)$ and $n_{n}(r_{\rm{ws}})$ can be understood
from the density dependence of the symmetry energy $E_{\rm sym}$.
In one set of generated models, $E_{\rm sym}$ has the same value at
$n_{\rm{fix}}=0.11\, \rm{fm}^{-3}$ for different $L$. However, a larger $L$
in one set of generated models corresponds to a larger $E_{\rm sym}$ at
higher density in the center region and to a smaller $E_{\rm sym}$ at lower
density in the neutron gas outside. Therefore, a larger $L$ favors a more
diffuse neutron distribution, which results in smaller $n_{n}(0)$
and larger $n_{n}(r_{\rm{ws}})$, as shown in Fig.~\ref{fig:12ws}(b).
This tendency can be seen more clearly in Fig.~\ref{fig:14wsD},
in which the density profiles are plotted with two extreme values
of $L$ in the set of TM1 at several average baryon densities.
We conclude that a larger $L$ in one set of generated models
predicts a higher neutron drip density $n_{\rm{drip}}$ due to its lower
neutron chemical potential $\mu_n$. Moreover, with increasing density,
neutrons drip out more easily for the model with a larger $L$ due to
its lower $E_{\rm sym}$ in the dilute neutron gas.
As a result, a larger value of $L$ predicts a higher neutron gas
density $n_{n}(r_{\rm{ws}})$ in the high-density region.
We show in Fig.~\ref{fig:14wsD} the density distributions of neutrons and
protons in the Wigner--Seitz cell at different average baryon density $n_b$
with two extreme values of $L$ in the set of TM1.
As $n_b$ increases, $r_{\rm{ws}}$ clearly decreases
and the neutron density outside becomes much larger.
On the other hand, the proton density at the center
decreases significantly with increasing $n_b$, while the neutron density
at the center does not change very much for different $n_b$.
However, the distributions of protons and neutrons
become more diffuse at higher density.
One can see that the differences between $L=40$ MeV and $L=110.8$ MeV
significantly increase with increasing $n_b$.
The nuclear size obtained with $L=40$ MeV is larger than that with
$L=110.8$ MeV, especially for the case of $n_b=0.05\,\rm{fm}^{-3}$,
which leads to larger $Z$ and $A_d$, as shown in Fig.~\ref{fig:11tfdrop}.
Furthermore, the neutron distributions with $L=110.8$ MeV are
more diffuse than those with $L=40$ MeV, which can be explained
by the density dependence of the symmetry energy $E_{\rm sym}$,
as discussed above. It is clearly seen that the neutron gas density
with $L=110.8$ MeV increases more rapidly than that with $L=40$ MeV,
which is also observed in Fig.~\ref{fig:12ws}(b).
Since the Coulomb interaction is self-consistently taken into account
in the TF approximation, it is seen that the proton distributions are influenced
by the Coulomb potential; namely, the proton densities at the center of the cell
are slightly lower than those at the surface region due to
the repulsive Coulomb potential.
\section{Conclusions}
\label{sec:5}
We have investigated the effects of the symmetry energy on the neutron
drip density and properties of nuclei in neutron star crusts.
The Wigner--Seitz approximation has been employed to describe the
nonuniform matter around the neutron drip density.
For the nuclear interaction, we have adopted the RMF theory
with several successful parametrizations.
We have considered and compared three different methods for
calculating properties of neutron star crusts,
namely, the self-consistent TF approximation,
the simple CP method with bulk Gibbs equilibrium conditions,
and the CLD model with equilibrium conditions determined by
including the surface and Coulomb energies.
It has been found that the simple CP method fails to describe the
nonuniform matter around the neutron drip density due to its higher
energies than that of homogeneous matter.
The failure of the CP method is mainly because the finite-size effects
due to the surface and Coulomb energies are too large to be treated
perturbatively at low densities, so that they should be included
self-consistently in determining the equilibrium state, as done in
the CLD and TF methods. The results of the CLD method have been
greatly improved by the inclusion of finite-size effects
compared to those of the CP method.
We have made a detailed comparison of the three methods
and concluded that the inclusion of surface and Coulomb
energies in determining the equilibrium state plays a
crucial role in the description of nonuniform matter
at low density.
We have examined the influence of the symmetry energy slope $L$
using two sets of models generated from the TM1 and IUFSU
parametrizations. All models in each set have the same isoscalar
saturation properties and fixed symmetry energy
at $n_{\rm{fix}}=0.11\, \rm{fm}^{-3}$, but they have different symmetry
energy slope $L$. The choice of $n_{\rm{fix}}=0.11\, \rm{fm}^{-3}$
can produce very similar binding energies for finite nuclei
within one set of generated models.
We have performed the self-consistent TF calculation to study
the influence of the symmetry energy slope $L$ on the neutron
drip density $n_{\rm{drip}}$. It has been found that $n_{\rm{drip}}$
increases with increasing $L$, which is related to the decrease of
the Wigner--Seitz cell radius. At the neutron drip point,
the proton fraction of the equilibrium nucleus is found to decrease
with increasing $L$, while the neutron skin thickness shows
an obvious increase with increasing $L$.
The $L$ dependence of the equilibrium nucleus at the neutron
drip density is qualitatively consistent with that obtained
in finite-nuclei calculations.
We have studied the effects of the symmetry energy on properties
of nuclei in the inner crust within the TF approximation.
It has been found that the proton number $Z$ and the nucleon
number $A_d$ of the droplet weakly depend on the average baryon
density $n_b$ in the low-density region, while they rapidly change
at relatively high densities. For the $L$ dependence of $Z$ and $A_d$,
it has been shown that $Z$ decreases monotonically with
increasing $L$, while $A_d$ is almost independent of $L$
at low densities. On the other hand, a strong $L$ dependence
has been observed for properties of the equilibrium nucleus
at high densities. The results obtained in the present
self-consistent TF calculation are qualitatively consistent
with those found in the literature~\cite{Oyam07,Bao14,Gril12}.
We note that nuclear shell and paring effects have been
neglected in the present work. It would be interesting
to consider these effects in future studies.
\section*{Acknowledgment}
This work was supported in part by the National Natural Science Foundation
of China (Grant No. 11375089).
\newpage
|
\section{Введение}
Упаковка шаров в пространстве называется {\it жесткой} или, иногда, {\it локально-жесткой}, если они расположены так, что каждый шар из упаковки зажат соседями и его нельзя сдвинуть в сторону с тем, чтобы увеличить минимальное расстояние между центром этого шара и центрами других шаров.
Рассмотрим $N$ не пересекающихся шаров одинакового радиуса $r$ в ${\Bbb R}^3$, которые расположены так, что все они касаются одного (центрального) шара единичного радиуса. Обозначим через $P:=\{A_{1},\ldots,A_{N}\}$ точки касания внешними шарами центрального шара. Соединим точки $A_{i}$ и $A_{j}$ ребром (минимальной дугой большого круга), если соответствующие внешние шары касаются. Полученный граф будем называть {\it контактным} и обозначать $\cg(P)$. Если же эта упаковка на ${\Bbb S}^2$ является локально-жесткой, то будем называть граф $\cg(P)$ {\it неприводимым}.
Таким образом, задача изучения жестких упаковок локально сводится к изучению неприводимых графов.
Имеется связь этой геометрической задачи с другими задачами упаковки шаров. Одно из основных приложений вне математики - это материаловедение, где рассматриваются локально-жесткие упаковки твердыми телами и наночастицами (см. например, \cite{AMB,Appl04, Appl10}). Заметим также, что большинство конфигураций физических частиц, задающих минимум потенциальной энергии тоже являются локально-жесткими.
В математике - В. Хабихт, К. Шютте, Б. Л. ван дер Варден, и Л. Данцер применяли неприводимые контактные графы к проблеме контактных чисел и проблеме Таммеса \cite{HabvdW, SvdW1, vdW, SvdW2, Dan}. Недавно с помощью этого метода мы решили проблему Таммеса для $N=13$ и $N=14$ \cite{MT,MT14}.
В этой статье мы показываем, что с помощью списка неприводимых контактных графов можно решать различные задачи об экстремальных упаковках. Здесь мы рассматриваем задачи Таммеса для сферы и проективной плоскости, задачу о наибольшем числе ребер у сферического контактного графа, задачи Данцера и другие.
\section{Неприводимые контактные графы}
\subsection{Основные определения}
Обозначим через ${\Bbb S}^{2}$ единичную сферу в ${\Bbb R}^3$. Для точек $x$ и $y$ на сфере, $\dist(x,y)$ - это расстояние в угловом измерении.
Пусть $X$ - конечное подмножество единичной сфере ${\Bbb S}^{2}$. Обозначим
$$\psi(X):=\min\limits_{x,y\in X}{\{\dist(x,y)\}}, \mbox{ где } x\ne y.$$
Пусть $d_N$ обозначает наибольшее значение $\psi(X)$, которое может достигаться для
$X\subset{\Bbb S}^{2}$ с $|X|=N$, т. е.
$$
d_N:=\max\limits_{X\subset{\Bbb S}^2}{\{\psi(X)\}}, \, \mbox{ при } \; |X|=N.
$$
\noindent{\bf Контактные графы.} Как и выше, $X$ - конечное подмножество ${\Bbb
S}^2$. {\it Контактным графом} $\cg(X)$ называется граф на ${\Bbb S}^2$ с вершинами в $X$ и ребрами (дугами) $xy, \, x,y\in X$, минимальной длины, т. е. с $\dist(x,y)=\psi(X)$.
\medskip
\noindent{\bf Сдвиг вершины.}
Будем говорить, что вершину $x\in X$ можно {\it сдвинуть в сторону}, если в любой открытой окрестности $x$ найдется такая точка $x'\in {\Bbb S}^2$, что $$\dist(x',X\setminus\{x\})>\dist(x,X\setminus\{x\}).$$
\medskip
\noindent{\bf Неприводимые графы.}
Назовем контактный граф $\cg(X)$ {\it неприводимым}, если ни одну из его вершин нельзя сдвинуть в сторону.
Этот термин используется в работах \cite{SvdW1,SvdW2,FeT,Dan}.
Давайте обозначим через $J_N$ семейство всех конечных множеств $X$ в ${\Bbb S}^2$ с $|X|=N$ таких что $\cg(X)$ является неприводимым.
\medskip
\noindent{\bf Д-неприводимые графы.}
Л. Данцер \cite[Sec. 1]{Dan} определил следующую операцию. Пусть
$x,y,z$ - вершины $\cg(X)$ с $\dist(x,y)=\dist(x,z)=\psi(X)$. Обозначим через $x^0$ зеркальный образ точки $x$
относительно дуги большого круга $yz$ (см. Рис.~\ref{fig3}). Мы назовем эту операцию
{\it Д-отражением}, если $\dist(x^0,X\setminus\{x,y,z\}) > \psi(X)$.
Неприводимый контактный граф $\cg(X)$ будем называть Д-неприводимым, если он не допускает ни одного Д-отражения.
\medskip
\begin{figure}[h]
\begin{center}
\includegraphics[clip,scale=1]{pics/13-6.mps}
\end{center}
\caption{Д-отражение}
\label{fig3}
\end{figure}
\medskip
\noindent{\bf Максимальные графы.} Предположим, что для $X\subset{\Bbb S}^{2}$ с $|X|=N$ имеет место равенство $\psi(X)=d_N$. Будем тогда называть контактный граф $\cg(X)$ - {\it максимальным.}
\subsection{Свойства неприводимых контактных графов.}
Здесь мы рассмотрим такие подмножества $X\subset {\Bbb S}^2$, что граф $\cg(X)$ является неприводимым, т.е. $X\in J_N$.
Следующие свойства неприводимых графов были опубликованы в работах \cite{SvdW1}, \cite{Dan}, и \cite{BS,BS14} (см. также \cite[Глава VI]{FeT}).
Пусть $a,b,x,y\in X$ с $\dist(a,b)=\dist(x,y)=\psi(X)$.
Тогда кратчайшие дуги ${ab}$ и ${xy}$ не пересекаются. В противном случае,
длина одной из дуг $ax, ay, bx, by$ будет меньше чем $\psi(X)$. Из этого вытекает
\begin{prop} Если $X$ - конечное подмножество ${\Bbb S}^2$, то $\cg(X)$ является планарным графом.
\end{prop}
\begin{prop} Если $X\in J_N$, то все грани $\cg(X)$ являются выпуклыми в ${\Bbb S}^2$.
\end{prop}
\begin{prop} Если $X$ является максимальным, то для $N>5$ граф $\cg(X)$ является Д-неприводимым и, в частности, неприводимым.
\end{prop}
\begin{prop} Если $X\in J_N$, то степени вершин графа $\cg(X)$ могут быть только $0$ (изолированные вершины), $3$, $4$, или $5$.
\end{prop}
\begin{prop} Если $X\in J_N$, то грани $\cg(X)$ являются многоугольниками не более чем с $\lfloor2\pi/\psi(X)\rfloor$ вершинами.
\end{prop}
Следующее свойство было найдено Бёрёцким и Сабо {\cite[Lemma 8 and Lemma 9(iii)]{BS}}.
\begin{prop} Пусть $X\in J_N, \, N>10$. Если $\cg(X)$ содержит изолированные вершины, то эти вершины лежат внутри граней $\cg(X)$ с шестью или более вершинами. Более того, внутри шестиугольника не может лежать две изолированные вершины.
\end{prop}
Комбинируя эти предложения вместе, получаем следующие комбинаторные свойства неприводимых контактных графов:
\begin{cor}\label{cor1} Если $X\in J_N$, то $G:=\cg(X)$ удовлетворяет следующим свойствам
\begin{enumerate}
\item $G$ является планарным графом;
\item У любой вершины $G$ степень равна $0,3,4,$ или $5$;
\item Если $N>10$ и у $G$ есть изолированная вершина $v$, то $v$ лежит в грани с $m\ge 6$ вершинами. Гексагональная грань $G$ не может содержать две изолированные вершины.
\end{enumerate}
\end{cor}
\subsection{Работа Л. Данцера по неприводимым графам}
В работе \cite{Dan} Людвиг Данцер приводит решение проблемы Таммеса для $N=10$ и $N=11$. Эта статья - английский перевод его докторской диссертации: ``Endliche Punktmengen auf der 2-sph\"are mit m\"oglichst gro{\ss}em Minimalabstand'', Universit\"at G\"othingen, 1963. В этой работе, в частности, к понятию приводимости было добавлено понятие Д-отражения и Д-приводимости. (В статье применяется другая терминология и мы используем здесь Д в честь Л. Данцера.)
\medskip
В своей работе \cite{Dan} Данцер также приводит список Д-неприводимых графов для $6\le N \le 10$. Так как максимальные контактные графы являются Д-неприводимыми, то в этот список включены и графы, дающие решение проблемы Таммеса для этих $n$. Заметим, что для случая $N=11$ Данцер рассматривал только максимальные контактные графы.
\subsection{Перечисление неприводимых контактных графов} Пусть конечное множество точек $X\subset {\Bbb S}^2$ такое, что контактный граф $\cg(X)$ является неприводимым. В Следствии 2.1 мы собрали вместе комбинаторные свойства $\cg(X)$. Имеется целый ряд геометрических свойств этих графов.
Напомним, что все грани $\cg(X)$ являются выпуклыми (Предложение 2.2). Поскольку все ребра $\cg(X)$ одинаковой длины $\psi(X)$, то его грани - выпуклые равносторонние сферические многоугольники с числом вершин не превосходящим $\lfloor2\pi/\psi(X)\rfloor$.
Рассмотрим теперь планарный граф $G$, с заданными гранями $\{F_k\}$, который удовлетворяет Следствию 2.1. Мы будем рассматривать вложения этого графа в ${\Bbb S}^2$ как неприводимого контактного графы $\cg(X)$ для некоторого $X\subset {\Bbb S}^2$.
Вложение графа $G$ в ${\Bbb S}^2$ однозначно задается следующим набором параметров (переменных):\\
(i) Длина ребра $d$;\\
(ii) Наборы углов $u_{ki}$, $i=1,\ldots,m_k$ граней $F_k$. (Здесь $m_k$ обозначает число вершин у грани $F_k$.)
В нашей работе по перечислению графов были рассмотрены основные геометрические соотношения между этими параметрами (\cite[Предложение 4.1]{MT2013}).
\medskip
Алгоритм перечисления неприводимых контактных графов состоит из двух частей:\\
(I) На первом этапе составляется список $L_{N}$, состоящий из всех графов с $N$ вершинами и удовлетворяющий Следствию 2.1;
Чтобы создать список $L_{N}$ мы используем программу {\it plantri}
(см. \cite{PLA1,PLA2}). Эта программа является генератором не изоморфных планарных графов различных классов, включая триангуляции и другие разбиения на выпуклые многоугольники.
Заметим. что число графов в $L_{N}$ с ростом $N$ быстро возрастает. Например, при $N=6,7,8$, $|L_N|=7,34, 257$, а уже
$|L_{13}|=94754965$
\medskip
\noindent(II) Используя линейную аппроксимацию соотношений из \cite[Предложения 4.1]{MT2013}, из $L_{N}$ удаляются все графы, которые не могут быть вложены в сферу. Оставшиеся графы после дополнительной проверки с помощью солверов и оценки границ изменения параметров заносятся в список неприводимых контактных графов.
\medskip
Основной результат работы \cite{MT2013} приведен в Приложении.
\section{Контактные числа и проблема Таммеса}
\subsection{Контактные числа}
{\it Контактным числом} $k(n)$ называют наибольшее число не пересекающихся шаров одинакового радиуса в ${\Bbb R}^n$, которые можно расположить так, чтобы все они касались одного (центрального) шара такого же радиуса.
Очевидно, что $k(2)=6$. В трехмерном пространстве, в задаче о контактных числах спрашивается: ``Как много белых бильярдных шаров могут одновременно касаться черного бильярдного шара?''
Наиболее симметричная конфигурация, 12 бильярдных шаров вокруг одного, это когда центры 12 шаров расположены в вершинах правильного икосаэдра, а центральный шар расположен в центре икосаэдра. Однако, эти 12 внешних шаров не касаются друг друга и могут свободно перемещаться по поверхности центрального шара. Таким образом, возможно, что эти 12 шаров можно сдвинуть в одну сторону, так что найдется место для 13-го шара?
Этот вопрос был предметом спора между И. Ньютоном и Д. Грегори в 1694 году. Ньютон считал, что $k(3)=12$, в то время как Грегори думал, что ответ может быть равен 13. Эту задачу Ньютона - Грегори часто называют {\it проблемой тринадцати шаров.}
Несложно видеть, что проблема тринадцати шаров сводится к следующей задаче: {\it Доказать, что на единичной сфере ${\Bbb S}^2$ нельзя расположить 13 точек так, чтобы расстояния между ними были не меньше чем $60^\circ$ в угловом измерении.}
Проблема тринадцати шаров оказалось достаточно трудной и была решена только в 1953 году. К. Шютте и Б.Л. Ван дер Варден \cite{SvdW2} доказали, что Ньютон был прав и $k(3)=12$. Доказательство Шютте -- ван дер Вардена основано на неприводимых контактных графах. Ими было показано, что что на единичной сфере ${\Bbb S}^2$ не найдется контактного графа ${\Gamma}$ с ребрами одинаковой длины, которая не меньше чем $60^\circ$.
В 1956 году Дж. Лич \cite{Lee} напечатал двухстраничный набросок элегантного доказательства. Это доказательство было приведено в первом издании известной книги М. Айгнера и Г. Циглера ``Доказательства из Книги'' \cite{AZ}. Однако эта глава была исключена из книги при втором издании, так как авторам не удалось привести подробное доказательство без громоздких вычислений, основанных на сферической тригонометрии. В последние 12 лет было опубликовано несколько новых решений этой старой проблемы
\cite{Hs, Ma, Manew, Bor,Ans, Mus13}.
Заметим, что проблема контактных чисел решена только для размерностей $n=3,4,8$ и $24$ (см. \cite{BDM, Mus1, Mus3, Mus4}). Доказательства в этих работах основаны на методе Дельсарта и его обобщениях.
\subsection{Проблема Таммеса}
У проблемы 13 шаров имеется естественное обобщение: найти расположение множества $X$, состоящего из $N$ точек на ${\Bbb S}^{2}$, такое что минимальное расстояние между точками $X$ - максимально возможное. Эту задачу впервые поставил голландский ботаник Таммес \cite{Tam} (см. также \cite[Section 1.6: Problem 6]{BMP}).
Задача Таммеса решена только для нескольких значений $N$: для $N=3,4,6,12$ ее решил Л. Фейеш Тот \cite{FeT0}; для
$N=5,7,8,9$ - Шютте и ван дер Варден \cite{SvdW1}; для $N=10,11$ - Данцер \cite{Dan} (для $N=11$ см. также \cite{Bor11}) и для $N=24$ - Робинсон \cite{Rob}. Недавно мы решили эту задачу для $N=13$ \cite{MT} и для $N=14$ \cite{MT14}.
В работе Л. Фейеша Тота \cite{FeT0} (см. также его книгу \cite{FeT}) была найдена верхняя оценка для $d_N$. (Напомним, что $d_N$ обозначает наибольшее значение $\psi(X)$, которое может достигаться для
$X\subset{\Bbb S}^{2}$ с $|X|=N$.)
$$
d_N\leqslant \arccos{\frac{\ctg^2{w_N}-1}{2}}, \; \mbox{ где } \; w_N:=\frac{\pi N}{6N-12}.
$$
Эта формула является точной для $N=3,4,6,12$ и решением проблемы Таммеса для этих $N$ соответственно являются правильный треугольник на экваторе, правильный тетраэдр, правильный октаэдр и правильный икосаэдр.
Р. М. Робинсон \cite{Rob} обобщил оценку Фейеша Тота и решил проблему Таммеса для $N=24$.
Оказывается, что $d_5=d_6=90^\circ$ \cite{SvdW1, FeT}. При $N=5$ максимальное расположение на сфере не является единственным с точностью до изометрии. Если взять две точки расположенные в северном и южном полюсах сферы, а оставшиеся три точки расположить на экваторе так, чтобы (угловое) расстояния между ними было бы не меньше $90^\circ$, то получим максимальное расположение.
Для решения проблемы Таммеса при $7\leqslant N \leqslant 11$ и при $N=13, 14$ применялись неприводимые контактные графы. Заметим, что контактный граф у максимальной конфигурации должен быть неприводимым. Поэтому, максимальный граф находится среди неприводимых контактных графов. Более того, он является Д-неприводимым \cite{Dan}.
Можно значительно сократить перебор и список ``допустимых'' неприводимых графов, если добавить условие: $d\geqslant\delta_N$, где $\delta_N$ обозначает длину ребра у гипотетического максимального графа. (Собственно, проблема состоит в доказательстве равенства $d_N=\delta_N$.) Большой набор примеров имеется в книге \cite[Глава VI, §4]{FeT} и в таблице Н. Слоэна: http://neilsloane.com/packings/dim3/. В этой таблице приведены возможные максимальные конфигурации вплоть до $N=130$.
Из Таблиц 7.1--7.6 в Приложении получаем следующую таблицу:
\medskip
\begin{tabular}{ccccccc}
$N$ & $6$ & $7$ & $8$ & $9$ & $10$ & $11$\\
$I_N$ & $2$ & $2$ & $4$ & $10$ & $30$ & $38$\\
\end{tabular}
\medskip
Здесь $I_N$ обозначает число неприводимых контактных графов с $N$ вершинами.
Поскольку, для $N=7,8,9$ числа $I_N$ не очень большие, то в работе \cite{SvdW1} удалось отбросить неподходящие графы без особой суеты. При $N=10$ число $I_N=30$, что довольно много. В этом случае, также как и при $N=11$, Л. Данцер рассматривал только Д-неприводимые контактные графы. В наших работах по проблеме Таммеса для $N=13$ \cite{MT} и для $N=14$ \cite{MT14} мы проводили компьютерный перебор подходящих, т. е. из списка $L_N$, Д-неприводимых контактных графов.
Интересно, что $d_{11}=d_{12}=\arccos{(1/\sqrt{5}})$. Для $N=11$ максимальная конфигурация получается из вершин правильного икосаэдра удалением одной из них.
Заметим, что во всех решенных случаях проблемы Таммеса, максимальное расположение оказалось единственным с точностью до изометрии. Можно, конечно, предположить, что так будет для всех $N$. Однако, в задаче, аналогичной проблеме Таммеса, для плоского тора (периодические упаковки конгруэнтных кругов) при $N=7$ оказалось, что имеется три разные максимальные упаковки \cite{MN}. Это обстоятельство делает проблему Таммеса для сферы еще более интригующей.
В заключении этого раздела отметим, что прямой подход по решению проблемы Таммеса, основанный на компьютерном перечислении неприводимых контактных графов, т. е. такой как в наших работах \cite{MT,MT14}, себя практически исчерпал. Для $N=15,16,...$ компьютерный перебор может занимать многие месяцы. Здесь нужны новые подходы и идеи.
\section{Проблема Таммеса для антиподальных конфигураций}
Если рассмотреть сферу ${\mathbb S}^d$ как множество единичных векторов $x$ в ${\mathbb R}^{d+1}$, то точки $x$ и $y=-x$ называются {\it антиподальными}, а отображение
$x \to -x$ называется {\it антиподальным} на ${\mathbb S}^d$. Множество $X\subset {\mathbb S}^d$ называется {\it антиподальным}, если антиподальное отображение переводит $X$ в себе, $X=-X$, иными словами, если $x\in X$, то и $-x\in X$.
Рассмотрим теперь проблему Таммеса для антиподальных множеств на ${\mathbb S}^2$. Пусть $X$ -- антиподальное множество на сфере. Тогда $X$ содержит четное число точек и, стало быть, $|X|=2M, \, M=1,2,...$. Определим для таких множеств аналог величины $d_N$.
$$
a_M:=\max\limits_{X=-X\subset{\Bbb S}^2}{\{\psi(X)\}}, \, \mbox{ при } \; |X|=2M.
$$
{\it Проблема Таммеса для антиподальных множеств} состоит в том, чтобы найти все конфигурации $X=\{x_1,-x_1,\ldots,x_M,-x_M\}$ на ${\mathbb S}^2$, чтобы $\psi(X)=a_M$.
Если отождествить на сфере ${\mathbb S}^2$ антиподальные точки, то получим проективную плоскость ${\Bbb R\Bbb P}^2$. Поэтому, по сути, эта проблема о конфигурациях на проективной плоскости у которых минимальное расстояние между точками является максимально возможным.
Рассмотрим простейшие свойства $a_M$ и оптимальных конфигураций.
\begin{lemma}
\begin{enumerate}
\item $a_M\leqslant d_{2M}$;
\item Если $|X|=2M, \, \psi(X)=d_{2M}$ и $X$ -- антиподальное множество, то $a_M = d_{2M}$. Следовательно, если $X=-X$ и $X$ является решением проблемы Таммеса для $N=2M$, то $X$ также является решением и задачи Таммеса для антиподальных конфигураций;
\item Если $X\subset{\Bbb S}^2$ -- антиподальное множество, $|X|=2M$ и $\psi(X)=a_{M}$, то $\cg(X)$ является Д-неприводимым контактным графом.
\end{enumerate}
\end{lemma}
\begin{proof} Доказательства 1 и 2 вытекают непосредственно из определений.
3. Будем доказывать это утверждение от противного. Для антиподального множества $X$ сдвиг вершины и Д-отражение мы будем делать одновременно для вершины $v$ и ее антипода $-v$, т. е. мы сохраняем антиподальность. Тогда, если конфигурация позволяет такую операцию, то сдвигая еще точки мы получим новое антиподальное $X'$ у которого $\psi(X')>\psi(X)$, -- противоречие.
\end{proof}
\begin{thm} Пусть $X_M\subset{\Bbb S}^2$ является решением задачи Таммеса для антиподальных конфигураций, т. е. $\psi(X_M)=a_M$. Тогда
\begin{enumerate}
\item $X_2$ -- множество вершин квадрата на экваторе, $a_2=90^\circ$;
\item $X_3$ -- множество вершин правильного октаэдра, $a_3=90^\circ$;
\item $X_4$ -- множества вершин куба, $a_4=\arccos{(1/3)}$;
\item $X_5$ -- множество, состоящее из пяти пар антиподальных вершин правильного икосаэдра, $a_5=\arccos{(1/\sqrt{5})}$.
\item $X_6$ -- множество вершин правильного икосаэдра, $a_6=\arccos{(1/\sqrt{5}})$.
\end{enumerate}
\end{thm}
\begin{proof} 1. В случае $M=2$ у нас имеется две пары антиподальных точек и поэтому эти четыре точки лежат на большой окружности. Следовательно, вершины квадрата являются оптимальным расположением этих точек.
Пункты 2 и 5 теоремы вытекают из Леммы 4.1 (2).
Лемма 4.1 (3) позволяет выбрать $X_M$ из списка неприводимых контактных графов. Таблицы 7.3 и 7.5 доказывают пункты 3 и 4 теоремы.
\end{proof}
\section{Задача о наибольшем числе контактов}
Пусть $X$ конечное множество в произвольном метрическом пространстве ${\bf M}$ с расстоянием $\dist$.
Как и раньше,
$$\psi(X):=\min\limits_{x,y\in X}{\{\dist(x,y)\}}, \mbox{ где } x\ne y.$$
Обозначим через $e(X)$ число пар $(x,y)$ в $X$ у которых $\dist(x,y)=\psi(X)$. (Иными словами, $e(X)$ это число ребер у графа $\cg(X)$) Определим максимальное контактное число (число ребер) для заданного числа точек.
$$
\ko({\bf M},N):=\max\limits_{X\in {\bf M}, |X|=N}{e(X)}.
$$
В {\it задаче о наибольшем количестве контактов} спрашивается как найти числа $\ko({\bf M},N)$ или хотя бы оценить их?
В случае когда ${\bf M}={\Bbb S}^2$ будем обозначать величину $\ko({\bf M},N)$ через $\ko_N$.
Это определение можно обобщить. Рассмотрим $N$ не пересекающихся кругов (сферических шапочек) диаметра $d$ на сфере ${\Bbb S}^2$. Обозначим через $\ko_N(d)$ максимально возможное число касаний этих кругов.
Заметим, что
$$
\ko_N=\max\limits_{d\leqslant d_N}{\ko_N(d)}.
$$
В работе \cite{P12} рассматривалась эта задача для $N=12$ и $d=60^\circ$. Было доказано, что $\ko_N(d)=24.$
Начнем с простейших свойств множеств с максимальным числом контактов.
\begin{lemma} Пусть $\Gamma_N$ обозначает контактный граф $\cg(X)$ множества $X$ на сфере, с $|X|=N$, у которого число ребер равно $\ko_N$. Тогда $\Gamma_N$ является планарным графом и степень любой вершины этого графа равна $2, 3, 4$ или $5$.
\end{lemma}
\begin{proof} Планарность следует из Предложения 2.1.
У графа $\Gamma_N$ не может быть ни изолированных ($\deg(v)=0$), ни висячих вершин ($\deg(v)=1$), так как в противном случае мы можем придвинуть эту вершину к другим и увеличить число контактов до двух.
Неравенство $\deg(v)<6$ следует из известного факта, что у равнобедренного сферического треугольника
$ABC$ с $|AB|=|AC|=d$ и $|BC|\ge d$ угол $\angle BAC > \pi/3=60^\circ$. Предположим, что у вершины $v$, $\deg(v)\geqslant 6$. Тогда $60^\circ\deg(v)\geqslant 360^\circ$. С другой стороны, сумма углов при вершине $v$ равна $360^\circ$, а поскольку каждый угол больше чем $60^\circ$, то $360^\circ>60^\circ\deg(v)$, - противоречие. Отсюда вытекает требуемое неравенство.
\end{proof}
Очевидно, что $\ko_2=1$. Для больших $N$ имеет место следующее неравенство.
\begin{thm} Пусть $N>2$. Тогда $$\ko_N\leqslant 3N-6$$ и равенство достигается только при $N=3,4,6,12$.
\end{thm}
\begin{proof} Рассмотрим контактный граф $\Gamma_N$ с максимальным числом контактов. Пусть у $\Gamma_N$ на сфере будет $F$ граней. Тогда по формуле Эйлера получаем
$$
N - \ko_N +F=2.
$$
Поскольку у каждой грани не менее трех сторон, то
$F\leqslant2\ko_N/3$. Из этого неравенства и формулы Эйлера вытекает требуемое.
Если $N=3,4,6,12$, то равенство соответственно достигается для правильного треугольника на экваторе, правильного тетраэдра, правильного октаэдра и правильного икосаэдра.
Заметим, что равенство получается только в том случае, если все грани вложения $\Gamma_N$ на сферу являются конгруэнтными правильными сферическими треугольниками. Степени всех вершин должны быть равны. Но тогда при $N>3$ вершины задают правильные многогранники с треугольными гранями. Известно, что всего у трех из пяти правильных многогранников грани треугольные и они перечислены выше.
\end{proof}
Поскольку из теоремы следует, что $\ko_5 < 9$, то получаем следующее утверждение:
\begin{cor} $\ko_5=8$. Такое число получается когда одна точка находится в ``северном полюсе'', а оставшиеся четыре лежат на экваторе и образуют квадрат.
\end{cor}
Таким образом, мы знаем решение задачи о максимальном количестве контактов на сфере для $N\leqslant 6$ и $N=12$. Разберем теперь случаи $7\leqslant N \leqslant 11$.
Множество с максимальным количеством контактов не обязано быть из $J_N$, т. е. его контактный граф не обязан быть неприводимым. Обозначим через $I_N$ множество состоящее из вершин правильного икосаэдра на сфере из которых удалено $12-N$ соседних вершин. Это означает, что $I_{11}$ это множество вершин икосаэдра без одной вершины, $I_{10}$ - множество $I_{12}$ из которого удалены две вершины соединенные ребром, а $I_{9}$ это вершины икосаэдра без трех вершин, лежащих на одной грани. Несложно видеть, что $e(I_N)=3N-9$ при $N<11$.
Поскольку $e(I_{10})=21$, то $\ko_{10}\geqslant 21$. С другой стороны, в таблице 7.5 - неприводимых графов для $N=10$, графами с наибольшим числом ребер являются 7.5.28 и 7.5.29 с $e=20.$ (Заметим, что эти графы, полученные удалением двух не соседних вершин икосаэдра, не являются максимальными для задачи Таммеса.) Следовательно, при $N=10$ максимальное число ребер достигается на графе, который является приводимым.
Определим еще одну величину $\ko_N^*$ - максимальное число ребер у контактного неприводимого графа с $N$ вершинами,
$$
\ko^*_N:=\max\limits_{X\in J_N}{e(X)}.
$$
\begin{thm}
\begin{enumerate}
\item $\ko^*_7=\ko_7=12;$
\item $\ko^*_8=\ko_8=16;$
\item $\ko^*_9=\ko_9=18;$
\item $\ko^*_{10}=20, \, \ko_{10}=21;$
\item $ \ko^*_{11}=\ko_{11}=25.$
\end{enumerate}
\end{thm}
Доказательство теоремы следует из двух лемм о контактных графах.
\begin{lemma} Пусть $X$ - конечное множество на сфере ${\Bbb S}^2$. Если любая грань контактного графа $\cg(X)$
является треугольником или четырехугольником, то этот граф неприводим.
\end{lemma}
\begin{proof} Заметим, что граф $\cg(X)$ на сфере является приводимым если у него найдется невыпуклая грань. Однако, у контактного графа все грани являются равносторонними, и в нашем случае это либо равносторонний треугольник, либо ромб. В обоих случаях это выпуклые многоугольники.
\end{proof}
\begin{lemma} Пусть $X\subset{\Bbb S}^2$ и $|X|=N,\, N>6$. Предположим, что $e(X)\geqslant 3N-8$. Тогда контактный граф $\cg(X)$ является неприводимым.
\end{lemma}
\begin{proof} Из формулы Эйлера следует, что у графа с $N$ вершинами на сфере, все грани которого являются треугольниками, число ребер $e$ равно $3N-6$. По условию $e=3N-6, \, 3N-7$ или $3N-8$. В первом случае, у нас все грани треугольные, а во втором одна из граней четырехугольная. В обоих случаях Лемма 5.2 доказывает неприводимость контактного графа.
Остается третий случай, когда $e=3N-8$. Здесь возможно два подслучая: либо имеется ровно две четырехугольных грани, но тогда опять можно применить Лемму 5.2, либо одна пятиугольная грань $ABCDE$.
Предположим, что граф $\cg(X)$ приводим. Тогда сферический равносторонний пятиугольник $ABCDE$ невыпуклый. Без ограничения общности, можно считать, что внутренний угол вершины $A$ больше или равен $180^\circ$, а все остальные углы меньше $180^\circ$. Тогда $A$ может быть соединено ребром только с одной вершиной $F$. Это следует из того факта, что все углы больше $60^\circ$ и в случае если бы были еще вершины, кроме $F$, $B$ и $E$, соединенные с $A$, то внешний угол $A$ был бы больше $180^\circ$. Поскольку все грани $\cg(X)$, кроме $ABCDE$, - треугольные, то все грани, которые сходятся в вершине $F$ являются треугольниками и $F$ соединена ребрами с $B$ и $E$, см. рисунок.
Предположим, что степень вершины $F$ равна $m$. Тогда $m=3,4$ или $5$. Из этого следует, что углы треугольных граней соответственно равны $120^\circ$, $90^\circ$ или $72^\circ$. Первых двух случаев не может быть при $N>6$, а в третьем -- треугольные грани будут такими же как у правильного икосаэдра.
Тогда, все внутренние углы пятиугольника $FBCDE$ равны $144^\circ$, а стало быть и вершина $A$ соединена с вершинами $C$ и $D$. Это противоречие завершает доказательство леммы.
\end{proof}
\medskip
\begin{center}
\begin{picture}(320,140)(-80,-70)
\put(-50,-65){ Рис. Случай, когда $ABCDE$ невыпуклый.}
\put(10,-20){\circle*{5}}
\put(90,-40){\circle*{5}}
\put(70,20){\circle*{5}}
\put(130,40){\circle*{5}}
\put(10,40){\circle*{5}}
\put(70,60){\circle*{5}}
\thicklines
\put(10,-20){\line(4,-1){80}}
\put(10,-20){\line(0,1){60}}
\put(70,20){\line(3,1){60}}
\put(70,20){\line(-3,1){60}}
\put(90,-40){\line(1,2){40}}
\put(70,60){\line(0,-1){40}}
\put(70,60){\line(3,-1){60}}
\put(70,60){\line(-3,-1){60}}
\put(-5,-21){$D$}
\put(97,-41){$C$}
\put(58,28){$A$}
\put(134,44){$B$}
\put(73,65){$F$}
\put(-3,45){$E$}
\end{picture}
\end{center}
\begin{proof} Докажем теперь теорему. При $N=8$ и $N=11$ у графов 7.3.4 и 7.6.22 число ребер равно $3N-8$. Тогда Лемма 5.3 гарантирует, что контактный граф с максимальным числом ребер является неприводимым и поэтому $\ko_N=\ko^*_N$. В этих случаях, множества с максимальным количеством контактов единственны с точностью до изометрии.
Для оставшихся $N$ множества $I_N$ дают примеры контактных графов с $3N-9$ ребрами. (Более того, для $N=7$ и $N=9$ у максимальных графов 7.2.2 и 7.4.7 тоже $3N-9$ ребер.) Докажем, что это правильный ответ. От противного, если предположить что у контактных графов бывает больше ребер, т. е. $e(X)\geqslant 3N-8$, то по Лемме 5.3 графы $\cg(X)$ являются неприводимыми. Однако, в таблицах 7.2, 7.4 и 7.5 графов с таким числом ребер нет.
\end{proof}
\section{Задачи о неприводимых контактных графах}
В работе Л. Данцера \cite{Dan} имеется множество открытых вопросов, связанных с неприводимыми и Д-неприводимыми контактными графами. Наши работы \cite{MT,MT2013,MT14} тоже поднимают большое число вопросов. Остановимся на некоторых из них.
Зафиксируем абстрактный граф $G$. Предположим, что найдется такое множество $X\subset{\Bbb S}^2$, что его контактный граф является неприводимым и изоморфен $G$.
Возможны два варианта:\\
(i) множество $X$ единственное с точностью до изометрии;\\
(ii) имеется степень свободы для $X$, т. е. на сфере существует $k$-параметрическое семейство множеств $X$ с $k\geqslant 1$.
\medskip
Доказано, что для $6\leqslant N\leqslant14$ и $N=24$ к (i) типу относятся максимальные графы (т. е. контактные графы, задающие решение проблемы Таммеса). Возникает вопрос:
\medskip
\noindent{\bf 1.} {\it Верно ли, что все максимальные графы лежат в классе $(i)$?}
\medskip
Этот вопрос можно усилить. Мы уже отмечали, что в торической проблеме Таммеса при $N=7$ оказалось, что имеется три неизоморфных максимальных графа \cite{MN}.
\medskip
\noindent
{\bf 2.} {\it Верно ли, что на сфере для заданного $N>5$ имеется единственный (с точностью до изоморфизма) максимальный граф?}
\medskip
Для $6\leqslant N\leqslant10$, $N=13$ и $N=14$ максимальные графы получаются из графов типа (ii) добавлением ребер. Вместе с тем, для $N=11,12$ максимальные графы изолированы и не получаются таким образом. В связи с этим появляется такой вопрос:
\medskip
\noindent
{\bf 3.} {\it Являются ли случаи $N=11,12$ исключением и все остальные максимальные графы получаются из класса $(ii)$?}
\medskip
Заметим, что в классе $(i)$ лежат не только максимальные графы. Например, к нему относятся графы 7.5.11, 7.5.21, 7.5.28, 7.5.29, 7.6.1. 7.6.5, 7.6.6, 7.6.15 и еще несколько графов из таблицы 7.6. Однако, большинство графов обладают степенью свободы. Л. Данцер \cite[p. 21]{Dan} обсуждая графы типа (ii), отмечает, что во всех рассмотренных им примерах, меняя параметры так, чтобы увеличивалось $d=\psi(X)$ мы достигнем максимума, когда к графу добавятся новые ребра. Однако, далее он пишет, что не решается сформулировать это наблюдение в качестве гипотезы. А мы все-таки зададим вопрос:
\medskip
\noindent
{\bf 4.} {\it Пусть граф $G$ относится к классу $(ii)$. Предположим, что $X\subset{\Bbb S}^2$ такое, что $\cg(X)=G$. Верно ли, что можно немного так изменить $X$ на $X'$, что $\cg(X')=G$ и $\psi(X')>\psi(X)$?}
\medskip
По крайней мере, на один из вопросов Данцера \cite[Question 5, p. 65]{Dan} мы можем ответить. Он спрашивает: {\it``Имеется ли на сфере такое множество $X$, что $|X|<12$, $\cg(X)$ является неприводимым контактным графом и $\psi(X)<d_{12}=\arccos{1/\sqrt{5}}$?''}
\medskip
Дело в том, что среди рассмотренных Данцером Д-неприводимых графов таких не оказалось. Однако, в нашем списке графы с $\psi(X)<d_{12}$ появляются уже при $N=9$. Это граф 7.4.3. Таких графов много для $N=10$ и такие все, кроме максимального, для $N=11$.
Поскольку максимальный граф является неприводимым, то
$$
d_N=\max\limits_{X\in J_N}{\psi(X)}.
$$
Рассмотрим теперь и минимум.
$$
\delta_N=\min\limits_{X\in J_N}{\psi(X)}.
$$
Тогда вопрос Данцера можно переформулировать как: ``Найти минимальное $N$ при котором $\delta_N<d_{12}$.'' (Ответ на этот вопрос, как мы отмечали выше, $N=9$.) По аналогии с проблемой Таммеса здесь возникает такая задача
\medskip
\noindent
{\bf 5.} {\it Найти все $N$--точечные конфигурации $X$ на ${\mathbb S}^2$, чтобы $\psi(X)=\delta_N$.}
\medskip
Таблицы 7.1--7.6 позволяют ответить на этот вопрос для $N<12$, но поскольку $d_{min}$ там найдены численно, необходим более детальный геометрический анализ минимальных конфигураций.
\medskip
В предыдущем разделе мы рассматривали $K_N^*$ -- максимальное число ребер у неприводимого контактного графы, а здесь мы рассмотрим минимальное число ребер.
\medskip
\noindent
{\bf 6.} {\it Найти нижние оценки на величину $\kappa_N$, где}
$$
\kappa_N:=\min\limits_{X\in J_N}{e(X)}.
$$
Из таблиц 7.1--7.6 вытекает следующая теорема:
\begin{thm}
\begin{enumerate}
\item $\kappa_6=9$;
\item $\kappa_7=11;$
\item $\kappa_8=12;$
\item $\kappa_9=12;$
\item $\kappa_{10}=14;$
\item $\kappa_{11}=15.$
\end{enumerate}
\end{thm}
\medskip
В заключение, остановимся еще на одной задаче из списка Данцера \cite[p. 64]{Dan}. Пусть точки $X=\{x_1,\ldots,x_N\}$ заданы сферическими координатами $(\theta_i, \varphi_i)$. Без ограничения общности можно считать, что $\theta_1=\pi/2, \, \varphi_1=0$ и $\theta_2=\pi/2$. Тогда сферические координаты задают пространство конфигураций $\Pi_N$ размерности $2N-3$.
Поскольку, для каждого $X$ на сфере у нас определена величина $\psi(X)$, то задана функция $\psi:\Pi_N\to {\Bbb R}$. Возникает вопрос:
\medskip
\noindent
{\bf 7.} {\it Найти условия при которых $X$ является максимумом функции $\psi$ на $\Pi_N$. }
\medskip
Нам кажется, что эта задача довольно сложная.
\section{Приложение: Список всех неприводимых контактных графов для $N\leqslant 11$}
Приведем здесь основной результат нашей работы \cite{MT2013}.
\begin{thm} Список всех неприводимых контактных графов для $N=6,7,8,9,10,11$
на сфере ${\Bbb S}^2$ приводится в таблицах 7.1--7.6.
\end{thm}
В таблицах $*$
означает, что соответствующий неприводимый граф является также
и Д-неприводимым, а $**$ означает, что этот граф - максимальный.
Заметим, что Л. Данцер перечислил Д-неприводимые графы до $N=10$, и поэтому в таблице для $N=11$ у нас не отмечены Д-неприводимые графы.
В таблицах также показаны и предельные значения $d$, $d_{min}\leqslant
d \leqslant d_{max}$. (Однако, отметим, что величины $d_{min}$
и $d_{max}$ найдены численно и могут немного отличаться от истинных.)
\subsection{Неприводимые графы для 6 вершин.}
\begin{tabular}{ccc}
$N$ & $d_{min}$ & $d_{max}$ \\
$1*$ & $1.4274$ & $1.5708$ \\
$2**$ & $1.5708$ & $1.5708$ \\
\end{tabular}
\begin{tabular}{cccc}
&
\includegraphics[clip,scale=0.4]{pics/irr-6.mps}
& \; \; \; &
\includegraphics[clip,scale=0.2]{pics/irr-4.mps}
\end{tabular}
\subsection{Неприводимые графы для 7 вершин.}
\begin{tabular}{ccc}
$N$ & $d_{min}$ & $d_{max}$ \\
$1*$ & $1.34978$ & $1.35908$ \\
$2**$ & $1.35908$ & $1.35908$ \\
\end{tabular}
\begin{tabular}{cc}
\includegraphics[clip,scale=0.5]{pics/seven1.mps}
~~~~
\includegraphics[clip,scale=0.5]{pics/seven2.mps}
\end{tabular}
\subsection{Неприводимые графы для 8 вершин.}
\begin{tabular}{ccc}
$N$ & $d_{min}$ & $d_{max}$ \\
$1$ & $1.17711$ & $1.18349$ \\
$2*$ & $1.28619$ & $1.30653$ \\
$3*$ & $1.23096$ & $1.30653$ \\
$4**$ & $1.30653$ & $1.30653$ \\
\end{tabular}
\begin{tabular}{cc}
\includegraphics[clip,scale=0.5]{pics/eight1.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eight2.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eight3.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eight4.mps}
\end{tabular}
\subsection{Неприводимые графы для 9 вершин.}
\begin{tabular}{ccc}
$N$ & $d_{min}$ & $d_{max}$ \\
$1$ & $1.14099$ & $1.14143$ \\
$2*$ & $1.22308$ & $1.23096$ \\
$3$ & $1.10525$ & $1.14349$ \\
$4$ & $1.17906$ & $1.18106$ \\
$5$ & $1.15448$ & $1.17906$ \\
$6$ & $1.17906$ & $1.17906$ \\
$7**$ & $1.23096$ & $1.23096$ \\
$8$ & $1.15032$ & $1.18106$ \\
$9*$ & $1.10715$ & $1.14342$ \\
$10$ & $1.17906$ & $1.18428$ \\
\end{tabular}
\begin{tabular}{cc}
\includegraphics[clip,scale=0.5]{pics/nine1.mps}
~~
\includegraphics[clip,scale=0.5]{pics/nine2.mps}
~~
\includegraphics[clip,scale=0.5]{pics/nine3.mps}
~~
\includegraphics[clip,scale=0.5]{pics/nine4.mps}
~~
\includegraphics[clip,scale=0.5]{pics/nine5.mps} \\
\includegraphics[clip,scale=0.5]{pics/nine6.mps}
~~
\includegraphics[clip,scale=0.5]{pics/nine7.mps}
~~
\includegraphics[clip,scale=0.5]{pics/nine8.mps}
~~
\includegraphics[clip,scale=0.5]{pics/nine9.mps}
~~
\includegraphics[clip,scale=0.5]{pics/iv_nine1.mps}
~~
\end{tabular}
\subsection{Неприводимые графы для 10 вершин.}
\begin{tabular}{ccccccc}
$N$ & $d_{min}$ & $d_{max}$ & & $N$ & $d_{min}$ & $d_{max}$ \\
$1$ & $1.0839$ & $1.09751$ & & $2$ & $1.08161$ & $1.08439$ \\
$3$ & $1.03067$ & $1.04695$ & & $4$ & $1.10715$ & $1.0988$ \\
$5$ & $1.07529$ & $1.09431$ & & $6$ & $1.09386$ & $1.12285$ \\
$7*$ & $1.15278$ & $1.15448$ & & $8$ & $1.10012$ & $1.10801$ \\
$9$ & $1.06344$ & $1.07834$ & & $10*$ & $1.15074$ & $1.15191$ \\
$11$ & $1.0843$ & $1.08442$ & & $12$ & $1.10055$ & $1.10889$ \\
$13$ & $1.09504$ & $1.10429$ & & $14$ & $1.06032$ & $1.09604$ \\
$15$ & $1.06278$ & $1.1098$ & & $16$ & $1.09567$ & $1.10715$ \\
$17**$ & $1.15448$ & $1.15448$ & & $18$ & $0.99865$ & $1.0467$ \\
$19$ & $1.0843$ & $1.0844$ & & $20$ & $1.08334$ & $1.09547$ \\
$21*$ & $1.15341$ & $1.15341$ & & $22$ & $1.0988$ & $1.10608$ \\
$23*$ & $1.14372$ & $1.15191$ & & $24$ & $1.09249$ & $1.1098$ \\
$25*$ & $1.15191$ & $1.15245$ & & $26$ & $1.09658$ & $1.10977$ \\
$27*$ & $1.15191$ & $1.15191$ & & $28*$ & $1.10715$ & $1.10715$ \\
$29*$ & $1.10715$ & $1.10715$ & & $30$ & $1.15103$ & $1.15341$ \\
\end{tabular}
\newpage
\begin{tabular}{c}
\includegraphics[clip,scale=0.5]{pics/ten1.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten2.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten3.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten4.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten5.mps}
\includegraphics[clip,scale=0.5]{pics/ten6.mps}
\end{tabular}
\begin{tabular}{c}
\includegraphics[clip,scale=0.5]{pics/ten7.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten8.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten9.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten10.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten11.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten12.mps} \\
\end{tabular}
\begin{tabular}{c}
\includegraphics[clip,scale=0.5]{pics/ten13.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten14.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten15.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten16.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten17.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten18.mps} \\
\end{tabular}
\begin{tabular}{c}
\includegraphics[clip,scale=0.5]{pics/ten19.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten20.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten21.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten22.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten23.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten24.mps} \\
\end{tabular}
\begin{tabular}{c}
\includegraphics[clip,scale=0.5]{pics/ten25.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten26.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten27.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten28.mps}
~~
\includegraphics[clip,scale=0.5]{pics/ten29.mps}
~~
\includegraphics[clip,scale=0.5]{pics/iv_ten1.mps}
\end{tabular}
\medskip
\subsection{Неприводимые графы для 11 вершин.}
\begin{tabular}{ccccccc}
$N$ & $d_{min}$ & $d_{max}$ & & $N$ & $d_{min}$ & $d_{max}$ \\
$1$ & $1.05601$ & $1.05602$ & & $2$ & $1.0538$ & $1.05842$ \\
$3$ & $1.05834$ & $1.05842$ & & $4$ & $1.04765$ & $1.05455$ \\
$5$ & $1.06975$ & $1.06974$ & & $6$ & $1.06306$ & $1.06308$ \\
$7$ & $1.0522$ & $1.06131$ & & $8$ & $1.06621$ & $1.06846$ \\
$9$ & $1.0538$ & $1.05531$ & & $10$ & $1.0795$ & $1.07961$ \\
$11$ & $1.05331$ & $1.0737$ & & $12$ & $1.07163$ & $1.07197$ \\
$13$ & $1.0404$ & $1.06635$ & & $14$ & $1.04759$ & $1.05637$ \\
$15$ & $1.06974$ & $1.06974$ & & $16$ & $1.02726$ & $1.06117$ \\
$17$ & $1.04712$ & $1.06167$ & & $18$ & $1.06043$ & $1.06209$ \\
\end{tabular}
\begin{tabular}{ccccccc}
$N$ & $d_{min}$ & $d_{max}$ & & $N$ & $d_{min}$ & $d_{max}$ \\
$19$ & $1.05386$ & $1.05947$ & & $20$ & $1.05846$ & $1.05882$ \\
$21$ & $1.0632$ & $1.0636$ & & $22**$ & $1.10715$ & $1.10715$ \\
$23$ & $1.05388$ & $1.06537$ & & $24$ & $1.05375$ & $1.0737$ \\
$25$ & $1.06167$ & $1.0636$ & & $26$ & $1.06506$ & $1.06673$ \\
$27$ & $1.04636$ & $1.05882$ & & $28$ & $1.05426$ & $1.06822$ \\
$29$ & $1.07832$ & $1.07836$ & & $30$ & $1.07886$ & $1.07962$ \\
$31$ & $1.05429$ & $1.06105$ & & $32$ & $1.00523$ & $1.05671$ \\
$33$ & $1.061$ & $1.06117$ & & $34$ & $1.02751$ & $1.05828$ \\
$35$ & $1.05447$ & $1.06679$ & & $36$ & $1.0561$ & $1.05627$ \\
$37$ & $1.05431$ & $1.05827$ & & $38 (iv)$ & $1.0064$ & $1.03613$ \\
\end{tabular}
\medskip
\begin{tabular}{c}
\includegraphics[clip,scale=0.5]{pics/eleven1.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven2.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven3.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven4.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven5.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven6.mps} \\
~~
\includegraphics[clip,scale=0.5]{pics/eleven7.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven8.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven9.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven10.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven11.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven12.mps} \\
~~
\includegraphics[clip,scale=0.5]{pics/eleven13.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven14.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven15.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven16.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven17.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven18.mps} \\
\end{tabular}
\begin{tabular}{c}
\includegraphics[clip,scale=0.5]{pics/eleven19.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven20.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven21.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven22.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven23.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven24.mps} \\
~~
\includegraphics[clip,scale=0.5]{pics/eleven25.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven26.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven27.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven28.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven29.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven30.mps} \\
~~
\includegraphics[clip,scale=0.5]{pics/eleven31.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven32.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven33.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven34.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven35.mps}
~~
\includegraphics[clip,scale=0.5]{pics/eleven36.mps} \\
~~
\includegraphics[clip,scale=0.5]{pics/eleven37.mps}
~~
\includegraphics[clip,scale=0.5]{pics/iv_eleven1.mps}
~~
\end{tabular}
|
\section{Introduction}
Given a large graph $G$, the goal in graph sparsification is to
compute a ``small'' graph $H$ that retains, exactly or approximately,
some key properties of $G$. Two such standard regimes are when
$V(H)=V(G)$ but $H$ is a sparse graph, or when $|V(H)| \ll |V(G)|$.
Sparsifiers for basic properties such as connectivity, distances, cuts
and flows have been extensively studied. For instance, cut sparsifiers
were introduced by Benczur and Karger \cite{BenczurK96}, and were more
recently generalized to spectral sparsifiers \cite{BatsonSST13}, and
to cut and flow sparsifiers for vertex subsets \cite{Moitra,LM}.
\iffalse A notable example is the work initiated by Benczur and Karger
on sparsifying to preserve cuts \cite{BenczurK96}, and the work of
Batson, Spielman and Srivastava \cite{BatsonSS} who showed the
following: for every fixed $\epsilon > 0$ and every edge-capacitated
undirected graph $G=(V,E)$ on $n$ vertices, there is an
edge-capacitated graph $H=(V,E')$ on the same vertex set such that $H$
has $O(n/\epsilon^2)$ edges and the capacity of every cut of $G$ is
preserved in $H$ to within a $(1+\epsilon)$-factor; in fact $H$ is also a
spectral sparsifier. \fi Graph sparsifiers are closely related to the
notion of kernelization used in fixed-parameter tractable algorithms,
where an input instance is first reduced to a much smaller instance
(called a kernel), whose size is ideally polynomial in the parameter
$k$, and then the problem is solved on the smaller instance.
Sparsification and sparse representations are also of great importance
for other objects such as signals, matrices, and geometric objects to
name just a few.
We say that a graph $H$ is a \emph{strong} sparsifier for the given
graph $G$, if additionally $H$ is a minor of $G$. Strong sparsifiers
are of particular interest, since they retain some of the structure of
$G$. For example, if $H$ contains some graph $H'$ as a minor, then so
does $G$; a collection ${\mathcal{P}}$ of disjoint paths (or cycles) in $H$
immediately translates to a collection of disjoint paths (or cycles)
in $G$, and so on.
In this paper we study sparsifiers for {\em treewidth}, a fundamental
graph parameter with a wide variety of applications in graph theory
and algorithms. The treewidth of a graph $G=(V,E)$ is typically
defined via tree decompositions. A tree-decomposition of $G$ consists
of a tree $T=(V(T),E(T))$ and a collection of vertex subsets $\{X_v
\subseteq V\}_{v \in V(T)}$ called bags, such that: (i) for each edge
$(a,b) \in E$, there is some node $v \in V(T)$ with both $a,b \in X_v$
and (ii) for each vertex $a \in V$, the set of all nodes of $T$ whose
bags contain $a$ form a non-empty connected subtree of $T$. The {\em
width} of a given tree decomposition is $\max_{v \in V(T)} |X_v| -
1$, and the treewidth of a graph $G$, denoted by $\mathrm{tw}(G)$, is the
width of a minimum-width tree decomposition for $G$. Treewidth is
known to be \mbox{\sf NP}\xspace-hard to compute \cite{treewidth-np-hard}. The best
known polynomial-time approximation algorithm, given a graph $G$ of
treewidth $k$, computes a tree decomposition of width $O(k \sqrt{\log
k})$ \cite{FeigeHL05}. It is also known that treewidth is
fixed-parameter-tractable~\cite{Bodlaender-tw-fpt}: for every fixed
$k$, there is a linear-time algorithm, that, given $G$, decides
whether $\mathrm{tw}(G) \le k$; the dependence of the running time on $k$ is exponential in $\operatorname{poly}(k)$. There are many important results on the structure
of large-treewidth graphs. Perhaps the most well-known of these is the
Grid-Minor Theorem of Robertson and Seymour that we discuss in more
detail later.
Informally, graph $H$ is a treewidth sparsifier for a given graph $G$,
if $H$ is sparse, $|V(H)|$ is small, and $\mathrm{tw}(H)$ is (approximately)
the same as $\mathrm{tw}(G)$. For $H$ to be useful as a replacement for $G$,
it needs to be a strong sparsifier --- that is, $H$ should be a minor
of $G$\footnote{Note that if all we wanted is a graph $H$ that has
similar treewidth as $G$ then it suffices to (approximately) compute
$\mathrm{tw}(G)$ and let $H$ be any graph from a well-known class such as
grids, cliques or expanders with the same treewidth.}. The notion
of treewidth sparsifiers is closely related to the notion of kernels
for treewidth. A polynomial kernel for treewidth is a map $f$, that,
given an instance $(G,k)$, returns an instance $(G',k')$, with the
property that $\mathrm{tw}(G) \le k$ iff $\mathrm{tw}(G') \le k'$, while ensuring that
the size of the graph $G'$ is polynomial in $k$. Unless $\mbox{\sf NP}\xspace
\subseteq \mbox{\sf coNP}\xspace/{\sf poly}$ there is no polynomial kernel for
treewidth which follows from the results of Bodlaender et
al.~\cite{tw-no-poly-kernel} and
Drucker~\cite{Drucker12}. Super-linear lower bounds for more general
forms of kernelization are also known \cite{Jansen13}.
Our main result shows that if one is willing to settle for a
poly-logarithmic factor approximation in the treewidth, then there
exist sparsifiers with very strong properties. To state our main
result we need a definition. A graph $H$ is a topological minor of $G$
if $H$ is obtained from $G$ by edge and node deletions, and by
suppressing degree-$2$ nodes\footnote{Note that $H$ is a minor of $G$
if it can be obtained by edge and node deletions and edge
contractions. A minor $H$ of a graph $G$ need not be a topological
minor $G$, however, if the maximum vertex degree in $H$ is at most
$3$, then $H$ is also a topological minor of $G$.}. Equivalently,
$H$ is a topological minor of $G$ iff a subdivision of $H$ is a
subgraph of $G$. Our main result is summarized in the following
theorem. \iffull
\begin{theorem}\label{thm: main-topological-minor}
There is a randomized algorithm, that, given a graph $G$ of
treewidth at least $k$, with high probability
computes a topological minor $H$ of $G$, such
that:
\begin{itemize}
\item the treewidth of $H$ is $\Omega(k/\operatorname{poly}\log k)$;
\item the maximum vertex degree in $H$ is $3$; and
\item $|V(H)|=O(k^4)$.
\end{itemize}
The running time of the algorithm is polynomial in $|V(G)|$ and $k$.
\end{theorem}
\fi
\ifabstract
\begin{theorem}\label{thm: main-topological-minor}
There is a randomized algorithm, that, given a graph $G$ of
treewidth at least $k$, with high probability
computes a topological minor $H$ of $G$, such
that: (i) the treewidth of $H$ is $\Omega(k/\operatorname{poly}\log k)$; (ii)
the maximum vertex degree in $H$ is $3$; and (iii) $|V(H)|=O(k^4)$.
The running time of the algorithm is polynomial in $|V(G)|$ and $k$.
\end{theorem}
\fi
Our result is close to optimal: degree $3$ cannot be reduced, and the
best one can hope for in terms of the size of the sparsifier is
$O(k^2/\operatorname{poly}\log k)$ (when $G$ is a $k \times k$ grid). We also recall that the
best currently known polynomial-time approximation algorithm can only certify
treewidth to within an $O(\sqrt{\log k})$-factor. We conjecture a
strengthening of the theorem to almost optimal parameters.
\begin{conjecture}
For every graph $G$ with treewidth at least $k$,
there \emph{exists} a topological minor $H$ of $G$
such that $\mathrm{tw}(H) = \Omega(k/\operatorname{poly}\log k)$, $|V(H)|=O(k^2)$ and maximum
vertex degree in $H$ is $3$.
\end{conjecture}
The existence of sparsifiers of size $\operatorname{poly}(k)$ that preserve the
treewidth to within a constant factor remains a very interesting open
question.
\subsection{Treewidth Sparsifiers and Grid Minors}
A fundamental result in Graph Minor Theory is the Grid-Minor Theorem of
Robertson and Seymour \cite{RS-grid}. The theorem states that there is an
integer-valued function $f$, such that any graph $G$ with treewidth at
least $f(g)$ contains a $g \times g$ grid as a minor. The theorem is
equivalent to showing that $\mathrm{tw}(G) \ge f(g)$ implies that $G$ contains
a {\em wall} of height and width $\Theta(g)$ as a subgraph; see
Figure~\ref{fig:wall}.
\begin{figure}[h]
\centering
\includegraphics[height=1in]{wall}
\caption{An elementary wall of height and width $5$. A wall is a
subdivision of an elementary wall.\label{fig:wall}}
\end{figure}
We observe that a wall has maximum vertex degree $3$. Thus, one way to
obtain a degree-$3$ treewidth sparisfier is via the Grid-Minor
Theorem. The original proof of Robertson and Seymour \cite{RS-grid}
showed the existence of $f$ with an iterated exponential dependence on
$g$. Very recently, the first polynomial bound on $f$ was shown in
\cite{CC13-grid}: namely, every graph of treewidth $k$ contains a wall
of size $k^{\delta}$ as a topological minor, where $\delta = 1/98 -
o(1)$. This result implies a degree-$3$ treewidth sparsifier, whose
treewidth is $k^{1/98 - o(1)}$. In contrast, the sparsifier from
Theorem~\ref{thm: main-topological-minor} has treewidth
$\Omega(k/\polylog k)$. Moreover, there are graphs with treewidth
$k$, such that the size of the largest wall they contain is
$O(\sqrt{k/\log k})$ \cite{RobertsonST94}. Therefore, one cannot hope
to obtain small sparsifiers that preserve treewidth to within
polylogarithmic factors via the Grid-Minor Theorem. Our construction
bypasses this limitation.
One of our motivations for studying treewidth sparsifiers is improving
the bounds for the Grid-Minor Theorem. Theorem~\ref{thm:
main-topological-minor} allows us to focus on subcubic graphs with
the additional property that $|V(G)|$ is polynomial in
$\mathrm{tw}(G)$. Degree-$3$ sparsifiers have particular advantages: in such a
graph, for several applications of interest, one can replace
node-disjoint routing with the easier edge-disjoint routing.
We anticipate that using Theorem~\ref{thm: main-topological-minor} as
a starting point, the bounds on the Grid-Minor Theorem from
\cite{CC13-grid} can be improved. We also mention
that the fact that $|V(H)|=\operatorname{poly}(k)$ simplifies some
technical parts in the current proof of~\cite{CC13-grid}.
A related application is to the notion of graph immersions (see
\cite{GraphMinors23,Wollan13}). A graph $G$ admits a strong immersion
of a graph $H$ iff there is an injective mapping $\tau:V(H)
\rightarrow V(G)$ and a mapping $\pi:E(H) \rightarrow \mathcal{P}_G$,
where $\mathcal{P}_G$ is a set of paths in $G$, such that (i) for each
$f=(a,b) \in E(H)$ the path $\pi(f)$ connects $\tau(a)$ and $\tau(b)$;
(ii) for any two edges $f,f' \in E(H)$ the paths $\pi(f)$ and
$\pi(f')$ are edge-disjoint; and (iii) for every $f \in E(H)$ the path
$\pi(f)$ intersects $\tau(V(H))$ only at its endpoints. Note that $G$
admits $H$ as a topological minor if additionally the paths $\pi(f)$
and $\pi(f')$ are internally {\em node}-disjoint for any
distinct pair $f,f'\in E(H)$. If $G$ is a sub-cubic graph, then
$G$ contains $H$ as a topological minor iff $G$ contains $H$ as a
strong immersion. Therefore, $G$ contains a wall $W$ iff it contains
it as an immersion. In recent work, Wollan~\cite{Wollan13} defined
the notion of tree-cut width of a graph and showed, using the
Grid-Minor Theorem, that there is a function $g$, such that every
graph with tree-cut width at least $g(r)$ admits an $r$-wall as a weak
immersion. Motivated by this connection, he raised the question of the
existence of degree-$3$ treewidth sparsifiers. Theorem~\ref{thm:
main-topological-minor} answers his question (Question 18 in
\cite{Wollan13}) in a near-optimal fashion and we refer the reader to
\cite{Wollan13} for the quantitative and qualitative implications
to immersions.
\iffalse
{\em Immersions:} We now discuss an application to the notion of graph
immersions and a question raised by Wollan \cite{Wollan13} on
degree-$3$ treewidth sparsifiers. Immersions are closely related to
topological minors. A graph $G$ admits a weak immersion of a graph $H$
if there is an injective mapping $\tau:V(H) \rightarrow V(G)$ and a
mapping $\pi:E(H) \rightarrow \mathcal{P}_G$ where $\mathcal{P}_G$ is
the set of paths in $G$ such that (i) for each $f \in E(H)$ the path
$\pi(f)$ connects $\tau(a)$ and $\tau(b)$ the end points of $f$ and
(ii) for any two edges $f,f' \in E(H)$ the paths $\pi(f)$ and
$\pi(f')$ are edge-disjoint. It is a strong immersion if for every $f
\in E(H)$ the path $\pi(f)$ intersects $\tau(V(H))$ only in its end
points. $G$ admits $H$ as a topological minor if the paths $\pi(f)$
and $\pi(f')$ are internally {\em node}-disjoint. Robertson and
Seymour \cite{GraphMinors23} proved that finite graphs are
well-quasi-ordered under the immersion relation which confirmed a
conjecture of Nash-Williams. There has been substantial recent
interest and activity on immersions (see~\cite{Wollan13} for
pointers). In \cite{Wollan13} Wollan defines the notion of tree-cut
width of a graph and shows that there is a function $f$ such that
every graph with tree-cut width at least $f(r)$ admits an $r$-wall as a
weak immersion. In order to prove this he uses the grid-minor
theorem. It is an easy observation that a cubic graph admits a wall of
as an immersion iff it admits a wall of similar size as a topological
minor. This prompted Wollan to ask about the existence of degree-$3$
treewidth sparsifiers. Theorem~\ref{thm: main-topological-minor}
answers Wollan's question (Question 18 in \cite{Wollan13}) in a
quantitatively near-optimal fashion. As he observed, this implies that
the relationship between tree-cut width of a graph and the size of the
largest wall it admits as an immersion, is closely related to
grid-minor theorem; we refer the reader to \cite{Wollan13} for further
details.
\fi
Our result can be viewed as providing an approximate kernel for
treewidth, and we hope that it will find applications in preprocessing
graphs for fixed-parameter tractable (FPT) algorithms, and in constructive aspects of
Erdos-P\"{o}sa type theorems.
\iffalse
Fixed-parameter tractable (FPT) algorithms for
treewidth-bounded parameters, and Erdos-P\"{o}sa type theorems rely on
the following observation: if $\mathrm{tw}(G) \ge k$ for some given parameter
$k$, then $G$ has at least $g(k)$, for some function $g$, disjoint
copies of certain structures, for example, cycles (see \cite{CC13}
and pointers therein). A constructive proof would then have to work
with the entire graph to exhibit these disjoint
copies. Theorem~\ref{thm: main-topological-minor} allows the algorithm
to work with a much sparser graph whose size depends only on $k$. This
can lead to faster algorithms once $G$ is preprocessed to find the
sparsifier. For many of these applications it is crucial that the
sparsifier is a topological minor (subgraph) of the original graph.
\fi
We now briefly discuss our techniques. We use a combinatorial object,
called a path-of-sets system, that was defined in~\cite{CC13-grid}
(see also Figure~\ref{fig:pos}). Using the construction of the
Path-of-Sets system from~\cite{CC13-grid}, together with the
Cut-Matching Game of Khandekar, Rao and Vazirani~\cite{KRV}, we can
immediately obtain a strong degree-$4$ treewidth sparsifier $H$, with
$\mathrm{tw}(H) = \Omega(k/\polylog k)$. However, the size of $V(H)$ can be
arbitrarily large. Our main technical contribution
is two-fold. First, we lower the degree of the sparsifier to $3$, by
carefully sub-sampling the edges of $H$. Second, we reduce the size of
the sparsifier to $\operatorname{poly}(k)$. For the second part, we crucially need a
new technical ingredient, that is related to strong vertex-cut
sparsifiers, that we discuss
below.
\subsection{Sparsifiers Preserving Vertex Cuts}
Suppose we are given any graph $G=(V,E)$ and a pair $S,T\subseteq V$
of vertex subsets, containing $k$ vertices each. We say that the pair
$(S,T)$ is \emph{routable} in $G$ iff there is a set ${\mathcal{P}}$ of $k$
disjoint paths connecting the vertices of $S$ to the vertices of $T$
in $G$, and we say that the set ${\mathcal{P}}$ of paths \emph{routes} the
pair $(S,T)$. Assume now that we have two pairs of vertex subsets:
$S_1,T_1$, containing $k_1$ vertices each, and $S_2,T_2$ containing
$k_2$ vertices each. We say that both pairs $(S_1,T_1)$, $(S_2,T_2)$
are \emph{separately routable}, or just routable, in $G$ iff there is
a set ${\mathcal{P}}$ of paths routing $(S_1,T_1)$, and there is a set ${\mathcal{Q}}$
paths routing $(S_2,T_2)$ in $G$. Note that a vertex of $G$ may belong
to a path in ${\mathcal{P}}$ and a path in
${\mathcal{Q}}$. Our second main result is summarized in the following theorem.
\begin{theorem}\label{thm: 2-flow-main}
Assume that we are given a graph $G$, two sets $S_1,T_1\subseteq
V(G)$ of $k_1$ vertices each, and two sets $S_2,T_2\subseteq V(G)$
of $k_2$ vertices each, such that $k_1\geq k_2$, and the pairs
$(S_1,T_1)$ and $(S_2,T_2)$ are (separately) routable in $G$. Then
there are two sets ${\mathcal{P}},{\mathcal{Q}}$ of paths routing $(S_1,T_1)$ and
$(S_2,T_2)$ respectively, such that, if $H$ is the graph obtained by
the union of the paths in ${\mathcal{P}}$ and ${\mathcal{Q}}$, then $\tau(H)\leq
8k_1^4+ 8k_1$, where $\tau(H)$ is the number of nodes of degree more
than two in $H$. Moreover, we can find ${\mathcal{P}}$ and ${\mathcal{Q}}$ in time
polynomial in $n$ and $k_1$.
\end{theorem}
The preceding theorem gives an upper bound on the size of a
topological minor of $G$ that preserves the vertex connectivity
between $S_1,T_1$ and $S_2,T_2$. There are results in the literature
on reduction operations that preserve edge
connectivity~\cite{Lovasz-splitting-off,edge-connectivity} (and also
element connectivity \cite{element-connectivity,ChekuriK09}), however
no such nice operations are available for preserving vertex
connectivity. We briefly discuss some related work on cut sparsifiers
and an open problem on a generalization of Theorem~\ref{thm:
2-flow-main} that would yield strong sparsifiers that preserve
vertex cuts.
There has been a large amount of work in the recent past on graph
sparsifiers that preserve cuts and flows for subsets of
vertices~\cite{Moitra,LM,CLLM,MM,EGK,vsparsifiers}. We discuss some
closely related work. Given an edge-capacitated graph $G$ and a
terminal set ${\mathcal T} \subseteq V(G)$, a graph $H$ is a quality-$q$
cut-sparsifier for ${\mathcal T}$ if (i) ${\mathcal T} \subseteq V(H)$ and (ii) for
any partition $(A,B)$ of ${\mathcal T}$, $\operatorname{MinCut}_G(A,B) \le \operatorname{MinCut}_H(A,B)
\le q \operatorname{MinCut}_G(A,B)$ where $\operatorname{MinCut}_F(A,B)$ is the minimum edge-cut
separating $A$ from $B$ in a graph $F$.
Quality-$1$ sparsifiers
have also been called {\em mimicking networks} in
prior work~\cite{mimicking0,mimicking1,mimicking2,mimicking3}.
Leighton and Moitra~\cite{LM} have shown that for any graph $G$, there is a quality-$q$ sparsifier $H$ for $G$ with $q = O(\log k/\log \log k)$ and
$V(H) = {\mathcal T}$ (here $k = |{\mathcal T}|$); in other words the sparsifier
does not use any non-terminal (or Steiner) vertices. There are
instances on which the best quality one can achieve is
$\Omega(\sqrt{\log k})$ if $H$ does not have Steiner
vertices~ \cite{MM}. Even a relatively small number of Steiner vertices can help substantially in improving the
quality of the sparsifier as shown in \cite{vsparsifiers}.
To simplify the discussion, we restrict our attention to the case where
the terminals in ${\mathcal T}$ have degree $1$ and all edge capacities are $1$. In
this case constant quality cut-sparsifiers are known with $V(H) =
O(k^3)$ \cite{vsparsifiers,KratschW12}. The result of Kratsch and
Wahlstr\"{o}m \cite{KratschW12}, in fact, applies in the more general
setting of vertex-cuts, and yields a quality-$1$
sparsifier; we call such a sparsifier a vertex-cut sparsifier to
distinguish it from an edge-cut sparsifier.
However, the sparsifer of \cite{KratschW12} is {\em not} a minor
of the original graph $G$. Sparsifiers that are minors of the original
graph have an advantage that they allow flows (fractional or integral) and minors
in the sparsifier to be transferred back to the original graph $G$ without
any loss. Theorem~\ref{thm: 2-flow-main} gives us a
small-sized minor that preserves the vertex connectivity between two pairs of vertex
subsets. A natural open question is to generalize this result to a larger number of pairs of vertex subsets.
\begin{question}
Assume that we are given a graph $G$, and $h$ pairs
of vertex subsets $(S_1,T_1), \ldots, (S_h,T_h)$, such that for
each $i$: (1) $S_i,T_i \subseteq V(G)$, (2)
$|S_i|=|T_i| = k_i \le k$, and (3) $(S_i,T_i)$ are routable in $G$.
What is the smallest function $f(k,h)$, such that, given any graph $G$ and $(S_1,T_1), \ldots, (S_h,T_h)$ as above, there is always a
(topological) minor $H$ of $G$ with the property that
each $(S_i,T_i)$ is routable in $H$ and $|V(H)| \le f(k,h)$?
\end{question}
The case when $h = \polylog k$ is of particular interest. We believe
that a bound on $f(k,h)$ from the preceding question can be
used to obtain a vertex-cut sparsifier $H$ for any graph $G$ and a set ${\mathcal T}$ of $k$ terminals,
such that $H$ is a minor of $G$, $|V(H)|\leq f(\operatorname{poly} k, h)$ for $h=\operatorname{poly}\log k$, and
the quality of $H$ is $\polylog k$
\paragraph{Organization}
We prove Theorem~\ref{thm: 2-flow-main} in Section~\ref{sec: 2-flow}.
Section~\ref{sec: PoS system} provides the necessary
background on treewidth and the path-of-sets system.
Theorem~\ref{thm: main-topological-minor} is proved in two steps.
Section~\ref{sec: top minor} gives the proof of a weaker result,
a degree-$4$ sparsifier.
Section~\ref{sec: degree-3} gives the proof for the degree-$3$ sparsifier.
\ifabstract
Due to space constraints, several proofs have been moved to the Appendix.
\fi
\label{---------------------------------------------sec: 2-flow-----------------------------------------}
\section{Routing Two Pairs of Vertex Subsets}\label{sec: 2-flow}
In this section we prove Theorem~\ref{thm: 2-flow-main}. Recall that
a graph $H$ is a \emph{minor} of a graph $G$, iff $H$ can be obtained
from $G$ by a series of edge deletion, vertex deletion, and edge
contraction operations. Equivalently, $H$ is a minor of $G$ iff there
is a map $f:V(H)\rightarrow 2^{V(G)}$ assigning to each vertex $v\in
V(H)$ a subset $f(v)$ of vertices of $G$, such that: (a) for each
$v\in V(H)$, the sub-graph of $G$ induced by $f(v)$ is connected; (b)
if $u,v\in V(H)$ and $u\neq v$, then $f(u)\cap f(v)=\emptyset$; and
(c) for each edge $e=(u,v)\in E(H)$, there is an edge in $E(G)$ with
one endpoint in $f(v)$ and the other endpoint in $f(u)$. A map $f$
satisfying these conditions is called \emph{a model of $H$ in
$G$}. Given any subset $X\subseteq V$ of vertices of $G$, we say
that $H$ is an $X$-respecting minor of $G$, iff $X\subseteq
V(H)$. More formally, there is a model $f$ of $H$, where for each
vertex $x\in X$, there is a distinct vertex $v_x\in V(H)$ with
$f(v_x)=\set{x}$. For each $x\in X$, we will usually identify such
vertex $v_x$ with $x$. In particular, every subset $S\subseteq X$ of
vertices of $X$ corresponds to a subset $S'=\set{v_x\mid x\in X}$ of
vertices in $H$, and we will not distinguish between $S$ and $S'$.
Assume that we are given a graph $G$ and two pairs $(S_1',T_1')$,
$(S_2',T_2')$ of vertex subsets, with $|S_1'|=|T_1'|$ and
$|S_2'|=|T_2'|$, that are separately routable in $G$. We say that a
minor $H$ of $G$ is $(S_1',T_1',S_2',T_2')$-good, iff $H$ is an
$X$-respecting minor for $X=S_1'\cup S_2'\cup T_1'\cup T_2'$, and
$(S_1',T_1')$, $(S_2',T_2')$ are each routable in $H$. We say that it
is $(S_1',T_1',S_2',T_2')$-minimal, iff it is
$(S_1',T_1',S_2',T_2')$-good, and for every edge $e$ of $H$, both the graph
obtained from $H$ by deleting $e$, and the graph obtained from $H$ by contracting $e$, are not
$(S_1',T_1',S_2',T_2')$-good. The main result of this section is the
following theorem.
\begin{theorem}\label{thm: 2-flow-minor}
Assume that we are given a graph $G$, and sets $S_1',T_1',
S_2',T_2'\subseteq V(G)$ of $k$ vertices each, such that the pairs
$(S_1',T_1')$ and $(S_2',T_2')$ are (separately) routable in
$G$. Assume further that vertices in $S_1',T_1',S_2',T_2'$ are
distinct, and have degree $1$ each in $G$. Let $H$ be any
$(S_1',T_1',S_2',T_2')$-minimal minor of $G$. Then $|V(H)|\leq
4k^4+4k$.
\end{theorem}
\ifabstract
Theorem~\ref{thm: 2-flow-main} easily follows from
Theorem~\ref{thm: 2-flow-minor} - see the Appendix for a formal proof.
\fi
\iffull
We start by showing that Theorem~\ref{thm: 2-flow-main} follows from
Theorem~\ref{thm: 2-flow-minor}. Let $G$ be the input graph, and
$(S_1,T_1),(S_2,T_2)$ the given pairs of vertex subsets. We denote
$k_1=k$, and add $\Delta=k_1-k_2$ new edges
$e_1=(a_1,b_1),\ldots,e_{\Delta}=(a_{\Delta},b_{\Delta})$, whose
endpoints are distinct, to the graph. The vertices
$\set{a_1,\ldots,a_{\Delta}}$ are then added to $S_2$, and the
vertices $b_1,\ldots,b_{\Delta}$ are added to $T_2$, so
$|S_2|=|T_2|=|S_1|=|T_1|=k$. The new graph then contains a set of
paths routing $(S_1,T_1)$, and a set of paths routing $(S_2,T_2)$.
We add a new set $S_1'$ of $k$ vertices to the graph, and connect each
vertex in $S_1'$ to a distinct vertex in $S_1$ with an edge. We
construct sets $S_2',T_1',T_2'$ of vertices and connect them to the
vertices in $S_2,T_1,T_2$, respectively, in a similar manner. Let $G'$
be this final graph. Then $G'$ contains a set of paths routing
$(S_1',T_1')$, and a set of paths routing $(S_2',T_2')$. The vertices
in $S_1',T_1',S_2',T_2'$ are distinct, and have degree $1$ each in
$G'$. We now compute any $(S_1',T_1',S_2',T_2')$-minimal minor $H$ of
$G'$. Let $f : V(H) \rightarrow 2^{V(G')}$ be a map to cerfity that $H$
is a minor of $G'$. Let ${\mathcal{P}}'$ be the set of paths routing
$(S_1',T_1')$, and ${\mathcal{Q}}'$ a set of paths routing $(S_2',T_2')$ in
$H$. We use the sets of paths ${\mathcal{P}}',{\mathcal{Q}}'$ to define the sets
${\mathcal{P}},{\mathcal{Q}}$ of paths routing $(S_1,T_1)$ and $(S_2,T_2)$,
respectively, in $G$. This mapping is the natural one; we extend a
path $P' \in {\mathcal{P}}' \cup {\mathcal{Q}}'$ in $H$ to a path $P$ in $G$ by replacing each
vertex $v \in P'$ by a path contained in $G[f(v)]$ (the connected
sub-graph of $G$ corresponding to $v$) that connects the two edges
$e,e'$ of $P'$ incident on $v$. Since only two paths from ${\mathcal{P}}' \cup
{\mathcal{Q}}'$ can contain a node $v \in V(H)$, it is not hard to find paths
through $f(v)$ for them with at most $2$ nodes of degree $3$ or more
in $f(v)$; this will ensure that the number of vertices whose degree
is more than $2$ in the resulting graph is at most twice their number
in $H$. The formal argument is given below.
Consider any path $P'\in {\mathcal{P}}'$, and assume
that $P'=(s=v_0,v_1,v_2,\ldots,v_r,v_{r+1}=t)$, so $s\in S_1',t\in
T_1'$. For each $0\leq i\leq r$, we denote the edge $(v_i,v_{i+1})$ by
$e_i^{P'}$. The new path $P$ contains the edges
$e_1^{P'},\ldots,e_{r-1}^{P'}$. Additionally, for each vertex $v_i$,
for $1\leq i\leq r$, it contains an arbitrary path $R_{v_i}(P)$,
connecting the endpoints of $e_{i-1}^{P'}$ and $e_{i}^{P'}$ that are
contained in $f(v_i)$, such that $R_{v_i}(P)\subseteq G[f(v_i)]$. Notice
that since $G[f(v_i)]$ is connected, such a path exist. Let
${\mathcal{P}}=\set{P\mid P'\in {\mathcal{P}}'}$ be the resulting set of paths. Since
the paths in ${\mathcal{P}}'$ are node-disjoint, so are the paths in
${\mathcal{P}}$. It is then immediate to see that ${\mathcal{P}}$ routes $(S_1,T_1)$
in $G$.
Consider now some path $Q'\in {\mathcal{Q}}'$, and assume that
$Q'=(s=v_0,v_1,v_2,\ldots,v_r,v_{r+1}=t)$, so $s\in S_2',t\in
T_2'$. If $v_1=a_j$ for some $1\leq j\leq \Delta$, then $v_2=b_j$ must
hold, since $e_j$ is the only edge incident on $a_j$ in $G'$, in
addition to $(s,a_j)$. Therefore, $r=2$, and $Q'=(s,a_j,b_j,t)$. We
discard $Q'$ from ${\mathcal{Q}}'$. Otherwise, $Q'$ cannot contain any
vertices in $\set{a_1,b_1,\ldots,a_{\Delta},b_{\Delta}}$. For each
$0\leq i\leq r$, we denote the edge $(v_i,v_{i+1})$ by $e_i^{Q'}$. The
new path $Q$ contains the edges
$e_1^{Q'},\ldots,e_{r-1}^{Q'}$. Additionally, for each vertex $v_i$,
for $1\leq i\leq r$, it contains some path $R_{v_i}(Q)$, connecting
the endpoints of $e_{i-1}^{Q'}$ and $e_{i}^{Q'}$ that are contained in
$f(v_i)$, such that $R_{v_i}(Q)\subseteq G[f(v_i)]$. The path
$R_{v_i}(Q)$ is constructed as follows. If $v_i$ does not belong to
any path in ${\mathcal{P}}'$, then $R_{v_i}(Q)$ is any path connecting the
endpoints of $e_{i-1}^{Q'}$ and $e_{i}^{Q'}$ that are contained in
$f(v_i)$, such that $R_{v_i}(Q)\subseteq G[f(v_i)]$. Otherwise, let
$P'\in {\mathcal{P}}'$ be the path containing $v$. Let $R_1$ be the
intersection of the corresponding path $P\in {\mathcal{P}}$ with $G[f(v_i)]$
(which must be a path), and let $R_2$ be any path contained in
$G[f(v_i)]$, that connects the endpoints of $e_{i-1}^{Q'}$ and
$e_{i}^{Q'}$ that belong to $f(v_i)$. If $R_1$ and $R_2$ are disjoint,
then we let $R_{v_i}(Q)=R_2$. Otherwise, let $u$ be the first vertex
on $R_2$ that belongs to $R_1$, and let $v$ be the last vertex on
$R_2$ that belongs to $R_1$. Let $R_2'\subseteq R_2$ be the segment of
$R_2$ from its beginning until the vertex $u$, and $R_2''\subseteq
R_2$ the segment of $R_2$ from $v$ to its end. Let $R_1'\subseteq R_1$
be the segment of $R_1$ between $u$ and $v$. We then let $R_{v_i}(Q)$
be the concatenation of $R_2',R_1'$ and $R_2''$. Notice that in the
graph obtained by the union of $R_1$ and $R_{v_i}(Q)$, there are at
most two vertices whose degree is more than $2$ --- the vertices $u$ and
$v$. For each vertex $z$ that serves as an endpoint of the paths $R_1$
and $R_{v_i}(Q)$, if $z\neq u,v$, then the degree of $z$ is $1$ in
this graph.
Let ${\mathcal{Q}}$ be the final set of paths obtained after processing all
the paths in ${\mathcal{Q}}'$. Then it is immediate to see that ${\mathcal{Q}}$ routes
the original pair $(S_2,T_2)$ of vertex subsets in $G$. Let $H'$ be the graph obtained from the union of all paths in ${\mathcal{P}}\cup {\mathcal{Q}}$. Then for each $v\in V(H)$, $f(v)\cap V(H')$
contains at most two vertices whose degree in $H'$ is more than $2$, so
$\tau(H')\leq 2|V(H)|\leq 8k_1^4+8k_1$. This completes the proof of
Theorem~\ref{thm: 2-flow-main}.
\fi
In the rest of this section, we focus on the proof of
Theorem~\ref{thm: 2-flow-minor}. For simplicity, we denote
$S_1',S_2',T_1',T_2'$ by $S_1,S_2,T_1$ and $T_2$, respectively. Let
$H$ be a $(S_1,T_1,S_2,T_2)$-minimal minor of $G$. Let ${\mathcal{R}}$ be a
set of paths routing $(S_1,T_1)$ in $H$. We will often refer to the
paths in ${\mathcal{R}}$ as \emph{red paths}, and we will think of these paths
as directed from $S_1$ towards $T_1$ (even though in general the graph
is undirected). Similarly, let ${\mathcal{B}}$ be the set of paths routing
$(S_2,T_2)$ in $H$. We refer to the paths in ${\mathcal{B}}$ as \emph{blue
paths}, and view them as directed from $S_2$ to $T_2$. Notice that a
vertex in $S_1\cup T_1$ cannot participate in a blue path, since its
degree is $1$, and all vertices in $S_1,T_1,S_2,T_2$ are
distinct. Similarly, a vertex in $S_2\cup T_2$ cannot participate in a
red path. An edge $e\in E(H)$ may belong to a red path, or to a blue
path, but not both, since otherwise we could contract $e$ and obtain a minor
that is still $(S_1,T_1,S_2,T_2)$-good, contradicting the minimality
of $H$; this is possible since $e$ is not incident on $S_1\cup S_2\cup
T_1\cup T_2$. The edges that belong to the paths in ${\mathcal{R}}$ are
called \emph{red edges}, and the edges that belong to the paths in
${\mathcal{B}}$ are called \emph{blue edges}. From the minimality of $H$,
every edge is either red or blue. We will refer to the vertices in
$S_1\cup S_2\cup T_1\cup T_2$ as the terminals of $H$. From the
minimality of $H$, every non-terminal vertex belongs to one red path
and one blue path, and is incident on exactly two red edges and
exactly two blue edges. Assume now that there is another set
${\mathcal{R}}'\neq {\mathcal{R}}$ of paths in $H$ routing $(S_1,T_1)$. Then there
must be some red edge in $H$ that does not belong to any path of ${\mathcal{R}}'$,
contradicting the minimality of $H$. Therefore, ${\mathcal{R}}$ is the unique
set of paths routing $(S_1,T_1)$ in $H$, and similarly, ${\mathcal{B}}$ is the
unique set of paths routing $(S_2,T_2)$ in $H$. We prove the following
theorem.
\begin{theorem}\label{thm: labeling}
We can efficiently compute an assignment of labels in
$L=\set{\ell_1,\ell_2,\ldots,\ell_{2k}}$ to the vertices of $V(H)$,
such that each vertex in $V(H)$ is assigned one label, and for every
pair $R\in {\mathcal{R}}$, $B\in {\mathcal{B}}$ of paths, if two vertices $v$ and
$v'$ belong to both $R$ and $B$, and are assigned the same label,
then they appear in the same order on $R$ and on $B$.
\end{theorem}
Before we prove Theorem~\ref{thm: labeling}, let us first complete the
proof of Theorem~\ref{thm: 2-flow-minor} assuming it. Let $\ell:
V(H)\rightarrow L$ be the labeling computed by Theorem~\ref{thm:
labeling}. Next, we switch $S_1$ and $T_1$, so that the directions
of the paths in ${\mathcal{R}}$ are reversed. We apply Theorem~\ref{thm:
labeling} again to this new setting, and obtain another labeling
$\ell': V(H)\rightarrow L'$, where
$L'=\set{\ell_1',\ell_2',\ldots,\ell_{2k}'}$.
Assume for contradiction that $|V(H)|\geq 4k^4+4k+1$. Every
non-terminal vertex $v$ can be associated with a quadruple
$(R,B,\ell_i,\ell'_j)$, where $R$ and $B$ are the red and the blue
paths on which $v$ lies, $\ell_i$ is the label assigned to $v$ by
$\ell$, and $\ell'_j$ is the label assigned to $v$ by $\ell'$. Since
the total number of such quadruples is $4k^4$, there is a pair $u,v$
of non-terminal vertices that have the same quadruple
$(R,B,\ell_i,\ell'_j)$. As $u$ and $v$ are assigned the same label by
$\ell$, they must appear in the same order on $R$ and $B$. Assume
w.l.o.g. that $u$ appears before $v$ on both these paths. However,
since both these vertices are assigned the same label by $\ell'$, and
since the red paths were reversed when computing $\ell'$, the order of
$u$ and $v$ on paths $R$ and $B$ must be reversed, a contradiction.
In order to complete the proof of Theorem~\ref{thm: 2-flow-minor}, it
now only remains to prove Theorem~\ref{thm: labeling}.
\subsection*{Proof of Theorem~\ref{thm: labeling}}
Let $\tilde{H}$ be the directed counterpart of the graph $H$, where we
direct all red edges along the direction of the red paths from $S_1$
to $T_1$, and we direct the blue edges similarly along the blue paths
from $S_2$ to $T_2$. The main combinatorial object that we use in the
proof is a chain. A chain $Z$ is a directed (not necessarily simple)
path in graph $\tilde{H}$, such that the edges of $Z$ are alternating red
and blue edges. In other words, if the edges of $Z$ are
$e_1,e_2,\ldots,e_r$ in this order, then all odd-indexed edges
are red and all even-indexed edges are blue, or vice versa. The rest
of the proof consists of three steps. First, we show that every chain
must be a simple path, so no vertex may appear twice on a chain. If
this is not the case, we will show that ${\mathcal{R}}$ is not a unique set of
paths routing $(S_1,T_1)$, or that ${\mathcal{B}}$ is not a unique set of
paths routing $(S_2,T_2)$, leading to a
contradiction. In the second step, we construct a collection of $2k$
chains using a natural greedy algorithm: start from some source, and
then follow alternatively red and blue edges, while possible. We will
show that every vertex of $H$ belongs to at least one chain (but may
belong to more than one). We then associate a separate label with each
chain, and assign all vertices that belong to a chain the same
label. If a vertex belongs to several chains, then one of the
corresponding labels is assigned arbitrarily. Finally, we prove that
for every path $P\in {\mathcal{R}}\cup {\mathcal{B}}$ and every chain $Z$, if $v$ and
$v'$ are two vertices that belong to both $P$ and $Z$, then they must
appear in the same order on $P$ and on $Z$.
Before we proceed, we define two auxiliary structures: red and blue
cycles. Let $C$ be a directed simple cycle in the graph $\tilde{H}$ (so
every vertex may appear at most once on $C$). We say that it is a
\emph{blue cycle} iff we can partition $C$ into an even number of
edge-disjoint consecutive segments
$\sigma_1,\sigma_2,\ldots,\sigma_{2r}$, where $r>0$; for all $1\leq
i\leq r$, $\sigma_{2i}$ consists of a single red edge, and
$\sigma_{2i-1}$ is a non-empty path that only consists of blue
edges. Every edge of $C$ belongs to exactly one segment, and every
consecutive pair of segments shares one vertex (if $r=1$ then the two
segments share two vertices --- the endpoints of the segments). A red
cycle is defined similarly, with the roles of the red and the blue
segments reversed. We start by showing that $\tilde{H}$ cannot contain a red
or a blue cycle.
\begin{lemma}
\label{lem:no-red-blue-cycle}
Graph $\tilde{H}$ cannot contain a red cycle or a blue cycle.
\end{lemma}
\iffull
\begin{proof}
We prove for blue cycles; the proof for red cycles is similar. Let
$C$ be a blue cycle in $\tilde{H}$, and let
$\sigma_1,\sigma_2,\ldots,\sigma_{2r}$ be the corresponding segments
of $C$. Let $H'$ be the graph obtained from $H$ by deleting all
edges participating in the segments $\sigma_{2i-1}$, for $1\leq
i\leq r$ (that is, the blue segments). We claim that both
$(S_1,T_1)$ and $(S_2,T_2)$ remain routable in $H'$, contradicting
the minimality of $H$. Since we only deleted blue edges, it is clear
that $(S_1,T_1)$ remains routable via the paths in ${\mathcal{R}}$. We now
show that $H'$ contains a collection of paths routing $(S_2,T_2)$.
Let $A$ denote the set of all vertices $a$, such that $a$ is the
last vertex of some blue segment $\sigma_{2i-1}$ of $C$, for $1\leq
i\leq r$, and let $B$ denote the set of all vertices $b$, such that
$b$ is the first vertex of some blue segment $\sigma_{2i-1}$ of
$C$. Then $|A|=|B|=r$, and the red edges of $C$ define a complete
matching between $A$ and $B$. Let $\Sigma$ be the collection of
paths obtained from ${\mathcal{B}}$, by deleting all blue edges that
participate in the cycle $C$ (we do not include $0$-length paths in
$\Sigma$). Then $\Sigma$ is a collection of disjoint paths, that
only contain blue edges, which route the pair $(S_2\cup A)$ and
$(T_2\cup B)$. Notice that every path in $\Sigma$ contains at
least two vertices: this is since the terminals cannot belong to $C$
as their degrees are $1$, and every vertex appears on $C$ at most
once. Therefore, all vertices in $S_1,S_2,A,B,T_1,T_2$ are distinct.
We now construct the following directed graph $F$: the vertices of
$F$ are $S_2\cup A\cup T_2\cup B$. There is a directed edge $(u,v)$,
for $u\neq v$, in $F$ iff there is a (directed) red edge $(v,u)$ in
$C$ (in which case we say that $(u,v)$ is a red edge), or there is
some path in $\Sigma$ that starts at $u$ and terminates at $v$ (in
which case we say that $(u,v)$ is a blue edge). Notice that in the
graph $F$, every vertex in $S_2$ has one outgoing edge and no other
incident edges; every vertex in $T_2$ has one incoming edge and no
other incident edges; every vertex in $A$ has one outgoing edge
(blue), and one incoming edge (red); and every vertex in $B$ has one
incoming edge (blue) and one outgoing edge (red). It is then easy to
see that $F$ contains $k$ directed disjoint paths connecting $S_2$
to $T_2$, and this gives a set of paths routing $(S_2,T_2)$ in $H'$,
contradicting the minimality of $H$.
\end{proof}
\fi
The claim below essentially follows from the preceding lemma.
\begin{claim}\label{claim: no cycles on chain}
If $Z$ is a chain, then every vertex of $V(H)$ may appear on $Z$ at most once.
\end{claim}
\iffull
\begin{proof}
Assume otherwise, and let $Z$ be any chain, such that some vertex
appears more than once on $Z$. Then there is a segment $Z'$ of $Z$,
such that both endpoints of $Z'$ are the same vertex $v$, but every
other vertex appears at most once on $Z'$, and $v$ is not an inner
vertex of $Z'$. Notice that $Z'$ must contain at least two edges.
Then $Z'$ defines a simple directed cycle in $\tilde{H}$. Moreover, if
both edges incident on $v$ in $Z'$ are blue, then $Z'$ is a blue
cycle; if both edges are red then $Z'$ is a red cycle; otherwise it
is both a red and a blue cycle. Since $\tilde{H}$ cannot contain a red or
a blue cycle, every vertex appears on $Z$ at most once.
\end{proof}
\fi
We define a collection ${\mathcal{Z}}$ of $2k$ chains in $\tilde{H}$,
and prove that every vertex of $\tilde{H}$ belongs to at least one chain.
Let $s\in S_1\cup S_2$, and let $e$ be the unique edge leaving $s$. We
start building the chain by adding $e$ to the chain. If the last edge
added to the chain $e'=(u,v)$ is a red edge, and there is a blue edge
leaving $v$ in $\tilde{H}$, then we add the unique blue edge leaving $v$ in
$\tilde{H}$ to the chain; if no such edge exists, we complete the
construction of the chain --- in this case, $v\in T_1\cup T_2$ must
hold. Similarly, if the last edge added to the chain $e'=(u,v)$ is a
blue edge, and there is a red edge leaving $v$ in $\tilde{H}$, then we add
the unique red edge leaving $v$ in $\tilde{H}$ to the chain; if no such edge
exists, we complete the construction of the chain. Overall, we
construct one chain starting from each vertex in $S_1\cup S_2$,
obtaining $2k$ chains. Let ${\mathcal{Z}}$ denote the resulting collection of
the chains.
\begin{claim}
\label{claim: chain covers all vertices}
Every vertex of $\tilde{H}$ belongs to at least one chain.
\end{claim}
\iffull
\begin{proof}
We prove a slightly stronger claim: that every edge of $\tilde{H}$ belongs
to at least one chain.
Let $e$ be any edge of $\tilde{H}$, and assume w.l.o.g. that it is a blue
edge. We construct a chain $P$ from the end to the beginning, and we
start by adding the edge $e$ to $P$ as the last edge of $P$. Assume
that the last edge added to $P$ was $e'=(u,v)$. If $u\in S_1\cup S_2$,
then we terminate the construction of $P$; otherwise, there must be a
red edge entering $u$ and a blue edge entering $u$. If $e'$ is a red
edge, then we add the unique blue edge entering $u$ to $P$, and
otherwise we add the unique red edge entering $u$ to $P$, and continue
to the next iteration. Since at every step, the current path $P$ is a
valid chain, and no vertex may appear twice on a chain, this process
will eventually stop at some vertex $s\in S_1\cup S_2$. Then the
unique chain $Z\in {\mathcal{Z}}$ that starts from $s$ must contain $P$ as a
sub-path, and hence must contain the edge $e$.
\end{proof}
\fi
Our final step is the following claim
\begin{claim}\label{claim: order}
Let $Z$ be a chain, and assume that it contains two vertices
$v,v'\in V(P)$, where $P\in {\mathcal{R}}\cup {\mathcal{B}}$. Assume further that
$v$ appears before $v'$ on $Z$. Then $v$ appears before $v'$ on $P$.
\end{claim}
\iffull
\begin{proof}
Assume otherwise. Then there must be two vertices $u,u'$ that appear
on both $Z$ and $P$, such that no other vertex of $P$ appears
between $u$ and $u'$ on $Z$, $u$ appears before $u'$ on $Z$, and it
appears after $u'$ on $P$. Indeed, consider the segment $Z^*$ of $Z$
between $v$ and $v'$. If this segment contains no other vertex of
$P$, then we are done. Otherwise, assume w.l.o.g. that $v$ appears
before $v'$ on $Z$, and let $v_0=v,v_1,\ldots,v_x=v'$ be the
vertices of $P\cap Z$, that appear on $Z^*$ in this order. Since $v$
appears after $v'$ on $P$, there must be a consecutive pair
$v_i,v_{i+1}$ of vertices, such that $v_i$ appears after $v_{i+1}$
on $P$. We then set $u=v_i$ and $u'=v_{i+1}$.
Let $Z'$ be the segment of the chain $Z$ between $u$ and $u'$, and
let $P'$ be the segment of $P$ between $u'$ and $u$. Observe that
$P'\cap Z'=\set{u',u}$, and $Z'$ contains at least one edge whose
color is opposite from the color of $P$. If $P$ is a blue path, then
$Z'\cup P'$ is a blue cycle; otherwise it is a red cycle, a
contradiction.
\end{proof}
\fi
We are now ready to assign labels to the vertices of $H$. Let
${\mathcal{Z}}=\set{C_1,C_2,\ldots,C_{2k}}$. Fix any vertex $v\in H$, and let
$C_i\in {\mathcal{Z}}$ be any chain that contains $v$. We then assign to $v$
the label $\ell_i$.
Consider now any pair $R\in {\mathcal{R}}$, $B\in {\mathcal{B}}$ of paths, and let
$v,v'$ be two vertices that have the same label $\ell_i$ and appear on
both $R$ and $B$. Assume w.l.o.g. that $v$ appears before $v'$ on
chain $C_i$. Then from Claim~\ref{claim: order}, $v$ must appear
before $v'$ on both $R$ and $B$. This completes the proof of
Theorem~\ref{thm: labeling}, and hence of Theorem~\ref{thm:
2-flow-minor} and Theorem~\ref{thm: 2-flow-main}.
\label{---------------------------------------------sec: PoS system-minor-----------------------------------------}
\section{Background on Treewidth and Path-of-Sets System}\label{sec: PoS system}
\ifabstract
In this section we define some graph-theoretic notions and summarize
some previous results that we use in the proof of Theorem~\ref{thm:
main-topological-minor}. We also define a combinatorial object that
plays a central role in the proof --- the path-of-sets system from \cite{CC13-grid}.
Given a graph $G=(V,E)$ and a set $A\subseteq V$ of vertices, we
denote by $E_G(A)$ the set of edges with both endpoints in $A$, and by
$\operatorname{out}_G(A)$ the set of edges with exactly one endpoint in $A$. For
disjoint sets of vertices $A$ and $B$, the set of edges with one end
point in $A$ and the other in $B$ is denoted by $E_G(A,B)$. For a
vertex $v$ in a graph $G$ we use $d_G(v)$ to denote its degree. We may
omit the subscript $G$ if it is clear from the context. Given a set
${\mathcal{P}}$ of paths in $G$, we denote by $V({\mathcal{P}})$ the set of all
vertices participating in paths in ${\mathcal{P}}$, and similarly, $E({\mathcal{P}})$
is the set of all edges that participate in paths in ${\mathcal{P}}$. We
sometimes refer to sets of vertices as \emph{clusters}. A path $P$ in
a graph $G$ is a \emph{$2$-path} iff every inner vertex $v$ in $P$ has
$d_G(v)=2$. It is a \emph{maximal $2$-path} iff the degrees of the
endpoints of $P$ are both different from $2$. Given a set ${\mathcal{P}}$ of
paths, we denote by $J({\mathcal{P}})$ the graph obtained by the union of all
the paths in ${\mathcal{P}}$. Given a graph $G$, we denote by $\tau(G)$ the number of vertices of $G$ whose degree is at least $3$.
We now define the notion of linkedness and the different notions of
well-linkedness that we use.
\begin{definition}
We say that a set ${\mathcal T}$ of vertices is $\alpha$-well-linked in
$G$, iff for any partition $(A,B)$ of the vertices of $G$ into two
subsets, $|E(A,B)|\geq \alpha\cdot \min\set{|A\cap {\mathcal T}|,|B\cap
{\mathcal T}|}$.
We say that a set ${\mathcal T}$ of vertices is
\emph{node-well-linked} in $G$, iff for any pair $({\mathcal T}_1,{\mathcal T}_2)$
of equal-sized subsets of ${\mathcal T}$, there is a collection ${\mathcal{P}}$ of
$|{\mathcal T}_1|$ {\bf node-disjoint} paths, connecting the vertices of
${\mathcal T}_1$ to the vertices of ${\mathcal T}_2$. (Note that ${\mathcal T}_1$,
${\mathcal T}_2$ are not necessarily disjoint, and we allow empty paths).
We say that two disjoint vertex subsets $A$ and $B$ are
\emph{linked} in $G$ iff for any pair of equal-sized subsets
$A'\subseteq A$, $B'\subseteq B$ there is a set ${\mathcal{P}}$ of
$|A'|=|B'|$ node-disjoint paths connecting $A'$ to $B'$ in $G$.
\end{definition}
The following Theorem follows from previous work. For completeness, we
include a proof sketch in the Appendix.
\begin{theorem}\label{thm: weak well-linkedness to tw}
Let $G$ be any graph with maximum vertex degree $\Delta$, and
${\mathcal T}$ a subset of $\kappa$ vertices, such that ${\mathcal T}$ is
$\alpha$-well-linked in $G$, for $\alpha< 1$. Then the treewidth of
$G$ is $\Omega(\alpha\kappa/\Delta)$.
\end{theorem}
\paragraph{Path-of-Sets System}
A central combinatorial object that we use in the proof of
Theorems~\ref{thm: main-topological-minor} is a path-of-sets system,
that was introduced in~\cite{CC13-grid}.
\ifabstract
\begin{figure}[htb]
\centering
\scalebox{0.8}{\includegraphics[width=6in]{pos}}
\caption{Path-of-Sets System}
\label{fig:pos}
\end{figure}
\fi
\begin{definition}
A path-of-sets system $({\mathcal{{S}}},\bigcup_{i=1}^{r-1}{\mathcal{P}}_i)$ of width
$r$ and height $h$ consists of:
\begin{itemize}[noitemsep,topsep=0pt]
\item A sequence ${\mathcal{{S}}}=(S_1,\ldots,S_r)$ of $r$ disjoint vertex
subsets of $G$, where for each $i$, $G[S_i]$ is connected;
\item For each $1\leq i\leq r$, two disjoint sets $A_i,B_i\subseteq S_i$ of $h$ vertices each, such that $A_i$ and $B_i$ are linked in $G[S_i]$;
\item For each $1\leq i<r$, a set ${\mathcal{P}}_i$ of $h$ disjoint paths,
routing $(B_i,A_{i+1})$, such that all paths in $\bigcup_i{\mathcal{P}}_i$
are mutually disjoint, and do not contain the vertices of
$\bigcup_{S_j\in {\mathcal{{S}}}}S_j$ as inner vertices,
\end{itemize}
We say that it is a \emph{strong} path-of-sets system, if additionally
for each $1\leq i\leq r$, $A_i$ is node-well-linked in $G[S_i]$, and
the same holds for $B_i$. (See Figure~\ref{fig:pos}.)
\end{definition}
The following theorem, that was proved in~\cite{CC13-grid}, is the
starting point of the proof of Theorem~\ref{thm:
main-topological-minor}.
\begin{theorem}[Theorem 3.2 in~\cite{CC13-grid}]
\label{thm: strong PoS system}
Let $G$ be any graph of treewidth $k$, and let $h,r>1$ be integral
parameters, such that for some large enough constants $c$ and $c'$,
$k/\log^{c'}k>chr^{48}$. Then there is an efficient randomized
algorithm, that, given $G,h$ and $r$, w.h.p. computes a strong
path-of-sets system of height $h$ and width $r$ in $G$.
\end{theorem}
{\bf Expanders and the Cut-Matching Game.}
We say that a (multi)-graph $G=(V,E)$ is an $\alpha$-expander, iff
$\min_{\stackrel{S\subseteq V:}{|S|\leq |V|/2}}\set{\frac{|E(S,\overline S)|}{|S|}}\geq \alpha$.
We use the cut-matching game of Khandekar, Rao and
Vazirani~\cite{KRV}. In this game, we are given a set $V$ of $N$
vertices, where $N$ is even, and two players: a cut player, whose goal
is to construct an expander $X$ on the set $V$ of vertices, and a
matching player, whose goal is to delay its construction. The game is
played in iterations. We start with the graph $X$ containing the set
$V$ of vertices, and no edges. In each iteration $j$, the cut player
computes a bi-partition $(A_j,B_j)$ of $V$ into two equal-sized sets,
and the matching player returns some perfect matching $M_j$ between
the two sets. The edges of $M_j$ are then added to $X$. Khandekar, Rao
and Vazirani have shown that there is a strategy for the cut player,
guaranteeing that after $O(\log^2N)$ iterations we obtain a
$\ensuremath{\frac{1}{2}}$-expander w.h.p. Subsequently, Orecchia et
al.~\cite{better-CMG} have shown the following improved bound:
\begin{theorem}[\cite{better-CMG}]\label{thm: CMG}
There is a probabilistic algorithm for the cut player, such that, no matter how the matching player plays, after $\ensuremath{\gamma_{\mbox{\tiny{\sc CMG}}}}(N)=O(\log^2N)$ iterations, graph $X$ is an $\ensuremath{\alpha_{\mbox{\tiny{\sc CMG}}}}(N)=\Omega(\log N)$-expander, with constant probability.
\end{theorem}
Our algorithms work by embedding an expander $X$ into a sub-graph of
$G$. The embedding of the expander is then used to certify the
treewidth. We use the following notion of embedding.
\begin{definition}
Let $G,X$ be graphs. An embedding $\phi$ of $X$ into $G$ maps every
vertex $v\in X$ to a connected subgraph $C_v\subseteq G$, and every
edge $e=(u,v)\in E(X)$ to a path $P_e$ in graph $G$, whose endpoints
belong to $C_v$ and $C_u$, respectively. We say that the
\emph{congestion} of the embedding is at most $c$, iff every edge of
$G$ belongs to at most $c-1$ paths in $\set{P_e\mid e\in E(X)}$ and
at most one graph $\set{C_v\mid v\in V(X)}$.
\end{definition}
In the next simple claim, we show that if we can embed a
$\kappa$-vertex expander with congestion at most $c$ into a graph $H$
with bounded vertex degree, then the treewidth of $H$ is large. The proof appears in the Appendix.
\begin{claim}\label{claim: embed an expander treewidth}
Let $X$ be an $\alpha$-expander on $\kappa$ vertices for $\alpha<1$,
with maximum vertex degree $\Delta'$, and let $H$ be a graph with
maximum vertex degree at most $\Delta$, such that that there is an
embedding of $X$ into $H$ with congestion $\eta$. Then $\mathrm{tw}(H) =
\Omega\left (\frac{\alpha\kappa}{\eta\Delta\Delta'}\right )$.
\end{claim}
\fi
\iffull In this section we define some graph-theoretic notions and
summarize some previous results that we use in the proof of
Theorem~\ref{thm: main-topological-minor}. We also define a
combinatorial object that plays a central role in the proof --- the
path-of-sets system from \cite{CC13-grid}.
Given a graph $G=(V,E)$ and a set $A\subseteq V$ of vertices, we
denote by $E_G(A)$ the set of edges with both endpoints in $A$, and by
$\operatorname{out}_G(A)$ the set of edges with exactly one endpoint in $A$. For
disjoint sets of vertices $A$ and $B$, the set of edges with one end
point in $A$ and the other in $B$ is denoted by $E_G(A,B)$. For a
vertex $v$ in a graph $G$ we use $d_G(v)$ to denote its degree. We may
omit the subscript $G$ if it is clear from the context. Given a set
${\mathcal{P}}$ of paths in $G$, we denote by $V({\mathcal{P}})$ the set of all
vertices participating in paths in ${\mathcal{P}}$, and similarly, $E({\mathcal{P}})$
is the set of all edges that participate in paths in ${\mathcal{P}}$. We
sometimes refer to sets of vertices as \emph{clusters}. A path $P$ in
a graph $G$ is a \emph{$2$-path} iff every inner vertex $v$ in $P$ has
$d_G(v)=2$. It is a \emph{maximal $2$-path} iff the degrees of the
endpoints of $P$ are both different from $2$. Given a set ${\mathcal{P}}$ of
paths, we denote by $J({\mathcal{P}})$ the graph obtained by the union of all
the paths in ${\mathcal{P}}$. Given a graph $H$, let $\tau(H)$ denote the number of vertices of $H$ whose degree is more than $2$ in $H$.
We now define the notion of linkedness and the different notions of
well-linkedness that we use.
\begin{definition}
We say that a set ${\mathcal T}$ of vertices is
$\alpha$-well-linked\footnote{This notion of well-linkedness is
based on edge-cuts and we distinguish it from node-well-linkedness
that is directly related to treewidth. For technical reasons it is
easier to work with edge-cuts and hence we use the term well-linked to mean
edge-well-linkedness, and explicitly use the term
node-well-linkedness when necessary.} in $G$, iff for any
partition $(A,B)$ of the vertices of $G$ into two subsets,
$|E(A,B)|\geq \alpha\cdot \min\set{|A\cap {\mathcal T}|,|B\cap {\mathcal T}|}$.
\end{definition}
\begin{definition}
We say that a set ${\mathcal T}$ of vertices is \emph{node-well-linked} in
$G$, iff for any pair $({\mathcal T}_1,{\mathcal T}_2)$ of equal-sized subsets of
${\mathcal T}$, there is a collection ${\mathcal{P}}$ of $|{\mathcal T}_1|$ {\bf
node-disjoint} paths, connecting the vertices of ${\mathcal T}_1$ to the
vertices of ${\mathcal T}_2$. (Note that ${\mathcal T}_1$, ${\mathcal T}_2$ are not
necessarily disjoint, and we allow empty paths).
\end{definition}
The two different notions of well-linkedness are closely related. In
particular, suppose ${\mathcal T}$ is $\alpha$-well-linked in a graph $G$ of
maximum degree $\Delta$. Then there is a large subset ${\mathcal T}'\subseteq
{\mathcal T}$ of vertices that is node-well-linked in $G$, as shown in the
following theorem.
\begin{theorem}[Theorem 2.2 in~\cite{CC13-grid}]
\label{thm: grouping}
Suppose we are given a connected graph $G=(V,E)$ with maximum
vertex degree $\Delta$, and a subset ${\mathcal T}$ of $\kappa$ vertices
called terminals, such that ${\mathcal T}$ is $\alpha$-well-linked in $G$,
for some $\alpha<1$. Then there
is a subset ${\mathcal T}'\subset {\mathcal T}$ of
$\Omega\left(\frac{\alpha\kappa }{\Delta}\right )$ terminals, such
that ${\mathcal T}'$ is node-well-linked in $G$.
\end{theorem}
The following well-known lemma summarizes an important connection
between treewidth and node-well-linkedness.
\begin{lemma}[\cite{Reed-chapter}]
\label{lem:tw-wl}
Let $k$ be the size of the largest node-well-linked set in $G$. Then
$k \le \mathrm{tw}(G) \le 4k$.
\end{lemma}
Combining Theorem~\ref{thm: grouping} with Lemma~\ref{lem:tw-wl}, we
obtain the following theorem.
\begin{theorem}\label{thm: weak well-linkedness to tw}
Let $G$ be any graph with maximum vertex degree $\Delta$, and
${\mathcal T}$ a subset of $\kappa$ vertices, such that ${\mathcal T}$ is
$\alpha$-well-linked in $G$, for $\alpha< 1$. Then the treewidth of
$G$ is $\Omega(\alpha\kappa/\Delta)$.
\end{theorem}
A notion closely related to well-linkedness is that of linkedness,
where we require good connectivity between a pair of disjoint vertex subsets.
\begin{definition}
We say that two disjoint vertex subsets $A$ and $B$ are \emph{linked} in
$G$ iff for any pair of equal-sized subsets $A'\subseteq A$,
$B'\subseteq B$ there is a set ${\mathcal{P}}$ of $|A'|$ node-disjoint
paths connecting $A'$ to $B'$ in $G$.
\end{definition}
\paragraph{Path-of-Sets System}
A central combinatorial object that we use in the proof of
Theorems~\ref{thm: main-topological-minor} is a path-of-sets system,
that was introduced in~\cite{CC13-grid} (a somewhat similar object,
called a grill, was introduced by Leaf and
Seymour~\cite{LeafS12}). See Figure~\ref{fig:pos}.
\begin{definition}
A path-of-sets system $({\mathcal{{S}}},\bigcup_{i=1}^{r-1}{\mathcal{P}}_i)$ of width
$r$ and height $h$ consists of:
\begin{itemize}
\item A sequence ${\mathcal{{S}}}=(S_1,\ldots,S_r)$ of $r$ disjoint vertex
subsets of $G$, where for each $i$, $G[S_i]$ is connected;
\item For each $1\leq i\leq r$, two disjoint sets $A_i,B_i\subseteq
S_i$ of $h$ vertices each, such that $A_i$ and $B_i$ are linked in
$G[S_i]$;
\item For each $1\leq i<r$, a set ${\mathcal{P}}_i$ of $h$ disjoint paths,
routing $(B_i,A_{i+1})$, such that all paths in $\bigcup_i{\mathcal{P}}_i$
are mutually disjoint, and do not contain the vertices of
$\bigcup_{S_i\in {\mathcal{{S}}}}S_i$ as inner vertices,
\end{itemize}
We say that it is a \emph{strong} path-of-sets system, if additionally
for each $1\leq i\leq r$, $A_i$ is node-well-linked in $G[S_i]$, and
the same holds for $B_i$.
\end{definition}
\begin{figure}[htb]
\centering
\includegraphics[width=6in]{pos}
\caption{Path-of-Sets System}
\label{fig:pos}
\end{figure}
The following theorem, that was proved in~\cite{CC13-grid}, is the
starting point of the proof of Theorem~\ref{thm:
main-topological-minor}.
\begin{theorem}[Theorem 3.2 in~\cite{CC13-grid}]
\label{thm: strong PoS system}
Let $G$ be any graph of treewidth $k$, and let $h,r>1$ be integral
parameters, such that for some large enough constants $c$ and $c'$,
$k/\log^{c'}k>chr^{48}$. Then there is an efficient randomized
algorithm, that, given $G,h$ and $r$, w.h.p. computes a strong
path-of-sets system of height $h$ and width $r$ in $G$.
\end{theorem}
{\bf Expanders and the Cut-Matching Game.}
We say that a (multi)-graph $G=(V,E)$ is an $\alpha$-expander, iff
$\min_{\stackrel{S\subseteq V:}{|S|\leq
|V|/2}}\set{\frac{|E(S,\overline S)|}{|S|}}\geq \alpha$. We use the
cut-matching game of Khandekar, Rao and Vazirani~\cite{KRV} to
construct an expander that can be appropriately embedded in a
graph. In this game, we are given a set $V$ of $N$ vertices, where $N$
is even, and two players: a cut player, whose goal is to construct an
expander $X$ on the set $V$ of vertices, and a matching player, whose
goal is to delay its construction. The game is played in
iterations. We start with the graph $X$ containing the set $V$ of
vertices, and no edges. In each iteration $j$, the cut player
computes a bi-partition $(A_j,B_j)$ of $V$ into two equal-sized sets,
and the matching player returns some perfect matching $M_j$ between
the two sets. The edges of $M_j$ are then added to $X$. Khandekar, Rao
and Vazirani have shown that there is a strategy for the cut player,
guaranteeing that after $O(\log^2N)$ iterations, no matter the
strategy of the matching player, the resulting graph is
a $\ensuremath{\frac{1}{2}}$-expander w.h.p. Subsequently, Orecchia et
al.~\cite{better-CMG} have shown the following improved bound:
\begin{theorem}[\cite{better-CMG}]\label{thm: CMG}
There is a probabilistic algorithm for the cut player, such that, no
matter how the matching player plays, after $\ensuremath{\gamma_{\mbox{\tiny{\sc CMG}}}}(N)=O(\log^2N)$
iterations, graph $X$ is an $\ensuremath{\alpha_{\mbox{\tiny{\sc CMG}}}}(N)=\Omega(\log N)$-expander,
with constant probability.
\end{theorem}
Our algorithms work by embedding an expander $X$ into a sub-graph of
$G$. The embedding of the expander is then used to certify the
treewidth. We use the following notion of embedding.
\begin{definition}
Let $G,X$ be graphs. An embedding $\phi$ of $X$ into $G$ maps every
vertex $v\in X$ to a connected subgraph $C_v\subseteq G$, and every
edge $e=(u,v)\in E(X)$ to a path $P_e$ in graph $G$, whose endpoints
belong to $C_v$ and $C_u$, respectively. We say that the
\emph{congestion} of the embedding is at most $c$, iff every edge of
$G$ belongs to at most $c-1$ paths in $\set{P_e\mid e\in E(X)}$ and
at most one graph $\set{C_v\mid v\in V(X)}$.
\end{definition}
In the next simple claim, we show that if we can embed a
$\kappa$-vertex expander with congestion at most $c$ into a graph $H$
with bounded vertex degree, then the treewidth of $H$ is large.
\begin{claim}\label{claim: embed an expander treewidth}
Let $X$ be an $\alpha$-expander on $\kappa$ vertices for $\alpha<1$,
with maximum vertex degree $\Delta'$, and let $H$ be a graph with
maximum vertex degree at most $\Delta$, such that that there is an
embedding of $X$ into $H$ with congestion $\eta$. Then $\mathrm{tw}(H) =
\Omega\left (\frac{\alpha\kappa}{\eta\Delta\Delta'}\right )$.
\end{claim}
\begin{proof}
For each vertex $v\in V(X)$, let $t_v$ be an arbitrary vertex in
$C_v$, and let ${\mathcal T}=\set{t_v\mid v\in V(X)}$. Since $|X|=\kappa$,
from Theorem~\ref{thm: weak well-linkedness to tw}, it is enough to
show that ${\mathcal T}$ is $\frac{\alpha}{2\eta\Delta'}$-well-linked in
$H$. Let $(A,B)$ be any partition of $V(H)$, denote ${\mathcal T}_A=A\cap
{\mathcal T}$, ${\mathcal T}_B=B\cap {\mathcal T}$, and $E'=E(A,B)$. Assume w.l.o.g. that
$|{\mathcal T}_A|\leq |{\mathcal T}_B|$. Then it is enough to show that $|E'|\geq
|{\mathcal T}_A|\cdot \frac{\alpha}{2\eta\Delta'}$.
We partition ${\mathcal T}_A$ into two subsets, ${\mathcal T}_A'$ and ${\mathcal T}_A''$,
as follows. For each vertex $t_v\in {\mathcal T}_A$, if $C_v\subseteq A$,
then we add $t_v$ to ${\mathcal T}_A''$, and otherwise we add it to
${\mathcal T}_A'$. We partition ${\mathcal T}_B$ into two subsets, ${\mathcal T}_B'$ and
${\mathcal T}_B''$ similarly. Let $\kappa'=|{\mathcal T}_A|$. Assume first that
$|{\mathcal T}_A'|\geq \kappa'/2$. Then for each vertex $t_v\in {\mathcal T}_A'$,
at least one edge of $C_v$ belongs to $E'$. Since every edge of $H$
may belong to at most one graph in $\set{C_u\mid u\in V(X)}$,
$|E'|\geq |{\mathcal T}_A'|\geq \kappa'/2$ must hold. Similarly, if
$|{\mathcal T}_B'|\geq \kappa'/2$, $|E'|\geq \kappa'/2$.
Therefore, we assume from now on that
$|{\mathcal T}_A'|,|{\mathcal T}_B'|<\kappa'/2$, and so
$|{\mathcal T}_A''|,|{\mathcal T}_B''|\geq \kappa'/2$. Let $U_A=\set{v\in V(X)\mid
t_v\in {\mathcal T}_A''}$, and $U_B=\set{v\in V(X)\mid t_v\in
{\mathcal T}_B''}$. Since $X$ is an $\alpha$-expander, there is a set
${\mathcal{P}}$ of at least $\kappa'/2$ paths connecting the vertices of
$U_A$ to the vertices of $U_B$ in $X$, such that the edge-congestion
of ${\mathcal{P}}$ is at most $1/\alpha$. Since the maximum vertex degree in
$X$ is $\Delta'$, by sending $\alpha/\Delta'$ flow units along each
path in ${\mathcal{P}}$, we obtain a flow from the vertices in $U_A$ to the
vertices of $U_B$ of value at least
$\frac{\kappa'\alpha}{2\Delta'}$, where the flow across each vertex
is at most $1$. From the integrality of flow, there is a set
${\mathcal{P}}'$ of at least $\frac{\kappa'\alpha}{2\Delta'}$ node-disjoint paths in
$X$, where each path connects a vertex of $U_A$ to a vertex of
$U_B$. We now build a new set ${\mathcal{P}}^*$ of
$\frac{\kappa'\alpha}{2\Delta'}$ paths in graph $H$, connecting the
vertices of ${\mathcal T}_A''$ to the vertices of ${\mathcal T}_B''$, as follows.
Consider some path $P\in {\mathcal{P}}'$, and assume that its endpoints are
$v$ and $u$, so $t_v\in {\mathcal T}_A''$ and $t_u\in {\mathcal T}_B''$. We build
a new graph $H_P$, that includes, for every edge $e\in P$, the path
$P_e$ into which $e$ is embedded, and for every vertex $v'\in P$,
the sub-graph $C_{v'}$, where $v'$ is embedded into $C_{v'}$. It is
easy to see that $H_P$ contains a path connecting $t_v$ to
$t_u$. Let $P^*$ be any such path. We then set ${\mathcal{P}}^*=\set{P^*\mid
P\in {\mathcal{P}}'}$. Since the paths in ${\mathcal{P}}'$ are node-disjoint, and
the embedding of $X$ into $H$ has congestion $\eta$, every edge of
$H$ belongs to at most $\eta$ paths in ${\mathcal{P}}^*$. Since every path
in ${\mathcal{P}}^*$ must contain an edge in $E'$, $|E'|\geq
\frac{\kappa'\alpha}{2\eta\Delta'}$. We conclude that ${\mathcal T}$ is
$\frac{\alpha}{2\eta\Delta'}$-well-linked, and from
Theorem~\ref{thm: weak well-linkedness to tw}, the treewidth of $H$
is $\Omega\left (\frac{\alpha\kappa}{\eta\Delta\Delta'}\right )$.
\end{proof}
\fi
\label{--------------------------------------------sec: building the minor----------------------------------------------}
\section{A Small Treewidth-Preserving Degree-4 Minor}\label{sec: top minor}
\iffull
In this section we prove the following theorem which gives a
degree-$4$ sparsifier.
\begin{theorem}\label{thm: degree-4 minor}
There is a randomized algorithm, that, given a graph $G$ of
treewidth at least $k$, w.h.p. computes a minor $H$ of $G$, such
that:
\begin{itemize}
\item the treewidth of $H$ is $\Omega(k/\operatorname{poly}\log k)$;
\item every vertex has degree at most $4$ in $H$; and
\item $|V(H)|=O(k^4)$.
\end{itemize}
The running time of the algorithm is polynomial in $|V(G)|$ and $k$.
\end{theorem}
In order to prove Theorem~\ref{thm: degree-4 minor}, it is sufficient to
find a subgraph $H$ of $G$, with $\tau(H)=O(k^4)$, such that the
maximum vertex degree in $H$ is at most $4$, and the treewidth of $H$
is $\Omega(k/\operatorname{poly}\log k)$. Indeed, by replacing every maximal
$2$-path in $H$ with an edge connecting its endpoints, we obtain the
desired minor.
\fi
\ifabstract In this section we prove a slightly weaker version of
Theorem~\ref{thm: main-topological-minor} that yields a degree-$4$
sparsifier instead of a degree-$3$ sparsifier; the other parameters
are the same. A formal statement can be found in the Appendix. To
prove the theorem it suffices to find a subgraph $H$ of $G$, with
$\tau(H)=O(k^4)$, such that the maximum vertex degree in $H$ is at
most $4$, and the treewidth of $H$ is $\Omega(k/\operatorname{poly}\log k)$. Indeed,
by replacing every maximal $2$-path in $H$ with an edge connecting its
endpoints, we obtain the desired topological minor.
\fi
We start by applying Theorem~\ref{thm: strong PoS system} with
$r=\gkrv(k)$ and $h=\Omega(k/\operatorname{poly}\log k)$, so that $h$ is an even
integer, and $\frac{k}{\log^{c'}k}>chr^{48}$ holds, where $c,c'$ are
the constants from Theorem~\ref{thm: strong PoS system}. Let
$({\mathcal{{S}}},\bigcup_{i=1}^{r-1}{\mathcal{P}}_i)$ be the resulting path-of-sets
system.
Our next step is to construct an expander graph $X$ on $h$ vertices,
and to embed it into a sub-graph $H$ of $G$. Following the previous
work on routing
problems~\cite{RaoZhou,Andrews,Chuzhoy11,ChuzhoyL12,ChekuriE13}, we
will embed $X$ into $G$ using the cut-matching game, and the
path-of-sets system $({\mathcal{{S}}},\bigcup_{i=1}^{r-1}{\mathcal{P}}_i)$. We start
with an intuitive high-level description of the algorithm. For each
$1\leq i\leq r$, let ${\mathcal{Q}}_i$ be any set of node-disjoint paths
connecting $A_i$ to $B_i$ in $G[S_i]$ (this set exists due to the
linkedness of $(A_i,B_i)$ in $G[S_i]$). Let ${\mathcal{H}}$ be the set of $h$
paths, obtained by concatenating
${\mathcal{Q}}_1,{\mathcal{P}}_1,{\mathcal{Q}}_2,{\mathcal{P}}_2,\ldots,{\mathcal{P}}_{r-1},{\mathcal{Q}}_r$. We
denote ${\mathcal{H}}=\set{P_1,\ldots,P_{h}}$. The high-level idea is to
construct an expander $X$ over a set $V=\set{v_1,\ldots,v_{h}}$ of $h$
vertices, and to embed it into $G$ using the cut-matching game, as
follows. For each $1\leq i\leq h$, we embed $v_i$ into $P_i$, that is,
$C_{v_i}=P_i$. We construct the edges of the expander, and embed them
into $G$, using the cut-matching game, where for each $1\leq j\leq r$,
we use cluster $S_j\in {\mathcal{{S}}}$ to route the $j$th matching, as
follows. A partition $(Y,Z)$ of the vertices of $X$ computed by the
cut player naturally defines a partition $({\mathcal{H}}_Y,{\mathcal{H}}_Z)$ of the
paths in ${\mathcal{H}}$ into two equal-sized subsets, which in turn defines a
partition $(A_j',A_j'')$ of $A_j$ into two equal-sized subset. Using
the fact that $A_j$ is node-well-linked inside $G[S_j]$, we can find a
set ${\mathcal{B}}_j$ of node-disjoint paths in $G[S_j]$ connecting $A_j'$ to
$A_j''$. This set of paths defines a matching $M_j$ between the paths
in ${\mathcal{H}}_Y$ and ${\mathcal{H}}_Z$, and hence between the vertices of $Y$ and
$Z$ in $X$. We view this matching as the response of the matching
player. After $\gkrv(h)\leq \gkrv(k)=r$ iterations, we obtain an
expander $X$ and its embedding with congestion $2$ into
$G$. Intuitively, we would like to define $H$ as the set of all edges
and vertices of $G$ used in this embedding, that is, the union of the
paths in ${\mathcal{H}}$ and $\bigcup_{j=1}^r{\mathcal{B}}_j$. It is easy to see that
the maximum vertex degree in $H$ is at most $4$, since in each cluster
$S_j$ we only route 2 sets of node-disjoint paths: ${\mathcal{Q}}_j$ and
${\mathcal{B}}_j$. However, $\tau(H)$ may not be bounded by $O(k^4)$. In
particular, the paths in ${\mathcal{Q}}_j$ and ${\mathcal{B}}_j$ may intersect at many
vertices. In order to overcome this difficulty, we can use
Theorem~\ref{thm: 2-flow-main} to find new sets ${\mathcal{Q}}_j'$ and
${\mathcal{B}}_j'$ of paths, routing the same pairs of vertex subsets, such
that, if $J=J({\mathcal{Q}}_j'\cup {\mathcal{B}}_j')$, then $\tau(J)=O(h^4)$. However,
this re-routing changes the paths in ${\mathcal{H}}$, and therefore the
mapping between the vertices in sets $A_{j'}$ for $j'\neq j$ and the
vertices in $X$ may be changed. Therefore, we need to execute this
procedure more carefully. In particular, we apply Theorem~\ref{thm:
2-flow-main} in the graph $G[S_j]$ after each iteration $j$ of the
cut-matching game; for iteration $j+1$ we exploit the
node-well-linkedness of the set $A_{j+1}$ in $G[S_{j+1}]$ to maintain
consistency in the mapping of paths in ${\mathcal{H}}$ to the vertices of the
expander. \ifabstract The formal description of the embedding
procedure can be found in the Appendix. \fi \iffull We now provide a
formal description of the embedding procedure.
We will gradually construct the set ${\mathcal{H}}$ of paths over the course
of $\gkrv(h)$ iterations. For each $1\leq j\leq \gkrv(h)$, at the
beginning of the $j$th iteration, we are given a set ${\mathcal{H}}^j$ of $h$
disjoint paths, connecting the vertices of $A_1$ to the vertices of
$A_j$, and a bijection $f: {\mathcal{H}}^j\rightarrow V(X)$. At the beginning,
${\mathcal{H}}^1$ consists of $h$ paths, where each path consists of a single
distinct vertex of $A_1$, and the mapping $f:{\mathcal{H}}^1\rightarrow V(X)$
is an arbitrary bijection. We also start with a graph $X$ on $h$
vertices, and $E(X)=\emptyset$. For $1\leq j\leq \gkrv(h)$, the $j$th
iteration is executed as follows.
We use the cut player on the current graph $X$ to compute a partition
$(Y_j,Z_j)$ of $V(X)$ into two equal-sized subsets. This naturally
defines a partition $({\mathcal{H}}^j_Y,{\mathcal{H}}^j_Z)$ of ${\mathcal{H}}^j$ where
${\mathcal{H}}^j_Y$ contains all paths $P\in {\mathcal{H}}^j$, such that $f(P)\in
Y_j$. In turn, this gives a partition $(A_j',A_j'')$ of $A_j$, where a
vertex $v\in A_j$ belongs to $A_j'$ iff the path $P\in{\mathcal{H}}^j$ on which $v$ lies
belongs to ${\mathcal{H}}^j_Y$. Since the set $A_j$ of vertices is
node-well-linked in $G[S_j]$, there is a collection of node-disjoint
paths routing $(A_j',A_j'')$ in $G[S_j]$. Since $A_j$ and $B_j$ are
linked in $G[S_j]$, there is a collection of node-disjoint paths
routing $(A_j,B_j)$ in $G[S_j]$. From Theorem~\ref{thm: 2-flow-main},
we can find a set ${\mathcal{B}}_j$ of paths routing $(A_j',A_j'')$, and a
set ${\mathcal{Q}}_j$ of paths routing $(A_j,B_j)$ in $G[S_j]$, such that, if
$J=J({\mathcal{B}}_j\cup {\mathcal{Q}}_j)$, then the maximum vertex degree in $J$ is
bounded by $4$, the degree of every vertex in $A_j\cup B_j$ is at most
$3$, and $\tau(J)\leq O(h^4)$. We let ${\mathcal{H}}^{j+1}$ be the
concatenation of the paths in ${\mathcal{H}}^j$, ${\mathcal{Q}}_j$, and ${\mathcal{P}}_j$. In
order to define the mapping $f:{\mathcal{H}}^{j+1}\rightarrow V(X)$, for each
$P\in {\mathcal{H}}^{j+1}$, let $P'\in {\mathcal{H}}^j$ be the sub-path of $P$. Then
we set $f(P)=f(P')$. Notice that the set ${\mathcal{B}}_j$ of paths defines a
complete matching between the vertices in $A_j'$ and $A_j''$, and by
extension, a complete matching between the paths in ${\mathcal{H}}^j_Y$ and
${\mathcal{H}}^j_Z$, which in turn naturally defines a matching $M_j$ between
$Y_j$ and $Z_j$ in $X$. We add the edges of the matching $M_j$ to
$X$. Each edge $e=(v_i,v_{i'})\in M_j$ is mapped to the corresponding
path in $P_e\in {\mathcal{B}}_j$, that connects the unique vertex of $P_i\cap
A_j$ to the unique vertex of $P_{i'}\cap A_j$.
Let graph $H$ be the union of all paths in ${\mathcal{H}}^{\gkrv(h)}$ and
$\bigcup_{j=1}^{\gkrv(h)}{\mathcal{B}}_j$. Then it is easy to see that the
maximum vertex degree in $H$ is at most $4$, and $\tau(H)\leq
O(h^4\gkrv(h))=O(k^4)$. Moreover, every edge of $H$ belongs to at most
one path in ${\mathcal{H}}$, and at most one path in
$\bigcup_{j=1}^{\gkrv(h)}{\mathcal{B}}_j$. Therefore, we have constructed an
embedding of an $h$-vertex $\ensuremath{\alpha_{\mbox{\tiny{\sc CMG}}}}(h)$-expander $X$, whose maximum
vertex degree is $\gkrv(h)$, into $H$ with congestion 2, and so from
Claim~\ref{claim: embed an expander treewidth}, the treewidth of $H$
is at least $\Omega(h/\gkrv(h))=\Omega(k/\operatorname{poly}\log k)$.
\fi
\label{---------------------------------------sec: degree-3----------------------------------------------}
\section{Building a Degree-$3$ Minor}\label{sec: degree-3}
In this section we complete the proof of Theorem~\ref{thm:
main-topological-minor}. We start with an informal overview to help
understand the high-level plan. A reader may wish to skip it and go
directly to the formal proof.
\subsection{Overview}
We use an algorithm, similar to the one used in Section~\ref{sec: top
minor}, in order to embed an expander into $G$, using the
path-of-sets system. The main difference is that, instead of embedding
a single expander $X$, we will embed $N$ expanders $X_1,\ldots,X_N$
where $N = \Theta(\log k)$. For this purpose we start with a longer
path-of-sets system $({\mathcal{{S}}},\bigcup_{i=1}^{r-1}{\mathcal{P}}_i)$ with
parameters $h=k/\operatorname{poly}\log k$ and $r=O(\log^3k)$ and partition it into
$N=O(\log k)$ smaller path-of-sets systems with parameters
$r^*=\gkrv(h)$ and $h$ (hence $r \simeq N r^*$). For $1\leq i\leq N$, we
embed an expander $X_i$ into the $i$'th path-of-sets system using the
approach in the preceding section. Recall that for each cluster $S_i$,
we construct two sets of paths contained in $G[S_i]$: one set,
${\mathcal{R}}_i$, that we call red paths, routes $(A_i,B_i)$, and another
set, ${\mathcal{B}}_i$, that we call blue paths, routes $(A_i',A_i'')$, where
$(A_i',A_i'')$ is the partition of $A_i$ defined by the cut
player. Let $H_i$ be the topological minor of $G[S_i]$ obtained by
taking the union of the paths in ${\mathcal{R}}_i$ and ${\mathcal{B}}_i$, and
suppressing all degree-$2$ vertices, except for $A_i\cup B_i$. We
assume that $H_i$ is minimal in the following sense: for each edge $e$
of $H_i$, either $(A_i,B_i)$ or $(A_i',A_i'')$ is not routable in
$H_i\setminus \set{e}$. Abusing the notation, we assume that ${\mathcal{R}}_i$
and ${\mathcal{B}}_i$ are the sets of the red and the blue paths, routing
$(A_i,B_i)$ and $(A_i',A_i'')$, respectively in $H_i$. Notice that
every vertex of $H_i$ must lie on some red path. Let ${\mathcal{H}}$ be the
set of $h$ paths obtained by concatenating the paths in
${\mathcal{R}}_1,{\mathcal{P}}_1,\ldots,{\mathcal{P}}_{Nr^*-1},{\mathcal{R}}_{Nr^*}$, and let $H$ be
the topological minor of $G$ obtained by taking the union of the
graphs $H_i$ and the paths $\bigcup_{i=1}^{Nr^*-1}{\mathcal{P}}_i$.
We say that an edge of $H$ is
a red edge if it belongs to a red path and no blue paths; it is a blue
edge if it belongs to a blue path and no red paths; and it is a
red-blue edge if it belongs to both a red and a blue path. We can view
the $N$ different expanders as sharing the same vertex set, where the
vertices correspond to the paths in ${\mathcal{H}}$.
Consider a vertex $v$ of
degree $4$ in graph $H$; it must be incident to two red edges and two
blue edges. In order to reduce the degree to $3$, we use random
sampling to pick one of the two blue edges incident to $v$ and eliminate
it. After this step the degree of every vertex is at most $3$. Let
$H^*$ be this final topological minor of $G$. The heart of the
analysis is to show that $H^*$ has treewidth $\Omega(k/\operatorname{poly}\log k)$.
This is done by showing that the set $A=A_1$ of vertices remains
$\alpha$-well-linked in $H^*$, for $\alpha=\Omega(1/\operatorname{poly}\log k)$, and
applying Theorem~\ref{thm: weak well-linkedness to tw}.
We start by observing that the set $A$ of vertices is
$\alphawl$-well-linked in $H$, for some constant $\alphawl$. This is
shown by using the embeddings of the expanders $X_1,\ldots,X_N$ into
$H$. Next, we carefully partition each path in ${\mathcal{H}}$ into a
collection of disjoint segments. Intuitively, each segment of a path
$P\in {\mathcal{H}}$ is a sub-path of $P$ of length $\Theta(\operatorname{poly}\log h)$. We
then contract each such segment $\sigma$ into a super-node
$v_{\sigma}$. Let $F$ be this contracted graph, and let $F^*$ be the
corresponding contracted graph of $H^*$. Equivalently, $F^*$ is
obtained from $F$ by deleting all the edges in $E(H)\setminus E(H^*)$.
Each vertex of $A$ belongs to a distinct contracted segment, and is
associated with the corresponding super-node in $F$. We do not
distinguish between the vertices of $A$ and their corresponding
super-nodes. It is easy to see that $A$ remains $\alphawl$-well linked
in $F$ since we only contracted edges. The most crucial property of
the contracted graph $F$ is that the value of the minimum cut in $F$
is at least $\Omega(\log |V(F)|)$. This allows us to use arguments
similar to those used in Karger's sampling
technique~\cite{Karger-skeleton} to show that all cuts are
approximately preserved in $F^*$. In particular, the vertices of $A$
remain $\alphawl/32$-well-linked in $F^*$. Since the length of every
segment used in the construction of the contracted graph $F$ is
$O(\operatorname{poly}\log h)$, this implies that the vertices of $A$ are
$\alpha$-well-linked in $H^*$, for $\alpha=\Omega(1/\operatorname{poly}\log k)$.
The most challenging part of the proof is to set up the partition of
the paths in ${\mathcal{H}}$ into segments, so that in the resulting
contracted graph $F$, the value of the minimum cut is $\Omega(\log
|V(F)|)$. At a high-level, the proof proceeds as follows. Assume for
contradiction, that there is a partition $(X,Y)$ of $V(F)$ with
$X,Y\neq \emptyset$, and $|E_F(X,Y)|<N$. Let $X'\subseteq V(H)$ be
obtained from $X$ by un-contracting all super-nodes in $X$, and let
$Y'\subseteq V(H)$ be obtained from $Y$ similarly. Then $(X',Y')$ is a
partition of $V(H)$, and $|E_{H}(X',Y')|<N$. Assume first that there
are two paths $P,P'\in {\mathcal{H}}$, such that $P\subseteq H[X']$ and
$P'\subseteq H[Y']$. We then use the embeddings of the expanders
$X_1,\ldots,X_N$ to argue that $|E_H(X',Y')|\geq N$, reaching a
contradiction. Therefore, we can assume w.l.o.g. that no path of
${\mathcal{H}}$ is contained in $H[X']$. We next show that for some $1\leq
i^*\leq Nr^*$, partition $(X',Y')$ of $V(H)$ defines a partition
$(X^*,Y^*)$ of $V(H_{i^*})$, such that $|X^*|,|Y^*|>200N^4$, while
$|E_{H_{i^*}}(X^*,Y^*)|<N$. We then consider the segments of the red
paths in ${\mathcal{R}}_{i^*}$ and the blue paths in ${\mathcal{B}}_{i^*}$ that are
contained in $H_{i^*}[X^*]$. Let ${\mathcal{R}}^*$ denote the corresponding
segments of the red paths, and ${\mathcal{B}}^*$ the corresponding segments of
the blue paths. Using Theorem~\ref{thm: 2-flow-minor}, we show that
there is some edge $e\in H[X^*]$, such that we can still route the
endpoints of the paths in ${\mathcal{R}}^*$ to each other, and the endpoints
of the paths in ${\mathcal{B}}^*$ to each other, even after deleting $e$ from
$H[X^*]$. This new routing implies that we can route both
$(A_{i^*},B_{i^*})$ and $(A_{i^*}',A_{i^*}'')$ in
$H_{i^*}\setminus\set{e}$, contradicting the minimality of $H_{i^*}$.
\subsection{Proof of Theorem~\ref{thm: main-topological-minor}}
We set $r=2^{15}\log k\cdot \gkrv(k) =\Theta(\log^3 k)$, and
$h=\Omega(k/\operatorname{poly}\log k)$, so that $h$ is an even integer, and
$k/\log^{c'}k>chr^{48}$, where $c$ and $c'$ are the constants from
Theorem~\ref{thm: strong PoS system}. We assume w.l.o.g. that $k$ is
large enough, so $h>72\log k$ and $h>\gkrv(h)$. We then apply
Theorem~\ref{thm: strong PoS system} to graph $G$, with parameters $r$
and $h$, to obtain a strong path-of-sets system
$({\mathcal{{S}}},\bigcup_{i=1}^{r-1}{\mathcal{P}}_i)$ of height $h$ and width $r$.
Let $r^*=\gkrv(h)$, and let $N=\ceil{3072\log(10h^4\cdot r^*)}$;
it is easy to see that $N=\Theta(\log h)$. We will assume
w.l.o.g. that $h$ is large enough, so $N>1536\log(10h^4\cdot r^*\cdot
N)$ holds. Finally, we let $r'=N \cdot r^*$. Note that $r'=r^* \cdot
\ceil{3072\log(10h^4\cdot r^*)}\leq 2^{15}\gkrv(h) \log h<r$.
We construct a new, smaller, path-of-sets system, of height $h$ and
width $r'$, using the clusters ${\mathcal{{S}}}'=(S_1,\ldots,S_{r'})$, and the
sets ${\mathcal{P}}_i$ of paths, for $1\leq i\leq r'-1$; in other words
we restrict attention to the first $r'$ clusters from the initial path-of-sets
system. Abusing notation, we denote $r'$ by $r$ and ${\mathcal{{S}}}'$ by ${\mathcal{{S}}}$.
We denote by $G'$ the following minor of $G$: start with the union of
$G[S_i]$ for all $1\leq i\leq r$; for each path $P\in
\bigcup_{i=1}^{r-1}{\mathcal{P}}_i$, add an edge connecting the endpoints of
$P$ to $G'$. We denote by $E_i$ the set of edges corresponding to the
paths in ${\mathcal{P}}_i$. Equivalently, we obtain $G'$ from graph $\left (\bigcup_{S_i\in {\mathcal{{S}}}}G[S_i]\right )\cup\left (\bigcup_{j=1}^{r-1}{\mathcal{P}}_j\right )$ by suppressing degree-$2$ internal nodes on the paths in $\bigcup_{i=1}^{r-1}{\mathcal{P}}_i$. It
is now enough to find a topological minor $H^*$ of $G'$ whose treewidth is
$\Omega(k/\operatorname{poly}\log k)$, maximum vertex degree is $3$, and
$|V(H^*)|=O(k^4)$. We do so via the following theorem:
\begin{theorem}\label{theorem: finding a minor in a super-cluster}
There is an efficient randomized algorithm, that
finds a topological minor $H^*$ of $G'$, such that
w.h.p.:
\begin{itemize}
\item $|V(H^*)|=O(h^4\cdot r)$;
\item The maximum
vertex degree in $H^*$ is $3$;
\item $A_1\subseteq V(H^*)$; and
\item The set $A_1$ of vertices is
$\alpha$-well-linked in $H^*$, for $\alpha=\Omega(1/\log^7 k)$.
\end{itemize}
\end{theorem}
Theorem~\ref{thm: main-topological-minor} follows easily from
Theorem~\ref{theorem: finding a minor in a super-cluster}. The
desired topological minor of $G$ is $H^*$. The only property that is
left to verify is that $\mathrm{tw}(H) = \Omega(k/\polylog k)$ which follows
from $\alpha$-well-linkedness of $A_1$ in $H^*$. Indeed,
Theorem~\ref{thm: weak well-linkedness to tw} implies that $\mathrm{tw}(H) =
\Omega(\alpha |A_1|/3) = \Omega(k/\polylog k)$ since
$|A_1| = h = \Omega(k/\polylog k)$, $\alpha = \Omega(1/\log^7 k)$
and $H^*$ has maximum degree $3$.
From now on we focus on proving Theorem~\ref{theorem: finding a minor in a super-cluster}.
In order to simplify the notation, we refer to the graph $G'$ as $G$.
Recall that we are given a path-of-sets system
$({\mathcal{{S}}}=(S_1,\ldots,S_r),\bigcup_{i=1}^{r-1}{\mathcal{P}}_i)$ of height $h$
and width $r=Nr^*$ in $G$, where for each $1\leq i< r$, each path in
${\mathcal{P}}_i$ consists of a single edge, and the corresponding set of
edges is denoted by $E_i$. Let $E'=\bigcup_{i=1}^{r-1}E_i$. We denote
$A_1$ by $A$. Our goal is to construct a topological minor $H^*$ of
$G$, such $|V(H^*)|=O(h^4 r)$, the maximum vertex degree of $H^*$ is
$3$, while ensuring that $A\subseteq V(H^*)$ and it is
$\alpha$-well-linked in $H^*$, w.h.p.
The rest of the proof consists of three steps. In the first step, we
define the sets ${\mathcal{B}}_i,{\mathcal{R}}_i$ of paths for $1\leq i\leq r$ by
playing the cut-matching games; in the second step we partition the
resulting red paths into segments; and in the third step we complete
the proof of the theorem.
\paragraph{Step 1: Cut-Matching Games}
In this step we construct $N$ expanders $X_1,\ldots,X_N$, and embed each of them separately into $G$. For each $1\leq i\leq N$, let ${\mathcal{{S}}}_i=(S_{(i-1)r^*+1},\ldots,S_{ir^*})$, let $E^i=\bigcup_{j=(i-1)r^*+1}^{ir^*-1}E_j$, and let $\hat E^i=E_{ir^*}$ (for $i=N$, $\hat E^i=\emptyset$). Let $G_i$ be the graph obtained from the union of $G[S_j]$ for all $S_j\in {\mathcal{{S}}}_i$ and the edges in $E^i$. For each $1\leq i\leq N$, we embed the expander $X_i$ into $G_i$, using the cut-matching game, as follows. For convenience, we denote $(i-1)r^*$ by $z$.
We will gradually construct a set ${\mathcal{H}}_i$ of paths over the course of $r^*$ iterations. For each $1\leq j\leq r^*$, at the beginning of the $j$th iteration, we are given a set ${\mathcal{H}}^j$ of $h$ disjoint paths, connecting the vertices of $A_{z+1}$ to the vertices of $A_{z+j}$, and a bijection $f: {\mathcal{H}}^j\rightarrow V(X_i)$. At the beginning, ${\mathcal{H}}^1$ consists of $h$ paths, each of which consists of a single distinct vertex of $A_{z+1}$, and the mapping $f:{\mathcal{H}}^1\rightarrow V(X_i)$ is an arbitrary bijection. We also start with a graph $X_i$ on $h$ vertices, and $E(X_i)=\emptyset$. For $1\leq j\leq r^*$, the $j$th iteration is executed as follows.
We use the cut player on the current graph $X_i$ to find a partition $(Y_j,Z_j)$ of $V(X_i)$ into two equal-sized subsets. This naturally defines a partition $({\mathcal{H}}^j_Y,{\mathcal{H}}^j_Z)$ of ${\mathcal{H}}^j$ where ${\mathcal{H}}^j_Y$ contains all paths $P\in {\mathcal{H}}^j$, such that $f(P)\in Y_j$. In turn, this gives a partition $(A_{z+j}',A_{z+j}'')$ of $A_{z+j}$, where a vertex $v\in A_{z+j}$ belongs to $A_{z+j}'$ iff the path $P$ on which $v$ lies belongs to ${\mathcal{H}}^j_Y$. Since the set $A_{z+j}$ of vertices is node-well-linked in $G[S_{z+j}]$, there is a collection of node-disjoint paths routing $(A_{z+j}',A_{z+j}'')$ in $G[S_{z+j}]$. Since $A_{z+j}$ and $B_{z+j}$ are linked in $G[S_{z+j}]$, there is a collection of node-disjoint paths routing $(A_{z+j},B_{z+j})$ in $G[S_{z+j}]$. From Theorem~\ref{thm: 2-flow-main}, we can find a set ${\mathcal{B}}_{z+j}'$ of paths routing $(A_{z+j}',A_{z+j}'')$ , and a set ${\mathcal{R}}_{z+j}'$ of paths routing $(A_{z+j},B_{z+j})$ in $G[S_{z+j}]$, such that, if $J=J({\mathcal{B}}_{z+j}'\cup{\mathcal{R}}_{z+j}')$, then the maximum vertex degree in $J$ is bounded by $4$, the degree of every vertex in $A_{z+j}\cup B_{z+j}$ is at most $3$, and $\tau(J)\leq 8h^4+8h$.
We will assume that $J$ is a minimal graph in which $(A_{z+j}',A_{z+j}'')$ and $(A_{z+j},B_{z+j})$ are both routable: that is, for every edge $e\in E(J)$, either $(A_{z+j}',A_{z+j}'')$, or $(A_{z+j},B_{z+j})$ are not routable in $J\setminus \set{e}$. We let $H_{z+j}$ be the graph obtained from $J$ by replacing every maximal $2$-path that does not contain the vertices of $A_{z+j}\cup B_{z+j}$ as inner vertices, by an edge connecting its two endpoints. Then $|V(H_{z+j})|\leq 8h^4+8h\leq 10h^4$, every vertex of $H_{z+j}$ has degree at most $4$, while the vertices in $A_{z+j}\cup B_{z+j}$ have degree at most $3$; there is a set ${\mathcal{B}}_{z+j}$ of paths routing $(A_{z+j}',A_{z+j}'')$ , and a set ${\mathcal{R}}_{z+j}$ of paths routing $(A_{z+j},B_{z+j})$ in $H_{z+j}$, and for every edge $e\in E(H_{z+j})$, either $(A_{z+j}',A_{z+j}'')$, or $(A_{z+j},B_{z+j})$ are not routable in $H_{z+j}\setminus \set{e}$.
We call the paths in ${\mathcal{R}}_{z+j}$ \emph{red paths}, and the paths in ${\mathcal{B}}_{z+j}$ \emph{blue paths}. An edge that belongs to a red path, but no blue paths is called a red edge. An edge the belongs to a blue path but no red paths is called a blue edge. An edge that lies on a red and a blue path is called a red-blue edge. Notice that a vertex of $H_{z+j}$ has degree $4$ only if it is incident on two blue edges. Each vertex in $A_{z+j}$ serves as a source of a red path and a source or a destination of a blue path, so it can only be incident on at most two edges in $H_{z+j}$. A vertex $v\in B_{z+j}$ serves as a destination of a red path; its degree is at most $3$, and it is equal to $3$ only if $v$ is incident on two blue edges.
We let ${\mathcal{H}}^{j+1}$ be the concatenation of the paths in ${\mathcal{H}}^j$, ${\mathcal{R}}_{z+j}$, and $E_{z+j}$. In order to define the mapping $f:{\mathcal{H}}^{j+1}\rightarrow V(X_i)$, for each $P\in {\mathcal{H}}^{j+1}$, let $P'\in {\mathcal{H}}^j$ be the sub-path of $P$. Then we set $f(P)=f(P')$. Notice that the set ${\mathcal{B}}_{z+j}$ of paths defines a matching between the paths in ${\mathcal{H}}^j_Y$ and ${\mathcal{H}}^j_Z$, which in turn naturally defines a matching $M_j$ between $Y_j$ and $Z_j$ in $X_i$. We add the edges of the matching $M_j$ to $X$. Each edge $e=(v_i,v_{i'})\in M_j$ is mapped to the corresponding path in ${\mathcal{B}}_{z+j}$, that connects the unique vertex in $A_j\cap f^{-1}(v_i)$ to the unique vertex in $A_j\cap f^{-1}(v_{i'})$.
Finally, we set ${\mathcal{H}}_i={\mathcal{H}}^{r^*}$. Let $\tilde H_i$ be the union of the graphs $H_{z+1},\ldots,H_{z+r^*}$, and the edges $E^i$. Then we have defined an $\ensuremath{\alpha_{\mbox{\tiny{\sc CMG}}}}(h)$-expander $X_i$ on $h$ vertices with maximum vertex degree $\gkrv(h)$, and embedded it with congestion $2$ into $\tilde H_i$, where each vertex of $X_i$ is embedded into a distinct path in ${\mathcal{H}}_i$.
Let $H$ be the union of the graphs $\tilde H_i$, for $1\leq i\leq N$
and $\bigcup_{i=1}^{N-1}\hat E^i$, and let ${\mathcal{H}}$ be the
concatenation of ${\mathcal{H}}_1,\hat E_1,\ldots,\hat E_{N-1},{\mathcal{H}}_N$. We
will sometimes refer to the paths in ${\mathcal{H}}$ as red paths. All
vertices in $H$ have degree at most $4$, and, as observed before, a
vertex of $H$ may have degree $4$ only if it is incident on exactly
two blue edges. Every vertex in $A$ has degree at most $2$. Our final
graph $H^*$ is obtained from $H$ as follows: for each vertex $v\in
V(H)$ that is incident on two blue edges, we independently choose one
of these two blue edges at random. Each blue edge that has been chosen
by at least one vertex is then deleted from the graph. This final
graph is denoted by $H^*$. Notice that each edge $e=(u,v)$ may be
deleted from $H$ due to the choice made by $u$, or the choice made by
$v$; the overall probability that $e$ is not deleted is at least
$1/4$. Moreover, if $e$ and $e'$ do not share endpoints, then the
events that $e$ is deleted and that $e'$ is deleted are independent.
It is immediate to see that $|V(H^*)|\leq Nr^*\cdot O(h^4)=O(r h^4)$; the vertices of $A$ are contained in $V(H^*)$, and the maximum vertex degree in $H^*$ is $3$. It now only remains to prove that w.h.p. the vertices of $A$ are $\alpha$-well-linked in $H^*$, for some $\alpha=\Omega(1/\log^7 k)$. We do so in the next two steps, using the following claim.
\begin{claim}\label{claim: A well-linked}
The set $A$ of vertices is $\alphawl$-well-linked in $H$, where $\alphawl=\min\set{\ensuremath{\frac{1}{2}},\frac{N\cdot\alphaCMG(h)}{4\gkrv(h)}}=\Omega(1)$.
\end{claim}
\begin{proof}
Let $(Y,Z)$ be any partition of $V(H)$, $A_Y=A\cap Y$, $A_Z=A\cap Z$, and assume that $|A_Y|\leq |A_Z|$. We denote $|A_Y|$ by $\kappa$. The it is enough to show that $|E_H(Y,Z)|\geq \alphawl \kappa$. We partition the set $A_Y$ of vertices into subsets: $A''_Y$ contains all vertices $v\in A$, such that the unique path $P\in{\mathcal{H}}$ on which $v$ lies is contained in $Y$, and $A'_Y$ contains the remaining vertices. We partition $A_Z$ into $A'_Z$ and $A''_Z$ similarly. Assume first that $|A'_Y|\geq \kappa/2$. Then $|E_H(Y,Z)|\geq \kappa/2$, since for every vertex $v\in A'_Y$, the corresponding path $P\in {\mathcal{H}}$ contributes at least one edge to $E_H(Y,Z)$. Similarly, if $|A'_Z|\geq \kappa/2$, then $|E_H(Y,Z)|\geq \kappa/2$. From now on we assume that $|A'_Y|,|A'_Z|<\kappa/2$, and so $|A_Z''|\geq |A_Y''|\geq \kappa/2$.
Let ${\mathcal{Y}}\subseteq {\mathcal{H}}$ be the set of all the paths $P$, such that the first vertex of $P$ belongs to $A''_Y$. Define ${\mathcal{Z}}\subseteq {\mathcal{H}}$ similarly for $A''_Z$.
Fix some $1\leq i\leq N$, and consider the expander $X_i$.
We define two subsets of vertices of $X_i$: $Y^*$ contains all vertices $v$ that are embedded into the sub-paths of ${\mathcal{Y}}$, and $Z^*$ contains all vertices that are embedded into the sub-paths of ${\mathcal{Z}}$. Since $X_i$ is an $\alphaCMG(h)$-expander, there are $\alphaCMG(h)\cdot |Y^*|\geq \alphaCMG(h)\cdot \kappa/2$ edge-disjoint paths connecting the vertices of $Y^*$ to the vertices of $Z^*$ in $X_i$. Since the maximum vertex degree in $X_i$ is $\gkrv(h)$, there is a collection ${\mathcal{L}}_i$ of at least $\frac{\alphaCMG(h)}{\gkrv(h)}\cdot \frac{\kappa}{2}$ node-disjoint paths in $X_i$ connecting the vertices of $Y^*$ to the vertices of $Z^*$. We construct a collection ${\mathcal{L}}_i'$ of paths, connecting the vertices of $V({\mathcal{Y}})$ to the vertices of $V({\mathcal{Z}})$, such that ${\mathcal{L}}_i'\subseteq \tilde H_i$, and each edge of $\tilde H_i$ participates in at most two such paths. For each path $P\in {\mathcal{L}}_i$, we build a graph $G_P$ as follows: for each edge $e\in E(P)$, the graph includes the blue path of $\tilde H_i$ into which the edge $e$ is embedded, and, for each vertex $v\in E(P)$, the graph includes the red path $P_v\in {\mathcal{H}}_i$ into which $v$ is embedded. It is then easy to see that $G_P$ contains a path $P'\subseteq \tilde H_i$ connecting a vertex on some path $Q\in {\mathcal{Y}}$ to a vertex on some path $Q'\in {\mathcal{Z}}$. We let ${\mathcal{L}}_i'=\set{P'\mid P\in {\mathcal{L}}}$. Since each edge of $\tilde H_i$ belongs to at most one red path and at most one blue path, and the paths of ${\mathcal{L}}_i$ are node-disjoint, each edge of $\tilde H_i$ is contained in at most two paths of ${\mathcal{L}}_i'$. Let ${\mathcal{L}}=\bigcup_{i=1}^N{\mathcal{L}}_i'$. Then ${\mathcal{L}}$ contains $\frac{N\cdot\alphaCMG(h)}{\gkrv(h)}\cdot\frac{\kappa}{2}$ paths, where each path connects a vertex in $V({\mathcal{Y}})$ to a vertex in $V({\mathcal{Z}})$, and each edge of $H$ belongs to at most two such paths. Each path of ${\mathcal{L}}$ connects a vertex of $X$ to a vertex of $Y$, and so $|E_H(X,Y)|\geq
\frac{N\cdot\alphaCMG(h)}{4\gkrv(h)}\cdot \kappa$.
\end{proof}
\paragraph{Step 2: Partitioning the Red Paths}
In this step, we will define a collection $\Sigma_P$ of disjoint segments for every path $P\in{\mathcal{H}}$.
Consider any such path $P\in {\mathcal{H}}$. A sub-path $P'$ of $P$ is called a \emph{heavy sub-path} iff for some $1\leq i\leq Nr^*$, $P'$ contains at least $200N^4=\Theta(\log^4h)$ vertices that belong to $H_i$.
If $P$ contains no heavy sub-paths, then $\Sigma_P=\set{P}$. Notice that $P$ contains at most $Nr^*\cdot O(\log^4h)=O(\log^7h)$ vertices in this case.
Otherwise, we perform a number of iterations. In each iteration, we start with some heavy sub-path $P'$ of $P$, where at the beginning of the first iteration, $P'=P$.
Let $P''$ be the minimum-length heavy sub-path of $P'$ containing the first vertex of $P'$. If $P'\setminus P''$ is a heavy path, then we add $P''$ to $\Sigma_P$, delete all vertices of $P''$ from $P'$, and continue to the next iteration. Otherwise, we add $P'$ to $\Sigma_P$ and finish the algorithm. Notice that in any case, the length of every path added to $\Sigma_P$ is at most $Nr^*\cdot O(\log^4h)=O(\log^7h)$. Overall, for each path $P\in {\mathcal{H}}$, we obtain a partition of $P$ into disjoint sub-paths of length at most $O(\log^7h)$ each. Moreover, if $|\Sigma_P|>1$, then each path in $\Sigma_P$ is a heavy sub-path of $P$. Let $\Sigma=\bigcup_{P\in {\mathcal{H}}}\Sigma_P$.
We obtain a contracted graph $F$ from $H$ by contracting, for each $\sigma\in \Sigma$, the vertices of $\sigma$ into a single super-node $v_{\sigma}$. For every vertex $u\in A$, let $g(u)$ be the super-node $v_{\sigma}$ such that $u\in V(\sigma)$. Notice that for $u\neq u'$, $g(u)\neq g(u')$. Let $U=\set{g(u)\mid u\in A}$.
Since, from Claim~\ref{claim: A well-linked}, the vertices of $A$ are $\alphawl$-well-linked in $H$, the vertices of $U$ are $\alphawl$-well-linked in $F$. Since every vertex of $H$ must belong to some red path, $V(F)=\set{v_{\sigma}\mid \sigma\in \Sigma}$.
We define a graph $F^*$ from $H^*$, by similarly contracting all segments in $\bigcup_{P\in{\mathcal{H}}}\Sigma_P$ into super-nodes. Equivalently, graph $F^*$ is obtained from $F$, by deleting all edges in $E(H)\setminus E(H^*)$.
We prove the following claim.
\begin{claim}\label{claim: well-linkedness in sampled graph}
Set $U$ is $\alphawl/32$-well-linked in $F^*$ w.h.p.
\end{claim}
Assume first that the above claim is correct. We claim that $A$ is $\alpha$-well-linked in $H^*$, for $\alpha=\Omega(1/\log^7h)$. Indeed, let $(X,Y)$ be any partition of vertices of $H^*$. Let $A_X=A\cap X$, $A_Y=A\cap Y$, and $E^*=E_{H^*}(X,Y)$. Assume w.l.o.g. that $|A_X|\leq |A_Y|$. It is enough to prove that $|E^*|\geq \alpha |A_X|$. In order to prove this, we show that there is a set ${\mathcal{Q}}'$ of $|A_X|$ paths in $H^*$ connecting the vertices of $A_X$ to the vertices of $A_Y$ with edge-congestion at most $1/\alpha$.
Let $U_X=\set{g(v)\mid v\in A_X}$ and $U_Y=\set{g(v)\mid v\in Y_X}$. Since set $U$ is $\alphawl/32$-well-linked in $F^*$, there is a set ${\mathcal{Q}}$ of $|U_X|=|A_X|$ paths in $F^*$, such that each path connects a distinct vertex of $U_X$ to a distinct vertex of $U_Y$, and each edge of $F^*$ participates in at most $32/\alphawl$ such paths. We use the paths in ${\mathcal{Q}}$ in a natural way, in order to define the set ${\mathcal{Q}}'$ of paths in graph $H^*$. Let $P\in {\mathcal{Q}}$ be any such path. Assume that the endpoints of $P$ are $s$ and $t$, and let $s'\in A_X$, $t'\in A_Y'$ be such that $g(s')=s$ and $g(t')=t$. Consider the following sub-graph $H_P$ of $H^*$: start with all the edges that belong to $P$; for each vertex $v_{\sigma}$ on $P$, add the path $\sigma$ to $H_P$. It is easy to see that graph $H_P$ contains a path connecting $s'$ to $t'$. Let $P'$ be any such path. We then set ${\mathcal{Q}}'=\set{P'\mid P\in {\mathcal{Q}}}$. Since every vertex $v_{\sigma}$ of $F$ corresponds to a path $\sigma$ of length $O(\log^7h)$ in graph $H$, the degree of each such vertex $v_{\sigma}$ is $O(\log^7h)$. Since the paths in ${\mathcal{Q}}$ cause edge-congestion at most $32/\alphawl=O(1)$ in $F^*$, each vertex $v_{\sigma}$ may belong to $O(\log^7h)$ such paths. Therefore, the paths in ${\mathcal{Q}}'$ cause edge-congestion $O(\log^7h)$ in $H^*$, and $|E^*|\geq \Omega(|A_X|/\log^7h)$. We conclude that $A$ is $\alpha$-well-linked in $H^*$, for $\alpha=\Omega(1/\log^7h)=\Omega(1/\log^7k)$.
\paragraph{Step 3: Finishing the Proof}
In this step we prove Claim~\ref{claim: well-linkedness in sampled graph}.
We will sometimes refer to a subset $S\subseteq V(F)$ of the vertices of $F$, with $S,V(F)\setminus S\neq \emptyset$, as a \emph{cut}. The value of the cut $S$ is $|\operatorname{out}(S)|$.
The crucial part of the proof is the following claim.
\begin{claim}\label{claim: large degree}
The value of the minimum cut in graph $F$ is at least $N$.
\end{claim}
We prove Claim~\ref{claim: large degree} below, and first complete the proof of Claim~\ref{claim: well-linkedness in sampled graph} using it.
Let $n'=|V(H)|$. Then $|V(F)|\leq n'\leq 10 h^4\cdot r^*\cdot N$, and since $N>1536\log(10h^4\cdot r^*\cdot N)\geq 1536\log n'$, the value of the minimum cut in $F$ is at least $1536\log n'$.
The number of edges in $F$ is bounded by $m\leq 4n'\leq 40h^4r^*N=O(h^4\log^3h)$.
We use the following theorem of Karger:
\begin{theorem}[Corollary A.6 in \cite{Karger-skeleton}] \label{thm: number of almost-minimum cuts} Let $G$ be any $n$-vertex graph, and assume that the value of the minimum cut in $G$ is $C$. Then for any half-integral $\beta$, the number of cuts of value at most $\beta C$ in $G$ is bounded by $n^{2\beta}$.
\end{theorem}
Since in graph $F$, the set $U$ of vertices is $\alphawl$-well-linked, it is enough to show that w.h.p., for any subset $S$ of vertices of $F$, $|\operatorname{out}_{F^*}(S)|\geq |\operatorname{out}_F(S)|/32$. We partition the cuts $S\subseteq V(F)$ into $\ceil{\log m}$ collections ${\mathcal{C}}_1,\ldots,{\mathcal{C}}_{\ceil{\log m}}$, where for each $1\leq i\leq \ceil{\log m}$, ${\mathcal{C}}_i$ contains all cuts $S$ with $2^{i-1}N< |\operatorname{out}_F(S)|\leq 2^iN$; set ${\mathcal{C}}_1$ also contains all cuts $S$ with $|\operatorname{out}_F(S)|=N$. Consider now some such collection ${\mathcal{C}}_i$. From Theorem~\ref{thm: number of almost-minimum cuts}, $|{\mathcal{C}}_i|\leq (n')^{2^{i+1}}$. Consider some set $S\in {\mathcal{C}}_i$. Let $S'\subseteq V(H)$ be obtained by un-contracting all super-nodes in $S$, that is, $S'=\bigcup_{v_{\sigma}\in S}V(\sigma)$. Notice that $\operatorname{out}_H(S')=\operatorname{out}_F(S)$, and $\operatorname{out}_{H^*}(S')=\operatorname{out}_{F^*}(S)$.
Let $E_1(S)\subseteq \operatorname{out}_H(S')$ contain all red and red-blue edges of $\operatorname{out}_H(S')$, and let $E_2(S)=\operatorname{out}_H(S')\setminus E_1(S)$. If $|E_1(S)|\geq |\operatorname{out}_H(S')|/8$, then, since all edges of $E_1(S)$ belong to $F^*$, $|\operatorname{out}_{F^*}(S)|\geq |\operatorname{out}_F(S)|/8$. We assume from now on that this is not the case, and so $|E_2(S)|\geq 7|\operatorname{out}_F(S)|/8$. Next, we construct a maximal set $E'\subseteq E_2(S)$ of edges, such that the edges in $E'$ do not share endpoints in graph $H$. This is done by a simple greedy algorithm: while $E_2(S)\neq \emptyset$, let $e\in E_2(S)$ be any edge. Add $e$ to $E'$, and delete from $E_2(S)$ edge $e$ and all edges sharing endpoints with $e$ in graph $H$. Since all edges in $E_2(S)$ are blue, and each vertex may be incident on at most two blue edges, for every edge added to $E'$, we delete at most three edges from $E_2(S)$. Therefore, eventually $|E'|\geq |E_2(S)|/3\geq 7|\operatorname{out}_H(S')|/24\geq |\operatorname{out}_H(S')|/4=|\operatorname{out}_F(S)|/4$ holds.
\iffalse
We we partition the edges of $\operatorname{out}_H(S')$ into three subsets:
\begin{itemize}
\item $E_1(S)$ contains all red and red-blue edges of $\operatorname{out}_H(S')$. It also contains all blue edges of $\operatorname{out}_H(S')$ that are incident on the vertices $v\in S'$, such that $v$ is incident on at most one blue edge in $H$.
\item $E_2(S)$ contains all blue edges $e\in \operatorname{out}_H(S')$ incident on the vertices $v\in S'$, such that $v$ is incident on two blue edges in $H$, and the other blue edge $e'$ incident on $v$ is in $\operatorname{out}_H(S')$.
\item $E_3(S)$ contains all the remaining edges of $\operatorname{out}_H(S')$: that is, all blue edges $e\in \operatorname{out}_H(S')$ incident on vertices $v\in S'$, such that $v$ is incident on two blue edges, and the other blue edge $e'$ incident on $v$ is not in $\operatorname{out}_H(S')$.
\end{itemize}
Notice that $\operatorname{out}_{F^*}(S)$ contains all edges of $E_1(S)$ and exactly half the edges of $E_2(S)$. Therefore, if $|E_1(S)|+|E_2(S)|\geq |\operatorname{out}_F(S)|/2$, then $|\operatorname{out}_{F^*}(S)|\geq |\operatorname{out}_F(S)|/4$ as required. We assume from now on that this is not the case, and so $|E_3(S)|\geq |\operatorname{out}_F(S)|/2$.
\fi
Each edge of $E'$ belongs to $\operatorname{out}_{F^*}(S)$ independently with probability at least $1/4$. The expected number of the edges of $E'$ that belong to $\operatorname{out}_{F^*}(S)$ is therefore at least $|E'|/4\geq |\operatorname{out}_F(S)|/16\geq N\cdot 2^{i-5}$.
We use the following standard Chernoff bound: let $X_1,\ldots,X_n$ be independent random variables in $\set{0,1}$, and let $\mu=\expect{\sum_{i=1}^nX_i}$. Then $\prob{\sum_{i=1}^nX_i<\mu/2}\leq e^{-\mu/12}$. Therefore, the probability that $|\operatorname{out}_{F^*}(S)|<|\operatorname{out}_F(S)|/32$ is at most $e^{-N\cdot 2^{i-5}/12}$.
Overall, the probability that for some $S\in {\mathcal{C}}_i$, $|\operatorname{out}_{F^*}(S)|<|\operatorname{out}_F(S)|/32$ is at most:
\[(n')^{2^{i+1}}\cdot e^{-2^{i-5}N/12}<1/(n')^2\]
since $N>1536 \log n'$. Using the union bound over all $1\leq i\leq \ceil{\log m}$, with probability at least $\frac{\ceil{\log m}}{(n')^2}$, for every set $S\subseteq V$, $|\operatorname{out}_{F^*}(S)|\geq |\operatorname{out}_F(S)|/32$. In particular, set $U$ is $\alphawl/32$-well-linked in $F^*$ w.h.p. This concludes the proof of Claim~\ref{claim: well-linkedness in sampled graph}. As observed above, this implies that $A$ is $\alpha$-well-linked in graph $H^*$, thus completing the proof of Theorem~\ref{theorem: finding a minor in a super-cluster}. It now only remains to prove Claim~\ref{claim: large degree}.
\proofof{Claim~\ref{claim: large degree}}
We prove that the value of the minimum cut in $F$ is at least $N$.
Assume otherwise. Let $(X,Y)$ be a partition of $V(F)$, such that $X,Y\neq \emptyset$, and $|E_F(X,Y)|<N$.
Let $X'\subseteq V(H)$ be obtained from $X$, by un-contracting all vertices $v_{\sigma}$: that is, $X'=\bigcup_{v_{\sigma}\in X}V(\sigma)$.
We construct $Y'$ from $Y$ similarly. Observe that $(X',Y')$ is a partition of $V(H)$, and $|E_H(X',Y')|<N$.
Assume first that there are two paths $P,P'\in {\mathcal{H}}$, such that $P$ is contained in $H[X']$, and $P'$ is contained in $H[Y']$. We claim that $|E_H(X',Y')|\geq N$ in this case, leading to a contradiction. Indeed, recall that we have constructed $N$ expanders $X_1,\ldots,X_N$. For each $1\leq i\leq N$, expander $X_i$ contains some path $P_i$ connecting a pair $v,v'$ of vertices of $X_i$, where $v$ is embedded into a sub-path of $P$, and $v'$ is embedded into a sub-path of $P'$. Using the embedding of $X_i$ into $\tilde H_i$, path $P_i$ naturally defines a path $P'_i\subseteq \tilde H_i$, connecting a vertex of $P$ to a vertex of $P'$. It is immediate to see that paths $\set{P_i}_{i=1}^N$ are completely disjoint, as each such path is contained in a distinct graph $\tilde H_i$. Therefore, $H$ contains $N$ edge-disjoint paths connecting the vertices of $P$ to the vertices of $P'$. Each such path must contribute an edge to $E_H(X',Y')$, and so $|E_H(X',Y')|\geq N$, a contradiction.
Therefore, for one of the vertex sets $X',Y'$, no path $P\in {\mathcal{H}}$ is
contained in the sub-graph of $H$ induced by that set.
We assume w.l.o.g. that this set is $X'$.
Let ${\mathcal{R}}$ be the set of paths, obtained from ${\mathcal{H}}$, by deleting the edges of $E_H(X',Y')$ from them. Each path in ${\mathcal{R}}$ is contained in either $H[X']$ or $H[Y']$, and we let ${\mathcal{R}}'\subseteq {\mathcal{R}}$ be the set of paths contained in $H[X']$.
We claim that $|{\mathcal{R}}'|<N$. Indeed, since no path of ${\mathcal{H}}$ is contained in $H[X']$, every path $P\in {\mathcal{R}}'$ contributes at least one edge to $E_H(X',Y')$. Consider now any path $P'\in {\mathcal{R}}'$, and let $P\in {\mathcal{H}}$ be the path such that $P'$ is a sub-path of $P$. Let $\sigma$ be any segment of $P'$, such that $v_{\sigma}\in X$. Since no path of ${\mathcal{H}}$ is contained in $H[X']$, $\sigma$ must be a heavy segment of $P$. Therefore, there is some index $1\leq i\leq Nr^*$, such that the path $\sigma\cap H_i$ contains at least $200N^4$ vertices. We fix any such index $i^*$.
We define a new set ${\mathcal{R}}^*$ of paths as follows: for each path $P\in {\mathcal{R}}'$, we add the path $P\cap H_{i^*}$ to ${\mathcal{R}}^*$, if $P\cap H_{i^*}\neq \emptyset$. From the above discussion, $|{\mathcal{R}}^*|< N$, and $|V({\mathcal{R}}^*)|\geq 200N^4$.
Recall that ${\mathcal{R}}_{i^*}$ is the set of the red paths in $H_{i^*}$,
connecting $A_{i^*}$ to $B_{i^*}$, and ${\mathcal{B}}_{i^*}$ is the set of
the blue paths in $H_{i^*}$, connecting $A_{i^*}'$ to $A_{i^*}''$
that we have constructed during the first step of the
algorithm. Clearly, each path in ${\mathcal{R}}^*$ is a subpath of
a path in ${\mathcal{R}}_{i^*}$.
Let ${\mathcal{B}}$ be the set of paths, obtained from ${\mathcal{B}}_{i^*}$, by deleting all edges of $E_H(X',Y')$ from them.
Let ${\mathcal{B}}^*\subseteq {\mathcal{B}}$ be the set of paths contained in $H[X']$.
We claim that $|{\mathcal{B}}^*|\leq 2N$. Indeed, recall that the paths of ${\mathcal{B}}_{i^*}$ originate and terminate at the vertices of $A_{i^*}$.
Since $|{\mathcal{R}}'|< N$, at most $N$ such vertices $a\in A_{i^*}$ belong to $X'$. Therefore, at most $N$ paths of ${\mathcal{B}}_{i^*}$ may be contained in $H[X']$. Every other path of ${\mathcal{B}}^*$ must contribute one edge to $E_H(X',Y')$, and so in total $|{\mathcal{B}}^*|\leq 2N$
Observe that for every vertex $v\in V({\mathcal{R}}^*)$, if $v$ belongs to any blue path, then it belongs to a path in ${\mathcal{B}}^*$. Similarly, for $v\in V({\mathcal{B}}^*)$, if $v$ belongs to any red path, then it belongs to a path in ${\mathcal{R}}^*$.
We will view the paths in ${\mathcal{R}}_{i^*}$ as directed from $A_{i^*}$ to $B_{i^*}$, and we will view the paths in ${\mathcal{B}}_{i^*}$ as directed from $A_{i^*}'$ to $A_{i^*}''$. Let $S_1$ be the set of all vertices $v$, such that $v$ is the first vertex on some path in ${\mathcal{R}}^*$, and let $T_1$ be the set of all vertices $v$, such that $v$ is the last vertex on some path in ${\mathcal{R}}^*$. We define the sets $S_2,T_2$ of vertices for ${\mathcal{B}}^*$ similarly. Let $J=J({\mathcal{R}}^*\cup{\mathcal{B}}^*)$. Then $|S_1|=|T_1|\leq N$, and $|S_2|=|T_2|\leq 2N$, while $|V(J)|\geq 200N^4$. Since every vertex in $V(H_{i^*})\setminus (A_{i^*}\cup B_{i^*})$ has degree at least $3$ in $H_{i^*}$, $\tau(J)\geq 200N^4-|S_1\cup S_2\cup T_1\cup T_2|\geq 200N^4-6N$.
From Theorem~\ref{thm: 2-flow-main}, there are sets ${\mathcal{R}}'$ and ${\mathcal{B}}'$ of paths, routing $(S_1,T_1)$ and $(S_2,T_2)$, respectively, in $J$, such that, if $J'=J({\mathcal{R}}'\cup {\mathcal{B}}')$, then $\tau(J')<8(2N)^4+16N+1<200N^4-6N$. Since every vertex in $J\setminus (S_1\cup S_2\cup T_1\cup T_2)$ has degree more than $2$ in $J$, this means that there is some edge $e\in E(J)$, such that $(S_1,T_1)$ and $(S_2,T_2)$ are still routable in $J\setminus\set{e}$, via the sets ${\mathcal{R}}'$ and ${\mathcal{B}}'$ of paths.
We will now show that both $(A_{i^*},B_{i^*})$, and $(A_{i^*}',A_{i^*}'')$ remain routable in $H_{i^*}\setminus\set{e}$, contradicting the minimality of $H_{i^*}$. We show this for $(A_{i^*},B_{i^*})$; the proof for $(A_{i^*}',A_{i^*}'')$ is similar.
We start with a directed graph containing the original set ${\mathcal{R}}_{i^*}$ of paths routing $(A_{i^*},B_{i^*})$, where the edges on these paths are oriented from $A_{i^*}$ towards $B_{i^*}$. We then delete from this graph all edges whose both endpoints are contained in $J$.
Notice that the edge $e$ does not belong to the new graph. Also, for each vertex $v$:
\begin{itemize}
\item If $v\in A_{i^*}\setminus S_1$, then there is one edge leaving $v$ and no edges entering $v$;
\item If $v\in S_1\cap A_{i^*}$, then there is no edge entering or leaving $v$;
\item If $v\in S_1\setminus A_{i^*}$, then there is one edge entering $v$, and no edges leaving $v$;
\item if $v\in B_{i^*}\setminus T_1$, then there is one edge entering $v$ and no edges leaving $v$;
\item If $v\in T_1\cap B_{i^*}$, then there is no edge entering or leaving $v$;
\item If $v\in T_1\setminus A_{i^*}$, then there is one edge leaving $v$, and no edges entering $v$;
\item for all other vertices $v$, either there is one edge entering and one edge leaving $v$, or there is no edge incident on $v$.
\end{itemize}
Finally, we add all edges lying on the paths in ${\mathcal{R}}'$ to the
resulting graph. In this final graph, every vertex in $A_{i^*}$ has
one outgoing and no incoming edges, and every vertex in $B_{i^*}$ has
one incoming and no outgoing edges. Every other vertex either has
exactly one incoming and one outgoing edge, or it has no edges
incident on it. It is then easy to see that $(A_{i^*},B_{i^*})$ is
routable in this graph. Since this final graph is contained in
$H_{i^*}\setminus\set{e}$, this contradicts the minimality of
$H_{i^*}$.
\paragraph{Acknowledgement:} We thank Paul Seymour for posing the
question of the existence of degree-3 treewidth sparsifiers to us.
\bibliographystyle{alpha}
|
\section*{References}}
\newtheorem{thm}{Theorem}
\newtheorem{cor}[thm]{Corollary}
\newtheorem{lmm}[thm]{Lemma}
\newtheorem{prop}[thm]{Proposition}
\newproof{pf}{Proof}
\journal{European Journal of Combinatorics}
\begin{document}
\begin{frontmatter}
\title{Enumeration and classification of self-orthogonal partial Latin rectangles by using the polynomial method.}
\author{Ra\'ul M. Falc\'on}
\ead{<EMAIL>}
\cortext[cor1]{Corresponding author. Phone: + 34 954 550158; Fax: +34 954 556683}
\address{School of Building Engineering, University of Seville,\\ Avenida de Reina Mercedes 4 A, 41012, Seville, Spain.}
\begin{abstract}
The current paper deals with the enumeration and classification of the set $\mathcal{SOR}_{r,n}$ of self-orthogonal $r\times r$ partial Latin rectangles based on $n$ symbols. These combinatorial objects are identified with the independent sets of a Hamming graph and with the zeros of a radical zero-dimensional ideal of polynomials, whose reduced Gr\"obner basis and Hilbert series can be computed to determine explicitly the set $\mathcal{SOR}_{r,n}$. In particular, the cardinality of this set is shown for $r\leq 4$ and $n\leq 9$ and several formulas on the cardinality of $\mathcal{SOR}_{r,n}$ are exposed, for $r\leq 3$. The distribution of $r\times s$ partial Latin rectangles based on $n$ symbols according to their size is also obtained, for all $r,s,n\leq 4$.
\end{abstract}
\begin{keyword}
Self-orthogonal partial Latin rectangle \sep Gr\"obner basis \sep isotopism.
\MSC 05B15 \sep 20N05
\end{keyword}
\end{frontmatter}
\section{Introduction.}
An {\em $r\times s$ partial Latin rectangle} based on $[n]=\{1,\ldots,n\}$ is an $r \times s$ array in which each cell is either empty or contains a symbol of $[n]$, such that each symbol occurs at most once in each row and in each column. Its number of filled cells is its {\em size}. Let $\mathcal{R}_{r,s,n}$ and $\mathcal{R}_{r,s,n:m}$ respectively denote the set of $r\times s$ partial Latin rectangles based on $[n]$ and its subset of partial Latin rectangles of size $m$. Given $P=(p_{ij})\in \mathcal{R}_{r,s,n}$, its {\em orthogonal array representation} is the set $O(P)=\{(i,j,p_{ij})\colon\, i\in [r], j\in [s], p_{ij}\in [n]\}$. Permutations of rows, columns and symbols of $P$ give rise to new $r\times s$ partial Latin rectangles based on $[n]$, which are said to be {\em isotopic} to $P$. If $S_m$ denotes the symmetric group on $m$ elements, then $\Theta=(\alpha,\beta,\gamma)\in S_r\times S_s\times S_n$ is an {\em isotopism} of $\mathcal{R}_{r,s,n}$ and it is defined the isotopic partial Latin rectangle $P^{\Theta}$ such that $O(P^{\Theta})=\{(\alpha(i),\beta(j),\gamma(p_{i,j}))\colon\, i\in [r], j\in [s], p_{ij}\in [n]\}$. Given a permutation $\pi\in S_3$, it is defined the {\em parastrophic} partial Latin rectangle $P^{\pi}$ such that $O(P^{\pi})=\{(p_{\pi(1)},p_{\pi(2)},p_{\pi(3)})\colon\, (p_1,p_2,p_3)\in O(P)\}$. If $P^{\pi}\in \mathcal{R}_{r,s,n}$, then $\pi$ is called a {\em parastrophism} of $\mathcal{R}_{r,s,n}$. The composition of an isotopism and a parastrophism is a {\em paratopism}. Two partial Latin rectangles are in the same {\em main class} if one of them is isotopic to a parastrophic partial Latin rectangle of the other.
Two partial Latin rectangles $P=(p_{ij}), Q=(q_{ij})\in \mathcal{R}_{r,s,n}$ are {\em orthogonal} if, given $i,i'\in [r]$ and $j,j'\in [s]$ such that $p_{ij}=p_{i'j'}\in [n]$, then $q_{ij}$ and $q_{i'j'}$ are not the same symbol of $[n]$. If $r=s$, then the partial Latin rectangle $P$ is {\em self-orthogonal} if it is orthogonal to its transpose $P^t$ (see Figure \ref{fig_PLR}).
\begin{figure}[h]
{\scriptsize
$$\begin{array}{|c|c|c|c|c|}\hline
1 & 3 & \ \ & \ \ \\ \hline
2 & \ \ & 3 & 1 \\ \hline
\ \ & 1 & 2 & \ \ \\ \hline
\ \ & 2 & \ \ & 3 \\ \hline
\end{array}$$}
\caption{Example of a self-orthogonal $4\times 4$ partial Latin rectangle based on $[3]$.}
\label{fig_PLR}
\end{figure}
Let $\mathcal{SOR}_{r,n}$ be the set of self-orthogonal $r\times r$ partial Latin rectangles based on $[n]$. Only those isotopisms of the form $(\alpha,\alpha,\gamma)\in S_r\times S_r\times S_n$ and those paratopisms based on $\overline{S}_3=\{(1)(2)(3),(12)(3)\}$ preserve always the set $\mathcal{SOR}_{r,n}$. Hence, the sets $S_r\times S_n$ and $S_r\times S_n\rtimes \overline{S}_3$ determine, respectively, the {\em isotopism} and {\em paratopism groups} of $\mathcal{SOR}_{r,n}$. The enumeration of isotopism and main classes of $\mathcal{SOR}_{r,n}$ has been studied for $r=n\leq 10$ \cite{Burger2010, Burger2010a, Graham2006}. However, there does not exist a similar study for self-orthogonal partial Latin rectangles of any order. In the current paper, we deal with this problem by adapting the {\em Combinatorial Nullstellensatz} of Alon \cite{Alon1995}, whose effectiveness in the study of Latin squares has been exposed in \cite{Falcon2013,Falcon2007}.
The paper is organized as follows. In Section 2, we indicate some preliminaries concepts and results on commutative algebra. In Section 3, the set $\mathcal{R}_{r,s,n}$ is identified with that of independent sets of a Hamming graph and with the set of zeros of a zero-dimensional radical ideal, whose reduced Gr\"obner bases and Hilbert series determine, respectively, the elements and cardinality of $\mathcal{R}_{r,s,n:m}$, for all natural $m$. This cardinality is explicitly shown for $r\leq s\leq n\leq 4$ and $m\leq rsn$. In Section 4, we consider new polynomials to be added to the above ideal in order to determine the set $\mathcal{SOR}_{r,n}$. Besides, two strategies are indicated that allow us to reduce the cost of computation of the Gr\"obner basis and Hilbert series of the new ideal. They are used to determine the cardinality of $\mathcal{SOR}_{r,n}$ for $r\leq 4$ and $n\leq 9$. Some general formulas about the cardinality of $\mathcal{SOR}_{r,n}$ are finally exposed, for $r\leq 3$.
\section{Preliminaries.}
We start with some basic concepts of commutative algebra (see \cite{Cox1998, Cox2007, Kreuzer2005} for more details). Let $R=k[{\bf x}]=k[x_1,\ldots,x_n]$ be a polynomial ring in $n$ variables over a field $k$ with the {\em standard grading} induced by the degree of polynomials, that is, $R=\bigoplus_{0\leq d} R_d$, where each $R_d$ is the set of homogeneous polynomials in $R$ of degree $d$. The largest monomial of a polynomial of $R$ with respect to a given term order $<$ is its {\em initial monomial}. Given an ideal $I$ of $R$, the ideal generated by the initial monomials with respect to $<$ of all the non-zero elements of $I$ is its {\em initial ideal} $I_<$. Any monomial of $R$ which is not contained in $I_<$ is called a {\em standard monomial} of $I$ with respect to $<$. The set of standard monomials of $I$ with respect to any given term order can be used to study the dimension of the quotient ring $R/I$. This ring inherits the natural grading of $R$ and can be written as the direct sum $\bigoplus_{0\leq d} R_d/I_d$, where $I_d=R_d\cap I$. In particular, the set of standard monomials of $I$ of degree $d$ with respect to a given term order constitutes a linear $k$-basis of $R_d/I_d$ and hence, its cardinality coincides with $\mathrm{dim}_k(R_d/I_d)$, regardless of the term order which has been chosen. The {\em Hilbert function} $\mathrm{HF}_{R/I}$ of $R/I$ maps each non-negative integer $d$ onto $\mathrm{dim}_k(R_d/I_d)$. Its {\em Hilbert series} is the generating function $\mathrm{HS}_{R/I}(t)=\sum_{0\leq d}\mathrm{HF}_{R/I}(d)\cdot t^d$, which can also be written as:
\begin{equation}
\mathrm{HS}_{R/I}(t)=\frac {P(t)}{(1-t)^n}=\frac{Q(t)}{(1-t)^{\mathrm{dim}_k(I)}},
\end{equation}
where $P(t)$ and $Q(t)$ are polynomials with integer coefficients. The former is called the {\em Hilbert numerator} of $R/I$. If $I$ is zero-dimensional, then the Hilbert series of $R/I$ coincides with the polynomial $Q(t)$. The number of standard monomials of $I$ is then finite and coincides with the dimension of $R/I$ over $k$. Moreover, it is always greater than or equal to the number of points of the {\em affine variety} $V(I)$ of $I$, that is, the set of points in $k^n$ which are zeros of all the polynomials of $I$. The ideal $I$ is zero-dimensional if and only if $V(I)$ is finite. Further, it is verified that $|V(I)|=\mathrm{dim}_k(R/I)$ whenever $I$ is {\em radical}, that is, if any polynomial $p$ belongs to $I$ whenever there exists a natural $n\in \mathbb{N}$ such that $p^n\in I$.
The problem of computing a Hilbert series is NP-complete \cite{Bayer1992}. The standard algorithm which is commonly used to compute the Hilbert series of a quotient ring $R/I$ by determining the corresponding Hilbert numerator was first proposed by Mora and M\"oller in \cite{Mora1983} and is based on the additivity of the Hilbert function in short exact sequences. Even if it only works when $I$ is homogeneous, it is not an inconvenient, because the Hilbert series of $R/I$ coincides with that of $R/I_<$ for any term order $<$ and the initial ideal $I_<$ is homogeneous because all its elements are monomials. It is therefore necessary to determine explicitly the initial ideal of $I$ for a given term order $<$. In this regard, a {\em Gr\"obner basis} of $I$ with respect to $<$ is any generating set $G$ of $I$ such that the initial monomials of its elements generate the initial ideal $I_<$. It is said to be {\em reduced} if all its polynomials are monic and no monomial of a polynomial of $G$ can be generated by the initial monomials of the other polynomials of the basis. There exists only one reduced Gr\"obner basis of an ideal and the algorithm which is most commonly used to obtain it is that given by Buchberger \cite{Buchberger2006}.
\section{Boolean ideals and Hilbert series related to $\mathcal{R}_{r,s,n}$.}
In the current section, we show how the set $\mathcal{R}_{r,s,n}$ of $r\times s$ partial Latin rectangles based on $[n]$ can be identified with that of zeros of a boolean ideal which is zero-dimensional and radical. The use of non-linear polynomials to solve combinatorial problems was established by Alon \cite{Alon1995}. Afterwards, Bernasconi et al. \cite{Bernasconi1997} exposed the possibility of solving counting problems in Combinatorics by using Gr\"obner bases of {\em boolean} ideals, that is to say, ideals on $k[x_1,\ldots,x_n]$ containing the polynomials $x_i\cdot (x_i-1)$, for all $i\in [n]$. Such a boolean structure facilitates the computation of the corresponding reduced Gr\"obner basis with respect to any given term order.
\renewcommand{\baselinestretch}{1}
\begin{figure}[h]
\begin{center}
\begin{tabular}{cc}
\begin{tabular}{c}
$P\equiv\ $\begin{tabular}{|c|c|c|}\hline
1 & & 4\\ \hline
& 3 & 2\\ \hline
\end{tabular}\\ \\
$O(P)=\left\{\begin{array}{c}(1,1,1),(1,3,4),\\(2,2,3),(2,3,2)\end{array}\right\}$
\end{tabular}
&
\begin{tabular}{c}
\includegraphics[width=0.5\textwidth]{Hamming.eps}
\end{tabular}
\end{tabular}
\caption{$2\times 3$ partial Latin rectangle based on $[4]$ related to an independent set of $\mathcal{H}_{2,3,4}$.}
\label{fig0}
\end{center}
\end{figure}
\renewcommand{\baselinestretch}{2}
Let $\mathcal{H}_{r,s,n}$ be the {\em Hamming graph} \cite{Hammack2011} defined as the cartesian product $K_r\Box K_s\Box K_n$ of the three complete graphs of $r$, $s$ and $n$ vertices. It is a $(r+s+n-3)$-regular graph of order $rsn$, whose vertices can be labeled with the triples of $[r]\times [s]\times [n]$ so that two vertices are adjacent if and only if their corresponding labels differ exactly in one component. The labels of any {\em independent set} of $\mathcal{H}_{r,s,n}$, formed by $m$ pairwise nonadjacent vertices, constitute the orthogonal array representation of an $r\times s$ partial Latin rectangle based on $[n]$ of size $m$ (see Figure \ref{fig0}). The reciprocal is also verified and hence, the number of independent sets of $m$ vertices of $\mathcal{H}_{r,s,n}$ coincides with the cardinality of $\mathcal{R}_{r,s,n:m}$. When $r=s=n$, the Hamming graph $\mathcal{H}_{n,n,n}$ is usually denoted as $H(3,n)$ and its number of independent sets is only known for $n\leq 4$ (see the integer sequences A027681 and A027682 in \cite{Sloane2014}). The following result holds.
\begin{prop} \label{prop0} It is verified that:
\begin{itemize}
\item[a)] $|\mathcal{R}_{r,s,n:0}|=1$.
\item[b)] $|\mathcal{R}_{r,s,n:1}|=rsn$.
\item[c)] $|\mathcal{R}_{r,s,n:2}|=\frac 12 rsn\cdot (rsn-r-s-n+2)$.
\end{itemize}
\end{prop}
\begin{pf} The first two assertions are immediate. To prove the third one, observe that, since the number of independent sets of size two of any graph of $v$ vertices and $e$ edges is $\binom{v}{2} - e$, we have that $|\mathcal{R}_{r,s,n:2}|=\binom{rsn}{2}-\frac 12 \cdot rsn\cdot (r+s+n-3)=\frac 12 rsn\cdot (rsn-r-s-n+2)$. \hfill $\Box$
\end{pf}
\hspace{0.5cm}
Let us consider now a boolean variable $x_{ijk}$ related to each vertex of $\mathcal{H}_{r,s,n}$, for all $(i,j,k)\in [r]\times [s]\times [n]$, such that, given an independent set $S$ of $\mathcal{H}_{r,s,n}$, it takes the value $1$ if the vertex labeled as $(i,j,k)$ belongs to $S$, or $0$, otherwise. Keeping in mind the adjacency in our graph, the next result is satisfied.
\begin{thm} \label{thm0} The set $\mathcal{R}_{r,s,n}$ can be identified with the set of zeros of the ideal $I_{r,s,n}=\langle\,x_{ijk}(x_{ijk} -1),\, x_{ijk}x_{i'jk},\, x_{ijk}x_{ij'k},\, x_{ijk}x_{ijk'}\colon\, i\in [r],\, j\in [s],\, k\in [n], i'\in \{i+1,\ldots,r\}, j'\in \{j+1,\ldots,s\}, k'\in \{k+1,\ldots,n\}\,\rangle\, \subseteq \mathbb{F}_2[{\bf x}]$, where ${\bf x}=\{x_{111},\ldots,x_{rsn}\}$. Besides, $|\mathcal{R}_{r,s,n}|= \mathrm{dim}_{\mathbb{F}_2}(\mathbb{F}_2[{\bf x}]/I_{r,s,n})$ and $|\mathcal{R}_{r,s,n:m}|=\mathrm{HF}_{\mathbb{F}_2[{\bf x}]/I_{r,s,n}}(m)$, for all $m\geq 0$.
\end{thm}
{\bf Proof.} Any partial Latin rectangle $P=(p_{ij})\in\mathcal{R}_{r,s,n}$ can be uniquely identified with a zero $(x_{111},\ldots,x_{rsn})$, where $x_{ijk}=1$ if $p_{ij}=k$ and $0$, otherwise. The finiteness of $\mathcal{R}_{r,s,n}$ implies $I_{r,s,n}$ to be zero-dimensional. Further, for all $(i,j,k)\in [r]\times [s]\times [n]$, the intersection between $I_{r,s,n}$ and the polynomial ring $\mathbb{F}_2[x_{ijk}]$ of polynomials in the variable $x_{ijk}$ is the ideal $\langle\,x_{ijk}\left(x_{ijk}-1\right)\,\rangle\subseteq I_{r,s,n}$. Hence, Proposition 2.7 of \cite{Cox1998} assures $I_{r,s,n}$ to be radical and thus, Theorem 2.10 of \cite{Cox1998} implies that $|\mathcal{R}_{r,s,n}|=|V(I_{r,s,n})|= \mathrm{dim}_{\mathbb{F}_2}(\mathbb{F}_2[{\bf x}]/I_{r,s,n})$.
The reduced Gr\"obner basis of $I_{r,s,n}$ with respect to the lexicographic order $<_{\mathrm{lex}}$ coincides with its set of generators. Moreover, if we change in such a set each polynomial $x_{ijk}(x_{ijk}-1)$ by its initial monomial $x_{ijk}^2$ with respect to $<_{\mathrm{lex}}$, then the new set generates the initial ideal ${I_{r,s,n}}_{<_{\mathrm{lex}}}$. Let ${\bf x^a}=x_{111}^{a_{111}}\ldots x_{rsn}^{a_{rsn}}$ be a standard monomial of $I_{r,s,n}$ with respect to $<_{\mathrm{lex}}$. Since such a monomial does not belong to ${I_{r,s,n}}_{<_{\mathrm{lex}}}$, it is $a_{ijk}\in\{0,1\}$, for all $(i,j,k)\in [r]\times[s]\times[n]$. The standard monomial ${\bf x^a}$ can then be identified with the partial Latin rectangle $P=(p_{ij})\in\mathcal{R}_{r,s,n}$, such that $p_{ij}=k$ if $a_{ijk}=1$ and $\emptyset$, otherwise. The degree of ${\bf x^a}$ coincides with the size of $P$ and hence, given $m\geq 0$, the cardinality of $\mathcal{R}_{r,s,n:m}$ is equal to the number of standard monomials of degree $d$ of $I_{r,s,n}$. \hfill $\Box$
\vspace{0.5cm}
The initial ideal ${I_{r,s,n}}_{<_{\mathrm{lex}}}$ which appears in the proof of Theorem \ref{thm0} is the {\em modified edge ideal} \cite{Dickenstein2012} of the graph of $rsn$ vertices labeled by the variables $x_{111},\ldots,x_{rsn}$ and such that there exists an edge connecting the vertices labeled by $x_{ijk}$ and $x_{i'j'k'}$ if and only if the monomial $x_{ijk}x_{i'j'k'}$ belongs to ${I_{r,s,n}}_{<_{\mathrm{lex}}}$. Its standard monomials can then be identified with the independent sets of such a graph. It is the fundamental of a specialized algorithm exposed by Dickenstein and Tobis \cite{Dickenstein2012} to compute the Hilbert series related to this kind of ideals. We have implemented this algorithm in a procedure called {\em PLR} in the open computer algebra system for polynomial computations {\sc Singular} \cite{Decker2014}. It has been included in the library {\em pls.lib}, which is available online on {\tt http://personal.us.es/raufalgan/LS/pls.lib} and has been used to obtain the distribution in Table \ref{table_hs1} of the partial Latin rectangles of $\mathcal{R}_{r,s,n}$ with $r\leq s\leq n\leq 4$ according to their sizes. The computation of the corresponding Hilbert series is immediate ($0$ seconds) in an {\em Intel Core i7-2600, 3.4 GHz} with {\em Ubuntu}, except for $\mathcal{R}_{4,4,4}$, for which the CPU running time is $50$ seconds. Remark that all the values of Table \ref{table_hs1} have also been checked by exhaustive search.
\renewcommand{\baselinestretch}{1}
\renewcommand{\tabcolsep}{0.1pt}
\begin{table}[htb]
\begin{center}{\tiny
\hspace{-0.7cm} \begin{tabular}{rrrrrrrrrrrrrrrrrrrrrrr} \hline
\ & \multicolumn{5}{l}{$|\mathcal{R}_{r,s,n:m}|$}\\ \cline{2-21}
\ & \multicolumn{5}{l}{$r.s.n$}\\ \cline{2-21}
$m$ & 1.1.1 & 1.1.2 & 1.1.3 & 1.1.4 & 1.2.2 & 1.2.3 & 1.2.4 & 1.3.3 & 1.3.4 & 1.4.4 & 2.2.2 & 2.2.3 & 2.2.4 & 2.3.3 & 2.3.4 & 2.4.4 & 3.3.3 & 3.3.4 & 3.4.4 & 4.4.4\\ \hline
0 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1\\
1 & 1 & 2 & 3 & 4 & 4 & 6 & 8 & 9 & 12 & 16 & 8 & 12 & 16 & 18 & 24 & 32 & 27 & 36 & 48 & 64\\
2 & & & & & 2 & 6 & 12 & 18 & 36 & 72 & 16 & 42 & 80 & 108 & 204 & 384 & 270 & 504 & 936 & 1,728\\
3 & & & & & & & & 6 & 24 & 96 & 8 & 48 & 144 & 264 & 768 & 2,208 & 1,278 & 3,552 & 9,696 & 25,920\\
4 & & & & & & & & & & 24 & 2 & 18 & 84 & 270 & 1,332 & 6,504 & 3,078 & 13,716 & 58,752 & 239,760\\
5 & & & & & & & & & & & & & & 108 & 1,008 & 9,792 & 3,834 & 29,808 & 216,864 & 1,437,696\\
6 & & & & & & & & & & & & & & 12 & 264 & 7,104 & 2,412 & 36,216 & 494,064 & 5,728,896\\
7 & & & & & & & & & & & & & & & & 2,112 & 756 & 23,760 & 691,200 & 15,326,208\\
8 & & & & & & & & & & & & & & & & 216 & 108 & 7,776 & 581,688 & 27,534,816\\
9 & & & & & & & & & & & & & & & & & 12 & 1,056 & 283,584 & 32,971,008\\
10 & & & & & & & & & & & & & & & & & & & 75,744 & 25,941,504\\
11 & & & & & & & & & & & & & & & & & & & 10,368 & 13,153,536\\
12 & & & & & & & & & & & & & & & & & & & 576 & 4,215,744\\
13 & & & & & & & & & & & & & & & & & & & & 847,872\\
14 & & & & & & & & & & & & & & & & & & & & 110,592\\
15 & & & & & & & & & & & & & & & & & & & & 9,216\\
16 & & & & & & & & & & & & & & & & & & & & 576\\ \hline
Total & 2 & 3 & 4
& 5 & 7 & 13 & 21 & 34 & 73 & 209 & 35 & 121 & 325 & 781 & 3,601 & 28,353 & 11,776 & 116,425 & 2,423,521 & 127,545,137\\ \hline
\end{tabular}
}
\end{center}
\vspace{-0.2cm}
\caption{Distribution of $\mathcal{R}_{r,s,n}$ according to the size, for $r\leq s\leq n\leq 4$.}
\label{table_hs1}
\end{table}
\section{Enumeration and classification of $\mathcal{SOR}_{r,n}$.}
The ideal $I_{r,r,n}$ of Theorem \ref{thm0} can be slightly modified to impose partial Latin rectangles to be self-orthogonal. In this regard, the proof of the next result is analogous to that of the mentioned theorem.
\begin{thm} \label{thm1} The set $\mathcal{SOR}_{r,n}$ can be identified with the set of zeros of the following zero-dimensional ideal of $\mathbb{F}_2[x_{111},\ldots,x_{rrn}]$.
$$I_{r,n}=I_{r,r,n}\cup \langle\,x_{ijp}\cdot x_{klp}\cdot x_{jiq}\cdot x_{lkq}\colon\, i,j,k,l\in [r],\, p,q\in [n],\,(i,j)\neq(k,l)\,\rangle.$$
Moreover, $|\mathcal{SOR}_{r,n}|=\mathrm{dim}_{\mathbb{F}_2} (\mathbb{F}_2[{\bf x}]/I_{r,n})$ and $|\mathcal{SOR}_{r,n:m}|=\mathrm{HF}_{\mathbb{F}_2[{\bf x}]/I_{r,n}}(m)$.
\hfill $\Box$
\end{thm}
\vspace{0.5cm}
\renewcommand{\baselinestretch}{1}
\renewcommand{\tabcolsep}{0.5pt}
\begin{table}[htb]
\begin{center}{\tiny
\begin{tabular}{rrrrrrrrrrrrrrrrrrrrrr} \hline
\ & \multicolumn{5}{l}{$|\mathcal{SOR}_{r,n:m}|$}\\ \cline{2-19}
\ & \multicolumn{5}{l}{$r.n$}\\ \cline{2-19}
$m$ & 2.1 & 2.2 & 2.3 & 2.4 & 2.5 & 2.6 & 2.7 & 2.8 & 2.9 & 3.1 & 3.2 & 3.3 & 3.4 & 3.5 & 3.6 & 3.7 & 3.8 & 3.9 \\ \hline
0 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1\\
1 & 4 & 8 & 12 & 16 & 20 & 24 & 28 & 32 & 36 & 9 & 18 & 27 & 36 & 45 & 54 & 63 & 72 & 81\\
2 & & 12 & 36 & 72 & 120 & 180 & 252 & 336 & 432 & 12 & 96 & 252 & 480 & 780 & 1,152 & 1,596 & 2,112 & 2,700\\
3 & & & 24 & 96 & 240 & 480 & 840 & 1,344 & 2,016 & 2 & 172 & 1,014 & 3,032 & 6,730 & 12,612 & 21,182 & 32,944 & 48,402 \\
4 & & & & 24 & 120 & 360 & 840 & 1,680 & 3,024 & & 108 & 1,836 & 9,720 & 31,320 & 77,220 & 161,028 & 299,376 & 511,920 \\
5 & & & & & & & & & & & 12 & 1,476 & 15,912 & 80,040 & 270,900 & 720,972 & 1,633,296 & 3,296,592 \\
6 & & & & & & & & & & & & 444 & 12,816 & 110,040 & 537,360 & 1,883,700 & 5,313,504 & 12,859,056 \\
7 & & & & & & & & & & & & 36 & 4,608 & 76,680 & 573,120 & 2,743,020 & 9,870,336 & 29,142,288 \\
8 & & & & & & & & & & & & & 720 & 24,480 & 295,920 & 2,005,920 & 9,444,960 & 34,655,040 \\
9 & & & & & & & & & & & & & 48 & 3,120 & 58,320 & 566,160 & 3,551,520 & 16,456,608 \\
\hline
Total & 5 & 21 & 73 & 209 & 501 & 1,045 & 1,961 & 3,393 & 5,509 & 24 & 407 & 5,086 & 47,373 & 333,236 & 1,826,659 & 8,103,642 & 30,148,121 & 96,972,688 \\ \hline
\end{tabular}}
\end{center}
\caption{Distribution of $\mathcal{SOR}_{r,n}$ according to the size, for $r\leq 3$ and $n\leq 9$.}
\label{table0}
\end{table}
In order to compute the reduced Gr\"obner basis and the Hilbert series of the ideal of Theorem \ref{thm1}, we have implemented the procedure {\em SOR} in the library {\em pls.lib}. It has been used to obtain the distribution in Table \ref{table0} of the partial Latin rectangles of $\mathcal{SOR}_{r,n}$ according to their size, for $r\leq 3$ and $n\leq 9$. Although the CPU running time is similar to that of the analogous cases of Table \ref{table_hs1}, our computer system runs out of memory to determine the case $r=4$. It is due to the fact that the computation of reduced Gr\"obner bases and Hilbert series is extremely sensitive to the number of variables and polynomials. We propose two possible strategies to reduce this cost of computation.
\subsection{First strategy. Direct sum.}
Given $P\in\mathcal{R}_{r,r,n}$ and $Q\in\mathcal{R}_{r',r',n}$, let $P\oplus Q\in\mathcal{R}_{r+r',r+r',n}$ be the partial Latin rectangle having $P$ as upper left corner, $Q$ as lower right corner and empty the rest of its cells. Let us consider the set $S_{r,r',n}=\{(P,Q)
\in\mathcal{SOR}_{r,n}\times\mathcal{SOR}_{r',n}\colon\, P\oplus Q\in\mathcal{SOR}_{r+r',n}\}$ and the ideal $I_{r,r',n}^{P,Q}=I_{r,r',n} + \sum_{(i,j,k)\in [r]\times [r']\times [n] }\langle\, x_{ijk}\mid\, \exists\ l\in [r] \text{ such that } p_{il}=k \text{ or } \exists\ l\in [r'] \text{ such that } q_{lj}=k\rangle$. The affine variety $V(I_{r,r',n}^{P,Q})$ coincides with that of partial Latin rectangles of $\mathcal{R}_{r,r',n}$ which can be included in the upper right corner of $P\oplus Q$ to obtain an element of $\mathcal{R}_{r+r',r+r',n}$. Now, given $A=(a_{ij})\in V(I_{r,r',n}^{P,Q})$, let $I_{r',r,n}^{P,Q,A}$ be the following subideal of $I_{r',r,n}$.
\begin{center}
$I_{r',r,n} +
\langle\, x_{ijk}x_{i'j'k}\colon i,i'\in [r'], j,j'\in [r],(i,j)\neq (i',j'),k\in [n], a_{ji}=a_{j'i'}\in [n]\rangle\, + \langle\, x_{ijk}x_{i'j'k'}\colon i,i'\in [r'], j,j'\in [r], k,k'\in [n], a_{ji}=k'\neq k=a_{j'i'}\rangle\, + \langle\, x_{ijk}\colon (i,j,k)\in [r']\times [r]\times [n],\, a_{ji}=k, \text{ or }\, \exists\ l,m\in[r] \text{ such that }a_{ji}=p_{lm},\, p_{ml}=k, \text{ or }\, \exists\ l,m\in[s] \text{ such that } a_{ji}=q_{lm},\, q_{ml}=k, \text{ or }\, \exists\ l\in [r] \text{ such that } p_{lj}=k, \text{ or }\, \exists\ l\in [r'] \text{ such that } q_{il}=k\rangle\,$
\end{center}
It is verified that
\begin{equation}\label{eqHS}
|\mathcal{SOR}_{r,n:m}|=\sum_{\small \substack{(P,Q)\in S_{r,r',n}\\ A\in V(I_{r,r',n}^{P,Q})}}
t^{|P|+|Q|+|A|}\cdot
\mathrm{HS}_{\mathbb{F}_2/I^{P,Q,R}_{r',r,n}}(t).
\end{equation}
This expression has been used to check the data of Table \ref{table0} and to expose in Table \ref{table0a} the cardinality of $\mathcal{SOR}_{4,n}$, for all $n\leq 9$.
\renewcommand{\baselinestretch}{1}
\renewcommand{\tabcolsep}{0.3pt}
\begin{table}[htb]
\begin{center}{\tiny
\begin{tabular}{rrrrrrrrrr}
\ & \multicolumn{5}{l}{$r.n$}\\ \cline{2-10}
$m$ & 4.1 & 4.2 & 4.3 & 4.4 & 4.5 & 4.6 & 4.7 & 4.8 & 4.9 \\ \hline
0 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\
1 & 16 & 32 & 48 & 64 & 80 & 96 & 112 & 128 & 144\\
2 & 60 & 360 & 900 & 1,680 & 2,700 & 3,960 & 5,460 & 7,200 & 9,180\\
3 & 56 & 1,792 & 8,568 & 23,744 & 50,680 & 92,736& 153,272 & 235,648 & 343,224\\
4 & 14 & 4,196 & 45,306 & 199,784 & 587,750 & 1,373,004 & 2,763,026 & 5,008,976 & 8,405,694\\
5 & & 4,560 & 137,520 & 1,046,880 & 4,428,960 & 13,552,560 & 33,783,120 & 73,106,880 & 142,655,040\\
6 & & 2,256 & 240,216 & 3,479,616 & 22,225,680 & 91,696,080 & 288,559,656 & 755,440,896 & 1,731,190,176\\
7 & & 480 & 237,888 & 7,350,912 & 74,983,680 & 430,875,360 & 1,748,093,760 & 5,618,070,528 & 15,283,095,552\\
8 & & 24 & 131,544 & 9,785,664 & 169,923,120 & 1,410,554,520 & 7,551,498,024 & 30,273,440,064 & 98,905,243,104\\
9 & & & 40,896 & 8,103,552 & 256,494,720 & 3,202,600,320& 23,211,048,000 & 118,117,015,296 & 469,324,461,312\\
10 & & & 7,056 & 4,147,584 & 254,539,680 & 4,988,125,440 & 50,312,927,280 & 331,193,485,056 & 1,622,312,241,984\\
11 & & & 576 & 1,332,864 & 163,762,560 & 5,241,536,640 & 75,710,577,600 & 657,677,857,536 & 4,029,212,001,024\\
12 & & & 48 & 283,200 & 67,632,480 & 3,633,984,960& 77,231,577,360 & 903,490,374,528 & 7,027,446,121,920\\
13 & & & & 43,008 & 17,850,240 & 1,613,064,960& 51,545,020,800 & 827,927,331,840 & 8,299,928,625,408\\
14 & & & & 5,760 & 2,975,040 & 437,253,120 & 21,258,498,240 & 476,757,469,440 & 6,249,614,071,680\\
15 & & & & 768 & 291,840 & 65,571,840 & 4,861,006,080 & 154,221,473,280 & 2,678,459,470,848\\
16 & & & & 48 & 14,160 & 4,127,760& 466,312,560 & 21,145,881,120 & 492,310,895,328 \\
\hline
Total & 147 & 13,701 & 850,567 & 35,805,129 & 1,035,763,371 & 21,134,413,357 & 314,221,824,351 & 3,527,256,198,417
& 30,984,678,831,619\\ \hline
\end{tabular}}
\end{center}
\caption{Distribution of $\mathcal{SOR}_{4,n}$ according to the size, for $n\leq 9$.}
\label{table0a}
\end{table}
\subsection{Second strategy. Number of distinct symbols.}
Given $s\leq n$, let $\mathcal{SOR}_{r,n;s}$ be the subset of partial Latin rectangles of $\mathcal{SOR}_{r,n}$ which contain exactly $s$ distinct symbols in their cells. Observe that any partial Latin rectangle of $\mathcal{SOR}_{r,n;s}$ is equal, up to permutation of symbols, to a partial Latin rectangle of $\mathcal{SOR}_{r,s;s}$. If $s=0$, then the set $\mathcal{SOR}_{r,0;0}$ only contains the empty $r\times r$ Latin rectangle. We have that
\begin{equation}\label{eqSO}
\mathcal{SOR}_{r,n}=\bigcup_{s=0}^n\mathcal{SOR}_{r,n;s}
\end{equation}
Hence, if $\sigma_{r,s}$ denotes the cardinality of $\mathcal{SOR}_{r,s;s}$, then
\begin{equation}\label{eq1}
|\mathcal{SOR}_{r,n}|=\sum_{s=0}^n\binom{n}{s}\cdot \sigma_{r,s},
\end{equation}
The following result holds.
\begin{lmm}\label{lmm2} It is verified that:
\begin{enumerate}[i.]
\item $\sigma_{r,0}=1$.
\item $\sigma_{r,r^2-1}=\frac 12\cdot (r^2+1)! \cdot \frac {r^3-2r^2+r+2}r$.
\item $\sigma_{r,r^2}=r^2!$.
\item $\sigma_{r,s}=0$, for all $s>r^2$.
\end{enumerate}
\end{lmm}
\begin{pf} Claims (i), (iii) and (iv) follows straightforward from the definition of $\sigma_{r,s}$. To prove the second expression, observe that any $r\times r$ partial Latin rectangle which contains exactly $r^2-1$ distinct symbols has at most one empty cell. There are $r^2!$ possible partial Latin rectangles with exactly one empty cell. Otherwise, there exist two cells with the same symbol, where at most one of them is in the main diagonal. Specifically, there are $(r^2-1)!\cdot r\cdot (r-1)\cdot (r-2)$ possible partial Latin rectangles with one such a cell in the main diagonal and $\frac 12 \cdot (r^2-1)!\cdot (r^2-r)\cdot (r-1)\cdot (r-2)$ ones without any of them in the main diagonal. The result follows from the addition of all these possibilities.\hfill $\Box$
\end{pf}
\vspace{0.5cm}
The enumeration of $\mathcal{SOR}_{r,s;s}$ can be based on that of $\mathcal{SOR}_{r,s-1;s-1}$, for all $s\in [n]$. To see it, given a partial Latin rectangle $P=(p_{ij})\in\mathcal{SOR}_{r,s-1;s-1}$, let us define the subset $\mathcal{SOR}^P_{r,s;s}$ of partial Latin rectangles $Q=(q_{ij})\in \mathcal{SOR}_{r,s;s}$ such that $q_{ij}=p_{ij}$ if $p_{ij}\in [s-1]$ and $q_{ij}=\emptyset$, otherwise. It is then verified that
\begin{equation}\label{eq00}
\mathcal{SOR}_{r,s;s}=\bigcup_{P\in\mathcal{SOR}_{r,s-1;s-1}} \mathcal{SOR}^P_{r,s;s}.
\end{equation}
The partial Latin rectangles of $\mathcal{SOR}^P_{r,s;s}\cup \{P\}$ can be identified with the zeros of the ideal $I_{r,s;s}^P$ based on the $r^2-|P|$ variables which result after substituting the following variables in the ideal $I_{r,s}$ of Theorem \ref{thm1}.
\begin{enumerate}
\item[a)] Each variable $x_{ijk}$ such that $i,j\in [r]$ and $k\in [s-1]$ is substituted by $1$ if $p_{ij}=k$, or by $0$, otherwise.
\item[b)] Each variable $x_{ijs}$ such that $i,j\in [r]$ and $p_{ij}\in [s-1]$ is substituted by $0$.
\end{enumerate}
The reduced Gr\"obner basis of this new ideal can then be computed to determine explicitly the set $\mathcal{SOR}^P_{r,s;s}$. If the same reasoning is done for each partial Latin rectangle of $\mathcal{SOR}_{r,s-1;s-1}$ and for each $s\leq n$, then we can enumerate the set $\mathcal{SOR}_{r,n}$. The reduction on the number of variables of each ideal $I_{r,s;s}^P$, in comparison with the $r^2\cdot n$ variables of the ideal $I_{r,n}$, implies a significant improvement of the initial cost of computation which was necessary to enumerate the set $\mathcal{SOR}_{r,n}$ by applying Theorem \ref{thm1}. Even if such an improvement is obtained at the expense of time of computation (observe that it would be necessary to compute $\sum_{s=0}^{n-1}\sigma_{r,s}$ distinct reduced Gr\"obner bases), this time can be reduced if we make use of the distribution of $\mathcal{SOR}_{r,n}$ into main classes. In this regard, given $s\leq n$, let $\mathfrak{P}_{r,s;s}$ be the set of main classes of $\mathcal{SOR}_{r,s;s}$ and let $\mathfrak{P}_{s,P}$ denote the main class of $P\in\mathcal{SOR}_{r,s}$, where the necessity of the subscript $s$ for the number of symbols is due to the fact that $P$ is also a self-orthogonal partial Latin rectangle of $\mathcal{SOR}_{r,t}$, for all $t>s$. Since $S_r\times S_s\rtimes \overline{S}_3$ is a finite group which acts on $\mathcal{SOR}_{r,s}$, Burnside's lemma implies that
\begin{equation}\label{eq0}
\sigma_{r,s}=\sum_{P\in\mathfrak{P}_{r,s;s}}|\mathfrak{P}_{s,P}|= \sum_{P\in\mathfrak{P}_{r,s;s}} \frac {2\cdot r!\cdot s!} {|\mathfrak{I}_s(P,P)|+|\mathfrak{I}_s(P,P^t)|},
\end{equation}
where given two partial Latin rectangles $P,Q\in \mathcal{SOR}_{r,s}$, the set $\mathfrak{I}_s(P,Q)$ denotes the set of isotopisms which transform $P$ into $Q$. The next result shows how the polynomial method can be used to determine this set and its cardinality.
\begin{thm} \label{thm3} Given $P=(p_{ij}), Q=(q_{ij})\in\mathcal{SOR}_{r,s}$, the set $\mathfrak{I}_s(P,Q)$ can be identified with the set of zeros of the zero-dimensional ideal of $\mathbb{F}_2[x_{11},\ldots,x_{rr},$ $y_{11},\ldots,y_{ss}]$.
$$I_{\mathfrak{I}_s(P,Q)}=\langle\,1-\sum_{j\in [r]}x_{ij}\colon\,i\in [r]\,\rangle + \langle\,1-\sum_{j\in [n]}y_{ij}\colon\,i\in [s]\,\rangle + \langle\,1-\sum_{i\in [r]}x_{ij}\colon\,j\in [r]\,\rangle +$$ $$\langle\,1-\sum_{i\in [n]}y_{ij}\colon\,j\in [s]\,\rangle +\langle\,x_{ij}\cdot\left(1-x_{ij}\right)\colon\, i,j \in [r]\,\rangle +\langle\,y_{ij}\cdot\left(1-y_{ij}\right)\colon\, i,j \in [s]\,\rangle +$$ $$\langle\,x_{ik}\cdot x_{jl} \cdot (y_{p_{ij}q_{kl}}-1)\colon\, i,j,k,l\in [r], \text{ such that } p_{ij}, q_{kl}\in [s]\,\rangle +$$ $$\langle\,x_{ik}\cdot x_{jl} \colon\, i,j,k,l\in [r],\text{ such that } (p_{ij}=\emptyset \text{ and } q_{kl}\in [s]) \text{ or }$$ $$ \text{ or }(p_{ij}\in [s] \text{ and } q_{kl}=\emptyset)\,\rangle.$$
Moreover, $|\mathfrak{I}_s(P,Q)|= \mathrm{dim}_{\mathbb{F}_2}(\mathbb{F}_2[x_{11},\ldots,x_{rr},y_{11},\ldots,y_{ss}]/I_{\mathfrak{I}_s(P,Q)})$.
\end{thm}
\begin{pf} Any isotopism $\Theta=(\alpha,\gamma)\in S_r\times S_s$ of $\mathcal{SOR}_{r,s}$ can be univocally identified with a zero $(x^{\Theta}_{11},\ldots,x^{\Theta}_{rr},y^{\Theta}_{11},\ldots, y^{\Theta}_{ss})$, where $x^{\Theta}_{ij}=1$, (respectively, $y^{\Theta}_{ij}=1$) if $\alpha(i)=j$ (respectively, $\gamma(i)=j$) and $0$, otherwise. The first six subideals of $I_{\mathfrak{I}_s(P,Q)}$ imply that $\Theta$ belongs to $S_r\times S_s$ and the following two subideals imply that $\Theta$ transforms $P$ into $Q$. The rest of the proof is analogous to that of Theorem \ref{thm0}. \hfill $\Box$
\end{pf}
\vspace{0.5cm}
It just remains to determine a representative partial Latin rectangle of each main class of $\mathcal{SOR}_{r,s;s}$. The following result shows how the enumeration of $\mathfrak{P}_{r,s-1;s-1}$ can determine that of $\mathfrak{P}_{r,s;s}$.
\begin{lmm}\label{lmm3} Let $\{P_1,\ldots,P_m\}$ be a set of representative partial Latin rectangles of $\mathfrak{P}_{r,s-1;s-1}$. Given $P\in \mathcal{SOR}_{r,s;s}$, there exist a natural $i\leq m$ and a partial Latin rectangle $Q\in \mathcal{SOR}_{r,s;s}^{P_i}$ such that $P$ is in the same main class than $Q$.
\end{lmm}
\begin{pf} Given $P\in \mathcal{SOR}_{r,s;s}$, let $P'\in \mathcal{SOR}_{r,s-1;s-1}$ be the self-orthogonal partial Latin rectangle which results after removing the symbol $s$ from the cells of $P$. There exist a natural $i\leq m$ and a paratopism $\Theta=(\alpha,\gamma,\pi)\in S_r\times S_{s-1}\rtimes \overline{S}_3$ such that $P_i=P'^{\Theta}$. It is enough to consider $Q=P^{\Theta'}\in\mathcal{SOR}_{r,s;s}^{P_i}$, where $\Theta'=(\alpha,\gamma',\pi)\in S_r\times S_s\rtimes \overline{S_3}$ is defined such that $\gamma'(i)=\gamma(i)$ if $i<s$ and $\gamma'(s)=s$.\hfill $\Box$
\end{pf}
\vspace{0.5cm}
Once we have determined a set $S\subseteq \mathcal{SOR}_{r,s-1;s-1}$ of representative partial Latin rectangles of $\mathfrak{P}_{r,s-1;s-1}$, it is enough to enumerate the set $\bigcup_{P\in S}\mathcal{SOR}^P_{r,s;s}$ and to distribute its elements according to their main classes. Theorem \ref{thm3} can be used to determine this last distribution, because two partial Latin rectangles $P,Q\in\mathcal{SOR}_{r,s}$ are in the same main class if and only if $|\mathfrak{I}_s(P,Q)|+|\mathfrak{I}_s(P,Q^t)|>0$. Previously, in order to enumerate each set $\mathcal{SOR}^P_{r,s;s}$, we can make use of the ideal $I_{r,s;s}^P$ that was defined at the beginning of the current subsection. All these considerations have been implemented in the procedure {\em SOR} mentioned after Theorem \ref{thm1}. It has also been implemented a procedure called {\em ortisot} that computes the reduced Gr\"obner basis related to the ideal of Theorem \ref{thm3}. With the joint use of both procedures we have determined the main classes of $\mathcal{SOR}_{r,s;s}$, for all $r\leq 3$ and $s\leq r^2$. The number of these classes are shown in Table \ref{table1} and have been used to obtain from Expression (\ref{eq0}) the corresponding values of $\sigma_{r,s}$. The use of these values in Expression (\ref{eq1}) allow us to prove the next result.
\begin{thm} \label{thmGF} It is verified that:
\begin{enumerate}
\item $|\mathcal{SOR}_{1,n}|=n+1$.
\item $|\mathcal{SOR}_{2,n}|=n^4-2n^3+5n^2+1$.
\item $|\mathcal{SOR}_{3,n}|= n^9-15n^8+122n^7-604n^6+1973n^5-4201n^4+5640n^3-4240n^2+1347n+1$. \hfill $\Box$
\end{enumerate}
\end{thm}
\vspace{0.25cm}
\renewcommand{\baselinestretch}{1}
\renewcommand{\tabcolsep}{1pt}
\begin{table}[htb]
\begin{center}{\tiny
\begin{tabular}{l|rrr|rrr}
\ & \multicolumn{3}{l|}{$|\mathfrak{P}_{r,s;s}|$} & \multicolumn{3}{l}{$\sigma_{r,s}$}\\ \cline{2-7}
\ & \multicolumn{6}{l}{$r$} \\ \cline{2-7}
$s$ & 1 & 2 & 3 & 1 & 2 & 3\\ \hline
1 & 1 & 2 & 5 & 1! & 4 & 23\\
2 & & 3 & 24 & & 12 & 360\\
3 & & 2 & 71 & & 24 & 3,936\\
4 & & 1 & 128 & & 4! & 29,376\\
5 & & & 122 & & & 143,280\\
6 & & & 67 & & & 442,080\\
7 & & & 22 & & & 826,560\\
8 & & & 4 & & & 846,720\\
9 & & & 1 & & & 9!\\
\end{tabular}}
\end{center}
\caption{Main classes of $\mathcal{SOR}_{r,s;s}$.}
\label{table1}
\end{table}
\section{Conclusions.}
Keeping in mind the results of the current paper, we can conclude that the Combinatorial Nullstellensatz of Alon is a good method to deal with the enumeration and classification of self-orthogonal $r\times r$ partial Latin rectangles based on $n$ symbols. Based on the adjacency and independent sets of a Hamming graph, we have identified these combinatorial objects with the zeros of a zero-dimensional radical ideal. Besides, we have exposed two distinct strategies to reduce the cost of computation of the reduced Gr\"obner basis and Hilbert series of such an ideal. They can be used to enumerate explicitly the set of self-orthogonal partial Latin rectangles or to obtain some general formulas about its cardinality. All our results have been implemented into distinct procedures that have been applied in particular to determine the number of $r\times s$ partial Latin rectangles based on $n$ symbols, for $r,s,n\leq 4$ and that of self-orthogonal partial Latin rectangles of order $r\leq 4$.
|
\section{Introduction}
\label{sec:introduction}
The investigation of the $\text{AdS}/\text{CFT}$ correspondence~\cite{Maldacena:1997re,Witten:1998qj,Gubser:1998bc} using integrability techniques has led to a remarkably successful quantitative description of the 't~Hooft, or planar, limit~\cite{'tHooft:1973jz} of certain classes of dual theories. The two best known examples are type IIB strings on~$\text{AdS}_5\times \mathrm{S}^5$ and the dual $\mathcal{N}=4$ Supersymmetric Yang-Mills (SYM) theory, and type IIA string theory on $\text{AdS}_4\times \mathrm{CP}^3$ and the dual ABJM Chern-Simons theory~\cite{Aharony:2008ug}.\footnote{See~\cite{Arutyunov:2009ga,Beisert:2010jr,Klose:2010ki} for reviews and a more complete list of references.} The integrability methods used to understand these dual pairs can also be extended to their deformations, orbifolds and orientifolds~\cite{Zoubos:2010kh,vanTongeren:2013gva}, suggesting that other classes of examples of the $\text{AdS}/\text{CFT}$ correspondence may also be amenable to this approach.
Superstrings on~$\text{AdS}_3\times \mathcal{M}_7$ backgrounds with 16 real supersymmetries~\cite{Pesando:1998wm, Rahmfeld:1998zn,Park:1998un, Metsaev:2000mv,Babichenko:2009dk} have been shown to be classically integrable~\cite{Babichenko:2009dk}, opening the possibility of understanding such $\text{AdS}_3/\text{CFT}_2$ dualities with integrability methods. However, unlike the more supersymmetric cases mentioned above, such backgrounds contain massive and massless worldsheet excitations. While the massive excitations could be investigated \cite{David:2008yk, Babichenko:2009dk,David:2010yg, OhlssonSax:2011ms, Borsato:2012ud,Borsato:2012ss,Abbott:2012dd,Beccaria:2012kb,Beccaria:2012pm,Borsato:2013qpa,Borsato:2013hoa,Abbott:2013ixa} using more conventional integrability methods developed in the context of $\text{AdS}_5/\text{CFT}_4$ and $\text{AdS}_4/\text{CFT}_3$,%
\footnote{%
See also \cite{Sfondrini:2014via} for a review and a more extensive list of references.
} massless modes appeared initially to be rather different and difficult to incorporate fully into the holographic integrability approach. For example, these massless excitations made it difficult to apply directly some of the integrability methods, such as the finite-gap techniques, to the non-perturbative theory~\cite{Babichenko:2009dk,OhlssonSax:2011ms}. Initially, progress was made by considering the massless modes at the weakly-coupled spin-chain point~\cite{Sax:2012jv} and in the finite-gap equations of classical string theory~\cite{Lloyd:2013wza}.
Recently, through the analysis of the off-shell symmetry algebra of the theory and its representations, a complete non-perturbative worldsheet S matrix of type IIB superstring theory~$\text{AdS}_3\times \mathrm{S}^3\times \mathrm{T}^4$ supported by R-R flux was constructed~\cite{Borsato:2014exa,Borsato:2014hja}. This provided a unified description of massive and massless worldsheet excitations in an integrable framework, where all worldsheet excitations are \emph{non-relativistic} and so massless scattering can take place. This allows one to circumvent the more abstract constructions of massless \emph{relativistic} S matrices found in the integrability literature~\cite{Zamolodchikov:1992zr, Fendley:1993wq,Fendley:1993xa}.
Unlike the higher-dimensional $\text{AdS}$ backgrounds, Type IIB string theory on $\text{AdS}_3\times \mathrm{S}^3\times \mathrm{T}^4$ has a large moduli space of parameters and can be supported by a mixture of Neveu-Schwarz-Neveu-Schwarz (NS-NS) and Ramond-Ramond (R-R) three-form fluxes.
The relations between these backgrounds are governed by U-duality transformations and were analysed extensively in ref.~\cite{Larsen:1999uk}. In particular, type IIB S-duality relates $\text{AdS}_3\times \mathrm{S}^3\times \mathrm{T}^4$ backgrounds supported by different three-form fluxes: the pure R-R flux background can be obtained from the near-horizon limit of D1- and D5-branes, while backgrounds supported by mixed three-form fluxes involve the near-horizon limit of NS5-branes and fundamental strings in addition to the D1- and D5-branes.
In the bosonic non-linear sigma model, turning on the NS-NS three form flux yields a Wess-Zumino-Witten (WZW) term in the action~\cite{Novikov:1982ei,Witten:1983tw, Witten:1983ar}.
In units where the $\text{AdS}$ radius is one, the R-R three form $F$ and the NS-NS three form $H$ are given by
\begin{equation}
F = \tilde{q}\, \bigl( \operatorname{Vol}(\text{AdS}_3) + \operatorname{Vol}(\mathrm{S}^3) \bigr) , \qquad
H = q\, \bigl( \operatorname{Vol}(\text{AdS}_3) + \operatorname{Vol}(\mathrm{S}^3) \bigr) ,
\end{equation}
where the coefficients $q$ and $\tilde{q}$ satisfy
\begin{equation}
q^2 + \tilde{q}^2 = 1.
\end{equation}
The parameter $q$ is related to the quantised coupling $k$ of the WZW model
\begin{equation}
k = q \sqrt{\lambda}\,.
\end{equation}
where~$\lambda$ is the 't~Hooft coupling, which in turn parameterises the string tension $\sqrt{\lambda}/2\pi$.
Note that since we have%
\footnote{The theory is well defined and supersymmetric when $|q|\leq 1$. For simplicity, we restrict to positive~$q$. A parity transformation on the worldsheet amounts to~$q\to-q$, and can be used to consider $-1\leqq<0$.}
$0 \leq q \leq 1$, the coupling $\sqrt{\lambda}$ satisfies $\sqrt{\lambda} \ge k$.
If the supersymmetric completions of such $\text{AdS}_3$ backgrounds are tractable by integrability, we will have the exciting possibility of studying families of integrable models with deformation parameters related to (some of) the string moduli. Indeed, in ref.~\cite{Cagnazzo:2012se} the classical superstring action for the mixed-flux $\text{AdS}_3\times \mathrm{S}^3\times \mathrm{T}^4$ backgrounds was shown to be integrable. This led to rapid progress in understanding the role integrability plays in the massive sector of the theory~\cite{Hoare:2013pma,Hoare:2013ida,Hoare:2013lja,Ahn:2014tua,Babichenko:2014yaa}.
In this paper we derive the complete non-perturbative asymptotic worldsheet S~matrix, including the massless modes, for all mixed-flux $\text{AdS}_3\times \mathrm{S}^3\times \mathrm{T}^4$ backgrounds. We do this by computing the off-shell symmetry algebra of the theory and using it to determine the two-body S~matrix, which satisfies the Yang-Baxter equation. This allows us to treat the massive and massless modes on the same footing and shows that the approach used in refs.~\cite{Borsato:2014exa,Borsato:2014hja} to tackle massless modes is likely to be applicable to more general $\text{AdS}/\text{CFT}$ integrability settings where such modes frequently appear~\cite{Babichenko:2009dk,Sorokin:2011rr,Wulff:2014kja}. When restricted to the massive sector, our S~matrix reduces to the one presented in ref.~\cite{Hoare:2013ida,Hoare:2013lja}.
This paper is structured as follows. In section~\ref{sec:off-shell-algebra} we derive the off-shell symmetry algebra of the theory from the type IIB superstring action for $\text{AdS}_3\times\mathrm{S}^3\times\mathrm{T}^4$ with mixed flux in light-cone gauge, including the exact form of the off-shell central charges. In section~\ref{sec:algebra} we present the representations of the symmetry algebra~$\mathcal{A}$ that enter the S~matrix construction and then deform these representations in a way that produces the shortening condition for these exact charges. In section~\ref{sec:smat} we use these representations to construct an invariant S~matrix for the all worldsheet excitations of the mixed-flux theories, up to a number of dressing factors which we constrain by crossing symmetry. We conclude in section~\ref{sec:conclusion}. We relegate some more technical details to the appendices.
\section{Superstrings on \texorpdfstring{$\text{AdS}_3\times \mathrm{S}^3\times \mathrm{T}^4$}{AdS3 x S3 x T4} with mixed three-form flux and their off-shell symmetry algebra}
\label{sec:off-shell-algebra}
In this section we write down the fully gauge-fixed action for type IIB superstring theory on $\text{AdS}_3\times \mathrm{S}^3\times \mathrm{T}^4$ with mixed flux, determine the classical charges of this theory and compute the off-shell Poisson-bracket algebra $\mathcal{A}$ that the charges satisfy.
The coset formulation~\cite{Cagnazzo:2012se} of type IIB superstring theory on $\text{AdS}_3\times \mathrm{S}^3\times \mathrm{T}^4$ with mixed flux is useful in investigating classical integrability of these theories. However, the coset action which can be obtained from a Green-Schwarz action~\cite{Grisaru:1985fv} by fully fixing the kappa symmetry to the so-called coset kappa gauge~\cite{Babichenko:2009dk}, does not allow for a straightforward computation of the Poisson brackets of the symmetries. This is because the massless fermions have non-standard kinetic terms in the bosonic light-cone gauge. Instead, one needs to work directly with the Green-Schwarz action~\cite{Grisaru:1985fv} in the BMN light-cone kappa gauge. While expressions for this action are known explicitly up to quartic order in fermions~\cite{Wulff:2013kga}, we will only work up to quadratic order in fermions and so can use the actions written down in ref.~\cite{Cvetic:1999zs}.
We begin by writing down expressions for the Killing spinors of this background in section~\ref{sec:Killing-spinors}.
In section~\ref{sec:action} we gauge fix the bosonic action of the theory and write down the explicit expressions for the non-dynamical fields.
In section~\ref{sec:GS-action} we use the Killing spinors to write down explicitly the superstring action in a mixed-flux background and impose the BMN light-cone kappa gauge. In section~\ref{sec:currents-and-algebra} we write down the Noether currents for the charges that generate the algebra $\mathcal{A}$ and compute the Poisson brackets of these charges off shell in order to determine the classical algebra.
\subsection{Killing spinors of IIB supergravity on \texorpdfstring{$\text{AdS}_3\times \mathrm{S}^3\times \mathrm{T}^4$}{AdS3 x S3 x T4} with mixed flux}
\label{sec:Killing-spinors}
Expressions for Killing spinors on $\mathrm{S}^n$ and $\text{AdS}_n$ can be found in ref.~\cite{Lu:1996rhb,Lu:1998nu} in a particular coordinate system. Throughout this paper we will find it useful to work in a different coordinate system---one that is well suited for expansion around the BMN ground state---and so present the expressions for Killing spinors in this coordinate system below. Explicitly, we take the $\text{AdS}_3\times \mathrm{S}^3\times \mathrm{T}^4$ metric to be
\begin{equation}
\label{eq:metric}
ds^2 = ds_{\text{AdS}_3}^2 + ds_{\mathrm{S}^3}^2 + dX_i dX_i ,
\end{equation}
where
\begin{equation}
\label{eq:s3metric}
ds^2_{\mathrm{S}^3} = +\Bigl(\frac{1 - \frac{y_3^2 + y_4^2}{4}}{1 + \frac{y_3^2 + y_4^2}{4}}\Bigr)^2 d\phi^2 + \Bigl(\frac{1}{1 + \frac{y_3^2 + y_4^2}{4}}\Bigr)^2 ( dy_3^2 + dy_4^2 )
\end{equation}
and
\begin{equation}
\label{eq:ads3metric}
ds^2_{\text{AdS}^3} = -\Bigl(\frac{1 + \frac{z_1^2 + z_2^2}{4}}{1 - \frac{z_1^2 + z_2^2}{4}}\Bigr)^2 dt^2 + \Bigl(\frac{1}{1 - \frac{z_1^2 + z_2^2}{4}}\Bigr)^2 ( dz_1^2 + dz_2^2 ) ,
\end{equation}
In these coordinates the NS-NS $B$ field is given by
\begin{equation}
B =
\frac{q}{\bigl( 1 - \frac{z^2}{4}\bigr)^2} ( z_1 \, dz_2 \wedge dt + z_2 \, dt \wedge dz_1 )
+ \frac{q}{\bigl( 1 + \frac{y^2}{4}\bigr)^2} ( y_3 \, dy_4 \wedge d\phi + y_4 \, d\phi \wedge dy_3) .
\end{equation}
This leads to the NS-NS three form
\begin{equation}
H = dB =
2 q \frac{\hphantom{\bigl(}1 + \frac{z^2}{4}\hphantom{\bigl)^3}}{\bigl( 1 - \frac{z^2}{4}\bigr)^3} dt \wedge dz_1 \wedge dz_2
+ 2 q \frac{\hphantom{\bigl(}1 - \frac{y^2}{4}\hphantom{\bigl)^3}}{\bigl( 1 + \frac{y^2}{4}\bigr)^3} dy_3 \wedge dy_4 \wedge d\phi .
\end{equation}
Hence, the R-R and NS-NS three forms have tangent space components
\begin{equation}
F_{012} = F_{345} = 2 \tilde{q} , \qquad H_{012} = H_{345} = 2 q.
\end{equation}
The Killing spinor equations take the form
\begin{equation}
\label{eq:AdS3-S3-T4-Killing-spinor-eq}
\bigl( \delta_{IJ}(\partial_m + \frac{1}{4} \slashed{\omega}_m) + \frac{1}{48}\sigma^3_{IJ} \slashed{F} \slashed{E}_m + \frac{1}{8} \sigma^1_{IJ}\slashed{H}_m\bigr) \tilde{\varepsilon}_J = 0 ,
\end{equation}
where $\omega_m$ is the spin-connection, whose explicit components can be found in ref.~\cite{Borsato:2014hja} and the fluxes can be written as
\begin{equation}
\label{eq:Hslash}
\slashed{H}_m\equiv H_{mAB} \Gamma^{AB} = 2 q \bigl( \slashed{E}_m ( \Gamma^{012} + \Gamma^{345} ) + ( \Gamma^{012} + \Gamma^{345} ) \slashed{E}_m \bigr) ,
\end{equation}
and
\begin{equation}
\label{eq:Fslash}
\slashed{F}\equiv F_{ABC} \Gamma^{ABC} = 12\tilde{q} ( \Gamma^{012} + \Gamma^{345} ) .
\end{equation}
As shown in appendix~\ref{app:killing-spinors}, the solutions of the Killing spinor equations~\eqref{eq:AdS3-S3-T4-Killing-spinor-eq} are
\begin{equation}\label{eq:killing-spinor-lin-combs}
\tilde{\varepsilon}_1 = \sqrt{\frac{1+\tilde{q}}{2}} \, \varepsilon_1 - \sqrt{\frac{1-\tilde{q}}{2}} \, \varepsilon_2 , \qquad
\tilde{\varepsilon}_2 = \sqrt{\frac{1+\tilde{q}}{2}} \, \varepsilon_2 + \sqrt{\frac{1-\tilde{q}}{2}} \, \varepsilon_1 ,
\end{equation}
where $\varepsilon_I$ are the pure R-R background Killing spinors found in ref.~\cite{Borsato:2014hja}. Recall that these latter spinors can be written as
\begin{equation}
\varepsilon^1
=
\hat{M} \varepsilon_0^1,
\qquad\qquad
\varepsilon^2
=
\check{M} \varepsilon_0^2,
\end{equation}
where $\varepsilon_0^I$ are constant 9+1 dimensional Majorana-Weyl spinors, which further satisfy
\begin{equation}
\frac{1}{2} ( 1 + \Gamma^{012345} ) \varepsilon^I =
\frac{1}{2} ( 1 + \Gamma^{012345} ) \varepsilon_0^I = 0.
\end{equation}
Explicit expressions for the matrices $\hat{M}$ and $\check{M}$ are given in equations~\eqref{eq:M0-def} and~\eqref{eq:Mt-def}. We will find it useful to separate the dependence of these matrices on the transverse and light-cone coordinates of $\text{AdS}_3\times \mathrm{S}^3$
\begin{equation}
\hat{M}(z_{\underline{i}},y_{\underline{i}},t,\phi) = M_0(z_{\underline{i}},y_{\underline{i}}) M_t(t,\phi) , \qquad
\check{M}(z_{\underline{i}},y_{\underline{i}},t,\phi) = M_0^{-1}(z_{\underline{i}},y_{\underline{i}}) M_t^{-1}(t,\phi) .
\end{equation}
Below, so as not to over-crowd the notation, we will drop the explicit coordinate dependence and simply write $\hat{M}$, $\check{M}$, $M_t^{\pm 1}$ and $M_0^{\pm 1}$.
Before ending this sub-section we would like to use the matrices $\hat{M}$ and $\check{M}$ to define tangent-space rotations $ \hat{\mathcal{M}}_B{}^A$ and $\check{\mathcal{M}}_B{}^A$ which will be useful in the following sub-sections
\begin{equation}\label{eq:MN-identities}
\hat{M}^{-1} \Gamma^A \hat{M} = \Gamma^B \hat{\mathcal{M}}_B{}^A ,
\qquad
\check{M}^{-1} \Gamma^A \check{M} = \Gamma^B \check{\mathcal{M}}_B{}^A .
\end{equation}
These matrices are block diagonal,
\begin{equation}
\label{eq:orth-rot-ads3-s3}
\hat{\mathcal{M}} = \hat{\mathcal{M}}_{\text{AdS}_3} \oplus \hat{\mathcal{M}}_{\mathrm{S}^3} \oplus \mathbf{1}_4 ,
\qquad
\check{\mathcal{M}} = \check{\mathcal{M}}_{\text{AdS}_3} \oplus \check{\mathcal{M}}_{\mathrm{S}^3} \oplus \mathbf{1}_4 ,
\end{equation}
and explicit expressions for them can be found in Appendix C of~\cite{Borsato:2014hja}.
\subsection{The mixed-flux bosonic action and gauge-fixing}
\label{sec:action}
In this sub-section we write down the bosonic action of the mixed-flux background and impose uniform light-cone gauge. The gauge-fixing determines the non-dynamical fields ($x^{\pm}$ and $\gamma^{\alpha\beta}$) in terms of the physical degrees of freedom. Below, when computing the symmetry algebra $\mathcal{A}$, we will work to quartic order in transverse bosons and quadratic order in transverse fermions. As a result, we will only need explicit expressions for the non-dynamical fields up to zeroth order in fermions and quadratic order in transverse bosons. The bosonic action is\footnote{%
We suppress the overall string tension $\sqrt{\lambda}/2\pi$ in the worldsheet action, and only reinsert it in the final result for the central charge.%
} %
\begin{equation}
S_B = -\frac{1}{2} \int_{-r}^{+r} d\sigma \bigl(
\gamma^{\alpha\beta} G_{mn} \partial_\alpha X^m \partial_\beta X^n + \epsilon^{\alpha\beta} B_{mn} \partial_\alpha X^m \partial_\beta X^n
\bigr),
\end{equation}
where the range of the worldsheet coordinate $\sigma$ is given by $-r < \sigma < +r$.
Introducing the canonical momenta
\begin{equation}
p_m = - \gamma^{00} G_{MN} \dot{X}^N - \gamma^{01} G_{MN} \pri{X}^N - B_{MN} \pri{X}^N
\end{equation}
the action can be written in the first order form
\begin{equation}
S_B = \int_{-r}^{+r} d\sigma \bigl( p_m \dot{X}^m + \frac{\gamma^{01}}{\gamma^{00}} C_1 + \frac{1}{2\gamma^{00}} C_2 \bigr) ,
\end{equation}
with
\begin{equation}
C_1 = p_m \pri{X}^m
\end{equation}
and
\begin{equation}
C_2 = G^{mn} p_m p_n + G_{mn} \pri{X}^m\pri{X}^n + 2 G^{mn} B_{nk} p_m \pri{X}^K + G^{mn} B_{mk} B_{nl} \pri{X}^k \pri{X}^l .
\end{equation}
The constraints $C_1=0$ and $C_2=0$ are equivalent to the Virasoro constraints
\begin{equation}
\gamma^{11} G_{mn} \dot{X}^m \pri{X}^n + \gamma^{01} G_{mn} \dot{X}^m \dot{X}^n = 0 , \qquad
\gamma^{00} G_{mn} \dot{X}^m \dot{X}^n - \gamma^{11} G_{mn} \pri{X}^m \pri{X}^n = 0.
\end{equation}
We want to fix uniform light-cone gauge in which $x^+ = \tau$ and $p_-$ is constant.\footnote{%
We set $x^{\pm} = \frac{1}{2} ( \phi \pm t )$.
} %
Solving the Virasoro constraints we find\footnote{We have checked that these equations are consistent using the equations of motion for the transverse bosons.}
\begin{equation}
\begin{aligned}
\dot{x}^- &= - \frac{1}{4} (\dot{z}^2 + \dot{y}^2 +\dot{x}^2+\pri{z}^2 + \pri{y}^2 + \pri{x}^2 -z^2 -y^2), \\
\pri{x}^- &= -\frac{1}{2}(\dot{z} \cdot \pri{z} + \dot{y} \cdot \pri{y} + \dot{x} \cdot \pri{x}).
\end{aligned}
\end{equation}
Using the $x^{\pm}$ equations of motion we further find that to quadratic order in fields the worldsheet metric is
\begin{equation}
\begin{aligned}
\gamma^{00} &= -1 + \frac{z^2-y^2}{2} - \frac{q}{2} \epsilon^{\underline{i}\underline{j}}(z_{\underline{i}}\pri{z}_{\underline{j}}-y_{\underline{i}}\pri{y}_{\underline{j}}) , \\
\gamma^{01} &= \frac{q}{2}\epsilon^{\underline{i}\underline{j}}(z_{\underline{i}}\pri{z}_{\underline{j}}-y_{\underline{i}}\pri{y}_{\underline{j}}) , \\
\gamma^{11} &=+1 + \frac{z^2-y^2}{2} - \frac{q}{2} \epsilon^{\underline{i}\underline{j}}(z_{\underline{i}}\pri{z}_{\underline{j}}-y_{\underline{i}}\pri{y}_{\underline{j}}) .
\end{aligned}
\end{equation}
Note in particular that for $q \neq 0$, the worldsheet metric is non-diagonal already at quadratic order in fields.
In the limit where the light-cone momentum
\begin{equation}
P_- = \int_{-r}^{+r} d\sigma \, p_- = 4r
\end{equation}
is infinite, the worldsheet becomes decompactified and we are effectively working on a plane rather than a cylinder.
In the transverse directions we impose periodic boundary conditions $x^i(-r) = x^i(+r)$ and $x^{\underline{i}}(-r) = x^{\underline{i}}(+r)$. Physical closed string states should further be periodic in the light-cone direction $x^-$. This leads to the condition
\begin{equation}
\Delta x^- = x^-(+r) - x^-(-r) = \int_{-r}^{+r} d\sigma \, \pri{x}^- = 0 .
\end{equation}
The quantity $\Delta x^-$ is directly related to the worldsheet momentum
\begin{equation}
p_{\text{ws}} = - \int_{-r}^{+r} d\sigma \, ( p_{\underline{i}} \pri{x}^{\underline{i}} + p_i \pri{x}^i ) = 2 \Delta x^- .
\end{equation}
Here we have assumed that there is no winding along the $\phi$ direction. For non-zero winding number $w \in \mathbb{Z}$, the level-matching condition takes the form
\begin{equation}
\label{eq:level-matching}
p_{\text{ws}} = 2\pi w .
\end{equation}
However, in the rest of this section we will work at zero winding. Moreover, we are mainly interested in studying the symmetries of the worldsheet theory when we go \emph{off shell} by allowing the worldsheet momentum to take arbitrary values.
\subsection{The Green-Schwarz action}
\label{sec:GS-action}
In this sub-section we describe the gauge-fixing of the Green-Schwarz action in a form that will be particularly suited to computing $\mathcal{A}$. This computation requires the action to quadratic order in fermions, and we give explicit expressions for the action to this order in appendix~\ref{app:lagrangian}. The Green-Schwarz action for type IIB superstrings in a general supergravity background was written down in terms of superfields in ref.~\cite{Grisaru:1985fv} and explicit expressions in an expansion of fermions up to quadratic~\cite{Cvetic:1999zs} and quartic~\cite{Wulff:2013kga} order are known.
The Lagrangian can be written as
\begin{equation}
L = L_B + L_{\text{\scriptsize kin}} + L_{\text{WZ}} ,
\end{equation}
with the bosonic part, $L_B$, discussed in the previous sub-section. The remaining part of the Lagrangian is split into two terms: a term dependent on the worldsheet metric, $L_{\text{\scriptsize kin}}$, and the Wess-Zumino term $L_{\text{WZ}}$. Up to quadratic order in fermions these are given by ~\cite{Cvetic:1999zs}
\begin{align}
L_{\text{kin}} &=
-i\gamma^{\alpha\beta} \bar{\tilde{\theta}}_I \slashed{E}_\alpha \bigl( \delta^{IJ} D_\beta + \frac{1}{48} \sigma_3^{IJ} \slashed{F} \slashed{E}_\beta + \frac{1}{8} \sigma_1^{IJ} \slashed{H}_\beta \bigr) \tilde{\theta}_J ,\\
L_{\text{WZ}} &=
+i\epsilon^{\alpha\beta} \bar{\tilde{\theta}}_I \sigma_1^{IJ} \slashed{E}_\alpha \bigl( \delta^{JK} D_\beta + \frac{1}{48} \sigma_3^{JK} \slashed{F} \slashed{E}_\beta + \frac{1}{8} \sigma_1^{JK} \slashed{H}_\beta \bigr) \tilde{\theta}_K .
\end{align}
It is helpful to perform a field re-definition of the fermions so as to end up with fermionic coordinates that are best adapted to the underlying integrable structure. It is easiest to understand this field redefinition as a combination of two transformations of the fermionic coordinates $\tilde{\theta}_I$ appearing above. Initially we ``rotate'' the fermions along the $I-J$ index
\begin{equation}
\label{eq:fermion-IJ-rot}
\tilde{\theta}_1 = \sqrt{\frac{1+\tilde{q}}{2}} \, \theta_1 - \sqrt{\frac{1-\tilde{q}}{2}} \, \theta_2 , \qquad
\tilde{\theta}_2 = \sqrt{\frac{1+\tilde{q}}{2}} \, \theta_2 + \sqrt{\frac{1-\tilde{q}}{2}} \, \theta_1 .
\end{equation}
This ensures that the kinetic term in the Lagrangian is diagonal in terms of the $\theta_I$. There are now two different field redefinitions that are useful to consider for different purposes. The first is to redefine
\begin{equation}\label{eq:theta-vartheta-def}
\begin{aligned}
\theta_1 &= \hat{M} \frac{1-\Gamma^{012345}}{2} \vartheta_1^- + \hat{M} \frac{1+\Gamma^{012345}}{2} \vartheta^+_1 , \\
\theta_2 &= \check{M} \frac{1-\Gamma^{012345}}{2} \vartheta^-_2 + \check{M} \frac{1+\Gamma^{012345}}{2} \vartheta^+_2 .
\end{aligned}
\end{equation}
In the resulting action, before kappa-gauge fixing, supersymmetry is realised as a shift on the fermions $\vartheta^-_I$. The expression for the resulting Lagrangian is written down in equations~\eqref{eq:pre-gauge-action-kin} and~\eqref{eq:pre-gauge-action-WZ}. However, we will be interested in a (suitably) kappa gauge fixed action, and so we will need to perform a different field redefinition to the one above. It turns out that a particular kappa gauge simplifies the computation of the algebra $\mathcal{A}$. This kappa gauge is the so-called BMN light-cone kappa gauge for fermions that are neutral with respect to the two $U(1)$'s associated to shifts along $t$ and $\phi$~\cite{Alday:2005ww}. As a result, in addition to~\eqref{eq:fermion-IJ-rot}, the second redefinition of the fermions we perform is
\begin{equation}\label{eq:theta-chi-eta-def-t}
\begin{aligned}
\theta_1 &=
\frac{1}{2} ( 1 + \Gamma^{012345} ) \mathrlap{M_0 \chi_1}\hphantom{M_0^{-1} \chi_2}
+ \frac{1}{2} ( 1 - \Gamma^{012345} ) M_0 \eta_1 ,
\\
\theta_2 &= \frac{1}{2} ( 1 + \Gamma^{012345} ) M_0^{-1} \chi_2 + \frac{1}{2} ( 1 - \Gamma^{012345} ) M_0^{-1} \eta_2 .
\end{aligned}
\end{equation}
where the matrix $M_0$ and its inverse are defined in equation~\eqref{eq:M0-def}. The fermions $\eta_I$ and $\chi_I$ correspond to the massive and massless fermions, respectively, of the integrable S matrix.
Having defined fermions that are neutral under shifts in $t$ and $\phi$, we impose the BMN light-cone kappa gauge
\begin{equation}\label{eq:bmn-lc-kappa-gauge}
\Gamma^+ \eta_I = 0 , \qquad
\Gamma^+ \chi_I = 0 , \qquad
\Gamma^{\pm} = \frac{1}{2} \bigl( \Gamma^5 \pm \Gamma^0 \bigr) .
\end{equation}
The resulting light-cone kappa gauged-fixed action is written down in equations~\eqref{eq:post-gauge-action-kin} and~\eqref{eq:post-gauge-action-WZ}.
\subsection{The algebra \texorpdfstring{$\mathcal{A}$}{A}}
\label{sec:currents-and-algebra}
In this section we give the algebra $\mathcal{A}$ of (super)charges which commute with the Hamiltonian. As in the case of pure R-R flux the algebra itself is given by\footnote{%
We use the direct sum to denote the sum of the subalgebras as vector spaces. This does not imply that they commute with each other as will be clear when we write down the full commutation relations in section~\ref{sec:algebra}.%
}%
\begin{equation}
\label{eq:Aalgebra}
\mathcal{A} = \alg{psu}(1|1)^4_{\text{c.e.}} \oplus \alg{so}(4) ,
\end{equation}
where the subscript $\text{c.e.}$ denotes a four-fold central extension. In section~\ref{sec:currents} we first give the supercurrents $j_I$ that generate the algebra. In section~\ref{sec:supercurrent-algebra} we consider the Poisson brackets of these supercurrents, see that we produce the expected Hamiltonian and compute the central charges of the algebra. In particular we find that the off-shell central charges are simple rescalings of the ones of the pure R-R theory. Throughout this section we give explicit results in the main text to quadratic order in fields, with higher order results given in the appendices. We use a ``hybrid'' expansion~\cite{Arutyunov:2006ak} in which we expand order by order in fermions and transverse bosons, but keep all factors of the light-cone coordinate $x^-$ exact. This allows us to compute the central charges exactly in momentum.
\subsubsection{Supercurrents}
\label{sec:currents}
The supercurrents of the algebra $\mathcal{A}$ to quadratic order are given by
\begin{align}
\label{eq:quadratic-currents}
j^{\tau}_1 =
i e^{+x^- \gamma^{34}} \bigl(
& ( \dot{z}^{\underline{i}} - \dot{y}^{\underline{i}} ) \gamma_{\underline{i}} \eta_1
+ ( z^{\underline{i}} + y^{\underline{i}} ) \gamma^{34} \gamma_{\underline{i}} \eta_1
- ( \pri{z}^{\underline{i}} - \pri{y}^{\underline{i}} ) \gamma_{\underline{i}} ( \tilde{q} \eta_2 + q \eta_1 ) \\\nonumber
+ & \dot{x}^i \gamma^{34} \tilde{\tau}_i \chi_1
- \pri{x}^i \gamma^{34} \tilde{\tau}_i ( \tilde{q} \chi_2 + q \chi_1 )
\bigr) ,
\\\nonumber
j^{\tau}_2 =
i e^{-x^- \gamma^{34}} \bigl(
& ( \dot{z}^{\underline{i}} - \dot{y}^{\underline{i}} ) \gamma_{\underline{i}} \eta_2
- ( z^{\underline{i}} + y^{\underline{i}} ) \gamma^{34} \gamma_{\underline{i}} \eta_2
- ( \pri{z}^{\underline{i}} - \pri{y}^{\underline{i}} ) \gamma_{\underline{i}} ( \tilde{q} \eta_1 - q \eta_2 ) \\\nonumber
+ & \dot{x}^i \gamma^{34} \tilde{\tau}_i \chi_2
- \pri{x}^i \gamma^{34} \tilde{\tau}_i ( \tilde{q} \chi_1 - q \chi_2 )
\bigr) ,
\\\nonumber
j^{\sigma}_1 =
i e^{+x^- \gamma^{34}} \bigl(
& ( \dot{z}^{\underline{i}} - \dot{y}^{\underline{i}} ) \gamma_{\underline{i}} ( \tilde{q} \eta_2 + q \eta_1 )
+ ( z^{\underline{i}} + y^{\underline{i}} ) \gamma^{34} \gamma_{\underline{i}} ( \tilde{q} \eta_2 + q \eta_1 )
- ( \pri{z}^{\underline{i}} - \pri{y}^{\underline{i}} ) \gamma_{\underline{i}} \eta_1 \\\nonumber
+ & \dot{x}^i \gamma^{34} \tilde{\tau}_i ( \tilde{q} \chi_2 + q \chi_1 )
- \pri{x}^i \gamma^{34} \tilde{\tau}_i ) \chi_1
\bigr) ,
\\\nonumber
j^{\sigma}_2 =
i e^{-x^- \gamma^{34}} \bigl(
& ( \dot{z}^{\underline{i}} - \dot{y}^{\underline{i}} ) \gamma_{\underline{i}} ( \tilde{q} \eta_1 - q \eta_2 )
- ( z^{\underline{i}} + y^{\underline{i}} ) \gamma^{34} \gamma_{\underline{i}} ( \tilde{q} \eta_1 - q \eta_2 )
- ( \pri{z}^{\underline{i}} - \pri{y}^{\underline{i}} ) \gamma_{\underline{i}} \eta_2 \\\nonumber
+ & \dot{x}^i \gamma^{34} \tilde{\tau}_i ( \tilde{q} \chi_1 - q \chi_2 )
- \pri{x}^i \gamma^{34} \tilde{\tau}_i \chi_2
\bigr) .
\end{align}
In appendix~\ref{app:quartic-currents} we give expressions for the supercurrents to cubic order in transverse bosons and leading order in fermions. These expressions for the supercurrents are given in terms of fermions written as bispinors of $\alg{so}(4)_1 \oplus \alg{so}(4)_2 \in \alg{so}(1,9)$,\footnote{We have suppressed the corresponding spinor indices. These can easily be put back in; as defined in~\ref{app:conventions} the fermions carry spinor indices $(\eta_I)^{\underline{\dot{a}}\dot{b}}$ and $(\chi_I)^{\underline{a}b}$, while the $\alg{so}(4)$ gamma matrices carry indices $(\gamma_{\underline{i}})^{\underline{a}}{}_{\underline{\dot{b}}}$, $(\tilde{\tau}_i)^{\dot{a}}{}_b$ and $(\gamma^{34})^{\underline{a}}{}_{\underline{b}}$.} corresponding to rotations of $(z^{\underline{i}}, y^{\underline{i}})$ and $x^i$, as defined in ref.~\cite{Borsato:2014hja}.
The Lagrangian of the theory does not preserve~$\alg{so}(4)_1$, which is in fact broken to $\alg{so}(2)^2$ corresponding to separate rotations of~$z^{\underline{i}}$ and~$y^{\underline{i}}$. Nevertheless we will find it useful to write expression in these notation. On the other hand, $\alg{so}(4)_2$ is unbroken, and is part of~$\mathcal{A}$ in equation~\eqref{eq:Aalgebra}.
Our conventions for the gamma matrices in these expressions are given in equation~\eqref{eq:so4-gamma}.
We have checked that the currents given in equation~\eqref{eq:quadratic-currents} and appendix~\ref{app:quartic-currents} satisfy the conservation equation
\begin{equation}
\partial_\tau j_I^\tau + \partial_\sigma j_I^\sigma = 0 ,
\end{equation}
to the required order using the equations of motion arising from the Lagrangian computed in the previous section.
\subsubsection{The algebra from the supercurrents}
\label{sec:supercurrent-algebra}
To find the off-shell symmetry algebra we need to compute Poisson brackets of the supercurrents, which in turn requires the Poisson brackets of the fermions. Explicit expressions for these are given in appendix~\ref{app:poisson}. For the Poisson brackets of two charges with the same index we find\footnote{Here $\epsilon$ symbols carry appropriate spinor indices, which we have suppressed for brevity.}
\begin{equation}
\begin{aligned}
\int d\sigma \, d\sigma' \acommPB{ j_1^{\tau}(\sigma) }{ j_1^{\tau}(\sigma') }
&= + \frac{i}{2} \int d\sigma ( \mathcal{H} + \mathcal{M} ) \, \epsilon \, \epsilon ,
\\
\int d\sigma \, d\sigma' \acommPB{ j_2^{\tau}(\sigma) }{ j_2^{\tau}(\sigma') }
&= + \frac{i}{2} \int d\sigma ( \mathcal{H} - \mathcal{M} ) \, \epsilon \, \epsilon .
\end{aligned}
\end{equation}
The bosonic Hamiltonian density $\mathcal{H}$ is given to quadratic order by
\begin{equation}
\mathcal{H} = \frac{1}{2} \bigl(
p_{\underline{i}} p^{\underline{i}} + p_i p^i
+ \pri{z}_{\underline{i}} \pri{z}^{\underline{i}} + \pri{y}_{\underline{i}} \pri{y}^{\underline{i}} + \pri{x}_i \pri{x}^i
+ z_{\underline{i}} z^{\underline{i}} + y_{\underline{i}} y^{\underline{i}}
- 2 q \epsilon^{\underline{ij}} ( z_{\underline{i}} \pri{z}_{\underline{j}} + y_{\underline{i}} \pri{y}_{\underline{j}} )
\bigr) .
\end{equation}
The full quartic bosonic Hamiltonian can be found in equation~\eqref{eq:quartic-hamiltonian}.
The ``mass'' term $\mathcal{M}$ is given by
\begin{equation}
\mathcal{M} =
- \epsilon^{\underline{ij}} ( p_{\underline{i}} z_{\underline{j}} + p_{\underline{i}} y_{\underline{j}} )
- q ( p_{\underline{i}} \pri{z}^{\underline{i}} + p_{\underline{i}} \pri{y}^{\underline{i}} + p_i \pri{x}^i ) .
\end{equation}
It is important to note that this expression does not receive any corrections at quartic order.
Calculating the Poisson bracket between the two charges with different index we find
\begin{align}
\int d\sigma \, d\sigma' \acommPB{ j_1^{\tau}(\sigma) }{ j_2^{\tau}(\sigma') }
&= - \frac{i\tilde{q}}{2} \int d\sigma \Bigl[
\partial_{\sigma} \bigl( e^{2\gamma^{34} x^-} \bigr)
- \frac{1}{8} e^{2\gamma^{34} x^-} \partial_\sigma \bigl[(z^2 - y^2)^2\bigr]
\\ \nonumber &\qquad
- \partial_\sigma \bigl( e^{2\gamma^{34} x^-} \frac{z^2 - y^2}{2}
- z^{\underline{i}} y^{\underline{j}} \bigl( 1 + \frac{z^2 - y^2}{4} \bigr) \gamma_{\underline{ij}} \bigr)
\Bigr] \gamma^{34} \epsilon \, \epsilon.
\end{align}
The total derivative on the second line integrates to zero. The second term in the first line can be integrated by parts. The result is of higher order in transverse bosons and can therefore be dropped. Hence, we are left with
\begin{equation}
\int d\sigma \, d\sigma' \acommPB{ j_1^{\tau}(\sigma) }{ j_2^{\tau}(\sigma') } =
- \frac{i \tilde{q}}{2} e^{+2\gamma^{34} x^-(-\infty)} \bigl( e^{+\gamma^{34} p_{\text{ws}}} - 1 \bigr) \gamma^{34} \epsilon \, \epsilon
\end{equation}
Hence, we find the central charge
\begin{equation}
C=\frac{i\zeta}{2} \frac{\tilde{q} \sqrt{\lambda}}{2\pi} \left( e^{ip_{\text{ws}}} - 1 \right)
\end{equation}
where we have reintroduced the string tension, and where $\zeta=\mathrm{exp}(2ix^-(-\infty))$. This is related to the central charge of the pure R-R theory by a rescaling by $\tilde{q}$. It is precisely the fact that $C$ depends non-linearly on the momentum which imposes a non-local coproduct on the symmetry algebra, which we will discuss in the section~\ref{sec:coproduct}.
\section{Symmetry algebra and representations}
\label{sec:algebra}
We have seen that the off-shell symmetry algebra~$\mathcal{A}$ takes the same form in the mixed-flux case as it did in the pure-R-R one. Furthermore, in the limit where the NS-NS flux vanishes, we expect to recover precisely the same representations that were described in detail in ref.~\cite{Borsato:2014hja}. For this reason, we begin in section~\ref{sec:RRreprs} by briefly recalling the representations arising in the pure-R-R case. Then in sections~\ref{sec:near-BMN-representations} and~\ref{sec:exact-repr-mixed}, we describe how these are deformed, first in the near-plane-wave limit, and then in the full theory. We will see that the deformation can be completely understood by suitably altering the representation parameters of the pure-R-R case. As we proceed we will encounter, and comment on, several new features of the mixed-flux background.
\subsection{Overview of the pure-R-R symmetries and representations}
\label{sec:RRreprs}
The off-shell symmetry algebra $\mathcal{A}$ for type IIB superstrings on the pure-R-R $\text{AdS}_3\times\mathrm{S}^3\times\mathrm{T}^4$ background has been found in ref.~\cite{Borsato:2014hja}. There it was found that $\mathcal{A}$ is a central extension of $\alg{psu}(1|1)^4\oplus\alg{so}(4)_{2}$. Writing down the supercharges in components we have
\begin{equation}
\label{eq:cealgebra}
\begin{aligned}
&\{\gen{Q}_{\mbox{\tiny L}}^{\ \dot{a}},\overline{\gen{Q}}{}_{\mbox{\tiny L} \dot{b}}\} = \frac{1}{2}\delta^{\dot{a}}_{\ \dot{b}}\,(\gen{H}+\gen{M}),
&\qquad &\{\gen{Q}_{\mbox{\tiny L}}^{\ \dot{a}},{\gen{Q}}{}_{\mbox{\tiny R} \dot{b}}\} = \delta^{\dot{a}}_{\ \dot{b}}\,\gen{C},\\
&\{\gen{Q}_{\mbox{\tiny R} \dot{a}},\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ \dot{b}}\} = \frac{1}{2}\delta^{\ \dot{b}}_{\dot{a}}\,(\gen{H}-\gen{M}),
&\qquad &\{\overline{\gen{Q}}{}_{\mbox{\tiny L} \dot{a}},\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ \dot{b}}\} = \delta^{\ \dot{b}}_{\dot{a}}\,\overline{\gen{C}}.
\end{aligned}
\end{equation}
The supercharges carry labels ``L'' and ``R'' corresponding to the left and right labels in the superisometry algebra $\alg{su}(1,1|2)_{\mbox{\tiny L}} \oplus \alg{su}(1,1|2)_{\mbox{\tiny R}}$. We also decompose
\begin{equation}
\alg{so}(4)_2=\alg{su}(2)_{\bullet}\oplus\alg{su}(2)_{\circ},
\end{equation}
so that the massive fermions are charged only under $\alg{su}(2)_{\bullet}$ .
The lower and upper dotted indices correspond to the fundamental and anti-fundamental representation of $\alg{su}(2)_{\bullet}$ respectively. Finally, the central charges on the one-particle representation are
\begin{equation}
\label{eq:allloop-centralcharges-RR}
\begin{aligned}
&\gen{C}=+\frac{i h}{2}(e^{+i p}-1),
\qquad\qquad &&
\overline{\gen{C}}=-\frac{i h}{2}(e^{-i p}-1),\\
&\gen{H}=\sqrt{m^2+4h^2\sin\bigl(\frac{p}{2}\bigr)^2},
\qquad\qquad &&
\gen{M}=m,
\end{aligned}
\end{equation}
where $p$ is the momentum, $m$ is an angular momentum taking values $\pm 1,0$ and $h$ is the coupling constant, which is expected to be a so-far undetermined function of the 't~Hooft coupling, $h=h(\lambda)$.
\subsubsection{Exact representations for the pure-R-R theory}
\label{sec:RR-exactrepr}
\begin{figure}[t]
\centering
\begin{tikzpicture}[%
box/.style={outer sep=1pt},
Q node/.style={inner sep=1pt,outer sep=0pt},
arrow/.style={-latex}
]%
\node [box] (PhiM) at ( 0 , 2cm) { $\ket{Y^{\mbox{\tiny L}}}$};
\node [box] (PsiP) at (-2cm, 0cm) { $\ket{\eta^{\mbox{\tiny L} 1}}$};
\node [box] (PsiM) at (+2cm, 0cm) { $\ket{\eta^{\mbox{\tiny L} 2}}$};
\node [box] (PhiP) at ( 0 ,-2cm) { $\ket{Z^{\mbox{\tiny L}}}$};
\newcommand{0.09cm,0cm}{0.09cm,0cm}
\newcommand{0cm,0.10cm}{0cm,0.10cm}
\draw [arrow] ($(PhiM.west) +(0cm,0.10cm)$) -- ($(PsiP.north)-(0.09cm,0cm)$) node [pos=0.5,anchor=south east,Q node] {\small $\gen{Q}^{\ 1}_{\mbox{\tiny L}},\overline{\gen{Q}}{}^{\ 1}_{\mbox{\tiny R}}$};
\draw [arrow] ($(PsiP.north)+(0.09cm,0cm)$) -- ($(PhiM.west) -(0cm,0.10cm)$) node [pos=0.5,anchor=north west,Q node] {};
\draw [arrow] ($(PsiM.south)-(0.09cm,0cm)$) -- ($(PhiP.east) +(0cm,0.10cm)$) node [pos=0.5,anchor=south east,Q node] {};
\draw [arrow] ($(PhiP.east) -(0cm,0.10cm)$) -- ($(PsiM.south)+(0.09cm,0cm)$) node [pos=0.5,anchor=north west,Q node] {\small $\overline{\gen{Q}}{}_{\mbox{\tiny L} 1},\gen{Q}_{\mbox{\tiny R} 1}$};
\draw [arrow] ($(PhiM.east) -(0cm,0.10cm)$) -- ($(PsiM.north)-(0.09cm,0cm)$) node [pos=0.5,anchor=north east,Q node] {};
\draw [arrow] ($(PsiM.north)+(0.09cm,0cm)$) -- ($(PhiM.east) +(0cm,0.10cm)$) node [pos=0.5,anchor=south west,Q node] {\small $\overline{\gen{Q}}{}_{\mbox{\tiny L} 2},{\gen{Q}}{}_{\mbox{\tiny R} 2}$};
\draw [arrow] ($(PsiP.south)-(0.09cm,0cm)$) -- ($(PhiP.west) -(0cm,0.10cm)$) node [pos=0.5,anchor=north east,Q node] {\scriptsize $
\gen{Q}^{\ 2}_{\mbox{\tiny L}},\overline{\gen{Q}}{}^{\ 2}_{\mbox{\tiny R}}$};
\draw [arrow] ($(PhiP.west) +(0cm,0.10cm)$) -- ($(PsiP.south)+(0.09cm,0cm)$) node [pos=0.5,anchor=south west,Q node] {};
\draw [arrow] (PsiM) -- (PsiP) node [pos=0.6,anchor=south west,Q node] {\small $\gen{J}^{\ \dot{a}}_{\bullet}$};
\draw [arrow] (PsiP) -- (PsiM);
\draw[rounded corners=5mm] (-3.2cm,-2.6cm)rectangle (3.2cm,2.6cm);
\node (reprlabel) at ( 0 , -3cm) { $\varrho_{\mbox{\tiny L}}\otimes\varrho_{\mbox{\tiny L}}$};
\end{tikzpicture}
\hspace{2cm}
\begin{tikzpicture}[%
box/.style={outer sep=1pt},
Q node/.style={inner sep=1pt,outer sep=0pt},
arrow/.style={-latex}
]%
\node [box] (PhiM) at ( 0 , 2cm) { $\ket{Z^{\mbox{\tiny R}}}$};
\node [box] (PsiP) at (-2cm, 0cm) { $\ket{\eta^{\mbox{\tiny R}}_{\ 1}}$};
\node [box] (PsiM) at (+2cm, 0cm) { $\ket{\eta^{\mbox{\tiny R}}_{\ 2}}$};
\node [box] (PhiP) at ( 0 ,-2cm) { $\ket{Y^{\mbox{\tiny R}}}$};
\newcommand{0.09cm,0cm}{0.09cm,0cm}
\newcommand{0cm,0.10cm}{0cm,0.10cm}
\draw [arrow] ($(PsiP.north)-(0.09cm,0cm)$) -- ($(PhiM.west) +(0cm,0.10cm)$) node [pos=0.5,anchor=south east,Q node] {\small $\gen{Q}_{\mbox{\tiny R} 2},\overline{\gen{Q}}{}_{\mbox{\tiny L} 2}$};
\draw [arrow] ($(PhiM.west) -(0cm,0.10cm)$) -- ($(PsiP.north)+(0.09cm,0cm)$) node [pos=0.5,anchor=north west,Q node] {};
\draw [arrow] ($(PhiP.east) +(0cm,0.10cm)$) -- ($(PsiM.south)-(0.09cm,0cm)$) node [pos=0.5,anchor=south east,Q node] {};
\draw [arrow] ($(PsiM.south)+(0.09cm,0cm)$) -- ($(PhiP.east) -(0cm,0.10cm)$) node [pos=0.5,anchor=north west,Q node] {\small $\overline{\gen{Q}}{}^{\ 2}_{\mbox{\tiny R}},\gen{Q}_{\mbox{\tiny L}}^{\ 2}$};
\draw [arrow] ($(PsiM.north)-(0.09cm,0cm)$) -- ($(PhiM.east) -(0cm,0.10cm)$) node [pos=0.5,anchor=north east,Q node] {};
\draw [arrow] ($(PhiM.east) +(0cm,0.10cm)$) -- ($(PsiM.north)+(0.09cm,0cm)$) node [pos=0.5,anchor=south west,Q node] {\small $\overline{\gen{Q}}{}^{\ 1}_{\mbox{\tiny R}}, \gen{Q}_{\mbox{\tiny L}}^{\ 1}$};
\draw [arrow] ($(PhiP.west) -(0cm,0.10cm)$) -- ($(PsiP.south)-(0.09cm,0cm)$) node [pos=0.5,anchor=north east,Q node] {\small $\gen{Q}_{\mbox{\tiny R} 1}, \overline{\gen{Q}}{}_{\mbox{\tiny L} 1}$};
\draw [arrow] ($(PsiP.south)+(0.09cm,0cm)$) -- ($(PhiP.west) +(0cm,0.10cm)$) node [pos=0.5,anchor=south west,Q node] {};
\draw [arrow] (PsiM) -- (PsiP) node [pos=0.6,anchor=south west,Q node] {\small $\gen{J}^{\ \dot{a}}_{\bullet}$};
\draw [arrow] (PsiP) -- (PsiM);
\draw[rounded corners=5mm] (-3.2cm,-2.6cm)rectangle (3.2cm,2.6cm);
\node (reprlabel) at ( 0 , -3cm) { $\varrho_{\mbox{\tiny R}}\otimes\varrho_{\mbox{\tiny R}}$};
\end{tikzpicture}
\caption{%
The massive excitations and their transformation properties under~$\mathcal{A}$.
The left and right panel depict the left and right representations respectively. The bosons $Z^{\mbox{\tiny L},\mbox{\tiny R}}$ are excitations on $\text{AdS}_3$ while $Y^{\mbox{\tiny L},\mbox{\tiny R}}$ are excitations on $\mathrm{S}^3$. Note that the massive fermions $\eta^{\mbox{\tiny L} \dot{a}},\eta^{\mbox{\tiny R} \dot{a}}$ are charged under~$\alg{su}(2)_{\bullet}$. The tensor products below each diagram indicate how each representation can be obtained from the short fundamental representations of $\alg{su}(1|1)^2_{\text{c.e.}}$ introduced in section~\ref{sec:repr-smallalgebra}.
}
\label{fig:repr:massive}
\end{figure}
The fundamental excitations of the theory are 8 bosons and 8 fermions, which arrange themselves into three irreducible representations of~$\mathcal{A}$. The $4+4$ massive excitations transform in two irreducible representations, that we call ``left'' and ``right'' and depict in figure~\ref{fig:repr:massive}. The remaining modes transform in the ``massless'' representation of~$\mathcal{A}$, depicted in figure~\ref{fig:repr:massless}. All these are short representations of $\alg{psu}(1|1)^4_{\text{c.e.}}$, \textit{i.e.}\xspace they satisfy the shortening condition
\begin{equation}
\label{eq:shortening}
\gen{H}^2=\gen{M}^2+4\,\gen{C}\,\overline{\gen{C}}.
\end{equation}
The left representation is four-dimensional and has $m=+1$. It is an irreducible representation of $\alg{psu}(1|1)^4_{\text{c.e.}}$ and it owes its name to the fact that on shell only the left supercharges act non-trivially on its module. We can write it as
\begin{equation}\label{eq:repr-massive-L}
\begin{aligned}
\gen{Q}_{\mbox{\tiny L}}^{\ \dot{a}} \ket{Y_p^{\mbox{\tiny L}}} &= a^{\mbox{\tiny L}}_p \ket{\eta^{\mbox{\tiny L} \dot{a}}_p},
\qquad
&\gen{Q}_{\mbox{\tiny L}}^{\ \dot{a}} \ket{\eta^{\mbox{\tiny L} \dot{b}}_p} &= \epsilon^{\dot{a}\dot{b}} \, a^{\mbox{\tiny L}}_p \ket{Z_p^{\mbox{\tiny L}}}, \\
\overline{\gen{Q}}{}_{\mbox{\tiny L} \dot{a}} \ket{Z_p^{\mbox{\tiny L}}} &= - \epsilon_{\dot{a}\dot{b}} \, \bar{a}^{\mbox{\tiny L}}_p \ket{\eta^{\mbox{\tiny L} \dot{b}}_p},
\qquad
&\overline{\gen{Q}}{}_{\mbox{\tiny L} \dot{a}} \ket{\eta^{\mbox{\tiny L} \dot{b}}_p}& = \delta_{\dot{a}}^{\ \dot{b}} \, \bar{a}^{\mbox{\tiny L}}_p \ket{Y_p^{\mbox{\tiny L}}}, \\[4pt]
\gen{Q}_{\mbox{\tiny R} \dot{a}} \ket{Z^{\mbox{\tiny L}}_p} &= - \epsilon_{\dot{a}\dot{b}} \, b^{\mbox{\tiny L}}_p \ket{\eta^{\mbox{\tiny L} \dot{b}}_p},
\qquad
&\gen{Q}_{\mbox{\tiny R} \dot{a}} \ket{\eta^{\mbox{\tiny L} \dot{b}}_p} &= \delta_{\dot{a}}^{\ \dot{b}} \, b^{\mbox{\tiny L}}_p \ket{Y^{\mbox{\tiny L}}_p},\\
\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ \dot{a}} \ket{Y^{\mbox{\tiny L}}_p} &= \bar{b}^{\mbox{\tiny L}}_p \ket{\eta^{\mbox{\tiny L} \dot{a}}_p},
\qquad
&\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ \dot{a}} \ket{\eta^{\mbox{\tiny L} \dot{b}}_p} &= \epsilon^{\dot{a}\dot{b}} \, \bar{b}^{\mbox{\tiny L}}_p \ket{Z^{\mbox{\tiny L}}_p}.
\end{aligned}
\end{equation}
The representation coefficients $a^{\mbox{\tiny L}}_p$ and $b^{\mbox{\tiny L}}_p$ are such as to reproduce the central charges~\eqref{eq:allloop-centralcharges-RR} and $b^{\mbox{\tiny L}}_p$ vanishes on shell, \textit{i.e.}\xspace, $b^{\mbox{\tiny L}}_{p=0}=0$. We will comment more on the $a^{\mbox{\tiny L}}_p$ and $b^{\mbox{\tiny L}}_p$ in the next subsection.
The right representation is also four-dimensional and has $m=-1$. It is given by
\begin{equation}\label{eq:repr-massive-R}
\begin{aligned}
\gen{Q}_{\mbox{\tiny L}}^{\ \dot{a}} \ket{Z_p^{\mbox{\tiny R}}} &= b^{\mbox{\tiny R}}_p \ket{\eta^{\mbox{\tiny R} \dot{a}}_p},
\qquad
&\gen{Q}_{\mbox{\tiny L}}^{\ \dot{a}} \ket{\eta^{\mbox{\tiny R} \dot{b}}_p} &=- \epsilon^{\dot{a}\dot{b}} \, b^{\mbox{\tiny R}}_p \ket{Y_p^{\mbox{\tiny R}}},\\
\overline{\gen{Q}}{}_{\mbox{\tiny L} \dot{a}} \ket{Y_p^{\mbox{\tiny R}}} &= \epsilon_{\dot{a}\dot{b}} \, \bar{b}^{\mbox{\tiny R}}_p \ket{\eta^{\mbox{\tiny R} \dot{b}}_p},
\qquad
&\overline{\gen{Q}}{}_{\mbox{\tiny L} \dot{a}} \ket{\eta^{\mbox{\tiny R} \dot{b}}_p} &= \delta_{\dot{a}}^{\ \dot{b}} \, \bar{b}^{\mbox{\tiny R}}_p \ket{Z_p^{\mbox{\tiny R}}},\\[4pt]
\gen{Q}_{\mbox{\tiny R} \dot{a}} \ket{Y_p^{\mbox{\tiny R}}} &= \epsilon_{\dot{a}\dot{b}} \, a^{\mbox{\tiny R}}_p \ket{\eta^{\mbox{\tiny R} \dot{b}}_p},
\qquad
&\gen{Q}_{\mbox{\tiny R} \dot{a}} \ket{\eta^{\mbox{\tiny R} \dot{b}}_p} &= \delta_{\dot{a}}^{\ \dot{b}} \, a^{\mbox{\tiny R}}_p \ket{Z_p^{\mbox{\tiny R}}}, \\
\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ \dot{a}} \ket{Z_p^{\mbox{\tiny R}}} &= \bar{a}^{\mbox{\tiny R}}_p \ket{\eta^{\mbox{\tiny R} \dot{a}}_p},
\qquad
&\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ \dot{a}} \ket{\eta^{\mbox{\tiny R} \dot{b}}_p} &= - \epsilon^{\dot{a}\dot{b}} \, \bar{a}^{\mbox{\tiny R}}_p \ket{Y_p^{\mbox{\tiny R}}}.
\end{aligned}
\end{equation}
Note that this right representation follows from the previous one by relabelling everywhere~L$\leftrightarrow$R. We will refer this $\mathbb{Z}_2$ symmetry as \emph{left-right symmetry} (LR symmetry).
Finally, the massless representation is eight-dimensional with four bosons $T^{\dot{a}a}$ and four fermions $\chi^{a},\widetilde{\chi}^{a}$. This representation has $m=0$ and is given by two irreducible representations of~$\alg{psu}(1|1)^4_{\text{c.e.}}$ that form a doublet under~$\alg{su}(2)_{\circ}\subset\alg{so}(4)_2$. Each of these $\alg{psu}(1|1)^4_{\text{c.e.}}$ representations can equivalently be obtained by using the left or right representations above and taking a massless limit of the coefficients $a^{\mbox{\tiny L}}_p,b^{\mbox{\tiny L}}_p$ or $a^{\mbox{\tiny R}}_p,b^{\mbox{\tiny R}}_p$. This is due to the fact that the left and right representations become isomorphic in the massless limit. For definiteness, let us take the representation coefficients to be inherited from the left representation. Then we have
\begin{equation}\label{eq:repr-massless}
\begin{aligned}
\gen{Q}_{\mbox{\tiny L}}^{\ \dot{a}} \ket{T^{\dot{b}a}_p}& = \epsilon^{\dot{a}\dot{b}} a^{\mbox{\tiny L}}_p \ket{\widetilde{\chi}^a_p},
\qquad
&\gen{Q}_{\mbox{\tiny L}}^{\ \dot{a}} \ket{\chi^{a}_p} \;&= a^{\mbox{\tiny L}}_p \ket{T^{\dot{a}a}_p}, \\
\overline{\gen{Q}}{}_{\mbox{\tiny L} \dot{a}} \ket{\widetilde{\chi}^{a}_p}\;& = -\epsilon_{\dot{a}\dot{b}} \bar{a}^{\mbox{\tiny L}}_p \ket{T^{\dot{b}a}_p},
\qquad
&\overline{\gen{Q}}{}_{\mbox{\tiny L} \dot{a}} \ket{T^{\dot{b}a}_p} &= \delta_{\dot{a}}^{\ \dot{b}} \bar{a}^{\mbox{\tiny L}}_p \ket{\chi^a_p}, \\[4pt]
\gen{Q}_{\mbox{\tiny R} \dot{a}} \ket{T^{\dot{b}a}_p} &= \delta_{\dot{a}}^{\ \dot{b}} b^{\mbox{\tiny L}}_p \ket{\chi^a_p},
\qquad
&\gen{Q}_{\mbox{\tiny R} \dot{a}} \ket{\widetilde{\chi}^a_p} \;&= -\epsilon_{\dot{a}\dot{b}} b^{\mbox{\tiny L}}_p \ket{T^{\dot{b}a}_p}, \\
\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ \dot{a}} \ket{\chi^a_p}\;& = \bar{b}^{\mbox{\tiny L}}_p \ket{T^{\dot{a}a}_p},
\qquad
&\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ \dot{a}} \ket{T^{\dot{b}a}_p} &= \epsilon^{\dot{a}\dot{b}} \bar{b}^{\mbox{\tiny L}}_p \ket{\widetilde{\chi}^a_p}.
\end{aligned}
\end{equation}
Note also that the highest weight state of the massless representations are fermionic, namely $\ket{\chi^{a}}$, in contrast with the ones of the left and right representations, that are $\ket{Y^{\mbox{\tiny L}}}$ and $\ket{Z^{\mbox{\tiny R}}}$ respectively.
\begin{figure}
\begin{tikzpicture}[%
box/.style={outer sep=1pt},
Q node/.style={inner sep=1pt,outer sep=0pt},
arrow/.style={-latex}
]%
\begin{scope}[xshift=-4.2cm]
\node [box] (PhiM) at ( 0 , 2cm) { $\ket{\chi^{1}}$};
\node [box] (PsiP) at (-2cm, 0cm) { $\ket{T^{11}}$};
\node [box] (PsiM) at (+2cm, 0cm) { $\ket{T^{21}}$};
\node [box] (PhiP) at ( 0 ,-2cm) { $\ket{\widetilde{\chi}^{1}}$};
\newcommand{0.09cm,0cm}{0.09cm,0cm}
\newcommand{0cm,0.10cm}{0cm,0.10cm}
\draw [arrow] ($(PhiM.west) +(0cm,0.10cm)$) -- ($(PsiP.north)-(0.09cm,0cm)$) node [pos=0.5,anchor=south east,Q node] {\small $\gen{Q}^{\ 1}_{\mbox{\tiny L}},\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ 1}$};
\draw [arrow] ($(PsiP.north)+(0.09cm,0cm)$) -- ($(PhiM.west) -(0cm,0.10cm)$) node [pos=0.5,anchor=north west,Q node] {};
\draw [arrow] ($(PsiM.south)-(0.09cm,0cm)$) -- ($(PhiP.east) +(0cm,0.10cm)$) node [pos=0.5,anchor=south east,Q node] {};
\draw [arrow] ($(PhiP.east) -(0cm,0.10cm)$) -- ($(PsiM.south)+(0.09cm,0cm)$) node [pos=0.5,anchor=north west,Q node] {\small $\overline{\gen{Q}}{}_{\mbox{\tiny L} 1}, \gen{Q}_{\mbox{\tiny R} 1}$};
\draw [arrow] ($(PhiM.east) -(0cm,0.10cm)$) -- ($(PsiM.north)-(0.09cm,0cm)$) node [pos=0.5,anchor=north east,Q node] {};
\draw [arrow] ($(PsiM.north)+(0.09cm,0cm)$) -- ($(PhiM.east) +(0cm,0.10cm)$) node [pos=0.5,anchor=south west,Q node] {\small $\overline{\gen{Q}}{}_{\mbox{\tiny L} 2}, \gen{Q}_{\mbox{\tiny R} 2}$};
\draw [arrow] ($(PsiP.south)-(0.09cm,0cm)$) -- ($(PhiP.west) -(0cm,0.10cm)$) node [pos=0.5,anchor=north east,Q node] {\small $\gen{Q}^{\ 2}_{\mbox{\tiny L}},\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ 2}$};
\draw [arrow] ($(PhiP.west) +(0cm,0.10cm)$) -- ($(PsiP.south)+(0.09cm,0cm)$) node [pos=0.5,anchor=south west,Q node] {};
\draw [arrow] (PsiM) -- (PsiP) node [pos=0.65,anchor=south west,Q node] {\small $\gen{J}^{\ \dot{a}}_{\bullet}$};
\draw [arrow] (PsiP) -- (PsiM);
\draw[rounded corners=5mm] (-3.2cm,-2.6cm)rectangle (3.2cm,2.6cm);
\node (reprlabel) at ( 0 , -3cm) { $\varrho_{\mbox{\tiny L}}\otimes\widetilde{\varrho}_{\mbox{\tiny L}} \cong\varrho_{\mbox{\tiny R}}\otimes\widetilde{\varrho}_{\mbox{\tiny R}}$};
\end{scope}
\begin{scope}[xshift=0cm]
\draw [arrow] (-0.9cm,0cm) -- (0.9cm,0cm) node [Q node] at (0cm,0.22cm) {\small $\gen{J}^{\ a}_{\circ}$};
\draw [arrow] (0.9cm,0cm) -- (-0.9cm,0cm);
%
\end{scope}
\begin{scope}[xshift=4.2cm]
\node [box] (PhiM) at ( 0 , 2cm) { $\ket{\chi^{2}}$};
\node [box] (PsiP) at (-2cm, 0cm) { $\ket{T^{12}}$};
\node [box] (PsiM) at (+2cm, 0cm) { $\ket{T^{22}}$};
\node [box] (PhiP) at ( 0 ,-2cm) { $\ket{\widetilde{\chi}^{2}}$};
\newcommand{0.09cm,0cm}{0.09cm,0cm}
\newcommand{0cm,0.10cm}{0cm,0.10cm}
\draw [arrow] ($(PhiM.west) +(0cm,0.10cm)$) -- ($(PsiP.north)-(0.09cm,0cm)$) node [pos=0.5,anchor=south east,Q node] {\small $\gen{Q}^{\ 1}_{\mbox{\tiny L}},\overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ 1}$};
\draw [arrow] ($(PsiP.north)+(0.09cm,0cm)$) -- ($(PhiM.west) -(0cm,0.10cm)$) node [pos=0.5,anchor=north west,Q node] {};
\draw [arrow] ($(PsiM.south)-(0.09cm,0cm)$) -- ($(PhiP.east) +(0cm,0.10cm)$) node [pos=0.5,anchor=south east,Q node] {};
\draw [arrow] ($(PhiP.east) -(0cm,0.10cm)$) -- ($(PsiM.south)+(0.09cm,0cm)$) node [pos=0.5,anchor=north west,Q node] {\small $\overline{\gen{Q}}{}_{\mbox{\tiny L} 1}, \gen{Q}_{\mbox{\tiny R} 1}$};
\draw [arrow] ($(PhiM.east) -(0cm,0.10cm)$) -- ($(PsiM.north)-(0.09cm,0cm)$) node [pos=0.5,anchor=north east,Q node] {};
\draw [arrow] ($(PsiM.north)+(0.09cm,0cm)$) -- ($(PhiM.east) +(0cm,0.10cm)$) node [pos=0.5,anchor=south west,Q node] {\small $\overline{\gen{Q}}{}_{\mbox{\tiny L} 2}, \gen{Q}_{\mbox{\tiny R} 2}$};
\draw [arrow] ($(PsiP.south)-(0.09cm,0cm)$) -- ($(PhiP.west) -(0cm,0.10cm)$) node [pos=0.5,anchor=north east,Q node] {\small $\gen{Q}^{\ 2}_{\mbox{\tiny L}}, \overline{\gen{Q}}{}_{\mbox{\tiny R}}^{\ 2}$};
\draw [arrow] ($(PhiP.west) +(0cm,0.10cm)$) -- ($(PsiP.south)+(0.09cm,0cm)$) node [pos=0.5,anchor=south west,Q node] {};
\draw [arrow] (PsiM) -- (PsiP) node [pos=0.65,anchor=south west,Q node] {\small $\gen{J}^{\ \dot{a}}_{\bullet}$};
\draw [arrow] (PsiP) -- (PsiM);
\draw[rounded corners=5mm] (-3.2cm,-2.6cm)rectangle (3.2cm,2.6cm);
\node (reprlabel) at ( 0 , -3cm) { $\varrho_{\mbox{\tiny L}}\otimes\widetilde{\varrho}_{\mbox{\tiny L}} \cong\varrho_{\mbox{\tiny R}}\otimes\widetilde{\varrho}_{\mbox{\tiny R}}$};
\end{scope}
\draw[rounded corners=5mm] (-7.8cm,-3.6cm)rectangle (7.8cm,3.1cm);
\end{tikzpicture}
\caption{%
The massless excitations transform in two irreducible representations of~$\alg{psu}(1|1)^4_{\text{c.e.}}$, which form a doublet under~$\alg{su}(2)_{\circ}$. Each of these representation can equivalently be taken to be the massless limit of a left or a right representation with a fermionic highest-weight state. For definiteness, here we take both of them to be given by left representations.
Below each $\alg{psu}(1|1)^4_{\text{c.e.}}$ diagram we indicate how each representation can be obtained from one of two (left or right) isomorphic tensor products of fundamental $\alg{su}(1|1)^2_{\text{c.e.}}$ representations, see section~\ref{sec:repr-smallalgebra}.
}
\label{fig:repr:massless}
\end{figure}
\subsubsection{Representation coefficients for the pure R-R theory}
\label{sec:RR-repr-coeff}
In absence of NS-NS fluxes, the representations coefficients are
\begin{equation}
\begin{aligned}
\label{eq:RR-repr-coeff}
&a^{\mbox{\tiny L}}_p = a^{\mbox{\tiny R}}_p = \eta_p e^{i\xi}, \qquad&& \bar{a}^{\mbox{\tiny L}}_p = \bar{a}^{\mbox{\tiny R}}_p = \eta_p e^{-ip/2} e^{-i\xi},\\
&b^{\mbox{\tiny L}}_p = b^{\mbox{\tiny R}}_p = -\frac{\eta_p}{x^-_p} e^{-ip/2} e^{i\xi}, \qquad\qquad &&\bar{b}^{\mbox{\tiny L}}_p = \bar{b}^{\mbox{\tiny R}}_p = -\frac{\eta_p}{x^+_p} e^{-i\xi},
\end{aligned}
\end{equation}
with
\begin{equation}
\eta_p = e^{ip/4} \sqrt{\frac{ih}{2}(x^-_p - x^+_p)}.
\end{equation}
The equality of the left- and right-representation coefficients in equation~\eqref{eq:RR-repr-coeff} indicates that left-right symmetry is particularly simple in the pure-R-R case.
The Zhukovski variables~$x^{\pm}_p$ are mass-dependent:
\begin{equation}
\label{eq:zhukovski}
\frac{x^+_p}{x^-_p}=e^{ip},
\qquad
x^+_p +\frac{1}{x^+_p} -x^-_p -\frac{1}{x^-_p} = \frac{2i \, |m|}{h}.
\end{equation}
In fact, the dependence of the representation parameters on $m=\pm1,0$ is entirely encoded in $x^{\pm}_p$.
The phase~$\xi$ is irrelevant for the one-particle representation, but is instrumental for defining the two-particle representation, \textit{i.e.}\xspace in order to define a non-trivial coproduct~\cite{Borsato:2014hja}, similarly to what happens for $\text{AdS}_5\times\mathrm{S}^5$ strings~\cite{Arutyunov:2006yd}.
Let us note that while the excitations in the left and right modules have kinematic properties similar to the one of $\text{AdS}_5\times\mathrm{S}^5$ excitations, new features emerge in the massless case. When $m=0$, the Zhukovski variables satisfy the additional constraint
\begin{equation}
x^{+}_p=\frac{1}{x^{-}_p}.
\end{equation}
Moreover, the vanishing of $\gen{M}$ imposes that the representation coefficients satisfy
\begin{equation}
\label{eq:massless-rep-shortening1}
|a^{\mbox{\tiny L},\mbox{\tiny R}}_p|^2=|b^{\mbox{\tiny L},\mbox{\tiny R}}_p|^2,\qquad
\text{at}\quad m=0,
\end{equation}
and the dispersion relation becomes non-analytic,
\begin{equation}
E(p)=2h\,\Big|\sin\bigl(\frac{p}{2}\bigr)\Big|.
\end{equation}
This last property can be physically interpreted as an indication that left- and right-movers on the worldsheet should be treated as two different species of particles, similarly to what is done in the relativistic case.
\subsubsection{Representations of \texorpdfstring{$\alg{su}(1|1)^2_{\text{c.e.}}$}{su(1|1)**2 c.e.}}
\label{sec:repr-smallalgebra}
It is useful to introduce the $\alg{su}(1|1)^2_{\text{c.e.}}$ algebra, whose anticommutation relations are
\begin{equation}
\label{eq:cealgebra-small}
\begin{aligned}
&\{\gen{Q}_{\mbox{\tiny L}},\overline{\gen{Q}}{}_{\mbox{\tiny L}}\} = \frac{1}{2}\,(\gen{H}+\gen{M}),
&\qquad &\{\gen{Q}_{\mbox{\tiny L}},{\gen{Q}}{}_{\mbox{\tiny R}}\} = \gen{C},\\
&\{\gen{Q}_{\mbox{\tiny R}},\overline{\gen{Q}}{}_{\mbox{\tiny R}} \}= \frac{1}{2}\,(\gen{H}-\gen{M}),
&\qquad &\{\overline{\gen{Q}}{}_{\mbox{\tiny L}},\overline{\gen{Q}}{}_{\mbox{\tiny R}}\} = \overline{\gen{C}}.
\end{aligned}
\end{equation}
The short representations of this algebra are two-dimensional, and have been studied in ref.~\cite{Borsato:2012ud}. Once again, we have a left representation $\varrho_{\mbox{\tiny L}}$
\begin{equation}\label{eq:su(1|1)2-repr1}
\begin{aligned}
\gen{Q}_{\mbox{\tiny L}} \ket{\phi^{\mbox{\tiny L}}_p} &= a^{\mbox{\tiny L}}_p \ket{\psi^{\mbox{\tiny L}}_p} , \qquad &
\overline{\gen{Q}}{}_{\mbox{\tiny L}} \ket{\psi^{\mbox{\tiny L}}_p} &= \bar{a}^{\mbox{\tiny L}}_p \ket{\phi^{\mbox{\tiny L}}_p} , \\
\gen{Q}_{\mbox{\tiny R}} \ket{\psi^{\mbox{\tiny L}}_p} &= b^{\mbox{\tiny L}}_p \ket{\phi^{\mbox{\tiny L}}_p} , \qquad & \overline{\gen{Q}}{}_{\mbox{\tiny R}} \ket{\phi^{\mbox{\tiny L}}_p} &= \bar{b}^{\mbox{\tiny L}}_p \ket{\psi^{\mbox{\tiny L}}_p}
.
\end{aligned}
\end{equation}
Similarly, we can consider a right representation~$\varrho_{\mbox{\tiny R}} $
\begin{equation}\label{eq:su(1|1)2-reprR}
\begin{aligned}
\gen{Q}_{\mbox{\tiny L}} \ket{\psi_p^{\mbox{\tiny R}}} &= b^{\mbox{\tiny R}}_p \ket{\phi_p^{\mbox{\tiny R}}} , \qquad &
\overline{\gen{Q}}{}_{\mbox{\tiny L}} \ket{\phi_p^{\mbox{\tiny R}}} &= \bar{b}^{\mbox{\tiny R}}_p \ket{\psi_p^{\mbox{\tiny R}}} , \\
\gen{Q}_{\mbox{\tiny R}} \ket{\phi_p^{\mbox{\tiny R}}} &= a^{\mbox{\tiny R}}_p \ket{\psi_p^{\mbox{\tiny R}}} , \qquad &
\overline{\gen{Q}}{}_{\mbox{\tiny R}} \ket{\psi_p^{\mbox{\tiny R}}} &= \bar{a}^{\mbox{\tiny R}}_p \ket{\phi_p^{\mbox{\tiny R}}} .
\end{aligned}
\end{equation}
Two more representations, which we denote by $\widetilde{\varrho}_{\mbox{\tiny L}}$ and $\widetilde{\varrho}_{\mbox{\tiny R}}$, can be obtained from the ones above by exchanging bosons with fermions.
As discussed in detail in ref.~\cite{Borsato:2013qpa}, appropriate tensor products of pairs of these representations are isomorphic to the~$\alg{psu}(1|1)^4_{\text{c.e.}}$ representations discussed above. In fact, a similar structure will be present also in the mixed-flux case, and we will exploit it to write down the S~matrix.
Let us sketch these isomorphisms. Firstly, note that we can obtain~$\alg{psu}(1|1)^4_{\text{c.e.}}$ supercharges from those of the tensor products of~$\alg{su}(1|1)^2_{\text{c.e.}}$ by setting%
\footnote{This tensor-product structure is similar to the one of $\alg{psu}(2|2)^2_{\text{c.e.}}$, which is the off-shell symmetry algebra of $\text{AdS}_5\times\mathrm{S}^5$ superstrings~\cite{Arutyunov:2009ga,Beisert:2010jr}.}
\begin{equation}
\begin{aligned}
\gen{Q}_{\mbox{\tiny L}}^{\ 1} = \gen{Q}_{\mbox{\tiny L}} \otimes \mathbf{1} , \qquad
\gen{Q}_{\mbox{\tiny L}}^{\ 2} = \mathbf{1} \otimes \gen{Q}_{\mbox{\tiny L}}, \qquad
\gen{Q}_{\mbox{\tiny R} 1} = \gen{Q}_{\mbox{\tiny R}} \otimes \mathbf{1} , \qquad
\gen{Q}_{\mbox{\tiny R} 2} = \mathbf{1} \otimes \gen{Q}_{\mbox{\tiny R}} ,
\end{aligned}
\end{equation}
and similarly for their conjugates. Clearly then ${\varrho}_{\mbox{\tiny L}}\otimes {\varrho}_{\mbox{\tiny L}}$, ${\varrho}_{\mbox{\tiny R}}\otimes {\varrho}_{\mbox{\tiny R}}$, etc.\@ are representations of $\alg{psu}(1|1)^4_{\text{c.e.}}$. What is more, one can check that the left representation given in equation~\eqref{eq:repr-massive-L} is isomorphic to~${\varrho}_{\mbox{\tiny L}}\otimes {\varrho}_{\mbox{\tiny L}}$, while the right one~\eqref{eq:repr-massive-R} is isomorphic to~${\varrho}_{\mbox{\tiny R}}\otimes {\varrho}_{\mbox{\tiny R}}$. As for the two~$\alg{psu}(1|1)^4_{\text{c.e.}}$ modules that constitute the massless $\mathcal{A}$ module, each of them can be given either by ${\varrho}_{\mbox{\tiny L}}\otimes \widetilde{\varrho}_{\mbox{\tiny L}}$ or by ${\varrho}_{\mbox{\tiny R}}\otimes \widetilde{\varrho}_{\mbox{\tiny R}}$. This is consistent with the equivalence of left and right representations when $m=0$, and with the fact that the massless modules have fermionic highest-weight states.
The details of the isomorphisms outlined above are reviewed in appendix~\ref{app:tensorprod}.
\subsection{Representations in the near-plane-wave limit for the mixed-flux theory}
\label{sec:near-BMN-representations}
The near-plane-wave limit of the symmetry algebra~\cite{Gava:2002xb,Berenstein:2002jq} can be read off the explicit expression of the supercurrents~\eqref{eq:quadratic-currents}, truncated at quadratic order in the fields. This yields a representation on the fields and conjugate momenta $X, P, \eta, \bar{\eta}$, and so on. For our purposes it is more useful to work in terms of the excitations. To this end, we introduce creation and annihilation operators in the usual way. For the bosons, we schematically write
\begin{equation}
\begin{aligned}
a^{\dagger}(p)\approx \int\frac{\textup{d}\sigma}{\sqrt{\omega(p,m,q)}} \bigl(\omega(p,m,q)\,X-i P\bigr) e^{+ip\sigma},\\
a(p)\approx \int\frac{\textup{d}\sigma}{\sqrt{\omega(p,m,q)}} \bigl(\omega(p,m,q)\,X+i P\bigr) e^{-ip\sigma}.
\end{aligned}
\end{equation}
This representation depends on the energy~$\omega(p,m,q)$, which is function of the mass~$m$ and on the flux parameter~$q$. For the fermions we write
\begin{equation}
\begin{aligned}
d^{\dagger}(p)\approx \int\frac{\textup{d}\sigma}{\sqrt{\omega(p,m,q)}} \bigl(f(p,m,q)\,\eta-i g(p,m,q)\, \bar{\eta}\bigr) e^{+ip\sigma},\\
d(p)\approx \int\frac{\textup{d}\sigma}{\sqrt{\omega(p,m,q)}} \bigl(f(p,m,q)\,\eta+i g(p,m,q)\, \bar{\eta}\bigr) e^{-ip\sigma},
\end{aligned}
\end{equation}
where we introduced the wave-function parameters $f(p,m,q)$ and~$g(p,m,q)$. The creation operators generate the space of fundamental excitations, which as we reviewed consists of sixteen particles
\begin{equation}
\begin{gathered}
\ket{Z^{\mbox{\tiny L},\mbox{\tiny R}}}=a^\dagger_{\mbox{\tiny L},\mbox{\tiny R}\, z}\ket{0},\quad
\ket{Y^{\mbox{\tiny L},\mbox{\tiny R}}}=a^\dagger_{\mbox{\tiny L},\mbox{\tiny R}\, y}\ket{0},\quad
\ket{\eta^{\mbox{\tiny L} \dot{a}}}=d^{\ \dot{a} \dagger}_{\mbox{\tiny L}}\ket{0},\quad
\ket{\eta^{\mbox{\tiny R}}_{\ \dot{a}}}=d^{\dagger}_{\mbox{\tiny R} \dot{a}}\ket{0},\\
\ket{T^{\dot{a}a}}=a^{\dot{a} a\dagger}\ket{0},\qquad
\ket{\chi^{a}}=d^{a\,\dagger}\ket{0},\qquad
\ket{\widetilde{\chi}^{a}}=\tilde{d}^{a\,\dagger}\ket{0}.
\end{gathered}
\end{equation}
In appendix~\ref{app:charges:osc} we extract the supercharges from the supercurrents constructed in section~\ref{sec:off-shell-algebra}, which indeed gives an algebra of the form~\eqref{eq:cealgebra}. Furthermore, we rewrite them in terms of oscillators, obtaining
\begin{equation}
\label{eq:supercharges-leading-order-rep}
\begin{aligned}
&\gen{Q}_{\mbox{\tiny L}}^{\ {\dot{a}}}= \int \textup{d} p \ \Bigl[
(d_{\mbox{\tiny L}}^{\ {\dot{a}}\,\dagger} a_{\mbox{\tiny L} y} + \epsilon^{{\dot{a}\dot{b}}}\, a_{\mbox{\tiny L} z}^\dagger d_{\mbox{\tiny L} {\dot{b}}})f_p^{\mbox{\tiny L}}
+ (a_{\mbox{\tiny R} y}^\dagger d_{\mbox{\tiny R}}^{\ {\dot{a}}} +\epsilon^{{\dot{a}\dot{b}}}\, d_{\mbox{\tiny R} {\dot{b}}}^\dagger a_{\mbox{\tiny R} z})\,g_p^{\mbox{\tiny R}} \\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
+ \left( \epsilon^{{\dot{a}\dot{b}}}\, \tilde{d}^{{a}\,\dagger}a_{{\dot{b}a}}+a^{{\dot{a}a}\,\dagger}d_{{a}}\right)\tilde{f}_p
\Bigr],\\
&\gen{Q}_{\mbox{\tiny R} {\dot{a}}}=\int \textup{d} p \ \Bigl[
(d_{\mbox{\tiny R} {\dot{a}}}^\dagger a_{\mbox{\tiny R} y} -\epsilon_{{\dot{a}\dot{b}}}\, a_{\mbox{\tiny R} z}^\dagger d_{\mbox{\tiny R}}^{\ {\dot{b}}})f_p^{\mbox{\tiny R}}
+ (a_{\mbox{\tiny L} y}^\dagger d_{\mbox{\tiny L} {\dot{a}}} -\epsilon_{{\dot{a}\dot{b}}}\, d_{\mbox{\tiny L}}^{\ {\dot{b}}\,\dagger} a_{\mbox{\tiny L} z})\,g_p^{\mbox{\tiny L}}\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
+ \left( d^{{a}\,\dagger}a_{{\dot{a}a}}-\epsilon_{{\dot{a}\dot{b}}}\, a^{{\dot{b}a}\,\dagger}\tilde{d}_{{a}}\right)\tilde{g}_p
\Bigr],
\end{aligned}
\end{equation}
and similarly for their conjugates. Note that we suppressed the dependence of $f_p^{\mbox{\tiny L},\mbox{\tiny R}}, \tilde{f}_p$ and $g_p^{\mbox{\tiny L},\mbox{\tiny R}}, \tilde{g}_p$ on $m$ and $q$ for ease of notation.
This representation is indeed of the form~(\ref{eq:repr-massive-L}--\ref{eq:repr-massless}) up to fixing the representation coefficients, and closely resembles near-plane wave limit of the pure-R-R one discussed in ref.~\cite{Borsato:2014hja}. Note however that wave-function parameters are $f_p^{\mbox{\tiny L}}\neq f_p^{\mbox{\tiny R}}$ and $g_p^{\mbox{\tiny L}}\neq g_p^{\mbox{\tiny R}}$, unlike what happened in the pure-R-R case. This is reflected by the values of the central charges
\begin{equation}
\label{eq:MH-near-BMN}
\gen{M}=\begin{cases}
q p+1 &\text{left}, \\
q p-1 &\text{right}, \\
q p &\text{massless},
\end{cases}
\qquad\qquad
\gen{H}=\begin{cases}
\sqrt{\tilde{q}^2+(p+q)^2} &\text{left}, \\
\sqrt{\tilde{q}^2+(p-q)^2} &\text{right}, \\
\sqrt{p^2} &\text{massless},
\end{cases}
\end{equation}
while the off-shell central charges take the same form for all representations, and are both real
\begin{equation}
\gen{C}=\overline{\gen{C}}=-\frac{\tilde{q}}{2}\, p.
\end{equation}
This is consistent with the tree-level analysis of symmetries in the massive sector~\cite{Hoare:2013ida} and with the leading-order massless dispersion relation~\cite{Berenstein:2002jq}.
\subsubsection{Representation coefficients}
\label{sec:repr-paraman-nfs}
The wave-function parameters play the role of representation coefficients. In fact, comparing~\eqref{eq:supercharges-leading-order-rep} with the pure-R-R representations of section~\ref{sec:RRreprs}, we see that the real parameter $f_p$ should be the near-plane-wave limit of~$a_p$ and~$\bar{a}_p$, while~$g_p$ should be the limit of~$b_p$ and~$\bar{b}_p$. Therefore, the precise form of $f_p$ and $g_p$ will be important in order to fix $a_p, \bar{a}_p, b_p$ and $\bar{b}_p$ in the full theory.
Let us begin from the massive representations. We have
\begin{equation}
\begin{aligned}
\omega^{\mbox{\tiny L}}_p=\sqrt{\tilde{q}^2+(p+q)^2},
\qquad
f^{\mbox{\tiny L}}_p=\sqrt{\frac{1+q p+\omega^{\mbox{\tiny L}}_p}{2}},
\qquad
g^{\mbox{\tiny L}}_p=-\frac{\tilde{q}\,p}{2\,f^{\mbox{\tiny L}}_p},\\
\omega^{\mbox{\tiny R}}_p=\sqrt{\tilde{q}^2+(p-q)^2},
\qquad
f^{\mbox{\tiny R}}_p=\sqrt{\frac{1-q p+\omega^{\mbox{\tiny R}}_p}{2}},
\qquad
g^{\mbox{\tiny R}}_p=-\frac{\tilde{q}\,p}{2\,f^{\mbox{\tiny R}}_p}.
\end{aligned}
\end{equation}
where the energy is related to the representation parameters by~$\omega_p=f_p^2+g_p^2$.
The different sign in front of the $p$-linear terms is explained by the necessity of reproducing~\eqref{eq:MH-near-BMN}, and ultimately is a consequence of the fact that the NS-NS flux breaks parity invariance. In particular, this implies that LR symmetry will require a non-trivial map of the representation coefficients too.
For the massless representation we have
\begin{equation}
\tilde{\omega}_p\equiv \tilde{\omega}^{\mbox{\tiny L}}_p=\sqrt{p^2},
\qquad
\tilde{f}_p\equiv \tilde{f}^{\mbox{\tiny L}}_p=\sqrt{\frac{q p+\tilde{\omega}_p}{2}}
\qquad
\tilde{g}_p\equiv \tilde{g}^{\mbox{\tiny L}}_p=-\frac{\tilde{q}\,p}{2\,\tilde{f}_p}.
\end{equation}
This may appear troubling: we have argued in section~\ref{sec:RRreprs} that at least at~$q=0$ massless modes can be equivalently obtained from the left or right representation, and indeed this seems to be the case looking at~\eqref{eq:MH-near-BMN}. However, the values of $\tilde{f}_p$ and~$\tilde{g}_p$ come from a massless limit of $f^{\mbox{\tiny L}}_p$ and $g^{\mbox{\tiny L}}_p$. To see how this is inessential, let us define new massless parameters, now as limit of $f^{\mbox{\tiny R}}_p$ and $g^{\mbox{\tiny R}}_p$:
\begin{equation}
\tilde{\omega}^{\mbox{\tiny R}}_p=\sqrt{p^2},
\qquad
\tilde{f}^{\mbox{\tiny R}}_p=\sqrt{\frac{-q p+\tilde{\omega}^{\mbox{\tiny R}}_p}{2}}
\qquad
\tilde{g}^{\mbox{\tiny R}}_p=-\frac{\tilde{q}\,p}{2\,\tilde{f}^{\mbox{\tiny R}}_p}.
\end{equation}
Let us rescale \textit{e.g.}\xspace the massless fermion creation operators in \eqref{eq:supercharges-leading-order-rep} as
\begin{equation}
d_{a}\to \frac{\tilde{f}_p^{\mbox{\tiny L}}}{\tilde{g}^{\mbox{\tiny R}}_p}\, d_a,\quad
d^{a\,\dagger}\to \frac{\tilde{f}_p^{\mbox{\tiny L}}}{\tilde{g}^{\mbox{\tiny R}}_p}\,d^{a\,\dagger},\qquad
\tilde{d}_{a}\to -\frac{\tilde{g}^{\mbox{\tiny R}}_p}{\tilde{f}^{\mbox{\tiny L}}_p}\,\tilde{d}_a,\quad
\tilde{d}^{a\,\dagger}\to -\frac{\tilde{g}^{\mbox{\tiny R}}_p}{\tilde{f}^{\mbox{\tiny L}}_p}\,\tilde{d}^{a\,\dagger},
\end{equation}
and note the identities
\begin{equation}
(\tilde{f}_p^{\mbox{\tiny L}})^2=(\tilde{g}^{\mbox{\tiny R}}_p)^2,
\qquad
(\tilde{g}_p^{\mbox{\tiny L}})^2=(\tilde{f}^{\mbox{\tiny R}}_p)^2,
\qquad
\tilde{f}^{\mbox{\tiny L}}_p\,\tilde{f}^{\mbox{\tiny R}}_p= \tilde{g}^{\mbox{\tiny L}}_p\,\tilde{g}^{\mbox{\tiny R}}_p.
\end{equation}
In this way we replace everywhere the parameters $\tilde{f}_p\equiv\tilde{f}_p^{\mbox{\tiny L}}$ and $\tilde{g}_p\equiv\tilde{g}_p^{\mbox{\tiny L}}$ with $\tilde{f}_p^{\mbox{\tiny R}}$ and $\tilde{g}_p^{\mbox{\tiny R}}$, and in fact obtain the massless limit of a right representation. In summary, also in the mixed-flux case the massless representation can equivalently be described as a left or a right one, at least in the near-plane-wave limit. For definiteness, we will adopt the first choice.
\subsection{Exact representations for the mixed-flux theory}
\label{sec:exact-repr-mixed}
When we go beyond the near-plane-wave limit, we expect the representations discussed in the previous subsection to be deformed. In particular, as we have computed in section~\ref{sec:supercurrent-algebra}, the off-shell central charges will be non-linear functions of the worldsheet momentum,
\begin{equation}
\label{eq:Cexact}
\gen{C}=+\frac{ih}{2}(e^{+i\,\gen{P}}-1),
\qquad
\overline{\gen{C}}=-\frac{ih}{2}(e^{-i\,\gen{P}}-1).
\end{equation}
Here we introduce the mixed-flux coupling constant $h=h(\lambda,\tilde{q})$, which enters as an overall normalisation of the central charge. In the worldsheet calculation we found that for large $\sqrt{\lambda}$
\begin{equation}
h(\lambda,\tilde{q}) \approx \frac{\tilde{q} \sqrt{\lambda}}{2\pi}\, .
\end{equation}
However, this relation might receive perturbative and non-perturbative corrections in $1/\sqrt{\lambda}$, analogously to what happens for string theory in $\text{AdS}_4\times\mathrm{CP}^3$~\cite{Nishioka:2008gz,Gaiotto:2008cg}.\footnote{%
Recently a proposal has been made for the all-loop $\lambda$-dependence of the function $h(\lambda)$ in $\text{AdS}_4 \times \mathrm{CP}^3$~\cite{Gromov:2014eha}.
} %
Note that we have absorbed a factor $\tilde{q}$ into the definition of $h$. This makes~\eqref{eq:Cexact} take the same form as in the pure R-R case, but differs from the conventions of previous literature.
In our discussion of pure-R-R representations at the start of this section the eigenvalue of~$\gen{M}$ was a real number. However, from the string theory computations of section~\ref{sec:currents} we see that it should be a function of the total worldsheet momentum~$\gen{P}$. This may appear surprising as we expect~$\gen{M}$ to be a \emph{quantised} angular momentum on a physical state. As we detail in section~\ref{sec:mass-comments}, this can be achieved if we take
\begin{equation}
\label{eq:Mexact}
\gen{M}=m+\k\,\gen{P} ,
\end{equation}
where $m=\pm1,0$ depending on which representation we are considering, as in equation~\eqref{eq:MH-near-BMN}.
The constant $\k$ is related to the WZW level $k$ by
\begin{equation}
\k = \frac{k}{2\pi} = \frac{q\sqrt{\lambda}}{2\pi} \,.
\end{equation}
In this way, using the shortening condition~\eqref{eq:shortening}, we conclude that the all-loop dispersion relation is
\begin{equation}
\label{eq:all-loop-dispersion}
E(p)=\sqrt{(m+\k\,p)^2+4h^2\,\sin^2\bigl(\frac{p}{2}\bigr)} \,,
\end{equation}
which for massive particles confirms what was found by the analysis of giant magnons~\cite{Hoare:2013lja, Ahn:2014tua, Babichenko:2014yaa}. It is interesting to note that in the massless case the dispersion relation is non-analytic also at $\k \neq 0$. To make this evident, we write
\begin{equation}
\label{eq:all-loop-mless-dispersion}
E(p) = \k |p| \sqrt{ 1 + \frac{4 h^2 \sin^2\bigl(\frac{p}{2}\bigr)}{\k^2 p^2}} , \qquad
\text{at }m=0 ,
\end{equation}
where the square root is analytic when $p$ is in the vicinity of the real line.
\subsubsection{Exact representation parameters}
\label{sec:exact-repr-mixed:params}
We now want to construct three irreducible representations of $\mathcal{A}$ that in the limit $q\to0$ coincide with the pure-R-R ones which we recalled in section~\ref{sec:RRreprs}, and whose near-plane-wave limit is the one we computed in section~\ref{sec:near-BMN-representations}. To this end, it will be sufficient to suitably deform the representation coefficients of equations \eqref{eq:repr-massive-L}--\eqref{eq:repr-massless}. In particular, we define
\begin{equation}\label{eq:abparam}
\begin{aligned}
a^{\mbox{\tiny L}}_p &= \eta_p^{\mbox{\tiny L}}\, e^{i\xi},
&\quad
\bar{a}^{\mbox{\tiny L}}_p &= \eta_p^{\mbox{\tiny L}}\, e^{-ip/2} e^{-i\xi},
&\quad
b^{\mbox{\tiny L}}_p &= -\frac{\eta_p^{\mbox{\tiny L}}}{x^-_{\mbox{\tiny L}\, p}} e^{-ip/2} e^{i\xi},
&\quad
\bar{b}^{\mbox{\tiny L}}_p &= -\frac{\eta^{\mbox{\tiny L}}_p}{x^+_{\mbox{\tiny L}\,p}} e^{-i\xi},\\
a^{\mbox{\tiny R}}_p &= \eta_p^{\mbox{\tiny R}}\, e^{i\xi},
&\quad
\bar{a}^{\mbox{\tiny R}}_p &= \eta_p^{\mbox{\tiny R}}\, e^{-ip/2} e^{-i\xi},
&\quad
b^{\mbox{\tiny R}}_p &= -\frac{\eta_p^{\mbox{\tiny R}}}{x^-_{\mbox{\tiny R}\, p}} e^{-ip/2} e^{i\xi},
&\quad
\bar{b}^{\mbox{\tiny R}}_p &= -\frac{\eta^{\mbox{\tiny R}}_p}{x^+_{\mbox{\tiny R}\,p}} e^{-i\xi},
\end{aligned}
\end{equation}
with
\begin{equation}
\eta_p^{\mbox{\tiny L}} = e^{ip/4}\sqrt{\frac{ih}{2}(x^-_{\mbox{\tiny L}\, p} - x^+_{\mbox{\tiny L}\, p})} ,
\qquad
\eta_p^{\mbox{\tiny R}} = e^{ip/4}\sqrt{\frac{ih}{2}(x^-_{\mbox{\tiny R}\, p} - x^+_{\mbox{\tiny R}\, p})}.
\end{equation}
Here we have introduced two sets of Zhukovski variables~$x^{\pm}_{\mbox{\tiny L}\,p}$ and $x^{\pm}_{\mbox{\tiny R}\,p}$, which satisfy
\begin{equation}
\label{eq:zhukovski2}
\begin{gathered}
\frac{x^+_{\mbox{\tiny L}\,p}}{x^-_{\mbox{\tiny L}\,p}}=e^{ip},
\qquad
x^+_{\mbox{\tiny L}\,p} +\frac{1}{x^+_{\mbox{\tiny L}\,p}} -x^-_{\mbox{\tiny L}\,p} -\frac{1}{x^-_{\mbox{\tiny L}\,p}} = \frac{2i \, (|m|+\k \, p)}{h},\\
\frac{x^+_{\mbox{\tiny R}\,p}}{x^-_{\mbox{\tiny R}\,p}}=e^{ip},
\qquad
x^+_{\mbox{\tiny R}\,p} +\frac{1}{x^+_{\mbox{\tiny R}\,p}} -x^-_{\mbox{\tiny R}\,p} -\frac{1}{x^-_{\mbox{\tiny R}\,p}} = \frac{2i \, (|m|-\k\, p)}{h}.
\end{gathered}
\end{equation}
These equations can be solved by setting
\begin{equation}
\begin{gathered}
x^{\pm}_{\mbox{\tiny L}\,p}=\frac{(|m| + \k p)+\sqrt{(|m| + \k p)^2 + 4h^2 \sin^2(\frac{p}{2})}}{2h\sin(\frac{p}{2})}e^{\pm\frac{i}{2}p} ,\\
x^{\pm}_{\mbox{\tiny R}\,p}=\frac{(|m| - \k p)+\sqrt{(|m| - \k p)^2 + 4h^2 \sin^2(\frac{p}{2})}}{2h\sin(\frac{p}{2})}e^{\pm\frac{i}{2}p} ,
\end{gathered}
\end{equation}
as usual with~$m=\pm1,0$.
In this way, we reproduce the central charges~(\ref{eq:Cexact}--\ref{eq:Mexact}) and the dispersion relation~\eqref{eq:all-loop-dispersion}.
It is interesting to note that for the massless modes it is no longer true that $x^+=1/x^-$. This identity is replaced by
\begin{equation}
\label{eq:magic-relation}
x^{\pm}_{\mbox{\tiny L}\,p}=\frac{1}{x^{\mp}_{\mbox{\tiny R}\,p}}\qquad
\text{at}\quad m=0.
\end{equation}
Owing to this equality, we can check the following identities for the representation coefficients of the massless representation
\begin{equation}
\label{eq:massless-rep-shortening2}
a^{\mbox{\tiny L}}_p\,\bar{a}^{\mbox{\tiny R}}_p=b^{\mbox{\tiny L}}_p\,\bar{b}^{\mbox{\tiny R}}_p=a^{\mbox{\tiny R}}_p\,\bar{a}^{\mbox{\tiny L}}_p=b^{\mbox{\tiny R}}_p\,\bar{b}^{\mbox{\tiny L}}_p,\qquad
|a^{\mbox{\tiny L}}_p|^2=|b^{\mbox{\tiny R}}_p|^2,
\qquad
|a^{\mbox{\tiny R}}_p|^2=|b^{\mbox{\tiny L}}_p|^2,
\end{equation}
that generalise~\eqref{eq:massless-rep-shortening1}. Note that it is not true that \textit{e.g.}\xspace $|a^{\mbox{\tiny L}}_p|^2=|b^{\mbox{\tiny L}}_p|^2$ at $\k\neq 0$.
\subsubsection{Equivalent representations for massless modes}
\label{eq:equiv-reprs-mless}
In equation~\eqref{eq:repr-massless} we chose to describe all massless modes by the massless limit of left representations. In terms of $\alg{su}(1|1)^2_{\text{c.e.}}$ representations, this corresponds to describing the massless module as~$(\varrho_{\mbox{\tiny L}} \otimes\widetilde{\varrho}_{\mbox{\tiny L}})^{\oplus2}$. As discussed in detail in ref.~\cite{Borsato:2014hja} and as we briefly recalled, equivalent alternative description are $(\varrho_{\mbox{\tiny R}} \otimes\widetilde{\varrho}_{\mbox{\tiny R}})^{\oplus2}$, $(\varrho_{\mbox{\tiny L}} \otimes\widetilde{\varrho}_{\mbox{\tiny L}})\oplus (\varrho_{\mbox{\tiny R}} \otimes\widetilde{\varrho}_{\mbox{\tiny R}})$ or $(\varrho_{\mbox{\tiny R}} \otimes\widetilde{\varrho}_{\mbox{\tiny R}})\oplus(\varrho_{\mbox{\tiny L}} \otimes\widetilde{\varrho}_{\mbox{\tiny L}})$; all these representations are isomorphic in the massless limit at $\k=0$. Here we expect the same to hold, as \textit{a priori} there is no reason to prefer any of these choices to describe massless modes.
Let us perform the redefinition
\begin{equation}
\ket{\chi^{a}}\to\frac{a^{\mbox{\tiny L}}_p}{b^{\mbox{\tiny R}}_p}\ket{\chi^{a}},
\qquad
\ket{\widetilde{\chi}^{a}}\to-\frac{b^{\mbox{\tiny R}}_p}{a^{\mbox{\tiny L}}_p}\ket{\widetilde{\chi}^{a}}.
\end{equation}
Using~\eqref{eq:massless-rep-shortening2} in the defining relations~\eqref{eq:repr-massless} we find that, as a result, the representation $(\varrho_{\mbox{\tiny L}} \otimes\widetilde{\varrho}_{\mbox{\tiny L}})^{\oplus2}$ is indeed isomorphic to $(\varrho_{\mbox{\tiny R}} \otimes\widetilde{\varrho}_{\mbox{\tiny R}})^{\oplus2}$.
We can then obtain the mixed cases $(\varrho_{\mbox{\tiny L}} \otimes\widetilde{\varrho}_{\mbox{\tiny L}})\oplus (\varrho_{\mbox{\tiny R}} \otimes\widetilde{\varrho}_{\mbox{\tiny R}})$ or $(\varrho_{\mbox{\tiny R}} \otimes\widetilde{\varrho}_{\mbox{\tiny R}})\oplus(\varrho_{\mbox{\tiny L}} \otimes\widetilde{\varrho}_{\mbox{\tiny L}})$ by performing the rescaling only on $\ket{\chi^2},\ket{\widetilde{\chi}^2}$ or $\ket{\chi^1},\ket{\widetilde{\chi}^1}$ respectively. It is also interesting to note that the rescaling coefficient is just a sign:
\begin{equation}
\frac{a^{\mbox{\tiny L}}_p}{b^{\mbox{\tiny R}}_p}=-\text{sgn}\Bigl[\sin\bigl(\frac{p}{2}\bigr)\Bigr].
\end{equation}
\subsubsection{Two-particle representations}
\label{sec:coproduct}
So far, we have described the action of the symmetries on the one-particle representations. In order to construct the S~matrix, we will also need to consider two-particle representations. These can be constructed by introducing a \emph{deformed coproduct}~\cite{Plefka:2006ze}, or equivalently by appropriately picking the phase~$\xi$ in the one-particle representations~\cite{Arutyunov:2006yd}, \textit{cf.}\xspace equation \eqref{eq:abparam}. A way to find such a coproduct is to require that the central charges~$\gen{C},\overline{\gen{C}}$ vanish on physical two-particle states~\cite{Arutyunov:2006yd}, so that they should be
\begin{equation}
\label{eq:C12}
\gen{C}_{(12)}=+\frac{ih}{2}(e^{+i\gen{P}}-1),\qquad
\overline{\gen{C}}_{(12)}=-\frac{ih}{2}(e^{-i\gen{P}}-1),
\end{equation}
where $\gen{P}$ is the \emph{total} worldsheet momentum,
\begin{equation}
\gen{P}\ket{p_1,\dotsc,p_n}=(p_1+\dotsb+p_n)\ket{p_1,\dotsc,p_n}\,.
\end{equation}
This then enforces, in the same way as in refs.~\cite{Borsato:2013qpa,Borsato:2014hja}, that the supercharges are%
\footnote{%
It is possible to pick different coproduct that are related to this one by a momentum-dependent change of the two-particle basis, as discussed in ref.~\cite{Borsato:2013qpa}.
}
\begin{equation}
\begin{gathered}
\gen{Q}^{\ \dot{a}}_{\mbox{\tiny L}}{}_{(12)}(p,q)=\gen{Q}^{\ \dot{a}}_{\mbox{\tiny L}}(p_1)\otimes \mathbf{1} + e^{+\frac{i}{2}p}\Sigma\otimes\gen{Q}^{\ \dot{a}}_{\mbox{\tiny L}}(q)\,,\\
\gen{Q}_{\mbox{\tiny R}\dot{a}}{}_{(12)}(p,q)=\gen{Q}_{\mbox{\tiny R}\dot{a}}(p)\otimes \mathbf{1} + e^{+\frac{i}{2}p}\Sigma\otimes\gen{Q}_{\mbox{\tiny R}\dot{a}}(q)\,,\\
\overline{\gen{Q}}{}_{\mbox{\tiny L}\dot{a}}{}_{(12)}(p,q)=\overline{\gen{Q}}{}_{\mbox{\tiny L}\dot{a}}(p)\otimes \mathbf{1} + e^{-\frac{i}{2}p}\Sigma\otimes\overline{\gen{Q}}{}_{\mbox{\tiny L}\dot{a}}(q)\,,\\
\overline{\gen{Q}}{}^{\ \dot{a}}_{\mbox{\tiny R}}{}_{(12)}(p,q)=\overline{\gen{Q}}{}^{\ \dot{a}}_{\mbox{\tiny R}}(p)\otimes \mathbf{1} + e^{-\frac{i}{2}p}\Sigma\otimes\overline{\gen{Q}}{}^{\ \dot{a}}_{\mbox{\tiny R}}(q)\,,
\end{gathered}
\end{equation}
where~$\Sigma$ is the fermion-sign matrix taking values $+1$, $-1$ on bosons and fermions respectively. Consequently, on the central charges we have
\begin{equation}
\begin{gathered}
\gen{C}_{(12)}(p,q)=\gen{C}(p)\otimes\mathbf{1}+e^{+ip}\mathbf{1}\otimes\gen{C}(q)\,,\\
\overline{\gen{C}}{}_{(12)}(p,q)=\overline{\gen{C}}(p)\otimes\mathbf{1}+e^{-ip}\mathbf{1}\otimes\overline{\gen{C}}(q)\,,
\end{gathered}
\end{equation}
consistently with~\eqref{eq:C12}, and finally
\begin{equation}
\label{eq:trivialcoproduct}
\gen{H}_{(12)}=\gen{H}\otimes\mathbf{1}+\mathbf{1}\otimes\gen{H}\,,
\qquad
\gen{M}_{(12)}=\gen{M}\otimes\mathbf{1}+\mathbf{1}\otimes\gen{M}\,.
\end{equation}
Similarly, the coproduct is trivial for the $\alg{so}(4)_2$ generators.
\subsubsection{A momentum-dependent mass?}
\label{sec:mass-comments}
It may appear unnatural that $\gen{M}$ depends on the momentum of the excitations, since in the algebra of superisometries it had the interpretation of an angular momentum.
The resolution of this apparent contradiction is recalling that $\gen{M}$ is supposed to be identified with an isometry for a \emph{physical state}, \textit{i.e.}\xspace, on shell. This means that we should expect $\gen{M}$ to be integer-valued only when applied to states that satisfy the level-matching condition~\eqref{eq:level-matching}.
Let us rewrite~\eqref{eq:Mexact} in term of the integer level of the WZW term in the string action~$k = 2\pi\k \in\mathbb{Z}$. For a one-particle state we then have
\begin{equation}
\label{eq:Mwinding}
\gen{M}=m\mathbf{1}+\frac{k}{2\pi}\gen{P}\,.
\end{equation}
A physical state has worldsheet momentum~$2\pi w$, where $w\in\mathbb{Z}$ is the winding number. This shows that~$\gen{M}$ is integer on shell \emph{even for states with non-trivial winding}.
Note that linearity in the worldsheet momentum~$\gen{P}$ is crucial to extend this property to any physical \emph{multi-particle} state. In fact in the off-shell algebra the coproduct of~$\gen{M}$ \eqref{eq:trivialcoproduct} remains undeformed, so that its action on a multi-particle states is just additive. Any non-linear function in equation~\eqref{eq:Mwinding} would have prevented us from rewriting the eigenvalue of~$\gen{M}$ on a multiparticle state in terms of the \emph{total} worldsheet momentum, which is what is quantised on shell.%
\footnote{%
It is interesting to note that the winding number affects the ``mass'' of excitations, so that \textit{e.g.}\xspace when $k=w=1$ a right-moving excitation has the kinematics of a massless one. It would be interesting to understand if this has deeper implications, which may require analysing in more detail the complete bound-state spectrum of the theory.}
The quantisation of the angular momentum $\gen{M}$ explains why we have introduced the two coupling constants $h$ and $\k$, even though they are both proportional to the string tension $\sqrt{\lambda}/2\pi$ to leading order at strong coupling. According to the above discussion, the quantisation of the momentum-dependent term in the mass follows from the fact that the WZW coupling is integer valued, a relation that should not get any quantum corrections.
The coupling $h$, on the other hand, appears as an overall factor in front of the central charge $\gen{C}$ and is expected to receive corrections at higher orders in $1/\sqrt{\lambda}$.
It is also interesting to see how the momentum-dependence is compatible with the other symmetries of the theory. In ref.~\cite{Borsato:2014hja} it was argued that~$\gen{M}$ could not receive quantum correction without spoiling either the~$\alg{su}(2)_{\circ}$ symmetry or crossing invariance. Let us see how that argument works in the present setting. Invariance under~$\alg{su}(2)_{\circ}$ dictates that $\gen{M}$ takes the same value on both~$\alg{psu}(1|1)^4_{\text{c.e.}}$ massless modules. If we write $\gen{M}$ in a block-matrix form, with each block corresponding to a $\alg{psu}(1|1)^4_{\text{c.e.}}$ module\footnote{Respectively, $\varrho_{\mbox{\tiny L}}\otimes\varrho_{\mbox{\tiny L}}$, $\varrho_{\mbox{\tiny R}}\otimes\varrho_{\mbox{\tiny R}}$ and the massless doublet $(\varrho_{\mbox{\tiny L}} \otimes\widetilde{\varrho}_{\mbox{\tiny L}})^{\oplus 2}$.}, we see that this indeed the case:
\begin{equation}
\gen{M}=\left(
\begin{array}{cccc}
+1+\k p & 0 & 0 & 0\\
0 & -1+\k p & 0 & 0\\
0 & 0 & \k p & 0 \\
0 & 0 & 0 & \k p
\end{array}\right).
\end{equation}
On the other hand, under a crossing transformation it should be possible to map every irreducible module to some other one for which $\gen{M}$ has an opposite sign. We can see that at $\k=0$ this means sending right to left movers, and massless modes to themselves. If we perform the crossing transformation at $\k>0$ we must account for the fact that $p$ flips sign, \textit{i.e.}\xspace
\begin{equation}
\gen{M}=\left(
\begin{array}{cccc}
+1-\k p & 0 & 0 & 0\\
0 & -1-\k p & 0 & 0\\
0 & 0 & -\k p & 0 \\
0 & 0 & 0 & -\k p
\end{array}\right),\qquad
{\text{at crossed }p}.
\end{equation}
This shows that indeed even at $\k>0$ one can implement crossing by swapping left and right movers and sending the massless modes to themselves. The condition for this to be possible is that the eigenvalue of~$\gen{M}$ in the massless sector is an odd function of~$\gen{P}$. This in particular rules out a constant correction to the mass.
These general considerations on the momentum-dependence of~$\gen{M}$ fit together nicely with our analysis of the $x^-$-dependence in the supercharges, which constrains the non-local coproduct to take the form discussed in section~\ref{sec:coproduct}. They are also consistent with the form of the dispersion relation found by studying semi-classical solutions~\cite{Hoare:2013lja, Ahn:2014tua, Babichenko:2014yaa} and with the analysis of the possible spectrum of bound states performed in ref.~\cite{Hoare:2013lja}.
\section{S matrix}
\label{sec:smat}
Our discussion of the S~matrix of fundamental particles for the mixed-flux backgrounds will be based on the one done in ref.~\cite{Borsato:2014hja} in the pure-R-R case. As we have seen, the particle content is the same in the two theories, and the symmetry representations of our case of interest are a deformation of the ones of~\cite{Borsato:2014hja}.
We define the S matrix as the operator~$\mathcal{S}_{(12)}(p,q)$ acting on the two-particle Hilbert space and relating in- and out-states as
\begin{equation}
\mathcal{S}_{(12)}(p,q)\;\ket{\mathcal{X}_p^{(\text{in})} \mathcal{Y}_q^{(\text{in})}}= \ket{\mathcal{Y}_q^{(\text{out})} \mathcal{X}_p^{(\text{out})}} ,
\end{equation}
where $\mathcal{X}_p^{(\text{in})}, \mathcal{Y}_q^{(\text{in})}$ are two arbitrary excitations and $\mathcal{X}_p^{(\text{out})}, \mathcal{Y}_q^{(\text{out})}$ are the product of their scattering---possibly a linear combination of several two-particle states.
For this S~matrix to be physical and for the underlying theory to be integrable, several requirements should be satisfied. The most obvious is the invariance of~$\mathcal{S}$ under all symmetries of the theory
\begin{equation}\label{eq:S-mat-invariance}
\mathcal{S}_{(12)}(p,q)\;\gen{Q}_{(12)}(p,q)= \gen{Q}_{(12)}(q,p)\;\mathcal{S}_{(12)}(p,q).
\end{equation}
Here~$\gen{Q}_{(12)}(p,q)$ is any (super)charge of~$\mathcal{A}$, acting on a two-particle state. Note that we impose commutation with the \emph{off-shell} symmetries as~$\mathcal{S}_{(12)}(p,q)$ acts on particles that generally do not satisfy the level-matching condition. Next, we require \emph{braiding} and \emph{physical unitarity}, which read
\begin{equation}
\label{eq:Smat-symmetry}
\mathcal{S}_{(12)}(q,p)\;\mathcal{S}_{(12)}(p,q)= \mathbf{1} ,
\qquad
\Bigl(\mathcal{S}_{(12)}(p,q)\Bigr)^{\dagger}\;\mathcal{S}_{(12)}(p,q)= \mathbf{1} .
\end{equation}
We also impose the \emph{Yang-Baxter equation} on the three-particle Hilbert space
\begin{equation}
\mathcal{S}_{(12)}(q,r)\;
\mathcal{S}_{(23)}(p,r)\;
\mathcal{S}_{(12)}(p,q)
=
\mathcal{S}_{(23)}(p,q)\;
\mathcal{S}_{(12)}(p,r)\;
\mathcal{S}_{(23)}(q,r),
\end{equation}
which ensures that factorised scattering can be consistently defined.
We will find that the Yang-Baxter equation \emph{automatically holds} for the $\alg{psu}(1|1)^{4}_{\text{c.e.}}$ invariant S~matrices, signalling that this is a good candidate to be an integrable theory.
Lastly, there is the requirement of invariance under the crossing transformation. We will come back to this in section~\ref{sec:crossing}.
We will start by briefly recalling the form of some invariant matrices which will then be useful to restrict the form of~$\mathcal{S}$ by means of~\eqref{eq:Smat-symmetry}. Since our charges take the same form as the ones in ref.~\cite{Borsato:2014hja} up to suitably redefining the Zhukovski variables~$x^{\pm}$, we expect the final result to be closely related to the one found there. We will see that this is the case, even if there are some new features here. Imposing~\eqref{eq:Smat-symmetry} will fix the S~matrix up to some \emph{dressing factors}, which we will discuss in section~\ref{sec:crossing}.
\subsection{Invariant ~\texorpdfstring{$\alg{su}(1|1)^2_{\text{c.e.}}$}{su(1|1)**2 c.e.} S matrices}
\label{sec:small-smat}
We start by considering operators that are invariant under $\alg{su}(1|1)^2_{\text{c.e.}}$, which were first studied in ref.~\cite{Borsato:2012ud}. Since in this case the representations are much smaller, the resulting S~matrices are more manageable. Additionally, if we can define \textit{e.g.}\xspace a matrix~$\mathcal{S}^{\mbox{\tiny L}\sL}$ which commutes with all the generators of the two-particle~$\varrho_{\mbox{\tiny L}}$ representation, we are guaranteed that~$\mathcal{S}^{\mbox{\tiny L}\sL}\otimes\mathcal{S}^{\mbox{\tiny L}\sL}$ will commute with all generators of the two-particle~$\varrho_{\mbox{\tiny L}}\otimes\varrho_{\mbox{\tiny L}}$ one, which is one of the $\alg{psu}(1|1)^4_{\text{c.e.}}$ representations which will be of interest to us. Clearly the same holds for all representations we need to consider.
\subsubsection{Same target-space-chirality scattering}
Let us start from the case where we have two excitations in the representation~$\varrho_{\mbox{\tiny L}}$. In~\cite{Borsato:2012ud} it was found that the invariant S~matrix takes the form
\begin{equation}\label{eq:Smat-LL-small}
\begin{aligned}
\mathcal{S}^{\mbox{\tiny L}\sL} \ket{\phi_p^{\mbox{\tiny L}} \phi_q^{\mbox{\tiny L}}} &= A_{pq}^{\mbox{\tiny L}\sL} \ket{\phi_q^{\mbox{\tiny L}} \phi_p^{\mbox{\tiny L}}},
\qquad
&\mathcal{S}^{\mbox{\tiny L}\sL} \ket{\phi_p^{\mbox{\tiny L}} \psi_q^{\mbox{\tiny L}}} &= B_{pq}^{\mbox{\tiny L}\sL} \ket{\psi_q^{\mbox{\tiny L}} \phi_p^{\mbox{\tiny L}}} + C_{pq}^{\mbox{\tiny L}\sL} \ket{\phi_q^{\mbox{\tiny L}} \psi_p^{\mbox{\tiny L}}}, \\
\mathcal{S}^{\mbox{\tiny L}\sL} \ket{\psi_p^{\mbox{\tiny L}} \psi_q^{\mbox{\tiny L}}} &= F_{pq}^{\mbox{\tiny L}\sL} \ket{\psi_q^{\mbox{\tiny L}} \psi_p^{\mbox{\tiny L}}},\qquad
&\mathcal{S}^{\mbox{\tiny L}\sL} \ket{\psi_p^{\mbox{\tiny L}} \phi_q^{\mbox{\tiny L}}} &= D_{pq}^{\mbox{\tiny L}\sL} \ket{\phi_q^{\mbox{\tiny L}} \psi_p^{\mbox{\tiny L}}} + E_{pq}^{\mbox{\tiny L}\sL} \ket{\psi_q^{\mbox{\tiny L}} \phi_p^{\mbox{\tiny L}}}.
\end{aligned}
\end{equation}
This is the case also for us, with the ratio of the S-matrix elements being a function of~$x^{\pm}_{\mbox{\tiny L}\,p}$ and $x^{\pm}_{\mbox{\tiny L}\,q}$, and of course the overall normalisation being arbitrary. We collect the expressions for these coefficients in appendix~\ref{app:smat-param}. It is interesting to note that those expressions depend on~$h,\k$ and~$m$ only through the Zhukovski variables.
From~\eqref{eq:Smat-LL-small} we can immediately find the invariant S~matrix describing the scattering of \textit{e.g.}\xspace two particles that both are in the~$\widetilde{\varrho}_{\mbox{\tiny L}}$ representation. In fact, since the representations~${\varrho}_{\mbox{\tiny L}}$ and $\widetilde{\varrho}_{\mbox{\tiny L}}$ are related by a change of basis, we will have that
\begin{equation}
\label{eq:Smat-tilde-def}
\mathcal{S}^{\tilde{\text{\tiny L}}\stL}=\Pi^{g}\; \mathcal{O}^{-1}\;\Pi^{g}\;\mathcal{S}^{\mbox{\tiny L}\sL}\;\mathcal{O} ,
\end{equation}
where $\mathcal{O}=\mathcal{O}^{-1}=\sigma_1\otimes\sigma_1$ is the change-of-basis matrix and $\Pi^{g}$ is the graded permutation that accounts for the fermion signs. Up to suitably choosing~$\mathcal{O}$, this also yields~$\mathcal{S}^{\tilde{\text{\tiny L}}\mbox{\tiny L}}$ and~$\mathcal{S}^{\mbox{\tiny L}\tilde{\text{\tiny L}}}$.%
\footnote{%
The explicit form of the matrices $\mathcal{S}^{\tilde{\text{\tiny L}}\stL},\mathcal{S}^{\tilde{\text{\tiny L}}\mbox{\tiny L}}$ and~$\mathcal{S}^{\mbox{\tiny L}\tilde{\text{\tiny L}}}$ is also spelled out in ref.~\cite{Borsato:2014hja}.
}
The case of $\mathcal{S}^{\mbox{\tiny R}\sR}$, \textit{i.e.}\xspace both particles being in the representation~$\varrho_{\mbox{\tiny R}}$, is similar and in fact follows from the previous one by left-right symmetry. All we have to do is relabel everywhere~L$\to$R and introduce new scattering elements~$A^{\mbox{\tiny R}\sR}_{pq}, B^{\mbox{\tiny R}\sR}_{pq}$, etc. These will now depend on $x^{\pm}_{\mbox{\tiny R}\,p}$ and~$x^{\pm}_{\mbox{\tiny R}\,q}$. In a similar way, $\mathcal{S}^{\tilde{\text{\tiny R}}\mbox{\tiny R}}, \mathcal{S}^{\mbox{\tiny R}\tilde{\text{\tiny R}}}$ and $\mathcal{S}^{\tilde{\text{\tiny R}}\stR}$ can be easily found.
\subsubsection{Opposite target-space-chirality scattering}
Let us now consider the case where one particle transforms in~$\varrho_{\mbox{\tiny L}}$ and one transforms in~$\varrho_{\mbox{\tiny R}}$. At $\k=0$, such a set-up gives a scattering process of the form~\cite{Borsato:2012ud, Borsato:2013qpa}
\begin{equation}
\mathcal{S} \ket{\mathcal{X}^{\mbox{\tiny L}}_p\mathcal{Y}^{\mbox{\tiny R}}_q}= T_{pq} \ket{\mathcal{Y}^{\mbox{\tiny R}}_q\mathcal{X}^{\mbox{\tiny L}}_p} + R_{pq} \ket{\mathcal{Y}^{\mbox{\tiny L}}_q\mathcal{X}^{\mbox{\tiny R}}_p} , \end{equation}
where $T_{pq}$ is the transmission amplitude and $R_{pq}$ is the reflection one. Then, imposing LR-symmetry and unitarity requires either amplitude to vanish, and comparison with perturbative calculations sets $R_{pq}=0$.
On the other hand when $\k\neq 0$ we have that
\begin{equation}
\gen{H}\ket{\mathcal{X}^{\mbox{\tiny L}}_p\mathcal{Y}^{\mbox{\tiny R}}_q} \neq
\gen{H}\ket{\mathcal{X}^{\mbox{\tiny R}}_p\mathcal{Y}^{\mbox{\tiny L}}_q} ,
\end{equation}
which immediately sets~$R_{pq}=0$ when imposing~\eqref{eq:S-mat-invariance} for the Hamiltonian. This is an additional \textit{a posteriori} validation of the choice of a pure-transmission S~matrix originally made in ref.~\cite{Borsato:2012ud}.
We can therefore write down the matrix~$\mathcal{S}^{\mbox{\tiny L}\mbox{\tiny R}}$ as
\begin{equation}\label{eq:su(1|1)2-Smat-LRgrad1}
\begin{aligned}
\mathcal{S}^{\mbox{\tiny L}\mbox{\tiny R}} \ket{\phi^{\mbox{\tiny L}}_p \phi^{\mbox{\tiny R}}_q} &= A^{\mbox{\tiny L}\mbox{\tiny R}}_{pq} \ket{\phi^{\mbox{\tiny R}}_q \phi^{\mbox{\tiny L}}_p} + B^{\mbox{\tiny L}\mbox{\tiny R}}_{pq} \ket{\psi^{\mbox{\tiny R}}_q \psi^{\mbox{\tiny L}}_p}, \qquad
&\mathcal{S}^{\mbox{\tiny L}\mbox{\tiny R}} \ket{\phi^{\mbox{\tiny L}}_p \psi^{\mbox{\tiny R}}_q} &= C^{\mbox{\tiny L}\mbox{\tiny R}}_{pq} \ket{\psi^{\mbox{\tiny R}}_q \phi^{\mbox{\tiny L}}_p} , \\
\mathcal{S}^{\mbox{\tiny L}\mbox{\tiny R}} \ket{\psi^{\mbox{\tiny L}}_p \psi^{\mbox{\tiny R}}_q} &= E^{\mbox{\tiny L}\mbox{\tiny R}}_{pq} \ket{\psi^{\mbox{\tiny R}}_q \psi^{\mbox{\tiny L}}_p}+F^{\mbox{\tiny L}\mbox{\tiny R}}_{pq} \ket{\phi^{\mbox{\tiny R}}_q \phi^{\mbox{\tiny L}}_p} , \qquad
& \mathcal{S}^{\mbox{\tiny L}\mbox{\tiny R}} \ket{\psi^{\mbox{\tiny L}}_p \phi^{\mbox{\tiny R}}_q} &= D^{\mbox{\tiny L}\mbox{\tiny R}}_{pq} \ket{\phi^{\mbox{\tiny R}}_q \psi^{\mbox{\tiny L}}_p} .
\end{aligned}
\end{equation}
Here the S-matrix elements are functions of~$x^{\pm}_{\mbox{\tiny L}\,p}$ and~$x^{\pm}_{\mbox{\tiny R}\,q}$, see appendix~\ref{app:smat-param}.
Just as before, we can use changes of basis such as~\eqref{eq:Smat-tilde-def} to write down $\mathcal{S}^{\mbox{\tiny L}\tilde{\text{\tiny R}}},\mathcal{S}^{\tilde{\text{\tiny L}}\mbox{\tiny R}}$ and~$\mathcal{S}^{\tilde{\text{\tiny L}}\tilde{\text{\tiny R}}}$.
Finally, we can once more use LR symmetry to write down~$\mathcal{S}^{\mbox{\tiny R}\mbox{\tiny L}},\mathcal{S}^{\tilde{\text{\tiny R}}\mbox{\tiny L}},\mathcal{S}^{\mbox{\tiny R}\tilde{\text{\tiny L}}}$ and $\mathcal{S}^{\tilde{\text{\tiny R}}\tilde{\text{\tiny L}}}$. Due to the relabelling L$\leftrightarrow$R, these will all depend on~$x^{\pm}_{\mbox{\tiny R}\,p}$ and~$x^{\pm}_{\mbox{\tiny L}\,q}$.
\subsubsection{Tensor-product structure}
As we have argued, the tensor product of any pair of S~matrices invariant under~$\alg{su}(1|1)^2_{\text{c.e.}}$ will yield a $\alg{psu}(1|1)^4_{\text{c.e.}}$-invariant S~matrix in a given representation. Let us consider a pair of particles, transforming in two $\alg{psu}(1|1)^4_{\text{c.e.}}$ representations which we call $\varrho_{X_1}\otimes\varrho_{Y_1}$ and $\varrho_{X_2}\otimes\varrho_{Y_2}$ respectively, where $X_i$ and $Y_i$ could be L, R, $\tilde{\text{L}}$ or $\tilde{\text{R}}$. The $\alg{psu}(1|1)^4_{\text{c.e.}}$-invariant S matrix will be given by the tensor product of two $\alg{su}(1|1)^2_{\text{c.e.}}$ S~matrices. In formulae,
\begin{equation}
\mathcal{S}_{\alg{psu}(1|1)^4}=\mathcal{S}^{X_1 X_2}_{\alg{su}(1|1)^2}\,\check{\otimes}\;\mathcal{S}^{Y_1 Y_2}_{\alg{su}(1|1)^2}.
\end{equation}
Note that we have to account for signs arising from swapping fermionic excitations. To this end we define the \emph{graded} tensor product~$\check{\otimes}$, given by
\begin{equation}
\left( \mathcal{A}\,\check{\otimes}\,\mathcal{B} \right)_{MM',NN'}^{KK',LL'} = (-1)^{\epsilon_{M'}\epsilon_{N}+\epsilon_{L}\epsilon_{K'}} \ \mathcal{A}_{MN}^{KL} \ \mathcal{B}_{M'N'}^{K'L'} ,
\end{equation}
where~$\epsilon=0$ for bosons and $\epsilon=1$ for fermions.
\subsection{Matrix part of~\texorpdfstring{$\mathcal{S}$}{S}}
Let us decompose the S~matrix in different sectors, depending on the mass of the incoming particles, which clearly will be conserved during the scattering. We denote by $\mathcal{S}^{\bullet\bullet}$ the sector of massive excitations, $\mathcal{S}^{\circ\circ}$ the one of massless excitations, and by $\mathcal{S}^{\bullet\circ},\mathcal{S}^{\circ\bullet}$ the S-matrix blocks scattering particles of mixed~mass.
\subsubsection{Massive sector}
We can decompose the massive sector depending on whether the incoming particles have left or right target-space chirality. According to the discussion in the previous section, the scattering in each block should be given by the graded tensor product of two $\alg{su}(1|1)^2_{\text{c.e.}}$-invariant S~matrices, and indeed
\begin{equation}
\mathcal{S}^{\bullet\bullet} =
\left(
\begin{array}{cc}
\sigma^{\bullet\bullet}_{\mbox{\tiny L}\sL}\; \mathcal{S}^{\mbox{\tiny L}\sL}\,\check{\otimes}\,\mathcal{S}^{\mbox{\tiny L}\sL} &
\widetilde{\sigma}^{\bullet\bullet}_{\mbox{\tiny R}\mbox{\tiny L}}\; \mathcal{S}^{\mbox{\tiny R}\mbox{\tiny L}}\,\check{\otimes}\,\mathcal{S}^{\mbox{\tiny R}\mbox{\tiny L}}\\\\
\widetilde{\sigma}^{\bullet\bullet}_{\mbox{\tiny L}\mbox{\tiny R}}\;\mathcal{S}^{\mbox{\tiny L}\mbox{\tiny R}}\,\check{\otimes}\,\mathcal{S}^{\mbox{\tiny L}\mbox{\tiny R}} &
\sigma^{\bullet\bullet}_{\mbox{\tiny R}\sR}\;\mathcal{S}^{\mbox{\tiny R}\sR}\,\check{\otimes}\,\mathcal{S}^{\mbox{\tiny R}\sR}\\
\end{array}
\right).
\end{equation}
Note here that we are writing down four undetermined factors, rather than the two (same-chirality and opposite-chirality) that we would have in the pure-R-R case. Still, these are related pairwise by left-right symmetry, so we take
\begin{equation}
\begin{gathered}
\sigma^{\bullet\bullet}_{\mbox{\tiny L}\sL}(p,q)=\sigma^{\bullet\bullet}(x^{\pm}_{p\,\mbox{\tiny L}},x^{\pm}_{q\,\mbox{\tiny L}}),
\qquad
\sigma^{\bullet\bullet}_{\mbox{\tiny R}\sR}(p,q)=\sigma^{\bullet\bullet}(x^{\pm}_{p\,\mbox{\tiny R}},x^{\pm}_{q\,\mbox{\tiny R}}),\\
\widetilde{\sigma}^{\bullet\bullet}_{\mbox{\tiny L}\mbox{\tiny R}}(p,q)=\widetilde{\sigma}^{\bullet\bullet}(x^{\pm}_{p\,\mbox{\tiny L}},x^{\pm}_{q\,\mbox{\tiny R}}),
\qquad
\widetilde{\sigma}^{\bullet\bullet}_{\mbox{\tiny R}\mbox{\tiny L}}(p,q)=\widetilde{\sigma}^{\bullet\bullet}(x^{\pm}_{p\,\mbox{\tiny R}},x^{\pm}_{q\,\mbox{\tiny L}}),
\end{gathered}
\end{equation}
where~$\sigma^{\bullet\bullet}$ and~$\widetilde{\sigma}^{\bullet\bullet}$ are two appropriately defined functions.
\subsubsection{Massless sector}
In the massless sector we have two irreducible representations of $\alg{psu}(1|1)^4_{\text{c.e.}}$ that form a doublet under~$\alg{su}(2)_{\circ}$. For this reason the S~matrix here is the tensor product of a $\alg{psu}(1|1)^4_{\text{c.e.}}$-invariant one with an $\alg{su}(2)$-invariant pre-factor:
\begin{equation}
\mathcal{S}^{\circ\circ}=\sigma^{\circ\circ}\;\mathcal{S}_{\alg{su}(2)}\otimes \Bigl(\mathcal{S}^{\mbox{\tiny L}\sL} \,\check{\otimes}\, \mathcal{S}^{\tilde{\text{\tiny L}}\stL}\Bigr),
\end{equation}
where
\begin{equation}
\mathcal{S}_{\alg{su}(2)}=\frac{1}{1+\varsigma_{pq}} \bigl(\mathbf{1}+\varsigma_{pq}\Pi\bigr).
\end{equation}
In fact, the Yang-Baxter equation implies that $\mathcal{S}_{\alg{su}(2)}$ should be precisely the S~matrix of the Heisenberg model, \textit{i.e.}\xspace
\begin{equation}
\varsigma(p,q)=i\,(w_p-w_q) ,
\end{equation}
where $w_p$ is an appropriate rapidity.
This all follows closely what was found in refs.~\cite{Borsato:2014exa,Borsato:2014hja}. However, a few differences emerge at~$\k\neq 0$. Firstly, since now~$x^+\neq 1/x^-$ at $m=0$, the kinematics is richer. Consequently, scattering processes that accidentally had the same amplitude at $\k =m=0$ may now differ. For example
\begin{equation}
\Bra{\chi_q^{c}\chi_p^{d}}\mathcal{S} \Ket{\chi_p^{a}\chi_q^{b}}=
\Bra{\widetilde{\chi}_q^{c}\widetilde{\chi}_p^{d}}\mathcal{S} \Ket{\widetilde{\chi}_p^{a}\widetilde{\chi}_q^{b}}
\qquad\text{only at }\k=0 ,
\end{equation}
while at $\k\neq 0$ the ratio of the two amplitudes is given by~$(A^{\mbox{\tiny L}\sL}_{pq}/F^{\mbox{\tiny L}\sL}_{pq})^2$. Additionally, both the dressing factor~$\sigma^{\circ\circ}$ and the rapidity~$w_p$ might have a more complicated form in the mixed-flux case.
\subsubsection{Mixed-mass sector}
Let us now consider the scattering of a massive particle with a massless one. On symmetry grounds we can write
\begin{equation}
\mathcal{S}^{\bullet\circ} =
\Bigl[\sigma^{\bullet\circ}_{\mbox{\tiny L}} (\mathcal{S}^{\mbox{\tiny L}\sL}\otimes\mathcal{S}^{\mbox{\tiny L}\tilde{\text{\tiny L}}})^{\oplus 2}\Bigr]
\oplus
\Bigl[\sigma^{\bullet\circ}_{\mbox{\tiny R}}(\mathcal{S}^{\mbox{\tiny R}\mbox{\tiny L}}\otimes\mathcal{S}^{\mbox{\tiny R}\tilde{\text{\tiny L}}})^{\oplus 2}\Bigr] ,
\end{equation}
where the subscript indices L, R on the dressing factors refer to the target-space chirality of the massive excitation.
The presence of two copies of the S matrix inside each square bracket is due the fact that the massless $\alg{psu}(1|1)^4_{\text{c.e.}}$ modules are doublets under~$\alg{su}(2)_\circ$. Once again, we have two dressing factors, which should be related to one another by replacing~$x^\pm_{p\,\mbox{\tiny L}}\leftrightarrow x^\pm_{p\,\mbox{\tiny R}}$ everywhere, \textit{i.e.}\xspace
\begin{equation}
\label{eq:mixed-ansatz}
\sigma^{\bullet\circ}_{\mbox{\tiny L}}(p,q)= \sigma^{\bullet\circ}(x_{\mbox{\tiny L}\,p}^\pm,x_{\mbox{\tiny L}\,q}^\pm),
\qquad
\sigma^{\bullet\circ}_{\mbox{\tiny R}}(p,q)= \sigma^{\bullet\circ}(x_{\mbox{\tiny R}\,p}^\pm,x_{\mbox{\tiny R}\,q}^\pm)= \sigma^{\bullet\circ}\bigl(x_{\mbox{\tiny R}\,p}^\pm,\frac{1}{x_{\mbox{\tiny L}\,q}^\mp}\bigr),
\end{equation}
where in the last equation we used that $x^{\pm}_{\mbox{\tiny R}}=1/x^{\mp}_{\mbox{\tiny L}}$ at $m=0$. The scattering elements and hence the expression of $\sigma^{\bullet\circ}$ should not depend on whether we represent the massless particles as left- or right-movers, which is a constraint on the form of the dressing factor.
In a similar way, we can also can write
\begin{equation}
\mathcal{S}^{\circ\bullet} =
\Bigl[\sigma^{\circ\bullet}_{\mbox{\tiny L}} (\mathcal{S}^{\mbox{\tiny L}\sL}\otimes\mathcal{S}^{\tilde{\text{\tiny L}}\mbox{\tiny L}})^{\oplus 2}\Bigr]
\oplus
\Bigl[\sigma^{\circ\bullet}_{\mbox{\tiny R}}(\mathcal{S}^{\mbox{\tiny L}\mbox{\tiny R}}\otimes\mathcal{S}^{\tilde{\text{\tiny L}}\mbox{\tiny R}})^{\oplus 2}\Bigr] ,
\end{equation}
with the same caveats for the dressing factors as above.
\subsection{Dressing factors}
\label{sec:crossing}
The linear symmetries that we used in the previous subsection cannot constrain the scalar factors. On the other hand, braiding and physical unitarity, and crossing symmetry will impose new constraints. Before discussing those, let us fix the normalisation of each block of the S~matrix.
\subsubsection{Normalisations}
The normalisation each S-matrix block can be read off from the elements listed below. In the massive sector we have chosen
\begin{equation}
\begin{aligned}
\bra{Y^{\mbox{\tiny L}}_q \, Y^{\mbox{\tiny L}}_p} \mathcal{S} \ket{Y^{\mbox{\tiny L}}_p \, Y^{\mbox{\tiny L}}_q} & = \frac{x^+_{\mbox{\tiny L}\, p}}{x^-_{\mbox{\tiny L}\, p}} \, \frac{x^-_{\mbox{\tiny L}\, q}}{x^+_{\mbox{\tiny L}\, q}} \, \frac{x^-_{\mbox{\tiny L}\, p} - x^+_{\mbox{\tiny L}\, q}}{x^+_{\mbox{\tiny L}\, p} - x^-_{\mbox{\tiny L}\, q}} \, \frac{1-\frac{1}{x^-_{\mbox{\tiny L}\, p} x^+_{\mbox{\tiny L}\, q}}}{1-\frac{1}{x^+_{\mbox{\tiny L}\, p} x^-_{\mbox{\tiny L}\, q}}} \, \frac{1}{\left(\sigma^{\bullet\bullet}_{\mbox{\tiny L}\sL\,pq} \right)^2 }, \\
\bra{Y^{\mbox{\tiny R}}_q \, Y^{\mbox{\tiny L}}_p} \mathcal{S} \ket{Y^{\mbox{\tiny L}}_p \, Y^{\mbox{\tiny R}}_q} & = \frac{x^+_{\mbox{\tiny L}\, p}}{x^-_{\mbox{\tiny L}\, p}} \, \frac{x^-_{\mbox{\tiny R}\, q}}{x^+_{\mbox{\tiny R}\, q}} \, \frac{1-\frac{1}{x^+_{\mbox{\tiny L}\, p} x^-_{\mbox{\tiny R}\, q}}}{1-\frac{1}{x^+_{\mbox{\tiny L}\, p} x^+_{\mbox{\tiny R}\, q}}} \, \frac{1-\frac{1}{x^-_{\mbox{\tiny L}\, p} x^+_{\mbox{\tiny R}\, q}}}{1-\frac{1}{x^-_{\mbox{\tiny L}\, p} x^-_{\mbox{\tiny R}\, q}}} \, \frac{1}{\left(\tilde{\sigma}^{\bullet\bullet}_{\mbox{\tiny L}\mbox{\tiny R}\,pq} \right)^2 }, \end{aligned}
\end{equation}
with two more equations following by LR symmetry when we exchange everywhere~L$\leftrightarrow$R.
In the massless sector we set
\begin{equation}
\begin{aligned}
\label{eq:unitarity-varsigma}
\bra{T^{\dot{a}a}_q \, T^{\dot{a}a}_p} \mathcal{S} \ket{T^{\dot{a}a}_p \, T^{\dot{a}a}_q} & = \frac{1}{\left(\sigma^{\circ\circ}_{pq} \right)^2 }.
\end{aligned}
\end{equation}
Finally, in the mixed-mass sector we choose
\footnote{%
The normalisation chosen here takes a different form to the one of~\cite{Borsato:2014hja} but reduces to it at~$\k =0$. It is chosen to simplify the form of the constraints imposed by unitarity.
}
\begin{equation}
\begin{aligned}
\bra{T^{\dot{a}a}_q \, Y^{\mbox{\tiny L}}_p} \mathcal{S} \ket{Y^{\mbox{\tiny L}}_p \, T^{\dot{a}a}_q}
&=
\left( \frac{x^+_{\mbox{\tiny L}\,p} - x^+_{\mbox{\tiny L}\,q}}{x^+_{\mbox{\tiny L}\,p} - x^-_{\mbox{\tiny L}\,q}} \, \frac{x^-_{\mbox{\tiny L}\,p} - x^-_{\mbox{\tiny L}\,q}}{x^-_{\mbox{\tiny L}\,p} - x^+_{\mbox{\tiny L}\,q}} \right)^{1/2} \, \frac{1}{\left(\sigma^{\bullet\circ}_{\mbox{\tiny L}\sL\,pq} \right)^2 }, \\
\bra{Y^{\mbox{\tiny L}}_q \, T^{\dot{a}a}_p} \mathcal{S} \ket{T^{\dot{a}a}_p \, Y^{\mbox{\tiny L}}_q}
&=
\left( \frac{x^+_{\mbox{\tiny L}\,p} - x^+_{\mbox{\tiny L}\,q}}{x^+_{\mbox{\tiny L}\,p} - x^-_{\mbox{\tiny L}\,q}} \, \frac{x^-_{\mbox{\tiny L}\,p} - x^-_{\mbox{\tiny L}\,q}}{x^-_{\mbox{\tiny L}\,p} - x^+_{\mbox{\tiny L}\,q}} \right)^{1/2} \, \frac{1}{\left(\sigma^{\circ\bullet}_{\mbox{\tiny L}\sL\,pq} \right)^2 },
\end{aligned}
\end{equation}
and using LR-symmetry and the relation~\eqref{eq:magic-relation} this implies
\begin{equation}
\begin{aligned}
\bra{T^{\dot{a}a}_q \, Y^{\mbox{\tiny R}}_p} \mathcal{S} \ket{Y^{\mbox{\tiny R}}_p \, T^{\dot{a}a}_q}
&=
\left( \frac{1-\frac{1}{x^+_{\mbox{\tiny R}\,p} x^-_{\mbox{\tiny L}\,q}}}{1-\frac{1}{x^+_{\mbox{\tiny R}\,p} x^+_{\mbox{\tiny L}\,q}}} \, \frac{1-\frac{1}{x^-_{\mbox{\tiny R}\,p} x^+_{\mbox{\tiny L}\,q}}}{1-\frac{1}{x^-_{\mbox{\tiny R}\,p} x^-_{\mbox{\tiny L}\,q}}} \right)^{1/2} \, \frac{1}{\left(\sigma^{\bullet\circ}_{\mbox{\tiny R}\mbox{\tiny L}\,pq} \right)^2 }, \\
\bra{Y^{\mbox{\tiny R}}_q \, T^{\dot{a}a}_p} \mathcal{S} \ket{T^{\dot{a}a}_p \, Y^{\mbox{\tiny R}}_q}
&=
\left( \frac{1-\frac{1}{x^+_{\mbox{\tiny L}\,p} x^-_{\mbox{\tiny R}\,q}}}{1-\frac{1}{x^+_{\mbox{\tiny L}\,p} x^+_{\mbox{\tiny R}\,q}}} \, \frac{1-\frac{1}{x^-_{\mbox{\tiny L}\,p} x^+_{\mbox{\tiny R}\,q}}}{1-\frac{1}{x^-_{\mbox{\tiny L}\,p} x^-_{\mbox{\tiny R}\,q}}} \right)^{1/2} \, \frac{1}{\left(\sigma^{\circ\bullet}_{\mbox{\tiny L}\mbox{\tiny R}\,pq} \right)^2 }.
\end{aligned}
\end{equation}
\subsubsection{Unitarity}
Owing to our choice of normalisation, the requirements of braiding an physical unitarity take a simple form
\begin{equation}
\begin{aligned}
\sigma^{\bullet\bullet}_{qp}=\frac{1}{\sigma^{\bullet\bullet}_{pq}}=\bigl(\sigma^{\bullet\bullet}_{pq}\bigr)^* ,
\qquad
\tilde{\sigma}^{\bullet\bullet}_{qp}=\frac{1}{\tilde{\sigma}^{\bullet\bullet}_{pq}}=\bigl(\tilde{\sigma}^{\bullet\bullet}_{pq}\bigr)^* ,
\qquad
\sigma^{\circ\circ}_{qp}=\frac{1}{\sigma^{\circ\circ}_{pq}}=\bigl(\sigma^{\circ\circ}_{pq}\bigr)^* ,\\
\sigma^{\bullet\circ}_{qp}=\frac{1}{\sigma^{\circ\bullet}_{pq}}=\bigl(\sigma^{\bullet\circ}_{pq}\bigr)^* ,
\qquad
\sigma^{\circ\bullet}_{qp}=\frac{1}{\sigma^{\bullet\circ}_{pq}}=\bigl(\sigma^{\circ\bullet}_{pq}\bigr)^* ,
\qquad
\varsigma_{qp}=-\varsigma_{pq}=\bigl(\varsigma_{pq}\bigr)^* ,
\end{aligned}
\end{equation}
where~$*$ denotes complex conjugation.
\subsubsection{Crossing symmetry}
The crossing transformation acts by flipping the sign of momentum and making the energy negative:
\begin{equation}
\label{eq:general-crossing}
p\to \bar{p}=-p ,
\qquad
E(p)\to E(\bar{p})=-E(-p).
\end{equation}
Note that in the latter relation we also flip the sign of~$p$ in~$E(p)$. That sign is irrelevant in a parity-invariant theory, but it does affect the sign of the linear terms in the momenta in our case.
In general, it is convenient to describe the crossing transformation by introducing a rapidity variable that uniformises the dispersion relation. In relativistic theories, this can be done using a hyperbolic parametrisation, while for $\text{AdS}_5\times\mathrm{S}^5$ strings one can use elliptic functions to describe a rapidity torus~\cite{Janik:2006dc}. In both scenarios, crossing then amounts to an imaginary shift of the rapidity.
It is less clear how to uniformise the dispersion relation here. On the other hand, we can realise~\eqref{eq:general-crossing} in terms of the Zhukovski variables by setting
\begin{equation}
\label{eq:crossingmap}
x^{\pm}_{\mbox{\tiny L}}(\bar{p})=\frac{1}{x^{\pm}_{\mbox{\tiny R}}(p)} ,
\qquad
x^{\pm}_{\mbox{\tiny R}}(\bar{p})=\frac{1}{x^{\pm}_{\mbox{\tiny L}}(p)}.
\end{equation}
The supercharges are not meromorphic in~$x^\pm$, so that we have to resolve a square-root ambiguity when performing crossing. We do this by setting
\begin{equation}
\eta^{\mbox{\tiny L}}(\bar{p})= \frac{i}{x^{+}_{\mbox{\tiny R}}(p)}\eta^{\mbox{\tiny R}}(p) ,
\qquad
\eta^{\mbox{\tiny R}}(\bar{p})= \frac{i}{x^{+}_{\mbox{\tiny L}}(p)}\eta^{\mbox{\tiny L}}(p).
\end{equation}
In order to write down the crossing equation, we need to define a charge conjugation matrix, which in our case will be the same as the one of~\cite{Borsato:2014hja}. Let us pick a basis
\begin{equation}
( Y^{\mbox{\tiny L}}, \eta^{\mbox{\tiny L} 1}, \eta^{\mbox{\tiny L} 2}, Z^{\mbox{\tiny L}} ) \oplus ( Y^{\mbox{\tiny R}}, \eta^{\mbox{\tiny R} 1}, \eta^{\mbox{\tiny R} 2}, Z^{\mbox{\tiny R}} ) \oplus ( T^{11}, T^{21}, T^{12}, T^{22} ) \oplus ( \widetilde{\chi}^1, \chi^1, \widetilde{\chi}^2, \chi^2 ).
\end{equation}
The the charge conjugation matrix is
\begin{equation}
\label{eq:chargeconj}
\newcommand{\color{black!40}0}{\color{black!40}0}
\renewcommand{\arraystretch}{1.1}
\setlength{\arraycolsep}{3pt}
\mathscr{C}_p=\!\left(\!
\mbox{\footnotesize$
\begin{array}{cccc|cccc}
\color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & 1 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 \\
\color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & -i & \color{black!40}0 \\
\color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & i & \color{black!40}0 & \color{black!40}0 \\
\color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & 1 \\
\hline
1 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 \\
\color{black!40}0 & \color{black!40}0 & i & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 \\
\color{black!40}0 & -i & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 \\
\color{black!40}0 & \color{black!40}0 & \color{black!40}0 & 1 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 \\
\end{array}$}\!
\right)
\oplus
\!\left(\!
\mbox{\footnotesize$
\begin{array}{cccc|cccc}
\color{black!40}0 & \color{black!40}0 & \color{black!40}0 & 1 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 \\
\color{black!40}0 & \color{black!40}0 & -1 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 \\
\color{black!40}0 & -1 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 \\
1 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 \\
\hline
\color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & -i c_p \\
\color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & i c_p & \color{black!40}0 \\
\color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & i c_p & \color{black!40}0 & \color{black!40}0 \\
\color{black!40}0 & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 & -i c_p & \color{black!40}0 & \color{black!40}0 & \color{black!40}0 \\
\end{array}$}\!
\right),
\end{equation}
where we note that the dependence on~$p$ in the massless sector comes through
\begin{equation}
c_p=\frac{a_{\mbox{\tiny L}}(p)}{b_{\mbox{\tiny R}}(p)}=-\text{sgn}\Bigl[\sin \frac{p}{2}\Bigr] ,
\end{equation}
where the last equality uses that $m=0$. This is reassuring as it indicates that if we treat separately left- and right-movers on the worldsheet, $\mathscr{C}$ is indeed a constant matrix. Note that, even if we are taking the massless modes to be in the left representation, we see the coefficients of the right representation appearing. This is not surprising as crossing exchanges the two kinematics.
We can now write down the crossing equation%
\footnote{%
There is one more crossing equation where crossing is performed in the second variable, which is equivalent to the one given here by unitarity.
}
\begin{equation}
\begin{aligned}
\mathscr{C}_p \otimes \mathbf{1} \cdot \mathbf{S}^{\text{t}_1}(\bar{p},q) \cdot \mathscr{C}^{-1}_p \otimes \mathbf{1} \cdot \mathbf{S}(p,q) &= \mathbf{1} \otimes \mathbf{1},
\end{aligned}
\end{equation}
where we have introduced the short-hand~$\gen{S}=\Pi\,\mathcal{S}$ defined in terms of the permutation~$\Pi$, and~$\text{t}_j$ denotes transposition in the $j$th space.
From this we can read off the constraints on the dressing factors.
In the massive sector we find
\begin{equation}
\begin{aligned}
\sigma^{\bullet\bullet}_{\mbox{\tiny L}\sL}(p,q)^2 \ \tilde{\sigma}^{\bullet\bullet}_{\mbox{\tiny R}\mbox{\tiny L}}(\bar{p},q)^2 &= \left( \frac{x^-_{\mbox{\tiny L}\,q}}{x^+_{\mbox{\tiny L}\,q}} \right)^2 \frac{(x^-_{\mbox{\tiny L}\,p}-x^+_{\mbox{\tiny L}\,q})^2}{(x^-_{\mbox{\tiny L}\,p}-x^-_{\mbox{\tiny L}\,q})(x^+_{\mbox{\tiny L}\,p}-x^+_{\mbox{\tiny L}\,q})} \frac{1-\frac{1}{x^-_{\mbox{\tiny L}\,p}x^+_{\mbox{\tiny L}\,q}}}{1-\frac{1}{x^+_{\mbox{\tiny L}\,p}x^-_{\mbox{\tiny L}\,q}}} ,
\\
\sigma^{\bullet\bullet}_{\mbox{\tiny L}\mbox{\tiny R}}(p,q)^2 \ \tilde{\sigma}^{\bullet\bullet}_{\mbox{\tiny R}\sR}(\bar{p},q)^2 &= \left( \frac{x^-_{\mbox{\tiny R}\,q}}{x^+_{\mbox{\tiny R}\,q}} \right)^2 \frac{\left(1-\frac{1}{x^+_{\mbox{\tiny L}\,p}x^+_{\mbox{\tiny R}\,q}}\right)\left(1-\frac{1}{x^-_{\mbox{\tiny L}\,p}x^-_{\mbox{\tiny R}\,q}}\right)}{\left(1-\frac{1}{x^+_{\mbox{\tiny L}\,p}x^-_{\mbox{\tiny R}\,q}}\right)^2} \frac{x^-_{\mbox{\tiny L}\,p}-x^+_{\mbox{\tiny R}\,q}}{x^+_{\mbox{\tiny L}\,p}-x^-_{\mbox{\tiny R}\,q}}.
\end{aligned}
\end{equation}
Note that the Zhukovski variables carry the appropriate flavours with respect to the phases. This is guaranteed by the fact that both the transformation~\eqref{eq:crossingmap} and the charge conjugation matrix~\eqref{eq:chargeconj} swap left with right in the massive sector. Two additional equations can be written down by LR symmetry, exhausting the constraints of crossing in this sector.
In the massless sector we have
\begin{equation}
\begin{aligned}
\sigma^{\circ\circ}(p,q)^2 \ \sigma^{\circ\circ}(\bar{p},q)^2 &= \frac{\varsigma_{pq}-1}{\varsigma_{pq}} \, \frac{x^-_{\mbox{\tiny L}\,p}-x^+_{\mbox{\tiny L}\,q}}{x^+_{\mbox{\tiny L}\,p}-x^+_{\mbox{\tiny L}\,q}} \frac{x^+_{\mbox{\tiny L}\,p}-x^-_{\mbox{\tiny L}\,q}}{x^-_{\mbox{\tiny L}\,p}-x^-_{\mbox{\tiny L}\,q}},
\qquad
\varsigma_{\bar{p}q} &= \varsigma_{pq}-1.
\end{aligned}
\end{equation}
It is interesting to note that
\begin{equation}
\frac{x^-_{\mbox{\tiny L}\,p}-x^+_{\mbox{\tiny L}\,q}}{x^+_{\mbox{\tiny L}\,p}-x^+_{\mbox{\tiny L}\,q}} \frac{x^+_{\mbox{\tiny L}\,p}-x^-_{\mbox{\tiny L}\,q}}{x^-_{\mbox{\tiny L}\,p}-x^-_{\mbox{\tiny L}\,q}}=\frac{x^-_{\mbox{\tiny R}\,p}-x^+_{\mbox{\tiny R}\,q}}{x^+_{\mbox{\tiny R}\,p}-x^+_{\mbox{\tiny R}\,q}} \frac{x^+_{\mbox{\tiny R}\,p}-x^-_{\mbox{\tiny R}\,q}}{x^-_{\mbox{\tiny R}\,p}-x^-_{\mbox{\tiny R}\,q}}
\qquad
\text{at}\quad m=0,
\end{equation}
so that the crossing equation for the massless phases does not depend on whether we decided to represent the massless modes as left or right particles.
Finally, in the mixed-mass sector we have
\begin{equation}
\begin{aligned}
\sigma^{\bullet \circ}_{\mbox{\tiny L}}(p,q)^2 \ \sigma^{\bullet \circ}_{\mbox{\tiny R}}(\bar{p},q)^2 &=
\frac{x^-_{\mbox{\tiny L}\,p}-x^+_{\mbox{\tiny L}\,q}}{x^+_{\mbox{\tiny L}\,p}-x^+_{\mbox{\tiny L}\,q}} \frac{x^+_{\mbox{\tiny L}\,p}-x^-_{\mbox{\tiny L}\,q}}{x^-_{\mbox{\tiny L}\,p}-x^-_{\mbox{\tiny L}\,q}} =
\sigma^{\circ\bullet}_{\mbox{\tiny L}}(q,p)^2 \ \sigma^{\circ\bullet}_{\mbox{\tiny L}}(\bar{q},p)^2 ,
\end{aligned}
\end{equation}
where in order to write down this equation only in term of \emph{left} massless Zhukovski variables we have used~\eqref{eq:magic-relation}. The same formula, together with LR symmetry, yields
\begin{equation}
\begin{aligned}
\sigma^{\bullet \circ}_{\mbox{\tiny R}}(p,q)^2 \ \sigma^{\bullet \circ}_{\mbox{\tiny L}}(\bar{p},q)^2 &=
\frac{1-\frac{1}{x^+_{\mbox{\tiny R}\,p}x^+_{\mbox{\tiny L}\,q}}}{1-\frac{1}{x^+_{\mbox{\tiny R}\,p}x^-_{\mbox{\tiny L}\,q}}}
\frac{1-\frac{1}{x^-_{\mbox{\tiny R}\,p}x^-_{\mbox{\tiny L}\,q}}}{1-\frac{1}{x^-_{\mbox{\tiny R}\,p}x^+_{\mbox{\tiny L}\,q}}} =
\sigma^{\circ\bullet}_{\mbox{\tiny R}}(q,p)^2 \ \sigma^{\circ\bullet}_{\mbox{\tiny R}}(\bar{q},p)^2.
\end{aligned}
\end{equation}
Note also that the form of this crossing equation is compatible with~\eqref{eq:mixed-ansatz}.
\section{Discussion and outlook}
\label{sec:conclusion}
In this paper we have determined the complete all-loop worldsheet S matrix of Type IIB string theory on $\text{AdS}_3\times\mathrm{S}^3\times\mathrm{T}^4 $ with mixed R-R and NS-NS three-form flux, up to the dressing factors. We further wrote down the crossing relations that these dressing factors have to satisfy. In constructing this S matrix we relied on the off-shell symmetry algebra $\mathcal{A}$ of the gauge-fixed theory and its representations. This method, initially advocated in the context of $\text{AdS}_5\times \mathrm{S}^5$~\cite{Arutyunov:2006ak}, has recently been shown to be particularly well-suited in the study of \emph{massless}, as well as massive, excitations and allowed for the construction of the non-perturbative S matrix for the pure R-R flux $\text{AdS}_3\times\mathrm{S}^3\times\mathrm{T}^4$ theory~\cite{Borsato:2014exa,Borsato:2014hja}. Our present work demonstrates the versatility of this approach and provides strong evidence that, by using these methods, one will be able to tackle other classes of backgrounds such as $\text{AdS}_3\times\mathrm{S}^3\times\mathrm{S}^3 \times\mathrm{S}^1$~\cite{Abbott:2012dd, Babichenko:2009dk, Beccaria:2012kb, Beccaria:2012pm,Abbott:2013ixa,Sundin:2012gc, Borsato:2012ud,Borsato:2012ss}, or the less-supersymmetric backgrounds discussed in the context of integrability in refs.~\cite{Sorokin:2011rr,Wulff:2014kja,Murugan:2012mf,Abbott:2013kka,Hoare:2014kma}. It would be particularly interesting to investigate these backgrounds as well as to solve the crossing relations proposed in this paper and we hope to return to this subject in the near future.
The analysis carried out here provides strong evidence for the validity of the $q$-dependent dispersion relation proposed in ref.~\cite{Hoare:2013lja}. We find in particular that this form of dispersion relation is satisfied by both massive and massless modes.
The S-matrix construction presented here elucidates a feature pure-R-R S matrix. There, for massive particles, it was found that symmetries and unitarity leave two choices for the all-loop S~matrix: one where the target space chirality is always transmitted, and one where it is always reflected~\cite{Borsato:2012ud,Borsato:2013qpa}. Only the former option satisfies the Yang-Baxter equation and is compatible with tree-level perturbative calculations~\cite{Rughoonauth:2012qd}. Interestingly, in the presence of NS-NS fluxes we find that the pure-reflection S~matrix is immediately ruled out by symmetries, further motivating the choice of a reflectionless S~matrix in refs.~\cite{Borsato:2012ud,Borsato:2013qpa}.
The S matrix presented in this paper provides a two-parameter family ($\lambda$ and $q$) of quantum integrable models. It is likely that this parameter space can be further enhanced by the study of so-called $\eta$-deformations~\cite{Beisert:2008tw,Delduc:2013qra,Arutyunov:2013ega,Kawaguchi:2014qwa,Arutynov:2014ota}. This large parameter space, and the presence of novel massless modes deserves a detailed investigation in the context of integrability and may provide us with new inputs into the relationship between integrability and holography~\cite{Pittelli:2014ria}.
From our previous work~\cite{Borsato:2014exa,Borsato:2014hja} and the present paper we are lead to conclude that the solution of the spectral problem in $\text{AdS}_3/\text{CFT}_2$ is likely to be within reach using integrable methods. In particular it would be important to understand the mirror Thermodynamical Bethe Ansatz and Quantum Spectral Curve for these backgrounds~\cite{Ambjorn:2005wa,Arutyunov:2007tc, Arutyunov:2009zu, Gromov:2009tv,Bombardelli:2009ns,Arutyunov:2009ur, Cavaglia:2010nm, Gromov:2013pga,Gromov:2014caa}.
Given this, it would be interesting to investigate other aspects of this class of dualities, such as scattering amplitudes, Wilson loops or entanglement entropy using integrable methods. It would also be important to connect the results presented here to the higher-spin holography of $\text{AdS}_3$ backgrounds investigated in recent papers such as~\cite{Gaberdiel:2013vva,Ahn:2013oya,Gaberdiel:2014yla, Creutzig:2014ula,Gaberdiel:2014cha}.
Finally, our work suggests that connections between the integrable approach and other studies of the $\text{AdS}_3/\text{CFT}_2$ correspondence, that deserve to be explored more fully. Two links naturally suggest themselves. Firstly, the mixed-flux theory investigated in the present paper was analysed some time ago in the hybrid formalism~\cite{Berkovits:1999im} and it would be interesting to establish connections between those results and the work presented in this paper. For example, might one be able to see integrable structure in the framework of ref.~\cite{Berkovits:1999im}? Secondly, type IIB string theory on $\text{AdS}_3\times\mathrm{S}^3\times\mathrm{T}^4$ with only NS-NS flux was investigated in detail from the point of view of a WZW theory in refs.~\cite{Maldacena:2000hw,Maldacena:2000kv,Maldacena:2001km}, where worldsheet CFT methods were used to powerful effect. It would be intriguing to see if there is a way to take a $q\rightarrow 1$ limit of our integrable structure in a controlled way.
Uncovering the role of integrability in the pure NS-NS theory, as recently investigated in ref.~\cite{Hernandez:2014eta}, might provide new connections between integrable and worldsheet CFT approaches.
\section*{Acknowledgements}
We would like to thank Gleb Arutyunov, Sergey Frolov, Ben Hoare, Alessandro Torrielli, Arkady Tseytlin, Dan Waldram and Kostya Zarembo for helpful discussions, and Ben Hoare and Arkady Tseytlin for their comments on the manuscripts.
We are very grateful to Riccardo Borsato for his collaboration on earlier related work, critical reading and invaluable comments on the manuscript and for many useful discussions.
T.L. is supported by an STFC studentship.
O.O.S.'s work was supported by the ERC Advanced grant No.~290456,
``Gauge theory -- string theory duality''.
A.S.'s work is funded by the People Programme (Marie Curie Actions) of the European Union, Grant Agreement No.~317089 (GATIS). A.S. also acknowledges the hospitality at APCTP where part of this work was
done.
B.S.\@ acknowledges funding support from an STFC Consolidated Grant ``Theoretical Physics at City University''
ST/J00037X/1.
|
\section{Introduction}
The statistical (two-dimensional, lattice) models constitute large
class of exactly solvable models~\cite{Baxter} whose development is
witnessed even in present days. The solvability in those models
implies the possibility of finding exact results for the partition
function, correlation functions etc. On the other hand, there exists
another important branch of modern statistical physics, namely the
study of deformed analogs of Bose (and Fermi) gas model in two,
three, etc. dimensions. Its development, from the earliest papers on
the
subject~\cite{LeeYu1990,GeSu1991,Martin-Delgado1991,Chaichian1993,Man'ko1993}
and till some most recent
ones~\cite{AlginIlik2013,GM_UJP2013,Rovenchak2014} demonstrates that
this still remains a hot topic (see
also~\cite{DaiXie_rev1,AlginSenay2012} for some, by no means
exhaustive, list of works published during the last two decades),
with many interesting problems
awaiting for their resolving. Among these, it is worth to point out
the task of deriving exact expressions for statistical quantities,
in particular the multi-particle distribution functions and
respective (momentum) correlation functions or their intercepts. At
present, to the best of our knowledge, for only two essentially
differing deformed analogs of the Bose gas model that task has been
performed. Namely, the $r$-particle ($r\ge 1$) distribution
functions and the respective $r$-particle correlation function
intercepts have been obtained (i) for the $p,q$-deformed Bose gas
model, see~\cite{AdGa}, and (ii) for the recently proposed
$\mu$-deformed analog of Bose gas model~\cite{GaMi2012}. The two
models principally differ: while the first one belongs to Fibonacci
class~\cite{Arik1992Fib}, the second one is not Fibonacci, but is a
typical representative of the class of quasi-Fibonacci
models~\cite{GKR_Fib}.
Recently, a new symmetric Tamm-Dancoff (STD) $q$-deformed oscillator has
been introduced~\cite{Chung_S-TD} and some of its major properties studied.
Note that usual Tamm-Dancoff $q$-oscillator~\cite{OdakaTD1,Chaturvedi1993_TD2} involves only
{\it real values} of the parameter $q$, and possesses diverse cases of accidental
degeneracies of energy levels~\cite{GR_TD3}. The STD $q$-oscillator admits, besides
real, also the {\it complex phase-like} values of deformation parameter $q$ and, moreover, there
exist definite values of $q=\exp(\rm i\theta)$ for which accidental degeneracies
of energy levels do occur~\cite{Chung_S-TD}. The possibility to deal with the complex-valued
deformation parameter, which we have in particular in the STD type $q$-deformed
model, gives some important advantages, as it was discussed in (the last paragraph of) ref.~\cite{Chung_S-TD}.
Let us emphasize that whereas usual Tamm-Dancoff $q$-oscillator belongs to the class of
Fibonacci oscillators, the STD $q$-deformed oscillator
does not. Namely, as shown in~\cite{Chung_S-TD}, it is a quasi-Fibonacci one (the notion
of quasi-Fibonacci oscillators was introduced in~\cite{GKR_Fib}).
In this letter our goal is to derive, for the STD type $q$-deformed
Bose gas model based on the set of
STD type $q$-deformed oscillators,
the exact expressions for the $r$-particle ($r\ge 1$) distribution
functions and the respective $r$-th order correlation function intercepts. By deriving
this result we demonstrate that there is one more, the third, $q$-deformed family of modified Bose-gas
like models for which it proves possible to obtain the exact analytical formulas
for the $r$th order distribution functions and $r$-particle correlation intercepts.
\section{Basics of the symmetric Tamm-Dancoff $q$-oscillator~\cite{Chung_S-TD}}\label{sec:std-basics}
The symmetric $q$-deformed bosonic Tamm-Dancoff oscillator algebra is defined as
\begin{align}
&aa^{\dagger} - a^{\dagger}a = \{N+1\}_q - \{N\}_q = \frac{1}{2} ( 1
+ (1-q^{-1} )N ) )
q^{N} + \frac{1}{2} ( 1 + (1-q) N) q^{-N}, \label{1.1}\\
&[N, a^{\dagger}] = a^{\dagger}, \quad [N, a]= -a,\nonumber
\end{align}
where
\begin{equation} \label{phi_STD}
a^\dag a = \{N\}_q \equiv\varphi_{\rm STD}(N)\equiv \frac{N}{2}(q^{N-1}+q^{-N+1})= \frac{N}{2} q^{-N+1} (1 + (q^2)^{N-1} )
\end{equation}
with $\{N\}_q$ denoting the STD type $q$-number ($q$-bracket) or
structure function, and $q$ is either real, \ $0< q \le \infty$, or
complex phase-like: $q=\exp({\rm i}\theta)$, \ $-\pi \le\theta\le
\pi$.
One can easily show that the $q$-analog Fock-type representation of
the algebra (\ref{1.1}) is valid:
\begin{align*}
&N|n\rangle = n |n\rangle, ~~~~ n=0, 1, 2, \ldots ,\\
&a|n\rangle = \sqrt{\frac{ n }{2} ( q^{n-1} + q^{-n+1} ) }|n-1\rangle = \sqrt{\{ n \}_q}|n-1\rangle,\\
&a^\dag |n\rangle = \sqrt{\frac{ n+1 }{2} ( q^{ n} + q^{-n} ) }|n+1\rangle = \sqrt{\{n+1\}_q}|n+1\rangle.
\end{align*}
Remark that this STD $q$-bracket can be also written as a
``multiplicative hybrid'' of the structure function
$\varphi(n)\!=\!n$ of usual oscillator and that of the Biedenharn -
Macfarlane $q$-deformed oscillator~\cite{Bied,Macfar} i.e. the
(obviously symmetric under $q\leftrightarrow q^{-1}$) structure
function $\varphi_{BM}(n)\equiv[n]_q=\frac{q^n-q^{-n}}{q-q^{-1}}$,
as follows:
\begin{equation}
\{ n\}_q=n\,\frac{[n]_q\!-\![n\!-\!2]_q}{2} \qquad {\rm or} \qquad
\{ n\}_q=n\,\frac{[2(n-1)]_q}{2\,[n-1]_q}.
\end{equation}
\section{Intercepts of 2nd and 3rd order correlation functions}\label{sec2}
The deformed Bose gas model constructed from the set of independent
modes of deformed oscillators with STD type structure function of
deformation $\varphi_{\rm STD}(N)$, see eq.~(\ref{phi_STD}), is
studied here. We start with the following defining expression for
the intercept (see e.g.~\cite{ChapmanHeinz1994} for a nondeformed
case, and~\cite{AdGa} for deformed one) of $r$th order momentum
correlation function, with fixed momentum~$\bf k$:
\begin{equation}\label{lambda_def}
\lambda^{(r)}({\bf k}) = \frac{\langle (a^\dag_{\bf k})^r (a_{\bf k})^r \rangle}{\langle a^\dag_{\bf k} a_{\bf k}\rangle^r}-1
= \frac{\langle \varphi(N_{\bf k})\varphi(N_{\bf k}-1)\cdot...\cdot\varphi(N_{\bf k}-r+1) \rangle}{\langle \varphi(N_{\bf k})\rangle^r}-1.
\end{equation}
The bracket $\langle...\rangle$ denotes statistical (thermal)
average. As seen, to find the intercepts $\lambda^{(2)}({\bf k})$
and $\lambda^{(3)}({\bf k})$ we have to calculate the averages
$\langle a^\dag_{\bf k} a_{\bf k}\rangle$, $\langle (a^\dag_{\bf
k})^2 (a_{\bf k})^2 \rangle$ and $\langle (a^\dag_{\bf k})^3 (a_{\bf
k})^3 \rangle$. In what follows, when performing the calculations
that involve the quantities for a fixed mode, the index $\bf k$ will
be omitted for the sake of simplicity.
Taking the Hamiltonian in the simplest linear (additive) form
\begin{equation}\label{H}
H= \sum_{\bf k} \varepsilon({\bf k}) N_{\bf k} = \sum_{\bf k} \hbar\omega_{\bf k} N_{\bf k},
\end{equation}
for the average $\langle a^\dag a\rangle$ we find:
\begin{align}
\langle a^\dag a\rangle &= \langle\varphi(N)\rangle = \sum_{n=0}^\infty \frac{n}{2}(q^{n-1}+q^{-n+1}) e^{-\beta \hbar\omega n} \Bigl/
\sum_{n=0}^\infty e^{-\beta \hbar\omega n} = \nonumber\\
&= \frac12 (1-e^{-x}) \Bigl(-\dd{}{x}\Bigr) \sum_{n=0}^\infty (q^{n-1}+q^{-n+1}) e^{-nx}
= \frac{(1-e^{-x})}{2} \Bigl(-\dd{}{x}\Bigr) \Bigl(\frac{q^{-1}}{1\!-\!q e^{-x}} + \frac{q}{1\!-\!q^{-1} e^{-x}}\Bigr) =\nonumber\\
&= \frac{e^{-x}(1\!-\!e^{-x})}{2} \Bigl(\frac{1}{(1\!-\!q e^{-x})^2} + \frac{1}{(1\!-\!q^{-1} e^{-x})^2}\Bigr) \label{<a^dag_a>}
\end{align}
where $x = \beta \hbar\omega_{\bf k}$, $\beta=\frac{1}{k_B T}$,
$k_B$ is Boltzmann's constant. In a similar way we calculate
$\langle (a^\dag)^2 (a)^2 \rangle$:
\begin{align*}
\langle (a^\dag)^2 a^2 \rangle &= \langle\varphi(N)\varphi(N-1)\rangle = \frac14 \langle N(N-1)(q^{N-1}+q^{-N+1})(q^{N-2}+q^{-N+2})\rangle =\\
&= \frac14 (1-e^{-x}) \Bigl(\dd{^2}{x^2}+\dd{}{x}\Bigr) \sum_{n=0}^\infty (q^{n-1}+q^{-n+1})(q^{n-2}+q^{-n+2}) e^{-nx} =\\
&= \frac14 (1-e^{-x}) \Bigl(\dd{^2}{x^2}+\dd{}{x}\Bigr) \sum_{n=0}^\infty (q^{2n-3} + q^{-2n+3}+q^{-1}+q) e^{-nx}=\\
&= \frac14 (1-e^{-x}) \Bigl(\dd{^2}{x^2}+\dd{}{x}\Bigr) \Bigl(\frac{q^{-3}}{1-q^2e^{-x}} + \frac{q^3}{1-q^{-2}e^{-x}} + \frac{q+q^{-1}}{1-e^{-x}}\Bigr).
\end{align*}
Utilizing the equality $\Bigl(\dd{^2}{x^2}+\dd{}{x}\Bigr)
\frac{1}{1-\alpha e^{-x}} = 2\frac{\alpha^2 e^{-2x}}{(1-\alpha e^{-x})^3}$ we
arrive at the desired expressions
\begin{equation}\label{r=2}
\begin{aligned}
&\langle (a^\dag)^2 a^2 \rangle = \frac12 e^{-2x} (1-e^{-x}) \Bigl[\frac{q}{(1-q^2e^{-x})^3} + \frac{q^{-1}}{(1-q^{-2}e^{-x})^3}
+ \frac{q+q^{-1}}{(1-e^{-x})^3}\Bigr],\\
&\lambda^{(2)}({\bf k}) = \frac{2 \bigl[\frac{q}{(1-q^2e^{-x})^3} + \frac{q^{-1}}{(1-q^{-2}e^{-x})^3}
+ \frac{q+q^{-1}}{(1-e^{-x})^3}\bigr]}{(1\!-\!e^{-x}) \bigl((1\!-\!q e^{-x})^{-2} + (1\!-\!q^{-1} e^{-x})^{-2}\bigr)^2}-1.
\end{aligned}
\end{equation}
Analogously, with account of equality
$\Bigl(\dd{}{x}+2\Bigr)\Bigl(\dd{}{x}+1\Bigr)\dd{}{x}\,
\frac{1}{1-\alpha e^{-x}} = -6\frac{\alpha^3 e^{-3x}}{(1-\alpha
e^{-x})^4}$, for $\langle (a^\dag)^3 a^3 \rangle$ we obtain the
third order distribution function and same order correlation
function intercept:
\begin{equation}\label{r=3}
\begin{aligned}
&\langle (a^\dag)^3 a^3 \rangle = \frac34 e^{-3x} (1\!-\!e^{-x}) \Bigl[\frac{q^3}{(1\!-\!q^3e^{-x})^4} + \frac{q^{-3}}{(1\!-\!q^{-3}e^{-x})^4}
+ \frac{q^3(1\!+\!q^{-2}\!+\!q^{-4})}{(1\!-\!qe^{-x})^4} + \frac{q^{-3}(1\!+\!q^2\!+\!q^4)}{(1\!-\!q^{-1}e^{-x})^4}\Bigr],\\
&\lambda^{(3)}({\bf k}) = \frac{6 \bigl[\frac{q^3}{(1\!-\!q^3e^{-x})^4} + \frac{q^{-3}}{(1\!-\!q^{-3}e^{-x})^4}
+ \frac{q^3(1\!+\!q^{-2}\!+\!q^{-4})}{(1\!-\!qe^{-x})^4} + \frac{q^{-3}(1\!+\!q^2\!+\!q^4)}{(1\!-\!q^{-1}e^{-x})^4}\bigr]}{(1\!-\!e^{-x})^2 \bigl((1\!-\!q e^{-x})^{-2} + (1\!-\!q^{-1} e^{-x})^{-2}\bigr)^3}-1.
\end{aligned}
\end{equation}
\section{Distributions and correlation function intercepts of $r$th order}\label{sec3}
The definition~(\ref{lambda_def}) of $r$th order intercepts
$\lambda^{(r)}$ involves the averages (deformed distributions)
$\langle (a^\dag)^r a^r \rangle$, that in the STD $q$-Bose gas case
gives
\begin{equation}
\langle (a^\dag)^r a^r \rangle = \langle \varphi(N)\cdot...\cdot\varphi(N-r+1) \rangle
= \langle N(N-1)...(N-r+1) \prod_{j=1}^r (q^{N-j}+q^{-N+j})\rangle.
\end{equation}
For the product appearing in this equality we obtain
\begin{multline} \label{prod_q^n}
\prod_{j=1}^r (q^{N-j}+q^{-N+j}) = \prod_{j=1}^r q^{N-j} \prod_{j=1}^r (1+(q^2)^{-N+j})
= q^{r N - r(r+1)/2} \sum_{k=0}^r (q^2)^{-kN} \sum_{1\le j_1<...<j_k\le r} (q^2)^{j_1+...+j_k} =\\
= q^{r N - r(r+1)/2} \sum_{k=0}^r (q^2)^{-kN}\sum_{s=k(k+1)/2}^{(2r-k+1)k/2} p(k,r,s) q^{2s}
\end{multline}
where $p(k,r,s)$ is the number of partitions of $s$ into $k$ distinct summands each of
which being not greater than $r$. Remark also that the product in~(\ref{prod_q^n}) can be
expressed through the Gaussian $q$-binomial coefficients $\Bigl({\scriptstyle n\atop \scriptstyle k}\Bigr)_q$
(see e.g.~\cite{KacCheung} for definition) by means of the identity
\begin{equation}
\prod_{j=1}^r (1+(q^2)^{-N+j}) = \prod_{j=0}^{r-1} (1+q^{-2N+2}\cdot q^{2j}) = \sum_{k=0}^r q^{k(k+1)}
\Bigl({\scriptstyle r\atop \scriptstyle k}\Bigr)_{q^2} (q^2)^{-kN}.
\end{equation}
Now perform the calculation that generalizes the case of $\langle (a^\dag)^2 a^2 \rangle$
or $\langle (a^\dag)^3 a^3 \rangle$ to any $r$th order:
\begin{multline}\label{eq1}
\langle (a^\dag)^r a^r \rangle = \langle N(N-1)...(N-r+1)\prod_{j=1}^r (q^{N-j}+q^{-N+j})\rangle
= (1-e^{-x}) \sum_{n=0}^\infty \frac{1}{2^r} n(n-1)...(n-r+1)\cdot\\
\cdot\prod_{j=1}^r (q^{n-j}+q^{-n+j}) e^{-nx}
= \frac{(-1)^r}{2^r} (1-e^{-x}) \Bigl(\dd{}{x}+r-1\Bigr)...\dd{}{x} \sum_{n=0}^\infty \prod_{j=1}^r
(q^{n-j}+q^{-n+j}) e^{-nx} =\\
= \frac{(-1)^r}{2^r} (1-e^{-x}) \Bigl(\dd{}{x}+r-1\Bigr)...\dd{}{x}
\sum_{n=0}^\infty q^{r n - r(r+1)/2} \sum_{k=0}^r (q^2)^{-kn} \sum_{s=k(k+1)/2}^{(2r-k+1)k/2} p(k,r,s) q^{2s} e^{-nx} =\\
= \frac{(-1)^r}{2^r} (1-e^{-x}) \sum_{k=0}^r \sum_{s=k(k+1)/2}^{(2r-k+1)k/2} p(k,r,s) q^{2s-r(r+1)/2} \Bigl(\dd{}{x}+r-1\Bigr)...\dd{}{x} \frac{1}{1-q^{r-2k} e^{-x}}.
\end{multline}
By induction, it can be verified that
\begin{equation}\label{n-der}
\Bigl(\dd{}{x}+r-1\Bigr)...\Bigl(\dd{}{x}+1\Bigr)\dd{}{x}\,
\frac{1}{1-q e^{-x}} = (-1)^r r! \frac{q^r e^{-rx}}{(1-q
e^{-x})^{r+1}}.
\end{equation}
Indeed,
\begin{multline}
\Bigl(\dd{}{x}+r\Bigr)...\dd{}{x}\, \frac{1}{1-q e^{-x}} = (-1)^r r! q^r \Bigl(\frac{-r e^{-rx}}{(1-q e^{-x})^{r+1}} - (r+1) \frac{q e^{-x} e^{-rx}}{(1-q e^{-x})^{r+2}}\Bigr) + (-1)^r r! r \frac{q^r e^{-rx}}{(1-q e^{-x})^{r+1}}
=\\
= (-1)^{r+1} (r+1)! \frac{q^{r+1} e^{-(r+1)x}}{(1-q e^{-x})^{r+2}}.
\end{multline}
From (\ref{n-der}) and (\ref{eq1}) we derive one of our major formulas, namely
\begin{multline}
\langle (a^\dag)^r a^r \rangle = \frac{r!}{2^r} (1-e^{-x})e^{-rx} \sum_{k=0}^r \sum_{s=k(k+1)/2}^{(2r-k+1)k/2} p(k,r,s) q^{2s-r(r+1)/2} \frac{q^{(r-2k)r}}{(1-q^{r-2k} e^{-x})^{r+1}} =\\
= \frac{r!}{2^r} (1-e^{-x})e^{-rx} \sum_{k=0}^r \frac{q^{(r-2k)r}}{(1-q^{r-2k} e^{-x})^{r+1}} \Bigl(\sum_{s=k(k+1)/2}^{(2r-k+1)k/2} p(k,r,s) q^{2s-r(r+1)/2}\Bigr)
\end{multline}
which in terms of well-known $q$-binomials takes the form
\begin{equation}\label{<(a^+)^r_a^r>}
\langle (a^\dag)^r a^r \rangle = \frac{r!}{2^r} (1-e^{-x})e^{-rx}
\sum_{k=0}^r q^{k(k+1)-r(r+1)/2} \Bigl({\scriptstyle r\atop
\scriptstyle k}\Bigr)_{q^2} \frac{q^{(r-2k)r}}{(1-q^{r-2k}
e^{-x})^{r+1}}.
\end{equation}
Using~(\ref{<a^dag_a>}), (\ref{<(a^+)^r_a^r>}) and definition~(\ref{lambda_def}) we
write out the expression for the $r${\it th} order intercept as our final result:
\begin{equation}\label{lambda_r}
\lambda^{(r)}({\bf k}) = \frac{r! q^{-r(r-1)/2}\sum_{k=0}^r
\bigl({\scriptstyle r\atop \scriptstyle k}\bigr)_{q^2}
\frac{q^{(k-r)(k-r+1)}}{(1-q^{r-2k}
e^{-x})^{r+1}}}{(1\!-\!e^{-x})^{r-1} \bigl((1\!-\!q e^{-x})^{-2} +
(1\!-\!q^{-1} e^{-x})^{-2}\bigr)^r} - 1 , \hspace{12mm}
x = \beta \hbar\omega_{\bf k}.
\end{equation}
Its $\beta\hbar\omega_{\bf k}\to\infty$ (large momentum or low
temperature) asymptotics takes the form:
\begin{equation}\label{lambda_r_as}
\lambda^{(r)}_{as} = \{r\}_q!-1 = \frac{r!}{2^r} \prod_{k=1}^r (q^{k-1}+q^{-k+1}) - 1.
\end{equation}
Let us stress that the obtained asymptotics depends, besides the order $r$, on the deformation
parameter $q$ {\it only}: neither mass of particle nor the temperature of $q$-Bose gas survive
in the asymptotics. In the particular case of $r=4$ we have:
\begin{multline}\label{r=4}
\langle (a^\dag)^4 a^4 \rangle = \frac32 (1-e^{-x})e^{-4x} \Bigl[\frac{q^6}{(1-q^4e^{-x})^5} + \frac{1+q^2+q^4+q^6}{(1-q^2e^{-x})^5} + \frac{q^{-4}+q^{-2}+2+q^2+q^4}{(1-e^{-x})^5} + \\
+ \frac{1+q^{-2}+q^{-4}+q^{-6}}{(1-q^{-2}e^{-x})^5} + \frac{q^{-6}}{(1-q^{-4}e^{-x})^5}\Bigr],
\end{multline}
and
\begin{equation}
\lambda^{(4)}({\bf k}) = \frac{24 \bigl[\frac{q^6}{(1-q^4e^{-x})^5} + \frac{1+q^2+q^4+q^6}{(1-q^2e^{-x})^5} + \frac{q^{-4}+q^{-2}+2+q^2+q^4}{(1-e^{-x})^5} + \frac{1+q^{-2}+q^{-4}+q^{-6}}{(1-q^{-2}e^{-x})^5} + \frac{q^{-6}}{(1-q^{-4}e^{-x})^5}\bigr]}{(1\!-\!e^{-x})^3 \bigl((1\!-\!q e^{-x})^{-2} + (1\!-\!q^{-1} e^{-x})^{-2}\bigr)^4}-1.
\end{equation}
\vspace{12mm}
It is worth noting that the symmetry under $q\to q^{-1}$, though
certainly valid, is not so obvious in the obtained general result
(\ref{<(a^+)^r_a^r>}), but, it is easily seen for each of the particular results
(\ref{r=2}), (\ref{r=3}), (\ref{r=4}) for $r=2, 3, 4$ respectively.
As a kind of consistency check, we take the $q\to 1$ limit of
general formula (\ref{<(a^+)^r_a^r>}) and obtain
\begin{equation}
\langle (a^\dag)^r a^r \rangle |_{q\to 1}=
\frac{r!}{2^r}
\, e^{-rx} \sum_{k=0}^r \Bigl({\scriptstyle r\atop
\scriptstyle k}\Bigr) \frac{1}{(1- e^{-x})^{r}} =
\frac{{r!}\, e^{-rx}}{(1-e^{-x})^r},\qquad \lambda^{(r)}_{as}|_{q\to 1} = r!-1,
\end{equation}
as it should be in the usual case of standard (ideal) Bose gas model.
Thus, the expressions (\ref{<(a^+)^r_a^r>})-(\ref{lambda_r_as})
constitute our {\it exact results} for the $r$-particle momentum
distributions and $r$-th order ($r\ge 2$) correlation function
intercepts, along with their large momentum asymptotics, established
in the symmetric TD type $q$-Bose gas model.
\vspace{12mm}
\section{Concluding remarks}
The results obtained in this letter (presented in Eqs.~(\ref{<a^dag_a>}), (\ref{<(a^+)^r_a^r>}), (\ref{lambda_r}), (\ref{lambda_r_as}))
represent the third particular case from among different deformed analogs of Bose gas model
wherein the exact expressions for the $r$-particle distribution
functions and for the respective $r$th order correlation functions
intercepts have been derived. It is of interest to compare (the form
of) these formulas with the two analogous previous results for the
$p,q$-Bose gas model and for the $\mu$-deformed analog of Bose gas
model. Whereas in ref.~\cite{AdGa} the $r$th order correlation intercept $\lambda^{(r)}_{p,q}+1$
appears as fully factorized one-term expression depending on $e^x\equiv e^{\beta\hbar\omega}$
and $p,\,q$, the analogous results for $\lambda^{(r)}_\mu+1$ in~\cite{GaMi2012} and $\lambda^{(r)}_q+1$
in the present work are obtained as non-factorized expressions consisting of $r+1$ terms, each
of which formed from elementary functions of $e^{\beta\hbar\omega}$ in this letter (see eq.~(\ref{lambda_r})
above), but, each one containing special function (Lerch transcendent) in ref.~\cite{GaMi2012}.
Anyway, with these particular three models of deformed Bose gas at hands,
we have to mention in conclusion that now an interesting problem arises of describing
the whole class of deformed Bose gas models which admit the obtaining of similar exact results.
\section*{Acknowledgements}
This work was partly supported by the Special Program of the Division of
Physics and Astronomy of NAS of Ukraine and by the Grant (Yu.A.M.)
for Young Scientists of NAS of Ukraine (No.~0113U004910).
|
\section{Introduction}
In this paper we compare an algebra coming from knot theory with an algebra defined using elliptic curves over finite fields. We use this to derive results about polynomials associated to algebraic knots, which are iterated torus knots that arise from a singular point of a planar algebraic curve. Below we briefly introduce both algebras and then give a precise statements of the results.
\subsection{The Homflypt skein algebra}
The \emph{framed Homflypt skein module }$H(M)$ of an oriented $3$-manifold $M$ consists of $R$-linear combinations of framed oriented links in $M$ up to isotopy, modulo the linear `skein relations'
\bc
$\Xor - \Yor=(s-s^{-1})\ \Ior$ \qquad (Switch and smooth)\\[2mm]
$\Rcurlor=v^{-1}\ \Idor\ ,\qquad \Lcurlor =v\ \Idor$ \qquad (Framing change)\\[2mm]
\ec
using the ring $R={\mathbb{Z}}[v^{\pm1},s^{\pm1}]$ with denominators $s^r-s^{-r},r>0$ as coefficient ring.
In these relations we use the convention that the framing is defined by a ribbon for each component of the link, and the skein relations apply in a ball in which the framing ribbons lie parallel to the strands as seen. This is sometimes termed the `local blackboard framing'. The second framing relation then introduces $v^{\pm1}$ when a framing ribbon acquires a single extra twist.
\begin{remark}
The term \emph{skein module} indicates that $H(M)$ is a module over $R$. Since many skeins are also modules over more complicated $R$-algebras, Morton and his co-workers often simply use the term \emph{skein}, and suppress the qualifications \emph{framed Homflypt} also when the context is clear.
\end{remark}
When $F$ is an orientable surface and $M=F\times I$ we adopt the notation $H(F)$ in place of $H(F\times I)$, and refer to $H(F)$ as the skein (module) of the surface $F$. Framed links in $F\times I$ can be represented by diagrams in $F$, with the global `blackboard framing' from $F$, up to isotopy and Reidemeister moves $ R_{II}, R_{III}$. We can then regard elements of $H(F)$ as diagrams in $F$ modulo $ R_{II}, R_{III}$ and the skein relations.
The skein $H(F)$ forms an algebra over $R$ under the product induced by placing one copy of $F\times I$ containing an element $D\in H(F)$ on top of another copy containing $E\in H(F)$ to determine their product $DE$ (see Section \ref{sec_diagrams}).
In the case $\mathcal C=H(F)$ where $F$ is the annulus $S^1\times I$ this algebra is commutative and has been studied for some time. A recent account of some of its properties can be found in \cite{MM08}. It has an interpretation as the algebra of symmetric functions in a large number of commuting variables $x_1,\ldots,x_N$, and contains an element $P_m$ for each $m$ corresponding to the power sum $x_1^m+\cdots+x^m_N$. One representation of this element, due originally to Aiston \cite{Ais96}, is a multiple of the sum of $m$ explicit closed $m$-braids (see Section \ref{sec_annulus}).
The case of primary interest to us is the skein $H=H(T^2)$ for the surface $T^2$. As an algebra $H$ is non-commutative, and can be generated by elements $P_{{\mathbf{x}}}$, one for each ${{\mathbf{x}}}\in {\mathbb{Z}}^2 \setminus \{0\}$, corresponding to free homotopy classes of curves in $T^2$.
For a primitive ${{\mathbf{x}}}=(m,n)\in {\mathbb{Z}}^2$ we represent $P_{{\mathbf{x}}}$ by the oriented embedded $(m,n)$ curve on $T^2$. It is an immediate consequence of the switch and smooth skein relation that the commutator $[P_{(1,0)}, P_{ (0,1)}]$ satisfies $$[P_{(1,0)}, P_{ (0,1)}]=(s-s^{-1})P_{(1,1)}.$$
The same switching and smoothing relation shows that $[P_{\bf x},P_{\bf y}] =(s-s^{-1})P_{{\bf x}+{\bf y}}$ when the primitive curves $\bf x$ and $\bf y$ cross once in the positive direction. Our main result is the following satisfyingly simple extension of these commutation relations (see Theorem \ref{thm_commutationrelations}).
\begin{theorem_noname}{\rm (Global switch and smooth)}. The commutator $[P_{{\mathbf{x}}},P_{{\mathbf{y}}}] $ in $H$ satisfies
\begin{equation}\label{eq_commintro}
[P_{\bf x},P_{\bf y}] =(s^d-s^{-d})P_{{\bf x}+{\bf y}}
\end{equation}
where $d=\det[{\bf x \ y}]$ is the signed crossing number of $\bf x$ with $\bf y$.\\
\end{theorem_noname}
In making the statement of Theorem \ref{thm_commutationrelations} we must also specify $P_{m\bf x}$ for any multiple of a primitive $\bf x$. This is defined by decorating the embedded curve $\bf x$ by the element $P_m$ from the skein $\mathcal C$ of the annulus. (See Definition \ref{def_px}.)
When $\bf x, \bf y$ and ${\bf x}+{\bf y}$ are all primitive the commutator $[P_{\bf x},P_{\bf y}] $ may be regarded as the difference of a switch of the curves $\bf x$ and $\bf y$, with $P_{{\bf x}+{\bf y}}$ appearing as the simultaneous smoothing at the $k$ crossings.
The proof of Theorem \ref{thm_commutationrelations} relies on a result in \cite{MM08} to establish that $$[P_{(m,0)},P_{(0,1)}] =(s^m-s^{-m})P_{(m,1)}.$$ Direct skein manipulation shows that $$[P_{(0,-1)},P_{(m,1)}] =(s^m-s^{-m})P_{(m,0)},$$ using Aiston's representation of $P_m$. The full theorem is proved from these two cases in Section \ref{sec_relations} using induction on $\det[ {\mathbf{x}} \ {\mathbf{y}}]$. Using a theorem of Przytycki in \cite{Prz92}, we prove the following (see Corollary \ref{cor_basis}).
\begin{theorem_noname}\label{thm_intropres}
As an abstract algebra, $H(T^2)$ has a presentation with generators $P_{\mathbf{x}}$ for ${\mathbf{x}} \in \mathbb{Z}^2$ and relations given by equation (\ref{eq_commintro}).
\end{theorem_noname}
In general, if $M$ is a $3$-manifold, then $H(M)$ is a module over the algebra $H(\partial M)$. In particular, the skein module $H(S^3 \setminus K)$ of a knot complement is a module over the algebra $H(T^2)$. In the case where $K$ is the trivial knot we use earlier work of Morton and coauthors \cite{MH02, HM06} to give an explicit description of $\mathcal{C} := H( S^1\times D^2)$ as a module over $H(T^2)$. In particular, as a module over the (commutative) `horizontal' subalgebra generated by $\{P_{m,0}\mid m \in \mathbb{Z}\}$ it is simultaneously diagonalizable with distinct eigenvalues. Over the `vertical' subalgebra generated by $\{P_{0,n}\mid n \in \mathbb{Z}\}$ the module $\mathcal{C}$ is free of rank 1. We give explicit formulas for this action in Theorem \ref{thm_Cactionfull}.
These theorems can be viewed as analogues of the work of Frohman and Gelca in \cite{FG00}. They give a presentation of the Kauffman bracket skein algebra $K_q(T^2)$ of the torus $T^2$ and describe its action on the (Kauffman bracket) skein of the annulus. It turns out that the algebra $K_q(T^2)$ is the $t=q$ specialization of the $\mathfrak{sl}_2$ spherical double affine Hecke algebra introduced by Cherednik in \cite{Che95}, which is an algebra that has attracted much attention in representation theory. We show that an analogous statement is true for the Homfly skein algebra $H$.
\subsection{The elliptic Hall algebra}
Let $X$ be a smooth elliptic curve over a finite field. In \cite{BS12}, Burban and Schiffmann gave an explicit presentation for the \emph{elliptic Hall algebra} $\mathcal{E}_{q,t}$, which is the Drinfeld double of the Hall algebra of the category $Coh(X)$ of coherent sheaves over $X$. They show that the structure constants are Laurent polynomials in $q,t$, where $q^2, t^2$ are the eigenvalues of the Frobenius operator on the first $l$-adic cohomology group of $X$. We may therefore view $q,t$ as formal parameters.
It turns out that $\mathcal{E}_{q,t}$ is closely related to a number of algebras that have been studied recently under different names. For example, this algebra (or one of its close cousins\footnote{By `cousin' we mean either the `positive half' $\mathcal{E}^+_{q,t}$ or a central extension of $\mathcal{E}_{q,t}$.}) has appeared as the following:
\begin{itemize}
\item a generalized quantum affine algebra in \cite{DI97},
\item `a $(q,\gamma)$ analog of the $W_{1+\infty}$ algebra' in \cite{Mik07}
\item the `shuffle algebra' of \cite{FT11},
\item the `spherical $\mathfrak{gl}_\infty$ double affine Hecke algebra' in \cite{SV11}, (see also \cite{FFJMM11a})
\item the `quantum continuous $\mathfrak{gl}_\infty$' in \cite{FFJMM11a}
\item an algebra of operators on $\oplus K^T(\mathrm{Hilb}_n(\mathbb{C}^2))$ in \cite{SV13}. (See also \cite{FT11}, \cite{FFJMM11a}.)
\end{itemize}
As a consequence of Theorem \ref{thm_intropres}, we add another algebra to this list (see Theorem \ref{thm_HisotoE}).
\begin{theorem_noname}\label{thm_introiso}
The algebra $H$ is isomorphic to $\mathcal{E}_{q,q}$. In particular, any knot $K \subset S^3$ provides a module over the algebra $\mathcal{E}_{q,q}$.
\end{theorem_noname}
This theorem may seem surprising because the definitions of $\mathcal{E}_{q,t}$ and $H$ seem completely unrelated. One rough heuristic explanation for this isomorphism is as follows. When all parameters are set equal\footnote{``Setting all parameters equal to one'' is a statement that requires some care to be made precise, but we will not discuss this.}
to $1$, the Homflypt skein algebra of a surface $F$ surjects onto $\mathcal{O}_{F,N} := \mathcal{O}({\mathrm{Char}}(\pi_1(F), \GL_N(\mathbb{C}))$, the ring of functions on the scheme parameterizing representations of $\pi_1(F)$ up to equivalence.
The $\mathfrak{gl}_N$ spherical double affine Hecke algebra $\mathrm {S} \H_{q,t}^N$ is an algebra depending on 2 parameters $q,t \in \mathbb{C}^*$, and when $q=t=1$, this algebra is isomorphic to $\mathcal{O}_{T^2,N}$. Finally, Schiffmann and Vasserot show in \cite{SV11} that $\mathcal{E}_{q,t}$ surjects onto $\mathrm {S} \H_{q,t}^N$ for any $N$. Summarizing, when the parameters are set to 1, both $H$ and $\mathcal{E}_{1,1}$ surject onto the (commutative) algebras $\mathcal{O}_{T^2,N}$ for any $N$.
We also remark that Theorem \ref{thm_introiso} seems analogous to Kontsevitch's homological mirror symmetry for an elliptic curve (see \cite{PZ98} for a precise statement and proof in this case). Very roughly, in this case mirror symmetry predicts that the derived category $D^b(Coh(X))$ of coherent sheaves over an elliptic curve $X$ over $\mathbb{C}$ is
equivalent to the Fukaya category of the symplectic torus $S^1 \times S^1$. Objects in this Fukaya category are simple closed curves `decorated' by a representation of $U(n)$, and under this equivalence, a sheaf on $X$ of slope $m/n$ is sent to the $(m,n)$ curve on $S^1\times S^1$, decorated by a certain representation. This is reminiscent of the isomorphism in Theorem \ref{thm_introiso}; however, it seems that there are significant difficulties (at best) in attempting to prove this isomorphism using mirror symmetry.
\subsection{Algebraic knots}
We next discuss an application of the isomorphism $H \cong \mathcal{E}_{q,q}$. If $K$ is an iterated cable of the unknot, then it is straightforward to give a \emph{cabling formula} for the $\lambda$-colored Homflypt polynomial $J^H(K,\lambda)$ of $K$ in terms of the action of the algebra $H$ on $\mathcal{C}$ and the Homflypt evaluation map ${\rm{ev}}^H: \mathcal{C} \to H(S^3)$ (see Proposition \ref{prop_cabling}). Using the isomorphism of Theorem \ref{thm_HisotoE} and a theorem in \cite{SV13}, we show that all objects used in the cabling formula have $t$-deformations. This allows us to define 3-variable polynomials $J^\mathcal{E}(K, \lambda)$ (see Definition \ref{def_JE}) that specialize to the Homflypt polynomial $J^H(K,\lambda)$. (Technically, these are rational functions for generic $t$ - see Remark \ref{rmk_rational}.)
In \cite{Sam14} similar polynomials were defined using the $\mathfrak{sl}_2$ spherical double affine Hecke algebra, and Cherednik and Danilenko generalized these to arbitrary $\mathfrak{g}$ shortly afterwards in \cite{CD14} (with a slightly simpler construction). For $\mathfrak{g} = \mathfrak{gl}_N$, we recall their definition of the polynomials $J^N(K,\lambda)$ in Definition \ref{def_JN}. We prove the following
in Theorem \ref{thm_iteratedcable}:
\begin{theorem_noname}
If $K$ is an iterated cable of the unknot, we have the following specializations:
\begin{eqnarray*}
J^\mathcal{E}(K,\lambda; q,t,u)\Big|_{q=s^{-2}, t=s^{-2}, u = v^2} &=& v^\bullet s^\bullet J^H(K,\lambda; v,s) \\
{ } u^\bullet J^\mathcal{E} (K,\lambda; q,t,u)\Big|_{u = t^N} &=& q^\bullet t^\bullet J^N(K,\lambda; q,t)
\end{eqnarray*}
(where the powers denoted by ``$\bullet$'' depend on $K$ and $\lvert \lambda \rvert $, but not on $N$). In particular, the Connection Conjecture \cite[Conj.\ 2.4(i)]{CD14} is true.
\end{theorem_noname}
We remark that the definition of $J^\mathcal{E}(K)$ depends \emph{a priori} on a choice of presentation of $K$ as an iterated cable of the unknot. It isn't clear whether different choices for this presentation produce the same polynomials. (However, it was shown that certain different choices do produce the same polynomials $J^N(K)$ in \cite[Thm.\ 2.1(i)]{CD14}.)
At this point, one natural question is whether the modules over $\mathcal{E}_{q,q}$ associated to knots can be deformed to modules over $\mathcal{E}_{q,t}$ for any $t \in \mathbb{C}^*$ (in the style of \cite{BS14}). A positive answer may be interesting even for the module $\mathcal{C}$ associated to the unknot. A central extension $\mathcal{E}_{q,t}^c$ of $\mathcal{E}_{q,t}$ acts on $\Lambda$, which is the `Fock space' with basis given by the set of all partitions (see, e.g., \cite{FFJMM11a} or \cite{SV13}). The module $\mathcal{C}$ has a natural basis indexed by \emph{pairs} of partitions, and so can be identified with $\Lambda \otimes_\mathbb{C} \Lambda$ as a vector space. It does not seem that a module of this `size' has appeared in the recent literature about the representation theory of $\mathcal{E}_{q,t}$, but it does seem reasonable to expect that $\mathcal{C}$ deforms. We hope to address this question in future work.
\subsection{Summary}
We now summarize the contents of the paper. In Section \ref{sec_background} we provide brief background and definitions. We then give a presentation for the algebra $H$ in Section \ref{sec_relations}. The module $\mathcal{C}$ associated to the solid torus is described explicitly in Section \ref{sec_solidtorus}. In Section \ref{sec_ehall} we prove that $H$ is the $t=q$ specialization of the elliptic Hall algebra $\mathcal{E}_{q,t}$. In Section \ref{sec_adaptations} we describe different specializations of the Homflypt skein relations and their relation to other knot invariants. In Section \ref{sec_iteratedcables} we use the elliptic Hall algebra to construct a 3-variable polynomial that specializes to the Homflypt polynomial for iterated torus knots and to the 2-variable polynomials for $\mathfrak{gl}_N$ constructed in \cite{CD14}.
\noindent\textbf{Acknowledgments:} We would like to thank I. Cherednik for helpful discussions about \cite{CD14}, and D. Muthiah for help with using \texttt{Sage} and for several enthusiastic conversations. We also thank Y. Berest, F. Bergeron, A. Oblomkov, V. Shende, O. Schiffmann, and E. Vasserot for enlightening discussions of their work and/or the present paper. The authors also benefited from the Research in Pairs program in Oberwolfach, where final editing on this paper was completed.
\section{Background}\label{sec_background}
In this section we first establish notation for `quantum numbers' and partitions, and we then give definitions and background for Homflypt skein modules.
\subsection{Notation}\label{sec_notation}
In this paper we will use the coefficient ring $R = \mathbb{C}[ v^{\pm 1},s^{\pm 1}, (s-s^{-1})^{-k}]$ (where $k$ ranges over $\mathbb{N}$), and we use the following `quantum numbers:'
\[ [d] := \frac{s^d-s^{-d}}{s-s^{-1}},\quad \quad \{d\} := s^d-s^{-d},\quad \quad \{d\}^+ := s^d + s^{-d}
\]
Let $\lambda = (\lambda_1,\cdots,\lambda_k)$ be a partition of length $k$. We will represent partitions by Young diagrams using the continental convention, so that a nonempty partition always has a box in its lower left corner. We will write $\lambda$ both for a partition and for its representation as a Young diagram. If $x$ is a box in $\lambda$ in row $i$ and column $j$, we will use the standard notations
\begin{equation}\label{eq_pnotation}
a(x) = \lambda_i - j,\quad l(x) = \lvert \{n \mid \lambda_n \geq j\}\rvert - i,\quad a'(x) = j,\quad l'(x) = i
\end{equation}
Here $a(x)$, $l(x)$, $a'(x)$, and $l'(x)$ are the arm length, leg length, coarm length, and coleg length of $x$, respectively. Graphically, they are the number of cells strictly to the right, strictly above, weakly to the left, and weakly below $x$, respectively. (For a picture see, e.g. \cite[Fig.\ 1]{SV13}.) We will also use the \emph{content} $c(x)$ and \emph{hook length} $hl(x)$, which are defined by
\begin{equation}
c(x) := j - i,\quad hl(x) := a(x) + l(x) + 1
\end{equation}
\subsection{Homflypt skein modules}
Let $M$ be an oriented 3-manifold. A \emph{framed oriented link} in $M$ is (an ambient isotopy class of) a smooth embedding $\sqcup (S^1 \times [0,1]) \hookrightarrow M$, where each copy of $S^1$ is oriented. Let $\mathcal L(M)$ be the free $R$-module spanned by framed oriented links in $M$, and let $\mathcal L'(M) \subset \mathcal L(M)$ be the $R$-submodule generated by the skein relations in Figure \ref{fig_skeinrelations}. (A \emph{skein relation} is a formal linear combination of links differing only inside a 3-ball as shown in the figure.)
\begin{center}
\begin{figure}[ht]
\bc
$\Xor - \Yor=(s-s^{-1})\ \Ior$ \qquad (Switch and smooth)
$\Rcurlor=v^{-1}\ \Idor\ ,\qquad \Lcurlor =v\ \Idor$ \qquad (Framing change)
\ec
\caption{The Homflypt skein relations}\label{fig_skeinrelations}
\end{figure}
\end{center}
\begin{definition}\label{def_homflyskeinmod}
The \emph{framed Homflypt skein module} $H(M)$ of the manifold $M$ is the quotient $\mathcal L(M) / \mathcal L'(M)$.
\end{definition}
In general, the Homflypt skein $H(M)$ has four important properties:
\begin{enumerate}
\item The skein $H(M)$ is graded by the homology group $\textrm{H}_1(M)$, since each skein relation involves only links in the same homology class.
\item\label{item_mult} If $M = F \times [0,1]$, then $H(M)$ is an algebra (which is typically noncommutative). The product is given by stacking links (see Section \ref{sec_diagrams} for an example). The grading is additive under the product.
\item If $\partial M = F$, then $H(M)$ is a module over the algebra $H(F \times [0,1])$. The action is given by ``pushing links from the boundary into $M$.'' If $a \in H(F \times [0,1])$ and $m \in H(M)$ are homogeneous, then $deg(a\cdot m) = \iota(deg(a)) + deg(m)$, where $\iota_*: \textrm{H}_1(F) \to \textrm{H}_1(M)$ is the map on homology induced by the inclusion $\iota:F \hookrightarrow M$.
\item An oriented embedding $f:M \hookrightarrow N$ induces an $R$-linear map $f_*:H(M) \to H(N)$. When $f$ is a homeomorphism the map $f_*$ is an isomorphism.
\end{enumerate}
\begin{definition}
Write $H := H(T^2)$ for the (framed Homflypt) skein of the torus $T^2$, and $\mathcal{C} := H(S^1 \times I^2)$ for the skein of the solid torus, with a choice of explicit homeomorphism from the solid torus to $(S^1\times I)\times I$ to specify $\mathcal{C}$ as an algebra (see the beginning of Section \ref{sec_solidtorus}).
\end{definition}
In this paper we study the algebra structure of $H$ and the $H$-module structure of $\mathcal{C}$. The algebra $H$ is graded by ${\mathbf{x}} \in \mathbb{Z}^2= H_1(T^2)$. Set \[H=\oplus_{{\mathbf{x}}\in\mathbb{Z}^2} H_{{\mathbf{x}}}\] where the degree of a link is given by its homology class. Similarly the algebra $\mathcal{C}$ has a $\mathbb{Z}$ grading. There is an action of $\SL_2(\mathbb{Z})$ by algebra automorphisms on $H$ induced by the mapping class group action on $T^2$.
\subsection{Diagrammatic representations of links in $T^2$.}\label{sec_diagrams}
Use the classical presentation of $T^2$ as a square with pairs of edges identified,
\bc \Torus
\ec as indicated by the coloured edges in the diagram above. Then a link in $T^2$ can be drawn as a diagram in the square with some arcs meeting the coloured edges in matching pairs.
For example here are diagrams in $T^2$ of a 2-component link
\bc $D\quad =\quad$\torusD\ec and a 1-component link \bc $E\quad =\quad$\torusE\ec along with diagrams for the products
\bc$DE\quad=\quad$\torusDE\ and \ $ED\quad=\quad$\torusED\ .\ec
Elements in $H$ are linear combinations of diagrams, so for example the commutator $[D,E]$ is the element $DE-ED\in H$.\\[2mm]
Using the convention that the square has sides along the usual axes we can draw the embedded $(m,n)$ curve in $T^2$ as \bc
$P_{\mathbf{x}}$\quad =\quad \pic{minusonek}{.5} \ec meeting the vertical edges in $m$ points and the horizontal edges in $n$ points. Here $(m,n) =(-1,k)={\mathbf{x}}\in\mathbb{Z}^2$ is the homology class of the curve, so that $P_{\mathbf{x}}$ lies in the graded subspace $H_{\mathbf{x}}\subset H$. The curve will cross the edges of the square according to the signs of $m$ and $n$; in this diagram $m=-1$ and $n=k>0$.
This defines elements $P_{\mathbf{x}}$ for each primitive ${\mathbf{x}}\in\mathbb{Z}^2$, in other words where $m$ and $n$ are coprime. As a simple example we have
\bc $ P_{1,0}$ \ = \ \Ponezero, \quad $P_{0,1} $\ = \ \Pzeroone \ .\ec Their commutator then satisfies
\begin{eqnarray*} [P_{1,0},P_{0,1}]&=& \Ppos\ -\ \Pneg\\[1mm]
&=& (s-s^{-1})\ \Poneone\\[1mm]
&=&(s-s^{-1} )P_{1,1},
\end{eqnarray*} using the switch and smooth skein relation.
Clearly if the embedded curve $P_{\mathbf{x}}$ crosses the embedded curve $P_{\mathbf{y}}$ once in the positive sense then the switch and smooth relation applied at the crossing results in the embedded curve $P_{{\mathbf{x}}+{\mathbf{y}}}$ and their commutator satisfies the equation
\[ [P_{\mathbf{x}},P_{\mathbf{y}}]=(s-s^{-1})P_{{\mathbf{x}}+{\mathbf{y}}}.\]
\begin{remark} The signed number of crossings of curves with homology ${\mathbf{x}}$ and ${\mathbf{y}}$ is equal to the determinant $\det[{\mathbf{x}}\ {\mathbf{y}}]$ of the $2\times 2$ matrix with columns ${\mathbf{x}}$ and ${\mathbf{y}}$.
\end{remark}
Before giving the definition of the elements $P_{(m,n)}\in H$ where $m$ and $n$ are not coprime we use diagrams in the square meeting the edges in matched pairs to represent elements of some other skeins.
\subsection{The skein $\mathcal{C}$ of the annulus}\label{sec_annulus}
If we identify the two horizontal edges only we get an annulus. Elements of its skein $\mathcal{C}$ can be represented by diagrams which do not meet the vertical edges. Any such diagram also gives a diagram in $T^2$ by identifying the two vertical edges. The elements arising in this way in $H$ will lie in the graded part $H_{(0,k)}$ where $k$ is the signed number of crossings with the horizontal edge.
For $i,j \ge0$, and $i+j+1=k>0$ define elements $A_{i,j}\in \mathcal{C}$ diagrammatically by\\[1mm]
\bc $A_{i,j}$\quad=\quad \labellist\small
\pinlabel {$i$} at 165 410
\pinlabel {$j$} at 275 410
\endlabellist\pic{aijannulus}{.50} \ .\\[4mm]
\ec
Set \[X_k:=\sum_{i=0}^{k-1} A_{i,j}\in\mathcal{C}\]
Aiston \cite{Ais96} showed that $X_k$ represents a multiple of the $k$th power sum when interpreting $\mathcal{C} $ in terms of symmetric functions. Further work by Aiston, Morton and subsequent collaborators has identified other nice algebraic and skein theoretic properties of $X_k$ and more particularly the exact power sum $P_k:=(s-s^{-1})/(s^k-s^{-k}) X_k$, which we shall also use in this paper.
\begin{definition}\label{def_px}
For $k > 0$ we define $P_{0,k}$ to be the result of decorating the $(0,1)$ curve in $T^2$ by $P_k$ from $\mathcal{C}$, or equivalently \[P_{0,k}=\frac{s-s^{-1}}{s^k-s^{-k}}\sum_{i=0}^{k-1} A_{i,j}\] after identifying the vertical edges as well in the square to give diagrams in $T^2$.
To define $P_{\mathbf{x}}$ in general write ${\mathbf{x}}=k(m,n)$ with $m$ and $n$ coprime and $k>0$. Then \emph{decorate} the embedded $(m,n)$ curve by the element $P_k\in\mathcal{C}$ from the skein of the annulus to specify $P_{\mathbf{x}}$.
\end{definition}
\subsection{Decoration and framing}
The term \emph{decoration} in our definition above has been used widely by Morton and others in describing satellite invariants of framed knots and links in manifolds.
What is meant by \emph{decorating} a framed oriented curve $K$ in $M$ by a framed curve $Q$ in $S^1\times I$ is to embed the thickened annulus $S^1\times I$ on the neighbourhood of $K$, respecting orientation \emph{and} framing. The image of $Q$ in $M$ is the satellite of $K$ with pattern $Q$, and is referred to as $K$ \emph{decorated by} $Q$. It carries a framing which is determined by the framing of $Q$.
The embedding from $S^1\times I$ to $M$ depends on the framing and orientation of $K$. It induces a linear map of skeins from $\mathcal{C}= H(S^1\times I)$ to $H(M)$, with image denoted by $\mathcal{C}_K$.
In the case where $M=F\times I$ and $K$ is an \emph{embedded} curve in the surface $F$, which is framed by its neighbourhood in $F$, the induced linear map is an algebra homomorphism. The image $\mathcal{C}_K$ is then a subalgebra of $H(F)$.
When $F=T^2$ there is an embedded curve $P_{\mathbf{x}}$ for each \emph{primitive} ${\mathbf{x}}\in\mathbb{Z}^2$. We have defined $P_{k{\mathbf{x}}}$ above to be the image of $P_k\in\mathcal{C}$ under the homomorphism $\mathcal{C}\to H$ determined by the framed embedded curve $P_{\mathbf{x}}$.
\begin{remark}
When $k=1$ the result $P_{\mathbf{x}}$ is just the oriented ${\mathbf{x}} = (m,n)$ curve, as defined earlier. For $P_{0,n}$ with $n<0$ we decorate the $(0,-1)$ curve by $P_{-n}$. Also, since $SL(2,\mathbb{Z})$ acts on $T^2$ preserving the framing annuli of embedded curves we see that its action on $H$ permutes the elements $P_{\mathbf{x}}$.
\end{remark}
\subsection{Hecke algebras of type $A$}
There is a family of algebras modelled by the skein of the square with diagrams meeting the horizontal edges in matching pairs, but not meeting the vertical edges. Where there are $n$ meeting points on each edge, all oriented upwards, this skein models the Hecke algebra $H_n$ as in \cite{MT90} and \cite{AM98}. With $n$ upwardly and $p$ downwardly oriented meeting points the skein is the mixed algebra $H(n,p)$ in \cite{KM93} and in \cite{MH02}. In both cases the algebra product comes from placing one square above another.
\subsection{Affine Hecke algebras of type $A$}
Use the square with the vertical edges identified, to give a cylinder, and take diagrams meeting the horizontal edges in $n$ matching pairs, oriented upwards, to give an algebra closely related to the affine Hecke algebra $\dot{H}_n$.
A description of the affine Hecke algebra in this way is given by Graham and Lehrer \cite{GL03}. They restrict the diagrams to be braids in the cylinder, without closed curves, and with all strings running monotonically in the vertical direction. They then have no need for the parameter $v$, nor the framing change relation. In their work the regular elementary braids $\sigma_i$ represent the elements conventionally called $T_i$ in presentations of $\dot{H}_n$. They give cylindrical braid representatives for the commuting elements in $\dot{H}_n$ known as $X_i$, and for the element $\tau$ which realizes a $1/n$ turn around the cylinder.
Because we are using $X_i$ in this paper in a different context, following the terminology of \cite{MM08}, we
refer here to Graham and Lehrer's braids in red. Their elements $\r{X_i}, i=1,\ldots,n$ are
shown here.
\begin{figure}[ht]\label{Murphycylinder}
\bc $\r{X_i}\quad= \quad $\labellist\small
\pinlabel {$i$} at 210 335
\endlabellist \Murphycylinder\\[4mm]
\ec
\caption{The commuting elements $\r{X_i}$ in the affine Hecke algebra $\dot{H}_n$.}
\end{figure}
Symmetric functions in these are central elements in $\dot{H}_n$. In this context, it is worth extending to the full skein of the cylinder without restricting to the use of braids. In the full skein there is a nice model of the sum $\sum \r{X_i}$ as a sort of commutator of the identity braid with the closed curve $P_{1,0}$. The exact result is the equation
\[\backid\quad - \quad \frontid\quad =\quad (s-s^{-1})\sum_{i=1}^n \r{X_i }\ ,\] proved by applying the switch and smooth relations on the $n$ crossing points of the first diagram in sequence. Both diagrams on the left hand side are then clearly central in the full skein.
\begin{remark} The full skein of the cylinder with $n=1$ is used by Lukac \cite{Luk01}, and by Morton and Hadji \cite{HM06}, under the name $\mathcal A$, in proving results about $\mathcal{C}$.
One relevant result from \cite{MM08} is that by using $P_{k,0}$ in place of $P_{1,0}$ the commutator when $n=1$ is $s^k-s^{-k}$ times the cylindrical $1$-braid going $k$ times around the cylinder. This is illustrated below, with the cylinder drawn as an annulus, in Equation (\ref{cylindercommutator}). It leads immediately to the representation of $ (s^k-s^{-k})\sum_{i=1}^n \r{X_i}^k$ for every $n>1$.
Very close analogues of $\r{X_i}$ in $H_n$ are used in \cite{Mor02, Mor02b} where they represent the Murphy operators. The notation $T_i$ is used in that paper, where a similar construction gives a quick proof that symmetric polynomials in the commuting elements $T_i$ are central in $H_n$.
\end{remark}
It is also worth considering the algebras $\dot{H}_{n,p}$ where the vertical edges are again identified to give a cylinder, and diagrams meet the horizontal edges in $n$ upward and $p$ downward matched pairs of points.
By work of Turaev \cite{Tur88} the elements $X_m$ generate $H(S^1\times [0,1])$ as an algebra over $R$.
\section{A presentation for $H$}\label{sec_relations}
In this section we give a presentation of $H$ using the elements $P_{\mathbf{x}}$ of Definition \ref{def_px}.
\begin{lemma}
The algebra $H$ is generated by the elements $P_{{\mathbf{x}}}$.
\end{lemma}
\begin{proof}
Any element of $H$ can be reduced to a sum of products of knots using the skein relations. Then a knot in $H$ is in some graded piece $H_{j,k}$, and it follows from \cite{Prz92} that the skein module of the annulus surjects onto any particular graded piece $H_{j,k}$ by embedding the annulus onto a neighborhood of the $(j/d, k/d)$ curve (where $d = gcd(j,k)$). The claim then follows from the fact that the $X_m$ generate the skein module of the annulus, which was proved by Turaev in \cite{Tur88}, and the fact that over $R$ the $X_m$ can be written in terms of the $P_{\mathbf{x}}$.
\end{proof}
\begin{theorem}\label{thm_commutationrelations}
The elements $P_{\mathbf{x}}$ for ${\mathbf{x}} \in \mathbb{Z}^2$ satisfy the following commutation relation:
\begin{equation}\label{formula_allrelations}
[P_{\mathbf{x}},P_{\mathbf{y}}] = \{\mathrm{det}[{\mathbf{x}}\, {\mathbf{y}}]\} P_{{\mathbf{x}}+{\mathbf{y}}}
\end{equation}
\end{theorem}
\begin{proof}
We separate this proof into two subsections. In Section \ref{proof_ofsomerelations} we prove the following relations using methods and results of \cite{MM08}:
\begin{align}
\,[P_{k{\mathbf{x}}}, P_{j{\mathbf{x}}}] &= 0\notag\\
\,[P_{1,0}, P_{-1,k}] &= \{k\} P_{0,k}\label{formula_somerelations}\\
\,[P_{1,0}, P_{0,k}] &= \{k\} P_{1,k}\notag
\end{align}
Then in Section \ref{proof_allrelations} we show that the relations (\ref{formula_somerelations}) imply (\ref{formula_allrelations}).
\end{proof}
Before we prove this theorem, we use a theorem in \cite{Prz92} to show that this gives a presentation of $H$. More precisely, let $A$ be the abstract algebra generated by $P_{j,k}$ subject to the relations in Theorem \ref{thm_commutationrelations}.
\begin{corollary}\label{cor_basis}
The natural surjection $A \twoheadrightarrow H$ is an isomorphism.
\end{corollary}
\begin{proof}
We begin by recalling a basis for $H$ constructed in \cite{Prz92}. We pick a linear ordering $\leq$ of elements in $\pi^0 := \mathbb{Z}^2 \setminus \{0\}$ according to the angle with the positive $x$-axis. Then for each $w = (m,n) \in \pi^0$ we pick a representative diagram $B_w$, which is defined to be the curve $A_{k,0}$ (defined in Section \ref{sec_annulus}) inserted on the $(m/k,n/k)$ curve on the torus, where $k = gcd(m,n)$. If $Sym(\pi^0)$ is the symmetric algebra of $R\pi^0$, then Przytycki defines a map $Sym(\pi^0) \to H$ which takes a monomial $w_1 \cdots w_n$ to the product $B(w_1)B(w_2)\cdots B(w_n)$, where $w_1 \leq w_2 \cdots \leq w_n$. Then \cite[Thm.\ 6.2]{Prz92} states that this is an $R$-isomorphism.
We then note that over $R$, the diagram $B(w)$ can be written uniquely in terms of diagrams $P_{c w}$ for $c \in \mathbb{N}$. Then the commutation relations in Theorem \ref{thm_commutationrelations} allow one to order the $P_{\mathbf{x}}$ according to the angle between ${\mathbf{x}}$ and the positive $x$-axis. This completes the proof of the corollary.
\end{proof}
\subsection{Certain commutation relations}
\label{proof_ofsomerelations}
\begin{proof}
The first equation of (\ref{formula_somerelations}) is obvious because the two elements in question lie in parallel annuli.
To prove the second equation, we work with diagrams in the square with edges identified, as discussed above. The commutator $[P_{1,0}, P_{-1,k}]$ is represented by the torus diagrams $$ \pic{back}{.5}\quad-\quad\pic{front}{.5} \ .$$
We introduce intermediate torus diagrams $C_{j,k-j}, j=0,\ldots.k$, in which the first $j$ crossings on the $(1,0)$ curve in $$\pic{back}{.5}$$ are switched from over to under.
Thus
\bc
\labellist\small
\pinlabel {$j$} at 140 315
\pinlabel {$k-j$} at 240 315
\endlabellist $C_{j,k-j}\quad =\quad $ \pic{cijtorus}{.5}\ .\\[6mm]\ec
Then $C_{0,k}=P_{1,0} P_{-1,k}$ and $C_{k,0}=P_{-1,k}P_{1,0} $.
The skein relation applied at the circled crossing on $C_{j,k-j}$ gives
$$C_{j,k-j}-C_{j+1,k-j-i} = (s-s^{-1}) D_{j,k-j-1}.$$
Here
\bc
\labellist\small
\pinlabel {$j$} at 142 325
\pinlabel {$i$} at 250 325
\endlabellist $D_{j,i}\quad =\quad$ \pic{smoothedtorus}{.5}\\[6mm]\ec
is the smoothed diagram with $j$ crossings on the string from the left-hand vertical edge and $i$ crossings on the string to the right-hand vertical edge.
Then the commutator $[P_{1,0}, P_{-1,k}] $ is represented by
\begin{eqnarray*} C_{0,k}-C_{k,0} &=& \sum_{j=0}^{k-1} (C_{j,k-j}-C_{j+1,k-j-1})\\
&=& (s-s^{-1}) \sum_{j=0}^{k-1} D_{j,k-j-1}.
\end{eqnarray*}
Now the diagram $D_{j,i}$ in $T^2$ can be isotoped by moving $j$ crossings horizontally to the left to get an equivalent diagram which does not meet the vertical edge. The result is the closed braid diagram \bc
\labellist\small
\pinlabel {$i$} at 162 420
\pinlabel {$j$} at 270 420
\endlabellist \pic{aijtorus}{.50} \ .\\[5mm] \ec This is the element $A_{i,j}\in\mathcal{C}$ described earlier, running in the direction of the $(0,1)$ curve on $T^2$. This establishes that $[P_{1,0}, P_{-1,k}]$ is represented by the sum of closed braids.$$(s-s^{-1})\sum_{i+j=k-1} A_{i,j}$$ following the $(0,1)$ curve in $T^2$.
Now in \cite{Mor02} it is shown that $$\sum_{i+j=k-1} A_{i,j}= (s^k-s^{-k})/(s-s^{-1}) P_k,$$ and the same result is repeated in \cite{MM08}, when the closed braids $A_{i,j}$ are represented as the closure of different but
conjugate braids. The element $P_k$ decorating the $(0,1)$ curve in $T^2$ represents $P_{0,k}$ in our current notation, and this proves the equation
$$[P_{1,0}, P_{-1,k}] =\{k\}P_{0,k}.$$
\\[2mm]
To prove the final equation of (\ref{formula_somerelations}), we first note that a diagram on $T^2$ cut open along a $(0,1)$ curve gives a diagram in the annulus with some matched point on the two boundary curves. The product $P_{1,0} P_{0,k}$ can thus be represented by the diagram
\bc
{\labellist\small
\pinlabel {$P_k$} at 60 140
\endlabellist} \pic{Ahe} {.60} \ec in the annulus with one input on the left boundary and one output on the right boundary. The commutator $[P_{1,0},P_{0,k}]$ is then represented in the annulus by
\bc {\labellist\small
\pinlabel {$P_k$} at 60 140
\endlabellist} \pic{Ahe} {.60}\ $\ -\ $ {\labellist\small
\pinlabel {$P_k$} at 60 140
\endlabellist} \pic{Aeh} {.600}\ec
We then apply directly Theorem 4.2 of \cite{Mor02b} which gives an equation in the skein of the annulus with one point on each boundary curve. In diagrammatic form, Theorem 4.2 of that paper shows that
\begin{equation} {\labellist\small
\pinlabel {$P_k$} at 60 140
\endlabellist} \pic{Ahe} {.60}\ - {\labellist\small
\pinlabel {$P_k$} at 60 140
\endlabellist} \pic{Aeh} {.600}\ = (s^k-s^{-k})\ \pic{Aaa} {.600}\ .\label{cylindercommutator}\end{equation}
The curve in the right-hand diagram circles the annulus $k$ times, here shown in the case $k=2$. When the annulus follows the $(0,1)$ curve in $T^2$, and its boundary curves are rejoined to form $T^2$ the curve in the right-hand diagram becomes the $(1,k)$ curve in $T^2$ while the other two diagrams yield $P_{1,0} P_{0,k}$ and $P_{0,k} P_{1,0}$ respectively. This leads immediately to the equation
$$[P_{1,0},P_{0,k}]=\{k\}P_{1,k}.$$
\end{proof}
\subsection{All commutation relations}
\label{proof_allrelations}
In this section we prove Proposition \ref{prop_allfromsome}, which shows that the equations (\ref{formula_allrelations}) follow from equations (\ref{formula_somerelations}). In what follows, we write $d({\mathbf{x}}, {\mathbf{y}}) = \det\left[{\mathbf{x}}\,\, {\mathbf{y}}\right]$ for ${\mathbf{x}}, {\mathbf{y}} \in \mathbb{Z}^2$ and $d({\mathbf{x}}) = gcd(m,n)$ when ${\mathbf{x}} = (m,n)$. We will also use the following terminology:
\[
({\mathbf{x}},{\mathbf{y}}) \in \mathbb{Z}^2 \times \mathbb{Z}^2 \textrm{ \emph{satisfies} (\ref{formula_allrelations}) if } [P_{\mathbf{x}},P_{\mathbf{y}}] = \{d({\mathbf{x}},{\mathbf{y}})\}P_{{\mathbf{x}}+{\mathbf{y}}}
\]
The idea of the proof is to induct on the determinant of the matrix with columns ${\mathbf{x}}$ and ${\mathbf{y}}$. To induct, we write ${\mathbf{x}} = {\mathbf{a}} + {\mathbf{b}}$ for carefully chosen vectors ${\mathbf{a}}, {\mathbf{b}}$ and then use the following lemma.
\begin{lemma}\label{lemma_trueforab}
Assume ${\mathbf{a}} + {\mathbf{b}} = {\mathbf{x}}$ and that $({\mathbf{a}},{\mathbf{b}})$ satisfies (\ref{formula_allrelations}). Further assume that the four pairs of vectors $({\mathbf{y}}, {\mathbf{a}})$, $({\mathbf{y}}, {\mathbf{b}})$, $({\mathbf{y}}+{\mathbf{a}},{\mathbf{b}})$, and $({\mathbf{y}}+{\mathbf{b}},{\mathbf{a}})$ satisfy (\ref{formula_allrelations}). Then the pair $({\mathbf{x}},{\mathbf{y}})$ satisfies (\ref{formula_allrelations}).
\end{lemma}
\begin{proof}
By the first assumption, we have $[P_{\mathbf{a}}, P_{\mathbf{b}}] = \{d({\mathbf{a}}, {\mathbf{b}})\}P_{\mathbf{x}}$. We then use the Jacobi identity and the remaining assumptions to compute
\begin{eqnarray*}
-\{d({\mathbf{a}}, {\mathbf{b}})\}[P_{\mathbf{x}},P_{\mathbf{y}}] &=& -[[P_{\mathbf{a}},P_{\mathbf{b}}],P_{\mathbf{y}}]\\
&=& [[P_{\mathbf{y}},P_{\mathbf{a}}],P_{\mathbf{b}}] + [[P_{\mathbf{b}},P_{\mathbf{y}}],P_{\mathbf{a}}]\\
&=& \{d({\mathbf{y}},{\mathbf{a}})\}[P_{{\mathbf{y}}+{\mathbf{a}}},P_{\mathbf{b}}] + \{d({\mathbf{b}},{\mathbf{y}})\}[P_{{\mathbf{b}}+{\mathbf{y}}},P_{\mathbf{a}}]\\
&=& \Big(\{d({\mathbf{y}},{\mathbf{a}})\}\{d({\mathbf{y}}+{\mathbf{a}},{\mathbf{b}})\} + \{d({\mathbf{b}},{\mathbf{y}})\}\{d({\mathbf{b}}+{\mathbf{y}},{\mathbf{a}})\}\Big) P_{{\mathbf{x}}+{\mathbf{y}}}\\
&=:& cP_{{\mathbf{x}}+{\mathbf{y}}}
\end{eqnarray*}
We now use the identity $\{m\}\{n\} = \{m+n\}^+ - \{m-n\}^+$ to simplify the coefficient $c$:
\begin{eqnarray*}
c &=& \{d({\mathbf{y}},{\mathbf{a}})\}\{d({\mathbf{y}}+{\mathbf{a}},{\mathbf{b}})\} + \{d({\mathbf{b}},{\mathbf{y}})\}\{d({\mathbf{b}}+{\mathbf{y}},{\mathbf{a}})\} \\
&=& \{d({\mathbf{y}},{\mathbf{a}}) + d({\mathbf{y}},{\mathbf{b}}) + d({\mathbf{a}},{\mathbf{b}})\}^+ -
\{d({\mathbf{y}},{\mathbf{a}}) - d({\mathbf{y}},{\mathbf{b}}) - d({\mathbf{a}},{\mathbf{b}})\}^+\\
&\,& + \{d({\mathbf{b}},{\mathbf{y}}) + d({\mathbf{b}},{\mathbf{a}}) + d({\mathbf{y}},{\mathbf{a}})\}^+ - \{d({\mathbf{b}},{\mathbf{y}})-d({\mathbf{b}},{\mathbf{a}}) - d({\mathbf{y}},{\mathbf{a}})\}^+\\
&=& \{d({\mathbf{y}},{\mathbf{x}}) + d({\mathbf{a}},{\mathbf{b}})\}^+ - \{d({\mathbf{x}},{\mathbf{y}}) - d({\mathbf{b}},{\mathbf{a}})\}^+\\
&=& \{-d({\mathbf{x}},{\mathbf{y}}) + d({\mathbf{a}},{\mathbf{b}})\}^+ - \{d({\mathbf{x}},{\mathbf{y}}) + d({\mathbf{a}},{\mathbf{b}})\}^+\\
&=& -\{d({\mathbf{a}},{\mathbf{b}})\}\{d({\mathbf{x}},{\mathbf{y}})\}
\end{eqnarray*}
This completes the proof of the lemma.
\end{proof}
We next prove the following elementary lemma (which is a slight modification of \cite[Lemma 1]{FG00}). This lemma is used to make a careful choice of vectors ${\mathbf{a}}, {\mathbf{b}}$ so that the previous lemma can be applied.
\begin{lemma}\label{lemma_diophantine}
Suppose $p,q \in \mathbb{Z}$ are relatively prime with $0 < q < p$ and $p > 1$. Then there exist $u,v,w,z \in \mathbb{Z}$ such that the following conditions hold:
\begin{eqnarray}
u + w &=& p,\quad v + z = q\notag\\
0 < u, w &<& p\label{equation_conditionsonuvwz}\\
uz - wv &=& 1\notag
\end{eqnarray}
\end{lemma}
\begin{proof}
Since $p$ and $q$ are relatively prime, there exist $a,b \in \mathbb{Z}$ with $ bq - ap = 1$. This solution can be modified to give another solution $a' = a + q$ and $b' = b + p$, so we may assume $0 \leq b < p$. We then define
\[
u=b,\quad v=a,\quad w=p-b,\quad z=q-a
\]
By definition, $u,v,w,z$ satisfy the first condition of (\ref{equation_conditionsonuvwz}), and the inequalities $0 \leq b < p$ and $p > 1$ imply the second condition. To finish the proof, we compute
\[
uz - wv = b(q-a) - a(p-b) = bq - ap = 1
\]
\end{proof}
\begin{remark}\label{remark_gl2z}
There is a natural $R$-linear anti-automorphism $\sigma:H \to H$ which ``flips $T^2 \times [0,1]$ across the $y$-axis and inverts $[0,1]$.'' In terms of the elements $P_{a,b}$, we have $\sigma(P_{a,b}) = P_{a,-b}$. We therefore have an \emph{a priori} action of $\mathrm{GL}_2(\mathbb{Z})$ on $H$, where elements of determinant $1$ act by algebra automorphisms, and elements of determinant $-1$ act by algebra anti-automorphisms.
\end{remark}
\begin{proposition}\label{prop_allfromsome}
Suppose $A$ is an algebra with elements $P_{\mathbf{x}}$ for ${\mathbf{x}} \in \mathbb{Z}^2$ that satisfy equations (\ref{formula_somerelations}). Furthermore, suppose that there is a $\GL_2(\mathbb{Z})$ action by (anti-)automorphisms on $A$ as in Remark \ref{remark_gl2z}, and that the action of $\GL_2(\mathbb{Z})$ is given by $\gamma(P_{\mathbf{x}}) = P_{\gamma({\mathbf{x}})}$ for $\gamma \in \GL_2(\mathbb{Z})$. Then the $P_{\mathbf{x}}$ satisfy the equations (\ref{formula_allrelations}).
\end{proposition}
\begin{proof}
The proof proceeds by induction on $\lvert d({\mathbf{x}},{\mathbf{y}})\rvert $, and the base case $d({\mathbf{x}},{\mathbf{y}}) = \pm 1$ is immediate from Remark \ref{remark_gl2z} and one application of the skein relation. Now assume
\begin{equation}\label{assumption1}
\textrm{for all } {\mathbf{x}}',{\mathbf{y}}' \in \mathbb{Z}^2\textrm{ with } \lvert d({\mathbf{x}}',{\mathbf{y}}')\rvert < d({\mathbf{x}},{\mathbf{y}}), \textrm{ we have } [P_{{\mathbf{x}}'},P_{{\mathbf{y}}'}] = \{d({\mathbf{x}}',{\mathbf{y}}')\} P_{{\mathbf{x}}' + {\mathbf{y}}'}
\end{equation}
We would like to show that $[P_{{\mathbf{x}}},P_{{\mathbf{y}}}] = \{d({\mathbf{x}},{\mathbf{y}})\} P_{{\mathbf{x}} + {\mathbf{y}}}$. By Remark \ref{remark_gl2z}, we may assume
\[
{\mathbf{y}} = \left(\begin{array}{c} 0\\r\end{array}\right),\quad {\mathbf{x}} = \left(\begin{array}{c} p\\mathbf{q}\end{array}\right),\quad d({\mathbf{x}}) \leq d({\mathbf{y}}),\quad 0 \leq q < p
\]
If $p=1$, then this equation follows from (\ref{formula_somerelations}), so we may also assume $p > 1$. \\[2mm]
\noindent \emph{Case 1:} Assume $0 < q$.
Let $p' = p / d({\mathbf{x}})$ and $q' = q / d({\mathbf{x}})$ - by the assumption $0 < q$, we see that $d({\mathbf{x}}) < p$, so $p' > 1$. We can therefore apply Lemma \ref{lemma_diophantine} to $p',q'$ to obtain $u,v,w,z \in \mathbb{Z}$ satisfying
\begin{equation}\label{assumption1.5}
uz - vw = 1,\quad uq' - vp' = 1,\quad u + w = p',\quad v + z = q',\quad 0 < u,w < p'
\end{equation}
We then define vectors ${\mathbf{a}}$ and ${\mathbf{b}}$ as follows (the properties listed follow from (\ref{assumption1.5})):
\begin{equation}\label{assumption2}
{\mathbf{a}} := \left( \begin{array}{c} d({\mathbf{x}})u\\ d({\mathbf{x}})v\end{array}\right),\quad
{\mathbf{b}} := \left( \begin{array}{c} d({\mathbf{x}})w\\ d({\mathbf{x}})z\end{array}\right),
\quad {\mathbf{a}} + {\mathbf{b}} = {\mathbf{x}},\quad d({\mathbf{a}}, {\mathbf{b}}) = d({\mathbf{x}})^2
\end{equation}
Using Lemma \ref{lemma_trueforab} and Assumption (\ref{assumption1}), it is sufficient to show that the absolute values of each of $d({\mathbf{a}}, {\mathbf{b}})$, $d({\mathbf{y}},{\mathbf{b}})$, $d({\mathbf{y}},{\mathbf{a}})$, $d({\mathbf{y}}+{\mathbf{a}},{\mathbf{b}})$, and $d({\mathbf{y}}+{\mathbf{b}},{\mathbf{a}})$ are strictly less than $\lvert pr \rvert = d({\mathbf{x}},{\mathbf{y}})$. First,
\[
d({\mathbf{a}},{\mathbf{b}}) = d({\mathbf{x}})^2 \leq d({\mathbf{x}})d({\mathbf{y}}) = d({\mathbf{x}})r < pr
\]
where the last inequality follows from the assumption $0 < q < p$. Second, the absolute values of $d({\mathbf{y}},{\mathbf{b}})$ and $ d({\mathbf{y}},{\mathbf{a}}) $ are strictly less than $pr$ by the inequalities in (\ref{assumption1.5}). Third, we compute
\begin{eqnarray*}
-d({\mathbf{y}}+{\mathbf{a}}, {\mathbf{b}}) &=& -d({\mathbf{y}},{\mathbf{b}}) - d({\mathbf{a}}, {\mathbf{b}})\\
&=& d({\mathbf{x}})wr - d({\mathbf{x}})^2\\
&<& d({\mathbf{x}})wr\\
&\leq& pr
\end{eqnarray*}
Finally, we compute
\begin{eqnarray*}
-d({\mathbf{y}}+{\mathbf{b}}, {\mathbf{a}}) &=& -d({\mathbf{y}},{\mathbf{a}}) - d({\mathbf{b}}, {\mathbf{a}})\\
&=& d({\mathbf{x}})ur + d({\mathbf{x}})^2\\
&\leq& \left(d({\mathbf{x}})u + d({\mathbf{x}})\right)d({\mathbf{y}})\\
&=& (u+1)d({\mathbf{x}})r
\end{eqnarray*}
Therefore, we will be finished once we show that the absolute value of $(u+1)d({\mathbf{x}})$ is strictly less than $p$. We now split into subcases:\\[2mm]
\noindent\emph{Subcase 1a:} If $u + 1 < p'$, then $(u+1)d({\mathbf{x}})r < p'd({\mathbf{x}})r = pr$, and we are done.
\noindent \emph{Subcase 1b:} Assume $u + 1 = p'$. By equation (\ref{assumption1.5}), we have
\[
1 = uq' - vp' = (p'-1)q' - vp' \implies p'(q'-v) = 1 + q' < 1 + p'
\]
Since $p' > 1$, the last inequality implies $q' - v = 1$, which implies $v = q'-1$ and $z = 1$. Since $uz-vw = 1$, this implies $(p'-1)-(q'-1) = 1$, which implies $q' = p'-1$. If we write $g = d({\mathbf{x}})$, we then have
\[
-d({\mathbf{y}} + {\mathbf{b}}, {\mathbf{a}}) = -\det\left[ \begin{array}{cc}g&p -g\\mathfrak{g} + r&p-2g\end{array}\right] = rp + g(g-r) = rp + d({\mathbf{x}})(d({\mathbf{x}})-r)
\]
We already assumed that $d({\mathbf{x}}) \leq r$, and if this inequality is strict, then we are done.
\noindent\emph{Subcase 1c:} In this subcase, we are reduced to proving that the vectors ${\mathbf{y}} = (0,r)$ and ${\mathbf{x}} = (rp', rp'-r)$ satisfy (\ref{formula_allrelations}). First, suppose $r = 1$. Then there is a matrix in $\SL_2(\mathbb{Z})$ that fixes ${\mathbf{y}}$ and sends ${\mathbf{x}} \mapsto (p', -1)$. Therefore, if $r=1$, the second equation of (\ref{formula_somerelations}) implies that $({\mathbf{x}},{\mathbf{y}})$ satisfies (\ref{formula_allrelations}).
Now we assume $r > 1$ and replace our previous choice of ${\mathbf{a}}$ and ${\mathbf{b}}$ with a choice which is better adapted to this particular subcase. We define
\[
{\mathbf{a}} := \left(\begin{array}{c} 0\\-1 \end{array}\right),\quad {\mathbf{b}} := \left(\begin{array}{c} rp'\\rp' - r + 1 \end{array}\right)
\]
Since $r > 1$, it is clear that the absolute values of the determinants of the matrices $[{\mathbf{a}} , {\mathbf{b}}]$, $[{\mathbf{y}},{\mathbf{a}}]$, $[{\mathbf{y}}+{\mathbf{a}},{\mathbf{b}}]$, and $[{\mathbf{y}}+{\mathbf{b}},{\mathbf{a}}]$ are all less than $r^2p'$. This together with Assumption (\ref{assumption1}) shows that these pairs of vectors satisfy (\ref{formula_allrelations}). Finally, since $rp'-r$ and $rp'-r+1$ have different parities, we see that $d({\mathbf{b}}) \not= d({\mathbf{x}}) = r$. This together with Subcase 1b shows that $({\mathbf{y}},{\mathbf{b}})$ satisfies (\ref{formula_allrelations}). Then Lemma \ref{lemma_trueforab} finishes the proof of this subcase, which finishes the proof of Case 1. \\[2mm]
\noindent \emph{Case 2:} In this case we assume $q=0$. We define ${\mathbf{a}}, {\mathbf{b}}$ similarly to Subcase 1c, so we have
\[
{\mathbf{y}} = \left(\begin{array}{c} 0\\r\end{array}\right),\quad {\mathbf{x}} = \left(\begin{array}{c} p\\0\end{array}\right),\quad {\mathbf{a}} := \left(\begin{array}{c} 0\\-1 \end{array}\right),\quad {\mathbf{b}} := \left(\begin{array}{c} p\\ 1 \end{array}\right)
\]
If $r=1$, then the third equation of (\ref{formula_somerelations}) implies that the pair $({\mathbf{x}},{\mathbf{y}})$ satisfies (\ref{formula_allrelations}). If $r > 1$, then an identical argument to Subcase 1c finishes the proof of this case and of the theorem.
\end{proof}
\section{The skein of the annulus as a module over the algebra $H$.}\label{sec_solidtorus}
We now describe the action of the algebra $H$ on the skein $\mathcal{C}$ of the annulus.
Draw the torus $T^2$ as the boundary of a standardly embedded solid torus $V\subsetR^3$. Once an orientation and a framing for the core of $V$ have been chosen, by choosing an oriented annular neighbourhood of the core curve, we can regard $V$ as a thickened annulus and get an explicit identification of the skein $H(V)$ of $V$ with the skein $\mathcal{C}$. In the diagram below
\bc
\CylinderAnnulus
\ec
we indicate the relation of the torus and the framed core of $V$.
Parametrize $T^2$ so that the core lies in the direction of the $(0,1)$ curve, and the framing of the core is chosen to agree with the neighbourhood framing of the $(0,1)$ curve in $T^2$. The $(1,0)$ curve in $T^2$ is then a meridian of the solid torus $V$.
As remarked in Section 2, the skein $H(V)\cong \mathcal{C}$ is a module over the algebra $H(\partial V) \cong H$. To describe $h\cdot c\in\mathcal{C}$ explicitly for $h\in H$ and $ c\in\mathcal{C}$, we represent $c$ by a framed diagram in the core annulus, and $h$ by a framed diagram in $\partial V =T^2$. The union of these two framed diagrams in $V$ then represents $h\cdot c$ in $H( V) =\mathcal{C}$. (Properly $h$ and $c$ are represented by some $R$-linear combination of framed diagrams, and we extend the construction bilinearly).
\subsection{The action}
We now give precise statements about the action of $H$ on $\mathcal{C}$. Because the core annulus of $V$ is parallel to the $(0,1)$ curve in $T^2$ the action of $P_{0,n}$ on $\mathcal{C}$ is simply multiplication by $P_n$ in $\mathcal{C}$.
By results of Morton and Hadji \cite{MH02,HM06}, the module $\mathcal{C}$ has an $R$-linear basis given by elements $Q_{\lambda,\mu}$, where $\lambda,\mu$ range over the set of partitions. The basis elements $Q_{\lambda,\mu}$ are shown there to be eigenvectors of the `meridian maps' from $\mathcal{C}$ to itself defined for each $m$ by $c\mapsto P_{m,0}\cdot c$. We will describe the action of $H$ in this basis. Then at the end of the section we will collect from the literature several facts about the $Q_{\lambda,\mu}$ for the reader's convenience.
For partitions $\lambda, \mu$ write
\[
s_{\lambda,\mu} := (s-s^{-1})\left( v^{-1} \sum_{x \in \lambda} s^{2c(x)} - v\sum_{x \in \mu} s^{-2c(x)}\right) + \frac{v^{-1}-v}{s-s^{-1}}
\]
where $c(x) = j - i$ is the \emph{content} of the cell $x$ in row $i$ and column $j$. We will use the continental convention for Young diagrams, so that the unique partition of $1$ corresponds to a cell in the lower-left corner of the diagram, which is in row and column $0$.
For a partition $\lambda$ define its \emph{content polynomial} $C_\lambda(t)\in \mathbb{Z}[t^{\pm1}]$ by
\[C_\lambda(t)=\sum_{x\in\lambda} t^{c(x)}.\]
Then $s_{\lambda,\mu} = (s-s^{-1})(v^{-1} C_\lambda(s^2)- vC_\mu(s^{-2}))+ \frac{v^{-1}-v}{s-s^{-1}}$
\begin{theorem}\label{thm_Caction}
In the basis $Q_{\lambda,\mu}$ of $\mathcal{C}$ described in \cite{MH02, HM06}, the action of $H$ is determined by the equations
\begin{eqnarray*}
P_{1,0}\cdot Q_{\lambda,\mu} &=& s_{\lambda,\mu}Q_{\lambda,\mu}\\
P_{-1,0}\cdot Q_{\lambda,\mu} &=& s_{\mu,\lambda}Q_{\lambda,\mu}\\
P_{0,1}\cdot Q_{\lambda,\mu} &=& \sum_{\alpha \in \lambda +1} Q_{\alpha,\mu} + \sum_{\beta \in \mu -1}Q_{\lambda,\beta}\\
P_{0,-1}\cdot Q_{\lambda,\mu} &=& \sum_{\beta \in \lambda -1} Q_{\beta,\mu} + \sum_{\alpha \in \mu +1}Q_{\lambda,\alpha}
\end{eqnarray*}
where $\lambda +1,\lambda -1$ are the set of partitions where one cell has been added to (subtracted from) $\lambda$, respectively.
\end{theorem}
\begin{proof}
The first two equations are the meridian map formulae in \cite[Thm.\ 3.9]{HM06}, and the last two are the product formulae in \cite{MR10}.
\end{proof}
\begin{lemma}
The statements in Theorem \ref{thm_Caction} completely determine the module structure of $\mathcal{C}$.
\end{lemma}
\begin{proof}
Since the $Q_{\lambda,\mu}$ form a basis for $\mathcal{C}$ the action of $P_{\pm 1,0}$ and $P_{0,\pm 1}$ on $\mathcal{C}$ is completely described by Theorem \ref{thm_Caction}. Over $R$ these elements generate the algebra $H$, which completes the claim.
\end{proof}
\begin{remark}
The skein $\mathcal{C}$ has a multiplicative identity $1$, represented by the empty diagram. Then $h\cdot 1$ is the element $h$ regarded as lying in the solid torus $V$. In particular, when $h$ is a meridian element, meaning that $h$ is $P_{1,0}$ decorated by some element $c\in\mathcal{C}$, the element $h\cdot 1$ is represented by the zero framed unknot in a ball inside $V$ decorated by $c$. Its value in $\mathcal{C}$ is the scalar multiple ${\rm ev}(c) 1$ of the identity, where ${\rm ev}(c)$ is the Homflypt polynomial of the unknot decorated by $c$.
Thus $P_{m,0}\cdot 1 ={\rm ev}(P_m) 1$, and it is known that $${\rm ev}(P_m) =\frac{v^{-m}-v^m}{s^m-s^{-m}}$$
Our commutation relation $[P_{m,0},P_{0,n}] =\{mn\}P_{m,n}$ leads to the calculation in \cite{MM08} of the effect in $\mathcal{C}$ of putting a meridian decorated by $P_m$ around the core decorated by $P_n$.
The core decorated by $P_n$ is $P_{0,n}\cdot 1\in\mathcal{C}$. Hence putting a meridian decorated by $P_m$ around this gives
\begin{eqnarray*} P_{m,0}\cdot(P_{0,n}\cdot 1)=P_{m,0}P_{0,n}\cdot 1
&=&\{mn\}P_{m,n}\cdot 1+P_{0,n}P_{m,0}\cdot 1\\
& =&\{mn\}P_{m,n}\cdot 1 +{\rm ev}(P_m) P_{0,n}\cdot 1 \\
&=&\{mn\}P_{m,n}\cdot 1 +{\rm ev}(P_m) P_n.
\end{eqnarray*}
In fact it was this equation in \cite[Thm.\ 18]{MM08}, illustrated there in figure 14, with $N, M$ in place of our $m,n$ respectively, which encouraged us to conjecture (and then prove) the commutation relations of Theorem \ref{thm_commutationrelations} in their complete generality.
\end{remark}
We now deduce formulae for the action of $P_{m,n}$ on the elements $Q_{\lambda,\mu}$. We first establish some notation.
\begin{definition}
We will use several statistics on pairs of partitions.
\begin{enumerate}
\item For $\lambda \subset \alpha$, write $\alpha - \lambda$ for the skew partition consisting of the cells contained in $\alpha$ but not in $\lambda$.
\item Write $\lambda + n$ for the set of partitions $\alpha\supset\lambda$ where $\alpha - \lambda$ is a \emph{border strip}\footnote{A \emph{border strip} is a (skew) partition that is connected and contains no $2\times 2$ squares.} of length $n$. Similarly $\lambda - n$ is the set of partitions $\beta\subset\lambda$ where $\lambda-\beta$ is a border strip of length $n$.
\item Write $\mathrm{ht}(\gamma)$ for the \emph{height} of a border strip $\gamma$, defined as the number of rows in $\gamma$ plus 1.
\item Extend the definition of content polynomial to cover skew partitions $\alpha-\lambda$ by setting \[C_{\alpha-\lambda}(t):=\sum_{x\in\alpha -\lambda} t^{c(x)}=C_\alpha(t)-C_\lambda(t).\]
Replace $s$ and $v$ by $s^m$ and $v^m$ in $ s_{\lambda,\mu}$ to define \[ s_{\lambda,\mu}(m):=\{m\}(v^{-m}C_\lambda(s^{2m})-v^mC_\mu(s^{-2m})) + \frac{v^{-m}-v^m}{s^m-s^{-m}}.\]
\item For a skew partition $\gamma$ write
\begin{eqnarray*}
b(m,\gamma)&:=&v^{-m}C_\gamma(s^{2m})\\
b^-(m,\gamma)&:=&(-1)^{\mathrm{ht}(\gamma)} v^{-m}C_\gamma(s^{2m})
\end{eqnarray*}
\end{enumerate}
\end{definition}
Using these definitions, we have the immediate relations that \[s_{\lambda,\mu}(m) =\{m\}b(m,\lambda)+\{-m\}b(-m,\mu)+\frac{v^{-m}-v^m}{s^m-s^{-m}}\] and $s_{\lambda,\mu}(-m)=s_{\mu,\lambda}(m)$.
\begin{remark}\label{strip} If $\gamma$ is a border strip of length $n$ it consists of a sequence of $n$ cells, each adjacent to the previous one, starting with a cell of least content $k$, say. The content increases by one for each cell in the sequence, so that $C_\gamma(t)=t^k(1+t+\cdots+t^{n-1})$. Then $b(1,\gamma)=v^{-1}s^{2k}(s^{2n}-1)/(s^2 -1)=v^{-1}s^{2k+n-1}\{n\}/\{1\}$. It follows that \[\frac{\{m\}}{\{mn\}}b(m,\gamma)=(v^{-1}s^{2k+n-1})^m.\]
\end{remark}
The following formulae generalise those of Theorem \ref{thm_Caction}. They apply even in the case of $n<0$, provided that when $\alpha\subset\lambda$ we set $\mathrm{ht}(\alpha-\lambda)=\mathrm{ht}(\lambda-\alpha)$, and still take $C_{\alpha-\lambda}(t)$ to mean $C_\alpha(t)-C_\lambda(t)$, so that $b(m,\alpha-\lambda)=-b(m,\lambda-\alpha)$.
\begin{theorem}\label{thm_Cactionfull} For $m,n \in \mathbb{Z}-\{0\}$ we have the following equalities:
\begin{eqnarray}
P_{m,0}\cdot Q_{\lambda,\mu} &=& s_{\lambda,\mu}(m)Q_{\lambda,\mu}\label{full1}\\
P_{0,n}\cdot Q_{\lambda,\mu} &=& \sum_{\alpha \in \lambda + n} (-1)^{\mathrm{ht}(\alpha - \lambda)}Q_{\alpha,\mu} + \sum_{\beta \in \mu - n}(-1)^{\mathrm{ht}(\mu - \beta)}Q_{\lambda,\beta}\label{full2}\\
P_{m,n}\cdot Q_{\lambda,\mu} &=& \frac{\{m\}}{\{mn\}} \left[
\sum_{\alpha \in \lambda + n} b^-(m,\alpha-\lambda)Q_{\alpha,\mu} +
\sum_{\beta \in \mu - n} b^-(-m,\mu - \beta) Q_{\lambda,\beta}\right] \label{full3}
\end{eqnarray}
\end{theorem}
\begin{remark}
Before proving the theorem, we remark that when $\mu = \emptyset$, these formulas were already known. The first follows from the identification of $P_{0,n}$ with the power sum function in \cite{Mor02, MM08}, and the second appears as \cite[Lemma 17]{MM08}.
\end{remark}
\begin{proof}
We know the first two equations for $m = \pm 1$ and $n = \pm 1$. Equation (\ref{full3}) follows from (\ref{full1}) and (\ref{full2}) by a straightforward calculation using the commutation relation of Theorem \ref{thm_commutationrelations}.
To prove equation (\ref{full1}), we proceed by induction on $m$, where in the inductive step we assume that the first and third equations are true for $0 \leq m \leq M$ and for $n \in \{-1,1\}$, and prove the first equation for $m = M+1$. (The case $m < 0$ follows by symmetry.) By the inductive assumption, we have
\begin{eqnarray*}
P_{m,1}\cdot Q_{\lambda,\mu} &=& \sum_{\alpha \in \lambda + 1} b(m,\alpha - \lambda)Q_{\alpha,\mu} +
\sum_{\beta \in \mu - 1}b(-m,\mu - \beta)Q_{\lambda,\beta}\\
P_{1,-1} \cdot Q_{\lambda,\mu} &=& \sum_{\alpha \in \lambda - 1} b(1,\lambda - \alpha)Q_{\alpha,\mu} +
\sum_{\beta \in \mu + 1}b(-1,\beta - \mu)Q_{\lambda,\beta}
\end{eqnarray*}
From these, we can compute the action of $P_{m+1,0}$ using the commutation relation
\begin{equation}\label{eq_commtimesQ1}
P_{m+1,0}\cdot Q_{\lambda,\mu} = \frac{-1}{\{m+1\}\, } [P_{m,1},P_{1,-1}]\cdot Q_{\lambda,\mu}
\end{equation}
To shorten the following computation, we first note that in the formula for $P_{m,1}P_{1,-1}\cdot Q_{\lambda,\mu}$ there will be 4 types of terms corresponding to whether each operator adds cells to $\lambda$ or remove cells from $\mu$. The ``cross terms" where both
$\lambda$ and $\mu$ change will cancel with the analogous cross terms from $-P_{1,-1}P_{m,1}\cdot Q_{\lambda,\mu}$. Also, equality (\ref{eq_commtimesQ1}) for the terms where just $\lambda$ changes is equivalent to the equality of the terms where just $\mu$ changes, by symmetry. Therefore, in the following computation we will just write the terms of the right hand side of (\ref{eq_commtimesQ1}) where only $\lambda$ changes, and denote the rest of the terms by ``$\cdots$''.
\begin{eqnarray}
P_{m,1}P_{1,-1}\cdot Q_{\lambda,\mu} &=& \sum_{\alpha \in \lambda - 1}b(1,\lambda - \alpha)\sum_{\alpha' \in \alpha + 1}b(m,\alpha' - \alpha)Q_{\alpha',\mu} + \cdots \label{eq_t1} \\
P_{1,-1}P_{m,1}\cdot Q_{\lambda,\mu} &=& \sum_{\beta \in \lambda + 1}b(m,\beta - \lambda)\sum_{\beta' \in \beta - 1}b(1,\beta - \beta')Q_{\beta',\mu} + \cdots\label{eq_t2}
\end{eqnarray}
We now examine the coefficients of $Q_{\gamma,\mu}$ in $-[P_{m,1},P_{1,-1}]\cdot Q_{\lambda,\mu}$, which is the difference of these two expressions. The terms where $\gamma \not= \lambda$ appear exactly once in both (\ref{eq_t1}) and (\ref{eq_t2}) with equal coefficients, so they cancel. The coefficient of $Q_{\lambda,\mu}$ in (\ref{eq_t1}) comes from the cases where a cell $x$ is removed from $\lambda$ to give $\alpha\in\lambda-1$, and then restored to get $\alpha'=\lambda$. This gives \[\sum_{\{ x\,\mid \, \alpha = \lambda - x\} } b(1,x)b(m,x) = v^{-m-1}\sum_{\{ x \,\mid \, \alpha = \lambda - x\} } s^{2(m+1)c(x)} .\] In (\ref{eq_t2}) we need $\beta'=\lambda$, and so $\beta$ arises by adding one cell $y$ to $\lambda$, to give the coefficient \[\sum_{\{ y \,\mid\, \beta = \lambda + y\} } b(1,y)b(m,y)=v^{-m-1}\sum_{\{ y \, \mid \, \beta = \lambda + y\} } s^{2(m+1)c(y)}.\]
The difference is then
\begin{eqnarray}
&{}& v^{-m-1}\sum_{\{ y \, \mid \, \beta = \lambda + y\} } s^{2(m+1)c(y)} -v^{-m-1}\sum_{\{ x \,\mid \, \alpha = \lambda - x\} } s^{2(m+1)c(x)} \label{eq_t3}
\end{eqnarray}
It is now enough to show that the expression in (\ref{eq_t3}) is equal to the terms with coefficient $v^{-m-1}$ in $\{m+1\}s_{\lambda,\mu}(m+1)$. (The terms with coefficient $v^{m+1}$ will come from the terms where a cell is added and subtracted from $\mu$, by symmetry). In other words, we must show that (\ref{eq_t3}) is equal to
\begin{equation}\label{eq_t4}
v^{-m-1} + \{m+1\}^2 b(m+1,\lambda) = v^{-m-1}\left[ 1+(s^{m+1} - s^{-m-1})^2 \sum_{z \in \lambda} s^{2(m+1)c(z)}\right]
\end{equation}
Finally, the equality of the expressions (\ref{eq_t3}) and (\ref{eq_t4}) is a well-known combinatorial identity. It can be proved by elementary means by first expanding the right hand side of (\ref{eq_t4}) along the rows and then along the columns of $\lambda$. The two powers of $\{m+1\}$ turn these two expansions into telescoping sums, and the leftover terms are exactly the ones in (\ref{eq_t3}). This completes the proof of equation \ref{full1} in theorem \ref{thm_Cactionfull}.\\[2mm]
We now proceed to the proof of equation (\ref{full2}), using a similar induction on $n$. In this case we use the commutation relation
\begin{equation}\label{eq_commtimesQ2}
P_{0,n+1}\cdot Q_{\lambda,\mu} = \frac 1 {\{ n+1\}} [P_{1,n},P_{-1,1}]\cdot Q_{\lambda,\mu}
\end{equation}
The induction assumption shows that
\begin{eqnarray*}
\frac{\{n\} }{\{1\} } P_{1,n}\cdot Q_{\lambda,\mu} &=& \sum_{\alpha \in \lambda + n}b^-(1,\alpha - \lambda)Q_{\alpha,\mu}
+ \sum_{\beta \in \mu - n}b^-(-1,\mu - \beta)
Q_{\lambda,\beta} \\
P_{-1,1}\cdot Q_{\lambda,\mu}&=& \sum_{\alpha \in \lambda + 1}b^-(-1,\alpha - \lambda)Q_{\alpha,\mu}
+ \sum_{\beta \in \mu - 1}b^-(1,\mu - \beta)
Q_{\lambda,\beta}
\end{eqnarray*}
When $P_{1,n}P_{-1,1}$ is applied to $Q_{\lambda,\mu}$ there will be four types of terms, depending on how cells are added to $\lambda$ or subtracted from $\mu$.
Thus
\begin{eqnarray*}
\frac{\{n\} }{\{1\} }P_{1,n}P_{-1,1}\cdot Q_{\lambda,\mu}&=&\sum_{ \alpha\in\lambda+1, \gamma\in\alpha+n }a(\gamma,\mu)Q_{\gamma,\mu} + \sum_{ \alpha\in\lambda+1, \beta\in\mu-n }a(\alpha,\beta)Q_{\alpha,\beta}\\
&&+\sum_{ \beta\in\mu-1, \alpha\in\lambda+n}a'(\alpha,\beta)Q_{\alpha,\beta}+\sum_{\beta\in\mu-1, \gamma\in\beta-n }a(\lambda,\gamma)Q_{\lambda,\gamma}
\end{eqnarray*}
There is a similar expansion
\begin{eqnarray*}
\frac{\{n\} }{\{1\} }P_{-1,1}P_{1,n}\cdot Q_{\lambda,\mu} &=&\sum_{ \alpha\in\lambda+n, \gamma\in\alpha+1 } d(\gamma,\mu)Q_{\gamma,\mu} + \sum_{ \beta\in\mu-n,\alpha\in\lambda+1 }d(\alpha,\beta)Q_{\alpha,\beta}\\
&& +\sum_{ \alpha\in\lambda+n,\beta\in\mu-1}d'(\alpha,\beta)Q_{\alpha,\beta}+\sum_{\beta\in\mu-n, \gamma\in\beta-1 } d(\lambda,\gamma)Q_{\lambda,\gamma}\end{eqnarray*} where the operators are applied in the opposite order.
The coefficients of the ``cross terms'' $Q_{\alpha,\beta}$ are
\begin{eqnarray*} a(\alpha,\beta)&=& b(-1,\lambda-\alpha)b^-(1,\beta-\mu)\\
a'(\alpha,\beta)&=& b^-(1, \lambda-\alpha)b(-1,\beta-\mu)\\
d(\alpha,\beta)&=&b^-(1,\beta-\mu) b(-1,\lambda-\alpha)=a(\alpha,\beta)\\
d'(\alpha,\beta)&=& b(-1,\beta-\mu)b^-(1, \lambda-\alpha)=a'(\alpha,\beta)
\end{eqnarray*}
These are unchanged when the operators are applied in the opposite order and thus
they will cancel to leave no cross terms on the right hand side of (\ref{eq_commtimesQ2}).
We now consider the coefficient $a(\gamma,\mu)-d(\gamma,\mu)$ of $Q_{\gamma,\mu}$ in $\frac{\{n\}}{\{1\}}[P_{1,n},P_{-1,1}]\cdot Q_{\lambda,\mu}$. We show that
\begin{eqnarray*}
a(\gamma,\mu)-d(\gamma,\mu)&=&\left\{ \begin{array}{ll} (-1)^{\mathrm{ht}(\gamma - \lambda)}\{n+1\}\frac{\{n\}}{\{1\} } & {\rm if } \ \gamma\in \lambda+(n+1)\\
0 & {\rm otherwise}
\end{array}\right.
\end{eqnarray*}
Together with the similar result for the terms where only cells are removed from $\mu$ , relation (\ref{eq_commtimesQ2}) will establish equation (\ref{full2}) for the coefficient of $P_{0,n+1}\cdot Q_{\lambda,\mu}$, and complete the induction step.
The coefficients of $Q_{\gamma,\mu}$ are \begin{eqnarray*} a(\gamma,\mu)&=&\sum_{(\alpha\in\lambda+1) \ \cap\ (\gamma-n)} b(-1,\alpha-\lambda)b^-(1,\gamma-\alpha)\\
d(\gamma,\mu)&=&\sum_{(\alpha\in\lambda+n)\ \cap\ (\gamma-1)} b^-(1,\alpha-\lambda)b(-1,\gamma-\alpha)
\end{eqnarray*}
We now consider the possible terms in each sum, based on the shape of $\gamma - \lambda$. By construction $\gamma$ arises from $\lambda$, maybe in more than one way, either by adding a cell followed by an $n$-strip or an $n$-strip followed by a cell.
If the cell $x$ and the $n$-strip $Y$ are disjoint then they can be added to $\lambda$ in either order. There will be only one term in each of $a(\gamma,\mu)$ and $d(\gamma,\mu)$, both equal to $b(-1,x)b^-(1,Y)$, unless $n=1$ when $Y$ is also a cell, giving two terms in each sum. In either case $a(\gamma,\mu)-d(\gamma,\mu)=0$.
Otherwise $\gamma-\lambda$ is connected and so it is either an $(n+1)$ border strip for $\lambda$ or it contains a single $2\times 2$ square.
In the latter case the separately added cell $x$ must be one of the two cells on the south-west to north-east diagonal of the square, since they can't both occur in the $n$-strip $Y$. Now take $x$ to be the bottom left cell in the square and $x'$ the top right cell. Their respective complements $Y$ and $Y'$ in $\gamma-\lambda$ are $n$-strips. We can first add $x$ and then $Y$ or first $Y'$ and then $x'$ to get $\gamma$. Then $a(\gamma,\mu)=b(-1,x)b^-(1,Y) =b(-1,x')b^-(1,Y')=d(\gamma, \mu)$, since the content of cells is constant on diagonals and $\mathrm{ht}(Y)=\mathrm{ht}(Y')$. Therefore only terms $Q_{\gamma,\mu}$ where $\gamma - \lambda$ is a border strip can have a non-zero coefficient.\\[2mm]
Consider now the coefficient $a(\gamma,\mu)-d(\gamma,\mu)$ when $\gamma - \lambda$ is a border strip.
There are two extreme cells in the border strip, $x$ at the top left, and $x'$ at the bottom right. Write $Y$ and $Y'$ respectively for their complements in $\gamma-\lambda$. Write $h:=\mathrm{ht}(\gamma-\lambda)$ and $k:=c(x)$. Then $c(x')=k+n$, while the least content of a cell in $Y$ is $k+1$ and in $Y'$ is $k$. By remark \ref{strip} we have $b^-(1,Y)=(-1)^{\mathrm{ht}(Y)}v^{-1}s^{2(k+1)+n-1}\{n\}/\{1\}$ and $b(-1,x)=vs^{-2k}$. In the same way $b^-(1,Y')=(-1)^{\mathrm{ht}(Y')}v^{-1}s^{2k+n-1}\{n\}/\{1\}$ and $b(-1,x')=vs^{-2k-2n}$.
We can add $x$ and $Y$ to $\lambda$ in \emph{exactly} one order to get $\gamma$. If $x$ lies above $Y$ then we can add $x$ last, but not first. We have $\mathrm{ht}(Y) =h-1$, giving a contribution of $b^-(1,Y)b(-1,x)=(-1)^{h-1}s^{n+1}\{n\}/\{1\}$ to $d(\gamma,\mu)$. If $x$ lies to the left of $Y$ then we can add $x$ first but not last. In this case $\mathrm{ht}(Y) =h$ and we get a contribution of $b^-(1,Y)b(-1,x)=(-1)^{h}s^{n+1}\{n\}/\{1\}$ to $a(\gamma,\mu)$. In either case we get a contribution of $(-1)^{h}s^{n+1}\{n\}/\{1\}$ to $a(\gamma,\mu)-d(\gamma,\mu)$.
In the same way we can add $x'$ and $Y'$ in exactly one order to get $\gamma$. When $x'$ lies below $Y'$ it can be added first. Then $\mathrm{ht}(Y')=h-1$ and we get the contribution $b^-(1,Y')b(-1,x')=(-1)^{h-1}s^{-n-1}\{n\}/\{1\}$ to $a(\gamma,\mu)$. If $x'$ lies to the right of $Y'$ then it can be added last, but not first, and we get the same contribution to $d(\gamma,\mu)$ with a changed sign, since $\mathrm{ht}(Y')=h$ in this case. Hence we have in either case a total of
\begin{eqnarray*}
a(\gamma,\mu)-d(\gamma,\mu)&=&(-1)^h (s^{n+1}-s^{-n-1}) \{n\}/\{1\}\\
&=& (-1)^{\mathrm{ht}(\gamma-\lambda)}\{n+1\}\frac{\{n\}}{\{1\}}
\end{eqnarray*}
as claimed. This completes the proof of equation (\ref{full2}) and the proof of the theorem.
\end{proof}
\subsection{Further properties of $\mathcal{C}$}
We collect here some results about $\mathcal{C}$ as an algebra over $R$, and about its basis $Q_{\lambda,\mu}$. The subalgebra $\mathcal{C}^+$ spanned by $Q_{\lambda,\emptyset}$ is isomorphic to the ring $\Lambda$ of symmetric functions. The Schur function $s_\lambda$ corresponds to $Q_{\lambda,\emptyset}$. The identity element $1$ of $\mathcal{C}$, represented by the empty diagram, is given by $Q_{\emptyset,\emptyset}$. The element $P_n\in\mathcal{C}$ lies in $\mathcal{C}^+$ for $n>0$, and corresponds to the power sum $p_n\in\Lambda$. This interpretation of $\mathcal{C}$ as symmetric functions was suggested in \cite{AM98}, with details established in \cite{Luk01,Luk05} and \cite{Mor02}.
\begin{remark}
In the case $\mu = \emptyset$, the formulae in theorem \ref{thm_Cactionfull} were already known. Equation (\ref{full1}) appears in the context of meridian maps as \cite[Lemma 17]{MM08}.
A known result in the theory of symmetric polynomials is the expansion of the product of the $n$th power sum $p_n$ and the Schur function $ s_\lambda$ as a signed sum of Schur functions $s_\alpha$ where $\alpha-\lambda$ is an $n$-strip. Equation (\ref{full2}) then follows from the interpretation in \cite{MM08} of $\mathcal{C}^+$ as symmetric polynomials in which $P_n$ corresponds to $p_n$, and $Q_{\lambda,\emptyset}$ to $s_\lambda$.
\end{remark}
The subalgebra $\mathcal{C}^+$ is spanned by closed braid diagrams in the annulus where all the strings go in the same direction. Closed braids with strings in the reverse direction span an isomorphic subalgebra $\mathcal{C}^-$, which is also spanned by $Q_{\emptyset,\mu}$. Reversing string direction carries $Q_{\lambda,\mu}$ to $Q_{\mu,\lambda}$.
Now $\mathcal{C}$ can be written as $\mathcal{C}\cong \mathcal{C}^+\otimes\mathcal{C}^-$, using Turaev's early description \cite{Tur88} of $\mathcal{C}$ as a polynomial algebra. His generators $A_n$ are $P_{1,n}\cdot 1$ in the notation above, giving $\mathcal{C}^+$ for $n>0$ and $\mathcal{C}^-$ for $n<0$.
We can then present the whole algebra $\mathcal{C}$ as $\Lambda\otimes_R\Lambda$. We have already noted that $P_k\in\mathcal{C}^+$ represents $p_k$ when $k>0$ and hence $p_k\otimes 1$ in $\Lambda\otimes\Lambda$ while $P_{-k}$ for $k>0$ becomes $1\otimes p_{k}$.
The construction of the elements $Q_{\lambda,\mu}$ makes use of elements $h_n\in\mathcal{C}^+$ corresponding to the complete symmetric functions in $\Lambda\otimes 1$ and $h_n^*\in\mathcal{C}^-$, given by reversing the string direction, which become the complete symmetric functions in $1\otimes\Lambda$.
The formula in \cite{HM06} for $Q_{\lambda,\mu} $ is an extension of the classical Jacobi-Trudy formula for $s_\lambda$ as a polynomial in the complete symmetric functions. The general construction can be illustrated by the case when $\lambda$ has parts ${2,2,1}$ and $\mu$ has parts ${3,2}$. Take a matrix with diagonal entries as shown, corresponding to the parts of $\lambda$ and $\mu$.
\[\begin{pmatrix} { h^*_2}&&&&\\
&{ h^*_3}&&&\\
&&{ h_2}&&\\
&&&{ h_2}&\\
&&&&{ h_1}
\end{pmatrix}\]
Complete the rows by shifting indices upwards for the parts of $\lambda$,
and downwards for the parts of $\mu$, to get
\[M=\begin{pmatrix} { h^*_2}&h^*_1&1&0&0\\
h^*_4&{ h^*_3}&h^*_2&h^*_1&1\\
1&h_1&{ h_2}&h_3&h_4\\
0&1&h_1&{ h_2}&h_3\\
0&0&0&1&{ h_1}
\end{pmatrix}\]
Then $Q_{\lambda,\mu}=\det M$.\\[2mm]
There is a further interesting interpretation for the whole of $\mathcal{C}$, where we can consider $\Lambda\otimes\Lambda$ as symmetric functions in two sets of commuting variables ${\mathbf{x}}$ and ${\mathbf{y}}$, with the symmetric functions of ${\mathbf{x}}$ representing the first copy of $\Lambda$ and the symmetric functions of ${\mathbf{y}}$ representing the second copy.
In this context there is a body of results stemming from work of King \cite{Kin70}, Koike \cite{Koi89} and subsequent authors in which such functions are studied, both as functions of two sets of variables, and in the special setting with $y_i=x_i^{-1}$ that deals with characters of $gl(N)$ for large $N$.
Besides the Schur functions $s_\lambda({\mathbf{x}})$ and $s_\lambda({\mathbf{y}})$ King \cite{Kin70} defines `compound' Schur functions $s_{\lambda;\overline\mu}({\mathbf{x}};{\mathbf{y}})$ by determinants that closely resemble those for $Q_{\lambda,\mu}$ in \cite{HM06}, or their counterpart in terms of elementary symmetric functions. As a result we can identify $Q_{\lambda,\mu}$ with the compound Schur function $s_ {\lambda:\overline\mu}$ of \cite{Kin70}.
Here are a few further facts about $\mathcal{C}$ and its isomorphism with $\Lambda\otimes\Lambda$:
\begin{enumerate}
\item The symmetry $a\otimes b \mapsto b\otimes a$ of $\mathcal{C}$ sends $Q_{\lambda,\mu} \mapsto Q_{\mu,\lambda}$.
\item The products $Q_{\alpha,\beta}Q_{\alpha',\beta'}$ expand as positive integer combinations in the basis $Q_{\lambda,\mu}$.
\item The set $\{Q_{\lambda,\mu} \mid \lvert \lambda \rvert \leq n,\, \lvert \nu \rvert \leq p,\,\, \lvert \lambda \rvert - \lvert \mu\rvert = n - p\}$ spans the subspace $\mathcal{C}^{n,p}$ defined as the closure of $(n,p)$ diagrams in the square.
\item $\mathcal{C} = \oplus \mathcal{C}^{n,p}$, where $\mathcal{C}^{n,p} \subset \mathcal{C}^{n+1,p+1}$ and $\mathcal{C}^{n,p} \cap \mathcal{C}^{n',p'} = 0$ if $n-p \not= n'-p'$.
\item $Q_{\lambda,\emptyset} = s_\lambda \otimes 1$, and $Q_{\emptyset,\mu} = 1\otimes s_\mu$.
\item We have $Q_{\lambda,\mu} = Q_{\lambda,\emptyset}Q_{\emptyset,\mu} + v = (s_\lambda \otimes 1)(1\otimes s_\mu) + x$, for some $x \in \mathcal{C}^{\lvert \lambda \rvert - 1,\lvert \mu \rvert - 1}$.
\end{enumerate}
\begin{remark}
Fact 5 is established by Lukac, \cite[Ch.\ 3]{Luk01} or \cite{Luk05}. The other facts appear in \cite{HM06} and \cite{MH02}. Fact 1 is immediate from the determinantal formula on reversing the orientations of all curves and rotating the matrix. Fact 2 is theorem 3.5 in \cite{HM06}, while facts 3,4 and 6 are in \cite{MH02}. Fact 6, along with a more detailed expression for $x$, can also be deduced from \cite{Kin70} (see also \cite{Koi89}), when $Q_{\lambda,\mu}$ is interpreted in terms of compound Schur functions. \end{remark}
\section{The elliptic Hall algebra}\label{sec_ehall}
In this section we recall from \cite{BS12} a presentation of the elliptic Hall algebra $\mathcal{E}_{q,t}$, which depends on two parameters $q,t \in \mathbb{C}^*$. For the convenience of the next section, we will switch $t \mapsto t^{-1}$ from the notation of \cite{BS12}. We then prove that the $t=q$ specialization $\mathcal{E}_{q,q}$ is isomorphic to the Homflypt algebra $H_{s=q^{-1/2},v}$.
\begin{remark}
Before giving a presentation, we recall a short description of the construction of the algebra $\mathcal{E}_{q,t}$ from the introduction of \cite{BS12}. First, we consider a smooth elliptic curve $X$ over $\mathbb F_p$, and the category $\mathrm{Coh}(X)$ of coherent sheaves over $X$. The \emph{Hall algebra} of this category is a (topological) bialgebra $\mathcal{E}^+_{\sigma,\bar \sigma}$, where $\sigma, \bar \sigma$ are the Frobenius eigenvalues on the $l$-adic cohomology group $H^1(X_{\bar{\mathbb{F}}_p},\overline{\mathbb{Q}_l})$. It is proved in \cite{BS12} that the relations can be written entirely in terms of Laurent polynomials in these parameters, so we rename the parameters $q,t$ and allow them to be formal (i.e. $\mathcal{E}^+_{q,t}$ is an algebra over $\mathbb{C}[q^{\pm 1},t^{\pm 1}]$). Then $\mathcal{E}_{q,t}$ is the Drinfeld double of the algebra $\mathcal{E}^+_{q,t}$.
\end{remark}
As before, we will write $d({\mathbf{x}}) = gcd(a,b)$ if ${\mathbf{x}} = (a,b)$, and $d({\mathbf{x}},{\mathbf{y}}) = \det[{\mathbf{x}}\,{\mathbf{y}}]$ for ${\mathbf{x}},{\mathbf{y}} \in \mathbb{Z}^2$. Define the constant
\[
\alpha_i := (1-q^{i})(1-t^{-i})(1-q^{-i}t^{i}) / i
\]
\begin{definition}
By \cite[Thm.\ 5.4]{BS12}, the \emph{elliptic Hall algebra} $\mathcal{E}_{q,t}$ is generated by elements $u_{\mathbf{x}}$ for ${\mathbf{x}} \in \mathbb{Z}^2$, with relations
\begin{enumerate}
\item If ${\mathbf{x}},{\mathbf{y}}$ belong to the same line in $\mathbb{Z}^2$, then
\[
[u_{\mathbf{x}},u_{\mathbf{y}}] = 0
\]
\item If ${\mathbf{x}},{\mathbf{y}} \in \mathbb{Z}^2$ are such that $d({\mathbf{x}}) = 1$ and $\Delta_{{\mathbf{x}},{\mathbf{y}}}$ has no interior lattice points, then
\begin{equation}\label{eq_hallrelation1}
[u_{\mathbf{y}},u_{\mathbf{x}}] = \epsilon({\mathbf{x}},{\mathbf{y}}) \frac{\theta_{{\mathbf{x}}+{\mathbf{y}}}}{\alpha_1}
\end{equation}
\end{enumerate}
where $\epsilon({\mathbf{x}},{\mathbf{y}}) := \mathrm{sign}(d({\mathbf{x}},{\mathbf{y}}))$ and the elements $\theta_{\mathbf{x}}$ are polynomials in the $u_{k{\mathbf{x}}}$ defined for $d({\mathbf{x}}_0)= 1$ by equating the following series:
\begin{equation}\label{eq_hallrelation2}
1 + \sum_{i > 0}\theta_{i {\mathbf{x}}_0}z^i = \mathrm{exp}\Big(\sum_{r \geq 1}\alpha_r u_{r {\mathbf{x}}_0}z^r\Big)
\end{equation}
where $z$ is a formal variable.
\end{definition}
\begin{remark}\label{rmk_sl2actiononEhall}
By \cite[Lemma 5.3]{BS12}, the group $\SL_2(\mathbb{Z})$ acts on $\mathcal{E}_{q,t}$ via $\gamma(u_{\mathbf{x}}) = u_{\gamma({\mathbf{x}})}$.
\end{remark}
These relations look somewhat similar to the commutation relations for $H$, but they are complicated by the definition of $\theta_{\mathbf{x}}$. The key observation is that if $t=q$, then the element $\theta_{\mathbf{x}}$ defined in (\ref{eq_hallrelation2}) simplifies substantially.
\begin{lemma}
If $t=q$, then
\[
\frac{\theta_{\mathbf{x}}}{\alpha_1} = \left([d({\mathbf{x}})]_{q^{1/2}}\right)^2 u_{\mathbf{x}}
\]
\end{lemma}
\begin{proof}
Each constant $\alpha_i$ has a zero of order 1 at $t=q$. If we write the RHS of (\ref{eq_hallrelation2}) as $\mathrm{exp}(a)$, then only the terms of degree 1 in $a$ have a simple zero at $t=q$. Therefore, if we specialize $t=q$, the only surviving term in $\mathrm{exp}(a) / \alpha_1$ is $a/\alpha_1$, so the identity $\left(\alpha_i / \alpha_1\right)|_{t=q} = \left( [i]_{q^{1/2}}\right)^2$ shows the claim.
\end{proof}
\begin{corollary}\label{lemma_relatteq1}
If $t=q$, then the following relations are satisfied:
\begin{align}
\,[u_{1,0}, u_{-1,k}] &= - \mathrm{sign}(k) \left([k]_{q^{1/2}}\right)^2 u_{0,k}\notag\\
\,[u_{1,0}, u_{0,k}] &= - \mathrm{sign}(k) u_{1,k}\notag
\end{align}
\end{corollary}
We now define renormalized generators
$
w_{\mathbf{x}} :=\left(q^{d({\mathbf{x}})/2} - q^{-d({\mathbf{x}})/2}\right) u_{\mathbf{x}}
$.
\begin{theorem}\label{thm_HisotoE}
If we specialize $q=t$ and identify $t = q = s^{-2}$, then the map $P_{\mathbf{x}} \mapsto w_{\mathbf{x}}$ extends to an $\SL_2(\mathbb{Z})$-equivariant $\mathbb{Z}^2$-graded isomorphism of algebras $H_{s,v} \to \mathcal{E}_{q=s^{-2},t=s^{-2}}$.
\end{theorem}
\begin{proof}
We first remark that $\SL_2(\mathbb{Z})$ acts by permutation on the generators $w_{\mathbf{x}}$ by Remark \ref{rmk_sl2actiononEhall} (since it preserves the $gcd$ of the entries of vectors). If we rewrite the relations of Corollary \ref{lemma_relatteq1} in terms of $w_{\mathbf{x}}$ and the parameter $s$, we obtain
\begin{align}
\,[w_{1,0}, w_{-1,k}] &= \{k\}_s w_{0,k}\notag\\
\,[w_{1,0}, w_{0,k}] &= \{k\}_s w_{1,k}\notag
\end{align}
These are the same as the relations (\ref{formula_somerelations}), so Remark \ref{rmk_sl2actiononEhall} combined with Proposition \ref{prop_allfromsome} shows that the $w_{\mathbf{x}}$ also satisfy the relations (\ref{formula_allrelations}). The map is clearly surjective, and it is injective because the description of the basis of $H$ in Corollary \ref{cor_basis} agrees with the PBW basis of $\mathcal{E}_{q,t}$ described in \cite[Thm.\ 4.8]{BS12}.
\end{proof}
\begin{remark}
There is $S_3$ symmetry in the parameters $\{q,t,qt^{-1}\} \in (\mathbb{C}^*)^3$, so if we specialize $s^2=q^{-1}$ and $t=1$ the previous theorem remains true.
\end{remark}
\section{Adaptations of the Homflypt skein relations}\label{sec_adaptations}
There are a number of instances, for example in the context of families of Hecke algebras of type $A$, or in relation to quantum $\SL_N$ modules and associated invariants, where Homflypt skeins can be used as models after a simple adaptation.
The simplest model of the Hecke algebra $H_n$ of type $A_{n-1}$ is the Homflypt skein of oriented framed $n$-tangles, using diagrams in a rectangle with $n$ inputs at the bottom and $n$ outputs at the top \cite{MT90}. Composition is induced by stacking diagrams and the algebra is generated by the elementary $n$-braids \bc $\sigma_i$\quad=\quad \labellist\small
\pinlabel {$i$} at 200 405
\pinlabel {$i+1$} at 250 405
\endlabellist \sigmaior.\\[6mm]\ec
Write $T_i$ for the element of the skein represented by $\sigma_i$. The basic skein relation gives the equation $T_i -T_i^{-1}=(s-s^{-1}) {\rm Id}$, and hence the quadratic relation \[(T_i -s)(T_i +s^{-1})=0,\] with roots $s, -s^{-1}$.
Many algebraic accounts use a version of the Hecke algebra where the quadratic has roots $q, -1$, so it is useful to adapt the skein theory to allow for roots $xs, -xs^{-1}$ with an extra parameter $x$. This is done by Aiston and Morton in \cite{AM98} for the Hecke algebra, and subsequently used in the form below for other skeins.
\subsection{The adaptable Homflypt skein}
Use $R[x^{\pm1}]$-linear combinations of framed oriented curves in a $3$-manifold $M$, possibly including arcs with fixed input and output points on $\partial M$, subject to the relations
\bc
$x^{-1}\Xor - x\Yor=(s-s^{-1})\ \Ior$ \qquad (Switch and smooth)\\[2mm]
$\Rcurlor=xv^{-1}\ \Idor\ ,\qquad \Lcurlor =x^{-1}v\ \Idor$ \qquad (Framing change)\\[2mm]
\ec
with the local blackboard framing convention. The resulting skein $H_x(M)$ provides the relations $x^{-1}T_i -xT_i^{-1}=(s-s^{-1}){\rm Id}$ and hence the quadratic relation with roots $xs, -xs^{-1}$.
This is useful in several instances.
\begin{itemize}
\item
Take $x=s$ and set $q=s^2$ to recover the algebraic version of the Hecke algebra with roots $q, -1$.
\item
Take $x=v$ to eliminate the framing dependence.
\item
Take $s=e^{h/2}, v=s^{-N}, x=e^{-h/2N}=s^{-1/N}$ to adjust for the quadratic relation satisfied by the fundamental $R$-matrix of the $SL_N$ quantum group, and the effect of framing change when constructing knot invariants. \cite{Ais96, MM08}
\end{itemize}
Much of Aiston's original work uses these adaptable relations, with $x$ as an indeterminate alongside $v$ and $s$ in the coefficient ring.
Clearly, knowing the skein $H_x(M)$ we can find the basic skein $H(M)=H_1(M)$ by setting $x=1$.
Lukac suggested how to reverse the process in many instances and recover $H_x(M)$ from $H(M)$, so that we can work without $x$, while still being in a position to adapt if needed.
\begin{theorem} When $M=F\times I$ is a thickened surface there is a linear isomorphism $f_x:H(M)\to H_x(M)$.
\end{theorem}
\begin{proof} Represent each union of framed curves in $M$ by a diagram $D$ on $F$ with the blackboard framing. The allowed changes in the curves alter $D$ by isotopy in $F$ and Reidemeister moves $R_{II}, R_{III}$. The \emph{writhe} of $D$, $w(D)$, defined as the sum of the signs of the crossings in $D$, then depends only on the curves in $M$ and not on the choice of representing diagram.
Define $f_x$ on diagrams by \[f_x(D)=x^{-w(D)} D.\]
To prove that this induces a well-defined map on $H(M)$ we must show that the skein relations are respected.
For the switch and smooth relation we must show that \[f_x(D_+)- f_x(D_-)=(s-s^{-1})f_x(D_0)\] in $H_x(M)$,
where three diagrams $D_+, D_-, D_0$ differ only by switching or smoothing a crossing.
Now the writhes of $D_{\pm}, D_0$ satisfy $w(D_+)=w+1, w(D_-)=w-1$ where $w=w(D_0)$, so
\[f_x(D_+)- f_x(D_-)=x^{-w-1}D_+-x^{-w+1}D_-=x^{-w}(s-s^{-1})D_0 =(s-s^{-1})f_x(D_0)\] in $H_x(M)$.
Similarly, for the framing change, $w\left(\Rcurlor\right)=w+1$ where $w=w\left(\Idor\right)$, so in $H_x(M)$ we have
\[f_x\left(\Rcurlor\right)=x^{-w-1}\ \Rcurlor=x^{-w}v^{-1}\ \Idor=v^{-1}f_x\left(\Idor\right).\]
\end{proof}
For example, if we need to adapt the element $P_m=(s-s^{-1})/(s^m-s^{-m}) X_m$ from our algebra $H$ above to $H_x$ we replace $X_m=\sum A_{i,j}$ by $\sum x^{j-i}A_{i,j}$ as in Aiston's original version. The product $P_{\mathbf{x}} P_{\mathbf{y}}$ in $H$ is replaced by $x^{-k}P_{\mathbf{x}} P_{\mathbf{y}}$ on passing to $H_x$ where we use the adapted $P_{\mathbf{x}},P_{\mathbf{y}}$ in $H_x$, and set $k=\det[{\mathbf{x}}\ {\mathbf{y}}]$. This implies the following corollary of Theorem \ref{thm_commutationrelations}:
\begin{corollary}
Using the general Homflypt skein relations in this section (with parameters $x,s,v$), the algebra $H$ is generated by elements $P_{\mathbf{x}}$ with relations
\[
x^{-k}P_{\mathbf{x}} P_{\mathbf{y}}- x^kP_{\mathbf{y}} P_{\mathbf{x}} = \{d\}P_{{\mathbf{x}},{\mathbf{y}}}
\]
where $d = \det [{\mathbf{x}}\,{\mathbf{y}}]$.
\end{corollary}
The quadratic relations used by Schiffman, Vasserot, and Cherednik correspond to the basic skein, so there is largely no need for adaptation. However, comparison with the results of Frohman and Gelca \cite{FG00} for the Kauffman bracket skein needs the adaptation, after orienting, of $x=-A^{-1}$, $s=A^{-2}$ and $v=A^{-4}$.
\section{Iterated Cables}\label{sec_iteratedcables}
Let $K$ be an iterated cable of the unknot and $\lambda$ a partition. In this section we use the isomorphism between the elliptic Hall algebra and the Homflypt skein algebra to construct a 3-variable polynomial that specializes to the $\lambda$-colored Homflypt polynomial of $K$ (up to a monomial $s^\bullet v^\bullet$). This can be considered to be an $\mathfrak{sl}_\infty$ version of the construction in \cite{Sam14} for $\mathfrak{g} = \mathfrak{sl}_2$, which was generalized in \cite{CD14} to arbitrary $\mathfrak{g}$. Our construction uses the work of Schiffmann and Vasserot \cite{SV13, SV11} in an essential way, and when restricted to torus knots, it is essentially the same as the construction of Gorsky and Negut \cite{GN13}. We will use our construction to prove a conjecture of Cherednik and Danilenko in \cite{CD14}.
We first establish some notation. Throughout the section, ${\mathbf{m}} = (m_1,\ldots,m_k)$ and ${\mathbf{n}} = (n_1,\ldots,n_k)$ will be sequences of integers with $m_i, n_i$ relatively prime and $m_i>0$. We will write $\Lambda^N := \mathbb{C}[x_1,\cdots,x_N]^{S_N}$ for the graded ring of symmetric polynomials, and $\Lambda$ for the (graded) ring of symmetric functions, which is $N \to \infty$ limit of the $\Lambda^N$. We write $\pi_N: \Lambda \to \Lambda^{N}$ for the natural projection.
There are three algebras that we will use in this section: the Homflypt skein algebra $H$, the elliptic Hall algebra $\mathcal{E}_{q,t}$, and the double affine Hecke algebra $\mathscr{H}^N_{q,t}$ (defined below). In general we will use superscripts $H$, $\mathcal{E}$, and $N$ to distinguish between objects associated to these three algebras. For example, associated to the sequences ${\mathbf{m}},{\mathbf{n}}$ and a partition $\lambda$ we will define two polynomials using the representation theory of $\mathcal{E}_{q,t}$ and $\mathscr{H}^N_{q,t}$, respectively:
\begin{equation*}
J^\mathcal{E}({\mathbf{m}},{\mathbf{n}},\lambda; q,t,u) \in \mathbb{C}[q^{\pm 1}, t^{\pm 1},u^{\pm 1}],\quad J^N({\mathbf{m}},{\mathbf{n}},\lambda; q,t) \in \mathbb{C}[q^{\pm 1}, t^{\pm 1}]
\end{equation*}
(Technically, we actually define rational functions - see Remark \ref{rmk_rational}.) We will relate these polynomials to the colored Homflypt polynomial of the iterated cable $K({\mathbf{m}},{\mathbf{n}})$ of the unknot determined by the sequences ${\mathbf{m}}$ and ${\mathbf{n}}$. We first define the notion of iterated cable that we will use:
\begin{definition}\label{def_ourcable}
Let $K$ be a framed knot, let $T$ be the torus which bounds a neighborhood of $K$, and let $L_{fr}$ be the longitude in $T$ determined by the framing of $K$.
\begin{enumerate}
\item The \emph{algebraic} $(m,n)$ cable of $K$ is the framed knot in $T$ such that
\[
K(m,n) \sim m L_{fr} + n M
\]
(In this notation, the symbol $\sim$ means `is homologous to' and $M$ is the meridian of $K$.) The framing of $K(m,n)$ is defined to be parallel to the torus $T$.
\item We then define a framed knot $K({\mathbf{m}},{\mathbf{n}})$ inductively as follows: $K(m_1,n_1)$ is the algebraic $(m_1,n_1)$ cable of the 0-framed unknot, and $K({\mathbf{m}}_k,{\mathbf{n}}_k)$ is the algebraic $(m_k,n_k)$ cable of $K({\mathbf{m}}_{k-1},{\mathbf{n}}_{k-1})$.
\end{enumerate}
\end{definition}
\begin{remark}
We note that the algebraic cabling procedure is not the standard topological construction of a cable of a knot. However, the resulting cabling formula (Prop. \ref{prop_cabling}) is particularly simple, which makes it convenient for our purposes. This cabling procedure is related to algebraic knots, which are the knots obtained by intersecting a (singular) irreducible algebraic curve in $\mathbb{C}^2$ with a small copy of $S^3$ around the singularity. In particular, if the ${\mathbf{m}}$ and ${\mathbf{n}}$ are the \emph{Newton pairs} of an algebraic plane curve (see \cite[Appendix to Ch.\ 1]{EN85}), then the algebraic knot obtained from this curve is $K({\mathbf{m}},{\mathbf{n}})$. The knots that arise this way are exactly those with $n_i > 0$. (These knots are determined by their Alexander polynomial.)
We will not need to discuss algebraic knots, but it is worth calling attention to \cite[Conj.\ 2.4(iii)]{CD14}, which states that if the ${\mathbf{m}}$, ${\mathbf{n}}$ are the Newton pairs of an algebraic knot, then the specialization $J^\mathcal{E}(\square;q=1,t,u=0)$ is related to the Betti numbers of the Jacobian factor of the curve. It is not clear if the skein-theoretic point of view in this paper can say anything about this conjecture.
\end{remark}
We will use the following as our definition of the Homflypt polynomial of $K({\mathbf{m}},{\mathbf{n}})$. This definition differs from the standard definition of the Homflypt polynomial of $K$ by a monomial $s^\bullet v^\bullet$ depending on ${\mathbf{m}}$, ${\mathbf{n}}$ and $\lvert \lambda \rvert$, but we will ignore this difference since the conjecture in \cite{CD14} is stated up to an overall constant.
\begin{definition}\label{def_ourhomfly}
The evaluation in the skein of $S^3$ of the framed knot $K({\mathbf{m}},{\mathbf{n}})$ colored by the element $Q_\lambda \in \mathcal{C}^+$ will be denoted as follows:
\begin{equation*}
J^H({\mathbf{m}},{\mathbf{n}},\lambda; v,s) \in Homflypt(S^3) = \mathbb{C}[v^{\pm 1}, s^{\pm 1},(s^k - s^{-k})^{-1}]
\end{equation*}
\end{definition}
In this section we will prove the following theorem:
\begin{theorem}\label{thm_iteratedcable}
For ${\mathbf{m}},{\mathbf{n}},\lambda$ as above, we have the following specializations:
\begin{eqnarray}
v^\bullet s^\bullet J^\mathcal{E}({\mathbf{m}},{\mathbf{n}},\lambda; q,t,u)\Big|_{q=s^{-2}, t=s^{-2}, u = v^2} &=& J^H({\mathbf{m}},{\mathbf{n}},\lambda; v,s) \label{eq_equality1}
\\
{ } u^\bullet J^\mathcal{E} ({\mathbf{m}},{\mathbf{n}},\lambda; q,t,u)\Big|_{u = t^N} &=& q^\bullet t^\bullet J^N({\mathbf{m}},{\mathbf{n}},\lambda; q,t)\label{eq_equality2}
\end{eqnarray}
(where the powers denoted by ``$\bullet$'' depend on ${\mathbf{m}}$,${\mathbf{n}}$, and $\lvert \lambda\rvert $, but not on $N$). In particular, the Connection Conjecture \cite[Conj.\ 2.4(i)]{CD14} is true.
\end{theorem}
\begin{remark}
The existence of a polynomial $J^\mathcal{E}$ satisfying the second specialization was announced as a theorem in \cite{CD14}. The proof of this is essentially identical to the proof in \cite{GN13} (which used the results in \cite{SV11} and \cite{SV13}). For the sake of completeness we will include this proof in Section \ref{sec_dahapoly}. We also remark that the stabilization variable $a$ in the Connection Conjecture of \cite{CD14} is $-u$ for us, so their specialization $a = -t^N$ becomes our specialization $u = t^N$.
\end{remark}
\subsection{Homflypt cabling formula}
In this section we give a cabling formula for the Homflypt polynomial $J^H({\mathbf{m}},{\mathbf{n}},\lambda; v,s)$ of the $\lambda$-colored framed knot $K({\mathbf{m}},{\mathbf{n}})$. In particular, this will give a algebraic formula for $J^H$ in terms of the action of $H$ on the skein $\mathcal{C}$ of the annulus. This formula will later be compared to a specialization of the formula given in Section \ref{sec_epoly} which defines $J^\mathcal{E}$, and this will imply the first equality of Theorem \ref{thm_iteratedcable}.
To simplify comparison to the elliptic Hall algebra, we will need to twist the action of $H$ on $\mathcal{C}$ by an automorphism. To try to make this section self-contained, we will recall the necessary facts about $H$ and $\mathcal{C}$ before giving the cabling formula.
\subsubsection{Notation}
We will twist the action of $H$ on $\mathcal{C}$ by the automorphism $P_{m,n} \mapsto P_{-n,m}$. To compare with the constructions in the following sections we will use the following definitions.
\begin{definition}
We define the following subalgebras of $H$:
\[
H^\geq := \langle P_{m,n} \mid m \geq 0\rangle,\quad\quad H^> := \langle P_{m,n} \mid m > 0 \rangle
\]
We also will use the following $R$-submodule of $\mathcal{C}$
\[
\mathcal{C}^+ := R\{Q_{\lambda,\emptyset}\} \stackrel{\sim}\to \Lambda,\quad\quad Q_{\lambda,\emptyset} \mapsto s_\lambda
\]
\end{definition}
The map $\mathcal{C}^+ \stackrel \sim \to \Lambda$ is an \emph{algebra} isomorphism by \cite[Thm.\ 8.2]{Luk05}. The action of $H^\geq$ preserves the subspace $\mathcal{C}^+ \subset \mathcal{C}$, so we can identify $\Lambda$ as an $H^\geq$-module. This module structure is described as follows.
\begin{lemma}\label{lemma_actiononLambda}
The action of $H$ on $\Lambda$ is given by
\begin{eqnarray*}
P_{m,0}\cdot s_\lambda &=& p_m s_\lambda\\
P_{0,n}\cdot s_\lambda &=& \left[\frac{v^{-n} - v^n}{s^n - s^{-n}} + v^n(s^{-n} - s^n)\sum_{x \in \lambda} s^{-2c(x)}\right] s_\lambda \\
&=& \left[\frac{v^{-n} - v^n}{s^n - s^{-n}} + v^n s^{-n} \sum_{i=1}^k (s^{-2n\lambda_i} - 1)s^{2ni}\right]s_\lambda
\end{eqnarray*}
where $\lambda = (\lambda_1,\ldots,\lambda_k)$.
\end{lemma}
\begin{proof}
The first equality follows from the fact that when $\mu = \emptyset$, the second equation of Theorem \ref{thm_Cactionfull} agrees with the Murnaghan-Nakayama rule for the product $p_m s_\lambda$ of a power sum times a Schur function. The second equality is translated from Theorem \ref{thm_Cactionfull}, and the third equality follows from the fact that the sum for each row is a telescoping sum.
\end{proof}
\begin{remark}\label{rmk_previouslyproved}
These equations have been proved previously by Morton and coauthors. The first follows from the identification of $P_{m,0}$ with the power sum function in \cite{Mor02, MM08}, and the second appears as \cite[Lemma 17]{MM08}.
\end{remark}
If $K$ is a framed knot, then the inclusion $N_K \hookrightarrow S^3$ induces an $R$-linear evaluation map ${\rm{ev}}^H_K: \mathcal{C} \to R$. For later use we recall an explicit formula for the evaluation map ${\rm{ev}}_U^H$ restricted to the subspace $\Lambda\subset \mathcal{C}$ when $U$ is the $0$-framed unknot.
\begin{lemma}[{\cite[eq.\ (12)]{ML03}}]
For the $0$-framed unknot $U$, the evaluation map is
\[
{\rm{ev}}^H_U: \Lambda \to R,\quad \quad {\rm{ev}}^H_U(s_\lambda) = \prod_{x \in \lambda} \frac{v^{-1}s^{c(x)} - vs^{-c(x)}}{s^{hl(x)} - s^{-hl(x)}}
\]
\end{lemma}
(If $x \in \lambda$ is a box, then $c(x)$ and $hl(x)$ are its content and hook length, see Section \ref{sec_notation}.)
\begin{remark}
It was shown in \cite[Thm.\ 1]{Mor07} that the evaluation ${\rm{ev}}^H_K(Q_{\lambda,\mu})$ is divisible by the evaluation ${\rm{ev}}_U^H(Q_{\lambda,\mu})$. (The latter evaluation is typically called the quantum dimension.)
\end{remark}
\subsubsection{The cabling formula}
Let $K$ be a framed knot with $N_K$ a tubular neighborhood of $K$ and $T_K$ the boundary of $N_K$. We use the framing of $K$ to identify $T$ with the standard torus, and we use this to identify $H$ with the skein of $T_K$ and $\mathcal{C}$ with the skein of $N_K$. Under these identifications, the framed knot $K$ is isotopic to $P_{1,0}\in H$ and is also isotopic to $Q_{(1),\emptyset} \in \mathcal{C}^+ \subset \mathcal{C}$.
Let $K(m,n)$ be the algebraic $(m,n)$ cable of $K$ and let $N_{m,n}$ be a tubular neighborhood of $K(m,n)$. We identify $\mathcal{C}$ with the skein of $N_{m,n}$ using the framing of $K(m,n)$ - in particular, under this identification, the element $Q_{(1),\emptyset}$ is isotopic to $K(m,n)$.
Finally, let $\Gamma_{m,n}: N_{m,n} \to N_K$ be the inclusion, let $\iota_{m,n}^H: \mathcal{C} \to H$ be the inclusion given by inserting the annulus along the $(m,n)$ curve. We choose
\begin{equation}\label{eq_gammamn}
\gamma_{m,n} \in \SL_2(\mathbb{Z})\textrm{ such that } \gamma_{m,n}\left(\begin{array}{c}1 \\ 0 \end{array}\right) = \left(\begin{array}{c} m \\ n \end{array}\right)
\end{equation}
We will write
\[ \Gamma^H_{m,n}: \mathcal{C} \to \mathcal{C}
\] for the $R$-linear map induced by $\Gamma_{m,n}$ (where we have identified $H(N_{m,n})$ and $H(N_K)$ with $\mathcal{C}$ as described above). Then the following lemma follows immediately from our choices of identification. (See, e.g. \cite[Lemma 2.20]{Sam14}.)
\begin{lemma}\label{lemma_cablingmap}
Under the identifications above, the $R$-linear map $\Gamma^H_{m,n}: \mathcal{C} \to \mathcal{C}$ is
\[
\Gamma^H_{m,n}(x) = \gamma_{m,n}(\iota_{1,0}(x))\cdot 1 = \iota_{m,n}(x)\cdot 1
\]
\end{lemma}
Given sequences ${\mathbf{m}},{\mathbf{n}}$ as before, we will use the composition
\[
\Gamma^H_{{\mathbf{m}},{\mathbf{n}}} := \Gamma_{m_1,n_1}^H \circ \cdots \circ \Gamma_{m_k, n_k}^H
\]
\begin{proposition}\label{prop_cabling}
If $U$ is the $0$-framed unknot we have the following equality:
\[
J^H(K({\mathbf{m}},{\mathbf{n}}),\lambda;s,v) = {\rm{ev}}^H_{K({\mathbf{m}},{\mathbf{n}})}(Q_\lambda) = \left( {\rm{ev}}^H_U \circ \Gamma^H_{{\mathbf{m}},{\mathbf{n}}}\right) (Q_\lambda)
\]
In particular, the $\lambda$-colored Homflypt polynomial $J^H({\mathbf{m}},{\mathbf{n}},\lambda; v,s)$ of the iterated cable $K({\mathbf{m}},{\mathbf{n}})$ (in our normalization, see Definition \ref{def_ourhomfly}) is equal to the right hand side.
\end{proposition}
\begin{proof}
This follows from Lemma \ref{lemma_cablingmap} and our choices of identification. First, the left hand side of the equation is induced from the inclusion of $N_{K({\mathbf{m}},{\mathbf{n}})}\hookrightarrow S^3$, and this inclusion defines the ($\lambda$-colored) Homflypt polynomial. Then the right hand side is induced from the sequence of inclusions
\[
N_{K({\mathbf{m}}_k,{\mathbf{n}}_k)} \hookrightarrow N_{K({\mathbf{m}}_{k-1},{\mathbf{n}}_{k-1})} \hookrightarrow \cdots \hookrightarrow N_{K({\mathbf{m}}_1,{\mathbf{n}}_1)} \hookrightarrow N_U \hookrightarrow S^3
\]
and the composition of these inclusions is equal to the inclusion $N_{K({\mathbf{m}},{\mathbf{n}})} \hookrightarrow S^3$.
\end{proof}
\begin{remark}
This proposition shows that the colored Homflypt polynomials of an iterated cable of the unknot can be evaluated using the skein algebra $H$, the $\SL_2(\mathbb{Z})$ action on $H$, the action of $H$ on $\Lambda$, and the evaluation map on $\Lambda$. The more standard way of writing the cabling formula gives an expression for the polynomials of the cable of $K$ in terms of the polynomials of $K$. Such an expression could be derived from our cabling formula, but we will not need to do this. However, in the simpler case of the Kauffman bracket skein module, both versions of the cabling formula have appeared in multiple places. (Precise statements appear in \cite[Cor.\ 2.15, Cor.\ 2.16]{Sam14}, together with references to other versions.)
In \cite[Prop.\ 4.2]{CD14}, the $\mathfrak{sl}_2$ version of the latter cabling formula was used to prove the analogue of the specialization in equation (\ref{eq_equality1}), and the same specialization was proved in \cite[Thm.\ A.8]{Sam14} using the former cabling formula (again for $\mathfrak{sl}_2$). It therefore should not be surprising that our version of the Homflypt cabling formula implies the specialization in equation (\ref{eq_equality1}), after the Homflypt skein algebra and elliptic Hall algebra have been shown to be isomorphic.
\end{remark}
\subsection{The Hall algebra}\label{sec_epoly}
In this section we define a 3-variable polynomial that specializes to the Homflypt polynomial ${\rm{ev}}^H_{K({\mathbf{m}},{\mathbf{n}})}(s_\lambda)$ of Proposition \ref{prop_cabling}. The key idea is that all the objects on the right hand side of the equality in Proposition \ref{prop_cabling} have $t$-deformations that come from the elliptic Hall algebra $\mathcal{E}_{q,t}$. We will use the work of Schiffmann and Vasserot in \cite{SV11,SV13}, where they constructed an action of a subalgebra of $\mathcal{E}_{q,t}$ on $\Lambda$. However, to simplify comparison to double affine Hecke algebras we will change $t \mapsto t^{-1}$ (as in Section \ref{sec_ehall}). We also use their renormalized generators $v_{\mathbf{x}} := (q^{d(x)} - 1)u_{\mathbf{x}}$.
We first define subalgebras
\[
\mathcal{E}^\geq_{q,t} := \langle v_{m,n} \mid m \geq 0\rangle,\quad \quad \mathcal{E}^>_{q,t} := \langle v_{m,n} \mid m > 0\rangle
\]
We now recall from \cite[Prop.\ 1.4]{SV13} an action of $\mathcal{E}_{q,t}^\geq$ on $\Lambda$. The action will be written in terms of \emph{Macdonald polynomials} $P_\lambda \in \Lambda$, which form a basis for $\Lambda$. The element $v_{1,0}$ acts by multiplication by the power sum $p_1$, and for $k \geq 1$ we have
\begin{eqnarray}\label{eq_Eaction}
v_{0,k}\cdot P_\lambda &:=& \left( \sum_i (q^{k \lambda_i} - 1)t^{-k(i-1)}\right) P_\lambda \\
v_{0,-k}\cdot P_\lambda &:=& q^k\left( \sum_i (q^{-k \lambda_i} - 1)t^{k(i-1)}\right) P_\lambda\notag
\end{eqnarray}
\begin{remark}\label{rmk_homogeneous}
The Macdonald polynomials $P_\lambda$ are a homogeneous basis of $\Lambda$. Then equation (\ref{eq_Eaction}) combined with the presentation of $\mathcal{E}_{q,t}$ implies that the operators $v_{m,n}$ are graded operators, in the sense that they take homogeneous elements to homogeneous elements. In particular, the action of $\mathcal{E}^+_{q,t}$ on $\Lambda$ is graded, which generalizes property (4) under Definition \ref{def_homflyskeinmod}.
\end{remark}
In fact, the subalgebra $\mathcal{E}_{q,t}^{hor} := \langle v_{k,0} \mid k \geq 0 \rangle$ is isomorphic to $\Lambda$ as a graded algebra, and it acts by multiplication operators. Given $m,n \in \mathbb{Z}$ relatively prime, we let $\gamma_{m,n} \in \SL_2(\mathbb{Z})$ be as in equation (\ref{eq_gammamn}). By \cite{BS12}, the group $\SL_2(\mathbb{Z})$ acts on $\mathcal{E}_{q,t}$, and we let
$\iota^\mathcal{E}_{m,n}: \Lambda \to \mathcal{E}_{q,t}$
be the composition of the isomorphism $\Lambda \to \mathcal{E}_{q,t}^{hor}$ with the automorphism $\gamma_{m,n}: \mathcal{E}_{q,t} \to \mathcal{E}_{q,t}$. Since $\iota^\mathcal{E}_{m,n}$ is a map of algebras, it is uniquely determined by the following formula:
\begin{equation}
\iota^\mathcal{E}_{m,n}(p_k) = v_{km,kn} \in \mathcal{E}_{q,t}
\end{equation}
We recall Macdonald's evaluation map \cite[eq.\ VI.6.17 and VI.8.8]{Mac95} in terms of $P_\lambda$:
\begin{equation}\label{eq_macev}
{\rm{ev}}^\mathcal{E}: \Lambda \to \mathbb{C}[q^{\pm 1},t^{\pm 1},u^{\pm 1}],\quad\quad {\rm{ev}}^\mathcal{E}(P_\lambda) := \prod_{x \in \lambda} \frac{t^{l'(x)} - u q^{a'(x)}}{1 - q^{a(x)}t^{l(x)+1}}
\end{equation}
\begin{definition}\label{def_JE}
Given $m,n$ relatively prime and ${\mathbf{m}},{\mathbf{n}}$ as above, define maps $\Lambda \to \Lambda$ via
\begin{eqnarray*}
\Gamma^\mathcal{E}_{m,n}(x) &:=& \iota_{m,n}^\mathcal{E}(x)\cdot 1\\
\Gamma^\mathcal{E}_{{\mathbf{m}},{\mathbf{n}}} &:=& \Gamma^\mathcal{E}_{m_1,n_1}\circ \cdots \circ \Gamma^\mathcal{E}_{m_k,n_k}
\end{eqnarray*}
We then define the polynomial
\begin{equation*}
J^\mathcal{E}({\mathbf{m}},{\mathbf{n}},\lambda; q,t,u) := {\rm{ev}}^\mathcal{E}(\Gamma^\mathcal{E}_{{\mathbf{m}},{\mathbf{n}}}(P_\lambda))
\end{equation*}
\end{definition}
Before we prove the first equality of Theorem \ref{thm_iteratedcable} we prove a proposition relating the actions of $\mathcal{E}_{s^{-2},s^{-2}}$ and $H$ on $\Lambda$. We recall that Theorem \ref{thm_HisotoE} states that there is an isomorphism $H \to \mathcal{E}_{s^{-2},s^{-2}}$ uniquely determined by $P_{\mathbf{x}} \mapsto s^{d({\mathbf{x}})}v_{\mathbf{x}}$. We twist this by a graded automorphism of $H$ to obtain the isomorphism
\begin{equation}\label{eq_newiso}
\tilde \varphi: H \stackrel \sim \to \mathcal{E}_{s^{-2},s^{-2}},\quad \quad \tilde \varphi(P_{m,n}) = v^{n}s^{-m+d(m,n)}v_{m,n}
\end{equation}
\begin{proposition}\label{prop_compareactions}
For $x \in \Lambda$ and $P \in H^>$, we have
\[
P\cdot x = \tilde \varphi(P)\cdot x
\]
\end{proposition}
\begin{proof}
We first compare the actions of $P_{m,n}$ and $\tilde \varphi(P_{m,n})$ on $\Lambda$ for certain $m,n$. First,
\begin{eqnarray*}
P_{m,0}\cdot s_\lambda &=& p_m s_\lambda \\
v^{-0}s^{-m+0+m}v_{m,0} &=& v_{m,0} s_\lambda = p_m s_\lambda
\end{eqnarray*}
Second, from Lemma \ref{lemma_actiononLambda}, for all $n$ we have
\[
P_{0,n}\cdot s_\lambda = \left[\frac{v^{-n} - v^n}{s^n - s^{-n}} + v^n s^{-n} \sum_{i=1}^k (s^{-2n\lambda_i} - 1)s^{2ni}\right]s_\lambda
\]
It is well-known that when $t=q$, the Macdonald polynomials specialize to the Schur functions. Then for $n > 0$, the $t=q=s^{-2}$ specialization of equation (\ref{eq_Eaction}) states
\begin{eqnarray*}
v^n s^{0+n}v_{0,n}\cdot s_\lambda
&=& v^n s^{-n} \left[ \sum_i (s^{-2n \lambda_i} - 1)s^{2ni}\right]\\
v^{-n} s^{0+n}v_{0,-n}\cdot s_\lambda
&=& v^{-n} s^n\left[ \sum_i (s^{2n \lambda_i} - 1)s^{-2ni}\right]
\end{eqnarray*}
We have therefore shown that $P_{m,0}\cdot s_\lambda = \tilde \varphi(P_{m,0})\cdot s_\lambda$ and that
\begin{equation}\label{eq_p0n}
P_{0,n}\cdot s_\lambda = \left[ \frac{v^{-n} - v^n}{s^n - s^{-n}} + \tilde \varphi(P_{0,n})\right]\cdot s_\lambda
\end{equation}
Using the commutation relations in $H$, it is clear that any element in $H^>$ can be written as sums of products of commutators of $P_{m,0}$ and $P_{0,n}$. In the commutator the constant term on the right hand side of equation (\ref{eq_p0n}) drops out, which shows that
\begin{equation}
P\cdot s_\lambda = \tilde \varphi(P)\cdot s_\lambda,\quad\quad\textrm{ for all } P \in H^>
\end{equation}
which completes the proof of the proposition.
\end{proof}
\begin{proof}(of equality (\ref{eq_equality1}) of Theorem \ref{thm_iteratedcable})
We first use the cabling formula in Proposition \ref{prop_cabling} to reduce the equality that we are supposed to prove to the following:
\begin{equation}\label{eq_wanttoshow}
{\rm{ev}}^\mathcal{E}(\Gamma^\mathcal{E}_{{\mathbf{m}},{\mathbf{n}}}(P_\lambda))\big|_{q=s^{-2}, t=s^{-2}, u = v^2} = v^\bullet s^\bullet {\rm{ev}}^H(\Gamma^H_{{\mathbf{m}},{\mathbf{n}}}(s_\lambda))
\end{equation}
where the powers depend on ${\mathbf{m}}$, ${\mathbf{n}}$, and $\lvert \lambda\rvert$. We recall that if $p_k \in \Lambda$ is a power sum, then $\iota^H_{m,n} = P_{km,kn} \in H^>$ and $\iota^\mathcal{E}_{m,n}(p_k) = v_{km,kn} \in \mathcal{E}^>_{s^{-2},s^{-2}}$. Then Proposition \ref{prop_compareactions} implies
\[
\iota^H_{m,n}(p_k) = v^{kn}s^{-km}s^kv_{km,kn}(p_k) \in \End_R(\Lambda)
\]
where the equality is equality of operators in $\End_R(\Lambda)$. Now the assignments $p_k \mapsto v^{kn}s^{-km}s^k p_k$ induce a graded algebra isomorphism $\Lambda \to \Lambda$, which shows that for any homogeneous $x \in \Lambda$ (and in particular for $x = s_\lambda)$, we have the equality of operators
\[
\iota^H_{m,n}(x) = v^{\lvert x \rvert n}s^{-\lvert x \rvert m}s^{\lvert x \rvert}v_{\lvert x \rvert m,\lvert x \rvert n}(x) \in \End_R(\Lambda)
\]
Since this equality is as operators on $\Lambda$, this shows that $\Gamma_{m,n}^H(x) = v^\bullet s^\bullet \Gamma^\mathcal{E}_{m,n}(x)$, where the powers depend on $m$, $n$, and $\lvert x \rvert$. Finally, the actions of $H$ and $\mathcal{E}$ on $\Lambda$ are graded: if $a \in H^>$, $b \in \mathcal{E}^>_{q,t}$, and $x \in \Lambda$ are homogeneous, then $a\cdot x$ and $b \cdot x$ are also homogeneous. (See Remark \ref{rmk_homogeneous}.) This implies that $\Gamma_{m,n}^H$ and $\Gamma_{m,n}^\mathcal{E}$ are homogeneous maps, which implies that for homogeneous $x \in \Lambda$,
\begin{equation}
\Gamma_{{\mathbf{m}},{\mathbf{n}}}^H(x) = v^\bullet s^\bullet \Gamma_{{\mathbf{m}},{\mathbf{n}}}^\mathcal{E}(x)
\end{equation}
where the powers depend on ${\mathbf{m}}$, ${\mathbf{n}}$, and $\lvert x \rvert$.
Now to finish the proof of equation (\ref{eq_wanttoshow}), all that remains is to compare the evaluation maps ${\rm{ev}}^\mathcal{E}$ and ${\rm{ev}}^H$. We then equate parameters $t = q = s^{-2}$ and $u = v^2$ in the formula (\ref{eq_macev}) and compute
\begin{eqnarray*}
{\rm{ev}}^\mathcal{E}(s_\lambda) &=& \prod_{x \in \lambda} \frac{s^{-2l'} - v^2 s^{-2a'} } {1 - s^{-2a}s^{-2l-2} }\\
&=& v \prod_{x \in \lambda} \left(s^{1-l'-a'+a+l}\right) \frac{v^{-1}s^{a'-l'} - vs^{l'-a'}}{s^{-a-l-1}-s^{a+l+1} }\\
&=& vs^{-\lvert \lambda\rvert} \prod_{x \in \lambda} \frac{v^{-1}s^{c(x)} - vs^{-c(x)}}{s^{hl(x)} - s^{-hl(x)}}\\
&=& vs^{-\lvert \lambda \rvert} {\rm{ev}}^H(s_\lambda)
\end{eqnarray*}
where we have written $l = l(x)$, etc. (The third equality is a straightforward combinatorial identity.) Since the $s_\lambda$ are a homogeneous linear basis of $\Lambda$ and the maps $\Gamma_{{\mathbf{m}},{\mathbf{n}}}^H$ and $\Gamma_{{\mathbf{m}},{\mathbf{n}}}^\mathcal{E}$ are homogeneous, this shows equality (\ref{eq_wanttoshow}) and completes the proof.
\end{proof}
\subsection{Double affine Hecke algebras}\label{sec_dahapoly}
We now briefly recall the construction of \cite{CD14}. Since the main theorem of this section will follow from comparing to \cite{SV13}, we will only introduce the notation necessary to make this comparison. (As in the previous section, the $t$ of \cite{SV13} is our $t^{-1}$.)
The double affine Hecke algebra $\mathscr{H}_{q,t}^N$, of $\textrm{GL}_N$, abbreviated DAHA, is the algebra generated by elements $T_i^{\pm 1}$ for $1 \leq i \leq N-1$, and $X_j^{\pm 1}$, $Y_j^{\pm 1}$ for $1 \leq j \leq N$, subject to some relations which we will not write down. This algebra is $\mathbb{Z}^2$-graded, with $deg(X_i) = (1,0)$, $deg(Y_i) = (0,1)$, and $deg(T_i) = 0$. There is an $\SL_2(\mathbb{Z})$ action on $\mathscr{H}_{q,t}^N$ which permutes the graded components.
Let $S$ be the symmetrizing idempotent in the finite Hecke algebra (which is generated by the $T_i$'s), which is characterized by $T_jS = S T_j = t^{1/2}\boldsymbol{\mathrm{e}}$ for all $j$. The spherical DAHA is the subalgebra $\mathrm {S} \H^N_{q,t} := S \mathscr{H}^N_{q,t} S$ of $\mathscr{H}^N_{q,t}$, and it is also $\mathbb{Z}^2$-graded. The $\SL_2(\mathbb{Z})$ action on $\mathscr{H}^N_{q,t}$ preserves the subalgebra $\mathrm {S} \H_{q,t}^N$.
Following \cite[Sec.\ 2.2]{SV11}, for $k > 0$ we define elements
\begin{equation*}
P^N_{0,k} = S \sum_i Y_i^k S
\end{equation*}
Elements $P^N_{\mathbf{x}}$ for ${\mathbf{x}} \in \mathbb{Z}^2$ are defined using the $\SL_2(\mathbb{Z})$ action. We define $\mathrm {S} \H_{q,t}^{N,>}$ to be the subalgebra of $\mathrm {S} \H_{q,t}^N$ generated by $P_{m,n}^N$ with $m > 0$.
Cherednik defined an action of $\mathrm {S} \H_{q,t}^N$ on $\Lambda^N$ using Demazure-Lusztig operators (see, e.g. \cite{Che95}). Instead of defining these operators, we recall the following theorem of Schiffmann and Vasserot. (This determines the action of $\mathrm {S} \H_{q,t}^{N,>}$ on $\Lambda^N$ uniquely up to scalars, which is enough for our purposes.)
\begin{theorem}[\cite{SV13, SV11}]\label{thm_sv}
The assignment $v_{\mathbf{x}} \mapsto P_{\mathbf{x}}^N$ extends uniquely to a $\mathbb{Z}^2$-graded $\SL_2(\mathbb{Z})$-equivariant surjective algebra homomorphism
\[
\phi^N: \mathcal{E}_{q,t} \twoheadrightarrow \mathrm {S} \H_{q,t}^{N}
\]
Furthermore, under the projection $\pi_N: \Lambda \to \Lambda^N$, the actions of the subalgebras $\mathcal{E}_{q,t}^>$ and $\mathrm {S} \H_{q,t}^{N,>}$ are related via the formula
\[
\left(q^{\bullet}t^{\bullet}u^{\bullet}\pi_N\circ v_{m,n}\right)\Big|_{u=t^N} = P_{m,n}^N \circ \pi_N
\]
where $v_{m,n}$ and $P^N_{m,n}$ are viewed as endomorphisms of $\Lambda$ and $\Lambda^N$, respectively, and where the powers denoted `$\bullet$' depend on $m$ and $n$ but not on $N$ or on $\lvert \lambda \rvert$.
\end{theorem}
\begin{proof}
The first statement is \cite[Thm.\ 3.1]{SV11}. The second statement follows from \cite[Lemma 1.3]{SV13} and the discussion directly preceding \cite[Prop.\ 1.4]{SV13}. We remark that the definition of the $P_{m,n}^N$ differs in \cite{SV13} and \cite{SV11} - we have chosen the latter because these make the map $\phi^N$ equivariant under the $\SL_2(\mathbb{Z})$ action. The analogous surjection used in \cite{SV13} is a twist of $\phi^N$ by a graded automorphism, which accounts for the factor $q^\bullet$ in our statement. The factors $u^\bullet$ and $t^\bullet$ come from the two formulas below \cite[eq.\ (2.12)]{SV13}.
\end{proof}
There is a natural algebra map $\iota^N: \Lambda^N \to \mathrm {S} \H_{q,t}^{N,>}$ which takes the power sum functions $p_k$ to the element $P^N_{k,0}$. Given $m,n \in \mathbb{Z}$ relatively prime with $m > 0$, we will write
\[ \iota_{m,n}^N := \gamma_{m,n} \circ \iota^N
\]
where $\gamma_{m,n} \in \SL_2(\mathbb{Z})$ is as in equation (\ref{eq_gammamn}). We remark that the automorphism $\gamma_{m,n}$ of $\mathrm {S} \H^N_{q,t}$ does not preserve the subalgebra $\mathrm {S} \H_{q,t}^{N,>}$ - however, since $m > 0$, the image of the elements $P_{k,0}^N$ \emph{is} contained in the subalgebra $\mathrm {S} \H_{q,t}^{N,>}$.
\begin{definition}[{\cite[eq.\ (2.13)]{CD14}}]\label{def_JN}
Given $m,n$ relatively prime with $m > 0$, sequences ${\mathbf{m}},{\mathbf{n}}$ as above, and a partition $\lambda$ with at most $N$ parts, define maps $\Lambda^N \to \Lambda^N$ via
\begin{eqnarray*}
\Gamma^N_{m,n}(x) &:=& \iota^N_{m,n}(x)\cdot 1\\
\Gamma^N_{{\mathbf{m}},{\mathbf{n}}} &:=& \Gamma^N_{m_1,n_1}\circ \cdots \circ \Gamma^N_{m_k,n_k}
\end{eqnarray*}
We then define the polynomial
\begin{equation*}
J^N({\mathbf{m}},{\mathbf{n}},\lambda; q,t) := {\rm{ev}}^N(\Gamma^N_{{\mathbf{m}},{\mathbf{n}}}(P_\lambda^N))
\end{equation*}
where $P^N_\lambda \in \Lambda^N$ is the Macdonald polynomial associated to the partition $\lambda$.
\end{definition}
\begin{remark}\label{rmk_rational}
The Macdonald polynomials are actually rational functions in $q$ and $t$, so the definition above actually produces a rational function. The definition in \cite{CD14} is a renormalization of the definition above (they divide by the evaluation of $P_\lambda^N$), and in their normalization the output of their formula is actually a polynomial, which is important for their purposes. The effect of this normalization is that the polynomials for the unknot are all 1. Since we work with the skein-theoretic normalization of the Homflypt polynomial, our choice of normalization is slightly more convenient for our purposes.
\end{remark}
\begin{proof}(of equation (\ref{eq_equality2}) of Theorem \ref{thm_iteratedcable})
We need to prove the equality
\[
q^\bullet t^\bullet u^\bullet {\rm{ev}}^\mathcal{E}(\Gamma^\mathcal{E}_{{\mathbf{m}},{\mathbf{n}}}(P_\lambda)) \Big|_{u=t^N} = {\rm{ev}}^N(\Gamma^N_{{\mathbf{m}},{\mathbf{n}}}(P_\lambda^N))
\]
where the powers ``$\bullet$'' do not depend on $N$. We will do this by relating the various maps involving $\Lambda$ and $\Lambda^N$ with the projection $\pi_N:\Lambda \to \Lambda^N$ and $\phi^N: \mathcal{E}_{q,t} \to \mathrm {S} \H_{q,t}^N$.
First, it is well known that $\pi_N(P_\lambda) = P^N_{\lambda}$. Since the surjection $\phi^N$ is $\SL_2(\mathbb{Z})$-equivariant, we see that $\phi_N(\iota_{m,n}^\mathcal{E}(x)) = \iota_{m,n}^N(\pi_N(x))$ for $x \in \Lambda$. This combined with the assumption $m > 0$ and the second statement of Theorem \ref{thm_sv} shows the equality
\begin{equation}\label{eq_usemgeq0}
\left( q^\bullet t^{\bullet} u^\bullet \pi_N \circ \Gamma^\mathcal{E}_{m,n}\right)\Big|_{u=t^N} = \Gamma^N_{m,n}\circ \pi_N
\end{equation}
where the powers depend on $m,n$ but not on $N$. Finally, the equality
\[
{\rm{ev}}^\mathcal{E}\Big|_{u=t^N} = {\rm{ev}}^N\circ \pi_N
\]
completes the proof of the theorem.
\end{proof}
|
\section{INTRODUCTION}
\label{interduction}
\begin{figure*}
\begin{center}
\includegraphics[width=150mm]{fig1.ps}
\caption{Schematic pictures of the $r$-process model (left) and the engine model (right).}
\label{figure:model}
\end{center}
\end{figure*}
Gravitational wave (GW) observations are expected to provide a new view of relativistic phenomena in the Universe. One of the most promising candidates for the direct detection of GWs is the merger of compact binaries such as binary neutron stars (NSs). The second generation of ground-based GW detectors, such as Advanced LIGO \citep{Aba+10}, Advanced VIRGO \citep{Ace+14} and KAGRA \citep{Kur+10}, will reach the sensitivity required to detect GWs from the inspiral and coalescence of compact binary systems including binary NSs within a few hundred Mpc. Statistical studies suggest that a few tens of merger events should be observed per year \citep{Aba+10b}.
Electromagnetic counterparts of GW emitters have been recently focused on to maximize a scientific return from the expected detection of GWs \citep[e.g., ][]{MB12}. Follow-up observations of these electromagnetic counterparts are important to confirm a GW detection and to investigate progenitors and environments. The electromagnetic detection also improves the localization of GW sources because the localization accuracy by photons is much better than that by the ground-based GW detectors $\sim10-100$ deg$^2$ \citep[e.g., ][]{Ess+14}.
Sophisticated simulations have revealed mass ejection associated with the mergers of binary NSs by several mechanisms. Significant mass is dynamically ejected by gravitational torques and hydrodynamical interactions during the mergers, called dynamical ejecta \citep[e.g., ][]{Ros+99, RJ01, Hot+13}. General relativistic simulations show that these ejecta distribute nearly isotropic compared to Newtonian simulations in the cases of binary NSs \citep{Hot+13}, while they are anisotropic for NS-black hole (BH) mergers \citep{KIS13}. Mass may be also ejected through winds driven by neutrinos \citep{Des+09}, magnetic fields of and/or amplified by the merged objects \citep{SSKI11, KKS12, KKSSW14}, viscous heating and nuclear recombination \citep{FM13, FKMQ14}.
A traditional electromagnetic counterpart is short-hard gamma-ray bursts \citep[GRBs; ][]{NPP92}. Recent simulations have revealed that a hypermassive NS is formed from the merger of a NS binary \citep[e.g., ][]{Hot+13}, which is believed to collapse into a BH at later time. Non-collapsed matter and some ejecta falling back to the BH form a torus around the BH \citep[e.g., ][]{Ros07}. Then, a relativistic jet may be launched from the BH-torus system, which is believed to be the central engine of short-hard GRBs. Another interesting possibility is a so-called macronova/kilonova, which is thermal emission from ejecta \citep[e.g., ][]{LP98, Kul05, BK13}. The radiative energy of a macronova is estimated between that of a classical nova and supernova. Ejecta can also produce non-thermal emission at later time similarly to supernova remnants \citep{NP11, PNR13, TKI14}. Ejecta may accompany an advanced relativistic part, producing early emission \citep[$\sim$ hours; ][]{KIS14, Met+14}. Emission from macronovae and NS binary merger remnants is almost isotropic and hence different from that of short GRBs which depends on the directions of their relativistic jets. Moreover, macronovae are closer in time to mergers than emission from merger remnants and do not depend on the properties of circumburst environments. Therefore, macronovae are expected to play a crucial role to localize a large sample of GW events \citep{MB12}.
Recently, a macronova candidate following GRB 130603B was discovered \citep{Tan+13, BFC13}. This candidate is widely interpreted as the results of the radioactive decay of $r$-process elements produced in the ejecta of a compact binary merger \citep{Tan+13, BFC13, Hot+13b, PKR14, Gro+14}. We call this scenario {\it an $r$-process model} throughout this paper. The ejecta from a merger of binary NSs is primarily neutron-rich. Then, heavy radioactive elements (mass number $\gtrsim 130$) are expected to form through neutron-capture onto nuclei ($r$-process nucleosynthesis) \citep[e.g., ][]{LS74}. Although the $r$-process nucleosynthesis ends a few hundred millisecond after a merger, synthesized elements release energy due to nuclear fission and beta decays up to $\sim100$ days \citep[e.g., ][]{Wan+14}. A schematic picture for this model is shown in the left panel of figure \ref{figure:model}. If this scenario is correct, the observations also give important insights into the enrichment of $r$-process elements in the galaxy evolution \citep[e.g., ][]{PKR14}. Although the $r$-process model explains the observed light curve of the macronova, it is based on the limited observational data and the nuclear heating rate with large uncertainties. Required mass of dynamical ejecta to explain the observations is relatively large compared with the simulation results \citep{Gro+14}. In addition, the occurrence of $r$-process nucleosynthesis needs the ejecta with low electron fraction ($Y_e\lesssim0.1$). However, relatively high electron fraction ($Y_e\sim0.2-0.5$) can be also realized, which has been discussed for neutrino-driven wind \citep[e.g., ][]{FM13}. It is worth considering other possibilities such as the scenarios of an external shock between ejecta and surrounding medium \citep{Jin+13}, a supramassive magnetar \citep{Fan+13} and dust grains \citep{TNI14}.
In this study, we consider another power source of macronovae, i.e., energy injection from the activity of the central engine, in addition to the radioactive decay of $r$-process elements. This is similar to the early evolution of core-collapse supernovae \citep[e.g., ][]{A80, P93}. We call this model {\it an engine model} throughout this paper. There are several motivations to consider that the activity of the central engine contributes to the heating of ejecta. One observational motivation is the extended emission following the prompt emission of short GRBs. The origin of extended emission is considered to be the activity of the central engine \citep{Bar+05} because the sharp drop of its light curve is difficult to be reproduced by afterglow emission \citep{IKZ05}. After the merger, a stable NS or a BH is formed. In the case that a BH with a torus (or disk) is formed, the energy injection to the ejecta is expected as a form of the jet and/or disk wind \citep[e.g., ][]{Nak+13}. In the case that a NS with strong poloidal magnetic field is formed as a result of a merger, the wind of relativistic particles is ejected \citep{Dai+06, MQT08, YZG13, WD13, MP14}. Then, the wind collides with the ejecta, and about half of the wind energy converts to the internal energy by the shock-heating. A schematic picture is shown in the right-hand side of figure \ref{figure:model}.
The ejecta emission powered by a stable magnetar has already been discussed \citep{YZG13, WD13, MP14}. They suggest that the magnetar-powered ejecta emit the brighter optical and X-ray emissions than that of the $r$-process model. However, they did not show that the magnetar-powered ejecta explain the detected infrared excess in GRB 130603B \citep{Tan+13, BFC13}.
The engine model can provide energy enough to reproduce the detected macronova candidate, GRB 130603B. We do not specify the specific heating sources. Alternatively, to estimate the luminosity and temperature, we assume that the internal energy $E_{\rm int0}\sim10^{51}$ erg is injected to the ejecta at the time $t_{\rm inj}\sim10^2$ s after the merger. These values are consistent with typical isotropic energy $E_{\rm iso}\sim10^{50}-10^{51}$ erg and duration $t_{\rm dur}\sim10-10^2$ s of the extended emission \citep{Sak+11}. Using the velocity of the ejecta $v$, the temperature at $t_{\rm inj}$ is $T_0\sim[E_{\rm int0}/(av^3t_{\rm inj}^3)]^{1/4}$, where $a$ is the radiative constant. If we only consider the adiabatic cooling for the cooling process of the ejecta, the evolution of the internal energy $E_{\rm int}$ and temperature $T$ is scaled as $E_{\rm int}\propto t^{-1}$ and $T\propto t^{-1}$. The luminosity is described as $L\sim E_{\rm int}/t$. Adopting the ejecta velocity $v\sim10^{10}$cm s$^{-1}$ \citep{Hot+13}, the luminosity $L$ and the temperature $T$ at $t\sim10^6$s are
\begin{eqnarray}\label{sec1:L}
L &\sim& \frac{E_{\rm int0}}{t}\left(\frac{t}{t_{\rm inj}}\right)^{-1} \nonumber \\
&\sim&10^{41}\left(\frac{E_{\rm int0}}{10^{51}{\rm erg}}\right)\left(\frac{t_{\rm inj}}{10^2{\rm s}}\right)\left(\frac{t}{10^6{\rm s}}\right)^{-2}~{\rm erg}~{\rm s}^{-1},
\end{eqnarray}
and
\begin{eqnarray}\label{sec1:T}
T&\sim& T_0\left(\frac{t}{t_{\rm inj}}\right)^{-1} \nonumber \\
&\sim&2\times10^3\left(\frac{E_{\rm int0}}{10^{51}{\rm erg}}\right)^{1/4}\left(\frac{t_{\rm inj}}{10^2{\rm s}}\right)^{1/4}\nonumber \\
& &\times\left(\frac{v}{10^{10}{\rm cm}~{\rm s}^{-1}}\right)^{-3/4}\left(\frac{t}{10^6{\rm s}}\right)^{-1}~{\rm K}.
\end{eqnarray}
The observations of macronova of GRB 130603B give J-band luminosity $\sim10^{41}$ erg s$^{-1}$ and the difference between J-band and B-band $\gtrsim2.5$ mag which corresponds to the temperature $\lesssim4\times10^3$ K at $t\sim7$ days after GRB 130603B in the source rest frame \citep{Tan+13, BFC13}. Therefore, in this estimate, the luminosity and temperature for the engine model is consistent with the observation of the macronova following GRB130603B.
We model the evolution of luminosity and temperature of a macronova. Unlike the previous studies, we treat the model in an analytical manner and formulate a light curve including the early phase ($\sim10^3-10^5$ s), which is important for the search of electromagnetic counterparts of GW emitters. We consider shock-heating due to the activity of a central engine as a heating mechanism of the ejecta. For comparison, the $r$-process model, which has been discussed in most papers \citep[e.g., ][]{LP98}, is also formulated. Then, we compare the results of our models with observations to constrain the model parameters such as the ejected mass and the velocity of the ejecta. Although our models are simplified, it is valuable to make comparison between two heating models. In section \ref{model}, we introduce our model assumptions. We describe the analytical models for the evolution of luminosity and temperature in section \ref{evolution}. Then, we compare our results with observations in section \ref{discuss}. Implications for the discrimination between two models are also discussed. We summarize our results in section \ref{summary}. In appendix A, we summarize the formulae for the observed temperature and bolometric luminosity.
\section{MODEL}
\label{model}
Significant mass of material $\sim10^{-3}-10^{-1}M_{\odot}$ is ejected during a binary merger. We model ejecta by following the results of the general relativistic simulations of NS-NS mergers in \citet{Hot+13}. The simulations show that ejecta expand in a nearly homologous manner \citep[see also ][]{Ros+14}. The morphology of the ejecta is quasi-spherical in the case of a merger of binary NSs. According to these results, we assume an isotropic and homologous expansion for the ejecta. Then, the velocity of ejecta $v$ is
\begin{eqnarray}\label{sec2:v}
v\sim r/t
\end{eqnarray}
where the radius $r$ originates the central engine and the time $t$ is measured from the time when a compact binary merges.
Note that in the case of a merger of NS-BH binary, the ejected mass expands with significant anisotropy \citep{KOST11, Fou+13, Fou+14, Lov+13, Dea+13, KIS13}. We do not consider such anisotropic ejecta in this work.
\begin{figure*}
\begin{center}
\includegraphics[width=150mm]{fig2.ps}
\caption{Schematic pictures of the thin- (left) and thick-diffusion phases (right). The horizontal and vertical axes show the radius which originates from the central engine and the mass density of the ejecta, respectively, in logarithmic scales. The radii $r_{\rm out}$ and $r_{\rm in}$ correspond to the outer and inner edges of the ejecta. Ejecta expand in a homologous manner ($v\sim r/t$). Material at the inner region up to the inner edge of the ejecta (dotted line) falls back to the central engine. The thick vertical lines show the diffusion radius $r_{\rm diff}$ at which the diffusion time equals to the dynamical time. Photons emitted from the right side of the thick vertical line (effectively thin region) can diffuse out from the ejecta. The effectively thick region corresponds to the shaded area. The size $\Delta r$ is the propagation distance to evaluate the diffusion time. Since most scatterings occur near the diffusion radius, we divide two phases whether the diffusion radius is larger than $0.5r_{\rm out}$ (thick vertical dashed lines) or not. The time $t_{\times}$ corresponds to the time when the diffusion radius equals to the half of the radius of the outer edge of the ejecta. See text for details.}
\label{figure:diffuse}
\end{center}
\end{figure*}
\subsection{Density Profile}
\label{density}
\citet{Nag+14} found that the profile of ejecta obtained from simulations by \citet{Hot+13} can be well fitted by a power-law function $\rho\propto v^{-\beta}$. The power-law index of snapshot density $\beta$ is more or less independent on the dynamics of mergers, which is in the range of $\beta \sim 3$--4 for $v_{\min}\le v\le v_{\max}$, where $v_{\max}$ and $v_{\rm min}$ are the velocities of the outer and inner edges of the ejecta, respectively. We choose the middle of this range $\beta=3.5$ in this study. We also fix the maximum velocity $v_{\max}=0.4c$ from simulation results \citep{Hot+13}. The maximum velocity $v_{\max}$ is comparable with the escape velocity of the system. The minimum velocity $v_{\min}$ is mainly determined by complicated dynamics at the initial stage of the merger $t\ll10^2{\rm s}$. For the mass density profile at the front of the ejecta, we assume the discrete boundary and the mass density $\rho=0$ at the region $r>v_{\max}t$. Although this profile may be far from the actual one \footnote{The outer ejecta may have a relativistic component \citep{KIS14} and/or an exponential or a power-law profile (see Sec. \ref{outer}). Such a profile is difficult to calculate precisely with current numerical calculations.}, our main aim is to compare two models for energy sources, so that our conclusions are not affected. In section \ref{outer}, we discuss the dependence on the mass density profile at the outer region of the ejecta for the observed light curve. Here, we only consider the evolution after the initial stage of the merger ($t\gg t_{\rm inj}$) and treat $v_{\min}$ as a model parameter. Because of homologous expansion, the density decreases as $\rho \propto t^{-3}$. Then, the density profile is described by
\begin{eqnarray}\label{sec2-1:rho}
\rho (t,v)=\rho_0 \left(\frac{t}{t_0}\right)^{-3}\left(\frac{v}{v_{\min}}\right)^{-\beta}.
\end{eqnarray}
where $\rho_0$ and $t_0$ are normalization factors. The factor $\rho_0t_0^3$ is related to the total mass of the ejecta $M_{\rm ej}$ as following,
\begin{eqnarray}\label{sec2-1:M_ej}
M_{\rm ej}&=&4\pi\int_{v_{\min}t_0}^{v_{\max}t_0}\rho(t_0,v)r^2dr \nonumber \\
&=&\frac{4\pi}{\beta-3}\rho_0(v_{\min}t_0)^3\left[1-\left(\frac{v_{\max}}{v_{\min}}\right)^{3-\beta}\right],
\end{eqnarray}
where we use $dr(t=t_0)=t_0dv$ from equation (\ref{sec2:v}). We also introduce the radius of ejecta outer edge
\begin{eqnarray}\label{sec2-1:r_out}
r_{\rm out}=v_{\max}t,
\end{eqnarray}
and their inner edge
\begin{eqnarray}\label{sec2-1:r_in}
r_{\rm in}=v_{\min}t.
\end{eqnarray}
\subsection{Diffusion Radius}
\label{diffusion}
The inner part of the ejecta is optically thick, and therefore the propagation of radiation in the ejecta can be regarded as a diffusion process. Photons can diffusively escape from the region which satisfies that the diffusion time, $t_{\rm diff}$, is smaller than the dynamical time $t$,
\begin{eqnarray}\label{sec2-1:diffuse}
t_{\rm diff}\le t.
\end{eqnarray}
The medium in this region is called to be effectively thin \citep{RL79}. For convenience, we introduce a diffusion radius $r_{\rm diff}(t)$ which is the radius satisfying the condition $t=t_{\rm diff}$. Furthermore, we divide ejecta into two regions called the effectively thin ($r\ge r_{\rm diff}$) and effectively thick ($r < r_{\rm diff}$) regions. Near the diffusion radius, the optical depth is $\tau\gg1$. We consider random walk for photons so that the mean number of scatterings to propagate for the distance $\Delta r$ is $(\Delta r/l_{\rm mfp})^2$, where $l_{\rm mfp}$ is the mean free path for a photon. Hence, the diffusion time $t_{\rm diff}$ for the propagation distance $\Delta r$ is
\begin{eqnarray}\label{sec2-1:t_diff}
t_{\rm diff}\sim\frac{l_{\rm mfp}}{c}\left(\frac{\Delta r}{l_{\rm mfp}}\right)^2\sim\tau\frac{\Delta r}{c}.
\end{eqnarray}
In the right hand of equation (\ref{sec2-1:t_diff}), we use $\tau\sim\Delta r/l_{\rm mfp}$.
We calculate the diffusion radius $r_{\rm diff}$ from the condition $t_{\rm diff}=t$. Since the mass density profile of the ejecta is described by a decreasing power-law function (equation \ref{sec2-1:rho}), the diffusion time $t_{\rm diff}$ is negligible in an outer part. Thus, in order to calculate the diffusion radius $r_{\rm diff}$, it is a good approximation to only consider scatterings near $r_{\rm diff}$ ($\Delta r\sim r_{\rm diff}$). However, in the early phase, the distance from the outer edge of the ejecta $r_{\rm out}$ to the diffusion radius $r_{\rm diff}$ is smaller than the diffusion radius $r_{\rm out}-r_{\rm diff}<r_{\rm diff}$. Therefore, we should take the propagation distance as
\begin{eqnarray}\label{sec2-1:Deltar}
\Delta r\sim \left\{ \begin{array}{ll}
r_{\rm out}-r_{\rm diff} &~~(r_{\rm diff}>0.5r_{\rm out}) \\
& \\
r_{\rm diff} &~~(r_{\rm diff}\le 0.5r_{\rm out}). \\
\end{array} \right.
\end{eqnarray}
We call the first the thin-diffusion phase and the second the thick-diffusion phase throughout this paper. We schematically show these two phases in figure \ref{figure:diffuse}. Note that in the thin-diffusion phase, since the size of the effectively thin region is much smaller than the size of the ejecta ($r_{\rm out}-r_{\rm in}$), the calculation of the radiative transfer using Monte Carlo technique \citep[e.g., ][]{BK13, TH13} requires a large number of realizations to follow the temporal evolution, which do not seem to have been considered properly so far.
To obtain the diffusion radius, we need to calculate the optical depth $\tau$ of photons which propagate a distance $\Delta r$. Using equations (\ref{sec2-1:rho}) and (\ref{sec2-1:Deltar}), the optical depth $\tau$ is described as,
\begin{eqnarray}\label{sec2-1:tau}
\tau&=&\int_{r_{\rm diff}}^{r_{\rm out}}\kappa\rho dr \nonumber \\
&=&\frac{(\beta-3)\kappa M_{\rm ej}}{4\pi(\beta-1)v_{\min}^2t^2}\left[1-\left(\frac{v_{\max}}{v_{\min}}\right)^{3-\beta}\right]^{-1} \nonumber \\
& &\times
{\displaystyle \left[\left(\frac{r_{\rm diff}}{v_{\min}t}\right)^{1-\beta}-\left(\frac{v_{\max}}{v_{\min}}\right)^{1-\beta}\right]},
\end{eqnarray}
in the thin-diffusion phase, and
\begin{eqnarray}\label{sec2-1:tau2}
\tau&=&\int_{r_{\rm diff}}^{2r_{\rm diff}}\kappa\rho dr \nonumber \\
&=&\frac{(\beta-3)\kappa M_{\rm ej}}{4\pi(\beta-1)v_{\min}^2t^2}\left[1-\left(\frac{v_{\max}}{v_{\min}}\right)^{3-\beta}\right]^{-1} \nonumber \\
& &\times
{\displaystyle \left(\frac{r_{\rm diff}}{v_{\min}t}\right)^{1-\beta}(1-2^{1-\beta})},
\end{eqnarray}
in the thick-diffusion phase, where $\kappa$ is the opacity of the ejecta. For simplicity, we use a grey approximation and a spatially uniform value of the opacity $\kappa$. From the results of \citet{TH13} \citep[see also ][]{KBB13} which consider the contribution from all $r$-process elements to the opacity of merger ejecta, the evolution of the bolometric luminosity can be approximately described by the constant value of the opacity, $\kappa\sim3-30$ cm$^2$g$^{-1}$. Following their results, we use this value for the opacity of the ejecta. Note that the exact value of the opacity of the ejecta has some uncertainties in the production efficiency of $r$-process elements and its spatial distribution. Moreover, if the ejecta temperature is low enough for dust formation ($T\lesssim2000$ K), the opacity significantly increases \citep{TNI14}. From these reasons, we consider the dependence on $\kappa$ in section \ref{evolution}.
Our model is based on the formulation of the light curves of supernovae \citep[e.g., ][]{C92, NS10, RW11}, but there are several differences. In the case of type II supernovae, the opacity is significantly reduced due to hydrogen recombination \citep[e.g., ][]{GNS14}. However, since the ionization potentials of the lanthanides included in the $r$-process elements are generally lower than that of hydrogen and the iron group, the opacity remains high at relatively low temperature \citep{KBB13}. Therefore, we do not consider the recombination effects for the opacity.
As far as we know, the supernova studies \citep[e.g., ][]{C92, NS10, RW11} have not taken into account the thin-diffusion phase, which is necessary for treating the thickness of the diffusion length appropriately and estimating the physical quantities by the values at the outer edge of the ejecta in the analytical formulae. This phase may be also important for the case of supernovae.
Some supernova studies consider the planar phase \citep{PCW10, NS10} in which the evolution of the ejecta is approximately planar as long as its radius do not double. In the case of the NS-NS merger, since the initial length scale of the merger system is small $\sim10^6$ cm and the velocity of the merger ejecta is subrelativistic, the planar phase is irrelevant for the observations.
\subsection{Heating Mechanisms}
\subsubsection{Radioactivity}
\label{radio}
One of the two heating mechanisms we consider is nuclear heating by $r$-process elements. Since the beta decay products of $r$-process elements produced in NS binary mergers naturally heat ejecta, this mechanism is considered to power the emission of a macronova \citep[e.g., ][]{LP98}. The nuclear heating rate is calculated in several works \citep{Met+10, Rob+11, Kor+12, Ros+14, Wan+14}. The derived heating rates per unit mass $\dot{\epsilon}(t)$ are described by the following formula
\begin{eqnarray}\label{sec2-1:dotepsilon}
\dot{\epsilon}=\dot{\epsilon}_0\left(\frac{t}{1\ {\rm day}}\right)^{-\alpha}.
\end{eqnarray}
In this study, we use $\alpha=1.3$ and $\dot{\epsilon}_0=2\times10^{10}~{\rm erg~s}^{-1}{\rm g}^{-1}$ obtained by \citet{Wan+14}. The value of $\dot{\epsilon}$ has been obtained by simulations under some simplified assumptions with only limited parameter regions. Thus, we should note that the value of $\dot{\epsilon}_0$ has uncertainties.
The injected internal energy by the nuclear decay is $\propto t^{1-\alpha}$ in the region $r < r_{\rm diff}$. On the other hand, the injected energy in this region is decreased by adiabatic cooling. The time-evolution of internal energy due to the adiabatic cooling is proportional to $t^{-1}$. Comparing the two temporal evolution, the index of the adiabatic cooling is smaller than that of the increase of internal energy due to the nuclear decay for $\alpha < 2$. Since we use $\alpha=1.3$, we neglect the injected internal energy in the region $r < r_{\rm diff}$.
\subsubsection{Engine-driven shock}
\label{shock}
Unlike the $r$-process model, energy injection occurs only within the time $t_{\rm inj}$ in the engine model. We only consider adiabatic cooling as a cooling process of ejecta after $t_{\rm inj}$, and therefore, the temperature distribution at time $t$ is,
\begin{eqnarray}\label{sec2-1:T}
T(t,v)=T_0\left(\frac{t}{t_{\rm inj}}\right)^{-1}\left(\frac{v}{v_{\rm min}}\right)^{-\xi},
\end{eqnarray}
where the index $\xi$ is a parameter for a snapshot distribution and $T_0$ is a normalization factor described later. The time dependence of $t^{-1}$ is the effect of adiabatic expansion.
The normalized value $T_0$ is determined by using the relation of total injected internal energy $E_{\rm int0}$ as
\begin{eqnarray}\label{sec2-1:E_int}
E_{\rm int0}&=&4\pi\int_{v_{\min}t_{\rm inj}}^{v_{\max}t_{\rm inj}}aT^4(t_{\rm inj},v)r^2dr \nonumber \\
&=&\frac{4\pi}{3-4\xi}aT_0^4(v_{\min}t_{\rm inj})^3\left[\left(\frac{v_{\max}}{v_{\min}}\right)^{3-4\xi}-1\right],
\end{eqnarray}
where we use $dr=t_{\rm inj}dv$. For the temperature index $\xi>0.75$, the innermost region of ejecta has dominant internal energy. As will be shown in Section \ref{evolution}, since the luminosity and temperature always depend on the product of $E_{\rm int0}$ and $t_{\rm inj}$, we treat $E_{\rm int0}t_{\rm inj}$ as a parameter. Thus, the engine model has two parameters, $\xi$ and $E_{\rm int0}t_{\rm inj}$ instead of $\dot{\epsilon}_0$ and $\alpha$ in the $r$-process model.
Energy injection is not always a single event and the shock does not always get through the whole ejecta. It is considered that the activity of the central engine accompanies violent time variability. In this case, multiple shocks propagate into the ejecta. Some of the shock may not catch up with the outer edge of the ejecta. Current general relativistic simulations cannot calculate the evolution of ejecta for such a long time after merger ($t_{\rm inj}\sim 10^2$ s), so that the index $\xi$ of temperature distribution is highly uncertain. Therefore, we treat the temperature index $\xi$ as a parameter.
Unlike the case of core-collapse supernova \citep{NS10}, it is difficult to determine the temperature distribution of heated ejecta by the activity of a central engine. In the case that the activity of a central engine injects the energy into the ejecta, the radiation-dominated shock (where the internal energy behind the shock is dominated by radiation) is formed in the ejecta. The ejecta are heated during the propagation of the shock. This situation is similar to the initial phase of core-collapse supernovae \citep[e.g., ][]{A80, P93}. In the cases of core-collapse supernovae, the kinetic energy of ejecta before the shock heating is much smaller than the injected internal energy. In such ejecta, the relation between velocity and mass density was obtained by \citet{S60} (in the non-relativistic case for the velocity of the ejecta). Using Sakurai's (1960) solution and the equipartition between the kinetic energy after the shock heating and the internal energy \citep{NS10}, the distribution of the temperature distribution is derived. However, in the case of compact binary mergers, the merger ejecta have a large velocity ($\sim0.01-0.1c$) before the shock heating \citep{Hot+13}. Then, injected internal energy is not always larger than the kinetic energy of ejecta so that it is not clear whether we can use the equipartition to estimate the distribution of internal energy or not.
The kinetic energy of the ejecta $E_{\rm kin}$ is described as
\begin{eqnarray}\label{sec2-1:E_kin}
E_{\rm kin}&=&\frac{1}{2}\times4\pi\int_{v_{\min}}^{v_{\max}}\rho(t,v)v^4t^3dv \nonumber \\
&=&{\displaystyle \frac{1}{2}M_{\rm ej}v_{\min}^2\frac{(\beta-3)\left[\left(\frac{v_{\max}}{v_{\min}}\right)^{5-\beta}-1\right]}{(5-\beta)\left[1-\left(\frac{v_{\max}}{v_{\min}}\right)^{3-\beta}\right]}}.
\end{eqnarray}
Note that if the injected internal energy $E_{\rm int0}$ is larger than the kinetic energy of the ejecta, it is expected that some of the internal energy converts to the kinetic energy of the ejecta. As a result, the internal energy and the kinetic energy are equal as in the case of core-collapse supernovae. Then, the mass density distribution and the maximum velocity of the ejecta derived from simulations may be changed because the injection time may be long $\sim10^2$s compared to that calculated by simulations $\lesssim0.1$s \citep{Hot+13}. For simplicity, we only consider the case $E_{\rm int0}\le E_{\rm kin}$.
\section{Evolution of Luminosities and Temperatures}
\label{evolution}
\begin{table*}
\begin{center}
\begin{tabular}{clccc}
\multicolumn{5}{c}{TABLE 1 Model parameters.} \\ \hline
Symbol & & Fiducial model & Minimum mass model & Hot interior model \\ \hline
$M_{\rm ej}$ & Ejecta mass & $0.10M_{\odot}$ & $0.022M_{\odot}$ & $0.08M_{\odot}$ \\
$v_{\min}$ & Minimum velocity & $0.15c$ & $0.13c$ & $0.18c$ \\
$v_{\max}$ & Maximum velocity & $0.40c$ & $0.40c$ & $0.40c$ \\
$\beta$ & Index of the density profile & 3.5 & 3.5 & 3.5 \\
$\kappa$ & Opacity & 10 cm$^2$ g$^{-1}$ & 30 cm$^2$ g$^{-1}$ & 10 cm$^2$ g$^{-1}$ \\
$\dot{\epsilon}_0$ & Nuclear heating rate at 1 day & $2\times10^{10}$ erg s$^{-1}$ g$^{-1}$ & $\cdots$ & $2\times10^{10}$ erg s$^{-1}$ g$^{-1}$ \\
$\alpha$ & Index of nuclear heating rate & 1.3 & $\cdots$ & 1.3\\
$E_{\rm int0}$ & Internal energy at $t_{\rm inj}$ & $1.3\times10^{51}$ erg & $0.9\times10^{51}$ erg & $0.8\times10^{51}$ erg \\
$t_{\rm inj}$ & Injection time & $10^2$ s & $10^2$ s & $10^2$ s \\
$\xi$ & Index of the temperature profile & 1.6 & 1.1 & 2.7 \\ \hline
\multicolumn{5}{l}{}%
\end{tabular}
\label{tab:parameter}
\end{center}
\end{table*}
In this section, we present the evolution of the observed temperature and luminosity of a macronova using our model introduced in the previous section. In sections \ref{thin} -- \ref{transparent}, we focus on the parameter dependence of the evolution using some approximations. In section \ref{fiducial}, we calculate the temperature and luminosity using the fiducial model with parameters summarized in the first column of table 1.
To calculate the luminosity and temperature, we assume that the emission is well described by the blackbody radiation \citep[e.g., ][]{BK13}. For simplicity, we assume that the observed temperature equals to the temperature at the diffusion radius $r_{\rm diff}$. We also assume that the temperature is not so different from the diffusion radius $r_{\rm diff}$ to $2r_{\rm diff}$ so that in the thick-diffusion phase ($r_{\rm out}>2r_{\rm diff}$), we only consider the emission from $r_{\rm diff}$ to $2r_{\rm diff}$ to calculate the observed luminosity for both the $r$-process and engine models. In some studies \citep[e.g., ][]{Met+14}, the observed temperature is approximated by the temperature at the radius of the photosphere $r_{\rm ph}$ where the optical depth is unity. Since the velocity of the ejecta is near the light speed, the optical depths at the diffusion radius $r_{\rm diff}$ and its twice $2r_{\rm diff}$ are $\tau\sim1-10^2$. Therefore, our assumed temperature approximately equals to the temperature at the photosphere.
In section \ref{diffusion}, we introduced two phases, the thin- and thick-diffusion phases (figure \ref{figure:diffuse}), depending on the size of the region where photons make the diffusion in the ejecta $\Delta r$. We also introduce another phase $r_{\rm diff}\le r_{\rm in}$, the transparent phase, in which photons can diffuse out from the entire of the ejecta. Thus, we divide the evolution into these three phases for the values of the diffusion radius $r_{\rm diff}$ as described below.
\subsection{Thin-diffusion phase}
\label{thin}
The size of the effectively thin region gets larger with time. At the early phase of a macronova, the diffusion radius $r_{\rm diff}$, which is the inner radius of the effectively thin region, is near the outer edge of the ejecta $r_{\rm out}$. In this early phase, we take the propagation distance $\Delta r$ of a photon as $\Delta r\sim r_{\rm out}-r_{\rm diff} (<r_{\rm diff})$. Since we assume that the density is a homologous function of the velocity $\rho\propto v^{-\beta}$, the density can be approximated as $\rho\sim\rho(v_{\max})$ in the region $r_{\rm diff}\gg\Delta r$. Using the escaping condition for the diffusing photons $t\sim t_{\rm diff}$, equation (\ref{sec2-1:t_diff}) and approximation on the optical depth $\tau\sim\Delta r\kappa\rho(t,v_{\max})$, the propagation distance $\Delta r$ can be estimated as
\begin{eqnarray}\label{sec2-2-1:Deltar}
\Delta r&\sim&\sqrt{\frac{ct}{\kappa\rho(t,v_{\max})}} \nonumber \\
&\propto&\kappa^{-1/2}M_{\rm ej}^{-1/2}v_{\min}^{\frac{3-\beta}{2}}v_{\max}^{\beta/2}t^2.
\end{eqnarray}
In the discussion of parameter dependence (sections \ref{thin} -- \ref{transparent}), we only consider the dominant term. For example, we neglect the second term in the right-hand side of equation (\ref{sec2-1:M_ej}) to derive the parameter equation (\ref{sec2-2-1:Deltar}) because the index of the mass density is $\beta>3$ in our model. In section \ref{fiducial}, we include the subdominant terms to calculate the light curves numerically.
First we consider the $r$-process model. The evolution of temperature $T_{\rm obs}$ is obtained by the internal energy density $\dot{\epsilon}t\rho$ at the radius $r=r_{\rm out}$. Using equations (\ref{sec2-1:rho}) and (\ref{sec2-1:M_ej}), the parameter dependence of the density is $\rho(t,v_{\max})\propto M_{\rm ej}v_{\min}^{\beta-3}v_{\max}^{-\beta}t^{-3}$. The observed temperature is
\begin{eqnarray}\label{sec2-2-1:T_obs,nuc}
T_{\rm obs}&\sim&\left(\frac{\dot{\epsilon}t\rho(t,v_{\max})}{a}\right)^{1/4} \nonumber \\
&\propto&M_{\rm ej}^{1/4}v_{\min}^{\frac{\beta-3}{4}}v_{\max}^{-\beta/4}t^{-\frac{2+\alpha}{4}}.
\end{eqnarray}
For $\alpha=1.3$, the observed temperature evolves as $T_{\rm obs}\propto t^{-0.875}$. This is because in the thin-diffusion phase the ejecta is effectively a single expanding shell with $\rho\sim\rho(t,v_{\max})$ and the injected energy $\dot{\epsilon}t\propto t^{-0.3}$ is almost constant so that the observed temperature approximately follow adiabatic cooling $T\propto t^{-1}$. Note that in this phase the observed temperature does not depend on the opacity. The bolometric luminosity $L_{\rm bol}$ for the radioactivity is described as the product of the mass within the thickness $\Delta r$ in equation (\ref{sec2-2-1:Deltar}) and the nuclear heating rate $\dot{\epsilon}$ in equation (\ref{sec2-1:dotepsilon}) so that
\begin{eqnarray}\label{sec2-2-1:L_bol,nuc}
L_{\rm bol}&\sim& 4\pi r_{\rm out}^2\Delta r\rho(t,v_{\max})\dot{\epsilon} \nonumber \\
&\propto&\kappa^{-1/2}M_{\rm ej}^{1/2}v_{\min}^{\frac{\beta-3}{2}}v_{\max}^{\frac{4-\beta}{2}}t^{1-\alpha}.
\end{eqnarray}
For $\alpha=1.3$, the evolution of the bolometric luminosity is $L_{\rm bol}\propto t^{-0.3}$.
Next we consider the engine model. We should take into account the freedom of the temperature index $\xi$ in the temperature distribution (equation \ref{sec2-1:T}). Since we only consider a dominant term in the right-hand side of equation (\ref{sec2-1:E_int}) (the first term for $\xi<0.75$ or the second term for $\xi > 0.75$) in this subsection, the parameter dependence of the temperature $T_0$ is described as
\begin{eqnarray}\label{sec2-2-1:T_0}
T_0\propto E_{\rm int0}^{1/4}t_{\rm inj}^{-3/4}\times\left\{ \begin{array}{ll}
v_{\min}^{-\xi}v_{\max}^{\frac{4\xi-3}{4}} & (\xi < 0.75) \\
& \\
v_{\min}^{-3/4} & (\xi > 0.75) \\
\end{array} \right. .
\end{eqnarray}
Substituting $v=v_{\max}$ into equation (\ref{sec2-1:T}), the observed temperature is described as
\begin{eqnarray}\label{sec2-2-1:T_obs,sh}
T_{\rm obs}&\sim&T_0\left(\frac{t}{t_{\rm inj}}\right)^{-1}\left(\frac{v_{\max}}{v_{\min}}\right)^{-\xi} \nonumber \\
&\propto&E_{\rm int0}^{1/4}t_{\rm inj}^{1/4}t^{-1}\times\left\{ \begin{array}{ll}
v_{\max}^{-3/4} & (\xi < 0.75) \\
& \\
v_{\min}^{\frac{4\xi-3}{4}}v_{\max}^{-\xi} & (\xi > 0.75). \\
\end{array} \right.
\end{eqnarray}
Since the observed temperature $T_{\rm obs}$ approximately equals to the temperature at the outer edge of the ejecta, $T_{\rm obs}\sim T(t,v_{\max})$, the evolution of the observed temperature and the luminosity are also determined by the adiabatic cooling. The bolometric luminosity in the effectively thin region is equal to the total radiation created by thermal emission in this region \citep{RL79}. Using equation (\ref{sec2-2-1:T_obs,sh}), the bolometric luminosity is described as
\begin{eqnarray}\label{sec2-2-1:L_bol,sh}
L_{\rm bol}&\sim&4\pi r_{\rm out}^2\Delta r\frac{aT_{\rm obs}^4}{t} \nonumber \\
&\propto&\kappa^{-1/2}M_{\rm ej}^{-1/2}E_{\rm int0}t_{\rm inj}t^{-1} \nonumber \\
& &\times\left\{ \begin{array}{ll}
v_{\min}^{\frac{3-\beta}{2}}v_{\max}^{\frac{\beta-2}{2}} & (\xi < 0.75) \\
& \\
v_{\min}^{\frac{-3-\beta+8\xi}{2}}v_{\max}^{\frac{4+\beta-8\xi}{2}} & (\xi > 0.75). \\
\end{array} \right.
\end{eqnarray}
The time evolution of bolometric luminosity for the engine model does not depend on the temperature index $\xi$ in the thin-diffusion phase.
Comparing the engine model with the $r$-process model in the thin-diffusion phase, the bolometric luminosity and the observed temperature decrease faster in the engine model than those in the $r$-process model. These time-dependence do not depend on the indices of the density and temperature.
Note that the light curve may depend on the detailed profile of the front of the ejecta in this thin-diffusion phase. The profile of the ejecta front is difficult to calculate by the numerical simulation due to its low density, and hence has large uncertain \citep{KIS14}. We discuss its dependence in section \ref{discuss}.
\subsection{Thick-diffusion phase}
\label{thick}
We consider diffusion to evaluate the diffusion radius in the thick-diffusion phase. We take the propagation distance $\Delta r\sim r_{\rm diff}$ after the time when the difference between the radius of the outer edge of the ejecta $r_{\rm out}$ and the diffusion radius $r_{\rm diff}$ is larger than the diffusion radius, $r_{\rm out}-r_{\rm diff}>r_{\rm diff}$, since the optical depth of the outer part is negligible for the density profile in equation (\ref{sec2-1:rho}). In this thick-diffusion phase, mass density significantly deviates from $\rho(v_{\max})$. Substituting equations (\ref{sec2-1:t_diff}) and (\ref{sec2-1:tau2}) into $t=t_{\rm diff}$, the diffusion radius $r_{\rm diff}$ is calculated as
\begin{eqnarray}\label{sec2-2-2:r_diff}
r_{\rm diff}&\sim&\left[\frac{(\beta-3)\kappa M_{\rm ej}v_{\min}^{\beta-3}t^{\beta-4}}{4\pi(\beta-1)c}\right]^{\frac{1}{\beta-2}} \nonumber \\
&\propto&\kappa^{\frac{1}{\beta-2}}M_{\rm ej}^{\frac{1}{\beta-2}}v_{\min}^{\frac{\beta-3}{\beta-2}}t^{\frac{\beta-4}{\beta-2}},
\end{eqnarray}
where we use the relations $\Delta r\sim r_{\rm diff}$ and $v\sim r_{\rm diff}/t$. The latter is obtained from the assumption of the homologous expansion. Regarding the optical depth $\tau$, the second term is neglected in the right-hand side of equation (\ref{sec2-1:tau}) to focus only on the dominant term to study parameter dependence in sections \ref{thin} -- \ref{transparent}. For $\beta=3.5$, the diffusion radius decreases with time ($r_{\rm diff}\propto t^{-1/3}$). Then, emission from the region with relatively high mass density can be observed progressively in this phase ($\rho\propto t^{-3}(r_{\rm diff}/t)^{-\beta}\propto t^{(4\beta-9)/3}=t^{1.667}$).
We introduce the transition time $t_{\times}$ between thin- and thick-diffusion phases, which satisfies the relation $r_{\rm diff}=0.5r_{\rm out}$. Substituting equation (\ref{sec2-2-2:r_diff}) and $r_{\rm out}=v_{\max}t$ into the relation $r_{\rm diff}=0.5r_{\rm out}$, we can obtain the transition time $t_{\times}$ as
\begin{eqnarray}\label{sec2-2-2:t_times}
t_{\times}&\sim&\sqrt{\frac{2^{\beta-4}(\beta-3)\kappa M_{\rm ej}}{\pi(\beta-1)cv_{\max}}\left(\frac{v_{\max}}{v_{\min}}\right)^{3-\beta}} \nonumber \\
&\sim&4.1~\kappa_{10}^{1/2}M_{\rm ej,0.1}^{1/2}v_{\min,0.1}^{\frac{\beta-3}{2}}v_{\max,0.4}^{\frac{2-\beta}{2}}~{\rm day},
\end{eqnarray}
where $\kappa_{10}\equiv \kappa/10$ cm$^2$ g$^{-1}$, $M_{\rm ej,0.1}\equiv M_{\rm ej}/0.1M_{\odot}$, $v_{\min,0.1}\equiv v_{\min}/0.1c$ and $v_{\max,0.4}\equiv v_{\max}/0.4c$. As seen above, the transition time $t_{\times}$ is typically several days. This timescale is expected to allow the follow-up observations \citep{Aasi+14}. Thus, we should consider both phases to predict something useful for follow-up observations. If we fix $\kappa$ and $v_{\max}$ and use $\beta=3.5$, equation (\ref{sec2-2-2:t_times}) gives $t_{\times}\propto M_{\rm ej}^{1/2}v_{\min}^{1/4}$. If we increase the total mass of the ejecta $M_{\rm ej}$ and the velocity at the inner edge of the ejecta $v_{\min}$, the mass density of the ejecta $\rho$ and hence the optical depth are increased. As a result, the transition time $t_{\times}$ becomes large.
First, we consider the $r$-process model. Here, we introduce the velocity $v_{\rm diff}=r_{\rm diff}/t$ based on the homologous relation. Using the velocity $v_{\rm diff}$ and equation (\ref{sec2-2-2:r_diff}) for the mass density (equation \ref{sec2-1:rho}), the evolution of temperature is
\begin{eqnarray}\label{sec2-2-2:T_obs,nuc}
T_{\rm obs}&\sim&\left(\frac{\dot{\epsilon}t\rho(t,v_{\rm diff})}{a}\right)^{1/4} \nonumber \\
&\propto&\kappa^{\frac{\beta}{4(2-\beta)}}M_{\rm ej}^{\frac{1}{2(2-\beta)}}v_{\min}^{\frac{\beta-3}{2(2-\beta)}}t^{\frac{1}{\beta-2}-\frac{\alpha}{4}}.
\end{eqnarray}
For $\alpha=1.3$ and $\beta=3.5$, the evolution is described by $T_{\rm obs}\propto t^{0.341}$ so that the observed temperature increases with time. The bolometric luminosity is described as the product of the mass between $r_{\rm diff}$ and $2r_{\rm diff}$ and the nuclear heating rate $\dot{\epsilon}$. Using equations (\ref{sec2-1:dotepsilon}) and (\ref{sec2-2-2:r_diff}), we obtain the bolometric luminosity as,
\begin{eqnarray}\label{sec2-2-2:L_bol,nuc}
L_{\rm bol}&\sim&4\pi r_{\rm diff}^3\rho(t,v_{\rm diff})\dot{\epsilon} \nonumber \\
&\propto&\kappa^{\frac{3-\beta}{\beta-2}}M_{\rm ej}^{\frac{1}{\beta-2}}v_{\min}^{\frac{\beta-3}{\beta-2}}t^{\frac{2(\beta-3)}{\beta-2}-\alpha}.
\end{eqnarray}
For $\alpha=1.3$ and $\beta=3.5$, the evolution of the bolometric luminosity is $L_{\rm bol}\propto t^{-0.633}$.
Next, we consider the engine model. Using equations (\ref{sec2-1:T}) and (\ref{sec2-2-2:r_diff}), the evolution of the observed temperature is described as
\begin{eqnarray}\label{sec2-2-2:T_obs,sh}
T_{\rm obs}&\sim&T_0\left(\frac{t}{t_{\rm inj}}\right)^{-1}\left(\frac{v_{\rm diff}}{v_{\min}}\right)^{-\xi} \nonumber \\
&\propto&\kappa^{-\frac{\xi}{\beta-2}}M_{\rm ej}^{-\frac{\xi}{\beta-2}}E_{\rm int0}^{1/4}t_{\rm inj}^{1/4}t^{\frac{-\beta+2\xi+2}{\beta-2}} \nonumber \\
& &\times\left\{ \begin{array}{ll}
v_{\min}^{\frac{\xi(\beta-3)}{2-\beta}}v_{\max}^{\frac{4\xi-3}{4}} & (\xi < 0.75) \\
& \\
v_{\min}^{\frac{3\beta-6-4\xi}{4(2-\beta)}} & (\xi > 0.75). \\
\end{array} \right.
\end{eqnarray}
For $\beta=3.5$, the value $\xi=0.75$ is the boundary whether the observed temperature increases with time ($\xi>0.75$) or not ($\xi<0.75$). The evolution of the luminosity equals to the total radiation created by thermal emission in the sphere with radius $r_{\rm diff}$. Using the relation $v(r_{\rm diff})\sim r_{\rm diff}/t$ and equations (\ref{sec2-2-2:r_diff}) and (\ref{sec2-2-2:T_obs,sh}), we obtain
\begin{eqnarray}\label{sec2-2-2:L_bol,sh}
L_{\rm bol}&\sim&4\pi r_{\rm diff}^3\frac{aT_{\rm obs}^4}{t} \nonumber \\
&\propto&\kappa^{\frac{3-4\xi}{\beta-2}}M_{\rm ej}^{\frac{3-4\xi}{\beta-2}}E_{\rm int0}t_{\rm inj}t^{\frac{2(\beta+1-4\xi)}{2-\beta}} \nonumber \\
& &\times\left\{ \begin{array}{ll}
v_{\min}^{\frac{(3-4\xi)(\beta-3)}{\beta-2}}v_{\max}^{4\xi-3} & (\xi < 0.75) \\
& \\
v_{\min}^{\frac{3-4\xi}{2-\beta}} & (\xi > 0.75). \\
\end{array} \right.
\end{eqnarray}
If we take $\beta=3.5$ and $\xi=1.0$, the evolution of the bolometric luminosity is $L_{\rm bol}\propto t^{-0.666}$. This is almost the same dependence as in the $r$-process model. Note that even if the inner part of the ejecta has the larger internal energy ($\xi>0.75$), the bolometric luminosity does not always increase with time. Using the relation $E_{\rm int}(v,t)\propto t^{-1}$, from the adiabatic cooling, the evolution of bolometric luminosity for a given mass shell with $v$ is $L_{\rm bol}\sim E_{\rm int}(v)t^{-1}\propto t^{-2}$, where $E_{\rm int}(v,t)$ is the total internal energy for the mass shell with a given expanding velocity $v$. Since $E_{\rm int}(v_{\rm difff})\propto v^{3-4\xi}$ and $v_{\rm diff}= r_{\rm diff}/t\propto t^{\frac{2}{2-\beta}}$, the bolometric luminosity increases with time for the value of the temperature index $\xi>(\beta+1)/4=1.125$.
\subsection{Transparent phase}
\label{transparent}
Once the diffusion radius reaches the inner edge of the ejecta $(r_{\rm diff}=r_{\rm in})$, all photons emitted from the ejecta can diffuse out within dynamical timescale. If energy is not injected into the ejecta in this transparent phase, the internal energy in the ejecta runs out immediately. The transition time from the thick-diffusion phase to the transparent phase $t_{\rm tr}$ is described as
\begin{eqnarray}\label{sec2-2-3:t_tr}
t_{\rm tr}&\sim&\sqrt{\frac{(\beta-3)\kappa M_{\rm ej}}{4\pi(\beta-1)cv_{\min}}} \nonumber \\
&\sim&6.9~\kappa_{10}^{1/2}M_{\rm ej,0.1}^{1/2}v_{\min,0.1}^{-1/2}~{\rm day},
\end{eqnarray}
where we use the diffusion radius $r_{\rm diff}= r_{\rm in}$.
First we consider the $r$-process model. The observed temperature equals to the temperature at the inner edge of the ejecta, $T_{\rm obs}\sim [\dot{\epsilon}t\rho(v_{\min})/a]^{1/4}$. Using equation (\ref{sec2-1:rho}), we obtain
\begin{eqnarray}\label{sec2-2-3:T_obs,nuc}
T_{\rm obs}&\sim&\left(\frac{\dot{\epsilon}t\rho(t,v_{\min})}{a}\right)^{1/4} \nonumber \\
&\propto&M_{\rm ej}^{1/4}v_{\min}^{-3/4}t^{-\frac{2+\alpha}{4}}.
\end{eqnarray}
Since the energy is continuously injected due to the nuclear heating in the $r$-process model, the bolometric luminosity from the entire ejecta is described as $L_{\rm bol}\sim M_{\rm ej}\dot{\epsilon}$. However, the outer part of the ejecta emits photons with lower temperature and/or X-rays and $\gamma$-rays produced directly in radioactive decays. Although such emission contributes to the bolometric luminosity, we here focus only on the optical and infrared emissions. In the thick-diffusion phase, the observed emission comes from the region between $\sim r_{\rm diff}$ and $\sim2r_{\rm diff}$. In the transparent phase, we assume that the time evolution of the diffusion radius $r_{\rm diff}$ is the same as the thick-diffusion phase until $2r_{\rm diff}=r_{\rm in}$ and the observed luminosity comes from the region from $r_{\rm in}$ to $2r_{\rm diff}$ for simplicity. Then, the bolometric luminosity is described as
\begin{eqnarray}\label{sec2-2-3:L_bol,nuc}
L_{\rm bol}&\sim&4\pi r_{\rm in}^3\rho(t,v_{\min})\dot{\epsilon} \nonumber \\
&\propto&M_{\rm ej}t^{-\alpha}.
\end{eqnarray}
Although it appears that this time evolution directly reflects the nuclear decay rate, when we calculate the mass between $r_{\rm in}$ and $2r_{\rm diff}$ the evolution of the upper limit of the integration $2r_{\rm diff}$ makes the decrease of the luminosity faster than $\propto t^{-\alpha}$ (see a dashed line in the middle panel of figure \ref{figure:evolution}). In addition, the evolution of $r_{\rm diff}$ depends on the index $\beta$ (see equation \ref{sec2-2-2:r_diff}), so that the mass between $r_{\rm in}$ and $2r_{\rm diff}$ also depends on the index $\beta$.
Next we consider the engine model. We assume that the internal energy is exhausted when the diffusion radius reaches $2r_{\rm diff}=r_{\rm in}$. For the observed temperature $T_{\rm obs}$, we assume the relation $T_{\rm obs}=T(t,v_{\min})$ and use equation (\ref{sec2-1:T}),
\begin{eqnarray}\label{sec2-2-3:T_obs,sh}
T_{\rm obs}&\sim&T_0\left(\frac{t}{t_{\rm inj}}\right)^{-1} \nonumber \\
&\propto&E_{\rm int0}^{1/4}t_{\rm inj}^{1/4}t^{-1} \nonumber \\
& &\times\left\{ \begin{array}{ll}
v_{\min}^{-\xi}v_{\max}^{\frac{4\xi-3}{4}} & (\xi < 0.75) \\
& \\
v_{\min}^{-3/4} & (\xi > 0.75). \\
\end{array} \right.
\end{eqnarray}
The bolometric luminosity is described as
\begin{eqnarray}\label{sec2-2-3:L_bol,sh}
L_{\rm bol}&\sim&4\pi \int_{r_{\rm in}}^{2r_{\rm diff}}\frac{aT_{\rm obs}^4}{t} \nonumber \\
&\propto&E_{\rm int0}t_{\rm inj}t^{-2} \nonumber \\
& &\times\left\{ \begin{array}{ll}
\kappa^{\frac{3-4\xi}{\beta-2}}M_{\rm ej}^{\frac{3-4\xi}{\beta-2}}v_{\min}^{\frac{(\beta-3)(3-4\xi)}{\beta-2}} & \\
~~\times v_{\max}^{4\xi-3}t^{\frac{2(3-4\xi)}{2-\beta}} & (\xi < 0.75) \\
& \\
1 & (\xi > 0.75). \\
\end{array} \right.
\end{eqnarray}
Since the internal energy at the innermost region almost equals to the total internal energy $E_{\rm int}(v_{\min})\sim E_{\rm int0}(t/t_{\rm inj})^{-1}$ and determines the bolometric luminosity $L_{\rm bol}\sim E_{\rm int}(v_{\min})/t$ for the temperature index $\xi>0.75$, the bolometric luminosity does not depend on the mass $M_{\rm ej}$ and velocities $v_{\max}$ and $v_{\min}$. This luminosity always corresponds to the maximum luminosity for $\xi > 0.75$, so that we can impose the lower limit on the parameter $E_{\rm int0}t_{\rm inj}$.
\subsection{Fiducial Model}
\label{fiducial}
\begin{figure}
\begin{center}
\includegraphics[width=70mm]{fig3.ps}
\caption{ Temporal evolution of the diffusion radius (top), bolometric luminosities (middle) and observed temperatures (bottom) in the fiducial model (first column of table 1). Thick dashed and solid lines show the evolution for the $r$-process model and the engine model, respectively. For comparison, we also plot the bolometric luminosity from the whole ejecta for the $r$-process model after the transparent phase ($t>t_{\rm tr}$ in equation \ref{sec2-2-3:t_tr}) as a blue long-dashed line in the middle panel.}
\label{figure:evolution}
\end{center}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=120mm]{fig4.ps}
\caption{Theoretical light curves calculated under the fiducial parameter set (Table 1) at a near-infrared band (F160W, {\it red}) and optical band (F606W, {\it blue}). Two models (the $r$-process model, {\it solid}; the engine model, {\it dashed}) are considered. The observational results of GRB 130603B \citep[z = 0.356; ][]{Tan+13, BFC13, Cuc13, de14} are also plotted. The thin dotted lines are light curves calculated from a GRB afterglow model \citep{Tan+13}. Both models can reproduce the observational data well. }
\label{figure:observation}
\end{center}
\end{figure*}
We show the temporal evolution of the diffusion radius $r_{\rm diff}$, the bolometric luminosity $L_{\rm bol}$ and the observed temperature $T_{\rm obs}$ in figure \ref{figure:evolution} under the fiducial parameter set. The parameters are summarized in the first column of table 1. Here, we do not use approximations $\rho(v)\sim\rho(v_{\rm max})$ and $T(v)\sim T(v_{\max})$ at the thin-diffusion phase as in section \ref{thin}. Instead, the diffusion radius $r_{\rm diff}$ is calculated from equations (\ref{sec2-1:t_diff}) -- (\ref{sec2-1:tau}) without approximations. Using the obtained diffusion radius $r_{\rm diff}$ and the relation $v_{\rm diff}=r_{\rm diff}/t$, we calculate the observed temperatures in the thin- and thick-diffusion phases, $T_{\rm obs}\sim[\dot{\epsilon}t\rho(t,v_{\rm diff})/a]^{1/4}$ (equation \ref{sec2-2-2:T_obs,nuc}), and $T_{\rm obs}\sim T_0(t/t_{\rm inj})^{-1}(v_{\rm diff}/v_{\min})^{-\xi}$ (equation \ref{sec2-2-2:T_obs,sh}) for the $r$-process and the engine models, respectively. In the transparent phase, the temperature in equations (\ref{sec2-2-3:T_obs,nuc}) and (\ref{sec2-2-3:T_obs,sh}) are evaluated with $v=v_{\min}$. Equations on observed temperature and bolometric luminosity for both models are summarized in appendix. The set of parameters we choose here explains the observed optical and infrared light curves of GRB 130603B (see next section). The vertical dash-dotted lines in figure \ref{figure:evolution} show the time $t=t_{\times}$ (equation \ref{sec2-2-2:t_times}). The diffusion radius is plotted only up to the transition time $t=t_{\rm tr}$ (equation \ref{sec2-2-3:t_tr}).
In the thick-diffusion phase, the diffusion radius ($r_{\rm diff}\propto t^{\frac{\beta-4}{\beta-2}}=t^{-1/3}$ for $\beta=3.5$) moves inward in the ejecta ($r\propto t$). Since the observed luminosity and temperature are determined at the diffusion radius $r_{\rm diff}$, the time evolution of luminosity and temperature strongly depends on the indices of the profile, $\beta$ and $\xi$. For the $r$-process model, the bolometric luminosity decreases with time ($L_{\rm bol}\propto t^{\frac{2(\beta-3)}{\beta-2}-\alpha}=t^{-0.633}$) in the thick-diffusion phase (see equation \ref{sec2-2-2:L_bol,nuc}), which is more rapid than that in the thin-diffusion phase ($L_{\rm bol}\propto t^{1-\alpha}=t^{-0.3}$, see equation \ref{sec2-2-1:L_bol,nuc}). Since the index of the mass density $\beta=3.5$ is close to 3, in which the mass of each shell with a certain size $\delta r$ is the same value in logarithmic scale, the mass between the diffusion radius $r_{\rm diff}$ and its doubled value $2r_{\rm diff}$ does not significantly change with time. The luminosity is mainly determined by that mass, so that the evolution of the luminosity is slow compared with the evolution of nuclear heating rate ($\propto t^{-\alpha}$) in the thick-diffusion phase. On the other hand, bolometric luminosity and observed temperature increase with time in the engine model with the parameter set of the fiducial model. These mainly reflect the profile of the temperature distribution ($\xi=1.6$). In fact, using equation (\ref{sec2-2-2:L_bol,sh}), the index of the time $t$ for the bolometric luminosity is $-2(\beta+1-4\xi)/(\beta-2)\sim2.53$ for the engine model.
After the transition time $t\ge t_{\rm tr}$, the luminosity and temperature are almost determined by the quantities at the inner edge of the ejecta. Then, the evolution of the luminosity and temperature does not significantly depend on the indices of profile $\beta$ and $\xi$ as in the case of the thin-diffusion phase (except for the case $\xi<0.75$ of the engine model, equation \ref{sec2-2-3:L_bol,sh}). Since our used profile of mass density has an artificially steep cut-off at the inner edge of the ejecta (figure \ref{figure:diffuse}), bolometric luminosity in both models rapidly declines after the time $t\ge t_{\rm tr}$. In the bottom panel of figure \ref{figure:evolution}, the observed temperature in both models has a steep cutoff at $2r_{\rm diff}=r_{\rm in}$. For comparison, we also consider the time evolution of bolometric luminosity from the whole ejecta $L_{\rm bol}=M_{\rm ej}t^{-\alpha}$ in the $r$-process model. Time evolution is shown in the middle panel of figure \ref{figure:evolution} as a blue long-dashed line. This luminosity evolution ($L_{\rm bol}\propto t^{-\alpha}=t^{-1.3}$) is significantly slower than that of the engine model in the transparent phase. In section \ref{implication}, we discuss the implication for discriminating the $r$-process model and the engine model using these temporal behaviors.
\section{DISCUSSION}
\label{discuss}
\begin{figure}
\begin{center}
\includegraphics[width=100mm]{fig5.ps}
\caption{Range of parameter space in order to explain the observations of the macronova, GRB 130603B ({\it blue} for the engine model and {\it red} for the $r$-process model). Since two regions are overlapped, the color looks like purple for the $r$-process model. These regions are only a schematic view. We fix $v_{\max}=0.4c$, $\beta=3.5$, $\dot{\epsilon}_0=2\times10^{10}$ erg s$^{-1}$ g$^{-1}$ and $\alpha=1.3$. For the opacity $\kappa$, we use the range $\kappa=3-30$ cm$^2$ g$^{-1}$. For the engine model, we treat $E_{\rm int0}t_{\rm inj}$ and $\xi$ as free parameters to fit the light curve. The circle, square and triangle denote the case of the fiducial model, minimum mass model and hot interior model, respectively (see table 1).}
\label{figure:parameter}
\end{center}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=170mm]{fig6.ps}
\caption{Theoretical light curves calculated under the hot interior model (table 1) at V ({\it purple}), R, F606W ({\it blue}) and F160W bands ({\it red}). Two models (the $r$-process model, {\it solid}; the engine model, {\it dashed}) are considered. The observational results of GRB 050509B \citep[left panel, $z=0.122$, ][]{Hjo+05}, GRB 080905A \citep[middle panel, $z=0.225$, ][]{Row+10} and GRB 130603B \citep[right panel, $z=0.356$, ][]{Tan+13, BFC13, Cuc13, de14} are also plotted. The engine model can reproduce all three observational data well.}
\label{figure:observation2}
\end{center}
\end{figure*}
\subsection{Comparison with GRB 130603B}
We compare the results with the optical and infrared observations of short GRB 130603B in figure \ref{figure:observation}. The fiducial parameter set in table 1 is adopted. The $r$-process model and the engine model result in similar light curves at the optical and infrared bands. Both of them satisfy the observational data of GRB 130603B. Note that the detection point at F606W band at $\sim10^5$ s is consistent with the afterglow of GRB 130603B modeled as a smoothly broken power law \citep[blue dashed line, ][]{Tan+13}. We regard this detected value as an upper limit for the luminosity of emission from the ejecta. The detection point at F160W band at $\sim10^6$ s exceeds the extrapolation of the afterglow emission \citep[red dashed line, ][]{Tan+13}, so that we regard this detected emission as a thermal radiation from the ejecta.
The range of the model parameters $v_{\min}$ and $M_{\rm ej}$ to satisfy the constraints obtained from the observation of GRB 130603B is shown in figure \ref{figure:parameter} as colored areas (red area for the $r$-process model and blue area for the engine model). Note that the red area has a completely overlap with the blue area. We fix the other model parameters $v_{\max}=0.4c$, $\beta=3.5$, $\dot{\epsilon}_0=2\times10^{10}$ erg s$^{-1}$ g$^{-1}$ and $\alpha=1.3$ as in the fiducial model. We take into account the uncertain range of the opacity, $\kappa=3-30$ cm$^2$ g$^{-1}$ to constrain the parameters, $v_{\min}$ and $M_{\rm ej}$. In the engine model, $\xi$ and $E_{\rm int0}t_{\rm inj}$ are additionally treated as free parameters to derive the allowed area in figure \ref{figure:parameter}.
\subsubsection{Limits on ejecta mass}
In the $r$-process model, the luminosity becomes smaller for smaller ejecta mass $M_{\rm ej}$. The small ejected mass $M_{\rm ej}\lesssim0.07M_{\odot}$ cannot reproduce the infrared excess of GRB 130603B (figure \ref{figure:observation}). The required ejecta mass is relatively large compared to the mass indicated by recent numerical simulations for a merger of binary NSs \citep[e.g., ][]{Hot+13, Ros+14, Jus+14}. Note that in \citet{BFC13}, $0.03 - 0.08 M_{\odot}$ is required to explain the observed infrared excess, which is a factor $\sim$2 smaller than our results. Their theoretical light curves are based on the study of \citet{BK13}. In \citet{BK13}, a broken power-law mass density profile with the index $-1$ for the inner layer and $-10$ for the outer layer of ejecta is adopted, in which the mass of the ejecta is efficiently concentrated at the transition point of the density index. Therefore, the luminosity of ejecta is evaluated as the heating rate multiplied by the total ejecta mass at the moment when the diffusion radius reaches the transition point. On the other hand, the index of our mass density profile of the ejecta is $\beta = 3.5$, which is indicated by general relativistic simulations by \citet{Hot+13}. This profile is quite different from the profile adopted in \citet{BK13}; the index is close to 3, in which the mass in each logarithmic radius is constant. Then, at the time $t = t_{\rm tr}$ (equation \ref{sec2-2-3:t_tr}), the mass contributing to the luminosity is about $\sim$60\% of the total ejecta mass in the case $v_{\min}=0.1c$. This profile predicts luminosity dimmer than that other studies. In fact, \citet{Hot+13b} tried to explain the observed infrared excess using the mass profile which is almost the same with ours. In the case of a binary NS merger with ejecta mass $\sim$0.02 $M_{\odot}$, even if they use a larger nuclear heating rate (larger by a factor of 2), their predicted luminosity is slightly smaller than the observed infrared excess (in the left panel of their figure 3). This result is consistent with our model, i.e., our model requires larger mass than most of previous studies.
In the engine model, the injected internal energy which determines the luminosity does not depend on the ejecta mass (except for the limit in equation \ref{sec2-1:E_int}). However, the luminosity declines rapidly after the transition time $t\gtrsim t_{\rm tr}$ which depends on the ejecta mass as in equation (\ref{sec2-2-3:t_tr}). The condition $t_{\rm tr}\gtrsim10^6$ s in the observer frame is required to reproduce the excess observed from GRB 130603B in the near-infrared band. This condition gives the lower limit for the ejecta mass in the engine model, $M_{\rm ej}\gtrsim0.02M_{\odot}$ with the opacity $\kappa\sim30$ cm$^2$ g$^{-1}$.
Note that the observed upper limit on the infrared luminosity at $\sim3\times10^6$ s in the observer frame (figure \ref{figure:observation}), which corresponds to $t_{\rm tr}\lesssim3\times10^6$ s, gives the upper limit on the ejecta mass for both models. However, this limit is not important for the range $M_{\rm ej}<0.2M_{\odot}$ in the range of the opacity $\kappa=3-30$ cm$^2$ g$^{-1}$.
\subsubsection{Limits on the minimum velocity}
The smaller minimum velocity $v_{\min}$ gives the smaller bolometric luminosity at certain time in the $r$-process model (see equations \ref{sec2-2-1:L_bol,nuc} and \ref{sec2-2-2:L_bol,nuc}). The small minimum velocity enlarges the size of ejecta (when we fix the maximum velocity $v_{\max}$). Then, the diffusion time $t_{\rm diff}$ of photons emitted from the inner region of the ejecta becomes long for the small velocity $v_{\min}$ (equation \ref{sec2-2-3:t_tr}). The mass between $r_{\rm diff}$ and $2r_{\rm diff}$ (or $r_{\rm out}$) increases toward inner region of the ejecta (as long as $\beta>3$) so that the mass is reduced for the small minimum velocity $v_{\rm min}$ at certain time. In fact, the dependence of the mass on the minimum velocity is $4\pi r_{\rm diff}^3\rho(t,v_{\rm diff})\propto v_{\min}^{\frac{\beta-3}{\beta-2}}=v_{\min}^{1/3}$. As a result, smaller minimum velocity gives smaller luminosity to reproduce the observed infrared excess of GRB 130603B. Moreover, smaller minimum velocity gives larger temperature $T_{\rm obs}$ at certain time (equations \ref{sec2-2-2:T_obs,nuc} and \ref{sec2-2-3:T_obs,nuc}) because mass density at a shell with small velocity is large. The difference between the detected luminosity at F160W band and the upper limit on the luminosity at F606W band at $\sim10^6$ s in the observer frame gives the upper limit on the observed temperature ($T_{\rm obs}\lesssim4\times10^3$ K). To satisfy the observed upper limit on the temperature from GRB 130603B, a lower limit of $v_{\min}\gtrsim0.1c$ is obtained for $M_{\rm ej}\sim0.1M_{\odot}$.
The smaller minimum velocity $v_{\min}$ gives higher temperature $T_{\rm obs}$ in the engine model (equations \ref{sec2-2-2:T_obs,sh} and \ref{sec2-2-3:T_obs,sh}). The observational limit for the temperature at $\sim10^6$ s in the observer frame indicates that the range of the minimum velocity $v_{\rm min}$ is limited in the engine model ($v_{\min}\gtrsim0.06c$ for $M_{\rm ej}\sim0.1M_{\odot}$).
\subsubsection{Dependence on opacity}
We discuss the dependence on the value of $\kappa$. As mentioned in section 2.2, we use the temperature-independent opacity $\kappa$ with the grey approximation. In general, the $r$-process line opacity depends on frequency and changes with temperature and ionization state of the ejecta \citep{KBB13, TH13}. The indicated grey opacity is $\kappa=3-30$ cm$^2$ g$^{-1}$.
In the case of the $r$-process model, the luminosity significantly depends on opacity $\kappa$. The larger opacity causes larger diffusion time $t_{\rm diff}$, so that larger time is required to observe the inner region of the ejecta for given ejecta mass $M_{\rm ej}$ and minimum velocity $v_{\min}$. In fact, two transition times $t_{\times}$ and $t_{\rm tr}$ are proportional to $\kappa^{1/2}$ (equations \ref{sec2-2-2:t_times} and \ref{sec2-2-3:t_tr}). Then, the mass around the diffusion radius $r_{\rm diff}$ is small at certain time, so that the luminosity is reduced. As a result, in order to explain the infrared excess observed in GRB 130603B, larger mass $M_{\rm ej}$ is required for the larger value of opacity $\kappa$. For the opacity $\kappa>30$ cm$^2$ g$^{-1}$, total ejecta mass $M_{\rm ej}\gtrsim0.2M_{\odot}$ is required to reproduce the observed excess, which is much larger than the simulation results of mergers of binary NSs \citep[e.g., ][]{Hot+13}. On the other hand, the transition time $t_{\rm tr}$ is smaller for the smaller value of the opacity. Then, the luminosity significantly increases at $\sim10^5$ s. For the opacity $\kappa\lesssim3$ cm$^2$ g$^{-1}$, there is no parameter set which gives smaller luminosity than the detection at F606W band ($\sim10^5$ s in the observer frame) and the luminosity comparable to the observed excess at F160W band simultaneously in the $r$-process model.
In the case of the engine model, a larger value of opacity $\kappa$ reduces the lower limit for the mass $M_{\rm ej}$ to explain the observed excess. For certain temperature and luminosity, the opacity $\kappa$ and the ejecta mass $M_{\rm ej}$ always degenerate in the form $\kappa M_{\rm ej}$ (see equations \ref{sec2-2-1:T_obs,sh}, \ref{sec2-2-1:L_bol,sh}, \ref{sec2-2-2:T_obs,sh}, \ref{sec2-2-2:L_bol,sh}, \ref{sec2-2-3:T_obs,sh} and \ref{sec2-2-3:L_bol,sh}). This dependence comes from the optical depth (equation \ref{sec2-1:tau}) because the internal energy in the ejecta does not depend on the opacity and the ejecta mass, contrary to the $r$-process model. We present a parameter set to give the minimum ejecta mass $M_{\rm ej}$ in table 1 as the minimum mass parameter set. We also plot the value of $M_{\rm ej}$ and $v_{\min}$ of this model in figure \ref{figure:parameter} as a square. This ejecta mass is naturally realized in general relativistic simulations \citep[e.g., ][]{Hot+13}. Although the larger value of the opacity $\kappa$ reduces the lower limit for the ejecta mass $M_{\rm ej}$, the kinetic energy is $E_{\rm kin}\sim1.1\times10^{51}$ erg (equation \ref{sec2-1:E_kin}) which is close to the initial injected energy $E_{\rm int0}=0.9\times10^{51}$ erg for the minimum mass parameter set. The lower ejecta mass $M_{\rm ej}$ reduces the kinetic energy of the ejecta, $E_{\rm kin}(\propto M_{\rm ej}v_{\max}^{5-\beta}v_{\min}^{\beta-3})$ in equation (\ref{sec2-1:E_kin}), so that the required energy $E_{\rm int0}$ may exceed the kinetic energy of the ejecta for the larger opacity.
For the small value of the opacity, larger mass and smaller minimum velocity is required to satisfy the condition $t_{\rm tr}\gtrsim10^6$ s ($t_{\rm tr}\propto\kappa^{1/2}M_{\rm ej}^{1/2}v_{\min}^{-1/2}$ in equation \ref{sec2-2-3:t_tr}). For the opacity $\kappa=3$ cm$^2$ g$^{-1}$, the condition corresponds to $(M_{\rm ej}/0.2M_{\odot})(v_{\min}/0.1c)^{-1}\gtrsim1$. The observational constraint for the temperature also requires a large value of the minimum velocity $v_{\min}$. Then, there is no solution to explain the observed excess within the parameter range shown in figure \ref{figure:parameter} for the opacity $\kappa\le3$ cm$^2$ g$^{-1}$. Therefore, for the small opacity $\kappa\le 3$ cm$^2$ g$^{-1}$ the engine model cannot explain the observed excess.
\subsubsection{Dependence on engine parameters}
Since the engine model has additional free parameters, $\xi$ and $E_{\rm int0}t_{\rm inj}$, the allowed region of the parameters is larger than that of the $r$-process model. We can impose the lower limit on the parameter $E_{\rm int0}t_{\rm inj}$ by regarding the infrared luminosity $\sim10^{41}$ erg s$^{-1}$ at $t\sim 7$ day as bolometric luminosity in the source rest frame with equation (\ref{sec2-2-3:L_bol,sh}). The derived limit is $(E_{\rm int0}/10^{51}{\rm erg})(t_{\rm inj}/10^2{\rm s})\gtrsim0.4$. To satisfy the optical upper limit at $\sim10^5$ s and the detected luminosity at $\sim10^6$ s in the observer frame (figure \ref{figure:observation}), we find the lower limit on the index of the temperature profile $\xi\gtrsim1.0$. For a smaller value of the index $\xi$, emission from the ejecta with relatively high temperature can be observed at time $\sim10^5$ s, so that luminosity at F606W band is larger than the observed upper limit of GRB 130603B. In addition, the smaller value of $\xi$ decreases relative internal energy in the inner edge of the ejecta. To reproduce the luminosity at time $\sim 10^6$ s in the observer frame when observed emission comes from the inner ejecta, the smaller $\xi$ requires the larger initial internal energy $E_{\rm int0}$ which exceeds the kinetic energy of the ejecta $E_{\rm kin}$ in some cases.
\subsection{Comparison with Other GRBs with Deep Optical Observations}
\label{other}
Several deep optical observations of short GRBs give stringent upper limits on the luminosity of macronovae \citep{Kann+11}. We compare the results with two deep optical observations of short GRBs, GRB 050509B and GRB 080905A. For the fiducial parameter set, the luminosity exceeds the observational upper limits on these two observations. In the engine model, we can reduce the luminosity in the early phase $\lesssim10^5$ s without reducing the luminosity in the late phase $\sim10^6$ s by utilizing the steep temperature profile (large $\xi$).
Here, we introduce the hot interior parameter set with larger value of index $\xi$ than that of the fiducial parameter set. Since emission from the inner part of the ejecta is observed at the later time, the luminosity at the early phase decreases and avoids the observational limits if most of the internal energy is injected to the inner part of the ejecta. We show the light curve of the hot interior parameter set in figure \ref{figure:observation2}. We choose the parameters as $M_{\rm ej}=0.08M_{\odot}, v_{\min}=0.18c, t_{\rm inj}=10^2{\rm s}, E_{\rm int0}=0.8\times10^{51}{\rm erg}$ and $\xi=2.7$ (the right column of table 1). From figure \ref{figure:observation2}, the light curves are consistent with all three observations using the same model parameters. A possible scenario for the hot interior parameter set is that the shock produced by the activity of the central engine may not be able to catch up with the outer part of the ejecta because the velocity of the ejecta is close to the light speed ($v_{\max}=0.4c$). Then, only the inner part of the ejecta will be heated. For comparison, we also show the light curves in the $r$-process model with the parameter set of the hot interior in figure \ref{figure:observation2} as dashed lines. The luminosity of the $r$-process model exceeds the observed upper limits in two observations, GRB 050509B and GRB 080905A (left and middle panels of figure \ref{figure:observation2}) if we choose the parameter set of the hot interior model. We are not able to find any parameter set in the $r$-process model, which simultaneously satisfies the observed limits of the three observations. Note that we do not argue that the $r$-process model is excluded from these results because we need to take into account the variations of the model parameters for each event.
Note that the extended emission was not detected in three short GRBs. However, it is not unreasonable to miss the extended emission of these bursts. One possibility is a selection effect. Observationally, the fraction of short GRBs with extended emission is significantly larger at softer energy bands: $\sim$25\% in the Swift BAT samples \citep[$>$15 keV; ][]{NGS10} and $\sim$7\% in the BATSE samples \citep[$>$20 keV; ][]{BKG13}. This suggests that observations with a low energy threshold may dramatically increase short GRBs with extended emission \citep{Nak+13}. This will be further tested by future soft X-ray survey facilities such as Wide-Field MAXI (0.7-10 keV) \citep{Kaw+14}. The three referred short GRBs were detected by Swift BAT and therefore Swift BAT could not detect extended emission by chance. Alternatively, the outflow following the main short GRB jet could not breakout the ejecta. \citet{Nag+14} and \citet{Mur+14} investigated the propagation of jets in merger ejecta. They found the cases that relativistic jets can penetrate merger ejecta and produce the prompt emission of short GRBs, but in the late energy injection cases, outflow fails to breakout the ejecta. Therefore, some extended emission may not be observed, although the central engine works actively.
\subsection{Outer Region of Mass Density Profile}
\label{outer}
\begin{figure}
\begin{center}
\includegraphics[width=80mm]{fig7.ps}
\caption{Dependence of theoretical light curves for the $r$-process model on the shape of the front of the ejecta. We plot the case of GRB 130603B. Thick-dashed lines denote the fiducial model. Thin-solid, thin-dashed, and thin-dot-dashed lines correspond to the exponential profiles of mass density (equation \ref{sec3:rho}) with $v_{\max}' = 0.4c, 0.5c$ and $0.6c$, respectively. We also plot the observational data from \citet{Tan+13, BFC13, Cuc13, de14}.}
\label{figure:exp_comp}
\end{center}
\end{figure}
In the thin-diffusion phase, the light curve strongly depends on the density profile of the ejecta surface. The density profile is determined by the complex merger dynamics \citep{Hot+13}, so that the density profile of the ejecta cannot be analytically derived as mentioned in section 2.3. Since the outer part of the density profile is difficult to calculate precisely, little attention has paid on the mass profile at the outer region in current numerical simulations. In order to investigate the dependence of the light curve on the mass profiles in the thin-diffusion phase, we consider other forms of the mass profile and compare the light curve with that of equation (\ref{sec2-1:rho}). We adopt an exponential profile
\begin{eqnarray}\label{sec3:rho}
\rho(t,v)&=&\rho_0\left(\frac{t}{t_0}\right)^{-3}\left(\frac{v}{v_{\rm min}}\right)^{-\beta} \nonumber \\
& & \times\exp\left(-\frac{v-0.5v_{\max}}{v_{\max}'-v}\right).
\end{eqnarray}
We introduce an additional free parameter $v_{\max}'(\ge v_{\max})$ and the ejecta expand $v_{\min}\le v\le v_{\max}'$. The calculations are the same except for using $\Delta r\sim v_{\max}'t - r_{\rm diff}$. We fix the mass with velocity larger than $0.5v_{\max}$ and calculate three models for $v_{\max}'=0.4c, 0.5c, 0.6c$. For the other parameters, we adopt from the fiducial parameter set in table 1. We show the results in the $r$-process model in figure \ref{figure:exp_comp}. Since we fix the mass with velocity larger than $0.5v_{\max}$, the bolometric luminosities at the time $t\sim t_{\times}$ are almost the same values. In the thin-diffusion phase ($t\ll t_{\times}$), both the luminosity and the temperature are smaller than the fiducial model. This is because the density at the front of the ejecta is reduced in this mass profile. Since the maximum velocity effectively becomes large and the adiabatic cooling becomes efficient, these effects for luminosity and temperature should be also seen in the engine model. We conclude that the luminosities in the thin-diffusion phase ($t\ll t_{\times}$) have uncertainties at least with $\sim1-2$ mag, which originates from the uncertainty of the outermost mass profile.
Note that the emission from the ejecta with the mass profile discussed here reduces the tension between the light curve in the $r$-process model and the upper limits of the deep optical observations (GRB 050509 and GRB 080905A) as discussed in section \ref{other}. Especially, the optical luminosity in the case $v_{\max}'=0.6c$ (thin-dot-dashed line) significantly decreases after $t\gtrsim10^5$ s. Therefore, the ejecta with relatively shallow mass distribution at the front of the ejecta is able to explain the current optical follow-up observations in the $r$-process model.
\subsection{Implications to Discriminate Two Models}
\label{implication}
In the fiducial parameter set, the light curves for two models in the optical and infrared bands are similar (figure \ref{figure:observation}). In figure \ref{figure:parameter}, the allowed parameter region to explain the observation of GRB 130603B for both models are also overlapped. Recall that emission from ejecta is described as blackbody radiation for the two models, whose spectrum is narrow in bands. Therefore, no excess at other wavelengths is expected and also the prediction is consistent with radio observations \citep{Fong+14}. Therefore, it is difficult to discriminate two models from the currently available observational data.
In the $r$-process model, the light curve with mass profile $\rho\propto v^{-\beta}$ ($3\lesssim\beta\lesssim4$) and a parameter set which explains the infrared excess detected from GRB 130603B cannot explain the upper limits obtained from the deep optical observations of some short GRBs. Therefore, if both stringent optical upper limits at $\sim10^5$ s and bright infrared emission at $\sim10^6$ s are simultaneously obtained from a single event (with a difference larger than two magnitudes $M_{\rm optical}(\sim10^5~{\rm s}) - M_{\rm infrared}(\sim10^6~{\rm s})\gtrsim2$ mag ), the $r$-process model is significantly restricted. For the engine model, these observations give a constraint for the temperature distribution in the ejecta, which may give new insights into the activity of the central engine.
As shown in the middle panel of figure \ref{figure:evolution}, the bolometric luminosity in the $r$-process model from the whole ejecta (blue long-dashed line), including low temperature and/or X-rays and $\gamma$-rays produced directly in radioactive decays \citep{Chu14}, declines more gradually than that for the engine model. This is because there is no energy injection after the time $t>t_{\rm inj}$ for the engine model. Then, the luminosity significantly decreases when photons at the inner edge of the ejecta begin to diffuse out (see the middle panel of figure \ref{figure:evolution} and figure \ref{figure:observation}). The luminosity from the whole ejecta can be described as $L_{\rm bol}\sim M_{\rm ej}\dot{\epsilon}$ in the transparent phase. The index of time $t$ is determined by the nuclear heating rate, $\alpha\sim1.3$. Therefore, the two models are distinguishable by observing the temporal evolution of bolometric luminosity from the whole ejecta in this phase.
\section{SUMMARY}
\label{summary}
We calculated the light curves of macronovae by developing analytical models. We modeled the ejecta based on the results of numerical simulations for a merger of binary NSs. In addition to the nuclear decay of $r$-process elements (the $r$-process model which is often discussed), we considered another heating mechanism for the ejecta, the engine-driven shock (engine model). We compared the results with the optical and infrared observations of the first macronova candidate associated with GRB 130603B, and showed that both models can explain the observations. In order to reproduce the observed light curve, the $r$-process model requires relatively large ejecta mass $M_{\rm ej}\gtrsim0.07M_{\odot}$ which is mainly determined by the observed infrared luminosity $\sim10^{41}$ erg s$^{-1}$ at $\sim10^6$ s. In the engine model, the internal energy of ejecta, which mainly determines the observed luminosity, does not depend on the ejecta mass. Then, unless the entire of the ejecta is effectively thin (the diffusion time is smaller than the dynamical time, $t_{\rm diff}<t$, at the inner edge of the ejecta) \footnote{Since there is no energy injection after $t>t_{\rm inj}\sim10^2$ s in the engine model, the luminosity rapidly decreases after the time $t_{\rm tr}\propto M_{\rm ej}^{1/2}$ when photons can diffuse out from the inner edge of the ejecta (section \ref{transparent}).}, the required ejecta mass is $M_{\rm ej}\gtrsim0.02M_{\odot}$, which is comparable to the recent numerical simulation results. The initial internal energy $E_{\rm int0}$ and the injection time $t_{\rm inj}$ are required as $(E_{\rm int0}/10^{51}{\rm erg})(t_{\rm inj}/10^2{\rm s})\gtrsim 1$, which is consistent with the observed extended emission of short GRBs, $E_{\rm iso}\sim10^{50}-10^{51}$ erg and $t_{\rm dur}\sim10-10^2$ s. The required minimum velocity is about $v_{\min}\gtrsim0.05c$ for both models, which is mainly determined by the constraint for the observed temperature $\lesssim4\times10^3$ K at $\sim10^6$ s. For the range of the opacity $\kappa\lesssim3$ cm$^2$ g$^{-1}
, it is difficult for both models to explain the observations of macronova associated with GRB 130603B by the ejecta mass less than $M_{\rm ej}<0.2M_{\odot}$.
If macronovae are identical, the upper limits on the luminosity obtained in the deep optical observations of other short GRBs give stringent constraints on the $r$-process model. On the other hand, the engine model satisfies these constraints if the temperature profile is centrally concentrated in the ejecta (large $\xi$). Thus, if the difference between the optical magnitude at $\sim10^5$ s and the infrared magnitude at $\sim10^6$ s is larger than $\sim2$ mag in a single event, the $r$-process model is difficult to explain the observations unless the front of the ejecta has much shallow mass distribution. Another difference in the light curves between two models is the bolometric luminosity at the transparent phase when dynamical time is smaller than the diffusion time at the inner edge of the ejecta $r_{\rm in}$. Although the optical and infrared luminosities rapidly decrease in the transparent phase, the bolometric luminosity from the whole ejecta, including lower frequency than near-infrared band and/or X-rays and $\gamma$-rays produced directly in radioactive decays, is determined by the energy injection rate of nuclear decay, $\dot{\epsilon}\propto t^{-\alpha}~(\alpha\sim1.3)$. For the engine model, the bolometric luminosity decreases rapidly in this phase (faster than $t^{-2}$). Therefore, we expect that the light curve of the bolometric luminosity from the whole ejecta can distinguish between two heating mechanisms.
Our results show that early light curves depend on the density profile of the outermost edge of the ejecta. It is necessary to develop a method to calculate the low-density region of the ejecta in either the analytical or numerical ways in order to precisely predict the early light curves of macronovae.
\acknowledgments
We thank K. Asano, K. Kashiyama, K. Kiuchi, H. Nagakura, T. Nakamura, Y. Sekiguchi, M. Shibata for fruitful discussions. This work is supported by KAKENHI 24103006 (S.K., K.I.), 24.9375 (H.T.), 24000004, 26247042, 26287051 (K.I.).
\renewcommand{\theequation}{A-\arabic{equation}}
\setcounter{equation}{0}
\section*{APPENDIX A. Analytic formulae for macronova light curves}
We summarize the formula for the observed temperature and bolometric luminosity. The detailed derivation of equations in this section is described in sections 2 and 3.
Since we assume that the observed luminosity and temperature approximate the luminosity and temperature at the diffusion radius $r_{\rm diff}$ (section 3), we need to calculate the diffusion radius. For the dynamics of the ejecta, we assume an isotropic and homologous expansion. Then, the velocity of ejecta $v$ is described by equation (\ref{sec2:v})
\begin{eqnarray}
v\sim r/t
\end{eqnarray}
where the radius $r$ originates the central engine and the time $t$ is measured from the time when a compact binary merges. As in section \ref{fiducial}, we calculate the diffusion radius $r_{\rm diff}$ from the condition that the diffusion time equals the dynamical time, $t_{\rm diff}=t$. The diffusion time is described by equation (\ref{sec2-1:t_diff}) as
\begin{eqnarray}
t_{\rm diff}\sim\tau\frac{\Delta r}{c},
\end{eqnarray}
where $c$ is the speed of the light, $\tau$ is the optical depth described by equation (\ref{sec2-1:tau}) as
\begin{eqnarray}
\tau=\left\{ \begin{array}{ll}
{\displaystyle \int_{r_{\rm diff}}^{r_{\rm out}}\kappa\rho dr} &~~(r_{\rm diff} > 0.5r_{\rm out}) \\
& \\
{\displaystyle \int_{r_{\rm diff}}^{2r_{\rm diff}}\kappa\rho dr} &~~(r_{\rm diff} \le 0.5r_{\rm out} ), \\
\end{array} \right.
\end{eqnarray}
and $\Delta r$ is the width of the diffusion region described by the equation (\ref{sec2-1:Deltar}) as
\begin{eqnarray}
\Delta r\sim \left\{ \begin{array}{ll}
{\displaystyle r_{\rm out}-r_{\rm diff}} &~~(r_{\rm diff}>0.5r_{\rm out}) \\
& \\
{\displaystyle r_{\rm diff}} &~~(r_{\rm diff}\le 0.5r_{\rm out}). \\
\end{array} \right.
\end{eqnarray}
In the calculation of the optical depth $\tau$, we use the spatially uniform value of the optical depth $\kappa$ with grey approximation and the ejecta mass density $\rho(t,v)$ described by equation (\ref{sec2-1:rho}) as,
\begin{eqnarray}
\rho (t,v)=\rho_0 \left(\frac{t}{t_0}\right)^{-3}\left(\frac{v}{v_{\min}}\right)^{-\beta},
\end{eqnarray}
where $\rho_0$ and $t_0$ are normalized factors, and $v_{\min}$ is the velocity at the inner edge of the ejecta. The radius $r_{\rm out}$ is the outer edge of the ejecta, described by equation (\ref{sec2-1:r_out}) as
\begin{eqnarray}
r_{\rm out}=v_{\max}t,
\end{eqnarray}
where the velocity $v_{\max}$ is at the outer edge of the ejecta. The radius at the inner edge of the ejecta $r_{\rm in}$ is described by equation (\ref{sec2-1:r_in}) as
\begin{eqnarray}
r_{\rm in}=v_{\min}t.
\end{eqnarray}
The normalization factor $\rho_0t_0^3$ in the profile of the mass density is determined by the ejecta mass $M_{\rm ej}$ (in equation \ref{sec2-1:M_ej}) as
\begin{eqnarray}
M_{\rm ej}=4\pi\int_{v_{\min}t_0}^{v_{\max}t_0}\rho(t_0,v)r^2dr.
\end{eqnarray}
We consider two heating sources of the ejecta. In the $r$-process model, the internal energy of the ejecta is determined by the nuclear heating rate of the $r$-process element described in equation (\ref{sec2-1:dotepsilon}) as
\begin{eqnarray}
\dot{\epsilon}=\dot{\epsilon}_0\left(\frac{t}{1{\rm day}}\right)^{-\alpha}.
\end{eqnarray}
For the engine model, we assume the temperature profile of the ejecta $T(t,v)$ as a result of the activity of the central engine, described in equation (\ref{sec2-1:T}) as
\begin{eqnarray}
T(t,v)\sim T_0\left(\frac{t}{t_{\rm inj}}\right)^{-1}\left(\frac{v}{v_{\min}}\right)^{-\xi},
\end{eqnarray}
where $T_0$ is the normalization factor. This factor is determined by injected internal energy $E_{\rm int0}$ at the time $t_{\rm inj}$, which is described in equation (\ref{sec2-1:E_int}) as
\begin{eqnarray}
E_{\rm int0}=4\pi\int_{v_{\min}t_{\rm inj}}^{v_{\max}t_{\rm inj}}aT^4(t_{\rm inj},v)r^2dr.
\end{eqnarray}
Observed temperatures in the $r$-process model and the engine model are given by
\begin{eqnarray}
T_{\rm obs}\sim\left\{ \begin{array}{l}
{\displaystyle \left[\frac{\dot{\epsilon}t\rho(t,v_{\rm diff})}{a}\right]^{1/4}} \\
~~~~~~~~~~~~~~~(r_{\rm diff}>r_{\rm in}) \\
\\
{\displaystyle \left[\frac{\dot{\epsilon}t\rho(t,v_{\min})}{a}\right]^{1/4}} \\
~~~~~~~~~~~~~~~(0.5r_{\rm in} < r_{\rm diff} \le r_{\rm in}) \\
\\
{\displaystyle 0} \\
~~~~~~~~~~~~~~~(r_{\rm diff} \le 0.5r_{\rm in}), \\
\end{array} \right.
\end{eqnarray}
and by
\begin{eqnarray}
T_{\rm obs}\sim \left\{ \begin{array}{l}
{\displaystyle T_0\left(\frac{t}{t_{\rm inj}}\right)^{-1}\left(\frac{v_{\rm diff}}{v_{\min}}\right)^{-\xi}} \\
~~~~~~~~~~~~~~~~~(r_{\rm diff} > r_{\rm in}) \\
\\
{\displaystyle T_0\left(\frac{t}{t_{\rm inj}}\right)^{-1}} \\
~~~~~~~~~~~~~~~~~(0.5r_{\rm in} < r_{\rm diff} \le r_{\rm in}) \\
\\
0 \\
~~~~~~~~~~~~~~~~~(r_{\rm diff} \le 0.5r_{\rm in}), \\
\end{array} \right.
\end{eqnarray}
respectively. Note that we do not use the approximation $\rho(v,t)\sim\rho(v_{\max},t)$ and $T(v,t)\sim T(v_{\max},t)$ in the thin-diffusion case. The bolometric luminosities for the $r$-process model and the engine model are given by
\begin{eqnarray}
L_{\rm bol}\sim\left\{ \begin{array}{l}
{\displaystyle 4\pi \int_{r_{\rm diff}}^{r_{\rm out}}\rho(v,t)\dot{\epsilon}r^2dr} \\
~~~~~~~~~~~~~~~(r_{\rm diff} > 0.5r_{\rm out}) \\
\\
{\displaystyle 4\pi \int_{r_{\rm diff}}^{2r_{\rm diff}}\rho(v,t)\dot{\epsilon}r^2dr} \\
~~~~~~~~~~~~~~~(r_{\rm in} < r_{\rm diff} \le 0.5r_{\rm out}) \\
\\
{\displaystyle 4\pi \int_{r_{\rm in}}^{2r_{\rm diff}}\rho(v,t)\dot{\epsilon}r^2dr} \\
~~~~~~~~~~~~~~~(0.5r_{\rm in} < r_{\rm diff} \le r_{\rm in}) \\
\\
{\displaystyle 0} \\
~~~~~~~~~~~~~~~(r_{\rm diff} \le 0.5r_{\rm in}), \\
\end{array} \right.
\end{eqnarray}
and
\begin{eqnarray}
L_{\rm bol}\sim\left\{ \begin{array}{l}
{\displaystyle 4\pi\int_{r_{\rm diff}}^{r_{\rm out}}\frac{aT_{\rm obs}^4}{t}r^2dr} \\
~~~~~~~~~~~~~~~(r_{\rm diff} > 0.5r_{\rm out}) \\
\\
{\displaystyle 4\pi\int_{r_{\rm diff}}^{2r_{\rm diff}}\frac{aT_{\rm obs}^4}{t}r^2dr} \\
~~~~~~~~~~~~~~~(r_{\rm in} < r_{\rm diff} \le 0.5r_{\rm out}) \\
\\
{\displaystyle 4\pi\int_{r_{\rm in}}^{2r_{\rm diff}}\frac{aT_{\rm obs}^4}{t}r^2dr} \\
~~~~~~~~~~~~~~~(0.5r_{\rm in} < r_{\rm diff} \le r_{\rm in}) \\
\\
0 \\
~~~~~~~~~~~~~~~(r_{\rm diff} \le 0.5r_{\rm in}), \\
\end{array} \right.
\end{eqnarray}
respectively. An example of the calculated result is shown in figure \ref{figure:evolution}.
We present the numerical values with the parameter dependence for later use. Unlike equations (\ref{sec2-2-2:t_times}) and (\ref{sec2-2-3:t_tr}), we include the contribution from subdominant terms to the numerical values when we integrate equations. Some of the subdominant terms include the ratio $v_{\max}/v_{\min}$. Hereafter, the value $v_{\max}/v_{\min}=4$ in subdominant terms are fixed and are not included in the parameter dependence. We introduce the normalized quantities $M_{\rm ej,0.1}\equiv M_{\rm ej}/0.1M_{\odot}$, $v_{\min, 0.1}\equiv v_{\min}/0.1c$, $v_{\max,0.4}\equiv v_{\max}/0.4c$, $\kappa_{10}\equiv\kappa/10$ cm$^2$ g$^{-1}$, $E_{\rm int0,51}\equiv E_{\rm int0}/10^{51}$ erg and $t_{\rm inj,2}\equiv t_{\rm inj}/10^2$ s. For other parameters, we fix the index of the mass density profile $\beta=3.5$ and the parameters of the nuclear heating rate $\dot{\epsilon}_0=2\times10^{10}$ erg s$^{-1}$ g$^{-1}$ and $\alpha=1.3$. We also introduce the normalized time $t_5\equiv t/10^5$ s and $t_6\equiv t/10^6$ s. The values of observed temperature and bolometric luminosity in the thin-diffusion phase are
\begin{eqnarray}\label{app:T_obs,thin}
T_{\rm obs}\sim\left\{ \begin{array}{l}
{\displaystyle 5.63\times10^3~~{\rm K}} \\
~~{\displaystyle \times M_{\rm ej,0.1}^{0.25}v_{\min,0.1}^{0.125}v_{\max,0.4}^{-0.875}t_5^{-0.825}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~(r-{\rm process}) \\
\\
{\displaystyle 6.72\times10^3~~{\rm K}} \\
~~{\displaystyle \times E_{\rm int0,51}^{0.25}t_{\rm inj,2}^{0.25}v_{\min,0.1}^{0.25}v_{\max,0.4}^{-1}t_5^{-1}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=1) \\
\\
{\displaystyle 2.34\times10^3~~{\rm K}} \\
~~{\displaystyle \times E_{\rm int0,51}^{0.25}t_{\rm inj,2}^{0.25}v_{\min,0.1}^{1.25}v_{\max,0.4}^{-2}t_5^{-1}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=2) \\
\\
{\displaystyle 6.77\times10^2~~{\rm K}} \\
~~{\displaystyle \times E_{\rm int0,51}^{0.25}t_{\rm inj,2}^{0.25}v_{\min,0.1}^{2.25}v_{\max,0.4}^{-3}t_5^{-1}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=3), \\
\end{array} \right.
\end{eqnarray}
and
\begin{eqnarray}\label{app:L_bol,thin}
L_{\rm bol}\sim\left\{ \begin{array}{l}
{\displaystyle 3.51\times10^{41}~~{\rm erg~s}^{-1}} \\
~~{\displaystyle \times \kappa_{10}^{-0.5}M_{\rm ej,0.1}^{0.5}v_{\min,0.1}^{0.25}v_{\max,0.4}^{0.25}t_5^{-0.3}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(r-{\rm process}) \\
\\
{\displaystyle 7.10\times10^{41}~~{\rm erg~s}^{-1}} \\
~~{\displaystyle \times \kappa_{10}^{-0.5}M_{\rm ej,0.1}^{-0.5}E_{\rm int0,51}t_{\rm inj,2}v_{\min,0.1}^{0.75}v_{\max,0.4}^{-0.25}t_5^{-1}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=1) \\
\\
{\displaystyle 1.04\times10^{40}~~{\rm erg~s}^{-1}} \\
~~{\displaystyle \times \kappa_{10}^{-0.5}M_{\rm ej,0.1}^{-0.5}E_{\rm int0,51}t_{\rm inj,2}v_{\min,0.1}^{4.75}v_{\max,0.4}^{-4.25}t_5^{-1}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=2) \\
\\
{\displaystyle 7.32\times10^{37}~~{\rm erg~s}^{-1}} \\
~~{\displaystyle \times \kappa_{10}^{-0.5}M_{\rm ej,0.1}^{-0.5}E_{\rm int0,51}t_{\rm inj,2}v_{\min,0.1}^{8.75}v_{\max,0.4}^{-8.25}t_5^{-1}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=3), \\
\end{array} \right.
\end{eqnarray}
respectively.
The transition time from the thin-diffusion phase to the thick-diffusion phase $t_{\times}$ is
\begin{eqnarray}\label{app:t_times}
t_{\times}&=&\sqrt{\frac{2^{\beta-4}(\beta-3)(1-2^{1-\beta})\kappa M_{\rm ej}}{\pi(\beta-1)[1-(v_{\max}/v_{\min})^{3-\beta}]cv_{\max}}\left(\frac{v_{\max}}{v_{\min}}\right)^{3-\beta}} \nonumber \\
\nonumber \\
&\sim&4.53\times10^5~\kappa_{10}^{0.5}M_{\rm ej,0.1}^{0.5}v_{\min,0.1}^{0.25}v_{\max,0.4}^{-0.75}~{\rm s}.
\end{eqnarray}
The values of the observed temperature and bolometric luminosity in the thick-diffusion phase are
\begin{eqnarray}
T_{\rm obs}\sim\left\{ \begin{array}{l}
{\displaystyle 3.89\times10^3~~{\rm K}} \\
~~{\displaystyle \times \kappa_{10}^{-0.583}M_{\rm ej,0.1}^{-0.333}v_{\min,0.1}^{-0.167}t_6^{0.342}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(r-{\rm process}) \\
\\
{\displaystyle 3.86\times10^3~~{\rm K}} \\
~~{\displaystyle \times \kappa_{10}^{-0.667}M_{\rm ej,0.1}^{-0.667}E_{\rm int0,51}^{0.25}t_{\rm inj,2}^{0.25}v_{\min,0.1}^{-0.083}t_6^{0.333}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=1) \\
\\
{\displaystyle 7.73\times10^3~~{\rm K}} \\
~~{\displaystyle \times \kappa_{10}^{-1.333}M_{\rm ej,0.1}^{-1.333}E_{\rm int0,51}^{0.25}t_{\rm inj,2}^{0.25}v_{\min,0.1}^{0.583}t_6^{1.667}} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=2) \\
\\
{\displaystyle 1.29\times10^4~~{\rm K}} \\
~~{\displaystyle \times \kappa_{10}^{-2}M_{\rm ej,0.1}^{-2}E_{\rm int0,51}^{0.25}t_{\rm inj,2}^{0.25}v_{\min,0.1}^{1.25}t_6^3} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=3), \\
\end{array} \right.
\end{eqnarray}
and
\begin{eqnarray}
L_{\rm bol}\sim\left\{ \begin{array}{l}
{\displaystyle 1.16\times10^{41}~~{\rm erg~s}^{-1}} \\
{\displaystyle ~~\times \kappa_{10}^{-0.333}M_{\rm ej,0.1}^{0.667}v_{\min,0.1}^{0.333}t_6^{-0.633}} \\
{\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(r-{\rm process})} \\
\\
{\displaystyle 9.59\times10^{40}~~{\rm erg~s}^{-1}} \\
{\displaystyle ~~\times \kappa_{10}^{-0.667}M_{\rm ej,0.1}^{-0.667}E_{\rm int0,51}t_{\rm inj,2}v_{\min,0.1}^{0.667}t_6^{-0.667}} \\
{\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=1)} \\
\\
{\displaystyle 5.96\times10^{41}~~{\rm erg~s}^{-1}} \\
{\displaystyle ~~\times \kappa_{10}^{-3.333}M_{\rm ej,0.1}^{-3.333}E_{\rm int0,51}t_{\rm inj,2}v_{\min,0.1}^{3.333}t_6^{4.667}} \\
{\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=2)} \\
\\
{\displaystyle 2.62\times10^{42}~~{\rm erg~s}^{-1}} \\
{\displaystyle ~~\times \kappa_{10}^{-6}M_{\rm ej,0.1}^{-6}E_{\rm int0,51}t_{\rm inj,2}v_{\min,0.1}^6t_6^{10}} \\
{\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=3),} \\
\end{array} \right.
\end{eqnarray}
respectively. Note that since the diffusion radius $r_{\rm diff}$ cannot be analytically described in the thin-diffusion phase, we use the approximations $\rho(v,t)\sim\rho(v_{\max},t)$ and $T(v,t)\sim T(v_{\max},t)$ in equations (\ref{app:T_obs,thin}) and (\ref{app:L_bol,thin}). These approximations make discontinuity at the transition time $t_{\times}$. The ratios of the temperature in the thick-diffusion phase to the temperature in the thin-diffusion phase for the $r$-process $A_{\rm T,r}$ and the engine model $A_{\rm T,e}$ at the time $t_{\times}$ are
\begin{eqnarray}
A_{\rm T,r}&=&2^{\beta/4} \nonumber \\
&\sim&1.83
\end{eqnarray}
and
\begin{eqnarray}
A_{\rm T,e}&=&2^{\xi} \nonumber \\
&\sim&\left\{ \begin{array}{ll}
2.00 &~~(\xi=1) \\
4.00 &~~(\xi=2) \\
8.00 &~~(\xi=3), \\
\end{array} \right.
\end{eqnarray}
respectively. The ratios of the luminosity in the thick-diffusion phase to the luminosity in the thin-diffusion phase for the $r$-process model $A_{\rm L,r}$ and the engine model $A_{\rm L,e}$ at the time $t_{\times}$ are
\begin{eqnarray}
A_{\rm L,r}&=&2^{\frac{\beta-4}{2}}\left(\frac{1-2^{3-\beta}}{\beta-3}\right)\sqrt{\frac{\beta-1}{1-2^{1-\beta}}} \nonumber \\
&\sim&0.858
\end{eqnarray}
and
\begin{eqnarray}
A_{\rm L,e}&=&2^{\frac{-\beta-4+8\xi}{2}}\left(\frac{1-2^{3-4\xi}}{4\xi-3}\right)\sqrt{\frac{\beta-1}{1-2^{1-\beta}}} \nonumber \\
&\sim&\left\{ \begin{array}{ll}
1.04 &~~(\xi=1) \\
6.42 &~~(\xi=2) \\
58.8 &~~(\xi=3), \\
\end{array} \right.
\end{eqnarray}
respectively.
The transition time from the thick-diffusion phase to the transparent phase $t_{\rm tr}$ is
\begin{eqnarray}\label{app:t_tr}
t_{\rm tr}&=&\sqrt{\frac{(\beta-3)(1-2^{1-\beta})\kappa M_{\rm ej}}{4\pi(\beta-1)[1-(v_{\max}/v_{\min})^{3-\beta}]cv_{\min}}} \nonumber \\
\nonumber \\
&\sim&7.62\times10^5~\kappa_{10}^{0.5}M_{\rm ej,0.1}^{0.5}v_{\min,0.1}^{-0.5}~{\rm s}.
\end{eqnarray}
We introduce another transition time $t_{\rm tr2}$ when the upper limit of the integral for the luminosity $2r_{\rm diff}$ reaches the inner edge of the ejecta $r_{\rm in}$,
\begin{eqnarray}
t_{\rm tr2}&=&2^{\frac{\beta-2}{2}}t_{\rm tr} \nonumber \\
&\sim&1.28\times10^6~\kappa_{10}^{0.5}M_{\rm ej,0.1}^{0.5}v_{\min,0.1}^{-0.5}~{\rm s}.
\end{eqnarray}
The values of the observed temperature and bolometric luminosity in the transparent phase ($t_{\rm tr}\le t<t_{\rm tr2}$) are
\begin{eqnarray}
T_{\rm obs}\sim\left\{ \begin{array}{l}
{\displaystyle 2.83\times10^3~~{\rm K}} \\
~~{\displaystyle \times M_{\rm ej,0.1}^{0.25}v_{\min,0.1}^{-0.75}t_6^{-0.825}} \\
~~~~~~~~~~~~~~~~~~~~(r-{\rm process}) \\
\\
{\displaystyle 2.69\times10^3~~{\rm K}} \\
~~{\displaystyle \times E_{\rm int0,51}^{0.25}t_{\rm inj,2}^{0.25}v_{\min,0.1}^{-0.75}t_6^{-1}} \\
~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=1) \\
\\
{\displaystyle 3.74\times10^3~~{\rm K}} \\
~~{\displaystyle \times E_{\rm int0,51}^{0.25}t_{\rm inj,2}^{0.25}v_{\min,0.1}^{-0.75}t_6^{-1}} \\
~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=2) \\
\\
{\displaystyle 4.33\times10^3~~{\rm K}} \\
~~{\displaystyle \times E_{\rm int0,51}^{0.25}t_{\rm inj,2}^{0.25}v_{\min,0.1}^{-0.75}t_6^{-1}} \\
~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=3) \\
\end{array} \right.
\end{eqnarray}
and
\begin{eqnarray}
L_{\rm bol}\sim\left\{ \begin{array}{l}
{\displaystyle 3.30\times10^{41}~~{\rm erg~s}^{-1}} \\
~~{\displaystyle \times M_{\rm ej,0.1}t_6^{-1.3}\left[1-\left(\frac{t}{t_{\rm tr2}}\right)^{0.667}\right]} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(r-{\rm process}) \\
\\
{\displaystyle 1.33\times10^{41}~~{\rm erg~s}^{-1}} \\
~~{\displaystyle \times E_{\rm int0,51}t_{\rm inj,2}t_6^{-2}\left[1-\left(\frac{t}{t_{\rm tr2}}\right)^{1.333}\right]} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=1), \\
\\
{\displaystyle 1.00\times10^{41}~~{\rm erg~s}^{-1}} \\
~~{\displaystyle \times E_{\rm int0,51}t_{\rm inj,2}t_6^{-2}\left[1-\left(\frac{t}{t_{\rm tr2}}\right)^{6.667}\right]} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=2), \\
\\
{\displaystyle 1.00\times10^{41}~~{\rm erg~s}^{-1}} \\
~~{\displaystyle \times E_{\rm int0,51}t_{\rm inj,2}t_6^{-2}\left[1-\left(\frac{t}{t_{\rm tr2}}\right)^{12}\right]} \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~({\rm engine},~\xi=3), \\
\end{array} \right.
\end{eqnarray}
respectively.
\newpage
|
\section{Introduction}
The fact that synaptic plasticity can depend on the precise
timing of pre- and postsynaptic spikes \citep{Bi, Rubin},
indicates that time has to be coded somehow in individual neurons.
If the concentration of a certain ion or molecule, which we
will refer to as a {\it trace}, decays in time after a
given event in a regular fashion, then the level of that
trace could serve as a time coder, in the same way as the
concentration of a radioactive isotope can be used to date
a fossil.
A range of models have been proposed in the past that
formulate long-term potentiation (LTP) and
long-term depression (LTD) in terms of traces in the
postsynaptic neurons \citep{Karmarkar, Badoual, Shouval,
Rubin, Graupner, Uramoto}. Several of these models successfully
reproduce a wide range of experimental results; including
pairwise STDP, triplet and even quadruplet nonlinearities.
Most models, however, require fitting of a large number of
parameters individually for each experimental setup and involve
heavily non-linear functions of the trace concentrations.
While possibly realistic in nature, the study of
neural systems modeled under these rules from a dynamical
point of view becomes a highly non-trivial task. At the
other end, the connection between predictions of
simplified models, constructed as phenomenological rules
\citep{Badoual,Froemke}, and the biological underpinnings
is normally hard to establish, as they usually aim to
reproduce only the synaptic change and do away with the
information stored in the the traces themselves.
In the present work we propose a straightforward model
formulating synaptic potentiation and depression in terms
of two interacting traces representing the fraction of
activated N-methyl-D-aspartate (NMDA) receptors and the
concentration of intracellular $Ca^{2+}$ at the postsynaptic
spine, with the intention of bridging these two worlds.
Having a low number of parameters and being composed
of only polynomial differential equations, the model is
able nonetheless to reproduce key features of LTP and LTD.
Moreover, since the parameters of the model are easily
related to the dynamical properties of the system, it
permits to make a connection between the observed synaptic
weight change and the behavior of the underlying traces.
\section{The model}\label{model}
Plasticity in our model will be expressed in terms of two
interacting traces on the postsynaptic site, which we denote
$x$ and $y$, representing the fraction of open-state
NMDA receptors (or NMDARs) and the $Ca^{2+}$
concentration in the dendritic spine of the postsynaptic
neuron, respectively. For a clarification we shortly recall
the overall mechanism of the synaptic transmission process
in a glutamatergic synapse,
as illustrated in Fig.~\ref{mechanism}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.6\textwidth]{NECO-03-14-2089R4-Figure1.eps}
\end{center}
\caption{Illustration of a glutamatergic synapse. Primary
AMPA receptors (AMPAR) are directly activated by
glutamate, voltage gated calcium channels (VGCC)
by the backpropagation action potential. NMDA receptors (NMDAR)
are also activated by glutamate and allow the influx of calcium
if they additionally unblock, which occurs when the
backpropagating action potential removes the blocking $Mg^{2+}$
ion.}
\centerline{\rule{0.8\textwidth}{0.4pt}}
\label{mechanism}
\end{figure}
A presynaptic spike results in the release of glutamate
molecules across the synaptic cleft, which will activate
a series of receptors on the postsynaptic spine, including
the already mentioned NMDA receptors, and
$\alpha$-amino-3-hydroxy-5-methyl-4-isoxazolepropionic
acid (AMPA) receptors (or AMPAR)\citep{Meldrum}.
$Na^{+}$ ions will then flow through the AMPAR channels
into the dendritic spine of the postsynaptic
cell, triggering a cascade of events which may eventually
lead to the activation of an axonal spike at the soma of the
postsynaptic cell, and of an action potential backpropagating
down the dendritic tree. This action potential has two
effects captured within our model: the first is
the activation of voltage-gated $Ca^{2+}$ channels (VGCC),
allowing an influx of $Ca^{2+}$ ions, resulting hence in
an increase of the $Ca^{2+}$ concentration $y$; the second
is the unblocking of NMDAR channels, as we detail in what
follows.
$Ca^{2+}$ ions may flow into the postsynaptic spine also
through the NMDAR channels \citep{Meldrum}, but for this
to happen two conditions need to be fulfilled. NMDARs
are activated when glutamate binds to them, which
in turn opens the receptor's $Ca^{2+}$ permeable channel.
The channels are said to be open when the protein's
conformational state permits ions to flow through
them, and closed otherwise. At resting membrane
potential, however, $Mg^{2+}$ ions are present in the
channel's pore, blocking the channel and preventing
$Ca^{2+}$ ions from permeating the neuron \citep{Mayer}.
This block is temporarily removed by a back-propagating
action potential. For $Ca^{2+}$ to flow into the
postsynaptic spine two conditions need hence to be fulfilled.
The presence of glutamate in the synaptic cleft, triggered by
a presynaptic spike, and a back-propagating action
potential, signaling a postsynaptic spike. The NMDA
receptors are hence the primary agents, within our model,
for the interaction of pre- and postsynaptic neural
activities in terms of axonal spikes. They are hence
also the primary agents for causality within the
STDP rule.
\subsection{Trace dynamics}\label{trace_dynamics}
We denote with $\{t_{pre}^\sigma\}$ and $\{t_{post}^\sigma\}$
the trains of pre- and postsynaptic spikes, respectively.
The update rules for the fraction $x$ of open but blocked
NMDA receptors and the concentration $y$ of postsynaptic
$Ca^{2+}$ ions are then given by
\begin{equation}
\left\{
\begin{array}{lcl}
\dot x = - \frac{x}{\tau_x} + \mathit{E_x}(x)
\sum_\sigma\delta(t-t_{pre}^\sigma) \\
\dot y = - \frac{y}{\tau_y} + (x+y_c)
\mathit{E_y}(y)
\sum_\sigma\delta (t-t_{post}^\sigma)
\end{array}
\right.
\label{dot_xy}
\end{equation}
where $\tau_x$ and $\tau_y$ represent the time constants
for the decay of $x$ and $y$ respectively. In absence
of presynaptic spikes, glutamate in the synaptic cleft
is cleared by passive diffusion and glutamate
transporters \citep{Huang}. $Ca^{2+}$ concentration in
the postsynaptic site will decay, in turn, in absence
of postsynaptic spikes \citep{Carafoli}. In our model,
each incoming presynaptic spike produces an increase
in the number $x$ of open NMDA channels due to glutamate
release, and the $Ca^{2+}$ concentration $y$ increases
only when a postsynaptic spike
is present, viz when a backpropagating action potential
reaches the postsynaptic spine. Calcium increase in
(\ref{dot_xy}) is composed of two terms; a constant value
$y_c$, representing the contribution of VGCCs, and a term
proportional to the fraction of open NMDA receptors. In
this simplified approach, every NMDAR channel still open
from the presynaptic spike, is then unblocked by the
backpropagating action potential. Therefore the transient
calcium current through NMDA receptors is modeled as
proportional to $x$.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.9\textwidth]{NECO-03-14-2089R4-Figure2.eps}
\end{center}
\caption{\textbf{a)} Plot of the limiting factor $\mathit{E}_x$, as
defined by (\ref{theta}), as a function of the trace concentration,
here exemplified by $x$ (the same applies to $y$). \textbf{b)} Plot
of the LTP threshold $(y-b)\theta(y-b)$ present in (\ref{dot_w}).}
\centerline{\rule{0.8\textwidth}{0.4pt}}
\label{limiting_factors}
\end{figure}
The efficacy factors $\mathit{E_x}$ and $\mathit{E_y}$ included
in (\ref{dot_xy}) are defined as:
\begin{equation}
\mathit{E_z}(z) = \theta(z_b-z)\, \left(1-\frac{z}{z_b}\right), \qquad
\theta(z) = \left\{
\begin{array}{lcl}
0 \qquad z \leq 0 \\
1 \qquad z > 0
\end{array}
\right.
\label{theta}
\end{equation}
where $z$ is either $x$ or $y$, and determine the efficacy
of spikes in increasing trace concentrations. For trace
levels above the respective
reference values $x_b$ and $y_b$ no further increase is
possible (see Fig.~\ref{limiting_factors}\,\textbf{a)} and
the trace concentration can only decay exponentially. This
determines a refractory period, as shown
in Fig.~\ref{refractory}. The duration of this period
is in this case a function of the decay constant of the
trace and the magnitude of the overshoot above the
reference value. Below this level, $\mathit{E}$ will
tend asymptotically to unity as the trace concentration decays.
In this way, previous spikes decrease the efficacy of
future spikes. Similar mechanisms of reduced
spike efficacy have been proposed in the past in models
of STDP \citep{Froemke}.
Two forces therefore compete to drive nonlinear plasticity
in our model: trace accumulation and spike suppression, the
latter formulated in the present effective model via a
saturation term.
The update rules
(\ref{dot_xy}) for the traces are reduced, in the
sense that all superfluous constants have been rescaled
away, as discussed further in the \emph{Appendix}.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.7\textwidth]{NECO-03-14-2089R4-Figure3.eps}
\end{center}
\caption{Illustration of the effect of the limiting factor
$\mathit{E}_x$ (dashed red line, upper panel),
compare Fig.~\ref{limiting_factors} and Eq.~(\ref{theta}),
on the trace dynamics (solid blue line, lower panel, compare
Eq.~(\ref{dot_xy})). Here for the $x$-trace (the behavior
is qualitatively the same for $y$). Two spikes $t^1_{pre}$
and $t^2_{pre}$ are present and indicated as
solid green vertical bars, with the height being proportional to $E_x$.
The system ignores further incoming spike whenever $x>x_b$,
resulting in respective refractory periods (shaded grey areas).
For finite values of $x<x_b$ the efficacy of incoming spikes
is reduced proportionally to $x_b-x$.}
\centerline{\rule{0.8\textwidth}{0.4pt}}
\label{refractory}
\end{figure}
\subsection{Update rules for the synaptic weight}
We now formulate the updating rules for the synaptic
weight, or synaptic strength, in terms of the
trace concentrations. To this end we consider the
contribution of two pathways mediated by distinct enzymes
\citep{Colbran}; which, for simplicity we will denote as
LTP and LTD pathways. Calcium is involved in both the LTP
and the LTD pathways \citep{Cormier, Neveu, Yang}, with
high levels of calcium resulting in LTP and moderate and
low levels resulting in LTD. We propose the following
rule for the plasticity of the synaptic weight $w$,
\begin{equation}
\dot w = \alpha x (y-b) \theta(y-b) \sum_\sigma\delta(t-t_{post}^\sigma) -
\beta xy \sum_\sigma\delta(t-t_{pre}^\sigma),
\label{dot_w}
\end{equation}
with $\theta$ being the same previously defined step function
which, in this case, serves as a lower bound. The first term in
(\ref{dot_w}) leads to an increase of the synaptic weight;
it is triggered in the presence of a postsynaptic spike
and by the calcium concentration $y$, but only if $y$
is larger than a given threshold $b$,
see Fig.~\ref{limiting_factors}\,\textbf{b)}. A threshold
in the calcium concentration $Ca^{2+}$ for LTP has been
experimentally observed \citep{Cormier} and its
dependence with the previous synaptic activity has been
studied \citep{Huang_2}. In the present work we will consider
a constant $b = y_c$, and we will show in the next section
that the standard STDP curve is obtained with this choice.
The second term in (\ref{dot_w}), in turn, leads to
decrease of the synaptic weight and needs a finite level
for both the calcium concentration $y$ and for the fraction
of open NMDA receptors $x$ (which can be taken as a measure
of the glutamate concentration in the synaptic cleft),
in addition to the presence of a presynaptic spike,
which acts as a second coincidence detector as proposed by
\cite{Karmarkar}. The parameters $\alpha>0$ and $\beta>0$
represent the relative strengths of these two contributions.
The pre- and postsynaptic spikes $t_{pre/post}^\sigma$
mark the timing of the synaptic update in our model
(\ref{dot_w}) for the synaptic plasticity. Here we
considered $\delta$-like spikes and one needs, for
numerical simulations using small but finite time steps,
to update the traces via (\ref{dot_xy}) before updating
the weights via (\ref{dot_w}).
\subsection{The pairwise STDP rule}\label{section_STDP}
In the limit of low frequencies, the traces decay to
zero in between the occurrence of two pairs of spikes,
which may hence be considered as isolated.
We denote with $\Delta t$ the time between the pre- and
the postsynaptic spike, with a positive value corresponding to
a causal pre-post order and a negative $\Delta t$ to an
anti-causal post-pre ordering. For an isolated pair of spikes
one can easily integrate (\ref{dot_xy}) and (\ref{dot_w}),
obtaining:
\begin{equation}
\Delta w =
\left\{
\begin{array}{lrcl}
\alpha \mathrm{e}^{-|\Delta t|/\tau_x}
\left(\mathrm{e}^{-|\Delta t|/\tau_x} + y_c - b
\right)\qquad &\Delta t &>& 0\\
-\beta y_c \mathrm{e}^{-|\Delta t|/\tau_y} \qquad &\Delta t &<& 0
\end{array}
\right.
\label{isolated_pair}
\end{equation}
The synaptic weight is always depressed for an anti-causal
time ordering of the spikes with $\Delta t<0$, and potentiated
for $y_c\geq b$ and a causal time ordering corresponding to
$\Delta t>0$. The LTP term becomes a simple exponential
decay for $b=y_c$. We have chosen in our model a fully
decoupled formulation for LTP and LTD. While the LTD term is
always negative, the restriction on the LTP term to be always
positive could be relaxed by removing the step function in
equation (\ref{dot_w}). Then, with the choice $b>y_c$, a
depression window would arise after the peak of potentiation.
This window has indeed been observed in the past in CA1 cells
from rat hippocampal slices \citep{Nishiyama}. By setting
$b<y_c$, on the other hand, the decay would be composed of
two exponentials. In the LTD term we have not included a
threshold. Alternatively, one could replace the calcium level
$y$ by an expression $(y-b_{LTD})\theta(y-b_{LTD})$, analogous
to the LTP term, which is identical to the case we present for
$b_{LTD}=0$ since $y$ is always positive. It is however worth
discussing the cases where $b_{LTD}\neq0$. Apart from the
step function $\theta$, the LTD threshold represents only a
vertical shift of the negative portion of the STDP window by
a factor $\beta b_{LTD}$. If $b_{LTD}<0$ the plot is shifted
downwards, which means depression occurs even for isolated
presynaptic spikes ($\Delta t\rightarrow -\infty$). This is
usually not the case, as seen in \cite{Bi2} and \cite{Froemke}.
If $b_{LTD}<0$, on the other hand, the plot is shifted upwards
but, because of the step function, the $LTD$ term is always
negative and then the tail of the exponential would be cut-off.
By looking at the experimental results in
Figs.~\ref{STDP_theory-vs-experiment_HC} and
\ref{STDP_theory-vs-experiment_CX}, one observes that the
data seems in fact quite noisy to determine the exact shape
of the decay functions. In absence of further detail,
we have chosen in the present work to keep $b=y_c$ and no
threshold (or a threshold at $0$) for LTD, therefore
respecting the exponential fits proposed in the original
papers \citep{Bi3,Froemke}.
Rewriting the constants $\alpha$, $\beta$, $\tau_x$, and
$\tau_y$ as $\alpha = A^+$, $\beta = A^-/y_c$,
$\tau_x =2\tau_+$, and $\tau_y =\tau_-$ where
$A^+$, $A^-$, $\tau_+$ and $\tau_-$ represent the maximal
intensities and timescales of LTP and LTD for isolated
spike pairs, we obtain with
\begin{equation}
\Delta w =
\left\{
\begin{array}{lrcl}
+A^+\mathrm{e}^{-|\Delta t|/\tau_+} \qquad &\Delta t&>&0\\
-A^-\mathrm{e}^{-|\Delta t|/\tau_-} \qquad &\Delta t&<&0
\end{array}
\right.
\label{Delta_w_isolated_pair}
\end{equation}
the classical fit for pairwise STDP proposed both in hippocampal
and cortical neurons \citep{Bi3, Froemke}. This result is independent
of $y_c$, $x_b$, and $y_b$ and these three parameters can be hence
be used to reproduce additional experimental observations. In what
follows we will use the amplitudes $A^{\pm}$ as primary parameters,
instead of $\alpha$ and $\beta$ and rewrite the plasticity
rule (\ref{dot_w}) as
\begin{equation}
\dot w = A^+ x (y-y_c)\theta(y-y_c) \sum_\sigma\delta(t-t_{post}^\sigma)
- \frac{A^-}{y_c} xy \sum_\sigma\delta(t-t_{pre}^\sigma).
\label{dot_w_2}
\end{equation}
This is the final shape of the equation for the evolution
of the synaptic strength that we will use throughout this work,
it allows to interpret the results for a variety
of spike pattern situations in terms of the known
spike-pair STDP parameters. The representations
(\ref{dot_w_2}) and (\ref{dot_w}) are, in any case,
equivalent.
\subsection{Spike triplets}\label{sec_triplets}
The effect of a pair of pre- and postsynaptic spikes has
been experimentally shown to depend, in a non-linear
fashion, not only on its inter-spike interval but also
on the presence of additional spikes temporally proximal to
the pair. The contribution of spike triplets, the simplest
case of spike-pair interactions, cannot be described as a
linear sum of two individual contributions of spike-pairs
\citep{Froemke, Wang}.
In the following sections, we will study the model's results
for either two pre- and one postsynaptic
spikes in a pre-post-pre order, or one pre- and two
postsynaptic spikes in a post-pre-post ordering. For
example, with 15Post5 we denote a pre-post-pre ordering,
\begin{equation}
15\mathrm{Post}5, \qquad\quad
\left\{t_{pre}^\sigma\right\} = \{-15,5\},\qquad\quad
\left\{t_{post}^\sigma\right\} = \{0\}
\label{eq:15Post5}
\end{equation}
and with 10Pre20 a post-pre-post ordering,
\begin{equation}
10\mathrm{Pre}20, \qquad\quad
\left\{t_{pre}^\sigma\right\} = \{0\},\qquad\quad
\left\{t_{post}^\sigma\right\} = \{-10,20\}~,
\label{eq:10Pre20}
\end{equation}
where the times $t_{pre/post}^\sigma$ of
the spikes are given in milliseconds.
As for spike pairs, the weight-change induced by
low-frequency triplets can be computed analytically,
obtaining
\begin{equation}
\begin{array}{rcl}
\Delta w = &+&A^+ exp\left(-\frac{\rvert\Delta t_1\rvert}{\tau_+}\right)\\
&- &A^- exp\left(-\frac{\rvert\Delta t_2\rvert}{\tau_-}\right)
\left[1+\frac{exp\left(-\frac{\rvert\Delta t_1\rvert}{\tau_x}\right)}{y_c}\right]
\left[1 + exp\left(-\frac{\rvert\Delta t_1\rvert+\rvert\Delta t_2\rvert}{\tau_x}\right)\left(1 - \frac{1}{x_b}\right)\right]
\end{array}
\label{analytic_PrePostPre}
\end{equation}
for pre-post-pre triplets, and
\begin{equation}
\begin{array}{rcl}
\Delta w = &-&A^- exp\left(-\frac{\rvert\Delta t_1\rvert}{\tau_-}\right)\\
&+& A^+ exp\left(-\frac{\rvert\Delta t_2\rvert}{\tau_+}\right)
\left[1 + y_c exp\left(-\frac{\rvert\Delta t_1\rvert+\rvert\Delta t_2\rvert}{\tau_y}+\frac{\rvert\Delta t_2\rvert}{\tau_x}\right)
\left(1 - \frac{exp\left(-\rvert\Delta t_2\rvert/\tau_x\right)+y_c}{y_b}\right)\right]
\end{array}
\label{analytic_PostPrePost}
\end{equation}
for post-pre-post triplets, where we have assumed that
the traces are below their respective reference levels,
$x_b$ and $y_b$ respetively, by the time a second spike
arrives (the case of the second spike arriving while the
trace is above the reference level is discussed later in
this section). The saturation effect can reduce the effect of
a new spike to zero but not reverse the sign, as seen
in expression (\ref{theta}).
We see that the first term in Eqs.~(\ref{analytic_PrePostPre})
and (\ref{analytic_PostPrePost}), corresponding to the first pair,
remains in both cases unchanged, by construction, with
non-linearities appearing in the second contribution. In the second
term of Eq.~(\ref{analytic_PrePostPre}), we find a first
factor (the first parenthesis) corresponding to a correction
produced by the interaction between the two traces (the
calcium inflow through NMDAR channels), and a second factor
corresponding to the balance between trace accumulation and
spike suppression. In Eq.~(\ref{analytic_PostPrePost})
we also find a term balancing trace accumulation and spike
suppression. The multiplicative factor
$exp(\rvert\Delta t_2\rvert/\tau_x)$ inside the brackets
comes from the way we have decided to factorize the equation,
since $\tau_x = 2\tau_+$.
If the third spike would instead come within the respective
refractory period (see Fig.~\ref{refractory}),
the expressions (\ref{analytic_PrePostPre}) and
(\ref{analytic_PostPrePost}) would reduce to
\begin{equation}
\begin{array}{rcl}
\Delta w = &+&A^+ exp\left(-\frac{\rvert\Delta t_1\rvert}{\tau_+}\right)\\
&- &A^- exp\left(-\frac{\rvert\Delta t_2\rvert}{\tau_-}\right)
\left(1+\frac{exp\left(-\frac{\rvert\Delta t_1\rvert}{\tau_x}\right)}{y_c}\right)
exp\left(-\frac{\rvert\Delta t_1\rvert+\rvert\Delta t_2\rvert}{\tau_x}\right)
\end{array}
\label{analytic_PrePostPre_2}
\end{equation}
for pre-post-pre triplets, and:
\begin{equation}
\Delta w = -A^- exp\left(-\frac{\rvert\Delta t_1\rvert}{\tau_-}\right)
\label{analytic_PostPrePost_2}
\end{equation}
for post-pre-post triplets, where in Eq.~(\ref{analytic_PostPrePost_2})
the second pair is directly inhibited by the LTP threshold.
While this last situation is not encountered for the low
frequency triplet configurations presented in this work,
it becomes relevant in high frequency scenarios. This
condition could be relaxed by replacing
the strict threshold by a smooth sigmoidal.
\subsection{Interpretation of the variables and parameters in the model}\label{parameters}
The here proposed model contains a relative small number of
variables and parameters and can be considered an effective
approach with the biological underpinnings of STDP being
governed by a substantially larger number of variables and
parameters whose functional interdependences
are naturally far more complex than the polynomial
descriptions here proposed. Any effective model will
however pool together within each effective variable or
parameter several effects which might depend on a variety
of different factors in the biological neuron.
In section \ref{trace_dynamics}, we defined $x$ as the
fraction of open but unblocked NMDAR channels. When
paired with a postsynaptic spike, and under the
simplifications assumed in the model, the value of
$x$ can be then associated with a transient calcium
current and a comparison with experimental results
of the parameters related to $x$ would reflect this
role. The time window for LTP, for instance, results
in our model from the value of $\tau_x = 2\tau_+$ (as we showed in
section \ref{section_STDP}). $\tau_x$ can then be
interpreted in this context as the decay time of the
transient calcium current. It has been argued by \cite{Hao}
that the narrow window for LTP results from AMPA-EPSP in
the postsynaptic spine. In fact, as reviewed in the
same article, the whole spine seems to work as an electrical
amplifier, locally prolonging the depolarization time at
the spine. It is therefore not surprising to find different
values of the time constants in different neurons or even
within different synapses within the same neuron. In our
model we do not compute AMPA currents directly and reduce the
overall effect of the spine to the effective value of
$\tau_x$. Similarly, $\tau_y$ represents the timescale for
decay of the effective calcium concentration at the spine.
We have included in this work saturation terms for both
variables $x$ and $y$. As it has been proposed in the past
\citep{Froemke}, triplet nonlinearities in visual cortical
neurons indicate strong suppression effects on future
spikes by previous spikes of the train. The saturation
terms included in the model provide one possible
effective way of dealing with spike suppression, reducing
a biological complex phenomena further down the cascade of
processes, leading eventually to LTP and to LTD respectively.
\section{Results for the Hippocampus}\label{Results_Hippocampus}
Our model, as defined by (\ref{dot_xy}) and
(\ref{dot_w_2}) contains overall seven adjustable
parameters. Four of these parameters, namely
$A^+=\alpha$, $A^-=\beta/y_c$, $\tau_+$, and $\tau_-$,
enter explicitly the isolated spike-pair STDP rule
(\ref{Delta_w_isolated_pair}) and are determined directly
by experiment. For cultured rat hippocampal neurons
\begin{equation}
A^+= 0.86/60,\qquad A^- = 0.25/60,\qquad \tau_+=19\,\mathrm{ms}, \qquad
\tau_-=34\mathrm\,{ms}
\label{parameters_A_tau}
\end{equation}
have been measured \citep{Bi3} and we will use these
experimental values throughout the hippocampus part of
this study. In Fig.~\ref{STDP_theory-vs-experiment_HC} we
present, as an illustration, both the experimental and the
theory results, with the latter reproducing, by construction,
the experimental fit. For the model simulation, the experimental
protocol of 60 repetitions spaced by one second has been used.
However, the $1\mathrm{Hz}$ frequency of spike
pairs is so low that (\ref{Delta_w_isolated_pair}) could be
directly used without any discernible difference.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.6\textwidth]{NECO-03-14-2089R4-Figure4.eps}
\end{center}
\caption{Weight change after a train of 60 pairs at a constant
frequency of $1\,\mathrm{Hz}$ as a function of the time delay
$\Delta t$ between pre- and postsynaptic spikes. The red open
circles are the experimental data for hippocampal neurons
\citep{Bi2}. The continuous blue line represents the model's
results when the parameters are set to $A^+= 0.86/60$,
$A^- = 0.25/60$, $\tau_+=19\mathrm{ms}$, $\tau_-=34\mathrm{ms}$,
which correspond to a fit of the experimental
data, as presented in \cite{Bi3}.
}
\centerline{\rule{0.8\textwidth}{0.4pt}}
\label{STDP_theory-vs-experiment_HC}
\end{figure}
Three parameters entering (\ref{dot_xy}) and (\ref{dot_w_2}),
namely $y_c$, $x_b$, and $y_b$ are to be
selected. In a continuous time evolution scenario, $x_b$
and $y_b$ determine strict maximal concentrations for
the traces. In the discrete time scenario, overshoots are
however possible, due to the finite increase in the traces
after every spike. In this context,
and in a low frequency situation, the first spike in the
stimulation pattern is unaffected by the limiting factor,
and only the efficacy of the following spike is reduced.
Since $x_b$ and $y_b$ then do not affect pairwise
STDP, they need to be selected from higher order
contributions to the weight-change. In
this case, we selected the values of $y_c$, $x_b$,
and $y_b$ from triplet results, as presented in
what follows.
In Fig.~\ref{triplets_theory_experiment_HC} we now
compare our results for triplets, as described in section
\ref{sec_triplets}, with experiments for cultured rat
hippocampal neurons \citep{Wang}. The triplet stimulation
experimental protocol consists of a regular train of 60
triplets with a repetition frequency of $1\,\mathrm{Hz}$
and we use the identical protocol for the theory simulations.
We also keep the pairwise STDP parameters (\ref{parameters_A_tau})
valid for cultured rat hippocampal neurons and adjust
the remaining three free parameters $y_c$, $x_b$, and $y_b$
by minimizing the standard deviation (SD) between the
numerical and the experimental results, obtaining
$y_c = 0.28$, $y_b = 0.66$, and $x_b = 0.62$
(with an SD of $6.76$).
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.6\textwidth]{NECO-03-14-2089R4-Figure5.eps}
\end{center}
\caption{Synaptic strength change in hippocampal neurons
induced by triplets composed of either two pre- and one postsynaptic
spike, left side of the diagram, compare (\ref{eq:15Post5}),
or one pre- and two postsynaptic spikes, right side of the
diagram, see (\ref{eq:10Pre20}). A total
of 60 triplets are presented, with a repetition frequency
of $1\,\mathrm{Hz}$. Full blue boxes correspond to the model's
results, empty red bars to experimental data \citep{Wang},
and the green lines
represent the linear addition of the PostPre and PrePost
pairs each triplet contains via (\ref{dot_w_2}).
Simulation parameters:
$A^+= 0.86/60$, $A^- = 0.25/60$, $\tau_+=19\mathrm{ms}$,
$\tau_-=34\mathrm{ms}$, $y_c = 0.28$, $x_b = 0.62$, and
$y_b = 0.66$.}
\centerline{\rule{0.8\textwidth}{0.4pt}}
\label{triplets_theory_experiment_HC}
\end{figure}
We found that the SD varies smoothly, and relatively weakly,
with the exact choice of the three free parameters, as can be
expected from the analytical expressions, and that
this freedom can be used to obtain a range of
functional dependencies of the synaptic plasticity
upon spiking frequencies, as discussed in Sect.~\ref{frequency}.
We have also included in Fig.~\ref{triplets_theory_experiment_HC}
the expected synaptic weight changes for the case of
a linear superposition of the two respective interspike
contributions via (\ref{dot_w_2}). One observes that the
discrepancy between the non-linear and the linear interactions
is much stronger for pre-post-pre than for post-pre-post
triplets. With the former leading to an overall reduced
synaptic weight change and the later configuration to
a substantial potentiation. It is interesting to observe here
that spike suppression, as proposed in \cite{Froemke} from
cortical neurons cannot explain nonlinearities in hippocampus.
Suppression of the second presynaptic spike in the triplet
would reduce depression and the overall result would be
supralinear potentiation, contrary to the experimental
observation. Trace accumulation is the dominant effect
driving nonlinearities in hippocampal neurons.
\section{Results for the Cortex}\label{Results_Cortex}
We repeat now the procedure presented previously for
the hippocampus, comparing the results of the proposed
plasticity rule to experimental data obtained from slices
of the visual cortex. As in the previous section,
the values of $A^+=\alpha$, $A^-=\beta/y_c$, $\tau_+$, and
$\tau_-$ are determined directly by experiment.
We use
\begin{equation}
A^+= 1.03/60,\qquad A^- = 0.51/60,\qquad \tau_+=13.3\,\mathrm{ms}, \qquad
\tau_-=34.5\mathrm\,{ms}
\label{parameters_A_tau_CX}
\end{equation}
as obtained by \cite{Froemke} for pyramidal
neurons in layer 2/3 (L2/3) of rat visual cortical slices.
Both experiment and the STDP curve
are shown in Fig.~\ref{STDP_theory-vs-experiment_CX},
where we have reproduced, for the simulation, the experimental
protocol, using 60 repetitions at $0.2\,\mathrm{Hz}$.
Once again, the frequency of spike pairs is so low that
(\ref{Delta_w_isolated_pair}) could be
directly used without any discernible difference.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.6\textwidth]{NECO-03-14-2089R4-Figure6.eps}
\end{center}
\caption{As in Fig.~\ref{STDP_theory-vs-experiment_HC}, now
for visual cortical neurons. The stimulation frequency is now
$0.2\,\mathrm{Hz}$, as in the experiment \citep{Froemke}.
The red open circles are the experimental data (courtesy of
Robert C. Froemke and Yang Dan) and the continuous blue
line represents the model's results when the parameters are
set to $A^+= 1.03/60$, $A^- = 0.51/60$, $\tau_+=13.3\,\mathrm{ms}$,
$\tau_-=34.5\mathrm\,{ms}$, corresponding to the fit of the
experimental data presented in \cite{Froemke}.
}
\centerline{\rule{0.8\textwidth}{0.4pt}}
\label{STDP_theory-vs-experiment_CX}
\end{figure}
To select $y_c$, $x_b$, and $y_b$ we once again
resort to triplet results. In \cite{Froemke}, the change
produced by triplets of either two pre- and one postsynaptic
spikes or one pre- and two postsynaptic spikes was also
measured. The data consist in this case however
of a large set of specific triplet timing configurations,
with every individual triplet configuration measured
once. We decided to treat all measurements on an
equal footing, fitting the complete set by minimizing
the mean square error without introducing
any further bias.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.6\textwidth]{NECO-03-14-2089R4-Figure7.eps}
\end{center}
\caption{As in Fig.~\ref{triplets_theory_experiment_HC}, now
for visual cortical neurons. The stimulation frequency is now
$0.2\,\mathrm{Hz}$ as in the experiment \citep{Froemke}.
Full blue boxes correspond to the model's results for the best fit
of the parameters, empty red bars to experimental data,
and the green lines represent
the linear addition of the two PostPre and PrePost pairs each
triplet contains. With diamonds, the model's results for
an alternative set of parameters is presented. While the
quantitative differences are larger with this parameter choice,
the model still qualitatively reproduces cortical triplet
nonlinearities. Simulation parameters:
$A^+= 1.03/60$, $A^- = 0.51/60$, $\tau_+=13.3\,\mathrm{ms}$,
$\tau_-=34.5\mathrm\,{ms}$, Best fit: $y_c = 11.6$, $y_b = 10.9$, and
$x_b = 0.5$. Diamond points: $y_c = 1.0$, $y_b = 0.9$, and
$x_b = 0.4$.Experimental data courtesy of Robert C.
Froemke and Yang Dan.}
\centerline{\rule{0.8\textwidth}{0.4pt}}
\label{triplets_theory_experiment_CX}
\end{figure}
We obtain in this case $y_c = 11.6$, $y_b = 10.9$, and
$x_b = 0.5$. The obtained SD of $37.4$ is,
in this case, much larger than the one found for hippocampus,
though that is partly due to the variance in the
experimental data themselves, corresponding to individual
data points and not to averaged results. Another consequence
of the large variance in the data is that the minimum in the
SD is relatively broad. We will discuss these points in detail
in what follows.
In order to compare the results for cortical neurons
with the previous section on hippocampal neurons, as
presented in In Fig.~\ref{triplets_theory_experiment_HC},
we have performed a smooth interpolation of the set of individual
experimental results for cortical triplets by means of
gaussian filters. In Fig.~\ref{triplets_theory_experiment_CX}
we compare the theory results with the interpolated
experimental data.
Contrary to hippocampal triplet results presented
in Fig.~\ref{triplets_theory_experiment_HC}, experiments in
cortical slices show that post-pre-post triplets lead
to strong depression and pre-post-pre triplets to
potentiation. Post-pre-post triplets deviate, in addition,
somewhat more from a linear superposition of the contribution
of the two inherent spike pairs than the pre-post-pre
configuration. While the predictions of the model presented
in Fig.~\ref{triplets_theory_experiment_CX}) are clearly
not as good as the ones obtained for hippocampal culture,
they are still qualiatively in agreement with the experimental
results, successfully capturing the asymmetry between
post-pre-post and pre-post-pre triplets. While, there is
still room for improvement in this regard, we
believe it is important that the model can switch from the
hippocampal to the cortical regime in terms of triplet
nonlinearities.
As we previously mentioned, the data itself has a much larger
variance in this case. To have an idea of of the variability
of the data, we computed the standard deviation of the data
to the smooth gaussian interpolation of width 5ms that we
used for the visual comparison of
Fig.~\ref{triplets_theory_experiment_CX}, which
yields an SD of $32.5$ (as compared to the SD of $37.4$ between
model and experiment). For this reason, we believe that a
reasonable goal in this case is to reproduce the distinct
qualitative feature of the triplet nonlinearities, more than
an accurate quantitative approximation.
The optimal value of $y_c=11.6$ obtained when fitting the
experimental triplet results, see
Fig.~\ref{triplets_theory_experiment_CX}, seems to be
too large, in particular when compared to the one obtained
for hippocampal neurons. This result can be traced back to
the occurrence of a broad minimum for the least-square
fit together with a relative high variability
of the experimental data. We have hence also examined
parameter configurations with lower values for $y_c$. Also
included in Fig.~\ref{triplets_theory_experiment_CX} is an
example with $y_c = 1.0$, also representing the observed
experimental features qualitatively. We find that the
particular cortical structure of triplets arises
from strong saturation, being a consequence of $y_b < y_c$.
\section{Frequency dependence}\label{frequency}
So far, we have considered only pairs or triplets of
pre- and postsynaptic spikes coming at low frequencies and
with very precise timings. This will not necessarily be the case
in a natural train of spikes. It is therefore interesting to
examine the model's prediction for spike trains with
different degrees of correlation between pre- and postsynaptic
spikes. A neuron usually receives input from about ten thousand
other neurons. While the correlation of the postsynaptic neuron
will be higher for a strong synapse driving the neuron, the
postsynaptic neuron will in general not be correlated with all
of its inputs. We therefore study both types of
connections.
We begin in section \ref{uncorrelated} by studying
the case of uncorrelated trains of pre- and postsynaptic spikes
and then analyze in section \ref{correlated} the case of a driving
synapse with different degrees of correlation. In these sections
we numerically evaluate the synaptic strength change as a
function of the pre- and postsynaptic neuronal firing rates.
\subsection{Plasticity induced by uncorrelated spikes}\label{uncorrelated}
We begin by evaluating the synaptic change produced by
uncorrelated trains of Poisson pre- and postsynaptic
spikes. In these simulations we use the same parameters as
fitted from pairwise and triplet experiments in Hippocampus
and Cortex, refering to hippocampal and cortical
neurons respectively.
The results of the simulations for hippocampal neurons
are presented in Fig.~\ref{poisson_spikes_HC}. We
present two kinds of plots in the diagram: a plot where
the pre- and postsynaptic firing rates are equal, and
plots of constant presynaptic frequency for varying
postsynaptic firing rate. We observe in this last type,
that the sign of the weight changes, as a function of
the postsynaptic activity for a constant presynaptic
frequency, generically switching from negative
to positive at a certain threshold $\theta_H$. This
threshold increases with rising presynaptic frequency,
resulting in a sliding threshold. In other rate-based
learning rules like BCM \citep{Bienenstock}, similar
thresholds for potentiation are determined by appropriate
long term averages of the postsynaptic activity. In our
model, $\theta_H$ is set by the level of the presynaptic
activity, as measured on timescales of the respective
traces. This feature would allow the neuron to adjust the
threshold of each synapse independently, setting in each
case the level of what constitutes a significant activity.
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.0\textwidth]{NECO-03-14-2089R4-Figure8.eps}
\end{center}
\caption{Average weight change, for hippocampal neurons,
produced by one-second trains of uncorrelated
Poisson-distributed pre- and postsynaptic spikes,
as a function of the postsynaptic firing frequency
$f_{post}$ and for various constant presynaptic firing
rates $f_{pre}$ (full lines).
Also included is the case for identical pre- and postsynaptic
firing rates (dashed line). The pair-STDP values
(\ref{parameters_A_tau}) have been used and two sets
of values for the remaining three free parameters,
yielding both otherwise very similar results for the spike
triplets.
\textbf{a)} $y_c = 0.28$, $y_b = 0.66$, and $x_b = 0.62$.
\textbf{b)} $y_c = 0.8$, $y_b = 1.34$, and $x_b = 1.82$.
}
\centerline{\rule{0.8\textwidth}{0.4pt}}
\label{poisson_spikes_HC}
\end{figure}
The overall synaptic change becomes Hebbian for large pre-
and post- firing rates $f_{pre}$ and $f_{post}$, in
the sense that it is then proportional to the product
$f_{pre}\cdot (f_{post}-\theta_H)$. This weight change is
influenced in a substantial way by the value of $y_c$ and
we have presented in Fig.~\ref{poisson_spikes_HC} two sets
of parameters, one with $y_c=0.28$ (left panel) and one
with $y_c=0.8$ (right panel), yielding otherwise
similar SDs when fitting the experimental
triplet data ($6.76$ and $7.37$ respectively).
Potentiation dominates for larger values of $y_c$,
as seen in Fig.~\ref{poisson_spikes_HC}\,\textbf{b)}.
These results seem, at first sight, counterintuitive
given the role of $y_c$ as a threshold for LTP.
Note however, that $y_c$ contributes to the increase
in $y$ through (\ref{dot_xy}) and both LTP and LTD are
dependent on $y$ in the plasticity rule
(\ref{dot_w_2}), with the LTD contribution being
proportional to $1/y_c$.
Comparing Figs.~\ref{poisson_spikes_HC}\,\textbf{a)}
and \textbf{b)} we observe that $y_c$ can be used to
regulate the threshold for potentiation in the rate-encoding
limit, without changing the behavior of isolated spike
triplets substantially. $y_c$ is hence a vehicle for
also adapting the overall postsynaptic activity
level and it would be interesting, for future
research, to study how this regulative mechanism
would interact with other known ways to regulate
the overall level of the postsynaptic neural activity,
such as intrinsic plasticity rules
\citep{Triesch, Markovic, Linkerhand}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.0\textwidth]{NECO-03-14-2089R4-Figure9.eps}
\end{center}
\caption{As in Fig.~\ref{poisson_spikes_HC}, now
for visual cortical neurons. Full lines show the weight change
for specific constant presynaptic frequencies $f_{pre}$, as
a function of the postsynaptic rate $f_{post}$. For the
dashed line, pre- and postsynaptic firing rates are equal.
The pair-STDP values (\ref{parameters_A_tau_CX}) have been used
and two sets of values for the remaining three free parameters,
yielding both otherwise similar results for the spike
triplets.
\textbf{a)} $y_c = 11.6$, $y_b = 10.9$, and $x_b = 0.5$.
\textbf{b)} $y_c = 1.0$, $y_b = 0.9$, and $x_b = 0.4$.}
\centerline{\rule{0.8\textwidth}{0.4pt}}
\label{poisson_spikes_CX}
\end{figure}
It has to be remarked that the full lines in
Fig.~\ref{poisson_spikes_HC}, representing weight changes
as a function of the postsynaptic frequency for a constant
presynaptic firing rate, while of theoretical interest
to understand the behaviour of $\theta_H$, will not
correspond to a usual physiological functional relationship
between the rates, at least for a driving synapse. If the
presynaptic synapse drives the postsynaptic neuron, the
postsynaptic activity will in general be an increasing
function of the presynaptic rate. Here we have chosen
$f_{pre} = f_{post}$ (the dashed lines in
Fig.~\ref{poisson_spikes_HC} and \ref{poisson_spikes_CX})
as an illustration, but a more detailed transfer function
should be selected for accurate and quantitative
comparisons with experimental results. In this sense,
the parameter configuration of
Fig.~\ref{poisson_spikes_HC}\,\textbf{b)} shows a better
agreement with experimental procedures, such as that of
\cite{Sjostrom}, where potentiation is shown to become
stronger with higher frequencies.
No complete set of experimental results has hitherto
been published, unfortunately, where all pairwise, triplet,
and frequency dependent plasticity have been measured for the
same type of synapse and with the same experimental
stimulation procedure. A full consistency check between
model and experiment is hence not possible to date.
In Fig.~\ref{poisson_spikes_CX} the results of numerical
simulations for L2/3 cortical neurons for the same
protocol of Fig.~\ref{poisson_spikes_HC} are presented.
In this case, depression is found for all combinations of
pre- and postsynaptic frequencies, a robust prediction
of the model. Different values of $y_c$ were selected
to test this behavior, and in each case the rest of the
parameters were fitted to the triplet results. In each
case, the value of $y_b$ obtained by this fitting turned
out to be lower than $y_c$. The $y$-trace has hence a hard
time to overcome the threshold $y_c$ for LTP, as calcium
increase by further spikes is prevented. As a test, if
$y_{b}$ was artificially set to values larger than $y_c$,
potentiation for larger frequencies was recovered but the
fit of the experimental triplet data deteriorated
substantially, obtaining potentiation for PostPrePost
triplets, contrary to the experimental results. This
indicates that triplet nonlinearities found in L2/3
cortical neurons result from spike suppression, contrary
to the predominant trace accumulation effect present in
hippocampal neurons.
These results, predicted
for L2/3 neurons as fitted from \cite{Froemke}, would
then be in stark contrast to those
of \cite{Sjostrom} for L5 neurons in visual cortex where
LTP dominates for large frequencies. It should be pointed
out, however, that the pairwise STDP plot presented in
\cite{Sjostrom} is already different from that of
L2/3 neurons, raising the question of to what extent
results coming from different neurons, or obtained via
different stimulation procedures, should be alike.
On the other hand, the prediction of overall depression
dominating for uncorrelated spike trains in certain cortical
neurons seems to be in line with, or at least does not
contradict, experimental findings for deprivation experiments.
In cortical areas, where topological maps are usually found,
deprivation of sensory input has been shown to
result in depression of the respective synaptic
connections \citep{Trachtenberg, Feldman}. At the same
time, correlation has been found to substantially decrease
after these procedures, in areas projecting to cortex
\citep{Linden}, suggesting that decorrelation of spike
trains could be responsible for the observed depression in
cortical neurons.
A possible reason behind the observed differences in these
studies might be the stimulation protocol employed. While in
\cite{Bi2} and \cite{Wang}, plasticity is triggered by
eliciting the firing of the pre- and postsynaptic neurons by
dual whole-cell clamp, in the cortical results from
\cite{Froemke}, extracellular presynaptic stimulation is
performed, clamping only the postsynaptic neurons. This creates
an asymmetry between Pre-Post-Pre and Post-Pre-Post triplets.
Moreover, in the case of extracellular stimulation, the
question remains to what extent other synapses are being
affected, potentially triggering, in turn, other forms
of plasticity such as local synaptic scaling.
It is important to stress that the robust depression found
here for higher frequencies is a direct consequence of
the triplet results, and indeed vanishes if one uses
hippocampal-like triplet results. The same suppression
effect present for triplets also affects higher frequency
trains, resulting in depression.
For lower frequencies, the pairwise contribution dominates when
determining the balance between potentiation and depression.
In \cite{Izhikevich}, the authors show how a straightforward
application of the pairwise rule to Poisson uncorrelated spike
trains (as in our simulation), adding up linearly the effect
of every pair in the train according to the pairwise STDP
rule with cortical parameters, always leads to depression,
since the pairs simply sample the STDP curve which has an overall
negative area (the opposite is true in hippocampal neurons, as
we show below). Our model is by construction, equivalent in the
low frequency limit to the linear pairwise model since isolated
spikes produce no synaptic change in our model and triplets and
higher order configurations become very infrequent if the
frequency is low. For low pre- and postsynaptic frequencies
the trains of Poisson spikes can be considered as pairs of
random duration that sample the pairwise STDP curve.
As previously mentioned, the overall integrated area of the
pairwise STDP curve for L2/3 cortical neurons is negative
while it is positive for hippocampal neurons. One can easily
integrate the exponentials and one obtains a relation $A$
for the areas:
\begin{equation}
A = \frac{A^+ \tau_+}{A^- \tau_-}.
\end{equation}
While for hippocampal neurons $A=1.92$, we find in cortex
$A=0.77$. This means that, in absence of higher order
contributions (which is true if both the pre- and the
postsynaptic frequencies are low), uncorrelated spikes will on
average lead to depression in cortical neurons and to
potentiation in hippocampal neurons. If the frequencies tend
to zero, the average interspike interval will be long
compared to the STDP window duration and the net amount of
synaptic change, whether positive or negative, will be low.
In the following section this fact will become clear when
the synaptic change of correlated and uncorrelated spikes
are compared.
As the frequencies of pre- and postsynaptic spikes increase,
the interspike period decreases and when this becomes comparable
to the timescale of the STDP window (which is related to the
trace timescale), the pairwise approximation will break down
since interactions can no longer be neglected. When both the pre-
and postsynaptic frequencies are on the order of $10\mathrm{Hz}$
the average time between a pre- and a postsynaptic spike is on
the order of $50\mathrm{ms}$ and interactions are to be expected.
This is where the particular models for the underlying dynamics
will differ. Also in the work by \cite{Izhikevich}, the authors
show with their \emph{Nearest-neighbor Implementation} that
synaptic change goes from general depression (in the all-to-all
implementation) to BCM-like, when they consider only the
closest previous and posterior postsynaptic spike to each
presynaptic spike to compute the linear sum of pairs.
This choice, which at first glance would seem an approximation
independent of any underlying dynamics, actually has
strong implications for the biological underpinnings that
could actually implement this algorithm. A first neighbor
approximation requires to hard reset any traces possibly
present, forgetting completely anything that happened
outside that window. The nearest-neighor implementation
has, in addition not the aim to explain triplet non-linearities
as the PrePostPre triplet protocol.
The interspike interaction is in our model driven by the
undelying traces. We have chosen in our simulation to use
for the frequency-dependence protocol the same parameters
obtained from pairwise and triplet fits. As observed in
cortical PostPrePost triplets, strong suppression
severely limits potentiation of further spikes
(compare PostPrePost to linear superposition results).
At high frequencies triplet interactions become relevant
and the same suppression should then be evidenced for
frequency dependent plasticity. We believe then, that any
model aiming to reproduce time-dependent plasticity up to
triplet order as measured by \cite{Froemke}, should show
depression also for high frequencies in cortical neurons.
\subsection{Plasticity induced by correlated spikes}\label{correlated}
So far we have analyzed the effect of
uncorrelated spikes on the synaptic weight change both
in hippocampal and in cortical neurons
(Figs.~\ref{poisson_spikes_HC} and \ref{poisson_spikes_CX}).
It is however interesting to see the predictions of the model
for a strong synapse driving the postsynaptic neuron. In this
case pre- and postsynaptic spikes should be correlated,
at least partially, together with a certain positive delay.
To reproduce this effect with our model, we
simulated trains of correlated spikes where, with each
presynaptic spike, a postsynaptic spike can occur with
probability $p$, after a certain delay $d$. As an example,
if every presynaptic spike triggers a postsynaptic spike
then $p = 1$. This would mean, however that the postsynaptic
frequency $f_{post}$ changes with $p$ ($f_{post} = p$ $f_{pre}$).
In order to compare our results for different values of $p$
and keep $f_{post}$ independent of $p$, we complete the
train of postsynaptic
spikes with Poisson spikes of frequency $(1-p)f_{pre}$.
In this way, $f_{post}$ is independent of $p$, which now
regulates the degree of correlation between pre- and
postsynaptic spikes: $p=0$ represents the fully decorrelated
case, since all the postsynaptic spikes are drawn from
the poisson distribution, and $p=1$ represents the fully
correlated case already mentioned.
In Fig.~\ref{correlated_spikes} we present the synaptic
changes produced by trains spikes for different values of
$p$. In this case, results for a delay of $5\mathrm{ms}$
are presented. The same tests were performed with delays from
$2\mathrm{ms}$ to $10\mathrm{ms}$ with only quantitative but
not qualitative differences.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=1.0\textwidth]{NECO-03-14-2089R4-Figure10.eps}
\end{center}
\caption{Simulation results of the synaptic weight change
induced by a one-second train of spikes with different
degrees of correlation, as a function of the frequency
$f = f_{pre} = f_{post}$. The fraction $p$ of correlated
spikes takes the values $0$, $0.5$, and $1$ in these
plots. The delay between pre- and postsynaptic spikes was
taken to be $5\mathrm{ms}$ in these simulations. The case
$p=0$ corresponds to the dashed lines in
Figs.~\ref{poisson_spikes_HC} and \ref{poisson_spikes_CX}.
\textbf{a)} The pairwise hippocampal parameters
(\ref{parameters_A_tau}) were used, together with
$y_c = 0.28$, $y_b = 0.66$, and $x_b = 0.62$,
corresponding to the triplet fit.
\textbf{b)} Here the pairwise cortical parameters
(\ref{parameters_A_tau_CX}) were used, together with
$y_c = 11.6$, $y_b = 10.9$, and $x_b = 0.5$,
corresponding to the triplet fit.}
\centerline{\rule{0.8\textwidth}{0.4pt}}
\label{correlated_spikes}
\end{figure}
We observe in Fig.~\ref{correlated_spikes}, both for
hippocampal and cortical neurons, that correlated spike
trains induce an increasing amount of potentiation for low
to intermediate frequencies ($\sim1 -10\mathrm{Hz}$). In
the correlated scenario, and since in this case we are
simulating a driving synapse, postsynaptic spikes follow
presynaptic spikes in a causal order. When the frequency
is higher than $10\mathrm{Hz}$, the interspike period
becomes comparable to the STDP time window and each
postsynaptic spike will also ''see'' the following
presynaptic spike, thus triggering the LTD term.
Depending on the trace saturation constants, LTD or
LTP will eventually dominate for large frequencies.
If LTD dominates, depression results and after a
certain reversal frequency the behavior is switched from
Hebbian to Anti-Hebbian. This is the case for the triplet
fitted values presented in Fig.~\ref{correlated_spikes}.
It is important to note that the model is also able to
produce Hebbian behavior within the entire physiological range
of activities by changing $y_c$. The smaller the saturation
effects are, the larger this reversal frequency becomes. In fact,
with the second set of parameters used in
Fig.~\ref{poisson_spikes_HC}\,\textbf{b)}, no such reversal is
found within physiological frequencies (not shown here).
If observed, such a reversal, which is yet another side of the
suppression effect, would have the benefit of being self-stabilizing,
tuning synaptic strength to help keep neural activities bound.
The kink observable for the fully correlated curve
($p = 1.0$) of \ref{correlated_spikes}~\textbf{b)}
results from the particularly strong suppression effects
in cortical neurons, captured in our model by $y_b < y_c$.
For frequencies above a certain threshold
($\sim28\mathrm{Hz}$) the trace concentration $y$ is
always below the $y_c$ and LTP never triggers.
A fundamental difference between plots \textbf{a)} and
\textbf{b)} is the different qualitative behavior between
correlated and uncorrelated spikes. While in hippocampal
neurons, Hebbian, increasing potentiation, is always
present for low to intermediate frequencies (whether the
spikes are correlated or not), in cortical neurons,
our model predicts uncorrelated spikes always produce
depression and therefor Hebbian learning requires the
neurons to be at least partially correlated.
\section{Comparison to other models}
The problem of formulating plasticity in terms of the
specific timing of pre- and postsynaptic spikes can be
approached at different levels of detail and accuracy,
ranging from simplistic phenomenological rules to detailed
and complex models describing the different steps of the
biological machinery responsible for STDP. In sections
\ref{Results_Hippocampus} and \ref{Results_Cortex}
the comparison of our model to simple forms of
phenomenological rules has already been established,
noting that linear combinations of spike pairs are
generically not sufficient to explain the experimentally
observed triplet non-linearities.
We have also shown that, while linear combinations of
pairs, plus additional suppression, is enough to explain
the triplet nonlinearities of cortical neurons, as shown in
\cite{Froemke}, hippocampal triplet non-linearities cannot
be explained by suppression and a trace accumulation
mechanism seems to be taking place in these synapses. In
any case, this kind of phenomenological rules are not
likely to generalize well to arbitrary spike patterns
since no information of the underlying plasticity
mechanism is present in the formulation.
Other models, like \cite{Albers} and \cite{Pfister}, present
interesting dynamical formulations of plasticity in
terms of generic decaying markers or traces, but do
not attempt to establish a link to the biological
underpinnings of STDP. Calcium concentration and NMDA
receptors have been shown to play a central role in
time-dependent LTP and LTD, and we therefore believe it is
important to formulate plasticity in those terms. Our
model, though simplified, is formulated in terms of these
key ingredients and may therefore help to bridge the worlds
of functional and realistic models.
An alternative approach has been proposed by
\cite{Appleby}, where plasticity is described
in an ensemble-based formulation. The authors there argue
that the observed synaptic changes produced by standard
protocols cannot be explained at a single synapse level,
but rather state that the observed results arise at a
population level. The authors then show how the pairwise
STDP curve can be recovered at the ensemble level from
all or nothing potentiation or depression at the single
synapse level. The dependency of the model to the specific
timing of triplets is however in this case not computed.
The model we present in this work belongs to the family
of calcium-based spiking-neuron models. Within this
family, models formulating synaptic plasticity exclusively
in terms of the calcium levels \citep{Uramoto,Graupner},
while tuned to reproduce a variety of
experimental results, tend to show paradoxical results when
tested in other setups. The model presented by \cite{Uramoto},
for example, predicts synaptic changes even when only
postsynaptic spikes are present. The model in
\cite{Graupner}, in turn, shows plasticity also when
either pre- or postsynaptic spikes are absent,
since both pre- and postsynaptic spikes contribute directly
to the calcium level in this model, without the need of
coincidence. To avoid this, in our model we demand the
simultaneous presence of both pre- and postsynaptic spikes
for plasticity to arise, being proportional to the products
of traces $x$ and $y$ in our rule. We believe this to be an
important feature for simulations in situations of complex
spike patterns where the pre- and postsynaptic firing rates
do not necessarily match.
\section{Discussion}
We propose a basic trace model for timing-dependent
plasticity that incorporates, in a first order
approximation, the fundamental mechanisms acknowledged to
be taking place in STDP. We show that the model
successfully captures several main features of time-dependent
plasticity, including the standard shape for low frequency
pairing, experimentally observed triplet nonlinearities,
and large frequency effects.
The decay constants for the two traces and the relative
intensities of LTP and LTD can be extracted directly from
the standard STDP curves, as measured for isolated pairs.
The model is left hereafter with only three further
parameters, which can be used to fit higher order
contributions to plasticity. The model successfully
reproduces the distinct and contrasting nonlinearities
found in both hippocampal cultures and cortical
slices.
While the model predicts a similar frequency dependence
for correlated (or partially correlated) pre- and postsynaptic
spikes both for hippocampal and cortical neurons, the effect
of uncorrelated spikes (although smaller) differs
qualitatively in these two types of neurons. In this case,
the sign of the resulting plasticity depends for lower
frequencies on the overall area of the pairwise STDP curve,
resulting in potentiation for hippocampal neurons and
depression in cortical ones, and for higher frequencies
on the balance between spike suppression and trace accumulation.
We show that the model is able to reproduce typical
frequency-dependencies for uncorrelated spikes, while
fitting pairwise and triplet hippocampal parameters.
We do also find, that fitting triplet results for
L2/3 cortical neurons invariably leads to depression,
for higher frequencies and of uncorrelated spikes, contrary
to observations in L5 neurons. The question then arises,
to which extent plasticity results for different neurons,
or performed under different stimulation conditions, can
be expected to match. It seems then essential to have available
a consistent sets of experiments where pairwise, triplet, and
frequency results are measured for the same type of neuron and
with the same stimulation protocol. Otherwise one runs the
risk of possibly trying to build a complete picture out of
mismatching parts. In this sense, we hope that our predictions
serve as a motivation to revisit and to complete triplet and
frequency dependent studies for different types of neurons.
In order to compare our results with rate encoding plasticity
models, we have also shown, by setting the presynaptic frequency
to a constant value, that the amount of synaptic change is
proportional, for hippocampal neurons, to the product of the
activities, with a threshold that depends on the presynaptic
firing rate. While the system lacks a longer term average
threshold, as the one present in BCM, the presynaptic activity
acts as a value of reference for the significant level of activity.
If the postsynaptic activity exceeds this level then
potentiation occurs, otherwise, depression arises.
For correlated spikes, we have shown that the model leads to
similar results for hippocampal and cortical neurons, with
an initial hebbian behavior for small to medium frequencies
and, depending on choice for the parameters, a reversal to
anti-hebbian behavior for large frequencies, which could have
the virtue of being self-limiting, avoiding runaway growth
of synaptic connections. It has been shown recently
\citep{Echeveste}, that this self-limitation results naturally,
for rate encoding neurons, from the stationarity principle for
Hebbian learning. By tuning the value of $y_c$, the reversal
frequency can be made larger, to the point of producing Hebbian
behavior within the entire physiological range of activities.
The simplicity of the here proposed model makes it a useful
tool for simulations and studies of the dynamical
properties of networks adapted via these rules. Firstly, the
calculations required are straightforward, making it
computing time efficient. The relative small number of free
parameters and their direct link to both the biophysical
properties of the postsynaptic complex and to the dynamical
features of the trace dynamics makes it suitable when
studying the interplay between neural and synaptic dynamics
in neural systems.
\section*{Acknowledgments}
The authors acknowledge Robert C. Froemke and Yang Dan for
the experimental data from cortical visual neurons.
\section*{Appendix: Dimensionality Analysis}
In section \ref{model} we could have started by initially
denoting by $x'$ and $y'$ the fraction of NMDA receptors and
the $Ca^{2+}$ concentration, respectively, where the time
evolution of these traces is written as:
\begin{equation}
\left\{
\begin{array}{lcl}
\dot x' = - \frac{x'}{\tau_x} + c_1 \mathit{E_x}
\sum_\sigma\delta (t-t_{pre}^\sigma) \\
\dot y' = - \frac{y'}{\tau_y} + (c_2 x'+y'_c) \mathit{E_y}
\sum_\sigma\delta (t-t_{post}^\sigma)
\end{array}
\right.
\label{dot_xy_2}
\end{equation}
where $\tau_y$ and $\tau_x$ represent the time constants in
the decay of $x'$ and $y'$, and now two additional
parameters $c_1$ and $c_2$ are included. $c_1$ represents
the increase in $x'$ caused by a single presynaptic spike
(which in this simplified model we assume constant) and
$c_2$ represents the increase, per unit of $x'$, in $y'$
concentration. $y'_c$ is the constant contribution per
postsynaptic spike to $y'$ of the voltage-gated $Ca^{2+}$
channels. Again a spike efficacy $\mathit{E}$ is included that
limits trace concentrations, where $\mathit{E}$ is still
calculated as in (\ref{theta}).
Now, by a change of variables:
\begin{equation}
x = x'/c_1, \qquad
x_b = x'_b/c_1, \qquad
y = y'/c_1c_2, \qquad
y_c = y'_c/c_1c_2, \qquad
y_b = y'_b/c_1c_2
\label{change}
\end{equation}
we re-write (\ref{dot_xy_2}) as:
\begin{equation*}
\left\{
\begin{array}{lcl}
\dot x = - \frac{x}{\tau_x} + \mathit{E_x}
\sum_\sigma\delta (t-t_{pre}^\sigma) \\
\dot y = - \frac{y}{\tau_y} + (x+y_c) \mathit{E_y}
\sum_\sigma\delta (t-t_{post}^\sigma)
\end{array}
\right.
\end{equation*}
which is the exactly the expression presented in section
\ref{model}. By this procedure we have reduced the number
of parameters for the trace evolution to 5.
|
\section{Conclusion}
\label{sec:conclusion}
In this paper, we have examined the problem of contaminated likelihood functions that arise due to
label noise in hypothesis testing. In contrast to previous work on the subject which derived consistency results for the case when the likelihoods are mutually irreducible, we deal with arbitrary likelihoods and obtain
decision rules robust to uncertainty in the contamination proportions. Toward this end, we have posed an optimization problem that is naturally subject to linear constraints and shown that its objective function is a linear-fractional function. Therefore, the optimization problem reduces to
linear programs that can be
simplified using the KKT conditions into
a search over certain vertices of the constraint set.
We have shown the method on two numerical examples.
\section{Introduction}
\label{sec:intro}
Label noise in hypothesis testing problems results in the cross-contamination of the likelihood functions and possible degradation in detection performance if not accounted for when determining a decision rule. In this paper, we propose a linear programming framework for robustly dealing with contaminated likelihoods. Specifically, we propose an algorithm for obtaining
a minimax optimal decision rule under label noise that is applicable under general likelihood models.
We are motivated by problems encountered in workforce analytics: data-driven decision making to manage the human capital of a corporation. For example, decision makers may want to use human resources data to predict whether or not an employee will voluntarily resign within the next 12 months \cite{SinghVWMGFE2012}, or decision makers may want to determine whether an employee from another division is a suitable candidate to fill an open position on a team in their division, based on skills and expertise data about the employee. We face label noise and contamination of hypotheses in both examples. In the voluntary resignation example, we can take all employees that resigned in the recent past as samples from the alternative hypothesis and all employees that are currently active as samples from the null hypothesis. However, among currently active employees, some will resign in the coming months. Therefore, we are not in a position to observe an uncontaminated null distribution. In the suitable candidate example, we can take all employees in the decision maker's team as samples from the alternate distribution and all other employees as samples from the null distribution. However, not all team members may be suitable for the open position and not all other employees are unsuitable (which is why this problem is posed in the first place). Thus in this example, we observe contaminated versions of both likelihoods.
The problem of contaminated likelihoods in binary hypothesis testing was recently studied in considerable generality in \cite{scott2013COLT,scott2013arXiv}. The theoretical framework in the present work is largely guided by \cite{scott2013COLT,scott2013arXiv}. These previous works assume that the true likelihoods have an irreducibility property (described more fully in Section~\ref{sec:theory}) that allows consistency results to be established. However, the assumption of irreducibility is restrictive. It is not satisfied for example by two Gaussian distributions with different variances, nor is it likely to be satisfied by real-world distributions such as may be encountered in workforce analytics. A contribution of the current paper in Section~\ref{sec:theory} is to remove the irreducibility assumption and extend the analysis to arbitrary true likelihoods. Furthermore, the approach taken herein, described in Section~\ref{sec:robust}, differs fundamentally from \cite{scott2013COLT,scott2013arXiv} in focusing not on consistent learning of a particular contamination model, but rather on designing hypothesis tests that are robust to uncertainty in the model. In Section~\ref{sec:numerical}, the utility of the robust viewpoint is demonstrated in two numerical examples.
More broadly, various types of label noise have been studied in the machine learning literature, including random, adversarial, and observation-dependent, and noise that affects different classes symmetrically and asymmetrically \cite{FrenayK2015}. However, the vast majority of that work has been devoted to classifiers learned from finite training data and has been specific to particular
supervised classification algorithms, see numerous references given in \cite{scott2013COLT,scott2013arXiv}. In contrast, our work deals with the regime encountered in signal detection theory and hypothesis testing, not the regime with finite training samples. Therefore, we work with likelihood ratio tests and true error probabilities rather than with specific classification algorithms and generalization bounds. Somewhat more related is the mixture modeling approach of \cite{lawrence2001,bouveyron2009}, which attempts to learn the contamination model using the EM algorithm. This approach however requires parametric assumptions on the true likelihoods that we do not make.
\section{Numerical Examples}
\label{sec:numerical}
In this section we illustrate the proposed minimax procedure via two examples with likelihoods that are not mutually irreducible: Gaussian distributions with different means and different variances, and exponential distributions with different inverse scale parameters. The Gaussians example provides a rough model for features that predict voluntary resignation, since features such as time since the last job promotion and annual performance rating tend to be approximately normal in many organizations. The exponentials example provides a rough model for abilities among a high-performing group, which arises when finding suitable candidates.
Consider $P_0 \sim \mathcal{N}(\mu_0,\sigma_0^2)$ and $P_1 \sim \mathcal{N}(\mu_1,\sigma_1^2)$ where $\mu_0 \neq \mu_1$ and, without loss of generality, $\sigma_0 < \sigma_1$. For this problem, the uncontaminated error probabilities for a likelihood ratio test with threshold value $\gamma$ are:
\begin{align*}
R_0(\gamma) &= Q\left(\tfrac{y^+ - \mu_0}{\sigma_0}\right) + Q\left(\tfrac{-y^- + \mu_0}{\sigma_0}\right) \\
R_1(\gamma) &= 1 - Q\left(\tfrac{y^+ - \mu_1}{\sigma_1}\right) - Q\left(\tfrac{-y^- + \mu_1}{\sigma_1}\right),
\end{align*}
where $Q(y) = \frac{1}{\sqrt{2\pi}}\int_y^\infty \exp(-y'^2/2)dy'$, and $y^+$ and $y^-$ are the solutions to the quadratic equation:
\begin{displaymath}
(\sigma_1^2 - \sigma_0^2)y^2 + 2(\mu_1\sigma_0^2 - \mu_0\sigma_1^2)y + \mu_0^2\sigma_1^2 - \mu_1^2\sigma_0^2 - 2\sigma_0^2\sigma_1^2\ln\left(\gamma\tfrac{\sigma_1}{\sigma_0}\right) = 0.
\end{displaymath}
We examine the situation in which $\mu_0 = 0$, $\mu_1 = 0.2$, $\sigma_0 = 1$, and $\sigma_1 = 2$. Additionally, for the Bayes risk, we consider the simple case when $q_0 = 0.5$ and $c_{01} = c_{10} = 1$. The true contamination proportions, unknown to an observer, are $\pi_0 = 0.2$ and $\pi_1 = 0.3$. These contamination proportions result in $\nu^\ast(\tP_0, \tP_1) = 0.2857$ and $\nu^\ast(\tP_1, \tP_0) = 0.7202$, which are observed. Additional information on the contamination gives us the constraints $\pi_0 \ge 0.05$ and $\pi_1 \ge 0.1$, as well as $\pi_0 + \nu^\ast(\tP_0, \tP_1)\pi_1 \ge 0.2$ and $\nu^\ast(\tP_1, \tP_0)\pi_0 + \pi_1 \ge 0.25$. The last two inequalities follow from Theorem~\ref{thm:pi01} and upper bounds on $\nu^\ast(P_0,P_1)$, $\nu^\ast(P_1,P_0)$. With these constraints, the polygon $\Pi$ has six vertices.
After performing the inner maximization of the minimax procedure, we find the vertex of $\Pi$ that maximizes the Bayes risk to be $(0.1619, 0.1334)$.
This maximum Bayes risk is shown in Fig.~\ref{fig:RB}(a) as a function of the threshold $\lambda$ applied to the contaminated likelihood ratio
($\lambda$ is related to $\gamma$ through a transformation derived in \cite{scott2013COLT}).
\begin{figure}
\centering
\begin{tabular}{c}
\includegraphics[width=0.4\linewidth]{gaussRB}\\
\footnotesize{(a)}\\
\includegraphics[width=0.4\linewidth]{expRB}\\
\footnotesize{(b)}
\end{tabular}
\caption{Bayes risk as a function of the threshold on the contaminated likelihood ratio
for (a) Gaussian example and (b) exponential example: using unknown true contamination proportions (dashed), max solution (solid), and $(0,0)$ contamination proportions (dash-dot).}
\label{fig:RB}
\end{figure}
The minimum value of this function, i.e.\ the minimax Bayes risk we seek, is $0.3845$.
The figure also shows the Bayes risk if we use the unknown true contamination proportions (which equals the uncontaminated Bayes risk) and the Bayes risk if we use the $(0,0)$ point, i.e., we do not account for contamination. The minimum Bayes risk using the true contamination proportions is $0.3372$ and the minimum when using $(0,0)$ is $0.4186$. The minimax solution is between these two values. Notably, it is less pessimistic than the default $(0,0)$ solution. The
solution under irreducibility \cite{scott2013COLT} is not selected under the minimax criterion as it is too optimistic about the Bayes risk value.
As a second example, consider $P_0 \sim \mathcal{E}(\alpha_0)$ and $P_1 \sim \mathcal{E}(\alpha_1)$ where without loss of generality, $\alpha_0 < \alpha_1$. For this problem, the uncontaminated error probabilities for a likelihood ratio test threshold value $\gamma$ are: $R_0(\gamma) = 1 - e^{-\alpha_0y^*}$ and $R_1(\gamma) = e^{-\alpha_1y^*}$, where
$y^* = \ln\left(\frac{\alpha_0}{\alpha_1}\gamma\right)/(\alpha_0 - \alpha_1)$. We set $\alpha_0 = 1$ and $\alpha_1 = 2$ and keep all other parameters the same as in the first example. With these exponential likelihoods and parameter settings, $\nu^\ast(\tP_0, \tP_1) = 0.7059$ and $\nu^\ast(\tP_1, \tP_0) = 0.3750$ and the resulting $\Pi$ has five vertices. The maximizing vertex is $(0.1619,0.1334)$ and the maximum Bayes risk is shown in Fig.~\ref{fig:RB}. The minimax Bayes risk is $0.4130$, which lies between the minimum Bayes risk with known contamination proportions, $0.3750$, and the minimum Bayes risk using proportions $(0,0)$, $0.4375$, in the same manner as the previous example.
\section{Problem Statement}
\label{sec:prob}
We consider the binary hypothesis testing problem of deciding between a null hypothesis $H = h_0$ and an alternative hypothesis $H = h_1$ based on observation of a random variable $Y$. Under hypothesis $H = h_0$, $Y$ follows the probability distribution $P_0$, while under $H = h_1$, $Y$ follows distribution $P_1$. A decision rule $\hH$ is desired that maps every possible observation $Y = y$ to either $h_0$ or $h_1$. For a rule $\hH$, define $R_0(\hH) = \Pr(\hH = h_1 \mid H = h_0)$ and $R_1(\hH) = \Pr(\hH = h_0 \mid H = h_1)$ to be the Type I and Type II error probabilities. In this paper we focus on the Bayesian formulation in which the hypotheses have prior probabilities $\Pr(H = h_0) = q_0$, $\Pr(H = h_1) = 1-q_0$, and the performance measure is the Bayes risk
\begin{equation}\label{eqn:BayesRisk}
R_B(\hH) = c_{01} q_0 R_0(\hH) + c_{10} (1-q_0) R_1(\hH),
\end{equation}
where $c_{01}$ and $c_{10}$ are the costs of Type I and Type II errors.
Given knowledge of the conditional distributions $P_0$ and $P_1$, it is straightforward to construct a likelihood ratio test that minimizes the Bayes risk \cite{VanTrees1968}. However, in the contaminated version of the problem considered herein, $P_0$ and $P_1$ are not known.
Instead, we have access to the contaminated distributions
\begin{subequations}\label{eqn:Ptilde01}
\begin{align}
\tP_0 &= (1 - \pi_0) P_0 + \pi_0 P_1,\label{eqn:Ptilde0}\\
\tP_1 &= (1-\pi_1) P_1 + \pi_1 P_0,\label{eqn:Ptilde1}
\end{align}
\end{subequations}
where the contamination proportions $\pi_0, \pi_1 \in [0,1]$ are also unknown. The following constraint is placed on $\pi_0$, $\pi_1$,
\begin{equation}\label{eqn:contamBound}
\pi_0 + \pi_1 < 1,
\end{equation}
to resolve an interchange ambiguity and with essentially no loss of generality.
Indeed, if $\pi_0 + \pi_1 > 1$, then as noted in \cite{scott2013COLT}, interchanging $P_0$ and $P_1$ yields complementary proportions $1-\pi_0$, $1-\pi_1$ satisfying $(1-\pi_0) + (1-\pi_1) < 1$. If $\pi_0 + \pi_1 = 1$, then \eqref{eqn:Ptilde01} implies that $\tP_0 = \tP_1$ and discrimination is not possible.
As discussed in \cite{scott2013COLT}, it is not possible in general to design a test $\hH$ that minimizes the Bayes risk \eqref{eqn:BayesRisk}, defined in terms of the true distributions $P_0$, $P_1$, given only the contaminated distributions $\tP_0$, $\tP_1$ and no knowledge of $P_0$, $P_1$, $\pi_0$, $\pi_1$. Therefore in this paper we revise the objective to that of choosing $\hH$
to be robust to the uncertainty in
$P_0$, $P_1$, subject to limited additional input.
We note that in the absence of further conditions,
there is a large range of possible solutions to \eqref{eqn:Ptilde01}. In particular, it cannot be ruled out that there is no contamination, i.e.~$\pi_0 = \pi_1 = 0$, $P_0 = \tP_0$, and $P_1 = \tP_1$. In the sequel, we seek to identify conditions that require minimal knowledge of or assumptions on $P_0$, $P_1$, $\pi_0$, $\pi_1$ while also restricting uncertainty in a meaningful way in terms of Bayes risk
We focus in this paper on the population setting where the distributions $\tP_0$ and $\tP_1$ are known exactly. Our results can be extended fairly straightforwardly to the finite-sample setting where $\tP_0$ and $\tP_1$ are approximated using training data, for example following the learning-theoretic approach of \cite{scott2013COLT}.
In the finite-sample case, the lack of knowledge of $P_0$, $P_1$ translates into an inability to draw samples from $P_0$, $P_1$.
\section{Contamination-Robust Hypothesis Testing}
\label{sec:robust}
This section discusses the determination of decision rules that are robust to uncertainty in the contamination proportions $\pi_0$ and $\pi_1$.
Defining $\mbpi = (\pi_0, \pi_1)$, we rewrite the Bayes risk \eqref{eqn:BayesRisk} as follows,
\begin{equation}\label{eqn:BayesRiskContam}
R_{\Bayes}(\hH,\mbpi) = c_{01} q_0 R_0(\hH,\mbpi) + c_{10} (1-q_0) R_1(\hH,\mbpi),
\end{equation}
to make explicit the dependence on the contamination proportions.
From \eqref{eqn:Ptilde01Alt}, the two error probabilities under the true distributions $P_0$, $P_1$ can be expressed as
\begin{subequations}\label{eqn:R01}
\begin{align}
R_0(\hH,\mbpi) &= \frac{(1-\pi_1) \tR_0(\hH) - \pi_0(1-\tR_1(\hH))}{1-\pi_0-\pi_1},\label{eqn:R0}\\
R_1(\hH,\mbpi) &= \frac{(1-\pi_0) \tR_1(\hH) - \pi_1(1-\tR_0(\hH))}{1-\pi_0-\pi_1}.\label{eqn:R1}
\end{align}
\end{subequations}
The performance thus depends on the error probabilities $\tR_0(\hH)$, $\tR_1(\hH)$ under the contaminated distributions, which can be determined for fixed decision rule $\hH$, and $\pi_0$, $\pi_1$, which are only partially known.
The set of possible $(\pi_0, \pi_1)$ values is constrained by knowledge of $\tP_0$ and $\tP_1$ as shown in Fig.~\ref{fig:feasiblepi}. In addition to these initial constraints, we also consider lower and/or upper bounds on $\pi_0$, $\pi_1$ and the maximal mixture proportions $\nu^{\ast}(P_0, P_1)$, $\nu^{\ast}(P_1, P_0)$ for the pure distributions. As seen from Theorem~\ref{thm:pi01} and Fig.~\ref{fig:feasiblepi}, bounds on $\nu^{\ast}(P_0, P_1)$, $\nu^{\ast}(P_1, P_0)$ correspond to linear inequalities in $\pi_0$, $\pi_1$. It follows that the feasible region for $(\pi_0, \pi_1)$ is in general a convex polygon, which we may represent as a system of linear inequalities:
\[
\Pi = \{\mbpi : \mba_i^T \mbpi \leq b_i, \; i=1,\dots,m \}
\]
with appropriate choices of $\mba_i \in \mathbb{R}^2$ and $b_i \in \mathbb{R}$.
The additional bounds on $\pi_0$, $\pi_1$, $\nu^{\ast}(P_0, P_1)$, $\nu^{\ast}(P_1, P_0)$ may be provided by application-specific knowledge and past experience. For example, with voluntary resignation, we can examine the resignation rate historically and use it to roughly characterize or bound $\pi_0$. Moreover, examining data from more than a year in the past, we can observe $P_0$ and $P_1$ without contamination because any employee who was active then and has not resigned yet is by definition not a contaminated sample. Such historical $P_0$ and $P_1$ can be used to bound present values of $\nu^{\ast}(P_0, P_1)$ and $\nu^{\ast}(P_1, P_0)$.\footnote{One may ask why historical $P_0$ and $P_1$ cannot simply be used to determine the decision rule in the present; this is not possible in dynamic business environments where the resignation rate within job roles, skill sets, professions, and organizational units\hspace{1pt}---\hspace{1pt}which are all observations to predict resignation\hspace{1pt}---\hspace{1pt}changes rapidly due to technology trends and management changes. It is the level of differentiation between the classes that we assume does not change much over time, allowing us to bound $\nu^{\ast}(P_0, P_1)$ and $\nu^{\ast}(P_1, P_0)$.} In the case of finding suitable internal candidates for openings, similar openings filled in adjacent groups can provide bounds on $\pi_0$, $\pi_1$, $\nu^{\ast}(P_0, P_1)$, $\nu^{\ast}(P_1, P_0)$.
In this paper, the decision rule $\hH$ is chosen to minimize the Bayes risk subject to worst-case uncertainty in $(\pi_0, \pi_1)$ within the set $\Pi$:
\begin{equation}\label{eqn:BayesTest}
\hH_{\Bayes} = \argmin_{\hH} \, \max_{\mbpi \in \Pi} \, R_{\Bayes}(\hH,\mbpi).
\end{equation}
Alternative formulations include minimizing the worst-case deviation from the true Bayes risk (instead of the absolute Bayes risk in \eqref{eqn:BayesTest}) and minimizing the average Bayes risk over $\Pi$ with respect to some distribution for $\mbpi$. We leave these alternatives for future work.
The inner maximization in \eqref{eqn:BayesTest} can be restricted to a subset of the vertices of $\Pi$. For a vertex $\mbpi \in \Pi$, define $I(\mbpi) \subseteq \{1,\dots,m\}$ to be the set of constraints $\mba_i^T \mbpi \leq b_i$ that are met with equality (active constraints), and $\cone\left(\{\mba_i, i \in I(\mbpi)\}\right)$ to be the cone formed by non-negative combinations of the corresponding $\mba_i$. We use $\mathbb{R}_{-}^2$ as a shorthand for the non-positive quadrant of $\mathbb{R}^2$.
\begin{lemma}\label{lem:vertexBayes}
Assume that $\hH$ satisfies $\tR_0(\hH) + \tR_1(\hH) \leq 1$. Let $\mbpi^{k}$, $k = 1,\dots,V$, be the vertices of $\Pi$ such that
\begin{equation}\label{eqn:coneCond}
\cone\bigl(\{\mba_i, i\in I(\mbpi^k) \}\bigr) \cap \mathbb{R}_{-}^2 \neq \emptyset.
\end{equation}
Then
\[
\max_{\mbpi \in \Pi} \, R_{\Bayes}(\hH,\mbpi) = \max_{k=1,\dots,V} \, R_{\Bayes}(\hH,\mbpi^k).
\]
\end{lemma}
\begin{IEEEproof}
The restriction to vertices of $\Pi$ follows from the fact that $R_{\Bayes}(\hH,\mbpi)$ is a linear-fractional function of $\mbpi$ for fixed $\hH$.
This property is seen by substituting \eqref{eqn:R01} into \eqref{eqn:BayesRiskContam} to obtain
\begin{equation}\label{eqn:BayesRiskQuasilinear}
R_{\Bayes}(\hH,\mbpi) = \frac{\mbc^T \mbpi + d}{1 - \pi_0 - \pi_1},
\end{equation}
where $\mbc \in \mathbb{R}^2$ and $d \in \mathbb{R}$ do not depend on $\mbpi$ (explicit expressions are omitted here).
Given \eqref{eqn:BayesRiskQuasilinear}, the maximization of $R_{\Bayes}(\hH,\mbpi)$ may be carried out as a search for the largest $t \geq 0$ for which the linear program
\begin{equation}\label{eqn:LPBayes}
\max_{\mbpi \in \Pi} \, \mbc^T \mbpi + d - t(1 - \pi_0 - \pi_1)
\end{equation}
has a non-negative optimal value, implying that the superlevel set $\{\mbpi \in \Pi : R_{\Bayes}(\hH,\mbpi) \geq t \}$ is non-empty. Since \eqref{eqn:LPBayes} is a linear optimization over a bounded polygon, there exists a vertex of $\Pi$ that is optimal \cite[Thm.~2.8]{bt1997}. This holds in particular for $t = \max_{\mbpi \in \Pi} R_{\Bayes}(\hH,\mbpi)$ and hence it is sufficient to consider only the vertices of $\Pi$ in maximizing $R_{\Bayes}(\hH,\mbpi)$.
The restriction to vertices satisfying \eqref{eqn:coneCond} is due to the KKT optimality condition for the maximization of $R_{\Bayes}(\hH,\mbpi)$:
\begin{equation}\label{eqn:KKTBayes}
\nabla_{\mbpi} R_{\Bayes}(\hH,\mbpi) = \sum_{i\in I(\mbpi)} \mu_i \mba_i, \quad \mu_i \geq 0,
\end{equation}
which is a necessary condition because $\Pi$ is defined by linear inequalities \cite[Prop.~3.3.7]{bertsekas1999}. Using \eqref{eqn:R0}, \eqref{eqn:linFracDeriv},
and the assumption $\tR_0(\hH) + \tR_1(\hH) \leq 1$, we find that
\begin{align*}
\frac{\partial R_0(\hH,\mbpi)}{\partial\pi_0} &= -\frac{(1-\pi_1)\bigl(1 - \tR_0(\hH) - \tR_1(\hH)\bigr)}{(1 - \pi_0 - \pi_1)^2} \leq 0,\\
\frac{\partial R_0(\hH,\mbpi)}{\partial\pi_1} &= -\frac{\pi_0 \bigl(1 - \tR_0(\hH) - \tR_1(\hH)\bigr)}{(1 - \pi_0 - \pi_1)^2} \leq 0,
\end{align*}
and similarly for $R_1(\hH,\mbpi)$. Since $R_{\Bayes}(\hH,\mbpi)$ is a non-negative combination of $R_0(\hH,\mbpi)$ and $R_1(\hH,\mbpi)$ from \eqref{eqn:BayesRiskContam}, we have $\nabla_{\mbpi} R_{\Bayes}(\hH,\mbpi) \in \mathbb{R}_{-}^2$ in \eqref{eqn:KKTBayes}, while the right-hand side of \eqref{eqn:KKTBayes} can range over $\cone\bigl(\{\mba_i, i\in I(\mbpi) \}\bigr)$. We conclude that it suffices to consider vertices satisfying \eqref{eqn:coneCond}.
\end{IEEEproof}
\begin{remark}
The condition $\tR_0(\hH) + \tR_1(\hH) \leq 1$ is satisfied by any decision rule $\hH$ that is at least as good as random guessing. Hence no generality is lost.
\end{remark}
Combining \eqref{eqn:BayesTest} and Lemma~\ref{lem:vertexBayes} yields
\begin{equation}\label{eqn:BayesTestVertex}
\hH_{\Bayes} = \argmin_{\hH} \, t \quad \text{s.t.} \quad R_{\Bayes}(\hH,\mbpi^k) \leq t, \quad k=1,\dots,V.
\end{equation}
In the two-dimensional case considered here, the number $V$ of vertices satisfying \eqref{eqn:coneCond} is very small and $\mbpi^1,\dots,\mbpi^V$ are easily enumerated. Therefore \eqref{eqn:BayesTestVertex} represents a significant simplification compared to \eqref{eqn:BayesTest}. However, enumeration becomes increasingly difficult in higher dimensions that would arise in hypothesis testing with more than two hypotheses
\section{Contamination Model Theory}
\label{sec:theory}
In this section we present
results that precisely characterize the possible solutions $(P_0, P_1, \pi_0, \pi_1)$ to the contamination model \eqref{eqn:Ptilde01}. These results generalize parallels in \cite{scott2013COLT} as discussed shortly.
First we recall some definitions from \cite{scott2013COLT}. For probability distributions $P$ and $Q$, define the maximal mixture proportion $\nu^\ast(P,Q)$ as
\begin{equation}\label{eqn:maxMixProp}
\nu^\ast(P,Q) = \max \{\alpha \in [0,1] : \exists \text{ probability distribution } S : P = \alpha Q + (1-\alpha)S\}.
\end{equation}
One way of interpreting $\nu^\ast(P,Q)$ is as the infimum of the ratio $p(x)/q(x)$ if $P$ and $Q$ have probability densities $p(x)$ and $q(x)$ \cite[Lem.~5]{scott2013arXiv}. From this it can be seen that $\nu^\ast(P,Q)$ is not necessarily symmetric. If $\nu^\ast(P,Q) = 0$, $P$ is said to be irreducible with respect to $Q$, and if $\nu^\ast(Q,P) = 0$ also, then $P$ and $Q$ are mutually irreducible. Many of the results in \cite{scott2013COLT} depend on the assumption that the true distributions $P_0$ and $P_1$ are mutually irreducible. This assumption is relaxed in the present paper.
The first result below relates maximal mixture proportions between $P_0$ and $P_1$ to mixed counterparts involving both pure and contaminated distributions.
\begin{lemma}\label{lem:maxMixProp}
Under condition \eqref{eqn:contamBound},
\begin{align*}
\nu^\ast(P_0, \tP_1) &= \frac{\nu^\ast(P_0, P_1)}{1 - \pi_1 + \pi_1 \nu^\ast(P_0, P_1)},\\
\nu^\ast(P_1, \tP_0) &= \frac{\nu^\ast(P_1, P_0)}{1 - \pi_0 + \pi_0 \nu^\ast(P_1, P_0)}.
\end{align*}
\end{lemma}
\begin{IEEEproof}
It is shown that a decomposition of $P_0$ in terms of $P_1$ and another distribution $Q$ implies a decomposition of $P_0$ in terms of $\tP_1$ and $Q$, and vice versa. Combining the implications yields the first equality in the lemma. The proof of the second equality is entirely analogous.
For the forward implication, let $\nu$ and $Q$ be such that
\begin{equation}\label{eqn:maxMixProp1}
P_0 = \nu P_1 + (1-\nu) Q,
\end{equation}
where $\nu \leq \nu^\ast(P_0, P_1)$ by definition \eqref{eqn:maxMixProp}. Given \eqref{eqn:contamBound}, \eqref{eqn:Ptilde1} can be solved for $P_1$ and the result substituted into \eqref{eqn:maxMixProp1} to yield
\begin{align}
P_0 &= \nu \left( \frac{1}{1-\pi_1} \tP_1 - \frac{\pi_1}{1-\pi_1} P_0 \right) + (1-\nu) Q,\nonumber\\
P_0 &= \frac{\nu}{1-\pi_1+\nu\pi_1} \tP_1 + \frac{(1-\pi_1)(1-\nu)}{1-\pi_1+\nu\pi_1} Q.\label{eqn:maxMixProp2}
\end{align}
Since the numerators in \eqref{eqn:maxMixProp2} are non-negative and their sum equals the denominator, \eqref{eqn:maxMixProp2} is a valid mixture decomposition of $P_0$ in terms of $\tP_1$ and $Q$. It follows from \eqref{eqn:maxMixProp} that
\begin{equation}\label{eqn:maxMixProp3}
\nu^\ast(P_0, \tP_1) \geq \frac{\nu}{1-\pi_1+\nu\pi_1}.
\end{equation}
Using the formula
\begin{equation}\label{eqn:linFracDeriv}
\frac{d}{dx} \frac{Ax + B}{Cx + D} = \frac{AD-BC}{(Cx + D)^2},
\end{equation}
it is seen that the right-hand side of \eqref{eqn:maxMixProp3} is increasing in $\nu$. Therefore the bound \eqref{eqn:maxMixProp3} is optimized at $\nu = \nu^\ast(P_0, P_1)$:
\begin{equation}\label{eqn:maxMixProp4}
\nu^\ast(P_0, \tP_1) \geq \frac{\nu^\ast(P_0, P_1)}{1 - \pi_1 + \pi_1 \nu^\ast(P_0, P_1)}.
\end{equation}
For the reverse implication, suppose that $P_0 = \nu \tP_1 + (1-\nu) Q$ for $\nu \leq \nu^\ast(P_0, \tP_1)$ and some $Q$. Substituting for $\tP_1$ using \eqref{eqn:Ptilde1} and re-solving for $P_0$ as above gives
\begin{equation}\label{eqn:maxMixProp5}
P_0 = \frac{\nu(1-\pi_1)}{1-\nu\pi_1} P_1 + \frac{1-\nu}{1-\nu\pi_1} Q,
\end{equation}
which is again a valid mixture decomposition with non-negative coefficients that sum to $1$. Furthermore, the coefficient in front of $P_1$ is increasing in $\nu$. The combination of \eqref{eqn:maxMixProp} and \eqref{eqn:maxMixProp5} with the maximizing choice $\nu = \nu^\ast(P_0, \tP_1)$ implies
\[
\nu^\ast(P_0, P_1) \geq \frac{(1-\pi_1) \nu^\ast(P_0, \tP_1)}{1-\pi_1 \nu^\ast(P_0, \tP_1)}.
\]
Solving the last inequality for $\nu^\ast(P_0, \tP_1)$ yields \eqref{eqn:maxMixProp4} but with the inequality reversed, completing the proof.
\end{IEEEproof}
Lemma~\ref{lem:maxMixProp} generalizes \cite[Lem.~3]{scott2013COLT}, which states that $\nu^\ast(P_0, \tP_1) = 0$
if and only if $\nu^\ast(P_0, P_1) = 0$,
and similarly for the second equation.
In the non-irreducible case, it can be seen that the maximal mixture proportion must increase with contamination according to the bounds below.
\begin{cor}
Under condition \eqref{eqn:contamBound},
\begin{align*}
\nu^\ast(P_0, P_1) &\leq \nu^\ast(P_0, \tP_1) \leq \frac{\nu^\ast(P_0, P_1)}{1 - \pi_1},\\
\nu^\ast(P_1, P_0) &\leq \nu^\ast(P_1, \tP_0) \leq \frac{\nu^\ast(P_1, P_0)}{1 - \pi_0}.
\end{align*}
Equality holds throughout the first line only if $\pi_1 = 0$ or $\nu^\ast(P_0, P_1) = 0$, and similarly for the second line.
\end{cor}
\begin{IEEEproof}
The left inequality in the first line follows from the first line of Lemma~\ref{lem:maxMixProp} by adding $\pi_1 (1 - \nu^\ast(P_0, P_1))$ to the denominator, while the second inequality in the first line follows from subtracting $\pi_1 \nu^\ast(P_0, P_1))$ from the denominator.
\end{IEEEproof}
Given condition \eqref{eqn:contamBound}, the contamination model \eqref{eqn:Ptilde01} has an equivalent representation as specified by \cite[Lem.~1]{scott2013COLT}:
\begin{subequations}\label{eqn:Ptilde01Alt}
\begin{align}
\tP_0 &= (1 - \tpi_0) P_0 + \tpi_0 \tP_1, \qquad \tpi_0 = \frac{\pi_0}{1-\pi_1} \in [0,1),\label{eqn:Ptilde0Alt}\\
\tP_1 &= (1-\tpi_1) P_1 + \tpi_1 \tP_0, \qquad \tpi_1 = \frac{\pi_1}{1-\pi_0} \in [0,1).\label{eqn:Ptilde1Alt}
\end{align}
\end{subequations}
This alternative form makes clear that once $(\tP_0, \tP_1)$ and the modified parameters $(\tpi_0, \tpi_1)$ (or equivalently $(\pi_0, \pi_1)$) are fixed, $(P_0, P_1)$ are also specified exactly.
Using \eqref{eqn:Ptilde01Alt}, \cite[Cor.~1]{scott2013COLT} shows that $\tpi_0$ and $\tpi_1$ are uniquely determined
under the irreducibility conditions $\nu^\ast(P_0, \tP_1) = \nu^\ast(P_1, \tP_0) = 0$. The next lemma provides general expressions for $\tpi_0$, $\tpi_1$ that do not require irreducibility.
\begin{lemma}\label{lem:pitilde01}
The contamination model \eqref{eqn:Ptilde01Alt} has a unique solution in $(\tpi_0, \tpi_1)$ in terms of maximal mixture proportions:
\begin{align*}
\tpi_0 &= \frac{\nu^\ast(\tP_0, \tP_1) - \nu^\ast(P_0, \tP_1)}{1 - \nu^\ast(P_0, \tP_1)},\\
\tpi_1 &= \frac{\nu^\ast(\tP_1, \tP_0) - \nu^\ast(P_1, \tP_0)}{1 - \nu^\ast(P_1, \tP_0)}.
\end{align*}
\end{lemma}
\begin{IEEEproof}
By \cite[Prop.~2]{scott2013COLT} (originally \cite[Prop.~5]{blanchard2010}), there exists a distribution $P_0'$ such that $\nu^\ast(P_0', \tP_1) = 0$ and
\begin{equation}\label{eqn:P0'1}
\tP_0 = (1 - \nu^\ast(\tP_0, \tP_1)) P_0' + \nu^\ast(\tP_0, \tP_1) \tP_1.
\end{equation}
(An explicit construction for $P_0'$ is given in the proof of \cite[Prop.~5]{blanchard2010}.) Combining \eqref{eqn:P0'1} with \eqref{eqn:Ptilde0Alt} and solving for $P_0$, we have
\begin{equation}\label{eqn:P0'2}
P_0 = \frac{1 - \nu^\ast(\tP_0, \tP_1)}{1 - \tpi_0} P_0' + \frac{\nu^\ast(\tP_0, \tP_1) - \tpi_0}{1 - \tpi_0} \tP_1,
\end{equation}
noting that $\tpi_0 < 1$. From definition \eqref{eqn:maxMixProp} and \eqref{eqn:Ptilde0Alt}, it is seen that both coefficients in \eqref{eqn:P0'2} are non-negative and sum to $1$. Hence \eqref{eqn:P0'2} is a valid mixture decomposition of $P_0$ into $P_0'$ and $\tP_1$. Furthermore, since $\nu^\ast(P_0', \tP_1) = 0$, we may apply \cite[Cor.~1]{scott2013COLT} to \eqref{eqn:P0'2} to obtain
\[
\frac{1 - \nu^\ast(\tP_0, \tP_1)}{1 - \tpi_0} = 1 - \nu^\ast(P_0, \tP_1).
\]
Solving for $\tpi_0$ results in the first line in the lemma statement. The expression for $\tpi_1$ is similarly obtained.
\end{IEEEproof}
Combining Lemmas~\ref{lem:maxMixProp} and \ref{lem:pitilde01} yields a characterization of the contamination proportions $\pi_0$, $\pi_1$.
\begin{theorem}\label{thm:pi01}
Under condition~\eqref{eqn:contamBound}, we have the relations
\begin{align*}
\pi_0 + \nu^\ast(\tP_0, \tP_1) \pi_1 &= \frac{\nu^\ast(\tP_0, \tP_1) - \nu^\ast(P_0, P_1)}{1 - \nu^\ast(P_0, P_1)},\\
\nu^\ast(\tP_1, \tP_0) \pi_0 + \pi_1 &= \frac{\nu^\ast(\tP_1, \tP_0) - \nu^\ast(P_1, P_0)}{1 - \nu^\ast(P_1, P_0)}.
\end{align*}
\end{theorem}
\begin{IEEEproof}
We substitute the first line of Lemma~\ref{lem:maxMixProp} into the first line of Lemma~\ref{lem:pitilde01} to obtain
\[
\tpi_0 = \frac{\bigl(1 - \pi_1 + \pi_1 \nu^\ast(P_0, P_1)\bigr) \nu^\ast(\tP_0, \tP_1) - \nu^\ast(P_0, P_1)}{1 - \pi_1 + \pi_1 \nu^\ast(P_0, P_1) - \nu^\ast(P_0, P_1)}.
\]
Using \eqref{eqn:Ptilde0Alt} and rearranging numerator and denominator,
\begin{align*}
\frac{\pi_0}{1-\pi_1} &= \frac{\nu^\ast(\tP_0, \tP_1) - \nu^\ast(P_0, P_1) - \pi_1 \nu^\ast(\tP_0, \tP_1) (1 - \nu^\ast(P_0, P_1))}{(1 - \pi_1)(1 - \nu^\ast(P_0, P_1))},\\
\pi_0 &= \frac{\nu^\ast(\tP_0, \tP_1) - \nu^\ast(P_0, P_1)}{1 - \nu^\ast(P_0, P_1)} - \pi_1 \nu^\ast(\tP_0, \tP_1),
\end{align*}
which is equivalent to the first relation in the theorem statement. The derivation of the second relation is again analogous.
\end{IEEEproof}
Since $\tP_0$, $\tP_1$ and hence $\nu^\ast(\tP_0, \tP_1)$, $\nu^\ast(\tP_1, \tP_0)$ are assumed to be known, Theorem~\ref{thm:pi01} can be interpreted as a system of equations relating $\pi_0$, $\pi_1$ to the maximal proportions $\nu^\ast(P_0, P_1)$, $\nu^\ast(P_1, P_0)$ for the pure distributions. If $\nu^\ast(\tP_0, \tP_1)$, $\nu^\ast(\tP_1, \tP_0) < 1$, i.e., if $\tP_0 \neq \tP_1$, then this system is invertible because the determinant $1 - \nu^\ast(\tP_0, \tP_1) \nu^\ast(\tP_1, \tP_0) > 0$, and Theorem~\ref{thm:pi01} describes a bijection.
Fig.~\ref{fig:feasiblepi} depicts the set of feasible $(\pi_0, \pi_1)$ values given the contaminated maximal proportions $\nu^\ast(\tP_0, \tP_1)$, $\nu^\ast(\tP_1, \tP_0)$. The solid outer lines correspond to the mutually irreducible case, namely $\nu^\ast(P_0, P_1) = \nu^\ast(P_1, P_0) = 0$ in Theorem~\ref{thm:pi01}, and the intersection of the lines is the solution characterized in \cite[Prop.~3]{scott2013COLT}. Theorem~\ref{thm:pi01} generalizes to the interior of the region by specifying solutions for nonzero values of $\nu^\ast(P_0, P_1)$, $\nu^\ast(P_1, P_0)$. In particular, the dashed lines in Fig.~\ref{fig:feasiblepi} are lines of constant $\nu^\ast(P_0, P_1)$ or $\nu^\ast(P_1, P_0)$ and are parallel to the boundary lines. This geometry is used in the next section to
describe uncertainty in $\pi_0$, $\pi_1$.
\begin{figure}
\centering
\includegraphics[width=0.4\linewidth]{constraints_figure}
\caption{Region of feasible contamination proportions $(\pi_0, \pi_1)$ given contaminated distributions $\tP_0$ and $\tP_1$.}
\label{fig:feasiblepi}
\end{figure}
|
\section{Related work}
\textit{K}-means is a classic clustering algorithm. This algorithm finds the closest cluster for each document by finding the smallest distance between the document and the existing clusters' centroids. The clusters' centroids are also updated at each iteration due to new cluster members. \textit{K}-means is based on the assumption that documents belonging to same topic should also be close to each other in the feature space. In a similar vein, Naive Bayes and Gaussian mixture model \cite{qin2012improving,baker1998distributional,liu2002document} are used based on different document distribution assumptions. One problem with these methods is that if the corpora properties don't following such assumptions, the performance of these algorithms may be at risk.
Latent Semantics Indexing (LSI) \cite{deerwester1990indexing} is the technique that converts the corpora from the original feature space into a latent semantics space. Each basis axis in the latent semantics space essentially represents one type of semantic information of the corpora. By doing so, each document is essentially a combination of multiple semantics information. Then we can apply the classic clustering algorithms on these new representations of documents in latent semantics space. One issue with this method is that the coefficients of the combination could be positive or negative. A negative coefficient is not such a natural way to interpret the document. Meanwhile, the bases that spanning the latent semantics space in some LSI algorithms, like Singular Vector Decomposition (SVD) \cite{furnas1988information}, are orthogonal, which means that every semantics bases are different from each other. However, in reality, this is not always the case.
Similar to LSI algorithms, NMF also maps the corpora into latent feature space. The differences are that: firstly, the bases in latent feature space don't need to be orthogonal. Also, each basis now corresponds to one topic of the corpora, which makes it very easy to determine the topic of document by simply choosing the largest component in the latent space. Meanwhile, every element in the two decomposed low-rank matrices are nonnegative. This additive combination makes it more natural to understand each document in an intuitive manner.
The benefits of using NMF in document clustering have been heavily investigated in many existed papers \cite{xu2003document,shahnaz2006document,lin2007projected,wang2012adaptive}. However, many of them are mainly targeting on minimizing the $l_2$ norm or KL divergence in the process of matrix decomposition. Correntropy-based decomposition methods have proved effective in many areas like cancer clustering \cite{wang2013non}, face recognition \cite{he2011maximum}, etc. Some other solutions can be found in \cite{sun2012unsupervised,cai2011graph}. However, we never find such technique in the document clustering research or be used for very high-dimension data with considerable number of clusters, which is another starting point of our work.
\section{Algorithm}
Assuming we have a matrix $X \in \mathbb{R}^{D \times N}$. NMF allows us to factorize $X$ into two nonnegative matrices $H \in \mathbb{R}^{D \times K}$ and $W \in \mathbb{R}^{K \times N}$, where the product $H*W$ approximates the original matrix X. Each column in $X$ is the feature vector of one document with $D$ elements. Thus, $X$ essentially represents the whole corpus with $N$ documents. Conventionally, we name $H$ as basis matrix that each column forms the basis vector of the semantic feature space, and $W$ as coefficient matrix. Hence, a document is further represented as the additive combination of weighted basis vectors in semantic space. $l_2$ norm and Kullback-Leibler (KL) divergence are two commonly-used measures of the similarity between original matrix $X$ and the product of $H$ and $W$. Based on different similarity measures, we are able to solve the factorization problem by minimizing the corresponding errors between $X$ and $H*W$.
In this paper, we propose a new method to quantify the NMF by maximizing correntropy criteria in document clustering. Correntropy measures the generalized similarity between two random variables. More precisely, it models the expected differences between two random variables after mapping through kernel function. Without knowing the joint distribution of $X$ and $Y$, we can simply estimate the expectation by taking average (shown in Equation~\ref{eq.correntropyEstimation}):
\begin{equation} \label{eq.correntropyEstimation}
\hat{V}_\sigma (x,y) = \frac{1}{D} \sum_{i=1}^D k_\sigma (x_i - y_i)
\end{equation}
where $k_\sigma (.)$ is the kernel function and $x_i$ and $y_i$ are the element in $X$ and $Y$, respectively.
Thus, instead of using $l_2$ distance or KL divergence, we try to find the basis matrix $H$ and coefficient matrix $W$, whose product $Y$ approximates $X$, by maximizing their correntropy on a feature-by-feature basis to allow for weighting each feature differently. For each feature, the kernel function can be calculated as:
\begin{equation}
k_\sigma \left(\sqrt{\sum_{n=1}^N (x_{dn} - \sum_{k=1}^K h_{dk} w_{kn})^2}\right)
\end{equation}
Hence, the correntropy maximization problem is expressed in Equation~\ref{eq.mccKernel}.
\begin{equation} \label{eq.mccKernel}
\max \limits_{h_{dk}>0,w_{kn}>0}\sum_{d=1}^D k_\sigma \left(\sqrt{\sum_{n=1}^N (x_{dn} - \sum_{k=1}^K h_{dk} w_{kn})^2}\right)
\end{equation}
To simplify the calculations without losing generality, we choose the Gaussian kernel function as $k_\sigma(.)$:
\begin{equation} \label{eq.gaussianKernel}
k_\sigma(x-y)=exp\left(-\gamma||x-y||^2\right)
\end{equation}
After substituting Equation~\ref{eq.gaussianKernel} back into Equation~\ref{eq.mccKernel}, the basis and coefficient matrices can be derived by solving:
\begin{equation}
\max \limits_{h_{dk}>0,w_{kn}>0} \sum_{d=1}^D exp \left(-\gamma \sum_{n=1}^N (x_{dn} - \sum_{k=1}^K h_{dk} w_{kn})^2\right)
\end{equation}
We introduce the convex conjugate function $\varphi(.)$ and auxiliary variables $\rho=[\rho_1,...,\rho_D]^\top$. Based on the theory of convex conjugate functions, the above optimization problem is equivalent to:
\begin{flalign}
&\max \limits_{H,W,\rho} \hat{F}(H,W,\rho) \nonumber \\
&s.t.\text{ } H \geq 0, W \geq 0& \\
&\hat{F}(H,W,\rho)=\sum_{d=1}^D \left(\rho_d \sum_{n=1}^N (x_{dn}-\sum_{k=1}^K h_{dk}w_{kn})^2-\varphi(\rho_d) \right)& \nonumber
\end{flalign}
The optimization problem can be solved by Expectation-Maximization-like method. Starting from the initial value of $H$ and $W$, we compute $\rho$ in expectation step (E-step). Conditional on the $\rho$ value, we update the $H$ and $W$ values in maximization step (M-step). The process is called one iteration. This iterative process stops until it converges. The proposed MCC has a good convergence performance. We direct the readers to refer similar convergence proof in \cite{wang2013non}. We often assign $H$ and $W$ with random numbers to start the algorithm if we have no prior information about the distribution of data.
\textbf{E-step}: Starting from the estimated $H$ and $W$ from last M-step (or random values in the 1st iteration), $\rho$ of the \textit{t-th} iteration is computed as:
\begin{flalign}
&\rho_d^t=-g\left(\sqrt{\sum_{n=1}^N\left(x_{dn}-\sum_{k=1}^K h_{dk}^t w_{kn}^t\right)}, \sigma^t \right)& \\
&\text{where} \text{ } g(z,\sigma)=\sup \limits_{\varrho \in \mathbb{R^-}}\left(\varrho\frac{||z||^2}{\sigma^2}-\varphi(\varrho)\right)& \nonumber \\
&\text{and} \text{ } \sigma^t = \sqrt{\frac{\theta}{2D} \sum_{d=1}^D \sum_{n=1}^N \left(x_{dn} - \sum_{k=1}^K h_{dk}^t w_{kn}^t \right)^2}& \nonumber
\end{flalign}
\textbf{M-step}: Conditional on the new $\rho$ from last step, we compute the new basis and coefficient matrix, denoted as $H^{t+1}$ and $W^{t+1}$ respectively, by maximizing the object function:
\begin{align}
&(H^{t+1},W^{t+1}) \nonumber \\
= &\arg\!\max_{H,W} \sum_{d=1}^D \left( \rho_d^t \sum_{n=1}^N \left(x_{dn}-\sum_{k=1}^K h_{dk} w_{kn} \right) ^2 \right) \nonumber \\
= &\arg\!\max_{H,W} Tr[(X-HW)^\top diag(\rho^t)(X-HW)] \nonumber \\
= &Tr[X^\top diag(-\rho^t) X]-2Tr[X^\top diag(-\rho^t) HW] \nonumber \\
&+Tr[W^\top H^\top diag(-\rho^t)HW] \nonumber \\
&s.t. H \geq 0, W \geq 0& \nonumber
\end{align}
where $Tr(.)$ means the trace of the matrix.
We apply the Lagrange method to solve the optimization problem. Let the elements of matrices $\Phi=[\phi_{dk}]$ and $\Psi=[\psi_{kn}]$ be the corresponding Lagrange multipliers for the nonnegative conditions of $h_{dk} \geq 0$ and $w_{kn} \geq 0$. Then we can express the Lagrange optimization problem as:
\begin{align}
L &= Trac[X^\top diag(-\rho^t) X]-2Trac[X^\top diag(-\rho^t) HW] \nonumber \\
&+Trac[W^\top H^\top diag(-\rho^t) HW] + Trac[\Phi H^\top] \nonumber \\
&+ Trac[\Psi H^\top]
\end{align}
The partial derivatives of $L$ with respect to $H$ and $W$ are:
\begin{align}
\frac{\partial L}{\partial H}&=-2diag(-\rho^t)XW^\top + 2diag(-\rho^t)HWW^\top+\Phi \nonumber \\
\frac{\partial L}{\partial W}&=-2H^\top diag(-\rho^t)X+2H^\top diag(-\rho^t)HW+\Psi \nonumber
\end{align}
Set them to 0 based on Karush-Kuhn-Tucker optimal conditions, we have:
\begin{align}
&-(diag(-\rho^t)XW^\top)_{dk}h_{dk}+(diag(-\rho^t)HWW^\top)_{dk}h_{dk}=0\nonumber \\
&-(H^\top diag(-\rho^t)X)_{kn}w_{kn}+(H^\top diag(-\rho^t)HW)_{kn}h_{kn}=0 \nonumber
\end{align}
Hence, the basis matrix $H$ and coefficient matrix $W$ can updated as shown in Equation~\ref{eq:updateH} and Equation~\ref{eq:updateW}.
\begin{equation} \label{eq:updateH}
h_{dk}^{t+1} \leftarrow h_{dk}^t \frac{(diag(-\rho^t)XW^{t^\top})_{dk}}{(diag(-\rho^t)H^tW^tW^{t^\top}_{dk})_{dk}}
\end{equation}
\begin{equation} \label{eq:updateW}
w_{kn}^{t+1} \leftarrow w_{kn}^t \frac{(H^{{t+1}^\top}diag(-\rho^t)X)_{kn}}{(H^{{t+1}^\top}diag(-\rho^t)H^{t+1}W^t)_{kn}}
\end{equation}
\section{Experiment settings}
We test the MCC algorithm on two datasets: Reuters21578 \footnote{http://kdd.ics.uci.edu/databases/reuters21578/reuters21578.html} and TDT2 \footnote{http://www.itl.nist.gov/iad/mig/tests/tdt/1998/}. These two datasets have been widely used in many places \cite{xu2003document,shahnaz2006document} for document clustering. Reuters21578 test collection contains 21578 documents from 135 topics in total. We exclude those documents that belong to more than 1 topics in our experiment since we are trying to cluster each document to one single topic. Meanwhile, we also exclude those topics with less than 5 documents. As a consequence, we use 9545 documents for 51 topics in our experiment. TDT2 dataset contains around 11201 documents for 96 topics. We also apply similar pre-processing ways to it. The largest 30 topics with 9394 documents are used.
We use tf-idf method to extract the feature for each document. Stopwords and stemming are applied. The number of elements for each document are 16777 and 36771 for Reuters21578 and TDT2, respectively.
After matrix decomposition, the matrix $W$ with dimension $K*N$ is essentially the new representation of the corpora in the way that each column is the feature vector of one document after dimension deduction. And the new dimension of the feature vector is $K$ now. To evaluate the decomposition performance, we directly apply \textit{K}-means clustering method to cluster $W$ into $K$ clusters. \textit{K}-means will assign each document with a label. We compare the label from \textit{K}-means to the original ground-truth label to calculate the clustering accuracy. The accuracy is defined in Equation~\ref{eq:accu}.
\begin{equation} \label{eq:accu}
Accuracy = \sum\limits_{i=1}^{N} \delta(kmeans\_label_i, topic_i)/N
\end{equation}
where $\delta(kmeans\_label_i, topic_i)$ is the delta function, which returns 1 if $kmeans\_label_i = topic_i$; otherwise 0. To find the correspondence between the topic from ground-truth data and the label by \textit{K}-means, we use Kuhn-Munkres algorithm \cite{kuhn1955hungarian}.
\section{Results}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{table*}[ht]
\centering
\caption{Clustering accuracies on Reuters21578 and TDT2 datasets} \label{tab.results}
\begin{tabular}{|c||c|c|c|c|c||c|c|c|c|c|}
\hline
\multirow{2}{*}{\tabincell{c}{Number of \\ clusters}} & \multicolumn{5}{|c||}{Reuters21578} & \multicolumn{5}{|c|}{TDT2}\\
\hhline{~----------}
& MCC & L2 & K-L & GS-CLS & PG & MCC & L2 & K-L & GS-CLS & PG \\ \hline
2 & 0.911 & 0.839 & 0.871 & 0.894 & 0.838 & 0.927 & 0.824 & 0.845 & 0.881 & 0.863 \\ \hline
3 & 0.907 & 0.671 & 0.813 & 0.866 & 0.769 & 0.886 & 0.761 & 0.754 & 0.788 & 0.740 \\ \hline
4 & 0.898 & 0.625 & 0.719 & 0.866 & 0.637 & 0.823 & 0.702 & 0.695 & 0.710 & 0.591 \\ \hline
5 & 0.890 & 0.596 & 0.675 & 0.864 & 0.585 & 0.758 & 0.683 & 0.616 & 0.666 & 0.574 \\ \hline
6 & 0.863 & 0.567 & 0.632 & 0.845 & 0.567 & 0.729 & 0.618 & 0.583 & 0.647 & 0.509 \\ \hline
7 & 0.816 & 0.515 & 0.610 & 0.795 & 0.513 & 0.698 & 0.544 & 0.574 & 0.640 & 0.422 \\ \hline
8 & 0.799 & 0.509 & 0.599 & 0.788 & 0.448 & 0.651 & 0.531 & 0.558 & 0.621 & 0.400\\ \hline
9 & 0.770 & 0.494 & 0.490 & 0.770 & 0.409 & 0.647 & 0.533 & 0.556 & 0.573 & 0.395 \\ \hline
10 & 0.736 & 0.403 & 0.457 & 0.719 & 0.415 & 0.631 & 0.468 & 0.518 & 0.555 & 0.382 \\ \hline
\end{tabular}
\end{table*}
One important parameter we need to control is the number of cluster $K$. Intuitively, the value of $K$ controls the way to decompose the matrix. More importantly, it determines the number of topics that NMF algorithm can handle. That's to say, if $K = 2$, then the NMF is a two-topic clustering problem. If $K>2$, then it's a multi-topic clustering problem. To fully investigate the efficiency of the MCC, we use the following candidate numbers of clusters: \{2-10, 20, 30, 40, 51\} for Reuters dataset and \{2-10, 15, 20, 25, 30\} for TDT2 corpus. We randomly select $K$ topics, and use all documents from those $K$ topics as the current run's testing documents. We repeat this process 20 times to reduce the potential effect of random errors on our experiment results since the performance of NMF algorithm is affected by the initial values of the iterative process. The average of 20 runs is used as the output of each algorithm.
Table~\ref{tab.results} summarizes the results of two datasets with less than or equal to 10 topics. Figure~\ref{fig:reuters} and Figure~\ref{fig:tdt2} demonstrate the accuracies on all chosen numbers of clusters. We firstly test the proposed MCC algorithm against two classic loss functions: $l_2$ distance and KL divergence. It's clear that MCC algorithm outperforms the $l_2$ distance and KL divergence in all cases of $K$ in two datasets. This shows the supremacy of the MCC algorithm against the others. One possible reason is that $l_2$ and KL distance are effective when dealing with linear separable data. However, if the data distribution is nonlinear manifold, it is considerably difficult for these two linear kernels to distinguish them.
Meanwhile, we observe that for all algorithms, the accuracy decreases as the number of clusters increases. Intuitively, more clusters inevitably increase the difficulties of finding the right label for each document. However, MCC is more robust to the increment of $K$, compared to other distance functions.
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\columnwidth]{reuters_all.eps}
\caption{Accuracies on Reuters21578.}
\label{fig:reuters}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\columnwidth]{tdt2_all.eps}
\caption{Accuracies on TDT2.}
\label{fig:tdt2}
\end{figure}
We also compare MCC against two variants of NMF algorithms: gradient descent-constrained least squares (GSCLS) \cite{shahnaz2006document}, and Projected Gradient nonnegative matrix factorization (PG) \cite{lin2007projected}. Based on the results of two datasets, we can see that MCC suppresses the rest NMF algorithms when the number of clusters is smaller or equals to 10 on Reuters21578. When it comes to TDT2 dataset, MCC achieves the best performance in all cases, which shows the benefit of introducing the correntropy into the factorization process. One potential reason is that MCC can self-learn different kernels for different features. This adaptive learning property somehow further improves the performance of MCC when facing with nonlinear datasets (e.g. document collection).
\section{Conclusion}
\vspace{-4.1mm}
In this paper, we propose a new method to decompose the matrix into two low-rank matrices by maximizing the correntropy between them, such that we can easily and effectively use the decomposed matrix to cluster high-dimension data. We test the proposed MCC algorithm in the application of document clustering. We compare our proposed method to other loss functions and NMF algorithms. The results demonstrate the supremacy of our method on Reuters21578 and TDT2 corpora in terms of accuracy. In future, we will investigate the possibility of our proposed method in medical instrument\cite{anderson2013non}, mechanical instrument\cite{cui2012interpreting} and other related areas \cite{yang2014fairness,yang2013broadcasting,gao2012facial}.
\cite{zhou2013adaptive,li2014graph}
\section*{Acknowledgements}
This work was partially supported by National Natural Science Foundation of China under Grant No.61273217, 61175011 and 61171193, the 111 project under Grant No.B08004.
\nocite{ex1,ex2}
\bibliographystyle{latex8}
|
\section{Introduction}
The comparison principle for differential equations tells us whether two solutions starting from two distinct initial conditions can compare with each other when the initial
conditions are comparable.
The {\it sample-path comparison principle} for stochastic differential equations (SDEs) and
also for stochastic partial differential equations (SPDEs) have been studied extensively;
see e.g. \cite[Chapter VI]{IkedaWatanabe89} and \cite[Chapter V. 40]{RogerWilliams00Vol2} for SDEs,
and \cite{Assing99Comparison,Kotelenez92Comparison,Milian02Comparison,Mueller91Support,Shiga94Two} for SPDEs.
A related problem is the {\it stochastic comparison principle}, which is of the form:
\[
\mathbb{E}\left(\Phi(u_t\right))\le \mathbb{E}\left(\Phi(v_t)\right),\quad\text{for all $t>0$,}
\]
where $\{u_t(x)\}$ and $\{v_t(x)\}$ solve SDEs or SPDEs, with the same initial data but comparable drift and diffusion coefficients. One looks for as large a class of functions $\Phi$ as
possible. See \cite{CoxFleischmannGreven96Comparison,Hajek85Mean,JackaTribe03Comparison,JosephDavarMueller14Strong}.
In this paper, we will focus on the pathwise comparison principle for
the following nonlinear stochastic fractional heat equation:
\begin{align}\label{E:FracHt}
\begin{cases}
\left(\displaystyle\frac{\partial}{\partial t} - {}_x D_\delta^a \right) u(t,x) =
\rho\left(u(t,x)\right) \dot{W}(t,x),& t\in \mathbb{R}_+^*:=\;]0,+\infty[\;,\: x\in\mathbb{R}\cr
u(0,\cdot) = \mu(\cdot)
\end{cases}
\end{align}
where $a \in \;]1,2]$ is the order of the fractional differential
operator ${}_x D_\delta^a$ and $\delta$ ($|\delta|\le 2-a$) is its
skewness, $\dot{W}$ is the space-time white noise on $\mathbb{R}_+\times\mathbb{R}$, $\mu$ denotes the
initial data (a measure), and the function $\rho:\mathbb{R}\mapsto \mathbb{R}$ is Lipschitz
continuous. Throughout this paper, we assume that $a$ and $\delta$ are fixed constants such that
\begin{align}\label{E:aDelta}
a\in \;]1,2] \quad\text{and}\quad|\delta| \leq 2-a,
\end{align}
unless we state otherwise (see Corollary
\ref{C:FI}).
When $a=2$ and $\delta=0$, the fractional operator ${}_x D_\delta^a$ reduces to the Laplacian on $\mathbb{R}$,
which is the infinitesimal operator for a Brownian motion.
On the other hand, when $a\in\;]1,2[\;$ and $|\delta|\leq 2-a$,
the operator ${}_x D_\delta^a$ is the infinitesimal generator of an $a$-stable process with skewness $\delta$.
In particular, $\lMr{x}{D}{0}^a = -(-\Delta)^{a/2}$.
This fractional Laplace operator has been paid many attentions for several decades because of its non-local property, and thus it is widely used in many areas such as physics, biology, and finance to
model non-local (anomalous) diffusions.
We refer to \cite{MainardiEtc01Fundamental,UchaikinZolotarev99,Zolotarev86} for more details on these fractional operator and the related stable random variables.
The existence and uniqueness of a random field solution to \eqref{E:FracHt} have been studied
in \cite{ChenDalang13Heat,ChenDalang14FracHeat,ConusEtc11InitialMeasure, Debbi06Explicit,DebbiDozzi05On,FoondunKhoshnevisan08Intermittence}.
In particular, the existence, uniqueness, and moment estimates under measure-valued initial data have been established recently in
\cite{ChenDalang13Heat,ChenDalang14FracHeat,ConusEtc11InitialMeasure}.
We now specify the \emph{weak} and \emph{strong} comparison principles. Let $u_1(t,x)$ and $u_2(t,x)$ be two solutions to \eqref{E:FracHt} with initial measures $\mu_1$ and $\mu_2$,
respectively.
We say that \eqref{E:FracHt} satisfies the {\it weak comparison principle} if
$u_1(t,x)\le u_2(t,x)$ for all $t>0$ and $x\in\mathbb{R}$, a.s., whenever $\mu_1\le \mu_2$ (i.e., $\mu_2-\mu_1$ is a nonnegative measure).
And the equation \eqref{E:FracHt} is said to satisfy the {\it strong comparison principle} if $u_1(t,x)<u_2(t,x)$ for all $t>0$ and $x\in\mathbb{R}$, a.s., whenever $\mu_1<\mu_2$ (i.e., the
measure $\mu_2-\mu_1$ is nonnegative and nonvanishing).
Note that the stochastic comparison principle for \eqref{E:FracHt} with $\Phi(z)=|z|^k$ for $k \geq2$ (so-called the \emph{moment} comparison principle) has been shown
lately by Joseph, Khoshnevisan and Mueller
\cite{JosephDavarMueller14Strong}.
When $a=2$, the equation \eqref{E:FracHt} reduces to the stochastic heat equation (SHE).
The special case when $\rho(u)=\lambda u$ for some constant $\lambda \ne 0$ is called
the {\it parabolic Anderson model}; see
\cite{BertiniCancrini94Intermittence,CarmonaMolchanov94,FoondunKhoshnevisan08Intermittence}.
The weak comparison principle can be derived readily from the Feynman-Kac formula; see \cite{BertiniCancrini94Intermittence}.
But the proof of the strong comparison principle requires some more efforts.
This question is important because, e.g., the {\em Hopf-Cole solution} to the famous {\em Kardar-Parisi-Zhang equation (KPZ) } \cite{KPZ86} is the logarithm of the solution to the SHE.
For general $\rho$ which is Lipschitz continuous, we do not have the Feynman-Kac formula.
The weak comparison principle is no longer obvious.
In this case,
Mueller \cite{Mueller91Support} proves the strong comparison principle for the SHE on $\mathbb{R}$
for the initial data being absolutely continuous with respect to the Lebesgue measure with a bounded density function. Mueller uses the discrete Laplacian and discretizes time to approximate the
solution to the SHE, which results in the weak comparison principle.
He then obtains the strong comparison principle by employing some large deviation estimates for the stochastic integral part of the solution.
Using Mueller's large deviation estimates, Shiga \cite{Shiga94Two} gives another proof of the strong comparison principle for the initial data being a so-called
$C_{\text{tem}}$ function, that is, a continuous function with both
tails growing no faster than $e^{\lambda|x|}$ for all $\lambda>0$.
He outlines a different approach for proving the weak comparison principle in the
appendix of his paper: smooth both the Laplace operator and the white noise so that one can apply the comparison principle for SDEs. We will follow his approach in our proof for the
weak comparison principle.
In both Mueller \cite{Mueller91Support} and Shiga \cite{Shiga94Two},
the initial data should be functions.
One natural question is whether the solution remains strictly positive if we run the system \eqref{E:FracHt} starting from a measure, such as the Dirac delta measure.
Using the polymer model and following a convergence result by Alberts, Khanin and Questel \cite{AlbertEtc11},
Moreno Flores \cite{Moreno14Pos} recently proved the strict positivity result for
the Anderson model (i.e., the case where $a=2$ and $\rho(u)=\lambda u$) with the delta initial data.
Our results below generalize their result to the stochastic fractional heat equation (i.e., $a\in \:]1,2]$), and moreover,
we consider general measure-valued initial data and allow $\rho$ to be any Lipschitz continuous function.
Recently, Conus, Joseph and Khoshnevisan \cite{ConusEct12Corr} give a more precise estimate on the strong comparison principle for the SHE. When the initial data is the Lebesgue measure, they prove
that for every $t>0$, there exist two finite constants $A>0$ and $B>0$ such that for all $\epsilon\in \:]0,1[\:$ and $x\in\mathbb{R}$,
\begin{align}\label{E:R-CJK}
P\left(u(t,x)<\epsilon\right)\le A \exp\left(-B \left[|\log(\epsilon)| \cdot\log\left(|\log(\epsilon)|\right)\right]^{3/2}\right).
\end{align}
Clearly, this result implies the strong comparison principle.
In \cite{MuellerNualart08}, Mueller and Nualart prove that when $a=2$ and the space domain is $[0,1]$ with the zero Dirichlet boundary condition, for some constants $C_0$ and $C_1$,
\begin{align}\label{E:R-MN}
P(u(t,x)<\epsilon)\le C_0\exp\left(-C_1 |\log\epsilon|^{3/2-\epsilon}\right).
\end{align}
We will generalize these results to the stochastic fractional heat equation \eqref{E:FracHt}
following \cite{ConusEct12Corr}.
This shows how close to zero the solution to \eqref{E:FracHt} can be.
In order to state our results, we need some notation. Let
$\mathcal{M}\left(\mathbb{R}\right)$ be the set of signed (regular) Borel measures on $\mathbb{R}$.
From the Jordan
decomposition, $\mu=\mu_+-\mu_-$ such that $\mu_\pm$ are two non-negative Borel
measures with disjoint support and denote $|\mu|=\mu_++\mu_-$.
As proved in \cite{ChenDalang14FracHeat}, the admissible initial data for \eqref{E:FracHt} is
\[
\mathcal{M}_a\left(\mathbb{R}\right):=\left\{
\mu\in\mathcal{M}(\mathbb{R}):\; \sup_{y\in\mathbb{R}}\int_\mathbb{R} |\mu|(\ensuremath{\mathrm{d}} x) \frac{1}{1+|y-x|^{1+a}}<+\infty
\right\},\quad\text{for $a\in \:]1,2]$.}
\]
Moreover, when $a=2$, the admissible initial data can be more general than $\mathcal{M}_2(\mathbb{R})$: It can be any measures from the following set
\[
\mathcal{M}_H(\mathbb{R}) := \left\{\mu\in\mathcal{M}(\mathbb{R})\: :
\int_\mathbb{R} e^{-c x^2} |\mu|(\ensuremath{\mathrm{d}} x)<+\infty,\: \text{for all $c>0$}\right\};
\]
see \cite{ChenDalang13Heat}.
Clearly, $\mathcal{M}_a(\mathbb{R})\subseteq \mathcal{M}_H(\mathbb{R})$.
In the following, a ``$+$" sign in the subscript means the subset of nonnegative
measures.
An important example in $\mathcal{M}_{a,+}\left(\mathbb{R}\right)$ is the Dirac delta
measure.
For simplicity, denote
\[
\mathcal{M}_a^*(\mathbb{R}):=\begin{cases}
\mathcal{M}_a(\mathbb{R})& \text{if $1<a<2$,}\cr
\mathcal{M}_H(\mathbb{R})& \text{if $a=2$}.
\end{cases}
\]
We will follow Shiga's arguments \cite{Shiga94Two} to prove the following weak comparison principle.
This result allows more general initial conditions than those in \cite{Mueller91Support} and \cite{Shiga94Two}.
\begin{theorem}[Weak comparison principle]
\label{T:WComp}
Let $u_1(t,x)$ and $u_2(t,x)$ be two solutions to \eqref{E:FracHt} with the initial data
$\mu_1$ and $\mu_2\in \mathcal{M}_{a}^*(\mathbb{R})$, respectively.
If $\mu_1\le \mu_2$, then
\begin{align}\label{E:Com}
P\left(u_1(t,x)\le u_2(t,x),\;\text{for all $t\ge 0$ and $x\in\mathbb{R}$}\right) =1.
\end{align}
\end{theorem}
Here is one example. Let $\delta_{z}$ be the Dirac delta function with unit mass at $x=z$.
Suppose that $\mu_1=\delta_0$ and $\mu_2=2 \delta_0$. Then $u_1(t,x)\le u_2(t,x)$ for all $t>0$ and $x\in\mathbb{R}$, a.s.
As a direct consequence of Theorem \ref{T:WComp}, one can turn {\it weak intermittency} statements
in \cite{ChenDalang14FracHeat, FoondunKhoshnevisan08Intermittence} into the {\it full intermittency}.
More precisely, define the {\it upper and lower Lyapunov exponents of order $p$} by
\begin{align}\label{E:Lypnv-x}
\overline{m}_p(x) :=\mathop{\lim\sup}_{t\rightarrow+\infty}
\frac{1}{t}\log\mathbb{E}\left(|u(t,x)|^p\right),\quad
\underline{m}_p(x) :=\mathop{\lim\inf}_{t\rightarrow+\infty}
\frac{1}{t}\log\mathbb{E}\left(|u(t,x)|^p\right),
\end{align}
for all $p\ge 2$ and $x\in\mathbb{R}$.
According to Carmona and Molchanov \cite[Definition III.1.1, on p. 55]{CarmonaMolchanov94},
$u$ is {\it fully intermittent} if $\inf_{x\in\mathbb{R}} \underline{m}_2(x)>0$ and $m_1(x)\equiv 0$ for all $x\in\mathbb{R}$.
\begin{corollary}\label{C:FI}
Suppose that $a\in\:]1,2[\:$, $|\delta|< 2-a$ (strict inequality), $\mu\in\mathcal{M}_{a,+}(\mathbb{R})$, and $\rho$ satisfies that for some constants $\lip_\rho>0$ and $\vip\ge 0$, $\rho(x)^2\ge
\lip_\rho^2\left(\vip^2+x^2 \right)$ for all $x\in\mathbb{R}$.
If either $\mu\ne 0$ or $\vip\ne 0$, then the solution to \eqref{E:FracHt} is fully intermittent.
\end{corollary}
\begin{proof}
By \cite[Theorem 3.4]{ChenDalang14FracHeat}, $\inf_{x\in\mathbb{R}} \underline{m}_2(x)>0$.
By Theorem \ref{T:WComp}, $u(t,x)\ge 0$ a.s. and so, $\mathbb{E}[|u(t,x)|]=\mathbb{E}[u(t,x)]=J_0(t,x)$ (see \eqref{E:J0}).
Therefore, $m_1(x)\equiv 0$ for all $x\in\mathbb{R}$.
\end{proof}
We adapt both Mueller and Shiga's arguments (see \cite{Mueller91Support, Shiga94Two}) to prove the strong comparison principle.
In the proof, following the idea of \cite[Theorem 5.1]{ConusEct12Corr}, we develop a large deviation result similar to \cite{Mueller91Support} using the Kolmogorov continuity theorem.
\begin{theorem}[Strong comparison principle]\label{T:SComp}
Let $u_1(t,x)$ and $u_2(t,x)$ be two solutions to \eqref{E:FracHt} with the initial data
$\mu_1$ and $\mu_2\in \mathcal{M}_{a}^*(\mathbb{R})$, respectively.
If $\mu_1< \mu_2$, then
\[
P\left(u_1(t,x)<u_2(t,x)\;\;\text{for all $t>0$ and $x\in\mathbb{R}$}\right) =1.
\]
\end{theorem}
The following theorem gives more precise information on the positivity of the solutions.
Let $\spt{f}$ denote the support of function $f$, i.e., $\spt{f}:=\left\{x\in\mathbb{R}: f(x)\ne 0\right\}$.
\begin{theorem}[Strict positivity]\label{T1:Rates}
Suppose $\rho(0)=0$ and let $u(t,x)$ be the solution to \eqref{E:FracHt} with the initial data $\mu\in\mathcal{M}_a^*(\mathbb{R})$. Then we have the following two statements:\\
(1) If $\mu>0$, then for any compact set $K\subseteq \mathbb{R}_+^*\times\mathbb{R}$,
there exist finite constants $A>0$ and $B>0$ which only depend on $K$ such that for small enough $\epsilon>0$,
\begin{align}\label{E:Rates1}
P\left(\inf_{(t,x)\in K}u(t,x) <\epsilon \right)&\le A \exp\left(-B |\log(\epsilon)|^{1-1/a} \log\left(|\log(\epsilon)|\right)^{2-1/a}\right).
\end{align}
(2) If $\mu(\ensuremath{\mathrm{d}} x)=f(x)\ensuremath{\mathrm{d}} x$ with $f\in C(\mathbb{R})$, $f(x)\ge 0$ for all $x\in\mathbb{R}$ and $\spt{f}\neq \emptyset$,
then for any compact set $D\subseteq\spt{f}$ and any $T>0$,
there exist finite constants $A>0$ and $B>0$ which only depend on $D$ and $T$ such that for all small enough $\epsilon>0$,
\begin{align}
P\left(\inf_{(t,x)\in \: ]0,T]\times D}u(t,x) <\epsilon\right)\le A \exp\left(-B \left\{ |\log(\epsilon)|\cdot \log\left(|\log(\epsilon)|\right)\right\}^{2-1/a}\right).
\end{align}
\end{theorem}
Theorem \ref{T1:Rates} shows that for all $t>0$, the function $x\mapsto u(t,x)$ does not have a compact support (see \cite{Mueller91Support} and
\cite[Section 6.3]{Mueller09Tools} for some other scenarios where the compact support property can be preserved). We also note that thanks to Theorem \ref{T1:Rates}, one can regard the solution $u(t,
x)$ to (1.1) as the density at location $x$ of a continuous particle system at time $t$, where particles move as independent $a$-stable processes but branch independently according to the noise term;
see \cite{JosephDavarMueller14Strong}.
Furthermore, Theorem \ref{T1:Rates} implies that for all compact sets
$K\subseteq\mathbb{R}_+^*\times\mathbb{R}$ and all $p>0$, the negative moments exists: $\mathbb{E}\left[|\inf_{(t,x)\in K} u(t,x)|^{-p}\right]<\infty$.
Note that part (2) of Theorem \ref{T1:Rates} gives essentially the same rate as those in \eqref{E:R-CJK} and \eqref{E:R-MN} when $a=2$.
\bigskip
The next three theorems will be used in the proofs of the above theorems.
Since they are interesting by themselves, we list them below.
The first one, which is used in the proof of Theorem \ref{T:WComp}, says that we can approximate a solution to
\eqref{E:FracHt} starting from $\mu \in \mathcal{M}_a^*(\mathbb{R})$ by a solution to \eqref{E:FracHt} starting from smooth initial conditions.
Define
\begin{align}\label{E:psi}
\psi_{\epsilon}(x)=
\begin{cases}
1&\text{if $|x|\le 1/\epsilon$},\cr
1+1/\epsilon-|x| &\text{if $1/\epsilon\le |x|\le 1+1/\epsilon$},\cr
0 &\text{if $|x|\ge 1+1/\epsilon$}.
\end{cases}
\end{align}
\begin{theorem}\label{T:Approx}
Suppose that $\mu\in\mathcal{M}_a^*(\mathbb{R})$.
Let $u(t,x)$ and $u_{\epsilon}(t,x)$ be the solutions to \eqref{E:FracHt} starting from
$\mu$ and
$((\mu\:\psi_\epsilon)*\lMr{\delta}{G}{a}( \epsilon,\cdot))(x)$, respectively. Then
\[
\lim_{\epsilon\rightarrow 0}\mathbb{E}\left[|u(t,x)-u_{\epsilon}(t,x)|^2\right] =0,\quad \text{for all $t>0$ and $x\in\mathbb{R}$.}
\]
\end{theorem}
The following theorem, which is used in the proof of Theorem \ref{T:SComp}, shows the
sample-path regularity for the solutions to \eqref{E:FracHt}.
When the initial data has a bounded density, this has been proved in \cite{DebbiDozzi05On}.
For general initial data, the case where $a=2$ is proved in \cite{ChenDalang13Holder}.
The theorem below covers the cases where $1<a<2$.
We need some notation: Given a subset $K\subseteq \mathbb{R}_+\times\mathbb{R}$ and positive constants
$\beta_1,\beta_2$, denote by $C_{\beta_1,\beta_2}(K)$ the set of functions $v:
\mathbb{R}_+ \times \mathbb{R} \mapsto \mathbb{R}$ with the property that for each compact set
$D \subseteq K$, there is a finite constant $C$ such that for all $(t,x)$
and $(s,y)$ in $D$,
\[
\vert v(t,x) - v(s,y) \vert \leq C \left[\vert t-s\vert^{\beta_1} + \vert x-y \vert^{\beta_2}\right].
\]
Denote
\[
C_{\beta_1-,\beta_2-}(D) := \cap_{\alpha_1\in \;\left]0,\beta_1\right[} \cap_{\alpha_2\in \;\left]0,\beta_2\right[} C_{\alpha_1,\alpha_2}(D)\;.
\]
\begin{theorem}\label{T:Holder}
Let $u(t,x)$ be the solution to \eqref{E:FracHt} starting from $\mu\in\mathcal{M}_a^*(\mathbb{R})$.
Then we have
\[
u\in
C_{\frac{a-1}{2a}-,\frac{a-1}{2}-}\left(\mathbb{R}_+^*\times\mathbb{R}\right),\;\;\text{a.s.}
\]
\end{theorem}
The last one, which is also used in the proof of Theorem \ref{T:SComp},
shows that the solution $u(t,x)$ to \eqref{E:FracHt} converges to the initial measure $\mu$
in the weak sense as $t\rightarrow 0$.
The case when $a=2$ is proved in \cite[Proposition 3.4]{ChenDalang13Holder}.
Let $C_c(\mathbb{R})$ be the set of continuous functions with compact support
and $\InPrd{f,g}$ be the inner product in $L^2(\mathbb{R})$.
\begin{theorem}\label{T:WeakSol}
Let $u(t,x)$ be the solution to \eqref{E:FracHt} starting from $\mu\in\mathcal{M}_a^*(\mathbb{R})$. Then,
\[
\lim_{t\rightarrow 0}\InPrd{u(t,\circ),\phi} \stackrel{L^2(\Omega)}{=} \InPrd{\mu,\phi}\quad\text{for all $\phi\in C_c(\mathbb{R})$.}
\]
\end{theorem}
In the following,
we first list some notation and preliminary results in Section \ref{S:Pre}.
Then we prove Theorem \ref{T:WComp} in Section \ref{S:WComp} with many
technical lemmas proved in the Appendix.
The proof of Theorem \ref{T:SComp} is presented in Section \ref{S:Scomp},
Theorem \ref{T1:Rates} is proved in Section \ref{S:Rates}.
Finally, the three Theorems \ref{T:Approx}, \ref{T:Holder} and \ref{T:WeakSol} are proved in
Sections \ref{S:Approx}, \ref{S:Holder}, and \ref{S:WeakSol}, respectively.
\section{Notation and some preliminaries}\label{S:Pre}
The Green function associated to the
problem
\eqref{E:FracHt} is
\begin{align}\label{E:Green}
\lMr{\delta}{G}{a}(t,x) := \mathcal{F}^{-1}
\left[\exp\left\{\lMr{\delta}{\psi}{a}(\cdot) t\right\} \right](x)
= \frac{1}{2\pi}\int_\mathbb{R} \ensuremath{\mathrm{d}} \xi\: \exp\left\{i \xi x - t |\xi|^a
e^{-i \delta \pi\: \text{sgn}(\xi)/2}\right\},
\end{align}
where $\mathcal{F}^{-1}$ is the inverse Fourier transform and
\[
\lMr{\delta}{\psi}{a}(\xi) = -|\xi|^a e^{-i \delta \pi\:
\text{sgn}(\xi)/2}\;.
\]
Denote the solution to the homogeneous equation
\begin{align}\label{E:FracHt-H}
\begin{cases}
\left(\displaystyle\frac{\partial}{\partial t} - {}_x D_\delta^a \right) u(t,x) = 0,&
t\in \mathbb{R}_+^*\;,\: x\in\mathbb{R},\cr
u(0,\cdot) = \mu(\cdot),
\end{cases}
\end{align}
by
\begin{align}\label{E:J0}
J_0(t,x) := \left(\lMr{\delta}{G}{a}(t,\cdot)*\mu\right)(x)
=\int_\mathbb{R} \mu(\ensuremath{\mathrm{d}} y)\:\lMr{\delta}{G}{a}(t,x-y),
\end{align}
where ``$*$'' denotes the convolution in the space variable.
Following notation in \cite{ChenDalang13Heat}, let $W=\left\{
W_t(A),\, A\in\mathcal{B}_b(\mathbb{R}),\, t\ge 0 \right\}$
be a space-time white noise
defined on a probability space $(\Omega,\mathcal{F},P)$, where
$\mathcal{B}_b\left(\mathbb{R}\right)$ is the
collection of Borel sets with finite Lebesgue measure.
Let $(\mathcal{F}^0_t,\, t\ge 0)$ be the natural filtration generated by $W$ and augmented by the $\sigma$-field $\mathcal{N}$ generated
by all $P$-null sets in $\mathcal{F}$:
\[
\mathcal{F}_t^0 = \sigma\left(W_s(A):0\le s\le
t,A\in\mathcal{B}_b\left(\mathbb{R}\right)\right)\vee
\mathcal{N},\quad t\ge 0.
\]
Define $\mathcal{F}_t := \mathcal{F}_{t+}^0 = \wedge_{s>t}\mathcal{F}_s^0$ for $t\ge 0$.
In the following, we fix this filtered
probability space $\left\{\Omega,\mathcal{F},\{\mathcal{F}_t:t\ge0\},P\right\}$.
We use $\Norm{\cdot}_p$ to denote the
$L^p(\Omega)$-norm ($p\ge 1$).
The rigorous meaning of the SPDE \eqref{E:FracHt} is the integral (mild) form
\begin{equation}
\label{E:WalshSI}
\begin{aligned}
u(t,x) &= J_0(t,x)+I(t,x),\quad\text{where}\cr
I(t,x) &=\iint_{[0,t]\times\mathbb{R}}
\lMr{\delta}{G}{a}\left(t-s,x-y\right)\rho\left(u(s,y)\right)W(\ensuremath{\mathrm{d}} s,\ensuremath{\mathrm{d}}
y),
\end{aligned}
\end{equation}
where the stochastic integral is the Walsh integral \cite{Walsh86}.
\begin{definition}\label{DF:Solution}
A process $u=\left(u(t,x),\:(t,x)\in\mathbb{R}_+^*\times\mathbb{R}\right)$ is called a {\it
random field solution} to
\eqref{E:FracHt} if:
\begin{enumerate}[(1)]
\item $u$ is adapted, i.e., for all $(t,x)\in\mathbb{R}_+^*\times\mathbb{R}$, $u(t,x)$ is
$\mathcal{F}_t$-measurable;
\item $u$ is jointly measurable with respect to
$\mathcal{B}\left(\mathbb{R}_+^*\times\mathbb{R}\right)\times\mathcal{F}$;
\item $\left(\lMr{\delta}{G}{a}^2 \star \Norm{\rho(u)}_2^2\right)(t,x)<+\infty$
for all $(t,x)\in\mathbb{R}_+^*\times\mathbb{R}$,
where ``$\star$'' denotes the simultaneous
convolution in both space and time
variables. Moreover, the function
$(t,x)\mapsto I(t,x)$ mapping $\mathbb{R}_+^*\times\mathbb{R}$ into
$L^2(\Omega)$ is continuous;
\item $u$ satisfies \eqref{E:WalshSI} a.s.,
for all $(t,x)\in\mathbb{R}_+^*\times\mathbb{R}$.
\end{enumerate}
\end{definition}
Throughout the paper, we assume that the function $\rho:\mathbb{R}\mapsto \mathbb{R}$ is Lipschitz
continuous with Lipschitz constant $\LIP_\rho>0$, and moreover, for some constants $\Lip_\rho>0$ and $\Vip \ge 0$,
\begin{align}\label{E:LinGrow}
|\rho(x)|^2 \le \Lip_\rho^2 \left(\Vip^2 +x^2\right),\qquad \text{for
all $x\in\mathbb{R}$}.
\end{align}
Note that the above growth condition \eqref{E:LinGrow} is a consequence of $\rho$ being Lipschitz continuous.
Let $a^*$ be the dual of $a$, i.e., $1/a+1/a^*=1$.
The following constant is finite:
\begin{align}\label{E:Cst-dLa}
\Lambda=\lMr{\delta}{\Lambda}{a} := \sup_{x\in\mathbb{R}} \lMr{\delta}{G}{a}(1,x)\;,
\end{align}
and in particular, $\lMr{0}{\Lambda}{a}=
\pi^{-1}\Gamma\left(1+1/a\right)$; see \cite[(3.10)]{ChenDalang14FracHeat}.
In the following, we often omit the dependence of this constant on $\delta$ and
$a$ and simply write $\lMr{\delta}{\Lambda}{a}$ as $\Lambda$.
This rule will also apply to other constants.
For all $(t,x)\in\mathbb{R}_+^*\times \mathbb{R}$, $n\in\mathbb{N}$ and $\lambda\in\mathbb{R}$, define
\begin{align}
\notag
\mathcal{L}_0\left(t,x;\lambda\right) &:= \lambda^2 \lMr{\delta}{G}{a}^2(t,x) \\
\label{E:Ln}
\mathcal{L}_n\left(t,x;\lambda\right)&:=
\underbrace{\left(\mathcal{L}_0\star \cdots\star\mathcal{L}_0\right)}_{\text{$n+1$
factors } \mathcal{L}_0(\cdot,\circ;\lambda)}
,\quad\text{for $n\ge 1$,}\\
\mathcal{K}\left(t,x;\lambda\right)&:= \sum_{n=0}^\infty
\mathcal{L}_n\left(t,x;\lambda\right).
\label{E:K}
\end{align}
We apply the following conventions to $\mathcal{K}(t,x;\lambda)$:
\begin{align*}
\mathcal{K}(t,x) := \mathcal{K}(t,x;\lambda),\quad
\overline{\mathcal{K}}(t,x) := \mathcal{K}\left(t,x;\Lip_\rho\right),\quad
\widehat{\mathcal{K}}_p(t,x) := \mathcal{K}\left(t,x;4\sqrt{p} \Lip_\rho\right),\quad\text{for $p\ge 2$}.
\end{align*}
The following theorem is from \cite[Theorem 3.1]{ChenDalang14FracHeat} for $1<a<2$ and \cite[Theorem 2.4]{ChenDalang13Heat} for $a=2$.
\begin{theorem}[Existence,uniqueness and moments]\label{T:ExUni}
Suppose that $\mu\in \mathcal{M}_a^*\left(\mathbb{R}\right)$, and $\rho$ is Lipschitz continuous and satisfies \eqref{E:LinGrow}.
Then the SPDE \eqref{E:FracHt} has a unique (in the sense of versions) random
field solution $\{u(t,x)\!: (t,x)\in\mathbb{R}_+^* \times \mathbb{R} \}$ starting from $\mu$. Moreover, for all even integers $p\ge 2$, all $t>0$ and $x\in\mathbb{R}$,
\begin{align}\label{E:MomUp}
\Norm{u(t,x)}_p^2 \le
\begin{cases}
J_0^2(t,x) + \left(\left[\Vip^2+J_0^2\right] \star \overline{\mathcal{K}} \right)
(t,x),& \text{if $p=2$}\;,\cr
2J_0^2(t,x) + \left(\left[\Vip^2+2J_0^2\right] \star \widehat{\mathcal{K}}_p \right)
(t,x),& \text{if $p>2$}\;.
\end{cases}
\end{align}
\end{theorem}
In order to use the moment bounds in \eqref{E:MomUp}, we need some estimates on $\mathcal{K}(t,x)$.
Recall that if the partial differential operator is the heat operator
$\frac{\partial }{\partial t}-\frac{\nu}{2} \Delta$ where $\nu>0$, then
\[
\mathcal{K}^{\mbox{\scriptsize heat}}(t,x;\lambda) = G_{\frac{\nu}{2}}(t,x)
\left(\frac{\lambda^2}{\sqrt{4\pi\nu t}}+\frac{\lambda^4}{2\nu}
\: e^{\frac{\lambda^4 t}{4\nu}}\Phi\left(\lambda^2
\sqrt{\frac{t}{2\nu}}\right)\right),
\]
where $\Phi(x)$ is the distribution function of the standard normal
random variable and $G_\nu(t,x)=\frac{1}{\sqrt{2\pi\nu t}}\exp\left(-x^2/(2\nu t)\right)$; see \cite{ChenDalang13Heat}. If the partial
differential operator is the wave operator
$\frac{\partial ^2 }{\partial t^2}-\kappa^2 \Delta$ where $\kappa>0$, then
\[
\mathcal{K}^{\mbox{\scriptsize wave}}(t,x;\lambda) = \frac{\lambda^2}{4}
I_0\left(\sqrt{\frac{\lambda^2((\kappa t)^2-x^2)}{2\kappa}}\right) \Indt{|x|\le
\kappa t},
\]
where $I_0(x)$ is the modified Bessel function of the first kind
of order $0$; see \cite{ChenDalang14Wave}.
Except these two cases, there are no explicit formulas for $\mathcal{K}(t,x)$.
The following upper bound on $\mathcal{K}(t,x)$ from \cite[Proposition 3.2]{ChenDalang14FracHeat} will be useful in this paper.
\begin{proposition}
\label{P:UpperBdd-K}
Let $\gamma:=\lambda^2\Lambda\:\Gamma(1-1/a)$.
For some finite constant $C=C(\lambda)>0$,
\begin{align}
\mathcal{K}(t,x;\lambda)
&\le
\frac{C }{t^{1/a}} \lMr{\delta}{G}{a}(t,x)
\left(
1+t^{1/a} \exp\left(\gamma^{a^*} t\right)
\right), \quad\text{for all $t\ge 0$ and $x\in\mathbb{R}$.}
\label{E:UpBd-K}
\end{align}
\end{proposition}
\section{Proof of Theorem \ref{T:WComp}}\label{S:WComp}
Before proving Theorem \ref{T:WComp}, we need some preparation.
One may view $\lMr{\delta}{G}{a}(t,x)$ as an operator, denoted by $\lMr{\delta}{\bold G}{a}(t)$ for clarity, as follows:
\[
\lMr{\delta}{\bold G}{a}(t) f (x) := \left(\lMr{\delta}{G}{a}(t,\cdot)*f\right)(x).
\]
Let $\bold I$ be the identity operator: $\bold I f(x)=(\delta *f)(x) = f(x)$.
Set
\[
\tlMr{\epsilon}{\delta}{D}{a} = \frac{\lMr{\delta}{\bold G}{a}(\epsilon)-\bold I}{\epsilon}.
\]
Let
\begin{align}
\tlMr{\epsilon}{\delta}{\bold G}{a}(t) = \exp(t\tlMr{\epsilon}{\delta}{D}{a})&=
e^{-t/\epsilon}\sum_{n=0}^\infty \frac{(t/\epsilon)^n}{n!} \lMr{\delta}{\bold G}{a}(n\epsilon)=e^{-t/\epsilon}\bold I + \tlMr{\epsilon}{\delta}{\bold R}{a}(t),
\end{align}
where the operator $\tlMr{\epsilon}{\delta}{\bold R}{a}(t)$ has a density, denoted by $\tlMr{\epsilon}{\delta}{R}{a}(t,x)$, which is equal to
\begin{align}\label{E:R}
\tlMr{\epsilon}{\delta}{R}{a}(t,x)&=
e^{-t/\epsilon}\sum_{n=1}^\infty \frac{(t/\epsilon)^n}{n!} \lMr{\delta}{G}{a}(n\epsilon,x).
\end{align}
One may also write the kernel of $\tlMr{\epsilon}{\delta}{\bold G}{a}(t)$ as
\begin{align} \label{E:Geps}
\tlMr{\epsilon}{\delta}{G}{a}(t,x) = e^{-t/\epsilon} \delta_0(x) + \tlMr{\epsilon}{\delta}{R}{a}(t,x).
\end{align}
The two operators $\tlMr{\epsilon}{\delta}{\bold G}{a}$ and $\lMr{\delta}{\bold G}{a}$ are close in many senses;
see Appendix for more details.
\begin{proof}[Proof of Theorem \ref{T:WComp}]
Denote $\phi_\epsilon(x)=(2\pi\epsilon)^{-1/2}\exp\left(-x^2/(2\epsilon)\right)$.
For $\epsilon>0$ and $x\in\mathbb{R}$, denote
\[
W_x^\epsilon (t) := \int_0^t \int_\mathbb{R} \phi_\epsilon(x-y)W(\ensuremath{\mathrm{d}} s,\ensuremath{\mathrm{d}} y),\quad\text{for $t\ge 0$}.
\]
Clearly, $t\mapsto W_x^\epsilon (t)$ is a one-dimensional Brownian motion.
Denote $\dot{W}_x^\epsilon(t) = \frac{\ensuremath{\mathrm{d}} }{\ensuremath{\mathrm{d}} t}W_x^\epsilon (t)$. Then the quadratic variation of $\ensuremath{\mathrm{d}} W_x^\epsilon(t)$ is
\begin{align}\label{E:Quad}
\ensuremath{\mathrm{d}} \InPrd{W_x^\epsilon(t)} = \int_\mathbb{R} \phi_\epsilon^2(x-y)\ensuremath{\mathrm{d}} y \ensuremath{\mathrm{d}} t =\frac{1}{\sqrt{4\pi\epsilon}}\: \ensuremath{\mathrm{d}} t.
\end{align}
Consider the following stochastic partial differential equation
\begin{align}\label{E:Ueps}
\begin{cases}
\displaystyle
\frac{\partial }{\partial t} u_\epsilon(t,x) = \tlMr{\epsilon}{\delta}{D}{a} u_\epsilon(t,x)
+ \rho(u_\epsilon(t,x)) \dot{W}_x^\epsilon(t), & \quad t>0, \; x\in\mathbb{R}, \\[0.7em]
u_\epsilon(0,x)= (\mu*\lMr{\delta}{G}{a}(\epsilon,\cdot))(x), \quad x\in\mathbb{R}.
\end{cases}
\end{align}
Since $\rho$ is globally Lipschitz continuous, \eqref{E:Ueps} has a unique strong solution
\[
u_\epsilon(t,x)= (\mu*\lMr{\delta}{G}{a}(\epsilon,\cdot))(x)
+ \int_0^t \ensuremath{\mathrm{d}} s \tlMr{\epsilon}{\delta}{D}{a} u_\epsilon(s,x) +
\int_0^t \rho(u_\epsilon(s,x))\ensuremath{\mathrm{d}} W_x^\epsilon(s).
\]
{\vspace{0.7em}\noindent \bf Step 1.}
Let $u_{\epsilon,i}(t,x)$ be the solutions to \eqref{E:Ueps} with initial data $\mu_i$, $i=1,2$, respectively.
Denote $v_\epsilon(t,x):= u_{\epsilon,2}(t,x)-u_{\epsilon,1}(t,x)$. We will prove that
\begin{align}\label{E:NonUe}
P\left(v_\epsilon(t,x)\ge 0,\; \text{for every $t>0$ and $x\in\mathbb{R}$}\right) =1.
\end{align}
Let $a_n=-2(n^2+n+2)^{-1}$, $n\ge 0$. Then $a_n\uparrow 0$ as $n\rightarrow \infty$ and
$\int^{a_n}_{a_{n-1}}x^{-2}\ensuremath{\mathrm{d}} x = n$. Let $\psi_n(x)$, $n=1,2,\dots$, be nonnegative continuous functions supported on $\:]a_{n-1},a_n[\:$ such that
\[
0\le \psi_n(x)\le \frac{2}{n x^2}\quad\text{and}\quad \int^{a_n}_{a_{n-1}}\psi_n(x) \ensuremath{\mathrm{d}} x=1.
\]
Define
\[
\Psi_n(x) := \int_0^x\ensuremath{\mathrm{d}} y\int_0^y \psi_n(z)\ensuremath{\mathrm{d}} z.
\]
Clearly, $\Psi_n(x)\in C^2(\mathbb{R})$ with $\Psi_n''(x)=\psi_n(x)$, $\Psi_n(x)=0$ for $x\ge 0$, and $-1\le \Psi_n'(x) =\int_0^x \psi_n(z)\ensuremath{\mathrm{d}} z\le 0$ for all $x\in\mathbb{R}$. Let ${{\rm 1\mkern-1.5mu}\!{\rm I}}(\cdot)$
denote the indicator function. Here are
three important properties: For all $x\in\mathbb{R}$, as $n\rightarrow+\infty$,
\begin{align}\label{E:PsiConv}
\Psi_n(x)\uparrow -(x\wedge 0)=:\Psi(x), \quad
\Psi_n'(x)\downarrow -{{\rm 1\mkern-1.5mu}\!{\rm I}} (x<0)
\quad\text{and}\quad
\Psi_n'(x)\:x\uparrow \Psi(x),
\end{align}
Because for each $x\in\mathbb{R}$ fixed, $u_\epsilon(t,x)$ is a semi-martingale, by It\^{o}'s formula,
\begin{align*}
\Psi_n(v_\epsilon(t,x))=& \int_0^t \Psi_n'(v_\epsilon(s,x)) \left[\rho(u_{\epsilon,2}(s,x))-\rho(u_{\epsilon,1}(s,x))\right] \ensuremath{\mathrm{d}} W_x^\epsilon(s)\\
&+\frac{1}{2}\int_0^t\Psi_n''(v_\epsilon(s,x)) \left[\rho(u_{\epsilon,2}(s,x))-\rho(u_{\epsilon,1}(s,x))\right]^2\frac{1}{\sqrt{4\pi\epsilon}}\ensuremath{\mathrm{d}} s\\
&+ \int_0^t \Psi_n'(v_\epsilon(s,x)) \tlMr{\epsilon}{\delta}{D}{a} v_{\epsilon}(s,x) \ensuremath{\mathrm{d}} s.
\end{align*}
By the Lipschitz condition on $\rho$,
\[
\Psi_n''(v_\epsilon(s,x)) \left[\rho(u_{\epsilon,2}(s,x))-\rho(u_{\epsilon,1}(s,x))\right]^2
\le \LIP_\rho^2 \Psi_n''(v_\epsilon(s,x)) v_\epsilon^2(s,x)\le
2\LIP_\rho^2/n.
\]
Hence,
\begin{align*}
\mathbb{E}\left[\Psi_n(v_\epsilon(t,x))\right]\le &\frac{\LIP_\rho^2 t}{n \sqrt{4\pi\epsilon}}
+ \mathbb{E}\left[\frac{1}{\epsilon}\int_0^t \ensuremath{\mathrm{d}} s \: \Psi_n'(v_\epsilon(s,x)) \int_\mathbb{R} \ensuremath{\mathrm{d}} y \lMr{\delta}{G}{a}(\epsilon,x-y)\left[v_\epsilon(s,y)-v_\epsilon(s,x)\right]\right].
\end{align*}
Now let $n$ go to $+\infty$, by \eqref{E:PsiConv} and the monotone convergence theorem,
\begin{align*}
\mathbb{E}\left[\Psi(v_\epsilon(t,x))\right]\le & \frac{1}{\epsilon}\int_0^t \mathbb{E}\left[{{\rm 1\mkern-1.5mu}\!{\rm I}}(v_\epsilon(s,x)<0)v_\epsilon(s,x)\right] \ensuremath{\mathrm{d}} s\\
&-\frac{1}{\epsilon}\int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R} \ensuremath{\mathrm{d}} y
\lMr{\delta}{G}{a}(\epsilon,x-y)
\mathbb{E}\left[{{\rm 1\mkern-1.5mu}\!{\rm I}}(v_\epsilon(s,x)<0)v_\epsilon(s,y)\right].
\end{align*}
Notice that
\begin{align*}
-\frac{1}{\epsilon}\int_0^t \ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y &\lMr{\delta}{G}{a}(\epsilon,x-y) \mathbb{E}\left[{{\rm 1\mkern-1.5mu}\!{\rm I}}(v_\epsilon(s,x)<0) v_\epsilon(s,y)\right]\\
\le&
-\frac{1}{\epsilon}\int_0^t \ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y \lMr{\delta}{G}{a}(\epsilon,x-y) \mathbb{E}\left[{{\rm 1\mkern-1.5mu}\!{\rm I}}(v_\epsilon(s,x)<0,v_\epsilon(s,y)<0) v_\epsilon(s,y)\right]\\
=&\frac{1}{\epsilon}\int_0^t \ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y \lMr{\delta}{G}{a}(\epsilon,x-y) \mathbb{E}\left[{{\rm 1\mkern-1.5mu}\!{\rm I}}(v_\epsilon(s,x)<0,v_\epsilon(s,y)<0) |v_\epsilon(s,y)|\right]\\
\le &\frac{1}{\epsilon}\int_0^t \ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y \lMr{\delta}{G}{a}(\epsilon,x-y) \mathbb{E}\left[{{\rm 1\mkern-1.5mu}\!{\rm I}}(v_\epsilon(s,y)<0) |v_\epsilon(s,y)|\right]
\end{align*}
Then using the fact that $|x|{{\rm 1\mkern-1.5mu}\!{\rm I}}(x<0) = \Psi(x)$, we have that
\[
\mathbb{E}\left[\Psi(v_\epsilon(t,x))\right]\le \frac{1}{\epsilon}\int_0^t \ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}}
y
\lMr{\delta}{G}{a}(\epsilon,x-y) \mathbb{E}\left[\Psi(v_\epsilon(s,y))\right].
\]
Therefore, by Gronwall's lemma applied to $\sup_{y\in\mathbb{R}}\mathbb{E}\left[\Psi(v_\epsilon(s,y))\right]$, one can conclude that
$\mathbb{E}\left[\Psi(v_\epsilon(s,y))\right]=0$ for every $t>0$ and $x\in\mathbb{R}$. This proves \eqref{E:NonUe}.
{\vspace{0.7em}\noindent \bf Step 2.} In this step, we assume that
$\mu(\ensuremath{\mathrm{d}} x) = f(x)\ensuremath{\mathrm{d}} x$ with $f\in L^\infty(\mathbb{R})$ and $f(x)\ge 0$ for all $x\in\mathbb{R}$.
Denote $f_\epsilon(x):= (\mu*\lMr{\delta}{G}{a}(\epsilon,\cdot))(x)$.
We will prove that
\begin{align}\label{E:uAprox}
\lim_{\epsilon\rightarrow 0} \sup_{x\in\mathbb{R}}
\Norm{u_\epsilon(t,x)-u(t,x)}_2^2 = 0,\quad \text{for all $t>0$},
\end{align} where $u_\epsilon$ is a solution to \eqref{E:Ueps} with $u_\epsilon(0,x)=f_\epsilon(x)$ and $u$ is a solution to \eqref{E:FracHt}.
Fix $T>0$.
Notice that $u_\epsilon(t,x)$ can be written in the following mild form using the kernel of $\tlMr{\epsilon}{\delta}{\bold G}{a}(t)$ in \eqref{E:Geps}:
\begin{align*}
u_\epsilon(t,x)=&\left(f_\epsilon*\tlMr{\epsilon}{\delta}{G}{a}(t,\cdot)\right)(x)
+\int_0^t e^{-(t-s)/\epsilon}\rho\left(u_\epsilon(s,x)\right)\ensuremath{\mathrm{d}} W_x^\epsilon(s)\\
&+\int_0^t\int_\mathbb{R}\tlMr{\epsilon}{\delta}{R}{a}(t-s,x-y) \rho\left(u_\epsilon(s,y)\right)\ensuremath{\mathrm{d}} W_y^\epsilon(s)\ensuremath{\mathrm{d}} y,
\end{align*}
where the last term equals to
\[
\int_0^t \int_\mathbb{R} \left(\int_\mathbb{R} \ensuremath{\mathrm{d}} z \tlMr{\epsilon}{\delta}{R}{a}(t-s,x-z)\rho(u_\epsilon(s,z)) \phi_\epsilon(y-z)\right) W(\ensuremath{\mathrm{d}} s,\ensuremath{\mathrm{d}} y).
\]
By \eqref{E:RGaprx2} below, the boundedness of the initial data implies that for all $t>0$,
\begin{gather}\label{E:AT}
A_t:=\sup_{\epsilon\in\:]0,1]}\sup_{s\in[0,t]} \sup_{x\in\mathbb{R}} \Norm{u_\epsilon(s,x)}_2^2\vee \Norm{u(s,x)}_2^2<+\infty.
\end{gather}
By the linear growth condition \eqref{E:LinGrow},
\begin{align*}
&\Norm{u_\epsilon(t,x)-u(t,x)}_2^2\\
&\hspace{1em}\le
6\left(f_{\epsilon}*\tlMr{\epsilon}{\delta}{G}{a}(t,\cdot)-f*\lMr{\delta}{G}{a}(t,\cdot)\right)^2(x) \\
&\hspace{2em} +6\Lip_\rho^2 \int_0^t \ensuremath{\mathrm{d}} s \: \frac{1}{\sqrt{4\pi\epsilon}}e^{-2(t-s)/\epsilon}\left(\Vip^2+\Norm{u_\epsilon(s,x)}_2^2\right)\\
&\hspace{2em}+6\int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y\: \mathbb{E}\left[\left|
\int_\mathbb{R} \ensuremath{\mathrm{d}} z \tlMr{\epsilon}{\delta}{R}{a}(t-s,x-z)
\left[\rho(u_\epsilon(s,z))-\rho(u(s,z))\right]\phi_\epsilon(y-z)
\right|^2\right]\\
&\hspace{2em}+6\int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y\: \mathbb{E}\left[\left|
\int_\mathbb{R} \ensuremath{\mathrm{d}} z \tlMr{\epsilon}{\delta}{R}{a}(t-s,x-z)
\left[\rho(u(s,z))-\rho(u(s,y))\right]\phi_\epsilon(y-z)
\right|^2\right]\\
&\hspace{2em}+6\int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y\: \mathbb{E}\left[\left|
\int_\mathbb{R} \ensuremath{\mathrm{d}} z
\left[\tlMr{\epsilon}{\delta}{R}{a}(t-s,x-z)-\lMr{\delta}{G}{a}(t-s,x-z)\right] \rho(u(s,y))\phi_\epsilon(y-z)
\right|^2\right]\\
&\hspace{2em}+6\int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y\: \mathbb{E}\left[\left|
\int_\mathbb{R} \ensuremath{\mathrm{d}} z
\left[\lMr{\delta}{G}{a}(t-s,x-z)-\lMr{\delta}{G}{a}(t-s,x-y)\right] \rho(u(s,y))\phi_\epsilon(y-z)
\right|^2\right]\\
&\hspace{1em}=: 6 \sum_{n=1}^6 I_n(t,x;\epsilon).
\end{align*}
Denote $C_f:=\sup_{x\in\mathbb{R}} f(x)\ge \sup_{x\in\mathbb{R}} f_\epsilon(x)$.
Using the semigroup property, we see that
\begin{align*}
& I_1(t,x;\epsilon) \\
\le& \left[\left(f_\epsilon*\left|\tlMr{\epsilon}{\delta}{G}{a}(t,\cdot)-\lMr{\delta}{G}{a}(t,\cdot)\right|\right)(x) +\left(f*|\lMr{\delta}{G}{a}(t+\epsilon,\cdot)-\lMr{\delta}{G}{a}(t,\cdot)|\right)
(x) \right]\\
& \times \left[\left(f_\epsilon*\tlMr{\epsilon}{\delta}{G}{a}(t,\cdot)\right)(x)+\left(f*\lMr{\delta}{G}{a}(t,\cdot)\right)(x)\right]\\
\le &
2\: C_f^2\left(e^{-t/\epsilon}+\int_\mathbb{R} \ensuremath{\mathrm{d}} y \left|\tlMr{\epsilon}{\delta}{R}{a}(t,y)-\lMr{\delta}{G}{a}(t,y)\right|+\int_{\mathbb{R}}\ensuremath{\mathrm{d}} y \left|
\lMr{\delta}{G}{a}{(t+\epsilon,y)}-\lMr{\delta}{G}{a}(t,y) \right| \right)\\
\le &
2\: C_f^2
\left(2e^{-t/\epsilon}+(C'+C'') \left(\epsilon/t\right)^{1/2} \right),
\end{align*}
where the last step is due to Lemma \ref{L:RGaprxL1}, \eqref{E:timediff} and the fact that $\log(1+x) \leq \sqrt{x}$ for all $x\geq 0$, and $C'$ and $C''$ are the constants defined in
\eqref{E:RGaprxL1} and \eqref{E:constantc'}. For simplicity, define $C_*:=C'+C''$.
As for $I_2(t,x;\epsilon)$,
\[
I_2(t,x;\epsilon)\le \Lip_\rho^2\: (\Vip^2+A_t)\frac{\sqrt{\epsilon}}{\sqrt{4\pi}}
\left(1- e^{-2t/\epsilon }\right),
\]
which implies
\[
\lim_{\epsilon\rightarrow 0}\sup_{0\le t\le T}\sup_{x\in\mathbb{R}}I_2(t,x;\epsilon) =0.
\]
By the H\"older inequality and the Lipschitz continuity of the function $\rho$,
\begin{align}\notag
I_3(t,x;\epsilon)&
\le \LIP_\rho^2 \int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y\int_\mathbb{R}\ensuremath{\mathrm{d}} z \tlMr{\epsilon}{\delta}{R}{a}^2(t-s,x-z)\Norm{u_\epsilon(s,z)-u(s,z)}_2^2 \phi_\epsilon(y-z)\\
&\le \LIP_\rho^2 \int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y\tlMr{\epsilon}{\delta}{R}{a}^2(s,x-y)\Norm{u_\epsilon(t-s,y)-u(t-s,y)}_2^2,
\label{E:I3}
\end{align}
and similarly,
\begin{align*}
I_4(t,x;\epsilon)
& \le \LIP_\rho^2 \int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y\int_\mathbb{R}\ensuremath{\mathrm{d}} z \tlMr{\epsilon}{\delta}{R}{a}^2(s,z)\Norm{u(t-s,x-z)-u(t-s,x-y)}_2^2 \phi_\epsilon(y-z).
\end{align*}
By Lemma \ref{L:MInc}, for some constant $C:=C(T,a,\delta,\mu)$,
\[
\Norm{u(t-s,x-z)-u(t-s,x-y)}_2^2 \le C (t-s)^{-1/a}|y-z| + C A_T |y-z|^{a-1}.
\]
Hence, integrating over $\ensuremath{\mathrm{d}} y$ first and then integrating over $\ensuremath{\mathrm{d}} z$ using \eqref{E:RGaprx2} give that
\begin{align*}
I_4(t,x;\epsilon)
&\le \LIP_\rho^2 C \int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} z \tlMr{\epsilon}{\delta}{R}{a}^2(s,z)
\left[(t-s)^{-1/a} \sqrt{2\epsilon/\pi}+\frac{A_T 2^{(a-1)/2}}{\sqrt{\pi}}\Gamma\left(a/2\right)\epsilon^{(a-1)/2}
\right]\\
&\le\frac{\LIP_\rho^2 C \: C_{a,\delta}}{\sqrt{\pi}} \int_0^t \ensuremath{\mathrm{d}} s
\: s^{-1/a}\left[(t-s)^{-1/a} \sqrt{2\epsilon}+A_T 2^{(a-1)/2}\Gamma\left(a/2\right)\epsilon^{(a-1)/2}
\right],
\end{align*}
where $C_{a,\delta}$ is defined in Lemma \ref{L:RGaprx}. Finally, integrating over $\ensuremath{\mathrm{d}} s$ using the Beta integral, we have that for some finite constant $C^*:=C^*(T,a,\delta,\mu,A_T)>0$,
\[
I_4(t,x;\epsilon)\le C^* \LIP_\rho^2 \left(t^{1-2/a} \epsilon^{1/2}+\epsilon^{(a-1)/2}\right).
\]
By H\"older inequality, \eqref{E:LinGrow} and \eqref{E:AT},
\[
I_5(t,x;\epsilon)\le \Lip_\rho^2 \left(\Vip^2+A_t\right)\int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y \int_\mathbb{R}\ensuremath{\mathrm{d}} z \: \phi_\epsilon(y-z)
\left[\tlMr{\epsilon}{\delta}{R}{a}(t-s,x-z)-\lMr{\delta}{G}{a}(t-s,x-z)\right]^2.
\]
Integrate $\ensuremath{\mathrm{d}} y$ and enlarge the integral interval for $\ensuremath{\mathrm{d}} s$ from $[0,t]$ to $[0,T]$,
\[
I_5(t,x;\epsilon)\le \Lip_\rho^2 \left(\Vip^2+A_T\right)\int_0^T\ensuremath{\mathrm{d}} s\int_\mathbb{R} \ensuremath{\mathrm{d}} z \:
\left[\tlMr{\epsilon}{\delta}{R}{a}(s,z)-\lMr{\delta}{G}{a}(s,z)\right]^2,
\]
and then apply \eqref{E:RGaprx1} to obtain
\[
\lim_{\epsilon\rightarrow 0} \sup_{0\le t\le T}\sup_{x\in\mathbb{R}} I_5(t,x;\epsilon) =0.
\]
Similarly to the case of $I_5$, we have that
\begin{align*}
I_6(t,x;\epsilon)\le & \Lip_\rho^2 \left(\Vip^2+A_T\right)\int_0^T\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y \int_\mathbb{R}\ensuremath{\mathrm{d}} z \: \phi_\epsilon(y-z)
\left[\lMr{\delta}{G}{a}(s,x-z)-\lMr{\delta}{G}{a}(s,x-y)\right]^2\\
= &
2\Lip_\rho^2 \left(\Vip^2+A_T\right)\int_0^T\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y\left[
\lMr{\delta}{G}{a}^2(s,y)
-\lMr{\delta}{G}{a}(s,y) \int_\mathbb{R}\ensuremath{\mathrm{d}} z \lMr{\delta}{G}{a}(s,z) \phi_\epsilon(y-z)\right].
\end{align*}
Define $F_\epsilon(s,y):=\lMr{\delta}{G}{a}^2(s,y)
-\lMr{\delta}{G}{a}(s,y) \int_\mathbb{R}\ensuremath{\mathrm{d}} z \lMr{\delta}{G}{a}(s,z) \phi_\epsilon(y-z)$. Clearly, $\lim_{\epsilon\rightarrow 0}F_\epsilon(s,y)=0$ for all $s>0$ and $y\in\mathbb{R}$.
On the other hand,
\[
F_\epsilon(s,y)\le
\lMr{\delta}{G}{a}^2(s,y) + \lMr{\delta}{G}{a}(s,y)s^{-1/a} \Lambda,
\]
where the constant $\Lambda$ is defined in \eqref{E:Cst-dLa}. In fact, this upper bound is integrable:
\begin{align*}
\int_0^T\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y &\left(\lMr{\delta}{G}{a}^2(s,y) + \lMr{\delta}{G}{a}(s,y)s^{-1/a} \Lambda\right)\\
&=\int_0^T \ensuremath{\mathrm{d}} s \left[\lMr{\delta}{G}{a}(2s,0)+ \Lambda s^{-1/a}\right]
= \left(\frac{\lMr{\delta}{G}{a}(1,0)}{2^{1/a}}+\Lambda\right) \frac{a}{a-1}\:T^{1-1/a}.
\end{align*}
Hence, the dominated convergence theorem implies that
\[
\lim_{\epsilon\rightarrow 0}\sup_{0\le t\le T}\sup_{x\in\mathbb{R}}I_6(t,x;\epsilon) =0.
\]
Now set $M(t;\epsilon):=\sup_{y\in\mathbb{R}} \Norm{u_\epsilon(t,y)-u(t,y)}_2^2$.
Fix $T>0$. Combining things together, we can get that for some constant $C_T>0$,
\[
M(t;\epsilon)\le C_T\int_0^t\ensuremath{\mathrm{d}} s\: (t-s)^{-1/a} M(s;\epsilon)
+ H(T;\epsilon) + \widehat{H}(t;\epsilon),
\]
where
\begin{align*}
H(T;\epsilon) & :=6\sum_{n=2,5,6}\sup_{0\le t\le T}\sup_{x\in\mathbb{R}} I_n(t,x;\epsilon),\\
\widehat{H}(t;\epsilon) &:=12\: C_f^2
\left(2e^{-t/\epsilon}+C_* \left(\epsilon/t\right)^{1/2}\right) + 6\: C^* \LIP_\rho^2 \left(t^{1-2/a} \epsilon^{1/2}+\epsilon^{(a-1)/2}\right).
\end{align*}
Then by Chandirov's lemma, which is a variation of Bellman's inequality (see \cite[Theorem 1.4, on p. 5]{BainovSimeonov92}), for $0<t\le T$,
\begin{align*}
M(t;\epsilon) &\le H(T;\epsilon)+ \widehat{H}(t;\epsilon)
+\int_0^t \ensuremath{\mathrm{d}} s \left(H(T;\epsilon)+ \widehat{H}(s;\epsilon)\right)(t-s)^{-1/a} \,
\exp\left(\int_s^t \ensuremath{\mathrm{d}} \tau \: (t-\tau)^{-1/a}\right)\\
&= H(T;\epsilon)+ \widehat{H}(t;\epsilon)
+\int_0^t \ensuremath{\mathrm{d}} s \left(H(T;\epsilon)+ \widehat{H}(s;\epsilon)\right)(t-s)^{-1/a}\,
\exp\left(\frac{a}{a-1}(t-s)^{1-1/a}\right)\\
&\le H(T;\epsilon)\left(1+\frac{a}{a-1} T^{1-1/a}\,\exp\left(\frac{a\:T^{1-1/a}}{a-1}\right)\right)\\
&\qquad\qquad\qquad+ \widehat{H}(t;\epsilon)+\int_0^t \ensuremath{\mathrm{d}} s\: \widehat{H}(s;\epsilon)(t-s)^{-1/a}\,
\exp\left(\frac{a\: (t-s)^{1-1/a}}{a-1}\right).
\end{align*}
Clearly, as $\epsilon \rightarrow 0$, the first two terms in the above upper bound go to zero.
The integral also goes to zero by applying the dominated convergence theorem.
This proves \eqref{E:uAprox}.
Finally, suppose that $\mu_i(\ensuremath{\mathrm{d}} x)=f_i(x)\ensuremath{\mathrm{d}} x$ with $f_i\in L^\infty(\mathbb{R})$, $i=1,2$.
If $f_1(x)\le f_2(x)$ for almost all $x\in\mathbb{R}$, then by Step 1 we know that $v_\epsilon(t,x):= u_{\epsilon,2}(t,x)- u_{\epsilon,1}(t,x)\ge 0$ for all $t>0$ and $x\in\mathbb{R}$, a.s. Then Step 2 implies
$v_\epsilon(t,x)$ converges to $v(t,x)=u_2(t,x) - u_1(t,x)$ in $L^2(\Omega)$ for all $t>0$ and $x\in\mathbb{R}$. Therefore, the nonnegativity of $v(t,x)$ is inherited from that of $v_\epsilon(t,x)$, that is,
\[
P(u_1(t,x)\le u_2(t,x),\:\text{for all $t>0$ and $x\in\mathbb{R}$}) =1.
\]
{\vspace{0.7em}\noindent \bf Step 3.}
Now we assume that $\mu_i \in\mathcal{M}_a^*(\mathbb{R})$.
Recall the definition of $\psi_\epsilon$ in \eqref{E:psi}.
Fix $\epsilon>0$. Let $u_{\epsilon,i}$, $i=1,2$, be the solutions to \eqref{E:FracHt} starting from
$\left([\mu_i \psi_\epsilon]*\lMr{\delta}{G}{a}(\epsilon,\cdot)\right)(x)$.
Denote $v(t,x)=u_2(t,x)-u_1(t,x)$ and $v_\epsilon(t,x)=u_{\epsilon,2}(t,x)-u_{\epsilon,1}(t,x)$.
Because $\psi_\epsilon$ is a continuous function with compact support on $\mathbb{R}$, the initial data for $u_{\epsilon,i}(t,x)$ is bounded:
\[
\sup_{x\in\mathbb{R}} |\left([\mu_i \psi_\epsilon]*\lMr{\delta}{G}{a}(\epsilon,\cdot)\right)(x)|
\le
\frac{\Lambda}{\epsilon^{1/a}} \int_\mathbb{R} \psi_\epsilon(y)|\mu_i|(\ensuremath{\mathrm{d}} y)<+\infty,
\]
where $\Lambda$ is defined in \eqref{E:Cst-dLa}. Hence, by Step 2, we have that
\[
P(v_\epsilon(t,x)\ge 0,\:\text{for all $t>0$ and $x\in\mathbb{R}$}) =1,\quad\text{for all $\epsilon>0$.}
\]
Applying Theorem \ref{T:Approx}, we obtain
\[
P\left(v(t,x)\ge 0,\:\text{for all $t>0$ and $x\in\mathbb{R}$}\right) =1.
\]
This completes the proof of Theorem \ref{T:WComp}.
\end{proof}
\section{Proof of Theorem \ref{T:SComp}} \label{S:Scomp}
We need several lemmas.
Lemma \ref{L:J0LInd} below plays a role to initialize the induction procedure.
\begin{lemma}\label{L:J0LInd}
Let $d>0$. For all $t>0$ and $M>0$, there exist some constants $1<m_0=m_0(t,M)<\infty$ and
$0<\gamma \leq 1/4$ such that for all $m\ge m_0$, all $s\in \left[t/(2m),t/m\right]$ and $x\in\mathbb{R}$,
\begin{align}\label{E:J0LInd1}
\left(\lMr{\delta}{G}{a}(s,\cdot)*1_{]-d,d[}(\cdot)\right)(x)
\ge \gamma\, 1_{]-d-M/m,d+M/m[}(x).
\end{align}
\end{lemma}
\begin{proof}
Let $Z$ be a random variable with the stable density $\lMr{\delta}{G}{a}(1,x)$.
Define $\gamma:=\min\{P(Z\le 0), P(Z\ge 0)\}/2$. Clearly, $0<\gamma\le 1/4$.
We first consider the case where $-d-M/m\leq x \le 0$.
Because $t/2 \le ms\le t$, we have
\begin{align*}
\left(\lMr{\delta}{G}{a}(s,\cdot)*1_{]-d,d[}(\cdot)\right)(x) &=
\int_{-d}^d \lMr{\delta}{G}{a}(s,x-y)\ensuremath{\mathrm{d}} y= \int^{\frac{x+d}{s^{1/a}}}_{\frac{x-d}{s^{1/a}}} \lMr{\delta}{G}{a}(1,z)\ensuremath{\mathrm{d}} z
\\
& \ge P\left(-d\, (2m)^{1/a} t^{-1/a} \le Z\le -M\,m^{(1-a)/a} t^{-1/a} \right).
\end{align*}
Similarly, when $0 \le x \le d+M/m$, we have
\begin{align*}
\left(\lMr{\delta}{G}{a}(s,\cdot)*1_{]-d,d[}(\cdot)\right)(x)
&\ge P\left(M\, t^{-1/a} m^{(1-a)/a} \le Z \le d (2m)^{1/a} t^{-1/a}\right).
\end{align*}
Therefore, when $m$ is large enough, the above probabilities are bigger than $\gamma$.
This completes the proof of Lemma \ref{L:J0LInd}.
\end{proof}
\begin{lemma}\label{L:MBds}
(1) If $\mu\in\mathcal{M}_a(\mathbb{R})$ and $\rho$ satisfies \eqref{E:LinGrow}, then for all $p\ge 2$,
there exists some finite constant $C:=C(a,\delta,\Lip_\rho,\Vip,\mu,p)>0$ such that,
\[
\sup_{x\in\mathbb{R}}
\Norm{u(t,x)}_p^2 \le C (t\vee 1)^{2(1+1/a)}t^{-2/a}\left[1+t^{1-1/a}+ t\exp\left( \gamma^{a^*} t\right)\right],
\]
for all $t>0$, where $\gamma:=8p\Lip_\rho\Lambda\Gamma(1/a^*)$.\\
(2) If $\mu(\ensuremath{\mathrm{d}} x) = c\,\ensuremath{\mathrm{d}} x$, $c\ne 0$ and $\rho(0)=0$, then for some constant $Q:=Q(c,a,\LIP_\rho,\Lambda)>0$,
\[
\sup_{x\in\mathbb{R}}\mathbb{E}\left(|u(t,x)|^p\right)\le Q^p \exp\left( Q p^{\frac{2a-1}{a-1} } t\right),\quad
\text{for all $p\ge 2$ and $t\ge 0$}.
\]
\end{lemma}
\begin{proof}
Part (1) is from \cite[Lemma 4.9 and (4.20)]{ChenDalang14FracHeat}.
As for part (2), notice that $J_0(t,x) \equiv c$. Then by \eqref{E:MomUp} and \eqref{E:UpBd-K}, for $p\ge 2$ and $p\in\mathbb{N}$,
\begin{align*}
\Norm{u(t,x)}_p^2 & \le 2c^2 +C \:c^2\int_0^t\ensuremath{\mathrm{d}} s\: \left(s^{-1/a}+\exp\left(\gamma_p^{a^*} s\right)\right)\\
&\le
2c^2 +C \:c^2\left(\frac{a}{a-1} t^{\frac{a-1}{a}} + \gamma_p^{-a^*} \exp\left(\gamma_p^{a^*} t\right)\right),
\end{align*}
where $\gamma_p=16\: p \Lip_\rho^2 \Lambda \Gamma(1/a^*)$ and the constant $C=C(\LIP_\rho)$ is defined in Proposition \ref{P:UpperBdd-K}. Notice that $\log(x)\le \beta x$ for all $x\ge 0$
whenever $\beta\ge e^{-1}$.
So by choosing $\theta= \frac{a-1}{a e}\gamma_2^{-a^*}$, we have that $\exp\left(\theta \gamma_p^{a^*} t\right) \ge t^{\frac{a}{a-1}}$ for all $t>0$.
Hence, if $c\ne 0$, then
\[
\Norm{u(t,x)}_p^2 \le \left[ 2c^2 +
C \: c^2\left(\frac{a}{a-1} + \gamma_2^{-a^*} \right)\right]\exp\left(\theta \gamma_p^{a^*} t\right).
\]
Then, raise both sides by a power of $p/2$. This completes the proof of Lemma \ref{L:MBds}.
\end{proof}
The following lemma proves the inductive step.
\begin{lemma}\label{L:LgDiv}
Let $d>0, t>0$ and $M>0$. If $\rho(0)=0$ and $\mu(\ensuremath{\mathrm{d}} x)=1_{[-d,d]}(x)\ensuremath{\mathrm{d}} x$, then there are some finite constants $Q:=Q(\beta, \LIP_\rho, \Lambda, t)>0$, $0<\beta\le 1/8$, and $m_0>0$ such that
for
all $m\ge m_0$,
\begin{align*}
P\Big(
u(s,x)\ge \beta 1_{]-d-M/m,d+M/m[}(x)\;\;&\text{for all $\frac{t}{2m}\le s \le \frac{t}{m}$ and $x\in\mathbb{R}$}
\Big)\\
&\ge 1 -\exp\left(- Q \: m^{1-1/a} [\log(m)]^{2-1/a}\right).
\end{align*}
\end{lemma}
\begin{proof}
Define $S:=S_{t,m,d,M}:=\{(s,y): t/(2m)\le s\le t/m, \; |x|\le d+M/m \}$.
By Lemma \ref{L:J0LInd}, for some constant $0<\beta\le 1/8$,
\begin{align}\label{E_:2beta}
\left(\mu * \lMr{\delta}{G}{a}(s,\cdot)\right)(x)
\ge 2\beta 1_{]-d-M/m,d+M/m[}(x)\quad \text{for all $s\in \left[t/(2m),t/m\right]$ and $x\in\mathbb{R}$.}
\end{align}
Hence,
\begin{align*}
P\Bigg(&
u(s,x) < \beta 1_{]-d-M/m,d+M/m[}(x)\;\;\text{for some $\frac{t}{2m}\le s \le \frac{t}{m}$ and $x\in\mathbb{R}$}
\Bigg)\\
& \le
P\Bigg(
I(s,x) < - \beta\;\;\text{for some $(s,x)\in S$}
\Bigg) \\
&\le
P\left(
\sup_{(s,x)\in S}|I(s,x)| >\beta
\right)\le
\beta^{-p}
\mathbb{E}\left[
\sup_{(s,x)\in S}|I(s,x)|^p\right],
\end{align*}
where we have applied Chebyshev's inequality in the last step.
Denote $\tau=t/m$ and $S':=\{(s,y): 0\le s\le t/m, \; |x|\le d+M/m \}$.
By the fact that $I(0,x)\equiv 0$ for all $x\in\mathbb{R}$, a.s., we see that for all $0<\eta< 1-\frac{2(a+1)}{p(a-1)}$,
\begin{align}
\mathbb{E}\left[
\sup_{(s,x)\in S}\left|\frac{I(s,x)}{\tau^{\frac{a-1}{2a}\eta}}\right|^p\right]&\le
\mathbb{E}\left[
\sup_{(s,x), (s',x')\in S'}\left|\frac{I(s,x)-I(s',x')}{\left(|x-x'|^{\frac{a-1}{2}}+|s-s'|^{\frac{a-1}{2a}}\right)^\eta}\right|^p\right].
\label{E_:Holder}
\end{align}
Let us find the upper bound of \eqref{E_:Holder}. By the Burkholder-Davis-Gundy inequality and \cite[Proposition 4.4]{ChenDalang14FracHeat}, for some universal constant $C_1>0$,
\[
\mathbb{E}\left[|I(s,x)-I(s',x')|^p\right]\le C_1 \left(|x-x'|^{\frac{a-1}{2}}+|s-s'|^{\frac{a-1}{2a}}\right)^{p/2} \sup_{(t,y)\in S'} \|u(t,y)\|_p^p\:,
\]
for all $(s,x)$ and $(s',x')\in S'$. Hence, by part (1) of Lemma \ref{L:MBds},
\[
\sup_{(t,y)\in S'} \|u(t,y)\|_p^p \le Q^p \exp\left(Q p^{\frac{2a-1}{a-1}} \tau\right)=:C_{p,\tau},
\]
for some constant $Q:=Q(a,\LIP_\rho,\Lambda)>0$ and for all $p\in [2,\infty[\:$.
Notice that when $m$ is sufficiently large, $S\subset [0,1]^2$. Hence, the right hand side of \eqref{E_:Holder} is bounded by the same quantity with $S$ replaced by $[0,1]^2$.
Then by Kolmogorov's continuity theorem (see \cite[Theorem 1.4.1]{Kunita90Flow} and \cite[Proposition 4.2]{ChenDalang13Holder}), for some universal constant $C>0$, the
expectation on the right hand side of \eqref{E_:Holder} is bounded above by $C^p C_{p,\tau}$.
We consider the case where $p=O\left([m\log m]^{1-1/a}\right)$ as $m\rightarrow \infty$ (see \eqref{E:pA} below).
In this case, we have $p^{a/(a-1)}\tau=O(\log m)$ as $m\rightarrow\infty$ since $\tau=t/m$.
This implies that there exists some constant $Q':=Q'(\beta, \LIP_\rho, \Lambda, t)$ such that
\begin{align*}
\beta^{-p}\mathbb{E}\left[
\sup_{(s,x)\in S}\left|I(s,x)\right|^p\right]
&\le
Q'\, \tau^{\frac{(a-1)\eta}{2a}p}\, \exp\left(Q' p^{\frac{2a-1}{a-1}} \tau\right)\\
&=
Q' \exp\left(Q' p^{\frac{2a-1}{a-1}} \tau+\frac{(a-1)\eta}{2a}p \log(\tau)\right).
\end{align*}
By denoting $\eta = \theta \left(1-\frac{2}{p}\:
\frac{a+1}{a-1}\right)$ with $\theta \in \:]0,1[\;$, the above exponent becomes
\[
f(p):=Q' \tau p^{\frac{2a-1}{a-1}}+\frac{\log (\tau )\:\theta\: \left(p[a-1]-2 [a+1]\right)}{2 a}.
\]
It is easy to see that $f(p)$ for $p\ge 2$ is minimized at
\begin{align*}
p= & \left(\frac{(a-1)^2 \theta \log (1/\tau )}{2 a
(2 a-1) Q' \tau }\right)^{1-1/a}=
\left(\frac{(a-1)^2 \theta \: m \log(m/t)}{2 a (2 a-1) Q' t }\right)^{1-1/a}.
\end{align*}
Thus, for some constants $A:=A(\beta, \LIP_\rho, \Lambda, t)$ and $Q'':=Q''(\beta, \LIP_\rho, \Lambda, t)$,
\begin{align}\label{E:pA}
\min_{p\ge 2} f(p) \le f(p') = - Q'' m^{1-1/a} [\log(m)]^{2-1/a},\quad\text{with $p'= A \left[m \log(m)\right]^{1-1/a}.$}
\end{align}
This completes the proof of Lemma \ref{L:LgDiv}.
\end{proof}
\begin{figure}[htbp]
\centering
\includegraphics{ComparisonSchema}
\caption{Induction schema for the strong comparison principle.}
\label{F:SComp}
\end{figure}
\begin{proof}[Proof of Theorem \ref{T:SComp}]
Let $u(t,x):=u_2(t,x)-u_1(t,x)$ and $\tilde \rho(u):=\rho(u+u_1)-\rho(u_1)$. Then $u(t,x)$ is, in fact, a solution to \eqref{E:FracHt} with the nonlinear function $\tilde\rho$ and the
initial data $\mu:=\mu_2-\mu_1$. We note that $\tilde\rho(0)=0$ and $\tilde\rho$ is a Lipschitz continuous function with the same Lipschitz constant as for $\rho$. For simplicity, we will use $\rho$
instead of $\tilde\rho$. By the weak comparison principle (Theorem \ref{T:WComp}), we only need to consider the case when $\mu$ has compact support and prove that $u(t,x)>0$ for all $t>0$ and
$x\in\mathbb{R}$, a.s.
\paragraph{Case I.} We first assume that $\mu(\ensuremath{\mathrm{d}} x)=f(x)\ensuremath{\mathrm{d}} x$ with $f\in C(\mathbb{R})$ and $f(x)\ge 0$ for all $x\in\mathbb{R}$.
Since $\mu>0$, there exists $x\in\mathbb{R}$ such that $f(x)>0$.
By the weak comparison principle (Theorem \ref{T:WComp}),
we only need to consider the case where $f(x)=1_{[-d,d]}(x)$ for some $d>0$ (this is because, if $f(x)=1_{[a,b]}(x)$, then we can use $f(x-(a+b)/2)$ as our initial function; in addition, if $f(x)=c1_{[-d,d]}(x)$, then we can consider $\tilde u(t,x):=c u(t,x)$ which is the unique solution to \eqref{E:FracHt} with the
initial function $1_{[-d,d]}(x)$ and with replacing $\rho(z)$ by $c\rho(z/c)$ which is also Lipschitz continuous with the same Lipschitz constant as $\rho(z)$).
Let $\gamma\in \:]0,1/4]$ be the constant defined in Lemma \ref{L:J0LInd} and let $\beta:=\gamma/2$.
For any $M>0$ and $k=0,1,\cdots,m-1$, define the events
\begin{align*}
A_k &:= \left\{u(s,x)\ge \beta^{k+1} 1_{S_k^m}(x)\:\text{for all $s\in\left[\frac{(2k+1)t}{2m},\frac{(k+1)t}{m} \right]$ and $x\in\mathbb{R}$}\right\},\\
B_k &:= \left\{u(s,x)\ge \beta^{k+1} 1_{S_k^m}(x)\:\text{for all $s\in\left[\frac{kt}{m},\frac{(2k+1)t}{2m} \right]$ and $x\in\mathbb{R}$}\right\},
\end{align*}
where
\[
S^m_k :=\: \left ]-d-\frac{Mk}{m},d+\frac{Mk}{m}\right[\:.
\]
See Figure \ref{F:SComp} for an illustration of the schema.
By Lemma \ref{L:LgDiv}, there are constants $Q>0$ and $m_0>0$ such that for all $m\ge m_0$,
\[
P(A_0)\ge 1-c(m),
\]
where
\begin{equation}\label{E:CM}
c(m):=\exp\left(- Q \: m^{1-1/a} [\log(m)]^{2-1/a}\right).
\end{equation}
By definition, on the event $A_{k-1}$, $k\ge 1$,
\[
u\left(\frac{k t}{m},x\right)\ge \beta^{k} 1_{S_{k-1}^m}(x),\quad\text{for all $x\in\mathbb{R}$.}
\]
Let $w_k(s,x)$ be the solution to the following SPDE:
\begin{align*}
\begin{cases}
\left(\displaystyle\frac{\partial}{\partial t} - {}_x D_\delta^a \right) w_k(s,x) =
\rho_k\left(w_k(s,x)\right) \dot{W}_k(s,x),& s\in \mathbb{R}_+^*:=\;]0,+\infty[\;,\: x\in\mathbb{R},\cr
w_k(0,x) = 1_{S_{k-1}^m}(x),
\end{cases}
\end{align*}
where $\rho_k(x):=\beta^{-k} \rho(\beta^{k} x)$ and $\{\dot{W}_k(s,x):=\dot{W}(s+kt/m,x)\}_{k\ge 1}$ is the time-shifted white noise.
Note that $\rho_k(x)$ is also a Lipschitz continuous function with the
same Lipschitz constant as for $\rho$ and $\rho_k(0)=0$. Thus, by Lemma \ref{L:LgDiv}, we see that by the same constants $Q$ and $m_0$ as in \eqref{E:CM}, for all $m\ge m_0$,
\begin{align}\label{E:wk}
P\left( w_k(s,x) \geq \beta 1_{S_{k}^m}(x) \: \text{for all $s \in \left[\frac{t}{2m}, \frac{t}{m} \right]$} \:\text{and $x\in\mathbb{R}$}\right) \geq 1-c(m).
\end{align}
Let $v(s,x)$ be a solution to \eqref{E:FracHt} with the initial data $\mu(\ensuremath{\mathrm{d}} x):=\beta^k 1_{S_{k-1}^m}(x)\ensuremath{\mathrm{d}} x$,
subject to the above time-shifted noise $\dot{W}_k$ with the same $\rho$. Then $v(s,x)=\beta^k w_k(s,x)$ a.s. for all $s\ge 0$ and $x\in\mathbb{R}$.
Since $u(s+kt/m,x) \ge v(s,x)$ for all $x\in\mathbb{R}$ and $s\ge 0$ by the Markov property and the weak comparison principle (Theorem \ref{T:WComp}), \eqref{E:wk} implies that
\[
P\left(A_k\mid \mathcal{F}_{kt/m}\right)\ge 1- c(m), \quad\text{a.s. on $A_{k-1}$ for $1\le k\le m-1$.}
\]
Hence,
\[
P\left(A_k\mid A_{k-1}\cap \cdots \cap A_0\right) \ge 1-c(m),\quad\text{for all $1\le k\le m-1$.}
\]
Furthermore, because $A_0\subseteq B_0$, on the event $A_0$, we see that
\[
P(B_0) \ge P(A_0) \ge 1-c(m).
\]
Similarly, one can prove that
\[
P\left(B_k\mid B_{k-1}\cap \cdots \cap B_0\right) \ge 1-c(m),\quad\text{for all $1\le k\le m-1$.}
\]
Then,
\begin{align}\notag
P\left(\cap_{0\le k\le m-1} \left[A_k \cap B_k\right] \right)
&\ge 1-\left(1-P\left(\cap_{0\le k\le m-1}A_k\right)\right)-\left(1-P\left(\cap_{0\le k\le m-1}B_k\right)\right)\\ \notag
&\ge (1-c(m))^{m-1}P(A_0)+ (1-c(m))^{m-1}P(B_0) -1\\
&\ge 2(1-c(m))^{m}-1.
\label{E_:abcup}
\end{align}
Therefore, for all $t>0$ and $M>0$,
\begin{align*}
P\left(u(s,x)> 0 \;\; \text{for all $t/2\le s\le t$ and $|x|\le M/2$}\right)
&\ge
\lim_{m\rightarrow\infty} P\left(\cap_{0\le k\le m-1} \left[A_k \cap B_k\right] \right)
\\
&\ge \lim_{m\rightarrow\infty} 2(1-c(m))^{m}-1 =1.
\end{align*}
Since $t$ and $M$ are arbitrary, this completes the proof for Case I.
\paragraph{Case II.} Now we assume that $\mu\in\mathcal{M}_{a,+}^*(\mathbb{R})$.
We only need to prove that for each $\epsilon>0$,
\begin{align}\label{E:uEpsilon}
P\left(u(t,x)>0\; \text{for $t\ge \epsilon$ and $x\in\mathbb{R}$}\right) =1.
\end{align}
Fix $\epsilon>0$. Denote $V(t,x):=u(t+\epsilon,x)$. By the Markov property, $V(t,x)$ solves \eqref{E:FracHt} with
the time-shifted noise $\dot{W}_\epsilon (t,x):=\dot{W}(t+\epsilon,x)$ starting from $V(0,x)=u(\epsilon,x)$, i.e.,
\begin{equation}
\label{E:VInt}
\begin{aligned}
V(t,x) &= \left(u(\epsilon,\circ)* \lMr{\delta}{G}{a}(t,\circ)\right)(x)
+ \iint_{[0,t]\times\mathbb{R}} \rho(V(s,y)) \lMr{\delta}{G}{a}(t-s,x-y)W_\epsilon(\ensuremath{\mathrm{d}} s,\ensuremath{\mathrm{d}} y)\\
&=: \widetilde{J}_0(t,x) + \widetilde{I}(t,x).
\end{aligned}
\end{equation}
We first claim that
\begin{align}\label{E:uepsilonx}
P\left(u(\epsilon,x)= 0, \;\text{for all $x\in\mathbb{R}$}\right) =0.
\end{align}
Notice that by Theorem \ref{T:Holder}, the function $x\mapsto u(t,x)$ is H\"older continuous over $\mathbb{R}$
a.s. The weak comparison principle (Theorem \ref{T:WComp}) shows that $u(t,x)\ge 0$ a.s.
Hence, if \eqref{E:uepsilonx} is not true, then by the Markov property and the strong comparison principle in Case I, at all times $\eta\in [0, \epsilon]$, with some strict positive probability,
$u(\eta,x)=0$ for all $x\in\mathbb{R}$, which contradicts Theorem \ref{T:WeakSol} as $\eta$ goes to zero.
Therefore, there exists a sample space $\Omega'$ with $P(\Omega')=1$ such that for each $\omega\in\Omega'$,
there exists $x\in\mathbb{R}$ such that $u(\epsilon,x,\omega)>0$.
Since $u(\epsilon,x,\omega)$ is continuous at $x$, one can find two nonnegative constants $c$ and $\beta$ such that
$u(\epsilon,y,\omega)\ge \beta 1_{[x-c,x+c]}(y)$ for all $y\in\mathbb{R}$.
Then Case I implies that
\[
P\left(V_{\omega}(t,x)>0\;\text{for all $t\ge 0$ and $x\in\mathbb{R}$}\right) =1,
\]
where $V_\omega$ is the solution to \eqref{E:VInt} starting from $u(\epsilon,x,\omega)$. Therefore, \eqref{E:uEpsilon} is true.
This completes the whole proof of Theorem \ref{T:SComp}.
\end{proof}
\section{Proof of Theorem \ref{T1:Rates}}\label{S:Rates}
We first prove part (1).
For any compact sets $K\subseteq \mathbb{R}_+^*\times\mathbb{R}$, one can find $\eta>0$, $T>0$ and $N>0$ such that
$K\subseteq [\eta,T]\times [-N,N]$. Then choose $M=2NT/\eta$.
If $\mu(\ensuremath{\mathrm{d}} x)=f(x)\ensuremath{\mathrm{d}} x$ with $f\in C(\mathbb{R})$ and $f(x)\ge 0$ for all $x\in\mathbb{R}$, then following the proof of Theorem \ref{T:SComp}, from \eqref{E_:abcup}, we see that
\begin{align*}
P\left(\inf_{(s,x)\in K}u(s,x)< \beta^{m} \right)&
\le 1- P\left(\cap_{0\le k\le m-1} \left[A_k \cap B_k\right] \right)\\
&\le 2\left[1-(1-c(m))^m\right],
\end{align*} where $c(m)$ is defined in \eqref{E:CM}.
Because $\log(1-x)\ge -2 x$ for $0<x\le 1/2$, when $m$ is sufficiently large, so that
\begin{align}\label{E:1/2}
m \exp\left(- Q \: m^{1-1/a} [\log(m)]^{2-1/a}\right) \le 1/2,
\end{align}
we have that
\[
(1-c(m))^m\ge \exp\left(-2\: m \exp\left(- Q \: m^{1-1/a} [\log(m)]^{2-1/a}\right) \right).
\]
Since $1-e^{-x}\le x$ for $x\geq 0$, if $m$ is sufficiently large such that \eqref{E:1/2} holds, then
\[
\left[1-(1-c(m))^m\right]\le
2\: m \exp\left(- Q \: m^{1-1/a} [\log(m)]^{2-1/a}\right).
\]
If $\mu\in\mathcal{M}_{a,+}^*(\mathbb{R})$, then we follow the notation of Case II in the proof of Theorem \ref{T:SComp} with
$\epsilon =\eta/2$. Using the Markov property, for each initial data $u(\epsilon,x,\omega)$, we apply the previous case
to get \eqref{E:Rates1} with $u(t,x)$ replaced by $V_\omega(t-\epsilon,x)$.
Because the upper bound of which does not depend on $\omega$ and $u(\epsilon,x)$ is independent of $V(t,x)$,
\eqref{E:Rates1} holds for $V(t-\epsilon,x) =u(t,x)$.
This completes the proof of part (1) of Theorem \ref{T1:Rates}.
\bigskip
Now we prove part (2).
Since $f$ is a continuous function, there exists finite constant $c>0$ such that $f(x)\ge c 1_{D}(x)$. Without loss of generality, we assume
that $c=1$. Let $v(t,x)$ be the solution to \eqref{E:FracHt} with the initial data $1_{D}(x)\ensuremath{\mathrm{d}} x$.
By Theorem \ref{T:WComp}, $u(t,x)\ge v(t,x)$ for all $t>0$ and $x\in\mathbb{R}$, a.s.
Hence, it suffices to prove that for all $n\ge 1$,
\[
P\left(\inf_{x\in D\:}\inf_{t\in\:]0,T]} v(t,x)\le e^{-n}\right) \le A \exp\left(-B(n\log(n))^{(2a-1)/a}\right).
\]
We define a set of $\{\mathcal{F}_t\}_{t\ge 0}$-stopping times as follows: $T_0:=0$, and
\[
T_{k+1}:=\inf\left\{s>T_k: \inf_{x\in \:D}v(s,x)\le e^{-k-1}\right\},
\]
where we use the convention that $\inf \phi=\infty$.
Similar to the proof of Theorem \ref{T:SComp}, let $\{\dot{W}_k(t,x): k\in\mathbb{N}\}$ be time-shifted space-time white noises
and let $v_k(t,x)$ be the unique solution to \eqref{E:FracHt} subject to the noise $\dot{W}_k$, starting from $v_k(0,x)=e^{-(k-1)}1_{D}(x)$. Then
\[
w_k(t,x):=e^{k-1}v_k(t,x)
\]
solves
\begin{align*}
\begin{cases}
\left(\displaystyle\frac{\partial}{\partial t} - {}_x D_\delta^a \right) w_k(t,x) =
\rho_k\left(w_k(t,x)\right) \dot{W}_k(t,x),& t\in \mathbb{R}_+^*:=\;]0,+\infty[\;,\: x\in\mathbb{R},\cr
w_k(0,x) = 1_{D}(x),
\end{cases}
\end{align*}
where $\rho_k(x):=e^{k-1}\rho(e^{-(k-1)}x)$.
From the definitions of the stopping times, we see that
\[
e^{k-1} v(T_{k-1},x)\ge 1_{D}(x),\quad \text{for all $x\in\mathbb{R}$, a.s. on $\{T_{k-1}<\infty\}$, for all $k\ge 1$}.
\]
Therefore, by the strong Markov property and the weak comparison principle in Theorem \ref{T:WComp},
we obtain that on $\{T_{k-1}<\infty\}$,
\begin{equation*}
P\left(T_{k}-T_{k-1} \leq \frac{2t}{n} \Big{|} \mathcal{F}_{T_{k-1}}\right)\leq P\left(\sup_{(t,x)\in \;]0,2T/n]\times D} \left| w_k(t,x)-w_k(0,x)\right| \geq 1-1/e\right).
\end{equation*}
Since $\rho_k$ is Lipschitz continuous with the same Lipschitz constant as $\rho$, a suitable form of the Kolmogorov continuity theorem (see the arguments in the proof of Lemma \ref{L:LgDiv})
implies that for all $\eta \in\;]0,1-2(a+1)/(p(a-1))[\;$,
there exists a finite constant $Q>0$, not depending on $p$, $n$ and $\tau$, such that for all $p\geq 2, n\geq 1,$ and $\tau \in \;]0,1[\;$,
\begin{equation}\label{E:HolderTime}
\mathbb{E} \left[ \sup_{(s,x)\in \;]0,\tau]\times D} \left|w_k(s,x)-w_k(0,x)\right|^p \right] \leq Q\, \tau^{p\eta(a-1)/2a}\exp\left(Q\,\tau p^{(2a-1)/(a-1)}\right).
\end{equation}
Letting $\tau:=2t/n$ for $0<t<T$ and minimizing the right hand side of \eqref{E:HolderTime} over $p$, we obtain that for some finite constant $Q'>0$, not depending on $n$,
\begin{equation*}
P\left(T_{k}-T_{k-1} \leq \frac{2t}{n} \Big{|} \mathcal{F}_{T_{k-1}}\right)\leq Q' \exp\left\{-Q' \,n^{(a-1)/a}(\log n)^{(2a-1)/a}\right\}.
\end{equation*}
Therefore, we obtain the following:
\begin{align*}
&P\left(\inf_{x\in D\:}\inf_{t\in\:]0,T]} v(t,x)\le e^{-n}\right)\leq P\{T_n\leq t\}\\
& \leq P\big(\text{at least $\lfloor n/2 \rfloor$-many distinct values $k\in \{1,2, \dots, n\}$ such that $T_k-T_{k-1}\leq 2t/n$} \big)\\
& \leq {n\choose \lfloor n/2 \rfloor}c_1^{\lfloor n/2 \rfloor} \exp\left\{-c_2{\lfloor n/2 \rfloor} \,n^{(a-1)/a}(\log n)^{(2a-1)/a} \right\}.
\end{align*}
This completes the proof of Theorem \ref{T1:Rates}. \hfill$\square$
\section{Proof of Theorem \ref{T:Approx}}\label{S:Approx}
\begin{proof}[Proof of Theorem \ref{T:Approx}]
Fix $\epsilon>0$. By Theorems \ref{T:ExUni}, both $u(t,x)$ and $u_\epsilon(t,x)$ are well-defined solutions to \eqref{E:FracHt}.
By Lipschitz continuity of $\rho$ and the moment formulas \eqref{E:MomUp},
\begin{align*}
\Norm{u(t,x)-u_\epsilon(t,x)}_2^2 \le&\quad
\left[\left((\mu \: \psi_\epsilon)*\lMr{\delta}{G}{a}(\epsilon,\cdot)*\lMr{\delta}{G}{a}(t,\cdot)\right)(x)
-\left(\mu*\lMr{\delta}{G}{a}(t,\cdot)\right)(x)\right]^2\\
&+ \LIP_\rho^2 \int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R}\ensuremath{\mathrm{d}} y
\Norm{u(s,y)-u_\epsilon(s,y)}_2^2 \lMr{\delta}{G}{a}^2(t-s,x-y).
\end{align*}
Denote the first part on the above upper bound as $I_\epsilon(t,x)$.
Let $\widetilde{\mathcal{K}}(t,x):=\mathcal{K}(t,x;\LIP_\rho)$ and denote $f_\epsilon(t,x):=\Norm{u(t,x)-u_\epsilon(t,x)}_2^2$. Then formally,
\[
(f_\epsilon\star \widetilde{\mathcal{K}})(t,x) \le \left(I_\epsilon\star \widetilde{\mathcal{K}}\right)(t,x)
+ \LIP_\rho^2 \left(f_\epsilon\star \lMr{\delta}{G}{a}^2\star \widetilde{\mathcal{K}}\right)(t,x).
\]
Using the fact that $ \left(\LIP_\rho^2\lMr{\delta}{G}{a}^2\star \widetilde{\mathcal{K}}\right)(t,x)= \widetilde{\mathcal{K}}(t,x)-\LIP_\rho^2\lMr{\delta}{G}{a}^2(t,x)$, one has that
\[
\left(f_\epsilon\star\lMr{\delta}{G}{a}^2\right)(t,x)\le \LIP_\rho^{-2} \left(I_\epsilon\star\widetilde{\mathcal{K}}\right)(t,x).
\]
Hence, it reduces to show that
\begin{align}\label{E_:IKlim0}
\lim_{\epsilon\rightarrow 0} \left(I_\epsilon\star\widetilde{\mathcal{K}}\right)(t,x) =0,\quad\text{for all $t>0$ and $x\in\mathbb{R}$}.
\end{align}
We first assume that $a\in \:]1,2[\:$. Notice that
\[
I_\epsilon(t,x) =
\left[
\left((\mu\:\psi_\epsilon)*\left[\lMr{\delta}{G}{a}(t+\epsilon,\cdot)-\lMr{\delta}{G}{a}(t,\cdot)\right]\right)(x)
+
\left([\mu\psi_\epsilon-\mu]*\lMr{\delta}{G}{a}(t,\cdot)\right)(x)
\right]^2.
\]
By \cite[(4.3)]{ChenDalang14FracHeat}, for $0<t\le T$ and $x\in\mathbb{R}$,
\begin{align}\label{E_:InitD}
(|\mu \psi_\epsilon| * \lMr{\delta}{G}{a}(t,\cdot))\left(x\right)\le
(|\mu| * \lMr{\delta}{G}{a}(t,\cdot))\left(x\right) \le C_T t^{-1/a},
\end{align}
with $C_T:=A_a\: K_{a,0}\: (T\vee 1)^{1+1/a}$, where $A_a$ is defined as
\begin{align}
\label{E:Aa}
A_a:=\sup_{y\in\mathbb{R}}\int_\mathbb{R} \frac{|\mu|(\ensuremath{\mathrm{d}} z)}{1+|y-z|^{1+a}}.
\end{align}
Hence, if $0<t +\epsilon\le T$, then
\[
I_\epsilon(t,x)\le 4 C_T t^{-1/a} \left[
\left(|\mu\:\psi_\epsilon|*\left|\lMr{\delta}{G}{a}(t+\epsilon,\cdot)-\lMr{\delta}{G}{a}(t,\cdot)\right|\right)(x)
+
\left(|\mu\psi_\epsilon-\mu|*\lMr{\delta}{G}{a}(t,\cdot)\right)(x)
\right].
\]
Now use the upper bound on $\widetilde{\mathcal{K}}(t,x)$ in \eqref{E:UpBd-K},
\begin{equation}
\label{E_:IcalK}
\begin{aligned}
\left(I_\epsilon\star \widetilde{\mathcal{K}}\right)(t,x)
\le& 2 C_T C' C_T' \int_0^t \ensuremath{\mathrm{d}} s \; s^{-1/a} (t-s)^{-1/a}\left[ g_1(t,s,\epsilon, x)+g_2(t,x,\epsilon,x)\right]
\end{aligned}
\end{equation}
where $C':=C'(a,\delta,\LIP_\rho)$ is defined in Proposition \ref{P:UpperBdd-K},
$C_T' =1+T^{1/a}\exp\left((\LIP^2_\rho\Lambda\Gamma(1/a^*))^{a^*} T\right)$,
$1/a^*+1/a=1$, and
\begin{align*}
g_1(t,s,\epsilon,x) &=\left(|\mu\:\psi_\epsilon|*\left|\lMr{\delta}{G}{a}(s+\epsilon,\cdot)-\lMr{\delta}{G}{a}(s,\cdot)\right|*\lMr{\delta}{G}{a}(t-s,\cdot)\right)(x) ,\\
g_2(t,s,\epsilon,x) &=\left(|\mu\psi_\epsilon-\mu|*\lMr{\delta}{G}{a}(s,\cdot)*\lMr{\delta}{G}{a}(t-s,\cdot)\right)(x).
\end{align*}
By the semigroup property and the dominated convergence theorem,
\[
g_2(t,s,\epsilon,x) =\left(|\mu\psi_\epsilon-\mu|*\lMr{\delta}{G}{a}(t,\cdot)\right)(x)\rightarrow 0,\quad\text{as $\epsilon\rightarrow 0$}.
\]
Clearly, $g_2(t,s,\epsilon,x)\le 2 (\mu*\lMr{\delta}{G}{a}(t,\cdot))(x)$.
Again, by the dominated convergence theorem and by bounding
$\lMr{\delta}{G}{a}(t+\epsilon,\cdot)$ using \cite[(4.3)]{ChenDalang14FracHeat}, one can show that $\lim_{\epsilon\rightarrow 0}g_1(t,s,\epsilon,x) =0$.
Then by the semigroup property and \eqref{E_:InitD},
for $0<t+\epsilon \le T$,
\begin{align*}
g_1(t,s,\epsilon,x)\le& \left(|\mu|*(\lMr{\delta}{G}{a}(s+\epsilon,\cdot)+\lMr{\delta}{G}{a}(s,\cdot))* \lMr{\delta}{G}{a}(t-s,\cdot)\right)(x)\\
=&\left(|\mu|*\lMr{\delta}{G}{a}(t+\epsilon,\cdot)\right) (x)
+\left(|\mu|*\lMr{\delta}{G}{a}(t,\cdot)\right) (x)\\
\le& \:C_T\left((t+\epsilon)^{-1/a}+t^{-1/a}\right)\le 2 C_T t^{-1/a}.
\end{align*}
Hence, both upper bounds on $g_1$ and $g_2$ are integrable over $\ensuremath{\mathrm{d}} s$ in \eqref{E_:IcalK}. Therefore, by another application of the dominated convergence theorem, we have proved
\eqref{E_:IKlim0}.
Since both functions $f_\epsilon(t,x)$ and $\lMr{\delta}{G}{a}(t,x)$ are nonnegative and the support of $\lMr{\delta}{G}{a}(t,x)$ is over $\mathbb{R}$, we
can conclude that $\lim_{\epsilon\rightarrow
0}f_\epsilon(t,x) =0$ for almost all $t>0$ and $x\in\mathbb{R}$.
When $a=2$, one can apply the dominated convergence theorem to show that $I_\epsilon(t,x)\rightarrow 0$ as $\epsilon\rightarrow 0$. Another application of the dominated convergence theorem shows that
\eqref{E_:IKlim0} is true. The rest is same as the previous case.
We leave the details for interested readers.
This completes the proof of Theorem \ref{T:Approx}.
\end{proof}
\section{Proof of Theorem \ref{T:Holder}}\label{S:Holder}
Without loss of generality, we assume that $\mu\ge 0$.
Let $u(t,x)$ be the solution to \eqref{E:FracHt} starting from $\mu\in\mathcal{M}_a(\mathbb{R})$.
Fix $T>0$ and $\epsilon\in\;]0,(T/2)\wedge 1]$.
Denote $V(t,x):=u(t+\epsilon,x)$. By the Markov property, $V(t,x)$ solves \eqref{E:FracHt} with
the time-shifted noise $\dot{W}_\epsilon (t,x):=\dot{W}(t+\epsilon,x)$ starting from $V(0,x)=u(\epsilon,x)$.
Recall the integral form $ V(t,x)= \widetilde{J}_0(t,x) + \widetilde{I}(t,x)$ in \eqref{E:VInt}.
\paragraph{Time increments}
Recall that $u(t,x)=J_0(t,x)+ I(t,x)$.
Let $0<\epsilon\le t\le t'\le T-\epsilon$.
So
\begin{align*}
\Norm{I(t+\epsilon,x)-I(t'+\epsilon,x)}_p^2
\le& \quad 2\Norm{u(t+\epsilon,x)-u(t'+\epsilon,x)}_p^2\\
&+ 2\left|J_0(t+\epsilon,x)-J_0(t'+\epsilon,x)\right|^2,
\end{align*}
with
\begin{align*}
\Norm{u(t+\epsilon,x)-u(t'+\epsilon,x)}_p^2 =&
\Norm{V(t,x)-V(t',x)}_p^2\\
\le &
2\Norm{\widetilde{I}(t,x)-\widetilde{I}(t',x)}_p^2 + 2\Norm{\widetilde{J}_0(t,x)-\widetilde{J}_0(t',x)}_p^2.
\end{align*}
Notice that for all $p\ge 2$, by the Burkholder-Davis-Gundy inequality (see \cite[Lemma 3.3]{ChenDalang13Heat}),
\begin{align*}
\Norm{\widetilde{I}(t,x)-\widetilde{I}\left(t',x\right)}_p^2
\le 2 z_p^2 \Lip_\rho^2 I_1\left(t,t',x\right)
+
2 z_p^2 \Lip_\rho^2 I_2\left(t,t',x\right)\;,
\end{align*}
where $z_p\le 2\sqrt{p}$ and $z_2=1$, and
\begin{gather*}
I_1\left(t,t',x\right)=\iint_{\left[0,t\right]\times\mathbb{R}}\ensuremath{\mathrm{d}} s \ensuremath{\mathrm{d}} y\:
\left(\lMr{\delta}{G}{a}\left(t-s,x-y\right)-\lMr{\delta}{G}{a}(t'-s,x-y) \right)^2 \left(\Vip^2+\Norm{V\left(s,y\right)}_p^2\right),\\
I_2\left(t,t',x\right)
= \iint_{\left[t,t'\right]\times\mathbb{R}}\ensuremath{\mathrm{d}} s \ensuremath{\mathrm{d}} y\:
\lMr{\delta}{G}{a}^2\left(t'-s,x-y\right)\left(\Vip^2+ \Norm{V\left(s,y\right)}_p^2\right).
\end{gather*}
By part (2) of Lemma \ref{L:MBds}, for some finite constant $Q:=Q(a,\delta,\Lip_\rho,\Vip,\mu,p,\epsilon,T)>0$,
\[
\sup_{(s,y)\in[0,t]\times\mathbb{R}} \Norm{V(s,y)}_p^2
=\sup_{(s,y)\in[\epsilon,t+\epsilon]\times\mathbb{R}} \Norm{u(s,y)}_p^2
\le Q.
\]
Then apply \cite[Proposition 4.4]{ChenDalang14FracHeat} to see that for some finite constant $C_1=C_1(a,\delta)>0$,
\begin{align}\label{E:ItidT}
\Norm{\widetilde{I}(t,x)-\widetilde{I}\left(t',x\right)}_p^2
\le C_1 z_p^2 \Lip_\rho^2 Q
\: |t'-t|^{1-1/a}.
\end{align}
By Minkowski's integral inequality and \eqref{E:timediff} below, for some finite constant $C_2:=C_2(a)>0$,
\begin{align*}
\Norm{\widetilde{J}_0(t,x)-\widetilde{J}_0(t',x)}_p^2
\le & \sup_{y\in\mathbb{R}}\Norm{u(\epsilon,y)}_p^2 \left(\int_\mathbb{R} \ensuremath{\mathrm{d}} y \: \left|\lMr{\delta}{G}{a}(t,y)-\lMr{\delta}{G}{a}(t',y)\right|\right)^2\\
\le & C_2 \: Q\: \left[\log(t'/t)\right]^2\le C_2 \: t^{-2} Q\: |t'-t|^2,
\end{align*}
where in the last step, we have applied the inequality $\log(1+x)\le x$ for $x>-1$.
Because $|t'-t|\le T^{\frac{a+1}{2a}} |t'-t|^{\frac{a-1}{2a}}$ and $t\ge \epsilon$, we have that
\begin{align}\label{E:JTilde}
\Norm{\widetilde{J}_0(t,x)-\widetilde{J}_0(t',x)}_p^2 \le
C_2\: \epsilon^{-2} T^{\frac{a+1}{a}} \: Q\: |t'-t|^{\frac{a-1}{a}}.
\end{align}
Similarly,
\begin{align*}
\left|J_0(t+\epsilon,x)-J_0(t'+\epsilon,x)\right|^2 \le &\sup_{y\in\mathbb{R}} J_0^2(\epsilon,y)
\left(\int_\mathbb{R}\ensuremath{\mathrm{d}} y \left| \lMr{\delta}{G}{a}(t,x)-\lMr{\delta}{G}{a}(t',x)\right|\right)^2\\
\le& C_2\: \epsilon^{-2} \: T^{\frac{a+1}{a}} \: Q\: |t'-t|^{\frac{a-1}{a}}.
\end{align*}
\paragraph{Space increments}
Fix $t\ge \epsilon$. Let $x, x'\in [-T,T]$. Then
\begin{align*}
\Norm{I(t+\epsilon,x)-I(t+\epsilon,x')}_p^2
\le& \quad 2\Norm{u(t+\epsilon,x)-u(t+\epsilon,x')}_p^2\\
&+ 2\left|J_0(t+\epsilon,x)-J_0(t+\epsilon,x')\right|^2,
\end{align*}
with
\begin{align*}
\Norm{u(t+\epsilon,x)-u(t+\epsilon,x')}_p^2 \le &
2\Norm{\widetilde{I}(t,x)-\widetilde{I}(t,x')}_p^2 + 2\Norm{\widetilde{J}_0(t,x)-\widetilde{J}_0(t,x')}_p^2.
\end{align*}
For $p\ge 2$, by the Burkholder-Davis-Gundy inequality and \cite[Proposition 4.4]{ChenDalang14FracHeat},
\begin{align*}
\Norm{\widetilde{I}(t,x)-\widetilde{I}\left(t,x'\right)}_p^2
\le& 2 z_p^2 \Lip_\rho^2 \iint_{\left[0,t\right]\times\mathbb{R}}\ensuremath{\mathrm{d}} s \ensuremath{\mathrm{d}} y\:
\left(\lMr{\delta}{G}{a}\left(t-s,x-y\right)-\lMr{\delta}{G}{a}(t-s,x'-y) \right)^2\\
&\hspace{3em}\times\left(\Vip^2+\Norm{V\left(s,y\right)}_p^2\right)\\
\le & 2 z_p^2 \Lip_\rho^2 Q |x'-x|^{a-1}.
\end{align*}
By the Minkowski's integral inequality and \eqref{E:GL1}, for some finite constant $C_3:=C_3(a)>0$,
\begin{align}\notag
\Norm{\widetilde{J}_0(t,x)-\widetilde{J}_0(t,x')}_p^2
\le & \sup_{y\in\mathbb{R}}\Norm{u(\epsilon,y)}_p^2 \left(\int_\mathbb{R} \ensuremath{\mathrm{d}} y \: \left|\lMr{\delta}{G}{a}(t,x-y)-\lMr{\delta}{G}{a}(t,x'-y)\right|\right)^2\\
\le & C_3 \: Q\: t^{-1/a}|x'-x|\le C_3 \: Q\: \epsilon^{-1/a}(2T)^{2-a}|x'-x|^{a-1}.
\label{E:JTildeX}
\end{align}
Similarly,
\begin{align*}
\left|J_0(t+\epsilon,x)-J_0(t+\epsilon,x')\right|^2 \le &\sup_{y\in\mathbb{R}} J_0^2(\epsilon,y)
\left(\int_\mathbb{R}\ensuremath{\mathrm{d}} y \left| \lMr{\delta}{G}{a}(t,x)-\lMr{\delta}{G}{a}(t,x')\right|\right)^2\\
\le& C_3 \: Q\: \epsilon^{-1/a}(2T)^{2-a}|x'-x|^{a-1}.
\end{align*}
Finally, combining the two cases, we see that for all compact sets $D\subseteq \mathbb{R}_+^*\times\mathbb{R}$, one can find
$T>0$, $\epsilon\in\;]0,(T/2)\wedge 1]$, such that
$D\subseteq K(\epsilon,T):=[2\epsilon,T]\times[-T,T]$. There is
some finite constant $Q':=Q'(a,\delta,\Lip_\rho,\Vip,\mu,p,\epsilon,T)>0$ such that
for all $(t,x)$ and $(t',x')\in D$,
\begin{align*}
\Norm{I(t,x)-I(t',x')}_p^2\le& Q'\left(|t'-t|^{1-1/a}+|x'-x|^{a-1}\right).
\end{align*}
Then the H\"older continuity follows from Kolmogorov's continuity theorem (see \cite[Theorem 1.4.1]{Kunita90Flow} and \cite[Proposition 4.2]{ChenDalang13Holder}).
Note that $J_0(t,x)$ belongs to $C^{\infty}(\mathbb{R}_+^*\times\mathbb{R})$ (see \cite[Lemma 4.9]{ChenDalang14FracHeat}).
This completes the proof of Theorem \ref{T:Holder}. \hfill$\square$
\section{Proof of Theorem \ref{T:WeakSol}} \label{S:WeakSol}
The case when $a=2$ is proved in \cite[Proposition 3.4]{ChenDalang13Holder}.
Assume that $1<a<2$. Fix $\phi\in C_c(\mathbb{R})$. For simplicity, we only prove the case where $\rho(u)=\lambda u$ and $\mu\ge 0$.
As in the proof \cite[Proposition 3.4]{ChenDalang13Holder},
we only need to prove that
\[
\lim_{t\rightarrow 0_+} \int_\mathbb{R} \ensuremath{\mathrm{d}} x \: I(t,x) \phi(x) = 0 \quad\text{in
$L^2(\Omega)$}.
\]
Denote $L(t):=\int_\mathbb{R} I(t,x) \phi(x)\ensuremath{\mathrm{d}} x$.
By the stochastic Fubini theorem (see \cite[Theorem 2.6, p. 296]{Walsh86}),
whose assumptions are easily checked,
\[
L(t) = \int_0^t \int_\mathbb{R} \left(\int_\mathbb{R} \ensuremath{\mathrm{d}} x\; \lMr{\delta}{G}{a}(t-s,x-y) \phi(x)\right)
\rho(u(s,y)) W(\ensuremath{\mathrm{d}} s,\ensuremath{\mathrm{d}} y).
\]
Hence, by It\^{o}'s isometry,
\begin{align*}
\mathbb{E}\left[L(t)^2\right]=
\lambda^2 \int_0^t \ensuremath{\mathrm{d}} s \int_\mathbb{R} \ensuremath{\mathrm{d}} y
\left(\int_{\mathbb{R}} \ensuremath{\mathrm{d}} x\; \lMr{\delta}{G}{a}(t-s,x-y)\phi(x)\right)^2
\Norm{u(s,y)}_2^2 .
\end{align*}
Assume that $t\le 1$. Since for some constant $C>0$, $|\phi(x)|\le C \lMr{\delta}{G}{a}(1,x)$ for all $x\in\mathbb{R}$, we can apply the semigroup property to get
\begin{align*}
\mathbb{E}\left[L(t)^2\right]\le
C^2\lambda^2\Lambda \int_0^t \ensuremath{\mathrm{d}} s \frac{1}{(t+1-s)^{1/a}}
\int_\mathbb{R} \ensuremath{\mathrm{d}} y
\lMr{\delta}{G}{a}(t+1-s,y)
\Norm{u(s,y)}_2^2,
\end{align*}
where the constant $\Lambda$ is defined in \eqref{E:Cst-dLa}.
Apply the moment formula \eqref{E:MomUp},
\begin{align*}
\mathbb{E}\left[L(t)^2\right]\le
C^2\lambda^2\Lambda \left[L_1(t)+L_2(t)\right],
\end{align*}
with
\[
L_1(t):= \int_0^t\ensuremath{\mathrm{d}} s \frac{1}{(t+1-s)^{1/a}} \int_\mathbb{R} \ensuremath{\mathrm{d}} y \: J_0^2(s,y) \lMr{\delta}{G}{a}(t+1-s,y),
\]
and
\[
L_2(t):= \int_0^t\ensuremath{\mathrm{d}} s \frac{1}{(t+1-s)^{1/a}} \int_\mathbb{R} \ensuremath{\mathrm{d}} y \: \left(J_0^2\star\mathcal{K}\right)(s,y) \lMr{\delta}{G}{a}(t+1-s,y).
\]
We first consider $L_1(t)$. By \cite[(4.20)]{ChenDalang14FracHeat}, for some constant $C_1:=C_1(a,\delta,\mu)>0$,
$J_0(t,x)\le C_1 t^{-1/a}$. Thus,
\begin{align*}
L_1(t)\le &C_1 \int_0^t\ensuremath{\mathrm{d}} s \frac{1}{(t+1-s)^{1/a}s^{1/a}} \int_\mathbb{R} \ensuremath{\mathrm{d}} y \: J_0(s,y) \lMr{\delta}{G}{a}(t+1-s,y)\\
=&C_1 J_0(t+1,0) \int_0^t\ensuremath{\mathrm{d}} s \frac{1}{(t+1-s)^{1/a}s^{1/a}} \\
\le &C_1 J_0(t+1,0) \int_0^t\ensuremath{\mathrm{d}} s \frac{1}{(1-s)^{1/a}s^{1/a}} \rightarrow 0, \quad\text{as $t\rightarrow 0$.}
\end{align*}
The case for $L_2(t)$ can be proved in a similar way, where one needs to apply \eqref{E:UpBd-K}. We leave the details for interested readers. This completes the proof of Theorem
\ref{T:WeakSol}. \hfill$\square$
\section*{Appendix}
Recall the kernel function $\tlMr{\epsilon}{\delta}{R}{a}(t,x)$ defined in \eqref{E:R}.
The following two lemmas \ref{L:RGaprxL1} and \ref{L:RGaprx} below show that $\tlMr{\epsilon}{\delta}{R}{a}(t,x)$
is an approximation of $\lMr{\delta}{G}{a}(t,x)$. The proofs of both Lemmas depend on Lemma \ref{L:GH3F3} below.
\begin{lemma}\label{L:GH3F3}
For all $b\in \mathbb{R}$,
\begin{align}\label{E:GH3F3}
\lim_{z\rightarrow\infty} e^{-z} z^{b+1} \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^b} =1.
\end{align}
If $b\ge -1$, then
\begin{align}\label{E:GH3F3CB}
C_b:= \sup_{z\ge 0} e^{-z} z^{b+1} \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^b}<+\infty.
\end{align}
\end{lemma}
Note that when $b\in\mathbb{N}$, the series in \eqref{E:GH3F3} converges to the {\it generalized hypergeometric function} (see \cite[Chapter
16]{NIST2010}):
\[
\sum_{k=1}^\infty\frac{z^{k-1}}{k!\: k^b}=\lMr{b}{F}{b}\Big((\:\underbrace{1,\dots,1}_{\text{$b+1$}}\:),(\:\underbrace{2,\dots,2}_{\text{$b+1$}}\:);z\Big),\quad\text{for $b\in\mathbb{N}$ and $z\in \mathbb{C}$.}
\]
\begin{proof}[Proof of Lemma \ref{L:GH3F3}]
Clearly, the series converges on $z\in\mathbb{C}$ and it defines an entire function.
We first assume that $b\in \mathbb{N}$. We will prove \eqref{E:GH3F3} by induction. Clearly, the case $b=0$ is true.
Suppose that \eqref{E:GH3F3} is true for $b$. Now let us consider the case $b+1$: Applying
l'H\^{o}pital's rule and the induction assumption, we obtain
\begin{equation}\label{E:lHopital}
\begin{aligned}
\lim_{z\rightarrow\infty} e^{-z} z^{b+2} \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^{b+1}}&=
\lim_{z\rightarrow\infty}\frac{\sum_{k=1}^\infty \frac{z^{k}}{k! \: k^{b+1}}}{\frac{e^z}{z^{b+1}}}
=
\lim_{z\rightarrow\infty}\frac{\sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^{b}}}{\frac{e^z(z^{b+1}-(b+1)z^b)}{z^{2(b+1)}}}\\
&=
\lim_{z\rightarrow\infty} e^{-z} z^{b+1} \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^b}=1.
\end{aligned}
\end{equation}
This proves Lemma \ref{L:GH3F3} for $b\in\mathbb{N}$.
Now assume that $1/2 \leq b<3/2$.
Because the function $f(x)=x^b$ for $x\ge 1$ is either concave or convex,
we have that for all $k\ge 1$ and $z\ge 1$,
\[
\left|z^b - k^b\right| \le b \left(z^{b-1}\vee k^{b-1}\right) |k-z|
\le
b \left(z^{b-1}+ k^{b-1}\right) |k-z|.
\]
Hence, for all $z\ge 1$ and $k\ge 1$,
\begin{align}
\left|\frac{1}{k^b}-\frac{1}{z^b}\right| =
\frac{\left|z^b-k^b\right|}{z^b \:k^b} \le
b |k-z|\left(\frac{1}{z\: k^b}+\frac{1}{z^b k}\right)
\le
b|k-z|\left(\frac{1}{z \: k^{1/2}}+\frac{1}{z^{1/2} k}\right).
\label{E:DeltaB}
\end{align}
Thus, for $z\ge 1$, by Cauchy-Schwartz inequality and \eqref{E:Square},
\begin{align*}
\left|\sum_{k=2}^\infty \frac{z^k}{k! k^b} - \frac{1}{z^b}\left(e^z-1-z\right)\right|&=
\left|\sum_{k=2}^\infty \frac{z^k}{k! }\left(\frac{1}{k^{b}}-\frac{1}{z^{b}}\right)\right|\\
&\le
\frac{b}{ z} \sum_{k=2}^\infty \frac{z^k}{k!}\frac{|k-z|}{\sqrt{k}}+ \frac{b}{\sqrt{z}} \sum_{k=2}^\infty \frac{z^k}{k!}\frac{|k-z|}{ k}\\
& \le
\frac{b}{z} \left(\sum_{k=0}^\infty \frac{z^k}{k!}|k-z|^2\right)^{1/2}\left(\sum_{k=1}^\infty \frac{z^k}{k! \:k}\right)^{1/2} \\
& \quad+
\frac{b}{\sqrt{z}} \left(\sum_{k=0}^\infty \frac{z^k}{k!}|k-z|^2\right)^{1/2}\left(\sum_{k=1}^\infty \frac{z^k}{k! \:k^2}\right)^{1/2} \\
&=
\frac{b}{z^{3/2}}\left(e^{z} z^2\sum_{k=1}^\infty \frac{z^{k-1}}{k! \:k}\right)^{1/2}
+
\frac{b}{z^{3/2}}\left(e^{z} z^3\sum_{k=1}^\infty \frac{z^{k-1}}{k! \:k^2}\right)^{1/2}
\end{align*}
Hence, by the previous proof for the case $b\in\mathbb{N}$, and because $b<3/2$,
\[
\lim_{z\rightarrow \infty} z^{b}e^{-z} \left|\sum_{k=2}^\infty \frac{z^k}{k! k^{b}} - \frac{1}{z^{b}}\left(e^z-1-z\right)\right|
\le
\lim_{z\rightarrow \infty}\frac{2 b}{ z^{3/2-b}} =0.
\]
Therefore,
\[
\lim_{z\rightarrow\infty} z^{b} e^{-z} \sum_{k=2}^\infty \frac{z^k}{k!\; k^{b}}
= \lim_{z\rightarrow\infty} z^{b} e^{-z} \frac{1}{z^{b}}\left(e^z-1-z\right)=1.
\]
Now assume that $b<1/2$. Let $c\in [1/2,3/2[\;$ and $n\in\mathbb{N}$ such that $b+n=c$.
Then apply l'H\^{o}pital's rule $n$ times as in \eqref{E:lHopital},
\[
1=\lim_{z\rightarrow\infty} e^{-z} z^{c+1} \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^{c}}
=\lim_{z\rightarrow\infty} e^{-z} z^{c} \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^{c-1}}
=\cdots
=\lim_{z\rightarrow\infty} e^{-z} z^{c-n+1} \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^{c-n}}.
\]
Similarly, if $b\ge 3/2$, then let $c\in [1/2,3/2[\;$ and $n\in\mathbb{N}$ such that $b=c+n$.
Then apply l'H\^{o}pital's rule $n$ times as in \eqref{E:lHopital},
\[
\lim_{z\rightarrow\infty} e^{-z} z^{b+1} \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^{b}}
=\lim_{z\rightarrow\infty} e^{-z} z^{b} \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^{b-1}}
=\cdots
=\lim_{z\rightarrow\infty} e^{-z} z^{b-n+1} \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^{b-n}}=1.
\]
This proves \eqref{E:GH3F3} for all $b\in\mathbb{R}$.
Finally, \eqref{E:GH3F3CB} follows from the fact that the function $f(z)=z^{b+1} e^{-z} \sum_{k=1}^\infty \frac{z^{k-1}}{k!\; k^{b}}$ is continuous over $\mathbb{R}_+\cup\{+\infty\}$ with $f(\infty)=1$ and
$|f(0)|<\infty$ if $b\ge -1$ (actually, $f(0)=0$ if $b>-1$ and $f(0)=1$ if $b=-1$).
This completes the proof of Lemma \ref{L:GH3F3}.
\end{proof}
\begin{lemma}\label{L:RGaprxL1}
There exists a finite constant $C>0$ such that
\[
\int_\mathbb{R} \ensuremath{\mathrm{d}} x \: \left|\tlMr{\epsilon}{\delta}{R}{a}(t,x)-\lMr{\delta}{G}{a}(t,x)\right|\le
e^{-t/\epsilon}+C\left(\frac{\epsilon}{t}\right)^{1/2},\quad\text{for all $\epsilon>0$ and $t>0$,}
\]
where the constant $C$ can be chosen as
\begin{align}
\label{E:RGaprxL1}
C=\frac{1}{a}\left(1+K_{a,1}\Gamma\left(\frac{a}{a+2}\right)\Gamma\left(\frac{a+4}{a+2}\right)\right)
\left[\sup_{z\ge 0}
e^{-z}z (4 z^2+7 z+1) \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^2}\right]^{\frac{1}{2}},
\end{align}
with the constant $K_{a,1}$ defined in \cite[(4.3)]{ChenDalang14FracHeat}.
\end{lemma}
\begin{proof}
From \cite[(4.3)]{ChenDalang14FracHeat},
\begin{align}\notag
\left|\frac{\partial }{\partial t} \lMr{\delta}{G}{a}(t,x)\right| =& \left|-\frac{1}{at}\left(\lMr{\delta}{G}{a}(t,x)+ x \frac{\partial
\lMr{\delta}{G}{a}(t,x)}{\partial x}\right)\right|\\
\notag
&
\le
\frac{1}{at}\left(\lMr{\delta}{G}{a}(t,x)+ \left|x \frac{\partial
\lMr{\delta}{G}{a}(t,x)}{\partial x}\right|\right)\\
\label{E:DtG}
&\le
\frac{1}{at}\left[\lMr{\delta}{G}{a}(t,x)+t^{-\frac{2}{a}}\frac{K_{a,1}|x|}{1+|t^{-1/a} x|^{2+a}}\right].
\end{align}
Thus, for $0<t\le t'$,
\begin{equation}\label{E:timediff}
\begin{aligned}
\int_\mathbb{R} \ensuremath{\mathrm{d}} x \left|\lMr{\delta}{G}{a}(t',x)-\lMr{\delta}{G}{a}(t,x)\right| &
\le \int_\mathbb{R}\ensuremath{\mathrm{d}} x \int_{t}^{t'}\ensuremath{\mathrm{d}} s \left|\frac{\partial }{\partial t} \lMr{\delta}{G}{a}(s,x)\right|\\&
\le \int_t^{t'}\ensuremath{\mathrm{d}} s \: \frac{1}{a s} \left(1+\int_\mathbb{R} \ensuremath{\mathrm{d}} y \frac{K_{a,1}|y|}{1+|y|^{2+a}}\right)
\le C' \log\left(\frac{t'}{t}\right),
\end{aligned}
\end{equation}
where
\begin{equation}\label{E:constantc'}
C':=\frac{1}{a}\left(1+\int_\mathbb{R}\ensuremath{\mathrm{d}} y \frac{K_{a,1}|y|}{1+|y|^{2+a}}\right)=
\frac{1}{a}\left(1+K_{a,1}\Gamma\left(\frac{a}{a+2}\right)\Gamma\left(\frac{a+4}{a+2}\right)\right),
\end{equation}
and the integral in \eqref{E:constantc'} is evaluated by Lemma \ref{L:Beta}.
Notice that
\[
\left|\tlMr{\epsilon}{\delta}{R}{a}(t,x)-\lMr{\delta}{G}{a}(t,x)\right|
\le e^{-t/\epsilon} \lMr{\delta}{G}{a}(t,x)
+ e^{-t/\epsilon} \sum_{k=1}^\infty \left(\frac{t}{\epsilon}\right)^k \frac{1}{k!}
\left|\lMr{\delta}{G}{a}(t,x)-\lMr{\delta}{G}{a}(k\epsilon,x)\right|.
\]
By the above inequality, we have that
\[
\int_\mathbb{R}\ensuremath{\mathrm{d}} x\left|\tlMr{\epsilon}{\delta}{R}{a}(t,x)-\lMr{\delta}{G}{a}(t,x)\right|
\le e^{-t/\epsilon} +C' e^{-t/\epsilon}\sum_{k=1}^\infty \left(\frac{t}{\epsilon}\right)^k \frac{1}{k!}
\left|\log (k\epsilon/t)\right|.
\]
Denote the summation over $k$ in above upper bound by $I(t/\epsilon)$.
Because the function $x\mapsto \log(x)$ is concave, $|\log(t'/t)|\le |t'-t|\left(\frac{1}{t'}\vee \frac{1}{t}\right)\le |t'-t|\left(\frac{1}{t'}+ \frac{1}{t}\right)$. So, by letting $z=t/\epsilon$,
\[
I(z) \le \sum_{k=1}^\infty \frac{z^k}{k!}\left|k-z\right|\left(\frac{1}{k}+\frac{1}{z}\right)
=\sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k}\left|k^2-z^2\right|.
\]
Then by Cauchy-Schwartz inequality,
\[
I(z)\le
\left(\sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^2}\right)^{1/2}
\left(\sum_{k=1}^\infty \frac{z^{k-1}}{k! }\left[k^2-z^2\right]^2\right)^{1/2}.
\]
Notice that
\begin{align}\label{E:ExpZ}
\sum_{k=1}^\infty \frac{z^{k-1}}{k! }\left[k^2-z^2\right]^2=
e^z( 4 z^2+7 z+1)-z^3 \le e^z( 4 z^2+7 z+1).
\end{align}
To prove the equality in \eqref{E:ExpZ}, one can write $k^2= P_k^2+P_k^1$ and
$k^4=P_k^4+6P_k^3+7P_k^2+P_k^1$, where $P_k^n:= k (k-1)\cdots(k-n+1)$.
Hence, $(k^2-z^2)^2=P_k^4+6P_k^3+(7-2 z^2)P_k^2+(1-2z^2)P_k^1+z^4$.
Then use the fact that for $n\ge 1$,
$\sum_{k=1}^\infty \frac{z^{k}}{k!} P_k^n=e^z z^n$.
Therefore,
\[
I(z)e^{-z}\le
\left(e^{-z}z (4 z^2+7 z+1) \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^2}\right)^{1/2} \: \frac{1}{\sqrt{z}}.
\]
By Lemma \ref{L:GH3F3}, the function $f(z)=e^{-z}z (4 z^2+7 z+1) \sum_{k=1}^\infty \frac{z^{k-1}}{k! \: k^2}$ is continuous over $\mathbb{R}_+\cup\{+\infty\}$ with $f(0)=0$ and $f(\infty)=4$. Thus,
$\sup_{z\ge 0}f(z)<+\infty$.
This completes the proof of Lemma \ref{L:RGaprxL1}.
\end{proof}
\begin{lemma}\label{L:RGaprx}
We have that
\begin{gather}
\label{E:RGaprx0}
\lim_{\epsilon\rightarrow 0}\tlMr{\epsilon}{\delta}{R}{a}(t,x) = \lMr{\delta}{G}{a}(t,0) \Indt{x= 0},
\\
\label{E:RGaprx1}
\lim_{\epsilon\rightarrow 0}\int_0^t\ensuremath{\mathrm{d}} s \int_\mathbb{R}\ensuremath{\mathrm{d}} x
\left[\tlMr{\epsilon}{\delta}{R}{a}(s,x) - \lMr{\delta}{G}{a}(s,x)\right]^2=0,
\quad\text{for all $t>0$,}
\end{gather}
and there is a nonnegative constant $C_{a,\delta}<+\infty$ such that
\begin{align}\label{E:RGaprx2}
\int_\mathbb{R} \ensuremath{\mathrm{d}} x \tlMr{\epsilon}{\delta}{R}{a}^2(t,x)\le C_{a,\delta} \: t^{-\frac{1}{a}},\quad\text{for all $t>0$.}
\end{align}
\end{lemma}
\begin{proof}
Fix $t>0$. Denote $A:= \lMr{\delta}{G}{a}(1,0)$.
Clearly,
\[
\int_0^t\ensuremath{\mathrm{d}} s \int_\mathbb{R}\ensuremath{\mathrm{d}} x
\left[\tlMr{\epsilon}{\delta}{R}{a}(s,x) - \lMr{\delta}{G}{a}(s,x)\right]^2
= I_1(t,\epsilon) -2 I_2(t,\epsilon) + I_3(t),
\]
where
\begin{align}
I_1(t,\epsilon) &= \iint_{[0,t]\times\mathbb{R}} \ensuremath{\mathrm{d}} s\ensuremath{\mathrm{d}} x \tlMr{\epsilon}{\delta}{R}{a}^2(s,x), \\
I_2(t,\epsilon) &= \iint_{[0,t]\times\mathbb{R}} \ensuremath{\mathrm{d}} s\ensuremath{\mathrm{d}} x \tlMr{\epsilon}{\delta}{R}{a}(s,x)\lMr{\delta}{G}{a}(s,x), \\
I_3(t) &= \iint_{[0,t]\times\mathbb{R}} \ensuremath{\mathrm{d}} s\ensuremath{\mathrm{d}} x \lMr{\delta}{G}{a}^2(s,x).
\end{align}
By the semigroup property and scaling property \cite[(4.1)]{ChenDalang14FracHeat}, we have that
\begin{align}\label{E_:I3}
I_3(t) =\int_0^t \ensuremath{\mathrm{d}} s \; \lMr{\delta}{G}{a}(2s,0)=
A \int_0^t (2s)^{-\frac{1}{a}}\ensuremath{\mathrm{d}} s = \frac{a A}{2^{1/a}(a-1)}\: t^{1-\frac{1}{a}}.
\end{align}
{\vspace{0.7em}\noindent \bf Step 1.} We first calculate $I_1$. Use the semigroup property and scaling property \cite[(4.1)]{ChenDalang14FracHeat}:
\begin{align*}
I_1(t,\epsilon) &=\int_0^t \ensuremath{\mathrm{d}} s \: e^{-2s/\epsilon}
\sum_{n=1}^\infty
\sum_{m=1}^\infty\left(\frac{s}{\epsilon}\right)^{n+m}\frac{1}{n! m!} \lMr{\delta}{G}{a}((n+m)\epsilon,0)\\
&=
A \int_0^t \ensuremath{\mathrm{d}} s \: e^{-2s/\epsilon}
\sum_{n=1}^\infty
\sum_{m=1}^\infty\left(\frac{s}{\epsilon}\right)^{n+m}\frac{1}{n! m! (n+m)^{1/a} \epsilon^{1/a}}.
\end{align*}
Then by change of variables $u=s/\epsilon$ and let $z=t/\epsilon$, we have that
\[
I_1(t,\epsilon) = A t^{1-\frac{1}{a}}
\frac{1}{z^{1-\frac{1}{a}}}\int_0^z \ensuremath{\mathrm{d}} u\: e^{-2u}\sum_{n,m=1}^\infty\frac{u^{n+m}}{n!m!(n+m)^{1/a}}.
\]
By l'H\^{o}pital's rule,
\[
I_1(t):=\lim_{\epsilon\rightarrow 0}I_1(t,\epsilon)
= \frac{a A}{a-1} t^{1-\frac{1}{a}} \lim_{z\rightarrow\infty} z^{1/a} e^{-2z}
\sum_{n,m=1}^\infty\frac{z^{n+m}}{n!m!(n+m)^{1/a}}.
\]
Because $\sum_{n=1}^{k-1}\frac{1}{n!(k-n)!} = \frac{1}{k!}(2^k-2)$,
\[
\sum_{n,m=1}^\infty\frac{z^{n+m}}{n!m!(n+m)^{1/a}}
=\sum_{k=2}^\infty \frac{z^k}{k^{1/a}}\sum_{n=1}^{k-1} \frac{1}{n!(k-n)!}
=\sum_{k=2}^\infty \frac{z^k}{k! k^{1/a} }(2^k-2).
\]
Hence,
\begin{align}\label{E_:I1}
I_1(t)=\frac{a A}{a-1} t^{1-\frac{1}{a}} \lim_{z\rightarrow\infty}
z^{1/a} e^{-2z} \sum_{k=2}^\infty \frac{z^k}{k! k^{1/a} }(2^k-2)
=\frac{a A}{2^{1/a}(a-1)}\: t^{1-\frac{1}{a}},
\end{align}
where the last equality is due to Lemma \ref{L:GH3F3} with $b=1/a\in [1/2,1]$.
{\vspace{0.7em}\noindent \bf Step 2.} In this step, we calculate $I_2(t):= \lim_{\epsilon\rightarrow\infty} I_2(t,\epsilon)$. Similarly to the Step 1, use the semigroup property and scaling property
\cite[(4.1)]{ChenDalang14FracHeat}, and then change the variables $u=s/\epsilon$ and $z=t/\epsilon$,
\begin{align*}
I_2(t,\epsilon) &=\int_0^t \ensuremath{\mathrm{d}} s\: e^{-s/\epsilon}\sum_{n=1}^\infty
\left(\frac{s}{\epsilon}\right)^n \frac{1}{n!}\lMr{\delta}{G}{a}(s+n\epsilon,0)= A t^{1-\frac{1}{a}} \frac{1}{z^{1-\frac{1}{a}}} \int_0^z\ensuremath{\mathrm{d}} u\; e^{-u} \sum_{n=1}^\infty
\frac{u^n}{n! (u+n)^{1/a}}.
\end{align*}
By l'H\^{o}pital's rule,
\[
I_2(t)= \frac{a A}{a-1} t^{1-\frac{1}{a}} \lim_{z\rightarrow\infty} z^{1/a} e^{-z}
\sum_{n=1}^\infty\frac{z^{n}}{n!(z+n)^{1/a}}.
\]
Apply the inequality \eqref{E:DeltaB} below with $b=1/a\in [1/2,1]$,
\[
\left|\frac{1}{(z+n)^{1/a}}-\frac{1}{(2z)^{1/a}}\right|
\le \frac{1}{a} |n-z| \left(\frac{1}{2z \sqrt{n+z}}+\frac{1}{(n+z)\sqrt{2z}}\right)
\le \frac{1+\sqrt{2}}{2a}\:\frac{|n-z|}{z^{3/2}},
\]
for $z\ge 1$ and $n\ge 1$.
Notice that (see the proof of \eqref{E:ExpZ}),
\begin{align}\label{E:Square}
\sum_{n=0}^\infty \frac{z^n}{n!} |n-z|^2= e^z z.
\end{align}
By Cauchy-Schwartz inequality and \eqref{E:Square}, for $z\ge 1$,
\begin{align*}
\left|\sum_{n=1}^\infty \frac{z^n}{n! (z+n)^{1/a}} - \frac{1}{(2z)^{1/a}}(e^z-1)\right|
& =
\left|\sum_{n=1}^\infty \frac{z^n}{n!}\left(\frac{1}{(z+n)^{1/a}} - \frac{1}{(2z)^{1/a}}\right)\right|\\
&\le
\frac{1+\sqrt{2}}{2a z^{3/2}}\sum_{n=1}^\infty \frac{z^n}{n!} |n-z|\\
&\le
\frac{1+\sqrt{2}}{2a z^{3/2}}\left(\sum_{n=0}^\infty \frac{z^n}{n!} |n-z|^2\right)^{1/2}\left(\sum_{n=0}^\infty \frac{z^n}{n!}\right)^{1/2}\\
&=
\frac{1+\sqrt{2}}{2a z}e^z.
\end{align*}
Therefore,
\begin{align}
\label{E_:I2}
I_2(t)= \frac{a A}{a-1} t^{1-\frac{1}{a}} \lim_{z\rightarrow\infty} z^{1/a} e^{-z}
\frac{1}{(2z)^{1/a}}(e^z-1) =
\frac{a A}{2^{1/a}(a-1)}\: t^{1-\frac{1}{a}}.
\end{align}
Finally, \eqref{E:RGaprx1}
is proved by combining \eqref{E_:I3}, \eqref{E_:I1} and \eqref{E_:I2}.
{\vspace{0.7em}\noindent \bf Step 3.} Now we prove \eqref{E:RGaprx0}.
Clearly, if $x\ne 0$, then $\lim_{\epsilon\rightarrow 0}\tlMr{\epsilon}{\delta}{R}{a}(t,x) =0$. Otherwise, by Lemma \ref{L:GH3F3} with $b=1/a$,
\[
\lim_{\epsilon\rightarrow 0}\tlMr{\epsilon}{\delta}{R}{a}(t,0) =
\frac{A}{t^{1/a}} \lim_{z\rightarrow\infty} z^{1/a}e^{-z}\sum_{k=1}^\infty\frac{z^k}{k!k^{1/a}}=\lMr{\delta}{G}{a}(t,0).
\]
{\vspace{0.7em}\noindent \bf Step 4.} As for \eqref{E:RGaprx2}, denote $I(t;\epsilon)=\int_\mathbb{R} \ensuremath{\mathrm{d}} x \tlMr{\epsilon}{\delta}{R}{a}^2(t,x)$.
Following the arguments in Step 1,
\begin{align}\label{E_:sup}
I(t,\epsilon)\le \frac{A}{(2t)^{1/a}} \sup_{z\in\mathbb{R}_+}
z^{1/a} e^{-z} \sum_{k=2}^\infty \frac{z^k}{k! k^{1/a} }.
\end{align}
Clearly, the function $f(z)=\sum_{k=2}^\infty \frac{z^k}{k! k^{1/a}}$ is an entire function over $\mathbb{C}$. By Lemma \ref{L:GH3F3} and
$\lim_{z\rightarrow 0}z^{1/a}e^{-z}f(z)=0$, we know that the supremum in \eqref{E_:sup}, which depends only on $a$, is finite.
This completes the proof of Lemma \ref{L:RGaprx}
\end{proof}
\begin{lemma}\label{L:MInc}
If $\mu(\ensuremath{\mathrm{d}} x) =f(x)\ensuremath{\mathrm{d}} x$ with $f\in L^\infty(\mathbb{R})$, then for all $0<t\le T$ and $x$, $y\in\mathbb{R}$,
\[
\Norm{u(t,x)-u(t,y)}_2^2 \le C t^{-1/a}|x-y|+A_T C_1|x-y|^{a-1} ,
\]
where $K_{a,1}$ is defined in \cite[(4.3)]{ChenDalang14FracHeat}, $C_1:=C_1(a,\delta)$ is defined in \cite[Proposition 4.4]{ChenDalang14FracHeat}, and
\begin{align*}
C&:=8 K_{a,1}\: \Gamma\left(\frac{a+1}{a+2}\right)\Gamma\left(\frac{a+3}{a+2}\right) \sup_{x\in\mathbb{R}}[f(x)]^2,\quad A_T :=\sup_{s\in [0,T]}\sup_{x\in\mathbb{R}} \Norm{\rho(u(s,x))}_2^2.
\end{align*}
\end{lemma}
\begin{proof}
By It\^{o}'s isometry,
\begin{multline*}
\Norm{u(t,x)-u(t,y)}_2^2= \left[J_0(t,x)-J_0(t,y)\right]^2+
\int_0^t\ensuremath{\mathrm{d}} s\int_\mathbb{R} \ensuremath{\mathrm{d}} z \Norm{\rho(u(s,z))}_2^2 \\
\times \left[\lMr{\delta}{G}{a}(t-s,x-z)-\lMr{\delta}{G}{a}(t-s,y-z)\right]^2.
\end{multline*}
Denote $C_f:=\sup_{x\in\mathbb{R}} |f(x)|$ and fix $0<t<T$. Then
\begin{align*}
\Norm{u(t,x)-u(t,y)}_2^2\le & 2\: C_f^2 \int_\mathbb{R}\ensuremath{\mathrm{d}} z \left|\lMr{\delta}{G}{a}(t,x-z)-\lMr{\delta}{G}{a}(t,y-z)\right| \\
& +A_T \int_0^t\ensuremath{\mathrm{d}} s \int_\mathbb{R}\ensuremath{\mathrm{d}} z \left(\lMr{\delta}{G}{a}(s,x-z)-\lMr{\delta}{G}{a}(s,y-z)\right)^2.
\end{align*}
By \cite[Proposition 4.4]{ChenDalang14FracHeat}, for some constant $C_1:=C_1(a,\delta)$, the second part of the above upper bound is bounded by $A_T C_1 |x-y|^{a-1}$. As
for the first part,
notice that by \cite[(4.3)]{ChenDalang14FracHeat}, for all $t> 0$ and $x$, $y\in\mathbb{R}$,
\begin{align*}
\left|\lMr{\delta}{G}{a}(t,x)-\lMr{\delta}{G}{a}(t,y)\right|
& = \left|\int_x^y \frac{\partial }{\partial x}\lMr{\delta}{G}{a}(t,z) \ensuremath{\mathrm{d}} z\right|\le \int_x^y \left|\frac{\partial }{\partial x}\lMr{\delta}{G}{a}(t,z)\right| \ensuremath{\mathrm{d}} z\\
&\le K_{a,1} t^{-2/a} \int_x^y \frac{\ensuremath{\mathrm{d}} z}{1+|t^{-1/a}z|^{2+a}}\\
&\le K_{a,1} t^{-2/a} \frac{|x-y|}{1+\left(|t^{-1/a}x|\wedge|t^{-1/a}y|\right)^{2+a}}.
\end{align*}
Thus,
\begin{align*}
\int_\mathbb{R}\ensuremath{\mathrm{d}} z &\left|\lMr{\delta}{G}{a}(t,x-z)-\lMr{\delta}{G}{a}(t,y-z)\right|
\le K_{a,1} t^{-2/a} \int_\mathbb{R}\frac{|x-y|\:\ensuremath{\mathrm{d}} z}{1+\left(|t^{-1/a}(x-z)|\wedge |t^{-1/a}(y-z)|\right)^{2+a}}\\
& \le K_{a,1} t^{-2/a} |x-y| \int_\mathbb{R}\ensuremath{\mathrm{d}} z \left[\frac{1}{1+|t^{-1/a}(x-z)|^{2+a}}+\frac{1}{1+|t^{-1/a}(y-z)|^{2+a}}\right]\\
& = 2 K_{a,1} t^{-2/a} |x-y| \int_\mathbb{R} \frac{\ensuremath{\mathrm{d}} z}{1+|t^{-1/a} z|^{2+a}}=
2 K_{a,1} t^{-1/a} |x-y| \int_\mathbb{R} \frac{\ensuremath{\mathrm{d}} z}{1+|z|^{2+a}} .
\end{align*}
Hence, by letting
\[
C_a:=\int_\mathbb{R} \frac{\ensuremath{\mathrm{d}} z}{1+|z|^{2+a}} = 2\Gamma\left(\frac{a+1}{a+2}\right)\Gamma\left(\frac{a+3}{a+2}\right),
\]
where the integral is evaluated by Lemma \ref{L:Beta}, we have that
\begin{align}
\label{E:GL1}
\int_\mathbb{R}\ensuremath{\mathrm{d}} z \left|\lMr{\delta}{G}{a}(t,x-z)-\lMr{\delta}{G}{a}(t,y-z)\right| \le 2 K_{a,1} C_a t^{-1/a} |x-y|,\quad\text{for all $x$, $y\in\mathbb{R}$.}
\end{align}
This completes the proof of Lemma \ref{L:MInc}.
\end{proof}
\begin{lemma}\label{L:Beta}
For $a>0$ and $b\in \:]-1,a+1[\:$, $\int_0^\infty \ensuremath{\mathrm{d}} y \frac{y^b}{1+y^{2+a}} =\frac{1}{b+1}\Gamma\left(\frac{a-b+1}{a+2}\right)\Gamma\left(\frac{a+b+3}{a+2}\right)$.
\end{lemma}
\begin{proof}
Let $1+y^{2+a}=r^{-1}$. Then $\ensuremath{\mathrm{d}} y=-\frac{1}{a+2}r^{-\frac{a+3}{a+2}}(1-r)^{-\frac{a+1}{a+2}}\ensuremath{\mathrm{d}} r$. So,
\[
\int_0^1 \ensuremath{\mathrm{d}} r\: r \: r^{-\frac{b}{a+2}}(1-r)^{\frac{b}{a+2}} \frac{1}{a+2} r^{-\frac{a+3}{a+2}}(1-r)^{-\frac{a+1}{a+2}}
=
\frac{1}{a+2}\int_0^1 \ensuremath{\mathrm{d}} r \: r^{\frac{a+1-b}{a+2}-1} (1-r)^{\frac{b+1}{a+2}-1}.
\]
Then apply the Beta integral and use the recursion $x\Gamma(x)=\Gamma(x+1)$.
\end{proof}
\section*{Acknowledgements}
\addcontentsline{toc}{section}{Acknowledgements}
The authors appreciate many stimulating discussions and supports from Davar Khoshnevisan,
and especially his several suggestions on the proof of the strict comparison principle.
The authors also thank Carl Mueller for many helpful suggestions.
The authors thank Tom Alberts for some interesting discussions and for his pointing out the reference \cite{Moreno14Pos}.
The first author thanks Robert Dalang and Roger Tribe for many interesting discussions.
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$'${$'$}
\def$'${$'$}
|
\section{Introduction}
The explosive growth of mobile traffic is driving the rapid network densification to provide high-speed wireless access to coverage regions. To rein in the escalating network cost and interference, wireless access networks are evolving towards having an architecture with centralized signal processing and minimum on-site hardware comprising merely antennas and RF units, called either cloud radio access networks or base-station virtualization \cite{C-RAN}. The network evolution as well as the advancements of other technologies such as small cells \cite{ChanAndrews:FemtocellSurvey:2008} and massive MIMO \cite{RusekLarssonMarz:ScaleUpMIMO:2012} will lead to future networks where antennas are ubiquitous and have network-wide coordination to form a gigantic array, which is called the \emph{ubiquitous array} (UA) and forms the theme for the paper.
\subsection{Prior Work and New Challenges}
The UA system is equivalent to a \emph{distributed antennas system} (DAS) with dense antennas and without cells. DASs refers to cellular systems where in each cell, antennas are distributed over the cell region and connected to a central processing unit \cite{Heath:DistributedAntennas:2013}. This technology was initially developed to reduce transmission power and improve network coverage by either simulcast over all distributed antennas or serve each user using the nearest antennas \cite{Choi:DownlinkDASMultiCell:2007, Zhang:DASwithRandomAntennas:2008}. Recent research on DASs focuses on increasing the sum rate based on multiuser MIMO transmission using the distributed antennas as a virtual array, addressing issues such as inter-cell interference distribution \cite{Heath:DistributedAntennasSysOutOfCellInterference:2011}, resource allocation \cite{Zhu:RadioResourceAllocationDAS:2013}, capacity analysis \cite{KKWong:LargeScaleMACDistributedAnt:2013, McKay:SumRateAnalysisZFRXDistributedMIMO:2013, Lin:CompareCoLocatDistributedAntennas:2014} and multi-cell coordination \cite{Feng:VirtualMIMODistributedAntCoordTX:2013}. Each MIMO channel in such systems has coefficients corresponding to heterogeneous path loss depending on antenna locations and cannot be modeled as i.i.d. random variable as for the case of co-located antennas (see e.g., \cite{Tel:CapaMultGausChan:99}). This complicates the distributions of signals and interference \cite{Heath:DistributedAntennasSysOutOfCellInterference:2011} and provides extra degrees of freedom, namely antenna locations, for sum-rate optimization \cite{Wang:AntennaLocDAS:2009, ParkLee:AntennaPlacementDistributedAntenna:2012}. Essentially, this work is an attempt to quantify the advantages of extremely dense distributed antennas for channel estimation and data transmission.
Given the proximity between the UA and users, the UA channel is typically over free space or at most contains sparse scatterers. Combining free space channels and ubiquitous antennas shifts the paradigm of MIMO communications in several aspects. First, a new approach is needed for analyzing the capacity of the UA channel. Rich scattering is commonly assumed in a conventional MIMO channel, allowing the channel to be modeled as a random matrix and its capacity analyzed using probability theory and linear algebra (see e.g., \cite{Tel:CapaMultGausChan:99}). This approach, however, is unsuitable for the UA channel with free space propagation since the channel capacity depends on the array geometry and the user locations. Thus, analyzing the capacity of the UA channel should rely on an approach merging geometry, electromagnetic wave theory and information theory in the same vein as \cite{PoonTse:DoFMultiAntennaChannels:2005, Gruber:EMInfoTheoryWireless:2008}. Next, despite the massive number of elements in the UA, the UA channel over free space is determined only by a few parameters such as the user locations and this fact can be exploited to dramatically reduce the complexity of channel estimation. In contrast, given rich scattering, deploying more antennas leads to continuous growth of the number of degrees of freedom (DoF) in the channel, which makes channel estimation a key challenge in designing massive MIMO systems \cite{Marzetta:CellularUnlimitedBSAntennas:2010}. Last, the classic technique for multiuser beamforming for free space channels computes nulls by solving a linear system where the number of variables is equal to that of transmit antennas \cite{VanVeen:Beamforming:1988}, and thus is inefficient for the UA system with a massive number of transmit antennas. This calls for the design of efficient transmission algorithms for the UA systems.
\subsection{Contributions and Organization}
The paper represents the first attempt on designing the UA communication systems and focuses on the signal-processing aspect, namely channel estimation and data transmission. For tractability, the work considers a particular coverage region represented by a simple geometric model which comprises a continuous circular UA communicating with $U$ single-antenna users at fixed locations near the UA center. In the model, propagation is constrained to be within the horizontal plane. The results are subsequently extended to the case with propagation in the three-dimensional space and a continuous spherical UA. Channels are assumed to be free space, narrow band, and reciprocal. The elements of the UA and user antennas are all assumed to be omnidirectional. The layout of the above model is similar to that of some existing ones for single-cell DASs (see e.g., \cite{Yang:MassiveMIMOCircularlyDistributedAntennas}) that, however, assumes discrete antennas and rich scattering. The continuity of the UA, modeling dense antennas, is a typical technique for avoiding consideration of antenna placement (see e.g., \cite{PoonTse:DoFMultiAntennaChannels:2005}). More important, it is instrumental for the new findings and the algorithmic designs as summarized in the sequel.
First, consider estimation of multiuser UA channels using only single pilot symbols. The channels are determined by user locations and thus called \emph{location induced channels} (LI-channels). The channel responses are non-linear functions of the locations. This makes the conventional linear (mimimm-mean-square-error or zero-forcing) estimation unsuitable and the optimal maximum-a-posteriori estimation intractable since it requires solving a set of non-linear equations \cite{VanTrees:DetectionEstimation:2004}. To address this issue, a novel low-complexity channel-estimation technique is proposed based on decomposing the receive multiuser circular training signal into a Fourier series for the circular UA or spherical harmonics for the spherical UA. This leads to a derived function of location, called the \emph{user-location profile}. The proposed estimation method is to detect the locations of the peaks of the profile that yield estimated user locations. The estimation procedure is shown to suppress noise by averaging and incur estimation errors only due to interference between multi-channel etimation. The error is shown to decay with the minimum user-separation distance following the power law with the exponent $1/3$ and $1$ for the circular and spherical UA, respectively. Therefore, without orthogonal pilot sequences, multiuser channel estimation in the UA system is enabled by sufficiently large user-separation distances instead of differentiation in multiuser angles of arrival as in the conventional MIMO systems (see e.g., \cite{YinGesbert:CoordinatedApproachLargeScaleMIMO:2013}). In addition, applying the method to the scenario where users deploy orthogonal pilot sequences leads to almost-perfect channel estimation.
Next, consider channel conjugate data transmission using the UA. The signal-to-interference-and-noise ratios (SINRs) are derived in closed-form. In particular, the power of interference between any two users is shown be proportional to their separation distance (in wavelength) raised to the power of $2/3$ and $2$, corresponding to the circular and spherical UA, respectively. Therefore, even given single-user transmission, interference can be suppressed by increasing users' separation distances. Moreover, the path lose is shown to be inversely proportional to the propagation distance or \emph{fixed} regardless of the distance for the circular and spherical UAs, respectively. In contrast, the loss for a conventional array with collocated elements is inversely proportional to the \emph{squared} distance.
Last, a novel low-complexity precoding design is proposed for multiuser transmission using the UA. Specifically, the precoders are designed in the form of Fourier series for the circular UA and spherical harmonics for the spherical UA. Their coefficients are controlled as derived to excite different \emph{phase modes} of the circular array so as to null multiuser interference. For this sub-optimal design, the number of DoF available for interference nulling is shown to be proportional to the minimum user separation distance or its square for the circular and spherical UAs, respectively.
The reminder of the paper is organized as follows. The UA system model is described in Section~\ref{Section:System}. Algorithms for channel estimation and data transmission for the circular-UA system are presented in Section~\ref{Section:ChanEst} and \ref{Section:DataTx}, respectively. The results are extended to the spherical-UA system in Section~\ref{Section:Spherical:UA}. Simulation results are provided in Section~\ref{Section:Sim} follows by concluding remarks in Section~\ref{Section:Conclusion}. In Appendix~\ref{App:Bessel}, Bessel functions and spherical harmonics are defined and their properties discussed. Last, Appendix~\ref{App:Proofs} contains the proofs for lemmas.
\section{System Model}\label{Section:System}
As illustrated in Fig.~\ref{Fig:UASystem}, the UA is modeled as either a \emph{circular array} or a \emph{spherical array}, denoted as $\mathds{O}$ and $\mathds{O}^2$, respectively, having the same radius denoted as $r_0$. The dense UA is assumed to be \emph{continuous} for tractable analysis but this assumption is relaxed in simulation. The communication system comprises the UA centered at the origin and a set of $U$ single-antenna users enclosed by the UA, represented by their fixed locations $X_1, X_2, \cdots, X_U$ in the horizontal plane. A user, $X_u$, and a particular element of the UA, $A$, are represented by their spherical coordinates $(r_u, \varphi_u, \tfrac{\pi}{2})$ and $(r_0, \varphi, \vartheta)$, respectively, where $(\varphi_u, \varphi)$ are azimuth angles and $(\tfrac{\pi}{2}, \vartheta)$ are polar angles with $\vartheta = \frac{\pi}{2}$ for the case of circular UA. In addition, there are no scatterers.
\begin{assumption}\label{AS:CenterUsers}\emph{Users are located near the center of the UA such that $r_u/r_0 \ll 1$ for $u = 1, \cdots, U$. }
\end{assumption}
\noindent The assumption allows tractable analysis as it simplifies the expression for the propagation distances. Specifically, given a user $X_u$ and a UA antenna $A$, the separation distance and angle are denoted as $|X_u - A|$ and $\psi_u(A)$, respectively, with
\begin{align}
|X_u - A| &= \sqrt{r_0^2 + r_u^2 - 2 r_0 r_u\cos\psi_u(A)}\nonumber\\
&= r_0 - r_u\cos\psi_u(A) + o \label{Eq:Dist:Approx}
\end{align}
where $o$ represents $O(\max_u r_u/r_0)$. Note that the above model can be extended to include a set of scatterers at given locations, which reflect communication signals between the UA and users in the absence of lines of sight. Remarks on the extensions of results to scattering channels are provided in the sequel.
\begin{figure}[t]
\centering
\subfigure[Circular-UA System ]{\includegraphics[width=7.5cm]{UA_Cir.pdf}}
\subfigure[Spherical-UA System]{\includegraphics[width=8.5cm]{UA_Sph.pdf}}
\caption{A geometric model of a communication system using a continuous circular/spherical UA. }
\label{Fig:UASystem}
\end{figure}
The wave transmitted by an antenna is assumed to propagate as a plane wave in the three-dimensional free space. All antennas are assumed to be omni-directional. Let $h_u(A)$ represents the response of the channel between $X_u$ and $A$. As a result,
\begin{equation}
h_{u}(A) = \frac{1}{\sqrt{4\pi} |X_u - A|}e^{-j\frac{2\pi}{\lambda}|X_u - A|},\qquad A \in \mathds{O} \ \text{or} \ \mathds{O}^2\nonumber
\end{equation}
where $\lambda$ denotes the carrier wavelength. Based on \eqref{Eq:Dist:Approx},
\begin{equation}
h_{u}(A) =\frac{1}{\sqrt{4\pi} r_0} e^{-j\frac{2\pi }{\lambda} r_0 + j\frac{2\pi}{\lambda}r_u\psi_u(A)} + \frac{o}{r_0}, \qquad A \in \mathds{O} \ \text{or} \ \mathds{O}^2. \label{Eq:Ch:Gain}
\end{equation}
Channel estimation at the UA is assisted by pilot signals transmitted by users. Time is divided into slots with unit symbol durations. Since the effective aperture for a omni-directional antenna is $\lambda^2/4\pi$ \cite{Friis:TransmFormula:1946}, the total pilot signal received at antenna $A$ in an arbitrary slot, denoted as $q(A)$, is given as
\begin{equation}\label{Eq:Sig:UL}
q(A) =\frac{\lambda}{\sqrt{4\pi}}\sum_{u=1}^U h_{u}(A) s_{u}+ z(A), \qquad A \in \mathds{O}\ \text{or} \ \mathds{O}^2
\end{equation}
where $s_{u}$ is a pilot symbol transmitted by user $u$ and the noise $z(A)$ is a spatial sample of the additive white Gaussian noise $\mathcal{CN}(0, \sigma^2)$ process at location $A$. The noise processes are assumed to be \emph{spatially white} as follows.
\begin{assumption}\label{AS:White}\emph{The noise processes $z(X)$ and $z(Y)$ at two locations $X$ and $Y$ are independent if $X \neq Y$.}
\end{assumption}
\noindent Substituting \eqref{Eq:Ch:Gain} into \eqref{Eq:Sig:UL} gives
\begin{equation}\label{Eq:Training:Sig}
q(A) =\frac{\lambda e^{ -j\frac{2\pi}{\lambda}r_0}}{4\pi r_0}\sum_{u=1}^U e^{ j\frac{2\pi}{\lambda}r_u\psi_u(A)} s_{u}+ z(A) +\frac{o}{r_0}, \qquad A \in \mathds{O} \ \text{or} \ \mathds{O}^2.
\end{equation}
Next, consider downlink data transmission. The data symbol intended for user $u$ is denoted as $x_u$ and assumed to be distributed as a $\mathcal{CN}(0, 1)$ random variable. The symbol is precoded by a continuous precoder represented by $f_u: \mathds{O} \ \text{or} \ \mathds{O}^2 \rightarrow \mathds{C}$. Let $P_{\textrm{t}}$ denote the transmission power per user. Then the incident field at location $u$ in an arbitrary slot, denoted as $g(X_u)$, can be written as
\begin{equation}
g(X_u) =\sqrt{\frac{P_t}{\partial\mathcal{A}}} \int_{\mathcal{A}}h_u(A) \sum_{k=1}^U f_k(A) \ dA \ x_{k}, \qquad \mathcal{A} = \mathds{O} \ \text{or}\ \mathds{O}^2 \label{Eq:Sig:DL}
\end{equation}
with $h_u(A)$ being the channel response in \eqref{Eq:Ch:Gain} and $\partial\mathcal{A}$ represents the circumference of a circular UA ($\mathcal{A} = \mathds{O}$) or the surface area of a spherical UA ($\mathcal{A} = \mathds{O}^2$). It follows that the corresponding received signal is
\begin{equation}
y_u = \frac{\lambda }{\sqrt{4\pi}}g(X_u) + z_u, \qquad u = 1, 2, \cdots, U
\end{equation}
where $\{z_{u}\}$ are i.i.d. $\mathcal{CN}(0, \sigma^2)$ random variables representing channel noise.
\section{Communication Using the Circular UA: Channel Estimation}\label{Section:ChanEst}
Estimation of the LI-channels is to infer the user locations from the training signal, namely
\begin{equation}
\{q(A)\mid A \in \mathds{O} \}\Longrightarrow \{X_u\}\nonumber
\end{equation}
with $\{q(A)\}$ given in \eqref{Eq:Training:Sig}. As mentioned, the linear or MAP estimation techniques are intractable since $q(A)$ is a nonlinear function of $\{X_u\}$.
To address this issue, a simple estimation scheme is proposed in the following sub-sections, which reduces channel estimation to the detection of the peaks of a given function.
\subsection{LI-Channel Estimation with Single Pilot Symbols}\label{subsec:chn_estm_single_pilot}
Consider the scenario where users simultaneously transmit \emph{single} pilot symbols $\{s_u\}$ to facilitate channel estimation at the UA. Without loss of generality, assume that the pilot symbols are all ones: $s_u = 1\ \forall \ u$. Let the training signal $q(A)$ in \eqref{Eq:Sig:UL} be re-denoted as $q(\varphi)$ since $A = (r_0, \varphi, \tfrac{\pi}{2})$.
The LI-channel estimation scheme is designed as follows. First, the proposed scheme is based on decomposing the received training signal $\{q(\varphi)\}$ into a Fourier series:
\begin{equation}
q(\varphi) = \sum_{k = -\infty}^\infty Q_k e^{-j k\varphi}, \qquad \varphi \in [0, 2\pi]
\end{equation}
where the Fourier coefficients $\{Q_k\}$ are defined as
\begin{equation} \label{Eq:Fourier}
Q_k = \frac{1}{2\pi} \int_0^{2\pi} q(\varphi)\ d\varphi.
\end{equation}
The Fourier coefficients contain all information on the signal and thus can replace it in channel estimation. To facilitate the algorithmic design, the structure of the coefficients is characterized as follows. To this end, a few notations are introduced. Let ${\mathbf{Q}}$ represent the infinite sequence: $\cdots, Q_{-1}, Q_0, Q_1, \cdots$. The product between two sequences, ${\mathbf{V}}_1$ and ${\mathbf{V}}_2$, is denoted and defined as
${\mathbf{V}}_1\circ{\mathbf{V}}_2 = \sum_{n=-\infty}^{\infty}[{\mathbf{V}}_1]^*_n [{\mathbf{V}}_2]_n$ where $[\cdot]_n$ yields the element with index $n$. Moreover, $J_n: \mathds{R}\rightarrow\mathds{R}$ represents Bessel function with an integer index $n$ as defined and discussed in Appendix~\ref{App:Bessel}.
\begin{lemma}[Training signal decomposition]\label{Lem:Fourier}The sequence ${\mathbf{Q}}$ can be decomposed as
\begin{equation}
{\mathbf{Q}} = \frac{\lambda}{4\pi r_0}e^{-j\tfrac{2\pi}{\lambda}r_0} \sum_{u=1}^U {\mathbf{V}}(X_u) + \frac{o}{r_0}, \qquad \text{a.s.}
\end{equation}
where ${\mathbf{V}}(X_u) \!=\! \cdots V_{-1}(X_u), V_0(X_u), V_1(X_u) \cdots$ with the function $V_n(Y)$ defined for a given location $Y= (r_Y, \varphi_Y, \tfrac{\pi}{2})$ as
\begin{equation}
V_n(Y) = j^n e^{jn\varphi_Y} J_n\left(\tfrac{2\pi }{\lambda}r_Y\right).
\end{equation}
\end{lemma}
The proof is provided in Appendix~\ref{Lem:Fourier:Proof}.
\begin{remark}\emph{The Fourier coefficients $\{Q_k\}$ of the received training signal are \emph{noiseless}. The noise suppression is the combined result of the noise spatial whiteness in Assumption~\ref{AS:White} and the integral operation in \eqref{Eq:Fourier}.}
\end{remark}
Next, based on the decomposition of ${\mathbf{Q}}$ in Lemma~\ref{Lem:Fourier}, the key component of the proposed scheme for LI-channel estimation is a function $\Phi: \mathds{R}^2 \rightarrow \mathds{R}_+$ defined as
\begin{equation}\label{Eq:Loc:Profile}
\Phi(Y) = \frac{4\pi r_0 }{\lambda}|{\mathbf{V}}(Y)\circ{\mathbf{Q}}|
\end{equation}
and called the \emph{channel observation profile }. To derive a closed-form expression for $\Phi(Y)$, a useful property for ${\mathbf{V}}$ directly follows from the Addition Theorem in Property (B2) of Bessel functions in Appendix~\ref{App:Bessel} as shown below.
\begin{lemma}\label{Lem:Proj}Given two locations $X, Y\in \mathds{R}^2$, the product between the sequences ${\mathbf{V}}(X)$ and ${\mathbf{V}}(Y)$ satisfies
\begin{equation}
|{\mathbf{V}}(X)\circ{\mathbf{V}}(Y)| = J_0\left(\tfrac{2\pi }{\lambda}|X- Y|\right). \nonumber
\end{equation}
\end{lemma}
Combining Lemmas~\ref{Lem:Fourier} and \ref{Lem:Proj} and the definition of $\Phi(Y)$ in \eqref{Eq:Loc:Profile} gives the following theorem.
\begin{theorem}[Channel observation] \label{Theo:Profile}The channel observation profile corresponding to the circular UA is noiseless and given as
\begin{equation} \label{Eq:Loc:Profile:a}
\Phi(Y) = \left |\sum\nolimits_{u=1}^U J_0\left(\tfrac{2\pi }{\lambda}|X_u - Y|\right)\right| + o, \quad \text{a.s.}
\end{equation}
\end{theorem}
An example of $\Phi(Y)$ is illustrated in Fig.~\ref{Fig:Profile}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=9cm]{m_profile.pdf}\vspace{10pt}
\caption{ An example of the channel observation profile corresponding to three users with equal separation distances of $2$ wavelengths where the peaks are identified by their Cartesian coordinates and the ripples arise from the tails of superimposed Bessel functions (see Theorem~\ref{Theo:Profile}). }
\label{Fig:Profile}
\end{center}
\end{figure}
{\bf LI-channel estimation scheme:} One can observe from Theorem~\ref{Theo:Profile} that the $U$ Bessel functions in $\Phi(Y)$ have their peaks at corresponding user locations since $J_0(x)$ is maximized at $x = 0$. Motivated by this fact, the proposed scheme for estimating the user locations is to detect the peaks in the channel observation profile $\Phi(Y)$. Since the profile is noiseless, the only source for estimation errors is the coupling (interference) between the Bessel functions in $\Phi(Y)$ due to finite separation distances between user locations.
\begin{remark}[Channel estimation error] \label{Re:EstErr} \emph{Channel estimation using the proposed scheme is close to perfect for a single-user system since the corresponding channel observation profile $\Phi(Y) \approx J_0\left(\tfrac{2\pi }{\lambda}|X_1 - Y|\right) $ that is maximized at $Y = X_1$ with $\Phi(X_1) \approx 1$. Next, consider the estimation of multiuser LI-channels. The accuracy for estimating the channel corresponding to user $X_u$ can be evaluated by the difference between $\Phi(X_u)$ and the value of $1$ for its single-user counterpart. This measures the interference due to the presence of multiuser channels and can be obtained from Theorem~\ref{Theo:Profile} as follows:
\begin{equation}\nonumber
|\Phi(X_u) - 1| \leq \left |\sum\nolimits_{k\neq u } J_0\left(\tfrac{2\pi }{\lambda}|X_k - X_u|\right)\right| + o, \quad \text{a.s.}
\end{equation}
Based on Property (B4) of Bessel functions in Appendix~\ref{App:Bessel},
\begin{align}
|\Phi(X_u) - 1| &\leq \frac{1}{\nu} \sum_{k \neq u} \left(\tfrac{2\pi}{\lambda} |X_k - X_u|\right)^{-\frac{1}{3}} + o, \quad \text{a.s.} \nonumber\\
& \leq \frac{U - 1}{\nu} \left(\tfrac{2\pi}{\lambda} \min_{u \neq k} |X_k - X_u|\right)^{-\frac{1}{3}} + o, \quad \text{a.s.} \label{Eq:ErrBnd}
\end{align}
The result shows that the interference magnitude diminishes with the increasing minimum user-separation distance approximately following a sub-linear function. Given a pair of users, setting the upper bound in \eqref{Eq:ErrBnd} to a small value e.g., $0.1$, a rule of thumb for the user-separation distance sufficiently large for accurate multiuser channel estimation can be computed as $77\lambda$, which is $23$ m for a carrier of $1$ Gz and $2.3$ m for $10$ Gz.
}
\end{remark}
\begin{remark}[Effect of wavelength on channel estimation] \emph{One can infer from \eqref{Eq:ErrBnd} that with user locations fixed, the accuracy of channel estimation can be improved by increasing the carrier wavelength $\lambda$. However, this leads to a denser UA in practice since its elements are required to be separated by no more than $\lambda/2$.
}
\end{remark}
\begin{remark}[Scattering Channels]\emph{Assuming the reflection at scatterers are isotropic, the channels between the UA and users are determined not only by the scatterers' locations but also by the channel coefficients that combine the gains of the channels between scatterers and users and reflection attenuation at scatterers. Estimation of the channels can follow a two-phase procedure. First, the scatter locations can be estimated following the same scheme as discussed earlier for estimating user locations. Next, resolving the scatter locations allows the estimation of the training signals reflected by individual scatterers. However, further estimating the individual channel coefficient between each pair of scatterer and user requires the use of pilot sequences and exploiting their orthogonality. In other words, unlike that of free-space channels, channel estimation using single pilot symbols is infeasible for scattering channels.}
\end{remark}
\subsection{LI-Channel Estimation with Pilot Sequences}
In this section, the results in the preceding section for single pilot symbols are extended to the case with pilot sequences. Let the pilot sequence for the $u$-th user be denoted as ${\mathbf{s}}_{u} = [s_{u}(1),\cdots, s_{u}(L)]^T$ with length $L$. In the channel training phase, the UA receives sequentially $L$ infinite sequences over $L$ symbol durations, denotes as ${\mathbf{Q}}(1), {\mathbf{Q}}(2), \cdots, {\mathbf{Q}}(L)$, where ${\mathbf{Q}}(\ell)$ is modified from its single-symbol counterpart in Lemma~\ref{Lem:Fourier} as
\begin{equation}
{\mathbf{Q}}(\ell) = \frac{\lambda e^{-j\tfrac{2\pi}{\lambda}r_0}}{4\pi r_0} \sum_{u=1}^U s_u(\ell){\mathbf{V}}(X_u) + \frac{o}{r_0}, \qquad \text{a.s.}\label{Eq:Q:ell}
\end{equation}
To estimate the user location $X_u$, $\{{\mathbf{Q}}(\ell)\}$ are coherently combined using ${\mathbf{s}}_u$ as $\sum_{\ell=1}^L s^*_{u}(\ell){\mathbf{Q}}(\ell)$. The result, denoted as ${\mathbf{Q}}_u$, follows from \eqref{Eq:Q:ell} as
\begin{equation}
{\mathbf{Q}}_u = \frac{\lambda e^{-j\tfrac{2\pi}{\lambda}r_0}}{4\pi r_0} \sum_{k=1}^U {\mathbf{s}}^\dagger_u
{\mathbf{s}}_k{\mathbf{V}}(X_k) + \frac{o}{r_0}, \qquad \text{a.s.}
\end{equation}
A corresponding channel observation profile $\Phi_u(Y)$ can be defined similarly as in \eqref{Eq:Loc:Profile}:
\begin{equation}\label{Eq:Loc:Profile:Seq}
\Phi_u(Y) = \frac{4\pi r_0}{\lambda}|{\mathbf{V}}(Y)\circ{\mathbf{Q}}_u|.
\end{equation}
The profile $\Phi_u(Y)$ is decomposed into
the desired and interference terms as follows:
\begin{align}
\Phi_u(Y) \!=\! \Big | J_0\left(\tfrac{2\pi}{\lambda}|X_{u}\!-\!Y|\right) \!+\! &\sum_{k\neq u} {\mathbf{s}}^\dag_{u}{\mathbf{s}}_{k} J_0\left(\tfrac{2\pi}{\lambda}|X_k\!-\! Y|\right)\Big |^2 \!+\!o. \nonumber
\end{align}
Then the deviation of $\Phi_u(Y)$ from its single-user counterpart can been bounded as:
\begin{equation}
\left|\Phi_u(Y)-J_0^2\left(\tfrac{2\pi }{\lambda}|X_{u}-Y|\right)\right| \leqslant \sum\nolimits_{k\neq u} |{\mathbf{s}}^\dag_{u}{\mathbf{s}}_{k}| J_0\left(\tfrac{2\pi}{\lambda}|X_k -Y|\right)+o, \qquad \forall \ u. \label{Eq:Loc:Profile:Seq:UB}
\end{equation}
This leads to the following main result of this section.
\begin{theorem}[Effect of pilot sequences] \label{Theo:Profile:Seq}
Given $L \geq U$ and orthogonal pilot sequences, the LI-channel estimation using the circular UA is almost perfect since the channel observation profile is approximately equal to the single-user counterpart:
\begin{align}
\Big |\Phi_u(Y) - & J_0\left(\tfrac{2\pi}{\lambda}|X_{u} - Y|\right) \Big |\leqslant o, \qquad u =1, 2, \cdots, U. \nonumber
\end{align}
\end{theorem}
Comparing Theorems~\ref{Theo:Profile} and \ref{Theo:Profile:Seq}, the advantage of pilot sequences over single pilot symbols lies in their capability of decoupling the estimation of multiuser LI-channels, thereby providing close-to-perfect channel estimation.
\subsection{Effect of Finite Elements in the UA}
Consider a discrete circular UA with $N$ antennas uniformly placed on the circle centered at the origin and with the radius $r_0$. The discrete UA can be interpreted as a quantized version of the continuous UA with the quantization error bounded by $\pi/N$. Based on this interpretation and assuming $N$ is large, the analysis for the continuous UA can be straightforwardly extended to the case of discrete UA by including the quantization error. As a result, for channel estimation with single pilot symbols, the channel observation profile for the discrete UA, denoted as $\hat{\Phi}(Y)$, can be written as
\begin{equation}
\hat{\Phi}(Y) = \Phi(Y) + O\left(\frac{1}{N}\right).
\end{equation}
where $\Phi(Y)$ for the continuous UA is given in Theorem~\ref{Theo:Profile}. Moreover, the result in Theorem~\ref{Theo:Profile:Seq} for the case of pilot sequences can be modified by replacing $o$ with $o + O(1/N)$. Similarly, for data transmission using the discrete UA, the receive SINR at user $u$, denoted as $\widehat{\mathsf{SINR}}_u$, can be shown to be $\widehat{\mathsf{SINR}}_u = \mathsf{SINR}_u + O(1/N)$ with $\mathsf{SINR}_u$ corresponding to the continuous UA.
The above results suggest that with respect to the continuous counterpart, the discrete UA causes additional fluctuation in the channel observation profile and receive SNRs, which can potentially degrades the performance of channel estimation and data transmission.
\section{Communication Using the Circular UA: Data Transmission}\label{Section:DataTx}
In this section, two precoding techniques are designed for the UA, namely the channel conjugate and the \emph{multiuser phase mode} (MU-PM) precoding. For simplicity, it is assumed that the UA has perfect knowledge of the LI-channels.
\subsection{Channel Conjugate Transmission}
For channel conjugate transmission, the precoder applies a phase shift to each antenna for compensating propagation delay to achieve coherent combining at the target user location, which is similar to beamforming using a phase array. Specifically, the precoder $f_u(\varphi)$ is given as
\begin{equation}
f_u(\varphi) = \frac{h_u^*(\varphi)}{|h_u(\varphi)|},\quad \varphi \in [0, 2\pi)\label{Eq:CC:Prec}
\end{equation}
where $h(\varphi)$ is given in \eqref{Eq:Ch:Gain}. It follows that
\begin{equation}\label{Eq:CCTx:Prec}
f_u(\varphi) = e^{j\frac{2\pi }{\lambda} (r_0 - r_u\psi_u(\varphi))}, \quad \varphi \in [0, 2\pi).
\end{equation}
The normalization $|f_u(\varphi)|^2 = 1$ facilitates the UA implementation under the per-element power constraint e.g., using a phase array.
The channel conjugate precoder $\{f_u(\varphi)\}$ shapes the distribution of the field power density such transmission power is concentrated in a small region centered at the target location $X_u$. Characterizing the distribution is useful for analyzing the precoder performance. To this end, consider the transmission of an unmodulated wave using the UA after precoding using $\{f_u(\varphi)\}$ in \eqref{Eq:CCTx:Prec}. Conditioned on the precoder, let $g(X_k\mid X_u)$ denote the resultant field measured at the location $X_k$. With the propagation distance in \eqref{Eq:Dist:Approx}, it can be obtained that
\begin{equation}
g(X_k\mid X_u) =\sqrt{\frac{P_{\textrm{t}}}{2\pi r_0}}\times \frac{1}{\sqrt{4\pi r_0^2}}\times \int_0^{2\pi} e^{j\frac{2\pi}{\lambda}(-r_0 + r_u\cos\psi_u(\varphi) ) } f_k(\varphi) r_0 d \varphi + \frac{o}{\sqrt{r_0}}. \label{Eq:Field:Dist}
\end{equation}
At the right-hand side of \eqref{Eq:Field:Dist}, the three factors of the dominant term correspond to the density of transmission power uniformly distributed over the UA, the propagation loss and the wave superposition at $X_u$, respectively. Substituting the precoder in \eqref{Eq:CCTx:Prec} into \eqref{Eq:Field:Dist} yields
\begin{align}
g(X_k\mid X_u) &= \sqrt{\frac{P_{\textrm{t}}}{2 r_0}}\times \frac{1}{2\pi} \int_0^{2\pi} e^{j\frac{2\pi}{\lambda}(r_u\cos\psi_u(\varphi)- r_k\cos\psi_k(\varphi)) } d \varphi + \frac{o}{\sqrt{r_0}}. \label{Eq:Field:Dist:cir}
\end{align}
The field power density at location $X_k$ can be represented by $p(X_k\mid X_u) = |g(X_k\mid X_u)|^2$. Using \eqref{Eq:Field:Dist:cir}, a closed-form expression for $p(X_k\mid X_u) $ is obtained as shown in the following lemma proved in Appendix~\ref{App:CC:Field}.
\begin{lemma}[Field power density distribution] \label{Lem:CC:Field}
Given the circular UA and channel conjugate precoding targeting user $X_u$, the field power density measured at the user location $X_k$ is given as
\emph{\begin{equation}
p(X_k\mid X_u) = \frac{P_{\textrm{t}}}{2 r_0} J_0^2\left(\tfrac{2\pi}{\lambda}|X_u - X_k|\right) + \frac{o}{r_0}. \label{Eq:Field:Dist:cir:a}
\end{equation}}
\end{lemma}
The result shows that the field power density function $p(X_k\mid X_u)$ depends only on the distance $|X_k - X_u|$ and thus can be rewritten as $p(d)$ with $d > 0$ denotes the the distance from the location targeted by the precoder. Then the distribution of the field power density can be characterized by the function $p(d)/p(0) = J_0\left(\tfrac{2\pi d}{\lambda}\right)$ that is plotted in Fig.~\ref{Fig:Bessel:Bnd}. It can be observed from the figure that the channel-conjugate precoder shapes the field distribution such that most power is concentrated within a circular region centered at the target location and having a radius of half wavelength. The tail of the distribution function is undesirable as it causes interference to nearby unintended receivers. However, the envelop of the tail decays with distance $d$, allowing interference suppression by spatial separation as further discussed in the sequel.
\begin{figure}[t]
\begin{center}
\includegraphics[width=10cm]{Power_dist.pdf}
\caption{Given channel-conjugate transmission, the distribution of field power density (normalized by its peak) as a function of distance from the peak location.}
\label{Fig:Bessel:Bnd}
\end{center}
\end{figure}
With the field distribution in \eqref{Lem:CC:Field}, the performance of the channel-conjugate precoding is readily analyzed in terms of receive SINRs as follows. Let $P_{\textrm{r}}$ and $P_{\textrm{i}}$ denote the signal and interference powers at user $X_u$, respectively. Since the effective aperture of the receive omni-directional antenna is $\lambda/4\pi$, $P_{\textrm{r}}$ and $P_{\textrm{i}}$ are given as
\begin{equation}
P_{\textrm{r}} = \frac{\lambda}{4\pi }p(X_u\mid X_u) \quad\textrm{and}\quad P_{\textrm{i}} = \frac{\lambda}{4\pi } \sum_{k \neq u}p(X_u\mid X_k). \label{Eq:Pr:Pi}
\end{equation}
The receive SINR for user $X_u$ can be written in terms $P_{\textrm{r}}$ and $P_{\textrm{i}}$ as
\begin{equation}
\mathsf{SINR}_u = \frac{P_{\textrm{r}}}{P_{\textrm{i}} + \sigma^2}. \label{Eq:SINRu}
\end{equation}
Substituting Lemma~\ref{Lem:CC:Field}, \eqref{Eq:Pr:Pi} into \eqref{Eq:SINRu} yields the following main result of this section.
\begin{theorem}[Receive SINRs] \label{Theo:ChanConj:Tx}
For channel conjugate transmission using the circular UA, the receive SINR for user $u$ is given as
\begin{equation}
\mathsf{SINR}_{u} = \frac{1}{\sum\limits_{k\neq u } J_0^2\left(\frac{2\pi}{\lambda}|X_{u} - X_k|\right) + \frac{1}{\mathsf{SNR}} }+o, \qquad u = 1, 2, \cdots, U \label{Eq:SINR:CC}
\end{equation}
where the receive SNR is given as
\begin{equation}
\mathsf{SNR} = \frac{P_{\textrm{t}} \lambda}{8\pi \sigma^2 r_0} + \frac{o}{r_0}. \label{Eq:SNR:Cir}
\end{equation}
\end{theorem}
\begin{remark}[High SNR]\label{Re:CC:SIR}\emph{
Applying the bound on the Bessel function in Property (B$4$) in Appendix~\ref{App:Bessel}, for a high SNR, the signal-to-interference ratio (SIR) at user $u$ scales with the distance to the nearest interferer, namely $\min_{k\neq u}|X_k - X_u|$, and the wavelength $\lambda$ as
\begin{equation}
\mathsf{SIR}_{u} \geqslant \frac{\nu^2}{U-1}\left(2\pi\min_{k\neq u}\frac{|X_k - X_u|}{\lambda}\right)^{\frac{2}{3}} + o, \qquad \forall \ u. \label{Eq:SIR}
\end{equation}
Thus the receive SIRs increase with the increasing minimum user-separation distance (in wavelength) following a \emph{sub-linear} function. Moreover, the result in \eqref{Eq:SIR} suggests that denser simultaneous users can be supported by reducing the wavelength without compromising the system throughput. Specifically, the user density can scale linearly with $1/\lambda^2$.
}
\end{remark}
\begin{remark}[Free space vs. scattering channels]\emph{Free-space channels allow the UA to focus signal energy at intended users such that the signal power decays rapidly with the distance from the target location. As a result, given single-user (channel-conjugate) transmission, interference can be suppressed by increasing user spatial separation as shown in \eqref{Eq:SIR}. However, this is infeasible in the scenario of scattering channels as scattering introduces additional cross coupling between multiuser signals. Suppressing the resultant interference cannot rely on increasing users' spatial separation and multiuser precoding has to be used for this purpose. }
\end{remark}
\begin{remark}[High Mobility]\emph{Given fixed user locations, the sum rate is thus $\sum_u \log_2(1 +\mathsf{SINR}_u)$ with $\mathsf{SINR}_u$ given in \eqref{Eq:SINR:CC}. For the scenario where users have high mobility, it is more appropriate to consider the ergodic sum rate given as
\begin{equation}
\bar{R} = \mathsf{E}\left[\sum_u \log_2(1 +\mathsf{SINR}_u(X_1, \cdots, X_U))\right]
\end{equation}
where $\{X_u\}$ are random. The ergodic sum rate can be analyzed by combining the results in Theorem~\ref{Theo:ChanConj:Tx} and stochastic geometry \cite{StoyanBook:StochasticGeometry:95}.
}
\end{remark}
Last, the received SNR given in \eqref{Eq:SNR:Cir} implies the following result.
\begin{corollary}[Propagation loss] \label{Cor:PathLoss}Given channel-conjugate transmission using the circular UA, the propagation loss is given as \emph{
\begin{equation}
\frac{P_{\textrm{r}}}{P_{\textrm{t}}} = \frac{\lambda}{8\pi r_0}+ \frac{o}{r_0}.
\end{equation}}
\end{corollary}
In other words, transmission using the circular UA reduces the loss such that it is inversely proportional to the propagation distance instead of its square as in the case of a conventional array of collocated antennas.
\subsection{Multiuser Phase Mode Precoding}
In this section, MU-PM precoders are designed under the zero-forcing constraints to avoid multiuser interference. Exploiting the circular structure of the UA, the precoder for each user, say $ f_u'(\varphi)$ for user $u$, can be expressed in terms of the Fourier series representing the sequence of phase modes:
\begin{equation}\label{Eq:Precod:ZF}
f_u'(\varphi) = f_u(\varphi) \sum_{m=-\infty}^\infty c_{u, m} e^{-jm\varphi}, \qquad \varphi \in [0, 2\pi)
\end{equation}
where $f_u(\varphi) $ is the channel-conjugate precoder in \eqref{Eq:CC:Prec} for compensating the channel phase shift and the summation is the said Fourier series. The precoder coefficients $\{c_{u, m}\}$ satisfy the power constraint: $\sum_{m=-\infty}^\infty |c_{u, m}|^2 \leq 1$ for all $u$.
The precoder coefficients are designed to avoid the inter-user interference as follows. Without multiuser interference, it is unnecessary to analyze the field spatial distribution as in the case of channel conjugate transmission and instead the remainder of the section focuses on the precoder design and the analysis of the received signal power. To this end, the received signal at user $u$ is obtained by substituting the precoder in \eqref{Eq:Precod:ZF} into \eqref{Eq:Sig:DL}:
\begin{equation}
y_{u} = \sqrt{\frac{\lambda P_t}{8\pi r_0}} \sum_{k=1}^U \left( \frac{1}{2\pi} \int_0^{2\pi} \frac{h^*_k(\varphi)h_u(\varphi)}{|h_{k}(\varphi)|^2} \sum_{m=-\infty}^\infty c_{k, m} e^{-jm\varphi}\ d\varphi\right) x_{k} + z_{u}. \nonumber
\end{equation}
The substitution of the channel response in \eqref{Eq:Ch:Gain} yields
\begin{equation}
y_{u} = \sqrt{\frac{\lambda P_t}{8\pi r_0}} c_{u, 0} x_u + \sqrt{\frac{\lambda P_t}{8\pi r_0}} \sum_{k\neq u}\sum_{m=0}^\infty c_{k, m} \mathcal{J}_{u, k, m} x_{k}+z_{u} +\frac{o}{r_0}\label{Eq:Sig:ZF}
\end{equation}
where the first term and the summation represent the signal and interference, respectively, and $\mathcal{J}_{u, k, m}$ is defined as
\begin{equation}\label{Eq:J:Def}
\mathcal{J}_{u, k, m} = \frac{1}{2\pi } \int_0^{2\pi} e^{j\frac{2\pi}{\lambda}\left(r_{{u}}\cos(\varphi_{{u}}-\varphi)- r_{{k}}\cos(\varphi_{{k}}-\varphi)\right)-jm\varphi} \ d\varphi.
\end{equation}
A closed-form expression for $\mathcal{J}_{u, k, m}$ can be obtained as follows, following similar steps as in the proof for Lemma~\ref{Lem:CC:Field} with the details omitted for brevity.
\begin{lemma} \label{Lem:J:Coeff}The coefficients $\{\mathcal{J}_{u, k, m}\}$ defined in \eqref{Eq:J:Def} can be written as
\begin{equation}
\mathcal{J}_{u, k, m} = e^{jm \beta_{u, k}}J_m\left(\frac{2\pi}{\lambda}|X_u - X_k|)\right) \label{Eq:CrossTerm}
\end{equation}
where $1 \leq u, k \leq U$ and $m$ is an integer.
\end{lemma}
The interference term in the signal in \eqref{Eq:Sig:ZF} can be nulled by enforcing the following zero-forcing constraints:
\begin{equation}\label{Eq:ZF}
\sum_{m=-\infty}^\infty c_{k, m} \mathcal{J}_{u, k, m} = 0 , \quad \forall \ k \neq u.
\end{equation}
In practice, given a discrete UA and a constraint on computation complexity, it is infeasible to implement MU-PM precoders with an infinite number of phase modes and only a finite set of modes is considered in the design. Let the corresponding coefficient set be denoted as $\{c_{u, m} \mid1\leq u \leq U, -M \leq m \leq M\}$ with $M$ being a fixed integer and other coefficients set to zero. Under the zero-forcing constraints in \eqref{Eq:ZF}, it is desirable to choose the precoder coefficients such that the correspond set of coefficients $\{\mathcal{J}_{u, k, m}\mid 1 \leq u, k\leq U, -M \leq m \leq M\}$ are significant. Then applying Property (B4) of Bessel functions in Appendix~\ref{App:Bessel} gives that
\begin{equation}\label{Eq:DoF}
M =2\pi \min\limits_{u, k}\left\lfloor\frac{ |X_u - X_k|}{\lambda}\right\rfloor
\end{equation}
where users are assumed to be separated by distances much larger than a single wavelength, yielding the following fact.
\begin{remark}[Degrees of freedom] \emph{
Given MU-PM transmission using a circular UA, the total number of degrees of freedom for interference avoidance is approximately equal to $(2M+1)$ with $M$ given in \eqref{Eq:DoF}, which is proportional to the minimum user-separation distance.
}\end{remark}
Transmission over only $(2M+1)$ phase modes allows the precoder coefficients for user $u$ to be represented by the vector ${\mathbf{c}}_{u} = \begin{bmatrix}c_{u, -M}, &\cdots, & c_{u, 0}, & \cdots, & c_{u, M} \end{bmatrix}^T$. Furthermore, the zero-forcing constraints in \eqref{Eq:ZF} can be written in a matrix form:
\begin{equation}\label{Eq:Approx:ZF}
\mathbfcal{J}_{u}{\mathbf{c}}_{u} ={\mathbf{0}}
\end{equation}
where the matrix $\mathbfcal{J}_{u}$ is defined as
\begin{equation}\label{Eq:ZF:Coeff}
\mathbfcal{J}_{u} = \begin{bmatrix} \mathcal{J}_{u, 1, -M}, &\cdots, & \mathcal{J}_{u, 1, 0}, & \cdots, & \mathcal{J}_{u, 1, M} \\
\vdots & \ddots & \vdots &\ddots& \vdots \\
\mathcal{J}_{u, u-1, -M}, &\cdots, &\mathcal{J}_{u, u-1, 0}, & \cdots, & \mathcal{J}_{u, u-1, M} \\
\mathcal{J}_{u, u+1, -M}, &\cdots, & \mathcal{J}_{u, u+1, 0}, & \cdots, & \mathcal{J}_{u, u+1, M} \\
\vdots & \ddots & \vdots &\ddots & \vdots \\
\mathcal{J}_{u, U, -M}, &\cdots, & \mathcal{J}_{u, U, 0}, &\cdots, & \mathcal{J}_{u, U, M}
\end{bmatrix}
\end{equation}
with the elements specified in Lemma~\ref{Lem:J:Coeff}.
Given the constraints obtained in \eqref{Eq:Approx:ZF}, the main result of the section can be readily stated as follows.
\begin{theorem}[Multiuser UA precoding] \label{theorem:SZF_solutions}
For MU-PM transmission using the circular UA, the precoder coefficients are given as
\emph{
\begin{equation}\label{Eq:ZF:Solution}
{\mathbf{c}}_{u} \in \textsf{null}(\mathbfcal{J}_{u}), \qquad u = 1, 2, \cdots, U
\end{equation}
}
where the matrix $\mathbfcal{J}_{u}$ is given in \eqref{Eq:ZF:Coeff} and \emph{$\textsf{null}(\mathbfcal{J}_{u}) $} denotes the null-space of $\mathbfcal{J}_{u}$.
\end{theorem}
With interference avoided, the receive SNR for user $u$ follows from \eqref{Eq:Sig:ZF} as
\begin{equation}
\mathsf{SNR}_u = \frac{\lambda P_t |c_{u, 0}|^2}{8 \pi r_0 \sigma^2 }+\frac{o}{r_0}=\frac{\lambda P_t |{\mathbf{e}}_0^\dag{\mathbf{c}}_{u}|^2}{8 \pi r_0 \sigma^2 } + \frac{o}{r_0}
\end{equation}
where ${\mathbf{e}}_0 = [0,\cdots,0,1,0,\cdots,0]^T$. Thus, to maximize $\mathsf{SNR}_u$, the precoder-coefficient vector ${\mathbf{c}}_{u}$ should be chosen as the projection of ${\mathbf{e}}_0$ onto $\textsf{null}(\mathbfcal{J}_{u})$ that contains ${\mathbf{c}}_{u}$ according to Theorem~\ref{theorem:SZF_solutions}. This gives the following corollary.
\begin{corollary}\label{Cor:RxSNR}\emph{For MU-PM transmission using the circular UA, the maximum receive SNRs are given as
\begin{equation}\label{Eq:SNR:ZF}
\mathsf{SNR}_u =\frac{\lambda P_t \left|{\mathbf{e}}_0^\dagger {\mathbf{b}}_{u}\right|^2}{8 \pi r_0 \sigma^2 } + \frac{o}{r_0}, \qquad u = 1, 2, \cdots, U
\end{equation}
where ${\mathbf{b}}_{u}$ is a basis of $\textsf{null}(\mathbfcal{J}_{u})$.
}
\end{corollary}
With respect to the case of channel-conjugate transmission, the term $\left|{\mathbf{e}}_0^\dagger {\mathbf{b}}_{u}\right|^2$ represents the loss in receive SNR due to interference nulling.
\section{Communication Using the Spherical UA}\label{Section:Spherical:UA}
In the preceding section, the UA is modeled as the circular array. The results are extended in this section to the spherical UA system, which shows its performance improvements with respect to the circular-UA counterpart. Essentially, the difference in analysis arises from the use of spherical harmonics as the tool in place of Fourier series.
\subsection{Spherical LI-Channel Estimation}\label{Section:ChanEst:Sph}
Assume that users transmit single pilot symbols: $\{s_u\} = \{1\}$. The estimation scheme in Section~\ref{Section:ChanEst} is extended to the spherical UA as follows. To this end, the training signal in \eqref{Eq:Training:Sig}, re-denoted as $q(\varphi, \vartheta)$, is expanded as a Laplace series as follows by using the harmonic functions $\{Y^m_{\ell}\}$, defined in \eqref{Eq:Sph:Basis} in Appendix~\ref{App:Bessel}, as the basis:
\begin{equation}\label{Eq:Train:Sph}
q(\varphi, \vartheta) = \sum_{\ell = 0}^\infty \sum_{m = -\ell}^\ell Q_{\ell}^m Y^m_{\ell}(\varphi, \vartheta), \qquad \vartheta \in [0, \pi), \varphi \in [0, 2\pi)
\end{equation}
where the Laplace coefficients $\{Q_{\ell}^m\}$ are defined as
\begin{equation}\label{Eq:Q:Sph}
Q_{\ell}^m = \sum_{u=1}^U\int_{\vartheta=0}^{2\pi} \int_{\psi=0}^\pi q(\varphi, \vartheta) Y^m_{\ell}(\varphi, \vartheta) \sin\varphi d\varphi d\vartheta.
\end{equation}
To derive a closed-form expression for $Q_{\ell}^m$, a useful result is given as follows.
\begin{lemma}\label{Lem:Sph:Expan} Given two points $(r_0, \varphi, \vartheta), (r_u, \varphi_u, \vartheta_u)\in \mathds{R}^3$,
\begin{equation}\label{equ:3d_first_term_expand}
e^{j\frac{2\pi}{\lambda} r_u\cos \psi_u } = (2\pi)^{\frac{3}{2}} \sum_{\ell=0}^\infty \sum_{m=-\ell}^\ell \frac{j^\ell J_{\ell+\frac{1}{2}}\left(\frac{2\pi}{\lambda}r_u\right)}{\left(\frac{2\pi}{\lambda}r_u\right)^{\frac{1}{2}}} Y^m_{\ell}(\varphi_u, \vartheta_u) Y^m_{\ell}(\varphi, \vartheta)
\end{equation}
where $\psi_u$ denotes the angle between the vectors corresponding to the two points.
\end{lemma}
\noindent The lemma is proved in Appendix~\ref{Lem:Sph:Expan:Proof}. Using the orthogonality of the basis functions $\{Y^m_{\ell}(\varphi, \vartheta)\}$, substitution of the expression for $q(\varphi, \vartheta)$ in \eqref{Eq:Training:Sig} and Lemma~\ref{Lem:Sph:Expan} into \eqref{Eq:Q:Sph} leads to the following lemma.
\begin{lemma}[Training signal decomposition]\label{Lem:Qmn}The training signal received at the spherical UA, namely $\{q(\varphi, \vartheta)\}$ in \eqref{Eq:Train:Sph}, has the Laplace coefficients given as follows:
\begin{equation}\label{Eq:Qml}
Q_{\ell}^m = \frac{\lambda (2\pi)^{\frac{3}{2}}}{4\pi r_0} \sum_{u=1}^U \frac{j^\ell J_{\ell+\frac{1}{2}}\left(2\pi r_u/\lambda\right)}{\left(2\pi r_u/\lambda\right)^{\frac{1}{2}}} Y^m_{\ell}(\varphi_u, \vartheta_u) + \frac{o}{r_0}
\end{equation}
where $\ell = 1, 2, \cdots$ and $-\ell \leq m \leq \ell$.
\end{lemma}
\noindent Define two sets $\tilde{{\mathbf{Q}}}$ and $\tilde{{\mathbf{V}}}(Y)$ as $\tilde{{\mathbf{Q}}} = \{Q_{\ell}^m\}$ and $\tilde{{\mathbf{V}}}(Y) = \{V_{\ell}^m(Y)\}$ with their elements $Q_{\ell}^m$ given in \eqref{Eq:Qml} and
$V_{\ell}^m: \mathds{R}^3\rightarrow \mathds{R}$ being a function defined as
\begin{equation}\label{Eq:Vml}
V_{\ell}^m(Y) = (2\pi)^{\frac{3}{2}} \frac{j^n J_{\ell+\frac{1}{2}}\left(\frac{2\pi}{\lambda}r_u\right)}{\left(\frac{2\pi}{\lambda}r_u\right)^{\frac{1}{2}}} Y^m_{\ell}(\varphi_u, \vartheta_u).
\end{equation}
Using these definitions, the channel observation profile corresponding to the spherical UA, represented by $\tilde{\Phi}(Y)$, can be defined similarly as in \eqref{Eq:Loc:Profile}:
\begin{align}
\tilde{\Phi}(Y) &=\frac{r_0}{\lambda }\left | \tilde{{\mathbf{V}}}(Y) \circ\tilde{{\mathbf{Q}}} \right |\nonumber\\
&=\frac{r_0}{\lambda }\left |\sum_{u=1}^U \sum_{\ell=0}^\infty\sum_{m=-\ell}^\ell [V_{\ell}^m(Y)]^*Q_{\ell}^m\right |.\nonumber
\end{align}
Substituting \eqref{Eq:Qml} and \eqref{Eq:Vml} gives
\begin{equation}
\tilde{\Phi}(Y) = \left|2\pi^2 \sum_{u=1}^U \sum_{\ell=0}^\infty \frac{J_{\ell+\frac{1}{2}}\left(\frac{2\pi r_u}{\lambda}\right)J_{\ell+\frac{1}{2}}\left(\frac{2\pi r_Y}{\lambda}\right)}{\left(\frac{ 2\pi r_u}{\lambda}\right)^{\frac{1}{2}}\left(\frac{ 2\pi r_Y}{\lambda}\right)^{\frac{1}{2}}} \sum_{m=-\ell}^\ell [Y^m_{\ell}(Y)]^* Y^m_{\ell}(X_u) \right| +o
\label{Eq:Prof:Sph}
\end{equation}
where noise varnishes for the same reason as for the case of circular UA.
By applying Addition Theorem in Property (S3) for spherical harmonics in Appendix~\ref{App:Bessel},
\begin{equation}
\tilde{\Phi}(Y) = \left|\frac{\pi}{2} \sum_{u=1}^U \sum_{\ell=0}^\infty \frac{(2\ell+1)J_{\ell+\frac{1}{2}}\left(\frac{2\pi r_u}{\lambda}\right)J_{\ell+\frac{1}{2}}\left(\frac{2\pi r_Y}{\lambda}\right)}{\left(\frac{ 2\pi r_u}{\lambda}\right)^{\frac{1}{2}}\left(\frac{ 2\pi r_Y}{\lambda}\right)^{\frac{1}{2}}}P_\ell\left(\cos\psi_u\right) \right|^2 +o.
\end{equation}
Next, applying Addition Theorem in Property (B3) of Bessel functions in Appendix~\ref{App:Bessel} further simplifies the expression as shown in the following theorem.
\begin{theorem}[Channel observation] \label{Theo:Profile:Sph} The channel observation profile corresponding to the spherical UA is given as
\begin{equation}
\tilde{\Phi}(Y) = \left|\sum_{u=1}^U\mathrm{sinc}\left(\tfrac{ 2\pi}{\lambda}|Y - X_u|\right) \right| + o, \qquad \textrm{a.s.}
\end{equation}
\end{theorem}
\begin{remark}[Channel estimation error] \emph{As in Remark~\ref{Re:EstErr} for the circular UA, the accuracy of channel estimation can be evaluated using the difference $ \left|\tilde{\Phi}(X_u) - 1 \right|$. Since $\left|\mathrm{sinc}(d) \right| \leqslant d^{-1}$, it can be obtained from Theorem~\ref{Theo:Profile:Sph} that
\begin{align}
\left|\tilde{\Phi}(X_u) - 1 \right|
& \leq 2\pi (U - 1) \min_{u \neq k} \left(\tfrac{|X_k - X_u|}{\lambda} \right)^{-1} + o, \quad \text{a.s.} \label{Eq:ErrBnd:Sph}
\end{align}
The error bound is observed to diminish inversely with the increasing minimum user-separation distance (in wavelength) following an inverse function that is faster than the circular-UA counterpart in \eqref{Eq:ErrBnd}. This quantifies the gain of increasing the UA by one dimension from the perspective of channel estimation.
}
\end{remark}
Next, consider the case where users transmit pilot sequences with length of $L$ symbols. Following the same procedure as for deriving Theorem~\ref{Theo:Profile:Seq} yields the following corollary.
\begin{corollary}[Effect of pilot sequences]\label{Theo:Profile:Seq:Sph}
Given $L \geq U$ and orthogonal pilot sequences, the LI-channel estimation using the spherical UA is almost perfect since the channel observation profile is approximately equal to the single-user counterpart:
\emph{
\begin{equation}
\Big |\tilde{\Phi}_u(Y) - \mathrm{sinc}\left(\tfrac{2\pi}{\lambda}|X_{u} - Y|\right) \Big |\leqslant o. \nonumber
\end{equation}
}
\end{corollary}
\begin{remark}\emph{If orthogonal pilot sequences are used, the spherical UA does not have an advantage over the circular counterpart in terms of channel estimation. However, the former improves the performance of channel estimation in the case of single pilot symbols as well as that of data communication as shown in the sequel.
}
\end{remark}
\subsection{Data Transmission Using the Spherical UA}
\subsubsection{Channel Conjugate Transmission}
For channel conjugate transmission using the spherical UA, the spherical precoder, denoted as $\tilde{f}_u(\varphi, \vartheta)$, is modified from the circular counterpart in \eqref{Eq:CCTx:Prec} as
\begin{equation}\label{Eq:CCTX:Sph}
\tilde{f}_u(\varphi, \vartheta) = \frac{h_u^*(\varphi, \vartheta)}{|h_u(\varphi, \vartheta)|},\quad \vartheta\in [0, 2\pi), \varphi \in [0, \pi).
\end{equation}
The resultant receive SNRs are derived as follows. Let $\tilde{q}(X_u \mid X_k)$ and $\tilde{p}(X_u \mid X_k)$ denote the field and its power density measured at location $X_u$ given a precoder targeting user $X_k$. Then $\tilde{q}(X_u \mid X_k)$ can be obtained by modifying the circular-UA counterpart in \eqref{Eq:Field:Dist} by replacing the integration over a circle with one over a sphere:
\begin{align}
\tilde{g}(X_u \mid X_k) &=\sqrt{\frac{P_{\textrm{t}}}{4\pi r_0^2}}\times \frac{1}{\sqrt{4\pi r_0^2}}\times \iint e^{j\frac{2\pi}{\lambda}(r_{u}\cos\psi_{u}- r_{k}\cos\psi_k) } r_0^2 \sin \vartheta \ d \vartheta \ d \varphi + o \nonumber\\
&= \frac{\sqrt{P_{\textrm{t}}}}{4\pi } \iint e^{j\frac{2\pi}{\lambda}(r_{u}\cos\psi_{u}- r_{k}\cos\psi_k) } \sin \vartheta \ d \vartheta \ d \varphi + o.
\label{Eq:Field:Dist:Sph}
\end{align}
To facilitate analysis, a set of coefficients $\{\mathcal{J}_{u, k, m, n}\}$ with integer indices $(u, k, m, n)$ are defined as
\begin{equation}
\mathcal{J}_{u, k, m, \ell} = \iint e^{j\tfrac{2\pi}{\lambda} (r_u\cos\psi_u - r_k\cos\psi_k) } Y^m_{\ell}(\varphi, \vartheta) \sin\varphi \ d\varphi \ d\vartheta \label{Eq:J:Sph:Def}
\end{equation}
with $1 \leq u, k \leq U$, $n \geq 0$ and $-\ell \leq m \leq \ell$, which are also used for designing multiuser precoders in the next sub-section. They can be written in a closed form as shown in the following lemma that is proved in Appendix~\ref{App:J:Spy:Proof}.
\begin{lemma}\label{Lem:J:Spy} The coefficient $\mathcal{J}_{u, k, m, \ell}$ defined in \eqref{Eq:J:Sph:Def} can be written as
\begin{equation}
\mathcal{J}_{u, k, m, \ell} = \frac{(2\pi)^{\frac{3}{2}}j^\ell J_{\ell+\frac{1}{2}}\left(\tfrac{2\pi}{\lambda} | X_u - X_k|\right)}{\sqrt{\tfrac{2\pi}{\lambda} | X_u - X_k|}} Y^m_{\ell}(\varphi_{u, k}, \vartheta_{u, k}) \nonumber
\end{equation}
where the angles $\varphi_{u, k}$ and $\vartheta_{u, k}$ are defined by the following equations:
\begin{align}
\sin\varphi_{u, k}\cos\vartheta_{u, k} & = \frac{r_u\sin\varphi_u\cos\vartheta_u - r_k\sin\varphi_k\cos\vartheta_k}{r_{u, k}}\nonumber\\
\sin\varphi_{u, k}\sin\vartheta_{u, k} &= \frac{r_u \sin\varphi_u\sin\vartheta_u - r_k \sin\varphi_k\sin\vartheta_k}{r_{u, k}}\nonumber\\
\cos \varphi_{u, k} &= \frac{r_u\cos \varphi_u - r_k\cos \varphi_k}{r_{u, k}}.
\end{align}
\end{lemma}
Using the fact that
\begin{equation}
J_{\frac{1}{2}}(x) = \sqrt{\frac{2}{\pi}} \frac{\sin x}{\sqrt{x}}, \qquad Y_{0, 0}(\varphi_{u, k}, \vartheta_{u, k}) = \frac{1}{\sqrt{4\pi}}
\end{equation}
and Lemma~\ref{Lem:J:Spy}, the field in \eqref{Eq:Field:Dist:Sph} can be obtained in a closed form, yield the following result.
\begin{lemma}[Field power density distribution] \label{Lem:CC:Field:Sph}
Given the spherical UA and channel conjugate precoding targeting user $X_u$, the field power density measured at the user location $X_k$ is given as
\begin{equation}
\tilde{p}(X_u \mid X_k) = P_{\textrm{t}}\ \mathrm{sinc}^2\left(\tfrac{2\pi}{\lambda}|X_u - X_k|\right) + o. \label{Eq:Field:Dist:a}
\end{equation}
\end{lemma}
As in the case of circular UA, $\tilde{p}(X_u \mid X_k)$ can be rewritten as $\tilde{p}(d)$ with $d$ being the distance from the location targeted by the precoder. Then the distribution of the field power density can be characterized by the function $\tilde{p}(d)/\tilde{p}(0) = \mathrm{sinc}^2\left(\tfrac{2\pi d}{\lambda}\right)$ that is plotted in Fig.~\ref{Fig:Bessel:Bnd}. Like the circular-UA counterpart, by channel-conjugate precoding, the spherical UA focuses the transmission power into a region within a distance of half wavelength from the target location. The advantage of the spherical UA is reflected in that the tail of the distribution function has an envelop diminishing with the increasing distance much faster than that corresponding to the circular UA. This reduces multiuser interference and leads to substantial performance improvements as shown in the analysis and observed from simulation results.
Given the SINR defined similarly as in \eqref{Eq:SINRu}, the result in Lemma~\ref{Lem:CC:Field:Sph} leads to the spherical-UA counterpart of Theorem~\ref{Theo:ChanConj:Tx} as follows.
\begin{theorem}[Receive SINRs] \label{Theo:ChanConj:Sph}
For channel conjugate transmission using the spherical UA, the receive SINR for user $u$, denoted as $\widetilde{\mathsf{SINR}}$ is given as
\begin{equation}
\widetilde{\mathsf{SINR}}_{u} = \frac{1}{\sum\limits_{k\neq u } \mathrm{sinc}^2\left(\frac{2\pi}{\lambda}|X_{u} - X_k|\right) + \frac{1}{\mathsf{SNR}} }+o, \qquad u = 1, 2, \cdots, U \label{Eq:SINR:CC:Sph}
\end{equation}
where the receive SNR is
\begin{equation}\label{Eq:RxSNR:Sph}
\mathsf{SNR} = \frac{P_{\textrm{t}} \lambda^2}{4\pi \sigma^2}.
\end{equation}
\end{theorem}
\begin{remark}[High SNR] \emph{Using the fact $\left|\mathrm{sinc}(d) \right| \leqslant d^{-1}$, the SIR lower bound in \eqref{Eq:SIR} for the circular UA can be modified for the current case as
\begin{equation}
\widetilde{\mathsf{SIR}}_{u} \geqslant \frac{1}{U-1}\left(2\pi\min_{k\neq u}\frac{|X_k - X_u|}{\lambda}\right)^2 + o.
\end{equation}
Thus, the receive SIRs increase with the minimum user-separation distance (in wavelength) at least following a super-linear function with the exponent $2$, which is faster than the sub-linear function for the circular UA (see Remark~\ref{Re:CC:SIR}). This specifies the gain of increasing the UA by one dimension from the perspective of received signal quality.
}
\end{remark}
Last, the received SNR in \eqref{Eq:RxSNR:Sph} suggests the following result.
\begin{corollary}[Propagation loss] \label{Cor:PathLoss:sph} Given channel-conjugate transmission using the spherical UA, the propagation loss is approximately equal to the receive antenna aperture: \emph{
\begin{equation}
\frac{P_{\textrm{r}}}{P_{\textrm{t}}} = \frac{\lambda^2}{4\pi} + o.
\end{equation}}
\end{corollary}
In other words, the path loss is a constant and independent with the propagation distance $r_0$. In contrast, the loss corresponding to the circular UA and the conventional array is inversely proportional to $r_0$ (see Corollary~\ref{Cor:PathLoss}) and $r_0^2$, respectively.
\subsubsection{Multiuser Phase Mode Transmission}
The phase modes for the spherical UA correspond to different spherical harmonics. Then the spherical MU-PM precoder for user $u$, denoted as $ \tilde{f}_u'$, is modified from the circular counterpart in \eqref{Eq:Precod:ZF} as
\begin{equation}\label{Eq:MuPrecod:Sph}
\tilde{f}_u'(\varphi, \vartheta) = \tilde{f}_u(\varphi, \vartheta) \sum_{\ell=0}^\infty\sum_{m=-\ell}^\ell c_{u, m, \ell} Y^m_{\ell}(\varphi, \vartheta)
\end{equation}
where $\{c_{u, m, \ell}\}$ are the precoder coefficients to be designed, $\tilde{f}_u(\varphi, \vartheta)$ is the channel conjugate precoder in \eqref{Eq:CCTX:Sph} and the spherical harmonic function $Y^m_{\ell}(\varphi, \vartheta)$ is defined in \eqref{Eq:Sph:Basis}. By substitution of \eqref{Eq:MuPrecod:Sph} into \eqref{Eq:Sig:DL}, the signal received at user $X_u$ is given as
\begin{align}
\tilde{y}_{u} & = \lambda\sqrt{\frac{P_t}{4\pi}} \sum_{k=1}^U \left( \frac{1}{4\pi } \iint \frac{h^*_k(\varphi, \vartheta)h_u(\varphi, \vartheta)}{|h_{k}(\varphi, \vartheta)|} \sum_{\ell=0}^\infty\sum_{m=0}^\ell c_{k, m, \ell} Y^m_{\ell}(\varphi, \vartheta) \sin \varphi d\varphi d\vartheta \right) x_{k} + z_{u} \nonumber\\
& = \lambda\sqrt{\frac{P_t}{4\pi}} c_{u, 0, 0} x_u + \lambda\sqrt{\frac{P_t}{4\pi}} \sum_{k\neq u }\left( \frac{1}{4\pi } \iint e^{j\tfrac{2\pi}{\lambda} (r_u\cos\psi_u - r_k\cos\psi_k) } \sum_{\ell=0}^\infty \sum_{m=-\ell}^\ell c_{k, m, \ell} Y^m_{\ell}(\varphi, \vartheta) \sin \varphi d\varphi d\vartheta \right) x_{k} \nonumber\\
&\qquad + z_{u} + o \nonumber\\
&= \lambda\sqrt{\frac{P_t}{4\pi}} c_{u, 0, 0} x_u + \lambda\sqrt{\frac{P_t}{4\pi}} \sum_{k\neq u}\sum_{\ell=0}^\infty \sum_{m=0}^\ell c_{k, m, \ell} \mathcal{J}_{u, k, m, \ell} x_{k}+z_{u} + o\label{Eq:Sig:ZF:Sph}
\end{align}
where the coefficients $\{\mathcal{J}_{u, k, m, \ell}\}$ are defined earlier in \eqref{Eq:J:Sph:Def}. It can be observed from Lemma~\ref{Lem:J:Spy} that $\mathcal{J}_{u, k, m, \ell}$ is proportional to $J_{\ell + \frac{1}{2}}(\tfrac{2\pi}{\lambda}|X_u - X_k|)$. Therefore, following the same reason as for the circular UA, only the set of coefficients $\{\mathcal{J}_{u, k, m, \ell}\mid 0 \leq \ell \leq M \}$ have significant values. Considering only these values reduces the precoder coefficients to a finite set $\{c_{k, m, \ell}\mid 0\leq \ell \leq M\}$, yielding the following fact.
\begin{remark}[Degrees of freedom]\label{Re:DoF:Sph}\emph{
Given MU-PM transmission using a spherical UA, the total number of degrees of freedom for interference avoidance is approximately equal to $(M+1)^2$ with $M$ given in \eqref{Eq:DoF}. As a result, since $M$ is much greater than one, the number of degrees of freedom generated by the spherical UA is approximately proportional to $M^2$ that is much larger than that, namely $2M$, for the circular counterpart.
}
\end{remark}
Next, the said finite set of precoder coefficients can be designed by applying the zero-forcing constraints that follow from \eqref{Eq:Sig:ZF:Sph} as
\begin{equation}\label{Eq:ZF:Sph}
\sum_{\ell = 0}^M\sum_{m=-\ell}^\ell c_{k, m, \ell} \mathcal{J}_{u, k, m, \ell} = 0 , \quad \forall \ k \neq u.
\end{equation}
The constraints can be written in the matrix form. To this end, define the matrix $\mathbfcal{H}_{u, \ell}$ as
\begin{equation}\label{Eq:ZF:Coeff:Sph}
\mathbfcal{H}_{u, \ell} = \begin{bmatrix} \mathcal{J}_{u, 1, -\ell, \ell}, & \mathcal{J}_{u, 1, -\ell+1, \ell}, & \cdots & \mathcal{J}_{u, 1, \ell, \ell} \\
\vdots & \vdots &\ddots& \vdots \\
\mathcal{J}_{u, u-1, -\ell, \ell}, &\mathcal{J}_{u, u-1, -\ell+1, \ell}, & \cdots & \mathcal{J}_{u, u-1, \ell, \ell} \\
\mathcal{J}_{u, u+1, -\ell, \ell}, & \mathcal{J}_{u, u+1, -\ell+1, \ell}, & \cdots & \mathcal{J}_{u, u+1, \ell, \ell} \\
\vdots & \vdots &\ddots & \vdots \\
\mathcal{J}_{u, U, -\ell, \ell}, & \mathcal{J}_{u, U, -\ell+1, \ell}, &\cdots & \mathcal{J}_{u, U, \ell, \ell}
\end{bmatrix}\nonumber
\end{equation}
and the row vector ${\mathbf{a}}_{u, \ell} = \begin{bmatrix}c_{u, -\ell, \ell}, & c_{u, -\ell+1, \ell}, & \cdots& c_{u, \ell, \ell} \end{bmatrix}$ where $\ell = 0, 1, \cdots $ . Moreover, using these matrices/vectors as elements, define
\begin{equation}
\tilde{\mathbfcal{J}}_u = \begin{bmatrix}\mathbfcal{H}_{u, 0}, \mathbfcal{H}_{u, 2}, \cdots, \mathbfcal{H}_{u, M}\end{bmatrix}, \qquad
{\mathbf{c}}_u = \begin{bmatrix} {\mathbf{a}}_{u, 1}, {\mathbf{a}}_{u, 2}, \cdots, {\mathbf{a}}_{u, \widetilde{M}}\end{bmatrix}^T.
\end{equation}
Using these definitions, the zero forcing constraints in \eqref{Eq:ZF:Sph} can be written as
\begin{equation}\label{Eq:Approx:ZF:Sph}
\tilde{\mathbfcal{J}}_u \tilde{{\mathbf{c}}}_u ={\mathbf{0}}, \qquad u = 1, 2, \cdots, U.
\end{equation}
The main result of the section is summarized in the following theorem.
\begin{theorem}[Multiuser UA precoding] \label{theorem:MPM:Tx:SPh}
For MU-PM transmission using the circular UA, the precoder coefficients under the zero-forcing constraints are given as
\emph{
\begin{equation}\label{Eq:ZF:Solution:Sph}
\tilde{{\mathbf{c}}}_{u} \in \textsf{null}(\tilde{\mathbfcal{J}}_{u}), \qquad u = 1, 2, \cdots, U
\end{equation}}
where $\textsf{null}(\tilde{\mathbfcal{J}}_{u}) $ denotes the null-space of $\tilde{\mathbfcal{J}}_{u}$.
\end{theorem}
The spherical-UA counterpart of Corollary~\ref{Cor:RxSNR} is as follows.
\begin{corollary}[Receive SNRs]\label{Cor:RxSNR:Sph}For MU-PM transmission using the circular UA, the maximum receive SNR for user $u$ is given as
\emph{
\begin{equation}\label{Eq:SNR:ZF:Sph}
\mathsf{SNR}_u = \frac{\eta^2 P_{\textrm{t}} \left|{\mathbf{e}}_0^\dagger {\mathbf{b}}_{u}\right|^2}{\sigma^2} + o, \qquad u = 1, 2, \cdots, U
\end{equation}}
where ${\mathbf{b}}_{u}$ is a basis of \emph{$\textsf{null}(\tilde{\mathbfcal{J}}_{u})$}.
\end{corollary}
\begin{remark}\emph{Theorem~\ref{theorem:MPM:Tx:SPh} and Corollary~\ref{Cor:RxSNR:Sph} are observed to have the same forms as Theorem~\ref{Theo:ChanConj:Tx} and Corollary~\ref{Cor:RxSNR}, respectively. However, the space $\textsf{null}(\tilde{\mathbfcal{J}}_{u})$ corresponding to the sperical UA is much larger than its circular-UA counterpart $\textsf{null}(\mathbfcal{J}_{u})$. The extra degrees of freedom allow the spherical-UA system to support a larger number of simultaneous users and reduce the received SNR loss due to interference avoidance.
}
\end{remark}
\begin{figure}[t]
\begin{center}
\subfigure[Circular UA]{\includegraphics[width=8cm]{UA_Ran_Cir.pdf}}
\subfigure[Spherical UA]{\includegraphics[width=8cm]{UA_Ran_Sph.pdf}}
\subfigure[Collocated array]{\includegraphics[width=8cm]{Conv_Array.pdf}}
\caption{(a) Circular UA with a radius of $r_0 = 20$ m. (b) Spherical UA with the same radius. (c) Collocated array with a radius of $r_d = 5$ m. The number of users is $10$. }
\label{Fig:UA}
\end{center}
\end{figure}
\section{Simulation Results}\label{Section:Sim}
In simulation, discrete UAs with finite numbers of antennas are considered. As illustrated in Fig.~\ref{Fig:UA}, the UA antennas are uniformly distributed on a circle for the case of circular UA or a sphere for the case of spherical UA, both of which have the fixed radius $r_0 = 20$ m modeling a small deployment region. For benchmarking, the conventional \emph{collocated array} is also included in simulation whose antennas are uniformly distributed in a horizontal disk with a radius denoted as $r_d$ and right above the origin with a vertical distance equal to $r_0$ as shown in Fig.~\ref{Fig:UA}(c). Note that this location and orientation of the collocated UA are found by simulation to yield the best performance among other configurations with the same disk radius and distance to the origin. The radius of the collocated UA is set as $r_d = 5$ m in Fig.~\ref{Fig:UA} for ease of illustration and $r_d = 1$ m for all other simulation results. In addition, relaxing Assumption~\ref{AS:CenterUsers}, the users are uniformly distributed in the horizontal disk with a radius of $0.5r_0$ instead of being near the origin. Moreover, the carrier frequency is $2.5$ GHz and the noise variance is $-100$ dBm.
\subsection{Channel Estimation}
Consider channel estimation using single pilot symbols and algorithms from straightforward extension of those in Sections~\ref{Section:ChanEst} and \ref{Section:ChanEst:Sph} to discrete arrays. The average channel estimation error is defined as the difference between the estimated and actual locations of a typical user as averaged over the random distributions of users and antennas. Fig.~\ref{Fig:ChanEst} displays the curves of average channel estimation error versus the number of antennas and those of average error versus the number of users in two separate sub-figures. Several observations can be made from the curves. As the number of antennas increases, the average errors for the circular and spherical UAs both diminishes rapidly and converges to a small constant corresponding to the continuous arrays analyzed in the preceding sections. Moreover, the errors grows rapidly as the number of users increases. For relatively small numbers of users (e.g, fewer than $9$) or large numbers of antennas (e.g., larger than $200$), the performance of channel estimation for the UAs is much better than that using a collocated UA; the spherical UA outperforms the circular UA. The conventional array's incapability of accurate LI-channel estimation is mainly due to its confined geometry that is more suitable for estimating signals' angles-of-arrival. For verification, it can observed from Fig.~\ref{Fig:ChanEst} that the average error for the collocated array is insensitive to the numbers of antennas and users, suggesting that they are not the performance limiting factors.
\begin{figure}[t]
\begin{center}
\subfigure[Effect of the number of antennas]{\includegraphics[width=8cm]{Chan_Est_Err_Vs_Ant.pdf}}\hspace{10pt}
\subfigure[Effect of the number of users]{\includegraphics[width=8cm]{Chan_Est_Err_Vs_Mobiles.pdf}}
\caption{Consider LI-channel estimation with single pilot symbols. (a) Effect of the number of antennas on channel estimation error for $10$ users. (b) Effect of the number of users on channel estimation error for $200$ transmit antennas.}
\label{Fig:ChanEst}
\end{center}
\end{figure}
\subsection{Data Transmission}
Consider data transmission assuming perfect channel-state-information at the transmitter. The curves of sum throughput versus transmission power per user are plotted in Fig.~\ref{Fig:SumCap:TxPwr} with the number of users fixed at $10$. The plots account for two transmission schemes including multiuser phase-mode (PM) and single-user channel-conjugate (CC) transmission and different numbers of antennas, namely $100$ and $400$. Several observations are made by comparing the curves. First, the combination of spherical UA and PM transmission yields much higher sum throughput than any other combination since distributing antennas over a larger area allows the spherical UA to generate a much higher number degrees of freedom for avoiding interference and enhancing received signal power compared with other arrays (see Remark~\ref{Re:DoF:Sph}). Second, the sum throughputs for CC transmission using three types of arrays are comparable and higher than those achieved by PM transmission using the circular UA and collocated UA in the power range of $-70$ to $-45$ dBm; beyond this range, interference resulting from CC transmission dominates noise, causing the corresponding sum throughputs to saturate. Last, increasing the number of antennas from $100$ to $400$ contributes approximately the same throughput gain, about $20$ b/s/Hz, for different combinations of array and transmission scheme.
\begin{figure}[t]
\begin{center}
\subfigure[Transmit array with $100$ antennas]{\includegraphics[width=8cm]{Cap_Vs_TxPower_UA100.pdf}}
\subfigure[Transmit array with $400$ antennas]{\includegraphics[width=8cm]{Cap_Vs_TxPower_UA400.pdf}}
\caption{Sum throughput versus transmission power per user for a transmit array with (a) $100$ or (b) $400$ antennas. The number of users is fixed at $10$. }
\label{Fig:SumCap:TxPwr}
\end{center}
\end{figure}
The curves of sum throughput versus number of users are plotted in Fig.~\ref{Fig:SumCap:User} with the transmission power per user fixed at $-40$ dBm. For PM transmission, the sum throughputs are observed to increase approximately \emph{linearly} with the increasing number of users or equivalently the increasing number of simultaneous data streams. Furthermore, the curve corresponding to the spherical UA has a slope much larger than those for the other types of arrays that are comparable. In contrast, operating in the interference limiting regime, the sum throughputs for CC transmission using $100$ antennas saturate and are observed to be insensitive to the increase of the number of users. This issue can be alleviated by deploying more antennas ($400$) that leads to a substantial throughput gain e.g., about $20$ b/s/Hz for the number of users equal to $18$. Nevertheless, the gains for PM transmission are much larger and about $60$ b/s/Hz at the same number of users.
\begin{figure}[t]
\begin{center}
\subfigure[UA with $100$ antennas]{\includegraphics[width=8cm]{Cap_Vs_User_UA100.pdf}}
\subfigure[UA with $400$ antennas]{\includegraphics[width=8cm]{Cap_Vs_User_UA400.pdf}}
\caption{Sum throughput versus number of users for a transmit array with (a) $100$ or (b) $400$ antennas. The transmission power per users is fixed at $-40$ dBm. }
\label{Fig:SumCap:User}
\end{center}
\end{figure}
\section{Conclusion}\label{Section:Conclusion}
Techniques have been designed for channel estimation and data transmission in the UA system by modeling the UA as a gigantic continuous circular/spherical array and assuming free-space propagation. It has been shown that the UA enables accurate estimation of multiuser channels even if single pilot symbols are used provided that the user-separation distances are sufficiently large. If orthogonal pilot sequences are used, channel estimation is found to be always close to perfect. For single-user data transmission using the UA, inter-user interference can be suppressed by increasing user-separation distances. Alternatively, interference can be nulled at the UA based a novel design of multiuser phase-mode precoders. The resultant number of available degrees of freedom for interference nulling is shown to be proportional to the minimum user-separation distance. Furthermore, the spherical UA provides performance gain compared with the circular counterpart.
This work opens up several interesting directions for further research. First, the current analysis is based on the model of a continuous circular/spherical UA and targets users near the UA center. Generalizing the model and user locations makes the analysis more challenge and requires the development of new analytical techniques. Second, it is important to address practical issues in the design and analysis of UA communication techniques such as delay and error in message exchange between the UA elements and the presence of sparse scatterers. Furthermore, it is also interesting to design algorithms/protocols for resource allocation, broadband transmission, and power control for the UA systems.
\appendices
\section{Mathematics Preliminary: Bessel Functions and Spherical Harmonics}\label{App:Bessel}
Bessel functions and spherical harmonics are extensively used in the subsequent analysis. In this appendix, the functions are defined and some key properties useful for the analysis are summarized.
\subsection{Bessel Functions and Their Properties}
Only Bessel functions of the first kind are needed in the analysis and referred to simply as Bessel functions. A Bessel function with an integer order $n$ can be defined in an integral form as \cite{arfken_2005_mathematical_physics}:
\begin{equation}
J_n(x) = \frac{1}{2\pi}\int_0^{2\pi} e^{j(x\sin \varphi - n\varphi)} d\varphi. \label{Eq:Bessel:Def}
\end{equation}
Several useful properties of Bessel functions are described as follows.
\begin{enumerate}
\item [(B1)] The Jacobi-Anger expansion decomposes an exponential function of a trigonometric function into its harmonics as follows \cite{arfken_2005_mathematical_physics}:
\begin{equation}
e^{jx\cos\varphi} = \sum_{n=-\infty}^\infty j^n J_n(x) e^{j k \varphi}.
\end{equation}
\item [(B2)] Addition Theorem I \cite[(6.61)]{andrews_1992_special}:
\begin{equation}
J_n(R) e^{jn\omega} = \sum_{k=-\infty}^\infty J_k(a)J_{n+k}(b)e^{jk\varphi}\nonumber
\end{equation}
where $R=\sqrt{a^2+b^2-2ab\cos\varphi}$ and $\sin\omega = (a/R)\sin\varphi$.
\item [(B3)] Addition Theorem II rewritten from \cite[(6.62)]{andrews_1992_special}:
\begin{equation}
\mathrm{sinc}(R) = \frac{\pi}{2} \sum_{n=0}^\infty \left(2n + 1\right)\frac{J_{n+\frac{1}{2}}(a)J_{n+\frac{1}{2}}(b)}{\sqrt{ab}}P_n(\cos \varphi)\nonumber
\end{equation}
where $P_n(x)$ with $x\in [-1, 1]$ is the Legendre polynomial defined as
\begin{equation}\label{Eq:Legendre}
P_n(x) = \frac{1}{2^n n !}\frac{d^n}{dx^n}(x^2 - 1)^n.
\end{equation}
\item [(B4)] The Bessel function $J_0(x)$ can be upper bounded as \cite{landau_2000_bound_besselfun}
\begin{equation}
J_0(x) \leqslant \nu x^{-\frac{1}{3}}
\end{equation}
where the constant $\nu = 0.7857\cdots$.
\item [(B5)] Given $0 < z \leq 1$, a Bessel function with a high order satisfies \cite[$9.3.5$ and $9.3.6$]{AbramowitzBook}\footnote{Two functions $f$ and $g$ are \emph{asymptotic equivalent}, denoted as $f(x) \sim g(x)$, if $\lim_{x\rightarrow\infty} f(x)/g(x) = 1$.}
\begin{equation}
J_n(z n ) \sim \frac{c(z)}{n^{\frac{1}{3}}}, \qquad n \rightarrow \infty
\end{equation}
where $c(z)$ is a positive constant whose value depends only on $z$. Consequently, for $x\gg 1$ and $|n| \geqslant x$, $J_n(x) \approx 0$ \cite{PoonTse:DoFMultiAntennaChannels:2005}.
\item[(B6)] Gegenbauer's generalization of Poisson's integral \cite[$10.1.14$]{AbramowitzBook}:
\begin{equation}\label{Eq:Poisson:Int}
\sqrt{\frac{\pi}{2x}} J_{n+\frac{1}{2}}(x) = \frac{1}{2}(-j)^n \int_{-1}^1 e^{jx\tau}P_n(\tau)d\tau
\end{equation}
where $P_n(x)$ is the Legendre polynomial defined in \eqref{Eq:Legendre}.
\end{enumerate}
\subsection{Spherical Harmonics and Their Properties}
The spherical harmonic functions denoted as $\{Y^m_{\ell}(\varphi, \vartheta)\}$ with integer indices $\ell = 0, 1, \cdots$ and $-\ell \leqslant m \leqslant \ell$ are defined as \cite{arfken_2005_mathematical_physics}:
\begin{equation}\label{Eq:Sph:Basis}
Y^m_{\ell}(\varphi, \vartheta) = \sqrt{\frac{2m+1}{4\pi}\frac{(\ell-m)!}{(\ell+m)!}}P^m_{\ell}(\cos \vartheta)e^{jm\varphi}, \qquad \vartheta\in [0, \pi], \varphi \in [0, 2\pi]
\end{equation}
where $P^m_{\ell}(\cos \vartheta)$ represents the \emph{associated Legendre polynomial} defined as
\begin{equation}\label{Eq:Legendre:Assoc}
P^m_{\ell}(x) = \frac{(-1)^m}{2^\ell \ell !}\left(1 - x^2\right)^{\frac{m}{2}} \frac{d^{m+\ell}}{dx^{m+\ell}} \left(x^2 - 1\right)^\ell, \qquad -1 \leq x \leq 1.
\end{equation}
Note that $P^m_{\ell}(\cos \varphi)$ with $m = 0$ reduces to the Legendre polynomial in \eqref{Eq:Legendre}.
Several useful properties of the spherical harmonics are described as follows.
\begin{enumerate}
\item [(S1)] The functions $\{Y^m_{\ell}(\varphi, \vartheta)\}$ are orthonormal over the spherical surface:
\begin{equation}\label{Eq:Orthogonal}
\int_{\varphi = 0}^{2\pi}\int_{\vartheta = 0}^\pi [(Y_{\ell}^m)(\varphi, \vartheta)]^* Y_{\ell'}^{m'}(\varphi, \vartheta) \sin\vartheta \ d\vartheta d\varphi = \delta_{m, m'}\delta_{\ell, \ell'}
\end{equation}
where $\delta_{m, m'} $ is equal to $1$ if $m = m'$ and $0$ otherwise.
\item[(S2)] Funk-Hecke Theorem \cite[Theorem~$3$]{Gumerov:FastMultipoleHelmholtzEq:2005}: Let $\psi_u(\varphi, \vartheta)$ denote the angle between the vectors $X_u = (r_u, \varphi_u, \vartheta_u)$ and $(1, \varphi, \vartheta)\in \mathds{R}^3$. The Laplace series of a function $w(\cos \psi_u(\varphi, \vartheta))$ is given as
\begin{equation}\nonumber
w(\cos \psi_u(\varphi, \vartheta)) = \sum_{\ell=0}^\infty \sum_{m=-\ell}^\ell c_{\ell}^m(\varphi_u, \vartheta_u) Y^m_{\ell}(\varphi, \vartheta)
\end{equation}
where the coefficient $c_{\ell}^m(\varphi_u, \vartheta_u) = c_\ell Y^m_{\ell}(\varphi_u, \vartheta_u)$
with
\begin{equation}
c_\ell = 2\pi \int_{-1}^1 w(\tau ) P_\ell(\tau)d\tau. \nonumber
\end{equation}
\item[(S3)] Spherical Harmonic Addition Theorem \cite[($16.57$)]{arfken_2005_mathematical_physics}:
\begin{equation}\label{Eq:Sph:Add}
\sum_{m=-\ell}^\ell [Y^m_{\ell}(A)]^* Y^m_{\ell}(X_u) = \frac{2\ell+1}{4\pi}P_\ell\left(\cos\psi_u(A)\right)
\end{equation}
where $A, X_u \in \mathds{R}^3$ and $\psi_u(A)$ denotes their separation angle.
\end{enumerate}
\section{Proofs of Lemmas}\label{App:Proofs}
\subsection{Proof of Lemma~\ref{Lem:Fourier}}\label{Lem:Fourier:Proof}
By substituting \eqref{Eq:Training:Sig} and $\varphi_u(A) = \varphi_u - \varphi$ into \eqref{Eq:Fourier},
\begin{align}
Q_k &= \frac{\lambda e^{-j \frac{2\pi}{\lambda}r_0}}{4\pi r_0} \sum_{u=1}^U \int_0^{2\pi} e^{j\frac{2\pi }{\lambda}r_u\cos(\varphi_u-\varphi)}e^{jk\varphi} d\varphi + \frac{1}{2\pi}\int_0^{2\pi} z(\varphi) e^{jk\varphi}d\varphi + \frac{o}{r_0}. \label{Eq:Spectrum}
\end{align}
The noise term can be written as
\begin{equation}\label{Eq:Noise}
\frac{1}{2\pi}\int_0^{2\pi} z(\varphi) e^{jk\varphi}d\varphi = \lim_{N\rightarrow\infty}\frac{1}{N}\sum_{n=1}^N z\left(\frac{2\pi n}{N}\right) e^{jk\frac{2\pi n}{N}}.
\end{equation}
For ease of notation, define $\tilde{z}_n= z\left(\frac{2\pi n}{N}\right) e^{jk\frac{2\pi n}{N}}$. Since $z_0, z_1, \cdots, \tilde{z}_N$ is an i.i.d. sequence of $\mathcal{CN}(0, \sigma^2)$ random variables under Assumption~\ref{AS:White}, by applying the law of large numbers, it follows from \eqref{Eq:Noise} that
\begin{equation}\label{Eq:Noise:a}
\frac{1}{2\pi}\int_0^{2\pi} z(\varphi) e^{jk\varphi}d\varphi = 0, \qquad \text{a.s.}
\end{equation}
Next, based on the Jacobi-Anger expansion in Property (B1) of Bessel functions in Appendix~\ref{App:Bessel}, the first exponential term in \eqref{Eq:Noise} can be decomposed as
\begin{equation}\label{Eq:Jacobi-Anger}
e^{j\frac{2\pi }{\lambda}r_u\cos(\varphi_u-\varphi)} = \sum_{n=0}^\infty j^{n}J_{n}\left(\tfrac{2\pi }{\lambda} r_u \right) e^{j n(\varphi_u-\varphi)}.
\end{equation}
By substituting \eqref{Eq:Noise:a} and \eqref{Eq:Jacobi-Anger} into \eqref{Eq:Spectrum}, it can be obtained that
\begin{equation}
Q_k = \frac{\lambda e^{-j \frac{2\pi}{\lambda}r_0}}{4\pi r_0} \sum_{u=1}^U \sum_{n=0}^\infty j^{n}J_{n}\left(\tfrac{2\pi }{\lambda}r_u\right) \int_0^{2\pi} e^{jn(\varphi_u-\vartheta)}e^{jk\vartheta} d\vartheta + \frac{o}{r_0}, \qquad \text{a.s.} \label{Eq:Spectrum:a}
\end{equation}
Based on the following equality
\begin{equation}\label{equ:exponentials_integration}
\frac{1}{2\pi}\int_0^{2\pi} e^{j(k-n)\vartheta} d\vartheta = \delta_{k, n}, \nonumber
\end{equation}
the desired result follows from \eqref{Eq:Spectrum:a}. \hfill$\Box$
\subsection{Proof of Lemma~\ref{Lem:CC:Field}}\label{App:CC:Field}
Using trigonometric identities, it can be obtained that
\begin{align}
r_u \cos(\varphi_u-\vartheta)-r_k\cos(\varphi_k-\vartheta) & = |X_u - X_k| \left(\cos \beta_{u, k} \cos\vartheta - \sin\beta_{u, k}\sin\vartheta\right)\nonumber\\
&= |X_u - X_k| \cos (\beta_{u, k}+\vartheta) \nonumber
\end{align}
where the angle $\beta_{u, k}$ is defined by
\begin{equation}
\tan \beta_{u, k} = \frac{r_u\cos(\varphi_u) - r_k\cos(\varphi_k)}{r_u\sin(\varphi_u) - r_k\sin(\varphi_k)}.
\end{equation}
It follows that
\begin{align}
\frac{1}{2\pi} \int_0^{2\pi}e^{j\frac{2\pi}{\lambda}\left(r_u \cos(\varphi_u-\vartheta)-r_k\cos(\varphi_k-\vartheta)\right)} \ d\vartheta &= \frac{1}{2\pi} \int_0^{2\pi}e^{j\frac{2\pi}{\lambda}|X_u - X_k| \cos (\beta_{u, k}+\vartheta) } \ d\vartheta\nonumber\\
&= \frac{1}{2\pi} \int_0^{2\pi}e^{j\frac{2\pi}{\lambda}|X_u - X_k| \sin\left(\beta_{u, k}+\vartheta + \tfrac{\pi}{2}\right)} \ d\vartheta. \nonumber
\end{align}
Using the definition of Bessel function in \eqref{Eq:Bessel:Def},
\begin{equation}
\frac{1}{2\pi} \int_0^{2\pi}e^{j\frac{2\pi}{\lambda}\left(r_u \cos(\varphi_u-\vartheta)-r_k\cos(\varphi_k-\vartheta)\right) } \ d\vartheta = j e^{j \beta_{u, k}} J_m\left(\tfrac{2\pi}{\lambda}|X_u - X_k|\right).
\label{Eq:Trigo}
\end{equation}
Substituting \eqref{Eq:Trigo} into \eqref{Eq:Field:Dist:cir:a} gives
\emph{\begin{equation}
g(X_k\mid X_u) = \sqrt{\frac{P_{\textrm{t}}}{2 r_0}} j e^{j \beta_{u, k}} J_0\left(\tfrac{2\pi}{\lambda}|X_u - X_k|\right) + \frac{o}{\sqrt{r_0}}. \nonumber
\end{equation}}
The desired result follows. \hfill$\Box$
\subsection{Proof of Lemma~\ref{Lem:Sph:Expan}}\label{Lem:Sph:Expan:Proof}
By substitution of the training signal in \eqref{Eq:Training:Sig}, its Laplace coefficients defined in \eqref{Eq:Q:Sph} are obtained as
\begin{align}
Q_{\ell}^m &= \frac{\eta}{r_0}\sum_{u=1}^U\int_{\vartheta=0}^{2\pi} \int_{\psi=0}^\pi e^{j\frac{2\pi}{\lambda}r_u\cos \psi_u }Y^m_{\ell}(\varphi, \vartheta) \sin\varphi d\varphi d\vartheta+\nonumber\\
&\qquad \qquad \int_{\vartheta=0}^{2\pi} \int_{\psi=0}^\pi z(\varphi, \vartheta) \sin\varphi d\varphi d\vartheta + \frac{o}{r_0}\nonumber\\
&= \frac{\eta}{r_0}\sum_{u=1}^U\int_{\vartheta=0}^{2\pi} \int_{\psi=0}^\pi e^{j\frac{2\pi}{\lambda}r_u\cos \psi_u }Y^m_{\ell}(\varphi, \vartheta) \sin\varphi d\varphi d\vartheta + \frac{o}{r_0}, \qquad \text{a.s.}\label{Eq:Q:Sph:App}
\end{align}
where the vanishment of noise follows similar analysis as in the proof for Lemma~\ref{Lem:Fourier}. Using the Funk-Hecke Theorem in (S$2$) , the exponential term in the last equation can be also expanded into a Laplace series as
\begin{equation}\label{Eq:Sph:Exp}
e^{j\frac{2\pi}{\lambda} r_u\cos \psi_u } = \sum_{\ell=0}^\infty \sum_{m=-\ell}^\ell c_\ell Y^m_{\ell}(\varphi_u, \vartheta_u) Y^m_{\ell}(\psi, \vartheta)
\end{equation}
where
\begin{equation}
c_\ell = 2\pi \int_{-1}^1 e^{j\frac{2\pi}{\lambda} r_u \tau } P_\ell(\tau)d\tau. \label{Eq:Cn}
\end{equation}
Using Property (B$6$) in Appendix~\ref{App:Bessel}, it follows from \eqref{Eq:Cn} that
\begin{equation}\label{Eq:Cn:more}
c_\ell(r_u) = \frac{(2\pi)^{\frac{3}{2}} j^\ell J_{\ell+\frac{1}{2}}\left(\frac{2\pi}{\lambda}r_u\right)}{\left(\frac{2\pi}{\lambda}r_u\right)^{\frac{1}{2}}} Y^m_{\ell}(\varphi_u, \vartheta_u).
\end{equation}
Combining \eqref{Eq:Sph:Exp} and \eqref{Eq:Cn:more} gives the desired result. \hfill$\Box$
\subsection{Proof of Lemma~\ref{Lem:J:Spy}}\label{App:J:Spy:Proof}
Since $\psi_u$ is the angle between two unit vectors with spherical coordinates $(1, \varphi, \vartheta)$ and $(1, \varphi_u, \vartheta_u)$,
\begin{equation}
\cos \psi_u = \sin\varphi\cos \vartheta\sin\varphi_u\cos\vartheta_u + \sin\varphi\sin \vartheta\sin\varphi_u\sin\vartheta_u + \cos\varphi\cos \varphi_u.
\end{equation}
It follows that
\begin{align}
r_u \cos \psi_u - r_k\cos \psi_k & = \sin\varphi\cos \vartheta(r_u\sin\varphi_u\cos\vartheta_u - r_k\sin\varphi_k\cos\vartheta_k) + \nonumber\\
&\qquad \sin\varphi\sin \vartheta(r_u \sin\varphi_u\sin\vartheta_u - r_k \sin\varphi_k\sin\vartheta_k)+\nonumber\\
&\qquad \cos\varphi(r_u\cos \varphi_u - r_k\cos \varphi_k).\label{Eq:PhaseDiff:Sph}
\end{align}
To rewrite the right-hand side of \eqref{Eq:PhaseDiff:Sph} in a compact form, define the spherical coordinates $(r_{u, k}, \varphi_{u, k}, \vartheta_{u, k})\in\mathds{R}^3$ such that
\begin{align}
r_{u, k}^2 &= (r_u\sin\varphi_u\cos\vartheta_u - r_k\sin\varphi_k\cos\vartheta_k)^2 + (r_u \sin\varphi_u\sin\vartheta_u - r_k \sin\varphi_k\sin\vartheta_k)^2 + \nonumber\\
&\qquad (r_u\cos \varphi_u - r_k\cos \varphi_k)^2\label{Eq:r_uk:Def}
\end{align}
and the angles $\varphi_{u, k}$ and $\vartheta_{u, k}$ as in the lemma statement. Moreover, let $\psi_{u, k}$ denote the angle between $(r_0, \varphi, \vartheta)$ and $(r_{u, k}, \varphi_{u, k}, \vartheta_{u, k})$. Using these definitions, \eqref{Eq:PhaseDiff:Sph} can be reduced to
\begin{align}
r_u \cos \psi_u - r_k\cos \psi_k & = r_{u, k} \cos \psi_{u, k}
\end{align}
It can be obtained from \eqref{Eq:r_uk:Def} that $r_{u, k}^2 = |X_u - X_k|^2$. Thus,
\begin{equation}
r_u \cos \psi_u - r_k\cos \psi_k = |X_u - X_k| \cos \psi_{u, k}.
\end{equation}
Substituting this result into \eqref{Eq:J:Sph:Def} yields
\begin{equation}
\mathcal{J}_{u, k, m, \ell} = \iint e^{j\tfrac{2\pi}{\lambda} |X_u - X_k| \cos \psi_{u, k}} Y^m_{\ell}(\varphi, \vartheta) \sin\varphi \ d\varphi \ d\vartheta. \label{Eq:J:Sph:Proof}
\end{equation}
Applying Lemma~\ref{Lem:Sph:Expan} gives the desired result. \hfill$\Box$
\bibliographystyle{ieeetr}
|
\section{Introduction}
From a physical point of view, turbulence is characterized by the presence of an energy flow which, through local nonlinear interactions, transfers all the energy of the velocity field from larger to smaller scales. The energy is then dissipated by a viscous term. The energy-cascade mechanism is a phenomenon widely observed in fluid dynamics experiments and at present lacking rigorous mathematical understanding.
Dyadic (or shell) models have been studied as toy models for the Euler and Navier-Stokes equations (inviscid and viscous case respectively). Even though much simpler, they display peculiar features of the nonlinear structure of fluid dynamic equations. In particular, they mimic the scale by scale local transfer of energy occurring in three-dimensional turbulent flows. Among the most remarkable applications it is worth to mention \cite{Tao14}, in which a dyadic model type approximation is used to show the emergence of blow-up for an averaged Navier-Stokes system.\\
The dyadic model was early introduced by Desnianskii and Novikov in 1974 \cite{DeNo74} and, in recent times, independently reintroduced by Katz and Pavlovi\'c in 2005 \cite{KaPa05}. Since then, it has met the interest of a wide scientific community and several versions have been extensively investigated: viscous \cite{BaMoRo11, BeFe12}; inviscid \cite{BaMo13, BaMoRo11, BeFe12, ChCoFrSh08, ChFrPa10, KiZl05}; stochastically forced, either with additive \cite{FrG-HVi, Rom14} or multiplicative noise \cite{BaFlMo10, BaFlMo11, BaMo13stochastic, Bia13}.
The present paper is concerned with the study of the following stochastically driven shell model
\begin{equation}\label{model_stochastic}
\left\{\begin{array}{l l }
du_0=-u_0u_1dt+\sigma dW(t) &\\
du_j= \left( -2^{cj}u_ju_{j+1}+2^{c(j-1)}u_{j-1}^2 \right) dt & \mbox{ for } j\geq1 \\
u(0) = \underline{u}, &
\end{array}\right.
\end{equation}
with $t \in [0,T]$, $\sigma \in \mathbb{R}_+$, $c \, \in \, [1,3]$, $\underline{u} \in \ell^2$ , $\underline{u}_j\geq 0$ for every $j\geq 1$, and where $\{W(t): t \geq 0\}$ is a one dimensional Brownian motion. \\
The equations describe an inviscid dyadic system, whose peculiarity is that the additive noise affects only the first component. The random perturbation is therefore one dimensional unlike what is treated in \cite{BaFlMo10, BaFlMo11, BaMo13stochastic, Bia13, Rom14} where the noise acts on all the components and is indeed infinite dimensional. \\
The constant $c$ is an intermittency parameter and, roughly speaking, represents the velocity of the energy transfer from shell to shell. The range of values $c \in \left[ 1, \frac{5}{2} \right]$ is essentially the one corresponding, within the simplification of the model, to the 3D Euler equations. The range arises from scaling arguments applied to the nonlinear term \cite{ChFr09}.\\
Observe that the peculiar expression of the quadratic terms in \eqref{model_stochastic} provides the \emph{formal} conservation property
\[
\sum_{j=0}^{+\infty} u_j \left( 2^{c(j-1)}u_{j-1}^2 -2^{cj}u_ju_{j+1} \right) = 0
\]
that gives an a priori bound on the solution $u$ in $ \ell^2$ uniformly in $c$. Additionally, the fact that the state variables lie on the positive real half-line forces the energy to move from lower to higher wave numbers only.
Model \eqref{model_stochastic} has been introduced in \cite{FrG-HVi}, where the authors establish that it admits invariant martingale solutions. In our paper we focus on strong solutions and prove that a strong stationary distribution exists and is unique in the class of positive solutions. We would like to stress here that the monotonicity of the energy flow in dyadic models allows to derive our statements by using arguments and techniques that are more immediate than those used for the stochastic 2D Navier-Stokes problem \cite{HaMa06,HaMa08}. In particular, the key point is having at hand a sort of pathwise contraction property \eqref{contrazione} of the dynamics that, in turn, enables to get uniqueness of the stationary distribution almost straightforwardly. An analogue of \eqref{contrazione} is only known to be valid for Burgers equations \cite{Bor13}; whereas, it is believed to be false for Euler equations.
More in detail, our manuscript is organized as follows. \\
\emph{Section~\ref{pathwise}.} We deal with existence and uniqueness of pathwise solutions for problem~\eqref{model_stochastic}. We start by showing that there exists a uniform energy bound, which ensures the existence of a global pathwise solution (see Subsection~\ref{subsect:existence}). In Subsection~\ref{subsect:regularity} we provide a regularity condition satisfied by the trajectories. The proof follows from a modification of an approach used in \cite{FrG-HVi}. Our idea consists in replacing expectations by integrals over a time interval so to get a pathwise, rather than on average, property. The statements in Subsections~\ref{subsect:existence} and \ref{subsect:regularity} are valid over the entire range of $c \in [1,3]$. At the end of the section we show continuity with respect to positive initial conditions and noise in the range of $c \in [1,3)$ (see Subsection~\ref{subsect:continuity}). The derivation of this result makes crucial use of the solution regularity obtained in the previous subsection. With this in hand, we straightforwardly get pathwise uniqueness of the solution. \\
\emph{Section~\ref{sect:strong:sol}.} This section is devoted to demonstrating existence and uniqueness in a strong sense both of the solution and of the invariant distribution. We begin by proving that the continuity of the unique pathwise solution of \eqref{model_stochastic} guarantees the adaptability of the trajectory with respect to the filtration generated by the initial datum and the Brownian motion, providing strong existence and uniqueness. At this stage, having uniqueness in law (due to the classical Yamada and Watanabe theorem) and the existence of a weak stationary solution (provided by \cite{FrG-HVi}), we are also able to ensure that a positive strong statistically stationary solution of \eqref{model_stochastic} exists. As for the uniqueness of this statistically invariant state, the proof is based on an optimal transport argument. Roughly speaking, we introduce a distance between probability measures as a cost function to be minimized. Then, we show that if two different stationary distributions existed, the minimality prescribed by Kantorovich's formulation of the problem would be violated. The statements in Section~\ref{sect:strong:sol} are valid in the restricted range of $c \in [1,3)$.\\[-.4cm]
To our knowledge we provide the first result of strong existence and pathwise uniqueness for a stochastic shell model in the inviscid case.
\section{Pathwise solution}\label{pathwise}
In this section we consider the deterministic system obtained from \eqref{model_stochastic} by fixing a realization of the Brownian motion. Observe that in system \eqref{model_stochastic} the stochastic integral appears only in the equation
$du_0(t)=-u_0(t)u_1(t)dt+\sigma dW(t)$, which is equivalent to
$$u_0(t)=u_0(0)-\int_0^tu_1(s)u_0(s)ds+\sigma W(t) \ \ a.s..$$
Therefore, for every $\omega \in \Omega$ such that $w(\cdot):=W(\cdot,\omega) \in C([0,T],\mathbb{R})$ with $w(0)=0$, it is natural to define the following infinite dimensional deterministic system:
\begin{equation}\label{model}
\left\{\begin{array}{l l }
\displaystyle{u_0(t)=u_0(0)-\int_0^tu_1(s)u_0(s)ds+\sigma w(t)} &\\
du_j= \left( -2^{cj}u_ju_{j+1}+2^{c(j-1)}u_{j-1}^2 \right) dt & \mbox{ for } j\geq1 \\
u(0) = \underline{u}\, , &
\end{array}\right.
\end{equation}
where $t \in [0,T]$, $\sigma \in \mathbb{R}_+$, $c \, \in \, [1,3]$, $\underline{u} \in H_+$ and $w \in C([0,T],\mathbb{R})$ with $w(0)=0$. Here $H_+$ is the set of sequences which are positive away from the component $j=0$, that is
$H_+= \left\{u \, \in \, \ell^2 \, : u_j\geq 0, \ j\geq1 \right\}$.
\begin{remark}
If the initial condition $\underline{u} \in H_+$, then $u(t) \in H_+$ for all $t \in [0,T]$. We will make use of this positivity condition in the forthcoming computations.
\end{remark}
In this section, we will obtain results concerning solutions of \eqref{model}. Since they are provided for a fixed element $\omega \in \Omega$, they are equivalent to pathwise results for solutions of the stochastic system \eqref{model_stochastic}. We start by introducing the definition of solution for \eqref{model}.
\begin{definition}[Solution of system \eqref{model}]
We say that $u$ is a \emph{solution} of system \eqref{model} on $[0,T]$ with initial condition $\underline{u} \in H_+$ and noise $w \in C([0,T],\mathbb{R})$ with $w(0)=0$,
if $u$ satisfies system \eqref{model}, $u_j \in C([0,T], \mathbb{R})$ for all $j$ and $u(t) \in H_+$ for all $t \in [0,T]$.
\end{definition}
In the sequel we will denote by $\| \cdot \|$ the $\ell^2$-norm and by $\| \cdot \|_{\infty}$ the sup-norm.
\subsection{Existence}\label{subsect:existence}
The aim of this subsection is to establish an existence result for system \eqref{model}.
\begin{theorem}[Existence]\label{pathwise_exist}
For every $T \in [0,+\infty)$, $\sigma \in \mathbb{R}_+$, $c \in [1,3]$, $\underline{u} \in H_+$ and $w \in C([0,T],\mathbb{R})$ with $w(0)=0$, system \eqref{model} admits at least a solution.
\end{theorem}
The idea of the proof consists in considering a truncated version of \eqref{model}, for which global existence is ensured by uniform energy estimates, and then taking the limit with standard arguments.
\begin{proof}
We first introduce a finite dimensional truncation of \eqref{model}. For any $N \in \mathbb{N}$, we consider the first $N$ equations of \eqref{model}, that is
\begin{equation}\label{modelN}
\left\{\begin{array}{l l }
\displaystyle{u^{(N)}_0(t)=u_0(0)-\int_0^tu^{(N)}_0(s)u^{(N)}_1(s)ds+\sigma w(t)} &\\
du^{(N)}_j(t)= \left[-2^{cj}u^{(N)}_j(t)u^{(N)}_{j+1}(t)+2^{c(j-1)} \left(u^{(N)}_{j-1}(t) \right)^2\right] dt & \mbox{ for } j=1,\dots,N\\
u_j^{(N)}(0) = \underline{u}_j
&
\mbox{ for } j=1, \dots, N \\
u^{(N)}_j(t)\equiv0 & \mbox{ for } j \geq N+1
\end{array}\right.
\end{equation}
with $t \in [0,T]$ and $\underline{u}^{(N)} \in H_+$.
System (\ref{modelN}) satisfies the hypotheses of Cauchy-Lipschitz theorem and then admits a local solution $\left(u_n^{(N)} \right)_{n \in \mathbb{N}}$ on $[0,\delta]$, for a certain $\delta>0$.
We start by showing that the component $u_0^{(N)}(t)$ in \eqref{modelN} is uniformly bounded in $N$. To do so, we consider equation
\begin{equation*}\label{u_0}
u_0^{(N)}(t)=u^{(N)}_0(0)-\int_0^tu^{(N)}_0(s)u^{(N)}_1(s)ds +\sigma w(t).
\end{equation*}
Without loss of generality, we may assume
\[
\max_{t \in [0,\delta] } \left| u^{(N)}_0(t) \right|= u^{(N)}_0(t^*) > 0,
\]
for some $t^* \in [0,\delta]$ (the case $u^{(N)}_0(t^*) < 0$ can be treated similarly by symmetry). Now, let $\bar{t} = \sup \left\{ t \leq t^*: u^{(N)}_0(t)=0 \right\}$; if the set $\left\{ t \leq t^*: u^{(N)}_0(t)=0 \right\}$ is empty, let $\bar{t}=0$. Then, we have
\begin{align*}
u^{(N)}_0(t^*) & \leq \left\vert u^{(N)}_0(0) \right\vert -\int_{\bar{t}}^{t^*}u^{(N)}_0(s)u^{(N)}_1(s) ds + \sigma [w(t^*)-w(\bar{t})] \\
&\leq \left\vert u^{(N)}_0(0) \right\vert + 2\sigma \sup_{t\in[0,T]}|w(t)| .
\end{align*}
Therefore, for every $t$, it holds
\begin{equation}\label{bound_u0}
\left| u_0^{(N)}(t) \right| \leq a\, ,\quad \mbox{ with } \quad a := \|\underline{u}\| + 2 \sigma \|w\|_{\infty} \, .
\end{equation}
Now, we apply \eqref{bound_u0} to get a bound for the energy associated with dyadic model \eqref{modelN}. We start by considering the inequality
\[
\left\| u^{(N)}(t) \right\|^2 = \sum_{j=0}^N \left( u_j^{(N)}(t) \right)^2 \leq a^2 + \sum_{j=1}^N \left( u^{(N)}_j(t) \right)^2 \,.
\]
It is easy to see that, by taking the derivative of the summation term in the right-hand side of previous formula, thanks to cancellations, we get
\[
\hspace{0.5cm }\frac{d}{dt}\left [\sum_{j=1}^N \left( u_j^{(N)}(t) \right)^2\right ]=2 \left( u_0^{(N)}(t) \right)^2 u_1^{(N)}(t) \leq 2 a^2 u^{(N)}_1(t) \leq 2 a^2 \sqrt{\sum_{j=1}^N \left( u_j^{(N)}(t) \right)^2} ,
\]
from which it follows by comparison
\[
\sum_{j=1}^N \left( u_j^{(N)}(t) \right)^2\leq \left( a^2 t +\sqrt{\sum_{j=1}^N \left( u_j^{(N)}(0) \right)^2}\right)^2\leq \left(a^2t+a\right)^2.
\]
Summarizing, we conclude
\begin{equation} \label{estimateN}
\left\|u^{(N)}(t)\right\|^2\leq a^2 + \left( a^2t +a \right)^2\leq \left(a^2T+2a\right)^2.
\end{equation}
From this bound we get global existence of the solution of system \eqref{modelN}.\\
We are left to prove existence for the infinite dimensional system \eqref{model}. We will obtain the result by means of Ascoli-Arzel\`a theorem and a standard diagonal argument. \\
For every fixed $j$ and $t \in [0,T]$, it holds:
\begin{itemize}
\item[i)] Uniform boundedness of $\left( u_j^{(N)} (t) \right)_{N \in \mathbb{N}}$ in both $N$ and $t$:
\begin{equation*}
\left\vert u_j^{(N)}(t) \right\vert \leq \left\|u^{(N)}(t)\right\| \leq a^2 T + 2a;
\end{equation*}
\item[ii)] Equi-Lipschitzianity of $\left( u_j^{(N)} (t) \right)_{N \in \mathbb{N}}$ with respect to $N$: by i), we have
\begin{equation*}
\left\vert \frac{d}{dt} u^{(N)}_j(t) \right\vert \leq 2^{cj+1} \, \left\|u^{(N)}(t)\right\|^2 \leq 2^{cj+1} \, \left( a^2 T + 2a \right)^2.
\end{equation*}
\end{itemize}
Ascoli-Arzel\`a theorem implies for each fixed $j$ the existence of a convergent subsequence in $C([0,T])$; i.e., it is possible to find indeces $\left\{ N^j_k, k \in \mathbb{N} \right\}$ such that
\[
\sup_{t \in [0,T]} \left\vert u_j^{(N^j_k)}(t) - u_j(t) \right\vert \, \stackrel{k\uparrow\infty}{\longrightarrow} \, 0 \quad \mbox{ for fixed $j$.}
\]
The sequences $N_\bullet^j$ can be chosen so that $N_\bullet^{j+1}$ is a subsequence of $N_\bullet^j$ itself. By a standard diagonal argument we can extend the convergence to all $j$. Indeed, if we consider indeces $N_k := N_k^k$, we are extracting a common sub-subsequence such that
\begin{equation*}
\sup_{t \in [0,T]} \left\vert u_j^{(N_k)}(t) - u_j(t) \right\vert \, \stackrel{k\uparrow\infty}{\longrightarrow} \, 0 \qquad \mbox{ for all } j\geq0.
\end{equation*}
By taking the limit in the integral representation of the solution of \eqref{modelN}, one can see that
the uniform limit $u=(u_j)_{j\geq0}$ in $C([0,T],\mathbb{R})$ is indeed a solution of \eqref{model}.
\end{proof}
We can exploit part of the proof of Theorem~\ref{pathwise_exist} to simply obtain an energy estimate analogous to \eqref{estimateN} but valid for \emph{all} the solutions of the original system \eqref{model}.
\begin{proposition}[Energy bound]
For every $T \in [0,+\infty)$, $\sigma \in \mathbb{R}_+$, $c \in [1,3]$, $\underline{u} \in H_+$ and $w \in C([0,T],\mathbb{R})$ with $w(0)=0$, there exists $K_1=K_1(\sigma\|w\|_{\infty}, \|\underline{u}\|, T)$ that depends polynomially on $\sigma\|w\|_{\infty}, \|\underline{u}\|$ and $T$ and such that any solution $u$ of system \eqref{model} satisfies the energy estimate
\begin{equation}\label{bound_norma}
\|u(t)\|^2\leq K_1 \ \ \ \mbox{ for all } t \in [0,T].
\end{equation}
\end{proposition}
\begin{proof}
Consider a solution $u$ of system \eqref{model} and, for any $N \in \mathbb{N}$, let $u_0, u_1, \dots, u_N$ its first $N+1$ components. Notice that, by repeating the arguments as to derive \eqref{estimateN}, we get
\[
\vert u_0(t) \vert \leq a\, ,\quad \mbox{ with } \quad a := \|\underline{u}\| + 2 \sigma \|w\|_{\infty}
\]
and
\[
\hspace{1cm} \frac{d}{dt} \left[ \sum_{j=1}^N u_j^2(t) \right ] = 2 u_0^2(t) u_1(t) - 2^{c(N+1)} u_N^2(t) u_{N+1}(t) \leq 2 a^2 \sqrt{\sum_{j=1}^N u_j^2(t)},
\]
from which it follows
\begin{equation}\label{estimate}
\sum_{j=0}^N u_j^2(t) \leq \left( a^2 T + 2a \right)^2 \quad \mbox{ for every } N \geq 0.
\end{equation}
From \eqref{estimate}, by taking the limit as $N \to +\infty$, we conclude.
\end{proof}
It is worth to mention that the most part of following results heavily relies on \eqref{bound_norma}.
\subsection{Regularity}\label{subsect:regularity}
We aim at proving continuity of the solution $u$ of \eqref{model} with respect to the initial condition and to the function $w$ (see Theorem~\ref{thm:continuity} in Subsection~\ref{subsect:continuity}). To this purpose, the following regularity result will be crucial.
\begin{theorem}[Regularity]\label{pathwise_regularity}
For every $T \in [0,+\infty)$, $\sigma \in \mathbb{R}_+$, $c \in [1,3]$, $\underline{u} \in H_+$ and $w \in C([0,T],\mathbb{R})$ with $w(0)=0$, there exists $K_2=K_2(\sigma\|w\|_{\infty}, \|\underline{u}\|, T)$ that depends polynomially on $\sigma\|w\|_{\infty}, \|\underline{u}\|$ and $T$ and such that any solution $u$ of system \eqref{model} satisfies the regularity condition
\begin{equation}\label{regularity}
\int_0^Tu_{j}^2(s)ds\leq K_2 \, 2^{-\frac{2}{3}cj}\, , \ \forall j \geq 0\,.
\end{equation}
\end{theorem}
In \cite{FrG-HVi} authors establish that a statistically stationary martingale solution $\bar{u}$ of \eqref{model_stochastic} satisfies the bound $\mathbb{E} ( \bar{u}_j^2 ) \leq k \, 2^{-\frac{2}{3}cj} $, with $k$ positive constant. To prove Theorem \ref{pathwise_regularity} we adapt the method they used to derive such an estimate. Our proof relies on the idea of replacing the average by an integral over the interval $[0,T]$. This trick allows to get regularity properties of pathwise type.
Before proving Theorem \ref{pathwise_regularity}, we need the following technical lemma on real sequences.
\begin{lemma}\label{lmm:sequence}
Let $(a_n)_{n \in \mathbb{N}}$ be a sequence of positive real numbers such that $\limsup_{n\rightarrow+\infty} a_n<+\infty$. Suppose that, for some constants $\lambda > 0$, $0<C_1<1$ and $C_2 > 0$, there exists $\hat{n}$ such that, for $n \geq \hat{n}$, either
\begin{itemize}
\item[i)]
$a_n\leq C_1 a_{n+1}+C_2 2^{-\lambda n}$
\end{itemize}
or
\begin{itemize}
\item[ii)] $a_n\leq C_1 a_{n+2}+C_2 2^{-\lambda n}$
\end{itemize}
is satisfied. Then, for any $n \geq \hat{n}$, it holds
\[
a_n \leq \frac{C_2}{1-C_1} \, 2^{-\lambda n}.
\]
\end{lemma}
\begin{proof} We start by proving the assertion under hypothesis i). Suppose by contradiction that there exists $\bar{n} \geq \hat{n}$ such that $a_{\bar{n}} > \frac{C_2}{1-C_1} \, 2^{-\lambda \bar{n}}$.
Consider the inequality in i) and divide both sides by $C_1 \, a_{\bar{n}}$. We get
$$
\frac{a_{\bar{n}+1}}{a_{\bar{n}}}\geq \frac{1}{C_1}-\frac{C_2 \, 2^{-\lambda \bar{n}}}{C_1 \, a_{\bar{n}}}>\frac{1}{C_1}-\frac{1-C_1}{C_1} = 1.
$$
Since $a_{\bar{n}+1} > a_{\bar{n}}$, if we repeat the reasoning for the next ratio, we get
\[
\frac{a_{\bar{n}+2}}{a_{\bar{n}+1}} \geq \frac{1}{C_1}-\frac{C_2 \, 2^{-\lambda (\bar{n}+1)}}{C_1 \, a_{\bar{n}+1}} > \frac{1}{C_1}-\frac{C_2 \, 2^{-\lambda \bar{n}}}{C_1 \, a_{\bar{n}}} > 1.
\]
Proceeding by induction on $j$, we obtain that the sequence $(a_{\bar{n}+j})_{j \in \mathbb{N}}$ is monotonically increasing in $j$ and moreover, for all $j$, it holds
\[
\frac{a_{\bar{n}+j+1}}{a_{\bar{n}+j}} \geq \frac{1}{C_1}-\frac{C_2 \, 2^{-\lambda \bar{n}}}{C_1 \, a_{\bar{n}}} =: k > 1,
\]
where $k$ is a constant independent of $j$. Previous inequality
implies that the sequence $(a_n)_{n \in \mathbb{N}}$ diverges, which is a contradiction.\\
If we assume hypothesis ii) instead, the result follows by repeating the same argument as above for odd (or even) subsequences.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{pathwise_regularity}]
For $T \in [0,+\infty)$ and for all $j\geq1$, a solution $u$ of system \eqref{model} satisfies
\begin{equation}\label{diff:u}
u_j(T)-u_j(0)=2^{c(j-1)}\int_0^T u^2_{j-1}(s)ds-2^{cj}\int_0^T u_j(s)u_{j+1}(s)ds
\end{equation}
and
\begin{equation}\label{diff:u2}
\hspace{0.7cm} u^2_j(T)-u^2_j(0)=2\cdot2^{c(j-1)}\int_0^T u^2_{j-1}(s)u_j(s)ds-2\cdot2^{cj}\int_0^T u^2_j(s)u_{j+1}(s)ds.
\end{equation}
We want to estimate the terms $\int_0^T u_j^2(s) \, ds$ and $\int_0^T u_j^2(s)u_{j+1}(s)\, ds$ for every $j\geq0$. Starting from equation \eqref{diff:u}, we obtain
\begin{align}\label{stima:1}
\int_0^T u^2_{j-1}(s)ds & = 2^{-c(j-1)} \left[u_j(T)-u_j(0)\right] + 2^{c}\int_0^T u_j(s)u_{j+1}(s)ds \nonumber\\
& \leq k_1 \, 2^{-cj} + k_2 \left[\int_0^T u^2_j(s)u_{j+1}(s)ds\right]^{1/2}\left[\int_0^T u^2_{j+1}(s)ds\right]^{1/4},
\end{align}
where $k_1$, $k_2$ are polynomials depending on $\sigma\|w\|_{\infty}$, $\|\underline{u}\|$ and $T$. The last inequality follows from the energy bound \eqref{bound_norma} and by applying twice H\"older inequality. Now we need an estimate for $\int_0^T u_j^2(s)u_{j+1}(s) \, ds$. Summing up terms in \eqref{diff:u2} from $j=1$ to $j=N$, we obtain
\[
\hspace{1cm} 2^{cN+1} \int_0^T u^2_N(s)u_{N+1}(s) \, ds = 2 \int_0^T u_0^2(s)u_1(s) \, ds - \sum_{j=1}^N \left[u^2_j(T)-u^2_j(0)\right] \leq k,
\]
again by \eqref{bound_norma}. Therefore, for all $j\geq1$, we have
\begin{equation}\label{stima:2}
\int_0^T u^2_j(s)u_{j+1}(s) \, ds \leq k 2^{-cj},
\end{equation}
with $k$ positive and independent of $j$. By using estimate \eqref{stima:2} in \eqref{stima:1} and then applying Young inequality, we get
\begin{align*}
\int_0^T u_{j-1}^2(s)ds & \leq k_1 \, 2^{-cj} + k_2 \left[\int_0^T u^2_j(s)u_{j+1}(s)ds\right]^{1/2}\left[\int_0^T u^2_{j+1}(s)ds\right]^{1/4}\\
&\leq k \, 2^{-\frac{2}{3}cj}
+\frac{1}{4}\int_0^T u^2_{j+1}(s) \, ds.
\end{align*}
The constants $k$'s appearing in the previous calculations may change from line to line, but always keep their polynomial nature. Since $k$ does not depend on $t$ and the energy bound \eqref{bound_norma} holds, we can apply Lemma~\ref{lmm:sequence} to conclude~\eqref{regularity}.
\end{proof}
\subsection{Continuity}\label{subsect:continuity}
Next theorem is a result of continuity with respect to initial condition and noise. This will imply, on the one hand, uniqueness of solution $u$ of system~\eqref{model}; on the other, it will guarantee existence and uniqueness of strong solutions for system \eqref{model_stochastic}.\\
Before giving the result we need some more notation. For any $\alpha \in \mathbb{R}$, we denote by $H^{\alpha}$ the Sobolev-type space
$$
H^{\alpha} := \left\{ u=(u_n)_{n\in\mathbb{N}}: \left\| {u} \right\|^2_{\alpha} < \infty \right\},
$$
with norm $\| \cdot \|_{\alpha}$ given by
$
\| u \|_{\alpha}^2 :=\sum_{j=0}^{\infty}2^{2\alpha j } u_j^2 \,.
$
Notice that $H^0=\ell^2$.
\begin{theorem}[Continuity]\label{thm:continuity}
Let $c\in[1,3)$, $T \in [0,+\infty)$, $\sigma \in \mathbb{R}_+$, $\underline{u} \in H_+$ , $w \in C([0,T],\mathbb{R})$ with $w(0)=0$. There exists a function $f:\mathbb{R}_+ \longrightarrow \mathbb{R}_+$ with $\lim_{\delta\rightarrow0}f(\delta)=0$ such that, for all $\underline{\tilde{u}}$ $\in$ $H_+$ and all $\tilde{w}$ $\in$ $C([0,T],\mathbb{R})$ with $\tilde{w}(0)=0$, if
\begin{equation*}
\left\| \underline{u}-\underline{\tilde{u}} \right\| < \delta \quad \mbox{ and } \quad \left\| w-\tilde{w} \right\|_{\infty} < \delta \, ,
\end{equation*}
then
\begin{equation*}
\sup_{t\in [0,T]}\|u(t)-\tilde{u}(t)\|_{-\frac{1}{2}}<f(\delta) \, ,
\end{equation*}
where $u$ (resp. $\tilde{u}$) is a solution of system \eqref{model} with initial condition
$\underline{u}$ (resp. $\underline{\tilde{u}}$) and noise $w$ (resp. $\tilde{w}$). Moreover, if $w=\tilde{w}$ and $\underline{u}\neq\underline{\tilde{u}}$, then
\begin{equation}\label{contrazione}
\| u(t)-\tilde{u}(t) \|_{-\frac{1}{2}}<\| u(0)-\tilde{u}(0) \|_{-\frac{1}{2}}.
\end{equation}
\end{theorem}
Estimate \eqref{contrazione} is a contraction property somehow unusual for models deriving from fluid dynamics. Nevertheless, in \cite{Bor13} it is possible to find an analogous estimate for the $L^1$-norm of the viscous Burgers model. Similarly to \cite{Bor13}, we use \eqref{contrazione} to study statistically stationary distributions.
\begin{proof}
We set
\[
y_j (t) :=u_j(t) + \tilde{u}_j(t) \quad \mbox{ and } \quad z_j (t) := u_j(t) - \tilde{u}_j(t).
\]
From \eqref{model}, we obtain a system of equations for $z_j$. It is readily seen that
\begin{equation}\label{eqs:Z}
\left\{
\begin{array}{lc}
z_0(t) = \displaystyle{z_0(0)-\frac{1}{2}\int_0^t \left[z_0(s)y_1(s)+z_1(s)y_0(s) \right]ds+\sigma \left[w(t)-\tilde{w}(t) \right]} & \\
z_j(t) = \displaystyle{z_j(0)+2^{c(j-1)} \int_0^t z_{j-1}(s) y_{j-1}(s) \, ds} & \\
\phantom{z_j(t) =} \qquad \displaystyle{- \frac{2^{cj}}{2} \int_0^t \left[ z_j(s) y_{j+1}(s) + z_{j+1}(s) y_j(s) \right] ds} & \mbox{ for } j \geq 1 \\
z_j(0)=\underline{u}_j-\underline{\tilde{u}}_j & \mbox{ for } j \geq 0 \\
\end{array}
\right.
\end{equation}
We borrow a trick from \cite{BaFlMo10} and we study the $H^{-\frac{1}{2}}$-norm of $z(t)$. We aim at getting an upper bound for
\begin{align*}
\psi_n (t) &:= \sum_{j=0}^n \frac{z_j^2(t)}{2^j}\\
&=\phi_n(t)+2\sigma z_0(0)\left[w(t)-\tilde{w}(t)\right]+\sigma^2 \left[w(t)-\tilde{w}(t)\right]^2\\
&\qquad \qquad - \sigma \left[ w(t)-\tilde{w}(t)\right] \int_0^t \left[z_0(s)y_1(s)+z_1(s)y_0(s)\right]ds \,,
\end{align*}
where
\begin{align*}
\phi_n(t) &:= z^2_0(0)+\frac{1}{4} \left( \int_0^t \left[z_0(s)y_1(s)+z_1(s)y_0(s)\right] ds\right)^2 \\
&\qquad\qquad -z_0(0)\int_0^t \left[z_0(s)y_1(s)+z_1(s)y_0(s)\right] ds +\sum_{j=1}^n\frac{z_j^2(t)}{2^j} \, .
\end{align*}
First we study the quantity $\phi_n$. By using equations \eqref{eqs:Z}, we compute the derivative
\begin{align*}
\frac{d}{dt} \phi_n (t) &=\frac{1}{2} \left[z_0(t)y_1(t)+z_1(t)y_0(t)\right] \int_0^t \left[z_0(s)y_1(s)+z_1(s)y_0(s)\right]ds \\
&\qquad \qquad -z_0(0)[z_0(t)y_1(t)+z_1(t)y_0(t)] + \sum_{j=1}^n \frac{2}{2^j} \, z_j (t) \, \frac{d}{dt} z_j (t) \\
&= \left[z_0(t)y_1(t)+z_1(t)y_0(t)\right]\left\{-z_0(t)+z_0(0)+\sigma \left[w(t)-\tilde{w}(t)\right]\right\} \\
&\qquad \qquad -z_0(0)[z_0(t)y_1(t)+z_1(t)y_0(t)] + \sum_{j=1}^n \frac{2}{2^j} \, z_j (t) \, \frac{d}{dt}z_j (t) \\
&= -\sum_{j=0}^n2^{(c-1)j}z_j^2(t)y_{j+1}(t)-2^{(c-1)n}z_n(t)z_{n+1}(t)y_n(t) \\
& \qquad \qquad +\sigma \left[w(t) -\tilde{w}(t)\right] \left[z_0(t)y_1(t)+z_1(t)y_0(t)\right].
\end{align*}
Now we estimate $\frac{d}{dt}\phi_n$. We observe that the first term is negative, since $y$ $\in$ $H_+$. Then, we are left to consider
\begin{align*}
\frac{d}{dt} \phi_n (t) &\leq - 2^{(c-1)n} \, z_n (t) \, z_{n+1} (t) \, y_{n} (t) +\sigma \left[ w(t)-\tilde{w}(t) \right] \left[z_0(t)y_1(t)+z_1(t)y_0(t)\right] \\
&\leq - 2^{(c-1)n} \, z_n (t) \, z_{n+1} (t)y_n(t)+2\sigma \delta \left(\|u(t)\|^2+\|\tilde{u}(t)\|^2\right) \\
&\leq - 2^{(c-1)n} \, z_n (t) \, z_{n+1} (t)y_n(t)+\delta k \,,
\end{align*}
where constant $k$ follows from \eqref{bound_norma}. In particular, $k$ is a polynomial in $\sigma \|w\|_{\infty}$, $\sigma \|\tilde{w}\|_{\infty}$, $\| \underline{u} \|$, $\|\underline{\tilde{u}} \|$ and $T$. In the sequel, the value of constants may change from line to line, but they are always polynomial functions of those quantities. By integrating the previous inequality, we get
\begin{align*}
\phi_n (t) & \leq \phi_n(0) -2^{(c-1)n}\int_0^t [u_n(s)-\tilde{u}_n(s)][u_{n+1}(s)-\tilde{u}_{n+1}(s)][u_n(s)+\tilde{u}_n(s)]ds + \delta k T \\
& \leq 2\delta^2+ k_1 \, 2^{(c-1)n} \left(\int_0^tu_n^2(s)ds+\int_0^t\tilde{u}_n^2(s)ds\right)+\delta k T \\
& \leq 2\delta^2+k_1 \, 2^{(\frac{c}{3}-1)n} +\delta k T,
\end{align*}
where the last inequality is due to the regularity condition \eqref{regularity}. Then, for $\psi_n(t)$ it holds
\begin{equation*}
\psi_n(t)\leq k_1 \, 2^{(\frac{c}{3}-1)n} + k_2 \delta T +k_3 \delta^2.
\end{equation*}
Thus, if $c<3$ we obtain
\begin{equation*}
\|u(t)-\tilde{u}(t)\|^2_{-\frac{1}{2}}=\lim_{n\rightarrow +\infty} \psi_n(t) \leq \underbrace{k_2 \delta T +k_3 \delta^2}_{\mbox{\footnotesize $=: f(\delta)$}} \, ,
\end{equation*}
which is arbitrary small for small $\delta$, as wanted.
Now let $u$ and $\tilde{u}$ be two solutions of system \eqref{model} with the same noise $w$, but with different initial conditions $\underline{u}$ and $\underline{\tilde{u}}$, respectively. It is easy to verify that, if we perform the same computations as above in the case $w=\tilde{w}$, we get the following identity
\begin{align*}
\frac{d}{dt} \psi_n (t) &= \sum_{j=0}^n \frac{2}{2^j} \, z_j (t) \, \frac{d}{dt}z_j (t) \\
&= - \sum_{j=0}^n 2^{(c-1)j} \, z_j^2 (t) \, y_{j+1} (t) - 2^{(c-1)n} \, z_n (t) \, z_{n+1} (t) \, y_{n} (t).
\end{align*}
By integrating over the interval $[0,t]$, it yields
\begin{equation}\label{12}
\psi_n (t)-\psi_n (0)=- \! \int_0^t \! \sum_{j=0}^n 2^{(c-1)j} \, z_j^2 (s) y_{j+1} (s)ds - \! \int_0^t \! 2^{(c-1)n} \, z_n (s) z_{n+1} (s) y_{n} (s) ds.
\end{equation}
The first term on the right-hand side of \eqref{12} is strictly negative and decreasing as $n$ goes to infinity. From this and taking the limit in $n$, it immediately follows
\begin{equation*}
\| u(t)-\tilde u(t) \|_{-\frac{1}{2}}<\| u(0)-\tilde u(0) \|_{-\frac{1}{2}}.
\end{equation*}
\end{proof}
A consequence of the previous theorem is the uniqueness of the solution of~\eqref{model}.
\begin{corollary}[Uniqueness]\label{pathwise_uniqueness} Let $u$ and $\tilde{u}$ be two solutions of system~\eqref{model} with the same initial condition $\underline{u} \in H_+$ and the same noise $w \in C([0,T], \mathbb{R})$ with $w(0)=0$, then
\[
u(t)=\tilde{u}(t) \ \ \ \forall t \in [0,T].
\]
\end{corollary}
\section{Strong solutions}\label{sect:strong:sol}
In the first part of this section we prove that a strong solution for problem \eqref{model_stochastic} exists and is unique.
We then tackle the question of stationary distributions. In this respect, we establish existence of a strong statistically stationary solution in $L^2 (\Omega, H^{\alpha} \cap H_+)$, for every $\alpha < \frac{c}{3}$, and further we show uniqueness in $L^2 (\Omega, H_+)$. All the results are valid in the range of $c \in [1,3)$.\\
We start by giving the definition of strong solution.
\begin{definition}[Strong solution]
Let $\left( \Omega, \mathcal{F}, \{\mathcal{F}_t\}_{t \in [0,T]},\mathbb{P} \right)$ be a filtered probability space and $W$ a Brownian motion on $(\Omega, \mathcal{F},\mathbb{P})$, with respect to the filtration $\{\mathcal{F}_t\}_{t \in [0,T]} $. We say that $u$ is a \emph{strong solution} of system \eqref{model_stochastic} with $\mathcal{F}_0$-measurable initial condition $\underline{u}$, if $u$ is a stochastic process on $(\Omega, \mathcal{F},\mathbb{P})$ with continuous sample paths, satisfying \eqref{model_stochastic} and adapted to the filtration generated by $W$ and $\underline{u}$. \\
Moreover, this solution is \emph{unique} if, given two solutions $u^{(1)}$ and $u^{(2)}$ with the same initial condition, it holds $u^{(1)}(t)=u^{(2)}(t)$ for all $t \in [0,T]$ a.s..
\end{definition}
Existence and uniqueness of a pathwise solution, together with continuity with respect to the noise and the initial data, will allow to readily get existence and uniqueness in the strong sense. \\
By Theorem~\ref{pathwise_exist}, given an $\mathcal{F}_0$-measurable positive initial condition $\underline{u}$ and the noise, we can construct an application
\begin{equation}\label{application:F}
\begin{array}{lccc}
F_T: & \ell^2 \times C([0,T],\mathbb{R}) &\longrightarrow & C \left( [0,T],\ell^2 \right) \\
& (\underline{u}(\omega), W(\cdot,\omega)) &\longmapsto & F_T(\underline{u}(\omega), W(\cdot,\omega)),
\end{array}
\end{equation}
that associates the pair $(\underline{u}(\omega), W(\cdot,\omega))$ with the corresponding solution
of system~\eqref{model} on $[0,T]$. For every $T \in [0,+\infty)$, the function $F_T$ is well-defined thanks to Corollary~\ref{pathwise_uniqueness} and it is continuous with respect to the initial datum and the noise by Theorem~\ref{thm:continuity}.
As a consequence, the map $u(\cdot,\omega) := F_T(\underline{u}(\omega), W(\cdot,\omega))$ from $\Omega$ to $C \left( [0,T],\ell^2 \right)$ is measurable with respect to the $\sigma$-field generated by $\underline{u}$ and the Brownian motion. Concluding, $u (\cdot, \omega)$ is a strong solution of \eqref{model_stochastic}.
Such a solution is unique by Corollary \ref{pathwise_uniqueness}. Hence, we obtain:
\begin{theorem}[Existence and uniqueness]\label{thm:strong:ex}
For every $T \in [0,+\infty)$, $\sigma \in \mathbb{R}_+$, $c \in [1,3)$ and initial condition $\underline{u}$, $\mathcal{F}_0$-measurable random variable with values in $H_+$, system \eqref{model_stochastic} admits a unique strong solution $u$.
\end{theorem}
With the above results in hand, we can now turn to the analysis of strong statistically stationary solutions.
\begin{definition}[Strong stationary solution]
We say that $u^*$ is a \emph{strong stationary solution} of system \eqref{model_stochastic}, if it is a strong solution and the distribution of $u^*(t)$ does not depend on $t$, for all $t \in [0,T]$.
\end{definition}
\begin{remark}
Weak existence of a statistically stationary solution $u^*$ for system \eqref{model_stochastic} is proved in \cite{FrG-HVi}. Furthermore, by \cite[Prop. 3.1 and Thm. 4.2]{FrG-HVi}, $u^*$ can be chosen so that it belongs to $H_+$ for all times a.s..
\end{remark}
Existence and uniqueness of a strong solution imply, by Yamada and Watanabe theorem, uniqueness in law for solutions of system \eqref{model_stochastic}. This plus weak existence of a statistically stationary solution in $H_+$ (see previous Remark) gives strong existence. Moreover, from a moment estimate in \cite[Thm. 4.2]{FrG-HVi} we can infer that such a solution belongs to $L^2 (\Omega,H^{\alpha} \cap H_+)$, for every $\alpha < \frac{c}{3}$. Thus, the following theorem remains proved.
\begin{theorem}[Existence of a strong stationary solution] \label{thm:strong:stationary_ex}
For every $T \in [0,+\infty)$, $\sigma \in \mathbb{R}_+$ and $c \in [1,3)$, system \eqref{model_stochastic} admits a strong stationary solution $u^*$ such that $u^* \in L^2 \left(\Omega,H^{\alpha} \cap H_+ \right)$, for every $\alpha < \frac{c}{3}$.
\end{theorem}
To conclude our analysis, we prove that only one strong statistically stationary solution may exist.
\begin{theorem}[Uniqueness of stationary distribution] \label{thm:strong:stationary_uni}
For every $T \in [0,+\infty)$, $\sigma \in \mathbb{R}_+$ and $c \in [1,3)$, if $u_1^*, u_2^* \in L^2(\Omega, H_+)$ are strong stationary solutions of \eqref{model_stochastic}, then $u^*_1$ and $u^*_2$ have the same distribution.
\end{theorem}
The proof relies on the Kantorovich's formulation for the optimal transport problem. We will introduce a distance between probability measures as a cost function to be minimized. Then, we will use formula \eqref{contrazione} to show that the existence of two different stationary distributions would contradict the minimality achieved by the optimal transport plan.
\begin{proof}
By contradiction, let $\mu_1$ and $\mu_2$ be two different stationary distributions in $H_+$ giving rise to strong stationary solutions of \eqref{model_stochastic}. We can define the $2$-plans with marginals $\mu_1$ and $\mu_2$ as
$$
\Gamma(\mu_1,\mu_2):=\left\{\eta \, \in \, \mathcal{M}_1(\ell^2\times \ell^2) : \int_{\ell^2}\eta(x,dy)=\mu_1 \mbox{ and } \int_{\ell^2}\eta(dx,y)=\mu_2 \right\}.
$$
Moreover, for every $\eta \in \Gamma(\mu_1, \mu_2)$, we introduce the functional
$$
\varphi (\eta):= \int_{\ell^2\times \ell^2} \| x-y\|^2_{-\frac{1}{2}}\, \eta(dx,dy).
$$
By Kantorovich's result on optimal transport problem, we know there exists an element $\eta_0$ in $\Gamma(\mu_1,\mu_2)$, such that
\begin{equation}\label{Kantorovich}
\varphi (\eta_0) \leq \varphi (\eta), \ \forall \eta \in \Gamma(\mu_1,\mu_2).
\end{equation}
On $\tilde{\Omega}:=\ell^2\times \ell^2\times \Omega$, where $\Omega$ is the sample space of Brownian motion, we construct the random vector $(u^*_1(0), u^*_2(0))$ with joint law $\eta_0$, the probability measure that realizes the minimum \eqref{Kantorovich}.
Let $u^*_1$ (resp. $u^*_2$ ) be the strong stationary solution with initial condition $u^*_1(0)\sim \mu_1$ (resp. $u^*_2(0)\sim \mu_2$). After a time $t>0$, the random vector $(u^*_1(t), u^*_2(t))$ will have joint law $\eta_t$ that, by stationarity, is still belonging to $\Gamma(\mu_1,\mu_2)$. Therefore, from \eqref{Kantorovich}, we get
\begin{equation}\label{contraddizione}
\hspace{0.5cm}\begin{array}{ccccc}
\varphi (\eta_0) && \leq && \varphi (\eta_t)\\
\rotatebox{90}{$\,=$} &&&&\rotatebox{90}{$\,=$} \\
\displaystyle{\int_{\tilde{\Omega}}\|u^*_1(0)-u^*_2(0)\|^2_{-\frac{1}{2}}d\tilde{\mathbb{P}}(\tilde{\omega}) }&&& &\displaystyle{\int_{\tilde{\Omega}}\|u^*_1(t)-u^*_2(t)\|^2_{-\frac{1}{2}}d\tilde{\mathbb{P}}(\tilde{\omega})}
\end{array}
\end{equation}
with $\tilde{\mathbb{P}}$ probability measure on $\tilde{\Omega}$. On the other hand, by taking expectation on both sides of \eqref{contrazione}, it yields
\begin{equation*}
\displaystyle{\int_{\tilde{\Omega}}\|u^*_1(0)-u^*_2(0)\|^2_{-\frac{1}{2}}d\tilde{\mathbb{P}}(\tilde{\omega}) }\quad > \quad \displaystyle{\int_{\tilde{\Omega}}\|u^*_1(t)-u^*_2(t)\|^2_{-\frac{1}{2}}d\tilde{\mathbb{P}}(\tilde{\omega})},
\end{equation*}
that contradicts \eqref{contraddizione}.
\end{proof}
\section*{Acknowledgments}
The authors wish to warmly thank Markus Fischer for valuable discussions and suggestions. DB has been partially supported by the University of Padova through the Project ``Stochastic Processes and Applications to Complex Systems'' (CPDA123182). FC acknowledges financial support of FIRB research grant RBFR10N90W. MF has been partially supported by GA\v{C}R grant P201/12/2613. LP acknowledges financial support of the research project ``Singular perturbation problems for differential operators'', Progetto di Ateneo of the University of Padova.
\bibliographystyle{abbrv}
|
\section{Introduction}
The quantum computer promises an impressive speedup in certain problems such as prime factorisation \cite{Nielsen00}.
Measurement-based quantum computation (MBQC) is one approach for processing quantum information, attractive due to its relative ease of use once a suitable resource state has been prepared.
In MBQC, unitary operations are performed via pre-prepared multi-partite entangled resource states, referred to as cluster states \cite{Briegel01,Raussendorf01,Menicucci06}.
Sufficiently large cluster states are first prepared before being appropriately reshaped for any specific operations.
Arbitrary unitary operations are implemented by the precise selection of measurement bases and outcome-dependent feed-forward operations.
To date there have been several demonstrations of MBQC, predominantly in quantum optics.
Optical experiments performed in a continuous-variable (CV) setting benefit from deterministic state generation as well as deterministic implementations of Gaussian operations. The cluster states that facilitate MBQC can be generated via linear optics \cite{Su2007,Yukawa08,Yokoyama13,Moran13}. Four-mode and six-mode cluster states have been already used to implement arbitrary single-mode Gaussian gates~\cite{Ukai11}, a two-mode Gaussian gate~\cite{Ukai11QND}, and a gate sequence of these two~\cite{Su13}. Reshaping a cluster state \cite{Miwa10} is possible through quantum erasing \cite{Filip03} and wire-shortening \cite{Gu09}, which correspond to erasing and preserving the interaction gains between the nodes of the cluster state, respectively. Recently, large-scale \cite{Moran13} and ultra-large-scale \cite{Yokoyama13} cluster states have been generated by multiplexing in the frequency and time domain, respectively, both based on the same theoretical proposal \cite{Menicucci08,Menicucci11}.
Present techniques for shaping a cluster are inherently inefficient due to the lack of control over the interaction strength. For example the fixed-strength entangling gate demonstrated in Ref.~\cite{Ukai11QND} cannot have its entanglement strength tuned, and therefore it cannot completely make use of the underlying structure of the cluster state \cite{Ukai10}.
In this paper, we present a fully tuneable entangling gate for CV one-way quantum computation and experimentally demonstrate a proof-of-principle implementation. Our tuneable gate can be interpreted as a generalized instance of cluster state reshaping, which we name cluster gain tuning. Our implementation involves propagating two independent quantised optical modes (qumodes) through a three-mode linear cluster state while implementing the gate at various different strengths. The tuneable interaction gain in the resource cluster state is teleported onto the two-mode input state \cite{Bartlett03}, thus appearing at the output and becoming manifest as a certain form of entanglement.
\begin{figure}[b!]
\centering
\includegraphics[scale=1.4,clip]{ExperimentalSetup.pdf}
\caption{
(Color online)
(a) Cluster state shaping. (top) An initial three-mode linear cluster state. (bottom left) Quantum erasing. (bottom center) Interaction gain tuning. (bottom right) Wire-shortening.
Hemispheric objects next to the nodes (circles) and arrows mean measurements and feedforwards regarding to measurement outcomes, respectively.
$\hat \Pi$ is the measurement observable.
(b) Abstract illustration of our experiment. Dashed lines represent beam-splitter coupling.
(c) Schematic of our experimental setup.
OPO, optical parametric oscillator;
HD, homodyne detector;
EOM, electro-optical modulator;
$r$\%R, $r$\%-reflectivity beam-splitter;
LO, local oscillator.
}
\label{fig:ExperimentalSetup}
\end{figure}
\section{protocol for tuneable entangling gates via cluster gain tuning}
Our quantum states are represented by the quadrature operators $(\xx_j, \pp_j )$ of an electric field (annihilation) operator $\hat a_j = \xx_j + i \pp_j$, where the subscript $j$ denotes the $j$-th optical mode.
These quadrature operators play the roles of position and momentum operators of the corresponding harmonic oscillator, and hence they are canonically conjugate variables:
$[\xx_j, \pp_k]=i/2\,\delta_{jk}\ (\hbar =1/2)$, where $\delta_{jk}$ is the Kronecker delta.
A CV cluster state is defined, in the ideal case, through its zero eigenvalues for certain linear combinations of the canonical operators, so-called nullifiers,
\begin{align}
\pp_{Cj}-\sum_{k\in N_j} \xx_{Ck} \equiv \hat \delta_{j},
\end{align}
where $N_j$ refers to the nearest-neighbour nodes of node $j$ in the sense of a general graph \cite{Menicucci11Graph}.
Arbitrary bonds in CV cluster states are generated by applying controlled-phase gates $\hat C_{\mathrm{Z}jk}=\ee^{2i\xx_j \xx_k}$ on pairs of nodes, which are initialised as momentum eigenstates with zero eigenvalues in the limit of infinite squeezing \cite{Menicucci06}.
This can be understood mathematically as the transformation of nullifiers,
\begin{align}
\label{Clusternullifier}
\Big(\sum_{k\in N_j}\hat C_{\mathrm{Z}jk}\Big)\hat p_{Cj}\Big(\sum_{k\in N_j}\hat C_{\mathrm{Z}jk}^\dagger\Big)=\hat \delta_{j}.
\end{align}
The controlled-phase gates will be generalized
to include arbitrary, real gain values, $\hat C_{\mathrm{Z}jk}(g)\equiv\ee^{2ig\xx_j\xx_k}$, leading to certain weighted (real-valued) graph states, with $g=1$ as the special case of unweighted graph states.
More generally, any physical graph state can be described by complex weights and a complex adjacency matrix (including self-loops),
corresponding to a set of non-Hermitian nullifiers, where the eigenvalue (nullifier) conditions are still
exactly fulfilled even for finite squeezing \cite{Menicucci11Graph}.
However, instead of complex weights in the following,
we describe physical, finitely squeezed cluster states allowing non-zero excess noise in the Hermitian nullifier operators $\hat \delta_{j}$ \cite{Ukai11QND}.
After the preparation of a generic cluster state, the undesired bonds and nodes of the cluster can be erased by means of measurement and feed-forward, applying the quantum eraser \cite{Filip03}.
For example, the three-mode linear cluster state shown in the top of \refFig{ExperimentalSetup}(a), which is the resource state for our demonstration of the tuneable entangling gate, has bonds ($C1$--$C2$) and ($C2$--$C3$).
By measuring the position operator of node $C2$ ($\xx_{C2}$) and subtracting the measurement outcome from the momentum operators of the nearest-neighbour nodes ($C1$ and $C3$), the bonds are erased and the two modes end up in a separable state [the bottom left of \refFig{ExperimentalSetup}(a)].
On the other hand, a node can be deleted while keeping the bond up to local phase rotations, which is called wire-shortening \cite{Gu09}.
By measuring the momentum operator of node $C2$ ($\pp_{C2}$) in the same three-mode linear cluster state and subtracting the measurement outcome from the position operator of a nearest-neighbour node (either $C1$ or $C3$), the resulting two-mode state becomes an Einstein-Podolsky-Rosen (EPR) state or a two-mode cluster state up to local phase rotations [the bottom right of \refFig{ExperimentalSetup}(a)].
The two procedures described above can then be regarded as two extreme cases of the cluster shaping.
Here we generalise these procedures by considering intermediate operations between them, where we can tune the cluster gain between two cluster nodes via the measurement of the center node up to local unitaries.
We now discuss our implementation of cluster gain-tuning on a three-mode linear cluster state.
Here we consider a measurement of the observable $\xx_{C2}\cos\theta-\pp_{C2}\sin\theta$ on cluster node $C2$, where $\theta=0^\circ$ and $90^\circ$ correspond to erasing and wire-shortening, respectively. By subtracting the measurement outcome rescaled by $1/\cos\theta$ from the momentum operators of nearest-neighbour nodes ($C1$ and $C3$), the nullifiers of the resulting state become
\begin{subequations}
\begin{align}
\hat \delta_1'\equiv\pp_{C1}-(\xx_{C1}+\xx_{C3})\tan\theta\\
\mathrm{and}\quad
\hat \delta_3'\equiv\pp_{C3}-(\xx_{C1}+\xx_{C3})\tan\theta
\end{align}
\end{subequations}
[the bottom center of \refFig{ExperimentalSetup}(a)] (see Appendix B).
In analogy with Eq.\eqref{Clusternullifier}, they correspond to the transformation of nullifiers:
\begin{subequations}
\begin{align}
\label{Tznullifier1}
\hat T_{\mathrm{Z}C1C3}(\theta )\hat p_{C1} \hat T^\dagger_{\mathrm{Z}C1C3}(\theta )&=\hat \delta_1'\\
\label{Tznullifier2}
\hat T_{\mathrm{Z}C1C3}(\theta )\hat p_{C3} \hat T^\dagger_{\mathrm{Z}C1C3}(\theta )&=\hat \delta_3',
\end{align}
\end{subequations}
where the definition of
the unitary operator is
\begin{align}
\hat T_{\mathrm{Z}jk}(\theta)\equiv \ee^{i(\xx_j +\xx_k)^2\tan\theta },
\end{align}
therefore the resulting state corresponds to the application of the gate on two momentum eigenstates with zero eigenvalues.
We name this operation the {\it fully tuneable entangling gate} $T_\mathrm{Z}$, which has the tunable interaction parameter $\tan\theta$.
Since the measurement angle $\theta$ can be set arbitrarily from $-90^\circ$ to $90^\circ$, the \Tz gate can have an arbitrary real value of the interaction parameter $\tan\theta$.
The \Tz gate consists of two quadratic phase gates for individual modes ($\ee^{i\xx_j^2\tan\theta},\ee^{i\xx_k^2\tan\theta} $) \cite{Miwa09} and a controlled-phase gate ($C_{\mathrm{Z}jk}(\tan\theta)=\ee^{2i\xx_j\xx_k\tan\theta}$)
with the arbitrary interaction parameter $\tan \theta$.
The above cluster gain tuning allows for the generation of weighted gain cluster states from larger unweighted cluster states, while additional single-mode operations can be absorbed in the measurements at the latter process in order to perform larger one-way quantum computations.
The tuneable entangling gate is constructed by combining the cluster gain tuning scheme with two input states as shown in \refFig{ExperimentalSetup}(b).
Two input states in modes $\alpha$ and $\beta$ are teleported to modes $C1$ and $C3$ by half Bell measurements and cluster gain tuning, resulting in the \Tz gate operation being teleported onto the input states \cite{Bartlett03}.
In the following we describe the above procedure taking into account the excess noises $\hat \delta_j$ due to finite squeezing.
Each input mode ($\alpha$ or $\beta$) is coupled with a side mode in the cluster state via a balanced beam-splitter (50\%-BS).
Then one output arm of each of the two mixing beam-splitters as well as the centre mode in the cluster state are measured by means of homodyne detection.
The measured observables correspond to
\begin{subequations}
\begin{align}
\hat s_1&\equiv\hat x_{\alpha'}=\tfrac{1}{\sqrt{2}}(\xx_\alpha-\xx_{C1}), \\
\hat s_3&\equiv\hat x_{\beta'}=\tfrac{1}{\sqrt{2}}(\xx_\beta-\xx_{C3}), \\
\mathrm{and} \quad
\hat s_2(\theta)&\equiv \xx_{C2}\cos\theta-\pp_{C2}\sin\theta,
\end{align}
\end{subequations}
where $\theta$ is the measurement angle of the homodyne detection
on the centre mode.
We use primes to mark the modes after each beam-splitter interaction.
The quadratures of the remaining parts are
\begin{subequations}
\begin{align}
\hat x_{C1'}=\dfrac{1}{\sqrt{2}} (\hat x_\alpha +\hat x_{C1}),
\qquad
\hat p_{C1'}=\dfrac{1}{\sqrt{2}} (\hat p_\alpha +\hat p_{C1}),
\\
\hat x_{C3'}=\dfrac{1}{\sqrt{2}} (\hat x_\beta +\hat x_{C3}),
\qquad
\hat p_{C3'}=\dfrac{1}{\sqrt{2}} (\hat p_\beta +\hat p_{C3}).
\end{align}
\end{subequations}
Based on the measurement outcomes, we perform the following feed-forward operations onto the rest of the states:
\begin{subequations}
\begin{align}
&\hat X_{C1'}(\hat s_1) \hat Z_{C1'}\big((\hat s_1+\hat s_3)\tan\theta-\tfrac{\hat s_2(\theta)}{\sqrt{2}\cos\theta}\big)\\
\mathrm{and}\quad&\hat X_{C3'}(\hat s_3) \hat Z_{C3'}\big((\hat s_1+\hat s_3)\tan\theta-\tfrac{\hat s_2(\theta)}{\sqrt{2}\cos\theta}\big),
\end{align}
\end{subequations}
where $\hat X_k(s)=\ee ^{-2is\hat p_k}$ and $\hat Z_k(s)=\ee ^{2is\hat x_k}$ are the Weyl-Heisenberg position and momentum displacement operators on the state labeled by $k$, respectively.
The effects of these displacement operators correspond to additions and subtractions for quadratures [see Appendix A]:
\begin{subequations}
\begin{align}
\notag
\hat x_\mu &\equiv \hat x_{C1'}+\hat s_1 \\
&=\sqrt{2}\hat x_{\alpha}\\
\notag
\hat p_\mu &\equiv \hat p_{C1'}+(\hat s_1+\hat s_3)\tan\theta- \tfrac{\hat s_2(\theta)}{\sqrt{2}\cos\theta}\\
&=\tfrac{1}{\sqrt{2}}\Big[\hat p_\alpha +(\hat x_\alpha +\hat x_\beta)\tan\theta +\hat \delta_1+\hat \delta_2\tan\theta\Big]\\
\notag
\hat x_\nu &\equiv \hat x_{C3'}+\hat s_3 \\
&=\sqrt{2}\hat x_{\beta}\\
\notag
\hat p_\nu &\equiv \hat p_{C3'}+(\hat s_1+\hat s_3)\tan\theta- \tfrac{\hat s_2(\theta)}{\sqrt{2}\cos\theta}\\
&=\tfrac{1}{\sqrt{2}}\Big[\hat p_\beta +(\hat x_\alpha +\hat x_\beta)\tan\theta +\hat \delta_3+\hat \delta_2\tan\theta\Big],
\end{align}
\end{subequations}
where we refer to the two output modes as $\mu$ and $\nu$ in order to distinguish them from the input modes denoted by $\alpha$ and $\beta$.
Consequently, the input-output relation in the Heisenberg picture is given by
\begin{align}
\label{eq:inputoutputrelation}
\hat{\bm{\xi}}_{\mu\nu}
&=
\begin{pmatrix}
\bm{S} & \bm{0} \\
\bm{0} & \bm{S}
\end{pmatrix}
\begin{pmatrix}
\bm{I}+\bm{T}(\theta) & \bm{T}(\theta) \\
\bm{T}(\theta) & \bm{I}+\bm{T}(\theta)
\end{pmatrix}
\hat{\bm{\xi}}_{\alpha\beta}
+\hat{\bm{\delta}}
\\
&=
\Big(\hat S_\alpha \hat S_\beta \hat T_{\mathrm{Z}\alpha\beta}(\theta) \Big)^\dagger
\hat{\bm{\xi}}_{\alpha\beta}
\Big(\hat S_\alpha \hat S_\beta \hat T_{\mathrm{Z}\alpha\beta}(\theta) \Big)
+\hat{\bm{\delta}},
\end{align}
where
\begin{align}
\bm{S}=\begin{pmatrix}\sqrt{2} & 0 \\0 & 1/\sqrt{2} \end{pmatrix},
\qquad \bm{T}(\theta)=\begin{pmatrix}0 & 0 \\\tan\theta & 0 \end{pmatrix},
\end{align}
$\bm{I}$ is the $2\times 2$ identity matrix,
and
\begin{align}
\label{delta}
\hat{\bm{\delta}}=
\begin{pmatrix}
0 \\
\hat \delta _1 +\hat \delta _2 \tan\theta \\
0 \\
\delta _3 +\hat \delta _2\tan\theta
\end{pmatrix}
\end{align}
are excess noise terms for imperfect resource squeezing.
There are local squeezing operations
$\hat S_j=\ee^{-i\ln 2 (\hat x_j\hat p_j+\hat p_j \hat x_j )/2}$ in addition to the teleported \Tz gate.
These $-3.0$~dB $p$-squeezing operations are due to the input coupling with a 50\%-BS.
A half teleportation with a beam-splitter coupling corresponds to a squeezing gate \cite{Filip05,Yoshikawa07}.
Note that it can be eliminated by adding an additional coupling node at the edge of cluster states, by which full quantum teleportation with full Bell measurements is performed into the cluster state instead of half teleporation with half Bell measurement \cite{Ukai10}.
In order to verify the entangling capability of the \Tz gate, we now consider the case where both input states are coherent states.
We evaluate the entanglement with the symplectic eigenvalues $\tilde\lambda_-$ of the partially transposed covariance matrix of the output state \cite{Adesso04}.
This corresponds to the logarithmic negativity, which gives $E_N=\max[0,-\ln(4\tilde\lambda_-)]$ for the case of Gaussian states and which is an entanglement measure invariant under local unitary operations \cite{Vidal02}. The covariance matrix is given by
$V\equiv
\tfrac{1}{2}
\bigl\langle \{
\hat{\bm{\xi}}
,
\hat{\bm{\xi}}
\}
\bigr\rangle$,
where $\{\hat{\bm{u}}, \hat{\bm{v}}\}\equiv \hat{\bm{u}} \hat{\bm{v}}^T +\bigl(\hat{\bm{v}} \hat{\bm{u}}^T\bigr)^T$ \cite{Menicucci11Graph}.
For our setup the symplectic eigenvalues become
\begin{align}
\notag
\tilde\lambda_-
&=\frac{1}{4}\Big[
1+2t^2+(2+3t^2)\ee^{-2r}\\
&-\sqrt{4t^2(1+t^2+(2+3t^2)\ee^{-2r})+((1+3t^2)\ee^{-2r})^2}\Big]^{1/2},
\end{align}
where $t=\tan\theta$.
It can be calculated by means of the excess noise terms with $\hat \delta_{1}=\sqrt{2}\ee^{-r}\pp_1^{(0)}$, $\hat \delta_{2}=\sqrt{3}\ee^{-r}\pp_2^{(0)}$, and $\hat \delta_{3}=\frac{1}{\sqrt{2}}\ee^{-r}\pp_1^{(0)}+\sqrt{\frac{3}{2}}\ee^{-r}\pp_3^{(0)}$, where $\ee^{-r}\pp_j^{(0)}$ is a squeezed quadrature of the $j$-th resource squeezed-vacuum mode before the beam-splitter network.
Here we assume that all three modes have the same level of squeezing $r$ for simplicity.
The asymmetric case is easily derived in a similar manner.
The ideal (unphysical) cluster state is obtained in the limit $r\to \infty$.
The positivity under partial transposition (PPT criterion) is a necessary (and sufficient in the case of two-mode Gaussian states) measure for the separability of a state \cite{Simon00}. Thus, the output states of our setup are entangled if $\tilde\lambda_-$ is below 1/4.
Furthermore, the closer to zero $\tilde\lambda_-$ is, the stronger is the entanglement in the output states. With respect to our \Tz gate, $\tilde\lambda_-$ becomes smaller as we increase the interaction parameter $\tan\theta$.
\section{Experiment}
The schematic of our experimental setup is shown in \refFig{ExperimentalSetup}(c).
The light source is a continuous-wave Ti:sapphire laser with a wavelength of 860 nm and a power of about 1.7~W. The quantum states to be processed are qumodes at 1~MHz sidebands of the laser beam.
The resource cluster state is prepared by combining three squeezed vacuum states on two beam splitters, each generated by a subthreshold optical parametric oscillator (OPO).
We mainly employ the experimental techniques described in Ref.~\cite{Yukawa08Tele} for the feed-forward of measurement results through classical channels.
Note that the tuneable interaction parameter $t=\tan\theta$ of the \Tz gate is accessed via the relative phase $\theta$ between the signal beam and the reference local oscillator beam at the homodyne-2 detection station (HD-2). The relative phase is precisely controlled via the voltage sent to a piezoelectric transducer (PZT) attached to a mirror. Squeezing levels of the resource squeezed vacuum states are about $-4.5$~dB. The propagation losses from the OPOs to the homodyne detectors are 3\% to 9\%.
The detectors' quantum efficiencies are 99\%, and the interference visibilities are 96\% on average.
In order to evaluate our gate we measure the powers of the quadratures at the homodyne detectors with a spectrum analyzer. The measured frequency is 1~MHz with a resolution bandwidth of 30~kHz, and video bandwidth of 300~Hz. For each quadrature, 101 data points are taken with a sweep time of 0.05~s, while this is repeated 10 times for averaging. Standard errors in these averaged measurements are less than 0.06~dB. In the case of coherent state inputs, we average over even more measurements, leading to standard errors less than 0.01~dB. Note that no corrections are applied for any experimental losses.
\begin{figure}[t!]
\centering
\includegraphics[scale=0.65,clip]{Vac.pdf}
\caption{
(Color online) Output states for a \Tz gate operating with several interaction parameters $t=\tan\theta\in\{0,\, 1/5,\, 1/2,\, 1/\sqrt{2},\, 1,\, \sqrt{2},\, 2\}$, employing two vacuum inputs.
(a)--(d) Phase-space distributions. The second moments of Gaussian Wigner functions are represented by ellipses.
(a) and (b) Theoretical predictions for the ideal case with infinite resource squeezing.
(c) and (d) Experimental results computed from the measured variances of $\hat x_j$, $\hat p_j$, and $(\hat x_j\pm\hat p_j)/\sqrt{2}$, where $j\in\{\mu, \nu\}$.
(e) The measured variances of $\hat x_j$, $\hat p_j$.
The horizontal axis is the relative phase $\theta$ between the signal beam and the local oscillator beam at HD-2, which determines the interaction parameter, as $t=\tan \theta$.
The coloured lines show the theoretical predictions of $\langle\hat x^2_j\rangle$ (i), $\langle\hat p^2_j\rangle$ without squeezing (ii), $\langle\hat p^2_j\rangle$ with $-4.5$~dB resource squeezing (iii), and $\langle\hat p^2_j\rangle$ with inifinite squeezing (unphysical, ideal case) (iv).
Error bars are omitted, because they are very small compared to the scale of the vertical axis.
}
\label{fig:PhaseSpace}
\end{figure}
In \refFigs{PhaseSpace}{(a)}{(d)}, we visualise the phase-space distributions of the output Gaussian states by ellipses
for seven different interaction parameters $t=\tan\theta\in\{0,\, \tfrac{1}{5},\, \tfrac{1}{2},\, \tfrac{1}{\sqrt{2}},\, 1,\, \sqrt{2},\, 2\}$,
for vacuum state inputs. These interaction parameters correspond to the following measurement angles, $\theta\in\{0.0^\circ,\, 11.3^\circ,\, 26.6^\circ,\, 35.3^\circ,\, 45.0^\circ,\, 54.7^\circ,\, 63.4^\circ\}$.
The second moments are expressed by the size of the phase-space ellipse, which corresponds to the cross section of the quantum state's Wigner function.
Local and short radii correspond to $\sqrt{2}$ times standard deviations in the corresponding directions.
The theoretical predictions of the ideal case with infinite resource squeezing ($r\to\infty$) are shown in \refFigsAnd{PhaseSpace}{(a)}{(b)}.
Here, we see that the $\hat{x}$ quadrature amplitudes remain fixed, while the $\hat{p}$ quadrature amplitudes increase with larger interaction parameter values.
The broadening in $\hat{p}$ is due to the uncorrelated quantum fluctuations of both $\hat{x}_\alpha$ and $\hat{x}_\beta$ being added to $\hat{p}_\alpha$ and $\hat{p}_\beta$ by the interaction parameter dependent $T_\mathrm{Z}$ gate. Note that the additional local squeezing operations decrease these fluctuations.
The variances of $\hat x_\alpha$ and $\hat x_\beta$ are fixed at twice the shot noise level (SNL) from the additional local squeezing and are not dependent on the interaction parameter.
The experimental results are shown in \refFigsAnd{PhaseSpace}{(c)}{(d)}, which are
calculated from the measured variances of $\hat x_j$, $\hat p_j$, and $(\hat x_j\pm\hat
p_j)/\sqrt{2}$, where $j\in\{\mu, \nu\}$. We assume a Gaussian distribution and zero mean
value. Each of the two output modes have a nearly identical phase-space distribution with respect to
each other, indicative of the high level of symmetry in our optical mode matching. We see a slight
broadening in $\hat p$ compared to the ideal case predicted by theory, due to the finite resource
squeezing which couples in excess noise, while $\hat{x}$ remains unaffected, in accordance with $\hat{\bm{\delta}}$ in Eq.\eqref{delta}.
In order to compare them with the following results, the measured variances of $\hat x_\alpha$, $\hat x_\beta$, $\hat p_\alpha$, and $\hat p_\beta$ are plotted in \refFig{PhaseSpace}(e).
The horizontal axis is the relative phase $\theta$ between the signal beam and the local oscillator beam at HD-2.
The variances of $\hat x$ are 3.0~dB above the SNL independent of the resource squeezing level $r$ and the interaction parameter $\tan\theta$ as expected from the theory expressed by the blue line (i),while $\hat p$ depends on them.
The green line (ii) represents the theoretical predictions for zero resource squeezing, while the orange line (iv) represents infinite squeezing. Finite squeezing values appear between these two extremes, and we find our experimental results are close to the theoretical prediction of $-4.5$~dB resource squeezing, as indicated by the red line (iii). These results indicate a good qualitative agreement with the theoretical predictions.
\begin{figure}
\centering
\includegraphics[scale=0.65,clip]{Coherent.pdf}
\caption{
(Color online) The powers at the outputs from two coherent inputs with several interaction parameters.
The horizontal axes are the relative phases $\theta$ between the signal beam and the local oscillator beam at HD-2, which are related to the interaction parameters $\tan \theta$.
(a)--(d) $(\langle\hat x_\alpha\rangle,\langle\hat p_\alpha\rangle,\langle\hat x_\beta\rangle,\langle\hat p_\beta\rangle)$ are $(a,0,0,0)$, $(0,a,0,0)$, $(0,0,b,0)$, and $(0,0,0,b)$,
where $a^2$ and $b^2$ correspond to 13.8~dB and 16.9~dB above the shot noise level, respectively.
(i), (ii), (iii), and (iv) show the theoretical predictions of $\langle\hat x^2_\mu\rangle$, $\langle\hat p^2_\mu\rangle$, $\langle\hat x^2_\nu\rangle$, and $\langle\hat p^2_\nu\rangle$ with $-4.5$~dB resource squeezing, respectively;
coh., coherent state.
Error bars are omitted, because they are very small compared to the scale of vertical axis.
}
\label{fig:coherent}
\end{figure}
Next, we replace one of the input vacuum states by coherent states, allowing us to verify the input-output relationship based on the assumption that the gate has a linear response.
The powers of the input amplitude quadratures are individually measured in advance, corresponding to 13.8~dB for mode $\alpha$ and 16.9~dB for mode $\beta$, respectively, compared to the SNL.
In analogy with \refFig{PhaseSpace}, \refFigure{coherent}(a) shows the powers of the output quadratures for an input coherent state $\alpha$ and an input vacuum state $\beta$.
The output quadrature powers are shown as a function of the relative phase $\theta$ between the signal beam and the local oscillator beam at HD-2, which determines the interaction parameter $\tan\theta$. Theoretical predictions are shown as lines and experimental data as markers. The predictions are calculated from the measured input coherent amplitude with a resource squeezing level of $-4.5$~dB.
We observe fixed power increases in $\hat{x}$ and $\theta$-dependent increases in $\hat{p}$. The power of $\hat{x}_\mu$ increases by 3.0~dB above the inital 13.8~dB (corresponding to about 17~dB above the SNL, blue markers), which is due to the additional local squeezing operation. The power of $\xx_\nu$ is the same as the case of two vacuum inputs (corresponding to 3.0~dB above the SNL, cyan markers).
$\pp_\mu$ and $\pp_\nu$ experience larger increases in power relative to the case of vacuum inputs in \refFig{PhaseSpace}(e), due to the increasing contribution of the nonzero coherent amplitude of $\xx_\alpha$ via the \Tz gate.
Similarly, \refFigures{coherent}{(b)}{(d)} show the results for a nonzero coherent amplitude in the $\hat{p}_\alpha$, $\hat{x}_\beta$, and $\hat{p}_\beta$ input quadratures, respectively. The \Tz gate behaves as predicted, with the sum of $\hat{x}_\alpha$ and $\hat{x}_\beta$ appropriately appearing in both $\hat{p}_\mu$ and $\hat{p}_\nu$ quadratures, as a function of the interaction parameters.
The small discrepancies between our experimental results and the theoretical predictions are caused by the (slightly unbalanced) propagation losses and non-unity homodyne detections.
\begin{figure}[t!]
\centering
\includegraphics[scale=0.85,clip]{PTSymplecticEigenvalue.pdf}
\caption{
(Color online)
The dependence of the entanglement at the outputs on the interaction parameters of the \Tz gate.
The horizontal axis corresponds to the relative phase $\theta$ between the signal beam and the local oscillator beam at HD-2, which determines the interaction parameter $\tan \theta$.
The vertical axis corresponds to the symplectic eigenvalues of the partially transposed covariance matrix of the output state, connected to a measure of entanglement (see text).
(i) without squeezing, (ii) with $-4.5$~dB resource squeezing, (iii) with infinite squeezing (unphysical, ideal case) and (iv) quantum boundary; values below satisfy a sufficient condition for entanglement.
Error bars show standard errors.
}
\label{fig:entanglement}
\end{figure}
Finally, the entanglement strength is quantified in \refFig{entanglement}. Shown there is the set of symplectic eigenvalues $\tilde{\lambda}_-$ of the partially transposed covariance matrices corresponding to the output states. These are calculated from the variances of the output quadratures for vacuum inputs (see Ref.~\cite{Yokoyama14} for details), and are displayed as a function of \Tz interaction parameter (as determined by the relative phase of homodyne detection).
Note that the results of covariance matrices satisfy the physicality condition $V+(i/4)\ \Omega \ge 0$, where
$\Omega $ is a direct sum of
$\left(\begin{smallmatrix}
0 & -1 \\
1 & 0
\end{smallmatrix}
\right)$ \cite{Simon94,Pirandola09}.
The theoretical predictions for the experiment with and without resource squeezing are represented by the theoretical curves (ii) and (i), respectively.
We observe the remarkable feature of an enhancement in entanglement strength dependent on the interaction parameter. The entangling criterion is satisfied for parameter values of $\tan\theta=\tfrac{1}{2}$, $\frac{1}{\sqrt{2}}$, 1, $\sqrt{2}$, and 2 when the resource state is squeezed. Conversely, without squeezing the symplectic eigenvalues never cross the quantum boundary for any value of interaction parameter.
\section{conclusion}
In conclusion, we have proposed and experimentally demonstrated a fully tuneable \Tz gate for continuous-variable one-way quantum computation. Our proof-of-principle demonstration employed a three-mode linear cluster state as a resource for implementing a new cluster gain tuning protocol.
The capability of the gate to produce entanglement at the output is verified via the symplectic eigenvalues of the partially transposed covariance matrix of the output for the case of two coherent input states.
The interaction parameter at the gate and accordingly the entanglement strength in the output state are accurately tuned by a corresponding tuning of the set of measurement bases. Since our gate can be directly incorporated into large-scale one-way quantum computation schemes, it may facilitate efficient implementations of MBQC with cluster states.
\section*{ACKNOWLEDGMENTS}
This work was partly supported by PDIS, GIA, APSA, and FIRST initiated by CSTP,
ASCR-JSPS, and SCOPE program of the MIC of Japan.
S.Y.\ acknowledges support from ALPS\null.
R.U.\ acknowledges support from JSPS\null.
S.A.\ acknowledges support from the Prime Minister's Award.
|
\section{Introduction}
In the last fifteen years a groundbreaking discovery has altered the way we view our universe; namely observations show that the universe is not only expanding but it is also accelerating \cite{Perlmutter:1998,Riess:1998,Riess:1998dv}. This fact gave birth to a plethora of propositions for explaining it; only to list a few: quintessence models \cite{Ratra:1987rm,Tsujikawa:2013fta}, which invoke an evolving canonical scalar field with a potential; Chameleon fields in which the scalar field couples to the baryon energy density and is homogeneous \cite{Khoury:2003rn,Lombriser:2014dua}; a scalar field with
a non-canonical kinetic term, known as K-essence \cite{Chiba:1999ka,ArmendarizPicon:2000ah} based on earlier work of
K-inflation \cite{ArmendarizPicon:1999rj}; Chaplygin gases, which attempt to unify dark energy and dark matter under one roof by allowing for a fluid with an equation of state which evolves between the two \cite{Kamenshchik:2001cp,Bilic:2001cg,Bento:2002ps}; phantom dark energy \cite{Caldwell:2003vq} or even direct anthropic arguments \cite{Linde:1984ir,Weinberg:1988cp,Efstathiou:1990xe}; for a comprehensive review see \cite{Sahni:2004ai,Copeland:2006wr}.
Another big field of research is devoted to modified theories of gravity and specifically to the scalar--tensor case; which is the subject of the current work. Scalar--tensor theories, with a non minimal coupling, are considered as the most general, since they incorporate a major amount of other theories. It is well known that $f(R)$ gravity theories are equivalent to many scalar--tensor cases, with the derivative of the function $f(R)$ playing the role of the Brans--Dicke scalar \cite{Higgs1959,Chiba2003,Teyssandier:1983zz,Wands1994}; fourth-order gravity theory \cite{Higgs1959,Whitt1984} are also equivalent to a scalar tensor theory and there is even a big analogy among the $f (R)$--gravity with torsion and scalar--tensor theories with torsion, as discussed, for example, in \cite{German1985,Sung-Won1986} (for a review of all of them see \cite{Capozziello:2011et}).
The use of Noether symmetries in minisuperspace, either in classical or in quantum level, is not new. This approach for classical Bianchi cosmologies has been to the best of our knowledge, initiated in \cite{Capozziello:1996ay} and then used in \cite{Capozziello:1999xr,Cotsakis:1998zk}; while work on the subject has been revived from numerous authors \cite{tsamp1,Basilakos:2011rx,Vakili:2011uz,PintoNeto:2012ug,Capozziello:2012hm,Sarkar:2012sx}.
The common feature of all the above works is that they were dealing with systems described by singular Lagrangians, since all of them admit a time reparametrization invariance. In \cite{CDT2014} the symmetry treatment of such Lagrangians was addressed and it was shown how one can find all the Noether symmetries possessed by these systems. The result is that we have to extend the infinitesimal criterion of symmetry in such a way that it includes the constraint that arises from the reparametrization invariance. This method was used in \cite{Christodoulakis:2001um,tchris_sch,Christodoulakis:2013sya,Christodoulakis:2014wba} for the quantization of various minisuperspace models and in \cite{Dimakis:2013oza} where a Noether analysis of FRLW cosmology in the context of $f(R)$--gravity was performed, resulting in the discovery of several exact new solutions.
In the present work we use the method developed in \cite{CDT2014}, to
investigate a general non--minimal coupling for a scalar field $\phi$ with gravity, which is proportional to the Ricci scalar $R$, see \eqref{action} below, embedded in a Friedmann--Lema\^itre--Robertson--Walker (FRLW) spacetime. The strategy we follow is to demand a maximal number of Noether symmetries of the action \eqref{action} in order to find the general solution for the scale factor $a(t)$, the coupling function $F(\phi)$, the potential $V(\phi)$ and the scalar field $\phi(t)$.
In order to infer the physical properties of the solutions we obtain, we start from the known duality between scalar fields and perfect fluids \cite{Madsen1, Madsen2, Pimentel}. The usual line of thought is to try to interpret the energy momentum tensor of the scalar field as an energy momentum tensor of a perfect fluid \cite{Unnikrishnan:2010ag,Christopherson:2008ry,DiezTejedor:2005fz, Arroja:2010wy}; of course this duality must be taken with caution e.g. at the level of the Lagrangian formulation problems may arise as recently noted \cite{Faraoni:2012hn}.
We, on the other hand, choose to make a slightly different identification; we rewrite the field equations of the scalar--tensor theory, as in General Relativity, i.e. $G_{ij}=T_{ij}$ and interpret the right hand side as the energy momentum tensor of an imperfect fluid. The nice outcome is that in the general case the imperfect fluid is actually a perfect one. In order to pick up physically acceptable perfect fluid solutions, one must demand a sort of energy conditions.
The structure of the paper is the following. In Sec. 2, we set up the field equations, perform the Noether analysis and calculate the general solutions. In Sec. 3, we calculate the parameters that characterize the universe expansion (Hubble, deceleration and jerk) and establish the correspondence between a scalar field and a perfect fluid. In Sec. 4, we present a number of special solutions, among them one that obeys the major energy conditions and describes an expanding universe suffering from a cosmic jerk (a deceleration epoch followed by an accelerating one). Finally Sec. 5 is devoted to discussion.
\section{Noether analysis and general solutions}
\subsection{Background geometry and minisuperspace}
Let us consider a FLRW space--time, which describes a homogeneous and spatially flat universe, i.e.
\bal\label{FLRW_metric}
\textrm{d} s^2=-N(u)^2 \textrm{d} u^2+a(u)^2\left( \textrm{d} r^2+r^2\, \textrm{d} \Omega^2 \right) ,
\eal
where $ \textrm{d} \Omega^2=\textrm{d}\theta^2+\sin^2\theta\, \textrm{d}\varphi^2$ and $N(u)$ is the lapse function which will play an essential role in the development of our treatment of the problem.
The action that describes the non--minimal coupling between gravity and the scalar field $\phi$ is taken as
\bal\label{action}
S=\int\!\sqrt{-g}\left( F(\phi)R+\frac{\epsilon}{2}g_{mn}\phi^{;m}\phi^{;n}-V(\phi) \right) \textrm{d}^4 x,
\eal
where $F(\phi)$ a function that represents the coupling, the constant $\epsilon$ equals $\epsilon=\pm 1$ allowing ghost fields ($\epsilon=+1$), $V(\phi)$ is a self--interaction potential and $R$ is the Ricci scalar.
In order to find the field equations along with the equation that the scalar field $\phi$ obeys, we must vary the action \eqref{action} with respect to $g_{ij}$ and $\phi$ respectively. The result is
\begin{subequations}\label{EoM}
\bal
F(\phi)\left(R_{ij}-\frac{1}{2}Rg_{ij} \right)-\nabla_i\nabla_j F(\phi)+g_{ij}\Box F(\phi)&=T_{ij}, \label{EoM1}\\
\Box\phi+\epsilon V'(\phi)&=\epsilon F'(\phi)R \label{EoM2},
\eal
\end{subequations}
where
\bal\label{Tij}
T_{ij}=-\frac{\epsilon}{2}\nabla_i\phi \nabla_j\phi+\frac{1}{4}g_{ij}\left(\epsilon \nabla^k\phi\nabla_k\phi-2V(\phi) \right),
\eal
is the energy--monentum tensor, $\Box=\nabla_k\nabla^k$ is the Laplace--Beltrami operator, $\nabla_k$ is the covariant derivative and the prime $'$ on a function, denotes the derivative with respect to its argument.
The Lagrangian treatment of the problem begins by inserting the values of $g_{ij}$ from \eqref{FLRW_metric} in \eqref{action}. The resulting Lagrangian is given by
\bal\label{Lag}
L=\frac{1}{2N}\,G_{\alpha\beta} (x^\alpha) x'^\alpha\, x'^\beta-N\,U(x^\alpha), \quad G_{\alpha\beta} =
\begin{pmatrix}
-12a F & -6 a^2 F_\phi\\
-6 a^2 F_\phi & -\epsilon a^3
\end{pmatrix},
\quad U=a^3 V,
\eal
where $x^\alpha=(a,\phi)$\footnote{Greek indices take the values $1,2$ while the range of the English ones is $1\dots 4$}, the subscript $\phi$ indicates the derivative with respect to $\phi$ and $G_{\alpha\beta}$ is the mini--supermetric of our problem. It is an essential requirement in differential geometry to check, that the field equations \eqref{EoM} and the Euler--Lagrange equations yielding from the reduced Lagrangian \eqref{Lag} are equivalent; something that is true in our case.
\subsection{Noether Symmetries}
One way to solve the equations of motion \eqref{EoM} resulting from \eqref{Lag} is to search for the Noether symmetries that the system possesses. The significant feature of this Lagrangian is that it is singular, since there is no $N'$ term; thus, in order to find its Noether symmetries we must take this fact into consideration.
The correct way of treating this sort of Lagrangians, in order to acquire all their Noether symmetries, was exhibited in \cite{CDT2014}; the result is that the Noether symmetries correspond to the conformal Killing fields of both $G_{\alpha\beta}$ and $U(x^\alpha)$ with opposite conformal factors, i.e.
\bal
\emph{\textsterling}_\xi G_{\alpha\beta}=\omega(x^\alpha)\,G_{\alpha\beta}, \quad \emph{\textsterling}_\xi U =-\omega (x^\alpha) \,U(x^\alpha)
\eal
The freedom of time re--parametrization, allows us to redefine the lapse function $N(t)$ in such a way so that the potential $U(x^\alpha)$ becomes constant; the recipe is to define a new lapse $\xbar{N}= N/U(x^\alpha)$, which in turn scales the mini--supermetric to $\xbar{G}_{\alpha\beta}= U(x^\alpha)G_{\alpha\beta}$. In this parametrization the symmetries of \eqref{Lag} corresponding to integrals of motion are constructed by all the Killing fields of the scaled supermetric $\xbar{G}_{\alpha\beta}= U(x^\alpha)G_{\alpha\beta}$; Additionally, its homothetic field (which is a Lie-point symmetry of the equations of motion) can be used to define a rheonomous integral of motion, the details are explained in \cite{CDT2014}.
The scaled mini--supermetric $G_{\alpha\beta}$ reads (we drop the bars hereafter)
\bal\label{scGdd}
G_{\alpha\beta} =a^3\,V
\begin{pmatrix}
-12a F & -6 a^2 F_\phi\\
-6 a^2 F_\phi & -\epsilon a^3
\end{pmatrix},
\eal
while the corresponding Ricci scalar is proportional to
\bal\label{Ricci}
-2 F V_\phi^2\left(\epsilon F-3 F_\phi^2\right)+\nonumber\\
V \left(-6 F_\phi^3 V_\phi+F F_\phi \left(V_\phi \left(6 F_{\phi\phi}+\epsilon \right)-6 F_\phi V_{\phi\phi}\right)+2 \epsilon F^2 V_{\phi\phi}\right)+\nonumber\\
2 \epsilon V^2 \left(F_\phi^2-2 F F_{\phi\phi}\right).
\eal
\subsubsection{Flat minisuperspace}
The proportionality factor of \eqref{Ricci} is a particular function of $a$. Thus, if one wants to have the maximum number of Noether symmetries, the only viable case is for the Ricci scalar to be zero, since it can not be a non zero constant. Therefore, one is led to the nihilism of the above expression, which can be achieved if $F(\phi),V(\phi)$ are assumed to satisfy
\bal\label{FVpar}
F(\phi)=\frac{1}{4}h^2(\phi),\, V(\phi)=e^{f(\phi)}h^4(\phi), \quad \text{where} \quad 3h'^2-\lambda^2h^2 f'^2=\epsilon.
\eal
The functions $f(\phi),h(\phi)$ are arbitrary and $\lambda$ is a constant.
In order to calculate the form of the Killing fields $\xi^\alpha$ and the homothetic field $\eta^\alpha$ of the scaled mini--supermetric we bring it to a diagonal form
\bal
G_{\alpha\beta}=64\exp\left( 2\sqrt{3}w+f(\phi) \right)
\begin{pmatrix}
-1 & 0\\
0 & \lambda\,f'^2(\phi)
\end{pmatrix}
\eal
with the aid of the transformation
\bal\label{trans}
a=\frac{1}{\sqrt{F(\phi)}}e^{w/\sqrt{3}}.
\eal
The resulting fields are
\begin{subequations}
\bal
\xi_{(1)}&=-\frac{1}{2}\exp \frac{ \left(2\sqrt{3\lambda}-1\right) \left( -w+\sqrt{\lambda}f(\phi) \right) }{2\sqrt{\lambda}}\left( \partial_w-\frac{1}{\sqrt{\lambda} f'(\phi)}\partial_\phi \right)\\
\xi_{(2)}&=\frac{1}{2}\exp \frac{ -\left(2\sqrt{3\lambda}+1\right) \left( w+\sqrt{\lambda}f(\phi) \right) }{2\sqrt{\lambda}}\left( \partial_w+\frac{1}{\sqrt{\lambda} f'(\phi)}\partial_\phi \right)\\
\xi_{(3)}&=-\frac{1}{2}\partial_w+\frac{\sqrt{3}}{f'(\phi)}\partial_\phi\\
\eta&=\frac{1}{2\sqrt{3}}\partial_w.
\eal
\end{subequations}
From the above fields we can form the constants of motion $Q_I=\xi_{(I)}^\alpha\pi_\alpha$, where $\pi_\alpha=\partial_{x'^\alpha}L$ are the momenta, along with the constant $Q_\eta=\eta^\alpha\pi_\alpha+\int N du$ and calculate the functions $w(u),f(u)$. In order to simplify the results we can switch to the time variable $\tau$ with $\textrm{d} \tau=N(u)\,\textrm{d} u$. Denoting with $\kappa_I$ the three constants of motion which correspond to the Killing fields and with $k_h$ the constant arising from the homothetic filed, we have
\begin{subequations}\label{eqk}
\bal
32\exp \frac{\left(1+2\sqrt{3\lambda}\right)\left( \sqrt{\lambda}f(\tau)+w(\tau) \right)}{2\sqrt{\lambda}} \left( \sqrt{\lambda}f'(\tau)+w'(\tau) \right)&=\kappa_1\\
32\exp \frac{\left(1-2\sqrt{3\lambda}\right)\left( \sqrt{\lambda}f(\tau)-w(\tau) \right)}{2\sqrt{\lambda}} \left( \sqrt{\lambda}f'(\tau)-w'(\tau) \right)&=\kappa_2\\
32\exp\left( f(\tau)+2\sqrt{3}w(\tau) \right) \left( 2\sqrt{3}\lambda f'(\tau)+w'(\tau) \right)&=\kappa_3\\
\frac{32}{\sqrt{3}}\exp\left( f(\tau)+2\sqrt{3}w(\tau) \right)w'(\tau)&=\tau-k_h,
\eal
\end{subequations}
The above four equations can be solved \emph{algebraically} for the functions $f(\tau),w(\tau)$ and their derivatives $f'(\tau),w'(\tau)$; but after that, we must demand validity of the consistency equations $f'(\tau)=\textrm{d} f(\tau)/\textrm{d} \tau, w'(\tau)=\textrm{d} w(\tau)/\textrm{d} \tau$.
From the form of the equations \eqref{eqk}, it is obvious that we have to consider two cases, where the constant $\lambda$ equals $\frac{1}{12}$ or not.
$\bullet$ {\bf Case I: $\lambda=\dfrac{1}{12}$}.
The consistency equations imply the following relations between the constants $\kappa_i$
\bal
\kappa_3=\kappa_1,\, \kappa_2=-\frac{32}{\kappa_1},
\eal
so the functions $f(\tau),\,w(\tau)$ are given by
\begin{subequations}
\bal
f(\tau)&=c_1-\frac{\sqrt{3}}{\kappa_1}\,\tau+\frac{1}{2}\ln\left( 2\sqrt{3}\tau-k \right)\\
w(\tau)&=-\frac{\sqrt{3}}{6}\left( c_1 +\ln\frac{32}{\kappa_1} \right)+\frac{1}{2\kappa_1}\,\tau+\frac{\sqrt{3}}{12}\ln\left( 2\sqrt{3}\tau-k \right),
\eal
\end{subequations}
where $k=2\sqrt{3}c_h+\kappa_1$. The values of the original functions $a(\tau), V(\tau), F(\tau)$ and $\phi(\tau)$ can be deduced from the parametrization \eqref{FVpar} and \eqref{trans}, i.e.
\begin{subequations}\label{fcaseI}
\bal
a(\tau)&=\frac{(2\kappa_1)^{1/6}}{h(\tau)}\left( 2\sqrt{3}\tau-k\right)^{1/12}\exp\left(\frac{\sqrt{3}}{6\kappa_1}\tau-\frac{c_1}{6}\right) \\
V(\tau)&=\left( 2\sqrt{3}\tau-k\right)^{1/2}h^4(\tau)\exp\left( c_1-\frac{\sqrt{3}}{\kappa_1}\,\tau\right) \\ \label{fcaseIc}
F(\tau)&=\frac{1}{4}h^2(\tau) \\ \label{fcaseId}
\phi'(\tau)^2&=\frac{3}{\epsilon}\,h'^2(\tau)- \frac{1}{\epsilon} \left(\frac{2\sqrt{3}\tau-k-\kappa_1}{2\kappa_1\left( 2\sqrt{3}\tau-k \right)} \right)^2 h^2(\tau).
\eal
\end{subequations}
Thus we have an \emph{infinite} number of coupling functions $F(\phi)$ and interacting potentials $V(\phi)$ resulting from the \emph{infinite} choices of the arbitrary function $h(\tau)$.
If the actual form $F(\phi)$ is needed, it can be derived as follows: choose a function $h(\tau)$, calculate the functional form of $\phi(\tau)$ from \eqref{fcaseId}, take the inverse of that function in order to get $\tau=r(\phi)$ and then substitute the result in \eqref{fcaseId}.
As an example let $\epsilon=1,k=0,\kappa_1=2\sqrt{3},\, h(\tau)=\sqrt{48/143}\,e^\tau/\tau$, then $\phi(\tau)=c\pm e^\tau/\tau$ (where the plus sign emerges when $\tau>1$ while the minus sign when $\tau<1$), then $h=\pm\sqrt{48/143}\left( \phi -c\right)$ and finally $F(\phi)=12/143(\phi-c)^2$ along with $V(\phi)=c_2\left( \phi - c\right)^{7/2}$.
As it is common in General Relativity, the constants that are appearing in the solution set, are not all essential, i.e. they can be eliminated by a proper redefinition of them, along with a coordinate transformation. In our case the redefinitions $k=2\sqrt{3}\gamma,\,k_1=1/(\sqrt{3}\alpha),\, \exp c_1=\alpha/(4\sqrt{2\sqrt{3}}\beta)$ and the transformation $r \mapsto e^{-c_1/3}\alpha^{2/3}/(2\sqrt{2}3^{1/12})\, r$, bring the solution space into the form
\begin{subequations}\label{final_caseI}
\bal
a(\tau)&=\frac{\beta}{h(\tau)}\, e^{\alpha\tau/2}\left( \tau-\gamma \right)^{1/12} \label{a_I}\\
V(\tau)&=\frac{\alpha\,h^4(\tau)}{4\beta^2}\, e^{-3\alpha\tau}\sqrt{\tau-\gamma} \label{V_I}\\
F(\tau)&=\frac{1}{4}h^2(\tau) \label{F_I}\\
\phi'(\tau)^2&=\frac{3}{\epsilon}\,h'^2(\tau)- \frac{1}{48\epsilon} \left(\frac{6\alpha\tau-6\alpha\gamma-1}{\tau-\gamma} \right)^2 h^2(\tau), \label{psi_I}
\eal
\end{subequations}
yielding the line element
\bal\label{ds_caseI}
\textrm{d} s^2=\frac{\beta^2\, e^{3\,\alpha\,\tau}}{h^2(\tau)\sqrt{\tau-\gamma}} \left( -\textrm{d} \tau^2+e^{-2\,\alpha\,\tau}\left( \tau-\gamma \right)^{5/3} \left( \textrm{d} r^2 + r^2\textrm{d} \theta^2+r^2 \sin^2\theta\, \textrm{d} \varphi^2 \right) \right),
\eal
with $\tau>\gamma$.
$\bullet$ {\bf Case II: $\lambda \neq \dfrac{1}{12}$}.
In this case the consistency equations imply only one relation for the constants $\kappa_i$
\bal
\kappa_2=-\frac{32}{\kappa_1},
\eal
yielding the functions $f(\tau),\,w(\tau)$
\begin{subequations}
\bal
f(\tau)&=\frac{1}{s+1}\ln\frac{3k_1(s+1)(\tau-\beta)}{32s}-\frac{1}{s-1}\ln\frac{(s-1)(\tau+\alpha)}{k_1s}\\
w(\tau)&=\frac{s}{2\sqrt{3}}\left( \frac{1}{s+1}\ln\frac{3k_1(s+1)(\tau-\beta)}{32s}+\frac{1}{s-1}\ln\frac{(s-1)(\tau+\alpha)}{k_1s} \right),
\eal
\end{subequations}
where the redefinitions of the various constants are $\lambda=s^2/12,\,\kappa_1=\sqrt{3}\,k_1$ and $c_h=(\alpha(1-s)+\beta(1+s))/(2s), \kappa_3=\sqrt{3}(s^2-1)(\alpha+\beta)/(2s)$. Once more the values of the original functions $a(\tau), V(\tau), F(\tau)$ and $\phi(\tau)$ can be deduced from the parametrization \eqref{FVpar}, i.e.
\begin{subequations}\label{fcaseII}
\bal
a(\tau)&=\frac{2}{h(\tau)} \left( \frac{3k_1(s+1)(\tau-\beta)}{32s} \right)^{s/6(1+s)} \left( \frac{(s-1)(\tau+\alpha)}{k_1s} \right)^{s/6(1-s)}\\
V(\tau)&= \left( \frac{3k_1(s+1)(\tau-\beta)}{32s} \right)^{1+s} \left( \frac{(s-1)(\tau+\alpha)}{k_1s} \right)^{1-s}h^4(\tau)\\
F(\tau)&=\frac{1}{4}h^2(\tau) \\
\phi'(\tau)^2&=\frac{3}{\epsilon}\,h'^2(\tau)- \frac{s^2\left(2\tau-(s-1)\alpha-(s+1)\beta \right)^2}{12\epsilon (s^2-1)^2(\tau-\beta)^2(\tau+\alpha)^2}h^2(\tau).
\eal
\end{subequations}
Exactly as in case I, we have an \emph{infinite} number of coupling functions $F(\phi)$ and interacting potentials $V(\phi)$ arising from the appearance of the arbitrary function $h(\tau)$.
The following redefinition and the transformation of $r$--coordinate
\bal
k_1&=2^{-(s-1)(s-4)/(2s)}3^{-(s+2)(s-1)/(2s)}\nonumber\\
&(s-1)^{-(s+1)(s-2)/(2s)}(s+1)^{-(s-1)(s+2)/(2s)}\gamma^{-(s^2-1)/s}\nonumber\\
r &\mapsto 2^{(3-2s)/(3+3s)}3^{-(3+2s)/(3+3s)}k_1^{(2s)/(3-3s^2)}\nonumber\\
&(s-1)^{(3-2s)/(-3+3s)}(s+1)^{-(3+2s)/(3+3s)}s^{(6-4s^2)/(3-3s^2)}\,r\nonumber
\eal
considerably simplifies the form of the solution space
\begin{subequations}\label{final_caseII}
\bal
a(\tau)&=\frac{\gamma}{h(\tau)}\left( \tau +\alpha\right)^{s/6(s-1)} \left( \tau -\beta\right)^{s/6(s+1)} \label{a_II}\\
V(\tau)&=\frac{s^2\,h^4(\tau)}{6\gamma^2\left(s^2-1\right)}\left( \tau +\alpha\right)^{1/(1-s)} \left( \tau -\beta\right)^{1/(s+1)} \label{V_II}\\
F(\tau)&=\frac{1}{4}h^2(\tau) \label{F_II}\\
\phi'(\tau)^2&=\frac{3}{\epsilon}\,h'^2(\tau)- \frac{s^2\left(2\tau-(s-1)\alpha-(s+1)\beta \right)^2}{12\epsilon (s^2-1)^2(\tau-\beta)^2(\tau+\alpha)^2}h^2(\tau). \label{psi_II}
\eal
\end{subequations}
yielding the line element
\bal\label{ds_caseII}
\textrm{d} s^2=\frac{\gamma^2\left( \tau +\alpha\right)^{(2-s)/(s-1)}}{h^2(\tau)\left( \tau -\beta\right)^{(s+2)/(s+1)}} \left( -\textrm{d} \tau^2+\frac{\left( \tau +\alpha\right)^{n}}{\left( \tau -\beta\right)^{m}} \left( \textrm{d} r^2 + r^2\textrm{d} \theta^2+r^2 \sin^2\theta\, \textrm{d} \varphi^2 \right) \right),
\eal
where $n=\dfrac{4s-6}{3s-3},\,m=-\dfrac{4s+6}{3s+3}$.
\subsubsection{Mini-superspace with lesser autonomous integrals of motion}
In this subsection we investigate what the result of the previous investigation would be if the assumption of maximal symmetry for $G_{\alpha\beta}$ was relaxed, i.e. if we demanded less than three autonomous integrals of motion. As it is well known, in two dimensions the general metric can be brought in a conformally flat form. We thus, need to investigate the case where the conformal factor is such that the mini--supermetric \eqref{scGdd} is not flat. In order to find its Killing/homothetic fields we first begin by enumerating all the possibilities. The maximum number of Killing fields for an $n$--dimensional metric is $n(n+1)/2$ thus in our case this number equals three.
\begin{itemize}
\item If the metric admits three Killing fields, then its either flat or maximally symmetric; the first possibility is already checked, while the second (as we have already proved) is not admissible.
\item If the metric admits two Killing fields $\xi_{(1)},\xi_{(2)}$, then the possible Lie algebras these fields can span, are either the Abelians $2A_1=\langle\partial_x,\partial_y\rangle$ and $2A_1=\langle\partial_x,y\partial_x\rangle$, or the non-Abelians $A_2=\langle\partial_x,e^x\,\partial_y\rangle$ and $A_2=\langle\partial_x,x\,\partial_x\rangle$, see e.g. \cite{Popovych:RealLieAlg,Terzis:2013Faith}.
In the Abelian case the second algebra yields a degenerate metric, while the first algebra reproduces a flat metric, since $G_{\alpha\beta}=const.$ (and of course admits a third Killing field).
In the non--Abelian case the second algebra yields a degenerate metric, while the first algebra reproduces a metric with a constant Ricci scalar, i.e. a maximally symmetric metric.
\end{itemize}
Finally the only case which is left to discuss is when the scaled supermetric \eqref{scGdd} admits only one Killing field. First of all, let us state some general facts; let $h_{\alpha\beta}=h_{\alpha\beta}(x,y)$ a two dimensional metric which admits a Killing field $\xi^\alpha$, then it is always possible to bring it into its normal form, i.e. $\xi=\partial_y$. As a result the metric can be put in the special conformal form
\bal \label{metonev}
h_{\alpha\beta}=\Omega(x)
\begin{pmatrix}
1 & 0\\ 0 & \epsilon
\end{pmatrix},
\eal
see Appendix \ref{2Dcase}. Obviously with the help of the transformation $x\to y,\, y\to x$ we can make the conformal factor $\Omega$ a function of $x$.
Let us now return to the supermetric \eqref{scGdd} and apply once more the transformation $a=\exp(w/\sqrt{3})/\sqrt{F(\phi)}$, which turns the line element into the form
\bal\label{confGdd}
\textrm{d} s^2=-\frac{4e^{2\sqrt{3}w}V(\phi)}{F^2(\phi)}\left( \textrm{d} w^2+\frac{\epsilon\,F(\phi)-3\,F'^2(\phi)}{4F^2(\phi)}\, \textrm{d}\phi^2 \right).
\eal
Employing the transformation $\phi=r(y)$ such that
\bal
\sqrt{ \left| \frac{\epsilon\,F(\phi)-3\,F'^2(\phi)}{4F^2(\phi)} \right|}\,\textrm{d}\phi=\textrm{d} y,
\eal
we bring the metric \eqref{confGdd} into the desired form
\bal \label{metwy}
\textrm{d} s^2=-\frac{4e^{2\sqrt{3}w}V(r(y))}{F^2(r(y))}\left( \textrm{d} w^2+\epsilon\,\textrm{d} y^2 \right).
\eal
In order for this line element to admit one Killing field, there must exist a transformation that brings \eqref{metwy} into the form \eqref{metonev}. When $\frac{V}{F^2} = c e^{\mu y}$ the space is flat and thus admits three Killing fields, this case has the space of solutions described by the sets \eqref{final_caseI} and \eqref{final_caseII}. For all other functional forms of $\frac{V}{F^2}$, the space is not flat and cannot be transformed into a form analogous to \eqref{metonev}.
\section{Physical interpretation and physical parameters}
In a FLRW universe some of the physical observation parameters, are the Hubble parameter $H$ and the dimensionless parameters deceleration $q$ and jerk $j$, see for example \cite{Blandford:2004,Nair:2011}. The Hubble parameter quantifies the expansion of the universe; the deceleration parameter nowadays measures the acceleration of the universe whereas the jerk parameter is needed since the universe was once decelerating and is now accelerating.
Their definitions in comoving coordinates ($\textrm{d} s^2=-\textrm{d} t^2+a^2(t)\textrm{d} r^2+a^2(t)r^2\textrm{d} \Omega^2$) are given by
\bal
H=\frac{a'(t)}{a(t)}, \quad q=-\frac{a(t)\,a''(t)}{a'^2(t)}, \quad j=\frac{a^2(t)\,a'''(t)}{a'^3(t)}.
\eal
The solution sets \eqref{final_caseI} and \eqref{final_caseII}, are not referring to comoving coordinates, but it is an easy task to make the transition, from the time coordinate $\tau$ to the desired one $t$. If we follow the redefinition of the lapse function $N(u)$ and the time re--parameterization $\textrm{d} \tau=N(u)\textrm{d} u$, we then see that these solutions, are expressed as the line element
\bal\label{ds_tau}
\textrm{d} s^2=-u(\tau)^2 \textrm{d} \tau^2 +a^2(\tau)\left( \textrm{d} r^2+r^2\,\textrm{d}\Omega^2 \right),
\eal
thus the two coordinates are connected by
\bal
u(\tau) \textrm{d} \tau=\textrm{d} t,
\eal
and the time derivatives of the scale factor $a(t)$ are
\bal
a'(t)=\frac{\textrm{d} a(\tau)}{u(\tau)\textrm{d} \tau},\quad a''(t)=\frac{\textrm{d}}{u(\tau) \textrm{d} \tau}\left( \frac{\textrm{d} a(\tau)}{u(\tau) \textrm{d} \tau}\right), \quad \dots
\eal
For each one of the two solutions the aforementioned parameters are quite cumbersome, due to the existence of the arbitrary function $h(\tau)$, but are quite straightforward to be calculated. We only present the form of the Hubble parameter for each case
\begin{subequations}
\bal
H_{I}&=\frac{1}{12\beta\,\left( \tau-\gamma \right)^{1/4}} e^{-3\alpha\tau/2}\left( -12\left( \tau-\gamma\right)h'(\tau) +\left( 6\alpha\tau-6\alpha\gamma+1 \right) h(\tau)\right) \\
H_{II}&=\frac{\left(\tau+\alpha \right)^{s/(2-2s)}\left(\tau-\beta \right)^{-s/(2+2s)}}{6\gamma\,\left(s^2-1 \right)}\Big( -6\left(s^2-1\right)\left(\tau-\beta\right) \left(\tau+\alpha\right) h'(\tau)+\nonumber\\
&\phantom{=} s\left( \left(s-1\right)\alpha-\left(s+1\right)\beta+2s\tau \right) h(\tau) \Big).
\eal
\end{subequations}
As we have mentioned in the Introduction, the duality of scalar field/fluid is widely used in cosmology, thus we are going to apply this procedure in our case, for details see \cite{Madsen1, Madsen2, Pimentel}.
The equations of motion \eqref{EoM1} can be rewritten as
\bal
R_{ij}-\frac{1}{2}Rg_{ij}=\frac{1}{F(\phi)}\left(T_{ij} +\nabla_i\nabla_j F(\phi)+g_{ij}\Box F(\phi)\right)\Rightarrow E_{ij}=T^{(\phi)}_{ij},
\eal
where $E_{ij}$ is the Einstein tensor and $T^{(\phi)}_{ij}$ is the effective energy--momentum associated with the
scalar field. With this energy--momentum tensor we want to associate an energy--momentum tensor of an imperfect fluid
\bal
T^{(imf)}_{ij}=\left( \rho+p \right)u_iu_j+p g_{ij}+2q_{(i}u_{j)}+\pi_{ij},
\eal
where $\rho$ is the energy density of the fluid, $u_i$ the 4--velocity, $q_i$ the heat flux vector, $p$
the pressure and $\pi_{ij}$ the anisotropic stress tensor. The relations that make the identification possible are
\begin{subequations}
\begin{alignat}{2}
\Pi_{mn}&=T^{(\phi)}_{ij} h^i{}_m h^j{}_n=p h_{mn}+\pi_{mn} &\quad\quad \pi_{mn}&=\Pi_{mn}-\frac{1}{3}\Pi_k{}^k h_{mn}=\Pi_{mn}-p h_{mn}\\
\rho&=T^{(\phi)}_{ij}u^i u^j & p&=\frac{1}{3}\Pi_i{}^i\\
q_k&=-T^{(\phi)}_{ij}u^i h^j{}_k,
\end{alignat}
\end{subequations}
where $h_{ij}$ is the projection tensor orthogonal to velocity $u_i$ defined by
\bal
h_{ij} = g_{ij}+ u_i u_j \quad \text{with} \quad u_i u^i=-1.
\eal
The natural choice of the 4--velocity $u_i$ is the one which is associated with the normalized derivative of the scalar field $\phi$, i.e.
\bal
u_i=\frac{\nabla_i\phi}{\sqrt{-\nabla_j\phi \nabla^j\phi}},
\eal
where we have assumed that $\nabla_i\phi$ is timelike in order to describe a physical fluid. Furthermore the kinematical quantities of the fluid that are of interest, and appear in the decomposition of the covariant derivative of the velocity \cite{ellis2012relativistic}
\bal
\nabla_i u_j=-\dot{u}_i u_j+\omega_{ij}+\sigma_{ij}+\frac{1}{3}\theta h_{ij},
\eal
are
\bal
\dot{u}_i=u^j\nabla_j u_i, \quad \theta=\nabla_i u^i, \quad \sigma_{ij}=\nabla_{(i} u_{j)}+\dot{u}_{(i}u_{j)}-\frac{1}{3}\theta h_{ij}, \quad \omega_{ij}=\nabla_{[i} u_{j]}+\dot{u}_{[i}u_{j]}
\eal
i.e. the acceleration, the expansion, the shear and the rotation of the fluid respectively.
It is quite remarkable that in our case, \emph{irrespectively} of the solution space the heat flow $q_i$ along with the anisotropic stress tensor $\pi_{ij}$ are zero, thus the stress tensor $T_{ij}^{(\phi)}$ mimics
a perfect fluid. Moreover the fluid has zero acceleration, is shear free, exhibits no rotation and the expansion is three times the Hubble parameter with an opposite sign. The forms of the pressure $p$ and the energy density $\rho$ for the case I are
\begin{subequations}
\bal
p&=\frac{e^{-3\alpha\tau}}{48\beta^2\sqrt{\tau-\gamma}}\Bigg(96\left(\tau-\gamma\right)^2 h(\tau)h''(\tau)- 144\left(\tau-\gamma\right)^2h'^2(\tau)\nonumber\\
&-8\left(\tau-\gamma \right)\left(6\left(\tau-\gamma\right)\alpha-11 \right)h(\tau) h'(\tau)\nonumber\\
&+\left(36\alpha^2\left(\tau-\gamma\right)^2-36\alpha (\tau-\gamma)+1 \right) h^2(\tau)\Bigg) \\
\rho&=\frac{e^{-3\alpha\tau}}{48\beta^2\sqrt{\tau-\gamma}}\Big(12\left(\tau-\gamma\right) h'(\tau)-
\left(6\alpha\tau-6\alpha\gamma +1\right) h(\tau)\Big)^2,
\eal
\end{subequations}
while for the case II are given by
\begin{subequations}
\bal
p&=d(\tau) \Bigg(6\left(s^2-1\right)(\tau+\alpha)(\tau-\beta) h(\tau)h''(\tau)-9\left(s^2-1\right)(\tau+\alpha)(\tau-\beta)h'^2(\tau)\nonumber\\
&+\left( (5s+6)(s-1)\alpha-(5s-6)(s+1)\beta+2\left( 5s^2-6\right)\tau \right) h(\tau)h'(\tau)\nonumber\\
&-\frac{s^2}{4}\left(4 -\frac{\left( 2\tau-(s+1)\beta-(s-1)\alpha\right)^2}{\left(s^2-1\right)(\tau+\alpha)(\tau-\beta)} \right) h^2(\tau) \Bigg) \\
\rho&=d(\tau) \Bigg(9\left(s^2-1\right)(\tau+\alpha)(\tau-\beta) h'^2(\tau)-3s\left(2s\tau+(s-1)\alpha-(s+1)\beta\right)h(\tau) h'(\tau)\nonumber\\
&+\frac{s^2\left( 2s\tau+(s-1)\alpha-(s+1)\beta \right)^2}{4\left(s^2-1\right)(\tau+\alpha)(\tau-\beta)} h^2(\tau)\Bigg),
\eal
\end{subequations}
where
\bal
d(\tau)=\frac{(\tau+\alpha)^{1/(1-s)}(\tau-\beta)^{1/(s+1)}}{3\gamma^2\left(s^2-1\right)} . \nonumber
\eal
The above procedure of evaluating the pressure $p$ and the energy density $\rho$ has already been criticized as being non--physical in the case of vacuum scalar--tensor theory. In \cite{RomeroBarros} the authors started from the solutions of the vacuum field equations of Brans-Dicke scalar-tensor theory of gravity and calculated the corresponding energy--momentum tensor for the perfect fluid. Their conclusion was that \emph{''The examples presented in this paper seems to
suggest that this sort of equivalence is sometimes purely formal and rather artificial.''}. Thus in order for one to be on the safe side he must demand a physical character of the presented values of the pressure $p$ and the energy density $\rho$. The minimum assertions that can guarantee that sort of physical acceptance is the various energy conditions, which for the perfect fluid can then be formulated in terms of the eigenvalues of this energy momentum tensor:
\begin{itemize}
\item The weak energy condition stipulates that $\rho \ge 0, \; \; \rho + p \ge 0 .$
\item The null energy condition stipulates that $\rho + p \ge 0 $.
\item The strong energy condition stipulates that $\rho + p \ge 0, \; \; \rho + 3 p \ge 0 $.
\item The dominant energy condition stipulates that $\rho \ge |p| .$
\end{itemize}
Below we present a solution that satisfies the above energy conditions, expect the strong energy one. The validation of the strong energy condition is being criticized nowadays; one of its violation can be seen from the recent observational data regarding the acceleration of our universe, for more details see \cite{Visser:1999}.
\section{Special solutions}
In this section we intend to explore some physical consequences of the solution space found. To this end, we select some particular forms for the free function $h(\tau)$ parameterizing the different solutions. We thus arrive at the three cases given below.
\subsection{Energy Complete Solution}
As we have already mentioned each solution space \eqref{final_caseI}, \eqref{final_caseII} is modeled by the existence of the arbitrary function $h(\tau)$. A natural choice would be to make the $00$ component of the line element \eqref{ds_tau} equal to minus one, thus bringing it to comoving coordinates; it is to be noticed that with this procedure we \emph{do not} apply any coordinate transformation but we only make a specific choice of $h(\tau)$. For case I the function $h(\tau)$ which accomplishes this is
\bal\label{h_caseI}
h(t)=\beta\,e^{3\alpha\, t/2}\left( t-\gamma \right)^{-3/4},
\eal
(for simplicity we write $t$ instead of $\tau$) and the line element \eqref{ds_tau} reads
\bal\label{ds_t}
\textrm{d} s^2=-\textrm{d} t^2+e^{-2\alpha t}t^{5/3}\left( \textrm{d} r^2+r^2\textrm{d} \theta^2 +r^2\sin^2\theta\textrm{d} \phi^2\right).
\eal
For the above line element the Hubble parameter $H$, the deceleration parameter $q$ and the jerk parameter $j$ are
\bal\label{Hqj}
H=\frac{5}{6t}-\alpha,\quad q=-1+\frac{30}{\left( 6\alpha t-5 \right)^2}\, \quad j=1+90\,\frac{ 6\alpha t-1 }{\left( 6\alpha t-5 \right)^3}.
\eal
Furthermore the pressure and the energy density of the perfect fluid are
\bal
p=\frac{20-\left( 6\alpha t-5\right)^2}{12t^2}, \quad \rho=\frac{\left( 6\alpha t-5\right)^2}{12t^2},
\eal
which yield an equation of state $p=w\rho$, with variable equation of state parameter $w$
\bal
w=\frac{20-\left( 6\alpha t-5\right)^2}{\left( 6\alpha t-5\right)^2}\Rightarrow w=-1+\frac{20}{\left( 6\alpha t-5\right)^2}.
\eal
\begin{figure}[t]
\centering
\begin{minipage}{0.45\linewidth}
\includegraphics[width=0.9\textwidth]{Fig_w-t.jpg}
\caption{The zeros of the state parameter $w$ occur at $t_\pm=\left(5\pm2\sqrt{5} \right)/(6\alpha)$.}
\label{fig:w-t}
\end{minipage}
\,
\begin{minipage}{0.45\linewidth}
\includegraphics[width=0.9\textwidth]{Fig_q-t.jpg}
\caption{The deceleration epoch $q>0$ is followed by an acceleration one $q<0$.}
\label{fig:q-t}
\end{minipage}
\end{figure}
From the above we conclude the following facts
\begin{enumerate}
\item It is easy to see that the energy conditions, (expect the strong energy one) are satisfied for any value of the parameter $\alpha$ and for $t \neq 0$. Thus the induced perfect fluid is a physical one, although the parameter $w$ is $t$--dependent.
\item For $t=0$ the parameter $w$ equals $-1/5$ for every value of the constant $\alpha$, while for $t \to \infty, w\to -1$, i.e. the fluid behaves like a cosmological constant see figure \eqref{fig:w-t}.
\item In order to have a universe that is expanding we must require that the constant $\alpha$ in line element \eqref{ds_t} to be negative, $\alpha<0.$
\item The deceleration parameter $q$, equals $q=1/5$ at $t=0$ and changes its sign at $t_-=\disfrac{n-\sqrt{n}}{\alpha},\, n=5/6$; thus we are describing a universe that initially is expanding though decelerating, and after $t_-$ the expansion is accelerating, a fact that is supported by recent observations see \cite{Riess:2004dec}. This behavior is shown in figure \ref{fig:q-t}.
\item The jerk parameter is always positive for $t>0$ and for all $\alpha$, thus the acceleration is always increasing.
\item Since both the pressure $p$ and the energy density $\rho$ are functions of $t$ we can find a relation between the two of them; the result is $p=\dfrac{12}{5}\left(\alpha\pm \sqrt{\dfrac{\rho}{3}} \right)^2-\rho$.
\end{enumerate}
We can use the above results to estimate the age of the universe $t_0$. From the observations of the Hubble Space Telescope Key project \cite{Freedman:2000cf}, the present Hubble parameter is constrained to be
\bal
H_0^{-1}= 9.776h^{-1}\, Gyr, \quad 0.64 < h < 0.80,
\eal
where the subscript $0$ indicates present time. We could insert this value into \eqref{Hqj} to estimate $t_0$, but we must know the value of the constant $\alpha$. In order to calculate $\alpha$ we will make use of the knowledge of the redshift $z_{cr}$ where the cosmic jerk happened (at $t_-$), i.e. the redshift where the deceleration parameter changed sign indicating that the current epoch of cosmic acceleration was preceded by a cosmic deceleration one. The Supernova Search Team constrained this value to $z_{cr}=0.46 \pm 0.13$ \cite{Riess:2004dec}. From the relation of the scale factor $a(t)$ and the redshift $a_0/a=1+z$ we
have
\bal\nonumber
t&=t_0: a_0=e^{-\alpha t_0}t_0^{5/6}\\
t&=t_-: a_0=e^{-\alpha t_-}t_-^{5/6}(1+z_{cr})
\nonumber,
\eal
and using \eqref{Hqj} $\alpha=5/(6t_0)-H_0$ we end up
\bal
e^{-\sqrt{n}+x}=\left( \frac{\sqrt{n}-n}{x-n} \right)^n(1+z_{cr}), \quad n=\frac{5}{6},\quad x=H_0\,t_0.
\eal
Using $z_{cr}=0.46,\, h=0.72$ we have $x=0.952737$ thus $t_0=12.9361\, Gyr$ a very fine result for this model, since the most recent WMAP3 data
produces a value of $t_0 = 13.73^{+0.13}_{-0.17}\, Gyrs$ (assuming an \textrm{$\Lambda$}CMD model) \cite{Spergel:2006hy}.
\subsection{Singular supermetric - A dust solution}
The previous analysis of the solution space is valid in the case where the mini-supermetric $G_{\alpha\beta}$ is not singular, i.e. $\det{G_{\alpha\beta}} \neq 0$. Thus we have to consider separately the case where $\det{G_{\alpha\beta}} = 0$.
The determinant of the scaled supermetric \eqref{scGdd} is
\bal
G=12a^{10}V^2(\phi) \left( \epsilon\, F(\phi)-3F'^2(\phi) \right) \nonumber
\eal
which is zero when
\bal
F(\phi)=\frac{\epsilon}{12}\left( \phi - c\right)^2.
\eal
Taking for simplicity $\epsilon=1$ and $N=1$, we can calculate $V(\phi)$ from the $00$ component of the field equations \eqref{EoM1}
\bal
V(\phi)=\frac{2}{a^4\left( \left(\phi-c \right) a'+a\, \phi'\right)^2} \nonumber.
\eal
Substituting the above $V(\phi)$ in the field equation \eqref{EoM2}, we can solve for $\phi''$ with the help of which all the components of \eqref{EoM1} are made proportional to
\bal
4\alpha \phi'+6\left( \phi-c\right) a' =0, \nonumber
\eal
which can be integrated to
\bal
\phi=c+c_1\,a^{-3/2},
\eal
where $c_1$ is a constant of integration. With the above information at hand, equation \eqref{EoM2} reads
\bal
2a a''-5a'^2=0\Rightarrow a=\frac{c_2}{\left( 3t+2c_3\right)^{2/3}},
\eal
where $c_2,c_3$ are constants of integration. The coordinate transformation $t\to -2c_3/3+c_1^6/(648\, c_2^3\, t)$ along with the redefinition $c_2 \to \left(c_1^{4/3}6^{-2/3}\right)\kappa^{1/3}$, makes $\kappa$ a multiplicative constant, and simultaneously brings the line element to the form
\bal\label{ds_p=0}
\textrm{d} s^2=-\textrm{d} t^2+t^{4/3}\left( \textrm{d} r^2+r^2 \textrm{d} \theta^2+r^2\sin^2\theta \textrm{d} \phi^2\right).
\eal
The overall constant $\kappa$ does not appear in the line element, since it admits the homothetic vector field $h=3t\partial_t+r\partial_r$ which can absorb it. The final form of the scalar field $\phi(t)$, the potential $V(t)$, and the function $F(t)$ are
\bal
\phi(t)=c+\frac{c_1}{\kappa^{3/2}t},\quad V(t)=\frac{c_1^2}{18\kappa^5 t^4},\quad F(t)=\frac{c_1^2}{12\kappa^3 t^2}.
\eal
We can calculate the values of the Hubble, the deceleration and the jerk parameters for line element \eqref{ds_p=0}
\bal
H=\frac{2}{3t}, \quad q=\frac{1}{2}, \quad j=1,
\eal
thus we are describing a universe that expands while decelerating. Furthermore if we apply the identification of the scalar field to the perfect fluid we find that $p=0$ and $\rho=4/(3\kappa^2 t^2)$; thus we have the dust solution of Friedmann \cite{Friedmann1924, Friedmann1924en}.
It is quite interesting that this General Relativity solution is found from the perspective of scalar--tensor gravity, as an exceptional case.
\subsection{A cosmological solution with a constant parameter of state $w$}
Solution sets \eqref{final_caseI} and \eqref{final_caseII} can be reduced into General Relativity's theme by demanding the constancy of the function $F(\phi)$. A wide class of cosmologies can be inferred from these solutions; we are going to present one with constant parameter of state $w$, i.e. $p=const. \, \rho$.
For the set \eqref{final_caseII} in order to have $F(\phi)=1$ we must take $h(\tau)=2$. If we choose $\beta=-\alpha$ it is easy to see that the parameter of state reads
\bal
w=-1+\frac{2}{s^2}.
\eal
Performing the coordinate transformation $\tau=z^{z/(z-2)t^z-\alpha},\, z=s^2-1$ and making the redefinition $\gamma=2z^{s^2/(s^2-3)}\kappa$, in order to make $\kappa$ an overall constant, we end up with the line element
\bal
\textrm{d} s^2=-\textrm{d} t^2+t^{2s^2/3} \left( \textrm{d} r^2+r^2 \textrm{d} \theta^2+r^2\sin^2\theta\, \textrm{d}\phi^2 \right),
\eal
where $t>0$. This line element admits the homothetic vector field $h=3t\partial_t+(3-s^2)r\partial_r$ which justifies the omission of $\kappa$ in front of it. The pressure $p$ and the energy density $\rho$ of the perfect fluid are
\bal
p=\frac{s^2\left( -s^2+2 \right)}{3\,t^2},\, \rho=\frac{s^4}{3\,t^2} \Rightarrow w=-1+\frac{2}{s^2}.
\eal
This solution can be found in \cite{Stephani2003exact}, p.212 equation (14.8b), and it describes a decelerating expanding universe with constant deceleration parameter, since
\bal
H=\frac{2\,s^2}{3\,t}, \quad q=-1+\frac{3}{2\,s^2}, \quad j=1+\frac{9}{2\,s^4}(1-s^2).
\eal
\section{Discussion}
The use of symmetries in the process of acquiring new solutions is widespread in mathematical cosmology. In our case this approach was applied, in the context of a spatially flat FLRW space--time, to the mini--superspace Lagrangian of the scalar tensor theory of gravity. In order to acquire all the existing autonomous, linear in the momenta integrals of motion for a given cosmological system, its singular nature must necessarily be taken into account \cite{CDT2014,tchris_sch}. One of the main aims of the present work is to highlight exactly this: the immense possibilities that can be explored by taking into account the reparametrization invariance of constrained systems.
Unfortunately, a common practise in the literature is, to gauge fix the lapse function (usually to $1$), so that the theory of Noether symmetries for regular systems can be applied. However, this process is misleading in what regards the properties of the system under consideration. As it is known, the mini--superspace Lagrangians ensuing from cosmological systems are singular and belong to the general form \eqref{Lag}. By gauge fixing the lapse, for example $N=1$, the new fixed Lagrangian reads
\begin{equation}
L_{\mathrm{fixed}}=\frac{1}{2} G_{\alpha\beta} x'^\alpha\, x'^\beta - U(x)
\end{equation}
and describes a system different from \eqref{Lag}. Of course the former can admit the same solution if one uses the constraint equation $\frac{\partial L}{\partial N}=0$ of the initial system, as an \emph{ad hoc} condition. Nevertheless, as far as the search of symmetries is concerned, this procedure becomes too restrictive. The fixing of the lapse annihilates the freedom of the reparametrization invariance that in itself, as shown in \cite{CDT2014} and \cite{tchris_sch}, is a source for the emergence of linear in the momenta integrals of motion which are not obtained in the theory of regular systems.
All the previous arguments can be made clearer in the context of scalar tensor gravity that we have treated in this paper. By comparing results with \cite{tsamp}, where the authors start from the same action \eqref{action}, but in the process follow the gauge fixing approach, one can see that: Under the same condition that we used here, i.e. the mini-superspace being maximally symmetric, they are led to a specific functional form for $F(\phi)$, say $F_{\mathrm{fixed}}(\phi)$, for which $G_{\alpha\beta}$ is flat. Subsequently, they apply each of the three killing fields to the potential $U(x)$ acquiring three different scalar field potentials. Each of them is used to describe a regular system that admits one autonomous integral of motion generated by the corresponding Killing field of $G_{\alpha\beta}$ \footnote{Some extra cases admitting rheonomic integrals of motion are also explored, for more details see \cite{tsamp}}. However, one can notice that all three scalar field potentials belong to the same functional form, namely the form that makes the scaled mini--supermetric $\xbar{G}_{\alpha\beta}= U\, G_{\alpha\beta}$ flat for the specific value $F_{\mathrm{fixed}}(\phi)$ of the coupling function.
In this work, we use the reparametrization invariance which leads to the consideration of the scaled mini--supermetric $\xbar{G}_{\alpha\beta}$ as the crucial element describing the geometry of the configuration space and the dynamics of the system. Consequently, the demand for maximal symmetry does not fix the coupling function, but yields a relation between $F(\phi)$ and the scalar potential $V(\phi)$ \eqref{FVpar}. For the specific value $F_{\mathrm{fixed}}(\phi)$, treated in \cite{tsamp}, the potential $V(\phi)$ assumes the general functional form in which the three potentials given in that paper belong. Thus, what is considered as three different cases admitting one autonomous integral of motion in the study of the regular Lagrangian, is really one case admitting three autonomous integrals of motion in the actual (singular) cosmological system. Moreover, this is just a single case in the study of the singular Lagrangian \eqref{Lag}, since $F(\phi)$ is not fixed to an explicit functional form. Thus, the particular function $F_{\mathrm{fixed}}(\phi)$ is a choice, not a necessity for satisfying the demand of a flat the mini--superspace. The result is an infinite set of scalar tensor theories admitting the maximal number of autonomous, linear in the momenta integrals of motion. As we proved, even if one requires less symmetries, i.e. one or two autonomous charges, one is led to the case here examined: Each choice of $F(\phi)$, yields though \eqref{FVpar} the appropriate potential for a maximally symmetric (eventually flat) mini--superspace.
For all infinite cases that arise from the condition of maximal symmetry, we were able to acquire the general solution space for an arbitrary coupling function $F(\phi)$. This is not to be taken lightly; it means that the obtained sets \eqref{final_caseI} and \eqref{final_caseII} represent the general analytic solutions of \underline{every} scalar tensor, spatially flat FLRW cosmological theory that admits an autonomous, linear in the momenta symmetry. We also calculated all the physically relevant parameters and the effective energy--momentum tensor associated with the scalar field, which is seen to be mimicking a perfect fluid behaviour from Einstein's gravity's perspective.
We would like to emphasize that the correspondence between the scalar field and the perfect fluid we use is not the usual: The common practice is to identify the energy momentum tensor of the scalar field \eqref{Tij} ($V(\phi)=0$) with the energy momentum tensor of a perfect fluid. Our line of thinking is to rewrite the field equations in the form $E_{ij}=T_{ij}^{\phi}$ and treat the rhs as an energy momentum tensor. This different approach is responsible for enabling us to arrive to physically meaningful results.
In order to exhibit the way the general relations can be used and to complete our analysis, we have given some specific examples: a) A solution for a particular choice of the coupling function, that satisfies all the energy conditions (apart from the strong energy condition) and whose behaviour considerably matches with many observational facts. b) For the shake of completeness, we investigated the case when the mini--supermetric is degenerate, the only instance that is not covered by the general theory. This led to a solution that is seen to be equivalent to the dust solution of Friedmann in the context of General Relativity. c) We also obtained the known solution of General Relativity for a perfect fluid with a constant equation of state parameter $w$. This happened by considering the case $F(\phi)=1$ (minimally coupled scalar field) and by a suitable choice of the parameters entering the effective energy--momentum tensor. Of course this is not a new solution, but it serves to exhibit that the general solution for an arbitrary $F(\phi)$ correctly correlates to Einstein's theory when one sets $F(\phi)$ to a constant.
Since the presented method is quite a general one, it would be interesting to apply it to a broader setting i.e. one could add an actual perfect fluid along with the scalar field or explore the possibility of the existence of two scalar fields or even in more general theory like Horndeski's \cite{Horndeski1974}.
\newpage
|
\section{#1}\setcounter{equation}{0}}
\makeatletter
\@addtoreset{equation}{section}
\makeatother
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\def\one{{\hbox{ 1\kern-.8mm l}}}
\def\zero{{\hbox{ 0\kern-1.5mm 0}}}
\def\hr{\hat{r}}
\def\hro{\hat{r_1}}
\def\tC{ { \tilde C} }
\def\rp{ {r^{\prime}} }
\def\hs{\hat{s}}
\def\mE{ \mathcal{E}}
\def\cA{{\cal A}} \def\cB{{\cal B}} \def\cC{{\cal C}}
\def\cD{{\cal D}} \def\cE{{\cal E}} \def\cF{{\cal F}}
\def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}}
\def\cJ{{\cal J}} \def\cK{{\cal K}} \def\cL{{\cal L}}
\def\cM{{\cal M}} \def\cN{{\cal N}} \def\cO{{\cal O}}
\def\cP{{\cal P}} \def\cQ{{\cal Q}} \def\cR{{\cal R}}
\def\cS{{\cal S}} \def\cT{{\cal T}} \def\cU{{\cal U}}
\def\cV{{\cal V}} \def\cW{{\cal W}} \def\cX{{\cal X}}
\def\cY{{\cal Y}} \def\cZ{{\cal Z}}
\def\mbX{ { \mathbb{X}} }
\def\g{ \gamma }
\def\s{ \sigma }
\def\L{ \Lambda }
\def\bT{ { \bf{T}} }
\def\mP{ \mathbb{P} }
\def\mbF{ \mathbb{F} }
\def\bT{ { \bf{T}} }
\def\mbC{ \mathbb{C}}
\def\mbN{ \mathbb{N} }
\def\mbZ{ \mathbb{Z} }
\def\mbI{ \mathbb{I} }
\def\tr{ {\rm tr } }
\def\Dim{ {\rm Dim} }
\def\Sym{ {\rm Sym} }
\def\tr{ {\rm tr } }
\def\Str{ {\rm Str } }
\def\des{ {\rm desc} }
\def\planar{ {\rm planar} }
\def\cK{ \cal{K} }
\def\hOmega{ \widehat { \Omega } }
\begin{document}
\makeatletter
\@addtoreset{equation}{section}
\makeatother
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\rightline{WITS-CTP-132}
\vspace{1.8truecm}
\vspace{15pt}
{\LARGE
\centerline{\bf Heavy Operators in}
\centerline{\bf Superconformal Chern-Simons Theory}
}
\vskip.5cm
\thispagestyle{empty} \centerline{
{\large \bf Robert de Mello Koch\footnote{{\tt <EMAIL>}},}
{\large \bf Rocky Kreyfelt\footnote{{\tt <EMAIL>}}
and Stephanie Smith\footnote{{\tt <EMAIL>}}}
}
\vspace{.4cm}
\centerline{{\it National Institute for Theoretical Physics ,}}
\centerline{{\it School of Physics and Centre for Theoretical Physics }}
\centerline{{\it University of Witwatersrand, Wits, 2050, } }
\centerline{{\it South Africa } }
\vspace{1.4truecm}
\thispagestyle{empty}
\centerline{\bf ABSTRACT}
\vskip.4cm
We study the anomalous dimensions for scalar operators in ABJM theory in the $SU(2)$ sector.
The operators we consider have a classical dimension that grows as $N$ in the large $N$ limit.
Consequently, the large $N$ limit is not captured by summing planar diagrams - non-planar contributions have to be included.
We find that the mixing matrix at two-loop order is diagonalized using a double coset ansatz, reducing it to the Hamiltonian
of a set of decoupled oscillators.
The spectrum of anomalous dimensions, when interpreted in the dual gravity theory, shows that the energy of the fluctuations
of the corresponding giant graviton is dependent on the size of the giant.
The first subleading corrections to the large $N$ limit are also considered.
These subleading corrections to the dilatation operator do not commute with the leading terms, indicating that integrability
may not survive beyond the large $N$ limit.
\setcounter{page}{0}
\setcounter{tocdepth}{2}
\newpage
\setcounter{footnote}{0}
\linespread{1.1}
\parskip 4pt
\section{Introduction}\label{intro}
Integrability has proven to be a powerful tool in analyzing the spectrum of anomalous dimensions in ${\cal N} = 4$
Super Yang-Mills theory in the planar limit\cite{Minahan:2002ve,Beisert:2010jr}.
An interesting question is whether or not there are other large $N$ limits that are also integrable.
This question has been the focus of a number of recent
studies\cite{Koch:2010gp,DeComarmond:2010ie,Carlson:2011hy,Koch:2011hb,deMelloKoch:2011ci,deMelloKoch:2012ck,deMelloKoch:2011vn,deMelloKoch:2012sv,Koch:2012sf,Koch:2013xaa,Koch:2013yaa,Lin:2014yaa}.
At this point there is evidence that suggests certain large $N$ limits, that are not captured by simply summing the planar diagrams, do enjoy integrability.
The studies described above have all focused on ${\cal N}=4$ super Yang-Mills theory.
In this article we extend existing studies by exploring a large $N$ but non-planar limit of the ABJM theory, which is an ${\cal N}=6$ superconformal Chern-Simons-matter
theory with gauge group $U(N)\times U(N)$ on $R^{1,2}$ and Chern-Simons levels $k$ and $-k$.
Almost all of the results that have been obtained in the planar limit of ${\cal N} = 4$ super Yang-Mills theory hold in an appropriately modified form for the ABJM theory\cite{Klose:2010ki}.
Further, the technology needed to study operators with anomalous dimensions that grow as $N$ (called ``heavy operators'') has
been developed\cite{deMelloKoch:2012kv,Pasukonis:2013ts,Koch:2014yga}.
It is thus very natural to search for possible large $N$ but non-planar limits of ABJM theory that enjoy integrability.
This is the primary motivation for the study reported in this article.
We confine attention to the $SU(2)$ sector of theory and work at two loops.
In this case, relying on results of \cite{Koch:2014yga}, we are able to give a simple description, which employs restricted Schur polynomials.
Concretely, \cite{Koch:2014yga} proved that a basis for the operators in this sector of the theory is provided by restricted
Schur polynomials in the adjoints (of one of the $U(N)$ factors) constructed out of the bifundamental scalars fields.
The delicate point, resolved in \cite{Koch:2014yga}, involves demonstrating that the finite $N$ constraints are correctly
accounted for.
Our polynomials employ two adjoints, called $\phi_{11}$ and $\phi_{12}$ below.
The number of $\phi_{11}$ fields is $n_{11}$ and the number of $\phi_{12}$ fields is $n_{12}$.
As we show in section 2, the structure of the one loop dilatation operator for ABJM theory differs from that of ${\cal N}=4$ super
Yang-Mills theory.
The operators we consider are labeled by Young diagrams with $O(1)$ rows or columns and a total of $O(N)$ boxes.
For these operators we can employ the displaced corners approximation of \cite{Koch:2011hb}.
This requires $n_{12}\gg n_{11}$.
In this approximation, the leading terms in the dilatation operator are diagonalized using a double coset ansatz\cite{deMelloKoch:2012ck} and the results of spring field theory\cite{deMelloKoch:2011ci}.
The dilatation operator reduces to a set of decoupled oscillators.
There are subleading terms of size ${n_{11}\over n_{12}}$ relative to the leading contribution, which represent corrections to
the large $N$ limit.
These subleading terms are not diagonalized by the ansatz of \cite{deMelloKoch:2012ck}, so that a careful treatment
of these terms would indicate whether the large $N$ but non-planar integrability is a property only of the large $N$ limit.
Our study shows that these subleading terms do not commute with the leading order, so that they are not diagonalized
by the ansatz of \cite{deMelloKoch:2012ck}.
Although this does not prove that the system is not integrable, it does suggest that the integrability we have found
is only a property of the large $N$ limit.
Given similar results obtained in the planar limit of the theory\cite{Kristjansen:2008ib,Caputa:2009ug}, this is not surprising.
In last section we summarize our results and point out some interesting directions in which this study can be extended.
There are a number of further works related to our study, with relevant background.
In particular, \cite{Berenstein:2008dc} lays the foundation for the description of membranes in ABJM using a group
theoretic perspective.
See also \cite{Berenstein:2006qk,de Mello Koch:2007uu} for background from the ${\cal N}=4$ super Yang-Mills
theory which is relevant for our study.
\section{$SU(2)$ Dilatation Operator in Adjoint Variables}\label{advars}
We are studying an ${\cal N}=6$ Chern-Simons gauge theory with $U(N)\times U(N)$ gauge group.
The generalized restricted Schur polynomials, introduced and studied in \cite{Pasukonis:2013ts} provide a basis for the local operators of any quiver gauge theory with gauge group built from unitary group factors.
In constructing our local operators we will use scalar fields $A_1,A_2$ both transforming in the $(N,\bar{N})$ of $U(N)\times U(N)$, as well
as $B_1^\dagger,B_2^\dagger$ which transform in the $(\bar{N},N)$.
Given these transformation properties, it is clear that the fields
$$
\phi_{11}{}^a_b =A_1{}^a_\alpha B_1^\dagger{}^\alpha_b \, ,
\qquad
\phi_{12}{}^a_b =A_1{}^a_\alpha B_2^\dagger{}^\alpha_b \, ,
$$
$$
\phi_{21}{}^a_b =A_2{}^a_\alpha B_1^\dagger{}^\alpha_b \, ,
\qquad
\phi_{22}{}^a_b =A_2{}^a_\alpha B_2^\dagger{}^\alpha_b\, .
$$
transform in the adjoint of the first $U(N)$ and as a singlet of the second.
In general, the description of the theory in terms of these adjoint fields does not correctly capture the finite $N$ physics.
Indeed, as explained in \cite{Koch:2014yga}, the constraints on local operators at finite $N$ arising from the fact that the adjoints are $N\times N$ matrices is a subset of the full set of constraints, arising
because both $A_I$ and $B_I^\dagger$ are $N\times N$ matrices.
However, if we restrict to the so called $SU(2)$ sector in which only $\phi_{11}$ and $\phi_{12}$ are used, the finite $N$ constraints resulting from the description employing adjoint scalars $\phi_{11}$ and $\phi_{12}$
agree with the constraints obtained from the original variables.
The description employing adjoints has the advantage that the restricted Schur polynomials of \cite{deMelloKoch:2012kv} provides a suitable basis, and the technology to work with these operators is well developed (see for example \cite{Koch:2011hb}).
The restricted Schur polynomials we use are
\bea
\chi_{R,\{ r\},\alpha\beta}(\phi_{11},\phi_{12})
=\frac{1}{n_{11}!n_{12}!}\sum_{\sigma \in S_{m_1+m_2}}{\rm Tr}_{\{r\},\alpha\beta}(\Gamma_R(\sigma)) Tr\Big(\sigma (\phi_{11})^{\otimes n_{11}}(\phi_{12})^{\otimes n_{12}}\Big)
\label{ourOps}
\eea
where we are considering an operator constructed using $n_{11}$ $\phi_{11}$ fields and $n_{12}$ $\phi_{12}$ fields.
$\{ r\}$ denotes an irreducible representation of $S_{n_{11}}\times S_{n_{12}}\subset S_{n_{11}+n_{12}}$.
It is useful to think of $\{r\}$ as a pair of Young diagrams, one with $n_{11}$ boxes and one with $n_{12}$ boxes.
The irreducible representation $\{r\}$ may appear more than once upon restricting the
representation $R$ of $S_{n_{11}+n_{12}}$ to the $S_{n_{11}}\times S_{n_{12}}$ subgroup.
The multiplicity labels $\alpha,\beta$ distinguish between these different copies.
The trace ${\rm Tr}_{\{r\},\alpha\beta}(\Gamma_R(\sigma))$ is an instruction to trace only over the $\{r\}$ subspace within the carrier space of $R$.
Further, row indices are traced over the $\alpha$ copy of $\{r\}$ while the column indices are traced over the $\beta$ copy.
To implement the restricted trace we introduce intertwining operators $P_{R,\{r\},\alpha\beta}$ defined so that
\bea
{\rm Tr}_R\Big(P_{R,\{r\},\alpha\beta}\Gamma_R(\sigma)\Big)={\rm Tr}_{\{r\},\alpha\beta}\Big(\Gamma_R(\sigma)\Big)
\eea
where the trace on the LHS now runs over the full carrier space of $R$.
Our conventions for the action of the symmetric group in the space $V^{\otimes n_{11}+n_{12}}$ on which the multilinear operators $(\phi_{11})^{\otimes n_{11}}(\phi_{12})^{\otimes n_{12}}$ act
are as follows
\bea
(\sigma)^I_J =\delta^{i_1}_{j_{\sigma(1)}}\cdots\delta^{i_{n_{11}+n_{12}}}_{j_{\sigma(n_{11}+n_{12})}}
\eea
The two point function of these operators is\cite{deMelloKoch:2012kv}
\bea
\langle \chi_{R,\{ r\},\alpha\beta}(\phi_{11},\phi_{12}) \chi_{S,\{ s\},\gamma\delta}(\phi_{11},\phi_{12})^\dagger \rangle
=\delta_{RS}\delta_{r_{11}s_{11}}\delta_{r_{12}s_{12}}\delta_{\alpha\gamma}\delta_{\beta\delta}{f_R^2{\rm hooks}_R\over{\rm hooks}_{r_{11}}{\rm hooks}_{r_{12}}}
\label{twopoint}
\eea
We will need this result below.
The dilatation operator, acting in this $SU(2)$ sector, is given by\cite{Kristjansen:2008ib}
\bea
D=-\left( {4\pi\over k}\right)^2 :{\rm Tr}
\Big[ \left(B_2^\dagger A_1 B_1^\dagger - B_1^\dagger A_1 B_2^\dagger\right)\left(
{\partial\over\partial B_2^\dagger}{\partial\over\partial A_1}{\partial\over\partial B_1^\dagger}
- {\partial\over\partial B_1^\dagger}{\partial\over\partial A_1}{\partial\over\partial B_2^\dagger}\right)\Big]:
\eea
A straightforward application of the chain rule allows us to rewrite this in terms of adjoint fields as\footnote{For the ABJ theory
with gauge group $U(N)\times U(M)$, assuming $M>N$, the only change in this formula is that the factor of $N$ in the third last
line of (\ref{newdil}) would be replaced by an $M$. The $\phi_{ij}$ would continue to be $N\times N$ matrices. If $N<M$ our
description changes as we would need to form $U(M)$ adjoints.}
\bea
&& :{\rm Tr} \Big[ \left(B_2^\dagger A_1 B_1^\dagger - B_1^\dagger A_1 B_2^\dagger\right)\left(
{\partial\over\partial B_2^\dagger}{\partial\over\partial A_1}{\partial\over\partial B_1^\dagger}
- {\partial\over\partial B_1^\dagger}{\partial\over\partial A_1}{\partial\over\partial B_2^\dagger}\right)\Big]:\cr
&&= :{\rm Tr}
\Big[\left(\phi_{12}\phi_{11}-\phi_{11}\phi_{12}\right)\left(
{\partial\over\partial\phi_{12}}\phi_{1j} {\partial\over\partial\phi_{1j}}{\partial\over\partial\phi_{11}}
- {\partial\over\partial\phi_{11}}\phi_{1j}{\partial\over\partial\phi_{1j}}{\partial\over\partial\phi_{12}}\right)\Big]:\cr
&&+N : {\rm Tr} \Big[\left(\phi_{12}\phi_{11}-\phi_{11}\phi_{12}\right)\left(
{\partial\over\partial\phi_{12}}{\partial\over\partial\phi_{11}}
- {\partial\over\partial\phi_{11}}{\partial\over\partial\phi_{12}}\right)\Big]:\cr
&&+:{\rm Tr}\left[(\phi_{12}\phi_{11}-\phi_{11}\phi_{12}){\partial\over\partial\phi_{12}}\right]{\rm Tr}\left[{\partial\over\partial \phi_{11}}\right]:\cr
&&-:{\rm Tr}\left[(\phi_{12}\phi_{11}-\phi_{11}\phi_{12}){\partial\over\partial\phi_{11}}\right]
{\rm Tr}\left[{\partial\over\partial \phi_{12}}\right]:
\label{newdil}
\eea
We now turn to the problem of evaluating the action of the dilatation generator on the operators (\ref{ourOps}).
The evaluation uses the technology developed in \cite{DeComarmond:2010ie,Koch:2011hb}.
The matrix derivatives are straight forward to evaluate; in manipulating the resulting expressions the identity
$$
{\rm Tr}(\rho\cdot\alpha\cdot\beta \phi^{\otimes n})=\prod_{A=1}^n\phi^{l_{\beta^{-1}(A)}}_{l_{(\alpha\rho)(A)}}
$$
is extremely useful.
To express the result of the action of $D$ as a linear combination of restricted Schur polynomials, a key ingredient is the identity
$$
{\rm Tr}(\tau \phi_{11}^{\otimes n_{11}}\phi_{12}^{\otimes n_{12}})
=\sum_{R,\{r\},\alpha\beta}
{d_R n_{11}!n_{12}!\over d_{r_{11}}d_{r_{12}}\, n!}
\chi_{R,\{ r\},\alpha\beta }(\tau) \chi_{R,\{ r\},\beta\alpha}
$$
where the sum over $R$ runs over all irreducible representations of $S_{n_{11}+n_{12}}$ and $\{ r\}$ is summed over all
irreducible representations of $S_{n_{11}} \times S_{n_{12}}$.
This identity is derived in \cite{Bhattacharyya:2008xy} in the context of $U(N)$ gauge theory and it applies without change to our description in terms of adjoints.
We are interested in operators with a bare dimension of order $N$.
We achieve this large dimension by taking $n_{12}$ order $N$ and $n_{11}$ order $\sqrt{N}$.
For these operator, not all terms in (\ref{newdil}) have the same size at large $N$.
The sizes of the different terms follow by noting that differentiating with respect to $\phi_{12}$ produces order $N$ terms
while differentiating with respect to $\phi_{11}$ produces order $\sqrt{N}$ terms. Consequently, in the first term of
(\ref{newdil}) the terms with $j=2$ dominate; the terms with $j=1$ are supressed by a relative factor of $\sqrt{N}$.
Apart from the leading term, we will also study this first subleading contribution in this work.
The second term in (\ref{newdil}) also contributes at the leading order.
The third and fourth terms in (\ref{newdil}) are subleading, supressed by ${1\over N}$ and will consequently not be considered further in our study.
It would not be consistent to evaluate these terms without also including the ${1\over N}$ correction to the leading terms.
Finally, it is useful to express our result in terms of operators normalized so that
\begin{equation}
\langle\hat{O}_{R,\{ r\},\alpha\beta} \hat{O}_{S,\{ s\},\gamma\delta}^\dagger \rangle=f_R \delta_{RS}\delta_{r_{11}s_{11}}\delta_{r_{12}s_{12}}\delta_{\alpha\gamma}\delta_{\beta\delta}
\label{norm}
\end{equation}
Clearly then
\bea
\hat{O}_{R,\{ r\},\alpha\beta} (\phi_{11},\phi_{12})=\sqrt{f_R {\rm hooks}_R\over {\rm hooks}_{r_{11}}{\rm hooks}_{r_{12}}} \chi_{R,\{ r\},\alpha\beta}(\phi_{11},\phi_{12})
\eea
The normalization in (\ref{norm}) has been chosen so that the leading contribution to the dilatation operator most
closely resembles the result obtained in \cite{DeComarmond:2010ie} for ${\cal N}=4$ super Yang-Mills theory.
Note that operators labeled by Young diagrams $R$ with different shapes, are not normalized in the same way.
Clearly, from (\ref{norm}) it follows that the ratio of their normalizations is given by the ratios of the factors of the boxes
that do not agree between the two labels.
For operators with a dimension of order $N$ and number of rows (or columns) of order 1, this ratio is always equal to 1 plus
${1\over N}$ corrections.
Putting these ingredients together, we find
\bea
D\hat{O}_{R,\{ r\},\alpha\beta} &&=\sum_{S,\{ s\}\gamma\delta}
\sqrt{f_S{\rm hooks}_S{\rm hooks}_{r_{11}}{\rm hooks}_{r_{12}}\over f_R{\rm hooks}_R{\rm hooks}_{s_{11}}{\rm hooks}_{s_{12}}}
M_{R,\{r\},\alpha\beta;S,\{s\},\gamma\delta}\hat{O}_{S,\{s\},\gamma\delta}\cr
&&\equiv \sum_{S,\{ s\}\gamma\delta}\, D_{R,\{ r\},\alpha\beta\, ;\, S,\{s\},\gamma\delta}\hat{O}_{S,\{s\},\gamma\delta}
\label{dilop1}\eea
where
\bea
& M_{R,\{r\},\alpha\beta;S,\{s\},\gamma\delta}=-\left( {4\pi\over k}\right)^2
\sum_{R'}{c_{RR'}d_S n_{11} n_{12}\over d_{s_{11}}d_{s_{12}}d_{R'}(n_{11}+n_{12})}\cr
&\times\Bigg[ (n_{12}-1){\rm Tr}_{R\oplus S} \bigg[I_{S'R'}(1,n_{11}+2)\big[ (1,n_{11}+1),
P_{R,\{r\}\alpha\beta}\big]I_{R'S'}\big[(1,n_{11}+1),P_{S,\{s\}\gamma\delta}\big]\bigg]\cr
&+(n_{11}-1) {\rm Tr} \bigg[I_{S'R'}(1,2)\big[ (1,n_{11}+1),P_{R,\{r\}\alpha\beta}\big]I_{R'S'}
\big[(1,n_{11}+1),P_{S,\{s\}\gamma\delta}\big]\bigg]\cr
&+ N {\rm Tr} \bigg[I_{S'R'}\big[ (1,n_{11}+1),P_{R,\{r\}\alpha\beta}\big]I_{R'S'}
\big[(1,n_{11}+1),P_{S,\{s\}\gamma\delta}\big]\bigg]\cr
&+ {\rm Tr} \bigg[I_{S'R'}\left(
P_{R,\{r\}\alpha\beta}-(1,n_{11}+1)P_{R,\{r\}\alpha\beta}(1,n_{11}+1)\right)
I_{R'S'}\big[(1,n_{11}+1),P_{S,\{s\}\gamma\delta}\big]\bigg]
\Bigg]\cr
\label{dilop2}
\eea
To obtain this result, the sum over the symmetric group appearing in (\ref{ourOps}) is evaluated using the fundamental
orthogonality theorem of group representation theory.
The sum that appears after the derivatives act is a sum over $S_{n_{11}+n_{12}-1}\subset S_{n_{11}+n_{12}}$, so that the sum is non-zero as long as one of the representations
suduced by $R$ upon restricting to $S_{n_{11}+n_{12}-1}$ agrees with one of the representations subduced by $S$
under the same restriction.
The sum then produces the maps $I_{S'R'}$ and $I_{R'S'}$ which map between subspaces of the carrier spaces of $R$ and $S$.
We have used cycle notation for elements of the symmetric group.
To completely spell out our notation, note that each element of the symmetric group is in the representation inherited from
the subspace it acts in.
Thus, for example,
\bea
{\rm Tr}_{R\oplus S} \bigg[I_{S'R'}(1,n_{11}+2)I_{R'S'}(1,n_{11}+1)\bigg]
={\rm Tr}_{R\oplus S} \bigg[I_{S'R'}\Gamma^{R}\left((1,n_{11}+2)\right) I_{R'S'}
\Gamma^{S}\left((1,n_{11}+1)\right)\bigg]\nonumber
\eea
where $\Gamma^{S}\left(\sigma\right)$ is the matrix representing $\sigma$ in irreducible representation $S$.
The formulas (\ref{dilop1}) and (\ref{dilop2}) are the key results of this section.
These are exact in the sense that we have not used any simplifications of the large $N$ limit to obtain this result.
We now consider the eigenproblem of $D$ which, as we explain in the next section, can be solved in a specific limit, after exploiting simplifications of large $N$.
At large $N$ the last line in (\ref{dilop2}) is subleading and will therefore be dropped in what follows\footnote{The last line in (\ref{dilop2}) corresponds to the
third and fourth terms in (\ref{newdil})}.
\section{Displaced Corners Approximation}
It is perhaps useful to begin with a discussion of some of the intricacies inherint in the problem of diagonalizing (\ref{dilop1}).
The key difficulty in constructing the restricted Schur polynomials (\ref{ourOps}) is in the construction of the intertwining operators $P_{R,\{r\},\alpha\beta}$.
To compute the two point function (\ref{twopoint}), after summing over the free field Wick contractions, we simply need to take a product of two of these intertwining operators and then compute their trace, which is a relativly simple computation.
Indeed, the result depends only on the dimensions of the representations $R$ and $\{ r\}$ which appear.
The expression in (\ref{dilop1}) involves computing commutators of the intertwining operators with symmetric group elements and then tracing over a product of these commutators.
This is a much more sophisticated operation for which the explicit form of $P_{R,\{r\},\alpha\beta}$ is required.
Fortunately there is a limit in which we can construct $P_{R,\{r\},\alpha\beta}$ in a straight forward way: this is the displaced corners limit of \cite{Koch:2011hb} (see also \cite{Carlson:2011hy}).
The idea is simply that for the vast majority of restricted Schur polynomials
$\chi_{R,\{ r\},\alpha\beta}(\phi_{11},\phi_{12})$ that can be written down, the distance
between the last box in each row of $R$ is order $N$.
Here by the distance between boxes $a$ and $b$ we mean the smallest number of boxes that one needs to pass through when moving, in the Young diagram, from box $a$ to box $b$.
When the distance between the last box in the different rows of $R$ is order $N$, the action of the symmetric group simplifies
dramatically, which greatly simplifies the construction of $P_{R,\{r\},\alpha\beta}$.
To guarantee this simplification it is necessary to assume in addition that $n_{12}\gg n_{11}$; for
further discussion and all the details see \cite{Koch:2011hb}.
In this article we accomplish $n_{12}\gg n_{11}$ by scaling $n_{12}$ as $N$ and $n_{11}$ as $\sqrt{N}$ as we take $N\to\infty$.
Our results would seem to hold with $n_{11}$ scaled as $N^{\alpha}$ with $\alpha<1$, but due to the formidable technical
computations needed, we have not managed to explore this important point in detail.
For a Young diagram $R$ with $p$ rows, the maps $I_{S'R'}$ and $I_{R'S'}$ can be identified with elements of $u(p)$.
The action of the symmetric group elements appearing in (\ref{dilop1}), on these maps, is easy to evaluate.
The intertwining operators themselves take a factorized form
\bea
P_{R,\{r\},\alpha\beta}=p_{r_{11}\alpha\beta}{\bf 1}_{r_{12}}
\eea
where $p_{r_{11}\alpha\beta}$ projects onto $S_{n_{11}}$ irreducible representation
$r_{11}$ and ${\bf 1}_{r_{12}}$ projects onto $S_{n_{12}}$ irreducible representation $r_{12}$.
The concrete construction of these intertwining operators, together with detailed examples, is given in \cite{Koch:2011hb}.
Since we have to take $n_{12}\gg n_{11}$ we know that the terms in (\ref{newdil}) with $j=2$ will dominate.
This is indeed the case: in (\ref{dilop2}) the terms with coefficient $n_{12}-1$ come from the $j=2$ term of (\ref{newdil}) while the terms with coefficient $n_{11}-1$ come from $j=1$.
In this section we will restrict our attention to large $N$, which implies that we should keep only the leading order in
${n_{11}\over n_{12}}$.
This amounts to keeping only the terms in (\ref{dilop2}) that have coefficient $n_{12}-1$ or coefficient $N$
\bea
D^{(0)}\hat{O}_{R,\{ r\},\alpha\beta} &&=\sum_{S,\{ s\}\gamma\delta}
\sqrt{f_S{\rm hooks}_S{\rm hooks}_{r_{11}}{\rm hooks}_{r_{12}}\over f_R{\rm hooks}_R{\rm hooks}_{s_{11}}{\rm hooks}_{s_{12}}}
M^{(0)}_{R,\{r\},\alpha\beta;S,\{s\},\gamma\delta}\hat{O}_{S,\{s\},\gamma\delta}\cr
&&\equiv \sum_{S,\{ s\}\gamma\delta}\, D^{(0)}_{R,\{ r\},\alpha\beta\, ;\, S,\{s\},\gamma\delta}\hat{O}_{S,\{s\},\gamma\delta}
\label{Ldilop1}
\eea
where
\bea
& M^{(0)}_{R,\{r\},\alpha\beta;S,\{s\},\gamma\delta}=-\left( {4\pi\over k}\right)^2\sum_{R'}{c_{RR'}d_S n_{11} n_{12}\over d_{s_{11}}d_{s_{12}}d_{R'}(n_{11}+n_{12})}\cr
&\times\Bigg[ (n_{12}-1){\rm Tr} \bigg[I_{S'R'}(1,n_{11}+2)\big[ (1,n_{11}+1),
P_{R,\{r\}\alpha\beta}\big]I_{R'S'}\big[(1,n_{11}+1),P_{S,\{s\}\gamma\delta}\big]\bigg]\cr
&+ N {\rm Tr} \bigg[I_{S'R'}\big[ (1,n_{11}+1),P_{R,\{r\}\alpha\beta}\big]I_{R'S'}
\big[(1,n_{11}+1),P_{S,\{s\}\gamma\delta}\big]\bigg]\Bigg]
\label{Ldilop2}
\eea
We will return to the term with coefficient $n_{11}-1$ in the next section.
In the displaced corners approximation, using the simplifcations just outlined, we obtain
\bea
D^{(0)}_{R,\{ r\},\alpha\beta\, ;\, S,\{s\},\gamma\delta}=
-\left( {4\pi\over k}\right)^2
\sqrt{f_S\over f_R}\sum_{R'}{c_{RR'}\over (n_{11}-1)!}(N+r_{12i}) \sqrt{{\rm hooks}_{r_{11}} {\rm hooks}_{s_{11}}}\times\cr
\Bigg[
{\rm Tr}(E^{(1)}_{kk}p_{r_{11}\alpha\beta}E^{(1)}_{ii}p_{s_{11}\gamma\delta})\delta_{r'_{12,i};s'_{12,k}}+
{\rm Tr}(E^{(1)}_{ii}p_{r_{11}\alpha\beta}E^{(1)}_{kk}p_{s_{11}\gamma\delta})\delta_{r'_{12,i};s'_{12,k}}\cr
-\Big({\rm Tr}(E^{(1)}_{kk}p_{r_{11}\alpha\delta})\delta_{\beta\gamma}+
{\rm Tr}(E^{(1)}_{kk}p_{r_{11}\gamma\beta})\delta_{\alpha\delta}\Big)\delta_{R;S}\delta_{r_{11};s_{11}}\delta_{r_{12};s_{12}}
\Bigg]
\eea
In this last formula, $r_{12i}$ is the length of row $i$ of Young diagram $r_{12}$, $R'$ is obtained from $R$ by dropping the last
box in row $i$ and $S'$ is obtained from $S$ by dropping the last box in row $k$.
$D^{(0)}_{R,\{ r\},\alpha\beta\, ;\, S,\{s\},\gamma\delta}$ is diagonalized by the double coset ansatz \cite{deMelloKoch:2012ck}.
To motivate what follows, recall that the label $\{ r\}=\{r_{11},r_{12}\}$ and that $r_{12}$ can be obtained by removing
a total of $n_{11}$ boxes from $R$.
Denote the number of rows in $R$ by $p$.
If we remove $a_1$ boxes from the first row, $a_2$ from the second and so on up to $a_p$ from row $p$, then the vector
$\vec{n}_{11}=(a_1,a_2,...,a_p)$ plays an important role: in the displaced corners approximation, operators with different
$\vec{n}_{11}$ do not mix at one loop \cite{Koch:2011hb}.
Of course, we have $a_1+a_2+\cdots+a_p = n_{11}$.
The vector $\vec{n}_{11}$ can be used to define a group $H$ which is a product of symmetric groups
\begin{equation}
H=S_{a_1}\times S_{a_2}\times \cdots \times S_{a_p}
\end{equation}
According to the double coset ansatz\cite{deMelloKoch:2012ck}, each eigenfunction of the dilatation operator is in
one-to-one correspondence with an element of the double coset $H\setminus S_{n_{11}}/ H$.
These double coset elements can also be put into correspondence with graphs whose edges are oriented and hence with
open strings states that obey the Gauss Law, providing a convincing connection with the dual D-brane plus open string
excited states; for background see \cite{Balasubramanian:2004nb,deMelloKoch:2012ck}.
The graph has a total of $p$ nodes and there are $n_{11}$ oriented edges stretching between the nodes.
For this reason we will refer to these operators as Gauss graph operators and to the associated oriented graphs as Gauss graphs.
The Gauss graph operators are\cite{deMelloKoch:2012ck}
\begin{equation}
O_{R,r_{12}}(\sigma)={|H|\over\sqrt{n_{11}!}}\sum_{j,k}\sum_{r_{11}\vdash n_{11}}\sum_{\mu_1,\mu_2}\sqrt{d_{r_{11}}}
\Gamma^{(r_{11})}_{jk}(\sigma)B^{r_{11}\to 1_H}_{j\mu_1}B^{r_{11}\to 1_H}_{k\mu_2}
\hat{O}_{R,\{ r\},\mu_1\mu_2}
\label{GGops}
\end{equation}
where $\sigma\in H\setminus S_{n_{11}}/ H$, $\Gamma^{(r_{11})}_{jk}(\sigma)$ is the matrix representing $\sigma$ in the
irreducible representation $r_{11}$ of $S_{n_{11}}$ and the branching coefficients $B^{r_{11}\to 1_H}_{j\mu_1}$ resolve
the projector from irreducible representation $r_{11}$ of $S_{n_{11}}$ to the trivial representation of $H$
\begin{equation}
{1\over |H|}\sum_{\gamma\in H}\Gamma^{(r_{11})}_{jk}(\sigma)=\sum_\mu B^{r_{11}\to 1_H}_{j\mu}B^{r_{11}\to 1_H}_{k\mu}
\end{equation}
Note that these operators are not normalized.
We have computed the norm of these operators in the Appendix.
The action of the dilatation operator is most easily written in terms of parameters read from the Gauss graphs.
Following \cite{deMelloKoch:2011uq}, a useful combinatoric description of a Gauss graph is obtained by dividing
each string into two halves with a label for each half.
Using the orientation of the string, label both the outgoing and the ingoing string endpoints with an integer $1,2,\cdots,n_{11}$.
A permutation is then determined by how the halves are joined and conversely, given a permutation, we can reconstruct the graph.
A graph is not associated to a unique permutation because the strings leaving the $i$'th node are indistinguishable,
and the strings arriving at the $i$'th node are indistinguishable.
As a result, graphs are in one-to-one correspondence with elements of the double coset $H\setminus S_{n_{11}}/ H$.
Divide the integers $1,2,\cdots,n_{11}$ into $p$ sets, ${\cal S}_i$ $i=1,2,\cdots,p$ such that the symmetric group that is the
i$^{th}$ factor in $H$ permutes the elements of ${\cal S}_i$.
In the graph corresponding to $\sigma$, the number of oriented edges stretching from node $i$ to node $j$ is
\begin{equation}
n_{ij}^+(\sigma)=\sum_{k\in{\cal S}_i}\sum_{l\in{\cal S}_j}\delta (\sigma (k),l)
\end{equation}
The number of strings stretching in the opposite direction, between the same two nodes, is
\begin{equation}
n_{ij}^-(\sigma)=\sum_{k\in{\cal S}_i}\sum_{l\in{\cal S}_j}\delta (\sigma (l),k)
\end{equation}
The total number of strings stretching between the two nodes is $n_{ij}(\sigma)=n_{ij}^+(\sigma)+n_{ij}^-(\sigma)$.
The action of the dilatation operator is naturally written in terms of an operator $\Delta_{ij}$ defined as follows:
$\Delta_{ij}$ is a sum of three terms
\begin{equation}
\Delta_{ij}=\Delta^+_{ij}+\Delta^0_{ij}+\Delta^-_{ij}
\end{equation}
To define the action of each of the above terms, we need to introduce two new Young diagrams,
$(r_{12})_{ij}^{\pm}$: $(r_{12})_{ij}^+$ is the Young
diagram obtained from $r_{12}$ by removing the last box from row $j$ and adding it to the end of
row $i$, while $(r_{12})_{ij}^-$ is the Young
diagram obtained from $r_{12}$ by removing the last box from row $i$ and adding to the end of row $j$.
$R^\pm_{ij}$ are defined in the same way.
The actions we need to define are\footnote{The $O(1)$ corrections added to $N$ in the expressions which follow
must be retained. After cancelations, these terms give the leading contribution.}
\begin{eqnarray}
&&\Delta^0_{ij}O_{R,r_{12}}(\sigma)=-(2N+r_{12i}+r_{12j}-3)O_{R,r_{12}}(\sigma)\cr
&&\Delta^+_{ij}O_{R,r_{12}}(\sigma)=\sqrt{(N+r_{12i}-1)(N+r_{12j}-1)}O_{R^+_{ij},(r_{12})^+_{ij}}(\sigma)\cr
&&\Delta^-_{ij}O_{R,r_{12}}(\sigma)=\sqrt{(N+r_{12i})(N+r_{12j}+2)}O_{R^-_{ij},(r_{12})^-_{ij}}(\sigma)
\end{eqnarray}
Recall that $r_{12k}$ is the number of boxes in row $k$ of Young diagram $r_{12}$.
A computation very similar to that of \cite{deMelloKoch:2012ck} now shows
\begin{equation}
D^{(0)}O_{R,r_{12}}(\sigma_1)=-\left( {4\pi\over k}\right)^2\sum_{\gamma_1,\gamma_2\in H}
\delta (\gamma\sigma_1\gamma^{-1}\sigma_2^{-1})\sum_{i<j}(N+r_{12,i})n_{ij}(\sigma_1)\Delta_{ij}O_{R,r_{12}}(\sigma_2)
\label{LeadDil}
\end{equation}
In the large $N$ limit we can introduce continuous variables $x_i$ defined by
\begin{equation}
x_i ={r_{12,i}-r_{12,p}\over \sqrt{N+r_{12,p}}}
\end{equation}
In terms of this continuous variable, the leading contribution to the action of the dilatation operator (\ref{LeadDil}) becomes
\bea
D^{(0)}O_{R,r_{12}}(\sigma_1)&=&-\left( {4\pi\over k}\right)^2\sum_{\gamma_1,\gamma_2\in H}
\delta (\gamma\sigma_1\gamma^{-1}\sigma_2^{-1})\cr
&\times&\sum_{i<j}(N+r_{12,i})n_{ij}(\sigma_1)
\left(\left({d\over dx_i}-{d\over dx_j}\right)^2-{(x_i-x_j)^2\over 4}\right)
O_{R,r_{12}}(\sigma_2)\cr
&&\label{fdil}
\eea
After diagonalizing $n_{ij}(\sigma)$ this is a sum of decoupled oscillators, which is an integrable system.
\section{Subleading term}\label{LstRSP}
In this section we will consider the subleading correction contained in
\bea
D^{(1)}\hat{O}_{R,\{ r\},\alpha\beta} &&=\sum_{S,\{ s\}\gamma\delta}
\sqrt{f_S{\rm hooks}_S{\rm hooks}_{r_{11}}{\rm hooks}_{r_{12}}\over f_R{\rm hooks}_R{\rm hooks}_{s_{11}}{\rm hooks}_{s_{12}}}
M^{(1)}_{R,\{r\},\alpha\beta;S,\{s\},\gamma\delta}\hat{O}_{S,\{s\},\gamma\delta}\cr
&&\equiv \sum_{S,\{ s\}\gamma\delta}\, D^{(1)}_{R,\{ r\},\alpha\beta\, ;\, S,\{s\},\gamma\delta}\hat{O}_{S,\{s\},\gamma\delta}
\label{Ldilop3}
\eea
where
\bea
& M^{(0)}_{R,\{r\},\alpha\beta;S,\{s\},\gamma\delta}=-\left( {4\pi\over k}\right)^2\sum_{R'}{c_{RR'}d_S n_{11} n_{12}\over d_{s_{11}}d_{s_{12}}d_{R'}(n_{11}+n_{12})}\cr
&\times (n_{11}-1) {\rm Tr} \bigg[I_{S'R'}(1,2)\big[ (1,n_{11}+1),P_{R,\{r\}\alpha\beta}\big]
I_{R'S'}\big[(1,n_{11}+1),P_{S,\{s\}\gamma\delta}\big]\bigg]
\label{Ldilop4}
\eea
These terms correspond to the terms with $j=1$ in (\ref{newdil}).
Evaluating the above trace in the displaced corners approximation, we find
\bea
D^{(1)}_{R,\{ r\},\alpha\beta\, ;\, S,\{s\},\gamma\delta}=
-\left( {4\pi\over k}\right)^2
\sqrt{f_S\over f_R}\sum_{R'}{c_{RR'}\over (n_{11}-2)!} \sqrt{{\rm hooks}_{r_{11}} {\rm hooks}_{s_{11}}}\times\cr
\Bigg[\sqrt{r_{12b}\over r_{12k}}
{\rm Tr}(E^{(1)}_{kk}E^{(2)}_{bi}p_{r_{11}\alpha\beta}E^{(1)}_{ib}p_{s_{11}\gamma\delta})\delta_{r'_{12,b};s'_{12,k}}+
{\rm Tr}(E^{(1)}_{id}E^{(2)}_{id}p_{r_{11}\alpha\beta}E^{(1)}_{kk}p_{s_{11}\gamma\delta})\delta_{r'_{12,i};s'_{12,k}}\cr
-\Big({\rm Tr}(E^{(1)}_{kb}E^{(2)}_{bk}p_{r_{11}\alpha\delta})
\delta_{ik}\delta_{r_{11}s_{11}}\delta_{\beta\gamma}\delta_{R;S}+
\sqrt{r_{12k}\over r_{12i}} {\rm Tr}(E^{(2)}_{ki}p_{r_{11}\gamma\beta}E^{(1)}_{ik}p_{s_{11}\gamma\delta})\Big)
\delta_{r_{12};s_{12}}
\Bigg]
\eea
We have not managed to perform the sums needed to rewrite the action of $D^{(1)}$ on Gauss graph operators.
It is however straight forward to study this problem numerically, for specific choices of $n_{11}$ and $p$.
The numerical study we will discuss is focused on operators labeled by Young diagrams $R$ that have a
total of $p=3$ long rows, and $n_{11}=3$.
The results of this example are rather typical.
A total of 21 operators can be defined, so that the dilatation operator is a $21\times 21$ dimensional matrix.
Acting on this space, $D^{(0)}$ decomposes into a block diagonal matrix with a total of 10 blocks.
Each block can be labeled by the vector $\vec{n}_{11}$. The possible blocks together with their dimension and
allowed $s$ labels are
\bea
&&\vec{n}_{11}=(1,1,1)\qquad d=6\qquad s={\tiny \yng(3)\quad \yng(2,1)\quad \yng(1,1,1)}\cr
&&\vec{n}_{11}=(2,1,0)\qquad d=2\qquad s={\tiny \yng(3)\quad \yng(2,1)}\cr
&&\vec{n}_{11}=(2,0,1)\qquad d=2\qquad s={\tiny \yng(3)\quad \yng(2,1)}\cr
&&\vec{n}_{11}=(0,2,1)\qquad d=2\qquad s={\tiny \yng(3)\quad \yng(2,1)}\cr
&&\vec{n}_{11}=(1,2,0)\qquad d=2\qquad s={\tiny \yng(3)\quad \yng(2,1)}\cr
&&\vec{n}_{11}=(0,1,2)\qquad d=2\qquad s={\tiny \yng(3)\quad \yng(2,1)}\cr
&&\vec{n}_{11}=(1,0,2)\qquad d=2\qquad s={\tiny \yng(3)\quad \yng(2,1)}\cr
&&\vec{n}_{11}=(3,0,0)\qquad d=1\qquad s={\tiny \yng(3)}\cr
&&\vec{n}_{11}=(0,3,0)\qquad d=1\qquad s={\tiny \yng(3)}\cr
&&\vec{n}_{11}=(0,0,3)\qquad d=1\qquad s={\tiny \yng(3)}
\eea
It is a simple exercise to write down the complete set of partialy labeled Young diagrams\cite{Koch:2011hb} and write
down the action of the symmetric group on these states.
We need to explicitely consider all 3 $\phi_{11}$-boxes as well as a single $\phi_{12}$ box when constructing the dilatation
operator numerically.
Within this space, the projectors $p_{r_{11}\gamma\beta}$ are $81\times 81$ dimensional matrices.
The only representation that carries a nontrivial multiplicity label is the $s={\tiny \yng(2,1)}$ representation in the
$\vec{n}_{11}=(1,1,1)$ subpace.
The multiplicity free projectors can immediately be written down as
\bea
p_{r_{11}\,\vec{n}_{11}}={d_{r_{11}}\over 3!}\sum_{\sigma\in S_3}\chi_{r_{11}}(\sigma)\Gamma^{\vec{n}_{11}}(\sigma)
\eea
with $\chi_{r_{11}}(\sigma)$ an $S_3$ character.
The matrix $\Gamma^{\vec{n}_{11}}(\sigma)$ represent $\sigma\in S_3$, in the displaced corners approximation and inside
the $\vec{n}_{11}$ subspace.
To construct the projectors for the $s={\tiny \yng(2,1)}$ representation in the $\vec{n}_{11}=(1,1,1)$ subpace, we need
to resolve this subspace into two $U(3)$ states in the ${\tiny \yng(2,1)}$ representation.
The two states are described by the Gelfand-Tsetlin patterns that have the same inner multiplicity.
For our problem here, the two states are
\bea
\left[
\begin {array}{ccccc}
2 & & 1 & & 0 \\\noalign{\medskip}
& 1 & & 1 & \\\noalign{\medskip}
& & 1 & &
\end {array} \right]\qquad\qquad
\left[
\begin {array}{ccccc}
2 & & 1 & & 0 \\\noalign{\medskip}
& 2 & & 0 & \\\noalign{\medskip}
& & 1 & &
\end {array} \right]
\eea
and are easily constructed using $U(3)$ Clebsch-Gordan coefficients.
The detailed computation appears in Appendix C of \cite{Koch:2011hb}.
We find that $D^{(1)}$ not diagonal in the Gauss graph basis and it does not commute with $D^{(0)}$.
Further, it does not reduce to a block diagonal matrix and indeed, it mixes operators from different $\vec{n}_{11}$ sectors.
This mixing is expected and has a natural interpretation in the gravity dual.
Specifying $\vec{n}_{11}$ specifies how many oriented edges start and terminate at each node.
Interpreting the nodes as giant gravitons and the oriented edges as open strings attached to the giant graviton system,
$\vec{n}_{11}$ can only change as a result of open string splitting and joining.
Thus, the mixing we see is a signal of open string splitting and joining.
This interpretation is also natural given the fact that $D^{(1)}$ is a correction to the large $N$ limit, so that
we should indeed be seeing the first effects of string splitting and joining when this correction is included.
Finally, a remarkable feature of $D^{(0)}$ is the appearance of the integers $n_{ij}(\sigma)$ when the diagonalization
problem is solved.
Numerically we find that the eigenvalues of $D^{(1)}$ are again integers suggesting there may be a nice combinatorial
description of the problem, presumably exploiting the combinatorics of string splitting and joining.
\section{Discussion}\label{discussion}
In the $SU(2)$ sector of the ABJM theory we have managed to diagonalize the two loop dilatation operator
by employing the double coset ansatz.
This problem was already considered in \cite{deMelloKoch:2012kv} where the dilatation operator was already evaluated,
but not diagonalized.
One of the results we have reported, is precisely the solution of this diagonalization problem.
The main progress achieved in this article follows from our rewriting of the dilatation operator, in terms of adjoint variables.
This gives a useful organization of the dilatation operator and in particular, has allowed us to cleanly identify two terms that
contribute at the leading order at large $N$ and two that are subleading.
With this organization in hand, the eigenproblem of the dilatation operator is a straight forward exercise that can be achieved
using existing techniques.
The leading terms are diagonalized bythe double coset ansatz, reducing the problem to the diagonalization of a collection of
decoupled oscillators, which is an integrable system.
We find a new ``conservation law'': the dilation operator does not mix operators with different $\vec{n}_{11}$ quantum number.
The resulting spectrum of anomalous dimensions differs from the corresponding spectrum in ${\cal N}=4$
super Yang-Mills theory in an important quantitative way.
In the ${\cal N}=4$ super Yang-Mills theory, the frequencies of the decoupled oscillators are set by the eigenvalues of
the matrix $n_{ij}(\sigma)$ which can be read straight from the permutation labeling the Gauss graph.
From (\ref{fdil}) we see that for ABJM the frequencies of the decoupled oscillators are set by the eigenvalues of
$(1+{r_{12,i}\over N})n_{ij}(\sigma)$.
Thus, the frequencies depend both on the matrix $n_{ij}(\sigma)$, determined by the Gauss graph, and on $r_{12,i}$
which are the row lengths of the Young diagram $r_{12}$.
Each row of $r_{12}$ corresponds to a giant graviton.
The number of boxes in the $i^{\rm th}$ row of $r_{12}$ determines an ${\cal R}$- charge which corresponds to the
angular momentum of the giant graviton.
Since the giant expands to a definite size by balancing a Lorentz type force (trying to expand the giant) with tension (trying to
shrink the giant), the angular momentum of the giant sets the size of the giant.
Consequently, our result implies that the excitation spectrum of the giant graviton picks up a dependence on the size
of the giant graviton.
The fact that the spectrum of the anomalous dimensions in ${\cal N}=4$ super Yang-Mills theory is independent of the
parameters of the Young diagram associated to the giant graviton system, matches the fact that the spectrum of small
fluctuations around the giant is independent of the size of the giant\cite{Das:2000fu}.
This independence of the size of the giant is understood as follows\cite{Balasubramanian:2004nb}: as the radius of the giant
increases, there is an increase in the energy of fluctuations due to blue-shifting, as well as a decrease in the energy of the states
because the fluctuations now move on a bigger sphere.
These two effects precisely cancel producing a size independent spectrum.
For the ABJM case, our results predict that although these two effects still operate, they do not precisely cancel so that
the spectrum does pick up a dependence on the size of the giant.
This is consistent with the small fluctuation spectrum around a giant graviton performed in \cite{Giovannoni:2011pn}.
By perturbing around the near-maximal giant and the ``small" giant these authors find a spectrum that is size-dependent.
In this article we have also given a simple formula for the normalization of the Gauss Graph operators.
This will be a useful technical input when computing the effects of Gauss Graph operator mixing, at subleading orders in a
large $N$ expansion.
Finally, we have also evaluated the largest of the subleading (in ${1\over N}$) terms.
Although we have not managed an analytic result, a numerical study has lead to some interesting conclusions.
The subleading correction does not commute with the leading order dilatation operator.
Further, it allows mixing between operators with different $\vec{n}_{11}$ quantum numbers, so that
it spoils the conservation law that was present at large $N$.
This is naturally interpreted as a consequence of open string splitting and joining.
The discusion of \cite{Kristjansen:2008ib,Caputa:2009ug} suggests that the failure of this conservation law may
be an indication that integrability does not persist beyond the large $N$ limit.
A numerical diagonalization of this term shows that it has integer eigenvalues, suggesting that there may be a nice
combinatorial description waiting to be developed.
{\vskip 0.2cm}
\noindent
{\it Acknowledgements:}
We thank Jeff Murugan for helpful correspondence.
RdMK and RK are supported by the South African Research Chairs
Initiative of the Department of Science and Technology and the National Research Foundation.
SS is supported by the National Institute for Theoretical Physics.
\begin{appendix}
\section{Normalization of the Gauss Graph Operators}
The two point function of Gauss Graph Operators is
\bea
\langle O_{R,r}(\sigma)^\dagger O_{R,r}(\sigma)\rangle =\sum_{\gamma_1,\gamma_2\in H}
\delta (\sigma^{-1}\gamma_1\sigma\gamma_2^{-1})
\eea
The right hand side of the above equation is simply counting the number of solutions $\gamma_1,\gamma_2\in H$ to
\bea
\sigma = \gamma_1\sigma\gamma_2^{-1}\label{sc}
\eea
Using $\gamma_1$ and $\gamma_2$ we are able to swap the endpoints of the open strings.
If we swap the labels of strings that have the same start and endpoints, we leave $\sigma$ unchanged
and hence have a solution to (\ref{sc}).
In this way, for $n$ strings stretching from the same start point to the same endpoint, we will pick up a factor of $n!$.
Denote the number of oriented line segments stretching from node $i$ to node $j$ by $n_{ij}$ and the number of
segments stretching from node $i$ back to node $i$ by $n_{ii}$.
We have
\bea
\langle O_{R,r}(\sigma)^\dagger O_{R,r}(\sigma)\rangle =\prod_{i=1}^p n_{ii}!\prod_{k,l=1,l\ne k}^p n_{kl}!
\eea
\end{appendix}
|
\section{Introduction}
The recently introduced model of \emph{Broadcasting Automata}\index{broadcasting automata} \cite{Geometric} connects many of the techniques contained in distributed
algorithms, ad-hoc radio networks, cellular automata and neighbourhood sequences. Much like cellular
automata with variably defined neighbourhoods Broadcasting Automata can be
defined on some form of grid or lattice structure and have a simple computational
primitives comparative to a finite state automata with the ability to receive
and send messages both from and to those automata which are within its transmission
radius. Neighbourhoods are defined in the same way as with an ad-hoc
network, all those points within a certain transmission radius, receive the message
from the sender.
Algorithms for the broadcasting automata model are in the same vain as those encountered in
distributed algorithms using a simple notion of waves, messages passed from automata
to automata throughout the topology, to construct computations \cite{Geometric2}. Wave
algorithms are enhanced further with notions of composition of the information
that is carried within each wave borrowed from the physical world and embellished
with the new found computational power of the automata.
The waves generated by activating processes in a digital environment can be used
for designing a variety of wave algorithms. In fact, even very simple finite functions for
the transformation and analysis of passing information provides more complex dynamics
than classical wave effects. In \cite{Geometric,Geometric2} we generalized the notion of the standing wave which is
a powerful tool for partitioning a cluster of robots on a non-oriented grid. In contrast
to classical waves where interference patterns generate nodal lines (i.e. lines
formed by points with constant values), an automata network can have more complex
patterns which are generated by periodic sequences of states in time.
Here we take a different direction and aim to study the geometrical shapes of
informational waves on integer grid generated in broadcasting automata model as well as their potential
use for metric approximation in a discrete space.
The repeated transmission to nodes, within certain radii, applied to
a certain network topology, or physical layout of nodes has been studied under the name, \emph{Neighbourhood Sequences}
(NS).
The concept of neighbourhood sequences is of importance in a number of practical applications and was originally applied for measuring distances in a digital world \cite{Hajdu20032597}.
Initially two classical digital motions (cityblock and chessboard)\footnote{ The cityblock motion allows movements only in horizontal and vertical directions, while the chessboard allows to move in diagonal directions.} were introduced. Based on these two types of motions
periodic neighbourhood sequences were defined in \cite{Das} by allowing arbitrary mixture of cityblock and chessboard motions. In 2D the distances based on cityblock and chessboard neighbourhood sequences deviate quite substantially from the ideal Euclidian distances, so instead their combination that form ``the octagon'' was more often employed and studied. Later the concept of neighbourhood sequences was extended to arbitrary finite and infinite dimensions, periodic and non-periodic sequences and then analysed in terms of their geometric properties.
The aggregation of two classical neighbourhood sequences based on Moore and Von Neumann neighbourhoods (which correspond to cityblock and chessboard) was recently proposed as an alternative method for self-organization, partitioning and pattern formation on the non-oriented grid environment in \cite{Geometric}. In particular the discrete analogs of physical standing wave phenomena were proposed to generate nodal patterns in the discrete environment by two neighbourhood sequences. The power of the primitives was illustrated by giving distributed algorithms for the problem of finding the centre of a digital disk of broadcasting automata.
The shapes that can be formed by neighbourhood sequences in dimension two are quite limited. However the basic notion of neighbourhood sequences can be naturally extended by relaxing the constraints on the initial definition of the neighbourhood in such a way that two points are neighbors (r-neighbours) if the Euclidean distance is less than or equal to
some $r$, used to denote the radius of a circle. Then by \emph{Broadcasting sequences} we understand the periodic application of the $r$-neighbours distance function.
The main result of this work is characterization of geometrical shapes that can be generated by Broadcasting Sequences on the square lattice. The shapes of r- neighborhoods correspond to Discrete Discs which are discrete convex polygons.
First we use the language of Chain codes (i.e. the code based on $8$ degrees of motion) to describe the shape of the Discrete Discs and their Broadcasting sequences. In particular we introduce the notion of Chain code segments and Line segments to express the shapes of the polygons corresponding to Discrete Discs. Then we characterize the shapes of polygons produced by Broadcasting sequences and provide linear time algorithm for the composition of two chain codes of Discrete Discs. Based on their composition properties we derive a number of limitations for produced polygons. For example we show that there exist an infinite number of gradients (of line segments in the polygons) that cannot be produced by Broadcasting Sequences. It also becomes clear that the set of line segments, and as such gradients, that compose any discrete circle are closed under composition.
Moreover, we provide an alternative method for enriching the set of geometrical shapes and neighbourhood sequences by aggregation of two Broadcasting Sequences. Initially we illustrate the idea on Moire and Anti-Moire aggregation function to produce an infinite family of polygons and polygonal shapes and characterize the gradients of their line segments. We have noticed that Anti-Moire aggregation function can be slightly modified to provide better approximation for the Euclidean distance on a square lattice then classical neighbourhood sequences.
Finally it is possible to observe the variety of effects that are the result of the application of an aggregation function
which are themselves shapes of some form.
\section{Broadcasting Automata Model}
\tn{
One of the fundamental models of computation is the automaton.
Finite state automata take as input a word, which may in some instances be referred to as a tape,
and output is limited to an accepting state which the automata is left in if it is said to accept the word. Traditionally automata used in cases where it is important to transform some input in to an output, beyond the use of a single state, is the use of Moore Machines. Such machines differ from finite state machines in that they are able to produce an output word from an input word.
\begin{definition}
A Moore machine\index{Moore machine} \cite{conway2012regular} is a 6-tuple, $A = (Q,\Sigma,\Lambda,\delta,\Delta,q_0)\ $, where:
\begin{itemize}
\item $Q$ is a finite set of states,
\item $\Sigma$ is the set of input symbols,
\item $\Lambda$ the set of output symbols,
\item $\delta:Q\times \Sigma \rightarrow Q$ is the transition function mapping a state $q\in Q$ and a symbol, or set of symbols, $\sigma \in \Sigma$ to a state $q\in Q$,
\item $\Delta:Q\rightarrow \Lambda$ is the output function which maps a state, $q\in Q$, to an output symbol, $\lambda \in \Lambda$, and
\item $q_0$ is the initial or quiescent state in which the automata starts.
\end{itemize}
\end{definition}
It is assumed that such a machine is connected to an input tape, or word, and an output tape, or word, where the result of the computation is read. In a situation, as is presented in distributed systems, whereby automata are connected to each other it is possible for the output of one automata to become the input of another automata. Such a model is known as a network of automata and connections from one automatons output to another's input may be represented as a directed graph, where direction represents the output going to input from automaton to automaton.
\begin{definition}
A network of finite automata\index{network of automata} is a triple, $(G, A, C_0)\ $, where:
\begin{itemize}
\item $G=(V,E)$ is a directed graph, with vertices, $V$, and edges, $E$, which are ordered pairs of vertices,
\item $A$, is a Moore machine, and
\item $C_0$ is an initial configuration which maps states, $Q$, of the automata, $A$, to vertices, $V$, such that, $C_0 : V \rightarrow Q$.
\end{itemize}
\end{definition}
In this model the topology is fixed as specified by the construction of the graph, $G$, which dictates the flow of inputs and outputs from the Moore machines. Such symbols, where a symbol is part of the input/output word of the Moore machine, are generated as response to some input, by an automaton at vertex, $v\in V$, and then sent to all of the adjacent vertices in the graph or to a particular adjacent vertex, $G$, where the automata that receive the symbols process them as they would their input word.
A less abstract model is considered by adding more details about the communication between automata,
where they are located in Euclidean space and what can be transmitted.
It will also be required to define more refined models that will reflect the new features and
constraints of the physical environment, but are still at a high level of abstraction.
Taking the inspiration from ad hoc wireless communication networks we introduced in \cite{Geometric}
a new model of \emph{Broadcasting Automata}\index{broadcasting automata},
which can be seen as a network of finite automata with a dynamic network topology.
Informally speaking, the model of Broadcasting Automata comprises nodes, which correspond to points in space, and connectivity between nodes depends upon the distances between the points and the
strengths of the transmissions generated by the automata as may be seen in ad-hoc
radio networks.
Transmission strength may vary from round to round for any particular automaton in the space and is dictated by the state of the automaton. In this model the topology, or connectivity graph, of the network of automata is able to change at each time step based on the states of the automata.
In order to give a formal definition of the Broadcasting Automata model on a metric space\index{metric space} it is first necessary to introduce the notion of a metric space and to modify the classical notion of the Moore machine.
\begin{definition}
A metric space is an ordered pair, $(M,d)$, where $M$ is a set and $d$ is a metric on $M$ such that $d: M\times M\rightarrow \mathbb R$.
\end{definition}
Here, $(M,d)$, is any metric space, later the two dimensional euclidean space shall be used,
but this is not a necessity simply that a notion of the distance between two points is required.
\begin{definition}
The \emph{Broadcasting Automaton} extends the Moore Machine by introducing a set of final states, $F$, along with a function, $\tau : Q \rightarrow \mathbb R$, which maps the state to a real number and represents the radius of transmission for the output symbol and the output alphabet, $\Lambda$, is extended by adding an empty symbol, $\epsilon$ and it is represented by an 8-tuple $A = (Q,\Sigma,\Lambda,\delta,\Delta,\tau,q_0,F) $.
\end{definition}
A network of \emph{Broadcasting Automata}, which will also be referred to as the \emph{Broadcasting Automata model}, can now be defined.
\begin{definition}
The \emph{Broadcasting Automata model} is represented by a triple $BA = ((M,d),A,C_0)$ where:
\begin{itemize}
\item $(M,d)$ is a metric space,
\item $A$ is a Broadcasting Automata, $A = (Q,\Sigma,\Lambda,\delta,\Delta,\tau,q_0,F)$,
\item $C_0$ is the initial configuration of the Broadcasting Automata model, it is a mapping from points, $M$, to states of the Broadcasting Automata, $Q$, such that, $C_0 : M \rightarrow Q$.
\end{itemize}
\end{definition}
In some cases it may be that the input and output symbols are drawn from the same alphabet in which case, $\Sigma = \Lambda$.
The communication between automata is organised by message passing, where \emph{messages} are symbols from the output alphabet, $\Lambda$, of the automata, $A$, to all of the automata within its transmission radius. Messages, \tn{symbols from the output alphabet, $\Lambda$, of the automata, $A$,} are generated and passed instantaneously at discrete time steps, \tn{generation of message is given by the function, $\Delta$, for the automata, $A$, resulting in synchronous steps}. Those automata that have received a message, \tn{for the first time in the computation}, are said to be \emph{activated}.
If several messages are transmitted to an automaton, $A$, it will receive only a set of {\bf unique messages}, i.e. for any multiset of transmitting messages, where the multiset represents a number of the same message being sent, received by $A$, over some number of rounds, the information about quantity of each type, within a single round, will be lost. This simply illustrates that the automaton may not dictate a number
of the same symbol in any transition, all transitions must operate upon a set of
distinct symbols.
More formally the concept of computation and communication is captured in the concept of configurations of the Broadcasting Automata model and its semantics presented here.
\begin{definition}\label{config}
The \emph{configuration} of the Broadcasting Automata is given by the mapping, $c:M\rightarrow Q$, from points in the metric space to states. It can be noted here that $C_0$ is an initial configuration for the Broadcasting Automata model.
\end{definition}
}
\tn{Automata are updated, from configuration to configuration, synchronously at discrete time steps. The next state of each automaton depends upon the states of all other automata which have in their neighbourhood the automaton that is to be updated. Where here the neighbourhoods are formed by a combination of their position in $M$ and the range of their transmission, dictated by $\tau(q)$ for all automata.
\begin{definition}\label{messageRec}
The set of messages received by the automaton at a point, $u,v \in M$, is expressed by the set $\Gamma_u= \{ \Delta(c(v)) | v\in M \wedge d(u,v)\leq \tau(c(v))\}$ for the metric space, $(M,d)$.
\end{definition}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.33]{transRadGraph0.eps}
\includegraphics[scale=0.33]{transRadGraph1.eps}
\caption{The above figures show the possible evolution over time of a network of Broadcasting Automata\index{network of automata}. Each figure shows, on the left, the broadcast range for the automata, depicted by a dotted circle with the node at the centre, on the right, the same connectivity is shown in graph form. \label{transGraph}}
\end{figure}
In Figure~\ref{transGraph}, it is possible to see an elucidation of message passing in the Broadcasting Automata model through the process of constructing a digraph representation of the broadcasting radii. The figure depicts the automatons broadcasting radius by way of a dashed circle with the automaton broadcasting at that range shown in the centre. The corresponding graph may be constructed from this by making a directed edge in the graph from the node that is broadcasting to all nodes that are within the broadcast range, where the range is dictated by the state, $q\in Q$ and the function, $\tau(q)=r$, where $r$ is the broadcasting radius. The construction of such a graph shows both the connection to the network automata model, which operates on a similar premise with a fixed graph, but also highlights how Definition~\ref{messageRec} is constructed. Incoming edges are generated by connecting nodes, with the arrow from transmitter to node, reachable from the transmitter. Naturally such a graph, in the Broadcasting Automata model, can change over time which is shown by images further down the page as increase time steps in Figure~\ref{transGraph}.
It should be noted that this set of messages can be empty. The new state of the automaton at $u$ is now given by applying the automaton's transition function $\delta_{u}(\Gamma_{u})$. This may be applied to all of the automata in the BA and as such determines the global dynamics of the system. It can be seen that in one time step it is possible for configuration $c$ to become configuration $e$ where, for all $v\in M$, $e(v)=\delta_{v}(\Gamma_u)$. Such a function is called the global transition function. The { set of all configurations for BA} is denoted as $\mathcal{C}$.
\begin{definition}\label{transition}
The global transition function, $\mathcal{G}:\mathcal{C}\rightarrow \mathcal{C}$, represents the transition of the system at discrete time steps such that for two configurations, $c\in \mathcal{C}$ and $e\in \mathcal{C}$, where $e$ is the configuration the time step directly after $c$ such that, $e = \mathcal{G}(c)$.
\end{definition}
\begin{definition}
A computation is defined as any series of applications of the function, $\mathcal{G}$, from an initial condition, $C_0$, that leads to the set of all automata in the system being in one of their accepting states from the set, $F$, or the computation halts such that there exists no defined relation from the current configuration, $c$, under the global transition function $G(c)$.
\end{definition}
Here a notion of reachability is given.
\begin{definition}
One configuration, $c$, is said to be {\bf reachable} from another, $e$, if from the initial configuration, $C_0$, there exists a series of valid intermediary configurations such that $c\rightarrow c'\rightarrow c''...\rightarrow e$. Where $c\rightarrow c'$ is equivalent to $\mathcal{G}(c) = c'$. It can also be stated, in a shorthand way, that where the sequence $c\rightarrow c'\rightarrow c''...\rightarrow e$ exists reachability from $c$ to $e$ may be shown as $c\leadsto e$.
\end{definition}
In the above context the \emph{time complexity} of an algorithm in the Broadcasting Automata model is determined by the number of applications of the global transition function during the computation.
}
Whilst in general Broadcasting Automata may be used represent any of the common network topologies (such as, linear array, ring, star, tree, near-neighbour mesh, systolic array, completely connected, chordal ring, 3-cubes and hyper-cubes \cite{44900, 1667197}), through varying the location and transmission radii of the automata in the space, a different model is employed.
The model used throughout the paper is a hybrid mixing the notion of transmission radius in the pathloss geometric random graph model and common network topologies. The method involves the physical placement of nodes on the Euclidean plane in a lattice structure such that the distances between points conform to the euclidean distance. Such regular lattices coincide with those structures that can be formed by previous works \cite{Spring08, Suzuki99distributedanonymous, JoRaM09} giving the basis of the forms of lattice that can be studied. This paper will strictly concern itself with the {\bf square grid lattice}\index{grid lattice} though it can be extended in to any of the other grid configurations such as the hexagonal or triangular lattices along with many others.
The locations of the points on the lattice allow the construction of the metric space, $(M,d)$ as such the locations of the automata. Assuming a lattice structure for positioning of the automata and transmission radii, $\mathcal{R}$, is in accordance with the principles of transmission used in the pathloss geometric random graph model put forth in Ad-Hoc networks.
A variety of topologies, the lattices that can be constructed as in \cite{Spring08} (square, hexagon and triangle), and the shapes of neighbourhoods generated by varying the transmission neighbourhood can be seen in Figure~\ref{neighLatt} where black dots represent the placement of automata on the plane according to the lattice here shown by the black lines. Neighbours are those within the dashed circle and the automata at the centre of the circle is the initial transmitter of any messages.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.55]{neighLatt}
\caption{Showing three differing lattices and how their neighbourhoods can be altered simply by varying the radii of the transmission neighbourhood which, here, is depicted as a dotted circle. All automata, shown as black dots, are able to receive messages from the automata at the centre of the circle.The three lattices depicted are (left to right) square, hexagon and triangle.\label{neighLatt}}
\end{figure}
The connectivity graph defined in ad-hoc radio networks denotes those that are able to send and receive messages with certainty however in distributed algorithms the possibility of communication errors is not ignored. In such models messages may be lost, duplicated, reordered or garbled where they must be detected and corrected by supplementary mechanisms which are mostly referred to as protocols for example the common network protocol of TCP. The Broadcasting Automata model sides with the former method of message passing. It is assumed that there will be no error in transmission, the receipt or sending of messages across the network is guaranteed however consideration is made to the synchronicity of the network due to the high sensitivity to the timing of message passing as will become clear later when discussing pattern formation with such protocols.
To this extent we consider two variants of Broadcasting Automata: {\bf synchronous} and {\bf asynchronous} (or reactive) models. In each of the variations both the transmission of a method and amount of time take for the automata to process the message is considered to be constant. It is this constant time that will afford synchronisation within the model. \tn{The two differing models are presented to show that there are possible complexity trade offs depending on how it is to be measured. If the size of the alphabet is a priority, such that it must be minimised, then the synchronous model is best, however this comes at a penalty of time complexity, where the asynchronous model fairs better.}
In the {\bf asynchronous} model upon the receipt of a message, an input symbol (or set of symbols) $\sigma\in \Sigma$, the automaton, $A$, becomes active. The automaton may only react to a non-empty set of messages, it may not make and epsilon transition. Once activated the automaton, $A$, reacts to the symbol(s), $\sigma$, according to the configuration of the automaton and responds with the transmission of an output symbol, $\lambda\in \Lambda$, to \tn{those in its transmission radius for the current state, $q\in Q$, $\tau(q)$}. The processing of the input symbols by the automaton is done in one discrete time step which is the same period for all automata in the graph. Once active the automaton must wait for another message in order to change its state it is not allowed epsilon transitions.
The {\bf synchronous} model has a singular alphabet and as such, $\Sigma = \Lambda$ with the allowance of epsilon transitions. Upon the receipt of a message, an input symbol (or set of symbols) $\sigma\in \Sigma$, the automaton, $A$, becomes active. Once active the automaton is synchronised with the rest of the graph by repeated epsilon transitions that are representative of a constant transition via input symbol as in the asynchronous model. As each such transition is considered a change in configuration of the automaton it must take a single time step. As such this guarantees the synchronisation of system. However the multiple alphabet can be simulated by associating different symbols to different time steps as shall be seen later.
The following outlines, informally, the methodology of message passing used by the two differing constructions of automata, synchronous and asynchronous.
\noindent In the {\bf synchronous} model messages are passed from automaton to automaton according to the following rules:
\begin{enumerate}
\item An automaton, $A$, receives a message from an activating source at time, $t$;
\item At time $t+1$, $A$ sends a message to all automata within its transmission radius dependent on its state, $q\in Q$, $\tau(q)$;
\item At time $t+2$, $A$ ignores all incoming messages for this round.
\end{enumerate}
In the {\bf asynchronous model} the following rules can be applied:
\begin{enumerate}
\item An automaton, $A$, receives a message $\sigma_i\in \Sigma$ from an activating source at time $t$;
\item The automaton, $A$, broadcasts a message $\sigma_{(i+1)\mod |\Sigma|}$ to all automata within its transmission radius, dependent upon its state, $q\in Q$, $\tau(q)$ at time $t+1$;
\item Ignore all incoming messages at time $t+2$.
\end{enumerate}
\tn{In both models Step 3 prevents an automaton from receiving back the message that has just been passed to the automatons neighbours by ignoring all transmissions received the round after transmission and ensuring that the messages are always carried away from the initial source of transmission. In both cases this rejection of all messages may be modelled by the addition of a state that simply does not accept input in that state, where epsilon transitions apply, or that the transition is made to another equivalent state independent of input received, where from here it is possible to receive transmissions that once again affect the logic of the program. }
Naturally, given these two models it is interesting to see if they are capable of the same computations and that indeed anything that can be done in one model can be done in the other. There are many ways with which to establish equivalence between automata and models in general. One important technique that has been used to show equivalence in connection with Turing machines is {simulation} \cite{linz2001introduction}. Here a similar proposition is made with respects to establishing the equivalence of the two models of automata that are given here, synchronous and asynchronous.
\begin{proposition}\label{equiv}~\cite{Geometric2}
Both the synchronous and asynchronous models are able to simulate the other.
\end{proposition}
It should also be noted that with some generalisation, to allow a variable radius neighbourhood interaction which has been explored in a limited sense \cite{Wolfram19841, Gerhardt1990392}, the Cellular Automata\index{cellular automata} (CA) model may also be used to simulate the Broadcasting Automata model. In a grid of CA it is possible to assign states that correlate to transmitters where each transmitter in the grid is assigned a unique state and all other CA are in the quiescent state. The initial transmitters state causes each of the automata who have the initial transmitter in their neighbourhood to change from the quiescent state to the transmitters state plus one over some modulo. This clearly simulates the transmission of some word over a modulo for all automata in the grid. Upon finding a CA with one of this initial set of states within its neighbourhood, and such that it is all encompassing in its neighbourhood, the second transmitter changes its state which triggers a second cascade of state changes that are equivalent to those from the first transmitter only this time the automata must take in to account the initial state that it is already in. This should provide an exact copy of the Broadcasting Automata model, assuming that CA is extended to allow such variable neighbourhoods. This does not mean however that there is an exact translation of Cellular Automata in to Broadcasting Automata and as such it is only possible to say that $BA \subseteq CA$.
\section{Variable Radius Broadcasting over $\mathbb Z^2$}
Whilst the model provided here covers many configurations for the underlying communication graph this paper is mainly considering the following case, where all automata are place in euclidean space according to a regular arrangement equivalent to that of a square lattice such that, using the Cartesian coordinate system, for any automata in the space its nearest neighbour is at distance $1$ and its next nearest neighbour is at distance $\sqrt{2}$. This idea is illustrated for two dimensions in Figure~\ref{radii}. \tn{This now leads to the model of Broadcasting Automata where $(M,d)$ is $M = \mathbb Z \times \mathbb Z$ and where for $\rho=(\rho_0,rho_1)$ and $\rho'=(\rho'_0,\rho'_1)$ the distance function is $d(\rho,\rho')=\sqrt{(\rho_0-\rho'_0)^2+(\rho_0-\rho'_0)^2}$, the Euclidean distance function for $\rho,\rho'\in M$.}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.22]{radii.eps}
\caption{A variety of transmission radii are shown (l-r) squared radii $r^2=\{1,2,4,5,8,9,10,13,16\}$. Crosses represent the centre of the respective discrete disc.\label{radii} }
\end{figure}
\tn{
Figure~\ref{propWaves} shows as example the automata that receive a message when the Euclidean space and square lattice are restricted to two dimensions but the radius of broadcast, $\tau(q)$ for $q\in Q$, is varied. The automaton at point $\rho\in M$ which is the source of the transmission is shown as the circle at the centre of the surrounding automata on the plane, those that are black are within the transmission range $\tau(q)$ for $q\in Q$. Successive larger collections of automata coloured black show what happens when range can be changed to alter the automata that are included in the transmission radius, $\tau(q)\in \mathcal{R}$. If the transmission radius, $\mathcal{R}$, is equal to 1, as in Figure~\ref{propWaves} diagram $a)$ then only four of the eight automata can be reached. If the radius is made slightly larger and is equal to $\sqrt{2}$, it can encompass all eight automata in its neighbourhood as shown in diagram $b)$. Such structures are identical to the well studied neighbourhoods von Neumann\index{von Neumann neighbourhood} and Moore\index{Moore neighbourhood} respectively and as such here it shall be considered that such constructions of neighbourhoods are a generalisation of these two neighbourhoods. As we will show later, iterative broadcasting within von Neumann and Moore neighbourhoods can distribute messages in the form of a diamond wave and a square wave as shown later in Figure~\ref{fig_MVN}.}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.3]{propWaves}
\caption{ Diagram $a)$ represents the propagation pattern for a diamond wave (Von Neumann neighbourhood) and
diagram $b)$ shows the propagation pattern for a square wave (Moore neighbourhood).
\label{propWaves}}
\end{figure}
The construction of distinct radii for the circles is defined by the numbers $n$ such that $n = x^2 + y^2$ has a solution in non-negative integers $x$, $y$ \cite{oeisA001481}. Here it can be seen that $r^2$ is given for convenience as $n$ always has a convenient representation in $\mathbb Z$ whereas it becomes cumbersome to write out either the root of the integer or its decimal representation. For an explanation as to how these numbers relate to the distinct discrete discs it must be noted that distinct discs are constructed such that $r^2$ contains a new solution in $x$ and $y$. As it can be considered that the digital disc represents all of the maximal combinations of Pythagorean triples such that $x^2+y^2\leq r^2$ then for one disc to differ from another there must be an increase of $r^2$ such that there exists a new, distinct, maximal solution for $x,y\in \mathbb Z$. Generating these in the inverse direction by choosing $x,y\in \mathbb Z$ such that it is a new maximal combination not contained in any of the previous, smaller discrete discs allows the construction of the list of $r^2$ that defines the sequence of distinct discrete discs.
\section{Neighbourhood and Broadcasting Sequences}
The idea of propagating a pattern of square ($r^2=2$) or diamond waves ($r^2=1$), generated by repeated application of Moore or Von Neumann waves respectively, in dimension two are commonly known as Neighbourhood Sequences (see Figure~\ref{fig_MVN}). Neighbourhood Sequences\index{neighbourhood sequences} are an abstraction used to study certain discrete distance metrics that are generated upon the repeated application of certain neighbourhoods to a lattice for example the Moore neighbourhood to a square lattice. As previously discussed such neighbourhoods as von Neumann and Moore are encompassed by the more general model used here whereby the transmission radius is varied on a square lattice the first two of such transmission radii to produce distinct objects equating to those of the von Neumann and Moore neighbourhoods.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.3]{transNeigh12fig.eps}
\caption{Wave propagation with Neighbourhood Sequences on the square grid: Moore neighbourhood (left) and Von Neumann neighbourhood (right).
\label{fig_MVN}}
\end{figure}
Definitions and notation concerning neighbourhood sequences as considered in \cite{FarBajNag06}, and many other works, are now
given here for completeness.
Let $p\in \mathbb Z^n$ where $n\in \mathbb{N}$ and such that the $i$th coordinate of $p$ is given by $Pr_i(p)$ for $1\leq i \leq n$.
\begin{definition}\label{mNeigh}
For $M\in \mathbb Z$ where $0\leq M \leq n$ the points $p,q\in \mathbb Z^n$ are $M-Neighbours$ when the following two conditions hold:
\begin{itemize}
\item $|Pr_i(p)-Pr_i(q)|\leq 1$ for $(1\leq i \leq n)$
\item $\sum_{i=1}^{n}|Pr_i(p)-Pr_i(q)|\leq M$
\end{itemize}
\end{definition}
\tn{An $n$-dimensional neighbourhood sequence is denoted $\mathcal{A}=(a(i))^{\infty}_{i=1}$, $\forall i\in \mathbb{N}$ where $a(i)\in {1,...,n}$ denotes an \emph{M-neighbourhood} by its value of $M$ such that for $a(1)$ denotes that the next neighbourhood set of points in the sequence is those that differ by at most one coordinate as given by Definition~\ref{mNeigh} where $M=1$ in this case the Von Neumann neighbourhood. If $\mathcal{A}$ is periodic then $\exists l\in \mathbb{N}$, $a(i+l)=a(i)$ $(i\in \mathbb{N})$. Such periodic sequences are given as $\mathcal{A} = (a(1),a(2),...,a(l))$.
\begin{definition}
The $\mathcal{A}$-distance, $d(p,q;\mathcal{A})$, of $p$ and $q$ is the length of the shortest A-path(s) between them.
\end{definition}
As the spreading of such neighbourhoods is translation invariant only an initial point of the origin need be considered w.l.o.g.
\begin{definition}
The region occupied after $k$ applications of the neighbourhood sequence $\mathcal{A}$ is denoted as $\mathcal{A}_k=\{p\in \mathbb Z^n:d(0,p;\mathcal{A})\leq k\}$ for $k\in \mathbb{N}$.
\end{definition}
Also, let $H(\mathcal{A}_k)$ be the convex hull, given in Definition~\ref{convHull}, of $\mathcal{A}_k$ in $\mathbb Z^n$.
\begin{definition}\label{convHull}
A convex hull is the smallest set of points that form a convex set as given in Definition~\ref{convSet}.
\end{definition}
\begin{definition}\label{convSet}
{\bf Convex set}. A set $C\subset R^d$ is convex if for every two points $x, y \in C$ the whole segment $xy$ is also contained in $C$. In other words, for every $t \in [0, 1]$, the point $tx + ( 1 - t ) y$ belongs to $C$.
\cite{mat2002lectures}
\end{definition}
}
Discrete discs are formed by the Broadcasting Automata as they are arranged on the plane at integral Cartesian coordinates, $(\mathbb Z \times \mathbb Z)\in M$, and as such a broadcast to all automata in range, for point $v\in M$, $\tau(c(v))=r$ will cause a change in state to all those automata that form a discrete disc of radius $r$ from the point $v$. In the broadcasting automata model it is assumed that there is no restriction on the radius of transmission where it is considered that as the Von Neumann and Moore neighbourhoods can be described as the first two radii in the set of distinct discrete discs, $r^2=1$ and $r^2=2$. This means that discrete discs are quite a natural extension to the basic notion of neighbourhood sequences merely relaxing the constraints on the initial definition of $M-neighbours$ in the following way.
\begin{definition}\label{rNeighbours}
Two points $p,q\in \mathbb Z^n$ are $r-neighbours$ if the Euclidean distance, $d(p,q)=\sqrt{\sum^{n}_{i=0}(q_i-p_i)^2}$, is less than some $r$, used to denote the radius of a circle, such that $d(p,q)\leq r$.
\end{definition}
Utilising the framework supplied by the work done on neighbourhood sequences it is no possible to outline {\bf Broadcasting sequences}\index{broadcasting sequences} which denote any sequence of radii, $r$, such that $R=(r_1,r_2,...,r_l)$. Labelling those points that are reachable by some application, $R_k$, of the neighbourhood sequence is another extension to the notation. All points such that $p\in R_1$ are labelled $0$, all points $p\in R_2\backslash R_1$ are labelled as $1$. More generally labels will be assigned $b$ where $k \equiv b \mod m$ and $m\in \mathbb{N}$. The work conducted here will also be restricted to the study of $\mathbb Z^2$.
Many of the questions asked and, indeed, answered, in neighbourhood sequences, shall be useful in the exposition of broadcasting automata. These include whether a certain sequence of neighbourhoods are metrical or not \cite{1521291}. Indeed when using a general form of broadcasting automata whereby alternation of the broadcasting radius is allowed at each step as suggested here. Considering that all notions of the construction of algorithms presented here in the broadcasting automata model rely on the distance from the transmitter to the node, being able to assure that this distance will be consistent for all automata is essential. Also studied in neighbourhood sequences is the possible resultant shapes which are categorised and examined for their various properties such as their use in approximation\index{metric approximation} of euclidean distances where the isoperimetric ratio is used to compare the sequences based on the shapes that they form on the plane \cite{2010arXiv1006.3404F}. This same method is used to again show and estimation for euclidean distance as well as estimations for certain $L_p$ distances, here the astroid which will be presented in Section~\ref{Lp_Metrics}. The characterisation of the shapes generated by these neighbourhood sequences yield results about the shapes of compositions of such discrete discs on the plane and the partitions that such compositions form.
\section{Geometrical Properties of Discrete Circles}
The characterization of broadcasting sequences on ${\mathbb Z}^2$ is closely related to the study of discrete circles on the square lattice
and their discrete representation. One of the efficient methods to describe the discrete circle is a chain coding.
The method of chain coding was first described in \cite{5219197} as a way in which to encode arbitrary geometric configurations where they were initially used, as they shall be here, to facilitate their analysis and manipulation through computational means. At the time there were many questions about the encoding of shapes and the methodologies which should be used to encode such shapes and chain coding presented an schema which was simple, highly standardised and universally applicable to all continuous curves. Whilst the definition given here differs from the definition given in \cite{5219197} it only differs for convenience when discussing discrete discs in the first octant allowing the use of only the numeric symbols $0$ and $1$ instead of $0$ and $7$ as Freeman suggested. This gives the following definition of chain coding that shall be used throughout.
\begin{definition}
From a starting point on the square lattice a {\bf chain code}\index{chain code} is a word from the alphabet $\{0,1,2,3,4,5,6,7\}$.
\end{definition}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.45]{chainCode.eps}
\caption{(Left) Showing the eight possibilities of motion from the single point of origin, circled, as they may be traced out on the plane (Centre) Shows the eight possible points of motion from a central point and (Right) gives and example of an arbitrary line traced out on the lattice with a chain code of form $70102435$ originating from the circled point. \label{chainCod}}
\end{figure}
From some starting point chain codes may be used to reconstruct a shape, $S$, by translating the code's motion on to the lattice where $0$ indicates positive movement along the $x-axis$ with increasing values moving clockwise through all $8$ possible degrees of motion on the lattice, labelled from $0$ to $7$ respectively. An integral shape, $S$, can be encoded using an inverse method whereby the shape is traced from some starting point and the motion from one point to the next connected point in the shape is recorded as a chain code. For example, a straight line would be represented by the infinite repetition of a single digit such as $00000000000000...$, a line that satisfies the equation, $x-y=0$, may be encoded $77777777...$ or as the example given in Figure~\ref{chainCod} and arbitrary line may be encoded numerically as $7010235$.
The discrete disc is the basic descriptor of a transmission in broadcasting automata. It represents the total set of all automata that could be reached on the square lattice after a single broadcast of radius, $r^2$. In the following sections reasoning will follow only from the first octant, which can be seen in Figure~\ref{octants}, of the discrete circle which shall now be defined and gives all the information that is required to categorise the discrete disc, its perimeter, but also has a convenient method of chain code construction. The need only to observe the first octant is derived from the symmetry that occurs such that it is possible to take a solution of the first octant and by successive reflection operations about the eight octants generate a full circle \cite{Bresenham:1977:LAI:359423.359432} under the assumption, which is also made here, that the circles are centred on a integral point on the plane.
\begin{definition}\label{circD}
A discrete circle is composed of points in the $\mathbb Z^2$ set $\zeta =\{(x,y) | x^2+y^2\leq r^2,\ x,y\in \mathbb Z,\ r\in \mathbb R \}$ that have in their Moore neighbourhood any point in the complement of $\zeta$.
\end{definition}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.60]{octants.eps}
\caption{An illustration of a circle that has been divided in to eight octants with the labelling of those octants applied accordingly.\label{octants}}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.80]{digiCirc.eps}
\caption{An illustration of a discrete disc and its accompanying discrete circle of radius squared 116, where the chain code for the first octant is labelled as $l_0l_1l_2$.}\label{digiCirc}
\end{figure}
In Definition~\ref{circD} it is shown that $r\in \mathbb R$ however many of these do not produce distinct discrete circles. The set of $r$ that includes all distinct circles are those which form Pythagorean triples which consist of two positive integers $a,b\in \mathbb N$, such that $a^2+b^2=r^2$.
A simple way to draw the discrete circle is to divide it into octants as can be seen in Figure~\ref{digiCirc}. First calculate only one octant of the circle and as the rest of the circle is "mirrored" from the first octant it can now be successively mirrored to allow the construction of the other sections of the circle. Let us divide the circle into eight octants labelled 1-8 clockwise from the y-axis. The notation $Octant(i)$ is used to refer to octant sector $i$. If an algorithm can generate the discrete circle segment on one octant, it can be used to generate other octants by using some simple transformations, such as rotation by $90^{\circ}$ and reflection around the $x$-axis, $y$-axis, and the $+45^{\circ}$ diagonal lines.
Some basic results are now given about chain codes of such discrete circles. It is possible to show that in the first octant and given the definition of chain coding presented here only two digits, $0$ and $1$ are required to fully represent the discrete circle.
\begin{lemma}\label{chainStruc}
A Chain coding $w$ for the circle in the first octant is a word in $\{0,1\}^*$.
\end{lemma}
\begin{proof}
First let us show that in the first octant for any given point $p_0=(x_0,y_0)$ on the circle at the point $p_1=(x_0+1,y_0-1)$ must also be in the circle as $|p_1|<|p_0|$. When $p_0$ is in the circle and in the first octant segment such that $x<y$ it follows that the magnitude of $p_1$ is less than $p_0$ as $|p_0|-|p_1| = (x_0^2+y_0^2)-(x_1^2+y_1^2) = (x_0^2-x_1^2) + (y_0^2-y_1^2) = (x_0^2-(x_0+1)^2) + (y_0^2-(y_0-1)^2) = (x_0^2-x_0^2 - 2x_0 - 1) + (y_0^2-y_0^2 + 2y_0 -1) = 2y_0-2x_0-2 \geq 0$. Since $p_1=(x_0+1,y_0-1)$ should be in the circle if $p_0=(x_0,y_0)$ is on the circle we have that any chain code for the circle in the first octant does not contain a subword ``22'' as it will lead to contradiction. Moreover in any subword a single symbol ``2'' should be followed by ``0'' and can be substituted by $1$.
\end{proof}
As only the first octant segment is considered the solutions can be represented using strings consisting of only $0$ and $1$ where $0$ is positive motion along the $x-axis$ and $1$ is positive motion along the $x-axis$ and negative motion along the $y-axis$.
Indeed in the generation of such chain codes the following method shall be used which is a modification of the algorithm that is given in \cite{Bhowmick20082381}. Starting with the equation for the circle $x^2+y^2=r^2$ where it is known, by definition, that $r^2\in \mathbb Z$. Now as solutions of the first octant are required and under the assumption that, without loss of generality, the centre of the circle is at the origin it is now possible to replace $y^2$ with $(r'-k)^2$, where $r' = \lfloor r \rfloor$, such that when $k=0$ and rearranging for $x^2$ such that $x^2=r^2-r'^2-k^2+2r'k$ the solution to the circle equation gives $x^2= r^2-r'^2$ where $r'\leq r$, as such the length of the first solution of such an equation is $\sqrt{r^2-r'^2}$. Successive values of $k$ give solutions to the length in the $x$-axis of that particular maximal Pythagorean triangle. All such Pythagorean triangles will give the final solution to the circle.
Here it is necessary to start with some terminology to say more precisely which parts of the chain code are considered to represent specific components in the resulting polygons. It will become much clearer later on why such distinctions are required as a single side of the polygon may differ in its chain code categorisation. The first of these categories of chain code, the most basic, is given here.
\begin{definition}
A {\bf chain code segment} is a subword of a chain code which is either a word $0^{|s_0|}$
or $10^{|s_i|-1}$.
\end{definition}
A discrete circle, $u$, is composed of chain code segments where each segment is represented as $u_i$, i.e $u=u_0u_1u_2...u_n$. Chain code segments can be expressed using a power notation for the number of repetitions e.g $100100$ may be rewritten as $(100)^2$ and if $u=u_0u_1u_2u_3= (00)(100)(100)(10)=(00)^1(100)^2(10)^1$.
It is possible to map the chain code, $u$, in $Octant(1)$, to the others octants through a function that maps the alphabet of the chain code $\{0,1\}$, to code to a new code in another octant segment. A chain code of a circle in an $Octant(n)$ can be constructing from an $Octant(1)$
first by applying the following mapping $\{0,1\} \rightarrow \{n-1 \mod 8, n \mod 8 \}$, where $0\leq n \leq 7$.
Then for the $Octant(n)$, where $n \equiv 1 \mod 2 $, or $n$ is odd, the code and the mapping itself must be reversed such that $\{0,1\} \rightarrow \{n \mod 8, n-1 \mod 8 \}$. For the rest of this paper we will only consider chain codes in the first octant.
\begin{proposition}
A circle in the first octant can be chain coded as follows:
\begin{center}
$0^{|s_0|} 10^{|s_1|-1} 10^{|s_2|-1} \ldots 10^{|s_i|-1} \ldots 10^{|s_{r'^2/2}|-1}$
\end{center}
where,
\begin{center}
$s_i=\{x |r^2 - r'^2 + 2(i-1)r' - (i-1)^2 < x^2 \leq r^2 - r'^2 + 2ir' - i^2, x \in \mathbb Z \}$
\end{center}
$r$ is the radius and $r'$ is the floor of the radius.
\end{proposition}
\begin{proof}
It follows from \cite{Bhowmick20082381} where such a method of chain coding is introduced.
\end{proof}
\begin{proposition}
Chain code segments, following the schema $$0^{|s_0|} 10^{|s_1|-1} 10^{|s_2|-1} \ldots 10^{|s_i|-1} \ldots 10^{|s_{r'^2/2}|-1}$$ have the following constraints to their lengths in the discrete circle: $$\lfloor \frac{s_{i}-1}{2}\rfloor-1 \leq s_{i-1} \leq s_{i}+1\ .$$
\end{proposition}
\begin{proof}
It, again, follows from \cite{Bhowmick20082381} where this property is proved.
\end{proof}
As previously noted terminology must be introduced to distinguish the varying parts of the chain code and in order to talk about what they represent. In order to characterise the discrete circle two notions are introduced, \emph{Line Segments} and \emph{Gradients of Line segments}.The following definition for the gradient gives a method for extracting such information from the series of digits that construe the line as a word.
\begin{definition}
A {\bf gradient}, $G(u)$, of a chain code subword in $\{0,1\}$, $u$, is given by $G(u)=\frac{\#_1(u)}{|u|}$, where the function $\#_1$ returns the number of $1's$ found in the chain code and $|u|$ is the length of word $u$.
\end{definition}
Line segments are in reference to actual lines that these sections of chain code would represent on the plane. That is if one were to draw a line and then attempt to translate this line in to a chain, under the assumption that the gradient of the line is rational, for reasons which can be seen in the definition of a gradient, it would be done via a periodic series of chain code segments. Taking a single period of this chain all the information that is required to identify and reconstitute this line to any length is now known. In this way the name may interpreted quite literally as a segment of a line in chain coded form such that any repetition of this object produces a line. With this in mind it will naturally be of benefit to be able to know a little more about the line that has been chain coded as a line segment.
\begin{definition}\label{lineSeg}
A {\bf Line segment}, $l_i=u_ju_{j+1}...u_{j+n}$, is the shortest contiguous subsequence of chain code segments which maintain an increasing gradient such that $ G(l_i)<G(l_{i+1})$, $\forall i>0$, \tn{where the last chain code segment in $l_i$, $u_j+n$, is contiguous to the first chain code segment in the next line segment, that is, $l_{i+1}=u_{j+n+1}u_{j+n+2}...u_{j+n+m}$.}
\end{definition}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.30]{chainDia.eps}
\caption{Showing the the relationship between chain codes, chain code segments and line segments.}
\end{figure}
As, ultimately, all of the discrete discs that are discussed here, and, more importantly, that are possible to construct, are polygons, such polygons are composed of line segments which represent the sides of the convex shape on the plane. Further, when categorising the repeated broadcast of such discrete discs, later discussed as composition, the shapes that are generated are known to be similar, in the strict geometric sense. A definition for chain codes, translating from the geometry of the Reals, is given.
\begin{definition}
For any two chain codes, $u,v$, comprising line segments such that, $u=l_0l_1...l_m$ and $v=l'_0l'_1...l'_n$, for $m=n$, given by Definition~\ref{lineSeg}, are similar (geometrically) if there is a constant $k \in \mathbb N$ such that $u = l'\mathrlap{_{0}}^{k} l'\mathrlap{_{1}}^{k} ...l'\mathrlap{_{n}}^{k} $.
\end{definition}
In this way it is possible to say that two shapes are similar in a geometric sense such that they contain the same number of distinct line segments but the number of each of these line segments is different by some constant.
With these definitions stated it is possible to begin the categorisation of the set of discrete circles with the first observation which extends an already known notion about the chain code of the discrete disc. Following \cite{Bhowmick20082381} it is known that in the chain code $u$ any chain code segments with increasing lengths, such that $|u_i|<|u_{i+1}|<...<|u_{i+n}|$, may increase it by at most $1$ from the length of a previous segment, i.e. $|u_{i+1}| \leq |u_i|+1$ for all $i$.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.5]{exis.eps}
\includegraphics[scale=0.5]{exis2.eps}
\includegraphics[scale=0.2]{segChord.eps}
\caption{(Left and Centre) Showing the initial point $(x,y)$ the chord and the point on the chord, the solid circle. (Right) Showing a chord on a circle from $(x,y)$ }\label{existence}
\end{figure}
By direct construction it is easy to check that a line segment can be of the form $10^n$ or $10^n 10^{n+1}$. However we can show that no line segment may be composed of more than two chain codes which increase in length by one.
\begin{lemma}\label{lineLength}
No line segment can be of the form $10^n 10^{n+1} 10^{n+2} \ldots 10^{n+i}$, where $i>1$.
\end{lemma}
\begin{proof}
Let us first show that any line segment which is part of the discrete circle cannot be of the form $10^n 10^{n+1} 10^{n+2}$ for any $n>1$. We will prove it by contradiction. Assume that the line segment $l_i=10^n 10^{n+1} 10^{n+2}$ with the gradient $1/n+1$
starting at a point $p$ with coordinates $(x,y)$ and finishing at $(x+3n+6,y-3)$. Therefore a point $(x+2(n+2),y-2)$ , i.e. a point which can be reached by a code $10^n 10^{n+2}$ from a point $(x,y)$, is above the discrete circle. By the definition of the chordal property of the circle any line that joins two points on the circle must bound, within the triangle formed from the two points on the circle and its centre, points which are again within the circle. On the other hand a point $(x+2(n+2),y-2)$, shown circled in Figure~\ref{existence}, belongs to a chord between end points of a chord $(x,y)$ and $(x+3n+6,y-3)$ and therefore should be within a discrete circle. So it is not possible to have a line segment with the following chain code $10^n 10^{n+1} 10^{n+2}$. The same argument holds for the general case $10^n 10^{n+1} 10^{n+2} \ldots 10^{n+i}$, where $i>1$ since the extension of the line segment still keep the point $(x-2,y+2(n+2))$ above the discrete circle and on the other hand this point will be in the triangle formed by a centre of the circle and two end points of the line segment since its gradient is equal to $\frac{i+1}{n(i+1)+(i+1)+i(i+1)/2}=\frac{1}{n+1+i/2} \leq \frac{1}{n+2}$, where $i \geq 2$.
\end{proof}
Further to this weak restriction, that the chain codes may not be monotonically increasing or indeed non-decreasing, it is possible to give a more definite generalisation of the structure of the chain codes. This structure, which must be adhered to for all discrete discs, is given here.
\begin{theorem}\label{lineSegFor}
Any line segments on the discrete circle with non-negative gradients
should be in one of the following forms
$(10^{n})^*$, $(10^{n})(10^{n+1})^*$, $(10^{n})^*(10^{n+1})$.
\end{theorem}
\begin{proof}
Following Lemma~\ref{lineLength} we can restrict number of cases
since any concatenations of more than two chain codes which increase in length by one
may not be part of a line segment.
The line segments are sub-words of a chain code with increasing gradients, so if a subword
$(10^{n})^m$ is surrounded by any chain code segments $10^{n_1}$ and $10^{n_2}$, where
$n_1, n_2 \ne {n \pm 1}$ , then $(10^{n})^m$ is a line segment.
In the case were $n_1, n_2 = {n \pm 1}$ a subword $(10^{n})^m (10^{n+1})^k$
can be surrounded by only chain codes $10^{n_3}$ and $10^{n_4}$, where
$n_3 \ne n - 1$ and $n_4 \ne n +2$ (by Lemma~\ref{lineLength}).
We will show now that the only line segments of the form
$(10^{n})^m (10^{n+1})^k$ can be where either $m=1$ or $k=1$.
The proof is similar to Lemma~\ref{lineLength}.
Let us first show that the line segment from a point $(x,y)$
cannot have the following chain code $(10^{n})^2 (10^{n+1})^2$.
If we assume that such chain code may correspond to a line segment
such that a point $(x-2, y+2n+3)$, shown in the centre diagram in Figure~\ref{existence}, is above the chain code (i.e. our of a circle) and at the same
time is on a chord between points $(x,y)$ and $(x-4, y + 4n +6)$, so should
be in a discrete circle. Extending the chain code $(10^{n})^2 (10^{n+1})^2$
from the left by $(10^{n})^{m_1}$ or from the right by $(10^{n+1})^{m_2}$
for any $m_1, m_2 >0$ will not change the property of the point $(x-4, y + 4n +6)$.
So it can be in one of the following forms either $(10^{n})^m 10^{n+1}$ or $10^{n} (10^{n+1})^k$.
\end{proof}
\section{Iterative Composition}
Iterative composition of discrete circles may create different polygons, which later can also be used for forming
non-convex shapes. The following definitions, and further characterisations, shall be required in order to discuss the properties of such composition and develop the proofs.
\tn{
\begin{definition}
A discrete disc\index{discrete disc} is composed of points in the $\mathbb Z^2$ set $\zeta^{r^2} =\{(x,y) | x^2+y^2\leq r^2,\ x,y\in \mathbb Z,\ r\in \mathbb R \}$.
\end{definition}
\begin{definition}\label{composition}
The composition of $l$ digital discs is defined as $\zeta^{r_0}\circ \zeta^{r_1}\circ...\circ\zeta^{r_l}$ for $r_i\in \mathbb N$, where, $\zeta^r \circ \zeta^{r'} = \{a+b|a\in \zeta^r,\ b\in\zeta^{r'}\}$ and is equivalent to $H(\mathcal{A}_l)$ where $\mathcal{A}=r_0,r_1,...,r_l$ as given in Definition~\ref{convHull} and Definition~\ref{rNeighbours} respectively.
\end{definition}}
\begin{definition}
The composition $u\circ v = w$ represents the composition of the chain codes in the first octant for their respective discs, $\zeta^u$ and $\zeta^v$, which results in $w$ the chain code for the first octant of the resultant disc of $\zeta^u\circ \zeta^v$. This method can be seen in Figure~\ref{exis}.
\end{definition}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.8]{compo.eps}
\caption{Showing the composition of squared radii $Disc(85)\circ Disc(17)$ where $Disk(85)$ is shown as a solid line and $17$'s first composition is shown dashed. The hatched area represents all contributions from the previous compositions.}\label{exis}
\end{figure}
\begin{example}
Iterative composition of discrete circles\index{discrete circles} may create different polygons. For example, in the following picture, Figure~\ref{91626}, there are two differing cases, in one case we iteratively apply the digital circle with squared radius $9$ resulting in the equilateral octagon. In the second case squared radii $16$ and $26$ are periodically applied starting with $16$.
\end{example}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.20]{a9.png}
\includegraphics[scale=0.20]{a1626.png}
\caption{The above figures illustrate (left) the constant iteration of the discrete disc of squared radius $9$ and (right) the alternation between the transmission of squared radii $16$ then $26$. }\label{91626}
\end{figure}
From these definitions it is now possible to describe a naive algorithm, and derive its correctness, for the composition of any number of chain codes that represent discrete discs. The result is again a convex polygon that represents the combination of the constituent parts. Such an algorithm merely systematically checks all possible combinations of the chain code and reports the largest that is found for that step. Later it shall be seen that if the chain code of the discrete disc is categorised further it is possible to give a simple linear time algorithm with a proof of correctness that is derived from the Minkowski sum along with a combinatorial proof.
\begin{lemma}\label{direct_composition}
Given two chain codes, $u=u_0u_1...u_n$ and $v=v_0v_1...v_m$, in form $\{0,1\}^*$ then $u \circ v=w=0^{|w_0|}10^{|w_1|-1}...10^{|w_{m+n-1}|-1}$, such that $$|w_k|=max(\{\displaystyle\sum\limits_{i=0}^{n} |u_i| + \displaystyle\sum\limits_{j=0}^{k-n} |v_j|\ |\ 0 \leq n \leq k \}).$$
\end{lemma}
\begin{proof}
The naive way of generating the composition $u\circ v$ is for all points on the chain code of the disc $u=0^{|u_0|}10^{|u_1|-1}...10^{|u_i|-1}...10^{|u_n|-1}$, to be the centre of the disc $v=0^{v_0}10^{|v_1|-1}...10^{|v_j|-1}...10^{|v_m|-1}$. This is equivalent to placing the centre of the chain code $v$ at every point defined by the chain code $u$ generating $w$. The maximal point in each chain code segment (i.e if $u_i=100$ then the coordinate of the final $0$) need only be considered, clearly covering all others. Let $u_i\circ^\prime v_j = w_{i+j}$ denote the centring $v$ at the coordinate reached by the maximal point of $u_i$ in which we consider the maximal point attained by chain code segment $v_j$ and represent a possible length of the chain code for $w$ at $w_{i+j}$. The maximal of all such combinations of lengths, those for which the sum of $i,j$ are equivalent, is required and is defined as the longest contiguous subsequence of length $k$ from $v=v_0v_1...v_j$ and $u=u_0u_1...u_i$ for all $i,j$ such that $i+j=k$ which represents $w_k$ for $0 \leq k < m+n$ i.e:
\begin{center}
$max
\begin{pmatrix}
\vert u_0 \vert + \vert v_0 \vert \\
\end{pmatrix}
= \vert w_0 \vert$
\hspace{1cm}
$max
\begin{pmatrix}
|u_0| + |v_0| + |v_1| \\
|u_0| + |u_1| + |v_0| \\
\end{pmatrix}
= \vert w_1 \vert$
\vspace{0.4cm}
$max
\begin{pmatrix}
|u_0| + |v_0| + |v_1| + |v_2|\\
|u_0| + |u_1| + |v_0| + |v_1|\\
|u_0| + |u_1| + |u_2| + |v_0|\\
\end{pmatrix}
= \vert w_2 \vert$
$$...$$
$max
\begin{pmatrix}
|u_0| + |v_0| + |v_1| +...+ |v_k|\\
|u_0| + |u_1| + |v_0| + ...+ |v_{k-1}|\\
...\\
|u_0| + |u_1| + ... |u_i| + |v_0| + ...+ |v_{k-i}|\\
...\\
|u_0| + |u_1| + ... + |u_k| + |v_0|\\
\end{pmatrix}
= \vert w_k \vert$
$$...$$
$|u_0| + |u_1| + ... + |u_n| + |v_0| + |v_1| +...+ |v_m| = |w_{m+n}|.$
\end{center}
\end{proof}
Such an algorithm allows a characterisation of the composition of discrete discs. Indeed, here, it can be shown that the composition of chain codes, which have been shown to be a word $W\in \{0,1\}^*$, is, again, a word, $W'\in \{0,1\}^*$.
\begin{corollary}
Given two chain codes $u$, $v\in \{0,1\}^*$. The composition $u\circ v\in \{0,1\}^*$
\end{corollary}
\begin{proof}
Each segment in $u$, $v$, is non-zero $\forall i\ |u_i|\neq 0, |v_i|\neq 0$. It is followed from Lemma~\ref{direct_composition} that by enlarging $k$, for $l(k+1)$, the previous set of solutions is extended by a chain code from $u$ or $v$ so $l(k) < l(k+1)$. Thus the composition will be in $\{0,1\}^*$.
\end{proof}
Also as a derivation from the algorithm an observation about the properties of the composition function $\circ$.
\begin{theorem}
The composition $\circ$ of two chain codes, $u$ and $v$ is commutative, i.e. $u \circ v = v \circ u$.
\end{theorem}
\begin{proof}
Following Lemma~\ref{direct_composition} it is possible to exchange $u$ and $v$ and still obtain the same result,
showing commutativity of composition of chain codes $u$ and $v$. The following two sets representing
$k$-th level of a new chain code are equal as well as their maximums:
$$\{S|S= \displaystyle\sum\limits_{i=0}^{n} |u_i| + \displaystyle\sum\limits_{j=0}^{k-n} |v_j| ,0 \leq n \leq k \} =
\{S|S= \displaystyle\sum\limits_{i=0}^{n} |v_i| + \displaystyle\sum\limits_{j=0}^{k-n} |u_j| ,0 \leq n \leq k \} $$
\end{proof}
\tn{It is also notable that having shown commutativity chain code composition it may be possible that it is an Abelian group, where the identity element is simply the circle of radius $0$, associativity is derived from the algorithms ordering of line segments by gradient, which is naturally the same irrespective of the order it is carried out in, the inverse is simply the removal of one set of line segments from a chain code. However, it is currently unknown whether or not the operation is closed.}
Further, a proof of the ordering of the line segments in the discrete circle is here required in order to validate the proof of the composition theorem that is given as one of the main tools and results of this section.
\begin{lemma}\label{lineSegOrd}
Given line segments, $l_i$, of a form produced by the digitisation of a circle, which are composed of chain code segments of its first octant $a=10^{m-1}$ and $b=10^{m}$, then the ordering of their gradient for combinations of $a$ and $b$ is $$G(b)<...<G(ab^*)<...<G(ab)<...<G(a^*b)<G(a)\ .$$
\end{lemma}
\begin{proof}
The proof is a simple comparison of the gradients which shows, where $|a|=m$, $\frac{1}{m+1}<\frac{n}{(n-1)(m+1)+m}<\frac{n-i}{(n-1-i)(m+1)+m} <\frac{n-i}{(n-1-i)(m)+m+1}<\\<\frac{n}{(n-1)(m)+m+1}< \frac{n}{(n-1)m+m+1} $, where $i<n-1$, for $G(b)< G(ab^{n-1})<G(ab^{n-1-i})<G(a^{n-1-i}b)<G(a^{n-1}b)<G(a)$ respectively.
\end{proof}
Given the preceding validation about the convexity of the discrete discs it is now possible to employ the Minkowski sum to conclude and validate the composition of any number of discrete discs using the composition function.
\begin{theorem}\label{compoThe}
{\bf Composition Theorem.} Given two chain codes $u$ and $v$ which contain line segments
$l^{u}_1 l^{u}_2 \ldots l^{u}_t$ and $l^{v}_1 l^{v}_2 \ldots l^{v}_{t'}$
with strictly increasing gradients.
The chain code of a composition $u \circ v$ can be constructed
by combining line segments of $u$ and $v$ and ordering them
by increasing gradient.
\end{theorem}
\begin{proof}
One of the ways to prove the above statement is to use a similar result about the Minkowski sum \cite{schneider1993convex} of convex polygons in $\mathbb{R}^2$ and then to prove that it holds for the ordering of the segments of the digital circles in $\mathbb{Z}^2$. Here, a purely combinatorial proof is given in terms of chain codes, chain code segments and line segments. The above statement is proved by induction. The fact that the base case for the composition of the first two line segments holds can be seen by directly checking the expression from Lemma~\ref{direct_composition}.
Assume that the statement of the lemma holds and the first $z$ line segments of $u \circ v$ were composed from a set $LS=\{l^u_1, l^u_2, \ldots , l^u_i, l^v_1, l^v_2, \ldots , l^v_j \}$ and contains $x$ chain code segments. Let us also denote a set with other line segments from $u$ and $v$ as $\bar{LS}$. Without loss of generality suppose that the last line segment that has been added is $ l^v_j$, so $G(l^v_j) \geq G(l')$, where $l' \in LS$ and $G(l^v_j) \leq G(l'')$, where $l'' \in \bar{LS}$. By adding the next line segment $l'''$ which will have a gradient larger or equal than all other line segments in $LS$ and smaller or equal than gradients of $\bar{LS}$ it must be shown that the extended chain code of $u \circ v$ is still correct, i.e. it satisfies to Lemma~\ref{direct_composition}.
First of all note that adding a new part of the chain code would not change the previous $x$ layers. If a current maximum is in the form $$|u_p| + |u_{p-1}|+ ... + |u_0|+ |v_0|+|v_1|+ ... + |v_q|$$ then the next $\#_1(l''')$ sums will be extended by chain codes from $l'''$. If $l'''$ is a line segment in $u$ the gradient of $G(l^v_j)$ is greater then $G(l''')$ and, taking into account Lemma~\ref{lineSegOrd} and Theorem~\ref{lineSegFor}, it can be seen that $$|u_{p+1}| +|u_p| + |u_{p-1}|+ ... + |u_0|+ |v_0|+|v_1|+ ... + |v_q|$$ is a maximum within the following set $$\{|u_{p+1+shift}| + ... +|u_p| + |u_{p-1}|+ ... + |u_0|+ |v_0|+|v_1|+ ... + |v_{q-shift}|, shift \in \mathbb N \}$$ since the shift will represent removal of a larger value from the $v$ component and appending the smaller value from the $u$ component.
Repeating the procedure and extending the sum by one value it can be seen that, again, $$|u_{p+2}|+|u_{p+1}| + |u_p| + |u_{p-1}|+ ... + |u_0|+ |v_0|+|v_1|+ ... + |v_q|$$ is a maximum within a set $$\{|u_{p+2+shift}| + ... +|u_p| + |u_{p-1}|+ ... + |u_0|+ |v_0|+|v_1|+ ... + |v_{q-shift}|, shift \in \mathbb N \}$$ following the same reasoning. Similar arguments can be applied for the case were $l''' \in v$.
\end{proof}
\begin{example}
Let us illustrate the composition of two chain codes corresponding to digital circles with squared radii 45 and 9 in the first octant. The chain code of the first octant is $\zeta^{45}= 0001$ and for $\zeta^{9} = 10$ under composition this yields $\zeta^{45}\circ \zeta^{9}=000101$.
\end{example}
From the preceding notions it is now possible to describe a linear time algorithm, which vastly out performs the exhaustive search of all combinations that has previously been given, which is able to form the composition of the chain codes of any two discrete circles.
As Lemma~\ref{lineSegOrd} shows that all of the line segments will be in a non-increasing order it is possible to construct such line segments by counting the number of $0's$ after each delimiting $1$ combining those that increase in size with the preceding smaller chain code segment. Having found all line segments for both chain codes it is a simple case of, starting with the chain code with the line segment with the largest gradient, adding that element to a new chain code which is the composition of the initial two. Continue choosing the line segment with the largest gradient from either of the initial chain codes until there are no elements left in either of the original circles. The result of the algorithm is the composition of the two circles.
The following proposition gives a proof of the algorithm and the notion that it is linear time computable.
\begin{proposition}
Two discrete circles chain codes can be composed into a single chain code in a linear time.
\end{proposition}
\begin{proof}
As Lemma~\ref{lineSegOrd} shows that all of the line segments will be in a non-increasing order it is possible to construct such line segments by counting the number of $0's$ after each delimiting $1$ combining those that increase in size with the preceding smaller line segment. Having found all line segments for both chain codes it is a simple case of starting with the chain code with the line segment with the largest gradient adding that to a new chain code which is the composition of the initial two.
\end{proof}
It is also now possible to state a more formal definition of `similar' such that it is possible to mathematically determine whether two objects are `similar', geometrically. The following algorithm is given as one application of such an algorithm and also hints at the notion of composing one discrete disc from other discrete discs which allows an insight in to the primality of discrete discs those that cannot be composed from any other set of discrete discs.
\begin{theorem}
Given a finite broadcasting sequence of radii $R=(r_1,r_2,..., r_l)$ and
a convex polygon\index{convex polygon} $P$, it is decidable whether there are radii such that the chain coding of the composition of is {\bf similar} to $P$.
\end{theorem}
\begin{proof}
The algorithm computes all line segments, and their corresponding gradients, for the chain codes of the set of digital disks with radii in, $\mathbb R$, and the convex polygon, $P$. Each digital disk $r_i\in R$ can be represented as a vector with $k$ values, where $k$ is the cardinality of the set of gradients and each vector component corresponds to a particular gradient with an integer value standing for the number of line segments with this gradient in $R$. Finally we would need to solve a system of linear Diophantine equation over positive integers to check whether there is a set of factors for defined vectors that may match a vector for $P$ with another unknown factor.
\end{proof}
\section{Broadcasting Sequences and their Limitations}
Although the composition of circles gives us a large range of polygonal shapes, an analysis of broadcasting sequences \index{broadcasting sequences}
shows that there certain limitations for such composition.
It's seen that by Lemma~\ref{lineLength} there are only three possible forms which the lines that comprise the discrete circle may take. Translating these in to their respective gradients gives the following three possible gradient forms
$$G_1 = G((10^{n-1})^m) = \frac{m}{mn} = \frac{1}{n} $$
$$G_2 = G((10^{n-1})(10^{n})^m) = \frac{m+1}{n+m(n+1)}$$
$$G_3 = G((10^{n-1})^m(10^{n})) = \frac{m+1}{nm+n+1}\ .$$
The above gradients are reduced to a minimal form in order to elucidate the exclusivity of the gradients.
\begin{proposition}\label{posGrad}
It is only possible to express gradients in the first octant of a circle in a reduced form $\frac{1}{n}$, $\frac{a}{a(n+1)-1}$ or $\frac{a}{an+1}$ for $a,n\in \mathbb N$.
\end{proposition}
\begin{proof}
As $G_1$ is already in a minimal form it is obvious that this can only express those gradients, $g$, of the form $\frac{1}{n}$. For $G_2$ and $G_3$ the following method is employed, $G_2=\frac{m+1}{n+m(n+1)}=\frac{a}{b}$, where $a\bot b$ from which it follows that $m=a-1$ such that through substitution of $m$ in to denominator, $\frac{a}{a(n+1)-1}$ is arrived at. Similarly for $G_3$ by substitution the equation $\frac{a}{an+1}$ arises.
\end{proof}
It now becomes more clear of what cannot be expressed and the limitations inherent in the hulls of the broadcasting sequences.
\begin{proposition}
Not all rational gradients, $0\leq g\leq 1$, are expressed by the lines that comprise the discrete circle in the first octant.
\end{proposition}
\begin{proof}
A counter example is given as proof. It is impossible to express any such rational of the form $\frac{5}{8}$. It is clear that it is not possible for $G_1$ to express such a fraction. For $G_2$ and $G_3$ the following suffices. $G_2=\frac{a}{a(n+1)-1}$ where $a=5$ such that $\frac{5}{5n+4}$ where there is no such $n\in \mathbb N$ such that $G_2=\frac{5}{8}$. Similarly for $G_3=\frac{a}{an+1}$ where $a=5$ and $\frac{5}{5n+1}$ there is no such $n\in \mathbb N$ such that $G_3=\frac{5}{8}$
\end{proof}
It also becomes clear from the composition theorem (Theorem~\ref{compoThe}) itself that it is not possible to generate any further gradients or new line segments through composition and in turn successive broadcasts of radii may not generate any polygons with chain codes that include gradients not previously present.
\begin{corollary}
The set of line segments, and as such gradients, that compose any discrete circle are closed under composition.
\end{corollary}
As noted here the shapes that are generatable are limited by their convexity as well as by the gradients of the lines that compose them. In the next section some of these limitations will be removed or relaxed using multiple broadcast sequences and an aggregation function with which to map all values to a single value.
\section{Reducing Restrictions Through Aggregation}
This section makes uses of the notion of aggregation to reduce the restrictions that are imposed by the continual composition, or simple construction of discrete discs. It can be shown that it is possible in certain cases, which shall be exposited both here and in further sections where it is employed to show the approximation of the astroid metric.
As all points may be labelled by their distance over some arbitrary modulo value an extension to the current work is proposed whereby two differing $r-neighbourhood$ sequences, $A$ and $A'$ are used to label the $\mathbb Z^2$ lattice from the same point $p$. At any point, $p'$, there are now two labels such that $p'=(i,j)$ where by $k \equiv i \mod m$ for the sequence $A$ and $k' \equiv j \mod m$ for the sequence $A'$. Here two differing functions for the aggregation of values of $i$ and $j$ which define new shapes on the lattice and in turn new metrics. The new shapes defined by the lattice are not necessarily convex.
As all of the discrete disks which make up the labellings of the lattice are composed of discrete lines it is possible to analyse the effects of the combinations of disks by considering only the intersections of the lines that make up the disks. Consider a series of parallel lines expressed in the form $y_0=m_0x_0+c_0$ where $c_i\in \mathbb Z$ is an arbitrary constant or offset, $m_i\in \mathbb Q$ is any gradient permitted by the line segments of a discrete circle and $x,y\in \mathbb Z$ are the usual Euclidean coordinates. These lines are such that each successive line is of a distance $w_i$ from the last which shall imitate the width between iterations of the discrete circles, for the first line segment this is equivalent to $\lfloor \sqrt{r^2}\rfloor$, and the discrete lines that they generate. All coordinates such that $m_0x_0+c_0+kw_0\leq y_0< m_0x_0+c_0+(k+1)w_0$ are labelled as $k_0$. A second set of parallel lines differing from the first are defined as $y_1=m_1x_1+c_1$ where all coordinates such that $m_1x_1+c_1+k_1w_1\leq y_1< m_1x_1+c_1+(k_1+1)w_1$ are labelled $k_1$. The intersection of these two areas is thus $(k,k')$. Increasing the offset of $k_0$ and $k_1$ to $k_0+1$ and $k_1+1$ naturally results in the labelling of the area of their intersection as $(k_0+1, k_1+1)$. The ordering of the tuple is not relevant to the functions that aggregate them.
\tn{The first of the functions here have previously been studied in \cite{Geometric} with regards to geometric computations on the lattice. It is of particular interest in the Broadcasting Automata model where the notion of waves as observed in nature are used to control a large number of distributed automata. Being able to predict the resultant shapes that are formed by the transmission of waves is, naturally, largely advantageous in that it affords the ability to predict and manipulate the formations on the plane. Such formations can then be used for a variety of computational duties such as partitioning and geometric computation.
}
Two functions for aggregation will now be introduced with the relevant equations for resultant patterning of the lattice, where $m = 4$ and the values of $i,j \in \{0,1,2,3\}$.
\begin{definition}
The {\bf moir\'{e}} aggregation function is given in the table below.
\begin{minipage}[t]{0.5\linewidth}
\vspace{20pt}
{\it moir\'{e}(i,j)} \ \ \ \ =
\end{minipage}
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{tabular}{|l|l|l|l|l|}
\hline
$\oplus$ & 0 & 1 & 2 & 3 \\ \hline
0 & a & b & c & b \\ \hline
1 & b & a & b & c \\ \hline
2 & c & b & a & b \\ \hline
3 & b & c & b & a \\
\hline
\end{tabular}
\end{minipage}
\end{definition}
\begin{definition}
The {\bf Anti-moir\'{e}} aggregation function can be expressed simply as addition over modulo 4 and is given in the table below.
\begin{minipage}[t]{0.5\linewidth}
\vspace{20pt}
{\it Anti-moir\'{e}(i,j)} \ \ \ \ =
\end{minipage}
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{tabular}{|l|l|l|l|l|}
\hline
$\oplus$ & 0 & 1 & 2 & 3 \\ \hline
0 & a & b & c & d \\ \hline
1 & b & c & d & a \\ \hline
2 & c & d & a & b \\ \hline
3 & d & a & b & c \\
\hline
\end{tabular}
\end{minipage}
\end{definition}
\begin{proposition}
The gradients of the new lines formed by the moir\'{e} equation can be predicted using the equation $$\frac{w_o\cdot m_1- w_1\cdot m_0}{w_0-w_1}\ .$$ Variables refer to the, line gradient, $m_i$, and the width of the line, $w_i$.
\end{proposition}
\begin{proof}
The form of the moir\'{e} function is such that if the intersection of one area is labelled as $(k,k')$ then the next area that is labelled similarly will be of the form $(k+1,k'+1)$. Encoding this as the intersections of lines $y=m_0x+c_0+k\cdot w_0$ with $y=m_1x+c_1+k'\cdot w_1$ and $y=m_0x+c_0+(k+1)\cdot w_0$ with $y=m_1x+c_1+(k'+1)\cdot w_1$. All that is left is to find the gradient of the two intersections. Solving the first case, $x_0=\frac{c_1-c_0-k\cdot w_0-k'\cdot w_1}{m_0-m_1}$ and $y_0=m_0(\frac{c_1-c_0-k\cdot w_0-k'\cdot w_1}{m_0-m_1})+c_0+k\cdot w_0$ . The second, $x_1=\frac{c_1-c_0-(k+1)\cdot w_0-(k'+1)\cdot w_1}{m_0-m_1}$ and $y_1=m_ 0(\frac{c_1-c_0-(k+1)\cdot w_0-(k'+1)\cdot w_1}{m_0-m_1})+c_0+(k+1)\cdot w_0$. The gradient of the two points can be found, $\frac{y_1-y_0}{x_1-x_0}$, which results in the gradient of the labelling applied to the lattice.
\end{proof}
\begin{proposition}
The gradients of the new lines formed by the Anti-moir\'{e} equation can be predicted using the equation $$\frac{w_o\cdot m_1+ w_1\cdot m_0}{w_0+w_1}\ .$$ Variables refer to, line gradient, $m_i$, and the width of the line, $w_i$.
\end{proposition}
\begin{proof}
The form of the moir\'{e} function is such that if the intersection of one area is labelled as $(k,k')$ then the next area that is labelled similarly will be of the form $(k+1,k'+1)$. Encoding this as the intersections of lines $y=m_0x+c_0+k\cdot w_0$ with $y=m_1x+c_1+(k'+1)\cdot w_1$ and $y=m_0x+c_0+(k+1)\cdot w_0$ with $y=m_1x+c_1+k'\cdot w_1$. All that is left is to find the gradient of the two intersections. Solving the first case, $x_0=\frac{c_1-c_0-k\cdot w_0-(k'+1)\cdot w_1}{m_0-m_1}$ and $y_0=m_0(\frac{c_1-c_0-k\cdot w_0-(k'+1)\cdot w_1}{m_0-m_1})+c_0+k\cdot w_0$ . The second, $x_1=\frac{c_1-c_0-(k+1)\cdot w_0-k'\cdot w_1}{m_0-m_1}$ and $y_1=m_0(\frac{c_1-c_0-(k+1)\cdot w_0-k'\cdot w_1}{m_0-m_1})+c_0+(k+1)\cdot w_0$. The gradient of the two points can be found, $\frac{y_1-y_0}{x_1-x_0}$, which results in the gradient of the labelling applied to the lattice.
\end{proof}
Composition of this form is given as example in Figure~\ref{aggregate}. Here the two lines are shown on the lattice correspond to the gradients of the two differing aggregation functions.
The following proposition is now noted.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.50]{aggregation}
\caption{The two forms of aggregation via the above functions. Here pattern A gives the line formed by the anti-moire function and the pattern b gives the line as formed by the moire function.}\label{aggregate}
\end{figure}
\begin{proposition}
For any two discrete lines on the $\mathbb Z^2$ lattice it is possible to vary the observed gradients for resultant from aggregate functions by varying the width of the lines.
\end{proposition}
\begin{proof}
The proof is an observation of the equations for the gradients of the lines resultant from aggregation. Observing the results of two aggregating functions, $\frac{w_o\cdot m_1+ w_1\cdot m_0}{w_0+w_1}$ and $\frac{w_o\cdot m_1- w_1\cdot m_0}{w_0-w_1}$, any alteration to the width of the line which represents results in a direct change in the gradient.
\end{proof}
It is now possible to formulate the following statement which shows the increase in expressivity that comes from using the composition of two broadcast sequences.
The number of lines formed by aggregation are not limited by the restrictions that are demonstrated in Proposition.~\ref{posGrad} where it is now possible to construct new lines through the varying of widths and indeed new polygons may now be formed with this technique such that they are non-convex.
\section{Formation of Polygons Through Aggregation.}
Having shown that it is possible to construct different gradients from line sections it is natural to now observe what happens when whole circles are intersected and the relationships that are formed by the aggregation functions that have been introduced. It is noted that a diagram for the formation of polygons in case of the moir\'{e} aggregation function is shown in Figure~\ref{moireComp}. Two digital discs, those with squared radii two and 25 such that the discs are $\zeta^2$ and $\zeta^{25}$, are composed generating, from the two convex polygons, and new, previously unreachable, polygon which is $non-convex$.
\begin{figure}[htp]
\includegraphics[scale=0.6]{moireCompNN.eps}
\caption{The above schematic depicts the broadcast of two discrete circles. The first, inner, circle (of squared radius two where $\zeta^2 = l_0$ with gradient $m_0$) and the second discrete circle (of squared radius $25$ where $\zeta^{25} = l'_0 l'_1$ with gradients $m_1$ and $m_2$ respectively) the outer construction shows the resulting moir\'{e} lines here labelled $m_3$ and $m_4$. Arrowed lines show line width measurements, the respective $w_i$, here, $w_0=1$, $w_1=5$, $w_2=7$, $w_3=\frac{5}{4}$ and $w_4=\frac{7}{6}$.}\label{moireComp}
\end{figure}
The following examples given in Figure~\ref{moantimo} also elucidate the differences between the two aggregation functions showing the non-convex polygon generated by the two digital discs that are represented by, $\zeta^2$ and $\zeta^9$.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.10]{mo.png}
\includegraphics[scale=0.10]{antimo.png}
\caption{Above gives an example of aggregation of two broadcast sequences, one of the discrete disc $ \zeta^{32}$ and the other $\zeta^{36}$, from a central point on the diagram and with the two aggregation functions (left) moir\'{e} and (right) Anti-moir\'{e}.}\label{moantimo}
\end{figure}
It is also possible to gain a better picture of the overall shapes produced by the anti moire equation by reducing the number of labels, and so the colours, further. This is done by a process of merging certain values or reducing the values over some modulo which in this example, Figure~\ref{antimocomb}, is two. The reduction is given by the following aggregation function:
\begin{minipage}[t]{0.5\linewidth}
\vspace{20pt}
{\it Anti-moir\'{e}-Mod2(i,j)} \ \ \ \ =
\end{minipage}
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{tabular}{|l|l|l|l|l|}
\hline
$\oplus$ & 0 & 1 & 2 & 3 \\ \hline
0 & a & b & b & a \\ \hline
1 & b & b & a & a \\ \hline
2 & b & a & a & b \\ \hline
3 & a & a & b & b \\
\hline
\end{tabular}
\end{minipage}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.2]{antimocomb.png}
\caption{The above figure depicts the anti moire pattern after the merging of two sets of two colours and initially generated by the discrete discs of squared radius $32$ and $36$ from the central point on the plane.}\label{antimocomb}
\end{figure}
\section{Approximating $L^p$ Metrics with Broadcasting Automata}\label{Lp_Metrics}
Previously, von Neumann and Moore neighbourhoods have been used to achieve an approximation of the Euclidean metric through some mixing of the neighbourhoods. This has been done in many different settings such as periodic, non-periodic combinations of the two neighbourhoods, regular and non-regular, i.e, hexagonal, triangular grids to which different sequences, i.e applying different definitions of neighbourhood, are applied, as well as a variety of methods for defining just how an approximation of Euclidean distance by neighbourhood sequences should be defined and measured, where most of these techniques discuss notions of digital circularity such as the isoperimetric ratio, perimeter comparisons, etc. In short this has been one of the main studies with regards to neighbourhood sequences. There is ultimately, as has been previously discussed, a large barrier to the extension of this body of work in the approximation of the more general $L^p$ metrics due to the impossibility of constructing any non-convex polygon from the composition of the two convex polygons which represent the Moore and von Neumann neighbourhoods. The astroid\index{astroid} is part of the `family' of $L^p$ metrics of which the Moore neighbourhood, $L^{\infty}$, the von Neumann neighbourhood, $L^1$, and the Euclidean metric, $L^2$ are the most well known. In general an $L^p$ space\index{$L^p$ metrics} is defined by the formula $\parallel x\parallel_p=(|x_1|^p+|x_2|^p+...+|x_n|^p)^{\frac{1}{p}}$ such that $\parallel x\parallel_p$ is the $p$-norm.
It is in this section that a new method for the generalisation of metric approximation\index{metric approximation} with broadcasting neighbourhoods shall be given and, using only two broadcasting radii, here given in terms of $r^2$, and the moir\'{e} aggregating function, explore the ability of this model to approximate the astroid and, as such, a new class of metrics outside of the reach of neighbourhood sequences.
It has previously been seen that the methodology of combining two broadcasting sequences using an aggregating function can be used to extend the possible resulting polygons. In this section it shall be seen that this can be employed in the solution to a practical problem. The problem, in this case, is the approximation of $L^p$ metric, $L^{2/3}$, which forms an \emph{astroid} \cite{yates1974curves}. The astroid can be expressed by the equation, $x^{2/3}+y^{2/3}=r^{2/3}$, where $x$ and $y$ are those of the Cartesian plane and, as with the equation for the circle from which this equation is generalised, the $r$ is the 'radius' of the astroid.
A few preliminaries with regards to the astroid which are taken from \cite{yates1974curves} will be required to understand the methods proposed for comparing the approximation, via broadcasting sequence and aggregation, and the actual astroid. The astroid was discovered in Roemer in 1674 whilst searching for the best form of gear teeth. It is a hypercycloid of four cups and can be described by a point on a circle of radius $\frac{3}{4}\cdot a$ rolling on the inside of a fixed circle of radius $a$. The resultant shape has a perimeter, $L$, of $L=6a$ where $a$ is the radius of the outer fixed circle and is the maximal point reached by the astroid. The area, $A$, encapsulated is given by $A=\frac{3}{8}\pi a^2$. Finally, the point at which the $x$ and $y$ coordinates are equivalent on the perimeter of the astroid can be expressed as $\frac{r}{2}$. The equation is known to have applications in magnetism where the Stoner-Wohlfarth astroid curve separates regions with two minima of free energy density from those with only one energy minimum and is a geometric representation of the model of the same name \cite{Thiaville19985}.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.32]{astroidr.png}
\includegraphics[scale=0.4]{r5.png}
\caption{The above figure depicts the astroid, here, rotated by $45^\circ$ (left) compared to its best, found, approximation with broadcasting sequences (right), an aggregation of the discrete discs of squared radius, $2$ and $5$. }\label{astroid}
\end{figure}
In \cite{springerlink:10.1007/117749388, Hajdu2004101} a number of methods for how well a particular neighbourhood sequence approximates the euclidean distance are given. The authors of \cite{springerlink:10.1007/117749388} consider any sequence, including those sequences which are non-periodic, of neighbourhoods with which to approximate the euclidean distance. In order to show how well, or poorly, the sequence approximates the euclidean distance it is compared to the euclidean circle through a variety of methods. One such method is through the use of the isoperimetric ratio or the noncompactness ratio. Here, the attempt is to measure $\frac{P^2}{A}$ where, $P$, is the perimeter, and, $A$, is the area. This measure is conjectured to be minimal for the circle where it is $4\pi$, however, this does not help when looking at non-convex shapes, such as those produced when approximating the an astroid, due to the measure being optimised for convex and symmetrical shapes.
Other ways of approximating the euclidean circle are suggested such as a perimeter based approximation and area based approximation. With respects to are based approximation there are two techniques used. The first is that of the inscribed circle based approximation. This method attempts to find a sequence that generates the polygon, generated by the neighbourhood sequence, which is closest to the polygon having the given circle as the inscribed circle. The second of these methods is the covering circle based approximation such that polygon must be covered by the circle. More methods are given but rely on properties distinct to the circle and so are not discussed here.
Further, in \cite{Hajdu2004101}, another measure of circularity is considered using three differing methods. They formulate three approximation problems whereby the aim is to find the neighbourhood sequence that minimises the error in each formulation. The problems may be informally described as the following: problem one requires finding the neighbourhood sequence that best minimises the size of the symmetric difference between some neighbourhood sequence at step $k$, given as $A_k$ and the Euclidean circle of radius $k$; the second problem attempts to find the sequence that best minimises the complement of the neighbourhood generated by $A_k$ with it's smallest inscribed circle; the third problem is a discretisation of the second problem where an $A_k$ must be found such that it minimises the complement with the largest discrete disc that can be inscribed within.
It is the second of these problems from \cite{Hajdu2004101} that shall be explored as a method for showing a best approximation of the astroid with combined broadcasting sequences. In this case the method is the most simplistic to calculate and also the least dependent on any of the properties that are only present in the circle. As such a more formal representation of the problem can now be posed, here, given as, $Area(H(f_k(A, B))\backslash G'_k\leq Area(H(f_{k'}(A, B'))\backslash G'_{k'})$, where, $f_k(A, B)$ is the $k$th polygon generated by the aggregation of the broadcasting sequences, $A$, which here will be the discrete disc, $r^2=2$, and the broadcasting sequence, $B,B'$, here a variable which is to be optimised to find the best approximation. There is a simple change to $G_k$, originally used in \cite{Hajdu2004101} to represent the circle of radius $k$, to convert it to $G'_k$ in that $G'_k=\{q\in \mathbb R^2:L^{\frac{2}{3}}(0,q)\leq k\}$ as simple conversion of the metric from that of the euclidean distance, $L^2$, to the astroid distance, $L^{\frac{2}{3}}$.
The complexity of calculating these compositions for the purposes of optimisation means that only experimental data shall be given here as a proof of validity of the approximation of the concave metrics, which the astroid represents. As the astroids require a rotation by $45^\circ$ in order to match the polygon that is generated by the aggregation, $f(A, B)$, as defined before. In matching the point that is at the largest euclidean distance from the origin, which for simplicity, and without loss of generality, is considered the initial point from which all broadcasting occurs. This point is matched to the same point on some polygon on some composition, $f(A, B)$, and the complement of the areas compared. The following table is produced with this method.
The function of aggregation must also be defined. Here the choice is moir\'{e} aggregation without the modulo restriction, although, such a restriction is retained in the images for simplicity. Such a function can now be defined as, $f(A_i,B_j) = |i-j|$ for the $i$th and $j$th iteration of the sequence $A$ and $B$ respectively. For the function $f_k(A,B)$ where there exists some $A_i$ and $B_j$ such that $|i-j|=k$.
The min point for the $L^{\frac{2}{3}}$ is calculated by $\frac{r}{2}$ where $r$ is the radius of the astroid, the $r$ in $x^{\frac{2}{3}}+y^{\frac{2}{3}}=r^{\frac{2}{3}}$. The $B$ is the second broadcasting sequence, here, given as the $r^2$ of the disc. Images for each of the resultant, aggregated images are given in Table~\ref{astpic1} and Table~\ref{astpic2}.
\begin{table}[position specifier]
\centering
\begin{tabular}{| c | c | c | c | c | c | c | c | c |} \hline
\multicolumn{3}{|c|}{Astroid - $L^\frac{2}{3}$} & \multicolumn{3}{|c|}{moir\'{e} - $f_k(A,B)$} & \multicolumn{3}{|c|}{}\\ \hline
Radius & Min & Area & Radius & Min & Area & B & k & Complement \\ \hline
17 & 8.5 & 340 & 17 & 11 & 461 & 5 & 5 & 121 \\ \hline
17 & 8.5 & 340 & 17 & 13 & 753 & 9 & 8 & 413 \\ \hline
17 & 8.5 & 340 & 17 & 13 & 873 & 10 & 9 & 533 \\ \hline
18 & 9 & 382 & 18 & 16 & 1121 & 13 & 11 & 739 \\ \hline
18 & 9 & 382 & 18 & 14 & 1033 & 16 & 11 & 651 \\ \hline
18 & 9 & 382 & 18 & 14 & 1001 & 17 & 11 & 619 \\ \hline
17 & 8.5 & 340 & 17 & 15 & 1041 & 37 & 13 & 701 \\ \hline
17 & 8.5 & 340 & 17 & 16 & 1141 & 45 & 14 & 801 \\ \hline
17 & 8.5 & 340 & 17 & 16 & 1093 & 61 & 14 & 753 \\ \hline
17 & 8.5 & 340 & 17 & 16 & 1181 & 82 & 15 & 841 \\ \hline
\end{tabular}
\caption{Illustrating the experiments done with regards to the approximation of the astroid.}\label{compAst}
\end{table}
\begin{table}[h!]
\centering
\begin{tabular}{c c}
\includegraphics[scale=0.35]{r5.png}
&
\includegraphics[scale=0.35]{r9.png}
\\
\includegraphics[scale=0.35]{r10.png}
&
\includegraphics[scale=0.35]{r13.png}
\\
\end{tabular}
\caption{The above diagrams show the patterns generated by the aggregation of the two discs, where one is the disc of squared radius $2$ and the other is varied, in these diagrams (from left to right and line by line) there are discs of squared radius, $5$, $9$, $10$ and $13$.}\label{astpic1}
\end{table}
\begin{table}[position specifier]
\centering
\begin{tabular}{c c}
\includegraphics[scale=0.35]{r16.png}
&
\includegraphics[scale=0.35]{r17.png}
\\
\includegraphics[scale=0.35]{r37.png}
&
\includegraphics[scale=0.35]{r45.png}
\\
\includegraphics[scale=0.35]{r61.png}
&
\includegraphics[scale=0.35]{r82.png}
\\
\end{tabular}
\caption{The above diagrams show the patterns generated by the aggregation of the two discs, where one is the disc of squared radius $2$ and the other is varied, in these diagrams (from left to right and line by line) there are discs of squared radius, $16$, $17$, $37$, $45$, $61$ and $82$.}\label{astpic2}
\end{table}
Tables~\ref{compAst} and ~\ref{compAst2} give a series of comparisons for $Area(H(f_k(A, B)))\backslash G'_k$. From this table it is possible to observe that the best approximation, from those constructed, though there is a trend towards worsening approximations as $B$ increases, that the simplest composition yields the best results. In this case this value for $B$ is the disc generated by the squared radius $r^2$ of $5$, this disc being constructed by the next smallest squared radius that constructs a new discrete disc. The following table now looks, again, experimentally, at how the approximation changes as $k$ increases. The table notes that the approximation weakens as it increases perhaps indicating a divergence between the astroid and the polygon generated by $f_K(A,B)$.
\begin{table}[position specifier]
\centering
\begin{tabular}{| c | c | c | c | c | c | c | c | c |} \hline
\multicolumn{3}{|c|}{Astroid - $L^\frac{2}{3}$} & \multicolumn{3}{|c|}{moir\'{e} - $f_k(A,B)$} & \multicolumn{3}{|c|}{}\\ \hline
Radius & Min & Area & Radius & Min & Area & B & k & Complement \\ \hline
2 & 1 & 5 & 2 & 1 & 13 & 5 & 1 & 8 \\ \hline
5 & 2.5 & 30 & 5 & 3 & 65 & 5 & 2 & 35 \\ \hline
8 & 4 & 75 & 8 & 5 & 157 & 5 & 3 & 82 \\ \hline
11 & 5.59 & 143 & 11 & 7 & 289 & 5 & 4 & 146 \\ \hline
18 & 9 & 382 & 18 & 9 & 461 & 5 & 5 & 79 \\ \hline
21 & 10.5 & 520 & 21 & 11 & 673 & 5 & 6 & 153 \\ \hline
24 & 12 & 679 & 24 & 13 & 925& 5 & 7 & 246 \\ \hline
27 & 13.5 & 859 & 27 & 15 & 1217& 5 & 8 & 358 \\ \hline
30 & 15 & 1060 & 30 & 17 & 1549& 5 & 9 & 489 \\ \hline
33 & 16.5 & 1283 & 33 & 19 & 1921& 5 & 10 & 628 \\ \hline
\end{tabular}
\caption{Illustrating the experiments done with regards to the approximation of the astroid where both $A$ and $B$ are fixed.}\label{compAst2}
\end{table}
\section{Pattern Formations and Periodic Structures in $\mathbb Z^2$}
It is natural after observing the variety of effects that are the result of the application of an aggregation function to look at functions which are themselves shapes of some form. Here, functions of this form and their resultant patterns, imposed on the grid according to their application, are given. Such functions are only restricted by their symmetry, a result of the unordered nature of the tuples to be aggregated.
The first function, depicted in Figure~\ref{discreteDisc882636} (Left) takes the form of a discrete disc itself, in this case one which is also represented by the discrete disc of squared radius five. The functions here have also been increased in size and the size of the modulo for the labels has been increased. The use of discrete discs of squared radius 8 is important here as it constructs, in some sections of the lattice, a perfect reproduction of the shape given in the aggregating function. The following table describes the aggregating function and the details of the image.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.12]{discreteDisc.png}
\includegraphics[scale=0.12]{discreteDisc2636.png}
\caption{(Left) Patterns generated by the aggregating function representing the discrete disc of squared radius five and the labelling of the lattice given by two broadcasting sequences of squared radius eight. (Right) Patterns generated by the aggregating function representing the discrete disc of squared radius five and the labelling of the lattice given by two broadcasting sequences, one of squared radius 26 and the other of squared radius 36. }\label{discreteDisc882636}
\end{figure}
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{description}
\item[Array Size:] 300
\item[Centre 1:] (100,150)
\item[Centre 2:] (200,150)
\item[Radius 1:] 8
\item[Radius 2:] 8
\item[Modular Labelling:] (0,1,2,3,4,5,6)
\item[Aggregation Function:] Shown right.
\item[Figure~\ref{discreteDisc882636} (Left)]
\end{description}
\end{minipage}
\hfill
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
$\oplus$ & 0 & 1 & 2 & 3 & 4 & 5 & 6\\ \hline
0 & a & a & a & a & a & a & a\\ \hline
1 & a & a & b & b & b & a & a\\ \hline
2 & a & b & b & a & b & b & a\\ \hline
3 & a & b & a & a & a & b & a\\ \hline
4 & a & b & b & a & b & b & a\\ \hline
5 & a & a & b & b & b & a & a\\ \hline
6 & a & a & a & a & a & a & a\\ \hline
\end{tabular}
\end{minipage}
\vspace{0.4cm}
This reproduction of the aggregating function is not always exact. It is possible to skew and deform the representation of the image described by altering the radii of the circles that are used to form the underlying labelling of the lattice. The following image, Figure~\ref{discreteDisc882636} (right) gives an example of such a deformation.
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{description}
\item[Array Size:] 300
\item[Centre 1:] (100,150)
\item[Centre 2:] (200,150)
\item[Radius 1:] 26
\item[Radius 2:] 36
\item[Modular Labelling:] (0,1,2,3,4,5,6)
\item[Aggregation Function:] Shown right.
\item[Figure~\ref{discreteDisc882636} (Right)]
\end{description}
\end{minipage}
\hfill
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
$\oplus$ & 0 & 1 & 2 & 3 & 4 & 5 & 6\\ \hline
0 & a & a & a & a & a & a & a\\ \hline
1 & a & a & b & b & b & a & a\\ \hline
2 & a & b & b & a & b & b & a\\ \hline
3 & a & b & a & a & a & b & a\\ \hline
4 & a & b & b & a & b & b & a\\ \hline
5 & a & a & b & b & b & a & a\\ \hline
6 & a & a & a & a & a & a & a\\ \hline
\end{tabular}
\end{minipage}
\vspace{0.4cm}
The two following figures, Figure~\ref{b}, demonstrates changes that occur when manipulating the number of colours, changing the aggregation function and altering the underlying tuples that generate the overall shape of the colourings.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.12]{b88.png}
\includegraphics[scale=0.12]{b2636.png}
\caption{Patterns generated by the aggregating function represented by the table and the labelling of the lattice given by two broadcasting sequences, (right) both of squared radius eight (left) one of squared radius 26 and the other of squared radius 36.}\label{b}
\end{figure}
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{description}
\item[Array Size:] 300
\item[Centre 1:] (100,150)
\item[Centre 2:] (200,150)
\item[Radius 1:] 8
\item[Radius 2:] 8
\item[Modular Labelling:] (0,1,2,3,4,5)
\item[Aggregation Function:] Shown right.
\item[Figure~\ref{b} (left)]
\end{description}
\end{minipage}
\hfill
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
$\oplus$ & 0 & 1 & 2 & 3 & 4 & 5 \\ \hline
0 & b & b & c & c & b & b \\ \hline
1 & b & c & a & a & c & b \\ \hline
2 & c & a & a & a & a & c \\ \hline
3 & c & a & a & a & a & c \\ \hline
4 & b & c & a & a & c & b \\ \hline
5 & b & b & c & c & b & b \\ \hline
\end{tabular}
\end{minipage}
\vspace{0.4cm}
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{description}
\item[Array Size:] 300
\item[Centre 1:] (100,150)
\item[Centre 2:] (200,150)
\item[Radius 1:] 26
\item[Radius 2:] 36
\item[Modular Labelling:] (0,1,2,3,4,5)
\item[Aggregation Function:] Shown right.
\item[Figure~\ref{b} (rigth)]
\end{description}
\end{minipage}
\hfill
\begin{minipage}[t]{0.5\linewidth}
\vspace{0pt}
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
$\oplus$ & 0 & 1 & 2 & 3 & 4 & 5 \\ \hline
0 & b & b & c & c & b & b \\ \hline
1 & b & c & a & a & c & b \\ \hline
2 & c & a & a & a & a & c \\ \hline
3 & c & a & a & a & a & c \\ \hline
4 & b & c & a & a & c & b \\ \hline
5 & b & b & c & c & b & b \\ \hline
\end{tabular}
\end{minipage}
\vspace{0.4cm}
\begin{figure}[htp]
\centering
\includegraphics[scale=0.9]{interferenceEx.png}
\caption{The above figure shows a number of distinct patterns that are here generated by four different transmissions where two, one of squared radius $2$ and the other of squared radius $12$, are placed at two separate points.}\label{intEx}
\end{figure}
Altering the aggregating function is clearly a powerful tool in pattern and polygon formation. The alteration of the discs that are used as the basis of the aggregation also show that any underlying shape generate by an aggregating function can be skewed and otherwise altered whilst retaining the gestalt representation. Methods of manipulating the hew and scale of the shapes that are generated through some aggregating function may be useful to pattern recognition and detection methods that are part of the Swarm Robotics cannon among others. Whilst altering the aggregation function is one an interesting concept, for the purpose of pattern formation, there is still a lot of possibility that remains when only considering the standard moir\'{e} function. Such as can be seen in the variety of patterns that are formed in Figure~\ref{intEx}.
\input{ref}
\printindex
\end{document}
|
\section{Introduction}
Bayesian nonparametric mixtures have become a standard tool for inference when a distribution of either observable or unobservable quantities is considered unknown. A more challenging problem, which arises in many applications, is to define a prior for a collection of related unknown distributions. For example, \cite{mulros04} consider informing the analysis of a study with results from previous related studies. They considered the CALGB 9160 \citep{CALGB9160} clinical study which looked at the response over time of patients to different anticancer drug therapies. \cite{mulros04} suggested improving the precision of their inference using the results of the related study
CALGB 8881\citep{CALGB8881}.
Figure~\ref{f:data1} shows bivariate plots of two subject-specific regression parameters ($\beta_0$ and $\beta_1$) for the two studies.
\begin{figure}[h!]
\begin{center}
\includegraphics[trim=0mm 0mm 120mm 230mm, clip]{calgb_data}
\end{center}
\caption{Scatter-plots of the subject-specific regression parameters $\beta_0$ and $\beta_1$ for the groups in CALGB 8881 and CALGB 9160.}\label{f:data1}
\end{figure}
The graphs suggest differences between the joint distribution of $\beta_0$ and $\beta_1$ which should be included in any analysis which combines these data sets. The results for CALGB9160 also suggest that a nonparametric model is needed to fully describe the shape of the density. A natural Bayesian approach would assume different distributions for each study but construct a dependent prior for these distributions.
In general, suppose that $x\in\mathcal{X}$ denotes the value of covariates then, in a Bayesian nonparametric analysis, a prior needs to be defined across a collection of correlated distributions $\{\tilde{p}_x\vert x\in\mathcal{X}\}$. This problem was initially studied in a seminal paper on dependent Dirichlet processes \citep{MacEachern} where generalisations of the Dirichlet process were proposed. Subsequent work used stick-breaking constructions of random measures as a basis for defining such a prior. This work is reviewed by \cite{dun10}. These priors can usually be represented as
\begin{equation}
\tilde{p}_x=\sum_{i=1}^{\infty} w_i(x)\delta_{\theta_i(x)}
\label{DDP}
\end{equation}
where
$w_1(x),w_2(x),\dots$ follow a stick-breaking process for all $x\in\mathcal{X}$.
A drawback with this approach is the stochastic ordering of the $w_i(x)$'s for any $x\in\mathcal{A}$ which can lead to strange effects in the prior as $x$ varies.
If $\mathcal{A}$ is countable, several other approaches to defining a prior on a collection of random probability measures have been proposed. The
Hierarchical Dirichlet process (HDP) \citep{tehjor06} assumes that $\tilde{p}_x$ are {\it a priori} conditionally independent and identically distributed according to a Dirichlet process whose centring measure is itself given a Dirichlet process prior. This construction induces correlation between the elements of $\{\tilde{p}_x\vert x\in\mathcal{A}\}$ in the same way as in parametric hierarchical models. This construction can be extended to more general hierarchical frameworks
\cite[see {\it e.g.}][for a review]{tehjord}. Alternatively, a prior can be defined using the idea of normalized random measures with independent increments which are defined by normalising a completely random measure. The prior is defined on a collection of correlated completely random measures $\{\tilde\mu_x\vert x\in\mathcal{A}\}$ which are then normalized for each of $x$, {\it i.e.} $\tilde{p}_x=\tilde\mu_x/\tilde\mu_x(\mathbb{X})$ where $\mathbb{X}$ is the support of $\tilde\mu_x$. Several specific constructions have been proposed including various forms of superposition \citep{GKS13, NL14,NLP14a,NLP14b, Chen13, BaCaLe}, the kernel-weighted completely random measures \citep{fotwil12, Gri11} and L\'evy copula-based approaches \citep{LL, LLS, ZL}. In this paper, we develop an alternative method for constructing correlated completely random measures which is tractable, whose properties can be derived and for which sampling methods for posterior inference without truncation can be developed. The construction also provides a unifying framework for previously proposed constructions. Indeed, the $\sigma$-stable and gamma vector of dependent random measures, studied in the recent works of \cite{LL}, \cite{LLS} and \cite{ZL} are special cases. Although these papers derive useful theoretical results, their application has been limited by the lack of a sampling methods for posterior inference. The algorithms proposed in this paper can also be used for posterior sampling for these nonparametric priors which is another contribution of the paper.
The paper is organized as follows. Section 2 introduces the concepts of completely random measures, normalized random measures and their multivariate extensions. Section 3 discusses the construction and some properties of a new class of multivariate L\'evy process, \emph{Compound Random Measures}, defined by a {\it score distribution} and a {\it directing L\'evy process}. Section 4 provides a detailed description of Compound Random Measures with a gamma score distribution. Section 5 considers the use of normalized version of Compound Random Measures in nonparametric mixture models including the description of a Markov chain Monte Carlo scheme for inference. Section 6 provides an illustration of the use of these methods in an example and Section 7 concludes.
\section{Preliminaries}
Let $(\Omega,\mathcal{F},\P)$ be a
probability space and $(\mathbb{X},\mathcal{X})$ a measure space, with $\mathbb{X}$
Polish and $\mathcal{X}$ the Borel $\sigma$--algebra of subsets of $\mathbb{X}$. Denote by $\mathbb{M}_{\mathbb{X}}$ the space of boundedly finite measures on $(\mathbb{X},\mathcal{X})$, {\it i.e.} this means that for any $\mu$ in $\mathbb{M}_{\mathbb{X}}$ and any bounded set $A$ in $\mathcal{X}$ one has $\mu(A)<\infty$. Moreover, $\cal{M}_{\mathbb{X}}$ stands for the corresponding Borel $\sigma$--algebra, see \cite{daley} for technical details.
The concept of a \textit{completely random measure} was introduced by
\cite{Kingman67}.
\begin{dfn}
Let $\tilde{\mu}$ be a measurable mapping from $(\Omega,\mathcal{F},\P)$ into ($\mathbb{M}_{\mathbb{X}}$,$\cal{M}_{\mathbb{X}}$) and such that for any $A_1,\dots,A_n$ in $\mathcal{X}$, with $A_i\cap A_j=\emptyset$ for any $i\neq j$, the random variables $\tilde{\mu}(A_1), \dots, \tilde{\mu}(A_n)$ are mutually independent. Then $\tilde{\mu}$ is called a \textit{completely random measure} (CRM).
\end{dfn}
\noindent A CRM can always be represented as a sum of two components:
$$\tilde{\mu}=\tilde{\mu}_c+\sum_{i=1}^{M}V_i\delta_{x_i}$$
where the fixed jump points $x_1,\dots, x_M$ are in $\mathbb{X}$ and the non-negative random jumps $V_1,\dots,V_M$ are both mutually independent and independent from $\tilde{\mu}_c$. The latter is a completely random measure such that
$$\tilde{\mu}_c=\sum_{i=1}^{\infty} J_i\delta_{X_i}$$
where both the positive jump heights $J_i$'s and the $\mathbb{X}$-valued jump locations $X_i$'s are random. The measure $\tilde{\mu}_c$ is characterized by the \textit{L\'evy-Khintchine} representation which states that
$$\mathbb{E}\left[e^{-\int_X f(x)\tilde{\mu}_c(dx)}\right]
=e^{-\int_0^{\infty}\int_X [1-e^{-sf(x)}]\bar\nu(\mathrm{d} s,\mathrm{d} x)}
$$
where $f:\mathbb{X}\rightarrow\mathbb{R}^{+}$ is a measurable function such that $\int f\tilde{\mu}_c<\infty$ almost surely and $\bar\nu$ is a measure on $\mathbb{R}^{+}\times\mathbb{X}$ such that
$$
\int_{\mathbb{R}^{+}}\int_B \min\{1,s\}\bar\nu(\mathrm{d} s,\mathrm{d} x)<\infty$$
for any $B$ in $\mathcal{X}$. The measure $\bar\nu$ is usually called the L\'evy intensity of $\tilde{\mu}_c$. Throughout the paper, we will consider completely random measures without the fixed jump component ({\it i.e.} $M=0$).
For our purposes, we will focus on the \textit{homogeneous} case, {\it i.e.} L\'evy intensities where the height and location contributions are separated. Formally,
$$ \bar\nu(\mathrm{d} s,\mathrm{d} x)=\rho(\mathrm{d} s)\alpha(\mathrm{d} x)$$
where $\rho$ is a measure on $\mathbb{R}^+$ and
$\alpha$ is a non-atomic measure on $\mathbb{X}$, which is usually called the {\it centring measure}. Some famous examples are the \textit{Gamma} process,
$$\bar\nu(\mathrm{d} s,\mathrm{d} x)=s^{-1}e^{-s}\mathrm{d} s\,\alpha(\mathrm{d} x),$$
the \textit{$\sigma$-stable} process,
$$\bar\nu(\mathrm{d} s,\mathrm{d} x)=\frac{\sigma}{\Gamma(1-\sigma)}s^{-1-\sigma} \mathrm{d} s\,\alpha(\mathrm{d} x),\quad 0<\sigma<1,$$
and the homogeneous \textit{Beta} process,
\[
\bar\nu(\mathrm{d} s,\mathrm{d} x)= \theta s^{-1}(1-s)^{\theta-1}\mathrm{d} s\,\alpha(\mathrm{d} x),\quad 0<s<1,\quad \theta>0.
\]
A general class of processes that includes the gamma and $\sigma$-stable process is the \textit{Generalized Gamma} process,
$$\bar\nu(\mathrm{d} s,\mathrm{d} x)=\frac{\sigma}{\Gamma(1-\sigma)}s^{-1-\sigma} e^{-as}\mathrm{d} s\,\alpha(\mathrm{d} x),\quad 0<\sigma<1,\quad a>0$$
Random measures are the basis for building Bayesian nonparametric priors.
\begin{dfn}
Let $\tilde{\mu}$ be a measure in ($\mathbb{M}_{\mathbb{X}}$,$\cal{M}_{\mathbb{X}}$). A \textit{Normalized Random Measure} (NRM) is defined as $\tilde{p}=\frac{\tilde{\mu}}{\tilde{\mu}(\mathbb{X})}$.
\end{dfn}
\noindent The definition of a normalized random measure is very general and does not require that the underlying measure is completely random. The Pitman-Yor process (see \cite{PY97}) is
a well-known example of a Bayesian nonparametric priors which cannot be derived by normalizing a completely random measure. In this particular case, the unnormalized measure is obtained through a change of measure of a $\sigma$-stable process.
However, many common Bayesian nonparametric priors can be defined as a normalization of a CRM and many other processes can be derived by normalising processes derived from CRMs, see \cite{rlp}. For instance,
it can be shown that the \textit{Dirichlet Process}, introduced by \cite{ferg}, is a normalized gamma process. Throughout the paper, we will assume that the underlying measure is a CRM and use the acronym NMRI (Normalized Random Measures with independent increments) to emphasize the independence of a CRM on disjoint intervals.
Although nonparametric priors based on normalization are extremely flexible, in many
real applications data arise under different conditions and hence
assuming a single prior can be too restrictive. For example,
using covariates, data may be divided into different units. In this
case, one would like to consider different distributions for different
units instead of a single common distribution for all the units. In these situations, it is more reasonable to consider vectors of dependent random probability measures.
\subsection{Vectors of normalized random measures} Suppose $\tilde\mu_1,\dots, \tilde\mu_d$ are homogeneous CRMs on $(\mathbb{X},\mathcal{X})$
with respective marginal L\'evy intensities
\begin{equation}\label{Homo}
\bar{\nu}_j(\mathrm{d} s, \mathrm{d} x)=\nu_j(\mathrm{d} s)
\,\alpha(\mathrm{d} x),
\qquad \quad j=1,\dots,d.
\end{equation}
where $\nu_j$ is a measure on $\mathbb{R}^+$ and
$\alpha$ is a non-atomic measure on $\mathbb{X}$. Furthermore, $\tilde\mu_1, \dots, \tilde\mu_d$ are dependent and the random vector
$(\tilde\mu_1,\dots,\tilde\mu_d)$ has independent increments, in the sense that for any $A_1,\dots,A_n$ in $\mathcal{X}$, with $A_i\cap A_j=\emptyset$ for any $i\neq j$, the random vectors $(\tilde\mu_1(A_i),\dots,\tilde\mu_d(A_i))$ and
$(\tilde\mu_1(A_j),\dots, \tilde\mu_d(A_j))$ are independent. This implies that
for any set of measurable functions $\bm{f}=(f_1,\dots,f_d)$ such that $f_j:\mathbb{X}\to\mathbb{R}^{+}$, $j=1,\dots,d$ and $\int |f_j|\,d\tilde{\mu}_j<\infty$, one has a multivariate analogue of the L\'evy-Khintchine representation (see \cite{sato}, \cite{daley} and \cite{epifani})
\begin{equation}
\label{eq:biLaplace}
\mathbb{E}\left[e^{-\tilde\mu_1(f_1)-\cdots-\tilde\mu_d(f_d)}\right]=
\exp\left\{-\psi^*_{\rho,d}(\bm{f})\right\}
\end{equation}
where $\tilde\mu_j(f_j)=\int f_j\,d\tilde{\mu}_j$,
\begin{equation}
\label{ExpLevy}
\psi^*_{\rho,d}(\bm{f})=\int_\mathbb{X}\int_{(0,\infty)^d}
\left[1-\mathrm{e}^{-s_{_1} f_1(x)-\cdots-s_{d} f_d(x)}\right]\;\rho_d(\mathrm{d} s_1,\dots,\mathrm{d} s_d)
\:\alpha(\mathrm{d} x)
\end{equation}
and
\begin{equation}
\label{eq:marginals1}
\int_{(0,\infty)^{d-1}}\rho_d(\mathrm{d} s_1,\dots,ds_{j-1},A,ds_{j+1},\dots,ds_d)=
\int_A \nu_j(ds).
\end{equation}
The representation (\ref{Homo}) implies that the jump heights of $(\tilde\mu_1,\dots,\tilde\mu_d)$ are independent from the jump locations. Moreover, these jump locations are common to all the CRMs and are governed by $\alpha$. It is worth noting that, since $(\tilde\mu_1,\ldots,\tilde\mu_d)$ has independent increments, its distribution is characterized by a choice of $f_1,\ldots,f_d$ in \eqref{eq:biLaplace} such that $f_j=\lambda_j\,1_{A}$ for any set $A$ in $\mathcal{X}$, $\lambda_j\in\mathbb{R}^+$ and $j=1,\ldots,d$. In this case
\[
\psi^*_{\rho,d}(\bm{f})=\alpha(A)\,\psi_{\rho,d}(\bm{\lambda})
\]
where $\bm{\lambda}=(\lambda_1,\ldots,\lambda_d)$ and
\begin{equation}
\psi_{\rho,d}(\bm{\lambda})=\int_{(\mathbb{R}^+)^d}
\left[1-\mathrm{e}^{-\langle\bm{\lambda},\bm{s}\rangle}
\right]\rho_d(ds_1,\dots,ds_d)
\label{eq:lapl_exp_star}
\end{equation}
where $\bm{s}=(s_1,\ldots,s_d)$ and $\langle\bm{\lambda},\bm{s}\rangle=\sum_{j=1}^d \lambda_j s_j $.
We close the section with the definition of \textit{vectors of normalized random measures with independent increments}.
\begin{dfn}\label{DefVec}
Let ($\tilde\mu_1,\dots, \tilde\mu_d$) be a vector of CRMs on $\mathbb{X}$ and let $\tilde{p}_j=\frac{\mu_j}{\mu_j(\mathbb{X})}$, $j=1,\dots,d$. The vector
\begin{equation}
\label{vettore}
\tilde{p}=(\tilde
p_1,\dots,\tilde p_d)
\end{equation}
is called a \textit{vector of dependent normalized random measures with independent increments} on
$(\mathbb{X},\mathcal{X})$.
\end{dfn}
\section{Compound Random Measures}
In this section, we will define a general class of vectors of NRMI that incorporates
many recently proposed priors built using normalization, see for instance \cite{LL}, \cite{LLS}, \cite{ZL}, \cite{GKS13} and \cite{NLP14a}. Before introducing the formal definition of Compound Random Measures, we want to provide an intuitive illustration of the model. Consider the following dependent random probability measures:
$$\tilde{p}_1=\sum_{i\geq 1} \pi_{1,i}\delta_{X_i}, \,\dots, \,\tilde{p}_d=\sum_{i\geq 1} \pi_{d,i}\delta_{X_i},$$
where
\begin{equation}\label{Corm:reviewer}
\pi_{j,i}=\frac{m_{j,i} J_i}{\sum_l m_{j,l} J_l}.
\end{equation}
The $m_{j,i}$'s are perturbation coefficients that identify specific features of the $j$-th random measure and they are independent and identically distributed across the random measures. The shared jumps $(J_i)_{i\geq 1}$ lead to dependence among the $\tilde{p}_j$. In the next section, we will provide a formal definition of Compound random measures in terms of its multivariate L\'evy intensity.
\subsection{Definition} Let $(\tilde\mu_1,\dots, \tilde\mu_d)$ be a vector of homogeneous CRMs on $\mathbb{X}$, {\it i.e.} the L\'evy intensity $\nu_j$ of the measure $\tilde{\mu}_j$ is
$$ \bar\nu_j(ds,dx)=\nu_j(ds)\,\alpha(dx),\quad j=1,\dots,d.$$
Following the notation in Eq. \eqref{ExpLevy}, we want to define a $\rho_d$ such
\begin{equation}
\label{eq:marginals}
\int_{(0,\infty)^{d-1}}\rho_d(\mathrm{d} s_1,\dots,ds_{j-1},A,ds_{j+1},\dots,ds_d)=
\int_A \nu_j(ds)
\end{equation}
for any $j=1,\dots,d$. In this setting we can define a compound random measure.
\begin{dfn}
A \textit{Compound random measure} (CoRM) is a vector of CRMs defined by a
\emph{score distribution} $h$ and a
\emph{directing L\'evy process} with intensity $\nu^*$ such that
\begin{equation}\label{compound}
\rho_d(ds_1,\dots,ds_d)=\int h(s_1,\dots,s_d\vert z)\,ds_1\cdots ds_d\,\nu^{\star}(dz)
\end{equation}
where $h(\cdot\vert z)$ is the probability mass function or probability density function of the score distribution with parameters $z$ and $\nu^{\star}$ is the L\'evy intensity of the directing L\'evy process which satisfies the condition
$$\int\int\min(1,\parallel\bm{s}\parallel)
h(s_1,\dots,s_d\vert z)\,d\bm{s}
\,\nu^{\star}(dz)
<\infty$$
where $\parallel\bm{s}\parallel$ is the Euclidean norm of the vector $\bm{s}=(s_1,\dots,s_d)$.
\end{dfn}
\medskip
The compound Poisson process with jump density $h$ is a compound random measure with a score density $h$ and whose directing L\'evy process is a Poisson process. Therefore, compound random measures can be seen as a generalisation of compound Poisson processes.
It is straightforward to show that $\tilde{\mu}_1,\dots,\tilde{\mu}_d$ can be expressed as
\begin{equation}
\tilde{\mu}_j=\sum_{i=1}^{\infty} m_{j,i}J_i\delta_{X_i}
\label{jump_mug}
\end{equation}
where $m_{1,i},\dots,m_{d,i}\stackrel{i.i.d.}{\sim} h$ are {\it scores} and
\[
\tilde{\eta}=\sum_{i=1}^{\infty} J_i\delta_{X_i}
\]
is a CRM with L\'evy intensity $\nu^{\star}(ds)\alpha(dx)$. This makes the structure of the prior much more explicit. The random measures share the same jump locations (which have distribution $\alpha/\alpha(\mathbb{X})$)
but the $i$-th jump has a height $m_{j,i}J_i$ in
the
$j$-th measure and so the jump heights are re-scaled by the score (a larger score implies a larger jump height).
Clearly, the shared factor $J_i$ leads to dependence between the jump heights in each measure.
\medskip
The construction can be seen in an alternative way, in terms of (augmented) dependent Poisson random measures. Indeed,
$$\mathbb{E}\left[e^{-\lambda \tilde{\mu}_j(A)}\right]=e^{-\alpha(A)\int_0^{\infty}(1-e^{-\lambda s})\int h_j(s|z)\nu^{*}(dz)ds}.$$
If in the exponent we set $s=mz$, then
$$\mathbb{E}\left[e^{-\lambda \tilde{\mu}_j(A)}\right]=e^{-\alpha(A)\int_0^{\infty}(1-e^{-\lambda mz})\int z h_j(mz|z)\nu^{*}(dz)dm}$$
which entails that $\tilde{\mu}_j(dx)=\int_0^{\infty}\int mz N_j(dm,dz,dx)$ where
$$N_j=\sum_{i\geq 1} \delta_{(m_{j,i},J_i,X_i)}$$
is a Poisson random measure with intensity on $(0,+\infty)^2\times\mathbb{X}$ and with L\'evy intensity given by $\alpha(dx)z h_j(mz|z)\nu^{*}(dz)$. This is identical to the distribution given by \eqref{jump_mug}. The Poisson processes $(N_1,\dots,N_d)$ are dependent because they share $\{(J_i,X_i): i\geq 1\}$. The term ``augmentation'' here refers to the fact that the Poisson random measures that characterize the CRMs are typically on $(0,+\infty)\times \mathbb{X}$. A third dimension is introduced to account the heterogeneity across different measures.
\bigskip
To ensure the existence of the vectors of normalized CoRM, as introduced in Definition \ref{DefVec}, the following condition must be satisfied for each $j=1,\dots,d$:
$$\nu_j((0,+\infty))=\int_0^{+\infty}\int h_j(s|z)\nu^{\star}(dz)ds=+\infty$$
where $h_j(s|z)=\int h(s_1,\dots,s_{j-1},s,ds_{j+1},\dots,s_{d}|z)ds_1\cdots ds_{j-1}ds_{j+1}\cdots ds_{d}$. If this condition does not hold true, then $\tilde{\mu}_j(\mathbb{X})=0$ with positive probability and the normalization does not make sense, see \cite{rlp}.
In this paper, we will concentrate on the sub-class of CoRMs with a continuous score distribution which has independent dimensions and a single scale parameter so that
\[
h(s_1,\dots,s_d\vert z)=z^{-d}\prod_{j=1}^d f(s_j/z)\,
\]
where $f$ is a univariate distribution. This implies that each marginal process has the same
L\'evy intensity of the form
\begin{equation}
\nu_j(ds)=\nu(ds)=\int z^{-1}f(s\vert z)\,ds\,\nu^{\star}(dz).
\label{link_eqn}
\end{equation}
In Section~\ref{sec:comp_method},
algorithms are introduced to sample from the posterior of a hierarchical mixture models whose parameters are driven by a vector of normalized compound random measures. These samplers depend crucially on knowing the form of the Laplace Exponent and its derivatives.
Some general results about the Laplace exponent and the dependence are available if we assume that the density $z^{-1}f(s_i/z)$ admits a moment generating function.
\begin{thm}
Let
$$M_z^f(t)=\int e^{ts}z^{-1}f(s/z)ds$$
be the moment generating function of $z^{-1}f(s_j/z)$ and suppose that it exists.
Then
\begin{equation}\label{MGM}
\psi_{\rho,d}(\lambda_1,\dots,\lambda_d)=\int \left(1-\prod_{j=1}^{d} M_z^f(-\lambda_j)\right)\nu^{\star}(z)dz.
\end{equation}
\label{mgf_thm}\end{thm}
The proof of the Theorem stated above is in the appendix as well as a further result about the derivatives of the Laplace exponent.
\section{CoRMs with independent gamma distributed scores}
In this paper, we will focus on exponential or gamma score distributions.
Throughout the paper we will write $\mbox{Ga}(\phi)$ to be a gamma distribution (or density) with shape $\phi$ and mean $\phi$ which has density
\begin{equation}\label{Jumps}
f(x)=\frac{1}{\Gamma(\phi)}x^{\phi-1}\exp\{-x\}.
\end{equation}
This implies that
$z^{-1}f(y/z)$ is the density of a gamma distribution with shape parameter equal to $\phi$ and mean $\phi\, z$.
The L\'evy intensities $\nu$ and $\nu^{\star}$ and the score density $f$ are linked by (\ref{link_eqn}) and a CoRM can be defined by either deriving $\nu^{\star}$ for a fixed choice of $f$ and $\nu$ or by directly specifying $f$ and $\nu^{\star}$. In this latter case, it is interesting to consider the properties of the induced $\nu$.
Standard inversion methods can be used to derive the form of $\nu^*$. Equation \eqref{link_eqn} implies that
$$\nu(s)=\int z^{-1}\frac{1}{\Gamma(\phi)}\left(\frac{s}{z}\right)^{\phi-1}\exp\left(-\frac{s}{z}\right)\nu^*(z) dz$$
The change of variable $t=z^{-1}$ leads to
$$\nu(s)=\frac{s^{\phi-1}}{\Gamma(\phi)}\int \exp\left(-st\right)t^{\phi-2} \nu^*\left(\frac{1}{t}\right) dt.$$
The above integral can be seen as the classical Laplace transform of the function $f(t)=t^{\phi-2} \nu^*\left(\frac{1}{t}\right)$. If we denote by $\mathcal{L}$ the Laplace transform then
$$\nu(s)=\frac{s^{\phi-1}}{\Gamma(\phi)}\mathcal{L}(f(t))(s)$$
This means that
$$\nu^*\left(\frac{1}{t}\right)=t^{2-\phi}\mathcal{L}^{-1}\left(\frac{\Gamma(\phi)}{s^{\phi-1}}\nu(s)\right)(t)$$
where $\mathcal{L}^{-1}$ is the inverse Laplace transform. This ensures the unicity of $\nu^{*}$.
The forms for some particular choices of marginal process are shown in Table~\ref{t:derived1}.
\begin{table}[h!]
\begin{center}
\begin{tabular}{rll}\hline
$\nu^*(z)$ & Support & Marginal Process\\ \hline
$z^{-1}(1-z)^{\phi-1}$ & $0<z<1$&$\mbox{Gamma}$\\
& \\
$z^{-\sigma-1}\frac{\Gamma(\phi)}{\Gamma(\phi+\sigma)\Gamma(1-\sigma)}$ & $z>0$&$\mbox{$\sigma$-stable}$\\
&\\
$
\frac{\sigma\Gamma(\phi)}{\Gamma(\phi+\sigma)\Gamma(1-\sigma)} z^{-\sigma-1}(1-a\,z)^{\sigma+\phi-1}$ & $0<z<1/a$
&$\mbox{Gen. Gamma}$ \\ \hline
\end{tabular}
\end{center}
\caption{The form of directing L\'evy intensity in a CoRM which leads to particular marginal processes.}\label{t:derived1}
\end{table}
The results are surprising. A gamma marginal process arises when the directing L\'evy process is a Beta process and a $\sigma$-stable marginal process arises when the directing L\'evy process is also a $\sigma$-stable process. Generalized gamma marginal processes lead to a directing L\'evy process which is a generalization of the Beta process (with a power of $z$ which is less than 1) and re-scaled to the interval $(0,1/a)$. In fact, if we
use a gamma score distribution with shape $\phi$ and mean $a\phi$ which has density
\begin{equation}\label{Jumps}
f(x)=\frac{1}{a^{\phi}\Gamma(\phi)}x^{\phi-1}\exp\{-x/a\},
\end{equation}
the directing L\'evy intensity is a stable Beta \cite{tehgor} of the form
\[
\nu^{\star}(z)=\frac{a^{\sigma+1}\sigma}{\phi} \frac{\Gamma(\phi+1)}{\Gamma(\phi+\sigma)\Gamma(1-\sigma)}
z^{-\sigma-1}(1-z)^{\sigma+\phi-1},\qquad 0<z<1.
\]
\begin{rmk}
Several authors have previously considered hierarchical models where $\tilde\mu_1,\dots,\tilde\mu_d$ followed i.i.d. CRM (or NRMI) processes whose centring measure are given a CRM (or NRMI) prior. This construction induces correlation between $\tilde\mu_1,\dots,\tilde\mu_d$ and the hierarchical Dirichlet process is a popular example but we will concentrate on a hierarchical Gamma process \cite[see {\it e.g.}][]{Palla}. In this case, $\tilde\mu_1,\dots,\tilde\mu_d$ follow independent Gamma processes with centring measure $\alpha$ which also follows a Gamma process. This implies that we can write
\[
\alpha = \sum_{i=1}^{\infty} s_i\delta_{\theta_i}
\]
and we can write
\begin{equation}
\tilde\mu_j=\sum_{i=1}^{\infty} J_{j,i}\delta_{\theta_i}
\label{eqn_J}
\end{equation}
where $J_{j,i}\sim\mbox{Ga}(s_i)$. This can be represented as a CoRM process where $\alpha$ is the directing L\'evy process and the score distribution is $\prod_{j=1}^d \mbox{Ga}(s_i)$ where $s_i$ controls the shape of the conditional distribution of $J_{j,i}$. This contrasts with the processes considered in this section with independent gamma scores which multiply the jumps in the directing L\'evy process and lead to a marginal gamma process for $\tilde\mu_j$ (unlike the hierarchical model). These processes can be written in the form of (\ref{eqn_J}) with $J_{j,i}$ having a gamma distribution with shape $\phi$ and mean $\phi s_i$, and $\alpha$ chosen to follow a beta process.
\end{rmk}
\begin{rmk}
This paper is focused on Gamma scores but the class of CoRMs is very wide and other choices can be considered. For instance, if $Beta(\alpha,1)$ scores are selected, i.e.
$$f(x)=\alpha x^{\alpha-1}\qquad \alpha>0,\quad 0<x<1$$
then it is possible to introduce a multivariate version of the Beta process. Let $\nu(s)=\theta s^{-1}(1-s)^{\theta-1}$, $0<s<1$, i.e. the L\'evy intensity of the jumps of a Beta process,
then $\nu^{\star}(z)$ is the solution of the integral equation
\[
\nu(s)=\int_s^1 f(s/z)s^{-1}\nu^{\star}(z)\,dz,\quad 0<s<1
\]
A simple application of the fundamental Theorem of Calculus leads to
\[
\nu^{\star}(z)=
\theta z^{-1}(1-z)^{\theta-1}+\frac{\theta(\theta-1)}{\alpha}(1-z)^{\theta-2}
\]
which is the sum of $\nu(\cdot)$, the L\'evy intensity of the original Beta process, and a compound Poisson process (if $\theta>1$) with intensity $\theta/\alpha$ and jump distribution $Beta(1,\theta-1)$. This is well-defined if $\theta>1$.
\end{rmk}
It is interesting to derive the resulting multivariate L\'evy intensities which can be compared with similar results in \cite{LL}, \cite{LLS} and \cite{ZL}.
\begin{thm}\label{Gamma-GammaLevy}
Consider a CoRM process with independent $\mbox{Ga}(\phi, 1)$ distributed scores. If the CoRM process has gamma process marginals then
\begin{equation}
\rho_d(s_1,\dots,s_d)=\frac{(\prod_{j=1}^ds_j)^{\phi-1}}{[\Gamma(\phi)]^{d-1}}|\bm{s}|^{-\frac{d\phi+1}{2}}e^{-\frac{|\bm{s}|}{2}}W_{\frac{(d-2)\phi+1}{2},-\frac{d\phi}{2}}(|\bm{s}|)
\end{equation}
where $|\bm{s}|=s_1+\dots+s_d$ and $W$ is the Whittaker function. If the CoRM process has $\sigma$-stable process marginals then
\begin{equation}
\rho_d(s_1,\dots,s_d)=\frac{(\prod_{j=1}^ds_j)^{\phi-1}}{[\Gamma(\phi)]^{d-1}}\frac{\Gamma(\sigma+d\phi)}{\Gamma(\sigma)\Gamma(1-\sigma)}|\bm{s}|^{-\sigma-d\phi}.
\end{equation}
\end{thm}
The result is proved in the appendix with the following corollary.
\begin{coro}\label{Gamma-GammaCoro}
Consider a CoRM process with independent exponentially distributed scores. If the CoRM has gamma process marginals we recover the multivariate L\'evy intensity of \cite{LLS},
$$\rho_d(s_1,\dots,s_d)=\sum_{j=0}^{d-1}\frac{(d-1)!}{(d-1-j)!}|\bm{s}|^{-j-1}e^{-|\bm{s}|}.$$
Otherwise, if $\sigma$-stable marginals are considered then we recover the multivariate vector introduced in \cite{LL} and \cite{ZL},
$$\rho_d(s_1,\dots,s_d)=\frac{(\sigma)_d}{\Gamma(1-\sigma)}|\bm{s}|^{-\sigma-d}.$$
\end{coro}
Alternatively, we can specify $\nu^{\star}$ and derive $\nu$. The forms for some particular processes are shown in Table~\ref{t:derived2}
\begin{table}[h!]
\begin{center}
\begin{tabular}{rl}\hline
$\nu(s)$ & Directing L\'evy process\\\hline
$\frac{\Gamma(\theta+1)}{\Gamma(\phi)}s^{-1}
\exp\{-s\}U(\theta-\phi,1-\phi, s)$ & Beta \\
$2\frac{1}{\Gamma(\phi)}\frac{\sigma}{\Gamma(1-\sigma)}s^{(\phi-\sigma)/2-1}
a^{(\sigma+\phi)/2}
K_{\sigma+\phi}\left(2\sqrt{a s}\right)$
& Gen. Gamma \\\hline
\end{tabular}
\end{center}
\caption{The L\'evy intensity of the marginal process in a CoRM with different directing L\'evy processes.}
\label{t:derived2}
\end{table}
where $U$ is the confluent hypergeometric function of the second kind and $K$ is the modified Bessel function of the second kind.
\begin{rmk}
There are several special cases if $\nu^{\star}$ is the L\'evy intensity of a Beta process. Firstly, $U(\theta-\phi,1-\phi,s)=1$ if $\theta=\phi$ and $\nu$ is the L\'evy intensity of a gamma process.
If $\phi = 2\theta-1$, $$U(\theta-\phi, 1-\phi, s)=\pi^{-1/2}\exp\{s/2\}s^{1/2-\theta+\phi}K_{\theta-1/2}(s/2).$$
When $\theta=1$, $U(1-\phi,1-\phi, s) = \exp\{s\}\int_{s}^{\infty} u^{-(1-\phi)}\exp\{-u\}\,du$. The limits as $s\rightarrow 0$ are
\[
U(\theta-\phi,1-\phi,s)\rightarrow\left\{
\begin{array}{lc}
\Gamma(\phi)/\Gamma(\theta)+O(|s|^{\phi}) & 0<\phi<1\\
1/\Gamma(1+\theta-\phi)+O(|s\log s|) & \phi=1\\
\Gamma(\phi)/\Gamma(\theta)+O(|s|) & \phi>1
\end{array}
\right.
\]
Therefore, these processes have a L\'evy intensity similar to the L\'evy intensity of the gamma process
close to zero for any choice of $\phi$ and $\theta$. The tails of the L\'evy intensity are exponential. Therefore, the process has similar
properties to the gamma process.
\end{rmk}
\begin{rmk}
The generalized gamma process contains some special cases and the L\'evy intensity of the marginal process for these process are shown in Table~\ref{t:derived3}.
\begin{table}[h!]
\begin{center}
\begin{tabular}{rl}\hline
$\nu(s)$ & Directing L\'evy process\\\hline
$2\frac{1}{\Gamma(\phi)}s^{\phi/2-1}
K_{\phi}\left(2\sqrt{s}\right)$ & Gamma Process\\
$\frac{\Gamma(\phi+\sigma)}{\Gamma(\phi)}\frac{\sigma}{\Gamma(1-\sigma)}s^{-1-\sigma}$ & $\sigma$-stable Process.\\\hline
\end{tabular}
\end{center}
\caption{The L\'evy intensity of the marginal process in a CoRM with different directing L\'evy processes.}
\label{t:derived3}
\end{table}
With a generalized gamma directing L\'evy process,
It is straightforward to show that
\[
\nu(s)\approx
\sigma\frac{\Gamma(\sigma+\phi)}{\Gamma(\phi)\Gamma(1-\sigma)}
s^{-\sigma-1}
\]
for small $s$.
Therefore, the L\'evy intensity close to zero is similar to the L\'evy intensity of $\sigma$-stable process with parameter $\sigma$. For large $s$, we have
\[
\nu(s)s\propto
\sqrt{\pi}\frac{1}{\Gamma(\phi)}\frac{\sigma}{\Gamma(1-\sigma)}(a s)^{(\phi+\sigma)/2-1/4}
s^{-1-\sigma}
\exp\{-2\sqrt{a}s^{1/2}\}.
\]
Therefore, the tails will decays like $\exp\{-s^{1/2}\}$.
\end{rmk}
The next Theorems will provide an expression of the Laplace exponent when the scores are gamma distributed with $\phi\geq 1$ such that $\phi\in\mathbb{N}$. We want to stress the importance of the the Laplace transform in the Bayesian nonparametric setting. Indeed, it is the basis to prove theoretical results of the prior of interest. For instance, \cite{LL}, \cite{LLS} and \cite{ZL} used the Laplace Transform to derive some distributional properties such as correlation, partition structure and mixed moments. Additionally, we will see that the Laplace transform plays a role in the novel sampler proposed in this paper.
\begin{thm}\label{LapFinal}
Consider a CoRM process with independent $\mbox{Ga}(\phi, 1)$ distributed scores.
Suppose $\phi\geq 1$ such that $\phi\in\mathbb{N}$. Let $\bm{\lambda}\in(\mathbb{R}^+)^d$ be a vector such that it consists of $l\le d$ distinct values denoted as $\bm{\tilde{\lambda}}=(\tilde{\lambda}_1,\ldots,\tilde{\lambda}_l)$ with respective multiplicities $\bm{n}=(n_1,\dots,n_l)$.
Then
\begin{equation*}
\label{eq:casogenerale}
\psi_{\rho,d}(\bm{\lambda})
=\psi_{\rho,d}(\bm{\tilde{\lambda}},\bm{n})=\frac{[\Gamma(\phi)]^l}{\prod_{i=1}^l [\tilde{\lambda}_i^{\phi-1}\Gamma(n_i\phi)]}\left(\prod_{i=1}^l\frac{\partial^{(n_i-1)\phi}}{\partial^{(n_i-1)\phi}\tilde{\lambda}_i}\right)\left(\Upsilon_l^{\phi}(\bm{\tilde{\lambda}})\prod_{i=1}^l \tilde{\lambda}_i^{n_i\phi-1}\right),
\end{equation*}
where
$$\Upsilon^{\phi}_l(\bm{\tilde{\lambda}})=\int \left(1-\prod_{i=1}^{l} \frac{1}{(1+z\tilde{\lambda}_i)^{\phi}}\right)\nu^{\star}(z)dz.$$
\end{thm}
The proof of the previous Theorem is based on the result provided in Theorem \ref{mgf_thm} since the moment generating of a Gamma distribution exists and it is explicit.
To compute the expression of $\Upsilon_l^{\phi}(\bm{\tilde{\lambda}})$ we need to define the following set
$$A_{\phi,j}=\{\bm{k}\in\{1,\ldots,\phi\}^j:\: |\bm{k}|=\phi\}\qquad \phi\geq j.$$
\begin{thm}
Consider a CoRM process with independent $\mbox{Ga}(\phi, 1)$ distributed scores.
Suppose $\phi\geq 1$ such that $\phi\in\mathbb{N}$.
Let $\Lambda(\bm{\tilde{\lambda}},\bm{z})=(1-\sum_{h=1}^{j-1}z_h)\tilde{\lambda}_{i_j}+\sum_{h=1}^{j-1} \tilde{\lambda}_{i_h}z_h$ be a function defined on the (j-1)-dimensional simplex $$\Delta_{j-1}=\{\mathbf{z}\in (0,1)^{j-1}:z_1+\cdots+z_{j-1}<1\}$$ with the convention that $\Delta_0=[0,1]$ . Let
$$a_i(\bm{\tilde{\lambda}})=\frac{\tilde{\lambda}_i^{l-1}}{\prod_{\substack{j=1\\j\neq i}}^l (\tilde{\lambda}_i-\tilde{\lambda}_j)}\qquad i=1,\dots,l.$$
then
\begin{equation*}
\Upsilon_l^{\phi}(\bm{\tilde{\lambda}})=\left\lbrace\begin{array}{ll}
\phi!\displaystyle\sum_{j=1}^{\phi}\sum_{\bm{k}\in A_{\phi,j}}\sum_{0<i_1<i_2<\cdots<i_j\leq l} \frac{a_{i_1}^{k_1}(\bm{\tilde{\lambda}})\cdots a_{i_j}^{k_j}(\bm{\tilde{\lambda}})}{k_1!\dots k_j!}C(i_1,\dots,i_j;\bm{k};\bm{\tilde{\lambda}})&\mbox{ if } l>1\\
\psi(\lambda_1)&\mbox{ if } l=1\end{array}\right.
\label{eq:phi}
\end{equation*}
where
$$
C(i_1,\dots,i_j;\bm{k};\bm{\tilde{\lambda}})=
\Gamma(\phi)\int_{\Delta_{j-1}} \left((1-\sum_{h=1}^{j-1}z_h)^{k_j}\prod_{h=1}^{j-1} \frac{z_h^{k_h-1}}{\Gamma(k_h)}\right)\psi\left(\Lambda(\bm{\tilde{\lambda}},\bm{z})\right)d\bm{z}
$$
For the above integral we assume the usual convention that $\sum_{i}^j=0$ and $\prod_{i}^j=1$ whenever $i>j$.
\end{thm}
In the following Corollary, the expression of the Laplace exponent is recovered for the special case of a CoRM with independent exponentially distributed scores.
\begin{coro}
Consider a CoRM process with independent exponentially distributed scores. It follows that
\begin{equation*}
\psi_{\rho,d}(\bm{\lambda})
=\psi_{\rho,d}(\bm{\tilde{\lambda}},\bm{n})=\left(\prod_{i=1}^l\frac{1}{\Gamma(n_i)}\frac{\partial^{(n_i-1)}}{\partial^{(n_i-1)}\tilde{\lambda}_i}\right)\left(\Upsilon_I(\bm{\tilde{\lambda}})\prod_{i=1}^l \tilde{\lambda}_i^{(n_i-1)}\right),
\end{equation*}
\begin{equation*}
\Upsilon_l(\bm{\tilde{\lambda}})=\left\lbrace\begin{array}{ll}
\sum_{i=1}^l a_{i}(\bm{\tilde{\lambda}})\psi(\lambda_i)&\mbox{ if } l>1\\
\psi(\lambda_1)&\mbox{ if } l=1\end{array}\right..
\label{eq:Expocase}
\end{equation*}
\end{coro}
The proof of the corollary is omitted since it is a direct application of the results of the previous Theorems. Note that, if the vector has Gamma process marginals, {\it i.e.} $\psi(\lambda_i)=\log(1+\lambda_i)$, then we recover the results in \cite{LLS}. If the vector has $\sigma$-stable process marginals, {\it i.e.} $\psi(\lambda_i)=\lambda_i^{\sigma}$, then we recover the result in \cite{LL} and \cite{ZL}. \\
Finally, we close the section with some results about the dependence structure of CoRM processes. A useful description of the dependence of a vector of CRMs is given by the L\'evy copula. A L\'evy Copula is a mathematical tool that allows the construction of multivariate L\'evy intensities with fixed marginals, see appendix.
The following Theorem displays the underlying L\'evy Copula of a compound random measure.
\begin{thm}
Let $\rho_d$ be the compound random measure defined in \eqref{compound} and let $F$ be the the distribution function of $f$. The underlying L\'evy Copula of the compound random measure is
$$C(s_1,\dots,s_d)=\int \nu^{\star}(z)\prod_{j=1}^d (1-F(z^{-1}U^{-1}(s_j)))dz$$
where $U^{-1}$ is the inverse of the tail integral function $U(x):=\int_x^\infty \nu(s)\,\mathrm{d} s$.
\end{thm}
Furthermore, it is possible to prove a result similar to Proposition 5 in \cite{LLS}. This result gives a close formula for the mixed moments of two dimensions of a CoRM process.
The result is expressed in terms of an ordering on sets $\bm{0} \prec\bm{s}_1\prec\,\cdots\,\prec\bm{s}_j$ which is defined in \cite{constsav}.
\begin{thm}
Consider a CoRM process with an independent $\mbox{Ga}(\phi, 1)$ distributed scores. Let $\bm{q}=(q_1,\dots,q_d)$ and let $p_j(\bm{q},k)$ be the set of vectors
$(\bm{\eta},\bm{s}_{1},\ldots,\bm{s}_j)$ such that the coordinates
of $\bm{\eta}=(\eta_1,\ldots,\eta_j)$ are positive and such
that $\sum_{i=1}^j \eta_i=k$. Moreover,
$\bm{s}_i=(s_{1,i},\dots, s_{d,i})$ are vectors such that
$\bm{0} \prec\bm{s}_1\prec\,\cdots\,\prec\bm{s}_j$ and $\sum_{i=1}^j \eta_i(s_{1,i}+\cdots+s_{d,i})=k=q_1+\dots+q_d$.
Then,
\begin{equation*}
\begin{split}
\mathbb{E}\left[\prod_{i=1}^d\{\tilde\mu_i(A)\}^{q_i}\right]&=q_1!\cdots q_d!\: \sum_{k=1}^{\mid \bm{q}\mid}[\alpha(A)]^k\:\times\:\\
&\times \sum_{j=1}^{\mid \bm{q}\mid} \:\sum_{p_j(\bm{q},k)}\:\prod_{i=1}^j \frac{1}{\eta_i!} \left[\left(\prod_{l=1}^d \frac{(\phi)_{s_{l,i}}}{s_{l,i}!}\right)\int z^{{s_{1,i}+\cdots + s_{d,i}}}\nu^{\star}(z)dz\right]^{\eta_i}
\end{split}
\end{equation*}
where $\mid \bm{q}\mid=q_1+\cdots+q_d$.
\end{thm}
\begin{rmk} For instance, suppose that the CoRM process has generalized gamma process marginals. Then,
\begin{equation*}
\begin{split}
\int z^{s_{1,i}+\cdots + s_{d,i}}\nu^{\star}(z)dz=\frac{\sigma a^{\sigma-(s_{1,i}+\cdots + s_{d,i})}}{\Gamma(1-\sigma)}B(k-\sigma-1,\sigma+\phi).
\end{split}
\end{equation*}
\end{rmk}
\section{Normalized Compound Random Measures}
Vectors of correlated random probability measures can be defined by normalizing each dimension of a CoRM process. This will be called a Normalized Compound Random Measure (NCoRM) and is defined by a score distribution, a directing L\'evy process and a centring measure of the CoRM. The results derived in
Table \ref{t:derived1}
can be used to define a NCoRM with a particular marginal process. For example, an NCoRM with Dirichlet process marginals arises by normalizing each dimension of a CoRM with gamma process marginals.
In specifying an NCoRM prior, it is useful to have a method of choosing the parameters of the score distribution to give a particular level of dependence. We describe two possible methods.
It is possible to compute the covariance of a two dimensions of an NCoRM process. Indeed, following \cite{LLS},
\begin{equation}
\begin{split}
\label{eq:cross_moment_simple}
\mbox{Cov}\left[\tilde p_1(A),\,\tilde p_2(B)\right]
&= \left\{\alpha(A\cap B)-
\frac{\alpha(A)\alpha(B)}{\alpha(\mathbb{X})}\right\} \\[7pt]
&\qquad \times\:\int_{(\mathbb{R}^+)^2}g_\rho(1,1;\lambda_1,\lambda_2)\,
\mathrm{e}^{-\alpha(\mathbb{X})\psi_\rho(\lambda_1,\lambda_2)} \:\mathrm{d} \lambda_1\,\mathrm{d} \lambda_2
\end{split}
\end{equation}
where $g_{\rho}$ is the function introduced in Equation \eqref{grho}.
This result can be used to specify any parameters of the score distribution (or a prior for those parameters). Alternatively, if the scores are independent, the ratio of the same jump heights in the $i$-th and $j$-th dimension has the same distribution as the ratio of two independent random variables following the score distribution. For example, if the scores are independent and follow a gamma distribution with shape $\phi$ is chosen, this ratio follows an $F$-distribution with $\phi$ and $\phi$ degrees of freedom.
\subsection{Links to other processes}
Corollary~\ref{Gamma-GammaCoro} shows how the priors described in
\cite{LL}, \cite{LLS} and \cite{ZL} can be expressed in the CoRM framework. The CNMRI process \citep{GKS13}\citep[see also][]{NLP14a, Chen13} can also be expressed in the NCoRM framework. The CNMRI prior express the random measure $\tilde\mu_g$ as
\[
\tilde\mu_j=\sum_{k=1}^q D_{jk} \tilde\mu^{\star}_k
\]
where $D$ is a $(d\times q)$-dimensional selection matrix (with elements either equal to 0 or 1) and $\tilde\mu^{\star}_1,\dots,\tilde\mu^{\star}_q$ are independent CRMs where $\tilde\mu^{\star}_k$ has L\'evy intensity $M_k\nu^{\star}(ds)\bar\alpha(dx)$ for a probability measure $\bar\alpha$.
A CNRMI process can be represented by a vector of CoRMs with score probability mass function
\[
g(s_1=D_{1i}z,\dots,s_d=D_{di}z\vert z)=\frac{M_i}{\sum_{k=1}^q M_k},
\]
directing L\'evy intensity $\nu^{\star}$ and centring measure $\bar\alpha\sum_{k=1}^q M_k$. A CoRM process with independent scores can be used to construct a sub-class of CNRMI processes. A CoRM has a score distribution of the form $f(s)=\pi\delta_{s=1} + (1-\pi)\delta_{s=0}$, directing L\'evy intensity $\nu^{\star}(ds)$ and centring measure $M\bar\alpha$
is identical to an unnormalized CNRMI process with $q=2^d$, a $D$ whose rows are the binary expansion of $\{0,1,\dots,2^d-1\}$ and $M_k=M\prod_{l=1}^d \pi^{D_{kl}}(1-\pi)^{1-D_{kl}}$.
A more general class of unnormalized CNRMI processes with
$M_k=M\prod_{l=1}^d \pi_l^{D_{kl}}(1-\pi_l)^{1-D_{kl}}$ which corresponds to a vector of CRMs such that
\begin{equation}\label{compound2}
\rho_d(ds_1,\dots,ds_d)=\int z^{-d}\prod_{j=1}^d f_j(s_j/z)\,ds_1\cdots ds_d
\,\nu^{\star}(dz)
\end{equation}
where $f_j(m) = \pi_j\delta_{m=1}+(1-\pi_j)\delta_{m=0}$.
\subsection{Computational Methods}\label{sec:comp_method}
We describe methods for fitting a nonparametric mixture model where the mixing measure is given a NCoRM prior. We assume that the data can be divided into $d$ groups and $y_{j,1},\dots,y_{j,n_j}$ are the observations in the $j$-th group. The data are modelled as
\[
y_{j,i}\stackrel{ind.}{\sim} k(y_{j,i}\vert \zeta_{j,i}),
\quad
\zeta_{j,i}\sim\tilde{p}_j,\quad i=1,2,\dots,n_j,\quad j=1,\dots,d
\]
where
$k(y\vert \theta)$ is a probability density function for $y$ with parameter $\theta$
and
$\tilde{p}_1,\dots,\tilde{p}_d$ are given an NCoRM prior.
Using the notation of (\ref{jump_mug}), we write
\[
\tilde{p}_j = \frac{\tilde\mu_j}{\tilde\mu_j(\mathbb{X})}
=\frac{\sum_{k=1}^{\infty} m_{j,k} \,J_k\, \delta_{\theta_k}}{\sum_{k=1}^{\infty} m_{j,k}\,J_k}.
\]
Direct simulation from the posterior distribution is impossible since there are an infinite number of parameters. Several MCMC methods have been introduced which circumvent this problem in the class of normalized random measure mixtures.
\cite{Favaro2013} describe an auxiliary variable method which involves integrating out the unnormalized random measure whereas
\cite{GW2011} introduce a slice sampling method. We consider extending both methods to NCoRM mixtures.
We use the notation $m=(m_{j,k})$, $J=(J_1,J_2,\dots)$ and $\theta=(\theta_1,\theta_2,\dots)$. The
posterior distribution can be expressed in a suitable form for MCMC by introducing latent variables.
Firstly, latent allocation variables $c=(c_{j,i})$ (for which $\zeta_{j,i}=\theta_{c_{j,i}}$) are introduced to give
\begin{align}
p(y,c\vert m, J, \theta)&=
\prod_{j=1}^d \prod_{i=1}^{n_j}
\left[ k\left(y_{j,i}\vert \theta_{c_{j,i}} \right)
\frac{m_{j,c_{j,i}} \,J_{c_{j,i}}}{\sum_{k=1}^{\infty} m_{j,k}\,J_k}\right]\nonumber\\
&= \prod_{j=1}^d \frac{\prod_{i=1}^{n_j}
k\left(y_{j,i}\vert \theta_{c_{j,i}} \right)
m_{j,c_{j,i}} \,J_{c_{j,i}}}{\left(\sum_{k=1}^{\infty} m_{j,k}\,J_k\right)^{n_j}}.
\label{post1}
\end{align}
Secondly, latent variables $v=(v_1,\dots,v_d)$ are introduced to define
\begin{align*}
p(y,c,v\vert m, J, \theta)
=& \prod_{j=1}^d \left[\prod_{i=1}^{n_j}
k\left(y_{j,i}\vert \theta_{c_{j,i}} \right)
m_{j,c_{j,i}} \,J_{c_{j,i}}\right]
\prod_{j=1}^d\left[\frac{1}{\Gamma(n_j)} v_j^{n_j-1}\right]\\
&\times
\exp\left\{-\sum_{j=1}^d v_j \sum_{k=1}^{\infty} m_{j,k}\,J_k\right\}.
\end{align*}
Integrating over $v$ (using the identity $\frac{1}{\Gamma(n)} v^{n-1}\exp\{-vx\}=x^{-n}$) gives the expression in (\ref{post1}).
\subsubsection{Marginal method}
The \cite{Favaro2013} approach relies on an analytical form for $p(y,v, c)$ which is available for the NRMI mixtures using results of \cite{JLP09}. Suppose $\{c_{j,i}\}$ takes $K$ distinct values, that $a_{j,k}$ is the number of observations in the $j$-th group allocated to the $k$-th distinct value and define $a_k=(a_{k,1},\dots,a_{k,d})$. Extending the results of
\cite{JLP09} and \cite{Favaro2013} to vectors of normalized random measures (as in Section 2.1) leads to
\[
p(y,v, c)=
\prod_{i=1}^d\frac{1}{\Gamma(n_i)}v_i^{n_i-1}
\exp\{-\psi_{\rho,d}(v)\}\prod_{k=1}^K \kappa_{a_k}(v) \prod_{k=1}^K g(
\{y_{j,i}\vert c_{j,i}=k\})
\]
where
\[
\psi_{\rho,d}(v) = \int \left(1 - \exp\left\{-\sum_{i=1}^d v_i s_i\right\}\right) \rho_d(ds_1,\dots,ds_d),
\]
\[
\kappa_a(v)=\int \prod_{j=1}^d s_j^{a_j}
\exp\left\{-\sum_{i=1}^d v_i s_i\right\}\rho_d(ds_1,\dots,ds_d)
\]
and
\[
g(y)=\int \prod k(y_{j,i}\vert \theta) \alpha(d\theta).
\]
If the vector of the normalized random measures is chosen to be an NCoRM with independent gamma scores then
\begin{align*}
\kappa_a(v)&=\int \prod_{j=1}^d s_j^{a_j}
\exp\left\{-\sum_{i=1}^d v_i s_i\right\}
z^{-d}\prod_{j=1}^d f(s_j/z)\,ds_1\dots ds_d\,
\nu^{\star}(dz)\\
&=
\int z^{\sum_{j=1}^d a_j}
\prod_{j=1}^d \int \left[s_j^{a_{j}}
\exp\left\{- v_j z s_j \right\}
f(s_j)\, ds_j\right]\nu^{\star}(dz)
\\
&=
\int z^{\sum_{j=1}^d a_j}
\prod_{j=1}^d \tau_{a_j}(z,v_j)\nu^{\star}(dz)
\end{align*}
where
\[
\tau_a(z, v)=\int s^{a}\exp\left\{- v z s\right\}
f(s)\, ds.
\]
and Theorem 3.1 provides the expression
\begin{align*}
\psi_{\rho,d}(v) &
=\int \left(1-\prod_{j=1}^{d} M_z^f(-s_j)\right)\nu^{\star}(z)dz.
\end{align*}
If $f$ is chosen to be a gamma distribution with shape parameter $\phi$,
\begin{align*}
\tau_a(z,v)
&=\int s^a\exp\left\{- v z s\right\} f(s)\, ds
=\frac{\Gamma(a+\phi)}{\Gamma(\phi)}(1+vz)^{-a-\phi}.
\end{align*}
Two algorithms can be defined. One is suitable for conjugate mixtures where $g(y)$ can be calculated analytically and a second algorithm is suitable for non-conjugate mixtures where $g(y)$ cannot be calculated analytically.
In the case of a conjugate mixture model, the steps of the algorithm are
\subsubsection*{Updating $c_{j,i}$}
Let $C^{-(j,i)}_k=\{y_{l,m}\vert c_{l,m}=k, (l,m)\neq (j,i)\}$ and $K^{-(j,i)}$ be the number of distinct values of $\{c_{l,m}\vert (l,m)\neq (j,i)\}$. The parameter $c_{j,i}$ is updated from the discrete distribution
\[
p(c_{j,i}=k)\propto
\left\{
\begin{array}{ll}
\frac{\kappa_{a_k+r}(v)g\left(C_k^{-(j,i)}\cup \{y_{j,i}\}\right)}{\kappa_{a_k}(v)g\left(C_k^{-(j,i)}\right)} & 1\leq k\leq K^{-(j,i)}\\
\kappa_r(v)g(y_{j,i}) & k = K^{-(j,i)} + 1
\end{array}
\right.
\]
where $r$ is a $d$-dimensional vector with $r_m=1$ if $m=j$ and $r_m=0$ otherwise.
For independent $\mbox{Ga}(\phi, 1)$ scores,
\[
\frac{\kappa_{a_k+r}(v)}{\kappa_{a_k}(v)}
= (a_{j,k}+\phi)
\frac{\int z^{\sum_{m=1}^d a_{m,k}+1}
(1+v_j z)^{-a_{j,k}-1-\phi} \prod_{m=1;m\neq j}^d (1+v_m z)^{-a_{m,k}-\phi}\nu^{\star}(z)\,dz}
{\int z^{\sum_{m=1}^d a_{m,k}}
\prod_{m=1}^d (1+v_m z)^{-a_{m,k}-\phi}\nu^{\star}(z)\,dz}
\]
and
\[
\kappa_r(v)=
\phi
\int z
(1+v_j z)^{-1-\phi} \prod_{m=1; m\neq j}^d (1+v_m z)^{-\phi}\nu^{\star}(z)\,dz.
\]
\subsubsection*{Updating $v_j$}
The full conditional distribution of $v_j$ is proportional to
\[
v_j^{n_j-1}
\exp\{-\psi_{\rho,d}(v)\}\prod_{k=1}^K \kappa_{a_k}(v).
\]
This parameter can be updated using an adaptive Metropolis-Hastings random walk \citep{atros}.
\subsubsection*{Updating parameters of $f$}
The full conditional distribution of the parameters of $f$ is proportional to
\[
\exp\{-\psi_{\rho,d}(v)\}\prod_{k=1}^K \kappa_{a_k}(v).
\]
This parameter can be updated using an adaptive Metropolis-Hastings random walk \citep{atros}.
In the case of non-conjugate mixtures, \cite{Favaro2013} define an auxiliary variable method which introduces the distinct values $\theta_1,\dots,\theta_K$ into the sampler and $M$ potential distinct values for empty clusters $\theta'_1,\dots,\theta'_M$.
\subsubsection*{Updating $c_{j,i}$}
A set of values $\theta_1,\dots,\theta_M$ is formed. If $c_{j,i}$ is a singleton ({\it i.e.} $c_{j,i}\neq c_{k,m}$ for $(j,i)\neq (k,m)$), set $\theta'_1=\theta_{c_{j,i}}$ and sample $\theta'_j\sim \alpha/\alpha(\mathbb{X})$ for $j=2,\dots,M$. Otherwise, sample
$\theta'_j\sim \alpha/\alpha(\mathbb{X})$ for $j=1,\dots,M$. The full conditional distribution of $c_{j,i}$ is
\[
p(c_{j,i}=k)\propto
\left\{
\begin{array}{ll}
\frac{\kappa_{a_k+r}(v)}{\kappa_{a_k}(v))}k(y_{j,i}\vert \theta_k) & 1\leq k\leq K\\
\frac{\alpha(\mathbb{X})}{M}\kappa_r(v) k\left(y_{j,i}\vert \theta'_{k-K^{-(j,i)}}\right) & k = K^{-(j,i)} + 1,\dots,K^{-(j,i)}+M.
\end{array}
\right.
\]
\subsubsection*{Updating $\theta_k$}
The full conditional density of $\theta_k$ is proportional to
\[
\alpha(\theta_k)\prod_{\{(j,i)\vert c_{j,i}=k\}}k(y_{j,i}\vert \theta_k).
\]
The full conditional distributions of $v_j$ and any parameters of $f$ are unchanged from algorithm for conjugate mixture models.
\subsubsection{Slice sampling method}
We introduce $u=(u_{j,i})$ and define
\begin{align*}
p(y,c,v,u\vert m, J, \theta)
=& \prod_{j=1}^d \left[\prod_{i=1}^{n_j}
k\left(y_{j,i}\vert \theta_{c_{j,i}} \right)
m_{j,c_{j,i}} \,\mbox{I}(u_{j,i}<J_{c_{j,i}})\right]
\prod_{j=1}^d\left[\frac{1}{\Gamma(n_j)} v_j^{n_j-1}\right]\\
&\times \exp\left\{-\sum_{j=1}^d v_j \sum_{k=1}^{\infty} m_{j,k}\,J_k\right\}.
\end{align*}
Integrating over $u$ and $v$ gives the expression in (\ref{post1}). A similar form is derived in \cite{GW2011}.
This form of the likelihood is still not suitable for MCMC since it involves all jumps. To avoid this,
we define $L=\min_{i=1,\dots,n_j; j=1,\dots,d}\left\{u_{j,i}\right\}$ and divide the jumps into two disjoints sets: $A^{\dagger}=\{(J^{\dagger}_k, m^{\dagger}_{1,k},\dots,m^{\dagger}_{d,k})\vert J^{\dagger}_k>L\}$
and $A^{\star}=\{(J^{\star}_k, m^{\star}_{1,k},\dots,m^{\star}_{d,k})\vert J^{\star}_k\leq L\}$. The set $A^{\dagger}$ has a finite number of elements which is denoted $K$ and $A^{\star}$ has an infinite number of elements.
Integrating over $A^{\star}$ leads to posterior which is suitable for MCMC and has the form
\begin{align}
&\prod_{j=1}^d
\left[\prod_{i=1}^{n_j}
k\left(y_{j,i}\vert \theta_{c_{j,i}}\right) m^{\dagger}_{j,c_{j,i}}\,\mbox{I}\left(u_{j,i}< J_{c_{j,i}}\right)
\right]
\prod_{j=1}^d
\left[\frac{1}{\Gamma(n_j)}v_j^{n_j-1}\right]
\nonumber\\
&\times\exp\left\{-\sum_{j=1}^d v_j\sum_{k=1}^K m^{\dagger}_{j,k}J^{\dagger}_k\right\}\mathbb{E}\left[
\exp\left\{-\sum_{j=1}^d v_j\sum_{k=1}^{\infty} m^{\star}_{j,k}J^{\star}_k\right\}
\right].
\label{post_MCMC}
\end{align}
An MCMC scheme using this form of likelihood leads to a random truncation of the NCoRM process at each iteration but does not introduce a truncation error since integrating over the latent variables leads to the correct marginal posterior.
The expectation in (\ref{post_MCMC}) can be expressed in terms of a univariate integral using a variation on
Theorem \ref{mgf_thm} giving
\[
-\log \mathbb{E}\left[
\exp\left\{-\sum_{j=1}^d v_j\sum_{k=1}^{\infty} m^{\star}_{j,k}J^{\star}_k\right\}
\right]=
\int_0^L \left(1-\prod_{j=1}^{d} M_z^f(-v_j)\right)\nu^{\star}(z)\,dz.
\]
The full conditional distributions and a general discussion of methods for updating parameters are given below. Details of the implementation for specific processes are given in the appendix.
\subsubsection*{Updating $v_1,\dots,v_d$}
The updating of $v_1,\dots,v_d$ uses a variation on the interweaving approach of
\cite{yumeng11}, which leads to better mixing than the standard full conditional distribution for $v_j$.
The parameter $v_j$ is updated in the following way. Firstly, we re-parameterize to $\tilde{m}^{\dagger}_{j,k}=v_j m^{\dagger}_{j,k}$ and update $v_j$ from the full conditional density (conditioning on $\tilde{m}^{\dagger}_{j,k}$ rather than $m^{\dagger}_{j,k}$) which is proportional to
\[
v_j^{-(K+1)}f\left(\frac{\tilde{m}^{\dagger}_{j,k}}{v_j}\right)
\mathbb{E}\left[
\exp\left\{-v_j\sum_{k=1}^K m^{\star}_{j,k} J^{\star}_k\right\}
\right].
\]
Secondly, we re-parameterized to $m^{\dagger}_{j,k}=\tilde{m}^{\dagger}_{j,k}/v_j$
and update $v_j$ from the full conditional density proportional to
\[
v_j^{n_j-1}\exp\left\{-v_j \sum_{k=1}^K m^{\dagger}_{j,k} J_k\right\}
\mathbb{E}\left[
\exp\left\{-v_j\sum_{k=1}^K m^{\star}_{j,k} J^{\star}_k\right\}
\right].
\]
Both full conditional densities are sampled using a Metropolis-Hastings algorithm with random walk and an adaptive proposal distribution.
\subsubsection*{Updating $J^{\dagger}$ and $m^{\dagger}$}
The density of the full conditional distribution of $J^{\dagger}_k$ is proportional to
\[
\mbox{I}\left(J^{\dagger}_k> \max\{u_{j,i}\vert c_{j,i}=k\}\right)\nu^{\star}\left(J^{\dagger}_k\right)
\exp\left\{-\sum_{l=1}^d v_l\sum_{r=1}^K m^{\dagger}_{l,r}J^{\dagger}_r\right\}
\]
where $n_{j,k}=\sum_{i=1}^{n_j} \mbox{I}(c_{j,i}=k)$ and the full conditional density of $m^{\dagger}_{j,k}$
is $\mbox{Ga}\left(\phi+n_{j,k}, 1 + v_j J^{\dagger}_k\right)$.
The elements of $A^{\dagger}$ are also updated using a reversible jump Metropolis-Hastings method with a birth and a death move which are proposed with equal probability. The birth move involves proposing a new jump $J^{\dagger}_{K+1}$ from a density proportional to $\nu^{\star}\left(J^{\dagger}_{K+1}\right)$ for $J^{\dagger}_{K+1}>L$
and $m^{\dagger}_{1,K+1},\dots,m^{\dagger}_{d,K+1}\stackrel{i.i.d.}{\sim} f$. The death move proposes to delete an element of the set of jumps to which no observationsa are allocated
$
B=\left\{\left(J^{\dagger}_k,m^{\dagger}_{1,k},\dots,m^{\dagger}_{d,k}\right)\left\vert\sum_{j=1}^d n_{j,k}=0 \right. \right\}
$
uniformly at random. If $b$ is the number of elements in $B$, the acceptance probability for the birth move is
\[
\min\left\{
1,\exp\left\{-\sum_{j=1}^d v_j J^{\dagger}_{K+1}m^{\dagger}_{j,K+1}\right\}
\frac{\int_L^{\infty} \nu^{\star}(z)\,dz}{b+1} \right\}
\]
and the acceptance probability if the $k$-th jump is proposed to be delete is
\[
\min\left\{
1,\exp\left\{\sum_{j=1}^d v_j J^{\dagger}_{k}m^{\dagger}_{j,k}\right\}
\frac{b} {\int_L^{\infty} \nu^{\star}(z)\,dz}\right\}.
\]
\subsubsection*{Updating $u$}
The full conditional distribution of $u_{j,i}$ is a uniform distribution on $\left(0,J^{\dagger}_{c_{j,i}}\right)$ for
$i=1,\dots,n_j$ and
$j=1,\dots,d$. Let $\kappa$ be the $\min\{u_{j,i}\}$ from the previous iteration and $\kappa^{\star}$ be the $\min\{u_{j,i}\}$ from the current iteration. If $\kappa^{\star}>\kappa$ then the jumps for which $J^{\dagger}_j<\kappa^{\star}$ are deleted. Otherwise, if $\kappa^{\star}<\kappa$,
a Poisson distributed number of jumps with mean
\[
\int_{\kappa^{\star}}^{\kappa} \nu^{\star}(z)\prod_{j=1}^d \int \exp\left\{- v_j m_j\right\}\, f(m_j)\,dm_j\,dz
\]
are simulated from the density of $z$ proportional to
\[
\nu^{\star}(z)\prod_{j=1}^d \int \exp\left\{- v_j m_j z\right\}\, f(m_j)\,dm_j
,\qquad \kappa^{\star}<z<\kappa
\]
and $p\left(m^{\dagger}_j\right)\propto \exp\{-v_j z\}f\left(m^{\dagger}_j\right)$
Details on simulation for NCoRMs with Dirichlet process and normalized generalized gamma process marginals are provided in Appendix B.
\subsubsection*{Updating $\theta$}
The full conditional distribution of $\theta_k$ is
\[
\alpha(\theta_k)
\prod_{j=1}^d
\prod_{\{i\vert c_{j,i}=k\}}
k\left(y_{j,i}\vert \theta_k\right),\qquad k=1,\dots,K
\]
\subsubsection*{Updating the parameters of the NCoRM prior}
The full conditional distribution of the parameters of the NCoRM prior are proportional to
\begin{align*}
&\prod_{j=1}^d \prod_{k=1}^K f\left(m_{j,k}^{\dagger}\right)\prod_{k=1}^K \nu^{\star}\left(J^{\dagger}_k\right)
\exp\left\{- \int_L^{\infty} \nu^{\star}(z) dz\right\}\\
&\times\exp\left\{
-\int_0^L
\left(1-\prod_{j=1}^{d} M_z^f(-v_j)\right)\nu^{\star}(z)\,dz
\right\}
\end{align*}
\subsubsection*{Updating $c_{j,i}$}
The full conditional distribution of $c_{j,i}$ is a discrete distribution with a finite number of possible states proportional to
\[
m^{\dagger}_{j, c_{j,i}}\mbox{I}\left(J^{\dagger}_{c_{j,i}}>u_{j,i}\right)
k\left(y_{j,i}\vert \theta_{c_{j,i}}\right),\qquad 1,\dots,n_j,\quad j=1,\dots,d.
\]
\section{Illustrations}
The clinical studies
CALGB 8881
\citep{CALGB8881}
and CALGB 9160 \citep{CALGB9160} looked at the response of patients to different anticancer drug therapies. The response was white blood cell count (WBC) and patients had between four and 25 measurements taken over the course of the trial. The data was previously analysed by \cite{mulros04} who fit a nonlinear random effects model for the patient's response over time. The model assumes that the mean response at time $t$ with parameters $\theta=(z_1,z_2,z_3,\tau_1,\tau_2,\beta_0,\beta_1)$ is given by
\[
f(\theta,t)=\left\{
\begin{array}{ll}
z_1 & t<\tau_1\\
rz_1+ (1 - r) g(\theta,\tau_2) & \tau_1\leq t<\tau_2\\
g(\theta,t) & t\geq \tau_2
\end{array}
\right.
\]
where $r=(\tau_2-t)/(\tau_2-\tau_1)$ and $g(\theta, t)=z_2 + z_3/[1+\exp\{\beta_0 - \beta_1(t-\tau_2)\}]$. There were nine different combinations of the anticancer agent CTX, the drug GM-CSF and amifostine (AMOF) which are summarized in Table~\ref{t:data}.
\begin{table}[h!]
\begin{center}
\begin{tabular}{rrrrrr}\hline
Group & CTX & GM-CSF & AMOF & Study & Number of patients\\\hline
1 & 1.5 & 10.0 & 0 & 1 & 6\\
2 & 3.0 & 5.0 & 0 & 2 & 28\\
3 & 3.0 & 5.0 & 1 & 2 & 18\\
4 & 3.0 & 2.5 & 0 & 1 & 6\\
5 & 3.0 & 5.0 & 0 & 1 & 6\\
6 & 3.0 & 10.0 & 0 & 1 & 6\\
7 & 4.5 & 5.0 & 0 & 1 & 12\\
8 & 4.5 & 10.0 & 0 & 1 & 10\\
9 & 6.0 & 5.0 & 0 & 1 & 6\\\hline
\end{tabular}
\end{center}
\caption{The levels of CTX (g $\mbox{m}^{-2}$), GM-CSF ($\mu$g k$\mbox{g}^{-1}$) and AMOF across the nine groups. CALGB 8881 is indicated as Study 1 and CALGB 9160 as Study 2.}\label{t:data}
\end{table}
Summaries of the data are available as part of the \verb+DPpackage+ in R where a non-linear regression model is fitted with $f(\theta_{j,i},t)$ as the mean for the $i$-th patient in the $j$-th group. We will consider the differences in the distribution of the estimated values $\hat\theta_{j,i}$'s across the nine studies. It is assumed that
\[
\hat\theta_{j,i}\sim\mbox{N}(\mu_{j,i},\Sigma_{j,i}),\qquad (\mu_{j,j},\Sigma_{j,j})\sim \tilde{p}_j
\]
where $\tilde{p}_1,\dots,\tilde{p}_9$ are given a NCoRM process prior with independent
$\Gamma(\phi,1)$-distributed scores and Dirichlet process marginals. The centring measure $\alpha$ is $\mbox{N}(\mu\vert \bar{\hat\theta},100 \Sigma)\mbox{IW}(\Sigma\vert 14,4/9\times\hat\Sigma)$ where $\bar{\hat\theta}$ and $\hat\Sigma$ are the sample mean and the sample covariance matrix of $\hat\theta$. This implies a prior mean of $1/9\times \hat\Sigma$.
The parameter $\phi$ is given an exponential prior with mean 1.
\begin{figure}[h!]
\begin{center}
\includegraphics[trim=20mm 0mm 0mm 210mm, clip]{CALGB_results}
\end{center}
\caption{The posterior mean marginal densities of each parameters in the CALGB example. The lines indicated a group in CALGB 8881 (solid line) and CALGB 9160 (dashed line).}\label{f:results}
\end{figure}
The results of the analysis are illustrated in Figure~\ref{f:results} which shows the posterior mean marginal density of each parameter. The results within each study are very similar with the main difference occurring between the two studies.
All densities are very similar for the parameters $z_1$, $z_2$, $z_3$ and $t_2$. There is a slight difference in the distribution for $t_1$ but much bigger differences for parameters $\beta_0$ and $\beta_1$. The results for CALGB 8881 are unimodal whereas CALGB9160 includes additional modes at 0.5 for $\beta_0$ and $-0.5$ and 2 for $\beta_1$. Figure~\ref{f:results2} shows the posterior mean joint density of $\beta_0$ and $\beta_1$ which shows a bimodal distribution for CALGB9160 with one mode at roughly $(-1.5, 0.5)$ (which is the mode for CALGB8881) and a second mode at roughly $(-0.5, 0)$. This suggests that CALGB9160 may contains two groups who responded differently.
The posterior median of $\phi$ was 1.03 with a 95\% highest posterior density region of $(0.46, 2.36)$.
\begin{figure}[h!]
\begin{center}
\includegraphics[trim=0mm 0mm 120mm 250mm, clip]{CALGB_results_2D}
\end{center}
\caption{The posterior mean joint densities of $\beta_0$ and $\beta_1$ in the CALGB example for the groups in CALGB 8881 and CALGB 9160.}\label{f:results2}
\end{figure}
\section{Discussion}
The modelling of dependent random measures has been an extremely active area of research for the past fifteen years beginning with the seminal work of MacEachern \citep{MacEachern}. Much of the work has concentrated on dependent random probability measures with several general approaches developed in the literature. Using the notation of (\ref{DDP}), initial work considered approaches where $w_i(x)=w_i$ and dependence is modelled through the atom location $\theta_i(x)$. This implies that cluster sizes will be similar for all values of $x$ and so leads to a specific form of dependence. Alternatively,
many authors used $\theta_i(x)=\theta$ for all $x$ with dependence modelled through the weights; often using a
stick-breaking construction where $w_i(x) = V_i(x)\prod_{j<i}(1-V_j(x))$, see {\it e.g.} \citep{dun10} for a review. This usually leads to computationally tractable methods which either extend random truncation methods such as retrospective sampling
\citep{Pap08} or slice sampling \citep{Kalli11}, or develop truncation ideas for Dirichlet process mixtures
\citep{IshJames}. However, stick-breaking approaches have some limitations for modelling. The construction implies a stochastic ordering so that $w_1(x)$ will tend to be the largest weight for all $x$. This can be inappropriate for some regression problems where we would like different component to have large weights for different values of $x$. The correlation is usually built on $V_j(x)$ and so $w_i(x)$ is a non-linear function of many correlated processes. This can lead to a dependence structure on $w_i(x)$ which is hard to interpret. Analytical results such as generalizations of the exchangeable partition probability function are usually impossible to derive for these priors. These methods can often be applied to problems where $\mathcal{X}$ is continuous or discrete. Other priors are restricted to a discrete $\mathcal{X}$. One approach builds a hierarchy of nonparametric processes (see \cite{tehjord} for a review) leading from the seminal work of \cite{tehjor06} on
hierarchical Dirichlet process (HDP). For example, a two level hierarchical model could be constructed by assuming that the distributions for each group are conditionally independent draws from a nonparametric prior which is centred on a process which is itself given a nonparametric prior. This leads to the same correlation {\it a priori} between the distribution for each value in $\mathcal{X}$ (although, more complicated hierarchical structures could be introduced to allow different correlation within subsets of $\mathcal{X}$). Posterior simulation is usually implemented using the Chinese restaurant franchise algorithm.
The CoRM in its most general form is very flexible and allows both hierarchical and regression models.
Normalized compound random measures includes many previously described priors which makes the links between these priors clearer.
This paper has concentrated on priors where the dimensions of the scores are independent.
The tractability of these measures allows their properties to be derived and we concentrate on the class where the dimension of the scores are gamma distributed.
If the moment generating function of the marginal score distributions is available analytically, posterior computation for NCoRM mixture model can be carried out using an augmented P\'olya urn scheme or a slice sampler and several useful analytical expressions can be derived.
This restricts modelling to hierarchical type structures. More general, CoRM-type models where the scores are given by a regression are discussed by \cite{RanBlei15} who use a truncation of the infinite dimensional parameter and variational Bayes to make inference. In future work, we intend to extend both the P\'olya urn scheme and slice sampler to regression models.
The compound random measure is defined using a completely random measure and a finite dimensional score distribution. For a given marginal process, the dependence between the distributions is controlled by the choice of finite dimensional score distribution. In this paper, we have concentrated on the case where the scores are independent and gamma distributed.
This allows the dependence between the measures in different dimensions to be modelled by the shape parameter of the gamma distribution. In this case, we show how compound random measures can be constructed with gamma, $\sigma$-stable and generalized gamma process marginals. Importantly, the modelling of dependence between random measures can be achieved by the modelling of dependence between random variables and so greatly reduces the difficulty of specifying a
prior for a particular problem. Future work will consider studying these classes of compound random measures.
\section*{Acknowledgements}
Fabrizio Leisen was supported by the European Community's Seventh Framework Programme [FP7/2007-2013] under grant agreement no: 630677 and Jim E. Griffin was supported by
EPSRC Novel Technologies for Cross-disciplinary Research grant EP/I036575/I. The authors would like to acknowledge CALGB for the data used in the illustration.
|
\section{Introduction}
According to the celebrated uniqueness theorems (see~\cite{Robinsoon:2004zz,Chrusciel:2012jk} for reviews), the Kerr-Newman (KN) solution~\cite{Newman:1965my} describes the most general stationary, regular (on and outside a horizon) single black hole (BH) configuration of Einstein-Maxwell theory. The solution is a 4-parameter family, described by mass $M$, angular momentum $J$, electric charge $Q$ and magnetic charge $P$. The magnetic charge, however, besides being absent in standard electrodynamics, can be removed by using the electromagnetic duality of the electrovacuum Einstein-Maxwell theory~\cite{Deser:1976iy}. As such, it is often neglected---as will be the case here.
Even though it is unlikely that the KNBH plays a relevant role in astrophysics~\cite{Gibbons:1975kk,Blandford:1977ds,Barausse:2014tra}, this solution has raised considerable interest since its discovery, as an arena for theoretical investigations. In particular, it provides an ideal testing ground for studying the interplay between gravity and electrodynamics at a nonlinear level and the extent to which fundamental properties of the Kerr space-time are modified by the electromagnetic field.
As for similarities, the KN line element is of course remarkably similar to that of the Kerr solution.
In particular, special properties of Kerr also apply to the
more general KN spacetime. For instance, the Liouville integrability of the geodesic equations observed in Kerr, is still present in KN~\cite{Carter:1968rr}. Indeed, KN possesses a hidden constant of motion, which permits the separability of test particle equations. Geometrically, this conserved quantity can be understood from the existence of an irreducible Killing tensor~\cite{Carter:1977pq}. Yet another consequence of this hidden symmetry is that scalar perturbations, obtained by solving the scalar wave equation in the KN background, are separable~\cite{Carter:1968ks}.
A different behaviour, on the other hand, is found when considering electromagnetic and gravitational perturbations. Electromagnetic fluctuations are decoupled from gravitational fluctuations in the Kerr geometry, and both separate in an elegant way when using the Newman Penrose formalism~\cite{Teukolsky:1972my}. These properties allowed for a number of significant results to be achieved for the Kerr geometry, most notably its mode stability~\cite{Whiting:1988vc}. In contrast, electromagnetic and gravitational perturbations do not decouple in the KN background
and need to be studied jointly: a small gravitational fluctuation in such background induces a perturbation in the electromagnetic field which is of the same order of magnitude. The separability of the relevant equations in the KNBH background is a formidable open problem~\cite{MTB}. This difficulty has prevented the analysis of various physical properties of the KNBH, most notably its mode stability and oscillation properties.
Understanding the stability of a solution to Einstein's equations plays a central role in assessing the solution's physical relevance. As such, considerable effort has been devoted towards establishing a proof of the stability of the Kerr solution beyond mode analysis~\cite{Dafermos:2010hd}. Also, mounting (but certainly partial) evidence for stability has been
furnished by a large body of numerical simulations performed over the last decades. These include
binary mergers of BHs and/or neutron stars as well as rotating stars undergoing collapse. These efforts have accumulated considerable support for this solution being stable at the nonlinear level as well, at least within the
time-scales and regimes probed by these simulations (we refer the reader to~\cite{Lehner:2014asa} for a further
discussion of this point and numerous representative references of relevant examples, and to~\cite{Cardoso:2014uka} for similar efforts in a broader context).
Much less is known about the stability of the KN solution. Indeed, due to the difficulties mentioned above, progress has only been made recently in studying electromagnetic and gravitational perturbations in either the slow or extreme rotation limits. For instance, thorough analysis of the behavior of perturbations in the slow rotation regime
indicate that KN is linearly stable for all values of the charge~\cite{Pani:2013ija,Pani:2013wsa}. The methods used for this purpose, however, are not able to probe fast rotating KN solutions and thus require a different strategy. In recent years, interesting perturbative approaches have
been developed that exploit the particular structure arising in the near-extreme limit either directly~\cite{Yang:2013uba,Yang:2014tla}
or through the Kerr/CFT correspondence and the expanded set of isometries arising in such scenarios~\cite{Porfyriadis:2014fja,Hadar:2014dpa}. These works
are providing incipient evidence for linear stability in near-extremal BHs.
In this work we shall explore the nonlinear stability of the KN solution using tools from numerical relativity, which allow us to probe the fast rotating limit (see also~\cite{East:2013mfa} for studies in the non-charged case).
The formalism we employ here has been described in Refs.~\cite{Zilhao:2012gp,Zilhao:2013nda}, which was previously
employed to study collisions of charged BHs with equal and with opposite charges in Einstein-Maxwell theory. With this formalism we are able, after choosing appropriate initial data, to analyse the behavior of perturbed KNBHs.
Anticipating some of the discussions, we will show that our evolutions reveal no evidence for instabilities.
This paper is organized as follows. In Sec.~\ref{sec:evol-eq} we briefly review the formalism used in~\cite{Zilhao:2012gp,Zilhao:2013nda} for evolving the Einstein-Maxwell system. Section~\ref{sec:init-data} addresses the construction of appropriate initial data to describe a (perturbed) KNBH. Section~\ref{sec_diagnosis} describes the diagnostic tools used to monitor the evolution and decide on whether instabilities are present. The numerical results are reported in Sec.~\ref{sec:numerical_results} and our conclusions and final remarks are made in Sec.~\ref{sec:final}.
\section{Formalism}
\label{sec:evol-eq}
Following our previous work on collisions of charged BHs in Refs.~\cite{Zilhao:2012gp,Zilhao:2013nda}, we consider the enlarged electrovacuum Einstein-Maxwell equations
\begin{equation}
\label{eq:EFE}
\begin{aligned}
R_{\mu \nu} - \frac{R}{2} g_{\mu \nu} & = 8\pi T_{\mu \nu} \ ,\\
\nabla_{\mu}\left( F^{\mu \nu} + g^{\mu\nu} \Psi
\right) & = -\kappa n^{\nu} \Psi \ , \\
\nabla_{\mu}\left(
\star \!{}F^{\mu \nu} + g^{\mu\nu} \Phi
\right) & = -\kappa n^{\nu} \Phi \ ,
\end{aligned}
\end{equation}
where $F_{\mu \nu}=\partial_{\mu}A_{\nu} - \partial_{\nu}A_{\mu}$ is the Maxwell tensor and $\star \!{}F^{\mu \nu}$ its Hodge dual, $\kappa$ is a constant and
$n^\mu$ is the 4-velocity of Eulerian observers. We recover the
standard Einstein-Maxwell system when $\Psi = 0 = \Phi$ and merely introduce these fields as a means to damp and control violations of the
magnetic and electric constraints during the numerical evolution~\cite{Komissarov:2007wk,Palenzuela:2008sf}.
The electromagnetic stress-energy tensor takes the usual form
\begin{equation}
\label{eq:Tmunu}
T_{\mu \nu} = \frac{1}{4\pi} \left[ F_{\mu}{}^{\lambda} F_{\nu \lambda}
- \frac{1}{4} g_{\mu \nu} F^{\lambda \sigma} F_{\lambda \sigma}
\right] \ .
\end{equation}
We perform a Cauchy (3+1) decomposition by introducing a 3-metric
$\gamma_{\mu\nu} = g_{\mu \nu} + n_{\mu} n_{\nu}$,
and decompose the Maxwell tensor and its dual into the electric and magnetic 4-vectors as
\begin{equation}
\begin{aligned}
F_{\mu \nu} & = n_{\mu} E_{\nu} - n_{\nu} E_{\mu}
+ \epsilon_{\mu\nu\alpha\beta} B^{\alpha} n^{\beta} \ ,\\
\star \! F_{\mu \nu} & = n_{\mu} B_{\nu} - n_{\nu} B_{\mu}
- \epsilon_{\mu\nu\alpha\beta} E^{\alpha} n^{\beta} \ ,
\end{aligned}
\label{eq:faraday}
\end{equation}
where we use the convention $\epsilon_{1230} = \sqrt{-g}$,
$\epsilon_{\alpha \beta \gamma}
= \epsilon_{\alpha \beta \gamma \delta} n^{\delta}$,
$\epsilon_{123} = \sqrt{\gamma}$.
\section{Initial data}
\label{sec:init-data}
As already mentioned in the Introduction, the KN solution is defined by three (physical) parameters: mass $M$, spin $a M$ and electric charge $Q$.
In Boyer-Lindquist coordinates $(t,r_\mathrm{BL},\theta,\phi)$, the metric and vector potential take the form (see, e.g., Ref.~\cite{wald1984general})
\begin{equation}
\begin{aligned}
ds^2 & = - \left( \frac{\Delta - a^2 \sin^2 \theta}{\rho^2} \right) dt^2
+ \frac{\rho^2}{\Delta} dr_\mathrm{BL}^2
+ \rho^2 d\theta^2 \\
& \quad + \frac{ (r_\mathrm{BL}^2 + a^2)^2
- \Delta a^2 \sin^2 \theta}{\rho^2} \sin^2 \theta d\phi^2 \\
& \quad - \frac{2a \sin^2 \theta(r_\mathrm{BL}^2 + a^2 - \Delta)}{\rho^2} dt d\phi \,, \\
\mathcal{A} & = -\frac{Q r_\mathrm{BL}}{\rho^2}\left(dt-a \sin^2 \theta d\phi \right) \,,
\end{aligned}
\label{eq:KN-BL-ds}
\end{equation}
where
\begin{align*}
\rho^2 & \equiv r_\mathrm{BL}^2 + a^2 \cos^2 \theta \,, \\
\Delta & \equiv r_\mathrm{BL}^2 - 2 M r_\mathrm{BL} + a^2 + Q^2 \,.
\end{align*}
In order to obtain initial data suitable for numerical evolutions using the
``moving punctures'' technique \cite{Campanelli:2005dd,Baker:2005vv},
we express the solution in terms of a quasi-isotropic radial coordinate $R$. Following Refs.~\cite{Brandt:1996si,Krivan:1998td,Cook:2000vr} we perform the coordinate transformation
\begin{equation}
r_\mathrm{BL} = R \left(1+\frac{M + \sqrt{a^2 + Q^2}}{2R}\right)
\left(1 + \frac{M - \sqrt{a^2 + Q^2}}{2R}\right)\,,\nonumber
\end{equation}
and the metric then takes the form [$\bar i, \bar j = (R,\theta,\phi)$ are spatial indices]
\begin{equation}
\label{eq:KN-3+1}
ds^2 = \left(-\alpha^2 + \beta_{\phi}\beta^{\phi}
\right) dt^2 + 2 \beta_\phi d\phi dt+ \gamma_{\bar i \bar j} dx^{\bar i} dx^{\bar j} \,,
\end{equation}
where
\begin{align*}
\gamma_{\bar i \bar j} dx^{\bar i} dx^{\bar j} &=
\psi^4 \Big[dR^2 + R^2 \left( d\theta^2 + \sin^2 \theta d\phi^2 \right) \\
& \qquad + a^2 h R^4 \sin^4\theta d\phi^2\Big] \,, \\
\alpha & = \frac{\left(R+R_H \right) \left(R-R_H \right)}
{R \sqrt{r_\mathrm{BL}^2 + a^2(1+\sigma \sin^2 \theta)} } \,, \\
\beta_{\phi} &= - a \sigma \sin^2 \theta \,, \qquad \beta^{\phi}=\beta_{\phi}/\gamma_{\phi \phi} \,, \\
\psi^{4} & \equiv \rho^2/R^2, \qquad
h \equiv (1+\sigma)/(\rho^2 R^2), \\
\sigma & \equiv (2M r_\mathrm{BL} - Q^2)/\rho^2.
\end{align*}
Here, $R_H \equiv \frac{1}{2} \sqrt{M^2-a^2-Q^2}$ is the location of the event horizon in the quasi-isotropic coordinate $R$.
The nonzero components of the extrinsic curvature $K_{\bar i \bar j}$ take the form
\begin{equation}
K_{R \phi} = \psi^{-2} \frac{H_E \sin^2 \theta}{R^2} \,,\quad K_{\theta \phi} = \psi^{-2} \frac{H_F \sin \theta}{R} \,,
\end{equation}
where
\begin{equation}
\label{eq:aux}
\begin{aligned}
H_E & \equiv \frac{a M G}{\rho^3 \sqrt{r_\mathrm{BL}^2 + a^2 (1 + \sigma \sin^2 \theta)} } \,, \\
H_F & \equiv -\alpha \sigma a^3 \frac{\cos\theta \sin^2 \theta}{\rho} \,, \\
G & \equiv (r_\mathrm{BL}^2 - a^2)\rho^2 + 2r_\mathrm{BL}^2 (r_\mathrm{BL}^2 + a^2) \\
& \quad - \frac{Q^2}{M} r_\mathrm{BL} \left( 2 \rho^2 + a^2 \sin^2 \theta
\right) \, .
\end{aligned}
\end{equation}
The electric and magnetic field can be computed from~(\ref{eq:faraday}). Its nonzero components are
\begin{equation}
\begin{aligned}
E^R & = \frac{Q R (2r_\mathrm{BL}^2 - \rho^2) (r_\mathrm{BL}^2 + a^2)}
{\rho^6 \sqrt{r_\mathrm{BL}^2 + a^2 (1 + \sigma \sin^2 \theta ) } } \,, \\
E^{\theta} & = - \frac{2 a^2 Q \alpha R \cos\theta \sin\theta}{\rho^6} \,, \\
B^R & = \frac{2 a Q R r_\mathrm{BL} (r_\mathrm{BL}^2 + a^2) \cos \theta }
{\rho^6 \sqrt{r_\mathrm{BL}^2 + a^2 (1 + \sigma \sin^2 \theta ) } } \,, \\
B^{\theta} & = \frac{a Q \alpha (2 r_\mathrm{BL}^2 - \rho^2) \sin\theta}{\rho^6} \,.
\end{aligned}\label{eq:EB}
\end{equation}
We finally transform to Cartesian coordinates $ x^{\bar i} = (t,R,\theta,\phi) \to x^{i} = (t,x,y,z)$. Our initial data then reads
\begin{align}
& \gamma_{ij} dx^i dx^j = \psi^4 \big[
dx^2 + dy^2 + dz^2 \notag \\
& \qquad \qquad \quad + a^2 h \left( y^2 dx^2 - 2xy dx dy + x^2 dy^2 \right)
\big] \,,\label{eq:gamma0ij} \\
& K_{ij} = \Lambda^{\bar i}{}_{i} \Lambda^{\bar j}{}_{j} K_{\bar i \bar j} \,, \quad
E^{i} = \Lambda^{i}{}_{\bar i} E^{\bar i} \,, \quad
B^{i} = \Lambda^{i}{}_{\bar i} B^{\bar i} \,,
\end{align}
where $\Lambda^{\bar i}{}_{i} = \frac{\partial x^{\bar i}}{\partial x^{i}} $, in a form analogous to that of Refs.~\cite{Hannam:2006zt,Shibata:2009ad,Shibata:2010wz}.
In order to study the stability of this solution, we follow~\cite{Shibata:2009ad,Shibata:2010wz} and introduce a small bar-mode perturbation to the 3-metric $\gamma_{ij}$ and specify the initial conditions for the 3-metric elements as
\begin{equation}
\hat \gamma_{ij} = \gamma_{ij} \left[
1 + A \frac{x^2 - y^2}{M^2} e^{-\frac{(R-R_0)^2}{2R_H^2}}
\right]^{-1} \,, \label{eq:pert}
\end{equation}
where $A \ll 1$, $\gamma_{ij}$ is the unperturbed solution given by~\eqref{eq:gamma0ij}, and $R_0$ is a tunable parameter that localizes the perturbation.
This perturbation is constraint violating\footnote{Constraint violations are an
inherent consequence of the numerical modelling of spacetimes in general
relativity at the level of the numerical discretization error. Following a
common approach (see e.g.~\cite{Shibata:2009ad,Shibata:2010wz}), we here add a
small perturbation to the initial data in order to trigger an instability more
rapidly (if one exists). This perturbation introduces an additional constraint
violation at a level well below that due to the discretization but we
significantly mitigate this effect by localizing the perturbation well within
the horizon.}; confining the fluctuation within the horizon (by choosing
$R_{0}\simeq 0$) will however produce only a weak gravitational wave
signal. Thus, we here choose to monitor quantities that describe the horizon
deformation, as explained in the next section. We find that our results,
described below, can also be used to understand the gravitational wave signal at
large distances. Note also that for our choices of perturbation amplitudes $A$,
when looking at, for instance, the Hamiltonian constraint violation, we see no
noticeable differences when comparing with non-perturbed cases.
\section{Diagnostics}
\label{sec_diagnosis}
We analyze the result of our numerical investigations using
the following quantities. (i) The (coordinate invariant) horizon ``areal'' radius
\begin{equation}
\label{eq:rar}
r_\mathrm{areal} = \sqrt{r_H^2 + a^2} \,,
\end{equation}
where $r_H \equiv M + \sqrt{M^2 - a^2 - Q^2}$. (ii) The ratio between the polar and equatorial horizon circumferences
\begin{equation}
\label{eq:circ-ratio}
\mathcal{C}_{p} / \mathcal{C}_{e} = \frac{r_H^2}{\pi(r_H^2 + a^2)}
\int_{0}^{\pi} \sqrt{1 + \frac{a^2}{r_H^2} \cos^2 \theta } \ d\theta \,,
\end{equation}
which, for known $M$ and $Q$, allows one to determine $a$.
Finally, we (iii) quantify the ``strength'' of the bar-mode perturbation in terms of the following distortion parameters~\cite{Saijo:2000qt,Franci:2013mma}
\begin{equation}
\label{eq:eta}
\eta_{+} \equiv \frac{I^{xx} - I^{yy}}{I^{xx} + I^{yy}} \,, \qquad \eta_{\times} \equiv \frac{2I^{xy}}{I^{xx} + I^{yy}} \,,
\end{equation}
where
\begin{equation}
\label{eq:Iij}
I^{ij} = \int_{H} d^3x \sqrt{\gamma} x^i x^j
\end{equation}
is the quadrupole moment of the apparent horizon. We also compute the radiation from
the system by computing the Newman-Penrose scalar $\Psi_4$ at distances far from the BH.
We have found, however, that the behavior of $\{\eta_{+},\eta_{\times}\}$ is better suited to
analyze the response of the near BH region to the perturbations---which, as discussed
in the previous section, are initially concentrated in that region. This is a natural observation as the
BH potential barrier essentially traps the induced perturbations in the BH's vicinity.
To monitor the evolution, we compute the relative difference of both the areal radius $r_\mathrm{areal}$ of the apparent horizon and the measured BH spin to the known analytic value
\begin{equation}
\label{eq:rel-err}
\delta[f(t)] \equiv \max_{t>5M} \frac{|f(t) - f^0|}{f^0} \,,
\end{equation}
where $f^0$ is the analytic value.
We choose to evaluate the maximum from $t\simeq 5M$ onward to remove possible large fluctuations due to our
initial perturbation~\eqref{eq:pert}. Finally we monitor convergence of the solution via standard numerical
analysis and the behavior of the constraints to ensure truncation errors remain small throughout the simulation's time span.
\section{Numerical results}
\label{sec:numerical_results}
We numerically integrate the Einstein-Maxwell system using fourth-order spatial
discretization with the \textsc{Lean} code~\cite{Sperhake:2006cy}. This code is based on the
\textsc{Cactus} Computational toolkit~\cite{cactus}, the \textsc{Carpet} mesh
refinement package~\cite{Schnetter:2003rb,carpet} and uses
\textsc{AHFinderDirect} for tracking apparent
horizons~\cite{Thornburg:2003sf,Thornburg:1995cp}. \textsc{Lean} uses the BSSN
formulation of the Einstein equations~\cite{Shibata:1995we,Baumgarte:1998te}
with the moving puncture method~\cite{Campanelli:2005dd,Baker:2005vv}. We refer
the interested reader to Ref.~\cite{Sperhake:2006cy} for further details on the
numerical methods, and to~\cite{Zilhao:2013nda} for the tests performed with the Einstein-Maxwell implementation.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{figs/exclusion_plot}
\caption[]{The simulations performed in this work are displayed as crosses in the parameter space spanned by the rotation parameter $a$ and the charge $Q$.
The dashed blue line shows the extremal limit $a=a_{\rm max}$.
\label{fig:exclusion_plot} }
\end{figure}
We evolve the Einstein-Maxwell system of equations~(\ref{eq:EFE})
for several different charge and spin
values until $t\simeq 120M$ and monitor
both the areal radius $r_\mathrm{areal}$ and the polar to equatorial horizon circumferences ratio.
As a practical measure, we consider a configuration to be \emph{stable} if: (i) during the course of
the numerical evolution, the BH areal radius and spin [the latter inferred
through Eq.~(\ref{eq:circ-ratio})] vary by less than a few
percent---consistent with the perturbation---with respect to the analytic value and (ii),
their time-dependence show an attenuating behavior.
For visual guidance of the parameter space explored in this work, we display in
Fig.~\ref{fig:exclusion_plot} the extremality curve $a=a_{\rm max} \equiv
\sqrt{M^2-Q^2}$ together with the distribution of spin and charge values of the
simulations performed. The simulations performed include several configurations close
to extremality plus additional ones far from this regime for comparison purposes.
\begin{table*}[tbhp]
\centering
\caption{List of simulations performed with parameters used, where $a_{\rm max}\equiv \sqrt{M^2-Q^2}$.
The error reported was measured according to Eq.~(\ref{eq:rel-err}).
For simulations with $a\geq0.99a_{\rm max}$, the numerical grid structure used (in the notation of Sec.~II~E
of~\cite{Sperhake:2006cy}) was the following $\{(256,176,64,32,16,8,4,2,1,0.5,0.125), M/512\}$.
\label{tab:runs}}
\input{runs_table.tex}
\end{table*}
In Table~\ref{tab:runs} we list the simulations performed with the corresponding
physical parameters used. Note that, except for two instances (runs \verb|a0.990_q0.1_A0.0005| and
\verb|a0.975_q0.2_A0.0005|, both of these being cases where $a=0.995a_{\rm
max}$), the relative variation in $r_\mathrm{areal}$ is always smaller than $1\%$ (and for
most cases even smaller than $0.1\%$), which gives us confidence in the accuracy
of our numerical evolution since these are consistent with corresponding results for the
Schwarzschild case ($a=0=Q$).
The larger variation observed in the two mentioned
cases (and, to a lesser degree, also in the \verb|a0.992_q0.0_A0.0005| and
\verb|a0.994_q0.0_A0.0005| runs) is due to a small but steady growth in
$r_\mathrm{areal}$ observed from $t\sim 80M$ onward. We saw similar behaviour in other simulations accompanied by a steady increase of
the Hamiltonian constraint violations with time. In all such circumstances, this
behaviour was successfully cured with an increase in the numerical resolution
used. We believe that this is happening in all aforementioned cases: the growth
in the measured horizon area is merely telling us that more resolution is needed
should we want to accurately evolve such near-extremal configurations ($a
\gtrsim 0.994 a_{\rm max}$) for longer times. The already very high resolution used in
such cases effectively limits our ability to do so, however.
\subsection{Nonlinear stability of Kerr-Newman spacetimes}
Figures~\ref{fig:eta}--\ref{fig:etaplus} summarize our results.
In Fig.~\ref{fig:eta} we plot the time evolution of the deformation parameters~\eqref{eq:eta} for a ``typical'' case corresponding to $(a/M,Q/M)=(0.907,0.4)$.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{figs/eta}
\caption[]{Measured deformation parameters $\eta_{+,\times}$ as given from~(\ref{eq:eta}),
as function of time for a simulation with $a=0.907M$, $Q=0.4M$,
$A=5\times 10^{-4}$. \label{fig:eta} }
\end{figure}
The behavior of $\{\eta_{+},\eta_{\times}\}$ consists of a sum of damped sinusoids and decays away on timescales of order $100M$, consistent with linearized predictions for the ringdown timescale~\cite{Berti:2009kk}. For neutral or static BHs,
the ringing frequency and damping times of the fluctuations match well linearized calculations of quasinormal frequencies~\cite{Berti:2009kk}.
All our simulations display this same behavior: initial fluctuations are damped away. This is one of the main messages of our work:
for the parameters we studied, the KN geometry appears to be
nonlinearly stable against such perturbations on the timescales examined herein---thus indicating any possible instability should have a secular growth associated to it.
In Fig.~\ref{fig:hc_a07_q07}, we show the Hamiltonian constraint violations for a $(a/M,Q/M)=(0.7,0.7)$ example. Note that after a brief transient early on, the constraint
violation is not significantly growing in time, and that it bears the overall pattern observed for typical numerical BH evolutions.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{figs/a07_q07_hc_z}
\caption[]{Snapshots of the Hamiltonian constraint violation along the $z$-axis at taken at three different values of the evolution time for a simulation with $a=0.7M$, $Q=0.7M$, $A=0.005$. The inset shows the same data in a region close to the horizon [$R_H(t=0) \simeq 0.0707M$,
$R_H(t=160M) \simeq 0.16M$]. \label{fig:hc_a07_q07} }
\end{figure}
\subsection{Universality of oscillation modes}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{figs/etaplus_compare3}
\caption[]{Measured deformation parameter $\eta_{+}$ for several different
simulations as function of time. All curves were normalized to their respective maximum
amplitude. \label{fig:etaplus} }
\end{figure}
Our results indicate a surprising universal relation between the oscillation frequency and damping times of the fluctuations,
namely that for large $a/Q$, spacetimes with the same $a/a_{\rm max}$ behave in a similar way.
This is summarized in Fig.~\ref{fig:etaplus} where we show the evolution of $\eta_+$ for three different values of $(a,Q)$
which share the same $a/a_{\rm max}=0.99$, and have $a/Q>1$. The lines corresponding to the different cases overlap almost perfectly.
For comparison, another value of $(a,Q)$ with $a/a_{\rm max}=0.99$, but with $a/Q<1$ is exhibited, for which the curve is slightly displaced from the previous ones. We note that if this agreement holds throughout the entire range
of charge and mass, this would imply that the characteristic or quasinormal frequencies of these BHs satisfy
\begin{equation}
\omega=\omega(a/\sqrt{M^2-Q^2})\,,\label{univ}
\end{equation}
which for small charge can also be written as $\omega=\omega(a/M+ay/M+...)$, where we defined $y \equiv 1-\sqrt{1-Q^2/M^2}$
following Refs.~\cite{Pani:2013ija,Pani:2013wsa}. This prediction was tested against linearized calculations in the slowly-rotating regime
from~\cite{Pani:2013ija,Pani:2013wsa}, where frequencies are expressed as $M\omega=M\omega(a=0)+a\left(f_0+f_1y+...\right)$.
Translated into this notation, universality as described by Eq.~\eqref{univ} would imply that $f_0=f_1$ for both the real and imaginary components,
which is to very good precision the result presented in Table~I of~\cite{Pani:2013ija,Pani:2013wsa} for $l=2$ modes.
While such universality seems to hold only for quadrupolar modes (and again, the linearized calculations of Refs.~\cite{Pani:2013ija,Pani:2013wsa} are also consistent with universality for $l=2$ only), the mere existence of such property is intriguing and adds to the isospectrality found
in linearized studies~\cite{Pani:2013ija,Pani:2013wsa}.
Such universality is not an artifact of horizon-deformation measures. Our results indicate that the gravitational-wave signal at large distances (in particular the $l=m=2,4$ components of the scalar $\Psi_{4}$) shares the same characteristics.
Recently, an analytical formalism to compute the quasinormal mode spectra of (weakly) charged black holes has been
introduced~\cite{Mark:2014aja}.
With it, the extent of this seemingly universal behavior can be scrutinized.
This has confirmed such behavior for large spin values (see also~\cite{Hod:2014uqa}), but it degrades
considerably at low ones~\cite{markhuan}.
\section{Conclusions}
\label{sec:final}
In this paper we have used the techniques developed in~\cite{Zilhao:2012gp,Zilhao:2013nda} for performing BH evolutions in Einstein-Maxwell theory to study the nonlinear stability of the Kerr-Newman BH for a variety of parameters and, in particular, for rapidly spinning BHs. On the timescales explored here (a few hundred $M$), we have seen no evidence for instabilities in any of the simulations performed. We are able to measure the spin of the BH with high accuracy (see Table~\ref{tab:runs}), which varies only within the expected margin for numerical error, indicating that the solution is stable.
In order to trigger potential instabilities, we have considered an initial perturbation of a particular type: a bar-mode perturbation in the metric coefficients. We do not expect, however, that other types of qualitatively different initial perturbations---like Brill or Teukolsky waves (see e.g.~\cite{Hilditch:2013cba}, for a recent study using this type of initial data in moving puncture gauge) will give different results; otherwise, an instability would appear to require very specific perturbations neither contained in our bar mode nor in the numerical noise of the initial data.
As such, our nonlinear analysis reinforces previous linear results~\cite{Pani:2013ija,Pani:2013wsa,Civin:2014bha} on the stability of the nonextremal KNBH. This contrasts with the instability found for extremal KNBHs~\cite{Aretakis:2012ei,Reiris:2013efa}. Thus, the latter, albeit continuously connected to nonextremal KNBHs in parameter space, seem qualitatively disconnected in terms of physical properties.
Our results have also uncovered, in the large rotation regime, a new class of universality for the quadrupolar quasinormal modes of these BHs:
they depend solely on the combination $a/a_{\rm max}$, a feature which had been observed previously in the perturbative regime
of slow-rotation. The significance of such results is unclear, but together with the isospectrality---observed also in the slow-rotation regime---hints at deeper relations at work also in rotating and charged geometries.
\begin{acknowledgments}
We would like to thank Paolo~Pani, Zachary~Mark and Huan~Yang for useful discussions.
M.Z.\ is supported by NSF grants OCI-0832606, PHY-0969855, AST-1028087, and PHY-1229173.
V.C.\ acknowledges financial support provided under the European Union's FP7 ERC Starting Grant ``The dynamics of black holes:
testing the limits of Einstein's theory'' grant agreement no.\ DyBHo--256667.
L.L.\ acknowledges support by NSERC through a Discovery Grant and CIFAR.
U.S.\ acknowledges support by
the FP7-PEOPLE-2011-CIG CBHEO Grant No. 293412,
the STFC Grant No. ST/I002006/1,
the XSEDE Grant No. PHY-090003 by the National Science Foundation,
the COSMOS Shared Memory system at DAMTP, University of Cambridge, operated on
behalf of the DiRAC HPC Facility and funded by BIS National E-infrastructure
capital Grant Nos.~ST/J005673/1 and ST/J001341/1
and STFC Grant Nos.~ST/H008586/1
and ST/K00333X/1, and
the Centro de Supercomputacion de Galicia (CESGA) under Grant No. ICTS-2013-249.
This research was supported in part by Perimeter Institute for Theoretical Physics.
Research at Perimeter Institute is supported by the Government of Canada through
Industry Canada and by the Province of Ontario through the Ministry of Economic Development
$\&$ Innovation.
This work was supported by the NRHEP 295189 FP7-PEOPLE-2011-IRSES Grant, and by FCT-Portugal through projects
PTDC/FIS/116625/2010, CERN/FP/123593/2011 and the IF program.
Computations were performed on the ``Baltasar Sete-Sois'' cluster at IST, the ``Blafis'' cluster at Universidade de Aveiro, the NICS Kraken Cluster,
the SDSC Trestles Cluster, Cambridge's COSMOS,
on the ``venus'' cluster at YITP, and CESGA's Finis Terrae.
\end{acknowledgments}
\bibliographystyle{myutphys}
|
\section{Introduction}
The $\nu$MSM (neutrino Minimal Standard
Model)~\cite{Asaka:2005an,Asaka:2005pn} is a simple extension of the
Standard Model (SM), explaining the origins of neutrino masses, dark
matter and baryon asymmetry of the universe at the same time. Three
right-handed neutrinos are introduced with Majorana masses below the
electroweak scale ${\cal O}(100)$~GeV, which realize the seesaw
mechanism~\cite{Seesaw} for neutrino masses with very suppressed
Yukawa couplings. The model predicts three heavy neutral leptons
$N_I$ ($I=1,2,3$) in addition to ordinary active neutrinos $\nu_i$
($i=1,2,3$).
The lightest heavy neutral lepton $N_1$ with ${\cal O}(10)$~keV mass
is a candidate for dark matter (see, for example, a
review~\cite{Boyarsky:2009ix}). The others $N_2$ and $N_3$ with
quasi-degenerate masses can generate baryon asymmetry of the universe
through the mechanism given in~\cite{Akhmedov:1998qx,Asaka:2005pn}.
Enough baryon asymmetry can be generated even if the degenerate mass
$M_N$ of $N_2$ and $N_3$ is as small as ${\cal
O}(1)$~MeV~\cite{Canetti:2010aw,Asaka:2013jfa}. However, the lower
bound on masses is further restricted to avoid constraints from direct
searches and cosmology~\cite{Gorbunov:2007ak}. The recent
analysis~\cite{Asaka:2013jfa} shows that $M_N > 163$~MeV for the
normal hierarchy (NH), while $M_N = 188-269$~MeV and $M_N>285$~MeV for
the inverted hierarchy (IH) of active neutrino masses. It is
remarkable that, thanks to the smallness of masses, the heavy neutral
leptons in the $\nu$MSM, especially $N_2$ and $N_3$, can be directly
tested by a variety of experiments and/or
observations~\cite{Kusenko:2004qc,Gorbunov:2007ak,Atre:2009rg,Asaka:2012hc,Asaka:2012bb}.
These heavy neutral leptons mix with flavor neutrinos and their mixing
elements are given by the ratios between Dirac and Majorana masses.
It is then possible to produce $N_I$ by decays of various mesons
through the mixing as the production of ordinary active neutrinos. As
an example, when they are sufficiently lighter than charged
kaon, the decays $K^+ \to e^+ N_I$ and $K^+ \to \mu^+ N_I$ are
possible. In fact, these channels are good targets for direct search
of heavy neutral leptons by using the technique of the so-called peak
search experiment~\cite{Shrock:1980vy}.
Furthermore, such decays may spoil lepton universality of charged
meson decay~\cite{Shrock:1980ct,Shrock:1981wq}. For instance, it is
possible that the ratio of decay rates ($M = \pi, K, \cdots$)
\begin{eqnarray}
\label{eq:LU}
R_M
= \frac{\Gamma (M^+ \to e^+ \nu)}{\Gamma (M^+ \to \mu^+ \nu)} \,,
\end{eqnarray}
is significantly different from the SM prediction. Although each
partial decay width receives considerable hadronic uncertainties, the
theoretical prediction can be very precise by taking the ratio, and
thus $R_M$ offers a promising test for physics beyond the SM. The
general expression for the contribution to $R_M$ from heavy neutral
leptons had already been presented in Ref.~\cite{Shrock:1980ct}.
Recently, Refs.~\cite{Abada:2012mc,Abada:2013aba} had revisited the
importance of this issue and violations of various universality
including $R_M$ had been extensively studied. Especially, the
numerical estimation of $R_M$ in the inverse seesaw model had been
performed. In addition, they had also pointed out that $R_M$ can be applied
in the $\nu$MSM.
In this letter, following these developments, we estimate the
possible deviation of $R_M$ induced by heavy neutral leptons in the
$\nu$MSM. The deviation strongly depends on masses and mixing
elements of $N_I$. The mixing elements $\Theta_{\alpha 1}$ of $N_1$
must be very suppressed in order to avoid various constraints of dark
matter. We then find that the contribution of $N_1$ to $R_M$ is
negligible. Thus, $N_2$ and $N_3$ for the seesaw mechanism and
baryogenesis give the dominant contributions to lepton universality.
The main purpose of this letter is to identify the possible deviations
of lepton universality in the $\nu$MSM. Hereafter, we first summarize
the constraints on heavy neutral leptons $N_2$ and $N_3$ and present
the allowed region of their mixing elements in Sec.~2.
We then consider in Sec.~3 lepton universality in decays of light
mesons, $R_K$ and $R_\pi$, in the $\nu$MSM and estimate the
deviations from the SM. Current status and future
perspective of experiments of lepton universality are also discussed.
Finally, Sec.~4 is devoted to conclusion.
\section{Heavy Neutral Leptons in the $\nu$MSM}
First of all, we explain briefly the $\nu$MSM. Three right-handed
neutrinos are introduced with Lagrangian
\begin{eqnarray}
{\cal L}
=
i \overline{\nu_{R I}} \gamma^\mu \partial_\mu \nu_{R I}
- F_{\alpha I} \overline {L_\alpha} \Phi \nu_{R I}
- \frac{M_I}{2} \overline{\nu_{R I}^c} \nu_{R I} + h.c. \
\end{eqnarray}
Here and hereafter, we follow the notation presented in
Ref.~\cite{Asaka:2011pb}. The seesaw mechanism works when Dirac
masses $F_{\alpha I} \langle \Phi \rangle$ are much smaller than
Majorana masses $M_I$. In this case mass eigenstates of neutrinos are
three active neutrinos $\nu_i$ with masses $m_i$ and three heavy
neutral leptons $N_I$ with masses $M_I$. Then, the neutrino mixing is
given by
\begin{eqnarray}
\nu_{L \alpha}
= U_{\alpha i} \, \nu_i + \Theta_{\alpha I} \, N_I^c \,,
\end{eqnarray}
where $U_{\alpha i}$ are elements of the PMNS
matrix~\cite{Pontecorvo:1958,Maki:1962mu}, and $\Theta_{\alpha I} =
F_{\alpha I} \langle \Phi \rangle/M_I$ are mixing elements of heavy
neutral leptons.
Heavy neutral lepton $N_1$ with $M_1 = {\cal O}(10)$~keV plays a role
of dark matter. The mixing elements of $N_1$ must be suppressed
enough since too large $|\Theta_{\alpha 1}|$ would lead to the
overclosure of the universe due to too much present abundance and also
would provide too much X-rays from its radiative decay $N_1 \to \nu
\gamma$%
\footnote{Recently, the unidentified line spectrum is observed
\cite{Bulbul:2014sua,Boyarsky:2014jta,Boyarsky:2014ska}, which can
be interpreted by X-ray lines emitted by sterile neutrino dark
matter ({\it i.e.}, $N_1$ in the considering model). }
(see Ref.~\cite{Boyarsky:2009ix}). It is then found that $N_1$ can
only give negligible contribution to the seesaw mass matrix of active
neutrinos and can essentially play no role in baryogenesis to avoid
these difficulties. In addition, as will be discussed later, $N_1$
contribution to the ratio $R_M$ in Eq.~(\ref{eq:LU}) can be neglected
compared with those from $N_2$ and $N_3$. Therefore, we shall take
$|\Theta_{\alpha 1}| = 0$ in this analysis for simplicity.
Heavy neutral leptons $N_2$ and $N_3$ are then responsible to the mass
matrix for active neutrinos via the seesaw mechanism and also the
baryogenesis via flavor oscillation. In this case, to realize the
seesaw mechanism Yukawa coupling constants $F_{\alpha I}$ of $N_2$ and
$N_3$ can be expressed as follows~\cite{Casas:2001sr}:
\begin{eqnarray}
\label{eq:FP}
F_{\alpha I} = \frac{i}{\langle \Phi \rangle} \,
\left[ U \, D_{\nu}^{\frac{1}{2}} \,
\Omega \, D_N^{\frac{1}{2}} \right]_{\alpha I}\,.
\end{eqnarray}
Here and hereafter we shall follow the notation in
Ref.~\cite{Asaka:2011pb}: $U$ represents the PMNS matrix,
\begin{eqnarray}
U=\left(
\begin{array}{ccc}
c_{12}c_{13} & s_{12}c_{13} & s_{13}e^{-i\delta} \\
-c_{23}s_{12}-s_{23}c_{12}s_{13}e^{i\delta}& c_{23}c_{12}-s_{23}s_{12}s_{13}e^{i\delta} & s_{23}c_{13} \\
s_{23}s_{12}-c_{23}c_{12}s_{13}e^{i\delta}& -s_{23}c_{12}-c_{23}s_{12}s_{13}e^{i\delta} & c_{23}c_{13} \\
\end{array}
\right) \times {\rm diag}(1, \ e^{i\eta}, \ 1 ),
\end{eqnarray}
with $s_{ij}= \sin \theta _{ij}$, $c_{ij}= \cos \theta _{ij}$,
$D_{\nu} ={\rm diag}(m_1,m_2,m_3)$ and $D_N ={\rm diag}(M_2, M_3)$. The matrix
$\Omega$ is given by
\begin{eqnarray}
\Omega = \left(
\begin{array}{ccc}
0 & 0 \\
\cos \omega & -\sin \omega \\
\xi \sin \omega&\xi \cos \omega \\
\end{array}
\right) \mbox{for the NH case}\,,~~
\Omega = \left(
\begin{array}{ccc}
\cos \omega & -\sin \omega \\
\xi \sin \omega& \xi \cos \omega \\
0 & 0 \\
\end{array}
\right)
\mbox{\rm for the IH case}\,.
\end{eqnarray}
The couplings are written in terms of parameters of active neutrinos
and heavy neutral leptons. The former ones consist of masses $m_i$ as
well as mixing angles $\theta_{ij}$, Dirac phase $\delta$ and Majorana
phase $\eta$ in the PMNS matrix.%
\footnote{ Since $N_1$ essentially decouples from the seesaw
mechanism, the lightest active neutrino obtains a mass smaller than
${\cal O}(10^{-5})$ eV~\cite{Asaka:2005an}. The number of Majorana
phases in the PMNS matrix is effectively reduced to be one (rather
than two in the usual case with three massive active neutrinos). }
The latter ones are a complex parameter $\omega$, masses $M_{2,3}$ and
the sign parameter $\xi$. As for the masses, the successful
baryogenesis requires that $N_2$ and $N_3$ are quasi-degenerate in
mass, and we write them in the form $M_3 = M_N + \Delta M/2$ and $M_2
= M_N - \Delta M/2$ with $\Delta M \ll M_N$.
The imaginary part of $\omega$ is important to determine
the typical size of the mixing elements since
$|\Theta_{\alpha I}| \propto X_\omega \equiv \exp (\mbox{Im}\omega)$.
In fact, as shown in Ref.~\cite{Asaka:2011pb},
$|\Theta_{\alpha I}|$ can be large
by taking $X_\omega \gg 1$ without changing masses of active neutrinos.
The mixing elements $\Theta_{\alpha I}$ characterize the strength of
interactions for heavy neutral leptons, and then receive constraints
from direct searches and cosmology. Interestingly, as pointed out in
Ref.~\cite{Gorbunov:2007ak}, the former ones place the upper bounds on
$|\Theta_{\alpha I}|$ while the latter one gives the upper bound on
lifetimes $\tau_{N_2}$ and $\tau_{N_3}$ leading to the lower bounds on
$|\Theta_{\alpha I}|$. Consequently, we may obtain the allowed range
of the mixing elements. Such regions have already been evaluated in
Refs.~\cite{Gorbunov:2007ak,Asaka:2013jfa}. Here we reconsider this
issue, especially taking into account for the first time
the preliminary result from the BNL-E949
experiment~\cite{E949:2014}. Notice that we shall restrict ourselves
for the case when $M_N < m_K - m_e$ because such heavy neutral
leptons, as we will show later, induce a significant deviation of the
lepton universality in kaon decay.
In deriving the allowed region, we construct Yukawa couplings of $N_2$
and $N_3$ by using the central values of $\theta_{ij}$ and $\Delta
m_{ij}^2$ from the global analysis of neutrino oscillations in
Ref.~\cite{Forero:2014bxa} and by varying all the possible ranges for
other free parameters. We then show the allowed range for the
combination of $\Theta_{\alpha I}$
\begin{eqnarray}
|\Theta|^2 \equiv \sum_{I=2,3} \sum_{\alpha = e, \mu, \tau}
|\Theta_{\alpha I}|^2 \,,
\end{eqnarray}
for a given $M_N$.%
\footnote{ The mass difference $\Delta M$ gives negligible corrections
to all the results in the present analysis, and hence we take
$\Delta M = 0$ for simplicity. } In our parameterization of Yukawa
couplings, it is written as
\begin{eqnarray}
| \Theta |^2 = \frac{ \sum_{i=1,2,3}m_i}{2 M_N}
(X_\omega^2 + X_\omega^{-2}) \,.
\end{eqnarray}
As for the bounds from direct search experiments, we first consider
the case when $M_N < 450$~MeV and use the results from the peak search
experiments~\cite{E104,Britton:1992xv,PIENU:2011aa,E949:2014} as well
as the beam-dump
experiments~\cite{Bernardi:1985ny,Bernardi:1987ek,Levy:1986w}. (See
the discussion later for the case in which $m_K - m_e > M_N >
450$~MeV.) Following Refs.~\cite{Kusenko:2004qc,Ruchayskiy:2011aa}
we have taken into account the corrections applying the bounds from
PS191 experiment~\cite{Bernardi:1985ny,Bernardi:1987ek,Levy:1986w} to
the $\nu$MSM, {\it i.e.}, the targets are two heavy neutral leptons $N_2$ and
$N_3$ which are Majorana particles (the target is one Dirac particle in the
original analysis), and the neutral current contributions for decays
of heavy neutral leptons are added (such a contribution is neglected in the
original analysis).
Moreover, the successful baryogenesis also gives the upper bounds on
the mixing elements in order to avoid the strong washout of the
produced asymmetry~\cite{Canetti:2010aw}. However, as shown in Ref.~\cite{Canetti:2010aw}, such bounds are much weaker than those from PS191 experiment in the considering mass range.
In this analysis, we also consider the recent
bound from BNL-E949 experiment~\cite{E949:2014}. It is the peak
search experiment in $K^+ \to \mu^+ \nu$ decay giving the upper bound
on $|\Theta_{\mu I}|^2$. Finally, to avoid the cosmological
difficulty we impose the lifetime bound $\tau_{N_{2,3}} < 0.1$
s~\cite{Dolgov:2000pj,Dolgov:2000jw}. Unfortunately, the analysis
in Refs.~\cite{Dolgov:2000pj,Dolgov:2000jw} has been done in the
different situation from the $\nu$MSM. We then also discuss the case
when the lifetime bound is relaxed as $\tau_{N_{2,3}} < 1$ s to make
the most conservative analysis. To evaluate $\tau _{N_{2,3}}$,
we use the formulae of the partial decay widths of heavy neutral leptons
given in Ref.~\cite{Gorbunov:2007ak}.
\begin{figure}[t]
\centerline{
\includegraphics[width=8cm]{MFIG_THsq_ALL_NHrev2.eps}%
\includegraphics[width=8cm]{MFIG_THsq_ALL_IHrev2.eps}%
}%
\caption{\it
Allowed region in the $M_N$-$|\Theta|^2$ plane for the NH
case (left panel) and IH case (right panel).
Allowed regions are shown by the shaded regions with
red-solid line or red-dashed line for the case with
the cosmological lifetime bound $\tau_{N_{2,3}}< 0.1$ s
or $\tau_{N_{2,3}}< 1$ s, respectively. The hatched regions are excluded by the
bounds from BNL-E949 experiment~\cite{E949:2014}.
}
\label{fig:AR_MN_THsq_1sec}
\end{figure}
The results are summarized in Fig.~\ref{fig:AR_MN_THsq_1sec}. We find
that BNL-E949 experiment gives the more stringent bound for
$M_N \simeq 180-260$~MeV compared with the bounds from PS191
experiment, which is seen by the hatched regions in Fig. 1.
(See also the result in Ref.~\cite{Asaka:2013jfa} for
comparison.%
\footnote{
Ref.~\cite{Asaka:2013jfa}
had used the data of global analysis of the neutrino oscillations in Ref.~\cite{Fogli:2012ua}
rather than Ref.~\cite{Forero:2014bxa} used in this analysis.
}) Especially, in the IH case, the lower bound on $M_N$
changes a lot by the inclusion of such a bound. We then find that the
allowed mass region when $\tau_{N_{2,3}} < 0.1$ s is%
\footnote{ For the NH case the small mass region $M_N = 208$--211~MeV
is excluded. }
\begin{eqnarray}
M_N >
\left\{
\begin{array}{l l}
173 \mbox{~MeV} & ~~\mbox{for the NH case}
\\
264\mbox{~MeV} & ~~\mbox{for the IH case}
\end{array}
\right. \,.
\end{eqnarray}
It should be noted that, if the cosmological upper bound of the
lifetime is relaxed as $\tau_{N_{2,3}} < 1$ s, the lower bound on
$M_N$ becomes smaller as $M_N > 122$~MeV and 136~MeV for the NH and IH
cases, respectively. See also Fig.~\ref{fig:AR_MN_THsq_1sec}.
Therefore, the cosmological bound on the lifetime is crucial
for determining the lower bound of the masses of $N_2$ and $N_3$.%
\footnote{
The lifetime bound for the case when $M_N < m_\pi$ had also been
discussed in Ref.~\cite{Ruchayskiy:2012si} and had shown that the
mass region $M_N < m_\pi$ is excluded. To make a very conservative
analysis, however, we also consider the case where the lifetime of
$N_{2,3}$ is longer than the limit in \cite{Ruchayskiy:2012si}.
}
It is seen that the allowed range in Fig.~\ref{fig:AR_MN_THsq_1sec}
is very limited for both NH and IH cases.
In practice, all such regions can be verified if the sensitivity
of $|\Theta|^2$ by future experiments will be improved by
a factor of ${\cal O}(10^2)$ or ${\cal O}(10^3)$
when applying the lifetime bound $\tau_{N_{2,3}}< $ 0.1
or 1~s, respectively. Such experiments will be
not only the peak search and beam-dump experiments,
but also the precision measurements of lepton universality
of light meson decays as shown below.
\section{Lepton Universality in the $\nu$MSM}
Let us discuss lepton universality of charged meson decays
shown in Eq.~(\ref{eq:LU}) in the context of the $\nu$MSM.
We first consider the universality in
charged kaon decay $R_K$.
The SM prediction of $R_K$ is
\begin{eqnarray}
R_K^{\rm SM} =
\left( \frac{m_e}{m_\mu} \right)^2
\left( \frac{ m_K^2 - m_e^2} {m_K^2 - m_\mu^2} \right)^2
\left( 1 + \delta R_K \right)
\,,
\end{eqnarray}
where $\delta R_K$ denotes the radiative correction.
Notice that $K^+ \to e^+ \nu_e$ and $K^+ \to \mu^+ \nu_\mu$ occur through
charged current interaction and their rates are
helicity-suppressed. It should be mentioned that
both decay rates receive the hadronic uncertainties,
{\it e.g.}, through the decay constant of parent meson, such uncertainties
cancel to a large extent by taking the ratio.
The theoretical prediction of the SM is thus very
precise as~\cite{Finkemeier:1995gi,Cirigliano:2007xi}
\begin{eqnarray}
R_K^{\rm SM } = (2.477 \pm 0.001) \times 10^{-5} \,.
\end{eqnarray}
In addition, the measurements at high precision have been done~\cite{Ambrosino:2009aa,Goudzovski:2010uk,NA62:2011aa,Lazzeroni:2012cx}.
The recent NA62 experiment provides~\cite{Lazzeroni:2012cx}
\begin{eqnarray}
R_K^{\rm exp} = (2.488 \pm 0.010) \times 10^{-5} \,.
\end{eqnarray}
It is seen that the observational data agrees with
the SM value at the 1$\sigma$ level.
Consequently, the deviation
\begin{eqnarray}
\Delta r_K = \frac{R_K}{R_K^{\rm SM}} - 1 \,,
\end{eqnarray}
is as small as
\begin{eqnarray}
\Delta r_K = (4 \pm 4) \times 10^{-3} \,,
\end{eqnarray}
and thus it provides a powerful probe for physics beyond the SM.
In the $\nu$MSM, $K^+$ is possible to decay into not only active
neutrinos $\nu_i$ but also heavy neutral leptons $N_I$
depending on $M_I$. Then, the ratio $R_K$ is given by
\begin{eqnarray}
R_K \eqn{=}
\frac{
\sum_{i=1,2,3} \Gamma (K^+ \to e^+ \nu_i)
+
\sum_{I=1,2,3} \Gamma (K^+ \to e^+ N_I)
}{
\sum_{i=1,2,3} \Gamma (K^+ \to \mu^+ \nu_i)
+
\sum_{I=1,2,3} \Gamma (K^+ \to \mu^+ N_I)
} \,.
\end{eqnarray}
The general expression of $R_M$ in the presence of heavy neutral
leptons has been given by Ref.~\cite{Shrock:1980ct}. (See Eq.~(3.2) in
Ref.~\cite{Shrock:1980ct}.) By neglecting the masses of active
neutrinos and the experimental energy thresholds of charged leptons in
kaon decays, the deviation is~\cite{Shrock:1980ct}
\begin{eqnarray}
\Delta r_K
\eqn{=}
\frac{
\sum_{i=1,2,3} |U_{e i}|^2 +
\sum_{I=1,2,3} |\Theta_{e I}|^2 G_{e I}
}{
\sum_{i=1,2,3} |U_{\mu i}|^2 +
\sum_{I=1,2,3} |\Theta_{\mu I}|^2 G_{\mu I}
}
- 1 \,,
\label{eq:DelrK0}
\end{eqnarray}
where $G_{\alpha I} = 0$ if $M_I > m_K - m_{\ell_\alpha}$; otherwise
\begin{eqnarray}
G_{\alpha I} =
\frac{ r_\alpha + r_I - (r_\alpha - r_I)^2 }
{ r_\alpha ( 1 - r_\alpha)^2}
\sqrt{ 1 - 2 (r_\alpha + r_I) +(r_\alpha - r_I)^2 } \,,
\end{eqnarray}
with $r_\alpha = m_{\ell_\alpha}^2/m_K^2$ and $r_I = M_I^2 /m_K^2$.
(See Ref.~\cite{Gorbunov:2007ak} for the expressions of $\Gamma(K^+\rightarrow
l_{\alpha}^+N_I)$.)
The physical importance of $\Delta r_K$ (and also $\Delta r_\pi$ in
the later discussion) had been readdressed in
Refs.~\cite{Abada:2012mc,Abada:2013aba}. The main origins of such
deviations are (i) the additional contributions to the kaon decay from
heavy neutral leptons and (ii) the deviation from the unitarity of the
PMNS mixing matrix of active
neutrinos~\cite{Shrock:1980ct,Shrock:1981wq,Abada:2012mc,Abada:2013aba}.
Refs.~\cite{Abada:2012mc,Abada:2013aba} had presented the possible
range of $\Delta r_{K, \pi}$ in the inverse seesaw model and also had
pointed out that $\Delta r_{K, \pi}$ in Eq.~(\ref{eq:DelrK0}) can be
applied to the $\nu$MSM.
Based on these analyses, we would like to derive the predicted range
of $\Delta r_K$ in the $\nu$MSM.
First of all, it should be noted that
the mixing elements of active neutrinos and heavy neutral leptons
satisfy the unitarity condition
\begin{eqnarray}
\label{eq:Unitarity}
\sum_{i=1,2,3} |U_{\alpha i}|^2
+ \sum_{I=1,2,3} |\Theta_{\alpha I}|^2 = 1 \,.
\end{eqnarray}
It is seen that the violation of the unitarity in the PMNS matrix $U$
is very suppressed at ${\cal O}(|\Theta_{\alpha I}|^2)$
in this framework (see Fig.~\ref{fig:AR_MN_THsq_1sec}).
From the above condition
$\Delta r_K$ in Eq.~(\ref{eq:DelrK0}) can be written as
\begin{eqnarray}
\Delta r_K
\eqn{=}
\frac{
1 +
\sum_{I=1,2,3} |\Theta_{e I}|^2
\left[ G_{e I} - 1 \right]
}{
1 +
\sum_{I=1,2,3} |\Theta_{\mu I}|^2
\left[ G_{\mu I} - 1 \right]
}
- 1
\,.
\label{eq:delRk}
\end{eqnarray}
Therefore, we find that the deviation $\Delta r_K$ in the $\nu$MSM is
determined by the masses $M_I$ and mixing elements $\Theta_{\alpha I}$
of heavy neutral leptons.
Note that $\Delta r_K$ does not depend
explicitly on the PMNS matrix elements, but it
depends on them implicitly through $\Theta_{\alpha I}$.
(See the parametrization of Yukawa couplings of $N_I$ in
Eq.~(\ref{eq:FP}).) Since the mixing elements of dark matter
$N_1$ must be very small, we can safely neglect its contribution to
$\Delta r_K$.
First, we consider the case when $M_N < m_K - m_\mu$, {\it i.e.}, both $K^+
\to \mu^+ N_I$ and $K^+ \to e^+ N_I$ are kinematically allowed. In
this case one might expect that the deviation $\Delta r_K$ is very
suppressed as ${\cal O}(10^{-9})$--${\cal O}(10^{-7})$ since
$|\Theta|^2$ should be in such a range as shown in
Fig.~\ref{fig:AR_MN_THsq_1sec}. Decay rate of $K^+ \to \ell_\alpha^+
N_I$ (and then $G_{\alpha I}$) is, however, enhanced by
$(M_I/m_{\ell_\alpha})^2$ compared with $K^+ \to \ell_\alpha^+
\nu_\alpha$ due to the helicity suppression~\cite{Shrock:1980ct}. Interestingly, since
this enhancement factor is much larger for the decay into $e^+$ than
that into $\mu^+$, the $\nu$MSM predicts a positive $\Delta r_K$ in
this mass region as
\begin{eqnarray}
\label{eq:DELRK_ap}
\Delta r_K \simeq \sum_{I=2,3} | \Theta_{eI} |^2 \,
\frac{M_N^2}{m_e^2}
\left(
1- \frac{M_N^2}{m_K^2} \right)^2 \,.
\end{eqnarray}
Moreover, the upper limit of $\Delta r_K$ is then
derived from the upper bounds on the mixing elements $|\Theta_{e I}|$.
Such elements are severely restricted by PS191 experiment looking for
the production and decay modes $K^+ \to e^+ N_I$ and $N_I \to e^+
\pi^-, e^- \pi^+$, {\it e.g.}, $|\Theta_{e I}|^2 < {\cal O}(10^{-9})$--${\cal
O}(10^{-8})$ for $M_N \simeq 200$--400~MeV. Therefore, we expect
$\Delta r_K <{\cal O}(10^{-4})$--${\cal O}(10^{-3})$ by taking into
account the enhancement factor of $(M_N/m_e)^2 \sim10^{5}$.
When $m_K - m_e >
M_N > m_K - m_\mu$, $K^+ \to \mu^+ N_I$ is forbidden, but the behavior
of the correction $\Delta r_K$ is very similar to the above case.
On the other hand, when $M_N > m_K - m_e$, the situation is changed.
We should note that, even if $K^+ \to \mu^+ N_I$
and $K^+ \to e^+ N_I$ are kinematically forbidden,
the correction of $\Delta r_K$ is induced due to the non-unitarity
of the PMNS matrix (see Eq.~(\ref{eq:Unitarity})) as
\begin{eqnarray}
\label{eq:DEL_rK_h}
\Delta r_K \simeq \sum_{I = 2,3}
\bigl( |\Theta_{\mu I}|^2 - |\Theta_{e I}|^2 \bigr) \,.
\end{eqnarray}
In this case the sign of $\Delta r_K$ is determined according to the
relative sizes of $|\Theta_{\mu I}|^2$ and $|\Theta_{e I}|^2$ and the
magnitude is $|\Delta r_K | \mathop{}_{\textstyle \sim}^{\textstyle <} |\Theta|^2 = {\cal O}(10^{-9})$--${\cal
O}(10^{-7})$.
Now, we are at the point to present the numerical prediction of
$\Delta r_K$ in the $\nu$MSM. As explained in Sec.~2, we impose the
constraints from direct search experiments and cosmological lifetime bound.
The possible range of $\Delta r_K$ by varying all the free parameters
is shown in Fig.~\ref{fig:DEL_RK_1sec}. It is found that $\Delta r_K
={\cal O}(10^{-7})$--${\cal O}(10^{-3})$ for the NH case, and $\Delta
r_K = {\cal O}(10^{-6})$--${\cal O}(10^{-3})$ for the IH case,
where we have considered $M_N < 450$ MeV and
$\tau_{N_{2,3}}<0.1$ s.
The predicted region becomes wider if the
lifetime bound is relaxed as $\tau_{N_{2,3}}<1$ s.
\begin{figure}[t]
\centerline{
\includegraphics[width=8cm]{MFIG_DELRK_NH_1sec.eps}%
\includegraphics[width=8cm]{MFIG_DELRK_IH_1sec.eps}%
}%
\caption{\it
$\Delta r_K$ in the $\nu$MSM for the NH case (left panel)
and IH case (right panel).
Possible regions are shown by the shaded regions with
red-solid line or red-dashed line for the case with
the cosmological lifetime bound $\tau_{N_{2,3}}< 0.1$ s
or $\tau_{N_{2,3}}< 1$ s.
The horizontal (black dotted) lines are
$\Delta r_K = 4 \times 10^{-3}$ (current central value~\cite{Lazzeroni:2012cx})and
$\Delta r_K = 10^{-3}$ (which will be reached by
the near future experiments).
}
\label{fig:DEL_RK_1sec}
\end{figure}
The search bounds place the upper limit while the lifetime bound places
the lower limit of $\Delta r_K$, and hence the
$\nu$MSM predicts $\Delta r_K$ in certain range. We find that the
predicted range is indeed consistent with the current upper bound at 3
$\sigma$ level, $\Delta r_K < 1.2 \times 10^{-2}$.
We have considered the mass range $M_N < 450$ MeV so far. When $M_N >
450$ MeV, there is no stringent constraint on the mixing elements from
PS191 experiment. So, we expect a large $\Delta r_K$ for $450$ MeV $<
M_N < m_K - m_e$. In such a case, the upper bounds on
$|\Theta_{\alpha I}|$ are placed from CHARM and CHARM
II~\cite{Bergsma:1983rt, Bergsma:1985is, Vilain:1994vg},
IHEP-JINR~\cite{Baranov:1992vq} and NuTeV\cite{Vaitaitis:1999wq}
experiments. When $M_N$ is just above 450~MeV, the most stringent
bound on $|\Theta_{eI}|^2$ is obtained from IHEP-JINR and the bound in
Fig.~4 of Ref.~\cite{Baranov:1992vq} is weaker than that of
PS191~\cite{Levy:1986w} by a factor of $\sim 40$. This means that
$\Delta r_K$ can be $\sim 4 \times 10^{-3}$ in such a value of $M_N$.
In addition, the successful scenario of baryogenesis also puts the
important bound of the mixing elements for such mass regions. We,
however, find that such a bound on $|\Theta|^2$ in
Ref.~\cite{Canetti:2012kh} is slightly weaker than the above IHEP-JINR
bound on $|\Theta_{eI}|^2$. This shows that
search for heavy neutral leptons with $M_N$ just above 450~MeV
is very interesting since the present bounds from
beam-dump experiments, baryogenesis and also lepton universality
are very competitive and may be possible to be cross-checked
in various ways.
More precise estimation of these bounds as well as
$\Delta r_K$ in this case will be done elsewhere~\cite{AET}.
Near future experiments (such as NA62 at
CERN~\cite{Goudzovski:2012gh}, ORKA at
FNAL~\cite{E.T.WorcesterfortheORKA:2013cya} and TREK/E36 at
J-PARC~\cite{Kohl:2013rma}) will achieve the sensitivity $\Delta r_K =
10^{-3}$. Therefore, it is very interesting that these experiments
will start to probe the predicted region in the $\nu$MSM.
In particular, large $\Delta r_K$ are obtained when $M_N \sim 180$ MeV
and just above 450 MeV. Such mass regions will also be tested by
experiments using different search techniques, like
the peak search and/or beam-dump experiments,
in decays of kaon and charmed-mesons, respectively.
Next, we turn to consider lepton universality in pion decay.
The theoretical prediction of the SM is~\cite{Cirigliano:2007xi}
\begin{eqnarray}
R_\pi^{\rm SM} = (1.2352 \pm 0.0001) \times 10^{-4} \,,
\end{eqnarray}
while the experimental value is~\cite{Beringer:1900zz}%
\footnote{ Here we have cited the averaged value of Particle Data
Group~\cite{Beringer:1900zz}. The recent measurements at TRIUMF and
PSI give $R_\pi = (1.2265 \pm 0.0034 \pm 0.0044) \times
10^{-4}$~\cite{Britton:1992pg} and $R_\pi= (1.2346 \pm 0.0035 \pm
0.0036) \times 10^{-4}$~\cite{Czapek:1993kc}, respectively. }
\begin{eqnarray}
R_\pi^{\rm exp} = ( 1.230 \pm 0.004 ) \times 10^{-4} \,.
\end{eqnarray}
The deviation is then given as
\begin{eqnarray}
\label{eq:DEL_rpi}
\Delta r_\pi = ( -4 \pm 3 ) \times 10^{-3} \,.
\end{eqnarray}
\begin{figure}[t]
\centerline{
\includegraphics[width=8cm]{MFIG_DELRP_1sec.eps}%
}%
\caption{\it $\Delta r_\pi$ in the $\nu$MSM. Possible region is
shown by the shaded region with red-solid line or blue-dashed
line for the NH or IH case, respectively. Here we impose the
cosmological lifetime bound $\tau_{N_{2,3}}< 1$ s. The
horizontal (black dotted) line is $\Delta r_\pi = 5 \times
10^{-4}$ (which will be reached by the near future experiments).
}
\label{fig:DEL_RP_1sec}
\end{figure}
In the considering model, $\pi^+ \to \mu^+ N_{2,3}$ are impossible even
if the lifetime bound is relaxed as $\tau_{N_{2,3}}< 1$ s and then
we restrict ourselves to the case with $M_N > m_\pi - m_\mu$. When
$\pi^+ \to e^+ N_{2,3}$ are available, the sizable correction to
$R_\pi$ is expected due to the enhancement factor of $(M_N/m_e)^2$ and
its maximal value is determined by the upper bounds of
$|\Theta_{eI}|^2$. Then, the approximate form of $\Delta r_\pi$
is given by
\begin{eqnarray}
\Delta r_\pi \simeq \sum_{I=2,3} | \Theta_{eI} |^2 \,
\frac{M_N^2}{m_e^2}
\left(
1- \frac{M_N^2}{m_\pi^2} \right)^2 \,,
\end{eqnarray}
similar to Eq.~(\ref{eq:DELRK_ap}).
Moreover, the sign of $\Delta r_\pi$
is positive as in the kaon decay, and thus, even if $N_{2}$
and $N_3$ were allowed to be lighter than pion (to be precise $m_\pi -
m_e$), they would be conflict with $R_\pi^{\rm exp}$ at 1 $\sigma$
level (see Eq.~(\ref{eq:DEL_rpi})). As shown in
Fig.~\ref{fig:DEL_RP_1sec}, we find numerically that the predicted
range is $\Delta r_\pi < {\cal O}(10^{-4})$, and then it is consistent
with $R_\pi^{\rm exp}$ at 2 $\sigma$ level.
Notice that the experiments like PIENU at
TRIUMF\cite{Malbrunot:2011zz} and PEN at PSI~\cite{Pocanic:2009zz}
will improve the sensitivity at the level $\Delta r_\pi \simeq 0.05
-0.06$ \% (see also Ref.~\cite{Bryman:2011zz}), which is slightly
above the predicted range. Thus, the further improvement may be
required to probe $\Delta r_\pi$ in the $\nu$MSM.
When $M_N$ becomes larger than $m_\pi - m_e$, the non-unitarity of the
PMNS mixing matrix for active neutrinos induces the correction
\begin{eqnarray}
\Delta r_\pi \simeq \sum_{I= 2,3}
\bigl( |\Theta_{\mu I}|^2 - |\Theta_{e I}|^2 \bigr) \,,
\end{eqnarray}
as in the kaon decay (see Eq. (\ref{eq:DEL_rK_h})). In this case the magnitude
of $\Delta r_\pi$ is too small to be probed in near future
experiments. It is, however, interesting to notice that $\Delta
r_\pi$ and $\Delta r_K$ become the same when $M_N > m_K - m_e$.
We have so far discussed the corrections to lepton universality in
kaon and pion decays. It should be noted that heavy neutral leptons
$N_2$ and $N_3$ may lead to violations of lepton universality in
decays of charmed mesons, beauty mesons and tauon. See the recent
analysis in Ref.~\cite{Abada:2013aba}. The comprehensive
study for the test of the $\nu$MSM by lepton universality will be
discussed elsewhere~\cite{AET}.
\section{Conclusions}
We have discussed lepton universality of charged meson decays in
the $\nu$MSM. Among three heavy neutral leptons, $N_2$ and $N_3$,
which explain the seesaw mechanism for active neutrino masses and the
baryogenesis via their flavor oscillation, may induce the violations
of such universality
due to the non-unitarity of the mixing matrix of active neutrinos and
the additional contributions to meson decays.
The deviation of lepton universality in kaon decay $R_K$ has been
found to be as large as $\Delta r_K ={\cal O}(10^{-3})$ when applying the
cosmological bound on lifetime as $\tau_{N_{2,3}}< 0.1$ s. Such a
large $\Delta r_K$ is possible when $M_N \sim 180$ MeV and just above
$450$ MeV. Further, if
the cosmological bound on the lifetime is weak as $\tau_{N_{2,3}}
\mathop{}_{\textstyle \sim}^{\textstyle <} 1$ s, $\Delta r_K$ can be larger as ${\cal O}(10^{-2})$.
Notice that the sign of $\Delta r_K$ is always positive in the case
when $K^+ \to e^+ N_{2,3}$ are open. Furthermore, we have also
discussed lepton universality in pion decay. When $\pi^+ \to e^+
N_{2,3}$ are allowed by relaxing the lifetime bound, the deviation can
be as large as $\Delta r_\pi ={\cal O}(10^{-4})$.
Such regions of the model will begin to be explored by near future
experiments; the experiments of lepton universality in kaon decay as
NA62, ORKA and TREK/E36 experiments and those in pion decay as PIENU
and PEN experiments. It should be noted that
such regions are also good targets of direct search experiments
using the different methods (the peak search experiments,
the beam-dump experiments, and so on).
These facilities might reveal physics of $N_2$
and $N_3$, namely the origins of neutrino masses and baryon asymmetry
of the universe.
\section*{Acknowledgments}
T.A. was supported by JSPS KAKENHI Grant Numbers 25400249 and 26105508,
and S.E. was supported by Sinergia grant of the Swiss National Science Foundation CRSII2 141939.
|
\section*{Introduction}
Given finite sets $A$ and $B$ of real numbers, the \emph{sum set} and \emph{product set} of $A$ and $B$ are defined as
$$A+B = \{ a+b : a\in A, b\in B \}, \quad AB = \{ ab : a\in A, b\in B \}.$$
Erd\H{o}s and Szemer\'{e}di conjectured that at least one of $|A+A|$ or $|AA|$ is large with respect to $|A|$. Specifically, they conjectured the following.\footnotemark
\begin{conjecture}[Erd\H{o}s - Szemer\'{e}di] \label{conj:erdsze}
For all $A \subseteq \mathbb Z$ a finite set, and for all $\epsilon > 0$, we have
$$|AA| + |A+A| \gg |A|^{2 - \epsilon}.$$
\end{conjecture}
\footnotetext{In this paper we use the standard notation $X \ll Y$ to mean that there exists an absolute constant $c$ with $X \leq cY$. We have $Y \gg X$ iff $X \ll Y$. We write $X\sim Y$ to denote the existence of constants $0 < c_1\leq c_2 $ so that $c_1X \leq Y \leq c_2X$.
Additionally, the symbols $\lesssim$ and $\gtrsim$ are used to suppress logarithmic factors.} This conjecture remains open, and has given rise to the study of the \emph{sum-product} phenomenon, which, loosely defined, is the notion that finite sets cannot be simultaneously additively and multiplicatively structured. Conjecture \ref{conj:erdsze} is believed to be true over the real numbers, where current progress is given by Rudnev and Stevens \cite{MishaSophie}.
There are many variants of this problem in the literature; one family of such variants are concerned with convex functions\footnote{In this paper all convex functions considered are \emph{strictly} convex functions. Furthermore, our results also apply to strictly \emph{concave} functions.}. Such results quantify the notion that \emph{convex functions destroy additive structure.}
Some examples of common problems in this area are the following:
For $A \subseteq \mathbb R$ a finite set, and $f$ be a convex function:
\begin{itemize}
\item Is the set $A + f(A)$ always large?
\item Is at least one of the sets $A+A$ or $f(A) + f(A)$ is always large?
\end{itemize}
Much research has been done towards these problems and their variants, see for instance \cite{elekesnathansonruzsa,higherconvexity, liolly}. This is also related to the notion of a \emph{convex set}, that is, a set $A = \{ a_1 < a_2 < ... <a_n \}$ such that $a_{i+1} -a_i > a_i - a_{i-1}$ for all $2 \leq i \leq n-1$. Any convex set is the image of the interval $[n]$ under some convex function $f$. Current progress for these problems is given, respectively, by Li and Roche-Newton, \cite{liolly} and Shkredov \cite{shkredovconvex}.
\begin{theorem}[Li, Roche-Newton] \label{convex2original}
Let $A \subseteq \mathbb R$ be a finite set, and let $f$ be a convex function. Then we have
$$|A + f(A)| \gtrsim |A|^{\frac{24}{19}}.$$
\end{theorem}
\begin{theorem}[Shkredov] \label{convexoriginal}
Let $A \subseteq \mathbb R$ be a finite set, and let $f$ be a convex function. Then we have
$$|A+A| + |f(A) + f(A)| \gtrsim |A|^{\frac{100}{79}}.$$
\end{theorem}
These problems are also related to \emph{expander} results. Results of this nature state that some set, defined by (typically polynomial) combinations of elements of $A$, is \emph{always} large. Two of the simplest examples of expanders are the sets
$$AA +A = \{ ab+c : a,b,c \in A \}, \qquad A(A+1) = \{a(b+1) : a,b \in A \}$$
which are both expected to have size $|A|^{2 - \epsilon}$ for all $\epsilon>0$. In fact, the expander $A(A+1)$ is a special case of the set $A + f(A)$ from above. The current bounds in the literature for these expanders are due to Roche-Newton and Warren \cite{ollyaudie} and Jones and Roche-Newton \cite{jonesolly}, respectively.
\begin{theorem}[Roche-Newton, Warren]\label{AA+Aoriginal}
For all $A \subseteq \mathbb R$ finite, we have
$$|AA+A| \gtrsim |A|^{\frac{3}{2} + \frac{1}{194}}.$$
\end{theorem}
\begin{theorem}[Jones, Roche-Newton]\label{A(A+1)original}
For all $A \subseteq \mathbb R$ finite, we have
$$|A(A+1)| \gtrsim |A|^{\frac{24}{19}}.$$
\end{theorem}
\subsection*{Main results}
The proof of the sum-product result in \cite{MishaSophie} makes use of a combination of techniques used previously in the real numbers, combined with a technique used to prove sum-product results in finite fields, see \cite{rss}. In this paper we extend these techniques to give both quantitative and qualitative improvements to the problems mentioned above. Note that we make no attempt to optimise the logarithmic factors in our results, since in all cases the polynomial factor exponents are not expected to be tight. Our main result is the following.
\begin{theorem}\label{thm:main}
Let $A, B \subseteq \mathbb R$ be finite sets, and let $f$ and $g$ each be either a convex or concave function. Then we have
$$|A+B|^{38}|f(A) + g(B)|^{38} \gtrsim (|A||B|)^{49}.$$
\end{theorem}
For certain choices of $A,B,f,$ and $g$, this theorem implies improvements to many of the problems mentioned above. Firstly, we can recover the following improvements to Theorems \ref{convex2original} and \ref{convexoriginal}.
\begin{corollary} \label{cor:convex functions}
For all $A \subseteq \mathbb R$ finite, and $f$ a convex function, we have
$$|A + f(A)| \gtrsim |A|^{49/38},$$
$$|A + A| + |f(A) + f(A)| \gtrsim |A|^{49/38}.$$
\end{corollary}
The first inequality follows from setting $B = f(A)$ and $g = f^{-1}$. The second follows from setting $B = A$ and $f = g$. By slightly adjusting the proof of Theorem \ref{thm:main}, we can obtain a better bound for differences.
\begin{corollary} \label{cor:convex diff}
For all $A \subseteq \mathbb R$ finite, and $f$ a convex function, we have
$$|A - A|^5|f(A) - f(A)|^5 \gtrsim |A|^{13}.$$
\end{corollary}
In the case $A = [1,\dots,n]$, this matches the bounds of Schoen and Shkredov \cite{schoenshkredov} and Rudnev and Stevens \cite{MishaSophie} for estimates on differences and sums of convex sets respectively. Furthermore we match the result of Li and Roche-Newton \cite{liolly} in the case of few differences, many convex differences.
Secondly, we find an asymmetric sum-product result.
\begin{corollary}\label{sum-productasym}
For all $A,B \subseteq \mathbb R$ finite, we have
$$|AB|^{38}|A+B|^{38} \gtrsim (|A||B|)^{49}$$
\end{corollary}
This follows from setting $A = X$, $B = Y$, $f = g = \log(x)$.
Corollary~\ref{sum-productasym} appears to be a little studied variant of the asymmetric sum-product problem: One example of a result in this direction is by Solymosi \cite{Solymosi}, who showed that $|A+A||B+B||AB|\gtrsim |A|^2|B|^2$. There has also been work towards the more difficult problem of finding a lower bound on $|A+B||AC|$, see for instance \cite{elekesnathansonruzsa}, or \cite[Theorem 10]{brendan-richlinesgrids},
where the results are rather of a qualitative nature.
The statement of Corollary~\ref{sum-productasym} is particularly interesting in the extremal cases of `few sums' or `few products': e.g. if $|A|=|B|=N$ and $|A+B|\lesssim N$, then $|AB|\gtrsim N^{\frac32 + \frac3{38}}$. Typically the exponent of $3/2$ is a barrier in sum-product estimates, and so in this sense, Corollary~\ref{sum-productasym} is threshold-breaking.
Thirdly we give some results demonstrating the principle that
`translation destroys multiplicative structure', in particular improving Theorem \ref{A(A+1)original}.
\begin{corollary} \label{cor:A(A+1)}
For all $A,B \subseteq \mathbb R$ finite, we have
$$|A(A+1)| \gtrsim |A|^{49/38},$$
$$|AB| + |(A+1)(B+1)| \gtrsim (|A||B|)^{49/76}.$$
\end{corollary}
Finally, by combining techniques used in the proof of Theorem \ref{thm:main} with the method of Roche-Newton and Warren, we can give an improvement and generalisation of Theorem \ref{AA+Aoriginal}.
\begin{theorem}\label{thm: AB+A}
Let $A,B \subseteq \mathbb R$ be finite sets with $|A| \sim |B|$. Then we have
\[ |AB+A| \gtrsim |A|^{\frac{3}{2} + \frac3{170}}.\]
\end{theorem}
\subsection*{Techniques}
Here we give an overview of the techniques that we use, hinting at the aspects of our method that are most amenable to future improvements.
These techniques can be summarised as follows:
\begin{enumerate}
\item The Szemerédi-Trotter theorem gives good bounds on $\mathsf{E}_3^+(A,B)$, especially if we have data of the form $r_{QR}(a)\geq T$ for each $a\in A$. Similarly,
the Szemerédi-Trotter theorem gives good bounds on $\mathsf{E}_3^+(f(A),B)$ for a convex function $f$, if we have data of the form $r_{Q-R}(a)\geq T$ for each $a\in A$.
\item Using a regularisation result, we can find a subset $C\subseteq A$ so that $|C|\gtrsim |A|$ and for which we have the additive data $r_{Q-R}(a) \geq T$ for each $c\in C$.
\item We can count solutions $(a,b,c)$ to a tautological equation of the form $a-b = (a + c) - (b + c)$, where we insist that $a-b,a + c$ are in certain (different) sets via third moment energy bounds. This gives an auxiliary energy bound, see Proposition \ref{prop: energy general} below.
\item A corollary of the regularisation result (see Corollary \ref{c:E_3(B,V)E_3(f(C),U)} below) allowing us to upper bound certain products of energies, together with this auxiliary energy bound, leads to the result.
\end{enumerate}
Underlying many results about expander sets in $\mathbb{R}$ (with few variables) is the Szemerédi-Trotter theorem. It is common knowledge that the Szemerédi-Trotter theorem is particularly strong for finding bounds on the third moment energy $\mathsf{E}_3^+(A,B)$, an idea first introduced by Schoen and Shkredov \cite{schoenshkredov}. This, in part, is due to the `trick' that every element of $A$ can be written as a product of elements of $AA$ and $A$ in at least $|A|$ ways: $a = (ab)/b$ for any choice of $b \in A$ (we assume here that $0\notin A$). However, if one has additional multiplicative structure on $A$, say $r_{QR}(a)\geq T$ for each $a\in A$ and some auxiliary sets $Q $ and $R$ and a number $T$, one can use this information in place of the aforementioned `trick'. This gives a third moment energy bound in terms of $Q,R$ and $T$, the strength of which depends on the strength of the multiplicative information. This is the idea behind the so-called \emph{Szemerédi-Trotter sets} introduced by Shkredov \cite{shkredovconvex}, for which the notation $d^+(A)$ (and variants thereof) is used. We note that an analogue of this idea takes place in $\mathbb{F}_p$ using the point-line incidence bound of Stevens-de Zeeuw \cite{sophiefrank} in place of the Szemerédi-Trotter theorem, which naturally produces a bound on the fourth moment energy. For a convex function $f$, this trick changes as follows: we can obtain bounds on $\mathsf{E}_3^+(f(A),B)$ if we have \emph{additive} structure on $A$, say $r_{Q-A}(a)\geq T$ for all $a\in A$.
To benefit from the `enhanced energy trick' described above, we need the appropriate data on $Q, R$ and $T$. A generic technique for this, first described in \cite{energy} and refined in \cite{shkredovbw}, yields a subset $C\subseteq A$ with suitable parameters: that is, if $\mathsf{E}_3^+(A)\sim |D_t|t^3$ for some $D_t\subseteq A-A$, then $r_{D_t - A}(c)\geq |D_t|t|A|^{-1}$, and $|C|\geq |D_t|t|A|^{-1}$. A recent expository lemma of Xue \cite{xue} enhances the strength of this result, to enable one to take $|C|\gtrsim |A|$ - we use an adaptation of this regularisation result.
We conclude this section by considering where improvements to these techniques may be found. Certainly for the real numbers, there is hope that one could find a more optimised subset of $A$, with the data on $Q,R$ and $T$ optimised for the specific applications within our paper. Indeed, such a `better subset' is present in the current bounds for the sum-product problem \cite{MishaSophie}). In \cite{MishaSophie}, an elementary, somewhat geometric, argument justifies the existence of the subset used in the context of the sum-product problem.
The third item of our list might also be improved as follows: we bound the number of solutions to $a-b = (a+c)- (b+c)$ in terms of the third moment energy. During this argument, we use Cauchy - Schwarz to bound a factor of $\mathsf{E}^+_{3/2}(A,\cdot)$ which appears as a by-product of H\"{o}lder's inequality. However, it may be possible to directly bound $\mathsf{E}^+_{3/2}(A,\cdot)$ using other methods. For example, if $A$ is a convex set, then Solymosi and Ruzsa \cite{solymosiruzsa} show that $\mathsf{E}^+_{3/2}(A,B) \ll |A+B|^{3/2}$ for any set $B$.
In the proof of Theorem \ref{thm: AB+A}, we (implicitly) turn to the recent technique of studying the \emph{line energy} (see e.g. \cite{affinegroup4, ollyaudie}). We would not be surprised if future developments of this concept provide further tools relevant to the results in this paper.
\section{Preliminaries}
We use the notation $r_{Q-R}(a)$ to denote the number of representations of the element $a$ as a product from $Q-R$, that is, $r_{Q-R}(a) = |\{(q,r)\in Q\times r: q-r = a\}|$, and similarly for $r_{QR}(a)$ etc. The $k$th moment additive energy between sets $A$ and $B$ is defined to be
\[
\mathsf{E}_k^+(A,B):= \sum_{x\in A-B}r_{A-B}^k(x)
\]
for $k\geq 1$. If $A=B$ we simply write $\mathsf{E}_k^+(A)$. Similarly, we define the multiplicative energy $\mathsf{E}_k^\times(A,B):= \sum_x r_{A/B}^k(x)$.
\subsection{Energy Bounds via Szemer\'{e}di - Trotter}
Before beginning the proofs, we require some technical lemmas. The first gives a bound for the additive energy of two sets $A$ and $B$, subject to multiplicative information on the set $A$, and can be found in \cite{MishaSophie}. We give the proof for completeness, noting that the proof for Lemma \ref{lem: energy convex} follows from a similar argument.
\begin{lemma}\label{lem:estimatinge3}
Let $A, B, C,\, Q,R\subset \mathbb{R}$ be finite sets with the property that $r_{QR}(a)\geq T$ for all $a\in A$ and some $T\geq 1$.
Then if $|R||C|\ll \left(|Q||B|\right)^2,$
\begin{equation} \label{e:bdone}
|\{c=a-b;\,a\in A, b\in B,\,c\in C\}| \ll \frac{(|Q||R||B||C|)^{2/3}}{T}\,.
\end{equation}
Furthermore, if $|R||A|\leq |Q|^2|B|,$
\begin{equation}
\label{e:e3(a,b)}
\mathsf{E}_3 (A,B) \ll \frac{|Q|^2 |R|^2|B|^2}{T^3}\log|A|\,.
\end{equation}
\end{lemma}
The next lemma bounds the additive energy of two sets $f(A)$ and $B$, where $f$ is a convex function, subject to additive information on the set $A$.
\begin{lemma} \label{lem: energy convex}
Let $A \subset \mathbb{R}$ be finite, and let $f$ be a convex (or concave) function. Suppose that there exist finite sets $Q, R\subseteq \mathbb{R}$ with $|Q|\geq |R|$ and a number $T\geq 1$ so that $r_{Q-R}(a)\geq T$ for all $a\in A$. Then for any set $B$ satisfying $|R||A|\ll |Q|^2|B|$, we have
\[
\mathsf{E}_3^+(f(A),B) \ll \frac{|Q|^2|R|^2 |B|^2}{T^3}\log|A|\,.
\]
\end{lemma}
We remark that we have stated Lemmas~\ref{lem:estimatinge3} and \ref{lem: energy convex} as a \emph{third} energy bound. The same technique with an additional interpolation argument gives us $k$th moment energy bounds, see e.g. \cite{MishaSophie} for details.
\begin{proof}
To prove the first bound, we note that by utilising the information on the sets $Q$ and $R$, we have
$$|\{c = a-b: a \in A, b \in B, c \in C \} | \leq \frac{1}{T}|\{c = qr-b: q \in Q, r \in R, b \in B, c \in C \} |,$$
which can be viewed as incidences between the set of lines $L$ given by $y = qx - c$ for $(q,c) \in Q \times C$, and the point set $P = R \times B$. Applying the Szemer\'{e}di - Trotter theorem, we have
$$|\{c = qr-b: q \in Q, r \in R, b \in B, c \in C \} | = I(P, L) \ll \left(|Q||R||B||C| \right)^{2/3} + |Q||B|.$$
Because of the constraint present in the statement of the lemma, the leading term dominates. We therefore have
$$|\{c = a-b: a \in A, b \in B, c \in C \} | \ll \frac{\left(|Q||R||B||C| \right)^{2/3}}{T}$$
as needed.
For the second part of the lemma, we decompose the support of $\mathsf{E}_3^+(A,B)$ into dyadic groups: for $i= 0, \dots \lfloor \log|A|\rfloor$, let $D_i:=\{d\in A-B: r_{A-B}(d)\in [2^i,2^i)\}\subseteq A-B$.
Then
\[
\mathsf{E}_3^+(A,B) = \sum_{i =0}^{\lfloor|A|\rfloor} \sum_{d\in D_i} r_{A-B}^3 (d) < \sum_i |D_i|2^{3i+3} \ll \log|A|\max_i |D_i|2^{3i}\,.
\]
With $D_i$ playing the role of $C$ in \eqref{e:bdone}, we have
\[
2^i |D_i|\leq |\{(a,b,d)\in A\times B \times D_i: d = a-b \}| \ll \frac{(|Q||R||B||D_i|)^{2/3}}{T}\,.
\]
The result then follows, and all that is left to do is to verify the condition required, for $C = D_i$, i.e. that $|Q||D_i| \ll \left(|R||B|\right)^2$. Note that since $D_i \subseteq A - B$, this is certainly true if we have
$$|Q||A||B| \ll \left(|R||B|\right)^2 \iff |Q||A| \ll |R|^2|B|$$
which is the stated condition.
\end{proof}
\subsection{Regularisation Results}
In this section we give some regularisation results required for the proof. The first is a lemma present in \cite{MishaSophie}. This lemma will be used to give a certain subset of $A$ on which much of the energy is supported, and with certain popularity properties.
\begin{lemma}\label{lem:regu}
Let ${\mathcal{R}_\epsilon}$ be a deterministic rule with parameter $\epsilon \in (0,1)$ that, to every sufficiently large finite additive set $X$, associates a subset ${\mathcal{R}_\epsilon}(X)\subseteq X$ of cardinality $|{\mathcal{R}_\epsilon}(X)|\geq (1-\epsilon)|X|$.
For any such rule ${\mathcal{R}_\epsilon}$, any $m>1$ and a sufficiently large finite set $A$, set $\epsilon = c_1\log^{-1}(|A|)$ for some $c_1\in (0,1)$.
Then there exists a set $B\subseteq A$ (depending on ${\mathcal{R}_\epsilon}, \,m$), with $|B|\geq (1-c_1) |A|$ such that
\[\mathsf{E}_m^+(\mathcal{R}_\epsilon(B))\geq c_2\,\mathsf{E}_m^+(B)\,,\]
for some constant $c_2=c_2(m, c_1)$ in $(0,1]$. \end{lemma}
We also require the following proposition. It is very similar to an expository lemma of Xue \cite[Lemma~5.1]{xue}, but has been amended to admit an asymmetric form. We present the rather technical proof of this proposition in the appendix, where we make the dependence on $\log(|A|)$ and $k$ hidden in the notation explicit.
\begin{proposition}\label{p:decomp}
Let $A, V$ be finite subsets of $\mathbb R$, let $k> 1$ be a real number and fix $c_1 \in (0,1)$.
Then there are sets $B, C$ with $C \subseteq B\subseteq A$ and $|C|\gtrsim_{k,c_1} |B|\geq (1-c_1) |A|$
such that the following property holds: there is a number $1\leq t\leq |B|$ and a set $D_t=\{x\in B-V:\, t\leq r_{B-V}(x)<2t\}$ such that
\[
\mathsf{E}_k^+(B,V) \sim_k |D_t|t^
\]
and
\[
r_{D_t+V}(c) \sim_k \frac{|D_t|t}{|V|
\]
for any $c\in C$.
\end{proposition}
On a high level, the proofs of Lemma~\ref{lem:regu} and Proposition~\ref{p:decomp} follow the same schemata: given a set $A$, we define a rule which extracts a positive proportion subset $A'\subseteq A$ with desirable properties according to the rule in question. In Lemma~\ref{lem:regu}, this rule is abstract, whereas in Proposition~\ref{p:decomp} it is explicit. We then iterate this procedure until some stopping condition is satisfied. In Lemma~\ref{lem:regu}, this stopping condition is relative to the $m$th energy; in Proposition~\ref{p:decomp}, the stopping condition is defined with respect to the \emph{support} of the $k$th energy. These two regularisation results differ primarily because of this subtlety. Finally, we argue that this procedure must terminate in an acceptable number of steps, thus eventually outputting a positive proportion subset $B\subseteq A$.
Proposition \ref{p:decomp} admits the following corollary, which is similar to a result of Shakan \cite[Theorem 1.10]{shakan}.
\begin{corollary}\label{c:E_3(B,V)E_3(f(C),U)}
Let $A, V\subseteq \mathbb{R}$ be finite, and $f$ be a convex (or concave) function. Then there are sets $B,C$ with $C\subseteq B\subseteq A$ and $|C|\gtrsim_k|B|\gg_k |A|$ such that
\[
\mathsf{E}_3^+(B,V)\mathsf{E}_3^+(f(C),U)\lesssim |U|^2|V|^2|A|^3
\]
for any set $U$ with $|U||V|\gg |A|$.
\end{corollary}
\begin{proof}
We apply Proposition~\ref{p:decomp} with $k = 3$ to obtain the sets $B$ and $C$, so that $\mathsf{E}_3^+(B,V)\sim |D_t|t^k$ where $r_{B-V}(d)\in [t,2t)$ for all $d\in D_t$ and $r_{D_t + V}(c)\gg |D_t|t|B|^{-1}$ for all $c\in C$.
We are able to obtain a bound on $\mathsf{E}_3^+(f(C),U)$ using Lemma~\ref{lem: energy convex}:
\[
\mathsf{E}_3^+(f(C),U) \lesssim \frac{|D_t|^2|V|^2|U|^2}{|D_t|^3t^3|B|^{-3}} \sim \frac{|V|^2|U|^2|A|^3}{\mathsf{E}_3^+(B,V)}\,.
\]
We remark that since $|U||V|\gtrsim |A|$, it follows that $\min(|D_t|,|V|)|C|\lesssim \max(|D_t|,|V|)^2|U|$ and so we may indeed apply Lemma~\ref{lem: energy convex}.
\end{proof}
\section{Auxiliary energy bounds}
A unifying idea behind the proofs in this paper is the following proposition:
\begin{proposition}\label{prop: energy general}
Let $A,C\subseteq \mathbb{R}$ be finite, and $k\geq 1$. Suppose that $\mathsf{E}_k^+(A) \sim |D|\Delta^k$ for some $D\subseteq A-A$ and $\Delta \geq 1$, where $r_{A-A}(d) \in [\Delta,2\Delta)$. Then we have
\begin{equation}
\label{e:prop eq}
|D|^9 \Delta^{12} \lesssim \frac{ |A + C|^6 \mathsf{E}_3^+(A)^4\mathsf{E}_3^+(C)^2\mathsf{E}_3^+(A,D) \mathsf{E}_3^+(C,A + C)^2}{|C|^{18}|A|^3}\,.
\end{equation}
\end{proposition}
The stated form of Proposition~\ref{prop: energy general} gives us a great deal of flexibility. For example, if we had multiplicative information on the set $A$ in the guise of Lemma~\ref{lem: energy convex} -- that is, if $r_{QR}(a)\geq T$ for all $a\in A$ -- then we obtain an energy estimate in terms of this data.
Proposition~\ref{prop: energy general} also admits a multiplicative form, in which $\mathsf{E}^\times_k(A) \sim |D|\Delta^k$. Then all instances of $\mathsf{E}_3$ in \eqref{e:prop eq} should be replaced by $\mathsf{E}_3^\times$, and $A+C$ by $AC$.
\begin{proof}
We begin by defining the popular set
\[P(A,C):= \left\{ x \in A+C : r_{A+C}(x) \geq \frac{|A||C|}{\log|A||A+C|} \right\}.\]
We also define the set
\[A' :=\left\{ a\in A : |\{ c \in C : a+c \in P(A,C) \}| \geq \frac{|C|}{2} \right\}.\]
We perform a refinement step at the beginning of the proof, making use of Lemma \ref{lem:regu}. We claim that Lemma \ref{lem:regu} can be applied with the deterministic rule being the subset $A' \subseteq A$ defined above. Firstly we prove that $|A'|$ is large with respect to $|A|$. We have
\begin{align*}\sum_{a \in A'}|\{ c \in C : a+c \in P(A,C) \}| + \sum_{a \in A \setminus A'}|\{ c \in C : a + c \in P(A,C) \}| &= |\{ (a,c) \in A\times C : a+c \in P(A,C) \}| \\
& \geq \left( 1 - \frac{1}{\log|A|}\right) |A||C|.
\end{align*}
By setting $|A \setminus A'| = c|A|$ and using the bounds
$$\sum_{a \in A'}|\{ c \in C : a + c \in P(A,C) \}| \leq (1-c)|A||C|$$
$$ \sum_{a \in A \setminus A'}|\{ c \in C : a + c \in P(A,C) \}| \leq \frac{c}{2}|A||C|$$
we conclude that $|A'| \geq \left( 1 - \frac{2}{\log|A|}\right) |C|$. We can therefore apply Lemma \ref{lem:regu} at the outset of the proof, obtaining a set $A'$ as above with the property that $|A'| \gtrsim |A|$, and $\mathsf{E}_k^+(A') \sim \mathsf{E}_k^+(A)$.
We now consider the number of solutions $(a,b,c) \in A^2 \times C$ to the trivial equation
\begin{equation}\label{trivialeq} a-b = (a+c) - (b+c)\end{equation}
where the difference $a-b$ comes from the set $D \subseteq A'-A'$ such that $|D|\Delta^k \sim\mathsf{E}_k^+(A') \sim\mathsf{E}_k^+(A)$, and such that the sum $a + c$ is popular, that is, $a+c \in P(A,C)$.
There are at least $\Omega( |C||D|\Delta)$ solutions to equation \eqref{trivialeq}. We partition solutions to \eqref{trivialeq} with the relevant conditions, via the following:
$$(a,b,c) \sim (a+ t,b+ t,c-t), \quad t \in \mathbb R\,,$$
and let $[a,b,c]$ represent this equivalence class. Since $t$ cancels out in equation \eqref{trivialeq}, these classes are non-trivial.
Let $N$ denote the number of solutions to equation \eqref{trivialeq}.
We have
$$|C||D|\Delta \ll N = \sum_{[a,b,c]}|[a,b,c]|$$
and so, after an application of the Cauchy-Schwarz inequality, we obtain
\begin{equation}\label{eq:twofactors} \left(|C||D|\Delta\right)^2 \leq |\{ \text{equivalence classes} \}| \cdot \sum_{[a,b,c]}|[a,b,c]|^2.\end{equation}
We now aim to bound the two factors in equation \eqref{eq:twofactors}.
To bound the number of equivalence classes, note that each equivalence class gives a solution to the equation
$$d = s_1 - s_2, \quad d \in D, s_1 \in P(A,C), s_2 \in A+B.$$
Therefore we have
$$|\{ \text{equivalence classes} \}| \leq |\{ (d,s_1,s_2) \in D \times P(A,C) \times A + C : d = s_1-s_2 \}|.$$
By the popularity of $s_1$, we have
\begin{align*}|\{ \text{equivalence classes} \}| & \lesssim \frac{|A+C|}{|A||C|}|\{ (d,a,c,s) \in D \times A \times C \times A + C : d - a = c - s \}| \\
& = \frac{|A + C|}{|A||C|} \sum_{x} r_{D - A}(x) r_{ C - (A+C)}(x)
\\
& \leq \frac{|A + C|}{|A||C|}\mathsf{E}_3^+(A,D)^\frac16 (|A||D|)^\frac12\mathsf{E}_3^+(C,A+C)^\frac13\,,
\end{align*}
where the final bound is an application of H\"older's inequality, followed by Cauchy-Schwarz.
We now aim to bound the sum
$$ \sum_{[a,b,c]}|[a,b,c]|^2$$
where is it understood that the sum is taken over equivalence classes satisfying the relevant conditions. Note that this sum counts pairs of triples from the same equivalence class, and for each pair we have
$$(a,b,c) \sim (a',b',c') \implies \exists
~t \text{ \ with \ } a-a'= b- b'= c'- c = t.$$
We therefore have
$$ \sum_{[a,b,c]}|[a,b,c]|^2 \leq \sum_t r_{A-A}(t)^2r_{C-C}(t) \leq \mathsf{E}_3^+(A)^\frac23\mathsf{E}_3^+(C)^\frac13 $$
where again, the final inequality is a result of Hölder's estimate.
Finally, from \eqref{eq:twofactors} we have
\[\left(|C||D|\Delta\right)^2 \lesssim
\frac{|A + C|}{|A||C|}\mathsf{E}_3^+(A,D)^\frac16 (|A||D|)^\frac12\mathsf{E}_3^+(C,A + C)^\frac13
\mathsf{E}_3^+(A)^\frac23\mathsf{E}_3^+(C)^\frac13
\,.
\]
Rearranging and raising both sides to the sixth power concludes the proof.
\end{proof}
\section{Proof of Theorem~\ref{thm:main}}
We actually prove the following, slightly more general theorem.
\begin{theorem}\label{t:|A+B||f(A)+g(B)|etc}
Let $f$ and $g$ be convex or concave functions. Let $A, B \subseteq \mathbb{R}$ be finite sets. Then we have
\begin{equation}\label{e:|A+B||f(A)+g(B)|}
|A|^{49}|B|^{49}\lesssim
|A+B|^{38} |f(A)+g(B)|^{38}
\end{equation}
and
\begin{equation}\label{e:|A+B||f(A)+g(B)||A-A|etc}
|A|^{39}|B|^{39}\lesssim |A\pm B|^{20}|f(A)\pm g(B)|^{20}|A-A|^5||B-B|^5|f(A)-f(A)|^5|g(B)-g(B)|^5\,.
\end{equation}
\end{theorem}
We clarify that in this theorem, one may take $f$ to be convex, and $g$ to be concave.
On a high level, the proof proceeds by apply two iterations of Corollary~\ref{c:E_3(B,V)E_3(f(C),U)} to $A$ and $B$ with judicious choices of $V$ in each case. Then we apply Proposition~\ref{prop: energy general} to the ensuing subsets and their convex (resp. concave) counterparts. This gives an additive energy relation.
We obtain the statements of Theorem~\ref{t:|A+B||f(A)+g(B)|etc} using Cauchy-Schwarz and H\"older inequalities.
Let us make two simple observations regarding the third moment energy of a set $X$ that we use in the subsequent argument
Firstly, note that $\mathsf{E}_3^+(-X,X-X) = \mathsf{E}_3^+(X,X-X)$ due to the symmetry of the difference set. Secondly, for any set $Y\subseteq X$ and any set $Z$, we have $\mathsf{E}_3^+(Y,Z) \leq \mathsf{E}_3^+(X,Z)$.
\begin{proof}
Here we prove the slightly more technical statement~\eqref{e:|A+B||f(A)+g(B)|}, and indicate the changes necessary to prove \eqref{e:|A+B||f(A)+g(B)||A-A|etc}.
We begin by applying Corollary~\ref{c:E_3(B,V)E_3(f(C),U)} to the set $A$ with $V = A$ to obtain sets
$A_2\subseteq A_1 \subseteq A$ with $|A_2|\gtrsim |A_1|\gg |A|$
and
\[E_3(A_1,A)E_3(f(A_2),U) \lesssim |A|^5 |U|^2 \text{ for any }U\,.\]
Note that if $|U|\gg 1$, then this follows from Corollary \ref{c:E_3(B,V)E_3(f(C),U)}; if $|U|\ll 1$, then it follows trivially.
We now apply the concave analogue of Corollary~\ref{c:E_3(B,V)E_3(f(C),U)}, this time to the set $f(A_2)$ with $V =f(A)$ and the function $f^{-1}$. We obtain \footnote{Strictly speaking, we obtain sets $f(A_4)\subseteq f(A_3)\subseteq f(A_2)$.} the sets $A_4\subseteq A_3\subseteq A_2$ with $|A_4|\gtrsim |A_3|\gg |A_2|\gtrsim |A|$
so that
\[E_3(f(A_3), f(A)) E_3(A_4,U) \lesssim |A|^5 |U|^2 \text{ for any }U\,.
\]
We repeat this argument for the set $B$ taking $V = B$ to obtain $B_2\subseteq B_1\subseteq B$
so that
\[E_3(B_1,B)E_3(g(B_2),U) \lesssim |B|^5 |U|^2 \text{ for any }U\,,\]
and then once more to $g(B_2)$ with $V = g(B)$ and function $g^{-1}$ to obtain
$B_4\subseteq B_3\subseteq B_2$ with $|B_4|\gtrsim |B|$ and
\[E_3(g(B_3), g(B)) E_3(B_4,U) \lesssim |B|^5 |U|^2 \text{ for any }U\,.\]
To prove \eqref{e:|A+B||f(A)+g(B)|}, we dyadically decompose the sets $A_4,B_4,f(A_4),g(B_4)$ according to the second moment energy to obtain sets $D_i$ and numbers $t_i\geq 1$ so that
\[
\mathsf{E}_2^+(A_4) \sim |D_1|t_1^2, \quad \mathsf{E}_2^+(B_4) \sim |D_2|t_2^2\]
\[\mathsf{E}_2^+(f(A_4)) \sim |D_3|t_3^2, \quad \mathsf{E}_2^+(g(B_4)) \sim |D_4|t_4^2\,.\]
To prove \eqref{e:|A+B||f(A)+g(B)||A-A|etc}, we would instead dyadically decompose the sets $A_4, B_4, f(A_4), g(B_4)$ according to the $12/7$th moment energy, so that e.g. $\mathsf{E}^+_{12/7}(A_4) \sim |D_1|t_1^{12/7}$. Note that e.g. $D_1 \subseteq A_4 - A_4$.
We now apply Proposition~\ref{prop: energy general} to each of the sets $A_4,B_4,f(A_4),g(B_4)$, choosing $C$ in \eqref{e:prop eq} to be $B_4,A_4,g(B_4),f(A_4)$ respectively.
We then multiply together the four instances of \eqref{e:prop eq}, and make liberal use of the simple observations noted at the beginning of this section together with the consequences of Corollary~\ref{c:E_3(B,V)E_3(f(C),U)}. This gives
\begin{equation}\label{e:product 12/7 energy}
\prod_{1\leq i \leq 4}|D_i|^7t_i^{12} \lesssim|A+B|^{20} |f(A)+g(B)|^{20}|A|^9|B|^9\,.
\end{equation}
To prove statement \eqref{e:|A+B||f(A)+g(B)||A-A|etc}, we recall that we had initially dyadically decomposed according to the $12/7$th energy and so, after an application of H\"older's inequality for $\mathsf{E}^+_{12/7}(A_4)$ etc., we are done.
To prove statement \eqref{e:|A+B||f(A)+g(B)|}, let us multiply \eqref{e:product 12/7 energy} on both sides by $(t_1t_2t_3t_4)^2$.
Note that
\[
|D_1|t_1^3|D_3|t_3^3 \leq \mathsf{E}_3^+(A_4,A)\mathsf{E}_3^+(f(A_3),f(A)) \lesssim |A|^7 \implies t_1t_2\lesssim \frac{|A|^7}{\mathsf{E}_2^+(A)\mathsf{E}_2^+(f(A))}
\]
and similarly
\[
(t_2t_4) \lesssim \frac{|B|^7}{\mathsf{E}_2^+(B)\mathsf{E}_2^+(g(B))}
\]\,.
Hence we obtain
\[
\left(\mathsf{E}_2^+(B)\mathsf{E}_2^+(g(B))\mathsf{E}_2^+(A)\mathsf{E}_2^+(f(A))\right)^9 \lesssim|A+B|^{20} |f(A)+g(B)|^{20}|A|^{23}|B|^{23}\,.
\]
Finally, we use the Cauchy-Schwarz relation
\[
\frac{|X|^2|Y|^2}{|X+Y|}\leq \mathsf{E}_2^+(X,Y) \leq \mathsf{E}_2^+(X)^{1/2}\mathsf{E}_2^+(Y)^{1/2}
\]
to complete the proof.
\end{proof}
\section{Proof of Theorem \ref{thm: AB+A}}
In this section we prove Theorem \ref{thm: AB+A}
proving two complementary bounds, using a combination of the methods found in \cite{rss}, \cite{MishaSophie}, and \cite{ollyaudie}.
\subsection{Proof of Theorem~\ref{thm: AB+A} - Bound 1}
The method of Roche-Newton and Warren \cite{ollyaudie} involved studying the \emph{line energy} of lines of a particular structure. Their results, combined with an incidence theorem of Rudnev and Shkredov \cite{rudnevshkredov} and an additive combinatorial result of Roche-Newton and Rudnev\footnote{The result of Roche-Newton and Rudnev is that the number of solutions to the equation
\[
\frac{a_1 - a_2}{a_3 - a_4} = \frac{a_5 - a_6}{a_7- a_8}
\]
with each $a_i \in A$ is at most $O(|A|^6\log|A|)$.} \cite{Q}
imply the following incidence bound. See also \cite{affinegroup4} for more information on line energy and its applications.
\begin{theorem}
\label{thm:incidence}
Let $L$ be a set of lines of the form $y = ax + a'$ for $a,a' \in A \subseteq \mathbb R\setminus\{0\}$ a finite set.
Let $B,C \subseteq \mathbb R$ be two finite sets. Then we have
\[\mathcal{I}(B \times C, L) \lesssim \mathsf{E}_4^\times(A)^\frac1{12}|A|^\frac76
|B|^\frac23|C|^\frac12
+ |A|^2 |C|^\frac12.
\]
\end{theorem}
We shall apply Theorem \ref{thm:incidence} to the point set $B \times (AB+A)$ and to the set of lines $L$ of the form $y = ax + a'$ with $a,a' \in A$. Note that without loss of generality we may remove $0$ from $A$ if it is present. For each line $y = ax + a'$, for each $b \in B$ the point $(b,ab+a')$ lies on this line, and so we have at least $|A|^2|B| \sim |A|^3$ incidences. Using Theorem \ref{thm:incidence} we obtain
\begin{equation}\label{Incidence1}
|A|^3 \lesssim\mathcal I(B \times (AB+A),L) \lesssim
\mathsf{E}_4^\times(A)^\frac1{12}|A|^\frac76
|B|^\frac23|AB+A|^\frac12
+ |A|^2 |AB+A|^\frac12.
\end{equation}
Note that if the second term dominates we have a much stronger result than claimed in the statement in Theorem~\ref{thm: AB+A}. Let us therefore assume the first term dominates. Hence we have the first of our two bounds:
\begin{equation}\label{e:bound1}
|A|^{7/3}\lesssim |AB+A|\mathsf{E}_4^\times(A)^{1/6} \leq |AB+A||A|^{\frac{1}{6}} \mathsf{E}_3^\times(A)^{1/6}.
\end{equation}
\subsection{Proof of Theorem~\ref{thm: AB+A} - Bound 2}
To find the second bound, let us apply the multiplicative version of Proposition~\ref{p:decomp} to the set $A$ with $V = A$ to obtain sets $A_2\subseteq A_1\subseteq A$ with $|A_2|\gtrsim |A_1|\gg |A|$ so that
\begin{equation}\label{e:mult version of cor}
\mathsf{E}_3^\times(A_1,A)\mathsf{E}_3^+(A_2,U)\lesssim |A|^5|U|^2\,.
\end{equation}
Equation~\eqref{e:mult version of cor} is a consequence of using Lemma~\ref{lem:estimatinge3} in place of Lemma~\ref{lem: energy convex} in the proof of Corollary~\ref{c:E_3(B,V)E_3(f(C),U)}.
We now apply Proposition~\ref{prop: energy general} to the set $A_2$, writing $\mathsf{E}^+(A_2)\sim |D|t^2$ and taking $C = \lambda A_2$ for $\lambda \neq 0$. Note that $\mathsf{E}_k^+(A_2,X) = \mathsf{E}_k^+(\lambda A_2, \lambda^{-1} X)$ for any set $X$ and any $k\geq 1$.
From Proposition~\ref{prop: energy general}, \eqref{e:mult version of cor}, and the inclusions $A_2\subseteq A_1\subseteq A$ we obtain
\begin{equation}\label{e: bound2 penultimate}
|D|^7 t^{12} \lesssim \frac{|A+\lambda A|^{10}}{|A|^9} \left(\frac{\mathsf{E}_3^+(A_2)}{|A|^2}\right)^{12} \frac{\mathsf{E}_3^+(A_2,D)}{|D|^2} \left(\frac{\mathsf{E}_3^+(A_2,\lambda^{-1}A + A}{|A+\lambda A|^2}\right)^2 \lesssim \frac{|A+\lambda A|^{10}}{|A|^9} \frac{|A|^{45}}{\mathsf{E}_3^\times(A_1,A)^9}\,.
\end{equation}
We have
\[|D|t^3 \leq \mathsf{E}_3^+(A_2) \lesssim \frac{|A|^7}{\mathsf{E}_3^\times(A_1,A)} \implies t \lesssim \frac{|A|^7}{\mathsf{E}_3^\times(A_1,A) \mathsf{E}_2^+(A_2)}\,,
\]
and so multiplying \eqref{e: bound2 penultimate} by $t^2$ and applying the Cauchy-Schwarz energy bound
\[\frac{|A_2|^4}{|A_2 + \lambda A_2|} \leq \mathsf{E}(A_2,\lambda A_2) \leq \mathsf{E}(A_2)\]
we conclude that
\begin{equation}
\label{e:bound 2 final}
|A+\lambda A|^{19} \gtrsim \frac{\mathsf{E}_3^\times(A_1,A)^{11}}{|A|^{14}}\,.
\end{equation}
\subsection{Proof of Theorem~\ref{thm: AB+A} - Conclusion }
Combining the bounds of the previous section we obtain
\[
|A+BA|\gtrsim \max \left(
\frac{\mathsf{E}_3^\times(A,A_1)^\frac{11}{19}}{|A|^{\frac{14}{19}}}
, \frac{|A|^\frac{13}{6}}{ \mathsf{E}_3^\times(A,A_1)^\frac16}\right)
\]
where the first bound has instead been applied to the set $A_1$
given above, making use of the inequalities $|A| \sim |A_1|$ and $\mathsf{E}_3^+(A_1) \leq \mathsf{E}_3^+(A,A_1)$. In the worst possible case, both maximands are equal. This happens if
\[ \mathsf{E}_3^+(A_1,A) = |A|^\frac{331}{85}\]
and so we shall assume that this is indeed the case. We then obtain
\[ |A+BA|\gtrsim |A|^\frac{129}{85} = |A|^{\frac32 +\frac3{170}}\]
as required.
\section*{Acknowledgements}
The authors were supported by Austrian Science Fund FWF Project P 30405-N32. We are especially grateful to Oliver Roche-Newton who pushed us to improve our results. We also thank Misha Rudnev for helpful suggestions.
|
\section{Introduction}
The $D^+_s$-meson, which is composed of a charm quark and a strange antiquark, has been discovered in year 1993. There are rich physics contents in $D^+_s$-meson decays. The $D_s^+$-meson semileptonic or exclusive decay processes provide important heavy-to-light theoretical bases for studying heavy quark decays, investigating light meson spectroscopy and supplying a bridge between weak and strong interaction couplings of quarks. More and more experimental results have been reported from the BABAR, the CLEO, the BESIII collaborations, and etc., such as the $D_s^+ \to (\eta, \eta', K^0, a_0(980), f_0(980), \phi, K^{*0})\ell^+\nu_\ell$ decays' branching fractions are within the range of $[0.12, 2.61]\%$~\cite{BaBar:2008gpr, CLEO:2009dyb, CLEO:2009ugx, BESIII:2017ikf, BESIII:2018xre, BESIII:2021duu}. Total semileptonic branching fractions provide useful discrimination on the different theoretical evaluations of hadronic matrix elements, which sizably affect the charm quark semileptonic decays. The $D_s \to \eta^{(\prime)}\ell^+\nu_\ell$ is different from the usually considered channels with final state containing light-quark composition only, and it has attracted much attention from both theoretical and experimental groups. Moreover, the $\eta^{(\prime)}$-mesons, composed by $s\bar s$ quark pair, are especially intriguing, since the $s$-quark plays an important role for the flavor physics. In deep leaning of those two processes, one can obtain useful information on the CKM matrix element $|V_{cs}|$ and the heavy-to-light transition form factors (TFFs).
Experimentally, the semileptonic decay processes for $D_s^+ \to \eta^{(\prime)} \ell^+\nu_\ell$ have been found by the CLEO collaboration early in year 1995, and their measured value of the ratio of branching fractions ${\cal B}(D_s^+\to \eta'e^+ \nu_e)/{\cal B}(D_s^+\to \eta e^+\nu_e)$ is $0.35\pm0.09\pm0.07$~\cite{Brandenburg:1995qq}. Then, the CLEO collaboration issued the measured value of the branching fractions in years 2009 and 2015, i.e. ${\cal B}(D_s^+\to \eta e^+ \nu_e)=(2.48\pm0.29\pm0.13)\%$, ${\cal B}(D_s^+\to \eta e^+ \nu_e )=(2.28 \pm0.14 \pm0.20)\%$, ${\cal B}(D_s^+\to \eta' e^+ \nu_e) = (0.91 \pm0.33 \pm0.05)\%$, and ${\cal B}(D_s^+\to \eta' e^+ \nu_e )=(0.68\pm0.15\pm0.06)\%$~\cite{Yelton:2009aa, Hietala:2015jqa}. In year 2017, the BESIII collaboration measured the branching fractions by using the same channels based on the integrated luminosity of $482~{\rm pb^{-1}}$ of the $e^+e^-$ collision at the center-of-mass energy $\sqrt s=4.009~{\rm GeV}$, and they issued ${\cal B} (D_s^+ \to \eta \mu^+\nu_\mu)=(2.42\pm 0.46\pm 0.11)\%$, ${\cal B} (D_s^+ \to \eta' \mu^+\nu_\mu)=(1.06\pm 0.54\pm 0.07)\%$, ${\cal B}(D_s^+\to \eta e^+\nu_e)=(2.30\pm 0.31\pm 0.08)\%$, and ${\cal B}(D_s^+\to \eta' e^+\nu_e)=(0.93\pm 0.30\pm 0.05)\%$~\cite{Ablikim:2017omq, Ablikim:2016rqq}. In year 2019, the BESIII collaboration finished the improved measurements on the branching fractions by using $e^+e^-$ annihilation data corresponding to an integrated luminosity of $3.19~{\rm fb}^{-1}$ collected at a center-of-mass energy of $4.178~{\rm GeV}$ and then gave the first determination of $D_s\to \eta^{(\prime)}$ TFF, e.g. ${\cal B}(D_s^+\to \eta e^+\nu_e)=(2.323\pm 0.063\pm 0.063)\%$, ${\cal B}(D_s^+\to \eta' e^+\nu_e)=(0.824\pm 0.073\pm 0.027)\%$, $f^{\eta}_+(0) |V_{cs}| = 0.4455\pm0.0053\pm0.0044$, and $f^{\eta'}_+(0) |V_{cs}| = 0.477\pm0.049\pm0.011$ ~\cite{Ablikim:2019rjz}. There are large discrepancies for $D_s^+\to \eta^{(\prime)} \ell^+\nu_\ell$ among different experimental collaborations. With more and more data accumulated in the near future, the experimental precision shall be greatly improved and the gap among different measurements could be shrunk.
Theoretically, the decay widths or branching fractions for the semileptonic decay $D_s^+\to\eta^{(\prime)}\ell^+\nu_\ell$ depends heavily on the precision of the $D_s\to\eta^{(\prime)}$ TFFs. At present, the $D_s\to\eta^{(\prime)}$ TFFs have been studies under various approaches, such as the lattice QCD (LQCD)~\cite{Bali:2014pva}, the traditional and covariant light-front quark model (LFQM)~\cite{Verma:2011yw, Cheng:2017pcq, Wei:2009nc}, the constituent quark model (CQM)~\cite{Melikhov:2000yu}, the covariant confined quark model (CCQM)~\cite{Soni:2018adu, Ivanov:2019nqd}, the light-cone sum rules (LCSR)~\cite{Duplancic:2015zna, Offen:2013nma}, the QCD sum rules (QCD SR)~\cite{Colangelo:2001cv}. The LCSR approach is based on the operator product expansion (OPE) near the light-cone $x^2\rightsquigarrow 0$ and parameterizes all the non-perturbative dynamics into the light-cone distribution amplitudes (LCDAs), which have been applied for dealing with many semileptonic decay processes~\cite{Duplancic:2008ix, Momeni:2019uag, Descotes-Genon:2019bud, Cheng:2019tgh, Momeni:2020zrb, Emmerich:2018rug, Momeni:2018tjf, Shen:2016hyv, Straub:2015ica, YanJun:2011rn, Wang:2010tz, Wang:2008bw, Beneke:2018wjp,Du:2003ja, Yang:2005bv, Singh:2018yvt, Fu:2018yin}. One may observe that the predicted values of $D_s\to\eta^{(\prime)}$ TFFs behave differently from various groups. Those discrepancies indicate that it is important to improve the accuracy of theoretical calculation. In this paper, we will calculate the $D_s\to\eta^{(\prime)}$ TFFs by using the LCSR approach up to next-to-leading order (NLO) QCD corrections.
It is well known that the states $\eta$ and $\eta^{\prime}$ are considered as candidates for mixing. The mixing among pseudoscalar mesons is of great theoretical interests and significance for understanding the dynamics and hadronic structures, which is caused by the QCD anomaly and related to the chiral symmetry breaking. Thus, one can gain a better insight into the dynamics if the mixing parameters are more accurately determined. The semileptonic decays $D_s^+ \to \eta^{(\prime)} \ell^+\nu_\ell$ probe the $s \bar s$ components of $\eta$ and $\eta^{\prime}$, which can well separate the strong and weak effects in theory, are expected to be sensitive to $\eta-\eta^{\prime}$ mixing angle~\cite{Anisovich:1997dz}. Many measurements on the processes, where $\eta$ and $\eta'$ are involved, have been carried out to fix the mixing parameters. The $\eta-\eta'$ mixing can be described in different forms. To investigate the $\eta-\eta'$ mixing, two schemes have been suggested in the literature~\cite{Ball:2007hb, Ball:1995zv, Feldmann:1999uf, Huang:2006as, Ke:2009mn, DeFazio:2000my, Choi:2010zb}, i.e. the singlet-octet (SO) mixing scheme and the quark-flavor (QF) mixing scheme. The SO mixing angle $\theta$ between $\eta$ and $\eta'$ is known to be in the range of $\theta\in[-10^\circ, -23^\circ]$. In this scheme, The $\eta$ and $\eta'$ are the mixtures of the flavor SU(3) octet $\eta_8$ and single $\eta_0$ states:
\begin{align}
\left( \begin{array}{l}
\eta \\
{\eta'}
\end{array} \right) = \left( \begin{array}{l}
\cos \theta ~~~ - \sin \theta \\
\sin \theta ~~~~~ \cos \theta
\end{array} \right)\left( \begin{array}{l}
{\eta _8}\\
{\eta _0}
\end{array} \right),
\end{align}
where $\eta_8=(u\bar u+d\bar d-2s\bar s)/\sqrt6$ and $\eta_0=(u\bar u+d\bar d+s\bar s)/\sqrt3$. Analogously, information could be gathered on the mixing scheme in the QF basis, which is consists with the form $\eta$ and $\eta'$ states as combinations of $|\eta_q\rangle=|\bar uu + \bar dd\rangle/\sqrt 2$ and $|\eta_s\rangle=|\bar ss\rangle$:
\begin{align}
&|\eta \rangle = \cos \varphi |{\eta _q}\rangle - \sin \varphi |{\eta _s}\rangle,
\nonumber\\
&|\eta '\rangle = \sin \varphi |{\eta _q}\rangle + \cos \varphi |{\eta _s}\rangle.
\end{align}
It has been shown that in this scheme a single angle is essentially required. In year 2007, the KLOE Collaboration provides the value $\varphi =(41.5\pm0.3_{\rm stat}\pm 0.7_{\rm syst}\pm 0.6_{\rm th})^\circ $ by extracting the pseudoscalar mixing angle $\varphi$ in the QF basis by measuring the radio ${\cal B}(\phi\to\eta'\gamma) / {\cal B}(\phi \to \eta \gamma)$~\cite{Ambrosino:2006gk}. Some theoretical groups have calculated the single mixing angle $\varphi$~\cite{Ball:2007hb, Duplancic:2015zna, Colangelo:2001cv, Azizi:2010zj}, their predicted values are within the range of $\varphi\in[39^ \circ, 41.8^\circ]$. One can put forward the ratio ${\cal R}_{\eta'/\eta} = {\cal B}(D_s\to\eta '\ell^+\nu_\ell)/{\cal B} (D_s\to\eta\ell^+\nu_\ell)$ to access the $\eta-\eta'$ mixing angle through the ratio of the TFFs $f_+^\eta(q^2)/f_+^{\eta'}(q^2)$~\cite{Colangelo:2001cv}, which are related to the $\eta-\eta'$ mixing scheme. In particular, information could be gathered on the mixing scheme in the QF basis. In this paper, we will use the QF basis with the single mixing angle $\varphi$ to analyze the $D_s\to\eta^{(\prime)}$ decay modes, and the corresponding TFFs satisfy the relation
\begin{align}
\tan \varphi = \frac{|f_+^\eta (q^2)|}{|f_+^{\eta'}(q^2)|}. \label{bt}
\end{align}
One usually takes the large recoil point of the squared momentum transfer $q^2 = 0$ to do the calculation, i.e. $\tan \varphi = {|f_+^\eta (0)|}/{|f_+^{\eta'}(0)|}$. A more accurate $D_s^+\to\eta^{(\prime)}\ell^+\nu_\ell$ LCSR analysis is important.
The $\eta$ and $\eta'$ mesons full of rich phenomenology, which are predominantly flavor-singlet states. This means that their wave functions are approximately symmetric in the three lightest quark types (up, down, and strange), which build up the light-hadron spectroscopy. Thus, the LCDAs for $\eta^{(\prime)}$-meson, as one of the most important parameters, composed by $s\bar s$ are significant to the $D_s\to\eta^{(\prime)}$ TFFs, which can be expanded as a Gegenbauer polynomial series:
\begin{eqnarray}
\phi_{2;\eta^{(\prime)}}(u,\mu) = 6 u \bar u \left[1 + \sum_{n=1}^\infty a^n_{2;\eta^{(\prime)}}(\mu) C^{3/2}_n(\xi) \right], \label{HPDA_CZ}
\end{eqnarray}
where $a^n_{2;\eta^{(\prime)}}(\mu)$ stands for the $n{\rm th}$-order Gegenbauer moment, $\bar{u}=(1-u)$ and $\xi=(2u-1)$. When the factorization scale $\mu$ is large enough, the twist-2 LCDAs $\phi_{2;\eta^{(\prime)}}(u,\mu)$ tends to the asymptotic form $\phi_{2;\eta^{(\prime)}}(u,\infty) = 6u\bar u$. There are some theoretical and experimental predictions for the Gegenbauer moments $a^n_{2;\eta}(\mu)$, such as the fitting results coming from CLEO collaboration $a^2_{2;\eta}(\mu_0) = -0.07\pm0.03$~\cite{Gronberg:1997fj}, the fitting results from BABAR result $a^2_{2;\eta}(\mu_0) = -0.05\pm0.02$~\cite{BABAR:2011ad}, the results predicted by Kroll and Passek-Kumericki $a^2_{2;\eta}(\mu_0) = -0.05\pm0.02$~\cite{Kroll:2013iwa}, and the fitting results from a sum rule analysis $a_{2;\eta}^2 (\mu_0)= 0.25\pm0.15$~\cite{Offen:2013nma}. By taking the approximation with $\pi, K$-meson, Ball and Zwicky predicted $a^2_{2;\eta}(\mu_0) = 0.115$ and $a^4_{2;\eta}(\mu_0) = -0.015$~\cite{Ball:2004ye}. At present, few works have been done to calculate the second and higher order moment. Particularly, the twist-2 LCDA of $\eta'$-meson is rarely studied. So it is important and meaningful to make a more accurate calculation on the second order moment and the higher order moments for $\eta^{(\prime)}$-meson LCDAs within the QCD sum rule approach.
An effective way to calculate the $n$th-order moments of the $\eta^{(\prime)}$-meson LCDAs is to use the following definition,
\begin{eqnarray}
\langle\xi_{2;\eta^{(\prime)}}^n\rangle|_\mu = \int_0^1 du ~\xi^n~ \phi_{2;\eta^{(\prime)}}(u,\mu). \label{Eq:xin}
\end{eqnarray}
$\langle\xi_{2;\eta^{(\prime)}}^n\rangle|_\mu$ can be calculated by using the Shifman-Vainshtein-Zakharov (SVZ) sum rules~\cite{Shifman:1978by, Huang:1986wm, Govaerts:1983ka, Huang:1989gv, Novikov:1983gd, Hubschmid:1982pa}. In which, the perturbative QCD is established on the assumption that the perturbative vacuum and the short-distance interaction are not affected by the long-distance structure of the non-Abelian gauge field. The QCD physical vacuum contains a series of vacuum condensates, such as the quark condensate $\langle q\bar q\rangle$, the gluon condensate $\langle G^2 \rangle$, and etc.. These vacuum condensates reflect the non-perturbation characteristics of QCD. The QCD sum rules based on the background field theory (BFTSR) method gives a possible way to consider the non-perturbation effect, which also provides a systematic description of these vacuum condensates from the field theory point of view. At present, the BFTSR has been used in calculating the twist-2 or twist-3 LCDAs for $\pi$, $K$, $D$, $D_s$, $a_0$, $K_0^*$, $f_0$, $\rho$, $J/\psi$, $a_1(1260)$-mesons~\cite{Zhong:2014jla, Fu:2016yzx, Fu:2018vap, Zhong:2014fma, Zhong:2016kuv,Zhong:2011jf, Han:2013zg, Huang:2004tp, Huang:2005av, Zhong:2011rg, Zhang:2017rwz, Zhang:2021wnv,Hu:2021lkl}. Besides, other methods in studying the LCDAs can be found in Refs.~\cite{Cloet:2013tta, Polyakov:2020cnc, Cheng:2020vwr, Wang:2019msf, Zhang:2017bzy, Zuo:2011sk, Wu:2010zc, Khodjamirian:2006st, Zuo:2006re}. In this paper, we will calculate the $\eta^{(\prime)}$-meson twist-2 LCDAs within the BFTSR approach for the first time.
The rest of the paper are organized as follows. In Sec.~\ref{sec:2}, we present the basic idea of the QCD background field theory, the detailed BFTSR procedures for calculating the moments of $\phi_{2;\eta^{(\prime)}}(u,\mu)$, the branching fractions and the transition form factors involved in the semileptonic decay $D_s^+\to \eta^{(\prime)} \ell^+ \nu_\ell$. In Sec.~\ref{sec:3}, we present our numerical results and make a detailed comparison with other experimental and theoretical predictions. Section~\ref{sec:summary} is reserved for a summery. The intermediate processes for calculating the moments of $\eta^{(\prime)}$-meson LCDA within the BFTSR and the basic definition of $\eta^{(\prime)}$-meson twist-2, 3, 4 LCDAs are given in the Appendixes~\ref{sec:appendixA}, \ref{sec:appendixB} and \ref{sec:appendixC}.
\section{Calculation Technology}\label{sec:2}
\subsection{Basic idea and formulas for background field theory}
The important aspect for the SVZ sum rules approach is the OPE, which has been introduced by using the QCD physical vacuum $ \langle 0|{\cal O}_N|0\rangle_{\rm phys.}$. Its special property is the non-perturbation effect, which can be described by the classical background field satisfying the equations of motion. The main idea of the background field theory method is to describe the non-perturbation effect with the classical background field satisfying the equation of motion and to describe the quantum fluctuation, namely the perturbation effect, on this basis within the frame of the quantum field theory. More specifically, one can use the following substitution in the theoretical Lagrangian and Green's functions~\cite{Makeenko:1979pb, Shifman:1978zq, Huang:1986wm}
\begin{align}
&{\cal A}_\mu^a (x)\to {\cal A}_\mu^a (x) + \phi_\mu^a(x),
\nonumber\\
&\psi(x)\to\psi(x)+\eta(x), \label{AF}
\end{align}
where ${\cal A}_\mu^a (x)$ and $\psi(x)$ stand for the gluon and quark background fields, $\phi_\mu^a(x)$ and $\eta(x)$ represent their quantum fluctuations, respectively. In the presence of background fields, the quantization of $\phi_\mu^a(x)$ and $\eta(x)$ has been completed in Ref.~\cite{Ambjorn:1982bp}. Among them, the gluon quantum field satisfies the background field gauge,
\begin{align}
& D_\mu ^{ab}({\cal A})\phi _b^\mu = 0,
\nonumber\\
& D_\mu ^{ab}({\cal A}) = \delta^{ab}\partial_\mu - g_sf^{abc}{\cal A}_\mu ^c , \label{ca}
\end{align}
where the color indices $a,b,c = (1,2,...,8)$, $g_s$ is the coupling constant of strong interactions, $f^{abc}$ is the structure constant of the SU$_f$(3) group. The advantage of choosing the background field gauge is that it makes the theory be invariant under the background field gauge, making the calculated physical quantity independent of the gauge. With the help of Eq.~\eqref{AF}, one can obtain the following effective Lagrangian~\cite{Huang:1989gv}
\begin{align}
{\cal L}_{\rm eff} & = {\cal L}_{\rm QCD}({\cal A},\psi ) + {\cal L}({\rm ghosts}) + \bar \eta (i\DSS D - m)\eta + \frac12\phi _\mu ^a \bigg\{g^{\mu \nu}D_{ac}^2-\Big(1-\frac1\alpha\Big) [D^\mu {D^\nu }]_{ac}
\nonumber\\
&+ 2{g_s}f^{abc} G_b^{\mu \nu }\bigg\}\phi_\nu^c ~+~ g_s(\bar \psi \DS\phi^a T^a\eta +\bar\eta \DS\phi^aT^a\eta) ~-~ g_s^2f^{adf}f_{abc}{\cal A}_d^\mu \phi _f^\nu \phi _\mu ^b\phi _\nu ^c - g_s f^{abc}
\nonumber\\
& \times (\partial_\mu \phi _\nu ^a)\phi _b^\mu \phi _c^\nu-\frac{1}{4}g_s^2 f^{abc} f_{acf}\phi _b^\mu \phi _c^\nu \phi _\mu ^d\phi _\nu ^f + {g_s}\bar \eta {\DS\phi ^a}{T^a}\eta, ~\label{la}
\end{align}
where the ${\cal L}_{\rm QCD}({\cal A},\psi )$ with ${\cal A}_\mu^a (x)$, $\psi(x)$ has the usual form of QCD Lagrangian and can be minimized to zero when the classical fields ${\cal A}_\mu^a (x)$ and $\psi(x)$ are the solutions of the equation of motion, and $\alpha$ is the gauge-fixing parameter. The ${\cal L}({\rm ghosts})$ is the contribution of the ghost particle term. Especially, in order to describe the various quark-antiquark pairs and gluons in a vacuum, one can follow the classical QCD Lagrangian~\cite{Huang:1989gv}
\begin{align}
{\cal L}_{\rm QCD} ({\cal A},\psi) = -\frac14 G_{\mu\nu}^a G^{a\mu\nu} + \bar\psi(i\DSS D - m)\psi,
\end{align}
with $ G_{\mu\nu}^a = \partial_\mu {\cal A}_\nu^a - \partial_\nu {\cal A}_\mu^a + g_s f^{abc} {\cal A}_\nu^b {\cal A}_\mu^c$ stands for the gluon field strength tensor. The gluon field ${\cal A}_\mu^a(x)$ and quark field $\psi(x)$ satisfy the QCD equations of motion,
\begin{align}
&(i/\!\!\!\! D - m)\psi (x)=0 , \nonumber\\
&\widetilde D_\mu^{ab} G^{\nu\mu}_b(x)= g_s \bar\psi(x)\gamma^\nu T^a \psi(x) , \label{eq:mot}
\end{align}
where $D_\mu = \partial_\mu - i g_s T^a {\cal A}_\mu^a(x)$ with $a,b,c = (1,2,...,8)$ and $\widetilde D_\mu^{ab} = \delta^{ab}\partial_\mu - g_s f^{abc} {\cal A}_\mu^c(x)$ are fundamental and adjoint representations of the gauge covariant derivative, respectively. As an advantage of using the background field theory, one can take different gauges for dealing with the quantum fluctuations and background fields. More specifically, one can adopt the background gauge, i.e. $\widetilde{D}^{AB}_\mu \phi^{B \mu}(x) = 0$ for the gluon quantum field~\cite{BG3,BG4}, the Schwinger gauge or the fixed-point gauge, i.e. $x^\mu \mathcal{A}^A_\mu(x) = 0$ for the background field~\cite{Shifman:1980ui}.
As the background field satisfies the motion equation~\eqref{eq:mot}, at least one background field is included in the coupling between the quantum fluctuation field and the background field. In the effective Lagrangian, there is no contribution from the vertex ${g_s}\bar \psi {\gamma ^\mu }\phi_\mu ^a{T^a}\psi$, and it only has the vertex ${g_s}\bar \psi \DS\phi _\mu ^a{T^a}\eta$ and its conjugate contributions. According to the effective Lagrangian, the quantum quark and gluon propagators in the background field are~\cite{Huang:2004tp}
\begin{align}
& S_F(x) = i[i\gamma^\mu D_\mu-m]^{-1}, \label{qu}
\\
& S_{\mu \nu }^{ab}(x) = i[g_{\mu \nu}({D^2})^{ab} + 2{g_s} f^{abc}G_{\mu \nu }^c]^{ - 1} \label{glu},
\end{align}
where the gauge-fixing parameter is taken as $\alpha=1$. Within the framework of BFT, the quark propagator will be affected by the background quark and/or gluon fields, which satisfies the equation
\begin{eqnarray}
(i \DSS D - m) S_F(x,0) = \delta^4 (x).
\label{quaprofun1}
\end{eqnarray}
If taking
\begin{eqnarray}
S_F(x,0) = (i\DSS D + m) \mathcal{D}(x,0),
\label{quaprofun2}
\end{eqnarray}
Eq.~(\ref{quaprofun1}) can be changed as
\begin{eqnarray}
(\Box - \mathcal{P}_\mu \partial^\mu - \mathcal{Q} + m^2) \mathcal{D}(x,0) = \delta^4(x),
\label{quaprofun3}
\end{eqnarray}
where $\Box = \partial^2$, and
\begin{align}
&\mathcal{P}_\mu = 2i\mathcal{A}_\mu(x), \nonumber\\
&\mathcal{Q} = \gamma^\nu \gamma^\mu \mathcal{A}_\nu(x) \mathcal{A}_\mu(x) + i \gamma^\nu \gamma^\mu \partial_\nu \mathcal{A}_\mu(x).
\end{align}
Moreover, after applying the fixed-point gauge, the gluon background field can be expressed by using the gauge invariant $G_{\mu\nu;\alpha_1 \cdots \alpha_n}$ as
\begin{align}
\mathcal{A}_\mu(x) &= \frac{1}{2} x^\nu G_{\nu\mu} + \frac{1}{3} x^\nu x^\alpha G_{\nu\mu;\alpha} + \frac{1}{8} x^\nu x^\alpha x^\beta G_{\nu\mu;\alpha\beta}
+ \frac{1}{30} x^\nu x^\alpha x^\beta x^\gamma G_{\nu\mu;\alpha\beta\gamma}
\nonumber\\
& + \frac{1}{144} x^\nu x^\alpha x^\beta x^\gamma x^\delta G_{\nu\mu;\alpha\beta\gamma\delta} + \cdots, \label{gluexp}
\end{align}
where $G_{\mu\nu;\alpha_1 \cdots \alpha_n}$ is the notation for $(D_{\alpha_1}\cdots D_{\alpha_n}G_{\mu\nu})(0)$, where the indexes $\alpha_1 \cdots \alpha_n$ indicates the covariant derivative up to $n$-th order. Substituting Eq.~(\ref{gluexp}) into Eq.~(\ref{quaprofun3}), we obtain the expressions for $\mathcal{D}(x,0)$. By further using Eq.~(\ref{quaprofun2}), we obtain the required quark propagators in the background field, i.e.,
\begin{equation}
S_F(x,0)=S_F^0(x,0)+S_F^2(x,0)+S_F^3(x,0)+ \sum_{i=1}^{2} S_F^{4(i)}(x,0) + \sum_{i=1}^{3} S_F^{5(i)}(x,0) + \sum_{i=1}^{5} S_F^{6(i)}(x,0). \label{prop4}
\end{equation}
We present the quark propagators with various gauge invariant tensors $G_{\mu\nu;\alpha_1 \cdots \alpha_n}$ that shall result in up to dimension-six operators in the Appendix~\ref{sec:appendix_I1}. Because the fixed-point gauge violates the translation invariance, the quark propagator from $x$ to $0$, $S_F(0,x)$, can not be directly obtained by applying the replacement $x \to -x$ in Eq.~(\ref{prop4}). However, it can be related with $S_F(x,0)$ via the relation~\cite{Hubschmid:1982pa}
\begin{eqnarray}
S_F(0,x|\mathcal{A}) = C S^{\rm T}_F(x,0|-\mathcal A^{\rm T}) C^{-1},
\label{revpro}
\end{eqnarray}
where $C$ stands for the charge conjugation matrix and the symbol $T$ indicates transposition of both the Dirac and the color matrices.
Furthermore, one will encounter the vertex operator $\Gamma(z\cdot \tensor{D})^n$ with $\Gamma$ indicates all kinds of Dirac matrices for heavy/light-meson twist-2, 3 LCDAs. Generally, we have the following expansion~\cite{Zhong:2014jla}
\begin{eqnarray}
(z\cdot \tensor{D})^n = (z\cdot \overrightarrow{D} - z\cdot \overleftarrow{D})^n = (z\cdot \tensor{\partial} + z\cdot B)^n + \cdots,
\end{eqnarray}
where the ellipsis denotes for the higher-order terms, which are irrelevant for our present analysis and
\begin{eqnarray}
z\cdot B &=& -2 i z\cdot \mathcal{A} \nonumber\\
&=& -i x^\mu z^\nu G_{\mu\nu} - \frac{2i}{3} x^\mu x^\rho z^\nu G_{\mu\nu;\rho} - \frac{i}{4} x^\mu x^\rho x^\sigma z^\nu G_{\mu\nu;\rho\sigma} - \frac{i}{15} x^\mu x^\rho x^\sigma x^\lambda z^\nu G_{\mu\nu;\rho\sigma\lambda} \nonumber\\
&-& \frac{i}{72} x^\mu x^\rho x^\sigma x^\lambda x^\tau z^\nu G_{\mu\nu;\rho\sigma\lambda\tau} + \cdots. \label{zB}
\end{eqnarray}
We can expand the operator $(z\cdot \tensor{D})^n$ into series of the operators $(z\cdot \tensor{\partial})^n$ and $G_{\mu\nu;\rho\cdots}$. For the purpose, we first expand $(z\cdot \tensor{D})^n$ as
\begin{eqnarray}
(z\cdot \tensor{D})^0 &=& 1, \nonumber\\
(z\cdot \tensor{D})^1 &=& z\cdot \tensor{\partial} + z\cdot B, \nonumber\\
(z\cdot \tensor{D})^2 &=& (z\cdot \tensor{\partial})^2 + 2(z\cdot \tensor{\partial}) (z\cdot \underline{B}) + (z\cdot B)^2, \nonumber\\
(z\cdot \tensor{D})^3 &=& (z\cdot \tensor{\partial})^3 + 3(z\cdot \tensor{\partial})^2 (z\cdot \underline{B}) + \left[ (z\cdot \partial)^2 (z\cdot B) \right] + 3 (z\cdot \tensor{\partial}) (z\cdot \underline{B})^2 + (z\cdot B)^3, \nonumber\\
&& \cdots \cdots ,
\end{eqnarray}
where, the operator ``underline'' below symbol ``$B$'' (or ``$x$'' in the Appendix~\ref{sec:appendix_I2}) indicates that the operation $\tensor{\partial}$ does not act on it. In deriving those equations, the following equation has been adopted,
\begin{align}
(z\cdot \tensor{\partial})^n (z\cdot B) = \sum^n_{k=0} \frac{n!}{k! (n-k!)} (z\cdot \tensor{\partial})^{n-k} \left[ (z\cdot \partial)^k (z\cdot \underline{B}) \right].
\end{align}
By keeping only those terms that shall leads to operators up to dimension-six, we can obtain the full expression for the vertex operator, which are listed in the Appendix~\ref{sec:appendix_I2}.
\subsection{SVZ sum rules for the moments of $\phi _{2;\eta^{(\prime)}}(x,\mu)$}
Following the traditional way of constructing the light pseudoscalar meson, we take the following definition
\begin{align}
\langle 0| \bar s (0){\cal C}_s\DS z\gamma_5[z,-z] {(iz\cdot\tensor D)^{n}}s(0)|\eta^{(\prime)}(q)\rangle =i{(z\cdot q)}^{n+1} f_{\eta^{(\prime)}} \langle\xi^n_{2;\eta^{(\prime)}}\rangle|_\mu, \label{Eq:xi_defination}
\end{align}
where $z_\mu$ stand for the light-like vector and $[z,-z]$ is the path-ordered gauge connection, $f_{\eta^{(\prime)}}$ are the $\eta$ and $\eta^{\prime}$-meson decay constant, $(iz \cdot \tensor D)^n=(iz\cdot \overrightarrow D - iz\cdot \overleftarrow D)^n$. Basis on the QF scheme, the flavour content is ${\cal C}_s = ({\cal C}_1 - \sqrt{2}{\cal C}_8)/ \sqrt{3}$ with SO basis ${\cal C}_1 = \mathbf{1}/\sqrt{n_f}$ and ${\cal C}_8 = \lambda_8/\sqrt{2}$~\cite{Duplancic:2015zna,Ball:2007hb}. In which the $\lambda_i$ is the standard SU$_f$(3) Gell-Mann matrix and $\mathbf{1}$ is the $3\times 3$ unit matrix. As a special case, the 0th-order LCDA's moment for Eq.~\eqref{Eq:xin} satisfies the normalization condition
\begin{align}
\langle \xi^0_{2;\eta^{(\prime)}}\rangle|_\mu =\int_0^1 du \phi_{2;\eta^{(\prime)}}(u,\mu)=1.
\end{align}
To derive the SVZ sum rules for the moments $\langle \xi^n_{2;\eta^{(\prime)}}\rangle|_\mu$, we introduce the following correlation function,
\begin{align}
\Pi_{2;\eta^{(\prime)}}^{(n,0)}(z,q) &=i \int d^4x e^{iq\cdot x}\langle 0 |T\{ J_n(x), J_0^\dagger (0) \}|0\rangle
\nonumber\\
&=(z\cdot q)^{n+2} I_{2;\eta^{(\prime)}}^{(n,0)}(q^2), \label{gp}
\end{align}
where $J_n(x)= \bar s(x){\cal C}_s\DS z\gamma_5 (iz \cdot\tensor D)^ns(x)$, $J_0^{\dagger}(0)= \bar s(0){\cal C}_s\DS z\gamma_5 s(0)$ and $ z^2 \rightsquigarrow 0$. Only even moments are non-zero and the odd moments of the LCDA are zero because of G-parity, i.e. $n=(0,2,4,6,...)$ will contribute to the final results~\footnote{This point can also been seen in Eq. (16) of Ref.~\cite{Ball:2007hb}.}. At one hand, in deep Euclidean region $q^2 \ll 0$, one can apply the OPE for the correlation function Eq.~\eqref{gp}.
\begin{align}
\Pi_{2;\eta^{(\prime)}}^{(n,0)}(z,q) =i\int d^4x e^{iq\cdot x}{\rm Tr}[{\cal C}_s{\cal C}_s]
\Big\{\!&-{\rm Tr}\langle 0|S_F^s(0,x)\DS z{{\gamma }_{5}}(iz\cdot \tensor D)^n S_F^s(x,0)\DS z\gamma_5|0\rangle
\nonumber\\
& +{\rm Tr}\langle 0|\bar{s}(x)s(0)\DS z{{\gamma }_{5}}{(iz\cdot \tensor{D})^{n}}S_F^s(x,0)\DS z{{\gamma }_{5}}|0\rangle
\nonumber\\
&+{\rm Tr}\langle 0|S_F^s(0,x)\DS z{{\gamma }_{5}}{(iz\cdot \tensor{D})^{n}}\bar{s}(0)s(x)\DS z{{\gamma }_{5}}|0\rangle
\nonumber\\
& +\cdots \Big\},
\end{align}
where ${\rm Tr}[{\cal C}_s{\cal C}_s] =1$. In the detailed OPE calculation, we adopt the $\overline{\rm MS}$-scheme to deal with the infrared divergences. Lorentz invariant scalar function $I_{2;\eta^{(\prime)}}^{(n,0)}(q^2)$ in Eq.~\eqref{gp} depends on the condensation parameter and will encounter the vacuum matrix elements of the following form,
\begin{align}
&\langle 0|G_{\mu\nu}^a G_{\mu\nu}^b|0\rangle,
&&\langle 0|G_{\mu \nu }^{a}G_{\rho\sigma }^{b}G_{\lambda \tau }^{c}|0\rangle ,
\nonumber\\
&\langle 0|G_{\mu \nu ;\lambda }^{a}G_{\rho \sigma ;\tau }^{b}|0\rangle, && \langle 0|G_{\mu \nu }^{a}G_{\rho \sigma ;\lambda \tau }^{b}|0\rangle ,
\nonumber\\
&\langle 0|G_{\mu \nu ;\lambda \tau }^{a}G_{\rho \sigma }^{c}|0\rangle, && \langle 0|\bar q_\alpha^a(x) q_\beta^b(y)|0\rangle,
\nonumber\\
&\langle 0|\bar q_\alpha^a(x)q_\beta^b(y)G_{\mu\nu}^A|0\rangle,
&& \langle 0|\bar q_\alpha^a(0) q_\beta^b(0) G_{\mu\nu;\rho}^A|0\rangle.
\end{align}
The full expression for the vacuum condensates which one may encounter in the following calculations are listed in the Appendix~\ref{sec:appendix_I3}, which can also be found in our previous work~\cite{Zhong:2014jla}. On the other side, the correlation function ~\eqref{gp} can be treated by inserting a complete set of intermediate hadronic states in physical region to obtain its hadronic representation
\begin{align}
{\rm Im}I_{2;\eta^{(\prime)},{\rm Had}}^{(n,0)}( q^2 )&=\pi \delta(q^2 - m_{\eta^{(\prime)}}^2) f_{\eta^{(\prime)}}^2 \langle \xi^n_{2;\eta^{(\prime)}} \rangle|_\mu +\pi \frac{3}{4\pi^2 (n+1)(n+3)}\theta ( q^2 -s_{\eta^{(\prime)}}), \label{rm}
\end{align}
in which $s_{\eta^{(\prime)}}$ stand for the continuum threshold for the lowest continuum state. The first term is the contribution of $\eta^{(\prime)}$-meson poles, and the second term is the contribution of continuum states above poles. Then both of the OPE part and the hadronic representation of the invariant function $I_{2;\eta^{(\prime)}}^{(n,0)}( q^2)$ can be marched with the following dispersion relation
\begin{align}
\frac{1}{\pi}\int_{4m_s^2}^{\infty }{ds}\frac{{\rm Im}I_{2;\eta^{(\prime)} ,{\rm Had}}^{(n,0)}(s)}{s-q^2}=I_{2;\eta^{(\prime)} ,{\rm QCD}}^{(n,0)}({{q}^{2}}). \label{Ds}
\end{align}
The Borel transform helps to reduce the contribution from the continuum states on the left of Eq.~\eqref{Ds} and the contribution of the higher-dimension condensates on the right, and finally the sum rules can be obtained,
\begin{align}
\frac1{\pi}\frac1{M^2}\int ds e^{-s/M^2}{\rm Im}I_{2;\eta^{(\prime)} , {\rm had}}^{(n,0)}(s)=\hat{\cal B}_{M^2} I_{2;\eta^{(\prime)} ,{\rm QCD}}^{(n,0)}(q^2),
\end{align}
where $M^2$ is Borel parameter, $\hat{\cal B}_{M^2}$ is Borel transformation operator,
\begin{align}
\hat{\cal B}_{M^2} = \mathop {\lim }\limits_{\scriptstyle - q^2, n\to\infty \hfill\atop
\scriptstyle{-q^2/n} = {M^2}\hfill} \frac{1}{(n - 1)!}(-q^2)^n \Bigg( - \frac{d}{d( -q^2)}\Bigg)^n.
\end{align}
In order to deal with the $s$-quark mass ($m_s$) contribution to $\langle\xi_{2;\eta^{(\prime)}}^{n}\rangle|_\mu$, we take the expansion according to different orders of $m_s^{k}$ with $k=(0,2,4,...)$, i.e. $I_{m_s^{k}}(n, M^2)$, since $m_s$ is closer to $\Lambda_{\rm QCD}$, which is different from our previous treatment for the heavy quark such as $q=(c,b)$ in Refs.~\cite{Fu:2018vap, Zhong:2014fma}. Here, we take the first two orders of the squared $s$-quark mass, i.e. $I_{m_s^0}(n, M^2)$ and $I_{m_s^2}(n, M^2)$. The reason lies in the suppression of $m_s^4 < 0.1\%$ for the third-order, which are quite small and can be safely neglected. Recently, we have suggested a new method for renormalization of various moments, i.e. the 0th-order moment $\langle\xi_{2;\eta^{(\prime)}}^0\rangle|_\mu$ should be considered to the total results. The reason lies in the accuracy is often up to dimension-six condensates and the NLO QCD corrections for the perturbative part instead of the infinite dimension or infinite-order perturbative parts~\cite{Zhong:2021epq}. The final expression for $\langle\xi_{2;\eta^{(\prime)}}^{n}\rangle|_\mu$ can then be written as
\begin{align}
&\frac{f_{\eta^{(\prime)}}^2 \langle\xi^n_{2;\eta^{(\prime)}}\rangle|_\mu \langle \xi^0_{2;\eta^{(\prime)}}\rangle|_\mu}{M^2 e^{m_{\eta^{(\prime)}}^2/M^2}} = \frac1\pi \frac1{M^2} \int_{4m_s^2}^{s_{\eta^{(\prime)}}} ds e^{-s/M^2} \frac{3 v^{n+1}}{8\pi(n+1)(n+3)} \left(1+ \frac{\alpha_s}{\pi }A'_n\right) \bigg\{[1+(-1)^n]
\nonumber
\\
&\hspace{1.8cm}\times(n+1)\frac{1-v^2}2 + [1+(-1)^n]\bigg\}+\frac{2m_s\langle\bar ss\rangle}{M^4}+\frac{\langle \alpha_s G^2\rangle}{12\pi M^4}~\frac{1+n\theta(n-2)}{n+1} \,-\, \frac{8n+1}9
\nonumber
\\
&\hspace{1.8cm}\times\frac{m_s\langle g_s\bar s\sigma TGs\rangle}{M^6} + \frac{\langle g_s \bar ss\rangle }{81M^6}4(2n + 1) -\frac{\langle g_s^3fG^3\rangle}{48\pi^2M^6}n\theta(n-2)+\frac{\sum\langle g_s^2\bar qq\rangle ^2}{486\pi^2M^6}\bigg\{\!\!-2(51n
\nonumber
\\
&\hspace{1.8cm}
+25)\bigg(-\ln\frac{M^2}{\mu^2}\bigg)+3\,(17n+35)+\theta(n-2)\bigg[\,2n\bigg(-\ln\frac{M^2}{\mu^2}\bigg)-25\,(2n+ 1)\,\tilde \psi (n)
\nonumber
\\
&\hspace{1.8cm} + \frac1n(49n^2+100n+56)\bigg]\bigg\}+ I_{m_s^2}(n,M^2).\label{xi2}
\end{align}
Due to the mass of $s$ quark is heavier than $u$, $d$-quark, the $I_{m_s^2}(n, M^2)$-terms should be considered in this paper, which are
\begin{align}
&I_{m_s^2}(n,M^2)= m_s^2\bigg\{-\frac{\langle\alpha_s G^2\rangle}{6\pi M^6}\left[\theta(n-2)(n\tilde\psi (n)-2) + 2n\left(-\ln\frac{M^2}{\mu^2}\right)-n-2\right]+\frac{\langle g_s^3fG^3\rangle}{288\pi^2M^8}
\nonumber\\
&\hspace{0.6cm} \times \bigg\{\!-10 \delta^{n0} + \theta (n - 2) \left[ 4n(2n-1)\left( -\ln\frac{M^2}{\mu^2}\right) - 4n\tilde \psi (n) + 8({n^2} - n + 1) \right] + \theta(n-4)
\nonumber\\
&\hspace{0.6cm}\times [2n(8n-1)\tilde\psi(n) - (19{n^2} + 19n + 6)]+ 8n\,(3n - 1)\left( -\ln\frac{M^2}{\mu^2} \right) - (21n^2+53n-6)\bigg\}
\nonumber\\
&\hspace{0.6cm}- \frac{\sum \langle g_s^2q\bar q\rangle^2 }{972\pi^2M^8}~\bigg\{ 6{\delta ^{n0}}\left[ {16\left( -\ln\frac{M^2}{\mu^2} \right) - 3} \right] + \theta (n - 2) \bigg[8(n^2+ 12n -12)\left( -\ln\frac{M^2}{\mu^2} \right) - 2
\nonumber\\
&\hspace{0.6cm} \times (29n + 22)\, \tilde \psi (n)~ +~ 4\left(5{n^2} - 2n - 33 + \frac{46}{n}\right)\bigg] + \theta (n - 4)~\bigg[2\left(56{n^2} - 25n + 24\right) \tilde \psi(n)
\nonumber\\
&\hspace{0.6cm} -\,\left(139 n^2 + 91n + 54\right)\bigg] ~+~ 8\left(27n^2-15n-11\right)~ \left(-\ln\frac{M^2}{\mu^2} \right) - 3(63n^2+159n-50) \bigg\}
\nonumber\\
&\hspace{0.6cm} +\frac{4(n-1)}3 \frac{m_s\langle\bar ss\rangle }{M^6} + \frac{8n-3}9 \frac{m_s\langle g_s\bar s\sigma TGs\rangle}{M^8} - \frac{4(2n+1)}{81}\frac{\langle g_s \bar ss\rangle^2}{M^8}\bigg\}. \label{ms}
\end{align}
For convenience, we put the detailed terms contribute to the BFTSR $\langle\xi_{2;\eta^{(\prime)}}^n\rangle|_\mu$ in the Appendix~\ref{sec:appendixB}. By taking the index $n$ to zero, we get the sum rule of the 0th-order moment, which takes the following form,
\begin{align}
&\frac{(\langle\xi^0_{2;\eta^{(\prime)}}\rangle|_\mu)^2 f_{\eta^{(\prime)}}^2}{M^2e^{m_{\eta^{(\prime)}}^2/M^2}}= \frac1\pi \frac1{M^2}\int_{4m_s^2}^{s_{\eta^{(\prime)}}}ds e^{-s/M^2} \frac{v(3-v^2)}{8\pi} + \frac{2m_s\langle\bar ss\rangle}{M^4}-\frac{m_s\langle g_s\bar s\sigma TGs\rangle}{9M^6}+ \frac{\langle \alpha_s G^2\rangle}{12\pi M^4}
\nonumber\\
&\hspace{1.05cm} + \frac{4\langle g_ss \bar s\rangle^2}{81M^6}+\frac{\sum\langle g_s^2\bar qq\rangle^2}{M^6} \frac1{486{\pi^2}}\bigg[-50\bigg(-\ln\frac{M^2}{\mu^2}\bigg) + 105\bigg] + m_s^2\bigg\{\frac{\langle \alpha_s G^2\rangle }{3\pi M^6}- \frac{\langle g_s^3fG^3\rangle}{72\pi^2M^8}
\nonumber\\
&\hspace{1.05cm} - \frac{\sum\langle g_s^2\bar qq\rangle^2}{972\pi ^2M^8} ~\Bigg[8\left( -\ln\frac{M^2}{\mu^2} \right)-132 \Bigg] - \frac43 \frac{m_s\langle \bar ss\rangle}{M^6} \,-\, \frac13\frac{m_s\langle g_s\bar s\sigma TGs\rangle}{M^8} \,-\, \frac4{81}\frac{\langle g_s\bar ss\rangle^2}{M^8}\bigg\},
\end{align}
where
\begin{align}
\widetilde\psi(n)=\psi\bigg(\frac{n+1}2\bigg)-\psi \bigg(\frac n2\bigg)+\ln4 ,
\end{align}
with $v^2 = 1 - 4m_s^2/s$ and $A'_0=0$, $A'_2=5/3$, $A'_4=59/27$, $A'_6=353/135$ are the NLO correction to the perturbative part~\cite{Ball:1996tb}. The 0th-order derivative of the digamma function $\psi(n+1) = \Sigma_{k=1}^n 1/k -\gamma_E$, where the Euler's constant $\gamma_E = 0.557216$. Furthermore, in order to get the Gegenbauer moment $a_{2;\eta^{(\prime)}}^n$, one can expand $\phi_{2;\eta^{(\prime)}}^n(u, \mu)$ into a Gegenbauer polynomial series by using Eq.~\eqref{HPDA_CZ} and the basic definition of $\langle\xi^n_{2;\eta^{(\prime)}}\rangle|_\mu$~\eqref{Eq:xin}. Then, one can get the following relations up to $6$th-order,
\begin{align}
&\langle\xi^2_{2;\eta^{(\prime)}}\rangle|_\mu=\frac15 + \frac{12}{35}a_{2;\eta^{(\prime)}}^2(\mu),
\nonumber\\
&\langle\xi^4_{2;\eta^{(\prime)}}\rangle|_\mu=\frac3{35} + \frac8{35}a_{2;\eta^{(\prime)}}^2(\mu) + \frac8{77}a_{2;\eta^{(\prime)}}^4(\mu),
\nonumber\\
&\langle\xi^6_{2;\eta^{(\prime)}}\rangle|_\mu=\frac1{21} + \frac{12}{77} a_{2;\eta^{(\prime)}}^2(\mu) + \frac{120}{1001}a_{2;\eta^{(\prime)}}^4(\mu) +\frac{64}{2145} a_{2;\eta^{(\prime)}}^6(\mu). \label{xi}
\end{align}
Following this method, one can get higher-order Gegenbauer moments.
\subsection{The semileptonic decay $D_s^+\to \eta^{(\prime)}\ell^+\nu_\ell $}
All the following calculations are performed under the Standard Model (SM). In order to derive the full analytical LCSR expressions for the TFFs, we use the traditional current method to calculate the TFFs. The correlation function for the TFFs of $D_s\to\eta^{(\prime)}$ is defined as~\cite{Duplancic:2008ix}:
\begin{align}
\Pi_\mu(p,q) & =i\int d^4 x e^{iqx} \langle \eta^{(\prime)}(p)|T\{\bar s(x)\gamma_\mu c(x),\bar c(0)i\gamma_5 s(0) \}|0 \rangle \nonumber\\
& = \Pi[q^2, (p+q)^2] p_\mu + \tilde \Pi[q^2, (p+q)^2] q_\mu.
\label{Correlation function}
\end{align}
Following the basic procedure of LCSR approach, the correlation function can be treated by inserting complete intermediate states with the same quantum numbers as the current operator $(\bar c i \gamma_5 s)$ in the time-like ${(p+q)}^2$-region. After isolating the pole term of the lowest pseudoscalar $D_s$-meson, one can reach the following expression,
\begin{align}
\Pi_\mu^{\rm had}(p,q)&=\frac{\langle\eta^{(\prime)}(p)|\bar s\gamma_\mu c|D_s(p+q)\rangle \langle D_s(p+q)|\bar ci\gamma_5q|0\rangle }{m_{D_s}^2-(p+q)^2}
\nonumber\\
&+\sum\limits_{\rm H}{\frac{\langle\eta^{(\prime)}(p)|\bar s\gamma_\mu c|D_s^{\rm H}(p+q)\rangle \langle D_s^{\rm H}(p+q)|\bar ci\gamma_5q|0\rangle}{m_{D_s^{\rm H}}^2-(p+q)^2}}\nonumber\\[1.2ex]
& = \Pi^{\rm had}[q^2,(p+q)^2]p_\mu+\widetilde\Pi^{\rm had}[q^2,(p+q)^2]q_\mu,
\label{Eq:Hadronic Expression}
\end{align}
with the superscript ``had'' and ``H'' stand for the hadronic expression of the correlation function and higher-excited state of $D_s$-meson, respectively. The decay constant of $D_s$-meson can be defined via the relation, $\langle D_s|\bar ci\gamma_5q|0\rangle = m_{D_s}^2f_{D_s}/m_c $. The definition of transition matrix element for $D_s\to \eta^{(\prime)}$ has the following form
\begin{align}
\langle\eta^{(\prime)}(p)|\bar s\gamma_\mu c| D_s(p+q)\rangle = 2f^{\eta^{(\prime)}}_+(q^2)p_\mu
+ \tilde f^{\eta^{(\prime)}}(q^2)q_\mu.
\end{align}
with $\tilde f^{\eta^{(\prime)}} = f^{\eta^{(\prime)}}_+(q^2)+f^{\eta^{(\prime)}}_-(q^2)$. Then one can take the imaginary part of the invariant amplitude $\Pi^{\rm had}[q^2,(p+q)^2]$ and $\widetilde\Pi^{\rm had}[q^2,(p+q)^2]$ which has the following form,
\begin{eqnarray}
&&{\rm Im}\Pi^{\rm had}(q^2,s) = \pi \delta(s-m_{D_s^2}) \frac{2m_{D_s^2}f_{D_s}f_+^{\eta^{(\prime)}}}{m_c} + \pi\rho^{\rm H}(s) \theta(s-s_0)
\nonumber\\
&&{\rm Im}\widetilde\Pi^{\rm had}(q^2,s) = \pi \delta(s-m_{D_s^2}) \frac{m_{D_s^2}f_{D_s}\tilde f_+^{\eta^{(\prime)}}}{m_c} + \pi\tilde\rho^{\rm H}(s) \theta (s-\tilde s_0)
\end{eqnarray}
where $\rho^{\rm H}(s)$ and $\tilde\rho^{\rm H}(s)$ denote the spectral density of higher resonance and the continuum states $D_s^{\rm H}$, which can be approximated by invoking the so-called quark-hadron duality ansatz $\varrho^{\rm H}(s) = \varrho^{\rm QCD}(s)$ with $\varrho(s) = (\rho(s), \tilde\rho(s))$ and the perturbative spectral density $\rho^{\rm QCD}$ and $\tilde \rho^{\rm QCD}$, the usual step-function $\theta(x)$. Here, the $s_0$ ($\tilde s_0$) are effective parameters, which characterize the lower limit of continuum state, namely the continuum threshold parameters. After extracting the pole terms, the $s_0(\tilde s_0)$ can separate the ground state $D_s$ and the excited state $D_s^{\rm H}$. When taking the limit for the interval between the two adjacent excited states, the sum of excited states is transformed into the integral of continuum states and $s_0(\tilde s_0)$ will be changed into the lower limit of the integration. Traditionally, the continuum threshold are often taken as the magnitude that is close to the squared mass of the first excited state. The ground state of $D_s$-meson is calculated by the LCSR approach. The excited states' contribution can be highly suppressed when making the Borel transformation. Meanwhile, the continuum states, contribution is usually required to be less than 30\% so as to make the LCSR calculation more accurate and reliable. Then, one can use a general dispersion relation in the momentum squared $(p+q)^2$, which can establish a relationship with the QCD parts, i.e.
\begin{align}
\Pi^{\rm QCD}(q^2,s)&=\frac1{\pi} \int_{m_c^2}^\infty \frac{{\rm Im}\Pi ^{\rm had}(q^2,s)}{s-(p+q)^2}ds,
\nonumber\\
\widetilde\Pi^{\rm QCD}(q^2,s)&=\frac1{\pi} \int_{m_c^2}^\infty \frac{{\rm Im}\tilde\Pi ^{\rm had}(q^2,s)}{s-(p+q)^2}ds.
\end{align}
Here, we shall only deal with ${\rm Im}\Pi^{\rm had}[q^2,s]$ for the TFFs $f^{\eta^{(\prime)}}_+(q^2)$, which are the only TFFs contribute to the required branching fractions.
On the other hand, when the correlation function~\eqref{Correlation function} is dominated by the light-like distances, it can be expanded around the light-cone. The light-cone expansion is performed by integrating out the transverse and minus degrees of freedom and leaving only the longitudinal momenta of the partons as the relevant degrees of freedom. The integration over the transverse momenta is done up to a cutoff, $\mu_{\rm IR}$, all momenta below which are included in the $\eta^{(\prime)}$-meson LCDAs. Large transverse momenta are calculated in perturbative theory. Thus, the correlation function can be separated into perturbative and nonperturbative contributions, both of which depend on the longitudinal parton momenta and the factorization scale $\mu_{\rm IR}$~\cite{Ball:2007hb}.
In order to make our result more accurate, we consider both the leading-order (LO) for all the LCDAs' part and gluon radiative corrections to the dominant twist-2, 3 parts of the correlation function. The OPE result for the correlation function $\Pi^{\rm OPE}$ is then represented as a sum of LO and NLO parts,
\begin{align}
\Pi^{\rm OPE}[q^2, (p+q)^2] = F_0(q^2,(p+q)^2)+\frac{\alpha_s C_F}{4\pi } F_1(q^2,(p+q)^2).
\end{align}
To calculate the invariant amplitude $F_0(q^2,(p+q)^2)$ and $F_1(q^2,(p+q)^2)$, one needs to know the expression for the $c$-quark propagator, i.e.
\begin{align}
\langle 0|T\{c(x)\bar c(0)\}|0\rangle &=i\int\frac{d^4k}{(2\pi)^4}{e^{-ik\cdot x}}\frac{ m_c + \DS k}{ k^2-m_c^2}-ig \int\frac{d^4 k}{(2\pi)^4}e^{-ik\cdot x} \int_0^1 dv\Bigg[ \frac{m_c + \DS k}{2(m_c^2-k^2)^2} \nonumber \\
& \times G^{\mu\nu}(vx)\sigma_{\mu\nu} +\frac{v}{m_c^2-k^2}x_\mu G^{\mu\nu}(vx)\gamma_\nu \Bigg].
\end{align}
To do the calculation, the expression of twist-2, 3, 4 LCDAs matrix elements are needed, which are displayed in the Appendix~\ref{sec:appendixC}. To get the final LCSR expression, we need to use the Borel transformation to transform the variable $(p+q)^2$ into Borel parameter $M^2$. Then the expression of the $D_s\to \eta^{(\prime)}$ TFFs up to NLO gluon radiation correction to the twist-2, 3 LCDAs can be obtained by equating the two types of representation of the correlation function and by subtracting the contribution from higher resonances and continuum states, i.e.,
\begin{align}
f^{\eta^{(\prime)}}_+(q^2)&=\frac{e^{m_{D_s}^2/M^2}} {2m_{D_s}^2 f_{D_s}} \bigg[F_0(q^2,M^2,s_0)+\frac{\alpha_s C_F}{4\pi } F_1(q^2,M^2,s_0)\bigg].
\end{align}
where $F_{0(1)}(q^2, M^2, s_0)$ originates from the OPE result for the LO (NLO) invariant amplitude $F_{0(1)}(q^2, (p+q)^2)$. Finally, the LCSR for $D_s\to \eta^{(\prime)}$ TFFs have the following form,
\begin{align}
&f^{\eta^{(\prime)}}_+(q^2) = \frac{m_c^2 f_{\eta^{(\prime)}}}{2m_{D_s}^2 f_{D_s}}e^{m_{D_s}^2/M^2}\!\!\int_{u_0}^1\! du e^{-s(u)/M^2}\bigg\{\frac{\phi_{2;\eta^{(\prime)}}(u)}u \!+ \frac1{2 m_s m_c}\bigg[\phi_{3;\eta^{(\prime)}}^p(u) \!+\! \frac16 \bigg(2\frac{\phi_{3;\eta^{(\prime)}}^\sigma (u)}u
\nonumber\\
&\hspace{0.28cm}
- \frac{m_c^2+q^2-u^2m_{\eta^{(\prime)}}^2}{m_c^2-q^2+u^2m_{\eta^{(\prime)}}^2} \frac d{du}\phi_{3;\eta^{(\prime)}}^\sigma (u)+ \frac{4um_{\eta^{(\prime)}} ^2m_c^2}{(m_c^2 - q^2 + u^2m_{\eta^{(\prime)}}^2)^2}\phi_{3;\eta^{(\prime)}}^\sigma (u)\bigg)\bigg]\! +\! \frac1{m_c^2-q^2+u^2 m_{\eta^{(\prime)}}^2}
\nonumber\\
&\hspace{0.28cm}
\times \bigg[u\psi_{4;\eta^{(\prime)}}(u) + \left(1 - \frac{{2{u^2}m_{\eta^{(\prime)}}^2}}{m_c^2 - q^2 + u^2 m_{\eta^{(\prime)}}^2}\right) \int_0^u dv \psi_{4;\eta^{(\prime)}}(v)- \frac{m_c^2}{4}\frac{u}{m_c^2 - q^2 + u^2m_{\eta^{(\prime)}} ^2}\,\bigg(\frac{d^2}{du^2}
\nonumber\\
&\hspace{0.28cm}
- \frac{6um_{\eta^{(\prime)}} ^2}{m_c^2-q^2+u^2m_{\eta^{(\prime)}}^2}\frac{d}{du} \,+ \frac{12um_{\eta^{(\prime)}}^4}{(m_c^2 - q^2 + u^2m_{\eta^{(\prime)}} ^2)^2}\bigg)\phi _{4;\eta^{(\prime)}}(u)
- \bigg(\frac d{du} - \frac{2u m_{\eta^{(\prime)}}^2}{m_c^2-q^2+u^2m_{\eta^{(\prime)}}^2}\bigg)
\nonumber\\
&\hspace{0.28cm}
\times \bigg(\frac{f_{3\eta^{(\prime)}}}{f_{\eta^{(\prime)}} m_c}~I_{3;\eta^{(\prime)} }(u) ~+ I_{4;\eta^{(\prime)}}(u)\bigg)- \frac{{2{u}m_{\eta^{(\prime)}} ^2}}{m_c^2-q^2+u^2m_{\eta^{(\prime)}}^2}~\bigg(u\frac{d}{du} - \frac{{2{u^2}m_{\eta^{(\prime)}} ^2}}{m_c^2-q^2+u^2m_{\eta^{(\prime)}}^2} + 1\bigg)
\nonumber
\\
&\hspace{0.28cm}
\times \bar I_{4\eta^{(\prime)} }(u) ~+~ \frac{2u m_{\eta^{(\prime)}}^2(m_c^2-q^2-u^2m_{\eta^{(\prime)}} ^2)}{(m_c^2-q^2+u^2m_{\eta^{(\prime)}}^2)^2}~~\bigg(\frac{d}{du} - \frac{6um_{\eta^{(\prime)}} ^2}{m_c^2-q^2+u^2m_{\eta^{(\prime)}}^2}\bigg)~\int_u^1 {d\xi} \bar I_{4\eta^{(\prime)}}(\xi )\bigg]\bigg\}
\nonumber
\\
&\hspace{0.28cm}
+\frac{m_c^4{f_{\eta^{(\prime)}} } e^{ - m_c^2/M^2}}{4(m_c^2 - q^2 + m_{\eta^{(\prime)}} ^2)^2} \bigg(\frac{d}{du}\phi_{4;\eta}(u)\bigg)\bigg|_{u\to 1} + \frac{\alpha_s C_F e^{m_{D_s}^2/M^2}}{8\pi m_{D_s}^2 f_{D_s}} F_1(q^2,m^2,s_0),
\nonumber
\\\label{Eq:fp}
\end{align}
with $s(u) = m_c^2-(q^2-m_{\eta^{(\prime)}}^2u)\bar u/u$. The LO invariant amplitudes include twist-2, 3, 4 contributions. The NLO QCD corrections to the invariant amplitudes $F_1(q^2, (p+q)^2)$ include twist-2, 3 contributions, which can be separated into the following form,
\begin{align}
F_1(q^2,m^2,s_0) &=\frac1\pi \int_{m_c^2}^{s_0}dse^{-s(u)/M^2}{\rm Im} F_1(q^2,s)
\nonumber\\
& =\frac{f_{\eta^{(\prime)}}}\pi \int_{m_c^2}^{s_0} ds e^{-s(u)/M^2}\int_0^1{du}\Big\{ {\rm Im}{T_1}(q^2,s,u)\phi_{2;\eta^{(\prime)}}(u)
\nonumber\\[2ex]
& +\frac{\mu_{\eta^{(\prime)}}}{m_c}\big[{\rm Im}T_1^p(q^2,s,u)\phi_{3;\eta^{(\prime)}}^p(u) + {\rm Im}T_1^{\sigma}(q^2,s,u)\phi_{3;\eta^{(\prime)}}^\sigma(u)\big] \Big\},
\end{align}
where ${\mu _{\eta^{(\prime)}} } = {m_{\eta^{(\prime)}} ^2} /{2{m_s}}$. The imaginary parts of the amplitudes $T_1(q^2,s,u)$, $T_1^p(q^2,s,u)$ and $T_1^\sigma(q^2,s,u)$ are the hard-scattering amplitudes calculated by the 6 diagrams for the gluon corrections. The final detailed expressions agree with those of Refs.~\cite{Duplancic:2008ix, Duplancic:2015zna}, which are not listed here. The lower limit of integral
\begin{eqnarray}
u_0 = \Big(q^2 - s_0 + m_{\eta^{(\prime)}}^2 + \sqrt{(q^2 - s_0 + m_{\eta^{(\prime)}}^2 )^2 -4 m_{\eta^{(\prime)}}^2 (q^2 - m_c^2)}\Big)/(2m_{\eta^{(\prime)}}^2)
\end{eqnarray}
and for the final expression, we need a brief introduction to the integrals over three-particle LCDAs, i.e.
\begin{align}
I_{3;\eta^{(\prime)}}(u)& =\frac d{du}\Bigg[\int_0^u d\alpha_1 \int_\Delta^1 dv \Phi_{3;\eta^{(\prime)}}(\alpha_i)\Bigg],
\nonumber\\
I_{4;\eta^{(\prime)}}(u)&=\frac d{du}\Bigg\{\int_0^u d\alpha_1 \int_\Delta^1 \frac{dv}{v}
\Bigg[2\Psi_{4;\eta^{(\prime)}}(\alpha_i)-\Phi_{4;\eta^{(\prime)}}(\alpha_i)
+2\widetilde \Psi_{4;\eta^{(\prime)}}(\alpha_i)-\widetilde \Phi_{4;\eta^{(\prime)}}(\alpha_i)\Bigg]\Bigg\},
\nonumber\\
\bar I_{4;\eta^{(\prime)}}(u)&=\frac{d}{du}\Bigg\{\int_0^u d\alpha_1
\int_\Delta^1 \frac{dv}{v}\Bigg[\Psi_{4;\eta^{(\prime)}}(\alpha_i) ~+~ \Phi_{4;\eta^{(\prime)}}(\alpha_i)
+\widetilde{\Psi}_{4;\eta^{(\prime)}} (\alpha_i) \,+\, \widetilde{\Phi} _{4;\eta^{(\prime)}} (\alpha_i)\Bigg]\Bigg\},
\label{eq:fplusBpiLCSR3part}
\end{align}
where $\Delta = {(u-\alpha_1)}/{(1-\alpha_1)}$, $\alpha_2=1-\alpha_1-\alpha_3$ and $\alpha_3=(u-\alpha_1)/v$. Due to the contributions from three-particle parts are quite small, i.e. $<0.1\%$, we can safely neglect these parts in this paper. We would like to figure out that the decay branching fraction for the considered decay. Using the parametrization of the transition matrix elements in terms of TFFs, in massless lepton case, we get
\begin{align}
\frac{d\Gamma}{dq^2}(D_s^+\to \eta^{(\prime)}\ell^+\nu_\ell)&= \frac {G_F^2 |V_{cs}|^2}{192\pi^3m_{D_s}^{3}} \Big[\left( m_{D_s}^2+m_{\eta^{(\prime)}}^2-q^2 \right)^2 -4m_{D_s}^2 m_{\eta^{(\prime)}}^2 \Big]^{3/2}|f^{\eta^{(\prime)}}_+(q^2)|^2, \label{Eq:dGamma}
\end{align}
where the fermi coupling constant $G_F=1.166\times{10}^{-5}~{\rm GeV}^{-2}$.
\section{Numerical Discussions}\label{sec:3}
\subsection{Input parameters}
We adopt the following parameters to do the numerical calculation. The current charm-quark mass is $m_c = 1.27\pm0.02~ {\rm GeV}$, the masses of $D_s$, $\eta$ and $\eta'$-meson $m_{D_s}=1.9685~{\rm GeV}$, $m_{\eta}=0.5478~{\rm GeV}$, $m_{\eta'}=0.9578~{\rm GeV}$ and $s$-quark mass $m_s =0.093~{\rm GeV}$. All of them are taken from the Particle Data Group (PDG)~\cite{Zyla:2020zbs}. The $D_s$, $\eta, \eta'$-meson decay constants are taken as $f_{D_s}=0.274\pm0.013\pm0.007~{\rm GeV}$~\cite{Azizi:2010zj}, $f_\eta=0.130\pm0.003 ~{\rm GeV}$~\cite{Ball:2004ye} and $f_{\eta'}=0.157\pm0.003 ~{\rm GeV}$~\cite{Ali:1998eb}. The values of non-perturbative vacuum condensates up to 6-dimension are taken as follows~\cite{Shifman:1980ui, Colangelo:2000dp, Narison:2014ska},
\begin{align}
&\langle \alpha_s G^2 \rangle = 0.038\pm0.011~{\rm GeV}^4, \nonumber\\
&\langle g_s^3fG^3\rangle = 0.045\pm0.007~{\rm GeV}^6, \nonumber\\
&\langle g_s\bar qq\rangle ^2 = (2.082_{-0.697}^{+0.734})\times 10^{-3} ~{\rm GeV}^6, \nonumber\\
&\langle g_s^2\bar qq\rangle ^2 = (7.420_{-2.483}^{+2.614})\times 10^{-3}~{\rm GeV}^6, \nonumber\\
&\langle q\bar q\rangle = (-2.417_{-0.114}^{+0.227})\times 10^{-2}~{\rm GeV}^3, \nonumber\\
&\langle \bar ss\rangle = (-1.789_{-0.084}^{+0.168})\times 10^{-2}~{\rm GeV}^3, \nonumber\\
&\langle g_ss \bar s\rangle ^2 = (1.541_{ - 0.516}^{ + 0.543})\times{10^{ - 3}}~{\rm GeV}^6, \nonumber\\
&\sum {\langle g_s^2\bar qq\rangle ^2} = (1.891_{ - 0.633}^{ + 0.665})\times{10^{ - 2}}~{\rm GeV}^6.
\end{align}
The quark-gluon mixture condensate $\langle g_s\bar q\sigma TGq\rangle = m_0^2\langle \bar qq\rangle$ with $m_0^2 = 0.80 \pm 0.02{\rm GeV}^2$, which leads to
\begin{align}
\langle g_s\bar s\sigma TGs\rangle = ( - 1.431_{ - 0.076}^{ + 0.139})\times{10^{ - 2}}~{\rm GeV}^5.
\end{align}
Here, the ratio $\kappa = \langle\bar ss\rangle/\langle\bar qq\rangle = 0.74\pm0.03$ has been used. Meanwhile, the typical scale in this paper is $\mu_{\rm IR} = (m_{D_s}^2 - m_c^2)^{1/2} \approx 1.5~{\rm GeV}$. So the renormalization group equations (RGEs) should be used for running the quark masses and each vacuum condensates appearing in the BFTSR from the initial scale $\mu_0 = 1~{\rm GeV}$ to the typical scale $\mu_{\rm IR}$. The RGE can be found in Refs.~\cite{Yang:1993bp, Hwang:1994vp, Lu:2006fr}, which are not listed here.
\begin{table}[t]
\renewcommand\arraystretch{1.2}
\centering
\caption{The criteria for determining Borel windows, and the resultant Borel windows and the corresponding values of the $\eta$ and $\eta'$-meson leading twist LCDA moments $\langle \xi_{2;\eta^{(\prime)}}^n\rangle|_\mu $. ``Con.'' represents the continuum contribution and ``Six.'' represents the dimension-six condensates' contribution.}\label{tab:m2}
\begin{tabular}{c l c c c c c}
\hline
~~~~~~~~~~~~~~~~~~~~~& $n$~~~~~~ &~~~~~~~~$0$~~~~~~~~ & ~~~~~~~~$2$~~~~~~~~ & ~~~~~~~~$4$~~~~~~~~ & $6$ \\ \hline
&Con. & $ < 20\%$ & $ < 25\%$ & $ < 30\%$ & $ < 35\%$ \\
&Six. & $ < 5\%$ & $ < 5\%$ & $ < 5\%$ & $ < 5\%$ \\
\raisebox {2.0ex}[0pt]{$\eta$-meson}&${M^2}$ & $[0.535, 1.188]$ & $[1.026,1.402]$ & $[1.368,1.759]$ & $[1.677,2.194]$ \\
&$\langle {\xi^n_{2;\eta }}\rangle|_\mu$ & $[0.952,1.168]$ & $[0.231,0.230]$ & $[0.110, 0.102]$ & $[0.067, 0.059]$
\\ \hline
&$n$ &~~~~~~~~$0$~~~~~~~~ & ~~~~~~~~$2$~~~~~~~~ & ~~~~~~~~$4$~~~~~~~~ & $6$ \\ \hline
&Con. & $ < 30\%$ & $ < 35\%$ & $ < 40\%$ & $ < 45\%$ \\
&Six. & $ < 5\%$ & $ < 5\%$ & $ < 5\%$ & $ < 5\%$ \\
\raisebox {2.0ex}[0pt]{$\eta'$-meson}&${M^2}$ & $[1.049, 1.137]$ & $[1.026,1.233]$ & $[1.368,1.627]$ & $[1.677,2.082]$ \\
&$\langle {\xi^n_{2;\eta' }}\rangle|_\mu$ & $[1.076,1.061]$ & $[0.221,0.201]$ & $[0.099, 0.086]$ & $[0.059, 0.047]$ \\ \hline
\end{tabular}
\end{table}
\subsection{Determination for the Gegenbauer moments of $\eta^{(\prime)}$-meson twist-2 LCDA}
One of the significant parameters in BFTSR is the continuum threshold $s_{\eta^{(\prime)}}$ for the moments of $\eta^{(\prime)}$-meson twist-2 LCDAs. Following our previous works, we can determine $s_{\eta}= 1.3\pm0.1 ~{\rm GeV}$ and $s_{\eta'}= 0.8\pm0.1 ~{\rm GeV}$ by setting the 0th-order of Gegenbauer moment into 1, i.e. $\langle\xi_{2;\eta}^0\rangle|_\mu=\langle\xi_{2;\eta^{\prime}}^0\rangle|_\mu = 1$. Meanwhile, in order to determine the allowable range of the Borel parameter $M^2$ (i.e. the Borel Window), we adopt the following three criteria
\begin{itemize}
\item The continuum contributions are less than $45\%$ of the total results;
\item The contributions from the dimension-six condensates do not exceed $5\%$;
\item We require the variations of $\langle\xi_{2;\eta^{(\prime)}}^n\rangle|_{\mu}$ within the Borel window to be less than $10\%$.
\end{itemize}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.43\textwidth]{Fig_1_xi2.eps}\includegraphics[width=0.43\textwidth]{Fig_2_xi2.eps}\\
\includegraphics[width=0.43\textwidth]{Fig_1_xi4.eps}\includegraphics[width=0.43\textwidth]{Fig_2_xi4.eps}\\
\includegraphics[width=0.43\textwidth]{Fig_1_xi6.eps}~\includegraphics[width=0.44\textwidth]{Fig_2_xi6.eps}\\
\end{center}
\caption{Contributions from the continuum state and dimension-six condensates for the $\eta^{(\prime)}$-meson leading-twist LCDA moments $\langle {\xi^n_{2;\eta^{(\prime)} }}\rangle|_{\mu}$ versus the Borel parameter $M^2$, where all input parameters are set to be their central values. }
\label{fig:con}
\end{figure}
For the first four Gegenbauer moments $\langle\xi_{2;\eta^{(\prime)}}^n\rangle|_{\mu}$ with $n=(0,2,4,6)$, we list the allowable Borel region and their corresponding $\langle\xi_{2;\eta^{(\prime)}}^n\rangle|_{\mu}$ in Table~\ref{tab:m2}. When $n=(0,2,4,6)$, we have set the continuum contributions to be less than $20\%$, $25\%$, $30\%$, $35\%$ for $\langle\xi_{2;\eta}^n\rangle|_{\mu}$ and $30\%$, $35\%$, $40\%$, $45\%$ for $\langle\xi_{2;\eta'}^n\rangle|_{\mu}$, respectively, and the dimension-six condensates' contributions to be less than $5\%$ for all the order of $\langle\xi_{2;\eta^{(\prime)}}^n\rangle|_{\mu}$. Then, the determined continuum's and dimension-six condensates' contribution for $\langle\xi_{2;\eta}^n\rangle|_{\mu}$ and $\langle\xi_{2;\eta^{\prime}}^n\rangle|_{\mu}$ with $n=(2,4,6)$ are shown in the left and right panel of Fig.~\ref{fig:con}, respectively. In which, the shaded region stand for the Borel windows.
In order to provide a deeper insight into the flatness of the LCDA moments versus the Borel parameter $M^2$, we present the first three curves for the moments of $\eta^{(\prime)}$-meson twist-2 LCDA at the initial scale, i.e. $\langle \xi^n_{2;\eta^{(\prime)}} \rangle|_{\mu}$ with $n=(2,4,6)$ in Figure~\ref{fig:xi}. The determined Borel window are $M^2\in[1.0,2.5]~{\rm GeV}^2$ for $\eta$-meson and $M^2\in[1.0,2.4]~{\rm GeV}^2$ for $\eta'$-meson. It is noted that within this range, the moments $\langle \xi^2_{2;\eta^{(\prime)}} \rangle|_{\mu}$, $\langle \xi^4_{2;\eta^{(\prime)}} \rangle|_{\mu}$ and $\langle \xi^6_{2;\eta^{(\prime)}}\rangle|_{\mu}$ are almost flat, which vary less than $10\%$ for the total results in the Borel window.
By taking the squared average of all the uncertainty sources into consideration and making use of the relations between the Gegenbauer moments $a^n_{2;\eta^{(\prime)}}(\mu)$ and the LCDA moments $\langle\xi^n_{2;\eta^{(\prime)}}\rangle|_\mu$, i.e. Eq.~\eqref{xi}, we obtain the first three $a^n_{2;\eta^{(\prime)}}(\mu_0)$ and $\langle \xi^n_{2;\eta^{(\prime)}}\rangle|_{\mu_0}$ with $n = (2,4,6)$ for the leading-twist $\eta^{(\prime)}$-meson LCDA $\phi_{2;\eta^{(\prime)}}(u,\mu_0)$ predicted from BFTSR in Table~\ref{Tab:anxin}. The factorization scale is taken as the initial scale ${\mu _0} = 1~{\rm GeV}$. As a comparison, we also list the LCSR given in year 2013~\cite{Offen:2013nma}, the CLEO fit~\cite{Gronberg:1997fj}, the BABAR fit~\cite{BABAR:2011ad}, the ones given by Kroll~\cite{Kroll:2013iwa} and Ball~\cite{Ball:2004ye}, respectively. For the Ball's results, it is calculated by using the approximation $a_{2;\eta}^n(\mu) = a_{2;K}^n(\mu) = a_{2;\pi}^n(\mu)$. Our results for the second and fourth order $\eta$-meson LCDA's moments, e.g. $\langle\xi_{2;\eta}^2\rangle|_{\mu_0}$ and $\langle\xi_{2;\eta}^4\rangle|_{\mu_0}$ agree with the Ball's predictions within errors. But there still exist discrepancy for $a_{2;\eta}^4(\mu_0)$ with Ball's prediction. The main reason lies in the fourth order of equations between Gegenbauer and LCDA moments, i.e. Eq.~\eqref{xi} have large coefficients which will enlarge the small discrepancy of LCDA moments, even appears the opposite sign. At present, there are few studies on $\eta'$-meson's twist-2 LCDA.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{Fig_1_xi.eps}\includegraphics[width=0.511\textwidth]{Fig_2_xi.eps}
\caption{The first three moments $\langle\xi^n_{2;\eta^{(\prime)}}\rangle|_{\mu}$ with $(n=2,4,6)$ versus the Borel parameter $M^2$. The darker shaded bands indicate the Borel windows for $\langle\xi^n_{2;\eta^{(\prime)}}\rangle|_{\mu}$, respectively.} \label{fig:xi}
\end{figure}
\begin{table}[tb]
\renewcommand\arraystretch{1.2}
\begin{center}
\caption{First three Gegenbaner and LCDA moments $a^n_{2;\eta^{(\prime)}}(\mu_0)$ and $\langle \xi^n_{2;\eta^{(\prime)}}\rangle|_{\mu_0}$ with $n = (2,4,6)$ for the leading-twist $\eta^{(\prime)}$-meson LCDA $\phi_{2;\eta^{(\prime)}}(u,\mu_0)$, where the errors are squared averages of those from all the input parameters. Other theoretical predictions are also given as a comparison.} \label{Tab:anxin}
\begin{tabular}{lcccccccc}
\hline\hline
~~~~~~~~~~~~~~~~~~~~~~~~& &$\eta$-meson & &~~& & $\eta'$-meson& \\ \cline{2-4} \cline{6-8}
&$\langle \xi^2_{2;\eta}\rangle|_{\mu_0}$
& $\langle \xi^4_{2;\eta}\rangle|_{\mu_0}$
& $\langle \xi^6_{2;\eta}\rangle|_{\mu_0}$
&& $\langle \xi^2_{2;\eta'}\rangle|_{\mu_0}$
& $\langle \xi^4_{2;\eta'}\rangle|_{\mu_0}$
& $\langle \xi^6_{2;\eta'}\rangle|_{\mu_0}$ \\ \hline
BFTSR (This work) & $0.231_{-0.013}^{+0.010}$ & $0.109_{-0.007}^{+0.007}$ & $0.066_{-0.006}^{+0.006}$ && $0.211_{-0.017}^{+0.015}$ & $0.093_{-0.009}^{+0.009}$ & $0.054_{-0.008}^{+0.008}$ \\
CLEO fit~\cite{Gronberg:1997fj} & $0.176\pm0.010$ &- &- &&- &- &- \\
BABAR fit~\cite{BABAR:2011ad} & $0.183\pm0.007 $ &- &- &&- &- &- \\
P. Kroll~\cite{Kroll:2013iwa} & $0.183\pm0.007 $ &- &- &&- &- &- \\
SR fit~\cite{Offen:2013nma} & $0.286\pm0.051$ &- &- &&- &- &-\\
P. Ball~\cite{Ball:2004ye} & 0.239 & 0.110 &- & &- &- &-\\
\hline\hline
~~~~~~~~~~~~~~~~~~~~~~~~& &$\eta$-meson & &~~& & $\eta'$-meson& \\ \cline{2-4} \cline{6-8}
& $a^2_{2;\eta}(\mu_0)$
& $a^4_{2;\eta}(\mu_0)$
& $a^6_{2;\eta}(\mu_0)$
&& $a^2_{2;\eta'}(\mu_0)$
& $a^4_{2;\eta'}(\mu_0)$
& $a^6_{2;\eta'}(\mu_0)$ \\ \hline
BFTSR (This work) & $0.090_{-0.037}^{+0.031}$ & $0.025_{-0.010}^{+0.003}$ & $0.033_{-0.058}^{+0.055}$
&& $0.033_{-0.050}^{+0.042}$ & $-0.002_{-0.016}^{+0.007}$ & $0.043_{-0.072}^{+0.067}$ \\
CLEO fit~\cite{Gronberg:1997fj} & $-0.07\pm0.03$ &-&-&&-&-&- \\
BABAR fit~\cite{BABAR:2011ad} & $-0.05\pm0.02$ &-&-&&-&-&- \\
P. Kroll~\cite{Kroll:2013iwa} & $-0.05\pm0.02 $ &-&-&&-&-&- \\
SR fit~\cite{Offen:2013nma} & $0.25\pm0.15$ &-&-&&-&-&- \\
P. Ball~\cite{Ball:2004ye} & $0.115$ & $-0.015$ &-&&-&-&- \\ \hline\hline
\end{tabular}
\end{center}
\end{table}
After considering the Gegenbauer moments $a_{2;\eta^{(\prime)}}^n(\mu)$ up to 6th-order into the conformal expansion of the Gegenbauer polynomial at initial scale, i.e. Eq.~\eqref{HPDA_CZ}, we present the $\eta$ and $\eta^{\prime}$-meson twist-2 LCDAs in Figure~\ref{fig:fi2}(a) and Figure~\ref{fig:fi2}(b) separately. For $\phi_{2;\eta}(u,\mu_0)$, we present the asymptotic form, the CLEO~\cite{Gronberg:1997fj}, the BABAR~\cite{BABAR:2011ad}, Kroll's prediction~\cite{Kroll:2013iwa}, the SR fit~\cite{Offen:2013nma} and the Ball's prediction~\cite{Ball:2004ye} as a comparison. Figure~\ref{fig:fi2}(a) shows that the LCSR 2013 and Ball's results prefer a double-peaked behavior. The reason lies in that they adopt the $\pi, K$-meson's LCDAs as those of $\eta$-meson LCDAs. Conversely, the CLEO~\cite{Gronberg:1997fj}, the BABAR~\cite{BABAR:2011ad} and the fitting results by Kroll indicate a single-peaked behavior. Our prediction tends to a double-peaked behavior. For $\phi_{2;\eta'}(u,\mu_0)$, we only exhibit the asymptotic form due to there are less results from references, which is shown in Figure~~\ref{fig:fi2}(b). Furthermore, in order to have a look at the evolution $\phi_{2;\eta^{(\prime)}}(u,\mu_0)$ with $n=(2,4,6)$, we present the different curves in Figure~\ref{fig:fi22}. If we take $n=(2,4)$, the behavior of LCDAs shall be closer to the asymptotic form. Furthermore, a small shake is observed when taking the 6th-order LCDA moment into consideration.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{Fig_1_DA.eps}\includegraphics[width=0.5\textwidth]{Fig_2_DA.eps}
\caption{The $\eta^{(\prime)}$-meson leading-twist LCDA $\phi_{2;\eta^{(\prime)}}(u,{\mu}_{0})$ predicted from the BFTSR. We make a comparison with the asymptotic form, the CLEO~\cite{Gronberg:1997fj}, SR fit~\cite{Offen:2013nma}, the BABAR~\cite{BABAR:2011ad}, and the predictions of Kroll~\cite{Kroll:2013iwa} and Ball~\cite{Ball:2004ye}.}
\label{fig:fi2}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{Fig_1_DA2.eps}\includegraphics[width=0.5\textwidth]{Fig_2_DA2.eps}
\caption{The curves of $\eta$ and $\eta^{\prime}$-meson twist-2 LCDA with $n = (2,4,6)$ respectively.}
\label{fig:fi22}
\end{figure}
Other two-particle Fock state twist-3 and twist-4 LCDAs $\phi _{3;\eta^{(\prime)} }^p(u)$, $\phi _{3;\eta^{(\prime)} }^\sigma (u)$, ${\psi _{4;\eta^{(\prime)} }}(u)$ and ${\phi _{4;\eta^{(\prime)} }}(u)$ are defined as follows
\begin{align}
\phi_{3;\eta^{(\prime)}}^p(u) & = 1 + \Big(30\eta_3^{\eta^{(\prime)}} - \frac52\rho_{\eta^{(\prime)}}^2 \Big) C_2^{1/2} (\xi) + \Big(-3\eta_3^{\eta^{(\prime)}} \omega_3^{\eta^{(\prime)}}
-\frac{{27}}{{20}}\rho _{\eta^{(\prime)}} ^2 - \frac{81}{10}\rho_{\eta^{(\prime)}}^2a^2_{2;\eta^{(\prime)}} \Big)
\nonumber\\
&\times C_4^{1/2}(\xi),
\\
\phi_{3;\eta^{(\prime)} }^\sigma (u) & = 6u \bar u\Big(1 + 5\eta _3^{\eta^{(\prime)}} - \frac{1}{2}\eta _3^{\eta^{(\prime)}} \omega _3^{\eta^{(\prime)}} - \frac{7}{{20}}\rho _{\eta^{(\prime)}} ^2
- \frac{3}{5}\rho _{\eta^{(\prime)}} ^2
a^2_{2;\eta^{(\prime)}}\Big)C_2^{3/2}(\xi),
\\
\psi_{4;\eta^{(\prime)}}(u) & = \frac52\varepsilon^2u^2\bar u^2 + \frac12\varepsilon \delta^2 \bigg [u\bar u(2+13u\bar u)+ 10u^3\ln u\ln\bar u (2-3u+\frac65u^2) + 10\,\bar u^3\, (2
\nonumber\\
&- 3\bar u+\frac{6}{5}{\bar u^2})\bigg],
\\
\phi_{4;\eta^{(\prime)}}(u) &= \frac{{10}}{3}{\delta^2}u \bar u(u - \bar u),
\end{align}
where, the values for the twist-3, 4 LCDAs parameters are taken from Refs.~\cite{Ball:2006wn, Huang:2001xb}. In order to run the hadronic parameters of the $\eta^{(\prime)}$-meson twist-2, 3, 4 LCDAs from the initial factorization scale to any other scale, especially for typical scale $\mu_{\rm IR}$, the renormalization group equation should be used, which has the form
\begin{align}
c_i(\mu_{\rm IR}) = {\cal L}^{\gamma_{c_i}/\beta_0}c_i(\mu_0) , \label{Ci}
\end{align}
where ${\cal L}=\alpha_s(\mu_{\rm IR})/\alpha_s (\mu_0)$, $\beta_0 = 11-2/3 n_f$, and the one-loop anomalous dimensions $\gamma_{c_i}$ can be seen in our previous work~\cite{Fu:2020uzy}. Taking the hadronic parameters at initial scale $\mu_0$ and using the renormalization function ~\eqref{Ci}, one can achieve the corresponding values at the typical scale $\mu_{\rm IR}$.
\begin{table}[t]
\renewcommand\arraystretch{1.2}
\centering
\caption{The TFFs $f_+^{\eta^{(\prime)}}(0)$ at the large recoil point $q^2=0$. As a comparison, we also present the predictions from various experimental and theoretical groups.}\label{Tab:fq0}
\begin{tabular}{l l l l}
\hline
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~&References ~~~~~~~~~ ~~~~~~~~& $f_ +^{\eta} (0)$ ~~~~~~~~~~~~~~& $f_ + ^{\eta '}(0)$ \\ \hline
&BESIII~\cite{Ablikim:2019rjz} & $0.4576(70)$ & $0.490(51)$\\
Experimental results &LQCD-I~\cite{Bali:2014pva} & $0.542(13)$ & $0.404(25)$ \\
&LQCD-II~\cite{Bali:2014pva} & $0.564(11)$ & $0.437(18)$ \\
\hline
&This work (LCSR) & $0.476_{-0.036}^{+0.040}$ & $0.544_{-0.042}^{+0.046}$ \\
&LFQM~\cite{Verma:2011yw} & $0.76$ & - \\
&CQM~\cite{Melikhov:2000yu} & $0.78$ & $0.78$ \\
&CCQM~\cite{Soni:2018adu} & $0.78(12)$ & $0.73(11)$ \\
\raisebox {2.0ex}[0pt]{Theoretical predictions}&CCQM~\cite{Ivanov:2019nqd} & $0.49(7)$ & $0.59(9)$ \\
&LCSR~2013~\cite{Offen:2013nma} & $0.432(33)$ & $0.520(80)$ \\
&LCSR~2015~\cite{Duplancic:2015zna} & $0.495_{-0.030}^{+0.029}$ & $0.558_{-0.045}^{+0.047}$\\
&QCD SR~\cite{Colangelo:2001cv} & $0.50(4)$ & - \\\hline
\end{tabular}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=0.52\textwidth]{Fig_fpM2.eps}
\caption{The TFFs $f_+^{\eta^{(\prime)}}(0)$ versus the Borel parameters $M^2$, where the shaded band is induced by the variations of squared average of all input parameters.}
\label{Fig:fqM2}
\end{figure}
\subsection{TFFs and series expansion}
In order to determine the continuum threshold $s_0$ for the $D_s\to\eta^{(\prime)}$ TFFs within LCSR approach, i.e. Eq.~\eqref{Eq:fp}, one can follow the four criteria
\begin{itemize}
\item The continuum contributions are less than $30\%$ of the total results;
\item The contributions from the twist-4 LCDAs do not exceed $5\%$;
\item We require the variations of the TFF within the Borel window be less than $10\%$;
\item The continuum threshold $s_0$ should be closer to the squared mass of the first excited state of $D_s$-meson.
\end{itemize}
Based on the fourth term of the criteria, we take $s_0$ to be close to the squared mass of the excited state of $D_s$-meson $D_{s1}(2460)$, i.e. $s_0 = 6.1(3)~{\rm GeV^2}$. Furthermore, the Borel parameter is taken as $M^2 = 25 (1)~{\rm GeV^2}$. Furthermore, we can obtain the sum rule for $m_{D_s}$ by differentiating the form factors $f_+^{\eta^{(\prime)}}(q^2)f_{D_s}$ with respect to $- 1/M^2$~\cite{Fu:2014pba}. The resultant $m_{D_s}^{\rm LCSR} = 1.9653~{\rm GeV}$ agrees with the measured value $m_{D_s}^{\rm PDG} = 1.9685~{\rm GeV}$. In order to show the degree of stability of the TFFs versus the Borel parameter, we present the curve of TFFs in Figure~\ref{Fig:fqM2}, in which the shaded region shows the errors from all input parameters. The solid line with blue shaded band represents $f_{+}^{\eta}(0)$ and the dashed line with green band represents $f_{+}^{\eta'}(0)$. This figure shows that $f_+^{\eta^{(\prime)}}(0)$ changes less than $0.5\%$ within the range of $M^2\in [23,27]~{\rm GeV^2}$, which satisfies the third term of the criteria. More definitely, we put the errors caused by different input parameters in the following,
\begin{align}
f_+^\eta(0) & = 0.476+ (_{-0.012}^{+0.011})_{s_0} + (_{-0.001}^{+0.001})_{M^2} + (_{-0.032}^{+0.036})_{m_c,f_{D_s}} + (_{-0.011}^{+0.011})_{f_\eta} + (_{-0.003}^{+0.004})_{a_{2;\eta}^2}
\nonumber\\
& + (_{-0.000}^{+0.000})_{a_{2;\eta }^4}+ (_{-0.000}^{+0.000})_{a_{2;\eta }^6}
\nonumber\\
& = 0.476_{-0.036}^{+0.040}
\\
\nonumber \\
f_+^{\eta'}(0)&= 0.544+(_{-0.016}^{+0.015})_{s_0} + (_{-0.001}^{+0.001})_{M^2} + (_{-0.037}^{+0.042})_{m_c,f_{D_s}} + (_{-0.010}^{+0.010})_{f_{\eta'}} + (_{-0.005}^{+0.006})_{a_{2;\eta'}^2}
\nonumber\\
&+(_{-0.001}^{+0.000})_{a_{2;\eta'}^4} + (_{-0.000}^{+0.000})_{a_{2;\eta '}^6}
\nonumber\\
& = 0.544_{-0.042}^{+0.046}
\end{align}
According to Eq.~\eqref{HPDA_CZ}, when the Gegenbauer moments are taken up to 2nd, 4th and 6th order levels, the central values of TFFs $f_{+}^{\eta^{(\prime)}}(0)$ are
\begin{align}
&f_+^{\eta}(0)|_{\rm 2nd} = 0.4745, && f_+^{\eta}(0)|_{\rm 4th} = 0.4755, &&f_+^{\eta}(0)|_{\rm 6th} = 0.4763, \nonumber\\
&f_+^{\eta'}(0)|_{\rm 2nd} = 0.5435, &&f_+^{\eta'}(0)|_{\rm 4th} = 0.5434, &&f_+^{\eta'}(0)|_{\rm 6th} = 0.5436. \nonumber
\end{align}
In comparing with the 2nd-order Gegenbauer moment's contribution, the 4th, 6th-order contributions shall be changed by about 0.211\%, 0.379\% for $f_+^{\eta}(0)$ and $-0.018\%$, 0.018\% for $f_+^{\eta'}(0)$, respectively. These ratios are really small, indicating the Gegenbauer series has good convergence over the moment expansion. Then, we list the TFFs for $D_s\to\eta^{(\prime)}$ at large recoil point, i.e. $f_+^{\eta^{(\prime)}}(0)$, in Table~\ref{Tab:fq0}, in which the uncertainties are from the squared average of all the mentioned error sources. As a comparison, we also present other theoretical and experimental predictions, such as the BESIII~\cite{Ablikim:2019rjz}, the LQCD~\cite{Bali:2014pva}, the LFQM~\cite{Verma:2011yw}, the CQM~\cite{Melikhov:2000yu}, the CCQM~\cite{Soni:2018adu, Ivanov:2019nqd}, the QCD SR~\cite{Colangelo:2001cv}, the LCSR at 2013 and 2015~\cite{Offen:2013nma, Duplancic:2015zna}, respectively. Our results agree with the BESIII, the CCQM, the LCSR, the LQCD within errors, but are lack of agreement with the Lattice QCD results.
The physically allowable ranges for the TFFs are $m_\ell^2 \le q^2\le (m_{D_s}-m_{\eta})^ 2 \approx 2~{\rm GeV^2}$ and $m_\ell^2 \le q^2\le (m_{D_s}-m_{\eta'})^ 2 \approx 1~{\rm GeV^2}$. Theoretically, the LCSRs approach for $D_s\to\eta^{(\prime)}$ TFFs are applicable in low and intermediate $q^2$-regions, i.e. $q^2\in[0,1.2]~{\rm GeV^2}$ of $\eta$-meson, $q^2\in[0,0.6]~{\rm GeV^2}$ of $\eta'$-meson. One can extrapolate it to whole $q^2$-regions via a rapidly $z(q^2, t)$ converging series expansion (SE)~\cite{Bharucha:2010im}:
\begin{align}
f^{\eta^{(\prime)}}_+(q^2)= \frac1{B(q^2)\phi_+^{\eta^{(\prime)}}(q^2)}\sum\limits_k^{K-1} \alpha_k z^k(q^2,t_0).
\end{align}
\begin{table}[t]
\centering
\renewcommand\arraystretch{1}
\caption{The fitting parameters $z(q^2,t_0^{\eta^{(\prime)}})$ of SSE for TFFs $f_ + ^{{\eta^{(\prime)}}}(q^2)$.} \label{Tab:zTFFs}
\begin{tabular}{c ccc cccc}
\hline
~~~~~~~~~~~~~~~~~~~&$q^2({\rm GeV^2})$ & 0 & 0.4 & 0.8 & 1.2 & 1.6 & 2.0 \\
\raisebox {2.0ex}[0pt]{$\eta$-meson} & $z(q^2,t_0^\eta)$ & $0.048$ & $0.032$ & $0.014$ & $-0.005$ & $-0.025$ & $-0.047$
\\ \hline
& $q^2({\rm GeV^2})$ & 0 & 0.2 & 0.4 & 0.6 & 0.8 & 1.0 \\
\raisebox {2.0ex}[0pt]{$\eta'$-meson} & $z(q^2,t_0^{\eta'})$ & $0.016$ & $0.004$ & $-0.009$ & $-0.022$ & $-0.036$ & $-0.051$ \\ \hline
\end{tabular}
\end{table}
Normally, the parameter $K$ stands for the order of expansion. The $z(q^2, t)$ is the function:
\begin{align}
z(q^2, t)=\frac{\sqrt{t_+ - q^2} - \sqrt{t_+ -t}} {\sqrt{t_+ - q^2} +\sqrt{t_+ - t}},
\end{align}
where $t = (t_0^{\eta^{(\prime)}}, t_-^{\eta^{(\prime)}}, m_{D_s})$ with $t_\pm^{\eta^{(\prime)}} =(m_{D_s}\pm m_{\eta^{(\prime)}})^2$. Here, the $0\leq t_0^{\eta^{(\prime)}} \leq t_-^{\eta^{(\prime)}}$ is a free parameter which can be optimised to reduce the maximum value of $|z(q^2,t_0^{\eta^{(\prime)}})|$ in the physical TFFs range, $t_0^{\eta^{(\prime)}}|_{\rm opt.} = t_+^{\eta^{(\prime)}} (1-\sqrt{1-t_-^{\eta^{(\prime)}}/t_+^{\eta^{(\prime)}}})$. Here, we list the value of $z(q^2, t_0^{\eta^{(\prime)}})$ with different $q^2$ cases in Table~\ref{Tab:zTFFs}. The function $\phi_+^{\eta^{(\prime)}}(q^2)$ can be expressed as
\begin{eqnarray}
\phi_+^{\eta^{(\prime)}}(q^2) &=& \sqrt{\dfrac{\varsigma}{48\pi\chi_+^{\eta^{(\prime)}}(n)}} \dfrac{q^2 - t_+}{(t_+ - t_0)^{1/4}} \bigg(\dfrac{z(q^2, 0)}{-q^2}\bigg)^{(3+n)/2} \bigg(\dfrac{z(q^2, t_0)}{t_0 -q^2}\bigg)^{-1/2} \nonumber\\
&\times& \bigg(\dfrac{z(q^2, t_-)}{t_- -q^2}\bigg)^{-3/4}\bigg|_{n=2},
\end{eqnarray}
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{Fig_1_fp2SSE.eps}\includegraphics[width=0.5\textwidth]{Fig_2_fp2SSE.eps}
\caption{The SSE for the TFFs $f^\eta_+(q^2)$ and $f^{\eta^{\prime}}_+(q^2)$ up to $K=3,4,5$ order.}
\label{Fig:TFFs_K345}
\end{figure}
where $\varsigma = 1 $ is an isospin-degeneracy factor for a given channel and $B(q^2) = \prod\limits_i z(q^2, m_{D_s^i}^2)$ stands for the Blaschke factor. The $m_{D_s^i}$ stands for the mass of each resonance state with $J^P = 0^-, 0^+, 1^-...$, respectively, which can be found in PDG~\cite{Zyla:2020zbs}. For the coefficients $\chi_+^{\eta^{(\prime)}}(n)$, it can be calculated by using QCD sum rules including perturbative LO and NLO results as well as the condensate contributions, which can be expressed as~\cite{Bharucha:2010im}
\begin{eqnarray}
\chi_+^{\eta^{(\prime)}} = \frac{3}{32\pi^2 m_c^2} \bigg[1+\frac{\alpha_s(m_c)C_F}{4\pi}\bigg(\frac{25}{6}+\frac{2\pi^2}{3} \bigg) \bigg]- \frac{\langle q\bar q\rangle}{m_c^5} - \frac{\langle \alpha_s G^2\rangle}{12\pi m_c^6} - \frac{\langle \bar qGq\rangle}{m_c^7}.
\end{eqnarray}
Furthermore, the coefficients $\alpha_k$ should satisfy the basic unitarity constraint,
\begin{eqnarray}
\sum\limits_k^{K-1} \alpha_k^2 <1
\end{eqnarray}
The simplified version of the series expansion (SSE) method is to replace the Blaschke factor $B(q^2)$ by a simple pole $P(q^2)$ to account for low-lying resonances, i.e.~\cite{Bourrely:2008za}
\begin{align}
f_+^{\eta^{(\prime)}}(q^2) =\frac1{P(t)}\sum\limits_k^{K-1} \beta_k z^k(q^2,t_0).
\end{align}
As for the SSE parameterization, imposing the unitarity bound and by comparing the SE and SSE parameterizations
\begin{align}
\alpha_i= \sum\limits_{k=0}^{\min [K-1,i]} \zeta_{i-k}\beta_k,~~~~~~~0\leq i \leq K-1, \label{Eq:abij}
\end{align}
we obtain the unitarity bound of the SSE
\begin{align}
\sum\limits_{j,k=0}^{\min [K-1]} C_{jk}\beta_j \beta_k\leq 1,
\end{align}
with the positive defined matrix
\begin{align}
C_{jk} = \sum\limits_{i=0}^{K-1-\max [j,k]} \zeta_i\zeta_{j-k}.
\end{align}
\begin{table}[t]
\renewcommand\arraystretch{1.2}
\caption{The fitting parameters for the central TFFs $f_ + ^{{\eta^{(\prime)}}(C)}( q^2)$, the upper TFFs $f_+^{{\eta^{(\prime)}}(U)}( q^2)$ and the lower TFFs $f_+^{{\eta^{(\prime)}}(L)}( q^2)$.} \label{Tab:fit}
\begin{tabular}{l l l l l l l }
\hline
& $f_+^{\eta(C)}(q^2)$ ~& $f_+^{\eta(U)}(q^2)$ & $f_+^{\eta(L)}(q^2)$ & $f_+^{\eta^{\prime}(C)}(q^2)$ & $f_+^{\eta^{\prime}(U)}( q^2)$ & $f_+^{\eta^{\prime}(L)}( q^2)$ \\ \hline
$\alpha_0$ & $-0.0003$ & $-0.0003$ & $-0.0003$ & $-0.00001$ & $-0.00001$ & $-0.00001$
\\
$\alpha_1$ & $-0.0015$ & $-0.0017$ & $-0.0015$ & $-0.00017$ & $-0.00018$ & $-0.00016$
\\
$\alpha_2$ & $-0.0061$ & $-0.0074$ & $-0.0052$ & $-0.00134$ & $-0.00143$ & $-0.00127$
\\
$\displaystyle\sum{\alpha_k^2}$ & $3.9 \times 10^{-5}$ & $5.8 \times 10^{-5}$ & $2.9 \times 10^{-5}$ & $1.8 \times 10^{-6}$ & $2.4 \times 10^{-6}$ & $1.6 \times 10^{-6}$
\\
\hline
$\beta_0$ & $0.512$ & $0.555$ & $0.476$ & $0.563$ & $0.611$ & $0.519$
\\
$\beta_1$ & $-1.450$ & $-1.211$ & $-0.976$ & $-1.653$ & $-1.801$ & $-1.534$
\\
$\beta_2$ & $17.26$ & $12.24$ & $8.73$ & $28.95$ & $30.61$ & $27.98$
\\
$\displaystyle\sum C_{i,j}\beta_i \beta_j$ & $0.020$ & $0.019$ & $0.008$ & $0.005$ & $0.005$ & $0.004$
\\
$\Delta$ & $0.075\%$ & $0.129\%$ & $0.099\%$ & $0.005\%$ & $0.004\%$ & $0.007\%$ \\ \hline
\end{tabular}
\end{table}
Theoretically, the order $K$ can be taken up to infinite order. It has been proven that the higher order expansion shall give the same result with very small errors. We take $K=3,4,5$ as explicit examples, which are shown in Figure~\ref{Fig:TFFs_K345}. It is found that three SSE curves with $K=3,4,5$ for the TFFs $f_+^{\eta^{(\prime)}}(q^2)$ are almost coincide with each other. So, we will take $K=3$ to do our expansion in the following calculations, which also agrees with the choices of most of the theoretical groups. At the same time, the $\alpha_k$ of SE and $\beta_k$ of SSE should also satisfy the condition $\Delta<1\%$. From which, the parameter $\Delta$ is used to measure the quality of extrapolation, which is defined as
\begin{align}
\Delta = \dfrac{\sum_t|F_i(t) - F_i^{\rm fit}(t)|}{\sum_t|F_i(t)|} \times 100, \label{Eq:fit}
\end{align}
where $t\in[0,1/100,\cdots,100/100]\times 1.2~{\rm GeV}$ for $\eta$-meson, $t\in[0,1/100,\cdots,100/100]\times 0.6~{\rm GeV}$ for $\eta'$-meson. Numerical values of $\alpha_k$, $\beta_k$ with $k=(0,1,2)$ for central values, upper and lower limit of $f_+^{\eta^{(\prime)}}(q^2)$ are listed in Table~\ref{Tab:fit}. Here, the $\sum \alpha_k^2$ for SE unitarity bound and $\displaystyle\sum C_{i,j}\beta_i \beta_j$ for SSE are less than 1. The quality of extrapolation $\Delta$ are less than 0.13\%. Since there have good unitary bound and a small $\Delta$ value, the SSE results are in high agreement with our LCSR results.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{Fig_1_fp.eps}\includegraphics[width=0.5\textwidth]{Fig_2_fp.eps}
\caption{The TFFs $f^{\eta^{(\prime)}}_+(q^2)$ together with its uncertainties. Results of the LCSR 2013~\cite{Offen:2013nma}, the LCSR 2015~\cite{Duplancic:2015zna}, the CCQM~\cite{Ivanov:2019nqd} and the BESIII~\cite{Ablikim:2019rjz} are also given as the comparison. }
\label{Fig:fq}
\end{figure}
The extrapolated TFFs in the whole $q^2$-region is shown in Figure~\ref{Fig:fq}, other theoretical and experimental results, such as those of the BESIII~\cite{Ablikim:2019rjz}, the LCSR 2013~\cite{Offen:2013nma}, the LCSR 2015~\cite{Duplancic:2015zna} and the CCQM~\cite{Ivanov:2019nqd} are present as a comparison. Two sets of BESIII are from the two different $\eta^{(\prime)}$ decay channels. For $\eta$-meson, the blue triangle stands for $\eta\to\gamma\gamma$ channel and the red diamond stands for $\eta\to\pi^0\pi^+\pi^-$ channel. For $\eta'$-meson the blue triangle stands for $\eta'\to\gamma\rho^0$ channel and red diamond stands for $\eta'\to\eta_{\gamma\gamma}\pi^+\pi^-$ channel. To compare with other theoretical and experimental groups, our results have the following characteristics,
\begin{itemize}
\item Comparing with $f^\eta_+(q^2)$, the $f^{\eta'}_+(q^2)$ is more flat in the whole $q^2$ region.
\item Our predictions of $f^\eta_+(q^2)$ are in good agreement with the recent BESIII predictions for the $\eta\to\gamma\gamma$ channel.
\item In the LCSR $q^2$-region, our results have good agreement with the LCSR 2013 and 2015, the CCQM, and the two sets of BES-III predictions within errors.
\item The SSE of $f^\eta_+(q^2)$ in the region of $q^2\in [1.2, 2.0]~{\rm GeV}^2$ have agreement with the LCSR in 2015 results within errors. However, our predictions are larger than the LCSR in 2013 and CCQM predictions due to the different $\eta$-meson distribution amplitudes or different method.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{Fig_1_width.eps}\includegraphics[width=0.5\textwidth]{Fig_2_width.eps}
\caption{Decay width for the $D_s^+\to \eta^{(\prime)}\ell ^+\nu_\ell (\ell =e,\mu )$ versus $q^2$ within uncertainties. The LCSR in 2013~\cite{Offen:2013nma} and 2015~\cite{Duplancic:2015zna}, the CCQM~\cite{Ivanov:2019nqd}, and two sets of BESIII collaboration~\cite{Ablikim:2019rjz} predictions are also present as a comparison.}
\label{Fig:dGamma}
\end{figure}
\subsection{Decay widthes and branching fractions for the semileptonic decay $D_s^+ \to \eta^{(\prime)} \ell^+ \nu_\ell$}
One can get the differential decay widths by using the formula ~\eqref{Eq:dGamma}. For the CKM matrix element $|V_{cs}|$, we take it to be the average value of leptonic and semileptonic decay $c\to s$ processes coming from PDG~\cite{Narison:2014ska}, i.e. $|V_{cs}| = 0.987\pm0.011$. After taking the derived $D_s\to{\eta^{(\prime)}}$ TFFs into the decay widths, we present the differential decay widths in Figure~\ref{Fig:dGamma}. As a comparison, we also give the LCSR in 2013~\cite{Offen:2013nma} and 2015~\cite{Duplancic:2015zna}, the CCQM~\cite{Ivanov:2019nqd}, and two sets of BESIII collaboration~\cite{Ablikim:2019rjz} predictions. The LCSR and the CCQM results are calculated by applying their TFFs into the width formula. Those figures show that our prediction for $D_s^+\to\eta\ell^+\nu_\ell$ is in agreement with LCSR 2015, CCQM, the BESIII-I and the BESIII-II results within errors, and the $D_s^+\to\eta'\ell^+\nu_\ell$ agrees with the LCSR 2013 and 2015, the CCQM, the BESIII-I and BESIII-II predictions within errors. All the results are convergence to zero at the small recoil region $q^2_{\max} = (m_{D_s}-m_{\eta^{(\prime)}})^2$, which indicates that our results are reasonable.
\begin{table}[t]
\renewcommand\arraystretch{1.2}
\centering
\caption{Branching factions of $D_s^+\to \eta^{(\prime)} \ell^+ \nu_\ell$ with $\ell =e$ and $\mu$ (in unit $10^{-2}$). The errors are squared averages of all the mentioned error sources. As a comparison, we also present the predictions for various methods.}\label{tab:BF}
\begin{tabular}{l l l l l l}
\hline\hline
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ &Mode~~~~~~~~~~~~~~~~~~~~~~~~~~ & ${\cal B}(D_s^+ \to\eta e^+\nu_e)$~~~~~~~~~~~~~~~~ & ${\cal B}(D_s^ + \to\eta \mu^+ \nu_\mu)$
\\\hline
&BESIII~\cite{Ablikim:2019rjz,Ablikim:2017omq} & $2.323\pm 0.063\pm 0.063$ & $2.42\pm 0.46\pm 0.11$
\\
&CLEO~\cite{Hietala:2015jqa} & $2.28\pm 0.14\pm 0.19$ &- \\
\raisebox {2.0ex}[0pt]{Experimental results}&CLEO~\cite{Yelton:2009aa} & $2.48\pm 0.29\pm 0.13$ &- \\
&PDG~\cite{Zyla:2020zbs} & $2.32\pm 0.08$ & $2.4\pm 0.5$
\\
\hline
&This work (LCSR) & $2.346_{-0.331}^{+0.418}$ & $2.320_{-0.327}^{+0.413}$ \\
&LFQM~\cite{Cheng:2017pcq} & $2.26\pm 0.21$ & $2.22\pm 0.20$ \\
&CCQM~\cite{Ivanov:2019nqd} & $2.24$ & $2.18$ \\
Theoretical predictions &LCSR~\cite{Offen:2013nma} & $2.00\pm 0.32$ &- \\
&LCSR~\cite{Duplancic:2015zna} & $2.40\pm 0.28$ &- \\
&QCD SR-I~\cite{Colangelo:2001cv} & $2.6\pm 0.7$ &- \\
&QCD SR-II~\cite{Colangelo:2001cv} & $2.3\pm 0.4$ &- \\
\hline\hline
&Mode & ${\cal B}(D_s^+ \to \eta' e^+\nu_e)$ & ${\cal B}(D_s^ + \to\eta'\mu^+\nu_\mu)$\\ \hline
&BESIII~\cite{Ablikim:2019rjz,Ablikim:2017omq} & $0.824\pm 0.073\pm 0.027$ & $1.06\pm 0.54\pm 0.07$
\\
&CLEO~\cite{Hietala:2015jqa} & $0.68\pm 0.15\pm 0.06$ &- \\
\raisebox {2.0ex}[0pt]{Experimental results}&CLEO~\cite{Yelton:2009aa} & $0.91\pm 0.33\pm 0.05$ &- \\
&PDG~\cite{Zyla:2020zbs} & $0.80\pm 0.07$ & $1.1\pm 0.5$
\\
\hline
&This work (LCSR) & $0.792_{-0.118}^{+0.141}$ & $0.773_{-0.115}^{+0.138}$
\\
&LFQM~\cite{Cheng:2017pcq} & $0.89\pm 0.09$ & $0.85\pm 0.08$
\\
&CCQM~\cite{Ivanov:2019nqd} & $0.83$ & $0.79$ \\
Theoretical predictions &LCSR~\cite{Offen:2013nma} & $0.75\pm 0.23$ &- \\
&LCSR~\cite{Duplancic:2015zna} & $0.79\pm 0.14$ &- \\
&QCD SR-I~\cite{Colangelo:2001cv} & $0.89\pm 0.34$ &- \\
&QCD SR-II~\cite{Colangelo:2001cv} & $1.0\pm 0.2$ &- \\
\hline\hline
\end{tabular}
\end{table}
After integrating over the whole $m_\ell^2\leq q^2\leq (m_{D_s}-m_{\eta^{(\prime)}})^2$ region for the differential decay widths, we obtain the total decay widths for $D_s^+ \to \eta^{(\prime)} \ell^+ \nu_\ell$ with two different channel $\Gamma(D_s^+\to{\eta^{(\prime)}} e^+\nu_e)$ and $\Gamma(D_s^+\to{\eta^{(\prime)}}\mu^+\nu_\mu)$, i.e.,
\begin{eqnarray}
&& \Gamma (D_s^{+}\to \eta e^+ \nu_e) = 30.634_{-4.323}^{+5.453} \times 10^{-15}~{\rm GeV} , \nonumber\\
&& \Gamma (D_s^{+}\to\eta\mu^+ \nu_\mu) = 30.298_{-4.275}^{+5.395} \times 10^{-15}~{\rm GeV} , \nonumber\\
&& \Gamma (D_s^{+}\to \eta' e^+ \nu_e) = 10.345_{-1.534}^{+1.844} \times 10^{-15}~{\rm GeV} , \nonumber\\
&& \Gamma (D_s^{+}\to\eta'\mu^+ \nu_\mu) = 10.096_{-1.498}^{+1.800} \times 10^{-15}~{\rm GeV} . \label{Eq:GammaNumerical}
\end{eqnarray}
Then, by using the lifetime of the initial state $D_s^+$-meson, $\tau_{D_s^+}=(0.504\pm 0.007)~ {\rm ps}$~\cite{Zyla:2020zbs}, the branching fractions for the two different semileptonic decay channels $D_s\to{\eta^{(\prime)}}{{\ell}^{+}}\nu_\ell$ with $\ell = (e,\mu)$ can be obtained, which are presented in Table~\ref{tab:BF}. Here, we also listed the BESIII~\cite{Ablikim:2019rjz,Ablikim:2017omq}, the PDG~\cite{Zyla:2020zbs}, the CLEO~\cite{Yelton:2009aa} for the experimental results, and the CCQM~\cite{Ivanov:2019nqd}, the LFQM~\cite{Cheng:2017pcq}, the QCDSR-I, II~\cite{Colangelo:2001cv}, the LCSR~\cite{Duplancic:2015zna, Offen:2013nma} for theoretical predictions. Our results are closer to the BESIII, PDG, CLEO results, all of which are within $1\sigma$ uncertainties.
\begin{figure}[t]
\centering
\includegraphics[width=0.52\textwidth]{Fig_MixingAngle.eps}
\caption{$f^{\eta}_{+}(0)/f^{\eta'}_{+}(0)$ as a function of the Borel parameter $M^2$, where the shaded band is induced by the variations of squared average of all input parameters}
\label{Fig:BT}
\end{figure}
\begin{table}[t]
\renewcommand\arraystretch{1.2}
\centering
\caption{The $\varphi$ with respect to mixing angle, ${\cal R}_{\eta'/\eta}^\ell$ for different models and experimental values. As a comparison, we also present the experimental and theoretical predictions.}\label{tab:RDs}
\begin{tabular}{l l l}
\hline
Mode~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~& Angle $\varphi$ ~~~~~~~~~~~~~~~~~~& ${\cal R}_{\eta'/\eta}^\ell$ \\ \hline
This work~(LCSR) with $\ell = e$ & ${41.2^ \circ }_{ - 0.06}^{ + 0.05}$ & $0.338_{ - 0.051}^{ + 0.057}$ \\
This work~(LCSR) with $\ell = \mu$ & ${41.2^ \circ }_{- 0.06}^{ + 0.05}$ & $0.333_{-0.058}^{+0.050}$ \\
CLEO~\cite{Brandenburg:1995qq} &- & $0.35\pm 0.09\pm 0.07$ \\
BESIII~\cite{Ablikim:2016rqq} &- & $0.40\pm 0.14\pm 0.02$ \\
LFQM~\cite{Wei:2009nc} & ${39}^\circ$ & $0.39$ \\
LCSR~\cite{Duplancic:2015zna} & ${41.8}^\circ$ & $0.33\pm 0.07$ \\
QCD SR~\cite{Colangelo:2001cv} & ${40}^\circ$ & $0.44\pm 0.01$ \\
KLOE~\cite{Ambrosino:2006gk} & ${41.4}^\circ$ &- \\
LCSR-I~\cite{Azizi:2010zj} & ${39.7}^\circ$ & $0.32\pm 0.02$ \\
LCSR-II~\cite{Azizi:2010zj} & ${41.5}^\circ$ & $0.27\pm 0.01$ \\
\hline
\end{tabular}
\end{table}
After substituting the corresponding terms of Eq.~\eqref{bt}, one can get the mixing angle of $\eta-\eta'$. The mixing angle $\tan \varphi$ with the parameter of Borel parameter $M^2$ are shown in Figure~\ref{Fig:BT}. In the whole Borel parameter $M^2$ region, the mixing angle is changed slightly, which also indicates that there have a stable Borel window for $\tan\varphi$. Numerical results for the single mixing angle $\varphi$ are present in Table~\ref{tab:RDs}. Our result is closer to the KLOE~\cite{Ambrosino:2006gk} and the LCSR predictions~\cite{Duplancic:2015zna, Azizi:2010zj}, which is slightly larger than other LFQM~\cite{Wei:2009nc} and QCD SR~\cite{Colangelo:2001cv} predictions.
Furthermore, it is useful to study the ratio for the different decay channel ${\cal R}_{\eta'/\eta}^\ell$ related to the mixing angle, which has the basic definition
\begin{eqnarray}
{\cal R}_{\eta'/\eta}^\ell = \frac{{\cal B}(D_s \to \eta'\ell^+ \nu_\ell)}{{\cal B}(D_s \to \eta \ell^+ \nu_\ell)}.
\end{eqnarray}
Numerical results together with different experimental and theoretical predictions are given in Table~\ref{tab:RDs}. Our results are in agreement with the CLEO~\cite{Brandenburg:1995qq} and the BESIII~\cite{Ablikim:2016rqq} predictions, and the LCSR predictions~\cite{Duplancic:2015zna,Azizi:2010zj}. This can be considered as a good test of the correctness of the considered internal structure for the $D_s$-meson as well as the mixing angle between $\eta$ and $\eta'$ states.
\section{Summary}\label{sec:summary}
In this paper, we have calculated the moments $\langle\xi_{2;\eta^{(\prime)}}^n\rangle|_\mu$ of $\eta^{(\prime)}$-meson LCDA with $n=(2,4,6)$ up to NLO correction and completely dimension-six condensates within BFTSR, which are shown in Eq.~\eqref{xi2}. Due to the $\eta^{(\prime)}$-meson should be considered as $s\bar s$ component in $D_s\to\eta^{(\prime)}$ decay processes, we have also taken the $I_{m_s^2}(n,M^2)$ corrections into consideration, i.e. Eq.~\eqref{ms}, the detailed terms of OPE are listed in the Appendix~\ref{sec:appendixB}.
Then, we have sought a reasonable continuum threshold $s_0$ and stable Borel windows for $\langle\xi_{2;\eta^{(\prime)}}^n\rangle|_\mu$ with $n=(2,4,6)$ by using the traditional three criteria for the SVZ sum rules, which are present in Figs.~\ref{fig:con}, \ref{fig:xi} and Table~\ref{tab:m2}. By using the expression between two different Gegenbauer and LCDA moments, i.e. Eq.~\eqref{xi}, we have presented the first three Gegenbaner and LCDA moments $a^n_{2;\eta^{(\prime)}}(\mu_0)$ and $\langle \xi^n_{2;\eta^{(\prime)}}\rangle|_{\mu_0}$ with $n = (2,4,6)$ for the $\phi_{2;\eta^{(\prime)}}(u,\mu_0)$ within errors in Table~\ref{Tab:anxin}. Our results are in agreement with the CLEO fit and the BABAR fit predictions. Meanwhile, we have exhibited the curves of $\phi_{2;\eta^{(\prime)}}(u,\mu_0)$ of our prediction with $n=(2,4,6)$ and compare with others.
Furthermore, the TFFs $f^{\eta^{(\prime)}}_+(q^2)$ have been given in Eq.~\eqref{Eq:fp} up to NLO QCD corrections for twist-2, 3 LCDA contributions. The TFFs at large recoil region have been presented in Table~\ref{Tab:fq0} with respect to other theoretical predictions. After extrapolating it to the whole physical $q^2$-region via simplified series expansion, we have shown the behavior of TFFs in Figure~\ref{Fig:fq}. The differential/total decay widths and branching fractions in this work have also been given. Our results are in agreement with the BESIII and the PDG average value within errors. Finally, we have presented the mixing angle $\tan\varphi$ and ratio for different decay channel ${\cal R}_{\eta'/\eta}^\ell$, which agree with theoretical and experimental results within errors. Thus, the QCDSR within BFTSR can be considered a good approach in dealing with the heavy-to-light semileptonic processes, and we hope more data can be achieved in the near future for more precise studies.
\section{Acknowledgments}
We are grateful to Dr. Xu-Chang Zheng for helpful discussions and valuable suggestions. Hai-Bing Fu would like to thank the Institute of Theoretical Physics in Chongqing University (CQUITP) for kind hospitality. This work was supported in part by the National Natural Science Foundation of China under Grant No.11765007, No.11947406, No.11625520, and No.12047564, the Project of Guizhou Provincial Department of Science and Technology under Grant No.KY[2019]1171, and No.ZK[2021]024, the Project of Guizhou Provincial Department of Education under Grant No.KY[2021]030 and No.KY[2021]003, the China Postdoctoral Science Foundation under Grant Nos.2019TQ0329, 2020M670476, the Chongqing Graduate Research and Innovation Foundation under Grant No.ydstd1912, the Fundamental Research Funds for the Central Universities under Grant No.2020CQJQY-Z003, and the Project of Guizhou Minzu University under Grant No. GZMU[2019]YB19.
|
\section{Introduction} \label{intro_sec}
The rapid growth of metropolitan population and car ownership has aggravated the problem of traffic congestion and air pollution. With the ability of serving mass travel demands with a relatively low cost, bus transit is regarded as an effective approach to relieve the negative impacts of the enormous private car usage (Levinson et al., 2003; Deng and Nelson, 2011). To improve the attractiveness and encourage the preference choice of buses compared to private cars, a variety of strategies for granting the bus system's quality of service have been proposed and implemented. The crux for guaranteeing the service quality is to alleviate the travel delay of buses, particularly at intersection, thereby enhancing their on-time performance and schedule adherence.\par
One prevailing strategy is the transit signal priority (TSP), which provides a priority right-of-way (ROW) for buses to pass through a signalized intersection (Smith et al., 2005). TSP can be roughly divided into three main categories, i.e., passive, active, and adaptive (Baker et al., 2002). Passive TSP reduces the intersection delay of buses by cycle length reduction or phase split strategy, and operates whether or not a bus is present; thus, it inevitably induces continuous waste of road capacity along the arterial road and is more suitable for application to the scenario where the buses operate frequently and predictably (Urbanik et al., 1977). Active TSP, such as green extension and red truncation, is triggered only when a signal priority request is generated (Baker et al., 2002); it shows promising benefits for the transit service when the traffic demand is not high, but it potentially causes significant disruptions to the regular traffic; thus, the absolute personal travel time savings of buses may be compromised or even
neutralized by the extra delays of regular vehicles (Garrow and Machemehl, 1999; Sunkari et al., 1995; Balke et al., 2000). By contrast, adaptive TSP considers both the service quality of buses and the impact on the regular traffic (Dion and Hellinga, 2002; Liao and Davis, 2007; Hu et al., 2014; Chang et al.,1996; Li et al., 2011). It requires the acquisition of real-time information of the approaching buses through automatic vehicle location systems or connected vehicle technology, followed by concise prediction and continuous updating of their arrival times. Personal delay is proposed to normalize the bus delay into passengers' delay and is applied as a performance index to balance the mobility enhancement of buses and regular vehicles, minimizing the efficiency damage to the regular traffic under the premise of bus priority fulfillment. However, these studies have mainly focused on the design of adaptive TSP at an isolated intersection, where the advanced green time of buses at the upstream intersection may be compensated by the extra delay at the downstream intersection. To guarantee the effectiveness of the priority treatments at intersections, coordinated TSP strategy is then proposed to enable the cooperation of the signal controls between adjacent intersections (He et al., 2012; He et al., 2014; Christofa et al., 2016; Ma et al., 2013; Hu et al., 2015). \par
To further enhance the performance of bus service, the TSP strategy is always bound to the implementation of dedicated bus lanes (DBLs). DBL provides an exclusive ROW for buses along their traveling, and the interference from regular traffic and the mobility damage caused by the traffic flow, especially in congestion situations, can be alleviated ated (Levinson et al., 2003; Deng andNelson, 2011; Currie, 2006). However, exclusive lanes for bus transit will inevitably invade the road resources for regular traffic and substantially reduce the road capacity, potentially degrading the overall traffic efficiency. Intermittent bus lane (IBL) is further proposed to relieve this issue such that it is transferred to a DBL upon the appearance of a bus and is switched back to serve regular traffic upon the bus leaving the lane (Viegas and Lu,1997) . Furthermore, combining the implementation of IBL with the design of TSP strategy advances the improvement of both the service quality of buses and the efficiency of whole traffic, compared to the pure TSP strategy or DBL implementation at an isolated intersection ion (Eichler and Daganzo, 2006; Ma et al., 2014), and the benefit becomes more significant under an integrated control scheme on the network scale (Viegas and Lu, 2004; Mesbah et al., 2008). \par
The benefit of implementing IBL implies that a permanent DBL is not indispensable for guaranteeing the operational performance of buses. Instead, a series of dedicated time-space slots are sufficient in the view of the infrequent presence of buses, and are more embraced for the whole-network mobility. The achievement of IBL benefits from the technology development of automatic vehicle location or wireless communication, which enables the real-time information of buses to be available, making their arrival times predictable. Furthermore, with the emergence of automated driving technology, vehicles can be controlled more accurately and organized centrally (Administration et al., 2013), enabling the operation of not only transit buses but also the whole traffic to be more predictable and reliable. Thus, leveraging the technologies of connected and automated vehicles (CAVs) has great potential to fulfill a more efficient and credible priority treatment of transit buses. Different from the approach of reserving DBLs for buses to guarantee the service quality, we propose an innovative idea of assigning “dedicated time-space slots” for buses under a CAV environment, which provides exclusive ROWs for buses to eliminate the potential delays along their traveling in the network. \par
Moreover, although the coordinated control strategies in an arterial or a network scale handle the problem of the transit priority effectiveness between adjacent intersections, most of these are only effective in the under-saturated conditions, while over-saturated traffic is still an intractable problem. Specifically, when the traffic becomes over-saturated, vehicle queues will form and overflow the road capacity, and then block the upstream intersection and even cause gridlocks; in this situations, green time is no longer effective as no traffic can be discharged, i.e., descending to de facto red (Abu-Lebdeh and Benekohal, 1997; Daganzo, 2007). To tackle the over-saturation problem, access control of the perimeter is required and should be combined with the coordinated control within the network, involving a more large-scale centralized control scheme. Chen et al. (2020b) proposed an innovative mixed-use BRT/AV lane model where the DBL is available for both BRT vehicles and automated private cars; to simultaneously avoid the interference from private cars on the operation of BRT vehicles and enhance the mobility of private cars, the access control and trajectory planning of private cars are incorporated; the results show that the proposed method can improve the overall efficiency of both the BRT/AV lane and regular lanes, under the premise of guaranteeing the operation of BRT system. The study of Chen et al. (2020b) illustrates the benefit of centralized organization of CAVs for the mobility of the heterogeneous traffic, even though it only focuses on a corridor with a relatively low market penetration ratio of CAVs. With the ongoing development of CAV technology and improvement of the customer acceptance of CAVs, more CAVs will be available on roads in the future, implying great potential to promote the overall mobility of a network. Therefore, a centralized organization and integrated control framework of heterogeneous traffic under a CAV environment is promising for enhancing both the transit service quality and the network mobility.\par
Rhythmic control (RC), recently proposed by Chen et al. (2020a) and Lin et al. (2020), is a novel control framework to improve the automated traffic at intersections and on networks. The key idea of this approach is to organize the traffic into a regularly recurring and quickly-changing manner, which is analogous to the rhythm in music. To embody the rhythmic organization, the concept of "virtual platoon", abbreviated as VP, is introduced to represent a time-space slot that is available for being occupied by vehicles. Under the logic of RC, vehicles can follow the pace of VPs to pass through a traffic facility without any stops or collisions. As the name implies, the VPs just provide potential time-space resources for vehicles to travel with and the actual vehicle occupations are determined by the traffic flow. To resolve the complicated conflict relations at intersections, RC adopts a one-way network, which possesses the potential to improve traffic throughput and alleviate computational complexity in a real-time control implementation. Fig.\ref{Fig_RCreview} shows the concept of VPs and the collision-free trajectories.\par
\begin{figure}[!ht]
\centering
\subfloat[][Virtual platoon]{\includegraphics[width=0.4\textwidth]{Figures/Fig_RCreview1.png}}\hspace{1em}
\subfloat[][Collision-free trajectory]{\includegraphics[width=0.45\textwidth]{Figures/Fig_RCreview2.png}}\\
\caption[]{The concept of RC scheme (Lin et al., 2020)}
\label{Fig_RCreview}
\end{figure}
Since RC is a highly centralized organization and control method, it is capable of meeting the requirement of the bus operation with regard to reliability. Moreover, as buses are generally operated in a cyclic manner, opportunely sharing the characteristic of rhythm with RC, there is great potential to incorporate bus travel into the RC scheme. However, it can be expected that the problem will become more complicated to design an RC scheme under the heterogeneous traffic scenario with both private cars and buses, due to the incompatibility of the rhythm for vehicles with distinctive travel features. Specifically, buses usually travel at a relatively lower speed than that of private cars, along with extra dwelling time at bus stations; thus, it is difficult for buses to catch up with the RC rhythm designed for the traffic of private cars, while accommodating private cars under the rhythm of buses will slow down the pace of private cars and incur enormous delays. \par
For the above research goal, in this study we propose an improved version of RC for heterogeneous CAVs, abbreviated as RC-H, to coherently organize and jointly control the traffic of both private cars and buses in a network. Inheriting the concept of “virtual platoons” proposed by Lin et al. (2020), “dedicated virtual platoons” are designed for buses, which are essentially space-time slots and provide exclusive ROWs for buses on their routes as priority treatment expected. Traveling within the dedicated VPs, buses will fulfill their trips with no collision and limited delays within the network, meanwhile on-time performance and schedule adherence can be guaranteed. In the meantime, “regular virtual platoons” are also designed for private cars, with whose paces the private cars can pass through the network without any stops and collisions. Both the dedicated VPs and regular VPs are unified as "realized VPs", which are extended from the notion of VP of the RC scheme and are regarded as the carriers of the background rhythm. Along with the design of realized VPs, the planning of bus itinerary, defined as a series of space-time trajectories along the travel path, and the traffic assignment of private cars are jointly optimized to acquire an optimal RC-H scheme. The optimization is aimed at minimizing the total personal travel cost and is formulated as a mixed-integer linear program (MILP). A bilevel solution method with a heuristic searching algorithm and tractable approximations is proposed to solve the MILP efficiently. To validate the performances of the RC-H scheme and the effectiveness of the heuristic algorithm, numerical experiments are conducted on both a toy example and a real-world network, while simulation experiments are carried out to compare the results of the RC-H scheme to those of the signal control strategies with and without DBLs.\par
For the remainder of the paper, Section \ref{RC-H_sec} first briefly reviews the RC scheme, and then introduces the concept of RC-H scheme. Section \ref{Optimal_design_sec} develops a joint optimization problem of RC-H scheme composed of realized VP design, bus itinerary planning and traffic assignment of private cars. A bilevel heuristic solution method is proposed to solve the optimization in Section \ref{Two_level_sec}. Section \ref{Numerical_sec} reports a series of numerical examples and simulation experiments for the performance validation of RC-H scheme, and a comparison between RC-H scheme and traffic signal control strategies. Finally, Section \ref{Concluding_sec} summarizes the paper and discusses some future research directions.
\section{Rhythmic control of heterogeneous automated traffic} \label{RC-H_sec}
This section introduces the concept of rhythmic control scheme of heterogeneous traffic of both buses and private cars, i.e., RC-H scheme. In the following, we first briefly review the RC scheme proposed in Lin et al. (2020), and then qualitatively describe how to coherently organize the traffic of both buses and private cars under the RC scheme.\par
\subsection{A brief review of RC} \label{RCreview_subsec}
Rhythmic control (RC) is a novel traffic control framework for vehicles on urban road networks. The basic idea of RC is to organize the whole network traffic into a regularly recurring and quickly-changing manner, where low delays and high traffic throughput can be achieved under the premise of collision avoidance. The concept of “virtual platoon”, i.e., VP, is proposed to represent a time-space slot that is available for being occupied by vehicles and is generated with a preset rhythm. Under the logic of RC, all of the vehicles are required to follow the paces of VPs along the vehicles’ paths to finish their trips and they will not encounter any stops or collisions after entering the network.
Furthermore, Lin et al. (2021) proposes a rhythm design method to optimally determine the RC scheme composed of the cycle length, the VPs' sizes and their arrival times at conflicting nodes (i.e., the pace). The cycle length is the unified rhythm length of the RC scheme, which holds a trade-off between the network capacity and vehicle delay, as a longer rhythm can accommodate higher traffic demands but incurs long average waiting times at entrances or junctions. The sizes of VPs embody the traffic capacities of the links, that is, the maximum numbers of vehicles occupying on the platoons, violating which could potentially result in collisions at conflict points. As discussed above, the platoon sizes are determined according to the traffic demand pattern. Relative arrival time is proposed to denote the arrival time of VPs at the conflicting nodes in each cycle, based on which the travel time on links can be determined and the collision avoidance condition can be guaranteed. The design problem is formulated as a mixed-integer linear program (MILP), and by solving the MILP, an optimal RC scheme for minimizing the total delay can be obtained.\par
The readers can refer to Lin et al. (2020) and Lin et al. (2021) for detailed technical descriptions of the RC scheme. Here, we list some results that lay the foundation of our study of the RC-H scheme described below.
\begin{itemize}
\item A VP represents a time-space slot that is available for being occupied by vehicles and is generated in a rhythmic manner.
\item All vehicles are required to follow the VPs employed on the links and transfer between different platoons along the paths to finish their trips.
\item Under the logic of RC, vehicles will not encounter any stops or collisions after entering the network.\par
\item The average vehicle delay through the entire network is almost negligible when the travel demand is not too high, and massive traffic demands can be well accommodated.
\item RC provably maximizes network traffic throughput under some mild conditions.
\end{itemize}
\begin{figure}[!ht]
\centering
\subfloat[][]{\includegraphics[width=0.45\textwidth]{Figures/Fig_Bus_intersect1.png}}\hspace{1em}
\subfloat[][]{\includegraphics[width=0.45\textwidth]{Figures/Fig_Bus_intersect2.png}}\\
\caption[]{Conflict relation of buses at an intersection}
\label{Fig_Businters}
\end{figure}
\subsection{Concept of RC-H } \label{concept_RCH_subsec}
In a network, conflict points are the main crux of vehicle delays and traffic collisions, which should be given great focus in carrying out the centralized control of the network. Under the logic of the RC scheme, VPs are preset in each link, and all of the vehicles should travel with the pace of the VPs to avoid collisions at conflict points. The same way for buses, Fig.\ref{Fig_Businters} illustrates an example of the conflict relation of buses with VPs at a conflict point composed of two intersecting one-way roads. In Fig.\ref{Fig_Businters}(a), the bus catches up with a VP when passing through the intersection, while in Fig.\ref{Fig_Businters}(b), as the bus falls out of the VP, collision may occur between the bus and the intersecting vehicles at the intersection.\par
Unfortunately, as discussed above, due to their lower speed and dwelling process, it is difficult for buses to catch up with the rhythm designed for accommodating private cars. In the meantime, due to the interference from buses, which can be regarded as “moving bottlenecks”, the travel of private cars in the mixed traffic may also be blocked. Therefore, the original RC scheme is incompatible in the heterogeneous traffic, and an improved RC scheme is required for accommodating both private cars and buses in a harmonized rhythmic manner. Motivated by Fig.\ref{Fig_Businters}(a), in the heterogeneous traffic, we can still coordinate the arrivals of private cars and buses at intersections to form "common paces". It comes to an idea that by re-designing the VPs in the mixed traffic with some customized adjustments, the heterogeneous traffic of buses and private cars can be accommodated simultaneously under the RC framework. \par
Specifically, taking an RC scheme obtained from the study of Lin et al. (2020) as the background, which is designed for private cars with the preset rhythm, platoon size and relative node arrival time, we draw the trajectories of VPs in Fig.\ref{Fig_RCH_bustrajetory}, where the grey bands denote the VPs on link $(i,j)$. For a vehicle intended to pass through the link, as long as it could catch up with some VP at the entrance (node $i$) and the exit (node $j$) of the link, collision avoidance can be guaranteed; Fig.\ref{Fig_RCH_bustrajetory} shows an example of a valid bus trajectory. Then, by combining different VPs at the entrance and exit, a set of new “distorted” VPs can be generated, as the blue and yellow bands in Fig.\ref{Fig_RCH_concept} shows. To avoid confusion, we define the newly generated VP as “realized VP”, and the preset VP as “background VP”. The implication is that the realized VP will be assigned with real traffic while the background VP only provides the background rhythm without actual traffic assignment.\par
\begin{figure}[!ht]
\centering
{\includegraphics[width=0.7\textwidth]{Figures/Fig_RCHconcept1.png}}
\caption[]{Illustration of a valid bus trajectory}
\label{Fig_RCH_bustrajetory}
\end{figure}
\begin{figure}[!ht]
\centering
{\includegraphics[width=0.7\textwidth]{Figures/Fig_RCHconcept2.png}}
\caption[]{Trajectories for different types of virtual platoons}
\label{Fig_RCH_concept}
\end{figure}
From Fig.\ref{Fig_RCH_concept}, the relationship between the realized VP and the background VP can be described as follows:
\begin{itemize}
\item The background VPs are predetermined and provide a fundamental rhythm for the realized VPs, while the realized VPs can be generated as different for the same link in order to accommodate vehicles with different speeds.
\item Under the mixed traffic, the realized VP can be regarded as a “stretched” background VP as the travel time of the former should be no smaller than that of the latter in the presence of buses. However, there is still a chance that the realized VP coincides with the background platoon, such as when the interference of buses can be dodged for private cars.
\item Under the homogeneous traffic of private cars, the realized VPs overlap with the background platoons, as the background VPs are designed for the traffic of private cars, and they can travel freely without the interference from buses.
\end{itemize}\par
Moreover, in general, the system of buses operates in a cyclic manner with a fixed cycle denoted as $H$, while the background VP is generated with a cycle of $T$ under the RC scheme as shown in Fig.\ref{Fig_RCH_bustrajetory}. Therefore, the proposed RC-H scheme also possesses periodicity with a cycle length of the lowest common multiple of $T$ and $H$, denoted as $C$, which characterizes the rhythm of the realized VPs. To simplify the problem, we assume that $H=QT,Q\in\mathbb{N}^+$, where $Q$ is the number of the RC cycles, equivalently the number of background VPs, for a single RC-H cycle\footnote{In fact, $H=QT$ can be easily realized as on the one hand, $T$ is always far smaller than $H$ ($T$ is no greater than 20 seconds while $H$ is generally on the order of hundreds of seconds) and the setting of $T$ can be adjusted as expected; on the other hand, the cycle of the bus system can be regarded as $C$, where the repeated buses for the same line are generated.}. To differentiate the two cycles, "RC cycle" refers to the cycle of the RC scheme (i.e.,$T$) while "RC-H cycle" refers to that of the RC-H scheme (i.e.,$H$). As the movements of vehicles are the same from one RC-H cycle to another, we can model the movements of vehicles in a single RC-H cycle only. \par
To provide exclusive ROWs of buses and avoid the interference from private cars, the VPs assigned for buses are separated from those for private cars, i.e., "dedicated VPs" for buses and "regular VPs" for private cars, both of which are realized VPs. With the introduction of realized VPs, the heterogeneous traffic of buses and private cars can be coherently organized into a harmonized rhythmic manner. Along with the design of realized VPs, the itinerary planning of buses and traffic assignment of private cars can be further jointly considered, which is expected to achieve the optimal design of the RC-H scheme, as discussed in the next section.
\section{Optimal design of RC-H} \label{Optimal_design_sec}
The RC-H scheme proposed in Section \ref{RC-H_sec} guarantees that by following the realized VP, vehicles would be collision free at all conflict points in the network, but may incur delays due to the "stretched deformation" of the background VPs. Therefore, in this section, we present a framework to minimize the total travel cost by designing the RC-H parameters, when given bus line and traffic demand of private cars information.\par
Since previous studies have addressed the design problem of RC scheme, the related parameters including the cycle length, platoon sizes and relative node arrival times of the RC scheme are not decision variables and are already provided prior to solving the RC-H design problem. Therefore, the RC-H design can be modeled as a combined optimization problem with three components; (i) realized VP design is carried out to determine the dedicated VPs for buses and the regular VPs for private cars; (ii) based on the dedicated platoons, the itinerary of the buses can be planned given the bus routes and time schedule; (iii) meanwhile, the itinerary of private cars can also be determined, along with the traffic assignment on the different paths and different types of lanes. Then, by jointly optimizing the three components, the operation of buses and the overall traffic efficiency can be enhanced concurrently.\par
\subsection{Basic settings} \label{basic_RC-H_subsec}
Consider a network composed of links with multiple lanes, where buses are only allowed on a bus lane although the bus lane is shared with other traffic \footnote{Note that here we reserve only one lane for buses as the number of buses is generally limited compared to the private cars, and this approach can be easily generalized to the scenario where multiple bus lanes are required.}. Note that the lane served for buses is not a dedicated lane, but some time-space slots on the lane are dedicated, following the concept of dedicated VPs. Thus, two types of lanes are defined as mixed-traffic lanes and normal lanes. The network can be represented by a directed graph, i.e., $\mathcal{G}=(\mathcal{A}, \mathcal{V})$, where $\mathcal{A}$ denotes the link set and $\mathcal{V}$ denotes the node set. We can also use two nodes to denote a link, i.e., $(i,j)\in\mathcal{A}$, where $i,j\in\mathcal{V}$. The node set can be further divided into origin node set $\mathcal{V}_o$, destination node set $\mathcal{V}_d$, bus station node set $\mathcal{V}_s$ and intersection node set $\mathcal{V}_{in}$.
The notations used for this study are listed in Appendix A.
\subsection{Realized virtual platoon design} \label{virtualpla_subsec}
We introduce a binary variable $\Theta_{q,\hat{q}}^{ij}$ to denote the VP on link $(i,j)$ arriving with platoon $q$ and leaving with platoon $\hat{q}$, where $q,\hat{q}\in\mathcal{Q}$, defined as $q\rightarrow\hat{q}$. Fig.\ref{Fig_realized_virtualpla} illustrates the trajectory of the realized VP. If the VP is realized, $\Theta_{q,\hat{q}}^{ij}=1$, otherwise, $\Theta_{q,\hat{q}}^{ij}=0$. Note that the size of the realized VP is identical to the background platoon and any arrival platoon $q$ or departure platoon $\hat{q}$ can be occupied for only one realized platoon, constrained by the lane capacity, that is:
\begin{align}
& \sum_{\hat{q}\in\mathcal{Q}}\Theta_{q,\hat{q}}^{ij}\leq 1 & \forall ij\in\mathcal{A},q\in\mathcal{Q}\\
& \sum_{q\in\mathcal{Q}}\Theta_{q,\hat{q}}^{ij}\leq 1 & \forall ij\in\mathcal{A},\hat{q}\in\mathcal{Q}
\end{align}\par
\begin{figure}[!ht]
\centering
{\includegraphics[width=0.7\textwidth]{Figures/Fig_realized_virtualpla.png}}\\
\caption[]{Trajectory of a realized virtual platoon}
\label{Fig_realized_virtualpla}
\end{figure}
\vspace{0.5em}
\noindent\underline{\textit{Travel time computation}}\par
Denote the travel time of the realized VP on link $(i,j)$ of $q\rightarrow\hat{q}$ as $t_{q,\hat{q}}^{ij}$. The travel time can be calculated as:
\begin{align}
\begin{split}
& t_{q,\hat{q}}^{ij}=\tau_j-\tau_i+(\hat{q}-q)T+\beta_{q,\hat{q}}^{ij}H\\
&\beta_{q,\hat{q}}=\min\{\beta\in\mathbb{N}|\tau_j-\tau_i+(\hat{q}-q)T+\beta H\geq t_a^{ij}\}
\end{split} & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}
\end{align}
where $\tau_i$ is the relative arrival time of the VPs at node $i$ as shown in Fig.\ref{Fig_realized_virtualpla}, and $t_a^{ij}$ is the travel time of the background VPs on link $(i,j)$. Therefore, the travel time of realized VPs either has several additional cycles of $T$ or is identical to that of the background platoon. Note that the travel times of the realized VPs are not variables and are already computed prior to solving the problem.\par
\vspace{0.5em}
\noindent\underline{\textit{FIFO constraints}}\par
Under the RC scheme, as the vehicles are organized to travel in a relatively tight manner to promote the network throughput, overtaking is not allowed in the mainline roads to reduce traffic risks. Thus, in the realization of the VP, the first-in-first-out (FIFO) constraint should be satisfied. Considering two VPs of $q_m\rightarrow\hat{q}_m$ and $q_n\rightarrow\hat{q}_n$, the FIFO constraint can be expressed as:
\begin{align}
& \Theta_{q_m,\hat{q}_m}^{ij}+\Theta_{q_n,\hat{q}_n}^{ij}\leq \zeta_{q_m,\hat{q}_m,q_n,\hat{q}_n}^{ij} & \forall ij\in\mathcal{A}, q_m,\hat{q}_m,q_n,\hat{q}_n\in\mathcal{Q}
\end{align}
where $\zeta_{q_m,\hat{q}_m,q_n,\hat{q}_n}^{ij}$ denotes whether or not the FIFO constraint between the two realized platoons is violated; if such a violation occurs, $\zeta_{q_m,\hat{q}_m,q_n,\hat{q}_n}^{ij}=1$; otherwise, $\zeta_{q_m,\hat{q}_m,q_n,\hat{q}_n}^{ij}=2$.
For determining the value of $\zeta_{q_m,\hat{q}_m,q_n,\hat{q}_n}^{ij}$, without loss of generality, we assume $q_m<q_n$ to obtain two cases differentiated by $\hat{q}_m>\hat{q}_n$ or $\hat{q}_m<\hat{q}_n$.\par
Case I: $q_m<q_n$ and $\hat{q}_m>\hat{q}_n$. Note that it is possible that $\hat{q}_m$ and $q_m$ are in different cycles, and it can be evaluated by comparing $\hat{q}_m$ and $q_m+\alpha_{ij}$, where $\alpha_{ij}$ denotes the number of cycle $T$ required by the background platoons to pass through link $(i,j)$. Therefore, three possible relations between the platoons of $q_m\rightarrow\hat{q}_m$ and $q_n\rightarrow\hat{q}_n$ exist, as shown in Fig.\ref{Fig_FIFO_case1}. It is observed that the FIFO constraint is violated in the first two relations, i.e., $\hat{q}_m\geq q_m+\alpha_{ij}, \hat{q}_n\geq q_n+\alpha_{ij}$ as shown in Fig.\ref{Fig_FIFO_case1}(a) and
$\hat{q}_m<q_m+\alpha_{ij}, \hat{q}_n<q_n+\alpha_{ij}$ as shown in Fig.\ref{Fig_FIFO_case1}(b), where $\zeta_{q_m,\hat{q}_m,q_n,\hat{q}_n}^{ij}$ should be set to $1$.\par
\begin{figure}[!ht]
\centering
\subfloat[][]{\includegraphics[width=0.45\textwidth]{Figures/Fig_FIFO_1_1.png}}\hspace{1em}
\subfloat[][]{\includegraphics[width=0.45\textwidth]{Figures/Fig_FIFO_1_2.png}}\\
\subfloat[][]{\includegraphics[width=0.45\textwidth]{Figures/Fig_FIFO_1_3.png}}\\
\caption[]{Case I of FIFO constraint}
\label{Fig_FIFO_case1}
\end{figure}
Case II: $q_m<q_n$ and $\hat{q}_m<\hat{q}_n$. For this case, four possible relations between the platoons of $q_m\rightarrow\hat{q}_m$ and $q_n\rightarrow\hat{q}_n$ exist, as shown in Fig.\ref{Fig_FIFO_case2}. The FIFO constraint is violated in the last two relations, i.e., $\hat{q}_m\geq q_m+\alpha_{ij}, \hat{q}_n< q_n+\alpha_{ij}$ as shown in Fig.\ref{Fig_FIFO_case2}(c) and $\hat{q}_m<q_m+\alpha_{ij}, \hat{q}_n\geq q_n+\alpha_{ij}$ as shown in Fig.\ref{Fig_FIFO_case2}(d).
\begin{figure}[!ht]
\centering
\subfloat[][]{\includegraphics[width=0.45\textwidth]{Figures/Fig_FIFO_2_1.png}}\hspace{1em}
\subfloat[][]{\includegraphics[width=0.45\textwidth]{Figures/Fig_FIFO_2_2.png}}\\
\subfloat[][]{\includegraphics[width=0.45\textwidth]{Figures/Fig_FIFO_2_3.png}}\hspace{1em}
\subfloat[][]{\includegraphics[width=0.45\textwidth]{Figures/Fig_FIFO_2_4.png}}\\
\caption[]{Case II of FIFO constraint}
\label{Fig_FIFO_case2}
\end{figure}
\subsection{Bus itinerary planning} \label{bus_itine_subsec}
As buses have distinct travel characteristics, such as fixed schedule, preset routes and dwelling processes, the traveling of buses should be modeled separately from that of private cars. Moreover, to provide exclusive ROWs for buses along their trips, dedicated VPs are introduced for buses to follow up, differentiating from that for private cars, i.e., regular VPs.\par
\vspace{0.5em}
\noindent\underline{\textit{Dedicated virtual platoon}}\par
Assume a bus traveling on link $(i,j)$ as shown in Fig.\ref{Fig_dedicated_virtualpla}. We introduce another binary variable $\hat{\Theta}_{q,\hat{q}}^{p,ij}$ to denote the dedicated VP for bus $p$, and $p\in\mathcal{P}_{ij}$ and $\mathcal{P}_{ij}$ denotes the bus set traveling on link $(i,j)$. $\hat{\Theta}_{q,\hat{q}}^{p,ij}=1$ denotes that the VP of $q\rightarrow\hat{q}$ is dedicated for bus $p$ as shown by the blue band in Fig.\ref{Fig_dedicated_virtualpla}; $\hat{\Theta}_{q,\hat{q}}^{p,ij}=0$ otherwise. For each bus, one and only one dedicated VP is required to load the bus on each link, that is:
\begin{align}
& \sum_{\hat{q}\in\mathcal{Q}}\sum_{q\in\mathcal{Q}}\hat{\Theta}_{q,\hat{q}}^{p,ij}=1, & \forall ij\in\mathcal{A}, p\in\mathcal{P}_{ij}
\end{align}\par
\begin{figure}[!ht]
\centering
{\includegraphics[width=0.7\textwidth]{Figures/Fig_dedicated_virtualpla1.png}}
\caption[]{Illustration of a dedicated virtual platoon}
\label{Fig_dedicated_virtualpla}
\end{figure}
To establish the relationship between the dedicated VP and the realized VP, we introduce $\hat{\Theta}_{q,\hat{q}}^{ij}$ to denote whether or not platoon $q\rightarrow\hat{q}$ is dedicated for buses. Note that one dedicated VP can load multiple buses, as long as the platoon capacity constraint is satisfied. Thus, as long as one bus occupies the platoon $q\rightarrow\hat{q}$, $\hat{\Theta}_{q,\hat{q}}^{ij}=1$; otherwise, $\hat{\Theta}_{q,\hat{q}}^{ij}=0$. Then, we have:
\begin{align}
& M\times\hat{\Theta}_{q,\hat{q}}^{ij}\geq \sum_{p\in\mathcal{P}_{ij}}\hat{\Theta}_{q,\hat{q}}^{p,ij}, & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q} \label{bus_size1}\\
& \hat{\Theta}_{q,\hat{q}}^{ij}\leq \sum_{p\in\mathcal{P}_{ij}}\hat{\Theta}_{q,\hat{q}}^{p,ij}, & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}
\end{align}
and the constraint of platoon capacity can be expressed as:
\begin{align}
& s_b\sum_{p\in\mathcal{P}_{ij}}\hat{\Theta}_{q,\hat{q}}^{p,ij}\leq s_a^{ij}, & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}\label{bus_size2}
\end{align}
where $M$ is a large number, and $s_b$ and $s_a^{ij}$ denote the size of buses and VPs on link $(i,j)$, respectively. Incorporating (\ref{bus_size1}) and (\ref{bus_size2}), we further yield:
\begin{align}
& s_a^{ij}\times\hat{\Theta}_{q,\hat{q}}^{ij}\geq s_b \sum_{p\in\mathcal{P}_{ij}}\hat{\Theta}_{q,\hat{q}}^{p,ij}, & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}
\end{align}\par
Moreover, the dedicated VPs are feasible only under the premise that it has been realized, which can be expressed as:
\begin{align}
& \hat{\Theta}_{q,\hat{q}}^{ij}\leq\Theta_{q,\hat{q}}^{ij}, & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}
\end{align}\par
For describing the dwelling process of buses at stations, two primary types of bus stations are considered; the first is placed in the mainline road, where the vehicles following the bus will be blocked in the dwelling of the bus; the second is placed as a side-platform, where the main road will be free for the following vehicles when the bus dwells at the station, that is, overtaking is allowed during the time of bus dwelling. For the first type of station, the dwelling time of buses can be embodied by adding to the traveling time, as shown in Fig.\ref{Fig_dedicated_dwell}(a); this is equivalent to a bus traveling at a much slower speed, as the impact on the traffic remains the same. For the second type of station, as the bus in the dwelling seems essentially “disappeared” and will not block the traffic on the main road, the VPs planned to pass through within the dwelling time are available. Thus, to describe the movements on the lane with a side-platform station, we introduce an additional node at the position of the station that divides the original link into two connecting links, as shown in Fig.\ref{Fig_dedicated_dwell}(b); then, the traveling of the bus will be modeled separately in the two links, and the VPs passing within the dwelling time can be released for loading other vehicles. \par
\begin{figure}[!ht]
\centering
\subfloat[][Station in the mainline]{\includegraphics[width=0.7\textwidth]{Figures/Fig_dedicated_dwell2.png}}\\
\subfloat[][Station at side-platform]{\includegraphics[width=0.7\textwidth]{Figures/Fig_dedicated_dwell3.png}}\\
\caption[]{Illustration of dedicated virtual platoons at stations}
\label{Fig_dedicated_dwell}
\end{figure}
\begin{figure}[!ht]
\centering
{\includegraphics[width=0.7\textwidth]{Figures/Fig_dedicated_connect.png}}\\
\caption[]{Illustration for link connectivity}
\label{Fig_dedicated_connect}
\end{figure}
\vspace{0.5em}
\noindent\underline{\textit{Link connectivity}}\par
Under the logic of the RC-H scheme, buses will pass through the network continuously without stops (except to dwell at stations), as shown in Fig.\ref{Fig_dedicated_connect}; that is, when the bus departs from link $(i,j)$, it will enter the next link in its route, i.e., link $(j,k)$, immediately. The number of the arrival platoon and the departure platoon of bus $p$ can be obtained as:
\begin{align}
& q_p^{ij}=\sum_{\hat{q}\in\mathcal{Q}}\sum_{q\in\mathcal{Q}}q\times\hat{\Theta}_{q,\hat{q}}^{p,ij}, & \forall ij\in\mathcal{A}, p\in\mathcal{P}_{ij}\\
& \hat{q}_p^{ij}=\sum_{\hat{q}\in\mathcal{Q}}\sum_{q\in\mathcal{Q}}\hat{q}\times\hat{\Theta}_{q,\hat{q}}^{p,ij}, & \forall ij\in\mathcal{A}, p\in\mathcal{P}_{ij}
\end{align}
and the link connectivity constraint can be expressed as:
\begin{align}
& q_p^{jk}-\hat{q}_p^{ij}=0, & \forall j\in\mathcal{V}_{in}, p\in\mathcal{P}_{ij}\cup\mathcal{P}_{jk}
\end{align}
where $\mathcal{V}_{in}$ denotes the intersection node set. \par
For a station node, as shown in Fig.\ref{Fig_dedicated_dwell}(c), the gap between the departure platoon $\hat{q}_p^{ij}$ and the arrival platoon $q_p^{jk}$ should reflect the dwelling time of the bus, that is:
\begin{align}
& \hat{t}_{p}^j=(q_p^{jk}-\hat{q}_p^{ij}+\hat{\phi}_{p}^jQ)\times T, & \forall j\in\mathcal{V}_s, p\in\mathcal{P}_{ij}\cup\mathcal{P}_{jk}\\
& \underline{\hat{t}}_{p}^j\leq \hat{t}_{p}^j\leq H-\epsilon, & \forall j\in\mathcal{V}_s, p\in\mathcal{P}_{ij}\cup\mathcal{P}_{jk}
\end{align}
where $\mathcal{V}_s$ denotes the set of station nodes and $\epsilon$ is a small number used to eliminate the strict inequality of $\hat{t}_{p}^j<H$ in the formulation; $\hat{t}_{p}^j$ denotes the dwelling time of bus $p$ at station $j$, and $\underline{\hat{t}}_{p}^j$ denotes the lower bound of the dwelling time; $\hat{\phi}_{p}^j\in\mathbb{N}$.
\vspace{0.5em}
\noindent\underline{\textit{Travel time computation}}\par
Similar to the computation for the travel time of realized VPs, the bus travel time can be calculated as:
\begin{align}
& t_p^{ij}=\tau_j-\tau_i+(\hat{q}_p^{ij}-q_p^{ij})T+\phi_p^{ij}H, & \forall ij\in\mathcal{A}, p\in\mathcal{P}_{ij}\\
& \underline{t}_p^{ij}\leq t_p^{ij}\leq H+t_{a}^{ij}-\epsilon, & \forall ij\in\mathcal{A}, p\in\mathcal{P}_{ij}
\end{align}
where $t_p^{ij}$ denotes the travel time of bus $p$ on link $(i,j)$, which is a variable as the arrival platoon and departure platoon are decision variables, and $\underline{t}_p^{ij}$ denotes the lower bound of the travel time; $\phi_p^{ij}\in\mathbb{N}$. The inequality $t_p^{ij}<H+t_{a}^{ij}$ implies that the travel time difference between buses and private cars on the link is below $H$ in order to maintain the FIFO constraint. This is reasonable as a dedicated VP with a too large travel time will not be utilized due to the objective of minimizing the travel delay. Furthermore, for the link with a long length where the travel time difference is expected to exceed $H$, we can double the system cycle $H$ or divide the link into two connecting shorter links to handle this issue.\par
\subsection{Traffic assignment of private cars} \label{traffic_assign_subsec}
Unlike for buses, all of the lanes are available for private cars. In regular lanes, private cars can follow the preset (background) VPs with no extra delay, while in mixed-traffic lanes, they are required to follow the regular VPs that is realized but not dedicated for the buses. Therefore, the traffic assignment will determine which path and which lane on each link will be used, along with which regular VP on each lane to follow for the traffic of private cars, as Fig.\ref{Fig_traffic assignment} shows.\par
\begin{figure}[!ht]
\centering
{\includegraphics[width=0.7\textwidth]{Figures/trafficassignment.png}}\\
\caption[]{Illustration of traffic assignment of private cars}
\label{Fig_traffic assignment}
\end{figure}
\vspace{0.5em}
\noindent\underline{\textit{Feasible path}}\par
First, the traffic of private cars of different O-D pairs should be assigned to the corresponding feasible paths. Assume that $d^w$ is the stationary demand for O-D pair $w$, and $f^r$ is the traffic flow on path $r$, we have:
\begin{align}
& \sum_{r\in\mathcal{R}^w}f^r=d^w, & \forall w\in\mathcal{W}
\end{align}
where $\mathcal{R}^w$ is the path set for O-D pair $w$. Denote the set of all feasible paths by $\mathcal{R}$, i.e., $\mathcal{R}=\cup_{w\in\mathcal{W}}\mathcal{R}^w$.\par
\vspace{0.5em}
\noindent\underline{\textit{Traffic assignment}}\par
For the traffic of private cars in the mixed-traffic lane, we introduce variables $\pi_{q,\hat{q}}^{r,ij}$, $\lambda_{q}^{r,ij}$ and $\mu_{\hat{q}}^{r,ij}$ to denote the traffic flow following the VP $q\rightarrow\hat{q}$, the traffic arriving with platoon $q$ and the traffic leaving with platoon $\hat{q}$ on link $(i,j)$ belonging to path $r$, respectively. Thus, we have:
\begin{align}
& \sum_{q\in\mathcal{Q}}\pi_{q,\hat{q}}^{r,ij}=\mu_{\hat{q}}^{r,ij}, & \forall r\in\mathcal{R}, ij\in\mathcal{A}, \hat{q}\in\mathcal{Q}\\
& \sum_{\hat{q}\in\mathcal{Q}}\pi_{q,\hat{q}}^{r,ij}=\lambda_{q}^{r,ij}, & \forall r\in\mathcal{R}, ij\in\mathcal{A}, q\in\mathcal{Q}
\end{align}\par
Note that in the mixed-traffic lane, the traffic can be only assigned to the regular VPs, which have been realized but not dedicated for buses, that is:
\begin{align*}
\sum_{r\in\mathcal{R}}{\pi_{q,\hat{q}}^{r,ij}}\leq M\times (\Theta_{q,\hat{q}}^{ij}-\hat{\Theta}_{q,\hat{q}}^{ij})
\end{align*}
implying that only when $\Theta_{q,\hat{q}}^{ij}=1$ and $\hat{\Theta}_{q,\hat{q}}^{ij}=0$, there is a chance for $\sum_{r\in\mathcal{R}}{\pi_{q,\hat{q}}^{r,ij}}>0$. Additionally, as the traffic should be constrained by the platoon size, the constraint can be further expressed as:
\begin{align}
& \sum_{r\in\mathcal{R}}{\pi_{q,\hat{q}}^{r,ij}}\leq s_a^{ij}\times (\Theta_{q,\hat{q}}^{ij}-\hat{\Theta}_{q,\hat{q}}^{ij}), & \forall ij\in\mathcal{A}, q, \hat{q}\in\mathcal{Q}
\end{align}\par
For the regular lanes, where private cars will follow the background VPs, we introduce $\tilde{\Theta}_{q,\hat{q}}^{ij}$ to denote the background VPs; if platoon $q\rightarrow\hat{q}$ is on the rhythm of the background RC, $\tilde{\Theta}_{q,\hat{q}}^{ij}=1$; otherwise, $\tilde{\Theta}_{q,\hat{q}}^{ij}=0$. Note that $\tilde{\Theta}_{q,\hat{q}}^{ij}$ is not a variable and is determined by the background RC scheme. In the same way, we introduce variables $\tilde\pi_{q,\hat{q}}^{r,ij}$, $\tilde{\lambda}_{q}^{r,ij}$ and $\tilde{\mu}_{\hat{q}}^{r,ij}$ with the similar definition for the traffic assigned to the mixed-traffic lanes. Thus, we have:
\begin{align}
& \sum_{q\in\mathcal{Q}}\tilde{\pi}_{q,\hat{q}}^{r,ij}=\tilde{\mu}_{\hat{q}}^{r,ij}, & \forall r\in\mathcal{R}, ij\in\mathcal{A}, \hat{q}\in\mathcal{Q}\\
& \sum_{\hat{q}\in\mathcal{Q}}\tilde{\pi}_{q,\hat{q}}^{r,ij}=\tilde{\lambda}_{q}^{r,ij}, & \forall r\in\mathcal{R}, ij\in\mathcal{A}, q\in\mathcal{Q}\\
& \sum_{r\in\mathcal{R}}{\tilde{\pi}_{q,\hat{q}}^{r,ij}}\leq s_a^{ij}\times(l_{ij}-1)\times\tilde\Theta_{q,\hat{q}}^{ij}, &\forall ij\in\mathcal{A}, q, \hat{q}\in\mathcal{Q}
\end{align}
where $l_{ij}$ is the lane number on link $(i,j)$.\par
\vspace{0.5em}
\noindent\underline{\textit{Node conservation}}\par
For the node that is neither an entrance nor an exit, the conservation constraint is:
\begin{align}
& (\lambda_{q}^{r,jk}+\tilde\lambda_{q}^{r,jk})-(\mu_{q}^{r,ij}+\tilde\mu_{q}^{r,ij})=0, & \forall r\in\mathcal{R}, j\in\mathcal{V}_{in}\cup\mathcal{V}_{s}, q\in\mathcal{Q}\label{nodeconse_equ}
\end{align}
where $\mathcal{V}_{in}$ and $\mathcal{V}_{s}$ denote the node sets of intersections and stations, respectively. Formula (\ref{nodeconse_equ}) also implies that private cars are not permitted to dwell within the network, similar to the link connectivity constraint of buses.\par
For the node conservation at entrances and exits, we have:
\begin{align}
& \lambda_{q}^{r,jk}+\tilde\lambda_{q}^{r,jk}=Tf^r, & \forall r\in\mathcal{R}, j\in\mathcal{V}_o, q\in\mathcal{Q}\\
& \sum_{\hat{q}\in\mathcal{Q}}{(\mu_{\hat{q}}^{r,ij}+\tilde\mu_{\hat{q}}^{r,ij})}=Hf^r, &\forall r\in\mathcal{R}, j\in\mathcal{V}_d
\end{align}
where $\mathcal{V}_o$ and $\mathcal{V}_d$ denote the node sets of the origins and the destinations.
\subsection{Model formulation} \label{Model_formu_subsec}
To enhance the efficiency of the whole network, the objective of the RC-H design is to minimize weighted travel costs of both buses and private cars. The travel time of private cars is composed of that in mixed-traffic lanes and that in regular lanes, given by:
\begin{align}
& O_{a}=\sum_{ij\in\mathcal{A}}\sum_{q\in\mathcal{Q}}\sum_{\hat{q}\in\mathcal{Q}}\sum_{r\in\mathcal{R}}(\pi_{q,\hat{q}}^{r,ij}\times t_{q,\hat{q}}^{ij}+\tilde\pi_{q,\hat{q}}^{r,ij}\times t_{a}^{ij})
\end{align}\par
The travel time of buses is composed of that moving on mixed-traffic lanes and dwelling at stations, as given by:
\begin{align}
& O_{b}=\sum_{p\in\mathcal{P}_{ij}}\gamma_p\times(\sum_{ij\in\mathcal{A}} t_{p}^{ij}+\sum_{j\in\mathcal{V}_s}\hat{t}_{p}^{j})
\end{align}
where $\gamma_p$ is the average number of passengers on bus $p$. To combine the travel time of buses with private cars, the travel time of buses are normalized into passengers’ travel time (Christofa et al.,2016). Moreover, a parameter denoting the efficiency priority between the buses and private cars is also introduced, i.e., $\omega_p\in[0,1]$. Then, the objective function can be expressed as:
\begin{align}
& O=(1-\omega_p)\times O_{a}+\omega_p\times O_{b}
\end{align}\par
Along with the constraints proposed in the models of realized VP design, bus itinerary planning and traffic assignment of private cars, a mixed integer linear program (MILP) can be formulated, i.e., \textbf{MILP-O}, as follows.
\begin{align}
\mathrm{\textbf{MILP-O}} \nonumber \\
& \min_{\boldsymbol{\Theta},\boldsymbol{\hat\Theta},\boldsymbol{q},\boldsymbol{\hat{q}},\boldsymbol{\phi},\boldsymbol{f},\boldsymbol{\pi},\boldsymbol{\tilde{\pi}},\boldsymbol{\lambda},\boldsymbol{\tilde{\lambda}},\boldsymbol{\mu},\boldsymbol{\tilde{\mu}}}(1-\omega_p)\times O_{a}+\omega_p\times O_{b}
\end{align}
\begin{align}
&\mathrm{s.t.} \nonumber\\
& \sum_{\hat{q}\in\mathcal{Q}}\Theta_{q,\hat{q}}^{ij}\leq 1 & \forall ij\in\mathcal{A},q\in\mathcal{Q}\\
& \sum_{q\in\mathcal{Q}}\Theta_{q,\hat{q}}^{ij}\leq 1 & \forall ij\in\mathcal{A},\hat{q}\in\mathcal{Q}\\
& \Theta_{q_m,\hat{q}_m}^{ij}+\Theta_{q_n,\hat{q}_n}^{ij}\leq \zeta_{q_m,\hat{q}_m,q_n,\hat{q}_n}^{ij} & \forall ij\in\mathcal{A}, q_m,\hat{q}_m,q_n,\hat{q}_n\in\mathcal{Q}\\
& \Theta_{q,\hat{q}}^{ij}\in\{0,1\} & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}\\
& \sum_{\hat{q}\in\mathcal{Q}}\sum_{q\in\mathcal{Q}}\hat{\Theta}_{q,\hat{q}}^{p,ij}=1 & \forall ij\in\mathcal{A}, p\in\mathcal{P}_{ij}\\
\begin{split}
& s_a^{ij}\times\hat{\Theta}_{q,\hat{q}}^{ij}\geq s_b \sum_{p\in\mathcal{P}_{ij}}\hat{\Theta}_{q,\hat{q}}^{p,ij}\\
& \hat{\Theta}_{q,\hat{q}}^{ij}\leq \sum_{p\in\mathcal{P}_{ij}}\hat{\Theta}_{q,\hat{q}}^{p,ij}
\end{split} & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}\\
& \Theta_{q,\hat{q}}^{ij}\geq \hat{\Theta}_{q,\hat{q}}^{ij} & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q} \\
& \hat{\Theta}_{q,\hat{q}}^{ij}, \hat{\Theta}_{q,\hat{q}}^{p,ij}\in\{0,1\} & \forall ij\in\mathcal{A}, p\in\mathcal{P}_{ij}, q,\hat{q}\in\mathcal{Q}\\
\begin{split}
& q_p^{ij}=\sum_{\hat{q}\in\mathcal{Q}}\sum_{q\in\mathcal{Q}}q\times\hat{\Theta}_{q,\hat{q}}^{p,ij}\\
& \hat{q}_p^{ij}=\sum_{\hat{q}\in\mathcal{Q}}\sum_{q\in\mathcal{Q}}\hat{q}\times\hat{\Theta}_{q,\hat{q}}^{p,ij}\\
& q_p^{ij},\hat{q}_p^{ij}\in\mathcal{Q}
\end{split} & \forall ij\in\mathcal{A}, p\in\mathcal{P}_{ij}\\
& q_p^{jk}-\hat{q}_p^{ij}=0 & \forall j\in\mathcal{V}_{in}, p\in\mathcal{P}_{ij}\cup\mathcal{P}_{jk}\\
\begin{split}
& \hat{t}_{p}^j=(q_p^{jk}-\hat{q}_p^{ij}+\hat{\phi}_{p}^jQ)\times T \\
& \underline{\hat{t}}_{p}^j\leq \hat{t}_{p}^j\leq H-\epsilon\\
& \hat{\phi}_{p}^j\in\mathbb{N}
\end{split} & \forall j\in\mathcal{V}_s, p\in\mathcal{P}_{ij}\cup\mathcal{P}_{jk}\\
\begin{split}
& t_p^{ij}=\tau_j-\tau_i+(\hat{q}_p^{ij}-q_p^{ij})T+\phi_p^{ij}H\\
& \underline{t}_p^{ij}\leq t_p^{ij}\leq H+t_a^{ij}-\epsilon\\
& \phi_{p}^{ij}\in\mathbb{N}
\end{split} & \forall ij\in\mathcal{A}, p\in\mathcal{P}_{ij}\\
& \sum_{r\in\mathcal{R}^w}f^r=d^w & \forall w\in\mathcal{W}\\
\begin{split}
& \sum_{q\in\mathcal{Q}}\pi_{q,\hat{q}}^{r,ij}=\mu_{\hat{q}}^{r,ij}\\
& \sum_{\hat{q}\in\mathcal{Q}}\pi_{q,\hat{q}}^{r,ij}=\lambda_{q}^{r,ij}
\end{split} & \forall r\in\mathcal{R}, ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}\\
& \sum_{r\in\mathcal{R}}{\pi_{q,\hat{q}}^{r,ij}}\leq s_a^{ij}\times (\Theta_{q,\hat{q}}^{ij}-\hat{\Theta}_{q,\hat{q}}^{ij}) & \forall ij\in\mathcal{A}, q, \hat{q}\in\mathcal{Q} \\
\begin{split}
& \sum_{q\in\mathcal{Q}}\tilde{\pi}_{q,\hat{q}}^{r,ij}=\tilde{\mu}_{\hat{q}}^{r,ij}\\
& \sum_{\hat{q}\in\mathcal{Q}}\tilde{\pi}_{q,\hat{q}}^{r,ij}=\tilde{\lambda}_{q}^{r,ij}
\end{split} & \forall r\in\mathcal{R}, ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}\\
& \sum_{r\in\mathcal{R}}{\tilde{\pi}_{q,\hat{q}}^{r,ij}}\leq s_a^{ij}\times(l_{ij}-1)\times\tilde\Theta_{q,\hat{q}}^{ij} &\forall ij\in\mathcal{A}, q, \hat{q}\in\mathcal{Q} \\
& (\lambda_{q}^{r,jk}+\tilde\lambda_{q}^{r,jk})-(\mu_{q}^{r,ij}+\tilde\mu_{q}^{r,ij})=0 & \forall r\in\mathcal{R}, j\in\mathcal{V}_{in}\cup\mathcal{V}_{s}, q\in\mathcal{Q}\\
& \lambda_{q}^{r,jk}+\tilde\lambda_{q}^{r,jk}=Tf^r & \forall r\in\mathcal{R}, j\in\mathcal{V}_o, q\in\mathcal{Q}\\
& \sum_{\hat{q}\in\mathcal{Q}}{(\mu_{\hat{q}}^{r,ij}+\tilde\mu_{\hat{q}}^{r,ij})}=Hf^r &\forall r\in\mathcal{R}, j\in\mathcal{V}_d
\end{align}
\section{Bilevel solution method} \label{Two_level_sec}
In \textbf{MILP-O}, the numbers of both variables and constraints are enormous and the relations among the variables are complicated; thus, it is burdensome to obtain the optimal solution by directly solving the MILP. In this section, we propose a bilevel solution method for \textbf{MILP-O} to relieve the computational cost, that can provide sub-optimal solutions with reasonable computational efforts and acceptable solution quality.
\subsection{Problem decomposition} \label{Decomp_subsec}
Considering the scale of the problem, we decompose the original model into two levels, i.e., the upper-level model and the lower-level model. The upper-level model is focused on designing the itinerary of buses along bus route, while the lower-level model solves the problem of traffic assignment of private cars with the given bus itinerary plan. An iterative loop between the two levels is built up as shown in Fig.\ref{Fig_twolevel}. Beginning from the initial solution, the upper level provides an itinerary plan of buses for the lower level, based on which the lower level then solves the traffic assignment problem; after obtaining the solution from the lower level, the performance of the bus planning can be evaluated and coupled back to the upper level; in turn, the upper level improves the planning according to the feedback from the lower level, and then triggers the next iteration. The loop will not be terminated until a preset maximum iteration number is reached, when both the bus itinerary planning and traffic assignment of private cars can be acquired concurrently.
\begin{figure}
\centering
\includegraphics[width=0.55\textwidth]{Figures/Fig_two_level.png}
\caption{Illustration for the bilevel solution structure}
\label{Fig_twolevel}
\end{figure}
\subsection{Variable neighborhood search for upper level} \label{upper_subsec}
Variable neighborhood search (VNS) algorithm is applied for the upper level; it is a metaheuristic method capable of solving problems with a large scale or a complicated structure. VNS has the advantages of both descent to local minima and escape from the valley valley (Hansen et al., 2006). The idea of VNS algorithm is straight-forward; it improves the solution iteratively by systematic changes of the neighborhood with a local search algorithm (Mladenovic and Hansen, 1997). Specifically, in each iteration, taking the current incumbent solution as the search origin, a solution in the neighborhood will be generated as a new solution through a local search routine; then the performance of the new solution will be measured to determine whether to accept it and instruct the search neighborhood for the next iteration. If the solution has not been improved for a certain number of iterations, the size of the neighborhood scale will be adjusted.\par
For applying the VNS algorithm to the upper level, a plan of the bus itinerary will be generated in every iteration. The variables for the planning problem include the arrival time, the travel time on links and the dwelling time at stations for each bus. From the results of numerical experiments presented in the next section, it can be observed that the travel time of buses on links is always optimized to the minimum value, that is, $t_p^{ij}=\underline{t}_p^{ij}$; it is reasonable as the smaller travel time of buses is, the less delay of buses will incur, and concurrently the less delay of private cars will also be induced due to the reduced interference from buses. Therefore, the travel time of buses is fixed as the minimum value in the upper level, reducing the number of variables remained to be determined.\par
\textbf{Algorithm 1} describes the local search routine of a neighbor solution of the upper level. $\bm{x}^{*}$ denotes the current incumbent solution, which are updated continuously toward the direction of performance improvement. $\bm{x}^{k}$ denotes the newly generated solution in iteration $k$ with the searching origin of $\bm{x}^{*}$ and the searching radius $\Delta_k$. $\Delta_k$ embodies the scale of the searching neighborhood and will be updated according to $\Delta_k=\Delta_k+\epsilon_\Delta$ when the incumbent solution $\bm{x}^{*}$ has no update within a preset iteration times; $p_l^k$ and $p_u^k$ are two “instruction parameters” that determine the variation direction of the variables. Note that $p_l^k$ and $p_u^k$ are updated with iteration number $k$ to promote the quality of the generated solution; specifically, in the early searching process, more freedom is provided for the solution variation to create more possibilities for the solution set and avoid the solution from falling into local optima too early; as the iteration number increases, the variation tendency becomes restricted to focus more on the current solution as most variables are already at their optimal values. Although the proposed searching method cannot guarantee the optimality of solution, it does indeed provide sub-optimal solutions with acceptable quality, as will be validated in the numerical experiments.\par
\noindent\rule{\textwidth}{0.2pt}
\noindent \textbf{Algorithm 1. Local search algorithm} \par
\hangafter 1
\hangindent 3.5em
\noindent \textbf{Step 1}. Generate a random number for each variable $x_i^k$, $p_i^k\in[0,1]$, and calculate the values of the instruction parameters in iteration $k$, $p_u^k$ and $p_l^k$; \par
\hangafter 1
\hangindent 3.5em
\noindent \textbf{Step 2}. Compare the generated number with the instruction parameters; if $p_i^k<p_l^k$, set $x_i^k=x_i^k-\Delta_k$; if $p_i^k>p_u^k$, set $x_i^k=x_i^k+\Delta_k$; otherwise, $x_i^k$ remains unchanged; \par
\hangafter 1
\hangindent 3.5em
\noindent \textbf{Step 3}. Output new solution $\bm{x}^k$.\par
\noindent\rule{\textwidth}{0.2pt}
\subsection{Linear programming approximation for lower level} \label{lower_subsec}
In the lower-level problem, with the given bus itinerary plan, the traffic assignment of private cars is optimized; the optimization is formulated as an MILP problem, as shown for \textbf{MILP-L}:
\begin{align}
\mathrm{\textbf{MILP-L}} \nonumber \\
& O_L=\min_{\boldsymbol{\Theta},\boldsymbol{f},\boldsymbol{\pi},\boldsymbol{\tilde{\pi}},\boldsymbol{\lambda},\boldsymbol{\tilde{\lambda}},\boldsymbol{\mu},\boldsymbol{\tilde{\mu}}}(1-\omega_p)\times O_a
\end{align}
\begin{align}
&\mathrm{s.t.} \nonumber\\
& \sum_{\hat{q}\in\mathcal{Q}}\Theta_{q,\hat{q}}^{ij}\leq 1 & \forall ij\in\mathcal{A},q\in\mathcal{Q}\\
& \sum_{q\in\mathcal{Q}}\Theta_{q,\hat{q}}^{ij}\leq 1 & \forall ij\in\mathcal{A},\hat{q}\in\mathcal{Q}\\
& \Theta_{q_m,\hat{q}_m}^{ij}+\Theta_{q_n,\hat{q}_n}^{ij}\leq \zeta_{q_m,\hat{q}_m,q_n,\hat{q}_n}^{ij} & \forall ij\in\mathcal{A}, q_m,\hat{q}_m,q_n,\hat{q}_n\in\mathcal{Q}\label{FIFO_eq}\\
& \Theta_{q,\hat{q}}^{ij}\in\{0,1\} & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}\\
& \Theta_{q,\hat{q}}^{ij}\geq \hat{\Theta}_{q,\hat{q}}^{ij} & \forall ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q} \\
& \sum_{r\in\mathcal{R}^w}f^r=d^w & \forall w\in\mathcal{W}\\
\begin{split}
& \sum_{q\in\mathcal{Q}}\pi_{q,\hat{q}}^{r,ij}=\mu_{\hat{q}}^{r,ij}\\
& \sum_{\hat{q}\in\mathcal{Q}}\pi_{q,\hat{q}}^{r,ij}=\lambda_{q}^{r,ij}
\end{split} & \forall r\in\mathcal{R}, ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}\\
& \sum_{r\in\mathcal{R}}{\pi_{q,\hat{q}}^{r,ij}}\leq s_a^{ij}\times (\Theta_{q,\hat{q}}^{ij}-\hat{\Theta}_{q,\hat{q}}^{ij}) & \forall ij\in\mathcal{A}, q, \hat{q}\in\mathcal{Q} \\
\begin{split}
& \sum_{q\in\mathcal{Q}}\tilde{\pi}_{q,\hat{q}}^{r,ij}=\tilde{\mu}_{\hat{q}}^{r,ij}\\
& \sum_{\hat{q}\in\mathcal{Q}}\tilde{\pi}_{q,\hat{q}}^{r,ij}=\tilde{\lambda}_{q}^{r,ij}
\end{split} & \forall r\in\mathcal{R}, ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}\\
& \sum_{r\in\mathcal{R}}{\tilde{\pi}_{q,\hat{q}}^{r,ij}}\leq s_a^{ij}\times(l_{ij}-1)\times\tilde\Theta_{q,\hat{q}}^{ij} &\forall ij\in\mathcal{A}, q, \hat{q}\in\mathcal{Q} \\
& (\lambda_{q}^{r,jk}+\tilde\lambda_{q}^{r,jk})-(\mu_{q}^{r,ij}+\tilde\mu_{q}^{r,ij})=0 & \forall r\in\mathcal{R}, j\in\mathcal{V}_{in}\cup\mathcal{V}_{s}, q\in\mathcal{Q}\\
& \lambda_{q}^{r,jk}+\tilde\lambda_{q}^{r,jk}=Tf^r & \forall r\in\mathcal{R}, j\in\mathcal{V}_o, q\in\mathcal{Q}\\
& \sum_{\hat{q}\in\mathcal{Q}}{(\mu_{\hat{q}}^{r,ij}+\tilde\mu_{\hat{q}}^{r,ij})}=Hf^r &\forall r\in\mathcal{R}, j\in\mathcal{V}_d
\end{align}
In the \textbf{MILP-L}, $\hat{\Theta}_{q,\hat{q}}^{ij}$ is the input from the upper level, which denotes the VPs dedicated for buses. Under the premise of a given $\hat{\Theta}_{q,\hat{q}}^{ij}$, the traffic assignment of private cars can be solved, which is much more tractable compared to \textbf{MILP-O} and is solvable for a large-scale problem.\par
Moreover, to accelerate the solution process of the lower level, several approximation techniques are applied to further reduce the solution space and simplify the problem structure in the loop iteration, as described below.\par
\begin{enumerate}
\item The VPs slower than the bus traveling on the same link are abandoned. As the objective of the optimization is to minimize the total travel cost, the likelihood for the VPs with low speed to be realized for private cars is limited due to the large delay. Then, the candidate set of VPs can be shrunk as shown in Fig.\ref{Fig_abadonslow}. \par
\begin{figure}[!ht]
\centering
\subfloat[][Before shrink]{\includegraphics[width=0.45\textwidth]{Figures/Fig_abadonslow1.png}}\hspace{1em}
\subfloat[][After shrink]{\includegraphics[width=0.45\textwidth]{Figures/Fig_abadonslow2.png}}\\
\caption[]{ Illustration of abandoning slow virtual platoons}
\label{Fig_abadonslow}
\end{figure}
\item The FIFO constraint (\ref{FIFO_eq}) is removed in solving the lower-level problem. As the FIFO constraint is imposed on any two different VPs for each link, it generates a large number of constraints for the problem and slows down the solution process to a great extent. From the numerical experiments in the next section, it can be validated that the removal of the FIFO constraint has limited impacts on the results, as few VPs violating the FIFO constraint are realized. Moreover, with some adjustment methods, the violation of the FIFO constraint can be eliminated without affecting the total cost. Fig.\ref{Fig_adjustment} illustrates an example of the adjustment method, where Nos.1 and 2 are the trajectories violating FIFO constraint and Nos.3 and 4 are the trajectories after re-assignment. Note that this adjustment method is only applicable for the vehicles with the same destination.\par
\begin{figure}[!ht]
\centering
\subfloat[][Before adjustment]{\includegraphics[width=0.45\textwidth]{Figures/Fig_FIFOadjust1.png}}\hspace{1em}
\subfloat[][After adjustment]{\includegraphics[width=0.45\textwidth]{Figures/Fig_FIFOadjust2.png}}\\
\caption[]{Illustration of an adjustment method}
\label{Fig_adjustment}
\end{figure}
\item The VPs are allowed to be split for different traffic flows. This technique aims to remove the binary variable of $\Theta_{q,\hat{q}}^{ij}$, and then the problem can be re-formulated as a linear program (LP), which can be solved with a high computational efficiency. Similar to technique (2), the impact of the LP approximation is validated in the numerical experiments, where it shows that the approximation has limited impacts as most realized VPs remain non-split.\par
\end{enumerate}
Then, using these three approximation techniques, the lower-level problem can be formulated as an LP with reduced solution space, as shown by \textbf{LP-L}:
\begin{align}
\mathrm{\textbf{LP-L}} \nonumber \\
& O_L'=\min_{\boldsymbol{f},\boldsymbol{\pi},\boldsymbol{\tilde{\pi}},\boldsymbol{\lambda},\boldsymbol{\tilde{\lambda}},\boldsymbol{\mu},\boldsymbol{\tilde{\mu}}}(1-\omega_p)\times O_a
\end{align}
\begin{align}
&\mathrm{s.t.} \nonumber\\
& \sum_{r\in\mathcal{R}^w}f^r=d^w & \forall w\in\mathcal{W}\\
\begin{split}
& \sum_{q\in\mathcal{Q}}\pi_{q,\hat{q}}^{r,ij}=\mu_{\hat{q}}^{r,ij}\\
& \sum_{\hat{q}\in\mathcal{Q}}\pi_{q,\hat{q}}^{r,ij}=\lambda_{q}^{r,ij}
\end{split} & \forall r\in\mathcal{R}, ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}\\
& \sum_{r\in\mathcal{R}}{\pi_{q,\hat{q}}^{r,ij}}\leq s_a^{ij}\times \dot{\Theta}_{q,\hat{q}}^{ij} & \forall ij\in\mathcal{A}, q, \hat{q}\in\mathcal{Q} \\
\begin{split}
& \sum_{q\in\mathcal{Q}}\tilde{\pi}_{q,\hat{q}}^{r,ij}=\tilde{\mu}_{\hat{q}}^{r,ij}\\
& \sum_{\hat{q}\in\mathcal{Q}}\tilde{\pi}_{q,\hat{q}}^{r,ij}=\tilde{\lambda}_{q}^{r,ij}
\end{split} & \forall r\in\mathcal{R}, ij\in\mathcal{A}, q,\hat{q}\in\mathcal{Q}\\
& \sum_{r\in\mathcal{R}}{\tilde{\pi}_{q,\hat{q}}^{r,ij}}\leq s_a^{ij}\times(l_{ij}-1)\times\tilde\Theta_{q,\hat{q}}^{ij} &\forall ij\in\mathcal{A}, q, \hat{q}\in\mathcal{Q} \\
& (\lambda_{q}^{r,jk}+\tilde\lambda_{q}^{r,jk})-(\mu_{q}^{r,ij}+\tilde\mu_{q}^{r,ij})=0 & \forall r\in\mathcal{R}, j\in\mathcal{V}_{in}\cup\mathcal{V}_{s}, q\in\mathcal{Q}\\
& \lambda_{q}^{r,jk}+\tilde\lambda_{q}^{r,jk}=Tf^r & \forall r\in\mathcal{R}, j\in\mathcal{V}_o, q\in\mathcal{Q}\\
& \sum_{\hat{q}\in\mathcal{Q}}{(\mu_{\hat{q}}^{r,ij}+\tilde\mu_{\hat{q}}^{r,ij})}=Hf^r &\forall r\in\mathcal{R}, j\in\mathcal{V}_d
\end{align}
In the formulation of \textbf{LP-L}, $\dot{\Theta}_{q,\hat{q}}^{ij}$ denotes whether the corresponding VP can be realized, which is determined by the results of bus itinerary planning, i.e., $\hat{\Theta}_{q,\hat{q}}^{ij}$, as shown in Fig.\ref{Fig_abadonslow}; if it can be realized, $\dot{\Theta}_{q,\hat{q}}^{ij}=1$; otherwise, $\dot{\Theta}_{q,\hat{q}}^{ij}=0$. As the bus itinerary planning is determined by the upper bound, $\dot{\Theta}_{q,\hat{q}}^{ij}$ has been preset when solving the lower-level model; thus, all of the variables are continuous as expected.\par
Note that the \textbf{LP-L} is only applied in the iteration loop to accelerate the solving process, and \textbf{MILP-L} will be solved after the termination of the iteration, as the exact solution provided by the bilevel solution procedure.
\subsection{Iterative algorithm} \label{iterative_subsec}
\textbf{Algorithm 2} describes the detailed iterative process of the bilevel solution method. $S^{*}$ denotes the current incumbent solution of the bilevel problem and $S^{k}$ denotes the newly generated solution in iteration $k$, obtained by solving the \textbf{LP-L} with the input of $\bm{x}^k$. If the new solution $S^{k}$ is better than $S^{*}$, $S^{*}$ will be updated by $S^{k}$, and $\bm{x}^{*}$ will be updated by $\bm{x}^k$ as the new searching origin for \textbf{Algorithm 1}; otherwise, $S^{k}$ will be abandoned and the next iteration will be triggered.\par
\noindent\rule{\textwidth}{0.2pt}
\noindent \textbf{Algorithm 2. Iterative algorithm} \par
\hangafter 1
\hangindent 3.5em
\noindent {\textbf{Step 1}}. Set the initial values of step length $\Delta_0$ and instruction parameters $p_u^0$, $p_l^0$, $0\leq p_l^0<p_u^0\leq1$; obtain the initial solution of upper level, $\bm{x}^0$, and input it to the lower level to obtain the initial solution, $S_0$. Set the iteration number $k=1$ and let $\Delta_k=\Delta_0, p_l^k=p_l^0,p_u^k=p_u^0,\bm{x}^{*}=\bm{x}^0,S^{*}=S_0$; \par
\hangafter 1
\hangindent 3.5em
\noindent {\textbf{Step 2}}. Generate a new solution $\bm{x}^{k}$ in the neighborhood of $\bm{x}^{*}$ by \textbf{Algorithm 1} with the parameters of $\Delta_k$, $p_l^k$ and $p_u^k$; \par
\hangafter 1
\hangindent 3.5em
\noindent {\textbf{Step 3}}. Input solution $\bm{x}^{k}$ to the lower level and solve the \textbf{LP-L} to obtain a new solution of the bilevel problem, $S_k$; \par
\hangafter 1
\hangindent 3.5em
\noindent {\textbf{Step 4}}. Compare the new solution $S_k$ with current incumbent solution $S^{*}$; if the new solution is better, update the current incumbent solution by $S^{*}=S_k$, $\bm{x}^{*}=\bm{x}^{k}$; otherwise, abandon the new solution;\par
\hangafter 1
\hangindent 3.5em
\noindent {\textbf{Step 5}}. Update the iteration number as $k=k+1$. If $k$ is beyond the maximum iteration time, stop the iteration and go to Step 6; otherwise, update the value of $\Delta_k$, $p_l^k$ and $p_u^k$, and go back to Step 2;\par
\hangafter 1
\hangindent 3.5em
\noindent {\textbf{Step 6}}. Input solution $\bm{x}^{*}$ to the lower level and solve the \textbf{MILP-L} to obtain the exact solution of the bilevel problem.\par
\noindent\rule{\textwidth}{0.2pt}
\section{Numerical experiments} \label{Numerical_sec}
In the numerical experiments, two scenarios are tested, i.e., a toy example and a real-world network implementation. In the toy example, we test the proposed RC-H scheme in a small-scale network. As the optimal solution of \textbf{MILP-O} can be obtained for the small-scale problem, the performance of the RC-H scheme will be measured and the performances characteristics of the bilevel solution method will be validated. In the real-world network scenario, the applicability of the proposed methods on a large-scale problem in practice will be verified, and the advantages compared with the traffic signal control strategies will also be demonstrated by simulation experiments.
\subsection{Toy example} \label{toy_exam_subsec}
\subsubsection{Experimental setups} \label{setup_toy_subsubsec}
\noindent\underline{\textit{Road network settings}}\par
We consider a simple scenario of one corridor as shown in Fig.\ref{Fig_toy_network}. In this scenario, there are 11 nodes composed of 5 intersection nodes, 5 station nodes and 1 virtual node, and 10 links with one lane on each link. Note that node 1 is designed as a virtual node and link 1 is a virtual link for the traffic organization in the entrance with no actual length, which can be recognized as a waiting zone for vehicles prior to their entry into the network.
\begin{figure}[!ht]
\centering
{\includegraphics[width=0.95\textwidth]{Figures/Fig_toy_network1.png}}
\caption[]{Illustration for toy example scenario}
\label{Fig_toy_network}
\end{figure}\par
\vspace{0.5em}
\noindent\underline{\textit{Traffic demand settings}}\par
Two bus lines are considered: bus line 1 from node 2 to node 11, and bus line 2 from node 6 to node 11; note that there are shared links for the two bus lines as shown in Fig.\ref{Fig_toy_network}. One O-D pair of private car traffic is also set from node 2 to node 11 with various demand levels.
\vspace{0.5em}
\noindent\underline{\textit{RC-H scheme settings}}\par
The cycles of background RC scheme and bus system are set as $T=10s$ and $H=120s$, respectively, so that the number of background VPs for one RC-H cycle is $Q=12$. The free-flow travel time of private cars and buses on each link, i.e., $\underline{t}_a^{ij}$ and $\underline{t}_p^{ij}$, are listed in Table \ref{toy_travtim_table}, and the minimum dwelling time at stations of buses is set as $\underline{\hat{t}}_{p}^{j}=40s$. The VP size and bus size are set as $s_a=4$ and $s_b=2$, respectively, implying that one VP can accommodate at most two buses. The number of passengers on buses is set as $\gamma_p=20$.
\begin{table}[]
\centering
\caption{Free-flow travel time of vehicles}
\label{toy_travtim_table}
\begin{tabular}{c|c c c c c c c c c c}
\hline
Link & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10\\
\hline
$\underline{t}_a^{ij}(s)$ & 0 & 10 & 20 & 10 & 20 & 10 & 10 & 20 & 10 & 20\\
$\underline{t}_p^{ij}(s)$ & 0 & 20 & 30 & 20 & 30 & 20 & 20 & 30 & 20 & 30\\
\hline
\end{tabular}
\end{table}
\vspace{0.5em}
\noindent\underline{\textit{Heuristic algorithm settings}}\par
The parameters of the upper-level VNS algorithm are set as follows: the initial step length is the minimum unit of variables, i.e, $\Delta_0=T$, and the variation of the step length is also set as $\epsilon_\Delta=T$, and the upper bound of the step length is set as $H$. For the initial solution of the upper-level model, to guarantee the feasibility of solving the lower-level problem, the buses with shared routes collaborate to travel together, reducing the number of dedicated VPs and then reserve more capacity for private cars. The maximum iteration number for the iterative algorithm is set to $2000$.
\subsubsection{Results of \textbf{MILP-O}} \label{MILP_toy_subsubsec}
First, we investigate the maximum traffic volume of private cars that can be accommodated by the RC-H scheme, defined as “maximum admissible traffic”. For one RC-H cycle, there are $Q$ background VPs for each lane, where at least two of them will serve for the dedicated VP of a single bus when the dwelling process exists, as shown in Fig.\ref{Fig_dedicated_dwell}(b); therefore, the maximum admissible traffic of private cars in a single cycle can be calculated as $(Q-2)\times s_a$. Taking the maximum admissible traffic as the upper bound of the traffic demand of private cars, i.e., $\beta_1=1$, three other demands are set as low level ($\beta_1=0.2$), medium level ($\beta_1=0.5$) and high level ($\beta_1=0.8$), respectively. Moreover, two priorities are considered as bus priority ($\omega_p=0.9$) and private car priority ($\omega_p=0.1$) in the tests. Therefore, a total of $8$ cases are tested as shown in Table \ref{toy_MILP_table)}. \par
\begin{table}[]
\centering
\caption{Results of \textbf{MILP-O}}
\label{toy_MILP_table)}
\begin{tabular}{c|c c c c c c c c}
\hline
Case & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8\\
\hline
$\omega_p$ & 0.9 & 0.9 & 0.9 & 0.9 & 0.1 & 0.1 & 0.1 & 0.1 \\
$\beta_1$ & 0.2 & 0.5 & 0.8 & 1 & 0.2 & 0.5 & 0.8 & 1\\
$O_a^{m}$ & 1040 & 2600 & 4160 & 5200 & 1040 & 2600 & 4160 & 5200\\
$O_a^{opt}$ & 1153 & 2960 & 6160 & 9600 & 1153 & 2860 & 5040 & 6800\\
$\Delta O_a$ & 10.90\% & 13.85\% & 48.08\% & 84.62\% & 10.90\% & 10.00\% & 21.15\% & 30.77\% \\
$O_b^{m}$ & 13200 & 13200 & 13200 & 13200 & 13200 & 13200 & 13200 & 13200\\
$O_b^{opt}$ & 13200 & 13200 & 13200 & 13200 & 13200 & 13600 & 16400 & 24400\\
$\Delta O_b$ & 0.00\% &0.00\% &0.00\% &0.00\% &0.00\% & 3.03\% & 24.24\% & 84.85\% \\
$O^{m}$ &14240 & 15800 & 17360 & 18400 & 14240 & 15800 & 17360 & 18400\\
$O^{opt}$ & 14353 & 16160 & 19360 & 22800 & 14353 & 16460 & 21440 & 31200\\
$\Delta O$ & 0.80\% & 2.28\% & 11.52\% & 23.91\% & 0.80\% & 4.18\% & 23.50\% & 69.57\% \\
\hline
\end{tabular}
\end{table}
The optimization problem is implemented by YALMIP and solved by CPLEX 12.6 solver using MATLAB 2016a on a desktop with a 3.6-GHz Intel Core and 16 GB of RAM. Table \ref{toy_MILP_table)} shows the results of \textbf{MILP-O}, where $O_a^m$ and $O_a^{opt}$ denote the free-flow total travel cost and the cost under the MILP solution of private cars, respectively, and $\Delta O_a$ denotes the gap between $O_a^m$ and $O_a^{opt}$, which reflects the traffic efficiency; $O_b^m$, $O_b^{opt}$ and $\Delta O_b$ have similar meanings but for buses, and $O^m$, $O^{opt}$ and $\Delta O$ are for the total cost of private cars and buses. As the travel time of buses is calculated with the unit of person, i.e., multiplied by the passenger number, $O^m$, $O^{opt}$ and $\Delta O$ can be regarded as the summation of personal travel cost. \par
From Table \ref{toy_MILP_table)}, the results are affected by both traffic demand level and efficiency priority. The main observations are concluded as follows:
\begin{itemize}
\item For private cars, the travel cost increases with traffic demand, and the rate of increase under bus priority is higher than that under private car priority as the VPs are designed preferentially for dedicated VPs under bus priority.
\item For buses, the travel cost also increases with traffic demand under private car priority as the relationship of buses and private cars is undoubtedly competitive for the limited VPs. However, under bus priority, the travel cost of buses holds the minimum value and has no variation with the traffic demand increase; it implies that by setting bus priority, the performance of bus operation can be guaranteed under all demand cases.
\item From the view of personal cost, the travel cost under bus priority is lower than that under private car priority, especially under high demand, indicating that giving priority for buses can reduce the personal delay and is consistent to the encouragement of the transit application. Hence, bus priority is recommended for the personal efficiency and is set for the following tests.
\end{itemize}\par
\begin{figure}[!ht]
\centering
\subfloat[][Under bus priority]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_MILP1.png}}\hspace{1em}
\subfloat[][Under private car priority]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_MILP5.png}}\\
\caption[]{Results of \textbf{MILP-O} under low demand}
\label{Fig_toy_resu_MILP1}
\end{figure}
\begin{figure}[!ht]
\centering
\subfloat[][Under bus priority]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_MILP2.png}}\hspace{1em}
\subfloat[][Under private car priority]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_MILP6.png}}\\
\caption[]{Results of \textbf{MILP-O} under medium demand}
\label{Fig_toy_resu_MILP2}
\end{figure}
\begin{figure}[!ht]
\centering
\subfloat[][Under bus priority]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_MILP3.png}}\hspace{1em}
\subfloat[][Under private car priority]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_MILP7.png}}\\
\caption[]{Results of \textbf{MILP-O} under high demand}
\label{Fig_toy_resu_MILP3}
\end{figure}
\begin{figure}[!ht]
\centering
\subfloat[][Under bus priority]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_MILP4.png}}\hspace{1em}
\subfloat[][Under private car priority]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_MILP8.png}}\\
\caption[]{Results of \textbf{MILP-O} under maximum admissible demand}
\label{Fig_toy_resu_MILP4}
\end{figure}
Figs.\ref{Fig_toy_resu_MILP1}-\ref{Fig_toy_resu_MILP4} draw the trajectories of buses and private cars from the results of \textbf{MILP-O} under different demand levels, where (a) and (b) correspond to bus priority and private car priority, respectively. The pink lines denote the flows of private cars, where a darker color implies more delay and a thicker line implies more traffic volume; the red and green lines denote the trajectories of buses 1 and 2, respectively. From Figs.\ref{Fig_toy_resu_MILP1}-\ref{Fig_toy_resu_MILP4}, several additional observations for the MILP solution are as follows:
\begin{itemize}
\item The travel time of buses on all links is equal to the minimum travel time. That means that the extra travel cost of buses is due solely to the extra dwelling time at stations, supporting the setting of the bus travel time on links as the minimum value when solving the upper-level model, as discussed in Section \ref{upper_subsec}.
\item Different buses collaborate to travel together for most of the shared links. It is expected as the collaboration could reduce the number of dedicated VPs and then reserve more capacity for private cars, in accordance with the idea of the initial solution generation of the upper-level model.
\item Under bus priority, more delayed VPs are realized for private cars, especially when the demand is high, while buses are persistently granted to fulfill their trips with the minimum travel time. By the contrary, under private car priority where the efficiency consideration of buses gives way to that of private cars, extra dwelling times of buses are generated. In particular, when the demand is sufficiently high ($\beta_1=1$), the dwelling time of buses approaches the maximum value, consequently generating an effect of “green wave bands” for private cars as shown in Fig.\ref{Fig_toy_resu_MILP4}(b), remarkably promoting the efficiency of private cars.
\item In Fig.\ref{Fig_toy_resu_MILP4} where $\beta_1=1$, there are some links in which all background VPs are either designed for dedicated VPs or regular VPs with full traffic volume of private cars, as implied by the definition of maximum admissible traffic.
\end{itemize}
\subsubsection{Results of heuristic algorithm} \label{heu_toy_subsubsec}
To test the effectiveness of the bilevel solving method and the proposed heuristic algorithm, comparisons are conducted between the results of the optimal solution and the heuristic solution. Table \ref{toy_heu_table} shows the results under different demand levels, where $O^{opt}$, $O_{ap}^{heu}$, and $O_{ex}^{heu}$ denote the total cost of buses and private cars under the optimal solution of \textbf{MILP-O}, heuristic solution with LP approximations and heuristic solution without LP approximations, respectively; $\Delta O^{heu}$ is the gap between $O^{opt}$ and $O_{ex}^{heu}$. For the heuristic algorithms, the cost of the solution with approximation techniques is identical to the exact solution, verifying the reasonability of the approximation techniques. Moreover, the gap between the optimal solution and the heuristic solution is also tiny, with the maximum value of \textbf{$1.46\%$}; it validates that the heuristic method can provide a sub-optimal solution for \textbf{MILP-O} with a high-quality performance.\par
\begin{table}[]
\centering
\caption{Results obtained by the heuristic method}
\label{toy_heu_table}
\begin{tabular}{c|c c c}
\hline
Case & 1 & 2 & 3\\
\hline
$\beta_2$ & 0.2 & 0.5 & 0.8 \\
$O^{opt}$ & 10791 & 11152 & 11792\\
$O_{ex}^{heu}$ & 10948 & 11152 & 11792\\
$O_{ap}^{heu}$ & 10948 & 11152 & 11792\\
$\Delta O^{heu}$ & 1.46\% & 0.00\% & 0.00\% \\
\hline
\end{tabular}
\end{table}
\begin{figure}[!ht]
\centering
\subfloat[][Approximate solution]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_appro1.png}}\hspace{1em}
\subfloat[][Exact solution]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_appro2.png}}\\
\subfloat[][Iteration curve]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_appro3.png}}
\caption[]{Results of heuristic algorithm under medium demand}
\label{Fig_toy_resu_appro}
\end{figure}
\begin{figure}[!ht]
\centering
\subfloat[][Under low demand]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_heu1.png}}\hspace{1em}
\subfloat[][Under high demand]{\includegraphics[width=0.45\textwidth]{Figures/Fig_toy_resu_heu3.png}}\\
\caption[]{Results of heuristic algorithm under different demands}
\label{Fig_toy_resu_heu}
\end{figure}
Taking the case of medium demand as an example, Fig.\ref{Fig_toy_resu_appro} shows the results of heuristic algorithms where (a) and (b) draw the solutions obtained with LP approximations and without approximations, respectively. Comparing Figs.\ref{Fig_toy_resu_appro} (a) and (b), we observe that only limited regular VPs violating the constraints of FIFO and the indivisibility of VPs have been realized, so their impacts on the results is slight, consistent with the results listed in Table \ref{toy_heu_table}. Fig.\ref{Fig_toy_resu_appro}(c) draws the curve of the objective value in the iteration process. It illustrates that the solution has been improved continuously by the bilevel model, and the optimal solution is obtained after approximately 1600 iterations.\par
Fig.\ref{Fig_toy_resu_heu} shows the results of the heuristic algorithm under low and high demand cases. It is worthy of noting that in some cases (e.g., high demand), although the solution provided by the heuristic algorithm achieves the minimum objective value as the optimal solution of \textbf{MILP-O} does, different itinerary plans of buses and private cars are observed, implying that the optimal solution is non-unique.\par
\subsection{Real-world network implementation} \label{real_net_subsec}
\subsubsection{Experimental setups} \label{setup_real_subsubsec}
\noindent\underline{\textit{Road network settings}}\par
Considering a road network from a part of Wangjing district, Beijing, China, for the implementation of the the RC-H scheme, the network is re-organized with one-way roads, as shown in Fig.\ref{Fig_real_network}. The network contains a total of 18 intersections, 5 entrances and 6 exits. The number of lanes is set to $2$ of one way for each link; thus, for the road with buses traveling, there is one mixed-traffic lane and one regular lane. Based on the real-world situation, 4 bus lines and 12 bus stations are set as shown in Fig.\ref{Fig_real_network}, where the routes of buses are denoted by colorful lines and bus stations are denoted by purple stars. The detailed settings of the bus lines can be found in Table \ref{real_bus_table}.\par
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{Figures/Fig_real_busline.png}
\caption{Road network of Wangjing district}
\label{Fig_real_network}
\end{figure}
\begin{table}[]
\centering
\caption{Settings for bus lines}
\label{real_bus_table}
\begin{tabular}{c|c c c}
\hline
Number & Name & Route & Station\\
\hline
1 & Line 421 & $1\rightarrow15$ & 4;9;11 \\
2 & Line 854 & $7\rightarrow30$ & 6;21;28 \\
3 & Line 536 & $7\rightarrow40$ & 6;21;32;37;39 \\
4 & Line 538 & $27\rightarrow14$ & 32;37;41;22;13 \\
\hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{Figures/Fig_extranode_inters.png}
\caption{Extra nodes and links within intersections}
\label{Fig_extranode_inters}
\end{figure}
According to the road design required by the RC scheme, the additional nodes and links at the intersections are introduced for the traffic of right-turning and left-turning movements. For each intersection, 4 addtional nodes are placed and then 4 links are formed for different movements as shown in Fig.\ref{Fig_extranode_inters}. Consequently, along with the virtual nodes and links for entrances, the whole network is composed of 101 nodes and 129 links, as shown in Fig.\ref{Fig_wholenetwork}.\par
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{Figures/Fig_real_network_whole.png}
\caption{Illustration for the overall network topology}
\label{Fig_wholenetwork}
\end{figure}
\vspace{0.5em}
\noindent\underline{\textit{Traffic demand settings}}\par
As the real traffic demand is limited and the network is transferred to one-way roads, we design various demand patterns in the tests, which can fully verify the performance of the RC-H scheme. Three types of O-D pairs are considered: (1) from an entrance to an exit; (2) from an entrance to a junction; (3) from a junction to an exit; the junction is defined as the node inside the network with traffic generation or absorption, e.g., parking facilities and buildings. Two different O-D pairs are set for each type and then a total of 6 O-D pairs for the demand of private cars is obtained as shown in Fig.\ref{Fig_real_network} and Table \ref{real_OD_table}. Moreover, to improve the computational efficiency, for each O-D pair, only the three shortest paths are candidates for the traffic assignment. In fact, as observed in the following test results, even the third shortest path is not utilized, validating that the shrinkage of the path space is reasonable and its impact on the traffic performance is negligible.\par
\begin{table}[]
\centering
\caption{Settings for the O-D pairs of private cars}
\label{real_OD_table}
\begin{tabular}{c|c c c}
\hline
Number & Type & Origin & Destination \\
\hline
1 & 1 & 1 & 15\\
2 & 1 & 23 & 30\\
3 & 2 & 7 & 26\\
4 & 2 & 34 & 17\\
5 & 3 & 21 & 40\\
6 & 3 & 25 & 14\\
\hline
\end{tabular}
\end{table}
\vspace{0.5em}
\noindent\underline{\textit{RC-H scheme settings}}\par
Prior to the design of the RC-H scheme, an RC scheme should be generated that provides the background rhythm and VPs for the RC-H scheme. The design method of the RC scheme is borrowed from Lin et al. (2021), except that the objective function is the total personal delay of both buses and private cars. Therefore, an MILP problem is formulated to optimize the design of the RC scheme as follows:
\begin{align}
& \min_{\boldsymbol{\tau},\boldsymbol{\alpha},\boldsymbol{r},\boldsymbol{\theta}}\sum_{ij\in\mathcal{A}}(Q\times s_a\times r_a^{ij}+\sum_{p\in\mathcal{P}}\gamma_p\times r_p^{ij})
\end{align}
\begin{align}
&\mathrm{s.t.} \nonumber\\
\begin{split}
& \tau_j-\tau_i+\alpha_a^{ij}T\geq\underline{t}_a^{ij}\\
& \tau_j-\tau_i+\alpha_a^{ij}T<\underline{t}_a^{ij}+T\\
& r_a^{ij}=\tau_j-\tau_i+\alpha_a^{ij}T-\underline{t}_a^{ij}\\
& \tau_i\in[0,T)\\
& \alpha_a^{ij}\in\mathbb{Z}
\end{split} & \forall ij\in\mathcal{A}\\
\begin{split}
& \tau_j-\tau_i+\alpha_p^{ij}T\geq\underline{t}_p^{ij}\\
& \tau_j-\tau_i+\alpha_p^{ij}T<\underline{t}_p^{ij}+T\\
& r_p^{ij}=\tau_j-\tau_i+\alpha_p^{ij}T-\underline{t}_p^{ij}\\
& \alpha_p^{ij}\in\mathbb{Z}
\end{split} & \forall ij\in\mathcal{A}, p\in\mathcal{P}_{ij}\\
\begin{split}
& \tau_{i_1}-\tau_{i_2}+T\theta_{i_1,i_2}=\frac{T}{2}\\
& \theta_{i_1,i_2}\in \{0,1\}
\end{split} & \forall (i_1,i_2)\in\mathcal{C}
\end{align}
where $r_a^{ij}$ and $r_p^{ij}$ denote the delay of VPs and bus $p$ on link $(i,j)$, respectively, $\underline{t}_a^{ij}$ and $\underline{t}_p^{ij}$ denote the free-flow travel time of VPs and bus $p$ on link $(i,j)$, respectively, and $\mathcal{C}$ denotes the set of conflict points. By solving the MILP problem, a background RC scheme can be acquired. Note that the RC scheme is a preliminary design with some simplifications, such as a preset cycle length and an equal platoon size on all links. As the focus of this study is the control scheme for handling heterogeneous traffic, the elaborate design of RC scheme is not investigated in depth, and readers can refer to Lin et al. (2021) for the more details. \par
The setting of the parameters related to RC-H scheme is identical to the toy example, including $T=10s$, $H=120s$, $Q=12$, $s_a=4$, $s_b=2$, $\gamma_p=20$. Additionally, the minimum travel time is calculated by the free-flow speed of vehicles, which is set as $15m/s$ and $10m/s$ for private cars and buses, respectively.
\subsubsection{Results of RC-H scheme} \label{results_real_subsubsec}
First, we calculate the maximum admissible traffic for the network. For a single RC-H cycle, the maximum traffic of a regular lane can be calculated by $Q\times s_a$, while the maximum traffic of a mixed-traffic lane is identical to the toy example, i.e., $(Q-2)\times s_a$. Thus, the total capacity of a road with traveling buses is $(2Q-2)\times s_a$, and that without buses is $2Q\times s_a$. The maximum admissible traffic is defined as the upper bound of the traffic demand, i.e., $\beta_2=1$.\par
To test the impact of traffic demand, three different demand levels are considered as $\beta_2=0.2$ (low), $\beta_2=0.5$ (medium) and $\beta_2=0.8$ (high) for all O-D pairs. Apart from the balanced demand scenarios, imbalanced demands are also considered to test the performance under different demand patterns. Three cases are tested where the demand of one O-D pair type is double of the other two types. The settings of the traffic demand are shown in Table \ref{real_results_table}.\par
\begin{table}[]
\centering
\caption{Results of RC-H scheme}
\label{real_results_table}
\begin{tabular}{c|c c c c c c}
\hline
Pattern & \multicolumn{3}{c}{Balanced} & \multicolumn{3}{c}{Imbalanced} \\
\hline
Case & 1 & 2 & 3 & 4 & 5 & 6 \\
\hline
$\beta_2$ (Type 1) & 0.2 & 0.5 & 0.8 & 1.2 & 0.6 & 0.6\\
$\beta_2$ (Type 2) & 0.2 & 0.5 & 0.8 & 0.6 & 1.2 & 0.6\\
$\beta_2$ (Type 3) & 0.2 & 0.5 & 0.8 & 0.6 & 0.6 & 1.2\\
$O_a^m$ & 7700 & 19250 & 30800 & 31126 & 31020 & 30254\\
$O_a^{heu}$ & 7700 & 19250 & 31395 & 31414 & 31960 & 31316\\
$\Delta O_a$ & 0.00\% & 0.00\% & 1.93\% & 0.92\% & 3.03\% & 3.51\%\\
$O_b^m$ & 36180 & 36180 & 36180 & 36180 & 36180 & 36180\\
$O_b^{heu}$ & 37180 & 37380 & 37580 & 37580 & 36980 & 37980\\
$\Delta O_b$ & 2.76\% & 3.32\% & 3.87\% & 3.87\% & 2.21\% & 4.98\% \\
$O^m$ & 43880 & 55430 & 66980 & 67306 & 67200 & 66434\\
$O^{heu}$ & 44880 & 56630 & 68975 & 68994 & 68940 & 69296 \\
$\Delta O$ & 2.28\% & 2.16\% & 2.98\% & 2.51\% & 2.59\% & 4.31\% \\
\hline
\end{tabular}
\end{table}
Heuristic method is applied in solving the optimization of the RC-H scheme as the MILP is insolvable for the large-scale problem, and the results are shown in Table \ref{real_results_table}. $O_a^m$ and $O_a^{heu}$ denote the free-flow total travel cost and the cost under the heuristic solution of private cars, respectively, and $\Delta O_a$ denotes the gap between $O_a^m$ and $O_a^{heu}$; $O_b^m$, $O_b^{heu}$ and $\Delta O_b$ have similar meanings but for buses, and $O^m$, $O^{heu}$ and $\Delta O$ are for the personal cost of both private cars and buses. It is observed that when the demand is not high, no extra delay of private cars is generated, implying that all private cars can pass through the network with the background rhythm and the interference from buses can be dodged by the RC-H scheme. Moreover, even when the demand is high (i.e., $\beta_2=0.8$), the gap to the minimum travel cost is only $1.93\%$, which embodies the advantage of RC-H in accommodating high demand traffic. Furthermore, although extra travel costs of buses are observed, the gaps are below $5\%$ in all cases, and the extra personal travel costs are also limited as implied by the results of $\Delta O$.\par
Figs.\ref{Fig_real_resu_case2B} and \ref{Fig_real_resu_case2A} show the results of the RC-H scheme under the medium demand (i.e., case 2), where the itineraries of buses 3 and bus 4, and the traffic of O-D pair 3 are shown as examples. Figs.\ref{Fig_real_resu_case2B}(a) and (b) show the bus itineraries of buses 3 and 4, respectively, along with the traffic of private cars on the bus paths to illustrate the traffic assignment results in the mixed lanes. Fig.\ref{Fig_real_resu_case2A} shows the traffic of O-D pair 3 on its shortest path, where the left column is the traffic in the mixed-traffic lane along with the bus trajectories on the related links to illustrate the interaction with buses, and the right column is the traffic on the regular lane. According to the results, most of the private cars are assigned to the regular lanes as no extra delay will be generated in the regular lanes; moreover, the traffic assigned to the mixed-traffic lane is either following the VPs with no generated extra delay or yielding to the capacity constraint of the corresponding regular lane. As the traffic demand is not high, the regular lanes accommodate most of the traffic, and the private cars assigned to the mixed-traffic lanes also dodge the interference from buses. Therefore, private cars follow the background VPs in both lanes without extra costs.\par
\begin{figure}[!ht]
\centering
\subfloat[][Bus 3]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH2_B3.png}}\hspace{1em}
\subfloat[][Bus 4]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH2_B4.png}}\\
\caption[]{Results on the bus itineraries under medium demand}
\label{Fig_real_resu_case2B}
\end{figure}
\begin{figure}[!ht]
\centering
\subfloat[][O-D pair 3 on regular lanes]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH2_A3.png}}\hspace{1em}
\subfloat[][O-D pair 3 on mixed lanes]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH2_A4.png}}\\
\caption[]{Results on private car movements under medium demand}
\label{Fig_real_resu_case2A}
\end{figure}
Figs.\ref{Fig_real_resu_case3B} and \ref{Fig_real_resu_case3A} shows the results under the high demand (i.e., case 3). Compared with case 2, as the demand increases, some delayed VPs are realized for the traffic assignment of private cars, inducing the extra travel cost as shown in Table \ref{real_results_table}. Moreover, according to Fig.\ref{Fig_real_resu_case3A}(a), buses 3 and bus 4 collaborate to travel together on some of the shared links, weakening the interference effect of the buses on the mobility of private cars.\par
\begin{figure}[!ht]
\centering
\subfloat[][Bus 3]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH3_B3.png}}\hspace{1em}
\subfloat[][Bus 4]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH3_B4.png}}\\
\caption[]{Results on the bus itineraries under high demand}
\label{Fig_real_resu_case3B}
\end{figure}
\begin{figure}[!ht]
\centering
\subfloat[][O-D pair 3 on regular lanes]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH3_A3.png}}\hspace{1em}
\subfloat[][O-D pair 3 on mixed lanes]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH3_A4.png}}\\
\caption[]{Results on private car movements under high demand}
\label{Fig_real_resu_case3A}
\end{figure}
In addition, different demand patterns generate quite distinct results as observed from the results presented in Table \ref{real_results_table}. It is observed that the private cars of different O-D pairs are affected by buses to different extents. Specifically, O-D pair 1 interacts with only bus 1 while O-D pairs 3 and 5 interact with buses 2, 3 and 4 in their trips. Therefore, the private cars of O-D pairs 3 and 5 have less admissible demand and more efficiency damage due to the interference from buses. From Table \ref{real_results_table}, it can be obtained that in cases 5 and 6 where O-D pairs 3 and 5 have double demands, greater extra costs are generated compared to case 4. Figs.\ref{Fig_real_resu_case5B} and \ref{Fig_real_resu_case5A} show the results under the imbalanced demand of case 5. In Fig.\ref{Fig_real_resu_case5B}, there are several links that reach the associated lane capacities since the demand is doubled for O-D pair 3, and then more delayed VPs are realized as shown in Fig.\ref{Fig_real_resu_case5A}(a). Moreover, as the demand level of O-D pair 3 is set to 1.2, which is even larger than the upper bound, some private cars are assigned to other paths with detours due to the capacity constraint, further increasing the vehicle costs.
\begin{figure}[!ht]
\centering
\subfloat[][Bus 3]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH5_B3.png}}\hspace{1em}
\subfloat[][Bus 4]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH5_B4.png}}\\
\caption[]{Results on the bus itineraries under imbalanced demand}
\label{Fig_real_resu_case5B}
\end{figure}
\begin{figure}[!ht]
\centering
\subfloat[][O-D pair 3 on regular lanes]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH5_A3.png}}\hspace{1em}
\subfloat[][O-D pair 3 on mixed lanes]{\includegraphics[width=0.45\textwidth]{Figures/Fig_real_resu_RCH5_A4.png}}\\
\caption[]{Results on private car movements under imbalanced demand}
\label{Fig_real_resu_case5A}
\end{figure}
\subsubsection{Simulation comparisons with traffic signal control} \label{simu_real_subsubsec}
As the traffic demand of private cars is assumed to be stationary, the RC-H scheme describes a stable state of the whole network during a single control cycle. Moreover, the discretization of vehicles is ignored in the traffic assignment as the flow of private cars is regarded as continuous variables. Therefore, to further validate the performance under a more general vehicle arrival scenario, simulation tests of online itinerary planning are conducted. The online planning is based on the results of the RC-H scheme, which provides the rhythm and realized VPs for each link. For comparison purpose, traffic signal control (TSC) strategies with the implementation of dedicated bus lanes (DBLs) and without DBLs are also tested in the simulations.\par
Simulation of Urban Mobility (SUMO) is chosen as the platform for the simulation experiments, which has been widely applied for the tests of traffic operations and management (Behrisch et al., 2011). For TSC strategies, impartially, the time lengths of different phases are set to identical values as the sizes of VPs in the RC-H scheme are the same for all links; moreover, two phase lengths are tested, i.e., $15s$ and $30s$, with the clearance time of $2s$; thus, two cycle lengths are generated as $T=34s$ and $T=64s$, respectively; along with the implementation of DBLs or not, a total of four TSC strategies are tested. The parameter settings of vehicles under TSC strategies are identical to the RC-H scheme, and the vehicle following mode is set as "Cooperative Adaptive Cruise Control (CACC)" in SUMO. Note that the minimum headway of private cars is set as $1s$, which is essentially equivalent to the headway setting within the VP of the RC-H scheme\footnote{In the RC-H scheme, the vehicle headway within the VP is set as $1s$ and the headway between the intersecting VPs is set as $2s$; as the platoon size is set as $4$, the total time for two intersecting VPs to pass through is calculated as $(3+2+3+2)s$, which is identical to the set of the RC cycle length.}. The arrival of private cars is assumed to be Poisson arrival for both TSC strategies and the RC-H scheme, and the traffic demand level varies from $0.1$ to $1.5$ with an interval of $0.1$. For each scenario, the simulation test is run for $1$ hour with $5$ repetitions to test the stable traffic and eliminate randomness. \par
\begin{figure}[!ht]
\centering
\subfloat[][Travel time of private cars]{\includegraphics[width=0.5\textwidth]{Figures/Fig_real_simu1.png}}\\
\subfloat[][Travel time of buses]{\includegraphics[width=0.5\textwidth]{Figures/Fig_real_simu2.png}}\\
\subfloat[][Throughput]{\includegraphics[width=0.5\textwidth]{Figures/Fig_real_simu3.png}}\\
\caption[]{Results of simulation experiments}
\label{Fig_real_simu}
\end{figure}
Fig.\ref{Fig_real_simu} shows the results of simulation experiments, where (a), (b) and (c) draw the average travel time of private cars, the average travel time of buses, and vehicle throughput of the network, respectively. The vehicle throughput is calculated as the number of vehicles that have completed their trips during the simulation time. From Fig.\ref{Fig_real_simu}, several observations are summarized as follows:
\begin{itemize}
\item For private cars, the travel time holds a value slightly higher than the free-flow travel time when the traffic demand is low, and increases with the traffic demand under all of the tested control strategies. However, the increasing trend under TSC strategies appears to be more aggressive than that for the RC-H scheme as the travel time increases exponentially under TSC strategies when the demand level exceeds $0.4$, while under the RC-H scheme, it begins to increase until the demand level exceeds $0.9$.
\item Comparing the four TSC strategies, a longer cycle induces more delays when the demand is low, but the opposite effect appears when the demand is high; that is because a longer cycle of TSC can accommodate more traffic due to the lower time loss, as can also be validated in Fig.\ref{Fig_real_simu}(c). The implementation of DBLs damages the efficiency of private cars to a great extent as the ROW of private cars has been deprived on the lanes with bus traveling.
\item By contrast, buses benefit strongly by the implementation of DBLs. Under the TSC without DBLs, the travel time of buses increases with the traffic demand of private cars, due to the interference from the traffic flow in the travel and the vehicle queue at intersections. On the other hand, with DBL implementation, the delay of buses can be alleviated to a bounded value, especially under the TSC with the longer cycle. The result is different for the TSC with the shorter cycle, where the delay of buses still increases with the traffic demand for a certain range; that is because traffic gridlocks are incurred at several intersections, where the buses will be trapped even on the DBLs. Moreover, the delay of buses holds the smallest value under the RC-H scheme among all control strategies, as the priority ROW can be granted for buses along the whole trip, and gridlock will not appear under the centralized control.\par
\item Finally, the network throughput first increases with the demand level, and then remains almost non-increasing when the demand level exceeds a certain value. The maximum throughput of RC-H is approximately $8000 veh/h$, while the maximum throughput is below $4000 veh/h$ for all TSC strategies. The result for the throughput further validates the capability of the RC-H scheme to handle a large traffic demand.
\end{itemize}
\section{Concluding remarks} \label{Concluding_sec}
Enhancing the service quality of bus transit is an important topic to promote the attractiveness of buses and leverage the advantages of high-occupancy vehicles in alleviating the urban traffic issues of congestion and pollution. With the assistance of automated driving and V2X communication technologies, vehicles can be controlled accurately and organized efficiently, enabling the whole traffic to be more predictable and reliable. Leveraging the emerging technologies, this study proposes an innovative control scheme for the heterogeneous automated traffic composed of buses and private cars in a network, i.e., RC-H. Inheriting the idea of rhythmic control scheme proposed by Lin et al. (2020), we organize the traffic of buses and private cars in a rhythmic manner to enhance the transit service quality and the network mobility. Specifically, "dedicated virtual platoons" are designed for buses to provide exclusive right-of-ways (ROWs) on their routes and guarantee the schedule adherence through the network, while "regular virtual platoons" are designed for the traffic flow of private cars to improve the traffic efficiency under the interference from buses. By following the paces of virtual platoons, both buses and private cars can pass through the network without any stops and collisions.\par
Along with the design of virtual platoons, the problems of bus itinerary planning and the traffic assignment of private cars are jointly optimized to minimize the total cost of all vehicles, which is formulated as a mixed integer linear program (MILP). By solving the MILP, an optimal design of RC-H scheme can be obtained. A bilevel heuristic solution method is proposed to alleviate the computational burden, where the upper level provides the plan of the bus itinerary by applying the variable neighborhood search (VNS) algorithm, and the lower level solves the traffic assignment problem, which is transformed into a linear program (LP) by approximation techniques. It is validated that the heuristic method provides a sub-optimal solution with a high-quality performance. \par
Numerical experiments are conducted under both a toy example and a real-work network implementation. The results show that the travel cost of buses and private cars are affected by both traffic demand level and efficiency priority. By setting bus priority, no extra delay of buses will be generated under all demand cases and the personal travel cost can be minimized. Furthermore, simulation experiments are tested to verify the performance under a more general scenario and to compare to traffic signal control (TSC) strategies with dedicated bus lane (DBL) implementation. The results confirm the capability of RC-H in alleviating delays of private cars and granting the schedule adherence of buses concurrently. Compared to the TSC strategies, the RC-H scheme has distinctive advantages for handling massive traffic demand, reflected in both vehicle delay reduction and network throughput promotion.\par
Several promising research directions can be suggested. First, the RC-H scheme is optimized based on a given RC scheme obtained from a simplified method in this study. A more elaborate RC scheme or the incorporation of the design of the RC scheme into the RC-H scheme could further improve the performance. Moreover, to solve the optimization efficiently, a bilevel solution model with a heuristic algorithm is proposed, but the iteration time is still restricted due to the computational cost. Thus, design of a more tractable algorithm or application of a high-quality computational tool such as parallel computing is also highly valuable. Furthermore, while buses have some unique features such as fixed routes and preset time schedules, accommodating the general heterogeneous traffic composed of various vehicle types is worthy of further investigation, which could be more complex due to the distinctive travel features of the different vehicles. Finally, a common cycle length is assumed for the overall network to maintain the rhythm harmonization in this study. It is expected that the relaxation of the rhythm union has the potential to handle various demand patterns in the different parts of the network, and to accommodate the heterogeneous traffic of various vehicle types simultaneously.
\section*{Acknowledgements}
The research is supported in part by the Tsinghua-Daimler Joint Research Center for Sustainable Transportation, and the Tsinghua University-Toyota Research Center.
\section*{References}
\noindent Abu-Lebdeh, G. and Benekohal, R. F. (1997). Development of traffic control and queue management procedures for oversaturated arterials. Transportation Research Record, 1603(1):119–127.\par
\noindent Administration, N. H. T. S. et al. (2013). Preliminary statement of policy concerning automated vehicles. Washington, DC, pages 1–14.\par
\noindent Baker, R. J., Collura, J., Dale, J. J., Head, L., Hemily, B., Ivanovic, M., Jarzab, J., McCormick, D., Obenberger, J., Smith, L., et al. (2002). An overview of transit signal priority. Technical report.\par
\noindent Balke, K. N., Dudek, C. L., and Urbanik, T. (2000). Development and evaluation of intelligent bus priority concept. Transportation Research Record, 1727(1):12–19.\par
\noindent Behrisch, M., Bieker, L., Erdmann, J., and Krajzewicz, D. (2011). Sumo–simulation of urbanmobility: an overview. In Proceedings of SIMUL 2011, The Third International Conference on Advances in System Simulation. ThinkMind.\par
\noindent Chang, G.-l., Vasudevan, M., and Su, C.-c. (1996). Modelling and evaluation of adaptive bus preemption control with and without automatic vehicle location systems. Transportation Research
Part A: Policy and Practice, 30(4):251–268.\par
\noindent Chen, X., Li, M., Lin, X., Yin, Y., and He, F. (2020a). Rhythmic control of automated traffic – part i: Concept and properties at isolated intersections. Transportation Science, Submitted, arXiv:2010.04322 [math.OC].\par
\noindent Chen, X., Lin, X., He, F., and Li, M. (2020b). Modeling and control of automated vehicle access on dedicated bus rapid transit lanes. Transportation Research Part C: Emerging Technologies, 120:102795.\par
\noindent Christofa, E., Ampountolas, K., and Skabardonis, A. (2016). Arterial traffic signal optimization: A person-based approach. Transportation Research Part C: Emerging Technologies, 66:27–47.\par
\noindent Currie, G. (2006). Bus rapid transit in australasia: Performance, lessons learned and futures. Journal of Public Transportation, 9(3):1.\par
\noindent Daganzo, C. F. (2007). Urban gridlock: Macroscopic modeling and mitigation approaches. Transportation Research Part B: Methodological, 41(1):49–62.\par
\noindent Deng, T. and Nelson, J. D. (2011). Recent developments in bus rapid transit: a review of the
literature. Transport Reviews, 31(1):69–96.\par
\noindent Dion, F. and Hellinga, B. (2002). A rule-based real-time traffic responsive signal control system with transit priority: application to an isolated intersection. Transportation Research Part B:
Methodological, 36(4):325–343.\par
\noindent Eichler, M. and Daganzo, C. F. (2006). Bus lanes with intermittent priority: Strategy formulae and an evaluation. Transportation Research Part B: Methodological, 40(9):731–744.\par
\noindent Garrow, M. and Machemehl, R. (1999). Development and evaluation of transit signal priority strategies. Journal of Public Transportation, 2(2):4.\par
\noindent Hansen, P., Mladenovic, N., and Urosevic, D. (2006). Variable neighborhood search and local branching. Computers $\&$ Operations Research, 33(10):3034–3045.\par
\noindent He, Q., Head, K. L., and Ding, J. (2012). Pamscod: Platoon-based arterial multi-modal signal
control with online data. Transportation Research Part C: Emerging Technologies, 20(1):164–184.\par
\noindent He, Q., Head, K. L., and Ding, J. (2014). Multi-modal traffic signal control with priority, signal actuation and coordination. Transportation research part C: emerging technologies, 46:65–82.\par
\noindent Hu, J., Park, B., and Parkany, A. E. (2014). Transit signal priority with connected vehicle technology. Transportation research record, 2418(1):20–29.\par
\noindent Hu, J., Park, B. B., and Lee, Y.-J. (2015). Coordinated transit signal priority supporting transit progression under connected vehicle technology. Transportation Research Part C: Emerging
Technologies, 55:393–408.\par
\noindent Levinson, H. S., Zimmerman, S., Clinger, J., and Gast, J. (2003). Bus rapid transit: Synthesis of case studies. Transportation Research Record, 1841(1):1–11.\par
\noindent Li, M., Yin, Y., Zhang, W.-B., Zhou, K., and Nakamura, H. (2011). Modeling and implementation of adaptive transit signal priority on actuated control systems. Computer-Aided Civil and Infrastructure Engineering, 26(4):270–284.\par
\noindent Liao, C.-F. and Davis, G. A. (2007). Simulation study of bus signal priority strategy: taking advantage of global positioning system, automated vehicle location system, and wireless communications. Transportation research record, 2034(1):82–91.\par
\noindent Lin, X., Li, M., Yin, Y., Shen, Zuojun, M., and He, F. (2020). Rhythmic control of automated
traffic part ii: Grid network rhythm and online routing. Transportation Science, Forthcoming,
arXiv:2010.05416 [math.OC].\par
\noindent Lin, X., Li, M., Yin, Y., Shen, Zuojun, M., and He, F. (2021). Rhythmic control of automated traffic on general road networks. Working paper.\par
\noindent Ma, W., Head, K. L., and Feng, Y. (2014). Integrated optimization of transit priority operation at isolated intersections: A person-capacity-based approach. Transportation Research Part C:
Emerging Technologies, 40:49–62.\par
\noindent Ma, W., Ni, W., Head, L., and Zhao, J. (2013). Effective coordinated optimization model for transit priority control under arterial progression. Transportation Research Record, 2366(1):71–83.\par
\noindent Mesbah, M., Sarvi, M., and Currie, G. (2008). New methodology for optimizing transit priority at the network level. Transportation Research Record, 2089(1):93–100.\par
\noindent Mladenovic, N. and Hansen, P. (1997). Variable neighborhood search. Computers $\&$ operations
research, 24(11):1097–1100.\par
\noindent Smith, H. R., Hemily, B., and Ivanovic, M. (2005). Transit signal priority (tsp): A planning and implementation handbook.\par
\noindent Sunkari, S. R., Beasley, P. S., Urbanik, T., and Fambro, D. B. (1995). Model to evaluate the impacts of bus priority on signalized intersections. Transportation Research Record, pages 117–123.\par
\noindent Urbanik, T., Holder, R., and Fitzgerald, A. (1977). Evaluation of priority techniques for high occupancy vehicles on arterial streets. Technical report, Texas Transportation Institute, Texas
A $\&$ M University.\par
\noindent Viegas, J. and Lu, B. (1997). Traffic control system with intermittent bus lanes. IFAC Proceedings Volumes, 30(8):865–870.\par
\noindent Viegas, J. and Lu, B. (2004). The intermittent bus lane signals setting within an area. Transportation Research Part C: Emerging Technologies, 12(6):453–469.\par
\newpage
\setcounter{table}{0}
\renewcommand{\thetable}{A-\arabic{table}}
\section*{Appendix A: Notations}
\begin{table}[!ht]
\centering
\label{Nomenclature_table}
\normalsize
\begin{tabular}{ l l }
\hline
\multicolumn{2}{l}{\textbf{Sets}} \\
$\mathcal{A}$ & Link set\\
$\mathcal{V}$ & Node set\\
$\mathcal{V}_s$ & Node set of station\\
$\mathcal{V}_o$ & Node set of origin\\
$\mathcal{V}_d$ & Node set of destination\\
$\mathcal{V}_{in}$ & Node set of intersection\\
$\mathcal{Q}$ & Virtual platoon set\\
$\mathcal{P}_{ij}$ & Bus set on link $(i,j)$\\
$\mathcal{W}$ & O-D pair set\\
$\mathcal{R}$ & Path set\\
\hline
\multicolumn{2}{l}{\textbf{Parameters}} \\
$T$ & Cycle of RC scheme\\
$H$ & Cycle of RC-H scheme\\
$Q$ & Cycle number of RC for one RC-H cycle \\
$\tau_i$ & Relative time of node $i$\\
$\underline{t}_a^{ij}$ & Minimum travel time of private cars on link $(i,j)$\\
$t_a^{ij}$ & Travel time of background virtual platoon on link $(i,j)$\\
$t_{q,\hat{q}}^{ij}$ & Travel time of the realized virtual platoon of $q\rightarrow \hat{q}$ on link $(i,j)$\\
$\underline{t}_{p}^{ij}$ & Minimum travel time of bus $p$ on link $(i,j)$\\
$\underline{\hat{t}}_{p}^{j}$ & Minimum dwelling time of bus $p$ at station $j$\\
$s_a^{ij}$ & Virtual platoon size on link $(i,j)$\\
$s_b$ & Bus size\\
$l_{ij}$ & Number of lanes on link $(i,j)$\\
$\gamma_p$ & Number of passengers on bus $p$\\
$d^w$ & Demand of O-D pair $w$\\
\hline
\multicolumn{2}{l}{\textbf{Decision variables}} \\
$\Theta_{q,\hat{q}}^{ij}$ & Realized virtual platoon of $q\rightarrow \hat{q}$ on link $(i,j)$ \\
$\hat{\Theta}_{q,\hat{q}}^{p,ij}$ & Dedicated virtual platoon of $q\rightarrow \hat{q}$ for bus $p$ on link $(i,j)$ \\
$q_p^{ij}$ & Number of arrival platoon for bus $p$ on link $(i,j)$\\
$\hat{q}_p^{ij}$ & Number of departure platoon for bus $p$ on link $(i,j)$\\
$t_p^{ij}$ & Travel time of bus $p$ on link $(i,j)$\\
$\hat{t}_{p}^{j}$ & Dwelling time of bus $p$ at station $j$\\
$f_r$ & Traffic flow of private cars on path $r$\\
$\pi_{q,\hat{q}}^{r,ij}$ & Traffic of path $r$ of $q\rightarrow \hat{q}$ on mixed-traffic lanes of link $(i,j)$ \\
$\tilde{\pi}_{q,\hat{q}}^{r,ij}$ & Traffic of path $r$ of $q\rightarrow \hat{q}$ on regular lanes of link $(i,j)$ \\
$\lambda_{q}^{r,ij}$ & Traffic of path $r$ of arriving with platoon $q$ on mixed-traffic lanes of link $(i,j)$ \\
$\tilde{\lambda}_{q}^{r,ij}$ & Traffic of path $r$ of arriving with platoon $q$ on regular lanes of link $(i,j)$ \\
$\mu_{\hat{q}}^{r,ij}$ & Traffic of path $r$ of departing with platoon $\hat{q}$ on mixed-traffic lanes of link $(i,j)$ \\
$\tilde{\mu}_{\hat{q}}^{r,ij}$ & Traffic of path $r$ of departing with platoon $\hat{q}$ on regular lanes of link $(i,j)$ \\
\hline
\end{tabular} \\
\end{table}
\end{document}
|
\section{Introduction}\label{introduction}
Convolutional neural networks (CNNs) have achieved tremendous successes in real-world applications, such as autonomous vehicles \cite{grigorescu2020survey, hu2020panda} and computer-aided medical diagnoses \cite{trebeschi2017deep, shu2020enhancing}. However, CNNs have be shown vulnerable to well-crafted (and even minute) adversarial perturbations to inputs \cite{szegedy_intriguing_2014, goodfellow_explaining_2015, ilyas_adversarial_2019}. This has become hazardous in high-stakes applications such as medical diagnoses and autonomous vehicles.\blfootnote{Code: \url{https://github.com/HanshuYAN/CIFS}}
Recently, many empirical defense methods have been proposed to secure CNNs against these adversarial perturbations, such as adversarial training (AT) \cite{madry_towards_2019}, input/feature denoising \cite{xie_feature_2019, du_rain_2020} and defensive distillation \cite{papernot_distillation_2016}. AT~\cite{madry_towards_2019}, which generates adversarial data on the fly for training CNNs, has emerged as one of the most successful methods. AT effectively robustifies CNNs but leads to a clear drop in the accuracies for natural data~\cite{tsipras_robustness_2018} and suffers from the problem of overfitting to adversarial data used for training \cite{rice_overfitting_2020,zhang_geometry-aware_2020,chen2021robust}. To ameliorate these problems, researchers have proposed variants of AT, including TRADES \cite{zhang_theoretically_2019} and Friendly-Adversarial-Training (FAT) \cite{zhang_attacks_2020}.
To further robustify CNNs under AT, many works attempt to propose novel defense mechanism to mitigate the effects of adversarial data on features \cite{xie_feature_2019, du_rain_2020, xu_interpreting_2019}. For example, \citet{xie_feature_2019} found that adversarial data result in abnormal activations in the feature maps and performed feature denoising to remove the adversarial effects. Most of these works improved robustness by identifying and suppressing abnormalities \tit{at certain positions} across channels~(commonly referred to as feature maps in CNNs), whereas the other direction, namely, the connection between robustness and irregular activations of certain \tit{entire channels}, has received scant attention.
Since channels of CNNs' deeper layers are capable of extracting semantic characteristic features \cite{zeiler_visualizing_2014}, the process of making predictions usually relies heavily on aggregating information from various channels \cite{bach2015pixel}. As such, anomalous activations of certain channels may result in incorrect predictions. Thus, it is imperative to explore
which channels are \tit{entirely} irregularly activated by adversarial data and which channels' activations benefit or degrade robustness.
By utilizing this connection, we will be able to further enhance the robustness of CNNs via suppressing or promoting certain vulnerable or reliable channels respectively.
In this work, we attempt to build such a connection by comparing the channel-wise activations of non-robust (normally trained) and robustified (adversarially trained) CNNs. The \tit{channel-wise activations} are defined as the average activation magnitudes of all features within channels \cite{bai_improving_2021}. To identify what types of channels appear to be abnormal under attacks, we regard \tit{channels' relevances} to prediction results (formally defined in Equation (\ref{eq:relevance-assessment}) as $g^l$) as the gradients of the corresponding logits \tit{w.r.t} channel-wise activations. The channels, whose relevances to prediction results are positive or negative ($g^l_{[i]}>0$ or $g^l_{[i]}\leq 0$), are called \textit{positively-relevant} (PR) or \textit{negatively-relevant} (NR) channels.
On the one hand, we observe that, AT robustifies CNNs by aligning adversarial data's channel-wise activations with those of natural data. However, we find that the NR channels of adversarially trained CNNs are still over-activated by adversarial data (see Figure \ref{fig: w-vs-act_adv-auto}). Thus, we wonder: \textit{If we suppress NR channels during AT to facilitate the alignment of channel’s activations, will it benefit CNNs' robustness?}
On the other hand, we find that adversarially trained classification models do not enjoy similar robustness across all the classes (see Figure \ref{fig: w-vs-act_adv-auto} and \ref{fig: w-vs-act_adv-cat}). For classes with relatively good robustness, channels' activations usually align well with their relevances, i.e., channels with larger activations are more PR to labels. Given this phenomenon, a natural question arises: \textit{If we align channels' activations with their relevances during AT, will it improve the robustness of CNNs?} Regarding these two questions, we propose a unified hypothesis on robustness enhancement, denoted as $\mathcal H$: \ul{Suppressing NR channels and aligning channels' activations with their relevances to prediction results benefit the robustness of CNNs}.
To examine this hypothesis, we propose a novel mechanism, called \underline{C}hannel-wise \underline{I}mportance-based \underline{F}eature \underline{S}election (CIFS), which adjusts channels' activations with an \tit{importance mask} generated from channels' relevances. For a certain layer, CIFS first takes as input the representation of a data point at this layer and makes a \textit{raw prediction} for the data point by a \textit{probe network}. The probe serves as the surrogate for the subsequent classifier (the composition of subsequent layers) in the backbone and is jointly trained with the backbone under supervision of true labels. Then, CIFS computes the gradients of the sum of the top-$k$ logits \textit{w.r.t.} the channels' activations. We can obtain the \tit{relevance} of each channel to the top-$k$ prediction results by accumulating the gradients within the channel. Finally, CIFS generates a mask of importance scores for each channel by mapping channels' relevances monotonically to non-negative values.
Through extensive experiments, we answer the two questions in the affirmative and confirm hypothesis $\mathcal H$. Indeed, our results show that CIFS clearly enhances the adversarial robustness of CNNs.
We comprehensively evaluate the robustness of CIFS-modified CNNs on benchmark datasets against various attacks. On the CIFAR10 dataset, CIFS improves the robustness of the ResNet-18 by $4$ percentage points against the PGD-100 attack. We also observe that CIFS ameliorates the overfitting during AT. In particular, the robustness at the last epoch is close to that at the best epoch. Finally, we conduct an ablation study to further understand how various elements of CIFS affect the robustness enhancement, such as the top-$k$ feedback and architectures of the probe network.
\input{_fig_adv_effects}
\section{Related Works}
This section briefly reviews relevant adversarial defense methods from two perspectives: adversarial training (AT)-based defense and robust network architecture design.
\paragraph{AT-based Defense} Adversarial training (AT) defends against adversarial attacks by utilizing adversarially generated data in model training \cite{goodfellow_explaining_2015}, formulated as a minimax optimization problem. Recently variants of AT \cite{cai_curriculum_2018,Wang_Xingjun_MA_FOSC_DAT,wang_improving_2020,wu2020adversarial,zhang_geometry-aware_2020} have been proposed. For example, the Misclassification-Aware-AdveRsarial-Training \cite{wang_improving_2020} modifies the process of generating adversarial data by simultaneously applying the misclassified natural data, together with the adversarial data for model training. Recent works have shown AT robustifies CNNs but degrades the natural accuracy \cite{tsipras_robustness_2018, zhang_theoretically_2019, lamb_interpolated_2019}. To achieve a better trade-off, \citet{zhang_theoretically_2019} decomposed the adversarial prediction error into the natural error and boundary error and proposed TRADES to control both terms at the same time. Besides, inspired by curriculum learning \cite{cai_curriculum_2018, bengio_curriculum_2009}, \citet{zhang_attacks_2020} proposed FAT to train models with increasingly adversarial data, which enhances generalization without sacrificing robustness.
In addition, some works introduced various types of regularization for training models, such as layer-wise feature matching \cite{sankaranarayanan_regularizing_2018, liao_defense_2018, kannan_adversarial_2018}, low-rank representations \cite{sanyal_robustness_2020, mustafa_adversarial_2019}, attention map alignment \cite{xu_interpreting_2019}, and Lipschitz regularity\cite{Virmaux_lipschitz_2018, cisse_parseval_2017}. These types of regularization can work in conjunction with AT and benefit the models' robustness.
\paragraph{Robust Network Design} Other than robust training strategies, some works explored robust network architectures \cite{yan_robustness_2020, hsieh_robustness_2019}. For instance, the work by \citet{yan_robustness_2020} showed neural ODE-based models are inherently more robust than conventional CNN models; \citet{guo_sparse_2018} demonstrated that appropriately designed higher model sparsity implies better robustness of nonlinear networks. Another line of works defended against adversarial attacks via gradient obfuscation, such as random or non-differentiable image/feature transformations \cite{xie_mitigating_2018, du_rain_2020, dhillon_stochastic_2018, xiao_enhancing_2019}. However, they have been shown to be insecure to adaptive attacks \cite{athalye_obfuscated_2018, tramer_adaptive_2020}. Recently, many researchers have attempted to develop novel mechanisms for robustness enhancement. \citet{xie_feature_2019} performed feature denoising to remove the adversarial effects on feature maps. \citet{zoran_towards_2020} utilized the spatial attention mechanism to identify highlight important regions of feature maps. Most of these works manipulated CNNs' intermediate representations in the \tit{spatial domain}, whereas our work studies the adversarial robustness from the channel-wise activation perspective.
Channel-wise Activation Suppressing (CAS) \cite{bai_improving_2021}, the most relevant work to ours, also studied the channel-wise activations of adversarial data. It showed channels are activated more uniformly by adversarial data compared to the natural ones, and AT improves the robustness by attempting to align the distributions of channels' activations of natural and adversarial data. However, there are still some channels that are over-activated by adversarial data. To suppress these channels, the authors proposed CAS to adjust channels' activations based on their importance. Although CAS empirically suppresses certain channels, the authors did not show that the suppressed channels correspond to the target ones; this means the primary objective of CAS may not have been met. Thus, there is no guarantee CAS can enhance the robustness of CNNs (see Section \ref{sec:exp-eva} for further evidence on this). \tit{In contrast}, our work first builds a connection between robustness and channels' activations via their relevances to predictions. Then, the proposed CIFS can explicitly control channels' activations based on their relevances. Finally, experiments demonstrate the effectiveness of CIFS on robustness enhancement.
\input{_sec_method_}
\input{_sec_exp_}
\section{Conclusion}
We developed the CIFS mechanism to verify the hypothesis that suppressing NR channels and aligning PR ones with their relevances to predictions benefits adversarial robustness. Empirical results demonstrate the effectiveness of CIFS on enhancing CNNs' robustness.
There are two limitations of our current work: 1) We empirically verify the hypothesis $\mathcal H$, but it is still difficult to explicitly, not intuitively, explain why the adjustment of channels improves robustness. 2) Although CIFS ameliorates the overfitting during AT and improves the robustness, it sometimes leads to a bit drop in natural accuracies on certain datasets. In the future, we will attempt to address these two limitations.
\section*{Acknowledgements}
HY and VYFT are funded
by a Singapore National Research Foundation (NRF) Fellowship
(R-263-000-D02-281).
JF is supported by the National Research Foundation Singapore under its AI Singapore Programme (Award Number: AISG-100E-2019-035)
JZ, GN, and MS are supported by JST AIP Acceleration Research Grant Number JPMJCR20U3, Japan. MS is also supported by the Institute for AI and Beyond, UTokyo.
\section{Details on Visualizing Channel-wise Activations} \label{apdx:visualization}
\subsection{Non-robust CNNs vs. Robustified CNNs} \label{apdx:visualization-normal-at}
We train ResNet-18 models to perform the classification task on the CIFAR10 dataset. The models are trained normally and adversarially. We use adversarial data generated by PGD-10 attack ($\epsilon=\nicefrac{8}{255}$, step size $\nicefrac{\epsilon}{4}$, and random initialization) for adversarial training.
The ResNet-18 network consists of one convolutional layer, eight residual blocks, and one linear fully-connected (FC) layer connected successively. Each residual block contains two convolutional layers for the residual mapping. We visualize the features of the penultimate layer (the output of the eighth residual block) and the weights of the last linear layer in Figure \textcolor{Blue}{1}. Specifically, the weights of the last FC layer for a certain class are sorted and plotted in descending order. We process the penultimate layer's features with the global average pooling operation to obtain the channel-wise activations. For a certain class, we calculate each channel's mean activation magnitude over all the test samples in this category. We normalize the mean channel-wise activations by dividing them by their absolute maximum. The mean channel-wise activations are plotted according to the indices of the sorted weights. We also record the activated frequency of each channel. Here, the channel is regarded to be activated if its activation magnitude is larger than a threshold (1\% of the maximum of all channels' activations).
\subsection{Channel-wise Activations of CIFS-modifed CNNs}
\label{apdx:visualization-cifs-modification}
We train CIFS-modified CNNs normally and adversarially by using the adaptive loss in Equation (\textcolor{Blue}{3}). We use the PGD-10 attack to generate adversarial data. We illustrate the channels' activations of CIFS-modified CNNs in Figure \textcolor{Blue}{3}. The implementation details are same as those in Appendix \ref{apdx:visualization-normal-at}.
In Figure \textcolor{Blue}{3}, we show the channels' activations of data in class ``airplane''. Here, we plot the channels activations of data in other classes. From Figure \ref{fig:apx-hyp}, we see that CIFS indeed suppresses negatively-relevant (NR) channels and promotes the positively-relevant (PR) ones.
\begin{figure}[h!]
\centering
\includegraphics[width=.9\linewidth]{images/chart.pdf}
\vspace{-1em}
\caption{Robust accuracies (\%) of PGD-20 adversarial data for various classes of the CIFAR10 dataset.}
\label{fig:apdx-imbal}
\end{figure}
Besides, we also observe that CIFS ameliorates the class-wise imbalance of robustness under AT. In Figure \ref{fig:apdx-imbal}, we can see that, for the data in class ``cat'' and class ``deer'', the robust accuracies of the vanilla ResNet-18 model are 16.70\% and 25.50\%. Modifying the vanilla model with CIFS-softmax, we can improve the robust accuracies by 5.6 and 3.3 percentage points, respectively.
\def \SubFigWidth {0.245}
\def 2.1 {1}
\begin{figure*}[t!]
\centering
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10___sorted_weight_vs_act_ship.png}
\vspace{-2em}
\caption{\scriptsize ``ship'': non-CIFS}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10_CSAFR_top2_sigmoid_sorted_weight_vs_act_ship.png}
\vspace{-2em}
\caption{\scriptsize ``ship'': CIFS-sigmoid}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10_CSAFR_top2_softplus_sorted_weight_vs_act_ship.png}
\vspace{-2em}
\caption{\scriptsize ``ship'': CIFS-softplus}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10_CSAFR_topk_sorted_weight_vs_act_ship.png}
\vspace{-2em}
\caption{\scriptsize ``ship'': CIFS-softmax}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10___sorted_weight_vs_act_automobile.png}
\vspace{-2em}
\caption{\scriptsize ``autmobile'': non-CIFS}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10_CSAFR_top2_sigmoid_sorted_weight_vs_act_automobile.png}
\vspace{-2em}
\caption{\scriptsize ``autmobile'': CIFS-sigmoid}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10_CSAFR_top2_softplus_sorted_weight_vs_act_automobile.png}
\vspace{-2em}
\caption{\scriptsize ``autmobile'': CIFS-softplus}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10_CSAFR_topk_sorted_weight_vs_act_automobile.png}
\vspace{-2em}
\caption{\scriptsize ``autmobile'', CIFS-softmax}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10___sorted_weight_vs_act_frog.png}
\vspace{-2em}
\caption{\scriptsize ``frog'': non-CIFS}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10_CSAFR_top2_sigmoid_sorted_weight_vs_act_frog.png}
\vspace{-2em}
\caption{\scriptsize ``frog'': CIFS-sigmoid}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10_CSAFR_top2_softplus_sorted_weight_vs_act_frog.png}
\vspace{-2em}
\caption{\scriptsize ``frog'': CIFS-softplus}
\end{subfigure}
\begin{subfigure}{\SubFigWidth\linewidth}
\centering
\includegraphics[width=2.1 \linewidth]{images/hyp/PAT_Res18_cifar10_CSAFR_topk_sorted_weight_vs_act_frog.png}
\vspace{-2em}
\caption{\scriptsize ``frog'', CIFS-softmax}
\end{subfigure}
\caption{The magnitudes of channel-wise activations (top) at the penultimate layer, their activated frequency (bottom), and the weights of the last linear layer (middle) \textit{vs.} channel indices. The robust accuracies against PGD-20 (on the whole dataset) are 46.64\% for non-CIFS, 49.87\% for the CIFS-sigmoid, 50.38\% for the CIFS-softplus, and 51.23\% for the CIFS-softmax respectively}
\label{fig:apx-hyp}
\vspace{-1em}
\end{figure*}
\section{Robustness Evaluation on CIFAR10} \label{apdx:cifar10}
\subsection{Robustness Enhancement of CIFS under AT} \label{apdx:cifar10-at}
\textbf{Training and Evaluation details}: On the CIFAR10 dataset, we train ResNet-18 and WRN-28-10 models with PGD-10 adversarial examples ($\epsilon=\nicefrac{8}{255}$, step size $\nicefrac{\epsilon}{4}$ with random initialization). The $\beta$ in CIFS is set to be $2$. For the ResNet-18 and its CIFS-modified version, we train models for $120$ epochs with the SGD optimizer (momentum $0.9$ and weight decay $0.0002$). The learning rate starts from $0.1$ and is multiplied with $0.1$ at epoch $75$ and epoch $90$. For the WRN-28-10, we train model for $110$ epochs with weight decay $0.0005$.
In Section \textcolor{Blue}{4.1}, we evaluate the robustness of CNNs against four white-box attacks with a perturbation budget $\epsilon=\nicefrac{8}{255}$ in $l_{\infty}$ norm --- FGSM, PGD-20 (step size $\nicefrac{\epsilon}{10}$), C\&W (optimized by PGD for 30 steps with a step size $\nicefrac{\epsilon}{10}$) and PGD-100 (step size $\nicefrac{\epsilon}{10}$).
\textbf{Robustness Evaluation with AutoAttack}:
Here, we also report the robust accuracies of defense methods against AutoAttack \cite{croce_reliable_2020}, which consists of both white-box and black-box attacks. AutoAttack regards models to be robust at a certain data point only if the models correctly classify all types of adversarial examples generated by AutoAttack of that data point. We consider the AutoAttack including one strong white-box attack (Auto-PGD \cite{croce_reliable_2020}) and one black-box attack (Square-Attack \cite{andriushchenko2020square}). Since the Square Attack requires many queries, we sample \num[group-separator={,}]{2000} images (200 per class) from the CIFAR10 for evaluation. The attack parameters are set according to the officially released AutoAttack\footnote{\url{https://github.com/fra31/auto-attack}}. From Table \ref{tab:apdx-aa-cifar}, we observe that CIFS enjoys better robustness against AutoAttack in comparison to the vanilla ResNet-18 model and its CAS-modified version.
\begin{table}[t!]
\centering
\caption{Robustness comparison of defense methods on CIFAR10. We report the last-epoch robust accuracies (\%) against AutoAttack.}
\label{tab:apdx-aa-cifar}
\vspace{1em}
\scalebox{.8}{
\begin{tabular}{cccc}
\toprule
\tbf{\tit{ResNet-18}} & Vanilla & CAS & CIFS \\
ResNet-18 & 44.00 & 42.70 & \tbf{46.20} \\
WRN-28-10 & 47.20 & 46.55 & \tbf{49.75} \\
\bottomrule
\end{tabular}
}
\end{table}{}
\textbf{Best-epoch robustness during training}:
Due to the susceptibility of overtrained models to overfitting \cite{rice_overfitting_2020}, it seems reasonable to compare the results at the end of the training (and not for the best epochs) \cite{madry_towards_2019, zhang_attacks_2020, rice_overfitting_2020}. In Section \textcolor{Blue}{4.1}, we report the robust accuracies of ResNet-18 and WRN-28-10 models at the last epochs. Here, we also provide the results at the \tbf{best} epochs for reference.
From Table \ref{tab:apdx-cifar-best}, we see that, for the ResNet-18 architecture, the CIFS-modified model results in the similar best-epoch robustness (PGD-100) to that of the vanilla ResNet-18. For the WRN-28-10, the vanilla model has the better best-epoch robustness compared to the CIFS-modified version. This may be due to the fact that CIFS suppresses redundant channels and reduces the model capacity.
By comparing results in Table \ref{tab:apdx-cifar-best} with those in Table \textcolor{Blue}{1}, we observe that CIFS indeed ameliorates the overfitting of AT. Specifically, the best-epoch robust accuracy of the vanilla WRN-28-10 (resp. ResNet-18) against PGD-100 attack is 54.17\% (resp.49.47\%), but the last-epoch accuracy drops to 47.08\% (resp. 44.72\%). In contrast, for the CIFS-modified versions, the last-epoch robust accuracies against PGD-100 attack are maintained around the best-epoch ones (for WRN-28-10, from 52.03\% to 51.51\%; for ResNet-18, from 49.76\% to 48.74\%).
\begin{table}[h!]
\centering
\caption{Robustness comparison of defense methods on CIFAR10. We report the best robust accuracies (\%) during training. For each model, the results of the strongest attacks are marked with an underline.}
\label{tab:apdx-cifar-best}
\vspace{1em}
\scalebox{.8}{
\begin{tabular}{cccccc}
\toprule
\tbf{\tit{ResNet-18}} & Natural & FGSM & PGD-20 & C\&W & PGD-100 \\
\hline
Vanilla & 83.63 & 56.73 & 50.64 & 49.51 & \underline{49.47} \\
CAS & 85.66 & 56.25 & 47.69 & 46.52 & \underline{45.69} \\
CIFS & 82.46 & \tbf{58.98} & \tbf{51.94} & \tbf{51.25} & \tbf{\underline{49.76}} \\
\bottomrule
\toprule
\tbf{\tit{WRN-28-10}} & Natural & FGSM & PGD-20 & C\&W & PGD-100\\
\hline
Vanilla & 86.53 & \tbf{61.43} & \tbf{55.69} & \tbf{54.45} & \tbf{\underline{54.17}} \\
CAS & 87.51 & 58.54 & 52.06 & 51.27 & \underline{50.69} \\
CIFS & 84.67 & 61.03 & 54.09 & 53.76 & \underline{52.03} \\
\bottomrule
\end{tabular}
}
\end{table}{}
\textbf{Robust accuracies for various values of $\beta_{\text{atk}}$:} In Section \textcolor{Blue}{4.1}, we evaluate the robustness of CIFS-modified models by using the adaptive loss in Equation (\textcolor{Blue}{3}). For each type of attack, we assign various values to $\beta_{\text{atk}}$ and report the worst robust accuracies. Here, for reference, we provide the defense results of the ResNet-18 model on CIFAR10 for different values of $\beta_{\text{atk}}$ that are used in Section \textcolor{Blue}{4.1}. The results in Table \textcolor{Blue}{1} (ResNet-18) are collected from Table \ref{tab:apdx-beta-attack}.
\begin{table}[t!]
\vspace{-1em}
\centering
\caption{Robust accuracies (\%) for values of $\beta_{\text{atk}}$ on CIFAR10. The value ``$\infty$'' means the attack only considers the second term in Equation (\textcolor{Blue}{3}). The value ``$\infty$-1'' (resp. ``$\infty$-2'') means the attacker completely focuses on the first (resp. second) CIFS-modifed layer. The bracketed numbers are those reported in Table \textcolor{Blue}{1} (ResNet-18).}
\vspace{.5em}
\scalebox{.8}{
\begin{tabular}{ccccccc}
\toprule
\tbf{\tit{ResNet-18}} & $\beta_{\text{atk}}$ & Natural & FGSM & PGD-20 & C\&W & PGD-100 \\
\midrule
Vanilla & - & [84.56] & [55.11] & [46.62] & [45.95] & [\underline{44.72}] \\
\hline
CAS & 0& [86.73] & 83.17 & 88.45 & 88.52 & 88.24 \\
~ & 0.1 & - & 58.61 & 61.36 & 85.51 & 62.40\\
~ & 1 & - & 56.36 & 52.86 & 62.34 & 56.02\\
~ & 2 & - & 56.06 & 49.76 & 54.94 & 50.62\\
~ & 10 & - & 56.03 & 47.47 & 49.35 & 47.70\\
~ & 100 & - & 56.02 & 47.04 & 48.36 & 46.74\\
~ & $\infty$ & - & 56.02 & 47.06 & 48.31 & 46.55\\
~ & $\infty$-1 & - & [55.99] & [45.29] & [44.18] & [\underline{43.22}] \\
~ & $\infty$-2 & - & 82.68 & 87.87 & 87.79 & 87.72\\
\hline
CIFS & 0 & [83.86] & 60.58 & 52.64 & 51.32 & 49.94 \\
~ & 0.1 & - & [\tbf{58.86}] & 51.40 & 50.88 & 49.42\\
~ & 1 & - & 59.20 & 51.28 & [\tbf{50.16}] & 48.74\\
~ & 2 & - & 59.24 & [\tbf{51.23}] & 50.28 & 48.79 \\
~ & 10 & - & 59.35 & 51.27 & 50.70 & [\underline{\tbf{48.70}}] \\
~ & 100 & - & 59.38 & 51.41 & 51.04 & 48.80\\
~ & $\infty$ & - & 59.43 & 51.45 & 51.08 & 48.82\\
~ & $\infty$-1 & - & 61.06 & 54.96 & 53.83 & 52.82\\
~ & $\infty$-2 & - & 60.03 & 52.30 & 50.92 & 50.03\\
\bottomrule
\end{tabular}
}
\label{tab:apdx-beta-attack}
\vspace{-1em}
\end{table}{}
\subsection{Robustness Enhancement under TRADES} \label{apdx:cifar10-trades}
To improve the robustness of CNNs, various training-based strategies have been proposed, including vanilla adversarial training (AT) \cite{madry_towards_2019}, friendly-adversarial training (FAT) \cite{zhang_attacks_2020}, and TRADES \cite{zhang_theoretically_2019}. In Section \textcolor{Blue}{4.1}, we show that CIFS can further enhance the robustness of CNNs under the vanilla AT and FAT. Here, we conduct more experiments to check whether TRADES is also suitable for CIFS.
\begin{table}[h!]
\centering
\caption{Robustness comparison of vanilla CNNs and their CIFS-modified version under various AT-based strategies. We report the robust accuracies (\%) on various types of adversarial data.}
\label{tab:apdx-trades}
\vspace{1em}
\scalebox{.8}{
\begin{tabular}{ccccc}
\toprule
\tbf{\tit{ResNet-18}} & Natural & FGSM & PGD-20 & PGD-100 \\
\hline
Vanilla-AT & 84.56 & 55.11 & 46.62 & \underline{44.72} \\
Vanilla-TRADES & 83.96 & 57.09 & 50.27 & \underline{{48.83}} \\
Vanilla-FAT & 87.16 & 56.43 & 47.64 & \underline{45.35} \\
\hline
CIFS-AT & 83.86 & {58.86} & {51.23} & \underline{{48.74}} \\
CIFS-TRADES & 85.20 & 54.76 & 46.13 & \underline{43.65} \\
CIFS-FAT & 86.35 & \tbf{59.47} & \tbf{51.68} & \underline{\tbf{49.52}}\\
\bottomrule
\end{tabular}
}
\end{table}{}
From Table \ref{tab:apdx-trades}, we observe that, for the vanilla ResNet-18 model, TRADES effectively robustifies the network and outperforms its counterparts by a large margin (e.g., 48.83\% of TRADES \tit{vs.} 44.72\% of AT against PGD-100 attack). However, for the CIFS-modified models, TRADES performs worse than AT and FAT. In general, CIFS-modification in combination with the FAT training strategy achieves the best robustness against various attacks.
\section{Robustness Evaluation on SVHN} \label{apdx:svhn}
\textbf{Training and Evaluation details}: On the SVHN dataset, we train the ResNet-18 model and its CIFS-modified version with PGD-10 adversarial examples ($\epsilon=\nicefrac{8}{255}$, step size $\nicefrac{\epsilon}{4}$ with random initialization). We train models for $120$ epochs with the SGD optimizer (momentum $0.9$ and weight decay $0.0005$). The learning rate starts from $0.01$ and is multiplied with $0.1$ at epoch $75$ and epoch $90$.
In Section \textcolor{Blue}{4.1}, we evaluate the robustness of CNNs against four white-box attacks with a perturbation budget $\epsilon=\nicefrac{8}{255}$ in $l_{\infty}$ norm --- FGSM, PGD-20 (step size $\nicefrac{\epsilon}{10}$), C\&W (optimized by PGD for 30 steps with a step size $\nicefrac{\epsilon}{10}$) and PGD-100 (step size $\nicefrac{\epsilon}{10}$).
\textbf{Robustness Evaluation with AutoAttack}:
Here, we also report the robust accuracies of defense methods against AutoAttack on SVHN (Table \ref{tab:apdx-aa-svhn}). The evaluation settings of AutoAttack follows those in Appendix \ref{apdx:cifar10-at}.
\begin{table}[h!]
\centering
\caption{Robustness comparison of defense methods on SVHN. We report the robust accuracies (\%) at the last epochs.}
\label{tab:apdx-aa-svhn}
\vspace{1em}
\scalebox{.8}{
\begin{tabular}{cccc}
\toprule
\tbf{\tit{ResNet-18}} & Vanilla & CAS & CIFS \\
AutoAttack & 40.60 & 39.30 & \tbf{42.10} \\
\bottomrule
\end{tabular}
}
\end{table}{}
\tbf{Best-epoch robustness during training:} In Section \textcolor{Blue}{4.1}, we report the robust accuracies of ResNet-18 models at the last epochs during training. Here, we report the best-epoch robustness for reference (Table \ref{tab:apdx-svhn-best}). We see that CIFS modified version enjoys the better best-epoch robustness in comparison to the vanilla ResNet-18 model.
\begin{table}[h!]
\centering
\caption{Robustness comparison of defense methods on SVHN. We report the best robust accuracies (\%) during training. For each model, the results of the strongest attack are marked with an underline.}
\label{tab:apdx-svhn-best}
\vspace{1em}
\scalebox{.8}{
\begin{tabular}{cccccc}
\toprule
\tbf{\tit{ResNet-18}} & Natural & FGSM & PGD-20 & C\&W & PGD-100 \\
\hline
Vanilla & 93.88 & 66.02 & 51.71 & 48.87 & \underline{47.59} \\
CAS & 93.90 & 65.53 & 50.52 & 48.39 & \underline{46.39} \\
CIFS & 93.27 & \tbf{67.36} & \tbf{52.67} & \tbf{50.20} & \tbf{\underline{48.36}} \\
\bottomrule
\end{tabular}
}
\end{table}{}
\section{More Results on FMNIST} \label{apdx:fmnist}
\textbf{Training and Evaluation details:} On the FMNIST dataset, we train ResNet-10 with PGD-20 adversarial examples ($\epsilon=0.3$, step size $0.02$ with random initialization). The $\beta$ in CIFS is set to be $2$. We train models for $120$ epochs with the SGD optimizer (momentum $0.9$ and weight decay $0.0002$). The learning rate starts with $0.1$ and is multiplied with $0.1$ at epochs $45$, $75$ and $90$.
We evaluate the robustness of the ResNet-10 models against FGSM, PGD-20, and PGD-100 white-box attacks. The perturbation is bounded by $\epsilon=0.3$ in $l_{\infty}$ norm. The step size of PGD-20 is set to be $0.01$, and that of PGD-100 is set to be $0.02$. Here, we report both the last-epoch robust accuracies and the best-epoch robust accuracies in Table \ref{tab:apdx-fmnist-robustness}.
\begin{table}[h!]
\centering
\caption{Robustness comparison of defense methods on FMNIST. For each model, the robust accuracies (\%) of the strongest attack are remarked with an underline. For each type of attack, the best defense results are highlighted in bold. Comparing the defense rates of the strongest attacks, we observe that CIFS outperforms other defenses by a large margin.}
\label{tab:apdx-fmnist-robustness}
\vspace{1em}
\scalebox{.8}{
\begin{tabular}{cccccc}
\toprule
\tbf{\tit{Last}} & Natural & FGSM & PGD-40 & PGD-100 \\
\hline
Vanilla & 85.19 & {80.52} & 66.47 & \underline{60.99} \\
CAS & 86.59 & \tbf{82.45} & 65.58 & \underline{59.51} \\
CIFS & 83.35 & 77.48 & \tbf{66.59} & \underline{\tbf{65.50}} \\
\bottomrule
\toprule
\tbf{\tit{Best}} & Natural & FGSM & PGD-40 & PGD-100 \\
\hline
Vanilla & 85.19 & {81.21} & 67.63 & \underline{63.36} \\
CAS & 86.63 & \tbf{83.59} & 68.73 & \underline{62.65} \\
CIFS & 83.32 & 78.55 & \tbf{69.05} & \underline{\tbf{67.21}} \\
\bottomrule
\end{tabular}
}
\end{table}
\section{More Results on Ablation Study} \label{apdx:ablation}
\subsection{Effects of $\beta$ in CIFS:} \label{apdx:ablation-beta}
Here, we train CIFS-modified ResNet-18 models on CIFAR10 with various values of $\beta$ in Equation (\textcolor{Blue}{3}). The coefficient $\beta$ balances the accuracies of raw predictions and the final prediction. From Table \ref{tab:apdx-cifar-beta}, we observe that $\beta$ values that are too small or too large values lead to drops in the accuracies of natural data and adversarial data. On the one hand, if the value of $\beta$ is too small, the raw predictions made by CIFS are not reliable. Thus, the channels selected by CIFS may not be the truly relevant ones with respect to the ground-truth class. On the other hand, if the value of $\beta$ is too large, the optimization procedure mostly considers the raw predictions, the final prediction (output) becomes unreliable. When $\beta=2$, we achieve the best robustness against various types of attack.
\begin{table}[h!]
\centering
\caption{Robustness accuracies (\%) on CIFAR10 for CIFS with various values of $\beta$.}
\label{tab:apdx-cifar-beta}
\vspace{1em}
\scalebox{.8}{
\begin{tabular}{ccccc}
\toprule
\tbf{\tit{ResNet-18}} & Natural & FGSM & PGD-20 & PGD-100 \\
\hline
Vanilla & 84.56 & 55.11 & 46.62 & \underline{44.72} \\
$\beta=0.1$ & 75.22 & 53.41 & 48.10 & \underline{46.28} \\
$\beta=1$ & 82.34 &58.15 & 50.50 & \underline{48.35}\\
$\beta=2$ & 83.86 & \tbf{58.86} & \tbf{51.23} & \underline{\tbf{48.74}} \\
$\beta=10$ & 82.97 & 57.62 & 49.34 & \underline{47.10}\\
$\beta=100$ & 75.41 & 52.90 & 45.00 & \underline{43.12}\\
\bottomrule
\end{tabular}
}
\end{table}{}
\subsection{Effects of the top-$k$ feature assessment}
In general, $k$ should be larger than 1 but not too large.
If we use the top-$1$, once adv. data fool probe nets, the channels relevant to true labels will be missed, and this will lead to wrong predictions (Table \textcolor{Blue}{5}, line top-1). Instead, we use top-$2$ for reliable channel selection. The efficacy is attributed to \tit{two} aspects: {Firstly}, the top-$2$ accuracies of adv. data are usually high (see Table \textcolor{Blue}{4}), thus channels relevant to top-$2$ logits \tit{include those relevant to the true class}.
{Secondly}, \textcolor{black}{\textcolor{Blue}{Tian et al.} (\textcolor{Blue}{2021})}\footnote{\scriptsize Q. Tian, K, Kuang, F. Wu, Y. Wang, Intriguing class-wise properties of adversarial training. OpenReview. 2021} reports that CNNs' predictions of adv. data usually belong to the superclass that contains true labels. Classes (e.g., cat, dog) in the same superclass (e.g., animals) usually share similar semantic features. Thus, {most of the top-$2$ selected channels are useful} for predicting the true class.
Although the top-$2$ selected channels may contain info about the other wrong class, the following layers (after CIFS) are capable of “purifying” features and make better predictions. This is verified by Table \textcolor{Blue}{5}, the results in the line top-2 (CIFS/CIFS 48.72\% vs. CIFS/Final 54.96\%) mean that around 6\% adv. data, which successfully fool probes, are still finally correctly classified. However, too large $k$ may degrade the relevance assessment due to too much noisy info (e.g., the effect of top-3 is worse than top-2 in Table \textcolor{Blue}{5}).
\subsection{Layers to be modified}
\tbf{Positions of CIFS modules}: Here, we try different combinations of the layers to be modified by CIFS. In CNNs, the features of deep layers are usually more characteristic in comparison to those in the shallower layers \cite{zeiler_visualizing_2014}, and each channel of the features captures a distinct view of the input. The predictions often depend only on the information of a few essential views of the inputs. CIFS improves adversarial robustness by adjusting channel-wise activations. Thus, we apply CIFS to the deeper layers instead of the shallower ones. Specifically, we modify the ResNet-18 by applying CIFS at the last (P1) and/or the second last (P2) residual blocks. The experimental results are reported in Table \ref{tab:positions}. We observe that simultaneously applying CIFS into P1 and P2 performs the best against various attacks. Intuitively, because the features can be progressively refined, applying CIFS at P1\&P2 better purifies the channels compared to applying it only at P1 or P2.
\begin{table}[h]
\centering
\caption{Robustness (\%) comparison of the positions where CIFS modules are placed.}
\scalebox{.8}{
\begin{tabular}{cccccc}
\toprule
\tbf{\tit{ResNet-18}} & Natural & FGSM & PGD-20 & PGD-100\\
\hline
Vanilla & 84.56 & 55.11 & 46.62 & \underline{44.72} \\
P1 & 84.02 & 57.60 & 48.45 & \underline{45.95} \\
P2 & 82.62 & 56.55 & 47.22 & \underline{44.81} \\
P1-P2 & 83.86 & \tbf{58.86} & \tbf{51.23} & \underline{\tbf{48.74}} \\
\bottomrule
\end{tabular}
}
\label{tab:positions}
\end{table}{}
\subsection{Architecture of Probe Networks}
\tbf{Linear vs. Non-linear Probe}: For a certain layer modified by CIFS, the probe network in CIFS serves as the surrogate classifier of the subsequent layers in the backbone model. Thus, the probe networks should be powerful enough to make correct predictions based on the features of this layer. For the CIFS in the last residual block, we use a linear layer network as the probe, while for the CIFS in the second last residual block, we compare the cases of using a linear layer versus using a two-layer MLP network. From Table \ref{tab:linear}, we observe that the MLP-Linear combination shows a similar performance compared to the combination of two linear layers against adversarial attacks, but enjoys a clear advantage on the natural data ($83.86$\% vs. $81.52$\%). This is because the features in the second last residual block are not as characteristic as those in the last block and cannot be linearly separated. The MLP can thus classify the features better than a pure linear layer.
\begin{table}[h]
\centering
\caption{Robustness comparison (\%) of the probe architectures in CIFS modules (at P1-P2).}
\vspace{.5em}
\scalebox{.8}{
\begin{tabular}{cccccc}
\toprule
\tbf{\tit{ResNet-18}} & Natural & FGSM & PGD-20 & PGD-100\\
\hline
Vanilla & 84.56 & 55.11 & 46.62 & \underline{44.72} \\
Linear-Linear & 81.52 & 58.33 & \tbf{51.32} & \underline{\tbf{49.07}} \\
MLP-Linear & 83.86 & \tbf{58.86} & 51.23 & \underline{48.74} \\
\bottomrule
\end{tabular}
}
\label{tab:linear}
\end{table}{}
\section{Experiments} \label{sec:exp}
\subsection{Robustness Evaluation} \label{sec:exp-eva}
We utilize the CIFS to modify CNNs in different architectures to perform classification tasks on benchmark datasets, namely a ResNet-18 and a WideResNet-28-10 on the CIFAR10 \cite{krizhevsky_learning_nodate} dataset, a ResNet-18 on the SVHN \cite{netzer_reading_nodate} dataset and
a ResNet-10 on the Fashion-MNIST \cite{xiao2017/online} dataset. We train the models with the standard PGD adversarial training (AT) \cite{madry_towards_2019} and its variants, such as FAT \cite{zhang_attacks_2020}, to show that CIFS can work under various AT-strategies. We compare CIFS-modified CNNs with the vanilla versions as well as the CAS-modifications, where CAS \cite{bai_improving_2021} also modifies CNNs by adjusting channels' activations. Here, we report the results on CIFAR10 and SVHN. Results on FMNIST are presented in Appendix \ref{apdx:fmnist}.
\paragraph{Adaptive Attacks} As mentioned in Section \ref{sec:cifs-mechanism}, each CIFS-modified layer of CNNs outputs a raw prediction. To generate adversarial examples that are as strong as possible, we follow the strategy used in CAS and attack CNNs via the adaptive loss function $\ell_{\beta}$ in Equation (\ref{eq:adaptive-loss}) that considers all the raw and final predictions. We let the value of $\beta$ be chosen by the attacker, i.e., the attacker can try various values of $\beta_{\text{atk}}$ and select one that results in the most harmful perturbations. Our setting is more challenging for defense than CAS where the \tit{same} value of $\beta$ is used for training and attack. Here, for each adversarial attack, we evaluate the robustness by choosing $\beta_{\text{atk}}$ from $\{0, 0.1, 1, 2, 10, 100, \infty\}$ and report the worst robust accuracy\footnote{Results of various values of $\beta$ are present in Appendix \ref{apdx:cifar10-at}. We observe that CAS can improve the robustness of CNNs in most cases but \tit{fail} when attackers completely focus on CAS modules.}. Setting $\beta_{\text{atk}}=\infty$ means the attacks completely focus on the CIFS-modified layers \footnote{For $\beta_{\text{atk}}=\infty$, we consider the cases of attacking both CIFS-modified layers simultaneously and attacking each separately.} and only consider the second term in Equation (\ref{eq:adaptive-loss}).
\subsubsection{Robustness Enhancement of CIFS under AT}
We adversarially train ResNet-18 and WRN-28-10 models with PGD-10 ($\epsilon=\nicefrac{8}{255}$) adversarial data.
CIFS is applied to the last two residual blocks of each model. The probes for the last and penultimate blocks are a linear layer and a multi-layer perceptron (MLP) respectively. Channels' relevances are assessed based on top-$2$ results and we use the softmax function with $T=1$ as the IMGF.
Other training details are provided in Appendix \ref{apdx:cifar10} and \ref{apdx:svhn}.
\paragraph{Defense Results} We evaluate the robustness of CNNs against four types of white-box attacks: FGSM \cite{szegedy_intriguing_2014}, PGD-20 \cite{madry_towards_2019}, C\&W \cite{carlini_towards_2017}, and PGD-100. The $l_{\infty}$-norm of the perturbations are bounded by the value of $\epsilon=\nicefrac{8}{255}$. Here, we report the robustness evaluated at the last epoch for each model. Detailed attack settings and more defense results~( AutoAttack\footnote{AutoAttack consists of both white and black-box attacks.}~\cite{croce_reliable_2020} and the best epochs' results), are present in Appendix \ref{apdx:cifar10} and \ref{apdx:svhn}.
The defense results on CIFAR10 are reported in Table \ref{tab:white-box-cifar}. We observe that, for both of the ResNet-18 and WRN-28-10 architectures, CIFS consistently outperforms the counterparts against various types of adversarial attacks. For example, the CIFS-modified WRN-28-10 can defend the PGD100 attack with a success rate of $48.74\%$, which exceeds the second best by more than 4 percentage points. In contrast, under the strong adaptive attack, we see that the baseline CAS cannot improve and even worsens the robustness of CNNs. The defense results on SVHN are reported in Table \ref{tab:white-box-svhn}. The results also verify the effectiveness of CIFS on improving robustness.
\begin{table}[h]
\vspace{-1em}
\centering
\caption{Robustness comparison of defense methods on CIFAR10. We report the accuracies (\%) for adversarial and natural data. For each model, the results of the strongest attack are marked with an underline.
}
\vspace{.5em}
\scalebox{.8}{
\begin{tabular}{cccccc}
\toprule
\tbf{\tit{ResNet-18}} & Natural & FGSM & PGD-20 & C\&W & PGD-100 \\
\hline
Vanilla & 84.56 & 55.11 & 46.62 & 45.95 & \underline{44.72} \\
CAS & 86.73 & 55.99 & 45.29 & 44.18 & \underline{43.22} \\
CIFS & 83.86 & \tbf{58.86} & \tbf{51.23} & \tbf{50.16} & \underline{\tbf{48.70}} \\
\bottomrule
\toprule
\tbf{\tit{WRN-28-10}} & Natural & FGSM & PGD-20 & C\&W & PGD-100 \\
\hline
Vanilla & 87.29 & 58.50 & 49.17 & 48.68 & \underline{47.08} \\
CAS & 88.05 & 57.94 & 49.03 & 47.97 & \underline{47.25} \\
CIFS & 85.56 & \tbf{61.34} & \tbf{53.74} & \tbf{53.20} & \underline{\tbf{51.51}} \\
\bottomrule
\end{tabular}
}
\label{tab:white-box-cifar}
\vspace{-1em}
\end{table}{}
\begin{table}[h]
\centering
\caption{Robustness comparison of defense methods on SVHN. The accuracies (\%) for natural and adversarial data are reported.
}
\vspace{.5em}
\scalebox{.8}{
\begin{tabular}{cccccc}
\toprule
\tbf{\tit{ResNet-18}} & Natural & FGSM & PGD-20 & C\&W & PGD-100 \\
\hline
Vanilla & 93.72 & \tbf{65.87} & 50.35 & 47.89 & \underline{45.81} \\
CAS & 94.08 & 65.24 & 48.47 & 46.15 & \underline{43.75} \\
CIFS & 93.94 & 66.24 & \tbf{52.02} & \tbf{50.13} & \underline{\tbf{47.49}} \\
\bottomrule
\end{tabular}
}
\label{tab:white-box-svhn}
\vspace{-.5em}
\end{table}{}
\paragraph{Training Procedure} We train CNN classifiers in an adversarial manner for $120$ epochs and adjust the learning rate with a multiplier $0.1$ at epoch $75$ and epoch $90$. We summarize the training procedure by plotting the curves of training losses and the PGD-20 accuracies \tit{w.r.t.} epochs in Figure \ref{fig:tr_procedure}. We observe the best adversarial robustness of the vanilla ResNet-18 ($50.64\%$ defense rate) appears around the $75^{\text{th}}$ epoch. After epoch $75$, the model starts to overfit to training data, i.e., the training loss continues decreasing, but the robust accuracy drops as well. In contrast, the overfitting problem is ameliorated by the application of CIFS. We can see that the best robust accuracy appears around the $90^{\text{th}}$ epoch; After the best epoch, the training loss continues decreasing, but the robustness is maintained around the peak. This phenomenon may result from the fact that CIFS suppresses redundant channels. The model redundancy can be controlled by selecting few highly relevant channels and deactivating others. In this way, the overfitting in training is ameliorated.
\begin{figure}[h!]
\vspace{-.3em}
\centering
\includegraphics[width=0.8\linewidth]{images/training/training.png}
\vspace{-1em}
\caption{Comparison on the training of a vanilla ResNet-18 and its CIFS-modified version. Training losses and accuracies against PGD-20 attack are plotted.}
\label{fig:tr_procedure}
\end{figure}
On the computation overhead, we report the training time and the evaluation time of a ResNet-18 classifier on the CIFAR10 dataset for reference. For PGD-10 adversarial training, the vanilla CNN takes 166s for each epoch while the CIFS-modified model takes 172s. For the PGD-20 evaluation, the vanilla CNN takes 53s for the CIFAR10 test set; the CIFS-modified model takes 56s instead. In short, the proposed CIFS does not result in too much extra computation.
\subsubsection{CIFS Working in Conjunction with Variants of AT}
CIFS improves the adversarial robustness by adjusting channels' activations, which is orthogonal to defense training strategies. Here, we train the CIFS-modified CNNs with variants of AT to examine whether CIFS can work in conjunction with other state-of-the-art training-based defense techniques. We consider the FAT \cite{zhang_attacks_2020} and TRADES \cite{zhang_theoretically_2019} strategies.
We report the defense results of FAT in Table \ref{tab:fat} and provide the results of TRADES in Appendix \ref{apdx:cifar10-trades}. We observe that FAT training strategy improves the natural accuracy and robustness of CNNs (compared to the results in Table \ref{tab:white-box-cifar}). Under the FAT strategy, CIFS also improves the adversarial robustness of the vanilla CNNs in both ResNet-18 and WRN-28-10 architectures.
\begin{table}[h]
\vspace{-1em}
\centering
\caption{Robustness comparison of defense methods trained with FAT on CIFAR10. Comparing the accuracies (\%) against the strongest attacks, we observe that CIFS clearly robustifies CNNs.}
\vspace{.5em}
\scalebox{.8}{
\begin{tabular}{cccccc}
\toprule
\tbf{\tit{ResNet-18}} & Natural & FGSM & PGD-20 & C\&W & PGD-100 \\
\hline
Vanilla & 87.16 & 56.43 & 47.64 & 46.01 & \underline{45.35} \\
CIFS & 86.35 & \tbf{59.47} & \tbf{51.68} & \tbf{51.84} & \underline{\tbf{49.52}} \\
\bottomrule
\toprule
\tbf{\tit{WRN-28-10}} & Natural & FGSM & PGD-20 & C\&W & PGD-100 \\
\hline
Vanilla & 88.37 & 58.81 & 49.62 & 48.49 & \underline{47.58}\\
CIFS & 86.74 & \tbf{60.67} & \tbf{51.99} & \tbf{52.34} & \underline{\tbf{49.87}}\\
\bottomrule
\end{tabular}
}
\label{tab:fat}
\end{table}{}
\subsection{Ablation Study} \label{sec:exp-ablation}
Here, we conduct an ablation study to further understand the robustness properties of CIFS. Specifically, we investigate the effects of the feedback from the top-$k$ predictions. The ablation experiments are conducted on CIFAR10 based on the ResNet-18 model. Besides, in Appendix \ref{apdx:ablation}, we also study cases in which the CIFS is applied to different layers, the probe networks are in different architectures, various values of $\beta$ are used for training.
\paragraph{Feedback from Top-$k$ Prediction Results}
As is well-known, the top-$k$ classification accuracy for $k>1$ is always not worse than the top-$1$.
For example, in Table \ref{tab:topk}, we can see that the top-$2$ accuracy of an adversarially trained ResNet-18 against the PGD-20 attack exceeds the top-$1$ accuracy by $25$ percentage points. This implies that, although adversarial data can usually fool the classifier (i.e., low top-$1$ accuracy), the prediction confidence of the true class is still high and the corresponding score highly likely lies among the top-$2$ or $3$ logits.
\begin{table}[h]
\vspace{-1em}
\centering
\caption{Top-$k$ accuracies (\%) against adversarial attacks on CIFAR10 of an adversarially trained ResNet-18.}
\vspace{.5em}
\scalebox{.8}{
\begin{tabular}{cccccc}
\toprule
\tbf{\tit{ResNet-18}} & top-$1$ & top-$2$ & top-$3$\\
\hline
FGSM & 55.11 & 76.22 & 85.20 \\
PGD-20 & 46.62 & 71.71 & 81.60 \\
\bottomrule
\end{tabular}
}
\label{tab:topk}
\end{table}{}
CIFS generates the importance mask from the raw prediction and uses it to suppress or promote channels at the current layer. The final prediction made by subsequent layers strongly depends on the channels selected by CIFS. To ensure the accuracy of final predictions, the logits used for generating importance scores should include the true label's logit for each input so that the truly important channels will be highlighted. According to Table \ref{tab:topk}, if we use the top-$1$ logit to assess the importance of channels for PGD-20 adversarial data, the probability of incorrect assessment is over $50$\%. Instead, if we use the feedback from top-$2$ or $3$ logits, the truly important channels can highly likely be promoted. The following table presents more experiments that justify this argument.
\begin{table}[h]
\vspace{-1em}
\centering
\caption{Robustness comparison (\%) of importance assessment based on top-$k$ results against the PGD-20 attack. The column header $*/\#$ represents the attack point and the output prediction. For example, CIFS/Final means the model outputs the final prediction and the attacker solely focuses on the CIFS's raw prediction.}
\vspace{.5em}
\scalebox{.8}{
\begin{tabular}{cccccc}
\toprule
\tbf{\tit{ResNet-18}} & Natural/Final & CIFS/CIFS & CFIS/Final & Adap/Final\\
\hline
Vanilla & 84.56 & - & - & 46.62 \\
top-$1$ & 87.63 & 47.24 & 47.24 & {47.24} \\
top-$2$ & 83.86 & 48.72 & 54.96 & {\tbf{51.23}} \\
top-$3$ & 83.49 & 47.59 & 55.39 & {49.91} \\
\bottomrule
\end{tabular}
}
\label{tab:purify}
\end{table}{}
From Table \ref{tab:purify}, we observe that, for the top-$1$ case, the defense rate of `CIFS/CIFS' is the same as that of `CIFS/Final' and that of `Adap/Final.' This implies that, once an adversarial example successfully fools the raw prediction of CIFS, the final prediction also will be incorrect. Thus, the attacker only needs to focus on the CIFS's raw predictions to break the model. In contrast, for the top-$2$ case, the defense rate of `CIFS/Final' exceeds the `CIFS/CIFS' by 6 percentage points. This means that nearly 6\% adversarial data mislead the CIFS's raw predictions. However, through the channel adjustment via CIFS, these adversarial data are ``purified'', and more relevant characteristic features are thus transmitted to subsequent layers of CNNs. As such, these adversarial data are finally classified correctly. In this case, the attacker has to exhaustively search for an adaptive loss function to generate attacks, and the CIFS-modified CNNs are safer and more reliable. More discussion on why the top-$k$ assessment performs better and how to choose $k$ is provided in Appendix \ref{apdx:ablation}.
\section{Channel-wise Importance-based Feature Selection}
In this section, we first study the adversarial robustness by comparing channels' activations of non-robust (normally trained) and robustified (adversarially trained) CNNs. Based on our observations of AT's effects, we propose a hypothesis on robustness enhancement via the adjustment of channels' activations (Section \ref{sec:cifs-adv-effects}). To examine this hypothesis, we then develop a novel mechanism, CIFS (Section \ref{sec:cifs-mechanism}), to manipulate channels' activation levels according to their relevances to predictions . Finally, we verify the proposed hypothesis through extensive experiments (Section \ref{sec:cifs-verify}).
\subsection{Non-robust CNNs vs. Robustified CNNs: a Channel-wise Activation Perspective} \label{sec:cifs-adv-effects}
We compare a non-robust ResNet-18 \cite{he_deep_2016} model with an AT-robustified one on the CIFAR10 dataset \cite{krizhevsky_learning_nodate}.
In ResNet-18, the representations of penultimate layer are spatially averaged for each channel, then fed into the last linear layer for making predictions. Thus, the weights of the last linear layer indicate channels' relevances to predictions (according to the definition of channels' relevances in Introduction).
We visualize the channel-wise activation magnitudes, the activated frequencies (counted via a threshold of 1\% of the largest magnitude among all channels) in the penultimate layer for both natural and adversarial data, as well as the weights of the last linear layer in Figure~\ref{fig: w-vs-act}. The details of implementation are provided in Appendix~\ref{apdx:visualization-normal-at}.
From Figures~\ref{fig: w-vs-act_normal-auto} and \ref{fig: w-vs-act_normal-cat}, we observe, for a non-robust ResNet-18, the activation distribution of the adversarial data is obviously mismatched with that of the natural data: natural data activate channels that are PR to predictions with high values and high frequency, while adversarial data tend to amplify the NR ones.
From Figures~\ref{fig: w-vs-act_adv-auto} and \ref{fig: w-vs-act_adv-cat}, we observe that AT robustifies the model by aligning the activation distribution of adversarial data with that of natural data. Specifically, when dealing with adversarial data, AT boosts the activation magnitudes of PR channels while suppressing the activations of NR ones. However, we observe that, for many NR channels (e.g., around 150 channels from $350^{\text{th}}$ to $512^{\text{th}}$), the activations of adversarial data are much higher than those of natural data. These over-activations decrease the prediction scores corresponding to their true categories. Given this observation, we wonder \textcolor{OrangeRed}{(Q1)}: \textit{if we suppress these NR channels to regularize the freedom of adversarial perturbations, will it further improve the model's robustness upon AT}?
Besides, an adversarially trained model does not enjoy similar robustness for all classes, i.e., the robust accuracy of a certain class may be much higher than another (e.g., ``automobile'' with 69.0\% vs. ``cat'' with 16.7\% against PGD-20). Comparing the activations of these two classes (Figures~\ref{fig: w-vs-act_adv-auto} and \ref{fig: w-vs-act_adv-cat}), we observe that, for the class with strong robustness (e.g., ``automobile''), channels' activations align better with their relevances to labels, i.e., the channel with a greater extent of activation usually corresponds to a larger weight in the linear layer. In contrast, this alignment does not hold for the class with relatively poor robustness (e.g., for class ``cat'', the most activated channels, lying between the $26^{\text{th}}$ and $125^{\text{th}}$, are sub-PR to predictions).
Given this phenomenon, we may ask another question \textcolor{OrangeRed}{(Q2)}: \textit{{If we scale channels' activations based on their relevances to predictions, will it improve the model's robustness?}}
Considering the two questions above, we propose the unified hypothesis $\mathcal H$, as stated in the Introduction.
\subsection{Importance-based Channel Adjustment} \label{sec:cifs-mechanism}
To examine the hypothesis $\mathcal H$, one needs a systematic approach to manipulate the channels, viz. selecting channels via suppressing NR ones but promoting PR ones. To this end, we introduce a mechanism, dubbed as Channel-wise Importance-based Feature Selection (CIFS). CIFS modifies layers of CNNs by adjusting channels' activations with importance scores that are generated from the channels' relevances to predictions.
For clearly state CIFS, we first introduce some notations: For a $K$-category classification problem, let $(X,Y) \sim P_{XY} $ denote the pair of the random input and its label, where $X\in \mathcal{X} \subset {\mathbb{R}}^{n_X}$ and $Y\in \mathcal{Y} = \{0,1,\dots,K-1\}$. We design an $L$-layer CNN-based classification model to make accurate predictions for data sampled from $P_{XY}$. The $l^{\text{th}}$ layer is denoted by $f^{l}(\cdot)$ and parametrized by $\theta^{l} \in \Theta^l$; the mapping from the input to the $l^{\text{th}}$ layer's output is denoted by $f^{[l]}=f^{l}\circ f^{l-1}\circ \cdots \circ f^{1}$ and the combination of all the first $l$ layers' parameters is denoted by $\theta^{[l]}$, i.e., $\theta^{[l]}=(\theta^1, \dots, \theta^l)$.
Let us examine the $l^{\text{th}}$ layer where an input $x$ is transformed into a high-dimensional representation $z^l=f^{[l]}(x) \in {\mathbb{R}}^{n^l_{\text{C}} \times n^l_{\text{F}}}$; $z^l$ has $n^l_{\text{C}}$ \underline{c}hannels and each channel is a \underline{f}eature vector of length $n^l_{\text{F}}$. With these notations, we elaborate the details of CIFS in \tit{\tbf{{three}}} steps (as shown in Figure \ref{fig:CIFS}).
\begin{figure}[t!]
\vspace{-.5em}
\centering
\includegraphics[width=1 \linewidth]{images/cifs/picture}
\vspace{-1.5em}
\caption{CIFS: \tit{1)}~Probe Network $A^l$ first makes a raw prediction $p^l$ for $z^l$. \tit{2)}~Channels' relevances $g^l$ are assessed (Relev.~Ass.) based on the gradients of the top-$k$ prediction results $y^{l,k}$. \tit{3)}~The IMGF generates an importance mask $m^l$ from $g^l$ for channel adjustment.
}
\label{fig:CIFS}
\vspace{-1.5em}
\end{figure}
\textcolor{Blue}{1) Surrogate Raw Prediction:}~To assess channels' relevances to predictions,
a naive strategy is to compute the gradients of the final prediction with respect to $z^l$, i.e., $\nabla_{z^l} f^{[l+1:L]}(z^l)$, where $f^{[l+1:L]}=f^L \circ \cdots \circ f^{l+1}$. Since we need to adjust $z^l$ with importance scores generated from $\nabla_{z^l} f^{[l+1:L]}(z^l)$ and send the adjusted feature $\bar z^l$ to $f^{[l+1:L]}$ again for making the final prediction, it will result in computing the second-order derivatives during the training phase. Moreover, in practice, we may apply CIFS into multiple layers,
the forward pass will involve at least the second-order gradients (the latter CIFS-modified layer is recursively called). Thus, the back-propagation has to deal with at least the third-order gradients during training. This will aggravate the problem of training instability.
Instead, inspired by the design of auxiliary classifiers in CAS \cite{bai_improving_2021}, CIFS builds a probe network $A^l$ as the surrogate of $f^{[l+1:L]}$ for a making raw prediction $p^l = A^l(z^l)$, so that we can use the gradients of $p^l$ to approximately assess the channels' relevances to the final prediction. The assessment does not involve other CIFS-modified layers. Thus, we can avoid the problem of back-propagation through high-order derivatives. The probe network $A^l$ is parameterized by $\theta^l_A$ and $p^l\in {\mathbb{R}}^K$ represents the vector of prediction scores/logits. We can jointly optimize $A^l$ with the backbone network during the training phase under the supervision of true labels.
\textcolor{Blue}{2) Relevance Assessment:}~
With the prediction $p^l$, we can compute the gradients of logits in $p^l$ \tit{w.r.t.} $z^l$ to assess the feature's relevances to each class.
We consider the top-$k$ prediction results ($k\geq 2$) for the assessment of channels' relevances. As data from two semantically similar classes (e.g., ``dog'' and ``cat'') usually share common features, the prediction for an input often assigns large scores to the classes similar to the true one and the top-$k$ results may include several of these similar classes \cite{jia_certified_2020}. In case the top-$1$ prediction is wrong, considering the top-$k$ results may help us reliably extract some common relevant features (see Section \ref{sec:exp-ablation} for more evidence).
Let $y^{l,k}$ denote indices of the $k$ largest logits of prediction $p^l$.
Let $\delta \in {\mathbb{R}}^{n^l_{\text{C}}}$ be the channel-wise perturbation added to $z^l$, giving the perturbed representation $z^l_{\delta} = z^l + \delta \cdot \mathbf{1^{\top}}$. Here $\mathbf{1} \in {\mathbb{R}}^{n^l_{\text{F}}}$ is the column vector with all elements as one, i.e., the features in the same channel are perturbed by a common value. We calculate the gradients of the sum of the top-$k$ logits \textit{w.r.t.} the channel-wise perturbation $\delta$:
\vspace{-0.5em}
\begin{equation}
\small
g^l
= \left.\nabla_{\delta} \sum_{i \in y^{l,k}} p^l(\delta)_{[i]} \right\vert_{\delta=\mathbf{0}}
= \left.\nabla_{z^l_{\delta}} \sum_{i \in y^{l,k}} A^l(z^l_{\delta})_{[i]} \right\vert_{z^l_{\delta} =z^l} \cdot \mathbf{1},
\label{eq:relevance-assessment}
\end{equation}
where $g^l=(g^l_{[0]},...,g^l_{[n^l_{\text{C}}-1]})$
represents the vector of channels' relevances to the top-$k$ logits. During the training phase, since the true label of $z^l$ is given, we replace the top-$1$ prediction in $y^{l,k}$ with the true label $y$ and keep other prediction results untouched.
\textcolor{Blue}{3) Importance Mask Generation:}~
As we want to suppress or promote channels based on their relevances, we need to design proper Importance Mask Generating Functions (IMGFs), which monotonically map relevances to \textit{non-negative} importance scores; of particular importance is to map negative relevances to values close to zero.
Here, we provide several feasible options: To answer the first question on whether suppressing NR channels enhances robustness, one can use the sigmoid function as the IMGF. With a large value of $\alpha$, the sigmoid function serves as a switch by mapping negative relevances to importance scores close to zero and the positive close to one.
To answer the second question concerning aligning channel activations with their relevances, we can use the softplus or softmax function as the IMGF. Both of them can map negative relevacnes to values close zero and map positive relevances monotonically to positive values.
These three functions are stated here for ease of reference:
\begin{itemize}
\vspace{-.8em}
\item sigmoid: $m^l_{[i]} = \frac{1}{1+\exp(-\alpha \cdot g^l_{[i]})}, \quad \alpha>0$. \vspace{-0.4em}
\item softplus: $m^l_{[i]} = \frac{1}{\alpha} \cdot \log(1+\exp (\alpha \cdot g^l_{[i]})), \quad \alpha>0$. \vspace{-0.4em}
\item softmax: $m^l_{[i]} = \frac{\exp(g^l_{[i]}/T)}{\sum_{j} \exp(g^l_{[j]}/T)}, \quad T>0$.
\vspace{-.8em}
\end{itemize}
The usage of these functions will be discussed in detail in Section~\ref{sec:cifs-verify}. CIFS selects channels by multiplying the importance mask $m^l$ with $z^l$ as follows:
\begin{equation}
\small \bar z^l = z^l \otimes \text{repmat}(m^l, 1, n^l_L),
\end{equation}
where the ``$\text{repmat}$'' operation replicates the column vector $m^l$ along the second axis $n^l_L$ times.
\paragraph{Training of CIFS} In practice, we may apply the CIFS mechanism into several layers of a CNN. Let $I$ denote the set of indices of these layers, and $\theta_A^{I}$ denote the parameters of all the probes in the CIFS-modified layers. For each input $x$, the modified model $\bar f^{[L]}$ outputs $|I|$ raw predictions and one final prediction $p=\bar f^{[L]}(x)$. Given this, we use an adaptive loss function \cite{bai_improving_2021} for training the model:
\begin{equation}
\small
\ell_{\beta}(x,y) = \frac{1}{1+\beta} \cdot \ell_{\mathrm{ce}}(p,y) + \frac{\beta}{(1+\beta)|I|} \cdot {\sum\limits_{l\in I} \ell_{\mathrm{ce}}(p^l, y)},
\label{eq:adaptive-loss}
\vspace{-1em}
\end{equation}
where $\ell_{\text{ce}}(\cdot, \cdot)$ denotes the cross-entropy loss and the coefficient $\beta>0$ balances the accuracy of raw predictions by CIFS and the final prediction. Since the subsequent decisions closely depend on the channels of features selected by the previous CIFS-modified layers, we should choose a proper value of $\beta$ to make sure that the raw predictions made by CIFS are reliable. In practice, we set $\beta$ to be $|I|$, and the effect of $\beta$ is discussed in the ablation study (see Appendix \ref{apdx:ablation}). To robustify the CNN model against malicious attacks, we can train $\bar f^{[L]}$ in an adversarial manner with a perturbation budget $\epsilon$. Namely, we solve the following optimization problem:
\begin{equation}
\small
\min_{\theta^{[L]}, \theta^{I}_A} \mathbb{E}_{P_{XY}} \left[ \max_{X'\in \mathcal{B}(X, \epsilon, l_{\infty})} \ell_{\beta}(X',Y) \right],
\label{eq:adv_cifs}
\vspace{-.5em}
\end{equation}
where $\mathcal{B}(x, \epsilon, l_{\infty})=\{x' ~|~ \|x'-x\|_{l_{\infty}} \leq \epsilon \}$.
\input{_fig_hyp}
\subsection{Verification of Hypothesis $\mathcal H$ on Robustness Enhancement} \label{sec:cifs-verify}
We verify the hypothesis $\mathcal H$ by answering the two questions, \textcolor{OrangeRed}{Q1} and \textcolor{OrangeRed}{Q2}, in Section~\ref{sec:cifs-adv-effects} respectively.
To answer \textcolor{OrangeRed}{Q1}, we applied the sigmoid function to generate the mask $m^l$ from $g^l$. Setting $\alpha$ to be large enough (here, $\alpha=10$), we can generate importance scores close to zeros for NR channels and scores close to one for the PR ones, so that we approximately annihilate the NR channels but leave the PR ones as unchanged. We adversarially trained a ResNet-18 model and its CIFS-modified version. As shown in Figure~\ref{fig:hyp-1-non-CIFS}, the NR channels (Channel 300-512) of the vanilla ResNet-18 model are clearly activated (the average activation magnitudes of these channels are larger than 0.1; the activation frequencies are over 0.4). In contrast, the ResNet-18 with CIFS-sigmoid effectively suppresses the activation of NR channels (Channel $100^{\text{th}}$--$512^{\text{th}}$ in Figure~\ref{fig:hyp-1-sigmoid}). Most of their mean activation magnitudes are smaller than 0.05, and their activation frequencies have clearly decreased. The experimental results show that, under AT, the vanilla ResNet-18 model results in a \underline{46.64\%} defense rate against the PGD-20 attack while its CIFS-sigmoid modified version achieves \underline{49.87\%}. More results can be found in Appendix \ref{apdx:visualization-cifs-modification}. Thus, we conclude that suppressing NR channels enhances the robustness of CNNs.
To answer \textcolor{OrangeRed}{Q2}, we applied the softplus and softmax functions as IMGFs to generate the mask $m^l$ from $g^l$ respectively. Here the coefficient $\alpha$ in the softplus is set to be $5$ and the temperature $T$ in the softmax is set to be $1$.
From Figures~\ref{fig:hyp-2-softplus} and \ref{fig:hyp-2-softmax}, we observe that, by generating importance scores positively correlated with the relevances, the model tends to completely focus on few relevant (positive and negative) channels. The channel of the greatest weight (most PR to predictions) is activated with the highest magnitude.
Most channels become irrelevant (small absolute values of weights) to the predictions and are activated at a low level.
Using softplus as the IMGF (Figure~\ref{fig:hyp-2-softplus}),
the irrelevant channels are sparsely activated, and the activation magnitudes are smaller than 5\% of the most important channel.
In Figure~\ref{fig:hyp-2-softmax}, this phenomenon is enhanced by using softmax as IMGF: most channels become irrelevant to predictions and are usually deactivated. We evaluated the robustness of these two CIFS-modified CNNs against PGD-20 attack. Both of them outperformed the CIFS-sigmoid and the vanilla ResNet-18 classifiers (robust accuracies of CIFS-softplus and CIFS-softmax are \underline{50.38\%} and \underline{51.23\%} respectively, \tit{vs.}, \underline{49.87\%} for CIFS-sigmoid and \underline{46.64\%} for the vanilla ResNet-18). We also found CIFS can ameliorate the class-wise imbalance of adversarial robustness (\tit{e.g.}, CIFS-softmax increases the PGD-20 accuracy from 16.7\% to 22.3\% for class ``cat''). More details are provided in Appendix \ref{apdx:visualization-cifs-modification}. Thus, we conclude that aligning channels activations with their relevances to predictions can further robustify CNNs upon suppressing NR ones.
Given these empirical results, we verified the hypothesis $\mathcal H$ and justified that CIFS is an effective mechanism to improve the adversarial robustness of CNNs. In the following section, we conduct extensive experiments to evaluate the robustness enhancement through CIFS and study CIFS in an ablation manner.
|
\section{Introduction}
This Introduction is structured as follows:
in Subsection \ref{Se:Scope}, we present the scope of our article; in Subsection \ref{Se:Goal}, we give an informal statement of our results while in Subsection \ref{Se:Biblio} we give several bibliographical references on qualitative properties for optimal control problems, shape derivatives for parabolic problems and quantitative inequalities.
In Subsection \ref{Se:Schwarz}, we give basic information regarding the Schwarz rearrangement, which will be a key tool in our analysis, and we give bibliographical references for parabolic isoperimetric inequalities.
In Subsection \ref{Se:Results}, we state our main results, Theorems \ref{Th:Td} and \ref{Th:Ti} (Theorem \ref{Th:Uniqueness} deals with the uniqueness of solutions to our optimal control problem and is also stated in this Section).
In Subsection \ref{Se:Plan} we present the plan of our paper and, finally, in Subsection \ref{Su:Notation}, we gather the notations we will use throughout the paper.
\subsection{Scope of the article}\label{Se:Scope}
\subsubsection{Goal of this article: informal statement of the problems and of the results}\label{Se:Goal}
In this article, our goal is to present two different approaches for obtaining \textit{quantitative inequalities for optimal control problems}, which will also be dubbed \emph{quantitative isoperimetric parabolic inequalities}. Before explaining how this fits in the growing field of qualitative questions in optimal control theory, let us vaguely state the type of results we wish to establish, and sketch the two approaches that will be put forth. By quantitative inequalities, we mean the following: we consider a controlled parabolic partial differential equation assuming the general form
\begin{equation}\label{Eq:Intro}
u_t-\mathcal Lu=f \text{ in }(0;T)\times {\Omega},
\end{equation}
$\mathcal L$ being an elliptic operator; this equation is
supplemented with some initial condition and some boundary conditions. In this setting, $f$ is the control and depends \textit{a priori} both on time and space. It is assumed to satisfy some constraints, which will be taken into account by assuming that $f\in \mathscr X$, where $\mathscr X$ is some subset of a function space. The cost to be optimised is some functional $\mathcal J_T:\mathscr X \ni f\mapsto \mathcal J_T(f)$. The control problems reads
\begin{equation}\label{Eq:PvIntro}
\fbox{$\displaystyle \max_{f\in \mathscr X}\mathcal J_T (f).$}\end{equation}
The quantitative inequality we aim at can take two different forms:
\begin{itemize}
\item \textbf{For time independent controls.} In the context where all controls $f\in \mathscr X$ write $f=f(x)$, and if the solution of \eqref{Eq:PvIntro} is some $\overline f$ (assumed to be unique for simplicity), the goal is to establish the following kind of estimate
\begin{equation}\label{Eq:IntroIndep}\fbox{$\displaystyle \forall f \in \mathscr X\,,
\mathcal J_T(f)-\mathcal J_T(\overline f)\leq - C(T)\Vert f-\overline f\Vert_{L^1({\Omega})}^2$}\end{equation} for some constant $C(T)>0$.
The right-hand side quantity is natural in the context of quantitative inequalities for shape optimisation problems \cite{FuscoMaggiPratelli} and optimal control problems \cite{MazariQuantitative}, and is akin to the Fraenkel asymmetry. We refer to Subsection \ref{Se:Biblio}.
\item \textbf{For time-dependent controls.} In the context where the controls are time dependent i.e. $f=f(t,x)$ and when the solution of \eqref{Eq:PvIntro} is some $f^*$, the goal is to establish something of the form
\begin{equation}\label{Eq:IntroDep}\fbox{$\displaystyle
\forall f\in \mathscr X\,,
\mathcal J_T(f)-\mathcal J_T(f^*)\leq- \int_0^T \omega(s)\Vert f(s,\cdot)-f^*(s,\cdot)\Vert_{L^1({\Omega})}^2$}
\end{equation}
for a function $\omega:[0;T]\rightarrow \textnormal{I\kern-0.21emR}_+$ such that for any $s\in (0;T)$, $\omega(s)>0$. As will be explained more in detail in Subsection \ref{Se:Biblio} and commented upon in the Conclusion, see Section \ref{Cl:Time}, this is a stronger norm than the usual one.
\end{itemize}
To the best of our knowledge, neither type of quantitative estimates have been derived despite their natural interest.
Obviously, one can not expect to prove \eqref{Eq:IntroIndep} or \eqref{Eq:IntroDep} for all optimal control problems. What we propose here is to \textit{establish both these inequalities for a linearly controlled heat equation in the ball under $L^1$ and $L^\infty$ constraints.} The main equation under consideration is set in ${\Omega}=\mathbb B(0;R)$ and writes
\begin{equation}
\label{Eq:MainIntro}
\begin{cases}
\frac{\partial u_f}{\partial t}-\Delta u_f=f\text{ in }{(0;T)\times \O}\,,
\\ u_f(t=0)=u^0\geq 0\text{ in }{\Omega}\,,
\\u_f(t,\cdot)=0\text{ on }\partial {\Omega}.\end{cases}\end{equation} We will also assume that the initial condition $u^0\in \mathscr C^{2}({\Omega})\cap W^{1,2}_0({\Omega})$, $u^0\geq 0$, which is fixed, is radially symmetric and non-increasing. In the time-independent case (when $f=f(x)$), the functional we seek to maximise is defined by
\begin{equation}
\mathcal J_T(f):=\frac12\iint_{(0;T)\times \O} u_f^2(t,x)dxdt.
\end{equation}
In the time dependent case (when $f=f(t,x)$), the functional we seek to maximise is
\begin{equation}
\mathcal J^{\varepsilon}_T(f):=\frac12\iint_{(0;T)\times \O} u_f^2(t,x)dxdt+\frac\e2 \int_{\Omega} u_f^2(T,x)dx
\end{equation}
for some ${\varepsilon}>0$. The main reason behind supplementing the functional with a final time term is to ensure the non-degeneracy of the switch function associated to the optimisation problem
As a final comment, let us remark that the constant $C(T)$ appearing in \eqref{Eq:IntroIndep} and the weight $\omega$ are constructed in a non-explicit way.
\begin{remark}
Although we prove our results for maximisation of functionals, we believe the same strategies work for the minimisation of the functional. For both problems, both inequalities may have interesting consequences for inverse problems.\end{remark}
\begin{remark}
Obviously, if the optimal control $f^*$ for the time-dependent case does not depend on time, which will be the case here, \eqref{Eq:IntroDep} implies \eqref{Eq:IntroIndep} with $C(T)=\int_0^T \omega(s)ds.$ However, the reason why we present two proofs is the possibility of generalising the methods used to prove \eqref{Eq:IntroIndep}. In the conclusion, we explain why we believe this inequality can be extended to other types of control problems, such as bilinear control problems, or how, in general domains, technical assumptions on second order shape derivatives may enable one to derive it. In short, the proof of \eqref{Eq:IntroIndep} relies on two properties of the control problem: the first one is shape derivatives, for which the trickiest part is to prove coercivity of second order derivatives in general domains; the second one is the convexity of the problem, which is something extremely general.
The proof of \eqref{Eq:IntroDep} is specific to the case of the ball and it is unclear whether or not it may be adapted to other domains. Indeed, the parabolic isoperimetric inequalities used in its proof \cite{Alvino1990,Bandle,RakotosonMossino} may not hold in general domains, in the sense that explicit characterisation of maximisers may not be attainable.\end{remark}
\subsection{Statement of the main results}\label{Se:Results}
Let ${\Omega}=\mathbb B(0;R)$ be a centred ball in dimension $n$. We assume that we are given an initial condition satisfying
\begin{equation}\label{Eq:A0}
u^0\geq 0\,, u^0\in \mathscr C^2({\Omega})\cap W^{1,2}_0({\Omega})\,, u^0\text{ is radially symmetric and non-increasing}.\end{equation} For a function $f\in L^2((0;T)\times {\Omega})$ we consider the solution $u_f$ of
\begin{equation}
\label{Eq:Main}
\begin{cases}
\frac{\partial u_f}{\partial t}-\Delta u_f=f\text{ in }(0;T)\times {\Omega}\,,
\\ u_f(t=0)=u^0\,,
\\u_f(t,\cdot)=0 \text{ in }(0;T)\times\partial {\Omega}.\end{cases}\end{equation}
The functional we wish to optimise in the time-independent case is
\begin{equation}\tag{$J$}\label{Eq:Fonc}
\mathcal J_T(f):=\frac12\iint_{(0;T)\times \O} u^2_f(T,\cdot).\end{equation} For a given $L^1$ constraint $V_0\in (0;\operatorname{Vol}({\Omega}))$, the sets of admissible controls are
\begin{equation}\label{Eq:AdmTI}\tag{$\bold{\overline{Adm}}$}
\overline{ \mathcal M}({\Omega}):=\left\{f\in L^\infty({\Omega})\,, 0\leq f\leq 1\,, \int_{\Omega} f=V_0\right\}\end{equation} and
\begin{equation}\label{Eq:AdmTD}\tag{$\bold{Adm}_T$}
\mathcal M_T({\Omega}):=\left\{f \in L^\infty\left((0;T)\times {\Omega}\right) \text{ for a.e. $t\in (0;T)$, } f(t,\cdot) \in \overline{\mathcal M}({\Omega}).\right\}\end{equation}
The first problem we address is
\begin{equation}\label{Eq:PvTi}\tag{$\bold I_1$}
\fbox{$\displaystyle \max_{f \in \overline{\mathcal M}({\Omega})}\mathcal J_T(f).$}\end{equation}
The second problem is set in the time-dependent case and the functional we seek to optimise is, for some ${\varepsilon}> 0$,
\begin{equation}\mathcal J^{\varepsilon}_T(f):=\frac12\iint_{(0;T)\times \O} u^2_f(T,\cdot)+\frac\e2\int_{\Omega} u_f^2(T,\cdot).\end{equation}
The second problem we address is
\begin{equation}\label{Eq:PvTd}\tag{$\bold I_2^{\varepsilon}$}
\fbox{$\displaystyle \max_{f \in {\mathcal M_T}({\Omega})}\mathcal J_T^{\varepsilon}(f).$}\end{equation}
Finally, we fix throughout the paper the notation
\begin{equation}
f^*:=\mathds 1_{{\mathbb B}^*}
\end{equation}
where $\mathbb B^*$ is the unique centred ball of volume $V_0$, and define $u^*$ as the solution of \eqref{Eq:Main} associated with $f\equiv f^*$.
\begin{remark}
The existence of solutions of \eqref{Eq:PvTi} and \eqref{Eq:PvTd} are easy consequence of the direct methods of the calculus of variations.\end{remark}
As an easy corollary of \cite[Theorem 3]{Alvino1990}, $f^*$ is a maximiser of both \eqref{Eq:PvTi} and \eqref{Eq:PvTd}. To prove our results, the first step is the following Theorem:
\begin{theorem} \label{Th:Uniqueness}$f^*$ is the unique solution of \eqref{Eq:PvTi} and of \eqref{Eq:PvTd}.\end{theorem}
It is not the main goal of this paper, but the result is in itself interesting and relies on topological properties of some classes of functions defined \emph{via} rearrangements. We refer to Section \ref{Se:Uniq} for the proof.
Let us now pass to the two main results of this article. We choose to state first the time-dependent case, as the result holds without any restriction on the dimension. In this case, we need to take ${\varepsilon}>0$. The reason behind this is technical, and amounts, to put it shortly, to forcing the switch function of the problem to be non-degenerate. We comment on this in the Conclusion.
\def{\mathbb S}{{\mathbb S}}
\begin{theorem}\label{Th:Td
If ${\varepsilon}>0$, there exists a constant $C({\varepsilon},T)>0$ such that
\begin{equation}\label{Eq:DpValTd}
\forall f \in \mathcal M_T({\Omega})\,, \mathcal J_T^{\varepsilon}(f)-\mathcal J_T^{\varepsilon}(f^*)\leq- C({\varepsilon},T)\int_0^T\left(-\frac{\partial p_{\varepsilon}^*}{\partial \nu}\right)_{\partial {\mathbb B}^*}\Vert f(t,\cdot)-f^*\Vert_{L^1({\Omega})}^2.\end{equation}
In the estimate above $p_{\varepsilon}^*$ solves
\begin{equation}
\begin{cases}
\frac{\partial p_{\varepsilon}^*}{\partial t}+\Delta p_{\varepsilon}^*=-u^*\text{ in }{(0;T)\times \O}\,,
\\ p_{\varepsilon}^*(T,\cdot)={\varepsilon} u^*(T,\cdot)\,,
\\ p_{\varepsilon}^*(t,\cdot)=0\text{ on }(0;T)\times \partial {\Omega}
\end{cases}
\end{equation}
and is the switch function of \eqref{Eq:PvTd}.
\end{theorem}
\begin{remark}
Actually, when ${\varepsilon}>0$, we can even prove that there exists a constant $A({\varepsilon},T)>0$ such that
\begin{equation}
\forall t \in (0;T)\,, \left(-\frac{\partial p_{\varepsilon}^*}{\partial \nu}\right)_{\partial {\mathbb S}^*}\leq -A({\varepsilon},T).
\end{equation}
We however choose to keep the partial derivative of the switch function as it seems to us to be a more precise result.
\end{remark}
We then pass to the time-independent case, where the main innovation will be the use of shape derivatives. We include this result not only for the sake of completeness but also because this method seems, at this stage, generalisable to other domains, while it may not be the case for Theorem \ref{Th:Td}.
\begin{theorem}\label{Th:Ti} Assume $n=2$.
For any $T>0$ there exists a constant $C(T)>0$ such that
\begin{equation}\label{Eq:DpValTi}
\forall f \in \overline{\mathcal M}({\Omega})\,, \mathcal J_T(f)-\mathcal J_T(f^*)\leq -C(T)\Vert f-f^*\Vert_{L^1({\Omega})}^2.
\end{equation}
\end{theorem}
\subsubsection{Bibliographical references}\label{Se:Biblio}
Let us now present the frameworks into which we think our present work fits.
\paragraph{Qualitative questions in optimal control problems}
The question of qualitative properties of optimal control problems has recently drawn a lot of attention. Indeed, in many situations, explicit computation of the optimal control is nearly impossible, and a line of research has emerged that deals with the question of knowing what optimal controls nearly look like, or whether or not these controls are (un)stable in a sense that has to be specified. Among all these qualitative queries, one may single out the following:
\begin{itemize}
\item \textbf{Insensitising controls.} The question of insensitising controls is a very natural one, and is a possible solution to the following question: given that it is often the case that one can not practically realise the exact control strategy and that some imperfections may arise, how can a \textit{robust} control strategy be constructed? In that context, the goal is to find an \textit{insensitising} optimal control. This question has been studied, for instance, in \cite{Alabau2,LissyPrivat} and, more recently, in \cite{LissyPrivatErvedoza}.
\item \textbf{The turnpike property.} The turnpike property states that, when dealing with time evolving optimal control problems, it is sometimes possible to actually find a \textit{nearly static} optimal strategy or, in other words, that the optimal control remains \textit{close}, in some sense that has to be quantified, to the solution of a stationary optimal control problem. First motivated by applications in economics \cite{Dorfman}, this field has been rapidly growing over the last decade and has found applications in many contexts (e.g. control of non-linear differential equations, control of the wave equation, control of semilinear heat equations, machine learning) \cite{2006,Cass1966,borjan,LanceTrelatZuazua,PighinSakamoto,TrelatZhang,TrelatZhangZuazua,Zuazua2017}. It has recently been derived for bilinear optimal control problems using quantitative inequalities for stationary optimisation problems \cite{MRB2020}.
\end{itemize}
\paragraph{Shape derivatives for time-evolving problems}
Our work presents what is to the best of our knowledge the first detailed analysis of a second order shape derivative for a time-evolving optimal control problems (in the sense that a coercivity norm for the second order shape derivative is obtained), albeit it deals with shape derivatives with respect to a subdomain. Although the literature devoted to time-evolving optimal control problems is scarce, we would like to point to \cite{Moubachir2006} where a speed-method approach is presented, and to the recent preprint \cite{Harbrecht} where shape derivatives (with respect to the underlying domain ${\Omega}$) are computed and used to obtain numerical simulations of a shape optimisation problem.
\paragraph{Quantitative inequalities} The study of quantitative inequalities in shape optimisation problems is an enormous field. To mention a few works, we point to the seminal \cite{FuscoMaggiPratelli} for the quantitative isoperimetric inequality, and to \cite{BDPV} for quantitative spectral inequalities. Regarding quantitative inequalities for (stationary) control problems we refer to \cite{BrascoButtazzo} for a quantitative inequality for the natural Dirichlet energy, to \cite{CarlenLieb} for a quantitative spectral inequality (with respect to the potential) in $\textnormal{I\kern-0.21emR}^n$ (both these works are done under $L^p$ constraints), to \cite{MazariQuantitative} for a quantitative spectral inequality in the ball under $L^1$ and $L^\infty$ constraints and to \cite{MRB2020} for a generalisation of this inequality to other domains, and for an application to the turnpike property.
Let us comment on the type of estimates usually obtained: given a functional $\mathcal F:{\Omega}\mapsto \mathcal F({\Omega})$, a typical problem reads
\begin{equation}\inf_{{\Omega}\,, \operatorname{Vol}({\Omega})=\overline V}\mathcal F({\Omega}).\end{equation} Let us assume that, up to a translation, the unique minimiser of this functional is a ball $\overline {\mathbb B}$ of volume $\overline V$ (this is the case when $\mathcal F({\Omega})=\operatorname{Per}({\Omega})$), then the inequality obtained in \cite{FuscoMaggiPratelli} reads: there exists a constant $C>0$ such that, defining the Fraenkel asymmetry of ${\Omega}$ as
\begin{equation}
\mathcal A({\Omega})=\inf_{x\in \textnormal{I\kern-0.21emR}^n}\operatorname{Vol}\left( (x+\overline {\mathbb B})\Delta {\Omega}\right)\end{equation} there holds
\begin{equation}
\mathcal F({\Omega})-\mathcal F(\overline {\mathbb B})\geq C\mathcal A({\Omega})^2.\end{equation}
In the case of estimate \eqref{Eq:IntroIndep}, the coercivity obtained is akin to this measure of asymmetry if the maximiser $\overline f$ writes $\overline f=\mathds 1_{\overline E}$: by defining $\mathcal J_T(E):= \mathcal J_T(\mathds 1_E)$ with a slight abuse of notation, if we choose a competitor $f$ of the form $f=\mathds 1_E$ then estimate \eqref{Eq:DpValTi} rewrites
\begin{equation}
\mathcal J_T(\overline E)-\mathcal J_T(E)\geq C(T)\operatorname{Vol}\left(\overline E\Delta E\right)^2.\end{equation}
On the other hand, \eqref{Eq:DpValTd} may seem more surprising. If, indeed, we assume that $f^*(t,x)=\mathds 1_{E^*(t)}(x)$ and if the competitor $f$ is chosen to assume the form $f(t,x)=\mathds 1_{E(t)}(x)$ for some subset $E$ of ${\Omega}$ then, seeing $\overline E:=\cup_{t\in (0;T)}\{t\}\times E(t)$ and $\overline E^*:=\cup_{t\in (0;T)}\{t\}\times E^*(t)$ as subsets of the cylindrical domain ${(0;T)\times \O}$, the "natural quantity" that one should obtain should be the squared asymmetry of $\overline E$ with respect to $\overline E^*$, that is, $\operatorname{Vol}(\overline E\Delta \overline E^*)^2$. The Jensen inequality enables to recover this discrepancy. This stronger norm may be a consequence of having chosen a volume constraint for every $t$. It is unclear at this stage whether or not replacing the constraint
\begin{equation}\text{ for a.e. $t\in (0;T)$,} \int_{\Omega} f(t,\cdot)=V_0\end{equation} with a global constraint
\begin{equation}\iint_{{(0;T)\times \O}}f=V_0\end{equation} would yield the coercivity norm
\begin{equation}\left(\iint_{(0;T)\times \O}\left|f^*-f\right|\right)^2.\end{equation} We refer to the Conclusion, Section \ref{Cl:Time}.
\subsection{Schwarz's rearrangement and isoperimetric inequalities for parabolic equations}\label{Se:Schwarz}
In order to be able to comment on our results and methods of proof, we need to give the basic definition underlying most of our methods, that of Schwarz's rearrangement. The three books we refer to for a comprehensive introduction to rearrangements are \cite{Kawohl,Kesavan,Rakotoson}. Here, since we are already working in a ball ${\Omega}=\mathbb B(0;R)$, we only give the definitions for functions defined on the ball.
\begin{definition}\label{De:Schwarz}
For a function ${\varphi}\in L^2({\Omega})\,, {\varphi}\geq 0$, its Schwarz rearrangement is the unique radially symmetric non-increasing function ${\varphi}^\#:{\Omega}\to \textnormal{I\kern-0.21emR}$ such that
\begin{equation}\forall t \in \textnormal{I\kern-0.21emR}_+\,, \operatorname{Vol}\left(\{{\varphi}> t \}\right)=\operatorname{Vol}\left(\{{\varphi}^\#> t\}\right).\end{equation} We define its one-dimensional counter part ${\varphi}^\dagger:[0;R]\to \textnormal{I\kern-0.21emR}$ as
\begin{equation}{\varphi}^\dagger(|x|):={\varphi}^\#(x).\end{equation}
\end{definition} The first property is that the Schwarz rearrangement preserves all the Lebesgue norms:
\begin{equation}\label{Eq:Equimeasurability}
\forall p\in (1;+\infty)\,, \forall u \in L^p({\Omega})\,, u\geq 0\,, \int_{\Omega} u^p=\int_{\Omega}(u^\#)^p.
\end{equation}
Of great importance to us are two inequalities. The first one, the so-called Poly\'a-Szeg\"{o} inequality asserts that
\begin{equation}
\forall {\varphi}\in W^{1,2}_0({\Omega})\,, {\varphi}^\#\in W^{1,2}_0({\Omega})\text{ and }\int_{\Omega} |{\nabla} {\varphi}^\#|^2\leq \int_{\Omega} |{\nabla} {\varphi}|^2.\end{equation}
The equality case in this equality was fully derived in \cite{Brothers1988} (see also \cite{Ferone2003}), and quantitative versions were given in \cite{Barchiesi2014,CianchiEsposito}. The second one is the Hardy-Littlewood inequality:
\begin{equation}\forall f\,, g\in L^2({\Omega})\,, f\,, g\geq 0\,, \int_{\Omega} fg\leq \int_{\Omega} f^\#g^\#.\end{equation} This inequality can be rewritten in the following form: for a.e. $\tau$,
\begin{equation}\int_{\{g>\tau\}}f\leq \int_{\{g^\#>\tau\}}f^\#.\end{equation}
A quantitative version of this inequality can be found in \cite{Cianchi2008} (and \cite{MRB2020} in a simpler case where smoothness of the involved function $f$ is assumed). In Propositions \ref{Pr:Bathtub} and \ref{Pr:BathtubTD}, we give uniform versions of this quantitative inequality for families of functions.
Comparison principle for parabolic equations started with the work of Bandle \cite{Bandle}, Vazquez \cite{Vazquez}, using the seminal ideas of Talenti \cite{Talenti}, and were later extended in a series of works by Alvino, Lions and Trombetti \cite{Alvino1986,Alvino1990} and Rakotoson and Mossino \cite{RakotosonMossino}. By "comparison principle for parabolic equations" we mean results that enable one to compare the solution $u$ of a parabolic equation of the form
\begin{equation}\frac{\partial u}{\partial t}-\Delta u=f(t,\cdot)\end{equation} with the solution $v$ of the symmetrised equation
\begin{equation}\frac{\partial v}{\partial v}-\Delta v=f^\#(t,\cdot).\end{equation} Both equations are supplemented with Dirichlet boundary conditions, and we wilfully ignore first order terms. The correct comparison relation $\prec$ used for such comparisons is defined as:
\begin{equation}\label{Eq:Prec}\text{$f\prec g$ if and only if for any }
r \in [0;R]\,, \int_{\mathbb B(0;r)} f^\#\leq \int_{\mathbb B(0;r)} g,
\end{equation}
and the typical result asserts that $u^\#(t,\cdot)\prec v(t,\cdot)$. In this paper, we will rely, for the uniqueness result, Theorem \ref{Th:Uniqueness}, on the method of proof of \cite{RakotosonMossino}, which enables more easily to encompass the equality case. We expand on their techniques in the proof of Theorem \ref{Th:Uniqueness}, see Section \ref{Se:Uniq}.
\subsection{Plan of the paper}\label{Se:Plan}This paper is structured as follows:
\begin{enumerate}
\item In Section \ref{Se:Preliminary} we gather several elementary information about the optimisation problems (adjoint, switch function, regularity of the solutions, convexity of the functionals).
\item In Section \ref{Se:Uniq}, we prove the uniqueness result stated in Theorem \ref{Th:Uniqueness}.
\item Section \ref{Se:Td} contains the proof of Theorem \ref{Th:Td} and is independent of Section \ref{Se:Ti}.
\item Section \ref{Se:Ti} corresponds to the proof of Theorem \ref{Th:Ti}. In it, we state our coercivity results for second order shape derivatives. This Section is independent of Section \ref{Se:Td}.
\item The Conclusion, Section \ref{Se:Concl}, contains discussion about possible extensions, as well obstructions for generalising the results presented here
\end{enumerate}
\subsection{Notational conventions}\label{Su:Notation}
\begin{itemize}
\item For any $g\in L^2({\Omega})$, $g^\#$ denotes its Schwarz rearrangement and $g^\dagger$ its one-dimensional counterpart.
\item ${\mathbb B}^*=\mathbb B(0;r^*)$ is the unique centred ball of volume $V_0$. In other words, it is the only centred ball satisfying $\mathds1_{{\mathbb B}^*}\in \overline{\mathcal M}({\Omega}).$
\item $u^*$ is the solution of \eqref{Eq:Main} associated with the static control $f\equiv f^*=\mathds 1_{B^*}$.
\item For a function $f$ that is discontinuous across a smooth hypersurface $\Sigma$ with oriented normal $\nu$, but continuous in ${\Omega}\backslash \Sigma$, the jump of $f$ across $\Sigma$ is
\begin{equation}\left.\llbracket f\rrbracket\right|_\Sigma:=\lim_{t\to 0^+}\left(f(x+t\nu(x))-f(x-t\nu(x))\right).\end{equation}
\end{itemize}
\section{Preliminary results}\label{Se:Preliminary}
We gather here several results that will be used throughout the paper. We begin with some basic regularity estimates on the solutions of the equation.
\begin{proposition}\label{Pr:Regularity}
For any $\alpha\in (0;1)$, there exists $M_\alpha>0$ such that, for any $f\in \mathcal M_T({\Omega})$, we have the estimate
\begin{equation}
\Vert u_f(t,\cdot)\Vert_{\mathscr C^{0,\alpha}({(0;T)\times \O})}\leq M_\alpha.
\end{equation}
Furthermore, for any $\alpha\in (0;1)$ and almost every $t\in (0;T)$, $u_f(t,\cdot)\in \mathscr C^{1,\alpha}({\Omega})$.
\end{proposition}
\begin{proof}[Proof of Proposition \ref{Pr:Regularity}]
For the first point, we use \cite[Corollary 7.31, p.182]{Lieberman} which ensures that for any $p>1$,
\begin{equation}\int_0^T \Vert u(t,\cdot)\Vert_{W^{2,p}({\Omega})}+\left\Vert \frac{\partial u}{\partial t}\right\Vert_{L^p({\Omega})}\leq C(\Vert f\Vert_{L^\infty}+\Vert u^0\Vert_{\mathscr C^2}),\end{equation} where $C$ depends on the dimension, on $p$ and on ${\Omega}$. It thus follows that, in particular, for any $p\in (1;+\infty)$ there exists $C_p$ such that
\begin{equation}\Vert u\Vert_{W^{1,p}({(0;T)\times \O})}\leq C_p.\end{equation} It then suffices to apply the Sobolev embedding $W^{1,p}({(0;T)\times \O})\hookrightarrow
\mathscr C^{0,\alpha}({\Omega})$ for $p$ large enough.
The second point follows from the fact that, from the same estimate, for any $p>1$ and almost every $t\in (0;T)$, $u_f(t,\cdot)\in W^{2,p}({\Omega})$. The conclusion follows by the Sobolev embedding $W^{2,p}({\Omega})\hookrightarrow \mathscr C^{1,\alpha}({\Omega})$.
\end{proof}
We then provide structural information about the functionals which we seek to optimise. In Proposition \ref{Pr:Convexity}, we establish convexity properties which will prove crucial while, in Proposition \ref{Pr:Adjoint}, we compute the adjoint and the switch function of the equation.
\begin{proposition}\label{Pr:Convexity}
The map $\mathcal J_T:\overline{\mathcal M}({\Omega})\ni f\mapsto \mathcal J_T(f)$ is strictly convex. In the same way, for any ${\varepsilon}> 0$, the map $\mathcal J_T^{\varepsilon}:\mathcal M_T({\Omega})\ni f\mapsto \mathcal J_T^{\varepsilon}(f)$ is strictly convex.
\end{proposition}
\begin{proof}[Proof of Proposition \ref{Pr:Convexity}]
We only prove the convexity of $\mathcal J_T$, the convexity of $\mathcal J_T^{\varepsilon}$ following along the same lines.
It follows from standard argument that the map $\overline{\mathcal M}({\Omega})\ni f\mapsto u_f$ is twice G\^ateaux-differentiable. The convexity of the functional is equivalent to requiring that the second order G\^ateaux derivative be non-negative. For any admissible perturbation $h$ at $f$ (that is, such that for every $t>0$ small enough $f+th\in \overline{\mathcal M}({\Omega})$) the G\^ateaux-derivative of $u_f$ in the direction $h$, denoted by $\dot u_f$, solves
\begin{equation}\label{Eq:Chanel}
\begin{cases}
\frac{\partial \dot u_f}{\partial t}-\Delta \dot u_f=h\text{ in }{(0;T)\times \O}\,, \\ \dot u_f=0\text{ on }(0;T)\times \partial {\Omega}\,, \\\dot u_f(0,\cdot)\equiv 0.\end{cases}\end{equation}
From this equation on $\dot u_f$, we deduce that the G\^ateaux-derivative of $\mathcal J_T$ at $f$ in the direction $h$ is given by
\begin{equation}\label{Eq:KL}
\dot{\mathcal J}_T(f)[h]=\iint_{(0;T)\times \O} \dot u_f u_f.\end{equation}
In the same way, the second order G\^ateaux-derivative of $u_f$ in the direction $h$, denoted by $\ddot u_f$, satisfies
\begin{equation}
\begin{cases}
\frac{\partial \ddot u_f}{\partial t}-\Delta \ddot u_f=0\text{ in }{(0;T)\times \O}\,, \\ \ddot u_f=0\text{ on }(0;T)\times \partial {\Omega}\,, \\\ddot u_f(0,\cdot)\equiv 0\end{cases}\end{equation} and thus
\begin{equation}\ddot u_f=0.\end{equation}
Furthermore, the second order G\^ateaux-derivative of $\mathcal J_T$ in the direction $h$, denoted by $\ddot{\mathcal J_T}(f)[h,h]$, is given by
\begin{equation}\ddot{\mathcal J_T}(f)[h,h]=\iint_{(0;T)\times \O} \left(\dot u_f\right)^2+\iint_{(0;T)\times \O}\ddot u_f u_f=\iint_{(0;T)\times \O} \left( \dot u_f\right)^2\geq 0\end{equation} and the last inequality is strict unless $h\equiv 0.$ Since the second-order G\^ateaux-derivative of the functional is non-negative, the functional is convex.
\end{proof}
This convexity property is one of the fundamental point to carry out the proof of Theorem \ref{Th:Ti}.
\begin{proposition}\label{Pr:Adjoint}Let $f\in \overline{ \mathcal M}({\Omega})$. Let $p_f$ be the unique solution of
\begin{equation}\label{Eq:Fendi}\begin{cases}
\frac{\partial p_f}{\partial t}+\Delta p_f=-u_f\text{ in }{(0;T)\times \O},
\\ p_f(T,\cdot)=0\,,
\\ p_f(t,\cdot)=0\text{ on }(0;T)\times \partial {\Omega}.\end{cases}\end{equation}
Then for any $f\in \overline{\mathcal M}({\Omega})$ and any admissible perturbation $h$ at $f$, the G\^ateaux-derivative of $\mathcal J_T$ at $f$ in the direction $h$ is given by
\begin{equation}\label{Eq:Coco}\dot{\mathcal J}_T(f)[h]=\iint_{(0;T)\times \O} h(x)p_f(t,x)dtdx.\end{equation} In the same way, let us consider a parameter ${\varepsilon}>0$. Let $f\in \mathcal M_T({\Omega})$ and define $p_{{\varepsilon},f}$ as the unique solution of
\begin{equation}\begin{cases}
\frac{\partial p_{{\varepsilon},f}}{\partial t}+\Delta p_{{\varepsilon},f}=-u_f\text{ in }{(0;T)\times \O},
\\ p_{{\varepsilon},f}(T,\cdot)={\varepsilon} u_f(T,\cdot)\,,
\\ p_{{\varepsilon},f}(t,\cdot)=0\text{ on }(0;T)\times \partial {\Omega}.\end{cases}\end{equation} Then for any $f\in {\mathcal M_T}({\Omega})$ and any admissible perturbation $h$ at $f$, the G\^ateaux-derivative of $\mathcal J_T^{\varepsilon}$ at $f$ in the direction $h$ is given by
\begin{equation}\dot{\mathcal J}_T^{\varepsilon}(f)[h]=\iint_{(0;T)\times \O} h(t,x)p_f(t,x)dtdx.\end{equation}
$p_f$ is dubbed the switch function for the functional $\mathcal J_T$, while $p_{{\varepsilon},f}$ is dubbed the switch function for the functional $\mathcal J_T^{\varepsilon}$.
\end{proposition}
\begin{proof}[Proof of Proposition \ref{Pr:Adjoint}]
We only prove this proposition in the case $f\in \overline{\mathcal M}({\Omega})$, the time-dependent case following along the same exact lines. Let us first note that, as a backward, linear heat equation, existence and uniqueness of a solution to \eqref{Eq:Fendi} is guaranteed.
To get \eqref{Eq:Coco}, we start from the expression \eqref{Eq:KL} of the first order G\^ateaux-derivative of the functional $\mathcal J_T$:
\begin{equation}\dot{\mathcal J}_T(f)[h]=\iint_{(0;T)\times \O} \dot u_f u_f,\end{equation} where $\dot u_f$ solves \eqref{Eq:Chanel}. If we multiply this equation by the solution $p_f$ of \eqref{Eq:Fendi} and integrate by parts, we get
\begin{equation}
\iint_{(0;T)\times \O} \dot u_f u_f=\iint_{(0;T)\times \O} h p_f.\end{equation} Since $\dot{\mathcal J}_T(f)[h]=\iint_{(0;T)\times \O} \dot u_f u_f,$ the conclusion follows.
\end{proof}
We conclude this section with some information about the function $u^*$ solution of \eqref{Eq:Main} with $f\equiv f^*$.
\begin{proposition}\label{Pr:Radial}
The solution $u^*$ of \eqref{Eq:Main} with $f\equiv f^*$ is radially symmetric. Furthermore, for any $r\in (0;R)$ and any $t\in (0;T)$
\begin{equation}
-\frac{\partial u^*}{\partial r}(t,r)>0.
\end{equation}
\end{proposition}
\begin{proof}[Proof of Proposition \ref{Pr:Radial}]
The radial symmetry of the solution is immediate. In radial coordinates, and with a slight abuse of notation, $u^*$ satisfies
\begin{equation}\label{Eq:MainEtoile}
\begin{cases}
\frac{\partial u^*}{\partial t}-\frac1{r^{n-1}}\frac{\partial }{\partial r}\left({r^{n-1}}\frac{\partial u^*}{\partial r}\right)=f^*\text{ in } (0;T)\times (0;R)\,,
\\ u^*(t,R)=\frac{\partial u^*}{\partial r}(t,0)=0\text{ for any }t\,,
\\ u^*(0,\cdot)=u^0.
\end{cases}
\end{equation}
From Proposition \ref{Pr:Regularity} above we can differentiate $u^*$ with respect to $r$. Let us write
\begin{equation}z:=\frac{\partial u}{\partial r}.\end{equation} It follows from \eqref{Eq:MainEtoile} that $z$ solves
\begin{equation}\frac{\partial z}{\partial t}-\frac1{r^{n-1}}\frac{\partial }{\partial r}\left({r^{n-1}}\frac{\partial z}{\partial r}\right)=-\frac{(n-1)z}{r^2}\text{ in } (0;T)\times (0;R)\end{equation} and that the following jump condition is satisfied at $r=r^*$:
\begin{equation}\left\llbracket \frac{\partial z}{\partial r}\right\rrbracket(t,r^*)=-\left\llbracket f^*\right\rrbracket(r^*)=1>0.\end{equation} Since $u^*\geq 0$, the parabolic Hopf Lemma implies that for any $t>0$,
\begin{equation}
z(t,R)< 0.\end{equation}Differentiating \eqref{Eq:MainEtoile} with respect to $r$ and remembering that $u^0$ is non-increasing, we obtain that $z$ solves
\begin{equation}\label{Eq:z}
\begin{cases}
\frac{\partial z}{\partial t}-\frac1{r^{n-1}}\frac{\partial }{\partial r}\left(r^{n-1}\frac{\partial z}{\partial r}\right)=-\frac{(n-1)z}{r^2}\text{ in } (0;T)\times (0;R)\,,
\\ z(t,R)<0\text{ for any }t\,,
\\ z(t,0)=0,
\\ \left\llbracket \frac{\partial z}{\partial r}\right\rrbracket(t,r^*)=1\,, t\in (0;T),
\\ z(0,\cdot)\leq0.
\end{cases}
\end{equation}
Let us then show that for any $(t,r)\in (0;T)\times (0;R)$
\begin{equation}z(t,r)<0.\end{equation}
First of all, multiplying \eqref{Eq:z} by the positive part $z_+$ of $z$ we get (keeping in mind that $z_+(t,R)=0$)
\begin{equation}
\frac12\frac\partial{\partial t}\int_0^R r^{n-1} z_+(t,r)^2dr+\int_0^R r^{n-1}\left(\frac{\partial z_+}{\partial r}\right)^2+(r^*)^{n-1}z_+(t,r^*)=-\int_0^R \frac{(n-1)z_+(t,r)^2}{r^{3-n}}dr
\end{equation} so that $z_+(t,\cdot)=0$. As a consquence, $z\leq 0$. To argue that $z<0$ in $(0;T)\times {\Omega}$, we follow the same procedure as for the strong parabolic maximum principle. If we first assume $z\leq 0$ satisfies an inequality rather than an equality, that is, that $z$ satisfies
$$\frac{\partial z}{\partial t}-\frac1{r^{n-1}}\frac{\partial }{\partial r}\left(r^{n-1}\frac{\partial z}{\partial r}\right)<-\frac{(n-1)z}{r^2}$$ then if by contradiction we assume that there exists $(t_0,r_0)$, with $r_0\in (0;R]$ and $t_0\in (0;T)$ such that $z(t_0,r_0)=0$, it follows that $r_0<R$. By the jump condition, $r_0\neq r^*$. Since $t_0<T$, plugging the optimality conditions, the contradiction follows. To exclude the case $t_0=T$ it suffices to consider the equation on $[0;T+\epsilon]\,, \epsilon>0$ and to carry out the same reasoning in $(0;T+\epsilon)$. To then pass from this case (strict inequality) to ours (the equality case), with
$$\frac{\partial z}{\partial t}-\frac1{r^{n-1}}\frac{\partial }{\partial r}\left(r^{n-1}\frac{\partial z}{\partial r}\right)=-\frac{(n-1)z}{r^2}$$ it suffices to consider $z_{\varepsilon}(t,r):=z(t,r)-{\varepsilon} t$ and the conclusion follows from passing to the limit ${\varepsilon}\to 0^+$.
\end{proof}
\section{Proof of Theorem \ref{Th:Uniqueness}: Uniqueness of maximisers}\label{Se:Uniq}
\begin{proof}[Proof of Theorem \ref{Th:Uniqueness}]
It follows from \cite{RakotosonMossino} that $f^*$ is a solution of \eqref{Eq:PvTi} and \eqref{Eq:PvTd}. The uniqueness property for \eqref{Eq:PvTd} implies uniqueness for \eqref{Eq:PvTi} so we focus on the time-dependent case. Let us define ${u^*}$ as the solution of \eqref{Eq:Main} associated with $f\equiv f^*$. We consider another solution $f$ of \eqref{Eq:PvTd} and the solution $u$ of \eqref{Eq:Main} associated. By convexity of the functional we can assume that $f$ is a characteristic function so that \begin{equation}f^\#=f^*.\end{equation}
We proceed along a series of claims. The first one is :
\begin{claim}\label{Cl:RM1}
If $f$ solves \eqref{Eq:PvTd} and if $u$ is the associated solution of \eqref{Eq:Main} then for almost every $t\in (0;T)$, there holds
\begin{equation}
u^\#(t,\cdot)={u^*}(t,\cdot).\end{equation}
\end{claim}
\begin{proof}[Proof of Claim \ref{Cl:RM1}]
It follows from \cite{RakotosonMossino} and the results recalled in the introduction that for almost every $t\in (0;T)$ we have
\begin{equation}\label{Eq:Int}
u^\#(t,\cdot)\prec {u^*}(t,\cdot).\end{equation}
The relation $\prec$ was defined in Equation \eqref{Eq:Prec}. Thus, from \cite[Proposition 2]{alvino1991} we have that for almost every $t\in (0;T)$ we have
\begin{equation}(u^\#)^2(t,\cdot)\prec {(u^*)}^2(t,\cdot).\end{equation} Integrating this inequality in time and in space yields
\begin{equation}\iint_{(0;T)\times {\Omega}} \left(u^\#\right)^2\leq \iint_{(0;T)\times {\Omega}} \left({u^*}\right)^2.\end{equation} However, by equimeasurability of the Schwarz rearrangement \eqref{Eq:Equimeasurability} we have
\begin{equation}
\iint_{(0;T)\times {\Omega}}u^2=\iint_{(0;T)\times {\Omega}} (u^\#)^2.\end{equation}
Since $f$ is a maximiser of \eqref{Eq:PvTd} it follows that equality holds for almost every $t$ in
\begin{equation}\int_{\Omega} \left(u^\#\right)^2(t,\cdot)\leq \int_{\Omega} \left({u^*}\right)^2(t,\cdot).\end{equation} Thus we have for almost every $t\in (0;T)$,
\begin{equation}\label{Eq:Pa}\int_{\Omega} \left(u^\#\right)^2(t,\cdot)= \int_{\Omega} \left({u^*}\right)^2(t,\cdot).\end{equation}
Let us now introduce the set $\mathscr K({u^*})$ defined as
\begin{equation}\mathscr K({u^*})=\left\{g\in L^2({\Omega})\,, g\prec {u^*}\right\}.\end{equation}
From \cite{Alvino1989} this is a compact (for the weak $L^\infty-*$ topology) and convex set whose set of extreme points is
\begin{equation}\mathscr C({u^*})=\left\{g\in L^2({\Omega})\,, g^\#={u^*}\right\}.\end{equation}
Since $x\mapsto x^2$ is strictly convex, the map $\mathscr K(v)\ni g\mapsto \int_{\Omega} g^2$ is strictly convex. Besides, once again because of the convexity of $x\mapsto x^2$, we have, for any $g\in \mathscr K(u^*)$, $$(g^\#)^2=(g^2)^\#.$$
As a consequence, the only solutions of the maximisation problem
\begin{equation}\label{Eq:LS}
\sup_{g\in \mathscr K({u^*})}\int_{\Omega} g^2
\end{equation} are exactly the elements of $\mathscr C({u^*})$.
On the other hand, \eqref{Eq:Pa} states that $u(t,\cdot)$ is a solution of \eqref{Eq:LS}, so it follows that for almost every $t\in (0;T)$ there holds
\begin{equation}
u^\#(t,\cdot)=u^*(t,\cdot).\end{equation}
\end{proof}
In particular, and this is the main point of this proof, the two following properties hold: first,
\begin{equation}\label{Eq:Simp} \text{ If $f$ solves \eqref{Eq:PvTd} then for a.e. $t\in (0;T)$, } u^\dagger(t,\cdot)=(u^*)^\dagger(t,\cdot).\end{equation} Second, we have, as a consequence the following fact:
\begin{equation}\text{If $f$ solves \eqref{Eq:PvTd} then for a.e. $t\in (0;T)$, }\int_{\Omega} u(t,\cdot)=\int_{\Omega} u^\#(t,\cdot)=\int_{\Omega} u^*(t,\cdot).\end{equation}
We then prove that if $f$ solves \eqref{Eq:PvTd}, then all the level sets of $u$ are balls.
\begin{claim}\label{Cl:Balls}
If $f$ solves \eqref{Eq:PvTd}, then all the level sets of $u$ are balls.
\end{claim}
\begin{proof}[Proof of Claim \ref{Cl:Balls}]
We follow the approach of \cite{RakotosonMossino}. We first recall \cite[Theorem 1.2]{RakotosonMossino}: if ${\varphi}\in W^{1,2}((0;T),L^2({\Omega}))$ then ${\varphi}^\#\in W^{1,2}((0;T),L^2({\Omega}))$ and moreover there holds, if ${\varphi}$ only has measure sets of measure zero,
\begin{equation}\label{Eq:RakoMossi}
\frac{\partial {\varphi}^\#}{\partial t}(t,s)=\frac{\partial w}{\partial s}(t,s)
\end{equation}
where $w$ is defined by
\begin{equation}
w(t,s)=\int_{\left\{{\varphi}(t,\cdot)\leq {\varphi}^\#(t,s)\right\}}\frac{\partial {\varphi}}{\partial t}.
\end{equation}
We then consider \eqref{Eq:Main}. For any $\tau \in \textnormal{I\kern-0.21emR}_+$, we multiply the equation by $(u-\tau)_+$ and integrate by parts in space. We obtain in a classical way
\begin{equation}\label{Eq:Lis}
0\leq -\frac{\partial}{\partial \tau}\int_{\{u>\tau\}} |{\nabla} u|^2(t,\cdot)=\int_{\{u>\tau\}} \left(f-\frac{\partial u}{\partial t}(t,\cdot)\right).\end{equation} We write the repartition function of $u$ as $\mu$:
\begin{equation}\mu(t,\tau)=\operatorname{Vol}\left(\{u(t,\cdot)>\tau\}\right).\end{equation}By the isoperimetric inequality and the co-area formula, taking $S_n:=n\operatorname{Vol}(\mathbb B(0;1))^{\frac1n}$, we obtain, as in \cite{RakotosonMossino},
\begin{align}\label{Eq:Co}
S_n
\mu(t,\tau)^{1-\frac1n}&\leq \left(-\frac{\partial }{\partial \tau}\int_{\{u(t,\cdot)>\tau\}}|{\nabla} u|\right)
\\&\leq \left(-\frac{\partial \mu}{\partial \tau}\right)^{\frac12}\left(-\frac{\partial}{\partial \tau}\int_{\{u(t,\cdot)>\tau\}}|{\nabla} u|^2\right)^{\frac12}.
\end{align}
This leads to
\begin{align}
S_n \mu(t,\tau)^{1-\frac1n}&\leq\left(-\frac{\partial \mu}{\partial \tau}\right)^{\frac12}\left(-\frac{\partial}{\partial \tau}\int_{\{u(t,\cdot)>\tau\}}|{\nabla} u|^2\right)^{\frac12}
\\&\leq \left(-\frac{\partial \mu}{\partial \tau}\right)^{\frac12}\left(\int_{\{u(t,\cdot)>\tau\}} \left(f-\frac{\partial u}{\partial t}(t,\cdot)\right)\right)^{\frac12}.
\end{align}
Hence,
\begin{equation}\label{Eq:Viv
S_n^2\mu(t,\tau)^{2-\frac2n}\leq \left(-\frac{\partial \mu}{\partial \tau}\right)\int_{\{u(t,\cdot)>\tau\}} \left(f-\frac{\partial u}{\partial t}\right).\end{equation} Here we recall that $\int_{\{u(t,\cdot)>\tau\}} \left(f-\frac{\partial u}{\partial t}\right)\geq 0$ by \eqref{Eq:Lis}.
As is customary we use the Hardy-Littlewood inequality to obtain
\begin{equation}\label{Eq:Hl}
\int_{\{u(t,\cdot)>\tau\}}f\leq \int_0^{\mu(t,\tau)}f^\dagger=:F(t,\mu(t,\tau)).\end{equation}
Let us now define
\begin{equation}k(t,\tau):=\int_0^\tau u^\dagger(t,s)ds\end{equation} and we obtain
\begin{equation}\int_{\{u(t,\cdot)>\tau\}}\frac{\partial u}{\partial t}=\frac{\partial k}{\partial t}(t,\mu(t,\tau)).\end{equation}
As such, for some constant $c_n>0$,
\begin{equation}
1\leq S_n^{-2}\left({-\frac{\partial \mu}{\partial \tau}}\right)\mu(t,\tau)^{\frac{2}n-2}\left(F(t,\mu(t,\tau))-\frac{\partial k}{\partial t}(t,\mu(t,\tau))\right).
\end{equation}
Integrating this equation between $\tau_0$ and $\tau_1$ for any $0\leq \tau_0\leq \tau_1$ yields
\begin{equation}
\tau_1-\tau_0\leq S_n^{-2}\int_{\mu(t,\tau_0)}^{\mu(t,\tau_1)}s^{-2+\frac2n} \left(F(t,s)-\frac{\partial k}{\partial t}(t,s)\right)ds.\end{equation} We hence get in a classical way \cite{Mossino} the following differential inequality
\begin{equation}
-\frac{\partial u^\dagger}{\partial \tau}(t,\tau)
-\frac{\partial^2 k}{\partial \tau^2}(t,\tau)\leq S_n^{-2}\tau^{-2+\frac2n}\left(F(t,\tau)-\frac{\partial k}{\partial t}(t,\tau)\right).
\end{equation}
Let us now define \begin{equation}
k_{u^*}(t,\tau):=\int_0^\tau \left(u^*\right)^\dagger(t,\cdot).\end{equation}
We recall that $u^*$ is the solution of \eqref{Eq:Main} associated with $f\equiv f^*$. Since $f$ is radially symmetric and decreasing, all the equalities in the above reasoning carried for $u$ hold for $u^*$ with equalities instead of inequalities and $k_{u^*}$ solves
\begin{equation}
\frac{\partial^2 k_{u^*}}{\partial \tau^2}+S_n^{-2}\tau^{-2+\frac2n}\frac{\partial k_{u^*}}{\partial t}= S_n^{-2}\tau^{-2+\frac2n}F(t,\tau).\end{equation}
Finally, we set $K=k-k_{u^*}$. From Equation \eqref{Eq:Simp}, we have, for any $t\in (0;T)$ and any $s\in (0;\operatorname{Vol}({\Omega}))$,
\begin{equation} K(t,s)=0.\end{equation}
Since $K\equiv 0$, every equality in the above reasoning must in fact be an equality. In particular, \eqref{Eq:Co} is an equality, and hence all the level-sets of $u$ are balls, which concludes the proof.
\end{proof}
\begin{remark}
It would be interesting to investigate whether or not using the quantitative isoperimetric inequality could lead to quantitative estimates, but it is not at this point clear how to do that. We refer to the Conclusion, Section \ref{SD}.\end{remark}
As is customary in the study of equality cases in Talenti-like inequalities, we need to check that the level sets are not just balls but rather concentric balls.
\begin{claim}\label{Cl:BouleConcentrique}
If $f$ solves \eqref{Eq:PvTd} then the level sets of the associated solution $u$ are concentric balls.\end{claim}
\begin{proof}[Proof of Claim \ref{Cl:BouleConcentrique}]
The core idea of the proof is similar to \cite{Kesavan1988}. Let us first consider the solution $w$ of
\begin{equation}\label{Eq:w}
\begin{cases}
\frac{\partial w}{\partial t}+\Delta w=-1\text{ in }(0;T)\times {\Omega}\,,
\\ w=0\text{ on }(0;T)\times {\Omega}\,,
\\ w(T,\cdot)=0.
\end{cases}\end{equation}
It follows from the same arguments as in the proof of Proposition \ref{Pr:Radial} that $w$ is radially symmetric and decreasing (for any $t<T$), and so we obtain by the Hardy-Littlewood inequality that for almost every $t\in (0;T)$,
\begin{equation}
\int_{\Omega} fw \leq \int_{\Omega} f^\#w=\int_{\Omega} f^*w.\end{equation}
However multiplying Equation \eqref{Eq:w} by $u$ and integrating by parts both in time and space yields
\begin{align*}
\iint_{(0;T)\times \O} fw&=\iint_{(0;T)\times \O} \left(\frac{\partial u}{\partial t}-\Delta u\right)w
\\&=-\int_{\Omega} w u^0-\iint_{(0;T)\times \O} u\left(\frac{\partial w}{\partial t}+\Delta w\right)
\\&=-\int_{\Omega} wu^0+\iint_{(0;T)\times \O} u
\\&=-\int_{\Omega} wu^0+\iint_{(0;T)\times \O} u^* \text{ because of Claim \ref{Cl:RM1}}
\\&=\iint_{(0;T)\times \O} f^* w \text{ by the same computations with $u^*$ instead of $u$}.
\end{align*}
However, and since $w$ is radially symmetric and increasing, the Hardy-Littlewood inequality implies that for almost every $t\in (0;T)$ and almost every $\tau$ we have
\begin{equation}\label{Eq:Kla}
\int_{\{w(t,\cdot)>\tau\}} f\leq \int_{\{w(t,\cdot)>\tau\}} f^\#.\end{equation} Hence it follows that \eqref{Eq:Kla} must be an equality for almost every $t$.
Thus since for almost every $t$ the function $w$ is symmetric and radially decreasing we get
\begin{equation}\label{Eq:Ke}
\forall r \in (0;R)\,, \int_{\mathbb B(0;r)} f=\int_{\mathbb B(0;r)} f^*.\end{equation}
For the final step, let $\phi_1$ be the first Dirichlet eigenvalue of the laplacian in ${\Omega}$. It is standard to see that $\phi_1$ is radially symmetric and decreasing. Introduce the solution $\phi$ of
\begin{equation}\label{Eq:phi}
\begin{cases}
\frac{\partial \phi}{\partial t}+\Delta \phi=-\phi_1\text{ in }(0;T)\times {\Omega}\,,
\\ \phi=0\text{ on }(0;T)\times {\Omega}\,,
\\ \phi(T,\cdot)=0.
\end{cases}\end{equation} The function $\phi$ is radially symmetric and decreasing as well for any $t<T$. As a consequence, all level-sets of $\phi(t,\cdot)$ are level-sets of $w(t,\cdot)$ and conversely, from which we deduce that, for almost every $t\in (0;T)$ and almost every $\tau$
\begin{equation}\int_{\{\phi(t,\cdot)>\tau\}}f=\int_{\{\phi(t,\cdot)>\tau\}}f^\#=\int_{\{\phi(t,\cdot)>\tau\}}f^*.\end{equation} This gives in turn
\begin{equation}\int_{\Omega} f\phi(t,\cdot)=\int_{\Omega} f^* \phi(t,\cdot).\end{equation}
Multiplying \eqref{Eq:phi} by $u$ and integrating by parts gives in the same way
\begin{equation}
\iint_{(0;T)\times \O} u\phi_1=\iint_{(0;T)\times \O} f\phi=\iint_{(0;T)\times \O} f^*\phi= \iint_{(0;T)\times \O} u^*\phi_1=\iint_{(0;T)\times \O} u^\# \phi_1.
\end{equation}
The last equality comes from \eqref{Eq:Simp}.
Invoking the Hardy-Littlewood inequality we obtain in the same fashion that for almost every $t\in (0;T)$
\begin{equation}
\forall r\in (0;R)\,, \int_{\mathbb B(0;r)} u=\int_{\mathbb B(0;r)} u^\#.\end{equation} It follows that $u=u^\#$ so that the conclusion is reached.
\end{proof}
\end{proof}
\section{Proof of Theorem \ref{Th:Td}}\label{Se:Td
\subsection{Plan of the proof and heuristics}
This theorem relies on the following fact: assuming that we have a competitor $f$, to be compared with $f^*$, and defining, for every $t\in [0;T]$,
\begin{equation}\delta(t):=\Vert f(t,\cdot)-f^*\Vert_{L^1({\Omega})}^2\end{equation} we can set
\begin{equation}\label{Eq:MTDelta}\mathcal M_T({\Omega},\delta)=\left\{g\in \mathcal M_T({\Omega})\,, \text{ for a.e. }t\in [0;T]\,, \Vert g(t,\cdot)-f^*\Vert_{L^1({\Omega})}=\delta(t)\right\}\end{equation} and replace $f$ with the solution $f_\delta^*$ of
\begin{equation}\label{Eq:DeltaTD}
\max_{f\in \mathcal M_T({\Omega},\delta)}\mathcal J_T^{\varepsilon}(f).
\end{equation}
That such a solution exists follows by the same argument as in Lemma \ref{Le:ExistenceDelta} below (see the proof in Appendix \ref{Ap:Technical}) but we can actually prove (and this is the part that is specific to ${\Omega}$ being a centred ball) that the solutions to \eqref{Eq:DeltaTD} admits the following explicit description: let, for any $\overline \delta>0$, ${\mathbb A}_{\overline\delta}$ be defined, in radial coordinates, as
\begin{equation}
{\mathbb A}_{\overline\delta}=\{r<r^*-r_{\overline\delta}^-\}\sqcup \{r^*<r<r^*+r_{\overline\delta}^+\}\end{equation} where $r_{\overline\delta}^-,r_{\overline\delta}^+$ are the unique parameters such that
\begin{equation}\operatorname{Vol}({\mathbb A}_{\overline \delta})=V_0\,, \operatorname{Vol}\left({\mathbb A}_{\overline \delta}\Delta {\mathbb B}^*\right)=\overline \delta.\end{equation} Then we will show (Proposition \ref{Pr:GoodGuys})
\begin{equation}\label{Eq:Fdelta}f_\delta:t\mapsto \mathds 1_{\mathbb A_{\delta(t)}}\end{equation} is a solution of \eqref{Eq:DeltaTD}. Throughout the rest of this introduction to the proof, we keep the notation $f_\delta$ for this function.
Let us formally assume that
\begin{equation}\int_0^T \delta(t)dt\ll 1\end{equation} and define, for any $\xi\in (0;1)$, $p_{{\varepsilon},\xi}$ the adjoint state associated with $f(t)=f^*+\xi \left(f_\delta-f^*\right)$. By parabolic regularity, $p_{{\varepsilon},\xi}$ should be a non-increasing function of $r$ since the adjoint state $p^*_{\varepsilon}$ associated to $f^*$ is decreasing. By the mean-value theorem, there exists $\xi \in [0;1]$ such that
\begin{equation}
\mathcal J_T^{\varepsilon}(f_\delta)-\mathcal J_T^{\varepsilon}(f^*)=\iint_{(0;T)\times \O} p_{{\varepsilon},\xi}\left(f_\delta-f^*\right). \end{equation} A natural step is then to try and apply the quantitative bathtub principle to this quantity: since $p_{{\varepsilon},\xi}$ is a radially symmetric, non-increasing function of $r$, then for any $t\in (0;T)$, $f^*$ is the only solution of
\begin{equation}
\sup_{f \in \overline{\mathcal M}({\Omega})}\int_{\Omega} fp_{{\varepsilon},\xi}(t,\cdot).
\end{equation}
The hope is then to prove that there exists a constant $C>0$ such that for any $t\in (0;T)$ there holds
\begin{equation}\forall f \in \overline{\mathcal M}({\Omega})\,, \int_{\Omega} (f-f^*)p_{{\varepsilon},\xi}\leq- C\left(-\frac{\partial p_{{\varepsilon},\xi}}{\partial r}\right)(t,r^*)\Vert f(t,\cdot)-f^*\Vert_{L^1({\Omega})}^2.\end{equation}
However, the existence of such a uniform constant relies, in a crucial way, on ${\varepsilon}$: when ${\varepsilon}>0$, it is possible while, when ${\varepsilon}=0$, other difficulties may arise. The key difficulty is that when ${\varepsilon}>0$ we can guarantee that
\begin{equation}\sup_{t\in [0;T]}\frac{\partial p_{\varepsilon}^*}{\partial r}(t,r^*)<0\end{equation} while for ${\varepsilon}=0$ we can only guarantee
\begin{equation}\forall \tau>0\,, \exists \alpha(\tau)>0\,, \sup_{t\in [0;T-\tau]}\frac{\partial p_{\varepsilon}^*}{\partial r}(t,r^*)\leq-\alpha(\tau).\end{equation}
To give a synthetic presentation, we isolate the main tool of this proof in the following paragraph.
\subsection{Uniform quantitative bathtub principle}
\begin{proposition}\label{Pr:BathtubTD}
Let $\beta>0$ and consider a family of function $\{p_i\}_{i\in I}\in \mathscr C^{1,\beta}({\Omega})$ such that:
\begin{enumerate}
\item There exists $M>0$ such that
\begin{equation}\label{Eq:HolderUniforme}
\sup_{i\in I}\Vert p_i\Vert_{\mathscr C^{1,\beta}}\leq M.\end{equation}
\item For any $i\in I$, $p_i$ is radially symmetric. Furthermore, there exists $\alpha>0$ such that, for any $r\in [0;r^*]$,
\begin{equation}\label{Eq:Mack}\forall i \in I\,, p_i(r)-p_i(r^*)\geq \alpha |r-r^*|.\end{equation} We also assume that for any $i\in I$, $p_i$ is decreasing in $(r^*;R)$. In particular, the unique level set of $p_i$ of volume $V_0$ is $\mathbb B(0;r^*)$: there exists $c_i$ such that
\begin{equation}{\mathbb B}(0;r^*)=\{p_i>c_i\}\,, \partial {\mathbb B}(0;r^*)=\{p_i=c_i\}.\end{equation} This in particular ensures that the minimum of $p_i$ in $\mathbb B(0;r^*)$ is only achieved on $\partial {\mathbb B}(0;r^*)$. As another consequence, for this constant $\alpha>0$, we have
\begin{equation}\label{Eq:DerivUniforme}
\forall i\in I\,, -\frac{\partial p_i}{\partial r}(r^*)\geq \alpha>0.
\end{equation}
\end{enumerate}
Then there exists a constant $\omega>0$ such that
\begin{equation}\label{Eq:BathtubTD}
\forall f \in \mathcal M({\Omega})\,, \forall i\in I\,, \int_{\Omega} p_i(f^*-f)\geq \omega\left(-\frac{\partial p_i}{\partial r}(r^*)\right)\Vert f-f^*\Vert_{L^1({\Omega})}^2.
\end{equation}
\end{proposition}
\begin{proof}[Proof of Proposition \ref{Pr:BathtubTD}]
Let us write $\mathscr T:=\{p_i\}_{i\in I}$. We first note that the assumption ensure that for any $p\in \mathscr T$, $f^*$ is the only solution of the problem
\begin{equation}
\sup_{f\in \overline{\mathcal M}({\Omega})}\int_{\Omega} fp.\end{equation}
We define
\begin{equation}
\mathscr G: \mathscr T\times \left(\overline{\mathcal M}({\Omega})\backslash \{f^*\}\right) \ni (p,f)\mapsto \frac{ \int_{\Omega} p(f^*-f)}{-\frac{\partial p_i}{\partial r}(r^*)\Vert f-f^*\Vert_{L^1({\Omega})}^2}
\end{equation}
and obviously proving \eqref{Eq:BathtubTD} boils down to proving
\begin{equation}\inf_{\mathscr T\times \left(\overline{\mathcal M}({\Omega})\backslash \{f^*\}\right)}\mathscr G>0.\end{equation}
Let us consider a minimising sequence $\{p_k\,, f_k\}\in\left( \mathscr T\times \left(\overline{\mathcal M}({\Omega})\backslash \{f^*\}\right)\right)^\textnormal{I\kern-0.21emN}$. Let us fix $\beta'\in (0;\beta)$. By \eqref{Eq:HolderUniforme} there exists $p_\infty\in \mathscr C^{1,\beta'}({\Omega})$ radially symmetric such that \begin{equation}
p_k\underset{k\to \infty}\rightarrow p_\infty\text{ in }\mathscr C^{1,\beta'}({\Omega})\,, \end{equation} and as a consequence we have
\begin{equation}\Vert p_\infty\Vert_{\mathscr C^{1,\beta'}}=\lim_{k\to \infty} \Vert p_k\Vert_{\mathscr C^{1,\beta'}}\leq M\end{equation}
and \eqref{Eq:DerivUniforme} holds for $p_\infty$. In the same way, and passing to the limit in \eqref{Eq:Mack}, $f^*$ is the only solution of
\begin{equation}\sup_{f \in \overline{\mathcal M}({\Omega})}\int_{\Omega} p_\infty f.\end{equation}
Up to a subsequence we also have that there exists $f_\infty\in \overline{\mathcal M}({\Omega})$ such that
\begin{equation}f_k\underset{k\to \infty}\rightarrow f_\infty\text{ weakly in $L^\infty-*$}.\end{equation}
We distinguish between two cases related to the sequence $\{\delta_k\}_{k\in\textnormal{I\kern-0.21emN}}$ defined by
\begin{equation}
\forall k \in \textnormal{I\kern-0.21emN}\,, \delta_k:=\Vert f_k-f^*\Vert_{L^1({\Omega})}.\end{equation} The first case corresponds to the case where, up to a subsequence,
\begin{equation}
\delta_k\underset{k\to \infty}\rightarrow \delta_\infty>0.
\end{equation}
In that case, we define
\begin{equation}
\overline{\mathcal M}_{>\delta_\infty}({\Omega}):=\left\{f\in \overline{\mathcal M}({\Omega})\,, \Vert f-f^*\Vert_{L^1({\Omega})}\geq \frac{\delta_\infty}2\right\}
\end{equation}
Following the same arguments as in \cite[Proposition 22]{MRB2020} we can see that the class $\overline{\mathcal M}_{>\delta_\infty}({\Omega})$ is closed under the weak $L^\infty-*$ convergence. Hence, it follows that
\begin{equation}\Vert f_\infty-f^*\Vert_{L^1({\Omega})}\geq \frac{\delta_\infty}2.\end{equation} This implies that
\begin{equation}\lim_{k\to \infty}\mathscr G(p_k,f_k)\geq \frac4{\delta_\infty^2}\mathscr G(p_\infty,f_\infty)>0\end{equation} since $f^*$ is the only maximiser of $f\mapsto \int_{\Omega} f p_\infty$ in $\overline{\mathcal M}({\Omega})$.
The second case is the difficult one. We henceforth work under the assumption that
\begin{equation}
\delta_k \underset{k\to \infty}\rightarrow 0.\end{equation}
We introduce the sequence of variational problem
\begin{equation}
\forall k \in \textnormal{I\kern-0.21emN}\,, \sup_{f\in \overline{\mathcal M}({\Omega})\,, \Vert f-f^*\Vert_{L^1({\Omega})}=\delta_k}\int_{{\Omega}} p_kf.
\end{equation}
From the same arguments as in \cite[Proposition 22]{MRB2020} there exists a solution to this variational problem. Furthermore since $p_k^\#=p_k$ the function $\mathds 1_{{\mathbb A}_{\delta_k}}$ is a solution of this problem, where $\mathbb A_{\delta_k}$ is defined, in radial coordinates
\begin{equation}
\overline{\mathbb A}_{\delta_k}=\{r<r^*-r_{\delta_k}^-\}\sqcup \{r^*<r<r^*+r_{\delta_k}^+\}\end{equation} and $r_{\delta_k}^-,r_{\delta_k}^+$ are the unique parameters such that
\begin{equation}\operatorname{Vol}(\mathbb A_{\delta_k})=V_0\,, \operatorname{Vol}(\mathbb A_{\delta_k}\Delta {\mathbb B}^*)=\delta_k.\end{equation}
Hence we assume that
\begin{equation}f_k=\mathds 1_{\mathbb A_{\delta_k}}.\end{equation}
For a general $\overline \delta>0$, we define $\mathbb A_{\overline \delta}$ in the same manner, that is,
\begin{equation}
{\mathbb A}_{\overline\delta}=\{r<r^*-r_{\overline\delta}^-\}\sqcup \{r^*<r<r^*+r_{\overline\delta}^+\}\end{equation} where $r_{\overline\delta}^-,r_{\overline\delta}^+$ are the unique parameters such that
\begin{equation}\operatorname{Vol}({\mathbb A}_{\overline \delta})=V_0\,, \operatorname{Vol}\left({\mathbb A}_{\overline \delta}\Delta {\mathbb B}^*\right)=\overline \delta.\end{equation}
We also recall that we have, for the same exponent $\beta'>0$,
\begin{equation}\label{Eq:HolderFaible}
\forall i\in I\,, \Vert p_i\Vert_{\mathscr C^{1,\beta'}}\leq M\end{equation} as this will be a crucial point. Let us then prove the following claim:
\begin{claim}\label{Cl:Monster}
There exists $\delta_1>0$ and $\omega_0>0$ such that for any $0\leq \delta\leq \delta_1$ there holds
\begin{equation}
\forall p\in \mathscr T\cup \{p_\infty\}\,,
\int_{\Omega} p(f^*-\mathds 1_{{\mathbb A}_\delta})\geq \omega_0\left(-\frac{\partial p}{\partial r}(r^*)\right)\Vert \mathds 1_{{\mathbb A}_\delta}-f^*\Vert_{L^1({\Omega})}^2.
\end{equation}
\end{claim}
Assuming this Claim holds it follows that for any $k$ large enough we have
\begin{equation}\mathscr G(p_k,f_k)\geq \omega_0,\end{equation} hence leading to the required contradiction. It thus only remains to prove Claim \ref{Cl:Monster}:
\begin{proof}[Proof of Claim \ref{Cl:Monster}]
Let us define, for any $\delta>0$,
\begin{equation}
h_\delta:=f^*-\mathds 1_{{\mathbb A}_\delta}=\mathds 1_{\{r^*-r_\delta^-<r<r^*\}}-\mathds 1_{\{r^*<r<r^*+r_\delta^+\}}.
\end{equation}
The quantity we want to bound from below is
\begin{equation}\int_{\Omega} h_\delta p_i.\end{equation}
First of all, explicit computations show that there exists a constant $c_0=c_0(d,r^*)$ such that
\begin{equation}\label{Eq:Rad}
r_\delta^+\,, r_\delta^-\underset{\delta \to 0}\sim c_0 \delta.
\end{equation}
We now write \eqref{Eq:Rad} in radial coordinates and obtain for any $p\in \mathscr T\cup\{p_\infty\}$
\begin{align*}
\frac1{(2\pi)^d}\int_{\Omega} h_\delta p=\int_{r^*-r_\delta^-}^{r^*}p(r)r^{n-1}{d}r-\int_{r^*}^{r^*+r_\delta^+}p(r)r^{n-1}{d}r.
\end{align*}
Let us first notice that from \eqref{Eq:HolderFaible} and \eqref{Eq:DerivUniforme}, there exists $\overline {\varepsilon}>0$ such that, for any $\delta\in (0;\overline {\varepsilon})$,
\begin{equation}\label{Eq:Burr}
\forall p\in \mathscr T\cup\{p_\infty\}\,, \inf_{(r^*-\delta;r^*+\delta)}\left(-\frac{\partial p}{\partial r}\right)\geq -\frac12\frac{\partial p}{\partial r}(r^*).
\end{equation}
We now have thanks to the mean value theorem, that for any $r\in (0;R)$, there exists $y_{r-r^*}\in (r;r^*)$ or $(r^*;r)$ such that
\begin{equation} p(r)=p(r^*)+p'(y_{r-r^*})(r-r^*).\end{equation}
As a consequence
\begin{align}
\frac1{(2\pi)^n}\int_{\Omega} h_\delta p&=p(r^*)\left(\int_{r^*-r_\delta^-}^{r^*}r^{n-1}{d}r-\int_{r^*}^{r^*+r_\delta^+}r^{n-1}{d}r\right)
\\&+\left(\int_{r^*-r_\delta^-}^{r^*}r^{n-1}|r-r^*|(-p'(y_{r-r^*})){d}r+\int_{r^*}^{r^*+r_\delta^+}r^{n-1}|r-r^*|(-p'(y_{r-r^*})){d}r\right)
\\&\label{I1}\geq p(r^*)\left(\int_{r^*-r_\delta^-}^{r^*}r^{n-1}{d}r-\int_{r^*}^{r^*+r_\delta^+}r^{n-1}{d}r\right)
\\&\label{I2}-\frac12\frac{\partial p}{\partial r}(r^*)\left(\int_{r^*-r_\delta^-}^{r^*}r^{n-1}|r-r^*|{d}r+\int_{r^*}^{r^*+r_\delta^+}r^{n-1}|r-r^*|{d}r\right)\text{ by \eqref{Eq:Burr}}
\end{align}The right hand side of \eqref{I1} is 0 because
\begin{equation}\int_{r^*-r_\delta^-}^{r^*}r^{n-1}{d}r-\int_{r^*}^{r^*+r_\delta^+}r^{n-1}{d}r=\frac1{(2\pi)^n}\int_{\Omega} h_\delta=0.\end{equation} Furthermore by explicit computations we obtain
\begin{align*}
\int_{r^*-r_\delta^-}^{r^*}r^{n-1}|r-r^*|dr\underset{\delta \to 0}\sim C\delta^2,
\end{align*}
and in the same way
\begin{equation}
\int_{r^*}^{r^*+r_\delta^+}r^{n-1}|r-r^*|dr\underset{\delta \to 0}\sim C\delta^2.
\end{equation}
The conclusion follows immediately.
\end{proof}
This concludes the proof of the Proposition.
\end{proof}
We then present, in the following paragraph, the proof of the aforementioned Proposition \ref{Pr:GoodGuys} that deals with the characterisation of solutions of a penalised problem.
\subsection{Characterisation of the solutions of an auxiliary problem}
Let us consider a function $\delta:[0;T]\rightarrow [0;\operatorname{Vol}({\Omega})]$ and the class $\mathcal M_T({\Omega},\delta)$ defined in \eqref{Eq:MTDelta}, as well as the function $f_\delta$ defined by \eqref{Eq:Fdelta}.
\begin{proposition}\label{Pr:GoodGuys}
For any ${\varepsilon}> 0$ and any positive function $\delta:[0;T]\to [0;\operatorname{Vol}({\Omega})]$, $f_\delta$ is a solution of the variational problem
\begin{equation}
\max_{g\in \mathcal M_T({\Omega},\delta)}\mathcal J_T^{\varepsilon}(g).
\end{equation}
\end{proposition}
\begin{proof}[Proof of Proposition \ref{Pr:GoodGuys}]
This is a straightforward adaptation of the proof of the parabolic isoperimetric inequality whose main steps were recalled in Section \ref{Se:Uniq}. Let us consider a function $g\in \mathcal M_T({\Omega},\delta)$ and $u$ the associated solution of \eqref{Eq:Main}. With the same notations as in Section \ref{Se:Uniq}, proof of Theorem \ref{Th:Uniqueness} we obtain
\begin{equation}
S_n^2\mu(t,\tau)^{2-\frac2n}\leq \left(-\frac{\partial \mu}{\partial \tau}\right)\int_{\{u(t,\cdot)>\tau\}} \left(g-\frac{\partial u}{\partial t}\right).\end{equation}
However, by the Hardy-Littlewood inequality, if we define $G(t,\mu(t,\tau)):=\int_{0}^{\mu(t,\tau)} \mathds 1_{{\mathbb A}_{\delta(t)}}^\dagger$ we obtain
\begin{equation}\int_{\{u(t,\cdot)>\tau\}}g\leq G(t,\mu(t,\tau)).\end{equation}
This is a penalised version of the Hardy-Littlewood inequality: it is indeed straightforward to see that, for any function $\overline g \in \overline{\mathcal M}({\Omega},\delta(t))=\{g\in \overline{\mathcal M}({\Omega})\,, \Vert g-f^*\Vert_{L^1({\Omega})}=\delta(t)\}$ and any measurable positive function $\ell$, there holds
\begin{equation}\int_{\Omega} \ell \overline g\leq \int_{\Omega} \mathds 1_{\mathbb A_{\delta(t)}}\ell^\#.\end{equation}As a consequence, for some constant $c_n>0$,
\begin{equation}
1\leq S_n^{-2}\left({-\frac{\partial \mu}{\partial \tau}}\right)\mu(t,\tau)^{\frac{2}n-2}\left(G(t,\mu(t,\tau))-\frac{\partial k}{\partial t}(t,\mu(t,\tau))\right).
\end{equation}
The rest of the proof follows along exactly the same lines.
\end{proof}
\subsection{Proof of Theorem \ref{Th:Td}}
In this subsection, we prove Theorem \ref{Th:Td} with a fixed, positive parameter ${\varepsilon}>0$.
\begin{proof}[Proof of Theorem \ref{Th:Td}]
We argue by contradiction and assume that there exists a sequence $\{f_k\}_{k\in \textnormal{I\kern-0.21emN}}\in(\mathcal M_T({\Omega})\backslash\{f^*\})^\textnormal{I\kern-0.21emN}$ such that
\begin{equation}\label{Eq:Contrad}\lim_{k\to \infty}\frac{\mathcal J_T^{\varepsilon}(f^*)-\mathcal J_T^{\varepsilon}(f_k)}{\int_0^T\left(-\frac{\partial p_{\varepsilon}^*}{\partial r}(t,r^*)\right)\Vert f_k(t,\cdot)-f^*\Vert_{L^1({\Omega})}^2}=0,\end{equation} where we recall that $p_{\varepsilon}^*$ is the solution of
\begin{equation}\label{Eq:Adjointepsilon}
\begin{cases}
\frac{\partial p_{\varepsilon}^*}{\partial t}+\Delta p_{\varepsilon}^*=-u^*\text{ in }{(0;T)\times \O}\,,
\\ p_{\varepsilon}^*(T,\cdot)={\varepsilon} u^*(T,\cdot)\,,
\\ p_{\varepsilon}^*(t,\cdot)=0\text{ on } (0;T)\times \partial {\Omega}.\end{cases}\end{equation}
In the same way, if $f\in \mathcal M_T({\Omega})$, $p_{{\varepsilon},f}$ stands for the solution of \eqref{Eq:Adjointepsilon} with $u^*$ replaced by $u_{f}$. By Proposition \ref{Pr:Adjoint}, the derivative of $\mathcal J_T^{\varepsilon}$ at $f$ in a direction $h$ is given by
\begin{equation}\label{Eq:DerivAdjointEpsilon}
\dot{\mathcal J}_T^{\varepsilon}(f)[h]=\iint_{(0;T)\times \O} h p_{{\varepsilon},f}.\end{equation}
Let us then begin with the following Claim:
\begin{claim}\label{Cl:NDA}
For any $T,{\varepsilon}>0$ and any $y_0\in (0;r^*)$, there exists $\alpha(y_0;{\varepsilon},T)$ such that
\begin{equation}
\inf_{(0;T)\times(y_0;R)}\left(-\frac{\partial p_{\varepsilon}^*}{\partial r}(t,r)\right)\geq \alpha(y_0;{\varepsilon},T)>0.\end{equation}\end{claim}
\begin{proof}[Proof of Claim \ref{Cl:NDA}]
We define $q_{\varepsilon}^*(t,\cdot):=p_{\varepsilon}^*(T-t,\cdot)$. Since $u^*$ is radially symmetric, $q_{\varepsilon}$ is radially symmetric as well and satisfies, in radial coordinates,
\begin{equation}\label{Eq:Qespilon}
\begin{cases}
\frac{\partial q_{\varepsilon}^*}{\partial t}-\frac1{r^{n-1}}\frac{\partial}{\partial r}\left(r^{n-1}\frac{\partial q_{\varepsilon}^*}{\partial r}\right)=u^*(T-t,\cdot)\text{ in } (0;T)\times (0;R)\,,
\\ q_{\varepsilon}^*(0,\cdot)={\varepsilon} u^*(T,\cdot)\,,
\\ q_{\varepsilon}^*(t,R)=\frac{\partial q_{\varepsilon}^*}{\partial r}(t,0)=0.
\end{cases}
\end{equation}
It follows from standard Schauder estimates \cite[Theorem 4.9, p.59]{Lieberman} and Proposition \ref{Pr:Regularity} that $q_{\varepsilon}^*\in \mathscr C^{1,\alpha}({(0;T)\times \O})$. Besides, since $u^*\geq 0$, we also have $q_{\varepsilon}\geq 0$ and, by the strong parabolic maximum principle, $q_{\varepsilon}^*>0\in {(0;T)\times \O}$.
As a consequence of the Hopf Lemma and of the fact that $\frac{\partial u^*}{\partial r}(T,R)<0$, defining $\Phi_{\varepsilon}^*:=\frac{\partial q_{\varepsilon}^*}{\partial r}$, we obtain
\begin{equation}
\forall t \in [0;T]\, , \Phi_{\varepsilon}^*(t,R)<0.\end{equation}
From Proposition \ref{Pr:Radial}, $\Phi_{\varepsilon}^*(0,\cdot)\leq 0\,, <0$ in $(0;R]$. Differentiating \eqref{Eq:Qespilon}, $\Phi_{\varepsilon}^*$ thus solves
\begin{equation}\label{Eq:PhiEpsilon}
\begin{cases}
\frac{\partial \Phi_{\varepsilon}^*}{\partial t}-\frac1{r{^{n-1}}}\frac{\partial}{\partial r}\left(r^{n-1}\frac{\partial \Phi_{\varepsilon}^*}{\partial r}\right)=\frac{\partial u^*(T-t,\cdot)}{\partial r}-\frac{(n-1)\Phi_{\varepsilon}^*}{r^2}\text{ in } (0;T)\times (0;R)\,,
\\ \Phi_{\varepsilon}^*(0,r)<0\text{ if }r>0\,, \Phi_{\varepsilon}^*(t,0)=0\,,
\\ \Phi_{\varepsilon}^*(t,R)<0.
\end{cases}
\end{equation}
Since (Proposition \ref{Pr:Radial}) $\frac{\partial u^*}{\partial r}<0$ for almost every $t,r>0$, $\Phi_{\varepsilon}^*$ solves, in ${(0;T)\times \O}$, the differential inequality
\begin{equation}
\frac{\partial \Phi_{\varepsilon}^*}{\partial t}-\frac1{r^{n-1}}\frac{\partial}{\partial r}\left(r^{n-1}\frac{\partial \Phi_{\varepsilon}^*}{\partial r}\right)<-\frac{(n-1)\Phi_{\varepsilon}^*}{r^2}\text{ in } (0;T)\times (0;R).
\end{equation} We can then apply the maximum principle, as was done in Proposition \ref{Pr:Radial}, to ensure that for any $t\in (0;T]$ and any $r>y_0$,
\begin{equation}
\Phi_{\varepsilon}^*(t,r)<0.
\end{equation}
As $\Phi_{\varepsilon}^*(0,r^*)={\varepsilon}\frac{\partial u_{\varepsilon}^*}{\partial r}(T,r^*)<0$ it follows that
\begin{equation}\forall t\in [0;T]\,, \Phi_{\varepsilon}^*(t,r^*)<0.\end{equation} Since $\Phi_{\varepsilon}$ is continuous in time, we can define
\begin{equation}\alpha(y_0;{\varepsilon},T):=\inf_{t\in [0;T], r\in (y_0;R)}\left(-\Phi_{\varepsilon}^*(t,r)\right)>0\end{equation} and the conclusion follows.
\end{proof}
Using this Claim we can come back to the sequence $\{f_k\}_{k\in \textnormal{I\kern-0.21emN}}\in(\mathcal M_T({\Omega})\backslash\{f^*\})^\textnormal{I\kern-0.21emN}$ satisfying \eqref{Eq:Contrad}. Since $f^*$ is the unique maximiser of $\mathcal J_T^{\varepsilon}$, we must have
\begin{equation}\label{Eq:Raise}
\int_0^T\left(-\frac{\partial p_{\varepsilon}^*}{\partial r}(t,r^*)\right)\Vert f_k-f^*\Vert_{L^1({\Omega})}^2\underset{k\to \infty}\rightarrow 0.
\end{equation}
If this were not the case, it would follow that the sequence $\{f_k\}_{k\in \textnormal{I\kern-0.21emN}}$ converges weakly in $\mathcal M_T({\Omega})$ to some $f_\infty \neq f^*$. As a consequence, the sequence $\{u_{f_k}\}_{k\in \textnormal{I\kern-0.21emN}}$ would converge in $\mathscr C^0({(0;T)\times \O})$ (using the uniform H\"{o}lder bounds from Proposition \ref{Pr:Regularity}) to $u_{f_\infty}$, and so
\begin{equation}\mathcal J_T^{\varepsilon}(f_k)\underset{k\to \infty}\rightarrow \mathcal J_T^{\varepsilon}(f_\infty)>\mathcal J_T^{\varepsilon}(f^*).\end{equation} This would yield\begin{equation}\lim_{k\to \infty}\frac{\mathcal J_T^{\varepsilon}(f^*)-\mathcal J_T^{\varepsilon}(f_k)}{\int_0^T\left(-\frac{\partial p_{\varepsilon}^*}{\partial r}(t,r^*)\right)\Vert f_k(t,\cdot)-f^*\Vert_{L^1({\Omega})}^2}=\frac{\mathcal J_T^{\varepsilon}(f^*)-\mathcal J_T^{\varepsilon}(f_\infty)}{\int_0^T\left(-\frac{\partial p_{\varepsilon}^*}{\partial r}(t,r^*)\right)\Vert f_\infty(t,\cdot)-f^*\Vert_{L^1({\Omega})}^2}>0,\end{equation} a contradiction.
Hence we work under the assumption that \eqref{Eq:Raise} holds. From Claim \ref{Cl:NDA} this implies
\begin{equation}\int_0^T\Vert f_k(t,\cdot)-f^*\Vert_{L^1({\Omega})}^2\underset{k\to \infty}\rightarrow 0.\end{equation} Hence, by Jensen's inequality,
\begin{equation}\Vert f_k-f^*\Vert_{L^1({(0;T)\times \O})}=\int_0^T\Vert f_k(t,\cdot)-f^*\Vert_{L^1({\Omega})}dt\underset{k\to \infty}\rightarrow 0.\end{equation}
As a consequence of standard parabolic estimates (Proposition \ref{Pr:Regularity}) we have, for any $\alpha\in (0;1)$,
\begin{equation}u_{f_k}\underset{k\to \infty}\rightarrow u_{f^*}=u^*\text{ in }\mathscr C^{0,\alpha}({\Omega})\end{equation} Defining, for any $k\in \textnormal{I\kern-0.21emN}$, $p_{\varepsilon}^{f_k}$ as the solution of
\begin{equation}\begin{cases}
\frac{\partial p_{\varepsilon}^{f_k}}{\partial t}+\Delta p_{\varepsilon}^{f_k}=-u_{f_k}\text{ in }{(0;T)\times \O}\,,
\\ p_{\varepsilon}^{f_k}(T,\cdot)={\varepsilon} u_{f_k}\text{ in }{\Omega}\,,
\\ p_{\varepsilon}^{f_k}(t,\cdot)=0\text{ in }(0;T)\times \partial {\Omega}\,, \end{cases}\end{equation} This in in turn implies, by Schauder's estimates \cite[Theorem 48.2]{Souplet}
\begin{equation}p_{\varepsilon}^{f_k}\underset{k\to \infty}\rightarrow p_{\varepsilon}^*\text{ in }\mathscr C^{1,\alpha}({(0;T)\times \O}).
\end{equation}
Hence, for any $y_0\in (0;r^*)$, there exists $k(y_0)>0$ such that for any $k\geq k(y_0)$, by Claim \ref{Cl:NDA}, there holds,
\begin{equation}
\forall (t,r) \in (0;T)\times (y_0;R)\,, \left(-\frac{\partial p_{\varepsilon}^{f_k}}{\partial r}(t,r)\right)\geq-\frac12 \frac{\partial p_{\varepsilon}^{*}}{\partial r}(t,r)>0,
\end{equation}
and, for any $k>0$ large enough, ${\mathbb B}(0;r^*)$ is a uniquely defined level set of $p_{\varepsilon}^{f_k}$: there exists $c_k$ such that
\begin{equation}{\mathbb B}(0;r^*)=\{p_{\varepsilon}^{f_k}>c_k\}.\end{equation} As a consequence, choosing $y_0$ small enough, we can ensure that all the assumptions of Proposition \ref{Pr:BathtubTD} are satisfied.
Finally, let us note that, by the same argument, these property also hold for any $p_{\varepsilon}^{f^*+\tau(f_k-f^*)}$ for any $\tau \in (0;1)$ and any $k$ large enough. In all the reasoning above, it suffices to add $\tau$ as another parameter in the family.
This allows us to apply Proposition \ref{Pr:BathtubTD}: there exists a constant $\underline \omega>0$ such that
\begin{multline}\label{Almost}
\forall f \in \mathcal M({\Omega})\,, \forall k\text{ large enough, }\forall t \in (0;T)\,, \forall \tau \in (0;1)\,,\\ \int_{\Omega} p_{\varepsilon}^{f^*+\tau(f_k-f^*)}(t,\cdot)(f^*-f_k(t,\cdot))\geq \underline\omega\left(-\frac{\partial p_{\varepsilon}^*}{\partial r}(t,r^*)\right)\Vert f_k(t,\cdot)-f^*\Vert_{L^1({\Omega})}^2.\end{multline}
Let us now apply, for any $k$ large enough, the mean value theorem to the map
\begin{equation}T_k=[0;1]\ni \tau\mapsto \mathcal J_T^{\varepsilon}(f^*+\tau(f_k-f^*)).
\end{equation}
There exists $\overline \tau \in (0;1)$ such that
\begin{equation}
\mathcal J_T^{\varepsilon}(f_k-f^*)=\iint_{(0;T)\times \O} p_{\varepsilon}^{f^*+\overline\tau(f_k-f^*)} \left(f_k-f^*\right).
\end{equation}
Using \eqref{Almost} we get
\begin{equation}
\mathcal J_T^{\varepsilon}(f_k-f^*)=\iint_{(0;T)\times \O} p_{\varepsilon}^{f^*+\overline \tau(f_k-f^*)} \left(f_k-f^*\right)\geq\underline \omega\int_0^T\left(-\frac{\partial p_{\varepsilon}^*}{\partial r}(t,r^*)\right)\Vert f_k(t,\cdot)-f^*\Vert_{L^1({\Omega})}^2.
\end{equation}
This is a contradiction, and the Theorem follows.
\end{proof}
\section{Proof of Theorem \ref{Th:Ti}: quantitative inequalities via shape derivatives and bathtub principle}\label{Se:Ti}
\subsection{Presentation and plan of the proof}
The proof relies on the use of shape derivatives and on the study of an auxiliary problem.
The structure of the proof is inspired by a previous work of the author \cite{MRB2020} and we will refer to this paper when needed. The main point is here to show an example of how shape derivatives may be used for parabolic problems.
Let us define, for any $\delta>0$, the class
\begin{equation}\tag{$\overline{\bold{Adm}}(\delta)$}\label{Eq:AdmDelta}\overline{\mathcal M}({\Omega},\delta):=\left\{ f\in \overline{\mathcal M}({\Omega})\,, \Vert f-f^*\Vert_{L^1({\Omega})}=\delta\right\}.\end{equation} We first consider the auxiliary variational problem
\begin{equation}\tag{$\bold{P}_\delta$}\inf_{f\in \overline{\mathcal M}({\Omega},\delta)} \mathcal J_T(f)\end{equation}
and prove that it admits a solution $f_\delta$ (Lemma \ref{Le:ExistenceDelta} below). Once this is done, we prove (Lemma \ref{Le:Intermediaire} below) that Theorem \ref{Th:Ti} is equivalent to proving that
\begin{equation}
\underset{\delta \to 0}{\lim \inf}\left(\frac{\mathcal J_T(f^*)-\mathcal J_T(f_\delta)}{\delta^2}\right)>0.
\end{equation}
\begin{remark}At this stage, one may argue to explicitly characterize $f_\delta$ as a radially symmetric solution, and thus bypass the part about shape derivatives. However, as our goal is also to provide a full analysis of shape hessians for time-dependent problems, and to present, in the Conclusion, possible generalisations to other settings where the explicit characterisation of optimisers of such a penalised problem are no longer available, we choose to not take advantage of that fact here.
\end{remark}
We then recall that $f^*=\mathds 1_{{\mathbb B}^*}$. We consider, for smooth enough vector fields $\Phi$, the deformed set ${\mathbb B}^*_\Phi:=(Id+\Phi)({\mathbb B}^*)$ and, with a slight abuse of notation, we write
$$\mathcal J_T({\mathbb B}^*_\Phi):=\mathcal J_T(\mathds 1_{{\mathbb B}^*_\Phi}).$$ We will prove (Proposition \ref{Pr:NormalDeformation}) that whenever $\Phi$ is "small" enough (in a sense made precise in the section devoted to shape derivatives) there holds
\begin{equation}\mathcal J_T({\mathbb B}^*)-\mathcal J_T({\mathbb B}^*_\Phi)\geq C\operatorname{Vol}({\mathbb B}^*_\Phi\Delta {\mathbb B}^*)^2\end{equation}
for some constant $C>0$.
We also prove a quantitative bathtub principle (Proposition \ref{Pr:Bathtub}), and finally conclude as in \cite{MRB2020} by comparing any competitor with one of the level sets of the switch function, and then this level set with the set ${\mathbb B}^*$. The key to conclude here is the convexity of the cost functional $\mathcal J_T$.
To proceed, we need some basic informations about the optimality conditions for Problem \eqref{Eq:PvTi}.
\paragraph{Optimality conditions for \eqref{Eq:PvTi}}
We recall, from Proposition \ref{Pr:Adjoint} that for any admissible perturbation $h\in L^\infty({\Omega})$ (that is, such that, for any ${\varepsilon}>0$ small enough, $f^*+{\varepsilon} h\in \overline{\mathcal M}({\Omega})$) the G\^ateaux-derivative of $u_f$ in the direction $h$, thereafter noted $\dot u$ solves
\begin{equation}\label{Eq:Udot}
\begin{cases}
\frac{\partial \dot u}{\partial t}-\Delta \dot u=h\text{ in }{(0;T)\times \O}\,,
\\ \dot u=0\text{ on }({\Omega};T)\times \partial {\Omega}\,,
\\ \dot u(0,\cdot)\equiv 0
\end{cases}
\end{equation}
and that, introducing the solution $p_f$ of\begin{equation}\label{Eq:Adjoint}\begin{cases}
\frac{\partial p_f}{\partial t}+\Delta p_f=-u_f\text{ in } {(0;T)\times \O}\,,
\\ p_f=0\text{ on }(0;T)\times \partial {\Omega}\,,
\\ p_f(T,\cdot)\equiv 0.\end{cases}
\end{equation}
we get the following expression for the G\^ateaux-derivative of $\mathcal J_T$:
\begin{equation}\label{Eq:Jdot2}
\dot{\mathcal J_T}(h)=\iint_{(0;T)\times \O} hp_f=\int_{\Omega} h(x)\left(\int_0^Tp_f(t,x)dt\right)dx.
\end{equation}
Let us define
\begin{equation}\Psi(x):=\int_0^T p_f(t,x)dt.\end{equation} Hence it follows that \begin{equation}\dot{\mathcal J_T}(h)=\int_{\Omega} h\Psi.\end{equation}
\subsection{Reduction to an auxiliary problem}
We now justify the study of the auxiliary problem
\begin{equation}\label{Eq:PvDelta}\tag{$\bold{P}_\delta$}\inf_{f\in \overline{\mathcal M}({\Omega},\delta)} \mathcal J_T(f)\end{equation} where $\overline{\mathcal M}({\Omega},\delta)$ was defined in \eqref{Eq:AdmDelta}.
\begin{lemma}\label{Le:ExistenceDelta}
For any $\delta>0$, the variational problem \eqref{Eq:PvDelta} has a solution $f_\delta$.
\end{lemma}
This Lemma is an adaptation of \cite[Proposition 22]{MRB2020}; for the sake of readability, its proof is only given in Appendix \ref{Ap:Technical}.
Throughout the rest of the proof of Theorem \ref{Th:Ti} we adopt the following notation:
\begin{equation}\text{ For any $\delta>0$, $f_\delta$ is a solution of \eqref{Eq:PvDelta}.}\end{equation}
We now explain why we will focus on the study of $f_\delta$ as a competitor; it is the subject of the following Lemma:
\begin{lemma}\label{Le:Intermediaire}
Theorem \ref{Th:Ti} is equivalent to proving that
\begin{equation}\label{Eq:Le}\underset{\delta \to0}{\lim\inf} \frac{\mathcal J_T(f^*)-\mathcal J_T(f_\delta)}{\delta^2}\geq C_0>0\end{equation} for some positive constant $C_0$.
\end{lemma}
The proof of this result is an adaptation of \cite[Lemma 23]{MRB2020} and mostly relies on the uniqueness of maximisers. We postpone the proof to Appendix \ref{Ap:Technical}.
The rest of the proof of Theorem \ref{Th:Ti} is going to be devoted to the proof of Estimate \eqref{Eq:Le}, see Proposition \ref{Pr:Le} below. To prove it, we need a local inequality for deformations of the optimal set ${\mathbb B}^*$ and a quantitative bathtub principle which will be used in combination with the convexity of the functional.
\subsection{Quantitative inequalities for deformations of ${\mathbb B}^*$: using shape derivatives}
Let us consider a $\mathscr C^1$ set $E$ of volume $V_0$ such that $E\cap \partial {\Omega}=\emptyset$ and a smooth, compactly supported in ${\Omega}$, vector field $\Phi$. We define
\begin{equation}E_\Phi:=(Id+\Phi)(E).\end{equation} We recall that we see $\mathcal J_T$ as a shape functional by defining, with a slight abuse of notations,
\begin{equation}\mathcal J_T(E):=\mathcal J_T(\mathds 1_E).\end{equation}Our goal is the following proposition:
\begin{proposition}\label{Pr:NormalDeformation}
There exist a constant $C>0$, a parameter $\eta>0$ and $p\in (1;+\infty)$ such that, for any compactly supported vector field $\Phi$ satisfying $\Vert \Phi\Vert_{W^{2,p}}$ there holds
\begin{equation}
\mathcal J_T({\mathbb B}^*)-\mathcal J_T({\mathbb B}^*_\Phi)\geq C\operatorname{Vol} \left({\mathbb B}^*_\Phi\Delta {\mathbb B}^*\right)^2.
\end{equation}
\end{proposition}
The proof of this Proposition follows the synthetic presentation of quantitative inequalities for deformations of optimal sets presented in \cite{DambrineLamboley}; their proof holds for shape optimisation of the domain ${\Omega}$, and we have presented in \cite{MRB2020} how to adapt their method to the optimisation of a subdomain $E\subset {\Omega}$. Let us present the main steps of the proof of Proposition \ref{Pr:NormalDeformation}:
\begin{enumerate}
\item The first one is to prove that ${\mathbb B}^*$ is a critical shape in the following sense: computing, for any compactly supported vector field $\Phi\in W^{2,p}$ the first order shape derivative $\mathcal J_T'({\mathbb B}^*)[\Phi]$ we need to prove that, if $\Phi$ additionally satisfies the linearised constraint
\begin{equation}\label{Eq:Linearvol}\int_{\partial {\mathbb B}^*} \left(\Phi\cdot \nu \right)=0\end{equation} then there holds
\begin{equation}\mathcal J_T'(E^*)[\Phi]=0.\end{equation}
This allows to consider, for the computation and analysis of second-order shape derivatives, vector fields $\Phi$ that are normal to $\partial {\mathbb B}^*$, and also to define a Lagrangian associated with a Lagrange multiplier
\begin{equation}
\mathcal L_T(E):=\mathcal J_T(E)+\mu\operatorname{Vol}(E),\end{equation} which satisfies, for any compactly supported vector field $\Phi\in W^{2,p}$ not necessarily satisfying \eqref{Eq:Linearvol}
\begin{equation} \mathcal L_T'({\mathbb B}^*)[\Phi]=0.\end{equation}
\item As a second step, we compute the second order shape derivative of the Lagrangian $\mathcal L_T''({\mathbb B}^*)[\Phi,\Phi]$ and prove an $L^2$ coercivity estimate, i.e that there exists a constant $c_0>0$ such that
\begin{equation}\forall \Phi \in W^{2,p}({\Omega};\textnormal{I\kern-0.21emR}^2)\,, \int_{\partial {\mathbb B}^*}\phi\cdot \nu=0\Rightarrow \mathcal L_T''({\mathbb B}^*)[\Phi,\Phi]\leq- c_0 \Vert \Phi \cdot \nu \Vert_{L^2(\partial {\mathbb B}^*)}^2.\end{equation}
This is done using a comparison principle previously used for elliptic equations \cite{MazariNadinPrivat,MazariQuantitative}, and our contribution here is to show how it extends to the case of parabolic equations.
\item We then define for a compactly supported vector field $\Phi \in W^{2,p}$ the map \begin{equation}j_\Phi:[0;1]\ni t\mapsto \mathcal L_T({\mathbb B}^*_{t\Phi})+C(\operatorname{Vol}({\mathbb B}_{t\Phi}^*)-V_0)^2\end{equation} for some $C$ large enough such that
\begin{equation}j_\Phi''(0)\leq -\tilde c_0\Vert \Phi \cdot \nu \Vert_{L^2(\partial {\mathbb B}^*)}^2\end{equation} and prove that there exists a modulus of continuity $\eta$, that is, a continuous function $\eta:\textnormal{I\kern-0.21emR}_+\to \textnormal{I\kern-0.21emR}_+$ such that $\omega(0)=0$, such that
\begin{equation}|j_\Phi''(t)-j_\Phi''(0)|\leq \eta\left(\Vert \Phi\Vert_{W^{2,p}}\right) \Vert \Phi \cdot \nu \Vert_{L^2(\partial {\mathbb B}^*)}^2,\end{equation} and conclude using the Taylor-Lagrange formula
\begin{equation}j_\Phi(1)-j_\Phi(0)=\int_0^1 j_\Phi''(s)ds\leq \left(-\tilde c_0+\omega\left(\Vert \Phi\Vert_{W^{2,p}}\right)\right)\Vert \Phi \cdot \nu \Vert_{L^2(\partial {\mathbb B}^*)}^2.\end{equation}
\end{enumerate}
All these steps rely on fine properties of first and second order shape derivatives. We begin with the computations of the shape derivatives of the Lagrange multiplier associated with the volume constraint and of the diagonalisation of the associated shape hessian at $E^*$.
\subsubsection{Computation of first and second order shape derivatives, computation of the Lagrange multiplier and diagonalisation of the shape Hessian}
\paragraph{Computation and analysis of the first order shape derivative}
Let us define, for any subdomain $E$ of ${\Omega}$ the function $u_E$ as the solution of \eqref{Eq:Main} associated with $f=\mathds 1_E.$ It should be noted that the shape differentiability of first and second order of the shape functional $\mathcal J_T:E\mapsto \mathcal J_T(E)$ follows from the same arguments as in \cite{Harbrecht}, and so does the computation of the first order shape derivative. The computations are a straightforward adaptation of \cite{Harbrecht} and we only give here a heuristic approach. Let us, then, consider a $\mathscr C^1$ shape, and a $W^{2,p}$ compactly supported vector field $E$. The shape derivative of $E\mapsto u_E$ in the direction $\Phi$ is denoted by $u'$ for the sake of notational simplicity. The differentiation of the main equation of \eqref{Eq:Main} gives, in a weak form, that, for any test function $v$,
\begin{equation}-\iint_{(0;T)\times \O} \frac{\partial v}{\partial t}u'+\iint_{(0;T)\times \O}\langle {\nabla} u',{\nabla} v\rangle=\iint_{(0;T)\times\partial E}v\left(\Phi \cdot \nu\right).\end{equation} \begin{remark} Alternatively, at a formal level: the differentiation of the initial condition yields
\begin{equation} u'(0,\cdot)\equiv 0.\end{equation} The differentiation of the main equation gives
\begin{equation}\frac{\partial u'}{\partial t}-\Delta u'=0.\end{equation}
Finally, the structural condition given by the weak formulation of \eqref{Eq:Main} is that there is no jump of the normal derivative on $\partial {\mathbb B}^*$ or, mathematically, that
\begin{equation}\label{Eq:NiJump}\left.\left\llbracket \frac{\partial u_E}{\partial \nu}\right\rrbracket\right|_{\partial E}=0.\end{equation}
We refer to Subsection \ref{Su:Notation} for the definition of the jump. Differentiating \eqref{Eq:NiJump} yields
\begin{equation}
\left.\left\llbracket \frac{\partial u'}{\partial \nu}\right\rrbracket\right|_{\partial E}=-\left(\Phi\cdot \nu\right).
\end{equation}
\end{remark}
In conclusion, $u'$ satisfies
\begin{equation}\label{Eq:Uprime}
\begin{cases}
\frac{\partial u'}{\partial t}-\Delta u'=0\text{ in }{(0;T)\times \O}\,,
\\ u'(t,\cdot)=0\text{ on }(0;T)\times \partial {\Omega}\,,
\\ \left.\left\llbracket \frac{\partial u'}{\partial \nu}\right\rrbracket\right|_{\partial E}=-\left(\Phi\cdot \nu\right),
\\ u'(0,\cdot)\equiv0.\end{cases}\end{equation}
Furthermore, if we consider the adjoint state $p_{\mathds 1_E}$, which we abbreviate as $p_E$ for notational simplicity, given by Equation \eqref{Eq:Adjoint} we obtain
\begin{align*}
\mathcal J_T'(E)[\Phi]&=\iint_{(0;T)\times \O} u_Eu'
\\&=-\iint_{(0;T)\times \O} \left(\frac{\partial p_E}{\partial t}+\Delta p_E\right)u'
\\&=-\iint_{(0;T)\times \partial E}\left\llbracket\frac{\partial u'}{\partial\nu}\right\rrbracket p_E
\\&=\iint_{(0;T)\times \partial E} \left(\Phi\cdot \nu\right)p_E.
\end{align*}
Let us single out this last identity:
\begin{equation}\label{Eq:JPrime}
\mathcal J_T'(E)[\Phi]=\int_{\partial E}\left(\Phi\cdot \nu\right)\left(\int_{0}^T p_E\right).\end{equation}
This allows us to obtain the following result:
\begin{lemma}\label{Le:SCritical}
${\mathbb B}^*$ is a critical shape in the following sense: for any compactly supported vector field $\Phi \in W^{2,p}({\Omega};\textnormal{I\kern-0.21emR}^2)$
\begin{equation}
\int_{\partial {\mathbb B}^*}\left(\Phi\cdot\nu\right)=0\Rightarrow \mathcal J_T({\mathbb B}^*)[\Phi]=0.\end{equation}\end{lemma}
\begin{proof}[Proof of Lemma \ref{Le:SCritical}]
From Proposition \ref{Pr:Radial}, $u^*$ is a radially symmetric function. Hence, the associated adjoint state $p^*=p_{\mathbb B^*}$ is also radially symmetric, so that the map
\begin{equation}\Psi:{\mathbb B}(0;R)\ni x\mapsto \int_{0}^T p^*(t,x)dt\end{equation} is radially symmetric. Letting $\overline \Psi_{\partial {\mathbb B}^*}:=\left.\Psi\right|_{\partial {\mathbb B}^*}$ we obtain
\begin{equation}\mathcal J_T'({\mathbb B}^*)[\Phi]=\overline \Psi_{\partial {\mathbb B}^*}\int_{\partial {\mathbb B}^*}\left( \Phi\cdot \nu\right)=0.\end{equation}
\end{proof}
It follows that the Lagrange multiplier associated with the volume constraint is $\mu=-\overline{\Psi}_{\partial {\mathbb B}^*}$ and we can hence define the Lagrangian
\begin{equation}
\mathcal L_{{\mathbb B}^*}(E):=\mathcal J_T(E)-\overline{\Psi}_{\partial {\mathbb B}^*}\operatorname{Vol}(E)\end{equation} and observe that, since $\operatorname{Vol}'(E)[\Phi]=\int_{\partial E} \left(\Phi\cdot \nu\right)$ we have, for any compactly supported vector field $\Phi\in W^{2,p}({\Omega};\textnormal{I\kern-0.21emR}^2)$
\begin{equation}
\mathcal L_{{\mathbb B}^*}'({\mathbb B}^*)[\Phi]=0.\end{equation}
As a consequence of \cite[Theorem~5.9.2 and the remark below]{HenrotPierre}, the second-order shape derivative in a direction $\Phi$ only depends on the normal trace of $\Phi$ and we hence work under the Assumption:
\begin{equation}\tag{$\bold{A}_\nu$}\label{As:Normal}\text{$\Phi$ is normal to $\partial {\mathbb B}^*$.}\end{equation}
\paragraph{Computation of the shape hessian and diagonalisation of the shape hessian at the ball}
We can now turn to the computation of the second order shape derivative. We once again choose a $\mathscr C^2$ shape $E$ and a compactly supported vector field $\Phi\in W^{2,p}({\Omega};\textnormal{I\kern-0.21emR}^2)$. It is well-known \cite[Proposition 5.4.18]{HenrotPierre} that
\begin{equation}\operatorname{Vol}''(E)[\Phi,\Phi]=\int_{\partial E} \mathscr H\left(\Phi\cdot \nu\right)^2,\end{equation} where $\mathscr H$ is the mean curvature of $\partial E$. Furthermore, differentiating \eqref{Eq:JPrime} and using once again \cite[Proposition 5.4.18]{HenrotPierre} we obtain
\begin{equation}\mathcal J_T''(E)[\Phi,\Phi]=\int_{\partial E}\left(\Phi\cdot \nu\right)\left(\int_{(0;T)} p'\right)+\int_{\partial E}\left(\Phi\cdot \nu\right)^2\left(\mathscr H \int_{0}^T p_E+\int_0^T\frac{\partial p_E}{\partial \nu}\right)\end{equation} where $p'$ satisfies
\begin{equation}\label{Eq:AdjointPrime}\begin{cases}
\frac{\partial p'}{\partial t}+\Delta p'=-u'\text{ in }{(0;T)\times \O},
\\ p'=0\text{ on }(0;T)\times {\Omega}\,,
\\ p'(T,\cdot)\equiv 0\text{ in }\Omega.\end{cases}\end{equation}
In particular, the shape hessian of the Lagrangian at the ball is given by
\begin{align*}
\mathcal L_{{\mathbb B}^*}''({\mathbb B}^*)[\Phi,\Phi]&=\int_{\partial {\mathbb B}^*}\left(\Phi\cdot \nu\right)\left(\int_{(0;T)} p'\right)+\int_{\partial {\mathbb B}^*}\left(\Phi\cdot \nu\right)^2\left(\mathscr H\left.\overline{\Psi}\right|_{\partial {\mathbb B}^*}+\int_0^T\frac{\partial p^*}{\partial \nu}\right)-\left.\overline{\Psi}\right|_{\partial {\mathbb B}^*}\int_{\partial {\mathbb B}^*}\mathscr H \left(\Phi\cdot \nu\right)^2
\end{align*}
so that simplifying the terms involving the mean curvature we are left with
\begin{equation}\label{Eq:ShapeHessianBall}
\mathcal L_{{\mathbb B}^*}''({\mathbb B}^*)[\Phi,\Phi]=\int_{\partial {\mathbb B}^*}\left(\Phi\cdot \nu\right)\left(\int_{(0;T)} p'\right)+\int_{\partial {\mathbb B}^*}\left(\Phi\cdot \nu\right)^2\int_0^T\frac{\partial p}{\partial \nu}.\end{equation}
Let us now diagonalise it. Since $\Phi$ is a vector field that is normal to $\mathbb S^*:=\partial {\mathbb B}^*$ from Assumption \eqref{As:Normal} it follows that we can decompose it, in angular coordinates, as
\begin{equation}\label{Eq:Decomposition}
\Phi\cdot \nu=\sum_{k=1}^\infty \alpha_k \cos(k\cdot)+\beta_k\sin(k\cdot).\end{equation}
\begin{remark} The fact that the sum involving the cosines starts at $k=1$ is a consequence of the fact that to compute the optimality condition for second order shape derivative we need to work in the space satisfying the linearised constraint or, in this case, to assume that
\begin{equation}\int_{\partial {\mathbb B}^*}\Phi\cdot \nu=0.\end{equation}\end{remark}
Let us first define $u_k'$ (resp. $v_k'$) as the solution of \eqref{Eq:Uprime} associated with $\Phi\cdot \nu=\cos(k\cdot)$ (resp. $\sin(k\cdot)$). It is straightforward to see that these two functions write
\begin{equation}
u_k'(r,\theta)=y_k(r)\cos(k\theta)\,, v_k'(r,\theta)=y_k(r)\sin(k\theta)\end{equation} where $y_k$ solves, for any $k\in \textnormal{I\kern-0.21emN}^*$,
\begin{equation}\begin{cases}
\frac{\partial y_k}{\partial t}-\frac1r\frac{\partial }{\partial r}(r\frac{\partial y_k}{\partial r})=-\frac{k^2}{r^2}y_k\text{ in }(0;R)\,,
\\ \left\llbracket y_k'\right\rrbracket(r^*)=-1\,,
\\ y_k(R,\cdot)=0\,,
\\ y_k'(0)=0.\end{cases}\end{equation}
Let us also introduce $g_k'$ (resp. $w_k'$) the solution of \eqref{Eq:AdjointPrime} associated with $\Phi\cdot \nu=\cos(k\cdot)$ (resp. $\sin(k\cdot)$). It is straightforward to see that these two functions write
\begin{equation}
g_k'(r,\theta)=z_k(r)\cos(k\theta)\,, w_k'(r,\theta)=z_k(r)\sin(k\theta)\end{equation} where $z_k$ solves, for any $k\in \textnormal{I\kern-0.21emN}^*$,
\begin{equation}\begin{cases}
\frac{\partial z_k}{\partial t}+\frac1r\frac{\partial}{\partial r}(r\frac{\partial z_k}{\partial r})=\frac{k^2}{r^2}z_k-y_k\text{ in }(0;R)\,,
\\ z_k(R,\cdot)=0\,,
\\ z_k(T)=0.\end{cases}\end{equation}
Furthermore, since $p^*$ is a radially symmetric function let us introduce the function $\overline p$ such that
\begin{equation}p^*(t,r,\theta)=\overline p(t,r).\end{equation}
This allows to recast the second order shape derivative \eqref{Eq:ShapeHessianBall} through the following Lemma:
\begin{lemma}\label{Le:Diagonalisation}
If $\Phi\cdot \nu$ is of the form \eqref{Eq:Decomposition} then there holds
\begin{equation}
\mathcal L_{{\mathbb B}^*}''({\mathbb B}^*)[\Phi,\Phi]=\frac{r^*}2\sum_{k=1}^\infty \omega_k \left\{ \alpha_k^2+\beta_k^2\right\} \end{equation} where for every $k\in \textnormal{I\kern-0.21emN}^*$ we have defined
\begin{equation}
\omega_k:=\int_0^T z_k(t,r^*)dt+\int_0^T \frac{\partial \overline p}{\partial r}(t,r^*)dt.
\end{equation}
\end{lemma}
\begin{proof}[Proof of Lemma \ref{Le:Diagonalisation}]
We can write \eqref{Eq:ShapeHessianBall} as
\begin{align*}
\mathcal L_{{\mathbb B}^*}''({\mathbb B}^*)[\Phi,\Phi]&=\int_0^{2\pi}\left(\Phi\cdot \nu\right)\left(\int_{(0;T)} p'\right)+\int_0^{2\pi}\left(\Phi\cdot \nu\right)^2\int_0^T\frac{\partial p^*}{\partial \nu}
\\&=\sum_{k,k'=1}^\infty\int_{0}^{2\pi}\int_0^T\left(\alpha_k\alpha_{k'}\cos(k\theta)\cos(k'\theta)+\beta_k\beta_{k'}\sin(k\theta)\sin(k'\theta)\right.
\\&+\left.+\alpha_k\beta_{k'} \cos(k\theta)\sin(k'\theta)\right)z_k(t,r^*)dtd\theta
\\&+\sum_{k=1}^\infty\int_0^{2\pi}\int_0^T \frac12\left(\alpha_k^2+\beta_k^2\right)\frac{\partial \overline p}{\partial r}(t,r^*)dt.
\end{align*}
All the crossed terms disappear for $k\neq k'$, and the conclusion follows by integrating in polar coordinates.
\end{proof}
We may now state the main result of this subsection:
\begin{proposition}\label{Pr:Coercivity}
There exists a constant $c_0>0$ such that for any $\Phi\in W^{2,p}$ satisfying \eqref{As:Normal} there holds
\begin{equation}\mathcal L_{{\mathbb B}^*}''({\mathbb B}^*)[\Phi,\Phi]\leq -c_0\int_{\partial {\mathbb B}^*} \left(\Phi\cdot\nu\right)^2.
\end{equation}
\end{proposition}
\begin{proof}[Proof of Proposition \ref{Pr:Coercivity}]
Given Lemma \ref{Le:Diagonalisation} it suffices to prove that there exists a constant $c_0>0$ such that
\begin{equation}\label{Eq:Goal}
\forall k\in \textnormal{I\kern-0.21emN}^*\,, \omega_k\leq -c_0.
\end{equation}
Equation \eqref{Eq:Goal} is obviously provided the following Claim holds:
\begin{claim}\label{Cl:Paty}
The sequence $\{\omega_k\}_{k\in \textnormal{I\kern-0.21emN}^*}$ is decreasing. Furthermore, $\omega_1<0$.
\end{claim}
Indeed, it then suffices to take $c_0=-\frac2{r^*}\omega_1$ and we can then bound
\begin{equation}\mathcal L_{{\mathbb B}^*}''({\mathbb B}^*)[\Phi,\Phi]=\frac{r^*}2\sum_{k=1}^\infty \omega_k(\alpha_k^2+\beta_k^2)\leq -c_0\sum_{k=1}^\infty (\alpha_k^2+\beta_k^2)=-c_0\int_{\partial {\mathbb B}^*}\left(\Phi\cdot\nu\right)^2.\end{equation} We now focus on the proof of this last Claim.
\begin{proof}[Proof of Claim \ref{Cl:Paty}]
Let us note that from Lemma \ref{Le:Diagonalisation} we have
\begin{equation}
\forall k \in \textnormal{I\kern-0.21emN}^*\,, \omega_{k}-\omega_1=\int_0^T \left(z_{k}-z_1\right)(t,r^*)dt.
\end{equation}
The fact that $\{\omega_k\}_{k\in \textnormal{I\kern-0.21emN}}$ is decreasing is thus guaranteed provided the following estimate holds:
\begin{equation}\label{Eq:Zk}
\forall k\in \textnormal{I\kern-0.21emN}^*\,, z_{k}\leq z_1.\end{equation}
\eqref{Eq:Zk} will be proved using a comparison principle. If we want to compare $z_k$ and $z_1$, we need to compare, for any $k\in \textnormal{I\kern-0.21emN}^*$, $y_k$ and $y_1$.
The first thing to observe is that
\begin{equation}\label{Eq:Y1positif}
y_1\geq 0.\end{equation}
\begin{proof}[Proof of \eqref{Eq:Y1positif}]
We already know that $y_1$ satisfies
\begin{equation}
\begin{cases}
\frac{\partial y_1}{\partial t}-\frac1r\frac{\partial }{dr}(r\frac{\partial y_1}{dr})=-\frac{1}{r^2}y_1\text{ in }(0;R)\,,
\\ \left\llbracket y_1'\right\rrbracket(r^*)=-1\,,
\\ y_1(R,\cdot)=0\,,
\\ y_1'(0,\cdot)=0.\end{cases}\end{equation} We consider the negative part $y_1^-$ of $y_1$. We have $$\llbracket(y_1^-)'\rrbracket(t,r^*)\begin{cases}=0\text{ if $y_1(t,r^*)>0$,}\\ =1\text{ if $y_1(t,r^*)<0$}, \\>0 \text{ if $y_1(t,\cdot)$ locally changes sign at $r^*$}.\end{cases}$$ In any case, we obtain
\begin{equation}\left\llbracket (y_1^-)'\right\rrbracket\geq 0.\end{equation}
Multiplying the equation by $y_1^-$ and integrating by parts in space and time as in the proof of Proposition \ref{Pr:Radial} gives
\begin{equation}\frac12\int_{\Omega} (y_1^-)^2(T,\cdot)+\iint_{(0;T)\times \O} |{\nabla} y_1^-|^2+\iint_{(0;T)\times\partial {\mathbb B}^*}y_1^-\left\llbracket (y_1^-)'\right\rrbracket+\iint_{(0;T)\times \O} \frac1{r}(y_1^-)^2=0.\end{equation} As a conclusion, $y_1^-\equiv 0$, which concludes the proof.
\end{proof}
Using this information, we can now prove:
\begin{equation}\label{Eq:Comparison1K}
\forall k\in \textnormal{I\kern-0.21emN}^*\,, y_k\leq y_1.\end{equation}
\begin{proof}[Proof of \eqref{Eq:Comparison1K}]
Let us define, for any $k\in \textnormal{I\kern-0.21emN}^*$,
\begin{equation}\Psi_k:=y_k-y_1.\end{equation}
Then, in ${(0;T)\times \O}$, $\Psi_k$ solves
\begin{equation}
\frac{\partial \Psi_k}{\partial t}-\Delta \Psi_k=-\frac{k^2}{r^2}y_k+\frac1{r^2}y_1\leq -\frac{k^2}{r^2} (y_k-y_1)=-\frac{k^2}{r^2}\Psi_k,\end{equation} where the last inequality comes from the fact that $y_1$ is non-negative. Furthermore,
\begin{equation}
\left\llbracket \Psi_k'\right\rrbracket(t,r^*)=0,
\end{equation}
so that, following exactly the main line of reasoning, we obtain
\begin{equation}\Psi_k\leq 0,\end{equation} which concludes the proof.
\end{proof}
We now pass to the next step:
\begin{equation}\label{Eq:Z1positive}
z_1\geq 0\text{ in }{(0;T)\times \O}.\end{equation}
\begin{proof}[Proof of \eqref{Eq:Z1positive}]
The function $z_1$ satisfies
\begin{equation}\begin{cases}
\frac{\partial z_1}{\partial t}+\frac1r\frac{\partial }{\partial r}(r\frac{\partial z_1}{\partial r})=\frac{1}{r^2}z_1-y_1\text{ in }(0;T)\times (0;R)\,,
\\ z_1(R,\cdot)=0\,,
\\ z_1(T,0)=\partial_r z_1(t,0)=0.\end{cases}\end{equation}
Since $y_1\geq 0$ from \eqref{Eq:Y1positif} $z_1$ solves, in particular,
\begin{equation}\label{Eq:Z1}\begin{cases}
\frac{\partial z_1}{\partial t}+\frac1r\frac{\partial }{\partial r}(r\frac{\partial z_1}{\partial r})\leq\frac{1}{r^2}z_1\text{ in }(0;T)\times (0;R)\,,
\\ z_1(R,\cdot)=0\,,
\\ z_1(T,0)=\partial_r z_1(t,0)=0.\end{cases}\end{equation}
Let us now define $\overline z_1=z_1(T-t,\cdot)$. Straightforward computations show that $\overline z_1$ solves \begin{equation}\partial_t \overline z_1-\frac{1}r\partial_r(r\partial_r\overline z_1)\geq -\frac1{r^2}\overline z_1.\end{equation} Multiplying this identity by $\overline z_1^-$ and integrating by parts, we obtain in the same way
\begin{equation}z_1\geq 0,\end{equation} as claimed.
\end{proof}
We are now in a position to prove \eqref{Eq:Zk}:
\begin{proof}[Proof of \eqref{Eq:Zk}]
We define, for any $k\in \textnormal{I\kern-0.21emN}$, $\mathscr Z_k:=z_k-z_1$. It is clear that $\mathscr Z_k$ solves
\begin{equation}
\frac{\partial \mathscr Z_k}{\partial t}+\frac1r\frac{\partial }{\partial r}\left(r\frac{\partial \mathscr Z_k}{\partial r}\right)=\frac{k^2}{r^2}z_k-\frac1{r^2}z_1+y_1-y_k.\text{ in }(0;T)\times (0;R)
\end{equation}
From Estimate \eqref{Eq:Comparison1K} there holds
\begin{equation}
\frac{\partial \mathscr Z_k}{\partial t}+\frac1r\frac{\partial }{\partial r}\left(r\frac{\partial \mathscr Z_k}{\partial r}\right)\geq\frac{k^2}{r^2}z_k-\frac1{r^2}z_1\text{ in }(0;T)\times (0;R)
\end{equation}
and so, from Estimate \eqref{Eq:Z1positive} we get
\begin{equation}
\frac{\partial \mathscr Z_k}{\partial t}+\frac1r\frac{d}{dr}\left(r\frac{d \mathscr Z_k}{dr}\right)\geq\frac{k^2}{r^2}z_k-\frac{k^2}{r^2}z_1=\frac{k^2}{r^2}\mathscr Z_k\text{ in }(0;T)\times (0;R).
\end{equation}
From the same reasoning, we obtain
\begin{equation}\mathscr Z_k\leq 0\end{equation} and so
\begin{equation}z_k\leq z_1\text{ in }(0;T)\times (0;R).\end{equation}
\end{proof}
The proof of the first part of Claim \ref{Cl:Paty} is thus finished, and it hence remains to prove that
\begin{equation}\label{Eq:Omega1}
\omega_1<0.\end{equation}
\begin{proof}[Proof of \eqref{Eq:Omega1}]
We recall that $$\omega_1=\int_0^T z_1(t,r^*)dt+\int_0^T \frac{\partial \overline p}{\partial r}(t,r^*)dt.$$ First of all, is is easy to see that $p$ is non-negative.
Let us define $\overline {\varphi}:=\frac{\partial \overline p}{\partial r}$. Straightforward computations show that $\overline {\varphi}$ solves
\begin{equation}\label{Eq:Phi}
\begin{cases}
\frac{\partial {\overline \p}}{\partial t}+\frac1r\frac{\partial }{\partial r}\left(r\frac{\partial {\overline \p}}{\partial r}\right)=-\frac{\partial \overline u}{\partial r}+\frac1{r^2}{\overline \p}\text{ in }(0;T)\times (0;R)\,,
\\ {\overline \p}(t,R)\leq 0\,,
\\ {\overline \p}(T,\cdot)\equiv 0.
\end{cases}\end{equation}
If we define $\overline \Phi:={\overline \p} +z_1$ we thus have
\begin{equation}
\frac{\partial \overline \Phi}{\partial t}+\frac1r\frac{\partial }{\partial r}\left(r\frac{\partial \overline \Phi}{\partial r}\right)=\frac1{r^2}\overline \Phi-\frac{\partial \overline u}{\partial r}\geq \frac1{r^2}\overline \Phi.
\end{equation}
The last inequality comes from Proposition \ref{Pr:Radial}. Furthermore we have $\overline \Phi(t,R)\leq 0$. As a consequence, we have
\begin{equation}\overline \Phi\leq 0\text{ in }(0;T)\times {\Omega}.\end{equation} Furthermore, we necessarily have $\overline \Phi(t,r^*)<0$ in a subset of positive measure of $(0;T)$, for otherwise we have $\frac{\partial \overline u}{\partial r}(t,r^*)= 0$ on this subset, which is absurd given Proposition \ref{Pr:Radial}. As a conclusion, we obtain
\begin{equation}\omega_1=\int_0^T\overline \Phi(t,r^*)dr<0,\end{equation} as claimed.
\end{proof}
\end{proof}
\end{proof}
With this Proposition available, we are in a position to prove Proposition \ref{Pr:NormalDeformation}. Let us recall that, for a normal deformation $\Phi\in W^{2,p}$ we have defined
\begin{equation}j_\Phi(\xi):=\mathcal L_{{\mathbb B}^*}({\mathbb B}^*_{t\Phi})+C(\operatorname{Vol}({\mathbb B}_{t\Phi}^*)-V_0)^2.\end{equation} Since ${\mathbb B}^*$ is a critical shape we obtain
\begin{equation}j_\Phi'(0)=0\end{equation} so that the Taylor-Lagrange formula with integral remainder writes, in the case where $\operatorname{Vol}({\mathbb B}^*_\Phi)=V_0$,
\begin{equation}\mathcal L_{{\mathbb B}^*}({\mathbb B}_\Phi^*)-\mathcal L_{{\mathbb B}^*}({\mathbb B}^*)=\int_0^1j''\leq j''_\Phi(0)+\int_0^1\left|j_\Phi''(\xi)-j_\Phi''(0)\right| d\xi.\end{equation}
The key is now to prove the following Lemma:
\begin{lemma}\label{Le:Control}
There exists a modulus of continuity, that is, a continuous function $\eta:\textnormal{I\kern-0.21emR}_+\to \textnormal{I\kern-0.21emR}_+$ such that $\omega(0)=0$, such that
\begin{equation}|j_\Phi''(t)-j_\Phi''(0)|\leq \eta\left(\Vert \Phi\Vert_{W^{2,p}}\right) \Vert \Phi \cdot \nu \Vert_{L^2(\partial {\mathbb B}^*)}^2.\end{equation}
\end{lemma}
Indeed, Lemma \ref{Le:Control} implies Proposition \ref{Pr:NormalDeformation} in the following way: assuming it holds then
\begin{align}\mathcal L_{{\mathbb B}^*}({\mathbb B}_\Phi^*)-\mathcal L_{{\mathbb B}^*}({\mathbb B}^*)&=\int_0^1j''\leq j''_\Phi(0)+\int_0^1\left|j_\Phi''(\xi)-j_\Phi''(0)\right| d\xi
\\ &\leq -c_0\Vert \Phi\cdot\nu\Vert_{L^2(\partial {\mathbb B}^*)}^2+ \eta\left(\Vert \Phi\Vert_{W^{2,p}}\right) \Vert \Phi \cdot \nu \Vert_{L^2(\partial {\mathbb B}^*)}^2
\\&\leq -\frac{c_0}2\Vert \Phi\cdot\nu\Vert_{L^2(\partial {\mathbb B}^*)}^2 \text{ for $\Vert \Phi\Vert_{W^{2,p}}$ small enough}
\\&\leq -\frac{c_0}2\frac{1}{\operatorname{Per(\partial {\mathbb B}^*)}}\left(\int_{\partial {\mathbb B}^*} |\Phi \cdot \nu|\right)^2 \text{ by the Cauchy-Schwarz inequality}
\\&\leq -\tilde c_0 \operatorname{Vol}\left({\mathbb B}^*_\Phi\Delta {\mathbb B}^*\right)^2.\end{align}
The proof of Lemma \ref{Le:Control} is extremely similar to the proof of \cite[Proposition 23]{MRB2020} and is mostly a technical adaptation of \cite{DambrineLamboley}. For this reason, we postpone it to Appendix \ref{Ap:Control} and briefly sketch here why this $L^2$ norm of $\Phi\cdot \nu$ is, in contrast to the $H^{\frac12}$ usually required in shape optimisation \cite{DambrineLamboley}, the optimal norm here. If we consider, for instance, at at given shape $E$ the second order shape derivative of the Lagrangian, we have
\begin{equation}\mathcal L_{{\mathbb B}^*}''(E)[\Phi,\Phi]=\underbrace{\iint_{(0;T)\times \partial E}p'\left(\Phi\cdot \nu\right)}_{=I_1}+\underbrace{\int_{\partial E}\left(\Phi\cdot \nu\right)^2\left(\mathscr H \int_{0}^T p_E+\int_0^T\frac{\partial p_E}{\partial \nu}\right)-\overline \Psi_{\partial {\mathbb B}^*}\int_{\partial E}\mathscr H \left(\Phi\cdot \nu\right)^2}_{=:I_2}\end{equation} where:
\begin{enumerate}
\item $\mathscr H$ is the mean curvature of $\partial E$,
\item $p_E$ solves
\begin{equation}\begin{cases}
\frac{\partial p_E}{\partial t}+\Delta p_E=-u_E\text{ in }{(0;T)\times \O}\,,
\\ p_E(T,\cdot)=0\,,
\\ p_E(t,\cdot)=0\text{ on }(0;T)\times \partial {\Omega},
\end{cases}
\end{equation}
\item $u'$ solves
\begin{equation}\begin{cases}
\frac{\partial u'}{\partial t}-\Delta u'=0\text{ in }{(0;T)\times \O}\,,
\\ u'(0,\cdot)=0\,,
\\\llbracket \partial_\nu u'\rrbracket=-1\text{ on }(0;T)\times \partial E,
\\ u'(t,\cdot)=0\text{ on }(0;T)\times \partial {\Omega},
\end{cases}
\end{equation}
\item $p'$ solves
\begin{equation}\begin{cases}
\frac{\partial p'}{\partial t}+\Delta p'=-u'\text{ in }{(0;T)\times \O}\,,
\\ p'(T,\cdot)=0\,,
\\ p'(t,\cdot)=0\text{ on }(0;T)\times \partial {\Omega},
\end{cases}
\end{equation}
\item and
$$\left.\overline \Psi_{\partial {\mathbb B}^*}=\int_0^T p^*(t,\cdot)\right|_{\partial {\mathbb B}^*}$$ is the Lagrange multiplier associated with the volume constraint.
\end{enumerate}
Now, by the regularity estimates of Proposition \ref{Pr:Regularity} and by standard Schauder estimates, it is natural to expect that
\begin{equation}
\Vert I_2\Vert\leq M \left\Vert \Phi \cdot \nu \right\Vert_{L^2(\partial E)}^2.
\end{equation}
To prove that the same estimate holds for $I_1$, it suffices, by continuity of the trace, to obtain
\begin{equation}\iint_{(0;T)\times \O} |{\nabla} p'|^2\leq M \int_{\partial E^*}(\Phi\cdot\nu)^2.\end{equation} However, this just follows from standard parabolic estimates, provided we can prove that
\begin{equation}\label{Eq:Guan}
\iint_{(0;T)\times \O} (u')^2\leq M \int_{\partial E^*}(\Phi\cdot\nu)^2.\end{equation} To prove \eqref{Eq:Guan}, we use $u'$ as a test function in the weak equation on $u'$ and obtain, by the Cauchy-Schwarz inequality and the continuity of the trace,
\begin{align}
\frac{\partial}{\partial t}\int_{\Omega} (u')^2(t,\cdot)+\int_{\Omega} |{\nabla} u'|^2&=\int_{\partial E}(\Phi\cdot \nu)u'
\\&\leq M\Vert \Phi \cdot\nu\Vert_{L^2(\partial E)}\Vert {\nabla} u'\Vert_{L^2({\Omega})}.
\end{align}
Integrating this inequality in time yields the required result and we hence obtain
\begin{equation}\left|\mathcal L_{{\mathbb B}^*}''(E)[\Phi,\Phi]\right|\leq M\Vert \Phi\cdot \nu \Vert_{L^2(\partial E)}^2.\end{equation}
As a consequence, the $L^2$ norm should be the optimal coercivity norm.
\subsection{Quantitative bathtub principle: using the convexity of the functional}
In this section, we will fully exploit the convexity of the functional. We first heuristically explain how we are going to make use of it.
\paragraph{Heuristics}
Let us assume that we are working with a competitor $f\in {\overline {\mathcal M}}({\Omega})$, and let us define $p_f$ as the adjoint state associated to $f$ (solution of \eqref{Eq:Adjoint}). Hence, for an admissible perturbation $h$ at $f$ (i.e, such that $f+th\in \overline{\mathcal M}({\Omega})$ for any $t\geq 0$ small enough), the derivative of $\mathcal J_T$ at $f$ in the direction $h$ is given by (Proposition \ref{Pr:Adjoint})
\begin{equation}\dot{\mathcal J_T}(f)[h]=\int_{\Omega} h(x)\left(\int_0^T p_f(t,x)dt\right)dx.\end{equation}
Since $\mathcal J_T$ is convex (Proposition \ref{Pr:Convexity}), we have
\begin{equation}\label{Eq:Heuristics}
\mathcal J_T(f+h)-\mathcal J_T(f)\geq \dot{\mathcal J_T}(f)[h].
\end{equation}
As a consequence, let us assume that $f=\mathds 1_E$. In order to maximise the right hand side of \eqref{Eq:Heuristics}, we need to choose $h$ such that, defining $\overline \Psi:=\int_0^T p_f(t,\cdot)dt$, and choosing $\overline c>0$ such that $\operatorname{Vol}\left(\{\overline \Psi>\overline c\}\right)=V_0$ (assuming this set is uniquely defined and regular),
\begin{equation}f+h=\mathds 1_{\{\overline \Psi>\overline c\}}=\mathds 1_{\overline E}\end{equation} and so we obtain the lower bound
\begin{equation}\label{Eq:FWF}\mathcal J_T(\mathds 1_{\overline E})-\mathcal J_T(\mathds 1_E)\geq \int_{\overline E} \overline \Psi-\int_{E}\overline \Psi.\end{equation} Now, as we will see, when $f$ is close enough to $f^*$, $\overline E$ should be a normal deformation of ${\mathbb B}^*$, and the only thing left is thus to quantify
\begin{equation}\label{Eq:Start}
\int_{\overline E} \overline \Psi-\int_{E}\overline \Psi.
\end{equation}
Indeed, using \eqref{Eq:FWF} we obtain
\begin{equation}\mathcal J_T(\mathds 1_{{\mathbb B}^*})-\mathcal J_T(\mathds 1_E)\geq J_T(\mathds 1_{{\mathbb B}^*})-\mathcal J_T(\mathds 1_{\overline E})+\mathcal J_T(\mathds 1_{\overline E})-\mathcal J_T(\mathds 1_{ E})\geq C \operatorname{Vol}(E^*\Delta \overline E)^2+\mathcal J_T(\mathds 1_{\overline E})-\mathcal J_T(\mathds 1_{ E}).\end{equation} Here, $C>0$ is given by Proposition \ref{Pr:NormalDeformation}.
Since $\overline f:=\mathds 1_{\overline E}$ is a maximiser of $T_{\overline \Psi}:f\mapsto \int_{\Omega} f\overline \Psi$ in $\overline{\mathcal M}({\Omega})$, it turns out that estimating \eqref{Eq:Start} amounts to providing a quantitative estimate for the linear optimisation problem
\begin{equation}\sup_{f\in \overline{\mathcal M}({\Omega})}T_{\overline \Psi}(f)\end{equation} which is exactly the quantitative version of the bathtub principle.
The goal of the present paragraph is to give a uniform bathtub principle that was presented in a slightly different form in the section devoted to Theorem \ref{Th:Td}, see Proposition \ref{Pr:BathtubTD} above.
\begin{proposition}\label{Pr:Bathtub}
Let $\beta>0$ and let $\{\psi_i\}_{i\in I}\subset \mathscr C^{1,\beta}({\Omega};\textnormal{I\kern-0.21emR}_+)^I$ be a closed subset of $\mathscr C^{1,\beta'}({\Omega})$ for some $\beta'<\beta$. We assume that:
\begin{enumerate}
\item For every $i\in I$ there exists a unique $c_i$ such that, up to a set of measure 0,
\begin{equation}{\Omega}_i:=\{\psi_i>c_i\}=\{\psi_i\geq c_i\}\end{equation} and
\begin{equation}\label{Eq:Downtown}\forall i \in I\,, \operatorname{Vol}({\Omega}_i)=V_0\,, \overline L=\sup_{i\in I}\operatorname{Per}({\Omega}_i)<+\infty.\end{equation}
We define, for any $i\in I$,
$$\overline f_i:=\mathds 1_{{\Omega}_i}.$$
\item There exists $M>0$ such that
\begin{equation}\sup_{i\in I}\Vert \psi_i\Vert_{\mathscr C^{1,\beta}}\leq M_I.\end{equation}
\item There exists $\underline \mu>0$ such that
\begin{equation}\label{Eq:Feather}\inf_{i\in I}\inf_{\partial {\Omega}_i}\left\{-\frac{\partial \psi_i}{\partial\nu}\right\}\geq \underline \mu.\end{equation}
\end{enumerate}
Then there exists a constant $\overline \omega>0$ such that \begin{equation}
\forall i\in I\,, \forall f\in {\overline{\mathcal M}}({\Omega})\,, \int_{\Omega} (\overline f_i-f)\psi_i\geq \overline \omega \Vert \overline f_i-f\Vert_{L^1({\Omega})}^2.
\end{equation}
\end{proposition}
The proof of this Proposition is very similar to that of Proposition \ref{Pr:BathtubTD}.
\begin{proof}[Proof of Proposition \ref{Pr:Bathtub}]
We just need to prove that, thanks to our assumption, we can bring ourselves back to the proof of Proposition \ref{Pr:BathtubTD}. This is done using the Schwarz rearrangement, as was done in \cite{MRB2020}.
From the bathtub principle we have, for any $i\in I$, that $f_i^*:=\mathds 1_{{\Omega}_i}$ is the unique solution of
\begin{equation}\sup_{f\in \mathcal M({\Omega})}\int_{\Omega} f\psi_i.\end{equation}
By the uniform H\"{o}lder continuity of $\{{\nabla} \psi\}_{i\in I}$ there exists $\overline {\varepsilon}>0$ that only depends on $M_I$ and $\underline \mu$ such that
\begin{equation}\label{Eq:OS}\forall x \in {\Omega}\,, \forall i\in I\,, \psi_i(x)\in \Big(c_i-\overline {\varepsilon};c_i+\overline {\varepsilon}\Big)\Rightarrow \left|{\nabla} \psi_i\right|(x)\geq \frac{\underline \mu}2\,, \sup_{{\varepsilon}\in (-\overline {\varepsilon};\overline {\varepsilon})}\sup_{i\in I}\operatorname{Per}(\{\psi_i=c_i+{\varepsilon}\})<+\infty.\end{equation} Let us fix such an $\overline {\varepsilon}$.
We now reduce ourselves to the case of radially symmetric function:
\paragraph{Reduction to radially symmetric functions}
For any $i\in I$, let us consider the distribution function $\mathscr L_i$ of $\psi_i$. From \eqref{Eq:OS}, $\mathscr L_i$ is $\mathscr C^1$ in $(c_i-\overline {\varepsilon};c_i+\overline {\varepsilon})$ and so, letting $\psi_i^\#$ be the Schwarz rearrangement of $\psi_i$, we have
$$\int_{\partial \{\psi_i>c_i\}} \frac1{\left|\frac{\partial \psi_i}{\partial \nu}\right|}=-\mathscr L_i'(c_i)=\int_{\partial {\mathbb B}(0;r^*)}\frac1{\left|\frac{\partial \psi_i^\#}{\partial \nu}\right|}.$$
Given the uniform perimeter bound \eqref{Eq:OS} on the level sets close to $\{\psi_i=c_i\}$, it thus follows that there exists a constant $C>0$ and $\underline {\varepsilon}>0$ such that $\{\psi_i^\dagger\}_{i\in I}$ satisfies, in a $(r^*-\underline {\varepsilon};r^*+\underline {\varepsilon})$,
\begin{equation}\label{Eq:V}\forall i\in I\,, \left|\frac{d\psi_i^\dagger}{dr}\right|\geq C\underline \mu. \end{equation}
We can then observe the following thing: by equimeasurability of the Schwarz rearrangement, we have, for every $f\in \mathcal M({\Omega})$, the following property: if $\Vert f-\overline f_i\Vert_{L^1({\Omega})}=\delta$ then, defining $\mathbb A_\delta$ as the unique annulus such that $\operatorname{Vol}(\mathbb A_\delta)=V_0\,, \operatorname{Vol}(\mathbb A_\delta \Delta {\mathbb B}^*)=\delta$, the Haryd-Littlewood inequality and the equimeasurability of the rearrangement ensure that
\begin{equation}
\int_{\Omega} (\overline f_i-f)\psi_i\geq \int_{\Omega} (f^*-\mathds 1_{\mathbb A_\delta})\psi_i^\#.\end{equation} Hence, it suffices to prove that
\begin{equation}\int_{\Omega} (f^*-\mathds 1_{\mathbb A_\delta})\psi_i^\#\geq \underline \omega \delta^2\end{equation} where $\underline \omega$ does not depend on $i$. Thanks to \eqref{Eq:V}, the rest of the proof follows along the same exact lines as Proposition \ref{Pr:BathtubTD}.
\end{proof}
\subsection{Combining the bathtub principle and shape derivatives}
To conclude the proof of Theorem \ref{Th:Ti}, it thus only remains to prove the following proposition:
\begin{proposition}\label{Pr:Le}
Estimate \eqref{Eq:Le} holds.\end{proposition}
\begin{proof}\label{Pr:Le}
We argue by contradiction and assume that Estimate \eqref{Eq:Le} does not hold. Let us then consider a sequence $\{f_k\}_{k\in \textnormal{I\kern-0.21emN}}$ such that
\begin{equation}\frac{\mathcal J_T(f^*)-\mathcal J_T(f_k)}{\Vert f_k-f^*\Vert_{L^1({\Omega})}^2}\underset{k\to \infty}\rightarrow 0.\end{equation}
As in the proof of Theorem \ref{Th:Td}, the only closure point of $\{f_k\}_{k\in \textnormal{I\kern-0.21emN}}$ (in a weak $L^\infty-*$ sense) is $f^*$. We introduce, for any $k\in \textnormal{I\kern-0.21emN}$,
\begin{equation}\delta_k:=\Vert f_k-f^*\Vert_{L^1({\Omega})}.\end{equation} Up to replacing $f_k$ with $f_{\delta_k}$, we can assume that $f_k=f_{\delta_k}$.
Let us define, for any $k\in \textnormal{I\kern-0.21emN}$, $p_k$ as the adjoint state (solution of \eqref{Eq:Adjoint}) with $f=f_k$. From standard parabolic regularity and Proposition \ref{Pr:Regularity}, for any $\beta\in(0;1)$ there exists $M_\beta$ such that for any $t\in [0;T]$
\begin{equation}
\Vert p_k\Vert_{\mathscr C^{2,\beta}({\Omega})}\leq M_\beta,
\end{equation}
and hence, since $f_k\underset{k\to \infty}\rightarrow f^*$, we obtain
\begin{equation}
p_k\underset{k\to \infty}{\overset{\mathscr C^{2,\beta}({(0;T)\times \O})}\rightarrow} p^*\end{equation}where $p^*$ is the adjoint state associated with $f=f^*$.
Let, for any $k\in \textnormal{I\kern-0.21emN}$,
\begin{equation}\Psi_k:=\int_0^T p_k.\end{equation} From the same arguments,
\begin{equation}\Psi_k\underset{k\to \infty}{\overset{\mathscr C^{2,\beta}({\Omega})}\rightarrow}\Psi^*:=\int_0^T p^*.\end{equation} $\Psi^*$ is radially symmetric, it is decreasing and its only level set of volume $V_0$ is $\mathbb B^*$. Furthermore, from the same arguments as in Claim \ref{Cl:NDA}, we also have
\begin{equation}\forall \eta>0\,, \inf_{r>\eta}\left|\frac{\partial \Psi^*}{\partial r}\right|(r^*)=\ell(\eta)>0.\end{equation} Let, for any $k\in \textnormal{I\kern-0.21emN}$, $c_k$ be such that $\operatorname{Vol}\left(\{\Psi_k\geq c_k\}\right)\geq V_0\,, \operatorname{Vol}\left(\{\Psi_k>c_k\}\right)\leq V_0$.
Since $f_k\underset{k\to \infty}\rightarrow f^*$ and since $\mathcal J_T(f_k)\geq\mathcal J_T(\mathds 1_{\{\Psi_k>c_k\}})$ by convexity of the functional, it follows that $\left\{\mathds 1_{\{\Psi_k>c_k\}}\right\}_{k\in \textnormal{I\kern-0.21emN}}$ converges weakly to $f^*$. Since $f^*$ is an extreme point of $\overline{\mathcal M}({\Omega})$, this convergence occurs in $L^1$.
We choose $\eta>0$ small enough so that, for any $k\in \textnormal{I\kern-0.21emN}$ large enough, $\{\psi_k=c_k\}\cap \{0<r<\eta\}=\emptyset$. This is possible because $\Psi^*$ is radially symmetric and decreasing: indeed, argue by contradiction and assume that there exists a sequence $\{x_k\}_{k\in \textnormal{I\kern-0.21emN}}$ converging to 0 such that for any $k\in \textnormal{I\kern-0.21emN}$, $\Psi_k(x_k)=c_k$. Let $c$ be the limit of the sequence $\{c_k\}$. Since $\operatorname{Vol}(\{\Psi_k>c_k\})=V_0$, there exists $\eta'>0$ such that for any $k\in \textnormal{I\kern-0.21emN}$ there exists $y_k\,, \Vert y_k\Vert>\eta'$ such that $\Psi_k(y_k)>c_k$. Passing to the limit, there exists $y'>0$ such that $\Psi^*(y')\geq c=\lim_{k\to \infty}\Psi_k(x_k)=\Psi^*(0)$ and so $\Psi^*$ can not be decreasing. Hence such an $\eta>0$ exists.
As a consequence, for such an $\eta$ we have, for any $k$ large enough,
\begin{equation}\inf_{x\,, \Vert x\Vert>\eta}\left| {\nabla} \Psi_k\right|(x)\geq \frac{\ell(\eta)}2.\end{equation} Thus, the level set $\{\Psi_k=c_k\}$ is a $\mathscr C^1$ curve and
\begin{equation}\inf_{\{\Psi_k=c_k\}}|{\nabla} \Psi_k|\geq\frac{\ell(\eta)}2.\end{equation} Since $\{\Psi_k\}_{k\in \textnormal{I\kern-0.21emN}}$ is uniformly bounded in $\mathscr C^{2}({\Omega})$, these sets have uniformly Lipschitz boundaries.
It follows that the sequence of sets $\left\{\{\Psi_k>c_k\}\right\}_{k\in \textnormal{I\kern-0.21emN}}$ converges in Hausdorff distance to $\{\Psi^*>c^*\}={\mathbb B}^*$ where $c^*=\Psi^*(r^*)$.
Finally, for any $k\in \textnormal{I\kern-0.21emN}$ large enough, $E_k:=\{\Psi_k=c_k\}$ is a normal deformation of ${\mathbb B}^*$. Indeed, assuming that it is not, there exists a sequence $\{x_k\}_{k\in \textnormal{I\kern-0.21emN}}\in( \partial {\mathbb B}^*)^\textnormal{I\kern-0.21emN}$ and two sequences $\{t_{i,k}\}_{i=1,2,k\in \textnormal{I\kern-0.21emN}}$ converging to 0 such that $\Psi_k(x_k+t_{i,k}\nu(x_k))=c_k$, $i=1,2$. This gives the existence of a $t_k$, converging to 0 as $k\to \infty$, such that $\langle {\nabla} \Psi_k(x_k+t_k\nu(x_k)),\nu(x_k)\rangle=0$, which yields a contradiction when passing to the limit. Thus, $\partial E_k$ converges $W^{2,p}$ to $\partial {\mathbb B}^*$ for all $p>1$, and in $\mathscr C^{2}$, $\beta\in (0;1)$, and the sequence $\{\operatorname{Per}(\{\Psi_k=c_k\})\}_{k\in \textnormal{I\kern-0.21emN}}$ is bounded.
We can now prove Estimate \eqref{Eq:Le}: from the convexity of the functional and the fact that, for any $k\in \textnormal{I\kern-0.21emN}$, $f_k$ solves $(P_{\delta_k})$, there exists a subset $F_k$ of ${\Omega}$ such that $f_k=\mathds 1_{F_k}$. Let $E_k=\{\Psi_k>c_k\}$ be the unique level-set of $\Psi_k$ of measure $V_0$. By convexity of the functional,
\begin{equation}\mathcal J_T(\mathds 1_{E_k})-\mathcal J_T(f_k)\geq \int_{\Omega} (\mathds 1_{E_k}-\mathds 1_{F_k})\Psi_k.\end{equation}
We are now in a position to apply Proposition \ref{Pr:Bathtub}: with the $\overline \omega$ given in Proposition \ref{Pr:Bathtub}, we thus have
\begin{equation}\mathcal J_T(\mathds 1_{E_k})-\mathcal J_T(f_k)\geq \overline \omega \operatorname{Vol}(F_k\Delta E_k)^2.\end{equation}
Then, as $E_k$ is a normal deformation of ${\mathbb B}^*$ we can apply Proposition \ref{Pr:NormalDeformation} and obtain, for $C>0$ given by Proposition \ref{Pr:NormalDeformation},
\begin{equation}\mathcal J_T(\mathds 1_{{\mathbb B}^*})-\mathcal J_T(\mathds 1_{E_k})\geq C\operatorname{Vol}(E_k\Delta {\mathbb B}^*)^2.\end{equation}
We obtain the existence a $C'>0$ such that
\begin{equation}\mathcal J_T(\mathds 1_{{\mathbb B}^*})-\mathcal J_T(\mathds 1_{E_k})\geq C'\left( \operatorname{Vol}(E_k\Delta {\mathbb B}^*)^2+\operatorname{Vol}(F_k\Delta E_k)^2\right).\end{equation} However, by the triangle inequality in $L^1$ and the arithmetic-geometric inequality,
\begin{equation}\operatorname{Vol}({\mathbb B}^*\Delta F_k)^2\leq \frac12\left(\operatorname{Vol}(F_k\Delta E_k)^2+\operatorname{Vol}(E_k\Delta {\mathbb B}^*)^2\right).\end{equation}The conclusion follows.
\end{proof}
\section{Conclusion}\label{Se:Concl}
\subsection{Structure of the problem, structure of the proof}
In this paper, we have investigated two possible approaches to quantitative inequalities for time-evolving optimal control problems. While Theorem \ref{Th:Td}, dealing with time-dependent controls, is more powerful than Theorem \ref{Th:Ti}, it is likely that its proof does not generalise easily to other domains. Indeed, the first step of the proof is to identify, explicitly, the maximisers of an auxiliary optimisation problem, which can not be done in general, non-spherical domains.
On the other hand, the proof of Theorem \ref{Th:Ti} is susceptible of applying to other cases. Let us specify what we mean: considering a controlled heat equation
\begin{equation}\frac{\partial u_f}{\partial t}-\Delta u_f=f\end{equation} with Dirichlet boundary conditions, and where $f\in \overline{\mathcal M}({\Omega})$, let $f^*$ be a solution of \eqref{Eq:PvTi}. The convexity of the functional $\mathcal J_T$ (Proposition \ref{Pr:Convexity}) holds independently of the geometry of the domain an so any maximiser $f^*$ writes $\mathds 1_{E^*}$ for some subset $E^*$ of ${\Omega}$. In order to carry out the proof of Theorem \ref{Th:Ti} in this new domain, several things are in order:
\begin{enumerate}
\item The regularity of optimal sets: each set $E^*$ such that $f^*=\mathds 1_{E^*}$ is a solution of \eqref{Eq:PvTi} needs to be smooth enough that shape derivatives of the criterion may be computed. It is unclear at this stage whether or not the classical regularity works valid in the stationary case may be applied to obtain such regularity.
\item The coercivity of shape Lagrangians: defining $I^*:=\{E^*\subset {\Omega}\,, \mathds 1_{E^*}\text{ solves \eqref{Eq:PvTi}}\}$ and assuming that each $E^*\in I^*$ is smooth enough to compute first and second order shape derivatives, one needs to check that, defining the Lagrangian $L_{E^*}$ associated with the volume constraint, there exists a constant $\alpha>0$ such that, for any $E^*\in I^*$ and any admissible vector field $\Phi$ at $E^*$, there holds
\begin{equation}
L_{E^*}(E^*)[\Phi,\Phi]\geq \alpha \Vert \Phi \cdot \nu\Vert_{L^2(\partial E^*)}^2.\end{equation} This kind of estimate seems to be extremely challenging to obtain in general, as indicated by the fact that, in this paper, such a coercivity was obtained by explicit diagonalisation of the shape hessian. Such diagonalisation may not be available in general.
\end{enumerate}
If these two assumptions are satisfied, then we believe that the method of proof of Theorem \ref{Th:Ti} may adapt.
\subsection{The optimal coercivity norm for other types of constraints}\label{Cl:Time}
As mentioned in the Introduction, an interesting question is that of knowing whether or not the coercivity norm obtained in Theorem \ref{Th:Td} remains unchanged when considering other types of $L^1$ constraints. Indeed, let us consider the following variation: defining
\begin{equation}\tilde{\mathcal M}({\Omega}):=\left\{f\in L^\infty({(0;T)\times \O})\,, 0\leq f\leq 1\text{ a.e.,} \iint_{(0;T)\times \O} f=V_0\right\}\end{equation} we investigate the optimisation problem
\begin{equation}\sup_{f\in \tilde{\mathcal M}({\Omega})}J_T(f).\end{equation} Here, the convexity of the functional $\mathcal J_T$ is still valid, so that a solution of this new problem writes $f^*=\mathds 1_{E^*}$, with $E^*$ a measurable subset of $E^*$. Then, if one were to compare $f^*$ with a competitor $f=\mathds 1_E$, it would be more natural to expect the \textquotedblleft classical \textquotedblright discrepancy norm
\begin{equation}\operatorname{Vol}(E^*\Delta E)^2=\left(\iint_{(0;T)\times \O} \left|f-f^*\right|\right)^2\end{equation} to be optimal. We do not believe this to be true, however, and we believe that the correct discrepancy norm remains
\begin{equation}\int_0^T\operatorname{Vol}(E^*(t)\Delta E(t))^2dt.\end{equation}
To give some explanation as to why we believe this is to be expected, we can once again consider the case of the ball ${\Omega}=\mathbb B(0;R)$. Once again, the rearrangement arguments used throughout the paper remain valid, and, for almost every $t\in (0;T)$, $E^*(t)$ is a centred ball of radius $r(t)\geq 0$. We expect several difficulties in treating this problem (most notably, we expect the (non)-degeneracy of $r$, or, in other terms, the control of the set $\{r=0\}$, to be very hard to obtain) but the methods of Theorem \ref{Th:Td} should once again provide a quadratic estimate at each time $t$, yielding the aforementioned stronger estimate. We underline once again that, at the present moment, it is unclear to us how one may fully analyse this type of global constraint.
\subsection{Theorem \ref{Th:Td}: on the Assumption ${\varepsilon}>0$}
One may also argue that the assumption ${\varepsilon}>0$ is artificial. At this stage, and since we use in a crucial manner the uniform non-degeneracy of the switch function (Claim \ref{Cl:NDA}), we are not yet in a position to give a proof that would bypass this assumption. However, it should be noted that our proof makes use of very strong regularity properties in order to derive the uniform bathtub principle. It would be interesting to see if, using the general quantitative Hardy-Littlewood inequality {\cite{Cianchi2008}} one could bypass the strength required in the present proof to obtain the case ${\varepsilon}=0$ (and, in general, it would be extremely interesting to use {\cite{Cianchi2008}} to see if Theorem \ref{Th:Td} could be obtained in more general domain).
\subsection{Using the quantitative isoperimetric inequality to obtain our results}\label{SD}
We touch on another way which it would be interesting to investigate, that of using the quantitative isoperimetric inequality in order to obtain Theorem \ref{Th:Td}. It would amount, in the approach of \cite{RakotosonMossino} (see Section \ref{Se:Uniq}), to supplementing the isoperimetric inequality in \eqref{Eq:Co}. We expect that this would lead to a control of the isoperimetric deficit $\mathcal A(t,\tau)$ of the level sets $\{u_f(t,\cdot)>\tau\}$ in the sense that we could give a lower bound of the form $\int_0^T \int_0^{\Vert u_f\Vert_{L^\infty}}\mathcal A(t,\tau)^2d\tau dt$, but is unclear how this would then translate to a control of the isoperimetric deficit of $f$.
\subsection{Minimisation problems}
We believe the proof for minimisation problems works in exactly the same way, as we also have an explicit description of minimisers using rearrangement techniques.
\subsection{Technical obstructions and possible generalisations for bilinear control problems}\label{Cl:Bilinear}
Finally, we touch upon bilinear control problems. Let us assume that we are working with the state equation
\begin{equation}\begin{cases}
\frac{\partial u_f}{\partial t}-\Delta u_f=fu_f\text{ in }{(0;T)\times \O}\,,
\\ u_f(t=0)=u_0\geq 0\,, u_0\neq 0\,,
\\ u_f(t,\cdot)=0\text{ on }(0;T)\times \partial {\Omega}.\end{cases}\end{equation} The maximisation problem reads the same:
\begin{equation}\sup_{f\in \mathcal M_T({\Omega})}\frac12\iint_{(0;T)\times \O} u_f^2.\end{equation} Here we con once again explicitly characterise the maximisers using rearrangement techniques. However: the convexity of the functional is no longer obvious, and it can be checked that the switch function is here given
\begin{equation}\Psi=p_fu_f\end{equation} where $p_f$ solves
\begin{equation}\begin{cases}
\frac{\partial p_f}{\partial t}+\Delta p_f=-fp_f-u_f\,,
\\ p_f(T,\cdot)=0\,,
\\ p_f(t,\cdot)=0\text{ on }(0;T)\times \partial {\Omega}.\end{cases}\end{equation} Here we see our first difference with our approach, which is that the switch function can merely be expected to be $\mathscr C^{0,\alpha}$, which is in contrast with the $\mathscr C^{2,\alpha}$ regularity we obtained in our paper. Maybe it is possible to bypass this problem using the tools of \cite{Cianchi2008}.
\bibliographystyle{abbrv}
|
\section{Introduction} \label{intro}
Biomedical knowledge acquisition is an important task in information retrieval and knowledge management.
Professionals as well as the general public need effective assistance to access, understand and consume complex biomedical concepts.
For example, doctors always want to be aware of up-to-date clinical evidence for the diagnosis and treatment of diseases under the scheme of Evidence-based Medicine \citep{sackett1997evidence}, and the general public is becoming increasingly interested in learning about their own health conditions on the Internet \citep{fox2012online}.
Traditionally, Information Retrieval (IR) systems, such as PubMed, have been used to meet such information needs.
However, classical IR is still not efficient enough \citep{hersh2002factors, jacquemart2003towards, lee2006beyond, russell2017expert}.
For instance, \citet{russell2017expert} reported that it requires 4 expert hours to answer complex medical queries using search engines. Compared with the retrieval systems that typically return a list of relevant documents for the users to read, Question Answering (QA) systems that provide direct answers to users' questions are more straightforward and intuitive.
In general, QA itself is a challenging benchmark Natural Language Processing (NLP) task for evaluating the abilities of intelligent systems to understand a question, retrieve and utilize relevant materials and generate its answer.
With the rapid development of computing hardware, modern QA models, especially those based on deep learning \citep{cheng2016long, seo2016bidirectional, chen2017reading, peters2018deep, devlin2019bert}, achieve comparable or even better performance than human on many benchmark datasets \citep{hermann2015teaching, rajpurkar2016squad, joshi2017triviaqa, rajpurkar2018know, yang2018hotpotqa} and have been successfully adopted in general domain search engines and conversational assistants \citep{qiu2017alime, zhou2020design}.
The Text REtrieval Conference (TREC) QA Track has triggered the modern QA research \citep{voorhees2001trec}, when QA models were mostly based on IR.
\citet{zweigenbaum2003question} first identified the distinctive characteristics of BQA over general domain QA.
Later, many classic BQA systems have been proposed, such as EPoCare \citep{niu2003answering}, PICO-(P: patient/problem, I: intervention, C: comparison, O: outcome) and knowledge-extraction-based BQA systems \citep{demner2005knowledge, demner2006answer, demner2007answering}, MedQA \citep{yu2007development}, \citet{terol2007knowledge}, \citet{weiming2007automatic}, Health on the Net QA (HONQA) \citep{cruchet2008supervised}, AskHERMES \citep{cao2011askhermes} etc.
Such systems employ complex pipelines with numerous question, document and answer processing modules, which is typically reflected by the IBM Watson system \citep{ferrucci2012watson}.
BioASQ \citep{tsatsaronis2015overview} is a cornerstone challenge that has been running annually since 2013 for the evaluation of biomedical natural language understanding systems.
A variety of BQA systems have been proposed in BioASQ, improving QA performance from approximately 20\% top factoid mean reciprocal rank and list F-measure in BioASQ 1 to approximately 50\% in BioASQ 8 \citep{Nentidis2020overview}.
Notably, the landscape of BioASQ participating models has been re-shaped by several NLP methodological revolutions:
1. The introduction of distributed word representations \citep{mikolov2013efficient, mikolov2013distributed};
2. Deep learning-based QA models such as Bi-Directional Attention Flow (Bi-DAF) \citep{seo2016bidirectional};
3. Large-scale Pre-trained Language Models (PLMs) represented by Embeddings for Language Models (ELMo) \citep{peters2018deep} and bidirectional encoder representations from transformers (BERT) \citep{devlin2019bert}.
Currently, almost all top-performing BQA systems use the biomedical PLMs (e.g. BioBERT \citep{lee2020biobert}), in their systems.
Furthermore, various other BQA challenges and datasets have been introduced to further facilitate BQA research in different directions, e.g.: LiveQA-Med \citep{abacha2017overview}, MEDIQA \citep{abacha2019overview, savery2020question} for consumer health, emrQA \citep{pampari2018emrqa} for clinical BQA, VQA-Rad \citep{lau2018dataset}, VQA-Med \citep{abacha2019vqa} and PathVQA \citep{he2020pathvqa} for visual BQA.
Despite the tremendous developments, BQA is still immature and faces several key challenges:
\begin{itemize}
\item \textbf{Dataset Scale, Annotation \& Difficulty}:
Most current BQA models utilize deep learning and are thus data-hungry.
However, annotating large-scale biomedical corpora or knowledge bases is prohibitively expensive.
As a result, current expert-annotated BQA datasets are small in size, with only hundreds to few thousand QA instances.
To build large-scale datasets, many works have attempted to automatically collect BQA datasets, but their utility is limited and their annotation quality is not guaranteed.
Furthermore, questions of most current BQA datasets do not require complex reasoning to answer.
\item \textbf{Domain Knowledge Not Fully Utilized}:
There are rich biomedical resources that encapsulate different types of domain knowledge, including large-scale corpora, various biomedical KBs, domain-specific NLP tools and images.
Unfortunately, most BQA models fail to utilize them effectively.
As a result, biomedical domain knowledge is not fully utilized, which can be potentially solved by fusing different BQA approaches.
\item \textbf{Lack of Explainability}:
Since biomedicine is a highly specialized domain, ideal systems should not only return the exact answers (e.g.: ``yes"/``no"), but also provide the explanations for giving such answers.
However, there are only a few BQA systems that are explainable.
\item \textbf{Evaluation Issues}:
Qualitatively, current evaluations mainly focus on certain modules, e.g. Machine Reading Comprehension (MRC), within a complete QA pipeline.
Quantitatively, most evaluation metrics do not consider rich biomedically synonymous relationships.
\item \textbf{Fairness and Bias}:
Most machine-learning-based BQA systems learn from historical data, such as scientific literature and electronic medical records, which
can be potentially biased and out of date.
However, studies on BQA fairness and model transparency are quite sparse.
\end{itemize}
This paper is organized as follows:
We first describe the scope of this survey in \S\ref{scope};
We then give an overview of the surveyed BQA approaches in \S\ref{approach}.
Various methods and datasets have been proposed for each BQA approach, and they are systematically discussed in \S4-8;
To conclude, we summarize several challenges of BQA and discuss potential future directions in \S\ref{challenge}.
\section{Survey Scope} \label{scope}
Biomedicine is a broad domain that covers a range of biological and medical sciences.
Since the term is often loosely used by the community, we specifically define several sub-domains of biomedicine, namely scientific, clinical, consumer health and examination, as the focus of this survey.
Each content type is defined by the most distinctive characteristics of their corresponding users, questions and expected answers, as shown in Table \ref{tab:content_type}.
It should be noted that the content types are not mutually exclusive, but most of our surveyed works belong to only one of them.
In this section, we introduce these contents in \S2.1-2.4, and we also describe some related surveys with the focus of different scopes in \S\ref{related_surveys}.
Several typical QA examples for each content type are shown in Table \ref{tab:content_eg}.
The datasets are selected from hand literature search on PubMed and Google Scholar with keywords such as “biomedical”, “biological”, “medical”, “clinical”, “health” and “question answering”. For each included dataset paper, we also checked their references and papers citing it.
We describe mostly the methods with state-of-the-art performance on the surveyed datasets.
\begin{table*}[ht!]
\small
\centering
\begin{tabular}{p{2.8cm}p{2.5cm}p{5cm}p{3.2cm}}
\toprule
\textbf{Content} & \textbf{Main User} & \textbf{Question motivation} & \textbf{Answer style} \\
\midrule
Scientific (\S\ref{scientific}) & -- & Learning cutting-edge scientific advances & Professional-level \\
\midrule
Clinical (\S\ref{clinical}) & Professionals & Assisting clinical decision making & Professional-level \\
\midrule
Consumer health (\S\ref{consumer}) & General public & Seeking advice or knowledge & Consumer-understandable \\
\midrule
Examination (\S\ref{examination}) & -- & Testing biomedical knowledge & Mostly choices \\
\bottomrule
\end{tabular}
\caption{Characteristics of different BQA contents. `--' denotes no specific users.}
\label{tab:content_type}
\end{table*}
\begin{table*}
\small
\centering
\begin{tabular}{p{2cm}p{4cm}p{4cm}p{3.5cm}}
\toprule
\textbf{Type / Dataset} & \textbf{Question} & \textbf{Context} & \textbf{Answer} \\
\midrule
\textbf{Scientific} \\
\midrule
BioASQ & Is the protein Papilin secreted? & [...] secreted extracellular matrix proteins, mig-6/papilin [...] & Yes \\
\midrule
Biomed-Cloze & Helicases are motor proteins that unwind double stranded \underline{ ? } into [...] & Defects in helicase function have been associated with [...] & nucleic acid \\
\midrule
\textbf{Clinical} \\
\midrule
emrQA & Has the patient ever had an abnormal BMI? & 08/31/96 [...] BMI: 33.4 Obese, high risk. Pulse: 60. resp. rate: 18 & BMI: 33.4 Obese, high risk \\
\midrule
CliCR & If steroids are used , great caution should be exercised on their gradual tapering to avoid \underline{ ? } & [...] Thereafter, tapering of corticosteroids was initiated with no clinical relapse. [...] & relapse \\
\midrule
\textbf{Consumer} \\
\midrule
MedQuAD & Who is at risk for Langerhans Cell Histiocytosis? & NA & Anything that increases your risk of [...] \\
\midrule
MEDIQA-AnS & What is the consensus of medical doctors as to whether asthma can be cured? And do you have [...] & Asthma Overview Asthma is a chronic lung disease that causes episodes of wheezing [...] & Asthma is a chronic disease. This means that it can be treated but not cured. [...] \\
\midrule
\textbf{Examination} \\
\midrule
HEAD-QA & The antibiotic treatment of choice for [...] is & 1. Gentamicin; 2. Erythromycin; 3. Ciprofloxacin; 4. Cefotaxime & 4. Cefotaxime \\
\bottomrule
\end{tabular}
\caption{Typical question-answer examples of different content types.}
\label{tab:content_eg}
\end{table*}
\subsection{Scientific} \label{scientific}
Scientific QA addresses cutting-edge questions whose answers need to be extracted or inferred from scientific literature, e.g.: ``Which cells express G protein-coupled receptors?".
Most of the new findings in the biomedical field are published in the form of scientific literature, whose size is growing at an unprecedented pace:
for example, MEDLINE\footnote{\url{https://www.nlm.nih.gov/bsd/medline.html}}, a bibliographic database of life sciences, contains references to over $30$M articles and about $2.7$k articles are added each day in 2019.
Given the huge number of scientific publications,
it's almost impossible to manually read all relevant studies and give comprehensive answers to scientific questions, so automatic answering of scientific questions is vital.
The BQA community's fight against COVID-19 is a great example of scientific QA.
There has been a surge of COVID-19-related publications \citep{li2020surging} that human experts find difficult to keep up with.
Consequently, it's important to develop automatic methods for natural language understanding of them.
To facilitate NLP studies on the COVID-19 literature, \citet{wang2020cord} release the CORD-19 corpus which contains more than 280k papers about the novel coronavirus.
Many BQA datasets have been introduced to help develop and evaluate models that answer COVID-19-related questions, e.g.: COVID-QA and COVIDQA datasets and the EPIC-QA challenge.
Several resources and methods \citep{sun2020analysis, poliak2020collecting, zhang2020cough} have been introduced to tackle the COVID-19 QA by the QE approach (\S\ref{qeapproach}).
The most distinctive feature of scientific BQA is that large-scale corpora like PubMed and PubMed Central are freely available, which contain $4.5$B and $13.5$B tokens, respectively.
In contrast, the entire English Wikipedia contains only $2.5$B tokens.
Besides, documents in PubMed and PubMed Central are semi-structured -- they have sections of background, introduction, methods, conclusion etc., which can be potentially exploited in building domain-specific datasets.
Consequently, the largest expert-annotated BQA dataset -- BioASQ, and most large-scale (semi-)automtically constructed BQA datasets are all scientific BQA datasets (discussed in \S\ref{automatic}).
Further exploiting the scale and structure of the scientific literature to design novel BQA tasks remains an interesting direction to explore.
\subsection{Clinical} \label{clinical}
Clinical QA focuses on answering healthcare professionals' questions about medical decision making for patients.
\citet{ely2000taxonomy} find the most frequent clinical questions are:
1. What is the drug of choice for condition x? ($11$\%);
2. What is the cause of symptom x? ($8$\%);
3. What test is indicated in situation x? ($8$\%);
4. What is the dose of drug x? ($7$\%);
5. How should I treat condition x (not limited to drug treatment)? ($6$\%);
6. How should I manage condition x (not specifying diagnostic or therapeutic)? ($5$\%);
7. What is the cause of physical finding x? ($5$\%);
8. What is the cause of test finding x? ($5$\%);
9. Can drug x cause (adverse) finding y? ($4$\%);
10. Could this patient have condition x? ($4$\%).
Most of the clinical questions shown above are generic (case 1-9) and largely non-specific to patients.
In this case, clinical QA is similar to consumer health QA (\S\ref{consumer}).
If the questions are specific to certain patients (e.g.: case 10), their Electronic Medical Records (EMRs) should be provided.
EMRs store all health-related data of each patient in both structured (i.e.: tables) and unstructured (i.e.: medical notes) formats.
Due to the complexity and size of the EMR data, it's time-consuming and ineffective for the doctors to manually check the EMRs for clinical questions about the patient.
Clinical QA systems can meet such information needs by quickly and accurately answering these questions.
The difficulty of clinical BQA largely lies in the annotation of QA pairs, where considerable medical expertise and reasoning across clinical notes should be required to answer the questions \citep{raghavan2018annotating}.
For this, \citet{pampari2018emrqa} use expert-annotated templates (e.g.: ``Has the patient ever been on \{medication\}?") with the existing i2b2 dataset annotations\footnote{\url{https://www.i2b2.org/NLP/DataSets/}} (e.g.: ``[...] Flagyl $<$medication$>$ [...]") to build the first large-scale EMR BQA dataset emrQA.
\citet{yue2020clinical} analyze the emrQA dataset and find: 1. the answers are usually incomplete; 2. the questions are often answerable without using domain knowledge.
Both are caused by the dataset collection approach of emrQA.
Another large-scale clinical QA dataset, CliCR \citep{suster2018clicr}, is built by cloze generation (\S\ref{automatic}).
\citet{roberts2017semantic} show that the structured information of EMRs can be effectively queried by semantic parsing, where the goal is to map the natural language questions to their logic forms \citep{kamath2018survey}, e.g.: Q: ``Was her ankle sprain healed?" Logic form: is\_healed(latest(lambda(ankle sprain))).
To tackle the clinical QA of structured EMR data, \citet{soni2019using} annotate a dataset of $1$k clinical questions with their logic forms.
Some paraphrasing-based data augmentation methods are also introduced to improve the performance of semantic parsers of EMR questions \citep{soni2019paraphrase, soni2020paraphrasing}.
\citet{wang2020text} propose TREQS, a two-stage generation model based on the sequence-to-sequence model and the attentive-copying mechanism, and show its effectiveness on their MIMICSQL dataset for the question-to-SQL (table-based) task.
Based on MIMICSQL dataset, \citet{park2020knowledge} propose a question-to-SPARQL (graph-based) dataset: MIMIC-SPARQL$*$.
TREQS also performs better on the graph-based dataset.
Radiology and pathology images play a vital role in the diagnosis and treatment of diseases.
Clinical QA also contains VQA tasks, e.g.: VQA-Rad \citep{lau2018dataset}, VQA-Med \citep{abacha2019vqa} and PathVQA \citep{he2020pathvqa}, which help doctors to analyze a large amount of images required for medical decision making and population screening.
\citet{ely2005answering} also study the obstacles that prevent physicians from answering their clinical questions, and find that doubting whether the answer exists is the most commonly (11\%) reported reason for not pursuing the answers and the most common obstacle in pursuing the answer is the failure to find the needed information in the selected resources (26\%).
Both problems can be solved by the clinical QA system.
Currently, the main challenge for building such systems is the lack of large-scale expert-annotated datasets that reflect the real demands in the clinic.
Apart from the high-price of deriving such annotations, there are also privacy and ethical issues for releasing them, especially when the datasets are based on EMRs.
Future clinical QA datasets should have larger scales, less noise and more diversity.
\subsection{Consumer Health} \label{consumer}
Consumer health questions are typically raised by the general public on search engines,
where online medical services provide people with great convenience as they are not limited by time and space.
As a result, rapidly increasing numbers of consumers are asking health-related questions on the Internet:
According to one report released by the Pew Research Center \citep{fox2012online}, over one-third of American adults have searched online for medical conditions that they might have.
Many try to find answers to their medical questions before going to a doctor or making decisions about whether to go to a doctor,
and their information needs range from self-diagnosis to finding medications.
It is vitally important to provide accurate answers for such questions,
because consumers are unable to judge the quality of medical contents.
Considering the contradiction between the great demands of consumers and the scarcity of medical experts, an automatic answering system is helpful for sharing medical resources to provide online medical service.
Some works \citep{zhang2017chinese, zhang2018multi,tian2019chimed} have exploited the doctors' answers to patients' questions on online medical consultation websites e.g.: XunYiWenYao\footnote{\url{http://xywy.com/}}, to build large-scale consumer health QA datasets.
These datasets are formatted as multi-choice BQA, where the task is to find the relevant or adopted answers.
However, the quality of such datasets is questionable since the answers are written by users from online communities and the forum data has intrinsic noise.
Remarkable efforts have been made by NLM's Consumer Health Information and Question Answering (CHIQA) project\footnote{\url{https://lhncbc.nlm.nih.gov/project/consumer-health-question-answering}}.
CHIQA \citep{demner2020consumer} is aimed at dealing with a vast number of consumer requests (over 90k per year) by automatically classifying the requests and answering their questions.
It also provides various datasets to develop consumer health BQA methods, including question decomposition and type, named entity and spelling error datasets.
For consumer health QA, understanding the questions of consumers is a vital but difficult step: such questions might contain many spelling and grammar errors, non-standard medical terms and multiple focuses \citep{roberts2016interactive,zhang2017chinese}.
For example, \citet{abacha2019summarization} find that consumers often submit long and complex questions that lead to substantial false positives in answer retrieval.
To tackle it, they introduce the MeQSum corpus\footnote{\url{https://github.com/abachaa/MeQSum}} that contains $1$k summarized consumer health questions and achieve the best $44.16$\% ROUGE-1 score using pointer-generator network \citep{see2017get} with semantic augmentation from question datasets.
On the other hand, most consumers have no biomedical domain knowledge, so the returned answers should be not only accurate but also explainable (\S\ref{explainability}), posing further challenges for consumer health QA.
\subsection{Examination} \label{examination}
Many general domain QA datasets that are extracted from examinations have been introduced \citep{shibuki2014overview, penas2014overview, khashabi2016question, lai2017race}.
Similarly, Examination BQA has that addresses automatic answering of medical examination questions also been explored.
For example, in many countries, medical licensure requires the passing of specific examinations, e.g.: USMLE\footnote{\url{https://www.usmle.org/}} in the US.
Test items in examinations often take the form of multi-choice questions,
and answering them requires comprehensive biomedical knowledge.
Several datasets have been released that exploit such naturally existing QA data, e.g.: HEAD-QA \citep{vilares2019head} and NLPEC \citep{li2020towards}.
Usually, no contexts are provided for such questions and automatic answering of them requires the systems to find supporting materials (e.g.: texts, images and KBs) as well as reason over them.
However, the real utility of examination QA is still questionable.
\subsection{Related Surveys} \label{related_surveys}
\citet{athenikos2010biomedical} systematically review BQA systems, mainly classic ones published before 2010.
Content-wise, they classify BQA into biological QA and medical QA, which roughly correspond to our scientific and clinical content types, respectively.
\citet{bauer2012usability} and \citet{sharma2015survey} briefly compare several classic BQA systems.
\citet{neves2015question} present a detailed survey of QA for biology, which we classify as scientific BQA in this paper.
This survey also discusses various biological BQA systems.
Recently, \citet{nguyen2019question} identifies several challenges in consumer health BQA, including the lack of publicly available datasets, term ambiguity, incontinuous answer spans and the lack of BQA systems for patients.
\citet{nguyen2019question} proposes a research plan to build a BQA system for consumer self-diagnosis.
\citet{kaddari2020biomedical} presents a brief survey that discusses several scientific BQA datasets and methods.
\section{BQA Approach Overview} \label{approach}
\begin{figure*}[h]
\centering
\includegraphics[width=\linewidth]{figs/revised_overview.pdf}
\caption{Overview of major biomedical question answering approaches. Boxes indicate methods or resources.}
\label{fig:arch}
\end{figure*}
\noindent The fundamental task of BQA is to answer a given questions about biomedicine.
In this survey, each BQA approach denotes a distinctive means to tackle the task.
We briefly define different BQA approaches in this section, and a high-level overview of them is shown in Figure \ref{fig:arch}.
We first define the \textbf{Classic BQA approach} from a historical perspective.
Mainly due to the lack of modular QA datasets (e.g. MRC BQA datasets), systems of this approach typically:
1. contain many sub-tasks and follow the pipeline of the question, document and answer processing, similar to IBM's Watson system \citep{ferrucci2012watson};
2. use many rule-based and/or sparse-feature-based machine learning modules in one system;
3. are evaluated on small-scale private datasets.
Since most of the classic BQA systems are surveyed in detail by \citet{athenikos2010biomedical}, we just briefly introduce them in \S\ref{classic}.
Besides the classic BQA approach, other BQA approaches tackle the task using specific supporting resources that are included in standard, public datasets.
They typically contains a collection of datasets and methods, and we define several BQA approaches below:
\begin{itemize}
\item \textbf{Information Retrieval (IR) approach}: where systems retrieve relevant documents to answer the questions;
\item \textbf{Machine Reading Comprehension (MRC) approach}: where systems read given contexts about the questions to predict the answers. The contexts of MRC approach can be provided by the IR approach;
\item \textbf{Knowledge Base (KB) approach}: where systems either explicitly translate the input questions to RDF queries to search the KBs or implicitly use the integrated knowledge from certain biomedical KBs to get the answers;
\item \textbf{Question Entailment (QE) approach}: where systems find similar questions that have been previously answered in a Q-A pair database and re-use their answers to answer the given question;
\end{itemize}
Characteristics of these approaches are summarized in Table \ref{tab:approach_type}.
\begin{table*}[ht!]
\small
\centering
\begin{tabular}{p{4cm}p{4.5cm}p{4.2cm}}
\toprule
\textbf{BQA Approach} & \textbf{Supporting resources} & \textbf{Answer form} \\
\midrule
IR BQA approach (\S\ref{irbqa}) & Document collections & Specific documents \\
\midrule
MRC BQA approach (\S\ref{mrcbqa}) & Specific documents (contexts) & Y/N; Extraction; Generation \\
\midrule
KB BQA approach (\S\ref{kb}) & Knowledge bases & Biomedical entities/relations \\
\midrule
QE BQA approach (\S\ref{qeapproach}) & Answered questions (FAQs) & Existing answers of similar questions \\
\bottomrule
\end{tabular}
\caption{Characteristics of different BQA approaches. Y/N: Yes/No.}
\label{tab:approach_type}
\end{table*}
\section{Classic BQA} \label{classic}
In this section, We breifly introduce several representative classic BQA systems, and point the readers to the BQA survey by \citet{athenikos2010biomedical} for more details.
\begin{table*}[ht!]
\small
\centering
\begin{tabular}{p{2.2cm}p{1cm}p{3cm}p{3.6cm}p{3.2cm}}
\toprule
\textbf{System} & \textbf{Content} & \textbf{Question Processing} & \textbf{Document Processing} & \textbf{Answer Processing} \\ \midrule
EPoCare \citep{niu2003answering} & Clinical & PICO format & Keyword-based retriever & -- \\ \midrule
\citet{takahashi2004question} & Scientific & Question type classification; Query formulation & MySQL retriever & -- \\ \midrule
\citet{demner2005knowledge, demner2006answer, demner2007answering} & Clinical & PICO-format; Query formulation& Knowledge extraction; Semantic matching& Semantic clustering \& summarization \\ \midrule
MedQA \citep{lee2006beyond, yu2007development}& Consumer & Question type classification & Lucene retriever & Answer extraction \& summarization \\ \midrule
BioSquash \citep{shi2007question}& Scientific & Semantic annotation & Semantic annotation; Graph construction& Sentence selection, clustering and post-processing \\ \midrule
\citet{terol2007knowledge} & Clinical & Question and answer type classification; Logic form extraction& -- & Answer generation based on logic forms \\ \midrule
\citet{weiming2007automatic} & Clinical & Concept and relation recognition& Lucene retriever & Semantic interpretation and clustering based on relations \\ \midrule
HONQA \citep{cruchet2008supervised}& Consumer & Question, expected answer and medical type classification& -- & -- \\ \midrule
\citet{lin2008biological} & Scientific & Question type classification; Query expansion& Google-interfacing retriever& NER- \& SRL-based \\ \midrule
EAGLi \citep{gobeill2009question}& Scientific & Query and target categorization& PubMed e-utils and EasyIR \citep{aronson2005fusion}& Extracting concepts \\ \midrule
AskHERMES \citep{cao2011askhermes}& Clinical & Topic classification with MetaMap \citep{yu2008automatically}& BM25 retriever; Longest common subsequence extractor & Content Clustering \\ \midrule
MiPACQ \citep{cairns2011mipacq}& Clinical & Semantic annotation& Lucene retriever & ML-based re-ranking \\
\bottomrule
\end{tabular}
\caption{An overview of classic BQA systems (listed in chronological order). \text{``--"}: no special processing steps.}
\label{tab:traditional}
\end{table*}
Traditionally, QA approaches consist of 3 main parts \citep{hirschman2001natural}:
1. Question processing, where systems (a) determine the type of the question and the corresponding type of the expected answer and (b) form queries that are fed to certain document retrieval systems;
2. Document processing, where systems (a) retrieve relevant documents from the queries generated in the previous step and (b) extract answer candidates from the relevant documents;
3. Answer processing, where systems rank the candidate answers based on certain criteria.
Although recently some of these modules have become independent QA approaches (e.g. IR, MRC), classic BQA still remains a distinctive class of approach in this survey because most of these works describe a whole system that includes all these subtasks.
We show an overview of classic BQA methods in Table \ref{tab:traditional} with their specific methods for question, document and answer processing.
\paragraph{\textbf{PICO-based:}}
\citet{niu2003answering} explore PICO-based BQA in the EPoCare project using simple keyword-based retrievers.
Demner-Fushman and Lin further study the PICO-based semantic BQA for the practice of Evidence-Based Medicine (EBM) in a series of works \citep{demner2005knowledge, demner2006answer, demner2007answering}, where the core step involves searching PubMed articles that are annotated with extracted medical knowledge.
\citet{huang2006evaluation} study the feasibility of using the PICO format to represent clinical questions and conclude that PICO is primarily focused on therapy type clinical questions and unsuitable for representing the others (e.g.: prognosis, etiology).
\paragraph{\textbf{Natural-language-based:}}
To tackle a broader range of topics, most other classic BQA systems accept natural language questions:
The medical definitional question answering system (MedQA, \citet{lee2006beyond, yu2007development}) is the first fully implemented BQA system that generates answers by extractive summarization for users' definitional questions from large text corpora.
BioSquash \citep{shi2007question} is adapted from the general domain summarizer Squash \citep{melli2005description} and is focused on QA-oriented summarization of biomedical documents.
\citet{terol2007knowledge} utilize logic forms for BQA, where the core step is to derive the logic forms of questions and utilize them to generate answers.
HONQA \citep{cruchet2008supervised} is a French/English bilingual BQA system that focuses on the supervised classification of question and answer types for question answering.
\citet{lin2008biological} explore answering biomolecular event questions with named entities using syntactic and semantic feature matching.
\citet{gobeill2009question} generate 200 questions from biomedical relational databases to evaluate their EAGLi platform.
\citet{cao2011askhermes} propose the askHERMES system, a BQA system that performs several semantic analyses, including question topic classification and content clustering, to provide extractive summaries for clinical questions.
The classic BQA approaches rely heavily on rule-based models and various ad-hoc modules in their complex pipelines.
Although these might be necessary in industry-level applications, they are hard to develop and maintain in academic settings.
In addition, most classic BQA systems have not been validated on large-scale public datasets.
With the introduction of various BQA datasets that are focused on specific BQA topics or steps, only a few BQA systems that tackle the complete question-to-answer BQA task have been proposed recently, which will be discussed as the modular evaluation issue in \S\ref{eval}.
\section{Information Retrieval BQA} \label{irbqa}
Information Retrieval (IR) BQA denotes the approach that uses \textit{IR BQA Methods} to retrieve relevant text snippets from certain \textit{Document Collections} for the given question, where the retrieved snippets can be either directly used as answers or further fed to MRC models (\S \ref{mrcbqa}).
We also discuss several \textit{IR BQA Datasets} that are summarized in Table \ref{tab:irbqa}.
\begin{table*}[ht!]
\small
\centering
\begin{tabular}{p{3.5cm}p{1.8cm}p{1cm}p{4cm}p{1.5cm}p{1.2cm}}
\toprule
\textbf{Dataset} & \textbf{Size} & \textbf{Metric} & \textbf{State-of-the-art (\%)} & \textbf{Content} & \textbf{Format}
\\ \midrule
BioASQ Task B Phase A \citep{tsatsaronis2015overview} & $3.7$k & MAP & 33.04 (document) / 68.21 (snippet)$^1$ & Scientific & Retrieval \\ \midrule
BiQA \citep{lamurias2020generating} & $7.4$k & MAP & -- & Consumer & Retrieval \\ \midrule
EPIC-QA$^2$ & 45 & MAP & -- & Sci. \& Con. & Retrieval \\ \midrule
HealthQA \citep{zhu2019hierarchical} & $7.5$k & MRR & 87.88 \citep{zhu2019hierarchical} & Consumer & Retrieval \\ \midrule
TREC Genomics \citep{hersh2006trec, hersh2007trec} & $28$ (06), $36$ (07) & MAP & 54.39 (06) / 32.86 (07) \citep{hersh2009trec} & Scientific & Retrieval \\
\bottomrule
\end{tabular}
\caption{An overview of the information retrieval biomedical question answering datasets (listed in alphabetical order). $^1$Batch 2 of BioASQ Task 8 b Phase A. $^2$\url{https://bionlp.nlm.nih.gov/epic_qa/}}
\label{tab:irbqa}
\end{table*}
\subsection{Document Collections}
PubMed\footnote{\url{https://pubmed.ncbi.nlm.nih.gov/}} and PubMed Central\footnote{\url{https://www.ncbi.nlm.nih.gov/pmc/}} are the most widely used corpora.
Both were developed and are maintained by the National Library of Medicine (NLM) of the US.
PubMed provides free access to more than $30$M citations for biomedical literature, where each citation mainly contains the paper title, author information, abstract and semantic indices like MeSH (introduced in \S\ref{existing_kb}).
PubMed Central (PMC) includes full-texts of over $6$M biomedical articles in addition to the information provided in the PubMed citations.
More specific corpora are typically used for sub-domain BQA to filter out potential noise in larger corpora, e.g.: CORD-19 \citep{wang2020cord} for EPIC-QA Task A and Alzheimer's Disease Literature Corpus for QA4MRE-Alzheimer \citep{morante2012machine}.
\subsection{IR BQA Datasets}
\textit{\textbf{BioASQ Task B Phase A:}} BioASQ Task B is named ``Biomedical Semantic Question Answering" \citep{tsatsaronis2015overview} and contains two phases correspond to the IR and MRC BQA approaches in our BQA classification:
in the phase A (IR phase), systems retrieve relevant documents for the given question;
in the phase B (MRC phase), systems use gold standard relevant documents and ontological data to answer the given questions (discussed in \S\ref{mrcbqa}).
Specifically, for the given question, BioASQ phase A participants shall return the relevant:
1. Concepts from certain ontologies such as MeSH;
2. Articles from PubMed and text snippets within the articles;
3. RDF triples from the Linked Life Data\footnote{\url{http://linkedlifedata.com/}}.
Mean average precision (MAP) is used as a main metric for the BioASQ retrieval phase, with slight modifications for snippet retrieval.
\citet{lamurias2020generating} introduces \textit{\textbf{BiQA}}, an IR BQA dataset containing $7.4$k questions and $14.2$k relevant PubMed articles collected from online QA forums (Stack Exchange\footnote{\url{https://stackexchange.com/}} and Reddit\footnote{\url{https://www.reddit.com/}}).
They show that adding BiQA in the training set can marginally boost BioASQ Task B phase A test performance.
The Epidemic Question Answering (\textit{\textbf{EPIC-QA}}) challenges\footnote{\url{https://bionlp.nlm.nih.gov/epic_qa/}} are also formatted as IR BQA, where the participants return a ranked list of sentences from expert and consumer corpora to answer questions about the COVID-19 pandemic.
\textit{\textbf{HealthQA}} has $7.5$k manually annotated questions that can be answered by one of the $7.3$k web articles \citep{zhu2019hierarchical}.
The \textit{\textbf{TREC Genomics Tracks}} in 2006 and 2007 tackle the BQA with IR approach \citep{hersh2006trec, hersh2007trec, hersh2009trec}:
28 and 36 topic questions (e.g.: ``What [GENES] are genetically linked to alcoholism?") are released in 2006 and 2007, respectively, and the participating systems are required to retrieve passages from 162k full-text articles from Highwire Press\footnote{\url{https://www.highwirepress.com/}}.
IR BQA systems typically return a ranked list of documents as answers, and mean average precision (MAP) is usually used as the evaluation metric:
\[\text{AP}=\frac{1}{m}\sum_{k=1}^nP@k\times\text{rel}(k),\qquad \text{MAP}=\frac{1}{N}\sum_{q=1}^{N}\text{AP}_q\]
where $m$ denotes the number of gold-standard relevant documents and $n$ denotes the number of the returned documents.
rel(k) is an indicator function that has the value 1 if the $k$-th document is relevant otherwise 0.
\subsection{IR BQA Methods}
\textit{\textbf{BioASQ Task B Phase A:}}
Wishart \citep{liu2013university} re-ranks and combines sentences from the retrieved documents to form the ideal answers for BioASQ task B phase B, and generate exact answers from the ideal answers according to the question type.
The USTB team \citep{jin2017multi} wins all batches in document, snippet and concept retrieval in BioASQ 5.
They use sequential dependence model \citep{bonnefoy2012social}, pseudo relevance feedback, fielded sequential dependence model \citep{zhiltsov2015fielded} and divergence from randomness model \citep{clinchant2009bridging}.
The AUEB team proposes a series of models \citep{brokos2018aueb, pappas2019aueb, pappas2020aueb} that win most of the Task B Phase A challenges since BioASQ 6.
At BioASQ 6, they \citep{brokos2018aueb} use the Position-Aware Convolutional Recurrent Relevance model \citep{hui2017pacrr} and the Deep Relevance Matching Model \citep{guo2016deep} for document retrieval, and use the Basic Bi-CNN model \citep{yin2016abcnn} for snippet retrieval.
They win 3/5 and 5/5 batches for retrieving documents and snippets in BioASQ 6, respectively.
At BioASQ 7, they \citep{pappas2019aueb} combine the document and snippet retrieval system by modifying their BioASQ 6 system to also output the sentence-level (i.e.: snippet) relevance score in each document.
They win 4/5 and 4/5 batches for retrieving documents and snippets in BioASQ 7, respectively.
In BioASQ 8, They \citep{pappas2020aueb} continue to use this system and win 2/5 for document and 4/5 batches for snippet retrieval.
\textit{\textbf{HealthQA:}}
\citet{zhu2019hierarchical} propose Hierarchical Attention Retrieval, a ranking model for biomedical QA that uses a deep attention mechanism at word, sentence and document levels to compute a relevance score of the given query with each candidate document.
With the proposed model, they achieve an MRR of 0.8788 on the HealthQA dataset.
\citet{zhou2007trec} win the \textit{\textbf{TREC 2006 Genomics Track}}, where they first identify query concepts and retrieve relevant documents with concept-level and word-level similarity measurements, and then extract the answers.
At \textit{\textbf{TREC 2007 Genomics Track}}, NLMinter \citep{demnerfushman2007combining} achieves the best performance.
NLMinter is an interactive retriever where the fusion retrieval results are boosted by the document relevance feedback, which is determined by expert PubMed search and occasional examination of the abstracts.
\subsection{Comments}
Though Classic BQA approaches usually contain a retrieval step,
IR BQA is still considered as a distinct approach because the retrieved documents are directly used as answers,
and are evaluated by IR metrics.
Traditional retrieval methods like TF-IDF have been well-studied and ubiquitously used in IR BQA approach.
Future studies can focus more on PLM-based (re-)ranking methods \citep{chakraborty2020biomedbert, lin2020pretrained}, and how to better bridge the IR and MRC models.
\section{Machine Reading Comprehension BQA} \label{mrcbqa}
Machine Reading Comprehension (MRC) is a well-studied BQA task, where the models answer questions about given textual contexts.
\textit{MRC BQA Datasets} are typically specialized in content and have predetermined answer format, so most \textit{MRC BQA Methods} developed on them are end-to-end neural models.
\subsection{MRC BQA Datasets}
Many MRC BQA datasets have been proposed, and we show an overview of them in Table \ref{tab:mrcbqa}.
\textit{\textbf{BioASQ Task B Phase B:}} it provide the largest and most widely used manually-annotated MRC BQA dataset:
Starting from 2013, BioASQ annotates about 500 test QA instances each year,
which will be included in the training set of the following years.
Currently, BioASQ 2020 consists of 3,243 training QA instances and at least 500 test instances.
Questions in BioASQ are typically scientific questions (\S\ref{scientific}).
There are 4 types of QA instances in BioASQ: factoid, list, yes/no and summary.
Factoid, list and yes/no instances have both \textit{exact} and \textit{ideal} answers:
\textit{Exact} answers are short answers that directly answer the questions, e.g.: single and multiple biomedical entities for factoid and list questions, respectively; ``yes" or ``no" for yes/no questions.
\textit{Ideal} answers are \textit{exact} answers written in complete sentences, e.g.: ``Yes, because [...]".
The main evaluation metrics for yes/no, factoid, list and summary questions are accuracy, MRR, mean F-score and manual score, respectively.
We show several examples of BioASQ instances in Table \ref{tab:bioasq_eg}.
\begin{table*}[ht!]
\small
\centering
\setlength{\tabcolsep}{0.9mm}{
\begin{tabular}{p{1.4cm}p{3.2cm}p{4cm}p{2cm}p{3.5cm}}
\toprule
\textbf{Type} & \textbf{Example Question} & \textbf{Example Context} & \textbf{Exact answer} & \textbf{Ideal answer} \\
\midrule
Yes / No & Is the protein Papilin secreted? & [...] and two genes encoding secreted extracellular matrix proteins, mig-6/papilin [...]. & Yes & Yes, papilin is a secreted protein\\
\midrule
Factoid & Name synonym of Acrokeratosis paraneoplastica. & Acrokeratosis paraneoplastic (Bazex syndrome) is a rare, but [...] & Bazex syndrome & Acrokeratosis paraneoplastic (Bazex syndrome) is a rare [...]\\
\midrule
List & List Hemolytic Uremic Syndrome Triad. & Atypical hemolytic uremic syndrome (aHUS) is a rare disease characterized by the triad of [...]
& anaemia, renal failure, thrombocytopenia
& Hemolytic uremic syndrome (HUS) is a clinical syndrome characterized by [...] \\
\midrule
Summary & What is the effect of TRH on myocardial contractility? & Thyrotropin-releasing hormone (TRH) improved [...] & NA & TRH improves myocardial contractility\\
\bottomrule
\end{tabular}}
\caption{Types of questions in BioASQ and respective examples}
\label{tab:bioasq_eg}
\end{table*}
\begin{table*}[ht!]
\small
\centering
\begin{tabular}{p{2.8cm}p{1.4cm}p{2cm}p{3.8cm}p{1cm}p{1.8cm}}
\toprule
\textbf{Dataset} & \textbf{Size} & \textbf{Metric} & \textbf{State-of-the-art (\%)} & \textbf{Content} & \textbf{Format}
\\ \midrule
BioASQ Task B Phase B \citep{tsatsaronis2015overview} & $3.7$k & F1, MRR, List F1, Manual & 90.3, 39.7, 52.3, 4.39/5 \citep{Nentidis2020overview} & Scientific & Y/N; Extraction; Generation \\ \midrule
Biomed-Cloze \citep{dhingra2018simple} & $1$M & -- & -- & Scientific & Extraction \\ \midrule
BioMRC \citep{pappas2020biomrc} & $812$k & Acc & 80.06 (dev) / 79.97 (test) \footnotemark[1] \citep{pappas2020biomrc} & Scientific & Extraction \\ \midrule
BioRead \citep{pappas2018bioread} & $16.4$M & Acc & 47.06 (dev) / 51.52 (test) \citep{pappas2018bioread} & Scientific & Extraction \\ \midrule
BMKC \citep{kim2018pilot} & 473k (T); 370k (LS) & Acc & T: 85.5 (val) / 83.6 (test); LS: 80.1 (val) / 77.3 (test) \citep{kim2018pilot} & Scientific & Extraction \\ \midrule
CliCR \citep{suster2018clicr} & $100$k & EM, F1 & 55.2, 59.8 \citep{pham2020machine} & Clinical & Extraction \\ \midrule
COVIDQA \citep{tang2020rapidly} & $124$ & P@1, R@3, MRR & 30.6 \citep{chakravarti2020towards}, 47.7 \citep{su2020caire}, 41.5 \citep{tang2020rapidly} & Scientific & Extraction \\ \midrule
COVID-QA \citep{moller2020covid} & $2$k & EM, F1 & 37.2, 64.7 \citep{reddy2020end} & Scientific & Extraction \\ \midrule
EBMSummariser \citep{molla2016corpus} & 456 & ROUGE-1,2,SU4 & 39.85, 24.50, 22.59 \citep{shafieibavani2016appraising} & Clinical & Generation \\ \midrule
emrQA \citep{pampari2018emrqa} & $455$k & EM, F1 & 76.1, 81.7 \citep{rongali2020improved} & Clinical & Extraction \\ \midrule
MASH-QA \citep{zhu2020question} & $34.8$k & EM, F1 & 29.49, 64.94 \citep{zhu2020question} & Consumer & Extraction \\ \midrule
MEDHOP \citep{welbl2018constructing} & $2.5$k & Acc & 63.2 \citep{huo2020sentence} & Scientific & Multi-choice \\ \midrule
MEDIQA-AnS \citep{savery2020question} & 552 (single); 156 (multi) & ROUGE-1,2,L; BLEU & Extractive\footnotemark[2]: 29, 15, 12, 9; Abstractive: 32, 12, 8, 9 \citep{savery2020question} & Consumer & Generation \\ \midrule
MEDIQA-QA \citep{abacha2019overview} & $3$k & Acc, P, MRR & 79.49 \citep{he2020infusing}, 84.02 \citep{he2020infusing}, 96.22 \citep{pugaliya2019pentagon} & Consumer & Multi-choice \\ \midrule
ProcessBank \citep{berant2014modeling} & 585 & Acc & 66.7 \citep{berant2014modeling} & Scientific & Multi-choice \\ \midrule
PubMedQA \citep{jin2019pubmedqa} & $212$k & Acc, F1 & 68.08, 52.72 \citep{jin2019pubmedqa} & Scientific & Y/N \\ \midrule
QA4MRE-Alz \citep{morante2012machine} & $40$ & $c@1$ & 76 \citep{bhaskar2012question} & Scientific & Multi-choice \\ \midrule
\end{tabular}
\caption{An overview of the machine reading comprehension biomedical question answering datasets (listed in alphabetical order).
$^1$Results on BioMRC lite.
$^2$Results on single document summarization.
}
\label{tab:mrcbqa}
\end{table*}
Question Answering for Machine Reading Evaluation (\textit{\textbf{QA4MRE}}) holds a sub-task on machine reading of biomedical texts about Alzheimer's disease \citep{morante2012machine}.
This task provides only a test dataset with 40 QA instances, and each instance contains one question, one context and 5 answer choices.
\textit{\textbf{Cloze-style}} questions require the systems to predict the missing spans in contexts (e.g.: Q: ``Protein X suppresses immune systems by inducing \underline{\hbox to 8mm{}} of immune cells."; A: ``apoptosis").
There are many large-scale cloze-style MRC BQA datasets that are automatically constructed, such as CliCR, Biomed-Cloze, BioRead, BMKC and BioMRC.
\textit{\textbf{COVIDQA}} \citep{tang2020rapidly} is a QA dataset specifically designed for COVID-19.
It has 124 question-article pairs translated from the literature review page of Kaggle’s COVID-19 Open Research Dataset Challenge \footnote{\url{https://www.kaggle.com/allen-institute-for-ai/CORD-19-research-challenge/tasks}}, where relevant information for each category or subcategory in the review is presented.
\textit{\textbf{COIVD-QA}} \citep{moller2020covid} is another COVID-19 QA dataset with $2$k question-answer pair annotated by biomedical experts.
The annotation is similar to that of SQuAD while the answers in COVID-QA tend to be longer as they generally come from longer texts.
\citet{molla2011development, molla2016corpus} build the \textit{\textbf{EBMSummariser}}, a summarization dataset of 456 instances for EBM, from the Clinical Inquiries section of the Journal of Family Practice\footnote{\url{https://www.mdedge.com/familymedicine/clinical-inquiries}}: each instance contains a clinical question, a long ``bottom-line" answer, the answer's evidence quality and a short justification of the long answer.
\textit{\textbf{MASH-QA}} \citep{zhu2020question}, a dataset based on consumer health domain, is designed for extracting information from texts that span across a long document.
It utilizes long and comprehensive healthcare articles as context to answer generally non-factoid questions.
Different from the existing MRC datasets with short single-span answers, many answers in MASH-QA are several sentences long and excerpted from multiple parts or spans of the long context.
\textit{\textbf{MEDHOP}} \citep{welbl2018constructing} is a multi-hop MRC BQA dataset, where each instance contains a query of a subject and a predicate (e.g.: ``Leuprolide, interacts\_with, ?"), multiple relevant and linking documents and a set of answer options extracted from the documents.
Reasoning over multiple documents is required for the model to answer the question.
\textit{\textbf{MEDIQA-QA}} \citep{abacha2019overview} is the dataset of the QA subtask of MEDIQA 2019 shared task that has $400$ questions and $3$k associate answers. It is obtained by submitting medical questions to the consumer health QA system CHiQA then re-rank the answers by medical experts. The task of MEDIQA-QA dataset is to filter and improve the ranking of answers, making it a multi-choice QA task. \textit{\textbf{MEDIQA-AnS}} \citep{savery2020question}, on the other hand, is a summarization dataset. It provides extractive and abstractive versions of single and multi-document summary of the answer passages from MEDIQA-QA.
\textit{\textbf{ProcessBank}} \citep{berant2014modeling} contains multi-choice questions along with relevant biological paragraphs. The paragraphs are annotated with "process", a directed graph $(\mathscr{T},\mathscr{A},\mathscr{E}_{tt},\mathscr{E}_{ta})$, where nodes $\mathscr{T}$ are token spans denoting the occurrence of events, nodes $\mathscr{A}$ are token spans denoting entities in the process, and the latter two are edges describing event relations and semantic roles respectively.
\citet{jin2019pubmedqa} build the \textit{\textbf{PubMedQA}} dataset from PubMed articles that use binary questions as titles (e.g.: ``Do preoperative statins reduce atrial fibrillation after coronary artery bypass grafting?") and have structured abstracts.
The conclusive parts of the abstracts are the long answers, while the main task of PubMedQA is to predict their short forms, i.e.: yes/no/maybe, using the abstracts without the conclusive parts as contexts.
\subsection{MRC BQA Methods}
In this section, we first introduce the top-performing systems in each year of the BioASQ challenge to reflect the landscape changes of MRC BQA methods.
We then briefly describe SoTA models of other surveyed MRC BQA datasets.
\textit{\textbf{BioASQ:}}
The first two BioASQ challenges \citep{partalas2013results, balikas2014results} use a Watson-motivated baseline \citep{weissenborn2013answering} that ensembles multiple scoring functions to rank the relevant concepts with type coercion to answer the given questions.
The Fudan system \citep{zhang2015fudan} of BioASQ 3B contains three major components:
1. A question analysis module that mainly extracts semantic answer types of questions;
2. Candidates generating by PubTator \citep{wei2013pubtator} and Stanford POS tools\footnote{\url{https://nlp.stanford.edu/software/tagger.shtml}};
3. Candidates ranking based on word frequency.
The SNU team \citep{choi2015snumedinfo} directly combines the retrieved relevant passages to generate the ideal answer and achieve state-of-the-art performance.
At BioASQ 4B:
the HPI team \citep{schulze2016hpi} proposes an algorithm based on LexRank \citep{erkan2004lexrank} to generate ideal answers, which only uses biomedical named entities in the similarity function.
They win 1/5 batch in the ideal answer generation.
At BioASQ 5B:
the UNCC team \citep{bhandwaldar2018uncc} uses lexical chaining-based extractive summarization to achieve the highest ROUGE scores for ideal answer generation, with 0.7197 ROUGE-2 and 0.7141 ROUGE-SU4.
The CMU OAQA team describes a series of works for BioASQ \citep{yang2015learning, yang2016learning, chandu2017tackling}.
At BioASQ 3B, they propose a three-layered architecture \citep{yang2015learning} where:
the first layer contains domain-independent QA components such as input/output definition, intermediate data objects;
the second layer has implementations of biomedical domain-specific materials like UMLS and MetaMap \citep{aronson2001effective};
the third design layer is BioASQ-specific, including the end-to-end training and testing pipeline for the task.
The core components of the pipeline are the answer type prediction module and the candidate answer scoring module based on supervised learning.
At BioASQ 4B, they extend their BioASQ 3B system with general-purpose NLP annotators, machine-learning-based search result scoring, collective answer re-ranking and yes/no answer prediction \citep{yang2016learning}.
The CMU OAQA team is focused on ideal answer generation \citep{chandu2017tackling} at BioASQ 5B, using extractive summarization tools like Maximal Marginal Relevance \citep{carbonell1998use} and Sentence Compression \citep{filippova2015sentence} with biomedical ontologies such as UMLS and SNOMED-CT.
BioAMA \citep{sharma2018bioama} further improves the ROUGE score by 7\% for ideal answer generation than \citet{chandu2017tackling} by combining effective IR-based techniques and diversification of relevant snippets.
The Macquarie University team has participated in BioASQ 5-8 with the focus of ideal answer generation by extractive summarization for yes/no, factoid, list and summary questions \citep{molla2017macquarie, molla2018macquarie, molla2019classification, molla2020query}.
At BioASQ 5B, \citet{molla2017macquarie} observes that a trivial baseline that returns the top retrieved snippets as the ideal answer is hard to beat.
At BioASQ 6B, \citet{molla2018macquarie} show that using LSTM-based deep learning methods that predict the F1 ROUGE-SU4 score of an individual sentence and the ideal answer achieves the best results.
At BioASQ 7B, \citet{molla2019classification} observe that sentence-level classification task works better than regression task for finding the extractive summary sentences.
In recent years of BioASQ, transfer learning has gained increasing attention, where models are first pre-trained on large-scale general domain QA datasets or BQA datasets and then fine-tuned on the BioASQ training set.
\citet{wiese2017neural} achieve state-of-the-art performance on factoid questions and competitive performance on list questions by transferring the FastQA model pre-trained by SQuAD to BioASQ.
\citet{dhingra2018simple} show significant performance improvement over purely supervised learning by pre-training the GA-Reader \citep{dhingra2017gated} on an automatically generated large-scale cloze BQA dataset (\S\ref{automatic}) and then fine-tuning it on BioASQ.
\citet{du2018hierarchical, du2019hierarchical} have similar observations with transfer learning from the SQuAD dataset.
\citet{kang2020transferability} show transfer learning from NLI datasets also benefits BioASQ performance on yes/no (+5.59\%), factoid (+0.53\%) and list (+13.58\%) questions.
Generally, two main components are ubiquitously used in top-performing systems of the current BioASQ 8 challenge \citep{Nentidis2020overview}:
1. domain-specific pre-trained language models \citep{yoon2019pre}, such as BioBERT;
2. task-specific QA datasets that can (further) pre-train the used models, such as SQuAD for extractive QA and PubMedQA for yes/no QA.
In summary, with the introduction of large-scale MRC datasets like SQuAD \citep{rajpurkar2016squad, rajpurkar2018know}, a variety of neural MRC models have been proposed that incrementally improve the task performance, such as DCN \citep{xiong2016dynamic}, Bi-DAF \citep{seo2016bidirectional}, FastQA \citep{weissenborn2017making}.
Contextualized word embeddings pre-trained by language models (LM) like ELMo \citep{peters2018deep} and BERT \citep{devlin2019bert} show significant improvements on various NLP tasks including MRC.
Pre-trained LMs on biomedical corpora, such as BioELMo \citep{jin2019probing}, BioBERT \citep{lee2020biobert}, SciBERT \citep{beltagy2019scibert}, clinical BERT \citep{alsentzer2019publicly, huang2019clinicalbert} and PubMedBERT \citep{gu2020domain}, further improve their in-domain performance.
Probing experiments and analyses by \citet{jin2019probing} indicate that better encoding of biomedical entity-type and relational information leads to the superiority of domain-specific pre-trained embeddings.
Various methods have also been developed for other MRC BQA datasets.
Here we briefly discuss their representative SoTA methods as shown in Table \ref{tab:mrcbqa}.
\textit{\textbf{BioRead:}} \citet{pappas2018bioread} train the AOA Reader \citep{cui2017attention} on BioReadLite, which computes the mutual information between query and context, and places another attention layer over the document-level attention to achieve attended attention for the final prediction. They achieve the best accuracy of 0.5152.
\textit{\textbf{BioMRC:}} It is the updated version of BioRead. \citet{pappas2020biomrc} use SciBERT \citep{beltagy2019scibert} and maximize scores of all mentions of each entity in the passage, achieving SoTA accuracy of 0.7997.
\textit{\textbf{BMKC:}} Based on Attention Sum Reader architecture \citep{kadlec2016text}, \citet{kim2018pilot} present a new model that combines pre-trained knowledge and information of entity types. They also develop an ensemble method to integrate results from multiple independent models, which gets the accuracy of 0.836 on BMKC$\_$T and 0.773 on BMKC$\_$LS.
\textit{\textbf{CliCR:}} \citet{pham2020machine} show that language models have better performance with systematic modification on cloze-type datasets.
They replace \textit{@placeholder} with [MASK] and trains BioBERT \citep{lee2020biobert} on the modified dataset to obtain the SoTA EM of 0.552 and F1-score of 0.598.
\textit{\textbf{COVIDQA:}} \citet{chakravarti2020towards} fine-tune pre-trained language models on the Natural Questions dataset \citep{kwiatkowski2019natural} with attention-over-attention strategy and attention density layer. They try its zero-shot transfer and achieve $P@1$ of 0.306. \citet{su2020caire} combine HLTC-MRQA \citep{su2019generalizing} with BioBERT to rank context sentences to get the evidence, and obtain $R@3$ of 0.477. \citet{tang2020rapidly} achieve MRR of 0.415 by fine-tuning T5 \citep{raffel2020exploring} on MS MARCO \citep{nogueira2020document}.
\textit{\textbf{COVID-QA:}} \citet{reddy2020end} propose an example generation model for the training of MRC, and fine-tune RoBERTa-large \citep{liu2019roberta} on SQuAD2.0 \citep{rajpurkar2018know}, NQ and their generated training examples, which achieves EM of 0.372 and F1-score of 0.647.
\textit{\textbf{EBMSummariser:}} \citet{sarker2013approach} extract three sentences using hand-crafted features such as sentence length, position and question semantics for the EBMSummariser dataset, achieving ROUGE-L F-score of 0.168. \citet{shafieibavani2016appraising} utilize both UMLS and WordNet to summarise medical evidence for queries, and achieve ROUGE-1 of 0.3985, ROUGE-2 of 0.2450 and ROUGE-SU4 of 0.2259 on EBMSummariser.
\textit{\textbf{emrQA:}} \citet{rongali2020improved} use rehearsal and elastic weight consolidation to improve domain-specific training, which can benefit the performance of models in both general domain and domain-specific tasks. They achieve EM of 0.761 and F1-score of 0.817.
\textit{\textbf{MASH-QA:}} \citet{zhu2020question} propose MultiCo to select sentences across the long contexts to form answers. MultiCo combines a query-based sentence selection approach with a inter sentence attention mechanism, and achieves EM of 0.2949 and F1-score of 0.6494 on single-span MASH-QA dataset.
\textit{\textbf{MEDHOP:}} \citet{huo2020sentence} propose a Sentence-based Circular Reasoning approach which establishes a information path with sentence representation. They also implement a nested mechanism to systematically represent semantics, which improves the model performance significantly and achieves an accuracy of 0.632.
\textit{\textbf{MEDIQA-AnS:}} \citet{savery2020question} train BART \citep{lewis2020bart} on the BioASQ data to achieve SOTA results.
\textit{\textbf{MEDIQA-QA:}} \citet{he2020infusing} infuse disease knowledge into pre-trained language models like BERT and achieve accuracy of 0.7949 and precision of 0.8402. \citet{pugaliya2019pentagon} train their end-to-end system in a multi-task setting, and use the pretrained RQE and NLU modules to extract the best entailed questions and best candidate answers. They achieve MRR of 0.9622.
\textit{\textbf{ProcessBank:}} \citet{berant2014modeling} first predict a structure representing the process in the given paragraph, then they map each question into queries and compare them with the predicted structure. They achieve the accuracy of 0.667.
\textit{\textbf{PubMedQA:}} \citet{jin2019pubmedqa} take the multi-phase fine-tuning schedule with long answer as additional supervision, and achieve accuracy of 0.6808 and F1-score of 0.5272.
\subsection{Comments}
BioASQ is still the well-recognized benchmark and the ``go-to" dataset for MRC BQA because of its careful design, expert annotations, large size and highly active community.
Future models could explore developing pre-training methods that utilize richer biomedical knowledge than the raw texts (\S\ref{utilization}).
Additionally, collecting harder datasets / datasets that require other types of reasoning still remains an interesting future direction (\S\ref{difficulty}).
\section{Knowledge Base BQA} \label{kb}
KBQA (Knowledge Base QA) refers to answering questions using entities or relation information from knowledge bases \citep{fu2020survey}.
In biomedical domain, various large-scale biomedical KBs have been introduced, and one of their objectives is to assist with BQA.
Typically, one can convert natural language questions to SPARQL queries\footnote{\url{https://www.w3.org/TR/rdf-sparql-query/}} and use them to search the KBs for the answers.
In this section, we first introduce the \textit{existing knowledge bases} that have been used for KB BQA, and then introduce the \textit{KB BQA datasets} and \textit{KB BQA methods} developed on them.
\subsection{Existing Knowledge Bases} \label{existing_kb}
We define biomedical KBs as databases that describe biomedical entities and their relations, which can usually be stored by subject-predicate-object triples.
Biomedical KBs can be used for enhancing text representations \citep{jin2019probing,yuan2021coder,yuan-etal-2021-improving} and improving performances for BQA \citep{li2020towards} (not only KB BQA).
Substantial efforts have been made towards building biomedical KBs, including ontologies such as Medical Subject Headings (MeSH)\footnote{\url{https://www.ncbi.nlm.nih.gov/mesh/}} for biomedical text topics,
International Classification of Diseases (ICD)\footnote{\url{https://www.who.int/classifications/icd/}} for diseases and Systematized Nomenclature of Medicine Clinical Terms (SNOMED-CT, \citet{stearns2001snomed}) for medical terms.
The Unified Medical Language System (UMLS)\footnote{\url{https://www.nlm.nih.gov/research/umls}} is a metathesaurus that integrates nearly 200 different biomedical KBs like MeSH and ICD.
Biomedical KB is a big topic and we refer the interested readers to following references \citep{kamdar2020empirical, nicholson2020constructing}.
\subsection{KB BQA Datasets}
KB BQA datasets provide a list of biomedical questions and several biomedical KBs.
One should generate a SPARQL query for each question, and the answers are evaluated by query results.
We summary existing KB BQA datasets and show an overview of them in Table \ref{tab:kbbqa}.
\begin{table*}[ht!]
\small
\centering
\begin{tabular}{p{3.5cm}p{1cm}p{1cm}p{2.5cm}p{1.5cm}p{1.5cm}}
\toprule
\textbf{Dataset} & \textbf{Size} & \textbf{Metric} & \textbf{State-of-the-art (\%)} & \textbf{Content} & \textbf{Format} \\
\midrule
Bioinformatics \citep{sima2021bio} & 30 & F1 & 60.0 \citep{sima2021bio} & Scientific & Generation \\ \midrule
QALD-4 task 2 \citep{unger2014question} & 50 & F1 & 99.0 \citep{marginean2017question} & Consumer & Generation \\
\bottomrule
\end{tabular}
\caption{An overview of KB BQA datasets (listed in alphabetical order).}
\label{tab:kbbqa}
\end{table*}
\textit{\textbf{QALD-4 task 2:}} \citet{unger2014question} provides 50 natural language biomedical question and request SPARQL queries to retrieve answers from SIDER\footnote{\url{http://sideeffects.embl.de/}}, Drugbank\footnote{\url{https://www.drugbank.com/}} and Diseasome\footnote{\url{http://wifo5-03.informatik.uni-mannheim.de/diseasome/}}, where most questions require integrating knowledge from multiple databases to answer.
An example natural language question is ``Which genes are associated with breast cancer?'', and a possible query can be:
\begin{lstlisting}[captionpos=b, basicstyle=\ttfamily,frame=tb,xleftmargin=5em,xrightmargin=5em]
SELECT DISTINCT ?x
WHERE {
diseases:1669 diseasome:associatedGene ?x .
}
\end{lstlisting}
\textit{\textbf{Bioinformatics}} contains 30 biomedical queries with different complexity, and the database searching are restricted in Bgee\footnote{\url{https://bgee.org/sparql}} and OMA\footnote{\url{https://sparql.omabrowser.org/sparql}}.
The natural language questions include multiple concepts which leads to longer and more complicated SPARQL queries.
\subsection{KB BQA Methods}
In this section, we introduce the well-performing KB BQA methods applies on \textit{\textbf{QALD-4 task 2}} and \textit{\textbf{Bioinformatics}}.
\textit{\textbf{QALD-4 task 2:}}
\citet{marginean2017question} wins QALD-4 task 2 by introducing the GFMed which is built with Grammatical Framework \citep{ranta2009gf} and Description Logic constructors, achieving 0.99 F1 on the test set.
GFMed performs extraordinary well in QALD-4 task 2 since it is highly customized to this dataset.
CANaLI \citep{mazzeo2016question} designs an semantic automaton to parse the questions in specified form and achieves F1 of 0.92 on QALD-4 task 2. Questions not in specified form are ignored by CANaLI.
\citet{zhang2016joint} exploit KBs to find out candidate relation, type, entity + relation triple patterns in the questions.
They select and align triple patterns using integer linear programming and achieves F1 of 0.88 on QALD-4 task 2.
\citet{hamon2017querying} establish a complex pipeline to translate questions using existing NLP tools and semantic resources, and it achieves F1 of 0.85 on QALD-4 task 2.
\textit{\textbf{Bioinformatics:}}
\citet{sima2021bio} propose Bio-SODA which converts natural language questions into SPARQL queries without training data.
Bio-SODA generates a list of query graphs based on matched entities in the question, and rank the query graphs considering semantic and syntactic similarity, and node centrality.
Bio-SODA achieves F1 of 0.60 on QALD-4 task 2 and 0.60 on Bioinformatics.
Classic BQA systems also require natural language question translation systems to query KB.
\citet{rinaldi2004answering} adapt the general domain ExtrAns system \citep{molla2000extrans} to genomics domain.
They first convert the documents to Minimal Logical Forms (MLFs) and use them to construct a KB during the offline phase.
In the online QA phase, the system also converts the given question to MLFs by the same mechanism, and then get the answer by searching the built MLFs KB.
\citet{abacha2012medical, abacha2015means} propose MEANS for medical BQA which converts questions to SPARQL queries with a pipeline of classifying question types, finding Expected Answers Types, question simplification, medical entity recognition, extracting semantic relations and constructing SPARQL based on entities and semantic relations.
\citet{kim2013natural} introduce Linked Open Data Question Answering system to generate SPARQL queries for SNOMED-CT by predicate-argument relations from sentences.
\subsection{Comments}
Existing KB BQA datasets are limited by size, making it hard to train learning-based methods.
As a result, most top-performing KB BQA methods apply complex pipeline methods including entity extraction, relation extraction, entity alignment and entity typing to construct queries.
To leverage the potential of end-to-end deep learning methods, more labeled datasets are required for training a supervised Seq2seq question translation model.
\section{Question Entailment BQA} \label{qeapproach}
\citet{harabagiu2006methods} show that recognizing textual entailment can be used to enhance open QA systems.
The QE approach for BQA is essentially a nearest neighbor method that uses the answers of similar and already answered questions (e.g.: Frequently Asked Questions, FAQs) to answer the given question.
We will discuss three main components of this approach in this section:
1. Models that can recognize similar questions, i.e.: \textit{QE BQA Methods};
2. datasets of similar \textit{Question-Question Pairs} for training QE models;
3. datasets of answered questions, i.e. \textit{Question-Answer Pairs}, that can be used to answer new questions.
\subsection{QE BQA Methods}
QE is formally defined by \citet{abacha2016recognizing} as: a question $\text{Q}_\text{A}$ entails a question $\text{Q}_\text{B}$ if every answer to $\text{Q}_\text{B}$ is also a correct answer to $\text{Q}_\text{A}$.
Natural language inference (NLI) is a relevant NLP task that predicts whether the relation of entailment, contradiction, or neutral holds between a pair of sentences.
In the general domain, predicting question-question similarity is an active research area with potential applications in question recommendation and community question answering \citep{nakov2016semeval, nakov2017semeval}.
\citet{luo2015simq} propose the SimQ system to retrieve similar consumer health questions on the Web using UMLS-annotated semantic and AQUA-parsed \citep{campbell2002transformational} syntactic features of the questions.
CMU OAQA \citep{wang2017cmu} use a dual entailment approach with bidirectional recurrent neural networks (RNN) and attention mechanism to predict question similarity;
\citet{abacha2019question} use feature-based logistic regression classifier and deep learning models that pass the concatenation of two question embeddings to multiple ReLU layers \citep{nair2010rectified} for recognizing QE;
\citet{Zhu2019PANLPAM} fine-tune pre-trained language models to classify question pairs and conduct transfer learning from NLI to boost the performance.
\subsection{Question-Question Pairs}
Training QE models needs datasets of question pairs annotated with entailment (similarity) labels.
Towards this end, \citet{abacha2016recognizing} introduce the clinical-QE dataset\footnote{\url{https://github.com/abachaa/RQE_Data_AMIA2016}} that contains over $8$k training biomedical question pairs with similarity labels.
The questions are from clinical questions collected by \citet{ely2000taxonomy}, Consumer Health Questions and FAQs from NLM and NIH websites, respectively.
This dataset is also used as the RQE dataset in the MEDIQA challenge with slight changes.
\citet{sun2020analysis, poliak2020collecting, zhang2020cough} build question-question relevance datasets along with their FAQ datasets on COVID-19.
In general, only limited efforts have been made to build biomedical QE datasets, which results in the lack of training instances.
Many works instead consider a transfer learning approach to pre-train the QE models on other text pair tasks, including biomedical natural language inference (NLI) datasets like MedNLI \citep{romanov2018lessons}, general domain QE datasets like SemEval-cQA \citep{nakov2016semeval, nakov2017semeval}, and general domain NLI datasets like SNLI \citep{bowman2015large} and MultiNLI \citep{williams2018broad}.
\subsection{Question-Answer Pairs}
QE approach relies heavily on large databases of question-answer pairs with high quality to answer unseen questions.
For this, \citet{abacha2019question} build MedQuAD, a collection of 47,457 question-answer pairs from trusted websites e.g.: \url{https://www.cancer.gov/}.
Using MedQuAD for BQA can protect users from misleading and harmful health information since most answers are well-curated.
Moreover, several FAQ datasets have been introduced for answering COVID-19 related questions \citep{sun2020analysis, poliak2020collecting, zhang2020cough}.
However, since expert-curated answers are expensive to collect, such question-answer pair datasets might be limited in size.
Online health and QA communities like WebMD\footnote{\url{https://www.webmd.com/}}, Quora\footnote{\url{https://www.quora.com/}} provide large amounts of QA pairs, and many large-scale BQA datasets have been built using online doctor-patient QA data \citep{zhang2017chinese, zhang2018multi, he2019applying, tian2019chimed}.
These materials can be potentially used in QE approach, but the quality of user-provided answers should be carefully checked.
\subsection{Comments}
The most important components for the QE BQA approach are the datasets of question-question (Q-Q) and question-answer (Q-A) pairs.
However, these datasets are currently limited in scale or quality.
To tackle this issue, methods for automatically collecting large-scale Q-Q and Q-A datasets with high quality should be explored (\S\ref{automatic}).
\section{Challenges and Future Directions} \label{challenge}
In this section, we discuss the challenges identified in \S \ref{intro} with the surveyed works.
We also propose some interesting and potential future directions to explore.
\S\ref{automatic} and \S\ref{difficulty} involve dataset scale and difficulty, respectively.
In \S\ref{vqa}, we discuss visual BQA, which is an active and novel research field that is gaining more and more attention.
We explain why domain knowledge is not fully utilized in BQA and how the fusion of different BQA approaches can potentially solve it in \S\ref{utilization}.
In \S\ref{explainability}, we study different forms of explainability of BQA systems.
In \S\ref{eval}, we discuss two main issues of BQA system evaluation: qualitatively, what parts of the systems are evaluated and quantitatively, how they are evaluated.
Last but not the least, we discuss the fairness and bias issues of BQA in \S\ref{fairness}.
\subsection{Dataset Collection} \label{automatic}
Annotating large-scale BQA datasets is prohibitively expensive since it requires intensive expert involvement.
As a result, the majority of BQA datasets are automatically collected.
There are 3 main approaches for automatically collecting BQA datasets:
question generation, cloze generation and exploiting existing QA pairs.
\paragraph{\textbf{Question Generation:}}
Question generation (QG) can automatically generate questions from given contexts \citep{du2017learning}, which can be utilized to build QA datasets.
\citet{yue2020cliniqg4qa} apply the QG approach to synthesize clinical QA datasets without human annotations, and show that the generated datasets can be used to improve BQA models on new contexts.
\paragraph{\textbf{Cloze Generation:}}
Cloze-type QA is the task of predicting a removed word or phrase in a sentence, usually using a detailed context.
Several biomedical QA datasets have been collected by cloze generation, such as CliCR, BioRead, BMKC and BioMRC.
Most of them follow a similar process to \citet{hermann2015teaching} for generating the CNN and Daily Mail datasets:
1. Recognizing biomedical named entities appearing in both summary sentences (e.g. article titles, article learning points) and their detailed contexts (e.g. article abstracts) with tools like MetaMap;
2. Masking the recognized named entities in the summary sentences;
3. The task is to predict the named entities using the masked summary sentences and the contexts.
The generated datasets are typically large-scale (ranging from $100$k to $16.4$M instances) and thus can be used for pre-training \citep{dhingra2018simple} or as a task itself \citep{pappas2018bioread, kim2018pilot, pappas2020biomrc}.
When used for pre-training, cloze-type QA is actually a special type of language modeling that predicts only biomedical entities and is conditioned on the corresponding contexts.
\paragraph{\textbf{Exploiting Existing QA Pairs:}}
Another widely used approach for dataset collection is to exploit naturally existing QA Pairs, or exploiting domain-specific corpora structures.
Biomedical question-answer pairs can be found in a variety of contexts:
For example,
PubMedQA \citep{jin2019pubmedqa} collects citations in PubMed whose titles are questions, and uses the conclusive part of the abstracts as ideal answers.
MedQA \citep{zhang2018medical}, HEAD-QA \citep{vilares2019head} and NLPEC \citep{li2020towards} are built from QA pairs in medical examinations.
MedQuAD collects expert-curated FAQs from trusted medical websites.
cMedQA \citep{zhang2017chinese, zhang2018multi} and ChiMed \citep{tian2019chimed} exploit doctor-patient QA pairs on online healthcare communities.
\subsection{Dataset Difficulty} \label{difficulty}
BQA datasets should be difficult to evaluate the non-trivial reasoning abilities of BQA systems.
In this section, we discuss three types of advanced reasoning abilities: answerability, multi-hop and numeric reasoning.
\paragraph{\textbf{Answerability reasoning:}}
Almost all current BQA datasets and methods focus on answerable questions.
However, not all questions in biomedicine are answerable, the fact that only answerable questions are evaluated can be exploited by BQA systems to get high performance without the expected reasoning process (e.g. by identifying the only text snippet in the context that is consistent with the expected lexical answer type).
In general domain, a strong neural baseline drops from $86$\% F1 on SQuAD v1.1 to $66$\% F1 after unanswerable questions are added \citep{rajpurkar2018know}.
It remains an interesting direction to add unanswerable questions in BQA datasets and test the robustness of BQA systems under such settings.
\paragraph{\textbf{Multi-hop reasoning:}}
Answering real biomedical questions often requires multi-hop reasoning.
For example, doctors might ask ``What tests shall we conduct for patients with [certain symptoms]?".
To answer this question, models must: 1. infer the possible diseases which the patient might have, and 2. find the available tests that are needed for the differential diagnosis.
However, to the best of our knowledge, only MEDHOP evaluates multi-hop reasoning abilities, while almost all other BQA datasets focus on singe-hop reasoning.
\paragraph{\textbf{Numeric reasoning:}}
Numbers are widely used in modern biomedical literature, so understanding the numeric contents in texts is necessary to correctly answer non-trivial scientific questions.
\citet{jin2019pubmedqa} show that nearly all questions from PubMed article titles require quantitative reasoning to answer.
While about 3/4 of them have text descriptions of the statistics, e.g. ``significant differences", about 1/5 only have numbers.
For this, \citet{wu2021bionumqa} re-annotate the PubMedQA and BioASQ datasets with numerical facts as answers, and show that adding numerical encoding scheme improves BioBERT performance on their dataset.
However, most current BQA systems treat numbers as text tokens and do not have specific modules to process them.
\subsection{Biomedical VQA} \label{vqa}
Biomedical VQA is a novel BQA task.
In biomedical VQA, questions are asked about images, which are ubiquitously used and play a vital role in clinical decision making.
Since manual interpretation of medical images is time-consuming and error-prone, automatically answering natural language questions about medical images can be very helpful.
VQA is a novel variant of QA task that require both NLP techniques for question understanding and Computer Vision (CV) techniques for image representation.
General VQA is an active research area and there has been many recent survey articles \citep{gupta2017survey, wu2017visual, srivastava2019visual}.
Here, we mainly focus on \textit{Biomedical VQA Datasets} and their corresponding \textit{Multi-Modal Methods} that fuse NLP and CV methods.
\begin{table*}[ht!]
\small
\centering
\begin{tabular}{p{2cm}p{1cm}p{3.5cm}p{3.5cm}p{1cm}p{1.2cm}}
\toprule
\textbf{Dataset} & \textbf{Size} & \textbf{Metric} & \textbf{State-of-the-art (\%)} & \textbf{Content} & \textbf{Format} \\
\midrule
PathVQA \citep{he2020pathvqa} & $32.8$k & Acc, BLEU-1, BLEU-2, BLEU-3 & 68.2, 32.4, 22.8, 17.4 \citep{he2020pathvqa} & Clinical & Generation \\ \midrule
VQA-Med \citep{abacha2019vqa} & $15.3$k & Acc, BLEU & 64.0, 65.9 \citep{ren2020cgmvqa} & Clinical & Generation \\ \midrule
VQA-Rad \citep{lau2018dataset} & $3.5$k & Acc & 60.0 (open) / 79.3 (close) \citep{zhan2020medical} & Clinical & Generation \\
\bottomrule
\end{tabular}
\caption{An overview of biomedical VQA datasets (listed in alphabetical order).}
\label{tab:vqa}
\end{table*}
\paragraph{{Biomedical VQA Datasets}}
We show an overview of biomedical VQA datasets in Table \ref{tab:vqa} and an instance in Figure \ref{fig:vqa}.
\textbf{VQA-Rad} \citep{lau2018dataset} is the first VQA dataset for radiology.
It contains $3.5$K QA pairs that are annotated by clinicians and the images are from MedPix\footnote{\url{https://medpix.nlm.nih.gov/home}}.
Questions in VQA-Rad are classified into modality, plane, organ system, abnormality, object/condition presence, etc.
Answer formats include multi-choices and generations in VQA-Rad.
The \textbf{VQA-Med} \citep{abacha2019vqa} dataset is automatically constructed using captions of the radiology images.
There are $15.3$K questions in VQA-Med that are restricted to be about only one element and should be answerable from the corresponding images.
VQA-Med concentrates on the most common questions in radiology, which include categories of modality, plane, organ system and abnormality.
Yes/no and WH-questions are included in VQA-Med.
\textbf{PathVQA} \citep{he2020pathvqa} semi-automatically extracts $32.8$K pathology images and generates the corresponding answers from textbooks.
\begin{figure}[ht!]
\centering
{\includegraphics[width=0.72\linewidth]{figs/vqa_instance.pdf}}
\caption{An instance of VQA-Med.}
\label{fig:vqa}
\end{figure}
\paragraph{{Multi-Modal Methods}}
Typically, for biomedical VQA, images and texts are separately encoded, and a multi-modal pooling mechanism is often used to obtain the mixed representations for generating the answers.
\textit{\textbf{Image encoders:}} VGGNet \citep{simonyan2014very} and ResNet \citep{he2016deep} are commonly used for image feature extraction.
\citet{yan2019zhejiang, ren2020cgmvqa} adopt global average pooling \citep{lin2013network} on VGGNet for image encoding to prevent over-fitting on small datasets.
To overcome data limitation of images, \citet{nguyen2019overcoming} apply model-agnostic meta-learning \citep{finn2017model} and convolutional denoising auto-encoder \citep{masci2011stacked} to initialize CNN layers on VQA-Rad, and achieve 43.9 and 75.1 accuracy on open-ended and close-ended questions, respectively.
\textit{\textbf{Text encoders:}} Questions are usually encoded by a recurrent network or a pre-trained language model similar other BQA approaches.
The co-attention mechanism is used for finding important words and regions to enhance both textual and visual representation.
Stacked attention networks \citep{yang2016stacked} use text representation to query visual representation multiple times for obtaining multi-step reasoning.
\textit{\textbf{Multi-modal pooling:}} it is crucial to combine features from the visual and textual encoders. Direct concatenation of them can serve as a baseline.
Multi-modal compact bilinear pooling \citep{fukui2016multimodal}, multi-modal factorized bilinear pooling \citep{yu2017multi} and multi-modal factorized high-order pooling \citep{yu2018beyond} are often used for feature fusion in VQA.
Recently, several multi-modal pre-trained models have been proposed that use transformers \citep{li2019visualbert, tan-bansal-2019-lxmert} to generate visual and textual representations in the general domain.
Similarly, \citet{ren2020cgmvqa} introduce the CGMVQA model that feeds VGGNet and word embedding features into a single transformer for classification or generation on VQA-Med, achieving the accuracy of 0.640 and BLEU of 0.659
\subsection{Domain knowledge Utilization} \label{utilization}
There are a variety of biomedical domain-specific materials and tools that can be used in BQA, including:
1. Large-scale corpora like PubMed and PMC that contain millions of freely available biomedical articles;
2. Various biomedical KBs like UMLS and DrugBank;
3. Many domain-specific NLP tools, e.g.: MetaMap and SemRep for identifying biomedical entities and relations, respectively.
Each kind of resource has its advantages and disadvantages:
Biomedical raw textual resources are extremely large-scale, but their quality cannot be assured.
Specific textual resources, e.g.: FAQs from NLM websites, are regularly maintained and thus of high quality, but they are limited in scale since maintaining and collecting them is expensive.
KBs have high quality and intensive knowledge, but most of them are sparse and incomplete.
However, the abovementioned resources have not been fully utilized by current BQA systems.
As shown in Table \ref{tab:approaches}, different BQA approaches use only one or two different types of resources, but not all of them.
For example, IR, MRC and QE BQA systems typically use textual data, while KB BQA systems mainly use the KBs.
Biomedical NLP tools are mostly used in classic BQA systems.
Since each resource only encodes certain types of biomedical knowledge, only by fusing different BQA approaches can systems fully utilize the domain knowledge.
\begin{table}[ht!]
\small
\centering
\begin{tabular}{p{4.5cm}p{3cm}p{1cm}p{1.5cm}p{3cm}}
\toprule
\textbf{BQA Approach} & \textbf{Texts} & \textbf{Images} & \textbf{KBs} & \textbf{BioNLP tools} \\
\midrule
Information Retrieval & Document collections & -- & -- & -- \\
Machine Reading Comprehension & Raw documents (Contexts) & -- & -- & -- \\
Knowledge Base & -- & -- & \checkmark & Used for KB construction \\
Question Entailment & Existing FAQs & -- & -- & -- \\
Visual Question Answering & -- & \checkmark & -- & -- \\
Classic & Document collections & -- & Ontologies & \checkmark \\
\bottomrule
\end{tabular}
\caption{Utilized domain knowledge by different BQA approaches.}
\label{tab:approaches}
\end{table}
The KMQA model \citep{li2020towards} combines the IR-MRC approach and the KB approach by using co-attention mechanism and a novel knowledge acquisition algorithm.
Their ablation experiments show that only using texts achieves 64.6\% accuracy on the development set; only using knowledge bases achieves 45.3\%; using both texts and knowledge bases achieves 71.1\%.
This shows the effectiveness of the fusion of different BQA approaches.
However, it still remains an underexplored area.
\subsection{Answer Explainability} \label{explainability}
Explainability is a vital property of healthcare applications.
An ideal BQA model should not only have high accuracy in predicting the exact answers, but also be able to provide explanations or evidence for giving such answers.
This improves the answer reliability and enables further fact checking.
Each BQA approach has its intrinsic way for answer explanation:
for the IR approach, the retrieved documents can be considered as evidence;
for the KB approach, the reasoning paths in the KBs provide explainability;
for the QE approach, users are directly pointed to similar questions that have already been answered.
Though controversial \citep{jain2019attention, wiegreffe2019attention}, the attention mechanism \citep{bahdanau2014neural} that is ubiquitously used in modern BQA systems provide at least some level of explainability.
To evaluate explicit answer explanations, Phase B of BioASQ challenges also require the participants to submit ``ideal answers", i.e. answers that include both the exact answers and explanations, in addition to exact answers.
\citet{zhang2019multi} generate explanations with medical KB paths that link the entities extracted from consumer health questions and doctors' answers.
The path representations are learned by a translation-based method and the weights of reasoning paths for specific QA pairs are generated by a hierarchical attention network.
They also use the entity figures return from Google for better entity representation and consumer understanding.
\citet{liu2020interpretable} presents the MURKE model to solve HEAD-QA which iteratively select the most relative document to reformulate the question, where the series of modified questions can be considered as an interpretable reasoning chain.
\subsection{Evaluation Issues} \label{eval}
\paragraph{\textbf{Modular evaluation:}}
Most current evaluations are modular because they only evaluate certain parts of the full BQA system, e.g. for the IR-MRC BQA approach, BioASQ Task B phase A only evaluates the IR methods and the Phase B provides gold standard contexts and only evaluates the MRC methods.
The majority of BioASQ teams only participate in one phase \citep{Nentidis2020overview}.
However, in real settings: 1. it's impossible to have the relevant documents; 2. state-of-the-art MRC BQA systems might not perform well given non-perfect retrieved documents \citep{lin2020pretrained}.
As a result, closing the gap between system modules by combining the evaluations is vital to test the real utility of BQA systems.
In the general domain, \citet{chen2017reading} propose the Machine Reading at Scale task for the complete IR-MRC QA evaluation.
They show that the performance of a complete QA system that reads all Wikipedia might have a large drop compared to its MRC component that reads only the gold standard contexts, e.g.: from 69.5\% EM to 27.1\% on the development set of SQuAD.
In the biomedical domain, many datasets that only contain questions and answers have been proposed.
We list these datasets in Table \ref{tab:openbqa}, most of which are related to Consumer health (5/10) or Examination (4/10), because their dataset sources typically have no supporting materials for the answers.
It should be noted that other types of BQA datasets can also be converted to such datasets by removing the supporting materials (document collections, contexts, FAQs etc).
\begin{table*}[ht!]
\small
\centering
\begin{tabular}{p{2cm}p{1cm}p{1.5cm}p{5.2cm}p{1.5cm}p{1.6cm}}
\toprule
\textbf{Dataset} & \textbf{Size} & \textbf{Metric} & \textbf{State-of-the-art (\%)} & \textbf{Content} & \textbf{Format}
\\ \midrule
ChiMed \citep{tian2019chimed} & $24.9$k & Acc & 98.32 (rel.) / 84.24 (adopt.) \citep{tian2019chimed} & Consumer & Multi-choice \\ \midrule
cMedQA \citep{zhang2017chinese} & $54$k & P@1 & 65.35 (dev) / 64.75 (test) \citep{zhang2017chinese} & Consumer & Multi-choice \\ \midrule
cMedQA v2 \citep{zhang2018multi} & $108$k & P@1 & 72.1 (dev) / 72.1 (test) \citep{zhang2018multi} & Consumer & Multi-choice \\ \midrule
HEAD-QA \citep{vilares2019head} & $6.8$k & Acc & 44.4 (supervised) / 46.7 (unsupervised) \citep{liu2020interpretable} & Examination & Multi-choice \\ \midrule
LiveQA-Med \citep{abacha2017overview} & $738$ & avgScore & 82.7 \citep{abacha2016recognizing} & Consumer & Generation \\ \midrule
MedQA \citep{zhang2018medical} & $235$k & Acc & 75.8 (dev) / 75.3 (test) \citep{zhang2018medical} & Examination & Multi-choice \\ \midrule
MEDQA \citep{jin2020disease} & $61$k & Acc & MC: 69.3 (dev) / 70.1 (test); TW: 42.2 (dev) / 42.0 (test); US: 36.1 (dev) / 36.7 (test) \citep{jin2020disease} & Examination & Multi-choice \\ \midrule
NLPEC \citep{li2020towards} & $2.1$k & Acc & 71.1 (dev) / 61.8 (test) \citep{li2020towards} & Examination & Multi-choice \\ \midrule
webMedQA \citep{he2019applying} & $63$k & P@1, MAP & 66.0, 79.5 \citep{he2019applying} & Consumer & Multi-choice \\
\bottomrule
\end{tabular}
\caption{An overview of the BQA datasets that contain no supporting materials (listed in alphabetical order).}
\label{tab:openbqa}
\end{table*}
Olelo \citep{kraus2017olelo} and Bio-AnswerFinder \citep{ozyurt2020bio} are complete QA systems that participate in the BioASQ challenge.
Olelo is proposed as an integrated web application for QA-based exploration of biomedical literature.
For each user question, Olelo uses the HPI system at BioASQ 2016 (\citealt{schulze2016hpi}, described in \S\ref{mrcbqa}) to retrieve relevant abstracts and return the answers, as well as the entity-supported summarizations of the abstracts \citep{schulze2016entity}.
Bio-AnswerFinder uses iterative document retrieval by LSTM-enhanced keyword queries and BERT-based answer ranking.
The system performance is comparable to a BioASQ 5 SoTA MRC system for factoid questions (38.1\% v.s. 40.5\% MRR, \citet{wiese2017neural}), but is still lower than BioBERT (38.1\%, 48.3\%).
The baselines of ChiMed, cMedQA and webMedQA use answer matching models without explicit supporting materials, and the baselines provided by HEAD-QA, MedQA and MEDQA are basically combined IR-MRC approach (\S\ref{irbqa} and \S\ref{mrcbqa}).
Since most of the current BQA evaluations only focus on the MRC, tasks that involve both retrieving relevant contents and comprehending over them should be further explored.
\paragraph{\textbf{Evaluation metrics:}}
In extractive and generative BQA, current metrics do not consider the synonyms of biomedical concepts.
For example, if the ground truth answer is ``kidney diseases", ``renal diseases" should conceptually be an exact match and ``renal insufficiency" should be rated as relevant.
However, if we use EM in practice, both ``renal diseases" and ``renal insufficiency" have a score of 0;
if we use F1, BLEU or ROGUE, ``renal diseases" is only a partial match and `renal insufficiency" has a score of 0.
\citet{wiese2017neural} report that their model predictions of 10 among 33 analyzed questions are synonyms to the gold standard answers,
but are not counted as right in BioASQ evaluation.
There are two potential approaches to solve this problem:
1. From the annotation side, we can manually annotate more gold standard answers. This approach is expensive but the quality is guaranteed;
2. From the metrics side, it's worth exploring to infuse domain knowledge (e.g.: UMLS ontologies) into current evaluation metrics.
For example, to consider the rich concept synonyms in biomedicine during evaluation, \citet{suster2018clicr} also evaluates QA models by a cosine similarity metric between the mean word vectors of the ground truth and the predicted answer.
\subsection{Fairness and Bias} \label{fairness}
Fairness and bias are serious and vital issues in machine learning.
One cannot be more cautious in removing all potential biases, e.g.: racial and gender biases, when developing healthcare applications like BQA.
Here we discuss the fairness and bias issues of BQA from the NLP and the biomedical side.
From the NLP side: Word embeddings \citep{mikolov2013distributed} are ubiquitously used in NLP models, but such embeddings result in biased analogies like: ``man" is to ``doctor" as ``woman" is to ``nurse".
Similar trends have been observed \citep{kurita2019measuring} in contextualized word representations like BERT.
From the biomedical side, since most current BQA models learn from historically collected data (e.g.: EMRs, scientific literature), populations that have experienced structural biases in the past might be vulnerable under incorrect predictions \citep{rajkomar2018ensuring}.
Some works have been done in general NLP and biomedical machine learning domains, but only a little progress has been made in the BQA domain, and the majority of them study non-English BQA:
Unlike English BQA, non-English BQA suffers additional challenges mainly from the lack of domain-specific resources: much less scientific literature are available in non-English languages; general NLP tools are scarce for non-English languages, let alone biomedical domain-specific ones.
\citet{jacquemart2003towards, delbecque2005indexing} present preliminary studies of BQA in French.
\citet{olvera2011multilingual} evaluate multilingual QA system HONQA and find that English questions are answered much better than French and Italian.
Researchers also introduce multi-lingual BQA datasets for low-resource languages: \citet{zhang2017chinese, zhang2018medical, zhang2018multi, tian2019chimed, he2019applying} for Chinese, \citet{vilares2019head} for Spanish and \citet{veisi2020persian} for Persian.
However, current works are far from enough and our community should seriously take fairness and bias issues into account when introducing new BQA datasets and algorithms in the future.
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[acmlarge]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[acmsmall]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[acmsmall,screen,review]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[acmsmall]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[acmtog]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigconf,authordraft]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigconf]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[manuscript,screen,review]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigconf]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigplan,screen]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigconf]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
\label{sec:intro}
\vspace{-3pt}
The analysis of various environmental sounds in everyday life has be come an increasingly important area in signal processing \cite{Imoto_AST2018_01}.
The automatic analysis of environmental sounds will give rise to various applications, such as anomalous sound detection systems \cite{abnormal}, automatic life-logging systems \cite{Stork_ROMAN2012_01}, monitoring systems \cite{Ntalampiras_ICASSP2009_01}, and bird-call detection systems \cite{Okamoto_crow}.
Sound event detection (SED) is the task of recognizing sound event labels and their timestamp from a recording.
In SED, the models need to recognize overlapped multiple sound events in a time frame.
Recently, neural-network-based SED models have seen increasingly rapid advances, such as the convolutional neural network (CNN) \cite{Hershey_ICASSP2017_01}, recurrent neural network (RNN) \cite{Hayashi_TASLP2017_01}, and convolutional recurrent neural network (CRNN) \cite{SED_CRNN}.
CNN is the structure that automatically extracts features and is robust to time and frequency shifts.
RNN is good at modeling the time structure in an audio stream.
Moreover, some works considering the relationship between sound events and scenes have been proposed.
As an example of the relationship, ``mouse clicking'' occurs indoors such as ``office,'' whereas, ``car'' tends to occurs outdoor such as ``city center.''
On the basis of this idea, SED using the information on the acoustic scene \cite{Mesaros_EUSIPCO2011_01,Heittola_JASM2013_01,Imoto_IEICE2016_01} and the model combining SED and acoustic scene classification (ASC) \cite{helen_interspeech2019,Tonami_WASPAA2019_01,imoto_tonami_icassp2020,Komatsu_icassp2020,dcasenet_arxiv} have been proposed.
Heittola $\textit{et al}$. \cite{Heittola_JASM2013_01} have proposed the SED model using the results of the ASC, where the ASC model is trained in the first stage and then the SED model is trained in the second stage with the ASC results.
Tonami $\textit{et al}$. \cite{Tonami_WASPAA2019_01} have proposed the multitask-learning-based models combining SED and ASC.
\begin{figure}[t!]
\centering
\includegraphics[width=1.00\columnwidth]{two_types_events_v02.eps}
\vspace{-20pt}
\captionsetup{labelformat=empty}
\caption{{\bf Fig. 1}. Example of difference in training difficulty between sound events}
\label{fig:difference_difficulty}
\vspace{-10pt}
\end{figure}
\begin{figure*}[t!]
\centering
\includegraphics[width=2.00\columnwidth]{procedure_train_overall_v04.eps}
\captionsetup{labelformat=empty}
\caption{{\bf Fig. 2}. Examples of early and late stage of training based on curriculum learning}
\label{fig:state_learning}
\vspace{-2pt}
\end{figure*}
In the conventional SED methods, two types of events, namely, those that are present and those that do not occur in an acoustic scene, are treated as the same type of the events.
The conventional SED methods cannot effectively utilize the difference between the two types of events.
The all time frames of events that do not occur in a scene only need to be treated as inactive in the acoustic scene, as shown in Fig. \ref{fig:difference_difficulty} (``elephant'' and ''birdsong'' in ``airplane''), i.e., the training of the easy-to-train events is considered as the task of recognizing one class.
On the other hand, the time frames of events that are present in an acoustic scene must be classified as active or inactive in the acoustic scene, as shown in Fig. \ref{fig:difference_difficulty} (``footsteps'' in ``airplane''), i.e., the training of the difficult-to-train events is regarded as the task of binary classification.
To utilize the difference in the difficulty of training between the sound events, we employ curriculum learning \cite{curriculum}.
Curriculum learning is a method of learning data effectively considering the difficulty of training, in which a model learns progressively from easy- to difficult-to-train data.
Recently, some works using the curriculum learning have been carried out \cite{Speech_noise_curriculum,SER_curriculum,Speech_trains_curriculum}.
Lotfian and Busso \cite{SER_curriculum} have proposed the speech emotion recognition method based on the curriculum learning, where the ambiguity of emotion is considered.
In this paper, we propose a SED method using the curriculum learning, in which strong labels are given for the training.
In the proposed method, the SED models are trained from the easy- to difficult-to-train events on the basis of the curriculum learning.
More specifically, we present a new objective function of SED considering the difficulty of the training of events based on the curriculum learning.
\vspace{-3pt}
\section{Conventional method}
\label{sec:conv}
\vspace{-3pt}
SED involves sound event labels and their onset/offset from an audio.
Recently, many neural-network-based methods have been studied.
In most of the neural-network-based methods, the acoustic features in the time-frequency domain are used for the input to the SED models.
To optimize the neural-network-based SED models, the binary cross-entropy loss is used as follows:
\vspace{-10pt}
\begin{align}
\hspace*{0pt} {\mathcal L}_{\rm BCE}
= - \! \sum^{N}_{n=1} \sum^{T}_{t=1} \! {\Big \{} z_{n,t} \log {\sigma}(y_{n,t}) \!
\nonumber\\[-1pt]
+ ~\! (1 \! - \! z_{n,t}) &\log {\big (} 1 \! - \! {\sigma}(y_{n,t}) {\big )} {\Big \}},
\label{eq:event_loss}
\vspace{-20pt}
\end{align}
\noindent where {\it N} and {\it T} indicate the numbers of sound event categories and time frames, respectively.
$z_{n,t} \in \{ 0,1 \}$ is a target label of an event {\it n} at time {\it t}.
If the event is active, $z_{n,t}$ is 1; otherwise, $z_{n,t}$ is 0.
$y_{n,t}$ represents the output of the network of an event {\it n} at time {\it t}.
$\sigma(\cdot)$ denotes the sigmoid function.
\vspace{-5pt}
\section{Proposed method}
\vspace{-3pt}
\vspace{0pt}
\subsection{Training difficulty of events considering scenes}
\vspace{-3pt}
In the conventional SED methods, two types of events, namely, those that exist and those that do not occur in an acoustic scene, are treated as the same type of the events.
The conventional SED methods cannot effectively employ the difference between the two types of sound events.
The all time frames of events that do not occur in an acoustic scene only need to be regarded as inactive in the acoustic scene, as seen in Fig. \ref{fig:difference_difficulty} (``elephant'' and ``birdsong'' in ``airplane'').
The training of the sound events is treated as the task that recognizes one class (inactive), that is, the events are easy-to-train.
On the other hand, the time frames of sound events that exist in an acoustic scene need to be classified as active in addition to inactive in the acoustic scene, as shown in Fig. \ref{fig:difference_difficulty}.
The training of the sound events is considered as the task that classifies two classes (active or inactive), that is, the events are difficult-to-train.
In short, the sound events that exist in an acoustic scene are hardly trained compared with the events that do not occur in the acoustic scene as shown in Fig. \ref{fig:state_learning}.
\begin{figure*}[t!]
\centering
\includegraphics[width=2.0\columnwidth]{each_event_each_scene_wide.eps}
\vspace{-10pt}
\captionsetup{labelformat=empty,labelsep=none}
\caption{{\bf Fig.~3}. Number of frames of sound events on development set used for our experiments}
\label{fig:frames_events}
\vspace{-5pt}
\end{figure*}
\subsection{Curriculum-learning-based objective function}
\label{sec:prop}
\vspace{-3pt}
As mentioned in Sect. 3.1., there are differences in training difficulty between the sound events when the acoustic scenes are considered.
In the proposed method, we employ the curriculum learning to take advantage of the difference in the difficulty of training between the sound events when the acoustic scenes are considered.
To incorporate the concept of the curriculum learning into the BCE, the following loss function is used instead of Eq. \ref{eq:event_loss}:
\vspace{-10pt}
\begin{align}
\vspace{-25pt}
\hspace*{0pt} {\mathcal L}_{\rm prop}
= - \! \sum^{N}_{n=1} \sum^{T}_{t=1} g_{n} \! &{\Big \{} z_{n,t} \log \sigma(y_{n,t}) \!
\nonumber\\[-1pt]
&+ ~ \! (1 \! - \! z_{n,t})\log {\big (} 1 \! - \! \sigma(y_{n,t}) {\big )} {\Big \}},
\label{eq:proposed_loss_01}
\vspace{-10pt}
\end{align}
\noindent where $g_{n}$ is a gate function that controls the weight of training of two types of events.
More specifically, the gate function is calculated as
\vspace{-15pt}
\begin{align}
\vspace{-15pt}
\hspace*{0pt} g_{n}=\alpha_{s} f_{n}+ (1-\alpha_{s} ) (1-f_{n}),
\vspace{-15pt}
\end{align}
\noindent where $\alpha_{s}$ is a progressive parameter, which is changed from 0 to 1 with time-step {\it s} (epoch) during training.
$f_{n}$ is an event-flag.
If an event $n$ occurs at least once in the acoustic scene of the input audio, the frag is 1; otherwise, it is 0.
As shown in Fig. \ref{fig:state_learning}, in the early stage of the training, only the events that do not occur in an acoustic scene are trained.
On the other hand, in the late stage of the training, only the events that are present in an acoustic scene are trained.
Note that whether an event is difficult- or easy-to-train is determined by each acoustic scene label of an audio clip.
For example, a dataset includes a scene {\tt A}.
Events {\tt a} and {\tt b} occur at least once in the scene {\tt A}.
An event {\tt c} does not occur in the scene {\tt A}.
When the scene label of input audio is {\tt A}, {\tt a} and {\tt b} are regarded as the difficult-to-train events.
{\tt c} is regarded as the easy-to-train event.
\begin{table}[t]
\small
\captionsetup{labelformat=empty,labelsep=none}
\caption{{\bf Table 1}. Experimental conditions}
\vspace{-5pt}
\label{tbl:parameter}
\centering
\scalebox{1.00}[1.00]{
\begin{tabular}{ll}
\wcline{1-2}
&\\[-7pt]
Acoustic feature & Log-mel energy (64 dim.)\\
Frame length \hspace{-3pt} / \hspace{-3pt} shift & 40 ms \hspace{-3pt} / \hspace{-3pt} 20 ms\\
Length of sound clip & 10 s\\\hline
&\\[-8pt]
Network architecture & 3 CNN + 1BiGRU + 1 fully con.\\
\# channels of CNN layers & 128, 128, 128 \\
Filter size & 3$\times$3 \\
Pooling size & 8$\times$1, 2$\times$1, 2$\times$1 (max pooling) \\
&\\[-9pt]
\# units in GRU layer & 32 \\
\# units in fully con. layer & 32 \\
\# units in output layer & 25 \\
Threshold & 0.5 \\
\wcline{1-2}
\end{tabular}
}
\vspace{-5pt}
\end{table}
\begin{table*}[t]
\small
\captionsetup{labelformat=empty,labelsep=none}
\caption{\textbf{Table 3}. SED performance for each event}
\vspace{-18pt}
\label{tbl:each_event}
\begin{center}
\scalebox{0.80}[0.80]{
\begin{tabular}{llcccccccccccc}
\wcline{1-14}
&&&&&&&&&&&&&\\[-8pt]
\multicolumn{2}{c}{\multirow{2}{*}{Event}} & (object) & (object) & (object) & (object) & (object) & bird & brakes & \multirow{2}{*}{breathing} & \multirow{2}{*}{car} & \multirow{2}{*}{children} & \multirow{2}{*}{cupboard} & \multirow{2}{*}{cutlery}\\[-1pt]
&& banging & impact & rustling & snapping & squeaking & singing & squeaking &&&&&\\
&&&&&&&&&&&&&\\[-20pt]
\\\hline
&&&&&&&&&&&&&\\[-10pt]
\multirow{4}{*}{BCE} & \multirow{2}{*}{F-score} & 0.00\% & 0.37\% & \bf 6.06\% & 0.00\% & 0.00\% & 46.31\% & \bf 1.68\% & 0.00\% & 43.46\% & 0.00\% & 0.00\% & 0.00\% \\
&&&&&&&&&&&&&\\[-12pt]
& & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.64 & \scriptsize$\pm$8.85 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$11.56 & \scriptsize$\pm$5.73 & \scriptsize$\pm$0.00 & \scriptsize$\pm$6.19 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 \\\cline{2-14}
&&&&&&&&&&&&&\\[-10pt]
& \multirow{2}{*}{Error rate} & 1.00 & \bf 1.01 & \bf 0.99 & 1.00 & 1.00 & \bf 0.91 & 0.99 & 1.00 & \bf 1.06 & 1.00 & 1.00 & 1.00\\
&&&&&&&&&&&&&\\[-12pt]
& & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.01 & \scriptsize$\pm$0.03 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.05 & \scriptsize$\pm$0.03 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.06 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 \\\hline
& \multirow{2}{*}{F-score} & 0.00\% & \bf 0.78\% & 0.77\% & 0.00\% & 0.00\% & \bf 48.88\% & 1.46\% & 0.00\% & \bf 44.83\% & 0.00\% & 0.00\% & 0.00\% \\
&&&&&&&&&&&&&\\[-12pt]
Proposed& & \scriptsize$\pm$0.00 & \scriptsize$\pm$1.15 & \scriptsize$\pm$1.70 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$5.55 & \scriptsize$\pm$3.86 & \scriptsize$\pm$0.00 & \scriptsize$\pm$7.50 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 \\\cline{2-14}
method & \multirow{2}{*}{Error rate} & 1.00 & 1.02 & 1.03 & 1.00 & 1.00 & 0.96 & 0.99 & 1.00 & 1.15 & 1.00 & 1.00 & 1.00\\
&&&&&&&&&&&&&\\[-12pt]
& & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.03 & \scriptsize$\pm$0.05 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.09 & \scriptsize$\pm$0.02 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.23 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 \\\hline
&&&&&&&&&&&&&\\[-12pt]
\wcline{1-14}
\end{tabular}
}
\end{center}
\vspace{-14pt}
\small
\label{tbl:each_event}
\begin{center}
\scalebox{0.78}[0.78]{
\begin{tabular}{llccccccccccccc}
\wcline{1-15}
&&&&&&&&&&&&&\\[-8pt]
\multicolumn{2}{c}{\multirow{2}{*}{Event}} & \multirow{2}{*}{dishes} & \multirow{2}{*}{drawer} & \multirow{2}{*}{fan} & glass & keyboard & large & mouse & mouse & people & people & washing & water tap & wind
\\[-1pt]
&& & & & jingling & typing & vehicle & clicking & wheeling & talking & walking & dishes & running & blowing\\
&&&&&&&&&&&&&\\[-20pt]
\\\hline
&&&&&&&&&&&&&&\\[-10pt]
\multirow{4}{*}{BCE} & \multirow{2}{*}{F-score} & 0.00\% & 0.00\% & 9.95\% & 0.00\% & 0.00\% & 16.93\% & 0.00\% & 0.00\% & 0.00\% & 2.67\% & 17.25\% & 40.78\% & \bf 0.55\% \\
&&&&&&&&&&&&&&\\[-12pt]
& & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$18.74 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$1.43 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.08 & \scriptsize$\pm$3.66 & \scriptsize$\pm$12.70 & \scriptsize$\pm$12.36 & \scriptsize$\pm$0.21 \\\cline{2-15}
& \multirow{2}{*}{Error rate} & 1.00 & 1.00 & 0.95 & 1.00 & 1.00 & 6.17 & 1.00 & 1.00 & 1.06 & 1.02 & \bf 1.19 & \bf 0.83 & \bf 1.00 \\
&&&&&&&&&&&&&&\\[-12pt]
& & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.13 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.67 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.09 & \scriptsize$\pm$0.03 & \scriptsize$\pm$0.17 & \scriptsize$\pm$0.10 & \scriptsize$\pm$0.00 \\\hline
& \multirow{2}{*}{F-score} & \bf 1.22\% & 0.00\% & \bf 50.63\% & 0.00\% & \bf 0.01\% & \bf 18.13\% & 0.00\% & 0.00\% & \bf 0.06\% & \bf 3.43\% & \bf 24.80\% & \bf 45.07\% & 0.07\% \\
&&&&&&&&&&&&&&\\[-12pt]
Proposed & & \scriptsize$\pm$3.15 \scriptsize$\pm$0.00 & \scriptsize$\pm$26.68 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.03 & \scriptsize$\pm$5.02 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.17 & \scriptsize$\pm$3.46 & \scriptsize$\pm$13.15 & \scriptsize$\pm$4.82 & \scriptsize$\pm$0.20 \\\cline{2-15}
method & \multirow{2}{*}{Error rate} & 1.00 & 1.00 & \bf 0.77 & 1.00 & 1.00 & \bf 4.54 & 1.00 & 1.00 & \bf 1.03 & 1.02 & 1.32 & 0.84 & 1.01 \\
&&&&&&&&&&&&&&\\[-12pt]
& & \scriptsize$\pm$0.01 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.19 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$2.27 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.00 & \scriptsize$\pm$0.03 & \scriptsize$\pm$0.03 & $\pm$\scriptsize0.26 & $\pm$\scriptsize0.04 & \scriptsize$\pm$0.02 \\
&&&&&&&&&&&&&&\\[-12pt]
\wcline{1-15}
\end{tabular}
}
\vspace{-10pt}
\end{center}
\end{table*}
\begin{table}[t]
\small
\captionsetup{labelformat=empty,labelsep=none}
\caption{\textbf{Table 2}. Overall performance of SED}
\vspace{-18pt}
\label{tbl:F-score}
\begin{center}
\scalebox{0.88}[0.88]{
\begin{tabular}{ccccccc}
\wcline{1-7}
&&&&&&\\[-7pt]
\multicolumn{3}{c}{\multirow{2}{*}{Method}} & \multicolumn{2}{c}{F-score} & \multicolumn{2}{c}{Error rate} \\
&&&&&&\\[-13pt]
&&& micro & macro & micro & macro \\\hline
&&&&&&\\[-9pt]
\multicolumn{3}{l}{\multirow{2}{*}{BCE}} & 25.30\% & 7.44\% & 1.00 & 1.21 \\[-2pt]
&&&&&&\\[-13pt]
&&& \scriptsize$\pm$4.72 & \scriptsize$\pm$1.21 & \scriptsize$\pm$0.04 & \scriptsize$\pm$0.03\\[0pt]
\multicolumn{3}{l}{\multirow{2}{*}{MTL of SED \& SAD}} & 26.62\% & 7.36\% & 1.02 & 1.20 \\[-2pt]
&&&&&&\\[-13pt]
&&& \scriptsize$\pm$2.68 & \scriptsize$\pm$0.65 & \scriptsize$\pm$0.03 & \scriptsize$\pm$0.09\\[0pt]
\multicolumn{3}{l}{\multirow{2}{*}{MTL of SED \& ASC}} & 26.12\% & 7.46\% & 0.97 & 1.18\\[-2pt]
&&&&&&\\[-13pt]
&&& \scriptsize$\pm$3.94 & \scriptsize$\pm$0.58 & \scriptsize$\pm$0.07 & \scriptsize$\pm$0.06 \\[0pt]\hdashline
&&&&&&\\[-8pt]
\multicolumn{3}{l}{\multirow{2}{*}{Proposed method}} & \bf 35.39\% & \bf 9.61\% & \bf 0.85 & \bf 1.15\\[-4pt]
&&& \scriptsize$\pm$6.06 & \scriptsize$\pm$1.24 & \scriptsize$\pm$0.07 & \scriptsize$\pm$0.09 \\[0pt]
\multicolumn{3}{l}{\multirow{2}{*}{Proposed+MTL of SED \& SAD}} & \bf 33.57\% & \bf 9.11\% & \bf 0.93 & \bf 1.17\\[-4pt]
&&& \scriptsize$\pm$4.86 & \scriptsize$\pm$0.81 & \scriptsize$\pm$0.07 & \scriptsize$\pm$0.07 \\[0pt]
\multicolumn{3}{l}{\multirow{2}{*}{Proposed+MTL of SED \& ASC}} & \bf 35.62\% & \bf 9.65\% & \bf 0.85 & \bf 1.15\\[-4pt]
&&& \scriptsize$\pm$6.35 & \scriptsize$\pm$1.31 & \scriptsize$\pm$0.09 & \scriptsize$\pm$0.09 \\[0pt]\wcline{1-7}
\end{tabular}
}
\vspace{-22pt}
\end{center}
\end{table}
\vspace{-5pt}
\section{Experiments}
\label{sec:exp}
\vspace{-2pt}
\subsection{Experimental conditions}
\label{sec:condition}
\vspace{-2pt}
To evaluate the performance of the proposed method, we conducted evaluation experiments using the TUT Sound Events 2016 \cite{Mesaros2016TUTDF}, TUT Sound Events 2017 \cite{Mesaros2017}, TUT Acoustic Scenes 2016 \cite{Mesaros2016TUTDF}, and TUT Acoustic Scenes 2017 \cite{Mesaros2017} datasets.
From these datasets, we selected sound clips including four acoustic scenes, ``home,'' ``residential area'' (TUT Sound Events 2016), ``city center'' (TUT Sound Events 2017, TUT Acoustic Scenes 2017), and ``office'' (TUT Acoustic Scenes 2016), which contain 266 min (development set, 192 min; evaluation set, 74 min) of audio.
Here, the acoustic scene ``office'' in TUT Acoustic Scenes 2016 and ``city center'' in TUT Acoustic Scenes 2017 did not have sound event labels.
We thus manually annotated the sound clips with sound event labels by the procedure described in \cite{Mesaros2016TUTDF,Mesaros2017}.
These sound clips include the 25 types of sound event labels.
Fig. \ref{fig:frames_events} shows the numbers of active time frames of sound events on the development set that we used.
The labels of events annotated for our experiment are available in \cite{Imoto_dataset2019_01}.
As acoustic features, we used 64-dimensional log-mel energies calculated for each 40 ms time frame with 50\% overlap.
This setting is from the baseline system of the DCASE2018 Challenge task4 \cite{setting01}.
As the baseline model of SED, we used the convolutional neural network and bidirectional gated recurrent unit (CNN--BiGRU) \cite{SED_CRNN}.
Moreover, to verify the usefulness of the proposed method, we used a model combining SED and sound activity detection (SAD) based on multitask learning (MTL), referred to as ``MTL of SED \& SAD'' \cite{SED_SAD}, and a model combining SED and ASC, referred to as ``MTL of SED \& ASC'' \cite{Tonami_WASPAA2019_01}.
The sound activity detection is the mechanism of recognizing any active events in a time frame.
The reason why we choose MTL of SED \& SAD is that this modern method, in which no information on the scene is considered, is simple but effective.
MTL of SED \& ASC is the multitask-learning-based SED with ASC, which uses scene labels by ASC.
Other experimental conditions are listed in Table~\ref{tbl:parameter}.
In Table~\ref{tbl:parameter}, X $\times$ Y denotes that the filter size is X along the frequency axis by Y along the time axis.
As the evaluation metric, the frame-based metric is used.
We conduct the experiments using ten initial values.
To evaluate the SED performance, a segment-based metric \cite{Mesaros2016_MDPI} is used.
In this work, the size of a segment is set to the frame length.
In this work, we adopt the following exponential scheduler as the progressive parameter in Eq. 3:
\vspace{-10pt}
\begin{align}
\vspace{-15pt}
\hspace*{0pt} \alpha_{s} = \biggl( \frac{s}{s_{max}} \biggl)^\lambda ,
\vspace{-15pt}
\end{align}
\vspace{-10pt}
\noindent where $s$ and $s_{max}$ represent the current and maximum epoch, respectively.
$\lambda$ is tuned using the development dataset and is set as 2.0.
\vspace{-8pt}
\subsection{Experimental results}
\label{sec:results}
\vspace{-3pt}
Table 2 shows the SED performances in terms of the segment-based F-score and error rate.
In Table 2, micro and macro indicate the overall and class-average scores, respectively.
The numbers to the right of $\pm$ represent standard deviations.
``BCE'' is the CNN--BiGRU using the BCE loss.
``Proposed method'' represents the SED performance using Eqs. 2 and 3 with CNN--BiGRU.
``Proposed+MTL of SED \& SAD'' indicates the SED performance using Eqs. 2 and 3 with SAD.
``Proposed+MTL of SED \& ASC'' denotes the multitask-learning-based SED with ASC using the proposed objective function for SED.
The results show that the proposed method achieves a more reasonable performance than the conventional BCE.
Moreover, when using SAD and the model combining SED and ASC with the proposed objective function, the SED performance is better than those of the conventional MTL of SED \& SAD and the MTL of SED \& ASC.
In particular, ``Proposed method'' improves the F-score of SED by 10.09 percentage points compared with that of the conventional SED using the BCE.
The results indicate that the proposed method considering the training difficulty of events enables more effective SED performances than the conventional method using the BCE.
To investigate in detail the SED performance, we observed the segment-based F-score and error rate for each event.
Table 3 indicates the SED performance for each event.
As shown in Table 3, the proposed method outperformed the conventional SED using the BCE for many events.
In particular, the F-scores for ``fan,'' ``washing dishes,'' and ``water tap running'' are more significantly improved in the proposed method than in the conventional method.
This might be because the active frames of these events occur continuously, that is, these events are relatively effortless to be detected compared with other events.
On the other hand, the F-scores for ``(object) rustling,'' ``brakes squeaking,'' and ``wind blowing'' do not improve.
This is because the numbers of the events of the active frames are too small as shown in Fig. \ref{fig:frames_events}.
In other words, the active frames are trained mainly in the late stage of training when using the proposed method.
This may also lead to the poor results for some events when using the proposed method.
\vspace{-12pt}
\section{Conclusion}
\label{sec:conc}
\vspace{-8pt}
In this paper, we proposed the curriculum-learning-based objective function for SED.
In the proposed method, we applied the training difficulty between sound events considering acoustic scenes to the conventional BCE loss.
More specifically, the SED models using the proposed method are trained from the easy-to-train to difficult-to-train events during training.
The experimental results indicate that the proposed method improves the F-score of the SED by 10.09 percentage points compared with that of the conventional CNN--BiGRU using the BCE loss.
In our future work, we will investigate a more effective method for SED considering the relationship between sound events and acoustic scenes.
\vspace{-4pt}
\section{Acknowledgement}
\label{sec:ack}
\vspace{-4pt}
This work was supported by JSPS KAKENHI Grant Number JP19K20304.
\vspace{-8pt}
\begin{spacing}{0.99}
\bibliographystyle{IEEEbib}
|
\section{Introduction}
Greenberger-Horne-Zeilinger~(GHZ) states constitute an important class of entangled many-body states. A general form of $n$-qubit GHZ states can be expressed as~\cite{Bell1990,Bayesian2020,Reiter2016,Omran570,One2001}
\begin{equation}\label{e000}
\frac{1}{\sqrt{2}}(|\alpha_1 \alpha_2 \alpha_3\cdots \alpha_n \rangle+e^{i\phi}|\beta_1 \beta_2 \beta_3 \cdots \beta_n \rangle).
\end{equation}
where $\alpha_j+\beta_j=1~(\alpha_j,\beta_j\in\{0,1\})$. These states play a key role in quantum science and technologies, including open-destination quantum teleportation~\cite{Yang2004}, concatenated error-correcting codes~\cite{Knill2005}, quantum simulation~\cite{Song2018}, and high-precision spectroscopy~\cite{Leibfried2004}. In principle, the benchmark for quantum information capability is the number of particles that can be reliably entangled in a quantum processor. In experiment, multi-body entanglement is achieved recently by capturing 20 trapped ions with around the fidelity of $63.2\%$~\cite{Friis2018}, 12 photons with $59.8\%$~\cite{Zhong2018}, 18 photonic qubits exploiting three degrees of freedom of six photons with $72.4\%$~\cite{Wang2018}, 12 superconducting qubits with $55.6\%$~\cite{Gong2019}, 18 superconducting qubits with $53.0\%$~\cite{Song2019} and 20 Rydberg atoms with $54.2\%$~\cite{Omran2019}. Among various physical systems, the ubiquitous noise and device imperfections are, however, unavoidable and limit the range where multi-body entanglement can be realized with a high fidelity.
Topological insulator~\cite{Qi2011,Chiu2016}, a new kind of novel state
of matters, is characterized by the conducting edge states and the insulating bulk states. The
special conducting edge states are protected by the energy gap of the topological system, leading the edge state to be insensitive to local perturbations and disorders \cite{Kitaev2001,Kane2005,Hasan2010}. To this end, the topological state of matter has been studied in many systems, such as optical lattice systems~\cite{Guo2016,Lin2015}, spin Fermi systems~\cite{Ren2015,Shih2016}, supercooled atoms~\cite{Jotzu2014,Aidelsburger2015} and synthetic materials~\cite{Sun2017}. These novel properties support many potential applications of topological insulators in quantum information processing and computing. For instance, several proposals for quantum state transfer (QST) have been presented ~\cite{Yao2013,Dlaska2017,Mei2018,Longhi2019,Qi2020,Tan2020,Zheng2020,Qi2020pra} with topologically protected channels. A topologically protected channel for QST between remote quantum nodes mediated by the edge mode of a chiral spin liquid was proposed and analyzed~\cite{Yao2013,Dlaska2017}. Mei $et~al.$ presented an experimentally feasible mechanism for implementing robust QST via the topological edge states by connecting superconducting Xmon qubits into a one-dimensional chain~\cite{Mei2018}. Also, topologically protected entangled photonic states and its transport via edge states have been reported~\cite{Mikael2016,Blanco2018,Michelle2019,Hu2020}. In experiment, the topological protection of spatially entangled biphoton states was demonstrated~\cite{Michelle2019}. The robustness in crucial features of the topological biphoton correlation map in the presence of deliberately introduced disorder was found in the silicon nanophotonic structure. Recently, a spatially entangled two-particle NOON state is proposed by topological Thouless pumping in one-dimensional disordered lattices~\cite{Hu2020}.
In this paper, we propose a superconducting-circuit model to generate large-scale GHZ states, where the model is a chain consisting of $N$ flux qutrits connected by $(N-1)$ resonators, and the entangled states are protected by topological zero-energy mode. We analytically derive the wave function of the topological edge state with zero energy of the qutrit-resonator chain, through which a state transfer between two ends of the chain, accompanied with the state flips of all intermediate qutrits, can be implemented by designing qutrit-resonator coupling strengths. The key advantage that topology offers in such processes is the inherent protection of boundary edge states lying in the band gap of the dispersion relation when the bulk is topologically nontrivial. Based on such a peculiar state transfer protected topologically, we show three schemes of generating large-scale GHZ states, providing feasible and visible methods to generate the robust large-scale GHZ states by meeting the performance requirements of different experimental devices in the superconducting qutrit-resonator chain. Meanwhile, we take into account the experimental consideration, such as the implementation of the model, preparation of the initial state, and realization of tunable couplings in the circuit-QED chain. In addition, we discuss the potential of realizing fast GHZ states by speeding up the adiabatic proscess, which provides more possibilities for obtaining high-fidelity multi-body entanglement.
Our work may facilitate the potential applications of topological matter in quantum information processing, due to the following advantages and interests. First, the topological qutrit-resonator chain can be used to create a large-scale GHZ state theoretically with the size being far more than $N=20$ that is the particle number of multi-body entanglement realized experimentally up to now~\cite{Friis2018,Zhong2018,Wang2018,Gong2019,Song2019,Omran2019}. Second, as the core foundation of implementing large-scale GHZ states, we derive theoretically the wave function of an edge state with zero-energy mode, whose form involves the state flips of all intermediate qutrits, different from that in the frequently studied standard Su-Schrieffer-Heeger (SSH) model~\cite{Mei2018,Qi2020,Tan2020,Zheng2020}.
Finally, there are three schemes proposed for generating large-scale GHZ states, which provide potential choices, depending on different device requirements, i.e., coherence times of qutrits and resonators.
\section{Physical model and wave function of an edge state}\label{S2}
\begin{figure*}\centering
\includegraphics[width=0.9\linewidth]{fig1.eps}
\caption{(a)~A superconducting qutrit-resonator chain. The $n$-th unit cell contains one flux qutrit and one single-mode resonator, labeled as $A_n$ and $B_n$, respectively, and holds an intra-cell qutrit-resonator coupling strength $J_1$. Between two adjacent cells, a qutrit $A_{n+1}$ is coupled to the resonator $B_n$ with an inter-cell coupling strength $J_2$. The resonator $B_n$ drives the transition $|L\rangle \leftrightarrow |e\rangle$ ($|R\rangle \leftrightarrow |e\rangle$) of the two nearest-neighbor qutrits $A_n$ and $A_{n+1}$ when $n$ is odd (even). (b)~Circuit schematic of one unit cell in the superconducting qutrit-resonator chain. The coupling strength can be dynamically tuned by a coupler of SQUID. The energy level space of the qutrit can be tuned by FBL. (c)~Schematics of energy level transitions for qutrits $A_1$, $A_n$~($1<n<N$), and $A_N$. The energy level structure of a flux qutrit holds two ground states ($|L\rangle$ and $|R\rangle$) and one excited state ($|e\rangle$). We denote that coupling strengths $J_L=J_1$ and $J_R=J_2$~($J_L=J_2$ and $J_R=J_1$) when $n$ is even (odd).}\label{f1}
\end{figure*}
\subsection{Physical model}\label{S2A}
The setup of the superconducting qutrit-resonator chain for generating large-scale GHZ states is illustrated in Fig.~\ref{f1}(a). The chain contains $(2N-1)$ lattice sites, $N$ qutrits and $(N-1)$ resonators. Each unit cell in the chain contains one flux qutrit $A_n$ and one resonator $B_n$, whose circuit schematic is described by Fig.~\ref{f1}(b). The energy levels of qutrit and the coupling strengths are adjustable via the magnetic flux provided by the flux-bias line~(FBL)~\cite{Schmidt2013} and superconducting quantum interference device (SQUID). As described in Fig.~\ref{f1}(c), each flux qutirt holds a three-level structure, involving two ground states $|L\rangle$ and $|R\rangle$, and one excited state $|e\rangle$.
The interaction in the chain can be described by the following interaction-picture Hamiltonian ($\hbar$ = 1)
\begin{equation}\label{e0}
H_{I}=\sum_{n=1}^{N-1}\big(J_1|e\rangle_{n} \langle j_n|+ J_2|e\rangle _{n+1}\langle j_n|\big)b_n+\rm{H.c.},
\end{equation}
where $j_n=L~(R)$ when $n$ is odd~(even), and $b_{{n}}$ the annihilation operator of the resonator $B_n$. $J_1$ and $J_2$ can be tuned through adopting controlled voltage pulses generated by an arbitrary waveform generator (AWG) to tune the flux threading the loop~\cite{Majer2009}. Further, $H_{I}$ can be rewritten as
\begin{equation}\label{e1}
H=\sum_{n=1}^{N-1}J_1b_n\sigma_{n}^+(\sigma_{n}^{x})^{n}+J_2b_{n}\sigma_{n+1}^+(\sigma_{n+1}^{x})^{n}+\rm{H.c.},
\end{equation}
where we define $\sigma_{n}^+=|e\rangle_n\langle R|$, $\sigma_{n}^{x}=|L\rangle_n\langle R|+|R\rangle_n\langle L|$. The existence of the $(\sigma_{n}^{x})^n$ renders two different transitions $|L\rangle \leftrightarrow |e\rangle$ and $|R\rangle \leftrightarrow |e\rangle$ for qutrit $A_n$ depending on the odevity of $n$. For instance, $\sigma_{n}^+(\sigma_{n}^{x})^{n}=|e\rangle_n \langle R|~(|e\rangle_n \langle L|)$ with $n$ being even~(odd).
Such a adjustable chain can be analogous to an SSH model~\cite{Soliton1980,short2016,Su1979} which describes quanta~(e.g., electrons, photons, or phonons) hopping on a chain (one-dimensional lattice), with staggered hopping amplitudes. The chain of SSH model consists of $N$ unit cells, each unit cell hosting two sites, one on sublattice $A$, and the other on sublattice $B$. The Hamiltonian of the standard SSH model is of the form $H_{SSH}=(\nu\sum_{m=1}^{N}|m,B\rangle\langle m,A|+\mu\sum_{m=1}^{N-1}|m+1,A\rangle\langle m,B|)+\rm{H.c.}$. Here $|m,A\rangle$ and $|m,B\rangle$, with $m \in {1,2,\cdots N}$, denote the states of the chain where the hopping quantum is on sites $A$ and $B$, respectively, in the unit cell $m$~\cite{Soliton1980,short2016}. The Hamiltonian~(\ref{e1}) holds a form of the SSH model whose topological phase is characterized by winding number~\cite{Ryu_2010}, except for an additional operator $(\sigma_{n}^{x})^{n}$ that works for flipping qutrit states and is the key to realizing GHZ states. Based on the bulk-edge correspondence~\cite{Qi2010,Hasan2010,Classification2016}, the SSH model possesses zero-energy edge modes at open boundaries in the topologically nontrivial phase, which are protected by the topological properties of the system.
\subsection{Wave function of an edge state}\label{S2B}
The appearance of topologically protected gapless edge states within the bulk gap is a manifestation of the topological insulator. The number of such gapless edge modes is specified by topological invariants. As for an SSH model, the zero energy mode, one of the characteristics of the topological nontrivial SSH phase, is regarded as the topological invariant in the edge state and will protect the edge state topologically from local disorders. Thus, it is critical to obtain the wave function of the edge state with zero energy mode in our model.
The edge states of the chain with a single excitation are exponentially localized at the boundaries. The wave function of an edge state can be described by the following ansatz, analogous to the standard SSH model~\cite{Soliton1980,short2016,Su1979}
\begin{equation}\label{e2}
|\varphi_e\rangle=\sum_{n=1}^{N}\lambda^{n}\Big[\gamma\sigma_{n}^+(\sigma_{n}^{x})^{n-1}\bigotimes^{n-1}_{l=1}\sigma_{l}^{x}+\eta b_{{n}}^{\dagger}\bigotimes^{n}_{l=1}\sigma_{l}^{x}\Big]|G\rangle,
\end{equation}
where
\begin{equation*}
|G\rangle=|\underbrace{RLR\cdots}_{N} \rangle_{A_{N}}\otimes |\underbrace{000\cdots}_{N-1}\rangle_{B_{N-1}}
\end{equation*}
denotes a decoupled state of the qutrit-resonator chain with all resonators in $|0\rangle$~(i.e., zero-photon Fock state), while qutrits $A_1$, $A_2$, $A_3$, $\cdots$ in states $|R\rangle$, $|L\rangle$, $|R\rangle$, $\cdots$, repsectively. $\lambda$ is the localized index, $\gamma$ and $\eta$ being the probability amplitudes of the gap states. When $\lambda \textless 1$ ($\lambda \textgreater 1$), the probability amplitude of the site $n$ decays (increases) exponentially with distance $n$ which means the left (right) edge state with the wave function normalized. When $\gamma=0$~($\eta=0$), according to Eq.~(\ref{e2}) the resonators~(qutrits) are occupied by the edge state whose eigenenergy is $E=0$. In particular, in order to generate large-scale GHZ states with binary quantum information carried by two states in qutrits, we choose $\gamma=1$ and $\eta=0$ to render the qutrit in each unit cell to occupy the left~($\lambda<1$) and right~($\lambda>1$) edge states with $E=0$. Thus, the wave function of an edge state occupied by qutrits can be written as $|\varphi'_e\rangle=\sum_{n=1}^{N}\lambda^{n}\sigma_{n}^+(\sigma_{n}^{x})^{n-1}\bigotimes^{n-1}_{l=1}\sigma_{{l}}^{x}|G\rangle$. Substituting $|\varphi_e'\rangle$ into the eigenvalue equation $E|\varphi_e'\rangle=H|\varphi_e'\rangle$, one can obtain
\begin{eqnarray}\label{e3}
&&E\sum_{n=1}^{N}\lambda^{n}\sigma_{n}^+(\sigma_{n}^{x})^{n-1}\bigotimes^{n-1}_{l=1}\sigma_{{l}}^{x}|G\rangle\nonumber\\
&&=\Big(J_1\lambda^{n} b_{{n}}^{\dagger}\bigotimes_{l=1}^{n}\sigma_{{l}}^{x}
+J_2\lambda^{n+1}b_n^{\dagger}\bigotimes_{l=1}^{n}\sigma_{{l}}^{x}\Big)|G\rangle.
\end{eqnarray}
Through these analyses, the edge state can be worked out as
\begin{equation}\label{e4}
|\varphi_E\rangle=\sum_{n=1}^{N}\lambda^{n}\sigma^+_{{n}}(\sigma_{n}^{x})^{n-1}\bigotimes^{n-1}_{l=1}\sigma_{{l}}^{x}|G\rangle,
\end{equation}
with $\lambda=-{J_1}/{J_2}$.
\begin{figure*}
\centering
\includegraphics[width=0.75\linewidth]{fig2.eps}
\caption{(a) Energy spectra of the chain versus $J_1/J_2$. (b) Distribution of the zero energy mode versus $J_1/J_2$. (c) Energy spectra of the chain versus $g_0t$. (d) Distribution of the zero energy mode versus $g_0t$. The size of the chain is $L=2N-1=49$.}\label{f2}
\end{figure*}
To illustrate the topological property of the qutrit-resonator chain more clearly, we take the size $L\equiv2N-1=49$ of the chain as an example to plot the energy spectrum of the system and the distribution of the topological edge states, respectively, in Figs.~\ref{f2}(a) and (b) with varying $J_1/J_2$. We find that the energy spectrum possesses a unique zero energy mode keeping unchanged with varying $J_1/J_2$, which denotes a non-evolutive state. The closest distance between the zero energy state and the bulk appears at the point $J_1/J_2=1$. In Fig.~\ref{f2}(b), the zero energy state is localized near the left extremity when $J_1/J_2 <1$, while for $J_1/J_2>1$ it is localized near the right extremity. It means that one can achieve the topological state transfer assisted by the zero energy mode between the first and the last qutrits via varying the parameter $J_1/J_2$, which is consistent with conclusion obtained from Eq.~(\ref{e4}). Therefore, the Hamiltonian $H$ is a modulated model if varying the coupling strength $J_1$ and $J_2$. Specifically, the shapes of the coupling strengths are engineered as Gauss functions
\begin{equation}\label{e5}
\begin{split}
J_1&=g_0\exp{[-(t-3\tau)^2/\tau^2]},\\
J_2&=g_0\exp{[-(t-2\tau)^2/\tau^2]},
\end{split}
\end{equation}
where $\tau=T/7$ is set as the value of the pulse width, as well as the delay of $J_1$ and $J_2$, with $T$ being the evolution time. The forms of $J_1$ and $J_2$ in Eq.~(\ref{e5}) satisfy the state transfer conditions $J_1/J_2|_{t\rightarrow0}=0$ and $J_1/J_2|_{t\rightarrow T}=+\infty$ well. The idea of engineering the coupling strengths as Gauss functions is to achieve a temporal soft quantum control starting from and ending at a zero amplitude, which enables on-resonant couplings among a desired set of target levels, while efficiently avoiding unwanted off-resonant contributions coming
from others~\cite{Haase2018,Wu2020}.
Based on time-dependent Gaussian coupling strengths, as shown in Fig.~\ref{f2}(c), there exists a zero energy mode among all eigenstates during the whole evolutionary process. Simultaneously, the topological edge state at zero energy is well separated from bulk states. In Fig.~\ref{f2}(d), we plot the state distribution of the zero mode. The qutrit-resonator chain has not only a bulk part but also boundaries (which we refer to ends or edges). The qutrits $A_1$ and $A_N$ are regarded as two edges, while the other qutrits and resonators are the bulk part. The distribution of the zero mode state decays exponentially on the qutrits under the condition of $g_0t<1800$. Particularly, the eigenstate of the zero energy is localized near the first lattice site $A_1$~(left edge) and the distribution of left edge state with zero mode is equal to unity when $g_0t<1800$. When $g_0t$ increases continuously, the distribution of zero mode increases exponentially on the qutrits. Similarly, the eigenstate is localized near the last lattice site $A_{25}$ (right edge) and the distribution of right edge state with zero mode is equal to unity when $g_0t>2200$.
\section{Large-scale GHZ states}\label{S3}
\subsection{Scheme A for generating large-scale GHZ states}\label{S3A}
Now we focus on the generation of large-scale GHZ states among the $N$ qutrits in the superconducting qutrit-resonator chain. In Fig.~\ref{f2}(d), the edge state concentrates towards the left (right) end when $g_0t<1800$~($g_0t>2200$). In particular, when $g_0t\rightarrow 0$ and $g_0t\rightarrow +\infty$, the edge states become
\begin{eqnarray}\label{e01}
|l\rangle&=&|eLR\cdots m\rangle_{A_{N}}\otimes|000\cdots0\rangle_{B_{N-1}},\nonumber\\
|r\rangle&=&|LRL\cdots e\rangle_{A_{N}}\otimes|000\cdots0\rangle_{B_{N-1}},
\end{eqnarray}
where $m=R~(L)$ when $N$ is odd (even). For generating GHZ states, the superconducting qutrit-resonator chain is assumed initially in the state
\begin{eqnarray}\label{e6}
|\phi_0\rangle&=&\frac{1}{\sqrt{2}}\left(|G\rangle+|l\rangle\right).
\end{eqnarray}
Going through the evolution, the first decoupled state component in the Eq.~(\ref{e6}) does not evolve because no photon in the resonators can be absorbed to excite the ground-state qutrits. From the Fig.~\ref{f2}(d), we learn that the second term of Eq.~(\ref{e6}) is essentially the left edge state with zero energy, which can evolve into the right edge state along the topologically protected process. Thus, the following evolution occurs
\begin{equation*}
|\phi_0\rangle~\mapsto~\frac{1}{\sqrt{2}}\left[|G\rangle-(-1)^{N}|r\rangle\right].
\end{equation*}
Here we set that $|R\rangle$ is used to carry the logical state 1 and $|L\rangle$ is used to carry the logical state 0 for the qutrit $A_n$~($1\leq n<N$), while $|e\rangle$ of the last qutrit $A_N$ is encoded as the logic state 1(0) when $N$ is even (odd). Except for the last qutrit $A_N$, the excitation state $|e\rangle$ is an auxiliary state without carrying any quantum information. Accordingly, the final state of $N$ qutrits after the evolution with omitting the zero-photon product state of resonators is
\begin{equation}\label{e7}
\frac{1}{\sqrt{2}}(|101\cdots 1(0)\rangle_{A_{N}}-(-1)^{N}|010 \cdots 0(1)\rangle_{A_{N}}),
\end{equation}
which is exactly an $N$-body GHZ state according to Eq.~(\ref{e000}).
As an example, with the full Hamiltonian Eq.~(\ref{e1}) we take $N=25$ to numerically plot in Fig.~\ref{f3} the population evolution of the initial state, the ideal state $|\phi_{ideal}\rangle=\frac{1}{\sqrt{2}}(|RLR\cdots R\rangle_{A_{25}}+|LRL \cdots e\rangle_{A_{25}})\otimes|000\cdots 0\rangle_{B_{24}}$, and two edge states $|l_{\rm A}\rangle=|eLR\cdots R\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}}$ and $|r_{\rm A}\rangle=|LRL\cdots e\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}}$.
Evidently, the population of $|\phi_{ideal}\rangle$ evolves from $0.25$ to $1$, which indicates the successful creation of a $25$-body GHZ state. When $g_0t\approx 1800$, the populations of $|\phi_{ideal}\rangle$ and $|r_{\rm A}\rangle$ suddenly increase, which shows the same result as the distribution of the zero mode in Fig.~\ref{f2}(d).
\begin{figure}[htb]\centering
\centering
\includegraphics[width=1\linewidth]{fig3.eps}
\caption{Time evolution of populations for $|\phi_0\rangle$, $|\phi_{ideal}\rangle$, $|l_{\rm A}\rangle$ and $|r_{\rm A}\rangle$ with the size of the superconducting qutrit-resonator chain $L=49$. We choose $g_0T=3600$ as a total evolution time.}\label{f3}
\end{figure}
\begin{figure*}[htb]\centering
\centering
\includegraphics[width=0.75\linewidth]{fig4.eps}
\caption{(a)~Time evolution of $\log_{10}(1-F)$ for generating $N$-body GHZ states with $N$ from 10 to 60 at intervals of 5. (b) The fitting function and numerical scatters between the number of qutrits and evolution time with $99.9\%$ fidelity. (c) The fidelity of GHZ state against the coupling strength with disorder $\delta$ corresponding to 11, 18 and 25 qutrits. (d) The fidelity of GHZ state versus the varying $\delta/g_0$ and $g_0t/10^{3}$ for $11$-body GHZ state.}\label{f4}
\end{figure*}
\subsection{Shortest time for generating high-fidelity $N$-body GHZ states and robustness against disorder of coupling strengths}
In the protocol above, the $N$-body GHZ state is achieved through an excitation displacement of the zero-energy mode along the chain. The evolution time $T$ must be chosen as long as possible in order to obtain an adiabatic excitation displacement from one extremity to the other of the qutrit-resonator chain. The adiabaticity condition can be written as $\dot{\theta} \ll \Delta E$, where $\theta=\arctan(J_1/J_2)$ and $\Delta E$ is the energy gap between the zero-energy mode and the bulk modes~\cite{Mei2018,Felippo2020}. In the above numerical simulation, the evolution time is appropriate but not the shortest time of generating a high-fidelity~($0.999$) ideal GHZ state.
In order to get the information of the shortest time of generating $N$-body GHZ states, we numerically calculate the fidelity of ideal $N$-body GHZ states, and plot in Fig.~\ref{f4}(a) the time evolution of $\log_{10}(1-F)$ with $N$ from 10 to 60 at intervals of 5. The numerical results exhibit that a higher fidelity needs a longer evolution time with increasing $N$. For example, the evolution time for $10$-body GHZ state is $g_0t=661$ with $99.9\%$ fidelity. However, $60$-body GHZ state with $99.9\%$ fidelity requires $g_0t = 2.5\times10^4$. In practice, the value of $g_0$ can be chosen about $2\pi\times$ $50$~MHz~\cite{Mundada2019}. Thus, the evolution time of realizing $60$-body GHZ state is $T \approx 76~\mu $s. Up to now, the superconducting resonator lifetimes can be achieved between $1$ and $10$ ms~\cite{Reagor2013,Reagor2016,Axline2016}. And the decoherence time of the superconducting magnetic flux qubtrit achieved on $1$ ms by designing a $\pi$-phase difference across the Josephson junction in circuit has been reported~\cite{Pop2014}, which is far more than the evolution time $T$. As shown in Fig.~\ref{f4}(b), by selecting different $N$ and corresponding evolution times of generating GHZ states with the fidelity of $99.9\%$ as numerical samples, $g_0t$ versus $N$ can be fitted by a quadratic function $g_0t=6.9419N^{2}+2.455N-59.8933$. The evolution time should be larger enough to satisfy the adiabatic condition for a longer size of the chain.
Because of the existence of some circuit imperfections in a superconducting qutrit-resonator chain such as the mutual inductance and self-inductance of the circuit, and the unstable magnetic flux threading to the SQUID loop, which may cause variation in ideal couplings. In order to study the robustness of the scheme, we add a random disorder into coupling strength $J'_{1(2)}=J_{1(2)}(1+\rm{rand}~[-\delta,\delta])$ for each qutrit-resonator coupling, where $\rm{rand}~[-\delta, \delta]$ denotes a random number within the range of $[-\delta,\delta]$. Figure \ref{f4}(c) shows the relationship between the fidelity of GHZ state and the disorder $\delta \in [0,0.5]$ for 11, 18 and 25 qutrits. Note that the disorder is randomly sampled 101 times, and then the fidelity is taken as the mean value of the 101 results. Corresponding to a shorter size of the superconducting qutrit-resonator chain, the fidelity decreases more slowly and the GHZ state is more robust to the disorder of coupling strength. On account of the process of generating the GHZ state via the topological zero mode protected by the energy gap, the width of the gap in a superconducting qutrit-resonator chain usually exhibits an exponential decay behavior with the size of chain increasing~\cite{Chiu2016}. Furthermore, the disorder has little effect on the GHZ state with respect to $\delta/g_0 <[0,0.2]$. Even if $\delta/g_0 \in[0.2,0.5]$, the fidelity still keeps over $0.993$, which reflects the topological protection property of the system. As shown in Fig.~\ref{f4}(d), we plot the fidelity versus the varying evolution time and the disorder of coupling strength $\delta$ for 11 qutrits. We can learn that the damage to fidelity caused by coupling defects $\delta/g_0 \in[0.2,0.5]$ can be compensated by longer evolution time. Under the condition of $\{g_0t=830, \delta=0.5\}$, the fidelity of 11-body GHZ state can be achieved with fidelity over $0.995$. The robustness of $N$-body GHZ state to the disorder and perturbation provides much more convenience for the experimental realization and the practical application of multi-particle entanglement state.
\subsection{Influence of the losses in the superconducting qutrit-resonator chain}
We now give a discussion for the effect of qutrit-resonator losses on the fidelity of generating GHZ states. Two dominant channels are considered in the loss mechanism: (i) The loss of qutrits with decay rates $\gamma_{n}$; (ii) The loss of resonators with decay rates $\kappa_{n}$.
The effect of losses during the evolution time can be evaluated by using a conditional Hamiltonian~\cite{Pachos2002,Huang2018}
\begin{eqnarray}\label{e8}
H_{cond}&=&H-\frac{i\kappa_n}{2}\sum_{n=1}^{N-1}b^{\dagger}_{n}b_{n}-\frac{i\gamma_n}{2}\sum_{n=1}^{N}|e\rangle_{n}\langle e|,
\end{eqnarray}
where $H$ is the lossless Hamiltonian for the system in Eq.~(\ref{e1}). The second and third terms represent the losses of resonators and qutrits, respectively. For convenience, we assume that $\kappa_{n}=\kappa$ and $\gamma_{n}=\gamma$. We take $N=11$, 18, and 25 as examples. The fidelity is formulated as $F=\langle \Psi_{ideal}|\rho|\Psi_{ideal}\rangle$, where $|\Psi_{ideal}\rangle=1\ \sqrt{2}(|G\rangle-(-1)^{N}|r\rangle)$ is the output state of an ideal system, and $\rho$ is the density operator of the system dominated by the conditional Hamiltonian Eq.~(\ref{e8}).
We now numerically simulate the fidelity of ideal GHZ states by solving the non-Hermitian Liouville equation $\dot{\rho}=-i(H_{cond} \rho-\rho H^{\dagger}_{cond})$. Figure \ref{f5} shows the relationship between the fidelity $F$ and decay of qutrits or resonators. As the decay rate of qutrits or resonators increases, the fidelity of the ideal state shows a trend of decline. Comparing those lines, the size of chain increases, the fidelity decreases slightly. And the higher fidelities indicate that the Scheme A for generating $N$-body GHZ states is insensitive to the value of $\kappa$. The fidelity of the 11-, 18- or 25-body GHZ state plummets to $25\%$ roughly with $\gamma/g_0=0.01$. Evidently, the value of $\gamma$ has a greater influence on the fidelity. Owing to the distribution of edge states $|l_{\rm A}\rangle$ and $|r_{\rm A}\rangle$ closed to 0.5 at end of the evolution time in Fig.~\ref{f2}(d), the loss of qutrits in the excited state $|e\rangle$ in edge states during the whole evolution time causes a significant destructive effect on the fidelity of GHZ states. As shown in Fig.~\ref{f5}, we plot the fidelity of GHZ states without considering the losses in left and right edge states, i.e, setting $\gamma_1=\gamma_n=0$. Under the condition of $\gamma_n/g_0=0.01$ for $1<n<N$, the fidelities of 11-, 18- and 25-body GHZ states are $55.24\%$, $44.72\%$ and $33.92\%$, respectively, which indicates a notable improvement of the fidelities than before. Consequently, how to suppress the loss of excited state $|e\rangle$ of qutrits, especially of $A_1$ and $A_N$, is the key to enhancing the fidelity of GHZ states. In the next section, we propose two alternative schemes that are of improved robustness against losses of qutrits.
\begin{figure}
\includegraphics[width=\linewidth]{fig5.eps}
\caption{Fidelities for GHZ states versus decay rates of qutrits or resonators for different $N$, where $\gamma$~($\kappa$) in the legend represents decoherence involving only the qutrit~(resonator) decay. The evolution time $T$ is chosen as $g_0T=6.9416N^{2}+2.455N-59.8933$ that ensures a $99.9\%$ fidelity for generating a lossless $N$-body GHZ state.}\label{f5}
\end{figure}
\begin{figure*}[htb]\centering
\centering
\includegraphics[width=0.7\linewidth]{fig6.eps}
\caption{(a) and (b): Schematics of energy level transitions for qutirts $A_1$ and $A_N$ in Scheme B. (c) and (d): Schematics of energy level transitions for qutirts $A_1$ and $A_N$ in Scheme C. The interaction diagram of other cells is the same as Fig.~\ref{f2} in both Scheme A and B. The choice of coupling strengths and driving fields of last qutrits $A_N$ depends on the odevity of $N$.}\label{f6}
\end{figure*}
\section{Alternative schemes}\label{S4}
In the process of generating GHZ states, the distribution of excitations in two edge states is 0.5 at the start and end of the evolution time~(see Fig.~\ref{f3}). The losses of qutrits in the excited states have a great destructive effect on the final fidelities of GHZ states. In this section, we propose two improvement protocols, labeled Scheme B and Scheme C, to focus on suppressing the loss of qutrits in the excited state.
\subsection{Scheme B for suppressing excitation of qutrits $A_1$ and $A_N$}
The energy level space of a qutrit can be readily adjusted via changing the external flux applied to the SQUID loop~\cite{Neeley2008,Yang2010,You2005} or/and provided by an FBL~\cite{Schmidt2013}. As shown in Figs.~\ref{f6}(a) and (b), an extra long-lived state $|P\rangle$ is introduced into the right potential well of the qutrit $A_1$, upper than and separated enough from the ground state $|R\rangle$. The transition between $|e\rangle \leftrightarrow |L\rangle$ in qutrit $A_1$ is coupled off-resonantly to the resonator $B_1$ with coupling strength $J_1'$ and detuning $\Delta_1$. For the qutrit $A_N$ with $N$ being odd~(even), the transition between $|e\rangle\leftrightarrow|R\rangle~(|L\rangle$) is coupled off-resonantly to resonator $B_{N-1}$ with the coupling strength $J_2'$ and detuning $\Delta_2$. In addition, an auxiliary classical field is introduced to drive the transition off-resonantly between $|e\rangle \leftrightarrow |P\rangle$ for qutrit $A_1$ with Rabi frequency $\Omega_1$ and detuning $\Delta_1$. The other classical field with Rabi frequency $\Omega_2$ and detuning $\Delta_2$ drives the transition of $A_N$ off-resonantly, $|e\rangle \leftrightarrow |L\rangle~(|R\rangle$), with $N$ being odd~(even). The interaction Hamiltonians for the first and last qutrits coupled to their adjacent resonators can be written as
\begin{subequations}
\begin{equation}\label{e9a}
H_{1}=J_1'b_1|e\rangle_{1}\langle L|e^{i\Delta_{1}t}+\Omega_1|e\rangle_{1}\langle P|e^{i\Delta_{1}t}+\rm{H.c.},
\end{equation}
\begin{equation}\label{e9b}
H_{N}=J_2'b_{N-1}|e\rangle_{N}\langle m|e^{i\Delta_{2}t}+\Omega_2|e\rangle_{N}\langle n|e^{i\Delta_{2}t}+\rm{H.c.},
\end{equation}
\end{subequations}
with $m=R~(L)$ and $n=L~(R)$ when $N$ is odd (even), while the interaction Hamiltonian for other cells is identical with Eq.~(\ref{e1}).
For suppressing losses of qutrits in the excited state $|e\rangle$, the value of $\Delta_{1(2)}$ is supposed to be as large as possible to satisfy the condition $\Delta_{1(2)}\gg\{J_{1(2)}', \Omega_{1(2)}\}$. According to the theory of second-order perturbation~\cite{James2007}, one can eliminate adiabatically the excited states of qutrits $A_1$ and $A_N$. Besides, off-resonant-interaction-induced ground-state Stark shifts can be offset by introducing auxiliary off-resonant fields~\cite{Zhao2017}, phase compensations~\cite{Vepsaaineneaau5999}, or detuning compensations~\cite{Han:20}. Then, the Hamiltonians in Eqs.~(\ref{e9a}) and (\ref{e9b}) can be reduced into, respectively
\begin{subequations}
\begin{equation}\label{e10a}
H_{1\rm{eff}}=J_{1\rm{eff}}(|P\rangle_{1}\langle L|+|L\rangle_{1}\langle P|),
\end{equation}
\begin{equation}\label{e10b}
H_{N\rm{eff}}=J_{2\rm{eff}}(|R\rangle_{N}\langle L|+|L\rangle_{N}\langle R|),
\end{equation}
\end{subequations}
with $J_{1(2)\rm{eff}}=J_{1(2)}'\Omega_{1(2)}/\Delta_{1(2)}$, which involve solely long-lived states. When the chain with the size $L=2N-1=49$ is initially in the left edge state $|l_{\rm B}\rangle=|\psi_1\rangle=|PLR\cdots LR\rangle_{A_{25}}\otimes|000\cdots00\rangle_{B_{24}}$ under the interaction Hamiltonians (\ref{e9a}) and (\ref{e9b}), the system evolves in the finite space $\{|\psi_n\rangle\}$
\begin{equation}\label{e05}
\begin{split}
&|\psi_1\rangle=|PLR\cdots LR\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}},\\
&|\psi_2\rangle=|eLR\cdots LR\rangle_{A_{25}}\otimes|000\cdots 0\rangle_{B_{24}},\\
&|\psi_3\rangle=|LLR\cdots LR\rangle_{A_{25}}\otimes|100\cdots 0\rangle_{B_{24}},\\
&|\psi_4\rangle=|LeR\cdots LR\rangle_{A_{25}}\otimes|000\cdots 0\rangle_{B_{24}},\\
&\quad\vdots\\
&|\psi_{49}\rangle=|LRL\cdots RR\rangle_{A_{25}}\otimes|000\cdots 1\rangle_{B_{24}},\\
&|\psi_{50}\rangle=|LRL\cdots Re\rangle_{A_{25}}\otimes|000\cdots 0\rangle_{B_{24}},\\
&|\psi_{51}\rangle=|LRL\cdots RL\rangle_{A_{25}}\otimes|000\cdots 0\rangle_{B_{24}}.
\end{split}
\end{equation}
To further evaluate the topological properties of zero energy states in the Scheme B, we plot the distribution of the zero mode on component states $|\psi_n\rangle$, as shown in Fig.~\ref{f7}(a). The components $|\psi_1\rangle$ and $|\psi_{51}\rangle$, which play roles of left and right edge states, $|l_{\rm B}\rangle$ and $|r_{\rm B}\rangle$, respectively, are populated with the maximal distributions in the regions of $g_0t \in [0,1800)$ and $g_0t \in(2200,4000]$. In comparison between Figs.~\ref{f2}(d) and \ref{f7}(a), the systematic degree of freedom in the evolution subspace changes from $2N-1$ to $2N+1$ due to the introductions of classical drives on the two extremity qutrits of the chain. In the Scheme B, an extra long-lived state $|P\rangle$ is introduced in the qutrit $A_1$, and the large detuning condition is satisfied so as to eliminate adiabatically the excited state $|e\rangle$. In other words, the state transformations of Scheme A and Scheme B are, respectively
\begin{eqnarray*}
&|l_{\rm A}\rangle=|eLR\cdots R\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}} \nonumber\\
&\Downarrow\nonumber\\
&|r_{\rm A}\rangle=|LRL\cdots e\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}},
\end{eqnarray*}
and
\begin{eqnarray*}
&|l_{\rm B}\rangle=|PLR\cdots LR\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}}
\nonumber\\
&\Downarrow\nonumber\\
&|r_{\rm B}\rangle=|LRL\cdots RL\rangle_{A_{25}}\otimes|000\cdots 0\rangle_{B_{24}}.
\end{eqnarray*}
Thus, the distribution and spacing of bright and dark fringes in Fig.~\ref{f7}(a) are different from that in Fig.~\ref{f2}(d). The dark fringes in Fig.~\ref{f7}(a) appearing on $|\psi_2\rangle$ and $|\psi_{50}\rangle$ indicate that the population elimination of the excited state $|e\rangle$ in qutrits $A_1$ and $A_{25}$.
\begin{figure*}
\centering
\includegraphics[width=0.7\linewidth]{fig7.eps}
\caption{Distributions of the zero energy mode for (a)~Scheme B and (b)~Scheme C, respectively, on component states $|\psi_n\rangle$ and $|\phi_n\rangle$. The size of the superconducting qutrit-resonator chain is $L=2N-1=49$.}\label{f7}
\end{figure*}
For sake of generating large-scale GHZ states, the superconducting qutrit-resonator chain is assumed initially in the state
\begin{equation}\label{e11}
|\Phi'_0\rangle=\frac{1}{\sqrt{2}}\Big(|G\rangle+|l'_{\rm B}\rangle\Big).
\end{equation}
where $|l'_B\rangle=|PLR\cdots m\rangle_{A_{N}}\otimes|000\cdots0\rangle_{B_{N-1}}$ and $m=R~(L)$ when $N$ is odd~(even). After the evolution along the topologically protected zero-energy mode, similar to the Scheme A, one can obtain the final state
\begin{equation}\label{e12}
|\Phi'_{ideal}\rangle=\frac{1}{\sqrt{2}}\Big(|G\rangle-(-1)^{N}|r'_B\rangle\Big).
\end{equation}
in which $|r'_B\rangle=|LRL\cdots n\rangle_{A_{N}}\otimes|000\cdots 0\rangle_{B_{N-1}}$ with $n=L~(R)$ when $N$ is odd~(even). Here $|R\rangle$ is used to carry the logical state 1 while $|L\rangle$ carries the logical state 0 for all qutrits. Accordingly, the initial state without considering the zero-photon resonators after the evolution becomes
\begin{equation}\label{e7}
\frac{1}{\sqrt{2}}\left(|101\cdots 1(0)\rangle_{A_{N}}-(-1)^{N}|010 \cdots 0(1)\rangle_{A_{N}}\right).
\end{equation}
which is exactly an $N$-body GHZ state.
In Fig.~\ref{f8}(a), with the same parameters as in Fig.~\ref{f3} we numerically plot the time evolution of populations for the ideal state $|\Phi_{ideal}\rangle=\frac{1}{\sqrt{2}}(|RLR\cdots R\rangle_{A_{25}}+|LRL\cdots L\rangle_{A_{25}})\otimes|000\cdots0\rangle_{B_{24}}$, the initial state $|\Phi_{0}\rangle=\frac{1}{\sqrt{2}}(|RLR\cdots R\rangle_{A_{25}}+|PLR\cdots R\rangle_{A_{25}})\otimes|000\cdots0\rangle_{B_{24}}$, and two edge states $|l_{\rm B}\rangle$ and $|r_{\rm B}\rangle.$ Obviously, the population of $|\Phi_{ideal}\rangle$ ($|\Phi_0\rangle$) reaches nearly 1 (0.25) and keeps steady at the end of evolution time, which proves the feasibility of Scheme B. The populations of two edge states have the identical trend with Fig.~\ref{f3}.
In order to verify the effectiveness of suppressing the losses of qutrits in the excited state, we simulate the fidelity of the ideal GHZ state $|\Phi_{ideal}\rangle$ by solving the non-Hermitian Liouville equation. In Fig.~\ref{f8}(b), the fidelities for 11-, 18-, and 25-body GHZ states hold on $98.3\%$, $95.8\%$, and $93.2\%$ even though $\kappa/g_0=0.05$. In addition, it is evident that the fidelity in Fig.~\ref{f8}(b) can reach values similar to that without taking account of edge states in Fig.~\ref{f5}. When $\gamma/g_0=0.01$, the fidelities of 11-, 18-, and 25-body GHZ states maintain $53.2\%$, $45.2\%$, and $33.6\%$, respectively. With regard to $\gamma$, the fidelity of the ideal state decreases more slightly and displays a more robust result than Scheme A because of the suppression of excitation populations in qutrits $A_1$ and $A_N$ during the whole evolution.
\begin{figure*}[htb]
\centering
\includegraphics[width=0.7\linewidth]{fig8.eps}
\caption{(a) and (c) for Scheme B and Scheme C, respectively: Time evolution of population for $|\Phi_0\rangle$, $|\Phi_{ideal}\rangle$, $|\varphi_0\rangle$, $|\varphi_{ideal}\rangle$, edge states $|l_{\rm B(C)}\rangle$ and $|r_{B(C)}\rangle$ for the size of superconducting qutrit-resonator chain $L=2N-1=49$. (b) and (d) for Scheme B and Scheme C, respectively: Fidelity for the GHZ states versus the decay rate $\gamma$ or $\kappa$ of the superconducting qutrit-resonator chain for different number of qutrits, where $\gamma$~($\kappa$) in the legend represents decoherence involving only the qutrit~(resonator) decay. Parameters for (a) and (b): $J'_{1(2)}=20J_{1(2)}$, $\Delta_{12}=400g_0$, and $\Omega_{1(2)}=20g_0$.}\label{f8}
\end{figure*}
\subsection{Scheme C for suppressing excitation of all qutrits}
As shown in Figs.~\ref{f6}(c) and (d), for the qutrit $A_1$, the transition between $|e\rangle \leftrightarrow |L\rangle$ is coupled resonantly to the resonator $B_1$ with coupling strength $J_{1}$. When $N$ is odd (even), the transition between $|e\rangle\leftrightarrow|R\rangle~(|L\rangle$) in the last qutrit $A_N$ is coupled resonantly to the resonator $B_{N-1}$ with the coupling strength $J_{2}$. Also, one classical field drives the transition resonantly between $|e\rangle \leftrightarrow |P\rangle$ for qutrit $A_1$ with Rabi frequency $\tilde{\Omega}_{1}$. Under the condition of $N$ being odd (even), the other classical field with Rabi frequency $\tilde{\Omega}_{2}$ drives resonantly the transition $|e\rangle \leftrightarrow |L\rangle ~(|R\rangle)$ in qutrit $A_N$. Therefore, the interaction Hamiltonian involving the first and last qutrits can be written as
\begin{equation}\label{e13}
\begin{split}
H_{1'}&=J_{1}b_1|e\rangle_{1}\langle L|+\tilde{\Omega}_{1}|e\rangle_{1}\langle P|+\rm{H.c.},\\
H_{N'}&=J_{2}b_{N-1}|e\rangle_{N}\langle m|+\tilde{\Omega}_{2}|e\rangle_{N}\langle n|+\rm{H.c.},
\end{split}
\end{equation}
where $m=R~(L)$ and $n=L~(R)$ when $N$ is odd~(even). And the interaction Hamiltonian of other cells still keeps consistent with Eq.~(\ref{e1}). The shapes of the coupling strengths are engineered as Gauss functions
\begin{equation*}
\begin{split}
\tilde{\Omega}_{1}=J_2=g_0\exp{[-(t-3\tau)^2/\tau^2]},\\
\tilde{\Omega}_{2}=J_1=g_0\exp{[-(t-2\tau)^2/\tau^2]},
\end{split}
\end{equation*}
where $J_1$ and $J_2$ are reverse with respect to Eq.~(\ref{e5}).
We take $N=25$ as an example, and when the chain is initially in the left edge state $|l_{\rm C}\rangle=|\phi_1\rangle=|PLR\cdots R\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}}$. The system evolves in the finite subspace $\{|\phi_n\rangle\}$
\begin{equation}\label{e14}
\begin{split}
&|\phi_1\rangle=|PLR\cdots R\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}},\\
&|\phi_2\rangle=|eLR\cdots R\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}},\\
&|\phi_3\rangle=|LLR\cdots R\rangle_{A_{25}}\otimes|100\cdots0\rangle_{B_{24}},\\
&|\phi_4\rangle=|LeR\cdots R\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}},\\
&\quad\vdots\\
&|\phi_{49}\rangle=|LRL\cdots R\rangle_{A_{25}}\otimes|000\cdots1\rangle_{B_{24}},\\
&|\phi_{50}\rangle=|LRL\cdots e\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}},\\
&|\phi_{51}\rangle=|LRL\cdots R\rangle_{A_{25}}\otimes|000\cdots0\rangle_{B_{24}}.
\end{split}
\end{equation}
Then we plot the distribution of the zero energy mode on component states $|\phi_n\rangle$ in Fig.~\ref{f7}(b). Distinctly, both the two ends of evolutionary states $|\phi_1\rangle$ and $|\phi_{51}\rangle$ are populated with the maximal distributions in the regions of $g_0t \in [0,1800)$ and $g_0t\in(2200,4000]$, respectively. In comparison between Figs.~\ref{f2}(d) and \ref{f7}(b), the system degree of freedom in the evolution subspace changes from $2N-1$ to $2N+1$ due to the introductions of classical drives on the two extremity qutrits of the chain. In Fig.~\ref{f7}, the distribution and spacing of bright and dark fringes are different from that of Scheme A or Scheme B. The bright fringes in Scheme A or Scheme B are distributed in the excitations of qutrits, while in Scheme C they are distributed in the excitations of resonators. Therefore, in Scheme C excitations of all qutrits are suppressed.
So as to obtain a large-scale GHZ states, the initial state of the chain is prepared in Eq.~(\ref{e11}). The first state component $|G\rangle$ does not evolve, while the second state component $|l'_{B}\rangle$ evolves into $(-1)^{N}|r'_{B}\rangle$. Hence, the final state is
\begin{eqnarray}
\begin{split}\label{e15}
|\Psi_F\rangle=&\frac{1}{\sqrt{2}}\left(|G\rangle+(-1)^{N}|r'_{B}\rangle\right).
\end{split}
\end{eqnarray}
which has a minus sign difference from Eq.~(\ref{e12}). In Fig.~\ref{f8}(c), we plot the time evolution of populations for the initial state $|\varphi_{0}\rangle=|\Phi_0\rangle$ , the ideal state $|\varphi_{ideal}\rangle=\frac{1}{\sqrt{2}}(|RLR\cdots R\rangle_{A_{25}}-|LRL\cdots L\rangle_{A_{25}})\otimes|000\cdots0\rangle_{B_{24}}$, two edge states $|l_{C}\rangle=|l'_{B}\rangle$ and $|r_{\rm C}\rangle=|r'_{B}\rangle$ for generating a 25-body GHZ state. As expected, the population of $|\varphi_{ideal}\rangle$ ~($|\varphi_{0}\rangle$) is close to unity (0.25) at the time $g_0t=3600$ and remains stable. As for the two edge states, the population in Scheme C has the same climate as that in Scheme A and Scheme B. The result reveals that the above theoretical analysis is correct and feasible. Figure \ref{f8}(d) shows the relationship between the fidelity and the decay rate $\gamma$ or $\kappa$. As the decay rates increase, the fidelity of the ideal GHZ state decreases. As opposed to the loss mechanisms of Scheme A and Scheme B, the decay rate $\kappa$ of resonators has a greater influence on the fidelity. When $\kappa/g_0=0.01$, the fidelity is reduced to $56.1\%$ for the 11-body GHZ state. In contrast among the six lines, the fidelity is relatively immune to the decay of qutrits. The fidelity can retain at $95.9\%$, $95.0\%$, and $93.6\%$ with $\gamma/g_0=0.05$, which indicates that Scheme C displays the most robust performance to resist the losses of qutrits in the superconducting qutrit-resonator chain among the three schemes.
In Scheme C, the fidelity of the ideal large-scale GHZ state is affected least by the losses of qutrits, compared with Scheme A and Scheme B. However, the cost is that the losses of resonators have a greatest influence on the fidelity, which indicates that the Scheme C requires resonators to hold a long coherence time. While according to the interaction of Scheme A and Scheme B, it is required that the qutrits have a long coherence time, but without a strict requirement on the quality factor of resonators. Thus, corresponding to the performance of experimental devices in the superconducting qutrit-resonator chain, we can choose different schemes to realize large-scale GHZ states with high fidelity in the case of loss mechanisms by adjusting the energy level structure of qutrits and coupling strengths.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{fig9.eps}
\caption{Fidelity of the GHZ state with the scale of entanglement, $N$, under the feasible coupling strengths (e.g., $g_0/2\pi=50$ MHz, 10 MHz and 1 MHz) and coherence times of qutrits and resonators in Scheme A. We choose coherence time of the qutrits and resonators as $\tau_a=\tau_b=1$~ms.}\label{f9}
\end{figure}
\section{Scale of GHZ states}\label{S5}
In experiment, by designing a $\pi$-phase difference across the Josephson junction in circuit to restrain the energy relaxation induced by quasiparticle dissipation, one can obtain a qutrit with coherence time over 1~ms~\cite{Pop2014}. As for a resonator, the coherence time of the photons in the resonator can be much longer~\cite{GU2017}. Up to now, the superconducting resonator lifetimes between 1~ms and 10~ms have been reported~\cite{Reagor2013,Reagor2016,Axline2016}. Generally, the typical feasible coupling strength can be modulated in the range of 1 MHz to 50 MHz~\cite{Mundada2019}, providing a considerable adjustability in experiment.
In this section, we discuss an accessible GHZ state scale $N$ by taking into account the experimentally available coherence times of qutrits and resonators under the condition of feasible coupling strengths. For convenience, we take Scheme A as an example. Figure \ref{f9} shows the relationship between fidelities of GHZ states and the scale of entanglement $N$ with feasible $g_0$ and coherence times of qutrits ($\tau_a=1/\gamma=1$~ms) and resonators $\tau_b=1/\kappa=1$~ms). Obviously, the fidelity exhibits the decreasing tendency with increasing $N$. Regarding to feasible coupling strengths, the fidelity declines at different rates. Under the condition of $g_0/2\pi=50$~MHz, fidelities of 10- to 50-body GHZ states keep above $90\%$. Even if the scale of a GHZ state is $N=150$, its fidelity would still reach $50.60\%$. As for $g_0/2\pi=10$~MHz, the fidelity of 67-body GHZ state stays above $50\%$. The fidelity can be over $90\%$ with $N$ less than 23. However, the GHZ state fidelity has the fastest rate of decline in the case of $g_0/2\pi=1$~MHz, where the fidelity of a GHZ state can achieve $50\%$ with the maximum entanglement scale $N=20$. Under the existing experimental conditions, the higher fidelity of large-scale GHZ states can be generated with the higher value of coupling strengths.
\section{Experimental consideration and prospective improvement}\label{S6}
\subsection{Device and initial state}
Benefiting from the rapid development in circuit-QED technologies, the circuit-QED system provides us an excellent experimental platform to realize large-scale GHZ states proposed in our work. We can construct a circuit-QED system via arranging the transmission line resonator and the superconducting qutrits in the space, as shown in Fig.~\ref{f1}(b). According to the existing circuit-QED technology, the qubit-resonator chain systems with on the order of 10–20 qubits have been demonstrated~\cite{Kelly2015,Kandala2017,Neill195,Otterbach2017,Verifying2020,Mooney2019}. Simultaneously, a chain of 72 superconducting resonators coupled via transmons can be realized ~\cite{Fitzpatrick2017}. The scale of qutrit-resonator chain manufacturing in a metal chip is generally from micron to millimeter~\cite{Majer2009,Hybrid2020,Circuit2010,Quantum2020}. In general, the flux qutrit is a superconducting circuit made up of Josephson junctions or/and capacitance~\cite{Manucharyan113}. While the flux qutrit can be operated at any applied external flux through the flux qutrit loop. The resonator is composed by a linear inductance in parallel with a capacitance, which can be fabricated from a NbN film deposited on a sapphire substrate~\cite{Fast2020}. The coupler is replaced by two Josephson junctions with a SQUID loop to realize tunable coupling: the magnetic flux that threads this loop determines its effective inductance $E_{J}(\phi)\sim E_{J}(0)\cos(2\pi\phi/\Phi_0)$~\cite{Garc2020}. Josephson junctions can be designed by the metallic chip (Al/AlO$_{x}$/Al) using the electron-beam lithography~\cite{Junction2011,Circuit2010}. In experiment, the circuit-QED device is usually operated in a single-shot liquid $^3$He with a base millikelvin temperatures and heavily filtered cryogenic microwave lines~\cite{Majer2009,Quantum2020}.
In the above schemes, specified initial states are needed to prepare in Eqs.~(\ref{e6}) and (\ref{e11}) to realize $N$-body GHZ states. These specified initial states denote $A_1$ in the superposition $(|R\rangle+|e(P)\rangle)/\sqrt{2}$, $B_1$ in $|0\rangle$, $A_2$ in $|L\rangle$, $\cdots$, $A_N$ in $|R\rangle$ ($|L\rangle$) when $N$ is odd (even). In experiment, the preparation of the initial states can be conducted through two steps: (i) As reported recently in Ref.~\cite{Driving2018}, the qutrit holding a $\Lambda$-type structure can be cooled to one of its two lowest energy eigenstates ($|R\rangle$ or $|L\rangle$) by resonator decay through spontaneous Raman scattering so as to obtain the state $|RLR\cdots R(L)\rangle_{A_{N}}$. (ii) A $\pi/2$ pulse is applied to the qutrit $A_1$ to create a superposition state of $(|R\rangle + |e(P)\rangle)/\sqrt{2}$~\cite{Experimental2019}. The resonator keeps in the zero photon Fock state. Finally, the qutrit-resonator chain can be prepared in $1/\sqrt{2}(|RLR\cdots R(L)\rangle_{A_N}+|P(e)LR\cdots R(L)\rangle_{A_N})\otimes|000\cdots0\rangle_{B_{N-1}}$.
\subsection{Tunable couplings}
In order to achieve high-fidelity GHZ states, tunable couplings among qutrits and resonators are required. Tunable couplers of both varieties have been realized in several experiments: between two tunable qubits~\cite{Chen2014,Harrabi2009}, between a qubit and a resonator~\cite{Allman2014,Srinivasan2011}, and between resonators~\cite{Wang2011,Zakka2011}. Particularly, a direct tunable coupler is realized by a tunable circuit element between the qutrit and the resonator, e.g., a flux-biased direct-current SQUID to generate strong resonant and nonresonant tunable interactions between a qubit and a lumped-element resonator~\cite{Allman2014}.
In this work, we adopt a direct tunable coupler replaced by SQUID between qutrits and resonators. The coupling strengths $J_1$ and $J_2$ can be tunable through adopting controlled voltage pulses generated by an AWG to tune the flux threading the SQUID loop~\cite{Majer2009}. For example, the resonator $B_{N}$ is coupled resonantly to transitions $|e\rangle \leftrightarrow |m\rangle$ ($m \in \{R,L\}$). The coupling strength $J_{1(2)}$ can be expressed by~\cite{Possible2003,Quantum2004}
\begin{equation}\label{e16}
J_{1(2)}=\frac{1}{L}\sqrt{\frac{\omega_{B_{N}}}{2\mu_{0}\hbar}}\langle e|\Phi|m\rangle \int_{S} \textbf{B}_{B_{N}}(\vec{r},t)\cdot d\textbf{S},
\end{equation}
where $\textbf{S}$ is the surface bounded by the loop of SQUID, $\omega_{B_{N}}$ the resonator frequencies of $B_{N}$. Accordingly, $\textbf{B}_{B_{N}}(\vec{r},t)$ is the the magnetic components of the $B_{N}$ resonator mode. For a standingwave resonators, $\textbf{B}_{B_{N}}(\vec{r},t)=\mu_{0}\sqrt{2/V_N}\cos k_{N}z_{N}$ ($k_N$ is the wave number of $B_{N}$ resonator, $V_N$ and $z_N$ are the $B_{N}$ resonator volume and the $B_{N}$ resonator anxis). In this case, a modulating field can be added when a qutrit works on its optimal frequency point~\cite{Reagoreaao3603}, and thus does not cause the shortening of qutrit coherent times.
\subsection{Accelerating adiabatic process}
Over the past decade, techniques of shortcuts to adiabaticity~(STA)~\cite{Chen2010,Odelin2019} receive a lot of attention, because STA can accelerate adiabatic processes but remain the robustness of adiabatic processes. Recently, a fast quantum state transfer from the left edge state to the right edge state in a topological SSH chain with next-to-nearest-neighbor~(NNN) interaction was presented, which provides the simplest one-dimensional lattice with protected edge state~\cite{Felippo2020}. The idea of this approach is based on an engineering of NNN interactions between the sites of the chain, which exactly cancels nonadiabatic couplings due to an imperfect condition of adiabatic evolution. Inspired by counterdiabatic driving methods~\cite{Berry2009,Demirplak2003}, the quantum chain without limitation of the adiabaticity constraint, combined with a dynamical control of NNN interactions, is thus governed by the following Hamiltonian $H(t)$:
\begin{eqnarray}\label{e17}
H(t)&=&H_0(t)+H_c(t),\nonumber\\
H_0(t)&=&\sum_{n=1}^{N-1}t_2(t)|B_n\rangle \langle A_n|+t_1(t)|A_{n+1}\rangle \langle B_n|+\rm{H.c.},\nonumber\\
H_c(t)&=&\sum_{n=1}^{N-1}i\alpha_{n}(t)|A_{n+1}\rangle \langle A_{n}|+\rm{H.c.},
\end{eqnarray}
where $H_c(t)$ is the control Hamiltonian literally canceling the nonadiabatic couplings. The time-dependent NNN hoppings $\alpha_n(t)$ is only to cancel nonadiabatic transitions from the time-dependent eigenvector $|\phi_0(t)\rangle=\mathcal{N}\sum^{N}_{n=1}(\frac{-t_2}{t_1})^{n-1}|A_n\rangle$, where $\mathcal{N}$ is the normalization constant.
Analogously, with the help of such a method for speeding up the adiabatic state transfer in the SSH model, it is possible to find a control Hamiltonian, added to the initial Hamiltonian, that literally cancels the nonadiabatic couplings in the topological model. Therefore, it is of great potential to realize fast preparation of a large-scale GHZ state by engineering the NNN interaction between qutrits, which may constrain the impact of systematic decoherence and thus enhance the fidelity of generating large-scale GHZ states.
In addition, an alternative method, Floquet-engineering STA~(FESTA)~\cite{Petiziol2018,Claeys2019}, may also be applied in speeding up the generation of a large-scale GHZ state. FESTA is an effective succedaneum of counterdiabatic driving methods, which does not require an additional control Hamiltonian but is combined with a periodic driving component to oscillate the initial Hamiltonian~\cite{Petiziol2018}. This Floquet-engineering periodic driving component will form an effective connection between transferred two states so as to offset exactly nonadiabatic couplings~\cite{Petiziol2020}. FESTA has been demonstrated for quantum state transfer in spin chains in a recent experiment~\cite{Peng2019SB}. Therefore, it is of great potential to use FESTA in our model for accelerating the adiabatic generation of GHZ states by replacing the adiabatic pulses with the Floquet-engineering oscillating pulses.
\section{Conclusion}\label{S7}
In conclusion, we have proposed a model of a superconducting qutrit-resonator chain, and the topological edge state with zero energy is analytically derived. Along this topological zero-energy mode, a state transfer from one extremity qutrit to the other of the chain can be implemented, accompanied with state flips of all intermediate qutrits. Three schemes are shown for generating large-scale GHZ states that are protected by the topological zero-energy mode and thus hold great robustness against disorder of the qutrit-resonator coupling strengths. The effect of losses induced by decay of qutrits and resonators on entanglement fidelity is investigated, and the results indicate that the three schemes meet different performance requirements of experimental devices. Furthermore, we study the accessible entanglement scale when taking into account the experimentally available coherence times of qutrits and resonators, and find that with the maximum of the qutrit-resonator coupling strengths $g_0/2\pi=50$~MHz, it is possible to achieve a $50$-body GHZ state with the fidelity $F>0.9$, and a $150$-body GHZ state with $F>0.5$. Finally, we discuss the experimental consideration of generating GHZ states, including the physical implementation of the model, preparation of the initial states, and tunable couplings in circuit-QED chain, and also show the potential to accelerate adiabatic process of generating GHZ states. The present work is expected to be helpful for promoting the experimental improvement of large-scale GHZ states.
\section*{ACKNOWLEDGEMENTS}
This work was supported by National Natural Science Foundation of China (NSFC) (Grant No.11675046), Program for Innovation Research of Science in
Harbin Institute of Technology (Grant No. A201412), and Postdoctoral Scientific Research Developmental Fund of Heilongjiang Province (Grant No. LBH-Q15060).
|
\section*{Disclaimer}
\vskip 3mm
The views expressed herein are those of the author and do not reflect the views of my current employer, Wells Fargo Securities, or affiliate entities.
\section{Lagrangian Formulation}\label{CLASSA}
In the Group Quantization formalism,
$ \Theta \rvert_{\mathcal{C}}$, the projection onto the base manifold of the vertical form $\Theta$ along the integral trajectories of the
characteristic module $C_\Theta$ provides {\footnote{
If the quantization group $\tilde{G}$ is finite, the quotient space $P = \tilde{G}/\mathcal{C}_\Theta$ and the quotient connection form $\Theta_P = \Theta /\mathcal{C}_\Theta$,
where $\mathcal{C}_\Theta$ is the characteristic module of $\Theta$,
define a fiber bundle $(P,U, \pi)$ where the curvature form $\Omega = d \Theta_P$ is a symplectic form over $P/U$. Taking the quotient by the integral flows of $\mathcal{C}_\Theta$ allows the definition of
local coordinates in $P/U$, $(p_i,q_i), i=1,2 \ldots \frac{1}{2} dim(P/U)$ representing the canonical momenta and canonical coordinates.
of the vertical form $\Theta$ along the integral flows (trajectories) of the
characteristic module $C_\Theta$.}}
the classical action ${\mathcal{S}}$ (\cite{aldaya0} )
\vskip 1mm
\begin{equation*}
{\mathcal{S}} = \int \Theta \rvert_{\mathcal{C}} = \int {\mathcal{L}} (x, \dot{x}) d t
\label{act1}
\end{equation*}
The classical Lagrangian ${\mathcal{L}} (x, \dot{x})$ is obtained by the projection onto the base manifold. Note that $\Theta$ and $ \Theta + d f$ generate equivalent Lagrangians that differ in a total derivative and do not change the action ${\mathcal{S}}$
\vskip 1mm
The vertical (connection) form $d \Theta$ restricted to the quotient space ${\tilde{G}}/U$
is analogous to the {\it{Poincar{\'e}-Cartan}} form of classical mechanics $\Theta_{PC}$ ,which can
be written in local coordinates ${\bf{q}}, {\bf{p}}$ as
\begin{equation}
\Theta_{PC} = \sum_i p_i d q_i - H({\bf{p}}, {\bf{q}}) dt
\label{mod1bb4pb}
\end{equation}
where $H({\bf{p}}, {\bf{q}})$ is the Hamiltonian function, and ${\bf{q}}, {\bf{p}}$ are conjugate pairs of coordinates and momenta. In classical mechanics,
the equations of motion are just the flows associated to the Hamiltonian field
\begin{equation*}
X_H \equiv \frac{\partial}{\partial t} + \sum_i \frac{\partial H}{\partial p_i} \frac{\partial}{\partial q_i} - \frac{\partial H}{\partial q_i} \frac{\partial}{\partial p_i}
\label{mod1bb66pb}
\end{equation*}
so along Hamiltonian trajectories
\begin{equation*}
{\dot{q_i}} \equiv \frac{d q_i}{ d t } = \frac{\partial H}{\partial p_i} \qquad {\dot{p_i}}\equiv \frac{d p_i}{d t } = - \frac{\partial H}{\partial q_i}
\label{mod1bb6ppp}
\end{equation*}
Note that $\Theta_{PC}$ has the structure of a Legendre transformation
if one identifies $p_i \equiv \partial {\mathcal{L}} / \partial {\dot{q_i}}$, where ${\mathcal{L}}$ is the Lagrangian function. Along the Hamiltonian flows
\begin{equation*}
\Theta_{PC} \rvert_{H}= \mathcal{L}({\dot{{\bf{q}}}}, {\bf{q}}) dt
\label{mod1bb4ppk}
\end{equation*}
The differential of $ \Theta \rvert_{PC}$ along the Hamiltonian trajectories is a symplectic form.
\begin{equation*}
d \Theta_{PC} \rvert_{H} = \sum_i d p_i \wedge d q_i
\label{mod1bb4}
\end{equation*}
The Lagrangian function is used for the application of {\it{path integral}} methods in computational finance ( \cite{baaquie1}, \cite{baaquie4}, \cite{baaquie3}, \cite{linetsky}).
\vskip 1mm
\subsection{Black-Scholes Lagrangian}\label{LAGBS}
We find the Black Scholes Lagrangian by restricting the Black-Scholes connection $\Theta$ \eqref{theta2bs} to the characteristic module trajectories on the base manifold.
\vskip 1mm
$\mathcal{C}_\Theta$ is generated by the time translations ${X_{t}^L}$ whose integral flows on the base manifold are (see \eqref{mod2bbbs})
\begin{align}
& \frac{dt} { ds} = 1 \quad \frac{dp} { ds} = 0 \quad \frac{dx} { ds} = \sigma^2 p + \mu \nonumber \\
& \rightarrow \quad t = s \quad p = p_0 \quad x = x_0 + (\sigma^2 p_0 + \mu )\, t
\label{mod3bs}
\end{align}
where $x_0$, $p_0$ , $x_0$, $\zeta_0$ are integration constants and $s$ is the integration parameter.
\vskip 1mm
We first add the total differential $d (p x) $ to $\Theta$ so the equations adopt the Poincar{\'e}-Cartan expression \eqref{mod1bb4pb}
\begin{equation}
\Theta \rightarrow d(x p) + \Theta = p \, dx - E(p) \, dt
\end{equation}
then
\begin{align*}
\Theta \rvert_{\mathcal{C}} & = p (\sigma^2 p + \mu) d t - \left( \frac{1}{2 } \sigma^2 p^2 + \mu p -r \right) d t = \nonumber \\
& \frac{1}{2 \sigma^2} ( \dot{x} - \mu)^2 dt - r dt
\label{theta4bs}
\end{align*}
where we have used equation \eqref{mod3bs}
\begin{equation*}
\frac{dx} { dt} \equiv \dot{x} = \sigma^2 p + \mu \rightarrow p = \frac{ \dot{x} - \mu}{\sigma^2}
\end{equation*}
with the result
\begin{equation}
\mathcal{L} (x, \dot{x}) = \frac{1}{2 \sigma^2} ( \dot{x} - \mu)^2 - r
\label{theta5bsa}
\end{equation}
The solution $x(\tau)$ of the Lagrangian \eqref{theta5bsa} coincides with the expected value of the coordinate (log-price) using the
$K_{BS}$ kernel in \eqref{kernellag1})
\begin{equation}
\langle \hat{x} \rangle = \, \frac{1}{\sqrt{2 \pi \sigma^2 \tau}} \, \int _{-\infty}^{ \infty } \, d x \, x \, \ e^{ -\frac{1}{2 \sigma^2 \tau } \, (x^\prime-x -\mu \tau)^2} = x^\prime + \mu \tau
\label{kernelBS1}
\end{equation}
\vskip 1mm
\subsection{Euclidean Oscillator Lagrangian}\label{LAGHOH}
\vskip 1mm
Following the same steps than in section \ref{LAGBS}, and using the expressions \eqref{theta2bs} and \eqref{energyHO}, we find that the $\mathcal{C}_\Theta$ trajectories
\begin{equation}
\frac{dt} { ds} = 1 \quad \frac{dp} { ds} = \Omeg \, \lambda^2 x \quad \frac{dx} { ds} = \Omeg \, \lambda^{-2} p
\end{equation}
generates the following Lagrangian for the Euclidean harmonic oscillator
\begin{equation}
\mathcal{L}_{H} (x, \dot{x}) = \frac{1}{2\sigma^2}\, \left( {\dot{x}}^2 + \frac{1}{2} \Omeg^2 \, x^2\right)
\end{equation}
\vskip 1mm
As expected, the mapping $\Omeg \rightarrow i \Omeg$ provides the Lagrangian for the Euclidean repulsive oscillator
\begin{equation}
\mathcal{L}_{R} (x, \dot{x}) = \frac{1}{2\sigma^2}\, \left( {\dot{x}}^2 - \frac{1}{2} \Omeg^2 \, x^2 \right)
\end{equation}
\section{Black-Scholes}\label{BSQUANT}
\vskip 1mm
\subsection{Quantization Group}
\vskip 1mm
We apply the Group Quantization formalism to the
$WSp(2, \mathbb{R})$ subgroup generated by the parabolic $SL(2, \mathbb{R})$ matrix
\begin{equation}
{\bf{M}}_{BS}= \begin{bmatrix}
1 & \, {\sigma^2} t\\
0 & 1
\end{bmatrix} \qquad \sigma^2 \in {\mathbb{R}} \qquad t \in {\mathbb{R}}
\label{genmatrixBSa}
\end{equation}
\vskip 1mm
${\bf{M}}_{BS}$ has only one eigenvalue $\lambda = +1$ and it cannot be diagonalized.
${\bf{M}}_{BS}$ is a {\it{shear}} transformation on $\mathbb{R}^2$: it leaves the upper plane invariant, while it displaces the lower plane by $ {\sigma^2} t$. This will be significant
when finding a first order polarization in section \ref{FPO}.
\vskip 1mm
\subsection*{Composition Law}
The $WSp(2, \mathbb{R})$ subgroup generated by the matrix \eqref{genmatrixBSa} constitutes
the Black-Scholes quantization group $\tilde{B}$. Its
composition law is obtained from the $WSp(2, \mathbb{R} )$ composition law
\eqref{slr2a} and the matrix ${\bf{M}}_{BS}$, with some cocycle modifications that we will explain below{\footnote{With no modifications, the Group Quantization formalism leads to the heat equation.}}
\begin{align}
t^{\prime \prime} & =t + t{^\prime} \nonumber \\
p^{\prime \prime} & = p^{\prime} + p \nonumber \\
x^{\prime \prime} & =x^{\prime } + x + \sigma^2 p^{\prime} t \nonumber \\
\zeta^{\prime \prime} & =\zeta^{\prime } \zeta e^{ \epsilon_{BS}(g,g^\prime) }
\label{cocyclebs1a}
\end{align}
where $ t,p,x\in {\mathbb{R}}$ and $\zeta \in {\mathbb{R}}^{+}$
When deriving the group law, we have used that
\begin{equation*}
{\bf{M}}_{BS}^{\prime \prime} = {\bf{M}}_{BS}^{\prime}\, {\bf{M}}_{BS} \quad \Rightarrow \quad t^{\prime \prime} =t + t{^\prime}
\end{equation*}
\vskip 1mm
The composition law for $(t,p,x)$ gives the familiar Galilean transformations: the Galilei group $G$ is the base group of $\tilde{B}$, $G \simeq \tilde{B}/ {\mathbb{R}}^{+} $,
with $\sigma^2 p$ corresponding to the Galilean boost..
\vskip 1mm
Our interpretation of the group coordinates is that $x$ represents the (dimensionless) logarithm of the stock price, $S \equiv S_0 \, e^x$,
$p$ is the conjugate momentum of $x$, $t$ is the calendar time, and
$\sigma$ is the stock volatility. Notice that interest rates are not explicitly present in the group law. Interest rates in the Black-Scholes theory are not dynamical quantities and they
will be incorporated in section \eqref{IRDEF} as a coordinate change in the fiber $\zeta$.
\vskip 1mm
The extension cocycle $\epsilon_{BS}$
is the sum of the true cocycle{\footnote{ All true Galilean cocycles with the same $\sigma$
differ in a coboundary from the Galilean cocycle \eqref{galileancocycle}, i.e., they can be {\it{undone}} by a change of coordinates.
The exception occurs only for the one dimensional Galilean group. In one dimension there is another true cocycle which will be used in section \ref{IRQUANT}.
}} $\epsilon_{G}$ and the coboundary $\epsilon_{N}$.
\begin{equation}
\epsilon_{BS}(g,g^{\prime}) = \epsilon_{G}(g,g^{\prime}) \, + \, \epsilon_{N}(g,g^{\prime})
\label{cocyclebs2a}
\end{equation}
For later convenience when defining the Laplace transform, we use as Galilean cocycle{\footnote{
$\epsilon_{G}$ equals
the $WSp(2, \mathbb{R} )$ cocycle in \eqref{slr2a}
plus the coboundary generated by $(p x)/2$ (see section \ref{coboundaries1}).
\vskip 1mm
The $WSp(2, \mathbb{R} )$ cocycle is
\begin{equation}
\epsilon_{W}(g,g^{\prime}) \equiv - \frac{1}{2} { \bf{u}^{\prime } {\bf{M_{BS}}} {\bf{\mathrm{\Omega}}} \bf{u}^{T} } =
\frac{1}{2}( x p^{\prime} - p x^{\prime} + \sigma^2 \, p {p^\prime} t )
\label{galileancocycle1}
\end{equation}
The coboundary generated by $(p x)/2$ is
\begin{equation}
\delta_{c} ( \frac{1}{2} \, p x)= \frac{1}{2} ( x^{\prime} p + x p^{\prime} + \sigma^2 {p^{\prime}}^2 t +\sigma^2 p p^{\prime} t )
\end{equation} }}
\begin{equation}
\epsilon_{G}(g,g^{\prime}) = p^{\prime} x + \sigma^2 \, p {p^\prime} t + \frac{1}{2} \sigma^2 {p^\prime}^2 t
\label{galileancocycle}
\end{equation}
\vskip 1mm
$\epsilon_{N}$ represents a numeraire choice{\footnote{This coboundary will provide a constant first derivative coefficient in the Black-Scholes equation.
It is possible to add a new group coordinate, much like a quantum gauge potential (see \cite{calixto}), which allows the introduction of numeraires that depend on stock price and time. This approach will be explored in a future work.}}
\begin{equation}
\epsilon_{N} (g,g^{\prime}) \equiv \delta_c(f_N) = \mu\, p^{\prime} t
\label{galileancob}
\end{equation}
and has the generating function (see section \ref{coboundaries1})
\begin{equation*}
f _N(g) = \frac{\mu}{\sigma^2} x
\end{equation*}
\vskip 1mm
\subsection{Stock Volatility} \label{variance}
\vskip 1mm
Bargmann ({\cite{bargmann1}, \cite{bargmann2}) used a central extension of the Galileo group with $U(1)$
for the description of the free non relativistic quantum mechanical particle.
The cocycle used by Bargmann was
\begin{equation}
\epsilon_{B} (g,g^{\prime}) = - \frac{i }{2 \hslash } m ( x v^{\prime} - v x^{\prime} + \, v {v^\prime} t )
\label{bargman1}
\end{equation}
which is labeled by the particle mass $m$. Bargmann proved that two cocycles of the form \eqref{bargman1}
with different values of $m$ are not equivalent, they cannot be transformed into each other by a coboundary.
\vskip 1mm
The cocycle \eqref{bargman1} equals the $WSp(2, \mathbb{R} )$ cocycle
\begin{equation}
- \frac{1}{2} { \bf{u}^{\prime } {\bf{M_{BS}}} {\bf{\mathrm{\Omega}}} \bf{u}^{T} } =
\frac{1}{2}( x p^{\prime} - p x^{\prime} + \sigma^2 \, p {p^\prime} t )
\label{galileancocycle1}
\end{equation}
with the substitutions
\begin{equation*}
\sigma^2 \rightarrow i \frac{\hslash}{m} \quad \qquad p \rightarrow \frac{v}{\sigma^2}
\label{barg1}
\end{equation*}
The cohomological importance of mass in in quantum mechanics has been extensively studied in the literature. The stock volatility plays a similar role in finance.
Formally, the Black-Scholes theory is the quantum mechanics of a free particle with an imaginary mass. The classical limit $\hslash \to 0$ corresponds to the zero volatility limit.
\vskip 1mm
\subsection{Lie Algebra}{\label{IVF}}
\vskip 1mm
\subsubsection*{Left Invariant Vector Fields}
\begin{align}
{X_{p}^L} & = \frac{\partial}{\partial p} + x \, \Xi \qquad \qquad {X_{x}^L} = \frac{\partial}{\partial x} \nonumber \\
{X_{t}^L} & = \frac{\partial}{\partial t} + \sigma^2 p \, \frac{\partial}{\partial x} + E(p) \, \Xi \nonumber \\
{X_{\zeta}^L} & \equiv \Xi = \zeta \frac{\partial}{\partial \zeta}
\label{leftXbs}
\end{align}
with
\begin{equation}
\qquad E(p) \equiv \frac{1}{2 } \sigma^2 \,p^2 + \mu p
\label{energy}
\end{equation}
\vskip 1mm
\subsubsection*{Commutators}
The non-zero commutators are
\begin{equation}
[{X_{t}^L} , {X_{p}^L} ] = - \sigma^2 {X_{x}^L} - \mu \Xi \qquad \quad [{X_{p}^L} , {X_{x}^L} ] = \, - \Xi
\label{liebrakets0}
\end{equation}
\vskip 1mm
\subsection{Connection}
\vskip 1mm
The expression for the
the vertical form $\Theta$ and the curvature form $ d \Theta $ are (see section \ref{canonical})
\begin{align}
& \Theta = -x d p - E(p) \, dt + d \, \Xi \nonumber \\
& d \Theta = d p \wedge d x - (\sigma^2 p + \mu) \, dp \wedge dt
\label{theta2bs}
\end{align}
with
\begin{equation}
\qquad \qquad d\, \Xi \equiv\frac{1}{ \zeta} d \zeta
\label{energy}
\end{equation}
\vskip 1mm
\subsubsection{Characteristic Module}\label{charmod}
\vskip 1mm
$\mathcal{C}_\Theta$
is generated by a unique field $X_C $
\begin{align}
&X_C = {X_{t}^L} + \mu \, {X_{x}^L} = \nonumber \\
&\qquad \frac{\partial}{\partial t} + (\sigma^2 p + \mu ) \frac{\partial}{\partial x} + E(p) \, \Xi
\label{mod2bbbs}
\end{align}
The restriction of $\Theta$ to the $\mathcal{C}_\Theta$ flows gives the Black-Scholes Lagrangian, see appendix \ref{LAGBS}.
\vskip 1mm
\subsubsection{Interest Rates}\label{IRDEF}
\vskip 1mm
We introduce interest rates by performing
a change of coordinates{\footnote{This is equivalent to a non-horizontal polarization \cite{aldaya22}, with constraints of the form $X \,\Psi = a \Psi, a \in \mathbb{R}$ rather than $X \, \Psi = 0$. }} in
the fiber $\zeta$. The mapping
\begin{equation*}
\zeta \mapsto \tilde{\zeta} = \zeta \, e^{ r t} \qquad r \in \mathbb{R}
\end{equation*}
does not change the group law \eqref{cocyclebs1a}, since $t^{\prime \prime} =t + t{^\prime}$,
however it
alters the separation between vertical and horizontal spaces.
Under this coordinate change, the structural group generator $\Xi$ becomes
\begin{equation*}
\Xi = \zeta \frac{\partial}{\partial \zeta} \quad \mapsto \quad \tilde{\Xi} = \tilde{\zeta} \frac{\partial}{\partial \tilde{\zeta}}
\end{equation*}
and it dual form $d \Xi$ acquires a horizontal component
\begin{equation}
d \, \Xi = \frac{1}{ \zeta} d \zeta \mapsto \, \frac{1}{ \tilde{\zeta}} d \tilde{\zeta} + + r \, dt \equiv d \, \tilde{\Xi } + r \, dt
\label{tilda1}
\end{equation}
The vertical connection form $\Theta$ is modified in a straightforward way by equation \eqref{tilda1}
\begin{equation*}
\Theta \mapsto \tilde{\Theta} = \Theta + r \, dt
\end{equation*}
\vskip 1mm
Therefore, the time component ${f_t^L}$ of the LIVFs (determined by $\tilde{\Theta} ( X^{L}) = 0$) picks a new vertical part
\begin{equation*}
{f_t^L}\frac{\partial }{\partial t} \mapsto {f_t^L}\frac{\partial }{\partial t} - r \tilde{\Xi }
\end{equation*}
which induces the following change in the left time generator
\begin{equation*}
{X_{t}^L} \mapsto {X_{t}^L} - r \tilde{\Xi }
\end{equation*}
Note that one obtains the same Lie algebra commutators \eqref{liebrakets0} in the new coordinates. The RIVFs are not modified by this coordinate change.
\vskip 1mm
In the rest of this section we use the expression of the vectors fields and connection in these new coordinates and we omit the notation $\tilde{\zeta}$ , $\tilde{\Theta}$, etc. for brevity.
\vskip 1mm
\subsection{First Order Polarization}{\label{FPO}
\vskip 1mm
\subsubsection{Black-Scholes Equation in Momentum Space}
\vskip 1mm
In section \ref{BSQUANT} we noted that ${\bf{M_{BS}}}$ is a shear mapping where only the $p$ space (upper plane) is invariant, and that ${\bf{M_{BS}}}$ cannot be diagonalized. These features
indicate that the only possible first order polarization is a polarization in $p$-space, as we will verify below.
\vskip 1mm
The first order polarization algebra $\mathcal{P}$ is spanned
{\footnote{${X_{p}^L}$ cannot belong to $\mathcal{P}$ because
\begin{equation*}
[{X_{t}^L} , {X_{p}^L} ] = - \sigma^2 {X_{x}^L} - \mu \, \Xi
\end{equation*}
}}
by the $x$-translations and the $\mathcal{C}_\Theta$ generator $X_C$
\begin{equation}
{\mathcal{P}} = \langle {X_{x}^L} , {X_C} \rangle
\label{pol0bs}
\end{equation}
\vskip 1mm
The polarized functions $\Psi \in {\mathcal{F_P}}$ are found by imposing the polarization constraints
on the functional space ${\mathcal{F_P}}$
\begin{equation*}
{\mathcal{F_P}} = \{\, \Psi : \tilde{G} \rightarrow \mathbb{R} \, / \, \Psi(x,p, t,\zeta) = \zeta \Psi(x,p,t) \, \}
\label{theta5}
\end{equation*}
One has
\begin{subequations}
\begin{align}
{X_{x}^L} \Psi = 0 & \rightarrow \Psi = \zeta \, \psi(p,t) \nonumber \\
X_C \Psi = 0 & \rightarrow \nonumber \\
& \frac{\partial \psi}{\partial t}+ \frac{1}{2} \sigma^2 p^2 \, \psi + \mu \, p \, \psi - r \, \psi = 0 \label{fpol1}
\end{align}
\end{subequations}
Equation \eqref{fpol1} is the Black-Scholes equation in momentum space.
\vskip 1mm
\subsubsection{Polarized Functions}
\vskip 1mm
A separable solution of the Black-Scholes equation \eqref{fpol1} is
\begin{equation*}
\tilde{\psi} (p, t)= e^{-E_{r}(p) t} \, \Phi(p)
\end{equation*}
with $\Phi(p)$ an arbitrary function of $p$ and
\begin{equation*}
E_{r}(p) \equiv \frac{1}{2 } \sigma^2 \,p^2 + \mu p -r
\end{equation*}
We can write a general polarized function as an inverse Laplace transform (a Linear Canonical Integral transform)
\begin{align}
\Psi (\zeta, x, t) & = \nonumber \\
& \zeta \, \frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, K(p, t) \exp( p x) \,\Phi(p) \, d p
\label{pol1bs0}
\end{align}
with
\begin{equation}
\qquad K(p, t) = \exp(- E_{r}( p) \, t)
\label{momkernel}
\end{equation}
The integration measure $d p$ is the dual form of ${X_{p}^L}$, the vector field absent from the polarization algebra.
We will justify the use of the inverse Laplace transform in section \ref{OPX}.
\vskip 1mm
In the Black-Scholes theory, polarized functions represent prices of financial derivatives, whereas
the $\Phi(p)$ correspond to terminal (payoff) conditions. Appendix \ref{BRAKET} shows how to price financial instruments in momentum space.
\vskip 1mm
\subsection{High Order Polarization}\label{HOP}
\vskip 1mm
\subsubsection{Black-Scholes Equation in Coordinate Space}
\vskip 1mm
The Black Scholes equation can be obtained directly in the coordinate representation by using a {\it{high-order}} polarization (section {\ref{SecondPO}}).
\vskip 1mm
The second order operator
\begin{equation}
C_{P} = {X_{t}^L} + \mu {X_{x}^L} + \frac{1}{2}\,\sigma^2 \, X_{x}^{L} \,X_{x}^{L}
\label{cas1bs}
\end{equation}
is a Casimir operator commuting with all Black-Scholes LIVFs. Since $X_{x}^{L}$ and $X_{p}^{L}$ do not commute, $C_{P}$ defines two higher order
polarizations. The $x$-space polarization is generated by $ \langle \, C_{P} , X_{p}^{L} \, \rangle$
\begin{subequations}
\begin{align}
{X_{p}^L} \Psi = 0 & \rightarrow \Psi = \zeta \, e^{ - p x } \, \psi(x,t) \label{hpolbs1} \\
X_C \Psi = 0 & \rightarrow \nonumber \\
& \frac{\partial \psi}{\partial t} + \frac{1}{2} \sigma^2 \frac{\partial^2 \psi}{\partial x^2} + \mu \, \frac{\partial \psi}{\partial x} - r \psi = 0 \label{hpolbs}
\end{align}
\end{subequations}
Equation \eqref{hpolbs} is the Black-Scholes equation in coordinate (price) space.
\vskip 1mm
\subsection{Operators in Momentum Space}\label{FullQuant}
\vskip 1mm
\subsubsection{Right Invariant Vector Fields}
\vskip 1mm
\begin{align}
{X_{t}^R} & = \frac{\partial}{\partial t} \qquad \qquad {X_{\zeta}^R} \equiv \Xi = \zeta \frac{\partial}{\partial \zeta} \nonumber \\
{X_{x}^R} & = \frac{\partial}{\partial x} + p \Xi \nonumber \\
{X_{p}^R} & = \frac{\partial}{\partial p} + \sigma^2 t \frac{\partial}{\partial x} + (\sigma^2 \,p + \mu )\, t\, \Xi
\label{rightXbs}
\end{align}
we have
\begin{align}
[{X_{t}^R} , {X_p}^R] & = \sigma^2 {X_{x}^R} + \mu \, \Xi &&[{X_{x}^R} , {X_p}^R] = \, \Xi
\label{liebrakets01}
\end{align}
All other brackets are zero{\footnote{
The inner product of the vertical form $\Theta$ and the RIVFs gives the (Galilean) Hamiltonian constants of motion
\begin{align}
\Theta ({X_{x}^R}) = 0 & \rightarrow p \equiv p_0 \nonumber \\
\Theta ({X_{t}^R}) = 0 & \rightarrow \frac{1}{2}\sigma^2 p^2+ \mu p -r \equiv E_0 \nonumber \\
\Theta ({X_{p}^R}) =0 & \rightarrow x - ( \sigma^2 p + \mu ) t \equiv x_0 \nonumber
\label{noether}
\end{align}
}}.
\vskip 1mm
Irreducibility is achieved
by considering the action of the right invariant fields on the {\it{constants of motion}} $\Phi(p)$
in \eqref{pol1bs0}.
After straightforward algebra, one finds
\begin{equation}
{X_{p}^R} : \Phi(p) \rightarrow \frac{\partial}{\partial p} \,\Phi(p) \quad
{X_{x}^R} : \Phi(p) \rightarrow \, p \, \Phi(p)
\label{pol3bs2a}
\end{equation}
and
\begin{equation}
- {X_{t}^R} : \Phi(p) \rightarrow ( \frac{1}{2 } \sigma^2 p^2 + \mu p \, - r ) \, \Phi(p)
\label{pol3bs29}
\end{equation}
From \eqref{pol3bs2a} the price operator, $\hat{x}$, the {\it{momentum}} operator
$\hat{p}$ and the time evolution operator (Hamiltonian) $\hat{H}$ are given by
\begin{equation}
\hat{x} = \frac{\partial}{\partial p} \quad \hat{p} = p \quad \hat{H} \equiv \frac{1}{2} \sigma^2 {\hat{p}^2} + \mu \hat{p} - \, r
\label{oper1}
\end{equation}
\vskip 1mm
\subsection{Operators in Coordinate Space}\label{OPX}
\vskip 1mm
The expressions \eqref{oper1} for the operators in momentum space
suggest the following definition for the
coordinate operator, $\hat{x}$ and the (non hermitian) momentum , $\hat{p}$ operator in $x$-space
\begin{equation}
\hat{x} = x \quad \hat{p} = \frac{\partial}{\partial x} \qquad [\hat{p} , \hat{x} ]= 1
\label{oper2}
\end{equation}
Equation \eqref{oper2} is a direct consequence of our extension of the Heisenberg-Weyl subgroup by $\mathbb{R}^{+}$ (equation \eqref{weyl3}).
The 2-dimensional translation subgroup of the $WSp(2, \mathbb{R} )$ group acts on real exponentials, thus justifying the
definitions \eqref{oper2} and the use for the inverse bilateral Laplace transform in \eqref{pol1bs0}.
This is an opposition to quantum mechanics, where
the Heisenberg-Weyl subgroup acts on the unit circle instead, making the Fourier transform
the mapping between the coordinate and momentum spaces.
\vskip 1mm
Using equations \eqref{oper2}, the Black-Scholes Hamiltonian{\footnote{The non hermiticity of the Black-Scholes Hamiltonian
has been extensively
studied in the literature (\cite{jana}, \cite{boson}, \cite{blasi}).
Non hermiticity is relatively mild, with eigenvalues either real or appearing in complex conjugate pairs. }}
in coordinate space is
\begin{equation}
{\hat{H}}_{BS} \equiv \frac{1}{2} \sigma^2 \frac{\partial^2}{\partial x^2} + \mu \frac{\partial}{\partial x} - \, r
\label{oper3}
\end{equation}
\vskip 1mm
\subsubsection{Numeraire Coboundary Value}\label{MART}
\vskip 1mm
The value of the numeraire parameter $\mu$ can be determined by requiring the stock price $S \equiv e^x$ to be
a zero eigenvalue{\footnote{This is actually a martingale condition \cite{baaquie1}.}}
of the Hamiltonian operator{\footnote{In momentum space, the zero eigenvalue condition reads
\begin{equation}
(\frac{1}{2 } \sigma^2 p^2 + \mu p - \,r) \Psi_{0} (p) = 0 \rightarrow \Psi_{0}(p) = \delta(\frac{1}{2 } \sigma^2 p^2 + \mu p - \,r)
\end{equation}
It is straightforward to prove, using the properties of the Dirac delta and the inverse Laplace transform, that if we identify $\Psi_{0}(x) $ with the stock price $S$,
$\Psi_{0}(x) \simeq \exp(x )$, this requires $ \mu = r -\frac{1}{2} \sigma^2$
}}.
\vskip 1mm
Using the Black-Scholes Hamiltonian \eqref{oper3}, we find
\begin{equation}
\qquad {\hat{H}}_{BS} \, e^x = 0 \quad \Rightarrow \quad \mu = r -\frac{1}{2} \sigma^2
\end{equation}
\vskip 1mm
\subsection{Pricing Kernel} \label{EXPECT}
\vskip 1mm
The kernel $K_{BS}(x,x^\prime, \tau)$ for the Black-Scholes equation \eqref{hpolbs}
\begin{equation*}
\Psi(\tau, x) = \int K_{BS}(x,x^\prime, \tau) \Psi(0, x^\prime) d x
\end{equation*}
is obtained by an in inverse Laplace transform
from the momentum representation in \eqref{momkernel}
\begin{equation}
K_{BS}( x, x^\prime, t) = \frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, d p \, e^{ - E_{r}(p) \, t} e^{p (x-x^\prime)}
\label{pkernel1}
\end{equation}
The integral \eqref{pkernel1} does not exist for $t > 0$, in accordance to the fact that pricing problems in finance are time irreversible final value problems.
\vskip 1mm
Let $t < 0$ and define $\tau \equiv -t $ .Then the integral \eqref{pkernel1} exists and we recover the well known Gaussian pricing kernel
\begin{equation}
\label{kernellag1}
K_{BS}(x,x^\prime, \tau) = e^{-r \tau} \, \, \frac{1}{\sqrt{2 \pi \sigma^2 \tau}} e^{ -\frac{1}{2 \sigma^2 \tau } \, (x^\prime-x -\mu \tau)^2}
\end{equation}
\vskip 1mm
\subsubsection{Derivation using LCTs}
\vskip 1mm
The Black-Scholes pricing kernel can also be obtained with the methods developed in references \cite{Miller2} and \cite{wolf2}, using the properties of the $WSp(2, \mathbb{R} )$ group
and its representation as Linear Canonical Transformations.
\vskip 1mm
From equation \eqref{ltcdef2},
the $SL(2, \mathbb{R})$ matrix \eqref{genmatrixBSa} generates the LCT
\begin{equation}
W({\bf{M_{BS}}},x,x^\prime) = \frac{1}{\sqrt{2 \pi \sigma^2 \,t }}\, \exp( - ( x^\prime -x )^2/ (2 \,\sigma^2 \, t))
\end{equation}
which is the heat equation kernel (or Weierstrass transform $W[]$).
\vskip 1mm
$K_{BS}(x,x^\prime, \tau)$ is obtained by multiplying the heat kernel $ W({\bf{M_{BS}}})$ by the discount factor $\exp (-r \tau)$ and setting $\tau = T-t$, $x \rightarrow x + \mu \tau$.
We note that $x \rightarrow x + \mu \tau$ is the Galilean transformation \eqref{bs1m2b} representing a numeraire change.
\vskip 1mm
$K_{BS}$ can also be written in terms of the pseudo-differential operators {\footnote{
For smooth functions
\begin{equation}
e^{{a\,{\frac{\partial}{\partial x}}}}f(x)=f(x+a)
\end{equation}
Combining the result above with the Gaussian integral
\begin{equation}
e^{ a^2} = {\frac {1}{{\sqrt {4\pi }}}}\int _{{-\infty }}^{\infty } e^{a \, y} \, e^{{-y^2}/4} dy
\end{equation}
we get
\begin{equation}
e^{ {\frac{\partial^2}{\partial x^2}} } f(x) = {\frac {1}{{\sqrt {4\pi }}}}\int _{{-\infty }}^{\infty } f(x-y) e^{-y^2} dy
\end{equation}
}}
associated with $ W({\bf{M_{BS}}})$ \cite{wolf2}
\begin{align}
\label{kernellag2}
e^{\tau \, {\hat{H}}_{BS}} \, f(x) & = \quad e^{- \, r \tau} \, e^{ \frac{1}{2} \sigma^2 \tau \frac{\partial^2}{\partial x^2}} \, e^{\mu \tau \frac{\partial}{\partial x} } \, f(x) = \nonumber \\
& \int _{- \infty}^{+ \infty } d x^\prime K_{BS} (x,x^\prime, \tau) f(x^\prime)
\end{align}
\section{Instrument Prices in Momentum Space} \label{BRAKET}
\vskip 1mm
In this section we present examples of derivatives pricing in the less familiar momentum (Laplace) space.
\vskip 1mm
The price of a financial instrument $ \Psi ( x, t)$ maturing at $T$ is found by performing an inverse Laplace transform
\begin{equation}
\Psi ( x, t) = \frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, d p \, e^{x p} K(p,\tau) \Psi(p,T)
\label{price1}
\end{equation}
where $\tau \equiv T-t, \, t \leq T$ and $\Psi(p,T) = \Psi(p) \exp(E(p) T)$ is the payoff at maturity in the Laplace space
\begin{equation}
\Psi ( x, T) = \frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, d p \, e^{x p} \Psi(p,T)
\label{price2}
\end{equation}
\vskip 1mm
\subsection{Instrument Prices using the Mellin Transform}
\vskip 1mm
The inverse Mellin transform{\footnote{${\mathcal {M}}^{-1}$ exists only for
complex values of $y$ so that $c \equiv \Re(y) > 0$ is within certain (possibly multiple) {\it{convergence strips}}. Each strip leads to different results for ${\mathcal {M}}^{-1}(f)$
(\cite{Mellin2}, \cite{Mellin2b}, \cite{Mellin4}, \cite{Mellin5}).}} gives the instrument's price in terms of the stock value instead than the log-stock. The pricing equations \eqref{price1} and \eqref{price2} read now
\begin{align}
V ( S, t) & = \frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, d p \, \exp( - E(p) \tau) V(p,T) \, S^{p} \nonumber \\
V ( S, T) & = \frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, d p \, S^{p} V(p,T)
\label{mellinprice}
\end{align}
with
\begin{equation*}
E_{r}(p) \equiv \frac{1}{2 } \sigma^2 \,p^2 + \mu p -r \qquad \mu = 1- \frac{1}{2 } \sigma^2
\end{equation*}
The final condition in Mellin space is given by
\begin{equation}
C(p,0) = \int _{0}^{+ \infty } V ( S, T) \, S^{-p-1} dS
\label{mellinfc}
\end{equation}
\vskip 1mm
\subsection{Call Option Price}
\vskip 1mm
We price a call option with strike $X$ and maturity $T$ using the Mellin transform.
The payoff is, from \eqref{price2}
\begin{equation*}
C(p,0) = \int _{0}^{+ \infty } (S-X)^{+} S^{-p-1} dS = \frac{X^{1-p}}{p (p-1)}
\end{equation*}
with $ \Re(p) > 1$. The expression of the call option price in the Mellin space is, using \eqref{mellinprice}
\begin{equation*}
C(p,\tau) = e^{- r \tau} \, \frac{X^{1-p}}{p (p-1)} \, \exp ( - ( \frac{1}{2 } \sigma^2 \,p^2 + \mu p) \tau)
\end{equation*}
where $\tau \equiv T-t, \, t \leq T$.
\vskip 1mm
The Black-Scholes call option formula in the price (coordinate) representation is obtained by the inverse Mellin transform
\begin{equation*}
V(S, t) = e^{- r \tau} \frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, C(p,\tau) \, S^p \, d p
\label{call00}
\end{equation*}
where $c \in (1, +\infty)$ and $ \Re(p) > 1$.
\vskip1mm
Details of the calculation of this integral can be found in reference \cite{Mellin0}.
\section{Definitions}\label{DEF}
\vskip 1mm
\subsection*{Vector Fields}
\vskip 1mm
Let M be a N-dimensional manifold with local coordinates $x_i , i = 1,2 \ldots N$. A vector field $X$
is an application that associates a first order differential operator $X(x)$ to a point $x\in M$.
$X(x)$ can be expressed in local coordinates as a linear combination of the base fields\
\begin{equation*}
e_i \equiv \frac{\partial}{\partial x_i} \qquad i = 1,2 \ldots N
\label{fields1}
\end{equation*}
that is
\begin{equation*}
X(x) = X_i(x) \frac{\partial}{\partial x_i}
\label{fields2}
\end{equation*}
with $X_i(x), i = 1,2 \ldots N$ differentiable functions on M. The space of the vector fields is called the
tangent space of M, T (M). For brevity we will write $X$ instead of $X(x)$
The integral curves of X are the solution to the set of ordinary differential equations
\begin{equation*}
\frac{d x_i}{d s} = X_i(x)
\label{fields3}
\end{equation*}
where s is an integration parameter. Note that the invariance condition $X f = 0$ implies that$f$ is constant along the integral curves of X.
\subsection*{Forms}
\vskip 1mm
A 1-form $\Gamma$ is an application that associates to every point $x \in M$ an element of the dual space of T (M).
\begin{equation*}
\Gamma: x \rightarrow \Gamma(x) \qquad / \quad \Gamma(x) (X(x)) = f(x)
\label{forms1}
\end{equation*}
with $f(x)$ a differentiable function.
As with vector fields, we write $\Gamma$ for $\Gamma(x)$. The space of 1-forms
is called the cotangent space of M and is denoted by $T^{*}(M)$.
A convenient representation for the basis of $T^{*}(M)$ is
\begin{equation*}
u_i \equiv d x_i \qquad i = 1,2 \ldots N
\label{forms2}
\end{equation*}
and its action on the basis of T (M) is
\begin{equation*}
u_i (e_j) =\equiv d x_i ( \frac{\partial}{\partial x_j}) = \delta_{i,j} \qquad i ,j = 1,2 \ldots N
\label{forms3}
\end{equation*}
2-forms, 3-forms , etc. are linear combinations of tensor products of 1-forms. A function f is considered a zero-form.
An important operation on forms is the differential d. The differential of a n-form is either a (n+1)-form
or zero. For a function f, d is the ordinary differential
\begin{equation*}
d f(x) = \frac{\partial f }{\partial x_i} d x_i
\label{forms4}
\end{equation*}
For a 1-form $\Gamma$
\begin{equation*}
d \Gamma = \frac{\partial \Gamma_i }{\partial x_j} d x_i \wedge d x_j
\label{forms5}
\end{equation*}
where the wedge operator $\wedge$ is the antisymmetric combination
\begin{equation*}
d x_i \wedge d x_j = d x_i \otimes d x_j - d x_j \otimes d x_i
\label{forms6}
\end{equation*}
One can define in an analogous way differentials of higher order forms. Note that the antisymmetry
of the wedge operator implies that $d^2 f = 0$ and $d^2 \Gamma = 0$. Also, i f a $n$-form acts on $n-k$ k vector fields one obtains a $k$-form. For instance the $1$-form $d f$ acting
on a field X gives a zero-form
\begin{equation*}
d f(X) = \frac{\partial f }{\partial x_i} d x_i ( X_j(x) \frac{\partial}{\partial x_j}) = X_i(x) \frac{\partial f }{\partial x_i} = X(f)
\label{forms7}
\end{equation*}
Analogously, one can check that $d \Gamma$ acting on X alone gives a 1-form
\begin{equation*}
d \Gamma(X,.) = \frac{\partial \Gamma_i}{\partial x_j}( X_i d x_j - X_j d x_i)
\label{forms8}
\end{equation*}
We use the inner product notation to denote the action of a n-for $\Omega$ on a vector field X
\begin{equation*}
i_X(\Omega) = \Omega(X)
\label{forms9}
\end{equation*}
\vskip 1mm
\subsection*{Lie Derivative}
\vskip 1mm
The Lie derivative evaluates the change of vector fields and forms
along the flow defined by another vector field.
\vskip 1mm
The Lie derivative of a function f with respect to a vector field X is
\begin{equation*}
L_X f = X(f) = d f (X)
\label{lie1}
\end{equation*}
For two vector fields X, Y , the Lie derivative of Y with respect to X, $L_X Y$ , is a vector field defined by
\begin{align*}
L_X Y & = - L_Y X \equiv [X,Y] = \\
& \left( X_i \frac{\partial Y_j}{\partial x_i} - Y_i \frac{\partial X_j}{\partial x_i} \right) \frac{\partial} {\partial x_j}
\label{lie2}
\end{align*}
The Lie derivative of a 1-form $\Gamma$ with respect to a vector field X,$L_X \Gamma$ is also a 1-form, and has the meaning of
the rate of change of $\Gamma$ along the integral lines (flow lines) of X. One finds, in local coordinates
\begin{equation*}
L_X \Gamma = (L_X \Gamma)_i d x_i
\label{lie3}
\end{equation*}
where
\begin{equation*}
(L_X \Gamma)_i = X_j \frac{\partial \Gamma_i}{\partial x_j} + \Gamma_j \frac{\partial X_j}{\partial x_i}
\end{equation*}
or, in a more concise notation
\begin{equation*}
L_X \Gamma = d (\Gamma(X)) + d\Gamma (X,.) \equiv i_X d\Gamma + d(i_X \Gamma)
\label{lie4}
\end{equation*}
\vskip 1mm
\subsection*{Lie Algebra}{\label{LieAlgebra}}
\vskip 1mm
A Lie algebra ${\mathfrak{G}}$ is a vector space over a field ${\mathbb{F}}$ equipped with a bilinear map $[,]: ({\mathfrak{G}}, {\mathfrak{G}}) \rightarrow {\mathfrak{G}}$
such that $[X,Y] = - [Y,X]$ and $[X,[Y,Z]] + [Y,[Z,X]]+ [Z,[X,Y]] = 0$ (Jacobi identity).This map is called {\it{Lie bracket}}.
The Lie algebra for a Lie group G is defined using the tangent vectors at the identity $e$ as
vector space of and the Lie derivative as bracket operator. One has
\begin{equation*}
[X_{i}, X_{j}] = c_{i,j}^{k}X_{k}^{L} \qquad i, j = 1,2 \ldots {\text{dim}} \, G
\label{lieb0}
\end{equation*}
where the coefficients
$c_{i,j}^{k} \in {\mathbb{F}}, i, j, k = 1,2 \ldots {\text{dim}}\, G $ are called {\it{ structure constants}}.
Since $X_g$ is a LIVF if and only if $X_g = {L_g^{T}} X_{e}$
and an RIVF if and only if $X_g = {R_g^{T}} X_{e}$,
there is a biunivocal correspondence between the LIVFs (RIVFs) and the set of the tangent vectors at the identity $e$ of $G$, with the following brackets
\begin{align}
[X_{i}^{L}, X_{j}^{L}] & = c_{i,j}^{k}X_{k}^{L} \qquad [X_{i}^{R}, X_{j}^{R}] = - c_{i,j}^{k}X_{k}^{R} \nonumber \\
[X_{i}^{R}, X_{j}^{L}] & = 0 \qquad \forall \, i, j = 1,2 \ldots {\text{dim}} \, G
\label{lieb1}
\end{align}
\vskip 1mm
\section{The Group Quantization Formalism}\label{qgroupform}
\vskip 1mm
We describe the main features of the Group Quantization formalism (\cite{aldaya0}, \cite{wolf11}, \cite{garcia0}, \cite{aldaya22}).
\vskip 1mm
\subsection{The Quantization Group}\label{fiber}
\vskip 1mm
A {\it{Quantization Group}} $\tilde{G}$ is a connected Lie group {\footnote{A Lie group is a differential manifold $G$ endowed with a group composition law $F: G \times G \rightarrow G$ such as $F$ and its inverse are smooth, differentiable applications. }} $\tilde{G}$ which is
a central extension of a Lie group $G$ (called the {\it{dynamical group}}) by another Lie group $U$ called the structural group.
\vskip 1mm
The central extension
defines a {\it{principal bundle}} $(\tilde{G}, G, \pi, U)$. The structure group acts on the fiber by left multiplication. $G$ is called {\it{ base manifold}} of the bundle and $U$ is called the {\it{ fiber}}.
The projection $\pi$ is a continuous and surjective map $\pi: \tilde{G} \rightarrow G$
\vskip 1mm
\begin{figure}[h]
\includegraphics[width=0.8\linewidth]{fiberbundle1}
\caption{ One can visualize $\tilde{G}$ by imagining that through each point in the base manifold $G$ there is a line (fiber) of points with different values of the fiber coordinate $\zeta \in U$.}
\centering
\end{figure}
\vskip 1mm
A Quantization Group ${\tilde{G}}$ with structural group $U$ is a Cartan geometry. The Cartan geometry is the geometry of spaces that are locally (infinitesimally) like quotient spaces ${G = \tilde{G}}/U$.
Although G is not a subgroup of $\tilde{G}$ , $\tilde{G}/U \simeq G$ as topological spaces any
$\tilde{g} \in \tilde{G}$ can be decomposed in two parts, $\tilde{g} = (g, u), \, g \in G, u \in U$. This means that, locally, $ \tilde{G}$ looks like the Cartesian product of $G$ and $U$.
\vskip 1mm
In this paper, we will only consider Quantization Groups with structural group $U = {\mathbb{R}}^{+}$.
\vskip 1mm
\subsection{Invariant Vector Fields}\label{invariantl}
\vskip 1mm
The Group Quantization method considers the action of the group on itself, as opposed to considering the group acting on an external manifold. This guarantees the existence of two sets of commuting group operators, the right invariant fields and the left invariant fields.
\vskip 1mm
The {\it{left}} and {\it{right}} translations, $L_g$ and $R_{g}$ are defined as
\begin{align}
L_g:G \rightarrow G & \qquad / \quad L_g(g^\prime) = g g^\prime \nonumber \\
R_g:G \rightarrow G & \qquad / \quad R_g(g^\prime) = g^\prime g
\label{ff2}
\end{align}
$L_g$ and $R_{g}$ are diffeomorphisms of G. Note that $L_g$ and $R_{g}$ commute.
\vskip 1mm
Let $g,g^\prime, g^{\prime \prime} \in G $. In an abuse of notation, we write $g^{\prime \prime} = g^{\prime} g$ for the
base group $G$ composition law. We denote an element $\tilde{g} \in \tilde{G}$ by $\tilde{g}= (g, \zeta )$, with
$ \zeta$ in the structural group ${\mathbb{R}}^{+} $.
The composition law for $\tilde{G}$ is
\begin{equation}
{\tilde{g}}^{\prime\prime} \equiv (g^{\prime \prime} , \zeta^{\prime \prime} )= (g^{\prime} g, \zeta^{\prime} \zeta \exp(\epsilon(g^\prime, g))
\label{law0}
\end{equation}
where $G\in G$, $\zeta \in {\mathbb{R}}^{+}$ and $\epsilon(g^\prime, g) \in {\mathbb{R}}$ is the extension cocycle.
\vskip 1mm
A vector field $X$ is called a left invariant vector field (LIVF) if
\begin{equation*}
{L_{\tilde{g}}^{T}} X_{{\tilde{g}}^\prime} = X_{{\tilde{g}} {\tilde{g}}^\prime}
\label{ff3}
\end{equation*}
and a right invariant vector field (RIVF) if
\begin{equation*}
{R_{\tilde{g}}^{T}} X_{{\tilde{g}}^\prime} = X_{ {\tilde{g}}^\prime {\tilde{g}}}
\label{ff4}
\end{equation*}
\vskip 1mm
In local coordinates, the LIVF are given by
\begin{equation}
{X_i^L} = \sum_k\,{X_{i,k}^L} \frac{\partial}{\partial g_k} + \lambda_i \, \Xi
\label{ivf1}
\end{equation}
with
\begin{align}
& \Xi = \zeta \frac{\partial}{\partial \zeta} \qquad \qquad \lambda_i \equiv \left. \frac{\partial \epsilon(g^\prime, g)}{\partial g_i} \right \rvert_{g = g^\prime, g^\prime=e} \\ \nonumber
& {X_{i,k}^L} \equiv \left. \frac{\partial({g^\prime g)}_k}{\partial g_i} \right \rvert_{g = g^\prime, g^\prime=e}
\label{ivf2}
\end{align}
\vskip 1mm
and the RIVFs
\begin{equation}
{X_i^R} = \sum_k\,{X_{i,k}^R} \frac{\partial}{\partial g_k} + \gamma_i \, \Xi
\label{rvf1a}
\end{equation}
where
\begin{align}
& \Xi = \zeta \frac{\partial}{\partial \zeta} \qquad \qquad \gamma_i \equiv \left. \frac{\partial \epsilon(g^\prime, g)}{\partial g_i} \right \rvert_{g^\prime= g, g=e} \\ \nonumber
& {X_{i,k}^L} \equiv \left. \frac{\partial({g^\prime g)}_k}{\partial g^{\prime}_i }\right \rvert_{g^\prime= g, g=e}
\label{rvf2a}
\end{align}
The i-th component of a LIVF ${X_i^L}$ is calculated as the derivative respect to {\it{unprimed}} coordinates evaluated at the identity: first set $g^\prime = g$ in the derivative, then set $g=e$. The i-th component of a LIVF ${X_i^L}$ is calculated as the derivative respect to {\it{primed}} coordinates evaluated at the identity: first set $g = g^\prime$ in the derivative, then set $g^\prime=e$.
\vskip 1mm
Note that, by construction, RIVFs and RIVFs commute.
\vskip 1mm
\subsection{Cocycles}\label{cocycle}
The function $\epsilon$ in equation \eqref{law0} is called a cocycle. Cocycles are restricted by the group law
properties.
\vskip 1mm
Associativity of the group law
implies that $\epsilon$ must satisfy the following functional equation
\begin{equation}
\epsilon(g^{\prime \prime}, g^\prime) + \epsilon(g^{\prime \prime} g^\prime, g) = \epsilon(g^{\prime \prime},g^\prime g ) + \epsilon( g^\prime, g)
\label{extension4}
\end{equation}
and existence of an inverse element
\begin{equation}
\epsilon(g ,e) = \epsilon(e,g) = \epsilon(e,e ) =0
\label{extension4b}
\end{equation}
where $e$ is the identity element of $G$.
\vskip 1mm
\subsection{Coboundaries}\label{coboundaries1}
\vskip 1mm
A coordinate change in the fiber $U$ generates a mathematically trivial cocycle.
If the fiber elements $\zeta$ undergo the coordinate change $\zeta \rightarrow\zeta\, e^{f(g)}$ in the group law \eqref{law0}, the extension cocycle $\epsilon$ gets an extra additive factor $\delta_c(f)$
\begin{equation*}
\epsilon(g, g^\prime) \rightarrow \epsilon(g, g^\prime)+ \delta_c(f)(g,g^\prime)
\end{equation*}
with
\begin{equation}
\delta_c(f)(g,g^\prime) = f(g g^\prime) - f(g) -f(g^\prime)
\end{equation}
Trivial cocycles are called {\it{coboundaries}}{\footnote{Not all coboundaries are generated by a coordinate change.}}. The coboundary $\delta_c(f)(g,g^\prime)$ can can be undone by a change of coordinates.
\vskip 1mm
Although group extensions whose cocycles differ in a coboundary mathematically define the same group
representation, the dynamical effects generated by coboundaries are not necessarily trivial.
\vskip 1mm
One useful analogy is a change of variables in a differential equation: differential equations have {\it{canonical forms}} to which they can be reduced, however it is often convenient to work in
non-canonical coordinates. For instance, even if the Black-Scholes equation can be reduced to the simpler heat equation, in practice its solution is calculated using variables which allow to implement naturally the instrument's pricing boundary conditions.
\vskip 1mm
\subsection{Connection. Vertical Form}\label{canonical}
\vskip 1mm
A connection on $\tilde{G}$ is a smooth choice of {\it{horizontal}} subspace $H$ and {\it{vertical}} subspace $V$
such that $\tilde{G}$ can be decomposed as a direct sum, $\tilde{G} = H \oplus V$.
\vskip 1mm
\begin{figure}[h]
\includegraphics[width=0.8\linewidth]{fiberbundle2}
\caption{Transformation along the fiber $U$ are {\it{vertical}}, quantities defined on
the base manifold $G \simeq \tilde{G}/U$ are {\it{horizontal}}.}
\centering
\end{figure}
\vskip 1mm
The Group Quantization formalism uses a Cartan{\footnote{The notion of an affine connection on a differential manifold ( {\it{Poincar\'e-Cartan}} form ) was introduced by the French mathematician \'Ellie Cartan in two papers
published in 1923 and 1924 \cite{cartan}.}}
connection,
where the geodesics coincide with the flows of an integrable
LIVF algebra. Concretely,
the connection form in the Group Quantization formalism is
the {\it{vertical}} part $\Theta$ of the {\it{Maurer-Cartan}} form{\footnote{The {\it{Maurer-Cartan}} form.
$\Gamma :\tilde{G} \rightarrow {\mathfrak{\tilde{G}}}$ takes values in the Lie algebra of $\tilde{G}$, ${\mathfrak{\tilde{G}}}$
and it is the unique left-invariant 1-form such that $\Gamma\rvert_e$ is the identity map.
$\Gamma$ can be written
in terms of the left invariant fields at the identity ${X_i^L}, i = 1,2 \ldots dim(\tilde{G})$ and their dual forms $\theta$
\begin{equation*}
\Gamma =\sum _{i}{X_i^L}\otimes \theta ^{i} \qquad \theta ^{i}({X_i^L}) = \delta_{i,j}
\end{equation*}
The {\it{ Maurer-Cartan equations}}
\begin{equation*}
d\theta ^{i}=-{\frac{1}{2}}\sum _{{jk}}{c^{i}}_{{jk}} \, \theta ^{j}\wedge \theta ^{k}.
\end{equation*}
express the differential of the form in terms of the ${c^{i}}_{{jk}}$, the structure constants of the Lie algebra $ {\mathfrak{\tilde{G}}}$. }}
\vskip 1mm
In this work, the vertical space is the structural group $U= {\mathbb{R}}^{+} $ with a unique generator $\Xi = \zeta \frac{\partial}{\partial \zeta}$ (see equation \eqref{law0}).
\vskip 1mm
$\Theta$ provides a natural definition of horizontality
by requiring that the horizontal fields belong to the kernel of $\Theta$.
\begin{equation}
i_{X^L}\, \Theta \equiv \Theta(X^L) = 0 \qquad i_\Xi \, \Theta \equiv \Theta(\Xi) = 1
\label{extension7b}
\end{equation}
where $X^L$ stands for all LIVFs (Left Invariant Vector Fields).
\vskip 1mm
In coordinates, equations \eqref{extension7b} read (see section \ref{invariantl})
\begin{equation}
\Theta = \sum _{i} \, \theta_i \, d g_i + d\, \Xi \quad \qquad d\, \Xi \equiv\frac{1}{ \zeta} d \zeta
\label{extension8}
\end{equation}
where the $\theta_i$ are solutions of a linear algebraic system
\begin{equation}
\sum _{k} \, \theta_k \, {X_{i,k}^L} + \lambda_i = 0
\label{extension8}
\end{equation}
\vskip 1mm
\subsection{Characteristic Module}\label{charmod}
\vskip 1mm
In the Group Quantization formalism, the elements
of the characteristic module $\mathcal{C}_\Theta$ are interpreted as evolution operators.
$\mathcal{C}_\Theta$ is an integrable system analogous to Hamiltonian fields in classical mechanics,
and
the integral flows of the $\mathcal{C}_\Theta$ elements are the geodesics of the Cartan connection.
$\mathcal{C}_\Theta$ is defined
as the set of LIVFs that leave the connection form (vertical form) $\Theta$ strictly invariant
{\footnote{A vector field X is a symmetry of a 1-form $\Gamma$ if X leaves $\Gamma$ semi-invariant, that is, the Lie derivative
of $\Gamma$ with respect to X is a total differential
\begin{equation*}
L_X \Gamma= d f^{X} \quad \longrightarrow \quad d(i_X \Gamma) + i_X (d \Gamma) = d f^{X}
\label{lie5}
\end{equation*}
where $f^X$ is some function associated with the vector field X. Then, $\mathcal{C}_\Theta = \ker \Theta \cap \ker d \Theta $.
}}
\begin{equation}
X \in \mathcal{C}_\Theta \quad \longrightarrow \quad d(i_X \Theta) = 0 \quad i_X (d \Theta) = 0
\label{mod1bb}
\end{equation}
The curvature form $\omega = d \Theta$ can be written in local coordinates as
\begin{align}
\omega & = \sum _{i,j} \, \frac{\partial \theta_i }{\partial g_j} d g_j \wedge d g_i = \\ \nonumber
& \qquad \sum _{i,j} \, (\frac{\partial \theta_i }{\partial g_j}-\frac{\partial \theta_j}{\partial g_i}) dg_j \otimes dg_i
\label{extension7bb}
\end{align}
then, the second of equations \eqref{mod1bb} reads, using the results in section \ref{canonical}
\begin{equation}
\sum _{j} \, (\frac{\partial \theta_i }{\partial g_j}-\frac{\partial \theta_j}{\partial g_i}) \, {X_j} =0
\label{mod1bb2}
\end{equation}
Note that from the Maurer-Cartan equations, $ \omega(X,Y) = 0$ implies $\Theta([X,Y]) = 0$.
\vskip 1mm
\subsection{Polarization Algebra}{\label{FPO}}
\vskip 1mm
The Group Quantization
formalism constructs an irreducible representation of $\tilde{G}$ starting from the ring ${\mathcal{F}}$ of functions
with domain $\tilde{G}$ and range ${\mathbb{R}}$.
\vskip 1mm
The reduction is achieved by restricting the arguments of the functional space ${\mathcal{F}}$
using a set of horizontal operators and building a polarized function space ${\mathcal{F_P}}$ that
provides a representation of the group.
\vskip 1mm
The action of vertical fields (generators of the structural group $U$) $ X^V$ on ${\mathcal{F_P}}$ is
\begin{equation*}
X^V \Psi = f(X^V) \Psi \qquad \forall \, \Psi \in {\mathcal{F_P}}
\end{equation*}
where $f(X^V)$ is a function associated to the generator $X^V$. We say that $ {\mathcal{F_P}}$ is a
$U$- invariant functional space. In this document, $U = \mathbb{R}^{+}$ and $f(X^V) = 1$.
\vskip 1mm
We will see later in this document that, for the Black-Scholes and the Ho-Lee groups,
polarized functions have the meaning of
financial instrument prices.
\vskip 1mm
\subsubsection{First Order Polarization}{\label{FirstPO}}
\vskip 1mm
The Group Quantization formalism provides a
natural choice for polarization by
requiring invariance respect to
a {\it{horizontal}} algebra $\mathcal{P}$ ( {\it{polarization}} algebra )
that imposes constraints on ${\mathcal{F}}$.
\vskip 1mm
A {\it{first-order polarization}} (or just {\it{polarization}}) $\mathcal{P}$ is defined as a maximal
horizontal commutative left invariant algebra containing the characteristic module $\mathcal{C}_\Theta$, so that
the constraints are compatible with the evolution operators.
\vskip 1mm
The
{\it{polarized}} functions ${\mathcal{F_P}}$ will be characterized by conditions of the form
\begin{equation*}
{\mathcal{F_P}} = \{ \Psi \in {\mathcal{F}} / X^P \Psi = 0 \quad \forall X^P \in \mathcal{P} \}
\label{theta55}
\end{equation*}
\vskip 1mm
Since the elements of the algebra $\mathcal{P}$ are integrable vector fields, a first order polarization defines a {\it{foliation}} of $\tilde{G}$.
This means that it is possible to select functional subspaces on
on $\tilde{G}$ by requiring them to be constant along integral leaves
of the foliation.
\vskip 1mm
\subsubsection{Higher Order Polarization}{\label{SecondPO}}
\vskip 1mm
When a first order polarization is not able to provide the required functional constraints,
a {\it{higher order polarization}} can be used.
\vskip 1mm
As opposed to first-order polarizations described in (\ref{FirstPO}),
higher-order polarizations \cite{aldaya6} contain higher-order differential operators belonging to the left enveloping algebra.
Higher order polarizations do not define a {\it{foliation}} of $\tilde{G}$.
\vskip 1mm
A higher-order polarization $\mathcal{P_H}$
is a maximal subalgebra of the left-invariant enveloping algebra that has no intersection with the generators of the
structural group $U$ and commutes with the first order polarization $\mathcal{P}$.
\vskip 1mm
Commuting with the first order polarization ensures
compatibility with the action of the dynamical operators (RIVFs). As we shall see in the next sections, higher order polarizations can be constructed by using a Casimir operator.
\vskip 1mm
\subsection{Operators}\label{operators}.
\vskip 1mm
Commutativity of the right and left generators makes the RIVFs good candidates for
{\it{quantum operators}}, that is, operators that can reduce the representation from phase space, with coordinates and momenta, to an
irreducible group representation with only coordinates or momenta.
\vskip 1mm
Note that, since $\mathcal{P}$ is spanned by LIVFs, if $X^R$ is a right generator and $\Phi$
is polarized
\begin{equation*}
X^P(X^R \Phi) = X^R(X^P \Phi) = 0 \qquad \forall X^P \in \mathcal{P} \quad \forall \Phi \in {\mathcal{F_P}}
\label{mod1bb5}
\end{equation*}
so the action of the RIVFs on the space of polarized functions is well defined. Hence
the quantum operators are the restriction of the RIVFs to the polarized $U$- invariant functional space.
\vskip 1mm
\subsection{Noether's theorem}
\vskip 1mm
One can easily verify that the conection $\Theta$ is right invariant, $\Theta(X^R) = 0$ for all RIVFs.
The inner product of the vertical form $\Theta$ and the RIVFs gives the classical Hamiltonian constants of motion.
\vskip 1mm
\subsection{Lagrangian}
\vskip 1mm
The classical Lagrangian ${\mathcal{L}} (x, \dot{x})$ is obtained by the projection onto the base manifold $G$ of the connection form $\Theta$ along the
$C_\Theta$ flows (trajectories). See appendix \ref{CLASSA} for a discussion of the {\it{Poincar{\'e}-Cartan}} form of classical mechanics and the Lagrangian function.
\vskip 1mm
\subsubsection{Group Composition Law}
\vskip 1mm
Using equation \eqref{slr2a}, ${\bf{M_H}}$ generates the following composition law:
\begin{align}
t^{\prime \prime} & =t + t^\prime \nonumber \\
p^{\prime \prime}& = p + p^\prime \, \cosh \Omeg t + \lambda^{2} x^\prime \sinh \Omeg t \nonumber \\
x^{\prime \prime} & = x + x^{\prime} \, \cosh \Omeg t + \lambda^{-2} \, p^\prime \, \sinh \Omeg t \nonumber \\
\zeta^{\prime \prime} & = \zeta^\prime \zeta \exp \left ( \epsilon_H(g^\prime, g) \right)
\label{cocyclebs1z}
\end{align}
with the cocycle $\epsilon_{H}$
\begin{align}
\epsilon_{H}(g,g^{\prime}) & = \frac{1}{2} ( p x^\prime - x p^\prime) \cosh \Omeg t+ \\ \nonumber
& \quad \frac{1}{2} (\lambda^{-2} p \, p^\prime - \lambda^{2} \, x \, x^\prime) \sinh \Omeg t
\label{cocyclehs2}
\end{align}
The matrix ${\bf{M_H}}$ reduces to the Black-Scholes generating matrix ${\bf{M_{BS}}}$ in the limit $w \rightarrow 0$, and correspondingly,
the quantization group $\tilde{H}$ contracts to the Black-Scholes quantization group $\tilde{G}$ (modulo coboundaries) in this limit.
For brevity, we have omitted the numeraire coboundary{\footnote{The generating function for this coboundary is $\mu \, x$, and the final expression is
$\mu ( \lambda^{-2} p^\prime \sinh \Omeg t + x^{\prime} \cosh \Omeg t- x^{\prime})$, with $\mu$ real.}} \eqref{galileancob} in the composition law \eqref{cocyclebs1z}.
\vskip1mm
\subsection{Polarization using Orthogonal Coordinates}\label{LADDERPOL}
\vskip1mm
The $SL(2, \mathbb{R})$ matrix ${\bf{R}}$
\begin{equation}
\qquad \quad {\bf{R}} = \frac{1}{\sqrt{2}} \, \begin{pmatrix*}[c]
\lambda & - 1/ \lambda \\
\lambda & \phantom{-} 1/ \lambda
\end{pmatrix*}
\label{diagonalM}
\end{equation}
transforms ${\bf{M_H}}$ into a change of scale (squeeze) operator
\begin{equation*}
{\bf{D_H}} = {\bf{R^{-1}}} \, {\bf{M_H}} \, {\bf{R}} = \begin{bmatrix}
e^{ \Omeg t} & 0 \\
0 & e^{ - \Omeg t}
\end{bmatrix}
\end{equation*}
The change of coordinates
\begin{align}
A =& \frac{1}{\sqrt{2}} ( \frac{1}{\lambda} \, p - \lambda \, x) \nonumber \\
B =& \frac{1}{\sqrt{2}} ( \frac{1}{\lambda} \, p + \lambda \, x )
\label{diagonalFrac}
\end{align}
splits the phase space (p,x) into orthogonal subspaces. The group law \eqref{cocyclebs1z} is written as
\begin{align*}
t^{\prime \prime} & =t + t^\prime \nonumber \\
A^{\prime \prime} & = A + A^{\prime} e^{-\Omeg t} \\
B^{\prime \prime} & = B + B^{\prime} e^{\Omeg t} \nonumber \\
\zeta^{\prime \prime} & = \zeta^\prime \zeta \, e^{ \epsilon_H(g^\prime, g) }
\end{align*}
with
\begin{equation*}
\epsilon_{H}(g,g^{\prime}) =\frac{1}{2} (B^{\prime} A \, e^{ \Omeg t }\, - \, B \, A^{\prime} \, e^{ - \Omeg t })
\end{equation*}
\vskip 1mm
\subsubsection*{Left Invariant Vector Fields}
\vskip 1mm
\begin{align*}
{X_{t}^L} & = \frac{\partial}{\partial t} -\Omeg \, A \frac{\partial}{\partial A} + \Omeg \, B \frac{\partial}{\partial B} \nonumber \\
{X_{A}^L} & = \frac{\partial}{\partial A} + \frac{1}{2} B \, \Xi \qquad {X_{B}^L} = \frac{\partial}{\partial B} - \frac{1}{2} A \, \Xi
\end{align*}
with Lie Brackets
\begin{align*}
[{X_{t}^L} , {X_{A}^L} ] & = w {X_{A}^L} \qquad \quad [{X_{t}^L} , {X_{B}^L} ] = - w {X_{B}^L} \nonumber \\
[{X_{A}^L} , {X_{B}^L} ] & = -\Xi
\end{align*}
\vskip 1mm
\subsection*{Connection}
\vskip 1mm
The vertical form $\Theta$ and the curvature form $ d \Theta $ are given by
\begin{align}
\Theta & = \frac{1}{2} (A \,d B - B \, d A ) - w A B \, d t + d \Xi \nonumber \\
d \,\Theta & = d A \wedge d B - w A \, d B \wedge d t - w B \, d A \wedge d t
\end{align}
\vskip 1mm
$\mathcal{C}_\Theta$ is spanned by the time generator ${X_{t}^L} $.
\vskip 1mm
\subsection*{First Order Polarization}
\vskip 1mm
We choose the first order polarization algebra spanned by
\begin{equation*}
{\mathcal{P}} = \langle {X_{B}^L} , {X_{t}^L} \rangle
\end{equation*}
The polarized functions $\Psi \equiv \zeta \, \Psi(A,B,t) $ are found by imposing
\begin{subequations}
\begin{align}
{X_{B}^L} \, \Psi & = 0 \quad \rightarrow \quad \Psi = \zeta \, e^{\frac{A B}{2}} \Psi(A,t) \nonumber \\
{X_{t}^L} \, \Psi & = 0 \quad \rightarrow \quad \frac{\partial \Psi}{\partial t} - \Omeg A \, \frac{\partial \Psi}{\partial A} = 0 \label{ladder1}
\end{align}
\end{subequations}
The readers familiar with quantum mechanics will find
equation \eqref{ladder1} formally similar to quantum harmonic oscillator coherent state equations in Fock space (\cite{garcia0}, \cite{bisquert}), except that here the coordinate $A$ is real, not complex.
\vskip 1mm
We give the solution of these polarization constraints in equation \eqref{coherent2}.
Equation \eqref{ladder1} coincides with equation \eqref{polho1a} obtained in the next section \ref{PHASEPOL}.
\vskip 1mm
\subsection{First Polarization in Phase Space }\label{PHASEPOL}
\vskip 1mm
We use now the original phase space coordinates $(p,q)$ in the composition law \eqref{cocyclebs1z}.
\vskip 1mm
Since the generating matrix $M$ does not leave invariant either the $p$ or the $x$ space, a first order polarization cannot map the phase space $(p, x)$ into either the momentum or the coordinate space.
However, the results in this section will be illustrative of a phase space formulation for the harmonic oscillator and will also be useful for finding a higher order polarization.
\vskip 1mm
\subsubsection*{Left Invariant Vector Fields}
\vskip 1mm
\begin{align}
{X_{t}^L} & = \frac{\partial}{\partial t} + \Omeg \, \lambda^2 x \, \frac{\partial}{\partial p} + \Omeg \, \lambda^{-2} p \, \frac{\partial}{\partial x} \nonumber \\
{X_{x}^L} & = \frac{\partial}{\partial x} - \frac{1}{2} p \, \Xi \nonumber \quad \nonumber \\
{X_{p}^L} & = \frac{\partial}{\partial p} + \frac{1}{2} x \, \Xi
\label{leftXHO1}
\end{align}
\begin{align}
[{X_{t}^L} , {X_{p}^L} ] & = - \Omeg \, \lambda^{-2} {X_{x}^L} \qquad [{X_{t}^L} , {X_{x}^L} ] =- \Omeg \, \lambda^2 \, {X_{p}^L} \nonumber \\
[{X_{p}^L} , {X_{x}^L} ] & = \,- \Xi
\label{liebraketsHO}
\end{align}
All other brackets are zero.
\vskip 1mm
\subsection*{Connection}
\vskip 1mm
\begin{align}
\Theta & = \frac{1}{2}(p dx - x dp) - E(p,x) dt+ d \, \Xi \nonumber \\
d \Theta & = d p \wedge d x - \Omeg \, \lambda^{-2} \, p \, dp \wedge dt + \Omeg \, \lambda^2 \, x \, dx \wedge dt
\label{theta2bs}
\end{align}
with
\begin{align}
& E(p,x) \equiv \frac{1}{2 } \Omeg ( \lambda^{-2} p^2 - \lambda^2 \, x^2 ) \nonumber \\
& d\, \Xi \equiv\frac{1}{ \zeta} d \zeta
\label{energyHO}
\end{align}
\vskip 1mm
The time translations ${X_{t}^L}$ are the only generator of the characteristic module $\mathcal{C}_\Theta$.
\vskip 1mm
\subsection*{First Order Polarization}
\vskip 1mm
There are two first order polarizations $ \mathcal{P}$ in phase space
\begin{subequations}
\begin{align}
&\langle {X_{t}^L}, Y_{+} \rangle \qquad / \quad Y_{+} \equiv \, \lambda \, {X_{p}^L} + \lambda^{-1} \, {X_{x}^L} \label{hopol1} \\
& \langle {X_{t}^L}, Y_{-} \rangle \qquad / \quad Y_{-} \equiv \,\lambda \, {X_{p}^L} - \lambda^{-1} {X_{x}^L}
\end{align}
\end{subequations}
We choose \eqref{hopol1} as polarization. Let $ \Psi$ a function of the form $ \Psi(\zeta, p, q, t) = \zeta \, \Psi( p, q, t) $
\begin{align}
& Y_{+} \Psi = 0 \rightarrow \nonumber \\
& \lambda \frac{\partial \Psi}{\partial p} - \lambda^{-1} \frac{\partial \Psi}{\partial x} + \frac{1}{2}\, (\lambda x - \lambda^{-1} p) \Psi = 0 \label{phasecons5a}
\end{align}
The solution of \eqref{phasecons5a} is
\begin{equation}
\Psi(x,p, t,\zeta) = \zeta \psi ( A, t )\, e^{ \frac{1}{2} A \, B}
\label{phasecons5}
\end{equation}
where $A = A(p,q)$ and $B=B(p,q)$ are the orthogonal coordinates \eqref{diagonalFrac}. Then
\begin{equation}
{X_{t}^L} \Psi = 0 \rightarrow
\frac{\partial \psi}{\partial t} - \Omeg \, A \, \frac{\partial \psi}{\partial A} = 0
\label{polho1a}
\end{equation}
The constraint \eqref{polho1a} is the previously found \eqref{ladder1} using orthogonal coordinates.
\vskip 1mm
\subsection{First Order Polarized Functions}\label{phasepol}
\vskip 1mm
Using the separable solution of \eqref{polho1a}
\begin{equation*}
\qquad \qquad \, e^{\Omeg \, n \, t} \, A^n e^{\Omeg n t} \qquad n \in \mathbb{N}^{+}
\end{equation*}
the polarized functions are expressed as an infinite series
\begin{equation}
\Psi(\zeta, p,x,t) = \zeta \, \sum_{n=0}^{\infty} e^{\Omeg \, n \, t} \, \Alp_n \, A^{n} e^{\frac{1}{2} A B}
\label{coherent2}
\end{equation}
where $\Alp_n \in \mathbb{R}$ are the expansion coefficients.
\vskip 1mm
First order polarization in non orthogonal $(p,q)$ coordinates leads naturally to a financial theory in phase space, with
\begin{equation*}
F(p, q) = e^{\frac{1}{2} A B} = e^{\frac{1}{4} (\lambda^{-2} \, p^2 - \lambda^2 \, x^2)}
\end{equation*}
the analog of a Husimi quasi-probability.
The expression for the quantum harmonic coherent states in the Fock basis and the Bargmann-Segal transform
can be obtained{\footnote{( The zero energy $1/2$ is the change or coordinates in the fiber, $\zeta\rightarrow \zeta \exp( t/2)$}}
from \eqref{coherent2} by mapping the phase space $(p,q)$ into $\mathbb{C}$ by the simple correspondence $p \rightarrow i p$.
\vskip 1mm
\subsection{High Order Polarization}\label{HOP}
\vskip 1mm
\subsubsection{Harmonic Oscillator Equation in Coordinate Space}
\vskip 1mm
The pricing equation can be obtained directly in $x$-space by using a high-order polarization.
\vskip 1mm
The second order operator
\begin{equation*}
X_{P} = X_ t + \frac{1}{2} \Omeg \lambda^{-2} \, X_{x}^{L} \,X_{x}^{L} - \frac{1}{2} \Omeg \lambda^{2} \, X_{p}^{L} \,X_{p}^{L}
\label{cas1ho}
\end{equation*}
is a Casimir operator commuting with all LIVFs.
\vskip 1mm
Since $X_{x}^{L}$ and $X_{p}^{L}$ do not commute, $X_{P}$ defines two higher order
polarizations, $\langle X_{P} , X_{x}^{L} \rangle$ and $\langle X_{P} , X_{p}^{L} \rangle$.
The coordinate space representation is generated by $\langle X_{P} , X_{p}^{L} \rangle$
\begin{subequations}
\label{highpolhoALL}
\begin{align}
X_{p}^{L} \Psi & = 0 \rightarrow \qquad \Psi (\zeta,p,q,t) = \zeta e^{- p x/2} \, \psi(x,t) \nonumber \\
X_{P} \Psi & = 0 \rightarrow \nonumber \\
& \frac{\partial \psi}{\partial t} + \frac{1}{2} \Omeg \left( \lambda^{-2} \frac{\partial^2 \psi}{\partial x^2} - \lambda^2 \, x^2 \, \psi \right)= 0 \label{highpolho}
\end{align}
\end{subequations}
\vskip 1mm
Since in finance one is usually interested in final value problems, we make the change
$\tau = T - t$, where $T$ is a {\it{maturity time}} and $t \le T$. Equation \eqref{highpolho} becomes
\begin{equation}
\qquad \qquad \frac{\partial \psi}{\partial \tau} = H_{I} \, \psi
\label{highpolho2}
\end{equation}
wit $ H_{I}$ the harmonic oscillator Hamiltonian in coordinate space
\begin{equation}
H_{I} \equiv \frac{1}{2} \Omeg \left( \lambda^{-2} \frac{\partial^2 }{\partial x^2} - \lambda^2 \, x^2 \right)
\label{highpolho3}
\end{equation}
$ H_{I}$ can be written in terms of raising and lowering (ladder) operators by using the diagonalizing $SL(2, \mathbb{R})$ matrix \eqref{diagonalM}
\begin{equation}
H_{I} = \frac{1}{2} \Omeg \, ( a^{\dagger} \, a + a \, a^{\dagger})
\end{equation}
with
\begin{equation}
a \equiv \frac{1}{\sqrt{2}} (\lambda^{-1} \hat{p} - \lambda \, \hat{x}) \qquad a^{\dagger} \equiv \frac{1}{\sqrt{2}} (\lambda^{-1} \hat{p} + \lambda \, \hat{x})
\end{equation}
where
\begin{equation}
\hat{p} = \frac{\partial }{\partial x} \qquad \quad \hat{x} = x
\end{equation}
Note that $a, a^{\dagger}$ are the regular ladder operators that are used for the construction of the harmonic oscillator Fock space in quantum mechanics.
The Heisenberg-Weyl commutation relations $[ \hat{p} , \hat{x}] = 1$ imply the commutator $[a, a^{\dagger}] = 1$ since $SL(2, \mathbb{R})$ is an automorphism of the Heisenberg-Weyl subgroup
(equation\eqref{auto1}).
\vskip 1mm
\subsubsection{Polarized Functions in Coordinate Space}
\vskip 1mm
It is well know that the solution of equation \eqref{highpolho2} can be represented as a series in Hermite functions $\varphi _{n}$
\begin{equation}
\psi(\tau, x) = \sum_{n=0}^{\infty} e^{-\Omeg (n +\frac{1}{2}) \tau} \, \Alp_n \, \varphi_n (\lambda x)
\label{highpolho3}
\end{equation}
with $\Alp_n \in \mathbb{R}$ the expansion coefficients.
The Hermite functions $\varphi _{n}$ have the following expression(\cite{abra})
\begin{equation}
\varphi _{n}(x) = \frac{1}{\sqrt {2^{n}n!{\sqrt {\pi }}}} H_{n}(x) e^{-\frac{1}{2} x^{2}}
\label{varphi}
\end{equation}
where $H_n$ are the Hermite polynomials. $\varphi _{n}$ form an orthonormal basis in $L_2(\mathbb{R})$ and fulfill the eigenvalue condition
\begin{equation*}
\left(\frac{\partial^2 }{\partial x^2} - \, x^2 \right) \, \varphi _{n }(x)= -(2 n +1 ) \varphi _{n}(x) \nonumber
\end{equation*}
\vskip 1mm
The relationship between the (first) polarized functions \eqref{coherent2} and the polarized functions \eqref{highpolho3} is given by a modified Bargmann transform
\begin{align}
\varphi _{n} (\lambda \, x) & = \frac{1}{\sqrt {n!{\sqrt {\pi }}}} \times \nonumber \\
& \frac{1}{2 \pi i} \, \int _{-i \infty}^{i \infty } \, dp \, \int _{- \infty}^{ \infty } \, d x \, A^n\, e^{-B^2/2 + \sqrt{2} \lambda \,x B - \lambda^2 \, x^2/2} \, e^ {-A\, B}
\end{align}
where $A=A(p,q)$ and $B=B(p,q)$ are the variables defined in \eqref{diagonalFrac}. The integral over the momentum $p$ is an inverse double sided Laplace transform.
\vskip 1mm
\subsection{Pricing Kernel}
\vskip 1mm
The
the orthogonality of the expansion \eqref{highpolho3} makes possible to express
the pricing kernel $K_I(x,x^\prime, \tau)$, defined by
\begin{equation*}
\Psi(\tau, x) = \int K_I(x,x^\prime, \tau) \Psi(0, x^\prime) d x
\end{equation*}
as a sum of products of Hermite polynomials in $x$ and $x^\prime$
\begin{align}
& K_I(x,x^\prime, \tau) = \lambda \, e^{-\frac{1}{2} \Omeg \tau} \, \sum_{=0}^{\infty} \frac{ (e^{- \Omeg\,\tau})^n} { 2^n n!} \nonumber \\
& \qquad \qquad \,\times \, H_n (\lambda x) \, H_n(\lambda x^\prime ) \, e^{ \frac{1}{2} \lambda^2 \, (x^2 + {x^\prime}^2)}
\label{mehler1}
\end{align}
By applying the {\it{Mehler formula}} (\cite{mehler})
\begin{align*}
& \sum_{=0}^{\infty} \frac{ \rho^n}{ 2^n n!} \, H_n(x) H_n(y) \exp(\frac{1}{2} (x^2 + y^2)) = \nonumber \\
& \quad \frac{1}{\sqrt{1-\rho^2}} \, \exp \left( \frac{4 \rho \, x \, y- (1+ \rho^2) (x^2 + y^2)} { 2 \, (1- \rho^2)} \right) \nonumber
\end{align*}
to the equation \eqref{mehler1}, with $\rho = \exp(- \Omeg \tau)$, one obtains the
Mehler kernel, a generalized bivariate Gaussian probability density
\begin{equation}
K_I(x,x^\prime, \tau) = \frac{\lambda}{\sqrt{ 2 \, \pi \, \sinh \Omeg \tau }} \, \exp({ \frac{1}{2} \,\lambda^2 \, \bf{x} {\bf{A}} {\bf{x}}^ \intercal })
\label{mehlerK}
\end{equation}
where we have made the change $\tau =T -t$, $\tau \ge 0$, and ${\bf{x}} \equiv (x, x^\prime)$. $\bf{A}$ is the $SL(2, \mathbb{R})$ matrix
\begin{equation}
\qquad \quad {\bf{A}} = \begin{pmatrix*}[c]
-\coth \Omeg \tau & \phantom{-}\csch \Omeg \tau \\
\phantom{-}\csch \Omeg \tau & -\coth \Omeg \tau
\end{pmatrix*}
\end{equation}
with $\coth$ and $\csch$ the hyperbolic cotangent and cosecant, respectively.
\vskip 1mm
$\bf{A}$ is singular when $\Omeg \rightarrow 0$. However
\begin{equation*}
\lim_{\Omeg \to 0} \Omeg \, \, {\bf{A}} = \frac{1}{\tau} \, \begin{pmatrix*}[l]
-1& \phantom{-}1\\
\phantom{-} 1& -1
\end{pmatrix*}
\end{equation*}
As expected from the group law \eqref{cocyclebs1z}, the Mehler kernel maps into the heat kernel (hence the Black-Scholes theory) when $\Omeg \rightarrow 0$
\begin{equation}
\lim_{\Omeg \to 0} K_I(x,x^\prime, \tau) = \frac{1}{\sqrt{2 \pi \Sig^2 \, \tau}} \, e^{ \frac{1}{2 \Sig^2 \tau}\, (x - x^\prime)^2}
\end{equation}
\vskip 1mm
\subsubsection{Derivation using LCTs}
\vskip 1mm
The Mehler kernel \eqref{mehlerK} can also be obtained using the LCT associated (\cite{Miller2}, \cite{wolf2} ) to the harmonic oscillator generating matrix \eqref{genmatrixHO}.
From \eqref{ltcdef2}
\begin{equation}
W({\bf{M_H}},x,x^\prime) = \frac{1}{\sqrt{2 \pi b}}\, \exp( -( a {x^\prime}^2 - 2 x x^\prime + d\, x^2)/(2 \,b))
\end{equation}
where $a,b,c,d$ refer to the elements of \eqref{genmatrixHO}. By direct substitution, we get
\begin{equation}
W({\bf{M_H}},x,x^\prime) = K_I(x,x^\prime, \tau)
\end{equation}
\vskip 1mm
A discussion of heat kernels and Mehler-type formulas based on group-invariant solutions can be found in
reference \cite{mehler4}. Chapter 9 of \cite{wolf2} gives Baker-Campbell-Hausdorff relations between pseudo-differential operators for the harmonic oscillator based on
composition of $SL(2, \mathbb{R})$ matrices.
\vskip 1mm
\subsection{Financial Interpretation}
\vskip 1mm
Using the Feynman-Kac formula, we can write the solution of \eqref{highpolho} as
\begin{equation}
\psi(x,t) = \mathbb{E} ( e^{ - \frac{1}{2} \gamma \int_t^{T} X(s) ^2 \, ds }\left. \psi(X_T, T) \right| X(t) = x)
\label{fkac2}
\end{equation}
with $ \gamma \equiv \Omeg^2 / \Sig^2$
and where the expectation is taken with respect to a normal process with volatility $\sigma$. This equation describes a derivative with a payoff that is discounted quadratically
with the oscillator level $x$. A drift term can be easily added with a coboundary generated by $x$.
\vskip 1mm
A more interesting approach is to use the harmonic oscillator as a process describing stocks with non-normal returns with a correlation given by the Mehler Kernel \eqref{mehlerK}.
The reference \cite{hos1} proposes a quantum harmonic oscillator as a model for the market force which draws a stock return
from short-run fluctuations to the long-run equilibrium.
\section*{Introduction}\label{INTRO}
\vskip 1mm
The Group Quantization formalism (\cite{aldaya0}, \cite{wolf11}, \cite{garcia0}, \cite{aldaya22}) is a scheme for constructing a functional space that is an irreducible infinite dimensional representation of the
Lie algebra belonging to a dynamical symmetry group.
\vskip 1mm
This formalism utilizes
Cartan geometries in a framework similar to the Hamiltonian framework, where, in finance, the {\it{coordinates}} represent prices, rates ... and the conjugate momenta operators are the corresponding deltas.
\vskip 1mm
We apply the Group Quantization formalism to a modified $WSp(2, \mathbb{R} )$ group.
$WSp(2, \mathbb{R} )$ is the semi direct product of the two-dimensional
real symplectic group {\footnote{The infinite-dimensional representation theory of $SL(2, \mathbb{R})$ was used by Bargmann ({\cite{bargmann1}, \cite{bargmann2})
for the description of the free non relativistic quantum mechanical particle.}}
$Sp(2, \mathbb{R}) \approx SL(2, \mathbb{R})$
by the Heisenberg-Weyl group.
Our modification consists in that the embedded 2-dimensional translation subgroup in the Heisenberg-Weyl group has
been extended by $\mathbb{R}^{+}$, rather than $U(1)$, which is the usual extension group in the physics and mathematical literature.
\vskip 1mm
Our interest in the $WSp(2, \mathbb{R} )$ group (called {\it{Group of Inhomogeneous
Linear Transformations}} in \cite{wolf2}) is that this group is the symmetry group of the second-order parabolic differential equations.
Using the $WSp(2, \mathbb{R} )$ symmetry, one can find coordinate systems and operators that map the
equations of motion and the corresponding solutions (\cite{Miller2}, \cite{wolf2}).
\vskip 1mm
The use in finance of mathematical methods previously developed in in physics is often based in the formal similarities
between
the Black-Scholes equation and the quantum mechanical Schrodinger equation. These formal similarities have been explored both from the point of view of Lie algebra invariance (\cite{Miller}, \cite{lie1},\cite{lie2}, \cite{kozlov} ) and global symmetries of the Black-Scholes Hamiltonian operator( \cite{conformal1}, \cite{conformal2},\cite{conformal3},\cite{conformal4}).
\vskip 1mm
However,
the mathematical properties of a wave equation solution such as the Schrodinger equation
are totally different than the properties exhibited by a parabolic differential equation.
Moreover, hermiticity and unitarity do not play a prominent role in finance:
unlike the case of quantum mechanics, in finance there is no probabilistic interpretation of solutions of the pricing equation, and time evolution is irreversible
and non-unitary.
\vskip 1mm
The Group Quantization formalism makes the most of the principal fiber bundle structure linked to the central extension of a Lie group. Although this formalism shares some
features with the {\it{Geometric Quantization}} scheme (\cite{geom1}, \cite{geom2}, \cite{geom3}), contrary to Geometric Quantization,
the Group Quantization formalism does not require the previous existence of a Poisson algebra. In both formalisms, the word {\it{quantization}} signifies {\it{irreducible representation}}.
\vskip 1mm
As an example of the applicability of this formalism to finance, we will obtain the Black-Scholes theory,
the Ho-Lee model and the Euclidean attractive and repulsive oscillators.
\vskip 1mm
\subsection{Outline}
\vskip 1mm
Section \ref{NUM} shows that a Galilean transformation on the space of Black-Scholes solutions constitutes a {\it{numeraire change}}. Strict Galilean invariance plays the role of
phase invariance in quantum mechanics.
\vskip 1mm
We describe the main features of Group Quantization in section \ref{qgroupform}. Particularly important are the definition of the connection form, the polarization algebra and the concept of higher polarization. Group Quantization considers the action of the group on itself, as opposed to the group acting on an external manifold. This guarantees the existence of two sets of commuting generators, the right invariant fields and the left invariant fields. Left invariant fields
provide naturally a set of {\it{polarization}} constraints that result in pricing equations, while the right invariant fields provide operators compatible with these constraints.
\vskip 1mm
Section \ref{SL2RGROUP} gives a brief survey of the $WSp(2, \mathbb{R})$ group, the $SL(2, \mathbb{R})$ group and the Linear Canonical transformations. By using a
$\mathbb{R}^{+}$ central extension to create the embedded Heisenberg-Weyl group, Group Quantization provides the appropriate commutator between the momentum
and coordinate operators $[\hat{p},\hat{x}]= 1$, compatible with $\hat{p}$ representing a delta, rather than the quantum-mechanical $[\hat{p},\hat{x}]= - i \hslash$, without resorting to analogies with any quantum theory or rotating to a fictitious {\it{Euclidean time}}.
\vskip 1mm
Section \ref{BSQUANT} applies the Group Quantization formalism to a parabolic $SL(2, \mathbb{R})$ subgroup in order to construct the Black-Scholes theory. We obtain polarization constraints, operators and pricing Kernel both in the momentum space and in coordinate space. This Heisenberg-Weyl commutator makes the bilateral Laplace transform the integral transform mapping momentum and coordinate spaces. We give examples of pricing in momentum space in appendix \ref{BRAKET}.
\vskip 1mm
Sections \ref{IRQUANT} to \ref{ROSGROUP} discuss the application of Group Quantization to theories with {\it{deformed}} Black-Scholes equations, i.e., Black-Scholes with a (at most quadratic) potential.
\vskip 1mm
The Group Quantization of the linear potential, that in finance represents the Hee-Lo interest rate theory, is discussed in section \ref{IRQUANT}. We obtain the polarized functions and pricing equations both in momentum space and in coordinate space. We compare some of our results with the similarity methods developed in \cite{Miller2} and \cite{wolf2}.
\vskip 1mm
The harmonic and repulsive oscillators are studied in sections \ref{OSGROUP} and \ref{ROSGROUP}. The Heisenberg-Weyl commutator $[\hat{p}, \hat{q}] = 1$ implies that the harmonic
oscillator is generated by a hyperbolic $SL(2, \mathbb{R})$ subgroup, a change of scale operator, not a rotation. We discuss in detail the polarization constraints in phase space and in a {\it{real}} analog of the Fock space. In spite of
our non-standard $[\hat{p}, \hat{q}] = 1$, we recover the usual ladder operators when we construct higher order polarized functions. The Mehler kernel is obtained first by building the Kernel from
the (Hermite) polarized functions, and later by applying the Linear Canonical Transformation technique described in \cite{wolf2}.
\vskip 1mm
From the standpoint of the Group Quantization formalism, the repulsive oscillator can be obtained from the harmonic and oscillator by making the
frequency parameter pure imaginary. Both oscillators, which as an interest rate theory
represent quadratic interest rates, have also been recently proposed as models for stock returns ( \cite{hos1}, \cite{repul1}).
\vskip 1mm
Appendix \ref{DEF} contains some general definitions in order to make this paper more self-contained and to establish notation.
Appendix \ref{CLASSA} shows examples of the Lagrangian formalism and the relationship
of the connection form in the Group Quantization formalism and
the Poincar{\'e}-Cartan form in classical mechanics.
\vskip 1mm
We have favored clarity over mathematical rigor, and important topics such as group cohomology are just glossed over.
For simplicity, we use coordinates as much as possible instead of a more compact notation.
We refer the reader to the publications in the bibliography for
a detailed and rigorous treatment of the mathematical concepts relevant to this article.
\section{Linear Potential. Ho-Lee Model} \label{IRQUANT}
\subsection{Quantization Group}
\vskip 1mm
The $SL(2, \mathbb{R})$ parabolic shear mapping \eqref{genmatrixBSa} that has been used in the Black-Scholes theory will be used in this section
for constructing the linear potential model.
\vskip 1mm
The composition law for the new quantization group $\tilde{I}$ is
\begin{align}
t^{\prime \prime} & =t + t^\prime \nonumber \\
p^{\prime \prime} & = p + p^{\prime } \nonumber \\
x^{\prime \prime} & =x + x^{\prime } + \sigma^2 p^{\prime} t \nonumber \\
\zeta^{\prime \prime} & =\zeta^{\prime } \zeta e^{ \epsilon_{IR}(g,g^\prime) }
\label{cocycleIRa}
\end{align}
where $ t,p,x\in {\mathbb{R}}$ and $\zeta \in {\mathbb{R}}^{+}$
with the cocycle $\epsilon_{IR}$
\begin{align}
\epsilon_{IR}(g,g^{\prime}) & = \epsilon_{G}(g,g^{\prime}) \, + \, \epsilon_{N}(g,g^{\prime}) + \, \epsilon_{I}(g,g^{\prime})
\end{align}
$\epsilon_{G}(g,g^{\prime})$ is the Galilean cocycle in equation \eqref{galileancocycle} and $\epsilon_{N}(g,g^{\prime}) $ is the coboundary term given in \eqref{galileancob}.
We have added a new true cocycle (not a coboundary) $\epsilon_{\beta}(g,g^{\prime}) $
\begin{equation}
\epsilon_{\beta}(g,g^{\prime}) = \beta\, t \, (x^{\prime} + \frac{1}{2} \sigma^2 {p^\prime} t ) \qquad \qquad \beta \in {\mathbb{R}}
\label{cocyclebs2ira}
\end{equation}.
\vskip 1mm
We interpret the group coordinates{\footnote{Identifying group parameters is usually done after analyzing the polarization, however in this case the identification is simple enough.}} as $x$ representing a short rate,
$p$ its conjugate momentum, and $t$ the calendar time.
The parameter $\sigma$ is the short rate volatility.
\vskip 1mm
We will verify in the next sections that this group describes the Ho–Lee interest rate model.
In quantum physics, the $U(1)$ extension of this group describes the free fall (linear potential).
The Group Quantization formalism applied to the linear potential in physics can be found in \cite{wolf11}.
\vskip 1mm
\subsection{Lie Algrebra}{\label{IRVF}}
\vskip 1mm
\subsubsection{Left Invariant Vector Fields}
\vskip 1mm
From the composition law \eqref{cocycleIRa} and \eqref{cocyclebs2ira}, we obtain
\begin{align}
{X_{\zeta}^L} & \equiv \Xi = \zeta \frac{\partial}{\partial \zeta} \nonumber \\
{X_{x}^L} & = \frac{\partial}{\partial x} \qquad \qquad {X_{p}^L} = \frac{\partial}{\partial p} + x \, \Xi \nonumber \\
{X_{t}^L} & = \frac{\partial}{\partial t} + \sigma^2 p \frac{\partial}{\partial x} +(E(p)+ \beta x )\, \Xi
\label{leftXir}
\end{align}
where
\begin{equation}
E (p) \equiv \frac{1}{2} \, \sigma^2 p^2 + \mu \, p
\end{equation}
with the non-zero Lie Brackets
\begin{align}
[{X_{t}^L} , {X_{x}^L} ] & = -\beta \, \Xi & [{X_{t}^L} , {X_{p}^L} ] & = - \sigma^2 {X_{x}^L} -\mu \Xi \nonumber \\
[{X_{p}^L} , {X_{x}^L} ] & = - \Xi & \qquad
\label{liebraketsir0}
\end{align}
Note that, as opposed to the Black-Scholes case (section \ref{IVF}), the time generator and the $x$-generator do not commute.
\vskip 1mm
\subsection{Connection}
\vskip 1mm
The vertical form $\Theta$ and the curvature form $ d \Theta $ are given by
\begin{align}
& \Theta = -x \,d p - E(p) \, dt -\beta \,x \,dt + d \Xi \nonumber \\
& d \Theta = d p \wedge d x - ( \sigma^2 p + \mu) \, dp \wedge dt -\beta \, d x \wedge d t
\label{theta2ir}
\end{align}
with
\begin{equation}
d\, \Xi \equiv\frac{1}{ \zeta} d \zeta
\end{equation}
\vskip 1mm
\subsubsection{Characteristic Module}\label{charmod}
\vskip 1mm
$\mathcal{C}_\Theta$
is spanned by a unique field $X_C $
\begin{align}
& X_C = {X_{t}^L} + \mu \, {X_{q}^L} - \beta \, {X_{p}^L} = \\ \nonumber
& \frac{\partial}{\partial t} - \beta \frac{\partial}{\partial p} + (p\,\sigma^2 + \mu) \, \frac{\partial}{\partial x} + ( \frac{1}{2} \,\sigma^2 \, p^2 + \mu p) \, \Xi
\label{mod2bbir}
\end{align}
\vskip 1mm
\subsection{First Order Polarization}{\label{FPOIR}
\vskip 1mm
\subsubsection{Ho-Lee Equation in Momentum Space}
\vskip 1mm
As in the Black-Scholes case, there is only one first order $p$-space polarization $\mathcal{P}$ spanned
by the $x$-translations and the $\mathcal{C}_\Theta$ generator $X_C$.
\vskip 1mm
The polarized functions $\Psi \in {\mathcal{F_P}}$ are found by imposing the polarization constraints
on the functional space ${\mathcal{F_P}}$
\begin{equation*}
{\mathcal{F_P}} = \{\, \Psi : \tilde{I} \rightarrow \mathbb{R} \, / \, \Psi(x,p, t,\zeta) = \zeta \Psi(p,q,t) \, \}
\end{equation*}
\vskip 1mm
One has
\begin{subequations}
\begin{align}
{X_{x}^L} \Psi = 0 & \rightarrow \Psi = \zeta \, \psi(p,t) \\
X_C \Psi = 0 & \rightarrow \nonumber \\
& \frac{\partial \psi}{\partial t} - \beta \, \frac{\partial \psi}{\partial p} + \frac{1}{2} \sigma^2 p^2 \psi + \mu \, p \, \psi = 0 \label{firstpolir}
\end{align}
\end{subequations}
Equation \eqref{firstpolir} is the Ho-Lee Equation in Momentum Space.
\vskip 1mm
The general solution{\footnote{We will show in section \ref{HOP2} that the polarized functions can be expressed in the $x$-space using Airy functions. This can be anticipated by the identity \cite{abra}
\begin{equation*}
\frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, d p \, e^{x p + t p^3} = \frac{1}{ (3 t)^{\frac{1}{3}}} \, Ai\left( \frac{-x}{ (3 t)^{\frac{1}{3}}} \right) \qquad c>0
\end{equation*}
The solutions that are not proportional to $Ai$ grow very rapidly at infinity and they need to be constructed
with different contours in the complex plane.}}
of \eqref{firstpolir} is
\begin{equation}
\psi(p, t) = \Phi(p+\beta t) \, e^{ \frac{1}{2 \beta} \mu p^2 + \frac{1}{6 \beta} \sigma^2 p^3}
\label{firstpolir0}
\end{equation}
where $\Phi(p)$ is an arbitrary function of the momentum $p$.
\vskip 1mm
The bilateral Laplace transform relates the $p$-space and the $x$-space. From equation \eqref{firstpolir0}, the general expression of a polarized function in $x$-space can be written as
\begin{equation}
\Psi(\zeta, x, t) = \zeta \, e^{ - \beta t x} \, \frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, dp\, G(p ,t) \, e^{ x p} \, \Phi(p)
\label{price1bb}
\end{equation}
where
\begin{equation}
G (p, ,t) \equiv e^{\frac{1}{2 \beta } \mu (p-\beta t)^2 + \frac{1}{6 \beta} \sigma^2 (p-\beta t)^3}
\label{price1bb2}
\end{equation}
The integration contour is a
vertical line on the complex plane, such that all singularities of the integrand lie on the left of it.
\vskip 1mm
\subsection{High Order Polarization.}\label{HOP2} \label{HOLEEEQ}
\vskip 1mm
\subsubsection{Ho-Lee Equation in Coordinate Space}
\vskip 1mm
The pricing (evolution) equation can be found directly in $x$-space by using a {\it{high-order}} polarization (see section {\ref{SecondPO}}).
\vskip 1mm
Let $X_C$ be the $\mathcal{C}_\Theta$ generator. The second order operator
\begin{equation}
X_{P} = X_C + \frac{1}{2}\,\sigma^2 \, X_{x}^{L} \,X_{x}^{L}
\label{cas1ir}
\end{equation}
is a Casimir operator commuting with all LIVFs (equation \refeq{leftXir}). Since $X_{x}^{L}$ and $X_{p}^{L}$ do not commute, $X_{P}^{L}$ defines two higher order
polarizations, the $p$-space polarization ${\mathcal{P}}_{p} = \{ \, X_{P}^{L} , X_{x}^{L} \}$, and the $x$-space polarization ${\mathcal{P}}_{x} = \{ \, X_{P}^{L} , X_{p}^{L} \}$.
\vskip 1mm
The $x$-space polarized functions are obtained by imposing the ${\mathcal{P}}_{x}$ polarization constraints on functions (sections) of the form $\Psi(z,p,x,t) = \zeta \Psi(p,x,t)$
\begin{subequations}
\begin{align}
{X_{p}^L} \Psi = 0 & \rightarrow \Psi = \zeta \, V (x,t) \\
X_P \Psi = 0 & \rightarrow \nonumber \\
& \frac{\partial V}{\partial t} + \frac{1}{2} \sigma^2 \frac{\partial^2 V}{\partial x^2} + \mu \frac{\partial V}{\partial x}+ \beta \, x \, V = 0 \label{highpolir}
\end{align}
\end{subequations}
Equation \eqref{highpolir} is the Ho-Lee equation in coordinate space.
\vskip 1mm
If we write
\begin{equation}
V(x,t) = \, \frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, dp\, \Phi(p,t) \, e ^{p x}
\label{highpolir2}
\end{equation}
and substitute in \eqref{highpolir}, we recover the first order polarization equations \eqref{firstpolir}.
\vskip 1mm
Using the Feyman-Kac{\footnote{
Consider the differential operator $L$
\begin{equation*}
L \equiv \frac{1}{2} {\omega^2}(x,t)\frac{\partial^2}{\partial x^2} + \mu(x,t)\frac{\partial}{\partial x}
\end{equation*}
where $\mu(x, t), \omega(x, t)$ and $r(x, t)$ are functions of $(x, t)$ , with $x$ defined in a real
domain $D \subset \mathbb{R}$ and $t$ a positive real number, $t \in [0, T] $
Then, subject to technical conditions, the unique solution of the PDE
\begin{equation*}
(\frac{\partial}{\partial t} + L + r(x,t) ) \, f(x,t) = 0 \quad x \in D, 0 \geq t \leq T
\end{equation*}
with terminal value $f(x, T ) = g(x)$, is given by the {\it{Feynman-Kac Formula}}
\begin{equation*}
f(x,t) = \mathbb{E} ( e^{- \int_t^{T} r(X,s) ds }\left. g(X_T) \right| X(t) = x)
\end{equation*}
where the expectation is taken with respect to the transition density induced
by the SDE
\begin{equation*}
d X = \mu(X, t) dt + \omega (X, t) d W
\end{equation*}
with $W$ a Brownian motion.
}} theorem, the the solution of \eqref{highpolir} can be written as
\begin{equation}
V(x,t) = \mathbb{E} ( e^{- \beta \int_t^{T} X(s) ds }\left. \phi(X_T) \right| X(t) = x)
\label{fkac}
\end{equation}
where the expectation is taken with respect to a normal process with volatility $\sigma$ and
drift $\mu$ ($W$ is a Brownian montion)
\begin{equation*}
d X = \mu dt + \sigma\, d W
\end{equation*}
\vskip 1mm
\subsection{Airy Expansion}
\vskip 1mm
After performing the following change of function and independent variable
\begin{align}
& V(x,t) \equiv e^{-\frac{\mu}{\sigma^2}\, x} \, e^{\lambda \,t} U(y) \\ \nonumber
& y \equiv \left(\frac{2}{\sigma^2} \right)^{\frac{1}{3}} \, ( \frac{\mu}{2 \sigma^2} - \lambda - \beta \, x )
\label{airy4}
\end{align}
the polarization equation \eqref{highpolir} becomes the Airy equation
\begin{equation}
\frac{\partial^2 U}{\partial y^2} - y \, U(y) = 0
\label{airy3}
\end{equation}
Therefore, the polarized functions can then be written as a superposition of Airy functions{\footnote{There is a distinguished solution of equation \eqref{airy3},
called $Ai$, that decays rapidly as $y \to + \infty$, while a second linearly independent solution $Bi$ grows rapidly in this limit. Also, like Bessel functions,
both $Ai$ and $Bi$ are oscillatory with a slow decay for large values of their arguments. See reference \cite{abra} }}
\begin{equation}
V(x,t) = e^{-\frac{\mu}{\sigma^2}\, x} \, \sum_{i=0}^{\infty} \, e^{\lambda_i \,t} \, ( a_i \, Ai(y_i) + b_i \, Bi(y_i))
\label{highpolir4}
\end{equation}
where $ a_i, b_i, \lambda_i \in {\mathbb{R}}$ and
\begin{equation*}
y_i \equiv (\frac{2}{\sigma^2})^{\frac{1}{3}} \, (\frac{\mu}{2 \sigma^2} - \lambda_i - \beta \, x)
\end{equation*}
\vskip 1mm
This form is convenient for the analysis of complex boundary conditions in bond pricing problems, as shown in \cite{goldstein}. The boundary conditions constrain the values
for the expansion
coefficients $a_i, b_i$ and $\lambda_i$.
\vskip 1mm
Note that
the group contraction $\beta \rightarrow 0$ is smooth and the group law \eqref{cocycleIRa} reduces to the Black-Scholes group law,
however $\beta = 0$ is a singular point in the integral for the momentum polarized functions, \eqref{firstpolir0}
and the polarization equation \eqref{firstpolir}.
\vskip 1mm
\subsection{Pricing Kernel}
\vskip 1mm
\subsubsection{Similarity Methods}
\vskip 1mm
One can use the methods developed in \cite{Miller2} and \cite{wolf2} in order to find coordinate changes
mapping Black-Scholes solutions into solutions of the equation \eqref{highpolir}
\vskip 1mm
For instance, it can be shown that if
$V_{BS}(x,t)$ a solution of the Black-Scholes equation \eqref{hpolbs} with $r = 0$, then
\begin{equation}
V(x,t) = \Omega_\beta(x,t) \, \, V_{BS}( x - \frac{1}{2} \beta \sigma^2 t^2 ,t) \
\end{equation}
with
\begin{equation}
\Omega_\beta(x,t) = e^{\beta x t + \frac{1}{6} \beta^2 \sigma^2 t^3 + \frac{1}{2} \mu \beta t^2 }
\label{omega1}
\end{equation}
is a solution of \eqref{highpolir}. Hence, the Ho-Lee pricing Kernel can be obtained from the Black-Scholes pricing Kernel \eqref{kernellag1} (with $r = 0$) in an analogous manner
\begin{equation}
\label{kernellag2}
K_I(x,x^\prime, \tau) = \, \Omega_\beta(x,\tau) \, K_{BS}(x - \frac{1}{2} \beta \sigma^2 \tau^2, x^\prime, \tau)
\end{equation}
where we have switched to the variable $\tau = T-t$.
\vskip 1mm
Notice that $\Omega_\beta(x,\tau)$ is a solution of \eqref{highpolir}. In fact, $\Omega_1(x,\tau) $ is the expression for the Ho-Lee bond maturing at $T$
\begin{equation}
\Omega_1(x,\tau) = e^{-x (T-t) - \frac{\sigma^2}{6} (T-t)^3 + \frac{\mu}{2} (T-t)^2} \qquad t \in [0, T]
\end{equation}
\vskip 1mm
\subsubsection{Pseudo-Differential Operators}\label{EXPECT2}
\vskip 1mm
The pseudo-differential operator methods in section \ref{EXPECT} can also be used
to compute the evolution operator $U$.
\vskip 1mm
From equation \eqref{highpolir}.
\begin{equation}
U(\tau) = \exp(\tau H_I) \qquad H_I \equiv \frac{1}{2} \sigma^2 \frac{\partial^2 }{\partial x^2} + \mu \frac{\partial }{\partial x} + \beta \, x
\end{equation}
where $\tau = -t$. We split $H_I$ into two operators
\begin{equation}
A \equiv \frac{1}{2} \sigma^2 \frac{\partial^2 }{\partial x^2} + \mu \frac{\partial }{\partial x} \quad \qquad B = \beta \, x
\end{equation}
Given that the only non zero commutators are
\begin{equation*}
[A,B] = \beta \sigma^2 \frac{\partial }{\partial x} + \mu\,\beta \qquad [B,[A,B]] = -\beta^2 \, \sigma^2
\end{equation*}
it is possible to apply the {\it{left oriented extended}} version of the
Baker-Campbell-Hausdorff formula \cite{zazen}
\begin{equation}
e^{\tau A + \tau B} = e^{\frac{1}{3!} \tau^3 ( 2 [B,[A,B]] + [A,[A,B]])} \, e^{\frac{1}{2} \tau^2 [A,B]} \, e^{\tau B} \, e^{\tau A}
\end{equation}
We obtain{\footnote{We have interchanged the exponential factors
\begin{equation*}
e^{\frac{1}{2} \tau^2 \beta \sigma^2 \frac{\partial }{\partial x}} \, e^{\beta \tau x}
\end{equation*}
since the commutator is a number
\begin{equation}
e^{\frac{1}{2} \tau^2 \beta \sigma^2 \frac{\partial }{\partial x}} \, e^{\beta \tau x} = e^{\frac{1}{2} \tau^3 \beta^2 \sigma^2 }\, e^{\beta \tau x} \, e^{\frac{1}{2} \tau^2 \beta \sigma^2 \frac{\partial }{\partial x}}
\end{equation}}}
\begin{align}
e^{\tau H_I} & = e^{- \frac{1}{3} \beta^2 \, \sigma^2 \tau^3} \, e^{\frac{1}{2} \mu\,\beta \tau^2} \nonumber \\
& \times \,\, \, e^{\frac{1}{2} \tau^2 \beta \sigma^2 \frac{\partial }{\partial x}} \, e^{\beta \tau x} \,
e^{\frac{1}{2} \sigma^2 \tau \, \frac{\partial^2 }{\partial x^2} } \, e^{\mu \tau \, \frac{\partial }{\partial x} } \nonumber \\
& = \Omega_{\beta}(x,\tau) \, e^{\frac{1}{2} \tau^2 \beta \sigma^2 \frac{\partial }{\partial x}} \, e^{\frac{1}{2} \sigma^2 \tau \, \frac{\partial^2 }{\partial x^2} }\, e^{\mu \tau \, \frac{\partial }{\partial x} }
\label{operlo}
\end{align}
where $\Omega_{\beta}(x,\tau) $ has been defined in \eqref{omega1}.
We apply the operators in \eqref{operlo} from right to left, and use the Weierstrass theorem to recover equation \eqref{kernellag2}
\begin{equation}
\label{kernellag2a}
K_I(x,x^\prime, \tau) = \Omega(x,\tau) \, K_{BS}(x - \frac{1}{2} \beta \sigma^2 \tau^2, x^\prime, \tau)
\end{equation}
\section{Harmonic Oscillator}\label{OSGROUP}
\input{harmonic_oscillator}
\section{Repulsive Oscillator}\label{ROSGROUP}
\input{repulsive_oscillator}
\section{Conclusions}
\vskip 1mm
We have presented a methodology, the Group Quantization formalism, for constructing a financial theory from symmetry arguments.
The $WSp(2, \mathbb{R} )$ group has been used to describe the Black-Scholes model, the Ho-Lee model
and the harmonic and repulsive oscillators.
\vskip 1mm
The choice of $\mathbb{R}^{+}$ as the structural group in the Heisenberg-Weyl group ensures the appropriate commutator between the momentum
and coordinate operators $[\hat{p},\hat{x}]= 1$. This choice is compatible with the financial interpretation of $\hat{p}$ as a delta.
\vskip 1mm
In addition, extending translations by $\mathbb{R}^{+}$
formulates the theory directly in real (calculation) time, without the need to
switch to an Euclidean time in order to derive pricing quantities.
\vskip 1mm
The case of Black-Scholes has been studied in detail. We have constructed the polarized functions and the position, momentum and Hamiltonian operators both in
coordinate and in momentum spaces. The role of variance as cohomological invariant has been identified.
\vskip 1mm
First polarizations for the harmonic oscillator have been derived in phase space and in orthogonal coordinates. A high polarization was necessary to obtain polarized functions
in coordinate (price) space.
\vskip 1mm
Group Quantization generates representations of a financial theory in different functional spaces, allowing for alternative pricing frameworks
such as the use of Laplace and Mellin transforms.
In the case of Black-Scholes and Ho-Lee,
the polarized functions have the meaning of prices of financial instruments, while their meaning for the oscillators is object of active research.
\vskip 1mm
As we have shown in this article, among other interesting features, the Group Quantization formalism
provides naturally the functional constraints (polarization algebra or higher polarization) for deriving the pricing equations.
This makes Group Quantization a versatile methodology
for constructing a financial theory from symmetry arguments alone, using solely the
principal bundle structure of a centrally extended Lie group.
\newpage
\begin{appendices}
\input{geometric}
\input{black_scholes_prices_in_laplace_space}
\input{black_scholes_lagrangian}
\end{appendices}
\section*{Acknowledgements}
\input{acknowledgements}
\include{references}
\end{document}
\section{Galilean Transformations as Numeraire Change }\label{NUM}
As a motivation for this work, we show that the action of the Galilei group on the space of Black-Scholes solutions constitutes a {\it{numeraire change}}
\vskip 1mm
The Black-Scholes equation for a stock that pays no dividends is
\begin{equation}
\frac{\partial V}{\partial t} = - \frac{1}{2} \sigma^2 S^2 \frac{\partial ^2 V}{\partial S ^2} - r S \frac{\partial V}{\partial S} + r \, V
\label{bs1m1}
\end{equation}
where
$\sigma$ is the stock volatility, $r$ is the risk free rate, $S$ is the stock price, $t$ is time, and $V = V(S, t)$ is the price of a financial instrument. For simplicity, let's assume that $r$ and $\sigma$ are constant.
Under the change of variables
\begin{equation}
S \rightarrow S^\prime = S e^{v^\prime t^\prime} \qquad t \rightarrow t^\prime
\label{bs1m2}
\end{equation}
where $v^\prime$ is constant, the Black-Scholes equation becomes
\begin{equation}
\frac{\partial V^{\prime}}{\partial t^\prime} = - \frac{1}{2} \sigma^2 {S^\prime}^2 \frac{\partial ^2 V^{\prime}}{\partial {S^\prime}^2} -
(r + v^\prime ) S^{\prime} \frac{\partial V^{\prime} }{\partial S^{\prime}} + r\,V^{\prime}
\label{bs1m3}
\end{equation}
where $ V^\prime = V(S^\prime, t^\prime)$.
\vskip 1mm
The transformations \eqref{bs1m2} can be expressed in log-stock coordinates as
\begin{equation}
\label{bs1m2b}
x^{\prime } = x + v^\prime, \, t \quad v^{\prime } = v \quad t^{\prime } = t
\end{equation}
were $x \equiv \ln(S)$. We recognize the familiar Galilean transformations, with the {\it{position}} $x$ representing the logarithm of the stock price, and the {\it{velocity}} $v$ the stock's growth rate.
Given that the action of the Galileo group is linear in the log-stock cooordinates, it is convenient to express the Black-Scholes equation \eqref{bs1m1} in terms of the logarithm of the stock price
\begin{equation}
\frac{\partial V}{\partial t} = - \frac{1}{2} \sigma^2 \frac{\partial ^2 V}{\partial x^2} - \mu \frac{\partial V }{\partial x} + r\, V
\label{bs1m1b}
\end{equation}
with $\mu = r - \frac{1}{2} \sigma^2$.
Equation \eqref{bs1m3} becomes
\begin{equation}
\frac{\partial V^{\prime}}{\partial t^\prime} = - \frac{1}{2} \sigma^2 \frac{\partial ^2 V^{\prime}}{\partial {x^\prime}^2} -
(\mu + v^\prime ) \frac{\partial V^{\prime} }{\partial x^{\prime}} + r \, V^{\prime}
\label{bs1m2}
\end{equation}
where $ V^{\prime} = V(x^\prime, t^\prime)$
\vskip 1mm
The Black-Scholes equation is only covariant under the action of the Galilei group.
Strict Galilean invariance can be restored by multiplying the solution $V^\prime$ by an exponential factor
\begin{align}
\bar{V} (x^\prime, t^\prime) & = e^{\epsilon(x^\prime, t^\prime )} \, V^\prime (x^\prime, t^\prime) \nonumber \\
& \epsilon(x^\prime, t^\prime ) = \frac{v^\prime } { \sigma^2} \left( \frac{1}{2}\, v^\prime \, t^\prime - \left( x^\prime - \mu t^{\prime} \right) \right)
\label{sch3d}
\end{align}
One has{\footnote{
Note that
$\Phi(x,t) \equiv \exp(r t^\prime + \epsilon(x^\prime, t^\prime)) $ is a solution of equation \eqref{bs1m2}
}}
\begin{equation*}
\frac{\partial \bar{V} }{\partial t^{\prime}}= - \frac{1}{2} \sigma^2 \frac{\partial ^2 \bar{V} }{\partial {x^\prime} ^2} - \mu \frac{\partial \bar{V} }{\partial {x^\prime}} + r \, \bar{V}
\label{bs3mb}
\end{equation*}
\vskip 1mm
Numeraire invariance is analogous to {\it{Phase Invariance}} in quantum mechanics,
However,
the relevant Black-Scholes gauge group is not the quantum mechanical $U(1)$, but rather the multiplicative positive real line
$ {\mathbb{R}}^{+}$.
\vskip 1mm
Exponential gauge factors ({\it{cocycles}}) analog to the one in equation \eqref{sch3d} play a pivotal role in the {\it{Group Quantization}} formalism.
\section{The $WSp(2, \mathbb{R} )$ Group}\label{SL2RGROUP}
\vskip 1mm
\subsection{Heisenberg-Weyl Group.}\label{HWGROUP}
\vskip 1mm
The Heisenberg-Weyl group $W$
is a central extension of the two dimensional Euclidean translation group. Its algebra
is generated by three elements, $\mathds{ P} $, $\mathds{ Q} $ and $\mathds{ 1}$, where $\mathds{ 1}$ is the identity operator. The
only non trivial commutator is
\begin{equation}
\label{weyl1}
[\mathds{ P},\mathds{ Q}] = \gamma \, \, \mathds{1} \qquad \qquad \gamma \, \in \mathbb{ C}
\end{equation}
It can be proven that the choices for $\gamma$ are equivalent to select $\gamma$ real or $\gamma$ pure imaginary.
\vskip 1mm
Let $p, q \in \mathbb{R}$ , $\theta \in \mathbb{C}$ and define
\begin{equation}
\label{weyl2}
\mathds{W}(\theta,x,q) \equiv \exp( \theta \mathds{1}+ x \, \mathds{ P} + p \mathds{ Q})
\end{equation}
The operators $\mathds{W}$
specify{\footnote{The action of $W$ on $sl(2,\mathbb{R}$ is
\begin{equation}
\mathds{W} (a\, \mathds{Q} + b \, \mathds{P} + \eta \, \mathds{I}) \mathds{W}^{-1} = a\, \mathds{Q} + b \mathds{P} + (\eta \, + w (a \, p - b\, x) ) \mathds{1}
\end{equation}
therefore, $W$ acts on ${\mathfrak{w}}$ as a three-dimensional space with
Cartesian coordinates $(p, x, \theta)$.}} a group composition law under multiplication{\footnote{ Exponentiation and composition of operators are to be considered in the context of formal operator series.}}.
\begin{equation}
\label{mult2}
\mathds{W}^{\prime \prime} ( p^{\prime \prime}, x^{\prime \prime},\theta^{\prime \prime})= \mathds{W}^\prime ( p^{\prime}, x^{\prime},\theta^{\prime}) \, \mathds{ W} ( p, x,\theta)
\end{equation}
Using the the {\it{Baker-Campbell-Hausdorff}} formula{\footnote{The {\it{Baker-Campbell-Hausdorff}} formula reads
\begin{align}
\label{weyl3b}
\ln(e^X e^Y) & = X + Y + \frac{1}{2} [X,Y] + \nonumber \\
& \frac{1}{12} ( [X,[X,Y]] - [Y,[X,Y]]) + \ldots
\end{align}
when $[X,Y]$ is a number
\begin{equation}
\label{weyl2}
e^X e^Y = e^{X + Y + \frac{1}{2} [X,Y]} \qquad \rightarrow \quad e^X e^Y = e^Y e^X e^{ [X,Y]}
\end{equation}
One can prove that
\begin{equation}
\label{weyl3}
e^X e^Y e^{-X} = e^X Y e^{-X }= Y + [X,Y] + \frac{1}{2}[X, [X,Y]] + \ldots
\end{equation}
}}
, we obtain
\begin{align}
p^{\prime \prime} & = p + p^\prime \nonumber \\
x^{\prime \prime} & = x+ x^\prime \nonumber \\
\theta^{\prime \prime} & = \theta + \theta^\prime + \frac{1}{2} \gamma \, (p x^\prime - x p^\prime ) \label{weyl3}
\end{align}
\vskip 1mm
The group law \eqref{weyl3} provides a coordinate representation{\footnote{These are {\it{right invariant}} generators, corresponding to the action of $\mathds{W}^\prime $.}} of the abstract operators $\mathds{ P}$, $\mathds{ Q}$ and $\mathds{1}$ acting on the functional space
\begin{equation}
\Psi(p , q , \theta) = \exp(\theta)\, \psi(p,q)
\end{equation}
that provides the required Heisenberg-Weyl commutation relations
\begin{align}
\mathds{ P} \mapsto X_p & = \frac{\partial}{\partial p} - \frac{1}{2} \gamma \, x \, \frac{\partial}{\partial \theta} \nonumber \\
\mathds{Q} \mapsto X_x & = \frac{\partial}{\partial x} + \frac{1}{2} \gamma \, p\, \frac{\partial}{\partial \theta} \nonumber \\
\mathds{1} \mapsto X_\theta & = \frac{\partial}{\partial \theta}
\end{align}
We get
\begin{equation}
\qquad [ X_p, X_x] = \gamma \, X_\theta
\end{equation}
\vskip 1mm
\subsection{The $SL(2, \mathbb{R})$ group}
\vskip 1mm
$ SL(2, \mathbb{R}) $, the special (unimodular) linear group in two real dimensions,
has a natural representation by $2 \times 2$ real matrices with determinant one
\begin{equation}
\label{amatrix}
{\bf{M}} = \begin{pmatrix}
a & b \\
c & d
\end{pmatrix} \in SL(2, \mathbb{R}) , \quad a d - b c = 1
\end{equation}
\vskip 1mm
The $SL(2, \mathbb{R})$ group is an automorphism of the $W$ group, preserving the algebra commutators. Define new operators $\mathds{P}^\prime$ and $\mathds{Q}^\prime$
as linear combinations of $\mathds{P}$ and $\mathds{Q}$, using the matrix \eqref{amatrix}
\begin{align}
\mathds{P}^{\prime} = & a \, \mathds{P} + b\, \mathds{Q} \nonumber \\
\mathds{Q}^{\prime} = & c \, \mathds{P} + d \, \mathds{Q}
\end{align}
then, by direct computation,
\begin{equation}
[\mathds{P}^{\prime}, \mathds{Q}^{\prime}] = [\mathds{P}, \mathds{Q}]
\label{auto1}
\end{equation}
\vskip 1mm
$ SL(2, \mathbb{R})$ belongs to a class of Lie groups called the symplectic groups $ Sp(2n, \mathbb{R})$
which leave invariant a skew-symmetric form and play
an important role in the geometry of phase space and Hamiltonian systems.
One can verify the isomorphism $SL(2, \mathbb{R}) \approx Sp(2, \mathbb{R})$ for the two-dimensional symplectic matrix $ {\bf{\mathrm{\Omega \, }}}$
\begin{equation}
{\bf{M^\top}} {\bf{\mathrm{\Omega \, }}} {\bf{M}} = {\bf{\mathrm{\Omega \, }}} \qquad \qquad {{\bf{\mathrm{\Omega \, }}}} \equiv \begin{bmatrix}
0 & 1 \\
-1 & 0
\end{bmatrix}
\label{omegadef}
\end{equation}
\vskip 1mm
\subsection{The $WSp(2, \mathbb{R} )$ group}
\vskip 1mm
$WSp(2, \mathbb{R} )$ is a real, non-compact, connected, simple Lie group that is the semi direct product of the two-dimensional
real symplectic group $Sp(2, \mathbb{R})$($\approx SL(2, \mathbb{R})$) by the Heisenberg-Weyl group $W$.
\vskip 1mm
The
$WSp(2, \mathbb{R} )$ group is a subgroup of the Schrodinger group in one dimension, and is called{\footnote{Our composition law differs from \cite{wolf2} in that the extension group in \eqref{slr2a} is $\mathbb{R}^{+}$, not $\mathbb{R}$.}}
{\it{Group of Inhomogeneous
Linear Transformations}} in reference \cite{wolf2}.
\vskip 1mm
Let an element $g \in WSp(2, \mathbb{R} )$ be parametrized by $g \left( {\bf{M}}, \bf{u}, \zeta \right)$
where $ {\bf{M}} \in SL(2, \mathbb{R})$, $ \zeta \in {\mathbb{R}}^{+}$ and
${\bf{u}} \equiv (p,x) \in {\mathbb{R}}^2 $.
The $WSp(2, \mathbb{R} )$ composition law is
\begin{equation}
g^{\prime\prime} \left( {\bf{M}}^{\prime\prime}, \bf{u}^{\prime\prime}, \zeta^{\prime\prime} \right) = g^{\prime} \left( {\bf{M}}^{\prime}, \bf{u}^{\prime}, \zeta^{\prime} \right) \, g \left( {\bf{M}}, \bf{u}, \zeta \right)
\end{equation}
with
\begin{align}
{\bf{M}}^{\prime \prime} & = {\bf{M}}^{\prime } {\bf{M}} \nonumber \\
\qquad \bf{u}^{\prime \prime} & = \bf{u}^{\prime } {\bf{M}} + \bf{u} \nonumber \\
\zeta^{\prime \prime} & = \zeta^\prime \, \zeta \, \exp \left(- \frac{1}{2} \gamma \, \bf{u}^{\prime } {\bf{M}} {\bf{\mathrm{\Omega \, }}} \bf{u}^{T} \right)
\label{slr2a}
\end{align}
where ${\bf{\mathrm{\Omega}}}$ is the two-dimensional symplectic matrix in \eqref{omegadef} and
$\gamma$ is given by the $[\mathds{ P},\mathds{ Q}]$ Heisenberg-Weyl commutator in
\eqref{weyl1}.
\vskip 1mm
\subsubsection{Weyl Commutation Relations}
\vskip 1mm
In the mathematics literature, such as \cite{wolf2} and \cite{Miller}, the usual convention is that the momentum and position operator commutator is $[\mathds{ P},\mathds{ Q}] = - i $,
which corresponds to an extension{\footnote{In quantum mechanics the cocycle in \eqref{slr2a} is multiplied
by a factor $1/\hslash$ for dimensionality reasons, and $[\mathds{ P},\mathds{ Q}] = - i \, \hslash$.}} of the 2-dimensional translations by $U(1)$.
Then, for diffusive equations such as the heat equation, one makes the theory {\it{Euclidean}} at the last step by setting the time variable $t$ to $i \, t$.
\vskip 1mm
In this work we set $\gamma \, = 1$, corresponding to a central extension of the Euclidean translations by $\mathbb{R}^{+}$. The reason is
that in finance we want $\mathds{ P}$ to represent a change (a delta) in the quantity $\mathds{ Q}$, wich usually represents a a price or a rate. This choice has the added advantage that there is
no need to consider a passage to a fictitious {\it{Euclidean}} time, since the time variable in the group represents the actual calculation time.
\vskip 1mm
When we quote results from to the mathematics literature (for instance in section \ref{WST}) we will assume,
unless otherwise indicated, that the results are obtained with the usual commutator $[\mathds{ P},\mathds{ Q}] = - i$.
\vskip 1mm
\subsubsection{Orbits and Subgroups}
\vskip 1mm
The
$WSp(2, \mathbb{R} )$ generates the dynamics for quadratic Hamiltonians (\cite{Miller2}, \cite{Miller}, \cite{wolf11}, \cite{wolf2} ) by the action of higher order operators on $W$.
The adjoint action
of the group consist of six distinct orbits \cite{wolf2}. Representative of these orbits are given by the following operators, that include quadratic elements from the enveloping algebra of $W$
\begin{equation*}
\mathds{P}^2 \, , \, \mathds{ P}^2 + \mathds{Q} \, , \, \mathds{ P}^2 + \mathds{Q}^2 \, , \, \mathds{ P}^2 - \mathds{Q}^2 \, , \, \mathds{P} \, , \, \mathds{1}
\end{equation*}
\vskip 1mm
We have the following isomorphisms: $g \left( {\bf{M}}, {\bf{0}}, 1 \right) \approx SL(2, \mathbb{R})$ and $g \left( {\bf{1}}, \bf{u}, \zeta \right) \approx W$.
The unit element is $g \left( {\bf{1}}, {\bf{0}}, 1 \right)$
and the inverse element is $g \left( {\bf{M}}^{-1}, -{\bf{u}}\, { {\bf{M}}^{-1}}, {1 / \zeta} \right)$
\vskip 1mm
The different quantization groups considered in this paper will be distinguished subroups of $WSp(2, \mathbb{R} )$.
\vskip 1mm
In physics, the
$WSp(2, \mathbb{R} )$ includes as subgroups the symmetry group of the free particle, the gravitational free-fall, as well as
the symmetry group of the ordinary harmonic oscillator and the {\it{repulsive}}
harmonic oscillator (with imaginary frequency).
In finance, we will obtain the Black-Scholes theory,
the Ho-Lee model and the Euclidean attractive and repulsive oscillators.
\vskip 1mm
\subsection{Linear Canonical Transformations}\label{WST}
\vskip 1mm
$WSp(2, \mathbb{R} )$ acts on functional spaces as
integral transforms{\footnote{These transforms also define pseudo-differential (hyperdifferential) operators associated to the corresponding $SL(2, \mathbb{R})$ kernel \cite{wolf2}.}}
called {\it{Linear Canonical Transformations}},
whose kernel is a $SL(2, \mathbb{R})$ matrix \cite{wolf2}
\begin{equation}
g \left( {\bf{M}}, \bf{u} \right) (f(x)) = \int_{ \mathbb{R} } W({\bf{M}}, x,x^\prime) f(x^\prime) d x^\prime
\label{trans1}
\end{equation}
with kernel (using the notation of equation \eqref{amatrix})
\begin{equation}
W({\bf{M}},x,x^\prime) = \frac{e^{-i \pi/4}}{\sqrt{2 \pi b}}\, \exp( i ( a {x^\prime}^2 - 2 x x^\prime + d\, x^2)/(2 \,b))
\label{ltcdef}
\end{equation}
\vskip 1mm
Linear Canonical Transformations (LICs),
have the important property that composition of the transforms is equivalent to
multiplication of their $SL(2, \mathbb{R})$ kernels (\cite{wolf11}, \cite{wolf2}).
\vskip 1mm
LICs kernels can be analytically continued to $SL(2, \mathbb{C})$,
subject to some restrictions \cite{wolf11}. In this work, we will use, when indicated, equation \eqref{trans1} with the mapping $b \rightarrow - i b$
\begin{equation}
W({\bf{M}},x,x^\prime) = \frac{1}{\sqrt{2 \pi b}}\, \exp( - ( a {x^\prime}^2 - 2 x x^\prime + d\, x^2)/(2 \,b))
\label{ltcdef2}
\end{equation}
which has been modified from \eqref{ltcdef} because our choice for the Heisenberg-Weyl commutator $[\mathds{ P},\mathds{ Q}] = 1 $.
\vskip 1mm
The Fourier transform, the Laplace transform, the Bargmann transform and the Mellin transform are examples of LICs.
\vskip 1mm
\subsubsection{Fourier Transform}\label{FT}
The Fourier transform
\begin{equation*}
\label{lm0}
{\mathcal {F}} ( f)(z) = \int _{-\infty}^{\infty } f(p) \, e^{i p z} dp
\end{equation*}
has the $SL(2, \mathbb{R})$ kernel
\begin{equation*}
{\bf{F}} = \begin{pmatrix*}[c]
\phantom{-}0 & 1 \\
-1 & 0
\end{pmatrix*}
\end{equation*}
\subsubsection{Double Sided Laplace Transform}\label{LMT}
The double sided Laplace transform
\begin{equation}
\label{lm0}
{\mathcal {L}} ( f)(z) = \int _{-\infty}^{\infty } f(p) \, e^{ p z} dp
\end{equation}
can be obtained from $ {\mathcal {F}}$ by the formal map $p \mapsto i p$. Its kernel is a $SL(2, \mathbb{C})$ matrix
\begin{equation*}
{\bf{L}} = \begin{bmatrix}
0 & i \\
i & 0
\end{bmatrix}
\end{equation*}
\vskip 1mm
In spite of formal similarities, existence of a Fourier transform exists does not automatically imply the existence of a Laplace transform \cite{wolf2}.
\vskip 1mm
The inverse transform is
\begin{equation}
{\mathcal {L}}^{-1} ( f)(x) = \frac{1 }{2 \pi i} \, \int _{c-i \infty}^{c+ \infty } \, dp \, e^{ x p} \, f(p)
\label{price1bb}
\end{equation}
The integration contour is a
vertical line on the complex plane, such that all singularities of the integrand lie on the left of it.
\vskip 1mm
\subsubsection{Bargmann Transform}\label{BAT}
The Bargmann transform has the complex kernel
\begin{equation*}
{\bf{B}} = \frac{1}{\sqrt{2}} \, \begin{pmatrix*}[c]
\phantom{-} 1 & -i \\
-i & \phantom{-}1
\end{pmatrix*}
\end{equation*}
\vskip 1mm
\subsubsection{Mellin Transform}\label{MT}
\vskip 1mm
The Mellin transform ${\mathcal {M}}$
\begin{equation}
{\mathcal {M}} (f) (z) \equiv F(z) = \int _{0}^{\infty }y^{-z-1} f(y)\,d y
\label{mellin1}
\end{equation}
is obtained from the Laplace transform ${\mathcal {L}}$ by the change of coordinates $p \rightarrow - \ln(y)$.
\vskip 1mm
The resolution of identity
\begin{equation*}
\frac{1}{2 \pi i} \, \int _{c- i \infty}^{c+ i \infty } x^{-s} y^{s} d s= y \delta(x-y)
\end{equation*}
leads to the Mellin inversion formula
\begin{equation*}
{\mathcal {M}}^{-1}(f) (y) \equiv f( y) = \frac{1}{2 \pi i} \, \int _{c-i \infty}^{c+ i \infty } y^{z} F (z) \,dz
\end{equation*}
\vskip 1mm
The Mellin transform has been more widely used in finance
than the double sided Laplace transform (\cite{Mellin0}, \cite{Mellin2}, \cite{Mellin4}, \cite{Mellin5}).
The sign of the transform variable $z$ in \eqref{mellin1} differs from the usual definition in the mathematics literature in order to
calculate transforms of positive powers of the stock price.
\subsection{Financial Interpretation}
\vskip 1mm
Using the Feynman-Kac theorem, the solution of \eqref{highpolhor} can be written as
\begin{equation}
\psi(x,t) = \mathbb{E} ( e^{ \frac{1}{2} \gamma \int_t^{T} X(s)^2 \, ds }\left. \psi(X_T, T) \right| X(t) = x)
\label{fkac3}
\end{equation}
where the expectation is taken with respect to a normal process with volatility $\sigma$. This equation describes a derivative with a payoff that grows quadratically
with the oscillator level $x$.
\vskip 1mm
Reference \cite{repul1} uses a repulsive anharmonic oscillator model to explain the distribution of financial
returns in a stock market when the market exhibits an upward trend.
|
\section{High-level Ideas}\label{sec: high level}
\subsection{High-level ideas}\label{subsec:ideas}
To get a high-level idea of our approach, first consider what could go wrong when running {\small\textsf{\textup{ALG}}}\xspace alone in a non-stationary environment and how to fix that intuitively.
Decompose the dynamic regret as follows:
\begin{align}
\underbrace{\sum_{\tau=1}^t \left(f_\tau^\star - \widetilde{f}_\tau\right)}_{\textbf{term}_1} + \underbrace{\sum_{\tau=1}^t \left(\widetilde{f}_\tau - R_\tau\right)}_{\textbf{term}_2}. \label{eq: initial decomposition}
\end{align}
As mentioned, in a stationary environment, {\small\textsf{\textup{ALG}}}\xspace ensures that $\textbf{term}_1$ is simply non-positive and $\textbf{term}_2$ is bounded by $C(t)$ directly.
In a non-stationary environment, however, both terms can be substantially larger.
If we can detect the event that either of them is abnormally large, we know that the environment has changed substantially, and should just restart {\small\textsf{\textup{ALG}}}\xspace.
This detection can be easily done for $\textbf{term}_2$ since both $\widetilde{f}_\tau$ and $R_\tau$ are observable, but not for $\textbf{term}_1$ since $f_\tau^\star$ is of course unknown.
Note that, a large $\textbf{term}_1$ implies that a policy, possibly suboptimal in the past, now becomes the optimal one with a much larger reward.
A single instance of {\small\textsf{\textup{ALG}}}\xspace run from the beginning thus cannot detect this because suboptimal polices are naturally selected very infrequently.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\textwidth, trim={0.5cm 4.3cm 0cm 3cm}, clip]{figure2.pdf}
\caption{An illustration of how we detect non-stationarity via multiple instances of {\small\textsf{\textup{ALG}}}\xspace}
\label{fig: detection}
\end{figure}
To address this issue, our main idea is to maintain different instances of {\small\textsf{\textup{ALG}}}\xspace to facilitates non-stationarity detection, illustrated via an example in \prettyref{fig: detection}.
Here, there is one distribution change that happens in interval $\mathcal{I}$, making the value of $f_\tau^\star$ (the blue curve) drastically increase.
If within this interval, we start running another instance of {\small\textsf{\textup{ALG}}}\xspace (the red interval),
then its performance (the black curve) will gradually approach $f_\tau^\star$ due to its regret guarantee in a stationary environment.
Hypothetically, if another instance of {\small\textsf{\textup{ALG}}}\xspace run from the beginning could coexist with this new instance, we would see that the latter significantly outperforms the former and infer that the environment has changed.
The issue is that we cannot have multiple instances running and making decisions simultaneously, and here is where the optimistic estimators $\widetilde{f}_\tau$'s can help.
Specifically, since the quantity $U_\tau=\min_{s\leq \tau}\widetilde{f}_s$ (the green non-increasing curve) should always be an upper bound of the learner's performance in a stationary environment, if we find that the new instance of {\small\textsf{\textup{ALG}}}\xspace significantly outperforms this quantity at some point (as shown in \prettyref{fig: detection}), we can also infer that the environment has changed, and prevent $\textbf{term}_1 \leq \sum_{\tau=1}^t (f_\tau^\star - U_\tau)$ from growing too large by restarting.
To formally implement the ideas above, we need to decide when to start a new instance, how long it should last, which instance should be active if multiple exist, and others.
In \prettyref{sec: multi-scale}, we propose a randomized multi-scale scheme to do so, which is reminiscent of the ideas of {sampling obligation} in~\citep{auer2019adaptively} and {replay phase} in~\citep{chen2019new},
although their mechanisms are highly specific to their algorithms and problems.
\section{Algorithm}
\label{sec: multi-scale}
In this section, we first introduce {\small\textsf{\textup{MALG}}}\xspace, an algorithm that schedules and runs multiple instances of the base algorithm {\small\textsf{\textup{ALG}}}\xspace in a multi-scale manner (\prettyref{subsec: multiscale}).
Then, equipping {\small\textsf{\textup{MALG}}}\xspace with non-stationarity detection, we introduce our final black-box reduction {\small\textsf{\textup{MASTER}}}\xspace (\prettyref{sec: final algorithm sec}).
\subsection{{\normalsize\textsf{MALG}}: ~Running the Base Algorithm with Multiple Scales}
\label{subsec: multiscale}
We always run {\small\textsf{\textup{MALG}}}\xspace for an interval of length $2^n$, which we call a \emph{block}, for some integer $n$ (unless it is terminated by the non-stationarity detection mechanism).
During initialization, {\small\textsf{\textup{MALG}}}\xspace uses \prettyref{proc: alg profile} to schedule multiple instances of {\small\textsf{\textup{ALG}}}\xspace within the block in the following way:
for every $m = n, n-1, \ldots, 0$, partition the block equally into $2^{n-m}$ sub-intervals of length $2^m$, and for each of these sub-intervals, with probability $\frac{\rho(2^n)}{\rho(2^m)} \leq 1$ schedule an instance of {\small\textsf{\textup{ALG}}}\xspace (otherwise skip this sub-interval).
We call these instances of length $2^m$ \emph{order-$m$} instances.
Note that by definition there is always an order-$n$ instance covering the entire block.
We use $\inst$ to denote a particular instance of {\small\textsf{\textup{ALG}}}\xspace, and use $\inst.s$ and $\inst.e$ to denote its start and end time.
After the initialization, {\small\textsf{\textup{MALG}}}\xspace starts interacting with the environment as follows.
In each time $t$, the unique instance covering this time step with the shortest length is considered as being \emph{active}, while all others are \emph{inactive}.
{\small\textsf{\textup{MALG}}}\xspace follows the decision of the active instance, and update it after receiving feedback from the environment.
All inactive instances do not make any decisions or updates, that is, they are paused (they might be resumed at some point though).
We use $\widetilde{g}_t$ to denote the scalar $\widetilde{f}_t$ output by the active instance.
See \prettyref{alg: multialg} for the pseudocode.
\begin{procedures}[t]
\caption{A procedure that randomly
schedules {\small\textsf{\textup{ALG}}}\xspace of different lengths within $2^n$ rounds}
\label{proc: alg profile}
{\renewcommand{\nl}{\let\nl} \textbf{input}: $n$, $\rho(\cdot)$} \\
\For{$\tau=0, \ldots, 2^{n}-1$}{
\For{$m=n, n-1, \ldots, 0$}{
If $\tau$ is a multiple of $2^m$, with probability $\frac{\rho(2^n)}{\rho(2^m)}$, schedule a new instance \inst of {\small\textsf{\textup{ALG}}}\xspace \\
{\renewcommand{\nl}{\let\nl} that starts at $\inst.s=\tau+1$ and ends at $\inst.e=\tau+2^m$.}
}
}
\end{procedures}
\begin{algorithm2e}[t]
\caption{{\small\textsf{\textup{MALG}}}\xspace (Multi-scale {\small\textsf{\textup{ALG}}}\xspace)}
\label{alg: multialg}
{\renewcommand{\nl}{\let\nl} \textbf{input}: $n$, $\rho(\cdot)$} \\
\textbf{Initialization}: run \prettyref{proc: alg profile} with inputs $n$ and $\rho(\cdot)$.
At each time $t$, let the unique active instance be $\inst$, output $\widetilde{g}_t$ (which is the $\widetilde{f}_t$ output by $\inst$), follow $\inst$'s decision $\pi_t$, and update $\inst$ after receiving feedback from the environment.
\end{algorithm2e}
\begin{figure}[H]
\includegraphics[width=\textwidth, trim={1cm 1cm 0.3cm 5.2cm}, clip]{figure1.pdf}
\caption{An illustration of {\small\textsf{\textup{MALG}}}\xspace with $n=4$ (see detailed explanation in Section~\ref{subsec: multiscale})}
\label{fig: multiple schedule}
\end{figure}
For better illustration, we give an example with $n=4$ in \prettyref{fig: multiple schedule}.
Suppose that the realization of the random scheduling by \prettyref{proc: alg profile} is:
one order-$4$ instance (red), zero order-$3$ instance, two order-$2$ instances (green), two order-$1$ instances (blue), and five order-$0$ instances (purple).
The bolder part of the segment indicates the period of time when the instances are active, while the thinner part indicates the inactive period.
For example, the red order-$4$ instance is active for the first round, then paused for the next $8$ rounds, and then resumed (from the frozen internal states) for another $3$ rounds before becoming inactive again.
The dashed black arrow marked with \scalebox{0.8}{\circled{1}} indicates that {\small\textsf{\textup{ALG}}}\xspace is executed as if the two sides of the arrow are concatenated. On the other hand,
as another example, the two purple instances on the two sides of the dashed line marked with \scalebox{0.8}{\circled{2}} are two \emph{different} order-$0$ instances, so the second one should start from scratch even though they are consecutive.
One can see that at any point of time, the active instance is always the one with the shortest length.
\paragraph{Regret analysis of {\small\textsf{\textup{MALG}}}\xspace}
The multi-scale nature of {\small\textsf{\textup{MALG}}}\xspace allows the learner's regret to also enjoy a multi-scale structure, as shown in the next lemma (proof deferred to \prettyref{app:multi-scale}).
\begin{lemma}
\label{lemma: multi-scale reg}
Let $\widehat{n}\xspace=\log_2 T+1$ and $\widehat{\rho}(t) = 6\widehat{n}\xspace\log(T/\delta)\rho(t)$.
{\small\textsf{\textup{MALG}}}\xspace with input $n \leq \log_2 T$ guarantees the following:
for any instance \inst that {\small\textsf{\textup{MALG}}}\xspace maintains and any $t \in [\inst.s, \inst.e]$,
as long as $\Delta_{[\inst.s, t]}\leq \rho(t')$ where $t' = t-\inst.s+1$,
we have with probability at least $1-\frac{\delta}{T}$:
\begin{align}
&\widetilde{g}_t \geq \min_{\tau\in [\inst.s, t]} f_\tau^\star - \Delta_{[\inst.s,t]},
\qquad
\frac{1}{t'}\sum_{\tau=\inst.s}^{t} \left(\widetilde{g}_\tau - R_\tau\right) \leq \widehat{\rho}(t') + \widehat{n}\xspace\Delta_{[\inst.s,t]}, \label{eq: complicated bound}
\end{align}
and the number of instances started within $[\inst.s, t]$ is upper bounded by $6\widehat{n}\xspace\log(T/\delta)\frac{C(t')}{C(1)}$.
\end{lemma}
Note that \prettyref{eq: complicated bound} is essentially the analogue of \prettyref{eq: general condition 2} (up to logarithmic terms) with the starting time changed from $1$ to $\inst.s$.
It shows that even if we have multiple instances interleaving in a complicated way,
the regret for a specific interval is still almost the same as running {\small\textsf{\textup{ALG}}}\xspace alone on this interval, thanks to the carefully chosen probability in~\prettyref{proc: alg profile}.
Recall that there is always an order-$n$ instance starting from the beginning of the block, so {\small\textsf{\textup{MALG}}}\xspace is always providing a stronger multi-scale guarantee compared to running {\small\textsf{\textup{ALG}}}\xspace alone.
This richer guarantee facilitates non-stationarity detection as we show next.
\subsection{{\normalsize\textsf{MASTER}}: ~Equipping {\normalsize\textsf{MALG}} with Stationarity Tests}
\label{sec: final algorithm sec}
We are now ready to present our final algorithm {\small\textsf{\textup{MASTER}}}\xspace, short for {\small\textbf{\textsf{MA}}\textsf{LG}} with \textbf{S}tationarity \textbf{TE}sts and \textbf{R}estarts (see \prettyref{alg: final adaptive alg}).
{\small\textsf{\textup{MASTER}}}\xspace runs {\small\textsf{\textup{MALG}}}\xspace in a sequence of blocks with doubling lengths ($2^0, 2^1, \ldots$).
Within each block of length $2^n$ (with $t_n$ being the starting time),
{\small\textsf{\textup{MASTER}}}\xspace simply runs a new instance of {\small\textsf{\textup{MALG}}}\xspace and records the minimum optimistic predictor thus far for this block $U_t = \min_{\tau \in [t_n, t]} \widetilde{g}_\tau$.
At the end of each time, {\small\textsf{\textup{MASTER}}}\xspace performs two tests (\textbf{Test~1}\xspace and \textbf{Test~2}\xspace),
and if either of them returns {\it fail},
{\small\textsf{\textup{MASTER}}}\xspace restarts from scratch.
The two tests exactly follow the ideas described in \prettyref{subsec:ideas} (recall \prettyref{fig: detection}).
Following \prettyref{eq: initial decomposition}, we decompose the regret on $[t_n, t]$ as $\textbf{term}_1 + \textbf{term}_2$ where $\textbf{term}_1 = \sum_{\tau=t_n}^t \left(f_\tau^\star - \widetilde{g}_\tau\right)$ and $\textbf{term}_2 = \sum_{\tau=t_n}^t \left(\widetilde{g}_\tau - R_\tau\right)$.
\textbf{Test~1}\xspace prevents $\textbf{term}_1 \leq \sum_{\tau=t_n}^t \left(f_\tau^\star - U_\tau\right)$ from growing too large by testing if there is some order-$m$ instance's interval during which the learner's average performance $\frac{1}{2^m}\sum_{\tau=\inst.s}^{\inst.e} R_{\tau}$ is larger than the promised performance upper bound $U_t$ by an amount of $9\widehat{\rho}(2^m)$.
On the other hand, \textbf{Test~2}\xspace presents $\textbf{term}_2$ from growing too large by directly testing if its average is large than something close to the promised regret bound $3\widehat{\rho}(t-t_n+1)$.
\begin{algorithm2e}[t]
\caption{{\small\textbf{\textsf{MA}}\textsf{LG}} with \textbf{S}tationarity \textbf{TE}sts and \textbf{R}estarts ({\small\textsf{\textup{MASTER}}}\xspace)}
\label{alg: final adaptive alg}
{\renewcommand{\nl}{\let\nl} \textbf{input}: $\widehat{\rho}(\cdot)$ (defined in \prettyref{lemma: multi-scale reg})}\\
\textbf{Initialize}: $t\leftarrow1$ \\
\For{$n=0, 1, \ldots$}{ \label{line: restart line}
Set $t_n\leftarrow t$
and initialize an {\small\textsf{\textup{MALG}}}\xspace (\prettyref{alg: multialg}) for the block $[t_n, t_n+2^n-1]$. \\
\While{$t< t_n + 2^n$}{ \label{line: start of block}
Receive $\widetilde{g}_t$ and $\pi_t$ from {\small\textsf{\textup{MALG}}}\xspace, execute $\pi_t$, and receive reward $R_t$. \\
Update {\small\textsf{\textup{MALG}}}\xspace with any feedback from the environment, and set $U_t = \min_{\tau \in [t_n, t]} \widetilde{g}_\tau$.
\label{line: tracking repeated} \\
Perform \textbf{Test~1}\xspace and \textbf{Test~2}\xspace (see below).
Increment $t\leftarrow t+1$. \\
\lIf{either test returns \textit{fail}}{
restart from \prettyref{line: restart line}.
}
}
}
\renewcommand{\nl}{\let\nl}
\textbf{Test~1}\xspace: \ \ If $t=\inst.e$ for some order-$m$ $\inst$ and
$\frac{1}{2^m}\sum_{\tau=\inst.s}^{\inst.e} R_{\tau} \geq U_t + 9\widehat{\rho}(2^m)$,
return \textit{fail}.
\textbf{Test~2}\xspace: \ \ If
$\frac{1}{t-t_n+1}\sum_{\tau=t_n}^{t} \left(\widetilde{g}_{\tau} - R_{\tau}\right) \geq 3\widehat{\rho}(t-t_n+1)$,
return \textit{fail}.
\end{algorithm2e}
It is now clear that {\small\textsf{\textup{MASTER}}}\xspace indeed does not require the knowledge of $L$ or $\Delta$ at all.
To analyze {\small\textsf{\textup{MASTER}}}\xspace, we prove the following key lemma that bounds the regret on a single block $[t_n, E_n]$ where $E_n$ is either $t_n+2^n - 1$ or something smaller in the case where a restart is triggered.
\begin{lemma} \label{lemma: bound term1}
With high probability, the dynamic regret of {\small\textsf{\textup{MASTER}}}\xspace on any block $\mathcal{J}=[t_n, E_n]$ where $E_n \leq t_n+2^n -1$ is bounded as
\begin{align}
\sum_{\tau \in \mathcal{J}} \left(f_\tau^\star - R_\tau \right) \leq \widetilde{\mathcal{O}}\left(\sum_{i=1}^{\ell} C(|\mathcal{I}_i'|) + \sum_{m=0}^n \frac{\rho(2^m)}{\rho(2^n)}C(2^m)\right). \label{eq: block regret bound}
\end{align}
where $\{\mathcal{I}_1', \ldots, \mathcal{I}_\ell'\}$ is any partition of $\mathcal{J}$ such that $\Delta_{\mathcal{I}_i'}\leq \rho(|\mathcal{I}_i'|)$ for all $i$.
\end{lemma}
See \prettyref{app: key proofs} for the proof.
When $\rho(t) = \Theta(1/\sqrt{t})$ (as in all our examples), the first term is $\widetilde{\mathcal{O}}(\sum_{i=1}^\ell \sqrt{|\mathcal{I}_i'|})=\widetilde{\mathcal{O}}(\sqrt{\ell|\mathcal{J}|})$ by Cauchy-Schwarz; the second term is of order $\widetilde{\mathcal{O}}(\sqrt{2^n})$.
To derive a bound in terms of $L$, we can simply choose the partition $\{\mathcal{I}_1', \ldots, \mathcal{I}_\ell'\}$ in a way such that $\Delta_{\mathcal{I}_i'} = 0$ and $\ell = L_{\mathcal{J}}$,
while to derive a bound in terms of $\Delta$, the partition needs to be chosen more carefully depending on the value of $\Delta_{\mathcal{J}}$.
Noting that the number of blocks between two restarts is always at most $\log_2 T$,
to finally prove \prettyref{thm: regret bound}, it remains to bound the number of restarts, which intuitively should scale with $L$ or $\Delta$ because by design a restart will not be triggered when the environment is stationary.
The complete proof is deferred to \prettyref{app: epoch regret}--\prettyref{app: omitted proof}.
\section*{Mappings}
\section{Omitted Algorithms and Main Results in \prettyref{sec: average-reward}}\label{app:ucrl_details}
\begin{algorithm2e}[H]
\caption{{\small\textsf{\textup{UCRL}}}\xspace with Adaptive Confidence Widening ({\small\textsf{\textup{UCRL-ACW}}}\xspace)}
\label{alg: base alg UCRL}
\textbf{input}: $\overline{D}\geq 1$ \\
$t\leftarrow 1$. \ \
$N_1(s,a)\leftarrow 0$ for all $s,a$. \ \highlight{$\Gamma\leftarrow 0$} \\
\For{episode $k=1,\ldots, $}{
Set $t_k=t$, $\nu_k(s,a)=0$ for all $s, a$. \\
Define for all $s,a$:
\begin{align*}
\widehat{p}_{k}(s'|s,a) &= \frac{\sum_{\tau=1}^{t-1}\mathbbm{1}[(s_\tau,a_\tau,s_{\tau+1}')=(s,a,s')]}{N_k^+(s,a)}, \\
\widehat{r}_{k}(s,a) &= \frac{\sum_{\tau=1}^{t-1}R_\tau\mathbbm{1}[(s_\tau,a_\tau)=(s,a)]}{N_k^+(s,a)}.
\end{align*}
\\
{\renewcommand{\nl}{\let\nl} and for any $\eta$: }
\begin{align*}
\mathcal{P}_{k}^\eta(s,a) &= \left\{\widetilde{p}(\cdot|s,a)\in\Delta_S: \|\widetilde{p}(\cdot|s,a)-\widehat{p}_{k}(\cdot|s,a)\|_1\leq \sqrt{S}\cdot\textit{conf}_{k}(s,a) + \eta \right\}\\
\mathcal{R}_{k}(s,a) &= \left\{\widetilde{r}(s,a)\in[0,1]: |\widetilde{r}(s,a)-\widehat{r}_{k}(s,a)|\leq \textit{conf}_{k}(s,a)\right\}
\end{align*}
{\renewcommand{\nl}{\let\nl} where $\textit{conf}_{k}(s,a)\triangleq 8\sqrt{\frac{\log(SAT/\delta)}{N_{k}^+(s,a)}}$ and $N_{k}^+(s,a) = \max\{1, N_k(s,a)\}$.}\\
{\renewcommand{\nl}{\let\nl} \ \ \\}
\highlight{
$\eta \leftarrow \frac{1}{T}$ \label{line: adaptive start}\\
\While{\textit{true}}{
Perform EVI on $(\mathcal{P}_{k}^\eta, \mathcal{R}_{k})$ with error parameter $\epsilon_{k}=\sqrt{\frac{1}{t}}$, and obtain $\widetilde{\pi}$, $\widetilde{h}$, $\widetilde{J}$. \\
\lIf{$\text{sp}(\widetilde{h}) \leq 2\overline{D}$}{\textbf{break}} \label{line: enough widening}
$\eta\leftarrow 2\eta$
} \label{line: active end}
$\pi_{k}\leftarrow \widetilde{\pi}, \ \ \widetilde{h}_k\leftarrow \widetilde{h}, \ \ \widetilde{J}_k\leftarrow \widetilde{J}, \ \ \eta_k\leftarrow \eta$ \myComment{Adaptive confidence widening} \\
{\renewcommand{\nl}{\let\nl} \ \ \\}
}
\While{$\nu_k(s,a)<N_k^+(s,a)$ for all $s,a$}{
Choose action $a_t\sim \pi_k(s_t)$. \\
$\nu_k(s_t,a_t)\leftarrow \nu_k(s_t,a_t) +1$ \\
\highlight{
\label{line: accumulate}$\Gamma \leftarrow \Gamma + \eta_k$ \label{line: accumulate} \\
\lIf{$\Gamma > 4S\sqrt{At\log(SAT/\delta)} $}{ \label{line: terminate condition}
\textit{terminate and signal restart} \myComment{Early termination} \label{line: return terminate}
}
}
Observe the reward $R_t$ with $\mathbb{E}[R_t]= r_t(s_t,a_t)$ \\
\highlight{Observe $s_{t+1}'\sim p_t(\cdot|s_t,a_t)$. \\
The next state $s_{t+1}$ is either equal to $s_{t+1}'$, or re-assigned as an arbitrary state} \\
\ \ \myComment{The next state might be re-assigned} \label{line: reassignment} \\
$t\leftarrow t+1$
}
$N_{k+1}(s,a)\leftarrow N_k(s,a) + \nu_k(s,a)$ for all $s,a$.
}
\end{algorithm2e}
\begin{algorithm2e}[h]
\caption{Multi-scale {\small\textsf{\textup{UCRL-ACW}}}\xspace ({\small\textsf{\textup{MUCRL}}}\xspace)}
\label{alg: multialg-ucrl}
{\renewcommand{\nl}{\let\nl} \textbf{input}: $n$, $\rho_{{\small\textsf{\textup{UCRL}}}\xspace}(\cdot~; \overline{D})$, $\overline{D}$} \\
\textbf{Initialization}: run \prettyref{proc: alg profile} with base algorithm {\small\textsf{\textup{UCRL-ACW}}}\xspace and inputs $n$ and $\rho_{{\small\textsf{\textup{UCRL}}}\xspace}$. \\
At each time $t$, let the unique active instance be $\inst$, output $\widetilde{g}_t$ (which is the quantity $\widetilde{J}_{k(t)}$ of $\inst$), follow $\inst$'s decision, and update $\inst$ after receiving feedback from the environment.
Additionally, \highlight{terminate if the $\inst$ signals restart}.
\end{algorithm2e}
\newpage
\begin{algorithm2e}
\caption{{\small\textsf{\textup{MASTER-UCRL}}}\xspace}
\label{alg: master ucrl}
\textbf{input}: $\rho_{{\small\textsf{\textup{UCRL}}}\xspace}(\cdot~; \overline{D})$, $\overline{D}$\\
\textbf{Initialize}: $t\leftarrow1$ \\
\For{$n=0, 1, \ldots$}{ \label{line: restart line2}
Set $t_n\leftarrow t$
and initialize an {\small\textsf{\textup{MUCRL}}}\xspace (\prettyref{alg: multialg-ucrl}) for the block $[t_n, t_n+2^n-1]$. \\
\While{$t< t_n + 2^n$}{
Receive $\widetilde{g}_t$ and $\pi_t$ from {\small\textsf{\textup{MUCRL}}}\xspace, execute $\pi_t$, and receive reward $R_t$. \\
Update {\small\textsf{\textup{MUCRL}}}\xspace with any feedback from the environment, and set $U_t = \min_{\tau \in [t_n, t]} \widetilde{g}_\tau$.
\label{line: tracking}\\
Perform \textbf{Test~1}\xspace and \textbf{Test~2}\xspace (see below).
Increment $t\leftarrow t+1$. \\
\lIf{either test returns \textit{fail} \highlight{or {\small\textsf{\textup{MUCRL}}}\xspace terminates} }{
restart from \prettyref{line: restart line2}.
}
}
}
\textbf{Test~1}\xspace: \ \ If $t=\inst.e$ for some order-$m$ $\inst$ and
$\frac{1}{2^m}\sum_{\tau=\inst.s}^{\inst.e} R_{\tau} \geq U_t + 9\widehat{\rho}_{{\small\textsf{\textup{UCRL}}}\xspace}(2^m; \overline{D})$,
return \textit{fail}.
\textbf{Test~2}\xspace: \ \ If
$\frac{1}{t-t_n+1}\sum_{\tau=t_n}^{t} \left(\widetilde{g}_{\tau} - R_{\tau}\right) \geq 3\widehat{\rho}_{{\small\textsf{\textup{UCRL}}}\xspace}(t-t_n+1; \overline{D})$,
return \textit{fail}.
\end{algorithm2e}
The following is the main result for the infinite-horizon MDP case. Its proof requires several lemmas in the rest of this section, in addition to those from \prettyref{app:multi-scale}--\prettyref{app: epoch regret seperate} whose ideas are mostly aligned with the standard setting. The final analysis is done in \prettyref{app: omit proof for RL avg} and \prettyref{app: borl discuss} (see \prettyref{thm: known Dmax case}, \prettyref{thm: doubling trick algo for RL}, and the discussions in \prettyref{app: borl discuss}).
Note that to be consistent with prior works in this setting, we adopt the notation $J_t(\pi)$, which is the expected average reward of executing $\pi$ under the MDP for time $t$, and corresponds to the notation $f_t(\pi)$ we use in our general framework.
Similarly, define $J^\star_t = \max_\pi J_t(\pi)$.
\begin{theorem}
\label{thm: main theorem for RL}
Define non-stationarity measures
\begin{align*}
\Delta &= \sum_{t=1}^{T-1} \left(\max_{s,a}|r_t(s,a)-r_{t+1}(s,a)| + \max_{s,a}\|p_t(\cdot|s,a)-p_{t+1}(\cdot|s,a)\|_1\right), \\
L &= 1 + \sum_{t=1}^{T-1} \mathbbm{1}\left\{\max_{s,a}|r_t(s,a)-r_{t+1}(s,a)| + \max_{s,a}\|p_t(\cdot|s,a)-p_{t+1}(\cdot|s,a)\|_1 \neq 0\right\}.
\end{align*}
There exists an algorithm that takes $D_{\max}$ as input and achieves
\begin{align*}
\sum_{t=1}^T \left(J^\star_t - R_t \right) = \widetilde{\mathcal{O}}\left(\min\left\{D_{\max}S\sqrt{ALT}, \ \ D_{\max}S^{\frac{2}{3}}A^{\frac{1}{3}}\Delta^{\frac{1}{3}}T^{\frac{2}{3}} + D_{\max}S\sqrt{AT}\right\}\right)
\end{align*}
without knowing $L$ or $\Delta$.
There is also an algorithm that takes $L$ or $\Delta$ as input and achieves
\begin{align*}
\sum_{t=1}^T \left(J^\star_t - R_t \right) = \widetilde{\mathcal{O}}\left(D_{\max}S\sqrt{ALT}\right) \quad \text{or}\quad
\sum_{t=1}^T \left(J^\star_t - R_t \right) = \widetilde{\mathcal{O}}\left(D_{\max}S^{\frac{2}{3}}A^{\frac{1}{3}}\Delta^{\frac{1}{3}}T^{\frac{2}{3}} + D_{\max}S\sqrt{AT}\right)
\end{align*}
respectively, without knowing $D_{\max}$. Finally, there is an algorithm that achieves
\begin{align*}
\sum_{t=1}^T \left(J^\star_t - R_t \right) = \widetilde{\mathcal{O}}\left( D_{\max}(S^2A)^{\nicefrac{1}{4}}T^{\nicefrac{3}{4}} + \min\left\{ D_{\max}S\sqrt{ALT},\ \ D_{\max}(S^2A)^{\frac{1}{3}}\Delta^{\frac{1}{3}}T^{\frac{2}{3}}\right\}\right)
\end{align*}
without knowing $L, \Delta$, or $D_{\max}$.
\end{theorem}
\subsection{Auxiliary Lemmas related to Extended Value Iteration and Bellman Equation}
In this subsection, we provide auxiliary lemmas related to EVI and Bellman Equation. The results are extracted from \citep{jaksch2010near, cheung2020reinforcement, ortner2020variational}. We restate them here for completeness.
\begin{lemma}[Properties 1 and 2 in \citep{cheung2020reinforcement}]
Let $\widetilde{J}$, $\widetilde{h}$, and $\widetilde{\pi}$ be the set of solution obtained from EVI with confidence set $\mathcal{R}$ and $\mathcal{P}$ for reward and transition respectively, and error parameter $\epsilon$. Then
\begin{align}
\widetilde{J} + \widetilde{h}(s)
&\geq \max_{a} \left(\max_{\widetilde{r}\in \mathcal{R}(s,a)} \widetilde{r}(s,a) + \max_{\widetilde{p}\in \mathcal{P}(s,a)} \sum_{s'} \widetilde{p}(s'|s,a) \widetilde{h}(s') \right), \label{eq: optimistic bellman}\\
\widetilde{J} + \widetilde{h}(s) &\leq \max_{\widetilde{r}\in \mathcal{R}(s,\widetilde{\pi}(s))} \widetilde{r}(s,\widetilde{\pi}(s)) + \max_{\widetilde{p}\in \mathcal{P}(s,\widetilde{\pi}(s))} \sum_{s'} \widetilde{p}(s'|s,\widetilde{\pi}(s)) \widetilde{h}(s') + \epsilon. \label{eq: optimistic bellman 2}
\end{align}
\end{lemma}
\begin{lemma}[Lemma 2 of \citep{cheung2020reinforcement}]
\label{lemma: bound diamete}
Let $\widetilde{J}$, $\widetilde{h}$, and $\widetilde{\pi}$ be the set of solution obtained from EVI with confidence set $\mathcal{R}$ and $\mathcal{P}$ for reward and transition respectively. If $\mathcal{P}$ and $\mathcal{R}$ contain an MDP with diameter upper bounded by $D$, then $\text{sp}(\widetilde{h})\leq 2D$.
\end{lemma}
\begin{lemma}[Eq. (16) of \citep{cheung2020reinforcement}]
\label{lemma: dual feasible}
Let $r, p$ define the reward function and the transition kernel for a communicating MDP, respectively. Let $\widetilde{J}\in\mathbb{R}, \widetilde{h}\in\mathbb{R}^S$ be bounded and satisfy
\begin{align*}
\widetilde{J} + \widetilde{h}(s) \geq r(s,a) + \sum_{s'} p(s'|s,a)\widetilde{h}(s')
\end{align*}
for all $s$ and $a$. Then $\widetilde{J}\geq J^\star$, where $J^\star$ is the average reward of the optimal policy under the MDP.
\end{lemma}
\subsection{Guarantees of the {\small\textsf{\textup{UCRL-ACW}}}\xspace Algorithm (when running alone with an input $\overline{D}$)}
\begin{definition} \label{def: RL measure}
Define $ \Delta^r(t) \triangleq \max_{s,a} |r_t(s,a) - r_{t+1}(s,a)|$, $\Delta^p(t) \triangleq \max_{s,a} \| p_t(\cdot|s,a) - p_{t+1}(\cdot|s,a)\|_1$, $\Delta^J(t) \triangleq \max_{\pi} |J_t(\pi)-J_{t+1}(\pi)|$. Similar to \prettyref{def: deviation}, define $\Delta_{\mathcal{I}}^{\Box}=\sum_{\tau=s}^{e-1}\Delta^{\Box}(\tau)$ for interval $\mathcal{I}=[s,e]$, where $\Box= r, p$, or $J$. Finally, we define $\Delta_{\mathcal{I}; \overline{D}} \triangleq \Delta_{\mathcal{I}}^r + 2\overline{D}\Delta_{\mathcal{I}}^p + \Delta_{\mathcal{I}}^J$.
\end{definition}
\begin{lemma}[Theorem 1 of \citep{ortner2020variational}]
\label{lemma: ortner lemma}
$\Delta^J(t)\leq \Delta^r(t) + D_{\max} \Delta^p(t)$.
\end{lemma}
\begin{lemma}[\textit{c.f.} \prettyref{assum:assump2}]
\label{lem: modified UCRL}
When run alone, \prettyref{alg: base alg UCRL} with input $\overline{D}$ guarantees for all $t$ before it terminates:
\begin{align*}
\widetilde{J}_{k(t)} &\geq \min_{\tau\in [1,t]} J^\star_\tau - \Delta_{[1,t]; \overline{D}} \\
\frac{1}{t}\sum_{\tau=1}^{t} \left( \widetilde{J}_{k(\tau)} - R_\tau \right) &\leq \rho_{{\small\textsf{\textup{UCRL}}}\xspace}\left(t; \overline{D}\right) + \frac{2}{t}\overline{D}\textit{disc}_{[1,t]} + \Delta_{[1,t]; \overline{D}},
\end{align*}
where $\rho_{{\small\textsf{\textup{UCRL}}}\xspace}(t; \overline{D}) = \widetilde{\Theta}\left(\min\Big\{\overline{D} S\sqrt{\frac{A}{t}} + \frac{\overline{D} SA}{t}, \ \ \overline{D}\Big\}\right)$, $\textit{disc}_{\mathcal{I}}\triangleq \sum_{t\in\mathcal{I}}\mathbbm{1}[s_t \nsim p_{t-1}(\cdot|s_{t-1},a_{t-1}) ]$ is the number of state re-assignments within $\mathcal{I}$ (\prettyref{line: reassignment} of \prettyref{alg: base alg UCRL}), and $k(t)$ is the index of the episode time $t$ belongs to.
\end{lemma}
\begin{proof}\ \ \
Suppose that at time $t$ the algorithm has not terminated. For any episode $k$ that starts before $t$, we have
\begin{align*}
\overline{r}_{k}(s,a) = \frac{\sum_{\tau=1}^{t_k-1} r_\tau(s,a)\mathbbm{1}[(s_\tau,a_\tau)=(s,a)]}{N_k^+(s,a)}, \quad
\overline{p}_k(s'|s,a) = \frac{\sum_{\tau=1}^{t_k-1} p_\tau(s'|s,a)\mathbbm{1}[(s_\tau,a_\tau)=(s,a)]}{N_k^+(s,a)}.
\end{align*}
By Azuma's inequality, $\overline{r}_k(s,a)\in\mathcal{R}_k(s,a)$ and $\overline{p}_k(\cdot|s,a)\in\mathcal{P}_k^{\eta_k}(s,a)$ with high probability for all $k, s,a$.
To show the first part of the lemma, we lower bound the right-hand side of \prettyref{eq: optimistic bellman}:
\begin{align}
\widetilde{J}_k + \widetilde{h}_k(s)
&\geq \overline{r}(s,a) + \sum_{s'} \overline{p}(s'|s,a) \widetilde{h}_k(s') \nonumber \\
&\geq r_\tau(s,a) + \sum_{s'} p_\tau(s'|s,a)\widetilde{h}_k(s') - \left(\Delta^r_{[1,t]} + 2\overline{D} \Delta^p_{[1,t]}\right), \tag{for any $\tau\in[1,t]$}
\end{align}
where in the last inequality we use $|\overline{r}(s,a) - r_\tau(s,a)|\leq \Delta_{[1,t]}^r$ and $\sum_{s'} |\overline{p}(s'|s,a) - p_\tau(s'|s,a)|\widetilde{h}_k(s') \leq \|\overline{p}(\cdot|s,a) -p_\tau(\cdot|s,a) \|_1 \text{sp}(\widetilde{h}_k)\leq 2 \Delta_{[1,t]}^p \overline{D}$.
Using \prettyref{lemma: dual feasible}, we get \[\widetilde{J}_k + \Delta_{[1,t]; \overline{D}} \geq \widetilde{J}_k + \left(\Delta^r_{[1,t]} +2 \overline{D} \Delta^p_{[1,t]}\right) \geq J_\tau^\star,\] implying the first part of the lemma.
To show the second part of the lemma, starting from \prettyref{eq: optimistic bellman 2}, we have with high probability
\begin{align}
&\widetilde{J}_k + \widetilde{h}_k(s)\\
&\leq \overline{r}_k(s,\pi_k(s)) + \sum_{s'}\overline{p}_k(s'|s,\pi_k(s))\widetilde{h}_k(s') + 2\overline{D} \sqrt{S}\cdot \textit{conf}_k(s,\pi_k(s)) + 2 \overline{D} \eta_k + \epsilon_k \tag{$\overline{r}_k(s,a) \in \mathcal{R}_k(s,a)$ and $\overline{p}_k(s,a) \in \mathcal{P}_k(s,a)$} \\
&\leq r_\tau(s,\pi_k(s)) + \sum_{s'}p_\tau(s'|s,\pi_k(s))\widetilde{h}_k(s') + 2\overline{D} \sqrt{S}\cdot \textit{conf}_k(s,\pi_k(s)) + \left( \Delta^r_{[1,t]} + 2\overline{D}\Delta^p_{[1,t]} \right) +2 \overline{D} \eta_k + \epsilon_k. \label{eq: to telescope}
\end{align}
Now, we apply \prettyref{eq: to telescope} with $(k,\tau,s) = \{k(\tau), \tau, s_\tau\}_{\tau=1}^t$ respectively, and sum them up. Notice that $a_\tau=\pi_{k(\tau)}(s_\tau)$. Then we get
\begin{align*}
\sum_{\tau=1}^t \left(\widetilde{J}_{k(\tau)} - R_{\tau}\right)
&\leq \sum_{\tau=1}^t \left(\sum_{s'} p_{\tau}(s'|s_\tau,a_\tau)\widetilde{h}_{k(\tau)}(s') - \widetilde{h}_{k(\tau)}(s_\tau)\right) + \sum_{\tau=1}^t \left(r_\tau(s_\tau, a_\tau) - R_\tau\right) \\
&\quad + \sum_{\tau=1}^t 2\overline{D}\sqrt{S}\cdot\textit{conf}_{k(\tau)}(s_\tau, a_\tau)+ t\Delta_{[1,t]; \overline{D}} + 2\overline{D}\sum_{\tau=1}^t \eta_{k(\tau)} + \sum_{\tau=1}^t \epsilon_{k(\tau)}.
\end{align*}
We bound the terms on the right-hand side individually: for the first term, notice that when there is no state-reassignment at time $\tau+1$, $\mathbb{E}_\tau [\widetilde{h}_{k(\tau)}(s_{\tau+1})] = \sum_{s'}p_{\tau}(s'|s_\tau,a_\tau)\widetilde{h}_{k(\tau)}(s')$. Therefore,
\begin{align*}
&\sum_{\tau=1}^t \left(\sum_{s'} p_{\tau}(s'|s_\tau,a_\tau)\widetilde{h}_{k(\tau)}(s') - \widetilde{h}_{k(\tau)}(s_\tau)\right) \\
&\leq \sum_{\tau=1}^t \left(\mathbb{E}_\tau \left[\widetilde{h}_{k(\tau)}(s_{\tau+1})\right] - \widetilde{h}_{k(\tau)}(s_\tau) \right) + 2\overline{D}\textit{disc}_{[1,t]} \\
&\leq 2\overline{D}\sqrt{t\log(SAT)} + 2\overline{D}\sum_{\tau=1}^t \mathbbm{1}\left[\widetilde{h}_{k(\tau)} \neq \widetilde{h}_{k(\tau+1)}\right] + 2\overline{D}\textit{disc}_{[1,t]} \tag{by Azuma's inequality} \\
&\leq 2\overline{D}\sqrt{t\log(SAT)} + 2\overline{D} SA\log_2 T + 2\overline{D}\textit{disc}_{[1,t]},
\end{align*}
where in the last inequality we use the fact that the number of episodes cannot exceed $SA\log_2 T$. For the other terms:
$\sum_{\tau=1}^t \left(r_\tau(s_\tau, a_\tau) - R_\tau\right) \leq \widetilde{\mathcal{O}}\left(\sqrt{t}\right)$ by Azuma's inequality; $\sum_{\tau=1}^t 2\overline{D}\sqrt{S}\cdot\textit{conf}_{k(\tau)}(s_\tau, a_\tau) = \widetilde{\mathcal{O}}\left( \overline{D} S\sqrt{At} \right)$ by the standard pigeonhole argument; $2\overline{D}\sum_{\tau=1}^t \eta_{k(\tau)} = \widetilde{\mathcal{O}}\left(\overline{D} S\sqrt{At}\right)$
by the termination condition specified in \prettyref{line: terminate condition}; $\sum_{\tau=1}^t \epsilon_{k(\tau)}$ is also upper bounded by $\widetilde{\mathcal{O}}\left(\overline{D} S\sqrt{At}\right)$ by the way we choose the error parameter. Combining all the above arguments, we get
\begin{align*}
\sum_{\tau=1}^t \left(\widetilde{J}_{k(\tau)} - R_{\tau}\right) \leq \widetilde{\mathcal{O}}\left(\overline{D} S\sqrt{At} + \overline{D} SA \right) + 2\overline{D} \textit{disc}_{[1,t]} + t\Delta_{[1,t]; \overline{D}}
\end{align*}
with high probability. On the other hand, $\sum_{\tau=1}^t \left(\widetilde{J}_{k(\tau)} - R_{\tau}\right) \leq \overline{D} t$ is trivially true. Combining them we get the second claim of the lemma.
\end{proof}
\section{Analysis for the Multi-scale Algorithms}\label{app:multi-scale}
\begin{proof}{\textbf{of \prettyref{lemma: multi-scale reg}} }
Below, we fix an $\inst$ and fix a $t\in[\inst.s, \inst.e]$, and consider the case $\Delta_{[\inst.s, t]}\leq \rho(t')$ as specified in the lemma statement. For the first part of the lemma, note that $\widetilde{g}_t$ of {\small\textsf{\textup{MALG}}}\xspace is defined as $\widetilde{f}_t^{\inst'}$ where $\inst'$ is the active instance of {\small\textsf{\textup{ALG}}}\xspace at round $t$. By \prettyref{proc: alg profile}, $\inst'$ can only be an instance that starts within $[\inst.s, t]$ (i.e., $\inst'.s\geq \inst.s$). Therefore, the distribution drift undergone by $\inst'$ up to $t$ is upper bounded by $\Delta_{[\inst.s, t]}\leq \rho(t')$, which is further upper bounded by $\rho(t'')$ where $t''$ is the number of active rounds $\inst'$ runs within $[\inst.s, t]$, because $\rho(\cdot)$ is a decreasing function. Therefore, the conditions in \prettyref{assum:assump2} is satisfied for this $\inst'$, and thus we have
\begin{align*}
\widetilde{g}_t = \widetilde{f}_t^{\inst'} \geq \min_{\tau\leq t:~\inst' \text{\ is active at $\tau$}} f_\tau^\star - \Delta_{[\inst'.s,t]} \geq \min_{\tau \in [\inst.s, t]}f_{\tau}^\star - \Delta_{[\inst.s, t]},
\end{align*}
proving the first part.
Next, we prove the second part of the lemma.
We use $S_m$ to denote the set of order-$m$ instances which start within $[\inst.s, t]$. Note that
\begin{align}
\sum_{\tau=\inst.s}^{t} \left(\widetilde{g}_\tau - R_\tau\right)
&= \sum_{\tau=\inst.s}^{t} \sum_{m=0}^n \sum_{\inst'\in S_m} \mathbbm{1}[\inst'\ \text{is active at\ } \tau]\left(\widetilde{f}_\tau^{\inst'}-R_\tau\right) \nonumber \\
&= \sum_{m=0}^n \underbrace{\sum_{\inst'\in S_m}\sum_{\tau=\inst.s}^{t} \mathbbm{1}[\inst'\ \text{is active at\ } \tau]\left(\widetilde{f}_\tau^{\inst'}-R_\tau\right)}_{(*)}. \label{eq: regret decompose}
\end{align}
The first equality holds because $\widetilde{g}_\tau$ of {\small\textsf{\textup{MALG}}}\xspace is defined as the $\widetilde{f}_\tau$ of the active instance at round $t$.
Next, we focus on a specific $m$, and bound the $(*)$ term in \prettyref{eq: regret decompose}. Let $|S_m|=\ell$ and $S_m=\left\{\inst_1',\ldots, \inst_\ell'\right\}$, and let $\mathcal{I}_i\triangleq [\inst_i'.s, \inst_i'.e]\cap [\inst.s, t]$ for $i=1,\ldots,\ell$ (i.e., $\mathcal{I}_i$ are the rounds within $[\inst.s, t]$ where $\inst_i'$ is scheduled).
Clearly, $|\mathcal{I}_i|\leq \min\{\inst'_i.e - \inst'_i.s +1 , t-\inst.s+1\} = \min\left\{2^m, t'\right\}$. By \prettyref{assum:assump2}, we have
\begin{align}
(*)&= \sum_{i=1}^\ell \sum_{\tau=\inst.s}^{t} \mathbbm{1}[\inst_i'\text{\ is active at $\tau$}]\left(\widetilde{f}_\tau^{\inst_i'}-R_\tau\right) \nonumber \\
&\leq \sum_{i=1}^\ell \left(C(|\mathcal{I}_i|) + |\mathcal{I}_i|\Delta_{\mathcal{I}_i}\right) \nonumber \\
&\leq \ellC(\min\{2^m, t'\}) + t'\Delta_{[\inst.s, t]}, \label{eq: bound star}
\end{align}
where in the first inequality we use \prettyref{assum:assump2}, and that $\inst_i'$ updates for no more than $|\mathcal{I}_i|$ rounds in the interval $[\inst.s, t]$ (also, the condition in \prettyref{assum:assump2} is satisfied because $\Delta_{\mathcal{I}_i}\leq \Delta_{[\inst.s,t]}\leq \rho(t')\leq \rho(|\mathcal{I}_i|)$). In the last inequality, for the first term, we use that $C(\cdot)$ is increasing; for the second term, we use $|\mathcal{I}_i|\leq t'$, and that $\Delta_{\mathcal{I}_1}+\cdots+\Delta_{\mathcal{I}_\ell}\leq \Delta_{[\inst.s, t]}$ since $\mathcal{I}_1, \ldots, \mathcal{I}_\ell$ are non-overlapping intervals lying within $[\inst.s,t]$.
By \prettyref{proc: alg profile},
for every $m$, the expected number of order-$m$ {\small\textsf{\textup{ALG}}}\xspace's that starts within the interval $[\inst.s, t]$ can be upper bounded as
\begin{align}
\mathbb{E}[|S_m|]
&\leq \frac{\rho(2^n)}{\rho(2^m)} \left\lceil \frac{t'}{2^m} \right\rceil
\leq \frac{\rho(2^n)}{\rho(2^m)}\left( \frac{t'}{2^m}+1\right) \leq \frac{\rho(2^n)}{\rho(2^m)} \frac{t'}{2^m}+1 \label{eq: bound Sm}
\end{align}
By Bernstein's inequality, with probability $1-\frac{\delta}{T}$, $|S_m|\leq \mathbb{E}[|S_m|]+\sqrt{2\mathbb{E}[|S_m|]\log(T/\delta)} + \log(T/\delta) \leq 2\mathbb{E}[|S_m|] + 2\log(T/\delta)$. Thus, continuing from \prettyref{eq: bound star}, we have with probability at least $1-\frac{\delta}{T}$,
\begin{align}
(*)
&\leq 2 \cdot \left(\frac{\rho(2^n)}{\rho(2^m)} \frac{t'}{2^m} +1\right) C(\min\{2^m, t'\}) + 2\log(T/\delta)C(\min\{2^m, t'\}) + t'\Delta_{[\inst.s, t]} \nonumber \\
&\leq 2 \left(\frac{C(t')}{C(2^m)} + 2 \right) \log(T/\delta) C(\min\{2^m, t'\}) + t'\Delta_{[\inst.s, t]} \tag{$\rho(2^n) t'\leq \rho(t')t'=C(t')$} \nonumber \\
&\leq 6 C(t') \log(T/\delta) + t'\Delta_{[\inst.s, t]} \tag{$C(\cdot)$ is an increasing function} \\
\label{eq: upper bound star}
\end{align}
Finally, using this in \prettyref{eq: regret decompose}, we get the second claim of the lemma: with probability at least $1-\frac{\delta}{T}$,
\begin{align}
\sum_{\tau=\inst.s}^t\left(\widetilde{g}_\tau-R_\tau\right)
&\leq 6(n+1) C(t') \log(1/\delta) + t'(n+1)\Delta_{[\inst.s, t]}. \label{eq: regret temp}
\end{align}
For the third part of the lemma, as we calculated above, with probability at least $1-\frac{\delta}{T}$, the number of instances started within $[\inst.s, t]$ is upper bounded by
\begin{align*}
\sum_{m=0}^n 2 \cdot \left(\frac{\rho(2^n)}{\rho(2^m)} \frac{t'}{2^m} +2\right) \log(T/\delta)\leq 2\widehat{n}\xspace\left(\frac{C(t')}{C(1)} +2\right)\log(T/\delta) \leq 6\widehat{n}\xspace\frac{C(t')}{C(1)}\log(T/\delta)
\end{align*}
where we use $\rho(2^m)2^m = C(2^m) \geq C(1)$ and $\rho(2^n)t'\leq \rho(t')t'=C(t')$.
\end{proof}
\begin{lemma}[\textit{c.f.} \prettyref{lemma: multi-scale reg}]
\label{lemma: ucrl aggregated regret}
Before {\small\textsf{\textup{MUCRL}}}\xspace terminates, for every \inst and $t\in[\inst.s, \inst.e]$, {\small\textsf{\textup{MUCRL}}}\xspace guarantees with high probability
\begin{align*}
\widetilde{g}_t &\geq \min_{\tau\in[\inst.s, t]} J_\tau^\star - \Delta_{[\inst.s, t]; \overline{D}} \\
\frac{1}{t'} \sum_{\tau=\inst.s}^{t} \left(\widetilde{g}_\tau - R_\tau\right)&\leq \widehat{\rho}_{{\small\textsf{\textup{UCRL}}}\xspace}\left(t'; \overline{D}\right) + \widehat{n}\xspace\Delta_{[\inst.s, t]; \overline{D}}
\end{align*}
where $t'=t-\inst.s+1$, $\widehat{n}\xspace=\log_2 T+1$, and $\widehat{\rho}_{{\small\textsf{\textup{UCRL}}}\xspace}\left(t; \overline{D}\right)=18\widehat{n}\xspace\log(T/\delta)\rho_{{\small\textsf{\textup{UCRL}}}\xspace}(t;\overline{D})$.
\end{lemma}
\begin{proof}
This proof is similar to that of \prettyref{lemma: multi-scale reg}. For the first part of the lemma, we can simply follow the proof of the first part of \prettyref{lemma: multi-scale reg}, with $\widetilde{f}_t$ replaced by $\widetilde{J}_{k(t)}$, and $\Delta_{[\inst.s, t]}$ by $\Delta_{[\inst.s, t];\overline{D}}$.
For the second part, the analysis still tightly follows that of \prettyref{lemma: multi-scale reg}, but we need to add the additional cost caused by state re-assignment (i.e., the $\overline{D} \textit{disc}_{[1,t]}$ term in \prettyref{lem: modified UCRL}).
Following the same arguments as in proof as in \prettyref{eq: regret decompose} and \prettyref{eq: bound star}, we get
\begin{align}
\sum_{\tau=\inst.s}^{t} \left(\widetilde{g}_\tau - R_\tau\right)
&= \sum_{m=0}^n \sum_{\inst'\in S_m}\sum_{\tau=\inst.s}^{t} \mathbbm{1}[\inst'\ \text{is active at\ } \tau]\left(\widetilde{f}_\tau^{\inst'}-R_\tau\right) \tag{$S_m\triangleq$ \text{the set of order-$m$ {\small\textsf{\textup{ALG}}}\xspace initiated within $[\inst.s, t]$}} \\
&\leq \sum_{m=0}^n \sum_{i=1}^{|S_m|}\sum_{\tau=\inst.s}^{t} \mathbbm{1}[\inst_{m,i}'\ \text{is active at\ } \tau]\left(\widetilde{f}_\tau^{\inst_{m,i}'}-R_\tau\right) \tag{Let $S_m=\{\inst'_{m,1}, \inst'_{m,2}, \ldots\}$}\\
&\leq \sum_{m=0}^n \sum_{i=1}^{|S_m|} \left(C_{{\small\textsf{\textup{UCRL}}}\xspace}(|\mathcal{I}_{m,i}|; \overline{D}) + |\mathcal{I}_{m,i}|\Delta_{\mathcal{I}_{m,i}; \overline{D}} + 2\overline{D}\textit{disc}_{\mathcal{I}_{m,i}}^{\inst_{m,i}'}\right) \tag{by \prettyref{lem: modified UCRL}} \\
&\label{eq: last expression MDP}
\end{align}
where in the last expression, we denote $\mathcal{I}_{m,i}=[\inst'_{m,i}.s, \inst'_{m,i}.e] \cap [\inst.s, t]$ (the time within $[\inst.s, t]$ where $\inst'_{m,i}$ is scheduled), and $\textit{disc}^{\inst'}_{\mathcal{I}}$ is the total number of times within $\mathcal{I}$ when $\inst'$ encounters state-reassignments.
For a fixed $m$, observe that all order-$m$ instances are non-overlapping. Also, the aggregated number of state re-assignment for all order-$m$ instances started within $[\inst.s, t]$ is upper bounded by the total number of new instances of order not larger than $m-1$ started within $[\inst.s, t]$. The latter is further upper bounded by $6\widehat{n}\xspace\log(T/\delta)\frac{C_{{\small\textsf{\textup{UCRL}}}\xspace}(t'; \overline{D})}{C_{{\small\textsf{\textup{UCRL}}}\xspace}(1; \overline{D})}$ according to the last claim of \prettyref{lemma: multi-scale reg}. In other words, for every $m$, with probability $1 - \frac{\delta}{T}$,
\begin{align*}
\sum_{i=1}^{|S_m|} \textit{disc}_{\mathcal{I}_{m,i}}^{\inst_{m,i}'} = 6\widehat{n}\xspace \log(T/\delta) \frac{C_{{\small\textsf{\textup{UCRL}}}\xspace}(t'; \overline{D})}{C_{{\small\textsf{\textup{UCRL}}}\xspace}(1; \overline{D})}.
\end{align*}
Following the same calculation as in \prettyref{eq: bound star}, \prettyref{eq: bound Sm} and \prettyref{eq: upper bound star}, we also have that for every $m$, with probability $1-\frac{\delta}{T}$,
\begin{align*}
\sum_{i=1}^{|S_m|} \left(C_{{\small\textsf{\textup{UCRL}}}\xspace}(|\mathcal{I}_{m,i}|; \overline{D}) + |\mathcal{I}_{m,i}|\Delta_{\mathcal{I}_{m,i}; \overline{D}} \right) \leq 6\widehat{n}\xspace \log(T/\delta)C_{{\small\textsf{\textup{UCRL}}}\xspace}(t'; \overline{D}) + t'\Delta_{[\inst.s, t]; \overline{D}}.
\end{align*}
Using the above two bounds in \prettyref{eq: last expression MDP}, we get
\begin{align*}
\sum_{\tau=\inst.s}^{t} \left(\widetilde{g}_\tau - R_\tau\right)
&= 6\widehat{n}\xspace \log(T/\delta)C_{{\small\textsf{\textup{UCRL}}}\xspace}(t'; \overline{D}) + \widehat{n}\xspace t'\Delta_{[\inst.s, t]; \overline{D}} +2\overline{D} \times 6\widehat{n}\xspace \log(T/\delta) \frac{C_{{\small\textsf{\textup{UCRL}}}\xspace}(t'; \overline{D})}{C_{{\small\textsf{\textup{UCRL}}}\xspace}(1; \overline{D})}\\
&= 18\widehat{n}\xspace\log(T/\delta)C_{{\small\textsf{\textup{UCRL}}}\xspace}(t';\overline{D}) + \widehat{n}\xspace t'\Delta_{[\inst.s, t]; \overline{D}}
\end{align*}
where we use $C_{{\small\textsf{\textup{UCRL}}}\xspace}(1; \overline{D})\geq \overline{D}$ (by the definition of $\rho_{{\small\textsf{\textup{UCRL}}}\xspace}(\cdot~,\overline{D})$ in \prettyref{lem: modified UCRL}). Dividing both sides by $t'$ finishes the proof.
\end{proof}
\section{Single-block Regret Analysis I}
\label{app: key proofs}
In this section, we focus on the regret in a block of index $n$. The analysis applies to both the standard case (\prettyref{lemma: bound term1}), and the infinite-horizon RL case summarized in the following lemma.
\begin{lemma}[\textit{c.f.} \prettyref{lemma: bound term1}]
\label{lemma: block lemma for RL}
In a block of index $n$ that starts from $t_n$ and ends on $E_n$ ($E_n$ could be equal to $t_n+2^n-1$, or smaller, if any stationarity test fails or {\small\textsf{\textup{MUCRL}}}\xspace terminates), we have
\begin{align*}
\sum_{\tau=t_n}^{E_n} \left(f_\tau^\star - R_\tau\right) \leq \widetilde{\mathcal{O}}\left( \sum_{i=1}^\ell C_{{\small\textsf{\textup{UCRL}}}\xspace}(|\mathcal{I}_i'|; \overline{D}) + \sum_{m=0}^n \frac{\rho(2^m; \overline{D})}{\rho(2^n; \overline{D})} C_{{\small\textsf{\textup{UCRL}}}\xspace}(2^m; \overline{D}) \right)
\end{align*}
where $\mathcal{I}_1', \ldots, \mathcal{I}_\ell'$ are intervals that partition $[t_n, E_n]$ such that $\Delta_{\mathcal{I}_i'; \overline{D}} \leq \rho_{{\small\textsf{\textup{UCRL}}}\xspace}(|\mathcal{I}_i'|; \overline{D})$ for all $i$.
\end{lemma}
Throughout this section, if infinite-horizon RL is considered, $\rho(\cdot)\triangleq \rho_{{\small\textsf{\textup{UCRL}}}\xspace}(\cdot~; \overline{D})$, $\widehat{\rho}(\cdot)\triangleq \widehat{\rho}_{{\small\textsf{\textup{UCRL}}}\xspace}(\cdot~; \overline{D})$, $\Delta_{\mathcal{I}}\triangleq \Delta_{\mathcal{I}; \overline{D}}=\Delta^r_{\mathcal{I}} + 2\overline{D} \Delta^p_{\mathcal{I}} + \Delta^J_{\mathcal{I}}$ with a fixed $\overline{D}$, and $f^\star_t\triangleq J^\star_t$.
For the purpose of conducting analysis, we divide $[t_n, t_n+2^n-1]$ into consecutive intervals $\mathcal{I}_1=[s_1, e_1], \mathcal{I}_2=[s_2, e_2], \ldots, \mathcal{I}_K=[s_K, e_K]$ ($s_1=t_n$, $e_{i}+1=s_{i+1}$, $e_K=t_n+2^n-1$) in a way such that for all $i$:
\begin{align}
\Delta_{\mathcal{I}_i} \leq \rho(|\mathcal{I}_i|) \label{eq: i smaller than alpha}
\end{align}
One simple way to divide the intervals is to let $\Delta_{\mathcal{I}_i}=0$ in each $\mathcal{I}_i$. Then the number of intervals $K$ would be upper bounded by the number of stationary intervals within $[t_n, t_n+2^n-1]$. Intuitively, the number of intervals can also be related to $\Delta_{[t_n, t_n+2^n-1]}$.
We defer the calculation of the required number of intervals to \prettyref{lemma: interval divide}. For now, we only need the fact that the partition satisfies \prettyref{eq: i smaller than alpha}. From a high level, this partition makes the distribution in each interval close to stationary. Notice that this partition is independent of the learner's behavior in block $n$.
For convenience, we further define the following quantities that depend on the learner's behavior in block $n$:
\begin{definition}
Define $E_n$ as the index of the last round in block $n$. Since the block might terminate earlier than planned, we have $E_n \leq t_n+2^n -1$.
Let $\ell\in[K]$ be such that $E_n\in\mathcal{I}_\ell$ (that is, $\ell$ is the index of the interval where block $n$ ends). Define $e_i'=\min\{e_i, E_n\}$ and $\mathcal{I}_i'=[s_i, e_i']$ (therefore, $\mathcal{I}_i'=\emptyset$ for $i>\ell$).
\end{definition}
Recall the definition of $\widehat{n}\xspace$ and $\widehat{\rho}(t)$ from \prettyref{lemma: multi-scale reg} (or \prettyref{lemma: ucrl aggregated regret}).
For simplicity, we define $\alpha_m\triangleq \rho(2^m)$, $\widehat{\alpha}_m\triangleq \widehat{\rho}(2^m)$, and also $\widehat{C}(t) \triangleq t\widehat{\rho}(t)$.
Furthermore, we define the following technical quantities.
\begin{definition}
For every $i\in\{1,\ldots,K\}$, and every $m\in\left\{0, 1, \ldots, n\right\}$, define
\begin{align*}
\tau_i(m) = \min\left\{\tau\in \mathcal{I}_i': ~ f_\tau^\star - \widetilde{g}_\tau \geq 12\widehat{\alpha}_m \right\};
\end{align*}
that is, $\tau_i(m)$ is the first time $\tau$ in $\mathcal{I}_i'=\mathcal{I}_i\cap [t_n, E_n]$ such that $f_\tau^\star-\widetilde{g}_\tau$ exceeds $12\widehat{\alpha}_m$. If such $\tau$ does not exist or $\mathcal{I}_i'$ is empty, we let $\tau_i(m)=\infty$.
Besides, we define $\xi_i(m)=[e_i'-\tau_i(m)+1]_+$ where $[a]_+ = \max\{0, a\}$ (which is the length of the interval $[\tau_i(m), e_i']$ when $\tau_i(m)$ is not $\infty$).
\end{definition}
The intuition for $\tau_i(m)$ and $\xi_i(m)$ is as follows. Suppose that block $n$ has not ended at $\tau$.
If there exists some $\tau\in\mathcal{I}_i$ such that $f_\tau^\star - \widetilde{g}_\tau \geq 12\widehat{\alpha}_m$ (which first happens at $\tau_i(m)$), and if $\mathcal{I}_i$ is long enough (i.e., $\xi_i(m)$ is large enough) so that after $\tau_i(m)$, an order-$m$ instance of {\small\textsf{\textup{ALG}}}\xspace can run entirely within $\mathcal{I}_i$, then the learner is able to discover the fact that $f_\tau^\star - \widetilde{g}_\tau$ is large, and then restart. This coincides with our explanation in \prettyref{fig: detection}. The derivation in this section will formalize this intuition.
\begin{lemma}\label{lemma: block regret hard term}
Let the high-probability events described in \prettyref{lemma: multi-scale reg} (or \prettyref{lemma: ucrl aggregated regret}) hold. Then with high probability,
\begin{align*}
\sum_{\tau=t_n}^{E_n}\left(\widetilde{g}_\tau-R_\tau\right) &\leq 4\widehat{C}(2^n), \\
\sum_{\tau =t_n}^{E_n} \left(f_\tau^\star - \widetilde{g}_\tau \right) &\leq 96\widehat{n}\xspace\sum_{i=1}^{\ell} \widehat{C}(|\mathcal{I}_i'|) + 60\sum_{m=0}^n \frac{\alpha_m}{\alpha_n}\widehat{C}(2^m)\log(T/\delta)
\end{align*}
(notations are defined at the beginning of this section).
\end{lemma}
\begin{proof}
$\sum_{\tau=t_n}^{E_n}\left(\widetilde{g}_\tau-R_\tau\right)$ is trivially upper bounded by $3\widehat{C}(E_n-t_n+1) + 1\leq 4\widehat{C}(2^n)$ because it is guarded by \textbf{Test~2}\xspace. Below we focus on the second claim.
Note that we can write for all $i=1,\ldots, K$,
\begin{align*}
&\sum_{\tau\in\mathcal{I}_i'} \left(f_\tau^\star - \widetilde{g}_\tau\right) \\
&\leq 12 \sum_{\tau\in\mathcal{I}_i'} \left( \mathbbm{1}
\Big[ f_\tau^\star - \widetilde{g}_\tau \leq 12\widehat{\alpha}_n \Big] \widehat{\alpha}_n
+ \sum_{m=1}^{n} \mathbbm{1}\Big[ 12\widehat{\alpha}_{m} < f_\tau^\star - \widetilde{g}_\tau \leq 12\widehat{\alpha}_{m-1} \Big] \widehat{\alpha}_{m-1}
+\mathbbm{1}\Big[ f_\tau^\star - \widetilde{g}_\tau > 12\widehat{\alpha}_0 \Big]1 \right)\\
&\leq 12\left(|\mathcal{I}_i'| \widehat{\alpha}_n + \sum_{m=1}^{n} \widehat{\alpha}_{m-1} \xi_i(m) + \rho(1)\xi_i(0)\right) \tag{$\rho(1)\geq 1$ by \prettyref{assum:assump2}} \\
&\leq 12|\mathcal{I}_i'| \widehat{\alpha}_n + 24\sum_{m=0}^{n} \widehat{\alpha}_m \xi_i(m) \tag{$\widehat{\alpha}_m=\frac{\widehat{C}(2^m)}{2^m}\leq \frac{\widehat{C}(2^{m+1})}{2^m} = 2\widehat{\alpha}_{m+1}$}
\end{align*}
where in the second-to-last inequality we use $\sum_{\tau\in\mathcal{I}_i'} \mathbbm{1}\big[f_\tau^\star - \widetilde{g}_\tau \geq 12\widehat{\alpha}_m\big] = \sum_{\tau\in[\tau_i(m), e_i']} \mathbbm{1}\big[f_\tau^\star - \widetilde{g}_\tau \geq 12\widehat{\alpha}_m\big] \leq
\xi_i(m)$ by the definition of $\tau_i(m)$.
Summing the above over intervals $i$ and notice that $\sum_{i=1}^\ell |\mathcal{I}_i'|\leq 2^n$, we get
\begin{align}
\sum_{\tau=t_n}^{E_n}\left(f_\tau^\star - \widetilde{g}_\tau\right)
&\leq 12\cdot 2^n\widehat{\alpha}_n + 24\sum_{m=0}^n \sum_{i=1}^\ell \widehat{\alpha}_m\xi_i(m) = 12\widehat{C}(2^n) + 24\sum_{m=0}^n \sum_{i=1}^\ell \widehat{\alpha}_m\xi_i(m). \label{eq: tmptmp}
\end{align}
Next, we upper bound $ \sum_{i=1}^\ell \widehat{\alpha}_m\xi_i(m)$ for each $m$.
\begin{align}
\sum_{i=1}^\ell \widehat{\alpha}_m\xi_i(m)
&= \sum_{i=1}^\ell \widehat{\alpha}_m \min\left\{\xi_i(m), 4\cdot 2^m\right\} + \sum_{i=1}^\ell \widehat{\alpha}_m \left[\xi_i(m)-4\cdot 2^m\right]_+. \label{eq: bound two individual term}\\
& \tag{using $a= \min\{a,b\} + [a-b]_+$}
\end{align}
The first term on the right-hand side of \prettyref{eq: bound two individual term} can be bounded as below:
\begin{align*}
\sum_{i=1}^\ell \widehat{\alpha}_m \min\left\{\xi_i(m), 4\cdot 2^m\right\}
&\leq 4\sum_{i=1}^\ell \widehat{\rho}(2^m) \times \min\left\{ \xi_i(m), 2^m\right\} \\
&\leq 4\sum_{i=1}^\ell \widehat{\rho}(\min\{\xi_i(m), 2^m\}) \times \min\left\{ \xi_i(m), 2^m\right\} \tag{$\widehat{\rho}(\cdot)$ is a decreasing function} \\
&= 4\sum_{i=1}^{\ell} \widehat{C}(\min\{\xi_i(m), 2^m\}) \\
&\leq 4\sum_{i=1}^{\ell} \widehat{C}(|\mathcal{I}_i'|). \tag{$\widehat{C}(\cdot)$ is an increasing function}
\end{align*}
The second term on the right-hand side of \prettyref{eq: bound two individual term} is bounded using \prettyref{lemma: key term reg bound} below. Combining them into \prettyref{eq: tmptmp} finishes the proof.
\end{proof}
\begin{lemma}
\label{lemma: key term reg bound}
Let the high probability events described in \prettyref{lemma: multi-scale reg} (or \prettyref{lemma: ucrl aggregated regret}) hold. Then with high probability,
\begin{align*}
\sum_{i=1}^\ell \widehat{\alpha}_m \left[\xi_i(m) -4\cdot 2^m\right]_+ \leq \frac{2\alpha_m}{\alpha_n}\widehat{C}(2^m)\log(T/\delta).
\end{align*}
\end{lemma}
\begin{proof}
Using the fact that $[[a]_+ - b]_+ = [a-b]_+$ when $b\geq 0$, we have
\begin{align}
[\xi_i(m) - 4\cdot 2^m]_+=\left[e_i'-\tau_i(m) +1 -4\cdot 2^m \right]_+. \label{eq: simple relation 1}
\end{align}
Next, we consider the following quantity: ``the number of rounds in the interval $[\tau_i(m), e_i'-2\cdot 2^m B]$ which are candidate starting points of an order-$m$ {\small\textsf{\textup{ALG}}}\xspace''. By \prettyref{proc: alg profile}, this quantity can be written and lower bounded as
\begin{align*}
A_i\triangleq \sum_{t\in\mathcal{I}_i} \mathbbm{1}\Big[ t \in [\tau_i(m),~ e_i'- 2\cdot 2^m], \quad (t-t_n) \text{\ mod\ } 2^m = 0\Big] \geq \frac{\left[e_i'-\tau_i(m) +1 - 4\cdot 2^m\right]_+}{2^m}
\end{align*}
where we use the fact in an interval of length $w$, there are at least $\frac{w+2-2u}{u}$ points whose indices are multiples of $u$.
Notice that the right-hand side is related to what we want to upper bound in the lemma according to \prettyref{eq: simple relation 1}. Thus we continue to upper bound the left-hand side above. We define the following events:
\begin{align*}
W_{t} &= \left\{\tau_i(m) \leq t \leq e_i-2\cdot 2^m \text{\ where\ }i\text{\ is such that\ } t\in\mathcal{I}_i \right\}, \\
X_t&= \left\{ t\leq E_n-2\cdot 2^m \right\}, \\
Y_t&=\left\{ t\leq E_n \text{\ and\ } (t-t_n)\text{\ mod\ }2^m=0 \right\}, \\
Z_t&= \left\{ \exists \text{\ order-$m$\ } \inst \text{\ such that\ } \inst.s=t \right\}, \\
V_t&= \left\{ \exists \tau\in[t_n, t] \text{\ such that\ } W_{\tau}\cap Y_{\tau}\cap Z_{\tau} \right\}.
\end{align*}
Then we can write (recall the definition of $K$ in the beginning of this section)
\begin{align*}
\sum_{i=1}^\ell A_i = \sum_{i=1}^K A_i
&= \sum_{t=t_n}^{t_n+2^n-1} \mathbbm{1}[W_t, X_t, Y_t] \leq \underbrace{\sum_{t=t_n}^{t_n+2^n-1} \mathbbm{1}[W_t, Y_t, \overline{V_t}]}_{\textbf{term}_3} + \underbrace{\sum_{t=t_n}^{t_n+2^n-1}\mathbbm{1}[X_t, V_t]}_{\textbf{term}_4}
\end{align*}
For $\textbf{term}_3$, notice that conditioned on $W_t\cap Y_t$, the event $Z_t$ happens with a constant probability $\frac{\alpha_n}{\alpha_m}$ (by \prettyref{proc: alg profile}). Therefore, $\textbf{term}_3$ counts the number of trials up to the first success in a repeated trial with success probabiliy $\frac{\alpha_n}{\alpha_m}$. Therefore, with probability $1-\frac{\delta}{T}$, $\textbf{term}_3\leq 1+ \frac{\log(T/\delta)}{-\log\left(1-\frac{\alpha_n}{\alpha_m}\right)} \leq \frac{2\alpha_m}{\alpha_n}\log(T/\delta)$.
Next, we deal with $\textbf{term}_4$. Below we show that $\textbf{term}_4=0$. The event $V_t$ implies that there exists some order-$m$ \inst which starts at $\inst.s=t^\star$, where $t^\star\leq t$ and $\tau_i(m) \leq t^\star\leq e_i-2\cdot 2^m $. Therefore, we have $\inst.e=\inst.s+2^{m}-1= t^\star + 2^m-1 \leq e_i-2^m-1 < e_i$, and thus $[\inst.s, \inst.e]\subseteq \mathcal{I}_i$. Together with $X_t$, the event $V_t\cap X_t$ implies that $\inst.e = \inst.s + 2^m-1 \leq t+2^m -1 < E_n$, and therefore, and time $\inst.e$, block $n$ has not ended.
Since at time $\inst.e$, block $n$ is still on-going, the learner performs \textbf{Test~1}\xspace. By \prettyref{lemma: multi-scale reg} (or \prettyref{lemma: ucrl aggregated regret} for the infinite-horizon RL case), with high probability, we have
\begin{align*}
\frac{1}{2^m}\sum_{\tau=\inst.s}^{\inst.e} R_\tau
&\geq \frac{1}{2^m } \sum_{\tau=\inst.s}^{\inst.e} \widetilde{g}_\tau - \widehat{\alpha}_m - \widehat{n}\xspace \Delta_{[\inst.s, \inst.e]} \tag{\prettyref{lemma: multi-scale reg} or \prettyref{lemma: ucrl aggregated regret}}\\
&\geq \min_{\tau\in \mathcal{I}_i } f_\tau^\star - \widehat{\alpha}_m - (\widehat{n}\xspace+1) \Delta_{\mathcal{I}_i} \tag{because $[\inst.s, \inst.e]\subseteq \mathcal{I}_i$} \\
&\geq f^\star_{\tau_i(m)} - \widehat{\alpha}_m - (\widehat{n}\xspace+3) \Delta_{\mathcal{I}_i} \tag{$| \min_{\tau\in \mathcal{I}_i } f_\tau^\star - f^\star_{\tau_i(m)}| \leq 2\Delta_{\mathcal{I}_i}$} \\
&\geq \widetilde{g}_{\tau_i(m)} +12\widehat{\alpha}_m - 2\widehat{\alpha}_m \tag{by the definition of $\tau_i(m)$ and $\Delta_{\mathcal{I}_i}\leq \rho(|\mathcal{I}_i|) \leq \rho(2^m)\leq \frac{\widehat{\alpha}_m}{6\widehat{n}\xspace}$} \\
&\geq U_{\inst.e} + 10\widehat{\alpha}_m \tag{Because $\inst.e\geq \tau_i(m)$, $U_{\inst.e}\leq \widetilde{g}_{\tau_i(m)}$ by the algorithm}
\end{align*}
This should trigger the restart at time $\inst.e < E_n$, contradicting the definition of $E_n$. Therefore, $\mathbbm{1}[X_t, V_t]=0$.
Finally, combining all previous arguments, we have that with high probability,
\begin{align*}
\sum_{i=1}^\ell \widehat{\alpha}_m \left[\xi_i(m)-4\cdot 2^m\right]_+ &= \sum_{i=1}^\ell \widehat{\alpha}_m \left[e_i'-\tau_i(m)+1-4\cdot 2^m\right]_+ \leq \widehat{\alpha}_m 2^m\sum_{i=1}^\ell A_i \\
&= \widehat{C}(2^m) \sum_{i=1}^\ell A_i \leq \frac{2\alpha_m}{\alpha_n}\widehat{C}(2^m)\log(T/\delta),
\end{align*}
finishing the proof.
\end{proof}
\section{Single-block Regret Analysis II (under a Special Form of $C(\cdot)$)}
\label{app: epoch regret}
In \prettyref{app: key proofs}, we have derived the regret bound in a single block for both the standard setting and the infinite-horizon MDP setting (\prettyref{lemma: bound term1} and \prettyref{lemma: block lemma for RL}). They are both of the form
\begin{align}
\sum_{\tau\in\mathcal{J}}(f_\tau^\star - R_\tau) = \widetilde{\mathcal{O}}\left( \sum_{i=1}^\ell C(|\mathcal{I}_i'|) + \sum_{m=0}^n \frac{\rho(2^n)}{\rho(2^m)}C(2^m) \right). \label{eq: block regret restate}
\end{align}
(replacing $C(\cdot)$ and $\rho(\cdot)$ by $C_{{\small\textsf{\textup{UCRL}}}\xspace}(\cdot; \overline{D})$ and $\rho(\cdot; \overline{D})$ for the case of infinite-horizon MDP).
In this section, we further derive more concrete dynamic regret bounds for both cases by assuming that $C(\cdot)$ is of some specific form. The form of $C(\cdot)$ we consider in this section is defined as follows:
\begin{definition}
\label{def: standard}
We define a form of $C(t)$ as $C(t)=\min\{c_1 t^p+c_2, c_3t\}$ for some $p\in[\frac{1}{2}, 1)$ and some $c_1$, $c_2$, $c_3$ ($c_3\geq 1$) that capture dependencies on $\log(T/\delta)$ and other problem-dependent constants.
\end{definition}
In fact, usually, a regret bound is only written in the form of $c_1 t^p + c_2$. However, since the reward is bounded between $0$ and $1$, the regret bound of $\min\{c_1 t^p + c_2, t\}$ is also trivially correct. \prettyref{def: standard} is slightly more general than this by allowing a coefficient $c_3\geq 1$ (the regret bound would still be trivially correct). In some cases, we make $c_1, c_2, c_3$ larger than their tightest possible values to make the final regret bound better --- notice that the choice of $c_1, c_2, c_3$ affects the probability specified in \prettyref{proc: alg profile}, and thus smaller $c_1, c_2, c_3$ does not necessarily make the final regret bound smaller. This subtle issue can be observed from the analysis.
To get a concrete bound, we also need to decide the number $\ell$ in the single-block regret bound above. In \prettyref{app: key proofs}, we have stated the condition (i.e., \prettyref{eq: i smaller than alpha}) that should be satisfied by $\mathcal{I}_1', \ldots, \mathcal{I}_\ell'$ (or $\mathcal{I}_1, \ldots, \mathcal{I}_K$). In the next lemma, we upper bound the value of $\ell$ that is required to fulfill the condition.
\begin{lemma}\label{lemma: interval divide}
Let $\mathcal{J}=[t_n, E_n]$. Then we have $\ell \leq L_{\mathcal{J}}$. Furthermore, if $C(t)$ is in the form specified in \prettyref{def: standard}, we also have $\ell\leq 1 + 2\left(c_1^{-1}\Delta_{\mathcal{J}}|\mathcal{J}|^{1-p}\right)^{\frac{1}{2-p}} + c_3^{-1}\Delta_\mathcal{J}$.
\end{lemma}
\begin{proof}
The fact that $\ell\leq L_{\mathcal{J}}$ is straightforward to see (and has been explained in \prettyref{app: key proofs}): to satisfy the condition \prettyref{eq: i smaller than alpha}, one way to divide the block is to make each $\mathcal{I}_i$ a stationary interval, which makes $\Delta_{\mathcal{I}_i}=0$ for all $i\in[K]$. This way of division leads to $\ell \leq L_{\mathcal{J}}$.
For the second claim, we follow the same procedure as decribed in the proof of Lemma 5 in \citep{chen2019new}. Basically, the procedure divides $[t_n, t_n+2^n-1]$ in a greedy way, making all $\mathcal{I}_i=[s_i, e_i]$ satisfy $\Delta_{[s_i, e_i]} \leq \rho(e_i-s_i+1)$ and $\Delta_{[s_i, e_i+1]} > \rho(e_i-s_i+2)$ for all $i\in[K-1]$ (i.e., except for the last interval). Then we have
\begin{align*}
\Delta_{\mathcal{J}} &\geq \sum_{i=1}^{\ell-1} \Delta_{[s_i, e_i+1]} \tag{by the definition of $\Delta_{[\cdot,\cdot]}$} \\
&> \sum_{i=1}^{\ell-1} \rho(e_i-s_i+2) \\
&\geq \sum_{i=1}^{\ell-1} \min\left\{c_1 (e_i-s_i+2)^{p-1}, c_3\right\} \tag{by \prettyref{def: standard}} \\
&\geq \sum_{i=1}^{\ell-1} \min\left\{\frac{1}{2}c_1 (e_i-s_i+1)^{p-1}, c_3\right\} \tag{$(x+2)^{p-1} \geq \left(2(x+1)\right)^{p-1}\geq \frac{1}{2}(x+1)^{p-1}$ for any $x\geq 0$ and $p\leq 1$} \\
&= \frac{1}{2}\sum_{i=1}^{\ell_1} c_1 (e_i-s_i+1)^{p-1} + \sum_{i=1}^{\ell_2} c_3
\end{align*}
where in the last equality we separate the intervals where $\min\left\{\frac{1}{2}c_1 (e_i-s_i+1)^{p-1}, c_3\right\}$ takes the former or the latter value. Note that $\ell_1+\ell_2=\ell-1$.
The above inequality implies that $\Delta_{\mathcal{J}} $ upper bounds both $\frac{1}{2}\sum_{i=1}^{\ell_1} c_1 (e_i-s_i+1)^{p-1} $ and $\sum_{i=1}^{\ell_2} c_3 $. Thus, $\ell_2\leq c_3^{-1}\Delta_{\mathcal{J}}$, and
by H\"{o}lder's inequality,
\begin{align*}
\ell_1 &\leq \left(\sum_{i=1}^{\ell_1} (e_i-s_i+1)^{p-1} \right)^{\frac{1}{2-p}} \left(\sum_{i=1}^{\ell_1} (e_i-s_i+1) \right)^{\frac{1-p}{2-p}} \leq \left(\frac{2\Delta_{\mathcal{J}}}{c_1}\right)^{\frac{1}{2-p}}|\mathcal{J}|^{\frac{1-p}{2-p}}.
\end{align*}
Combining them finishes the proof.
\end{proof}
In the following \prettyref{lemma: single block regret form}, we bound the regret within a block by combining \prettyref{eq: block regret restate} and \prettyref{lemma: interval divide}.
We will frequently use the following two properties: let $\{\mathcal{S}_1, \mathcal{S}_2, \ldots, \mathcal{S}_K\}$ be a partition of the interval $\mathcal{S}$. Then
\begin{align}
\sum_{i=1}^K L_{\mathcal{S}_i} &\leq L_{\mathcal{S}} + (K-1), \label{eq: L partition} \\
\sum_{i=1}^K \Delta_{\mathcal{S}_i} &\leq \Delta_{\mathcal{S}}. \label{eq: dev partition}
\end{align}
They can be derived using the definitions of $L_{[\cdot,\cdot]}$ and $\Delta_{[\cdot,\cdot]}$.
\begin{lemma}
\label{lemma: single block regret form}
If $C(t)$ is of the form specified in \prettyref{def: standard}, then
\begin{align*}
\sum_{\tau=t_n}^{E_n} (f_\tau^\star - R_\tau) \leq \widetilde{\mathcal{O}}\left( \min\Big\{\text{Reg}\xspace_L(\mathcal{J}), \text{Reg}\xspace_{\Delta}(\mathcal{J})\Big\}+ c_12^{np} + \frac{c_2c_3}{c_1}2^{n(1-p)} + \frac{c_2^2}{c_3} \right),
\end{align*}
where $\text{Reg}\xspace_L(\mathcal{J}) \triangleq c_1 L_\mathcal{J}^{1-p} |\mathcal{J}|^{p} + c_2L_\mathcal{J}$ and
\begin{align*}
\text{Reg}\xspace_{\Delta}(\mathcal{J}) &\triangleq \left(c_1 \Delta_{\mathcal{J}}^{1-p}|\mathcal{J}|\right)^{\frac{1}{2-p}} + c_1|\mathcal{J}|^p + c_1(c_3^{-1}\Delta_{\mathcal{J}})^{1-p} |\mathcal{J}|^{p} + c_2\left(c_1^{-1}\Delta_{\mathcal{J}}|\mathcal{J}|^{1-p}\right)^{\frac{1}{2-p}} + c_2 + c_2c_3^{-1}\Delta_\mathcal{J}.
\end{align*}
\end{lemma}
\begin{proof}
We bound each term in \prettyref{eq: block regret restate} using \prettyref{def: standard}. First, notice that
\begin{align}
\widetilde{\mathcal{O}}\left( \sum_{i=1}^\ell C(|\mathcal{I}_i'|) \right)
&= \widetilde{\mathcal{O}}\left( \sum_{i=1}^\ell \min\left\{c_1|\mathcal{I}_i'|^p + c_2, \ c_3t \right\} \right) \nonumber \\
&\leq \widetilde{\mathcal{O}}\left(\sum_{i=1}^\ell (c_1|\mathcal{I}_i'|^p + c_2) \right)
\leq \widetilde{\mathcal{O}}\left(c_1 \ell^{1-p} |\mathcal{J}|^{p} + c_2\ell\right). \label{eq: regret first term}
\end{align}
Using the first upper bound for $\ell$ given in \prettyref{lemma: interval divide}, \prettyref{eq: regret first term} can be bounded by $ \widetilde{\mathcal{O}}\left(\text{Reg}\xspace_L(\mathcal{J})\right) $;
using the second upper bound, \prettyref{eq: regret first term} can be bounded by
$\widetilde{\mathcal{O}}\left(\text{Reg}\xspace_{\Delta}(\mathcal{J})\right)$.
Next, we have
\begin{align*}
&\widetilde{\mathcal{O}}\left(\frac{\rho(2^m)}{\rho(2^n)}C(2^m)\right)
= \widetilde{\mathcal{O}}\left(c_1 2^{np} + \frac{c_2c_3}{c_1}2^{n(1-p)} + \frac{c_1^2}{c_3} 2^{m(2p-1)} + \frac{c_2^2}{c_3} 2^{-m} \right).
\end{align*}
by \prettyref{lem: alpha calculation} below.
Notice that because $c_3\geq 1$ and $p\geq \frac{1}{2}$, $\frac{c_1^2}{c_3} 2^{m(2p-1)} \leq c_1^2 2^{n(2p-1)} \leq c_12^{np}$ when $c_1\leq 2^{n(1-p)}$. This is indeed the regime we care about since if $c_1 > 2^{n(1-p)}$ then the first term $c_1 2^{np} > 2^n$, which is a vacuous bound for the regret of block $n$.
Therefore, we can drop this term. Thus, the dynamic regret in block $n$ can be summarized as the following based on \prettyref{eq: block regret restate}:
\begin{align}
\widetilde{\mathcal{O}}\left( \min\Big\{\text{Reg}\xspace_L(\mathcal{J}), \text{Reg}\xspace_{\Delta}(\mathcal{J})\Big\}+ c_12^{np} + \frac{c_2c_3}{c_1}2^{n(1-p)} + \frac{c_2^2}{c_3} \right),
\end{align}
finishing the proof.
\end{proof}
\begin{lemma}\label{lem: alpha calculation}
Let $C(t)$ be of the form in \prettyref{def: standard}. Then
\begin{align*}
\frac{\rho(2^m)}{\rho(2^n)}C(2^m)=\mathcal{O}\left(c_1 2^{np} + \frac{c_2c_3}{c_1}2^{n(1-p)} + \frac{c_1^2}{c_3} 2^{m(2p-1)} + \frac{c_2^2}{c_3} 2^{-m} \right).
\end{align*}
\end{lemma}
\begin{proof}
This is by direct calculation:
\begin{align*}
\frac{\rho(2^m)}{\rho(2^n)}C(2^m)
&= \frac{C(2^m)^2 }{C(2^n)} 2^{n-m}\\
&=\mathcal{O}\left( \frac{\min\{c_1^2 2^{2mp} + c_2^2,\ \ c_3^2 2^{2m} \}}{c_12^{np} + c_2}2^{n-m}+\frac{\min\{c_1^2 2^{2mp} + c_2^2,\ \ c_3^2 2^{2m} \}}{c_3 2^n}2^{n-m}\right) \\
&=\mathcal{O}\left( \min\left\{c_1 2^{np}2^{(n-m)(1-2p)} + \frac{c_2^2}{c_1}2^{n(1-p)-m}, \ \ \frac{c_3^2}{c_1}2^{n(1-p)+m} \right\} + \frac{c_1^2}{c_3} 2^{m(2p-1)} + \frac{c_2^2}{c_3} 2^{-m} \right) \\
&= \mathcal{O}\left(c_1 2^{np}+ \min\left\{\frac{c_2^2}{c_1}2^{n(1-p)-m}, \ \ \frac{c_3^2}{c_1}2^{n(1-p)+m} \right\}+ \frac{c_1^2}{c_3} 2^{m(2p-1)} + \frac{c_2^2}{c_3} 2^{-m} \right) \\
&= \mathcal{O}\left(c_1 2^{np} + \frac{c_2c_3}{c_1}2^{n(1-p)} + \frac{c_1^2}{c_3} 2^{m(2p-1)} + \frac{c_2^2}{c_3} 2^{-m} \right).
\end{align*}
\end{proof}
\section{Single-epoch Regret Analysis}\label{app: epoch regret seperate}
We call $[t_0, E]$ an \emph{epoch} if $t_0$ is the first step after restart (or $t_0=1$), and $E$ is the first time after round $t_0$ when the restart is triggered.
In this section, we continue the discussion in \prettyref{app: epoch regret} and bound the regret in a single epoch. Recall that the we consider cases where the single-block regret can be written as \prettyref{eq: block regret restate} and $C(\cdot)$ is in the form of \prettyref{def: standard}. This holds both for the case of the standard setting and the infinite-horizon MDP setting.
\begin{lemma}\label{lemma: sum epoch regret}
Let $\mathcal{E}$ be an epoch. Then
\begin{align*}
\sum_{\tau\in\mathcal{E}}\leq \widetilde{\mathcal{O}}\left(\min\big\{ \text{Reg}\xspace_L(\mathcal{E}), \text{Reg}\xspace_{\Delta}(\mathcal{E}) \big\} + \frac{c_2c_3}{c_1}|\mathcal{E}|^{1-p} + \frac{c_2^2}{c_3}\right)
\end{align*}
($\text{Reg}\xspace_L(\cdot)$ and $\text{Reg}\xspace_\Delta(\cdot)$ are defined in \prettyref{lemma: single block regret form})
\end{lemma}
\begin{proof}
Let $\mathcal{E}$ be an epoch whose last block is indexed by $n$. Then $|\mathcal{E}|=\Theta(2^n)$. Let $\mathcal{J}_1, \ldots, \mathcal{J}_n$ be blocks in $\mathcal{E}$. Then by \prettyref{lemma: single block regret form}, the dynamic regret in $\mathcal{E}$ is upper bounded by
\begin{align*}
&\widetilde{\mathcal{O}}\left(\min\left\{ \sum_{m=0}^n \text{Reg}\xspace_L(\mathcal{J}_m), \ \sum_{m=0}^n \text{Reg}\xspace_\Delta(\mathcal{J}_m) \right\}+ c_1\sum_{m=0}^n 2^{mp} + \frac{c_2c_3}{c_1}\sum_{m=0}^n 2^{m(1-p)} + \sum_{m=0}^n \frac{c_2^2}{c_3}\right).
\end{align*}
By H\"{o}lder's inequality,
\begin{align*}
\sum_{m=0}^n \text{Reg}\xspace_L(\mathcal{J}_m)
&= c_1\left(\sum_{m=0}^n L_{\mathcal{J}_m} \right)^{1-p}\left(\sum_{m=0}^n |\mathcal{J}_m|\right)^p + c_2 \sum_{m=0}^n L_{\mathcal{J}_m} \\
&\leq c_1 \left(L_{\mathcal{E}} + n \right)^{1-p} |\mathcal{E}|^p + c_2\left(L_{\mathcal{E}} + n \right) \tag{using \prettyref{eq: L partition}}\\
&\leq \widetilde{\mathcal{O}}\left(c_1 L_\mathcal{E}^{1-p} |\mathcal{E}|^{p} + c_2 L_\mathcal{E} \right)
= \widetilde{\mathcal{O}}\left(\text{Reg}\xspace_L(\mathcal{E})\right) \tag{because $n=\mathcal{O}(\log T)=\widetilde{\mathcal{O}}(1)$}
\end{align*}
Similarly, $\sum_{m=0}^n \text{Reg}\xspace_{\Delta}(\mathcal{J}_m) = \widetilde{\mathcal{O}}\left(\text{Reg}\xspace_{\Delta}(\mathcal{E})\right)$.
On the other hand, $c_1\sum_{m=0}^n 2^{mp} + \frac{c_2c_3}{c_1}\sum_{m=0}^n 2^{m(1-p)} + \sum_{m=0}^n \frac{c_2^2}{c_3} = \widetilde{\mathcal{O}}\left(c_1 2^{np} + \frac{c_2c_3}{c_1} 2^{n(1-p)} + \frac{c_2^2}{c_3}\right) = \widetilde{\mathcal{O}}\left(c_1 |\mathcal{E}|^p + \frac{c_2c_3}{c_1} |\mathcal{E}|^{1-p} + \frac{c_2^2}{c_3} \right)$. In summary, the dynamic regret within an epoch is of order
\begin{align}
\widetilde{\mathcal{O}}\left(\min\big\{ \text{Reg}\xspace_L(\mathcal{E}), \text{Reg}\xspace_{\Delta}(\mathcal{E}) \big\} + \frac{c_2c_3}{c_1}|\mathcal{E}|^{1-p} + \frac{c_2^2}{c_3}\right) \label{eq: epoch regret}
\end{align}
(the $c_1|\mathcal{E}|^p$ term is absorbed into $\min\left\{ \text{Reg}\xspace_L(\mathcal{E}), \text{Reg}\xspace_{\Delta}(\mathcal{E}) \right\}$).
\end{proof}
\section{Proof of \prettyref{thm: regret bound}} \label{app: omitted proof}
We are now ready to prove \prettyref{thm: regret bound} after showing the following two lemmas.
\begin{lemma}\label{lem: infreq restart}
Let $t$ be in an epoch starting from $t_0$. If $\Delta_{[t_0, t]}\leq \rho(t-t_0+1)$, then with high probability, no restart would be triggered at time $t$.
\end{lemma}
\begin{proof}
We first verify that \textbf{Test~1}\xspace would not fail with high probability. Let $t=\inst.e$ where $\inst$ is any order-$m$ {\small\textsf{\textup{ALG}}}\xspace in block $n$. Then with high probability,
\begin{align*}
U_t
&= \min_{\tau\in [t_n, t]} \widetilde{g}_\tau \\
&\geq \min_{\tau\in[t_n, t]} f^\star_\tau - \Delta_{[t_n, t]} \tag{by \prettyref{lemma: multi-scale reg}} \\
&\geq \frac{1}{2^m}\sum_{\tau\in[\inst.s, t]} f^\star_\tau - 3\Delta_{[t_n, t]} \tag{$[\inst.s, t]\subseteq [t_n, t]$} \\
&\geq \frac{1}{2^m} \sum_{\tau\in[\inst.s, t]} R_\tau - 2\sqrt{\frac{\log(T/\delta)}{2^m}} - 3\rho(t-t_0+1) \tag{$\mathbb{E}[R_\tau]=\mathbb{E}[f_\tau(\pi_t)]\leq f_\tau^\star$ and we use Azuma's inequality}\\
&\geq \frac{1}{2^m} \sum_{\tau\in[\inst.s, t]} R_\tau - \widehat{\rho}(2^m) - 3\rho(t-t_0+1) \tag{By \prettyref{assum:assump2}, $\widehat{\rho}(2^m)\geq 6\log(T/\delta)\rho(2^m)\geq 6\log(T/\delta)\sqrt{\frac{1}{2^m}}$}\\
&\geq \frac{1}{2^m} \sum_{\tau\in[\inst.s, t]} R_\tau - 2\widehat{\rho}(2^m). \tag{$\rho(t-t_0+1)\leq \rho(2^m)$ because $\rho(\cdot)$ is decreasing}
\end{align*}
So with high probability, \textbf{Test~1}\xspace will not return fail.
Furthermore, by \prettyref{lemma: multi-scale reg}, with high probability,
\begin{align*}
\frac{1}{t-t_n+1}\sum_{\tau=t_n}^t \left(\widetilde{g}_\tau - R_\tau\right)\leq \widehat{\rho}(t-t_n+1) + \Delta_{[t_n, t]} \leq 2\widehat{\rho}(t-t_n+1).
\end{align*}
Therefore, with high probability, \textbf{Test~2}\xspace will not return fail either.
\end{proof}
\begin{lemma}
\label{lemma: epoch upper}
With high probability, the number of epochs is upper bounded by $L$. If $C(\cdot)$ is in the form of \prettyref{def: standard}, the number of epochs is also upper bounded by $1+2\left(c_1^{-1}\Delta T^{1-p}\right)^{\frac{1}{2-p}} + c_3^{-1}\Delta$.
\end{lemma}
\begin{proof}
By \prettyref{lem: infreq restart}, if $[t_0, E]$ is not the last epoch, then $\Delta_{[t_0, E]}> \rho(E-t_0+1)$ with high probability. Then following the exact same arguments as in \prettyref{lemma: interval divide} proves the lemma.
\end{proof}
\begin{proof}[Proof of \prettyref{thm: regret bound}]\ \ If $C(t)=c_1t^p + c_2$ satisfies \prettyref{assum:assump2}, then $C(t)=\min\{c_1 t^p + c_2, t\}$ also satisfies it (since the reward is bounded in $[0, 1]$). Below we use $C(t)=\min\{c_1 t^p + c_2, t\}$ as the input to our algorithm. Notice that this is in the form of \prettyref{def: standard} with $c_3=1$. Let $\mathcal{E}_1, \ldots, \mathcal{E}_N$ be epochs in $[1, T]$. Then by \prettyref{lemma: sum epoch regret}, the dynamic regret in $[1, T]$ is upper bounded by
\begin{align}
&\widetilde{\mathcal{O}}\left( \min\left\{ \sum_{i=1}^N \text{Reg}\xspace_L(\mathcal{E}_i), \sum_{i=1}^N \text{Reg}\xspace_\Delta(\mathcal{E}_i) \right\} + \frac{c_2}{c_1}\sum_{i=1}^N |\mathcal{E}_i|^{1-p} + c_2^2 N \right). \label{eq: epoch regret bound ha}
\end{align}
By H\"{o}lder's inequality and \prettyref{eq: L partition},
\begin{align*}
\sum_{i=1}^N \text{Reg}\xspace_L(\mathcal{E}_i) \leq \widetilde{\mathcal{O}}\left( c_1 \left( L + N-1 \right)^{1-p} T^p + c_2 (L+N-1)\right) \leq \widetilde{\mathcal{O}}\left(c_1L^{1-p}T^{p} + c_2L\right),
\end{align*}
where in the last inequality we use \prettyref{lemma: epoch upper} to bound $N$.
Similarly,
\begin{align*}
&\sum_{i=1}^N \text{Reg}\xspace_\Delta(\mathcal{E}_i) \\
&\leq \widetilde{\mathcal{O}}\left( \left(c_1 \Delta^{1-p}T\right)^{\frac{1}{2-p}} + c_1N^{1-p}T^p + c_1\Delta^{1-p} T^{p} + c_2\left(c_1^{-1}\Delta T^{1-p}\right)^{\frac{1}{2-p}} + c_2 N + c_2\Delta\right) \\
&\leq \widetilde{\mathcal{O}}\left( \left(c_1 \Delta^{1-p}T\right)^{\frac{1}{2-p}} + c_1T^p +c_1\Delta^{1-p} T^{p} + c_2\left(c_1^{-1}\Delta T^{1-p}\right)^{\frac{1}{2-p}} + c_2 + c_2\Delta\right). \tag{using \prettyref{lemma: epoch upper} to bound $N$}
\end{align*}
Then we deal with the second term in \prettyref{eq: epoch regret bound ha}:
\begin{align*}
\frac{c_2}{c_1} \sum_{i=1}^N |\mathcal{E}_i|^{1-p} \leq \frac{c_2}{c_1} N^{p}T^{1-p},
\end{align*}
which can be either bounded by $\widetilde{\mathcal{O}}\left(\frac{c_2}{c_1} L^p T^{1-p} \right)$ or
\begin{align*}
\widetilde{\mathcal{O}}\left(\frac{c_2}{c_1}T^{1-p} + \frac{c_2}{c_1}\left(c_1^{-p}\Delta^{p}T^{2-2p}\right)^{\frac{1}{2-p}} + \frac{c_2}{c_1}\Delta^p T^{1-p}\right)
\end{align*}
using the upper bound for $N$ in \prettyref{lemma: epoch upper}. Finally, the third term in \prettyref{eq: epoch regret bound ha} can be upper bounded either by $\widetilde{\mathcal{O}}\left(c_2^2 L\right)$ or
\begin{align*}
\widetilde{\mathcal{O}}\left(c_2^2 + c_2^2\left(c_1^{-1}\Delta T^{1-p}\right)^{\frac{1}{2-p}} + c_2^2\Delta \right).
\end{align*}
With all terms expanded, below, we collect the dominant terms for the cases of $p=\frac{1}{2}$ and $p>\frac{1}{2}$. We say term $a(T)$ is dominated by $b(T)$ if $\lim_{T\rightarrow \infty} a(T)/b(T)=0$ under any sublinear growth rate of $L$ or $\Delta$ (e.g., $\sqrt{\Delta T}$ is dominated by $\Delta^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}$ and $L$ is dominated by $\sqrt{LT}$). And below we only write down terms that are not dominated by other terms.
\paragraph{The case for $p=\frac{1}{2}$:}
\begin{align*}
\widetilde{\mathcal{O}}\left( \min\left\{\left(c_1 + \frac{c_2}{c_1}\right)\sqrt{LT}, \quad \left(c_1^{\nicefrac{2}{3}} + c_2c_1^{-\nicefrac{4}{3}}\right)\Delta^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}} + \left(c_1 + \frac{c_2}{c_1}\right)\sqrt{T} \right\} \right) ;
\end{align*}
\paragraph{The case for $p>\frac{1}{2}$:}
\begin{align*}
\widetilde{\mathcal{O}}\left(\min\left\{c_1 L^{1-p} T^{p}, \quad \left(c_1\Delta^{1-p}T\right)^{\frac{1}{2-p}} + c_1T^{p}\right\}\right).
\end{align*}
This finishes the proof.
\end{proof}
\section{Main Results for Infinite-horizon MDP}\label{app: omit proof for RL avg}
\begin{lemma}[\textit{c.f.} \prettyref{lem: infreq restart}]
\label{lemma: infrequent RL restart}
Let $t$ be in an epoch started from round $t_0$. If $\Delta_{[t_0, t]; \overline{D}} < \overline{D} S\sqrt{\frac{A}{t-t_0+1}}$ and $\overline{D}\geq D_{\max}$, then with high probability, no restart will be triggered at time $t$.
\end{lemma}
\begin{proof}
To verify that \textbf{Test~1}\xspace will not fail with high probability, we follow very similar steps as in \prettyref{lem: infreq restart}. Let $t=\inst.e$ where $\inst$ is an order-$m$ {\small\textsf{\textup{ALG}}}\xspace in block $n$. Then with high probability (the following calculation is same as that in the proof of \prettyref{lem: infreq restart} except for the third inequality),
\begin{align*}
U_t
&= \min_{\tau\in [t_n, t]} \widetilde{g}_\tau \\
&\geq \min_{\tau\in[t_n, t]} J^\star_\tau - \Delta_{[t_n, t]; \overline{D}} \tag{by \prettyref{lemma: ucrl aggregated regret}} \\
&\geq \frac{1}{2^m}\sum_{\tau\in[\inst.s, t]} J^\star_\tau - 3\Delta_{[t_n, t]; \overline{D}} \tag{$[\inst.s, t]\subseteq [t_0, t]$} \\
&\geq \frac{1}{2^m} \sum_{\tau\in[\inst.s, t]} R_\tau - 4\overline{D} \sqrt{\frac{\log(T/\delta)}{2^m}} - 4\Delta_{[t_n, t]; \overline{D}} \tag{explained below}\\
&\geq \frac{1}{2^m} \sum_{\tau\in[\inst.s, t]} R_\tau - 4\widehat{\rho}_{\small\textsf{\textup{UCRL}}}\xspace(2^m; \overline{D}) - 4\rho_{\small\textsf{\textup{UCRL}}}\xspace(t-t_0+1; \overline{D}) \\
&\geq \frac{1}{2^m} \sum_{\tau\in[\inst.s, t]} R_\tau - 8\widehat{\rho}_{\small\textsf{\textup{UCRL}}}\xspace(2^m; \overline{D}). \tag{$\rho(t-t_0+1)\leq \rho(2^m)$ because $\rho(\cdot)$ is decreasing}
\end{align*}
where the third inequality is based on the following calculation: for all $\tau\in[\inst.s, t]$,
\begin{align*}
J_t^\star
&= r_t(s_\tau, a_\tau) + \sum_{s'} p_t(s'|s_\tau, a_\tau) h_t^\star(s') - h_t^\star(s_\tau) \\
&\geq r_\tau(s_\tau, a_\tau) + \sum_{s'} p_\tau(s'|s_\tau, a_\tau) h_t^\star(s') - h_t^\star(s_\tau) - \left(\Delta^r_{[\inst.s, t]} + D_{\max} \Delta^p_{[\inst.s, t]}\right) \\
&\geq r_\tau(s_\tau, a_\tau) + \sum_{s'} p_\tau(s'|s_\tau, a_\tau) h_t^\star(s') - h_t^\star(s_\tau) - \left(\Delta^r_{[\inst.s, t]} + \overline{D} \Delta^p_{[\inst.s, t]}\right) \tag{by the assumption $\overline{D}\geq D_{\max}$}
\end{align*}
and thus
\begin{align*}
\frac{1}{2^m}\sum_{\tau\in[\inst.s, t]} J^\star_\tau
&\geq J^\star_t - \Delta_{[\inst.s, t]}^J \\
&\geq \frac{1}{2^m} \sum_{\tau\in[\inst.s, t]} \left(r_\tau(s_\tau, a_\tau) + \sum_{s'} p_\tau(s'|s_\tau, a_\tau) h_t^\star(s') - h_t^\star(s_\tau)\right) - \Delta_{[\inst.s, t]; \overline{D}} \\
&\geq \frac{1}{2^m}\sum_{\tau\in[\inst.s, t]} \Big(R_\tau + h_t^\star(s_{\tau+1}) - h_t^\star(s_\tau)\Big) - 2D_{\max}\sqrt{\frac{2\log(SAT/\delta)}{2^m}} - \Delta_{[\inst.s, t]; \overline{D}} \tag{Azuma's inequality} \\
&\geq \frac{1}{2^m}\sum_{\tau\in[\inst.s, t]} R_\tau - 4\overline{D}\sqrt{\frac{\log(SAT/\delta)}{2^m}} - \Delta_{[\inst.s, t]; \overline{D}}. \tag{$D_{\max}\leq \overline{D}$}
\end{align*}
So with high probability, \textbf{Test~1}\xspace will not return fail.
Furthremore, by \prettyref{lemma: ucrl aggregated regret}, with high probability,
\begin{align*}
\frac{1}{t-t_n+1}\sum_{\tau=t_n}^t \left(\widetilde{g}_\tau - R_\tau\right)\leq \widehat{\rho}_{\small\textsf{\textup{UCRL}}}\xspace(t-t_n+1; \overline{D}) + \Delta_{[t_n, t]; \overline{D}} \leq 2\widehat{\rho}_{{\small\textsf{\textup{UCRL}}}\xspace}(t-t_n+1; \overline{D})
\end{align*}
where the last inequality is by the condition on $\Delta_{[t_0, t], \overline{D}}$.
Therefore, with high probability, \textbf{Test~2}\xspace will not return fail either.
It remains to show that the {\small\textsf{\textup{UCRL-ACW}}}\xspace will not terminate and call for restart under the specified condition. By \prettyref{lemma: bound diamete}, if $\mathcal{P}_k^\eta$ contains an MDP whose diameter is upper bounded by $\overline{D}$, then the span of the output bias vector is upper bounded by $2\overline{D}$, and then the if-statement in \prettyref{line: enough widening} of \prettyref{alg: base alg UCRL} will be triggered. Therefore, to show that {\small\textsf{\textup{UCRL-ACW}}}\xspace will not terminate, we upper bound the $\eta_k$ that needs to be added to $\mathcal{P}_k$ in order to make at least one true MDP (whose diameter is upper bounded by $D_{\max}\leq \overline{D}$) lie in $\mathcal{P}_k^{\eta_k}$. Then we further argue that $\sum_{\tau=t_0}^{t} \eta_{k(\tau)}$ is not large enough to reach the condition in \prettyref{line: terminate condition} of \prettyref{alg: base alg UCRL}.
For all episode $k$ that starts before $t$, by Azuma's inequality,
\begin{align*}
\left\|\overline{p}_k(\cdot|s,a) - \widehat{p}_k(\cdot|s,a)\right\|_1 \leq 2\sqrt{\frac{S\log(1/\delta)}{N_k^+(s,a)}}.
\end{align*}
By the condition on $\Delta_{[t_0,t]; \overline{D}}$, we have
\begin{align*}
\left\|p_{t_0}(\cdot|s,a) - \overline{p}_k(\cdot|s,a) \right\|_1 \leq \frac{1}{\overline{D}}\Delta_{[t_0,t]; \overline{D}} \leq S\sqrt{\frac{A}{t-t_0+1}}.
\end{align*}
Combining them, we get
\begin{align*}
\left\|p_{t_0}(\cdot|s,a) - \widehat{p}_k(\cdot|s,a)\right\|_1 \leq 2\sqrt{\frac{S\log(1/\delta)}{N_k^+(s,a)}} + S\sqrt{\frac{A}{t-t_0+1}}.
\end{align*}
Therefore, we see that in \prettyref{line: terminate condition} of \prettyref{alg: base alg UCRL}, as long as $\eta \geq S\sqrt{\frac{A}{t-t_0+1}}$, $p_{t_0}$ is contained in $\mathcal{P}^\eta_k$. Then we have $\text{sp}(\widetilde{h})\leq 2D_{\max}\leq 2\overline{D}$ by \prettyref{lemma: bound diamete}, and the for-loop will be broken at this $\eta$.
Thus we conclude that $\eta_k\leq 2S\sqrt{\frac{A}{t-t_0+1}}$ for all episode $k$ started before $t$. Thus, $\sum_{\tau=t_0}^t \eta_{k(\tau)} \leq (t-t_0+1)\times 2S\sqrt{\frac{A}{t-t_0+1}}=2S\sqrt{A(t-t_0+1)}$, and thus the algorithm will not terminate and call for restart at time $t$.
\end{proof}
\begin{lemma}[\textit{c.f.} \prettyref{lemma: epoch upper}]
\label{lemma: RL number epoch}
If $\overline{D}\geq D_{\max}$, then the number of epochs is upper bounded by $\min\left\{L,\ \ 1+ 3\left(\frac{\Delta^r + \Delta^p}{S\sqrt{A}}\right)^{\frac{2}{3}} T^{\frac{1}{3}}\right\}$.
\end{lemma}
\begin{proof}
Let $\mathcal{E}_1, \ldots, \mathcal{E}_N$ be the epochs. By \prettyref{lemma: infrequent RL restart}, for $i\leq N-1$, we must have $\Delta_{\mathcal{E}_i; \overline{D}}\geq \overline{D} S\sqrt{\frac{A}{|\mathcal{E}_i|}}$. By H\"{o}lder's inequality,
\begin{align*}
N-1
&\leq \left(\sum_{i=1}^{N-1}\frac{1}{\sqrt{|\mathcal{E}_i|}}\right)^{\frac{2}{3}} \left(\sum_{i=1}^{N-1} |\mathcal{E}_i| \right)^{\frac{1}{3}}\leq \left(\frac{\Delta_{[1, T]; \overline{D}}}{\overline{D} S\sqrt{A}}\right)^{\frac{2}{3}}T^{\frac{1}{3}}.
\end{align*}
We can further upper bound the term $\frac{1}{\overline{D}} \Delta_{[1,T]; \overline{D}}$ as follows:
\begin{align*}
\frac{1}{\overline{D}} \Delta_{[1,T]; \overline{D}}
&= \frac{1}{\overline{D}}\left(\Delta^r + 2\overline{D} \Delta^p + \Delta^J \right) \\
&\leq \frac{1}{\overline{D}}\left(2\Delta^r+ (2\overline{D} + D_{\max}) \Delta^p \right) \tag{by \prettyref{lemma: ortner lemma}} \\
&\leq \frac{3}{\overline{D}}\left(\Delta^r + \overline{D} \Delta^p \right) \\
&\leq 3(\Delta^r + \Delta^p).
\end{align*}
Thus we get
\begin{align*}
N\leq 1+ 3\left(\frac{\Delta^r + \Delta^p}{S\sqrt{A}}\right)^{\frac{2}{3}} T^{\frac{1}{3}}.
\end{align*}
Also, by \prettyref{lemma: infrequent RL restart}, when $\overline{D} \geq D_{\max}$, an epoch is created only when the reward function or the transition function changes. Thus the number of epochs is also upper bounded by $L$.
\end{proof}
\begin{lemma}\label{lemma: single epoch RL regret}
In every epoch $\mathcal{E}$, the dynamic regret of {\small\textsf{\textup{MASTER}}}\xspace-{\small\textsf{\textup{UCRL}}}\xspace is upper bounded by
\begin{align*}
\widetilde{\mathcal{O}}\left(X + \overline{D} S\sqrt{A|\mathcal{E}|} + \overline{D} S^2A^2 \right),
\end{align*}
where $X$ is the minimum of the following two terms:
\[
\overline{D} S\sqrt{AL_{\mathcal{E}}|\mathcal{E}|} + \overline{D} SA L_{\mathcal{E}}
\]
and
\[
\left(\overline{D}^2 S^2 A\Delta_{\mathcal{E}; \overline{D}} |\mathcal{E}|^2\right)^{\nicefrac{1}{3}} + SA\sqrt{\overline{D} \Delta_{\mathcal{E}; \overline{D}}|\mathcal{E}|} + \left(\overline{D} S A^2 \Delta_{\mathcal{E}; \overline{D}}^2|\mathcal{E}|\right)^{\nicefrac{1}{3}} + SA\Delta_{\mathcal{E}; \overline{D}}.
\]
\end{lemma}
\begin{proof}
Let $\mathcal{E}_1, \ldots, \mathcal{E}_N$ be the epochs.
By \prettyref{lemma: sum epoch regret}, we know that the regret within an epoch $\mathcal{E}$ is $\widetilde{\mathcal{O}}\left(\min\left\{\text{Reg}\xspace_L(\mathcal{E}), \text{Reg}\xspace_{\Delta}(\mathcal{E})\right\} + \frac{c_2c_3}{c_1}|\mathcal{E}|^{1-p} + \frac{c_2^2}{c_3}\right)$
with
\begin{align*}
\text{Reg}\xspace_{L}(\mathcal{E})&=c_1 L_{\mathcal{E}}^{1-p} |\mathcal{E}|^{p} + c_2L_{\mathcal{E}}, \\
\text{Reg}\xspace_{\Delta}(\mathcal{E}) &= \left(c_1 \Delta_{\mathcal{E}; \overline{D}}^{1-p}|\mathcal{E}|\right)^{\frac{1}{2-p}} + c_1|\mathcal{E}|^p + c_1(c_3^{-1}\Delta_{\mathcal{E}; \overline{D}})^{1-p} |\mathcal{E}|^{p} + c_2\left(c_1^{-1}\Delta_{\mathcal{E}; \overline{D}}|\mathcal{E}|^{1-p}\right)^{\frac{1}{2-p}} + c_2 + c_2c_3^{-1}\Delta_{\mathcal{E}; \overline{D}}
\end{align*}
when $C(t)$ is in the form of \prettyref{def: standard}. In our case $C_{{\small\textsf{\textup{UCRL}}}\xspace}(t; \overline{D})$ is in this form with $c_1=\overline{D} S\sqrt{A}$, $c_2=\overline{D} SA$, $c_3=\overline{D}$, and $p=\frac{1}{2}$. Using them in the bound above, we get that in an epoch, the dynamic regret is upper bounded by
\begin{align*}
\widetilde{\mathcal{O}}\left(\min\left\{\text{Reg}\xspace_L(\mathcal{E}), \text{Reg}\xspace_{\Delta}(\mathcal{E})\right\} + \overline{D}\sqrt{A|\mathcal{E}|} + \overline{D} S^2A^2\right)
\end{align*}
where
\begin{align*}
\text{Reg}\xspace_L(\mathcal{E}) &=\overline{D} S\sqrt{AL_{\mathcal{E}}|\mathcal{E}|} + \overline{D} SA L_{\mathcal{E}}\\
\text{Reg}\xspace_{\Delta}(\mathcal{E}) &= \left(\overline{D}^2 S^2 A\Delta_{\mathcal{E}; \overline{D}} |\mathcal{E}|^2\right)^{\nicefrac{1}{3}} + \overline{D} S\sqrt{A|\mathcal{E}|} + SA\sqrt{\overline{D} \Delta_{\mathcal{E}; \overline{D}}|\mathcal{E}|} + \left(\overline{D} S A^2 \Delta_{\mathcal{E}; \overline{D}}^2|\mathcal{E}|\right)^{\nicefrac{1}{3}} + \overline{D} SA + SA\Delta_{\mathcal{E}; \overline{D}}.
\end{align*}
Collecting terms finishes the proof.
\end{proof}
\begin{theorem}
\label{thm: known Dmax case}
If $D_{\max}\leq \overline{D} \leq 2D_{\max}$, then {\small\textsf{\textup{MASTER-UCRL}}}\xspace guarantees the following dynamic regret bound:
\begin{align*}
\widetilde{\mathcal{O}}\left( \min\left\{D_{\max}S\sqrt{ALT}, D_{\max}\left(S^2A\right)^{\nicefrac{1}{3}}(\Delta^r + \Delta^r)^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}} + D_{\max}S\sqrt{AT} \right\} \right).
\end{align*}
\end{theorem}
\begin{proof}
Let $\mathcal{E}_1, \ldots, \mathcal{E}_N$ be the epochs. The per epoch dynamic regret is given by \prettyref{lemma: single epoch RL regret}. Combining them with H\"{o}lder's inequality and \prettyref{eq: L partition}, \prettyref{eq: dev partition}, the dynamic regret in $[1, T]$ can be upper bounded by
\begin{align}
\widetilde{\mathcal{O}}\left(\min\{\text{Reg}\xspace_L, \text{Reg}\xspace_{\Delta}\} + \overline{D} S\sqrt{ANT} + \overline{D} S^2A^2 N \right) \label{eq: tmptmp6}
\end{align}
where
\begin{align}
\text{Reg}\xspace_{L}=\overline{D} S\sqrt{A(L+N)T} + \overline{D} SA(L+N) \label{eq: tmptmp7}
\end{align}
and
\begin{align}
\text{Reg}\xspace_{\Delta}
&=\left(\overline{D} ^2S^2A \Delta_{[1, T]; \overline{D}} T^2\right)^{\nicefrac{1}{3}} + SA\sqrt{\overline{D} \Delta_{[1,T]; \overline{D}} T} + \left(\overline{D} SA^2 \Delta_{[1,T]; \overline{D}}^2 T\right)^{\nicefrac{1}{3}} + SA\Delta_{[1,T]; \overline{D}}. \label{eq: tmptmp8}
\end{align}
Since $\overline{D}\geq D_{\max}$, the number of epochs can be bounded using \prettyref{lemma: RL number epoch}:
\begin{align*}
N\leq \min\left\{L,\ \ 1+ 3\left(\frac{\Delta^r + \Delta^p}{S\sqrt{A}}\right)^{\frac{2}{3}} T^{\frac{1}{3}}\right\}.
\end{align*}
With $N\leq L$, \prettyref{eq: tmptmp6}, and \prettyref{eq: tmptmp7}, the dynamic regret in $[1,T]$ can bounded by (omitting lower order terms)
\begin{align}
\widetilde{\mathcal{O}}\left(\overline{D} S\sqrt{ALT}\right). \label{eq: tmp11}
\end{align}
With $N\leq 1+ 3\left(\frac{\Delta^r + \Delta^p}{S\sqrt{A}}\right)^{\frac{2}{3}} T^{\frac{1}{3}}$, \prettyref{eq: tmptmp6}, and \prettyref{eq: tmptmp8}, the regret can alternatively be upper bounded by (omitting lower order terms)
\begin{align}
\widetilde{\mathcal{O}}\left( \left(\overline{D} ^2S^2A \Delta_{[1, T]; \overline{D}} T^2\right)^{\nicefrac{1}{3}} + \overline{D} S\sqrt{AT} + \overline{D} (S^2A)^{\nicefrac{1}{3}}(\Delta^r + \Delta^p)^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}} \right). \label{eq: tmp10}
\end{align}
Then notice that $\overline{D} \leq 2D_{\max}$ and thus $\Delta_{[1, T]; \overline{D}} = \Delta^r + 2\overline{D} \Delta^p + \Delta^J = \mathcal{O}(\Delta^r + D_{\max}\Delta^p)$ where we use \prettyref{lemma: ortner lemma}. Using these in \prettyref{eq: tmp11} and \prettyref{eq: tmp10} finishes the proof.
\end{proof}
\begin{theorem}\label{thm: doubling trick algo for RL}
The doubling trick strategy described in Section~\ref{subsec:MUCRL} for the unknown $D_{\max}$ and known $L$ case has a dynamic regret bound of $\widetilde{\mathcal{O}}\left(D_{\max}S\sqrt{ALT}\right)$; for the unknown $D_{\max}$ and known $\Delta$ case, the bound is
\begin{align*}
\widetilde{\mathcal{O}}\left(D_{\max} S\sqrt{AT} + D_{\max}(S^2A)^{\nicefrac{1}{3}}(\Delta^r+ \Delta^p)^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}\right).
\end{align*}
\end{theorem}
\begin{proof}
For the known $L$ case, when $\overline{D} \leq D_{\max}$, recall that the number of epochs is forced to be $N\leq L$. Similar to the proof of \prettyref{thm: known Dmax case}, the regret in any of these epochs is upper bounded by
\begin{align*}
\widetilde{\mathcal{O}}\left(\overline{D} S\sqrt{ANT}\right) = \widetilde{\mathcal{O}}\left(\overline{D} S\sqrt{ALT}\right).
\end{align*}
Summing the above over $\overline{D} = 1,2,4,\ldots, D_{\max}$, we get $\widetilde{\mathcal{O}}\left(D_{\max} S\sqrt{ALT}\right)$.
When $\overline{D}$ first enters $[D_{\max}, 2D_{\max}]$, we use \prettyref{thm: known Dmax case} to bound the regret in the rest of the rounds, which is still of order $\widetilde{\mathcal{O}}\left(D_{\max} S\sqrt{ALT}\right)$.
For the case of known $\Delta = \Delta^r + \Delta^p$, the analysis is similar: when $\overline{D}\leq D_{\max}$, we force $N=1+3(S^{-2}A^{-1}\Delta^2 T)^{\nicefrac{1}{3}}$, and thus the regret within any of these epochs is upper bounded by (similarly to the proof of \prettyref{thm: known Dmax case})
\begin{align*}
\widetilde{\mathcal{O}}\left( \left(\overline{D} ^2S^2A \Delta_{[1, T]; \overline{D}} T^2\right)^{\nicefrac{1}{3}} + \overline{D} S\sqrt{AT} + \overline{D} (S^2A)^{\nicefrac{1}{3}}(\Delta^r + \Delta^p)^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}} \right).
\end{align*}
Summing this over $\overline{D}=1, 2, \ldots, D_{\max}$ and using $\Delta_{[1, T]; \overline{D}}=\Delta^r + 2\overline{D} \Delta^p + \Delta^J = \mathcal{O}(\Delta^r + D_{\max}\Delta^p)$ for $\overline{D}=\mathcal{O}(D_{\max})$, we get $ \widetilde{\mathcal{O}}\left( D_{\max}\left(S^2A \Delta T^2\right)^{\nicefrac{1}{3}} + D_{\max} S\sqrt{AT} \right)$. When $\overline{D}$ first enters $[D_{\max}, 2D_{\max}]$, we use \prettyref{thm: known Dmax case} to bound the regret in the rest of the rounds, which is still of the same order.
\end{proof}
\section{Bandit-over-Reinforcement-Learning Approach}\label{app: borl discuss}
The idea of the BoRL framework is to run a multi-armed bandit algorithm over a set of sub-algorithms each using a different parameter. In our case, each sub-algorithm is a {\small\textsf{\textup{MASTER-UCRL}}}\xspace with a different guess on $D_{\max}$. The set of $\overline{D}$ only needs to span the range of $[1,\sqrt{T}]$, since if $D_{\max}=\Omega(\sqrt{T})$, the regret bound would be vacuous.
We divide the horizon into $\frac{T}{B}$ equal-length intervals each of length $B=S\sqrt{AT}$. In each interval, sub-algorithm $i$ restarts a {\small\textsf{\textup{MASTER-UCRL}}}\xspace with $\overline{D}=2^{i-1}$. The reward of sub-algorithm $i$ in interval $b\in[\frac{T}{B}]$ is its total reward gained in the MDP for this interval. We denote $i^\star$ as the sub-algorithm that uses $\overline{D} \in [D_{\max}, 2D_{\max}]$.
On top of these sub-algorithms, we run the EXP3.P algorithm \citep{auer2002nonstochastic}. The ``arms'' are the sub-algorithms. From the above description, for this EXP3.P, there are $M=\lceil \log_2 \sqrt{T}\rceil$ arms, the algorithm proceeds for $\frac{T}{B}$ rounds, and in each round the reward range is $B$. By the standard regret bound of EXP3.P, the learner's regret against sub-algorithm $i^\star$ is of order
\begin{align*}
\widetilde{\mathcal{O}}\left(B\sqrt{M\frac{T}{B}} + BM\right) = \widetilde{\mathcal{O}}\left(\sqrt{BT}\right).
\end{align*}
with high probability.
On the other hand, in each interval $b\in\left[\frac{T}{B}\right]$, since sub-algorithm $i^\star$ uses a correct guess of $\overline{D}$, by \prettyref{thm: known Dmax case}, its regret against the best sequence of policy in that interval is
\begin{align*}
\widetilde{\mathcal{O}}\left( \min\left\{D_{\max}S\sqrt{AL_b B}, \ \ D_{\max}(S^2A)^{\frac{1}{3}}(\Delta_b)^{\frac{1}{3}}B^{\frac{2}{3}} + D_{\max}S\sqrt{AB}\right\} \right)
\end{align*}
where we abuse notations and denote $L_b=L_{[(b-1)B+1, bB]}$, $\Delta_b=\Delta_{[(b-1)B+1, bB]}$.
Combining the two bounds above, we get that the regret of the learner against the best sequence of policies in $[1,T]$ is
\begin{align*}
&\widetilde{\mathcal{O}}\left(\sqrt{BT} + \sum_{b=1}^{\frac{T}{B}} \min\left\{D_{\max}S\sqrt{AL_b B}, \ \ D_{\max}(S^2A)^{\frac{1}{3}}(\Delta_b)^{\frac{1}{3}}B^{\frac{2}{3}} + D_{\max}S\sqrt{AB}\right\} \right) \\
&= \widetilde{\mathcal{O}}\left(\sqrt{BT} + \min\left\{ D_{\max}S\sqrt{A\left(L + \frac{T}{B}\right)T},\ \ D_{\max}(S^2A)^{\frac{1}{3}}(\Delta)^{\frac{1}{3}}T^{\frac{2}{3}} + D_{\max}S\sqrt{AB}\times \frac{T}{B} \right\}\right) \tag{using \prettyref{eq: L partition} and \prettyref{eq: dev partition}}\\
&= \widetilde{\mathcal{O}}\left(\sqrt{BT} + D_{\max}S\sqrt{\frac{A}{B}}T + \min\left\{ D_{\max}S\sqrt{ALT},\ \ D_{\max}(S^2A)^{\frac{1}{3}}\Delta^{\frac{1}{3}}T^{\frac{2}{3}}\right\}\right).
\end{align*}
Using the $B$ that we specified above, we get
\begin{align*}
\widetilde{\mathcal{O}}\left( D_{\max}(S^2A)^{\nicefrac{1}{4}}T^{\nicefrac{3}{4}} + \min\left\{ D_{\max}S\sqrt{ALT},\ \ D_{\max}(S^2A)^{\frac{1}{3}}\Delta^{\frac{1}{3}}T^{\frac{2}{3}}\right\}\right).
\end{align*}
\section{Verifying \prettyref{assum:assump2} for Several Algorithms}
\label{app: verify example}
To prove \prettyref{eq: general condition 2}, it suffices to prove the following.
\paragraph{Assumption 1'} \textit{There exist universal constants $c_1, c_2, c_3, c_4, c_5, c_6>0$ such that for all $t=1,2,\ldots$, as long as $\Delta_{[1,t]}\leq c_1\rho(t)$, the following holds with probability $1-\frac{\delta}{T}$:
\begin{align}
&\widetilde{f}_t \geq \min_{\tau\in [1,t]} f_\tau^\star - c_2\Delta_{[1,t]} \label{eq: general condition 3} \\
&\frac{1}{t}\sum_{\tau=1}^t \left(\widetilde{f}_\tau - R_\tau\right) \leq c_3\rho(t) + c_4\Delta_{[1,t]}. \label{eq: general condition 4}
\end{align}
Furthermore, $\rho(t)\geq \frac{c_5}{\sqrt{t}}$, $\Delta(t) \geq c_6 \max_{\pi}|f_t(\pi)-f_{t+1}(\pi)|$.
}\\
This is because for an algorithm satisfying Assumption~1', we can redefine $\Delta(t)\leftarrow (c_3/c_1 + c_2 + c_4 + 1/(c_1c_5)+ 1/c_6) \Delta(t)$ and $\rho(t)\leftarrow (c_3+c_1c_2 + c_1 c_4 + 1/c_5 + c_1/c_6)\rho(t)$. Then \prettyref{eq: general condition 2} is satisfied. Our verification below is thus mostly based on Assumption~1' for simplicity.
The following proofs are brief (some of them are just sketches) since they follow standard analysis and mostly appear in previous works. Please find more details in the references. We sometimes make minor modifications to the original algorithm to make them more aligned with our framework.
\subsection{UCB1 for Multi-armed Bandits}
\begin{exampl}
\caption{UCB1 for multi-armed bandits}
\label{ex: ucb1}
\textbf{input}: $A$ (number of arms), $T, \delta$. \\
\For{$t=1, \ldots, T$}{
Choose $a_t = \argmax_{a\in[A]} \left(\widehat{r}_{t,a} + c\sqrt{\frac{\log(T/\delta)}{N_{t,a}^+}}\right)$ \myComment{$c>0$ is some universal constant} \\
where
\begin{align}
\widehat{r}_{t,a}= \frac{\sum_{\tau =1}^{t-1} R_\tau \mathbbm{1}[a_\tau=a]}{N_{t,a}^+}, \qquad N_{t,a}^+=\max\left\{1, \sum_{\tau=1}^{t-1} \mathbbm{1}[a_\tau=a]\right\}. \label{eq: hatr and N}
\end{align} \\
Receive $R_t$ with $\mathbb{E}[R_t]=r_{t,a_t}$.
}
\end{exampl}
In this subsection, we consider the multi-armed bandit problem and the UCB1 algorithm by \cite{auer2002finite}. Suppose there are $A$ arms, and let $r_{t,a}$ denote the expected reward of arm $a$ at time $t$. Then the multi-armed bandit problem fits in our framework with $\Pi=[A]$ and $f_t(a)=r_{t,a}$. Below, we show that the UCB1 algorithm satisfies Assumption 1'.
The pseudocode of UCB1 is presented in \prettyref{ex: ucb1}.
At time $t$, UCB1 chooses the arm that has the highest optimistic reward estimator $\widetilde{r}_{t,a}\triangleq \widehat{r}_{t,a} + c\sqrt{\frac{\log(T/\delta)}{ N_{t,a}^+}}$, where $\widehat{r}_{t,a}$ is the empirical mean of the reward of arm $a$ up to time $t-1$, $N_{t,a}$ is the cumulative number of pulls of arm $a$ up to time $t-1$ and $N_{t,a}^+=\max\{1, N_{t,a}\}$, all defined in \prettyref{eq: hatr and N}; $c>0$ is some universal constant that is determined by Azuma's inequality.
To see that UCB1 satisfies Assumption 1', we define
\begin{align}
\Delta(t) = \max_{a}|r_{t,a}-r_{t+1,a}|, \qquad
\widetilde{f}_t = \max_{a} \widetilde{r}_{t,a}, \qquad
\rho(t) = \sqrt{\frac{A\log(T/\delta)}{t}} + \frac{A\log(T/\delta)}{t}. \label{eq: UCB1 choices}
\end{align}
Furthermore, denote $\overline{r}_{t,a}= \frac{\sum_{\tau=1}^{t-1} r_{\tau, a} \mathbbm{1}[a_t=a]}{N_{t,a}}$ (define $\overline{r}_{t,a}=1$ if $N_{t,a}=0$ for simplicity). Note that with high probability,
\begin{align*}
\widetilde{f}_t \geq \max_{a} \overline{r}_{t,a} \geq \max_a \max_{\tau\leq t} r_{\tau,a} - \Delta_{[1,t]} \geq \min_{\tau\leq t} \max_a r_{\tau,a}- \Delta_{[1,t]},
\end{align*}
where the first inequality is because with high probability, $\widetilde{r}_{t,a}\geq \overline{r}_{t,a}$ by Azuma's inequality. This verifies \prettyref{eq: general condition 3}.
On the other hand, by the selection rule $a_t=\argmax_a \widetilde{r}_{t,a}$, we have with probability $1-\delta$,
\begin{align*}
\sum_{\tau=1}^t (\widetilde{f}_\tau - R_\tau)
&\leq \sum_{\tau=1}^t (\widetilde{r}_{\tau,a_\tau} - r_{\tau, a_\tau}) + \sum_{\tau=1}^t (r_{\tau, a_\tau}-R_\tau)\\
&= \sum_{\tau=1}^t \left(\overline{r}_{\tau,a_\tau} - r_{\tau, a_\tau} + c\sqrt{\frac{\log(T/\delta)}{N_{\tau,a}^+}}\right) + \sum_{\tau=1}^t (r_{\tau, a_\tau}-R_\tau)\\
&\leq t\Delta_{[1,t]} + \mathcal{O}\left(\sqrt{At\log(T/\delta)}+A\log(T/\delta)\right)
\end{align*}
where in the last inequality we use $\overline{r}_{\tau,a_\tau} - r_{\tau, a_\tau}\leq \Delta_{[1,t]}$ and the standard pigeonhole argument, and use Azuma's inequality to bound $\sum_{\tau=1}^t (r_{\tau, a_\tau}-R_\tau)$. This proves \prettyref{eq: general condition 4}. Note that the condition $\Delta_{[1,t]}=\mathcal{O}(\rho(t))$ in Assumption~1' is even not needed.
\subsection{OFUL for Linear Bandits}
\begin{exampl}
\caption{OFUL for linear bandits}
\label{ex: oful for linear bandit}
\textbf{input: } $\mathcal{A} \subset \mathbb{R}^d$ (action set), $T, \delta$.\\
\For{$t=1, 2, \ldots, T$}{
Choose $a_t=\argmax_{a\in\mathcal{A}}\left(a^\top \widehat{\theta}_t + 2\beta\|a\|_{\Lambda_t^{-1}}\right)$, \\
where
\begin{align}
\beta = 4\sqrt{d\log(T/\delta)},\qquad \Lambda_{t}=I + \sum_{\tau=1}^{t-1}a_\tau a_\tau^\top, \qquad \widehat{\theta}_t = \Lambda_{t}^{-1}\sum_{\tau=1}^{t-1} R_\tau a_\tau. \label{eq: oful definitions}
\end{align}\\
Receive $R_t$ with $\mathbb{E}[R_t]=a_t^\top \theta_t$.
}
\end{exampl}
In this subsection, we consider linear bandits with a fixed action set, and the OFUL algorithm by \cite{abbasi2011improved}. The original OFUL algorithm handles the case where the action set can change over time (also known as the linear contextual bandit setting), but this is beyond the main focus of this paper. Let $\mathcal{A}$ be the action set, and $\theta_t$ be the reward vector at time $t$. Then the linear bandit problem fits in our framework with $\Pi=\mathcal{A}$ and $f_t(a)=a^\top \theta_t$.
The pseudocode of OFUL (with a fixed action set) is presented in \prettyref{ex: oful for linear bandit}.
For simplicity, assume that for all actions $a\in\mathcal{A}$, $\|a\|_2\leq 1$, and for all $t$, the reward vector $\theta_t$ satisfies $\|\theta_t\|_2\leq 1$. The OFUL algorithm chooses the action $a_t=\argmax_{a} a^\top \widehat{\theta}_t + 2\beta\|a\|_{\Lambda_t^{-1}}$ at time $t$,
where $\beta$, $\Lambda_{t}=I + \sum_{\tau=1}^{t-1}a_\tau a_\tau^\top$, and $\widehat{\theta}_t$ are defined in \prettyref{eq: oful definitions}.
Then we define
\begin{align}
\Delta(t) =d\sqrt{\log (T/\delta)}\|\theta_t-\theta_{t+1}\|_2, \;\;
\widetilde{f}_t = \max_{a\in\mathcal{A}} \left(a^\top \widehat{\theta}_t + 2\beta \|a\|_{\Lambda_{t}^{-1}}\right), \;\;
\rho(t) = \beta\sqrt{\frac{d\log (T/\delta)}{t}}. \label{eq: oful choices}
\end{align}
Below, we verify that OFUL satisfies Assumption 1' with the choices in \prettyref{eq: oful choices}. Under the assumption that $\Delta_{[1,t]}\leq \rho(t)$, for any action $a$, by similar arguments as in \cite[Lemma 1]{pmlr-v108-zhao20a},
\begin{align}
\left|a^\top (\theta_t - \widehat{\theta}_t)\right|
&\leq \left|a^\top \Lambda_{t}^{-1}\sum_{\tau=1}^{t-1}a_\tau a_\tau^\top (\theta_s - \theta_t)\right| + \beta \|a\|_{\Lambda_{t}^{-1}} \nonumber\\
&\leq \sum_{\tau=1}^{t-1}\left| a^\top \Lambda_{t}^{-1}a_\tau \right| \left| a_\tau^\top (\theta_\tau - \theta_t) \right| + \beta \|a\|_{\Lambda_{t}^{-1}} \nonumber\\
&\leq \frac{\Delta_{[1,t]}}{d\sqrt{\log (T/\delta)}}\times \left( \sum_{\tau=1}^{t-1}\|a\|_{\Lambda_{t}^{-1}}\|a_\tau\|_{\Lambda_{t}^{-1}}\right) + \beta \|a\|_{\Lambda_{t}^{-1}} \tag{ $a_\tau^\top (\theta_\tau-\theta_t) \leq \|\theta_\tau-\theta_t\|_2\leq \frac{\Delta_{[1,t]}}{d\sqrt{\log (T/\delta)}}$} \nonumber\\
&\leq \frac{\Delta_{[1,t]}}{d\sqrt{\log (T/\delta)}} \times \|a\|_{\Lambda_{t}^{-1}} \times \sqrt{(t-1)\sum_{\tau=1}^{t-1} \|a_\tau\|_{\Lambda_{t}^{-1}}^2} + \beta \|a\|_{\Lambda_{t}^{-1}} \tag{Cauchy-Schwarz} \nonumber\\
&\leq \left(\beta + \Delta_{[1,t]}\sqrt{\frac{t}{d\log (T/\delta)}}\right) \|a\|_{\Lambda_{t}^{-1}} \tag{$\sum_{\tau=1}^{t-1}\|a_\tau\|_{V_{t-1}^{-1}}^2=\text{tr}(\Lambda_{t}^{-1}\sum_{\tau=1}^{t-1}a_\tau a_\tau^\top)\leq d$} \nonumber\\
&\leq 2\beta \|a\|_{\Lambda_{t}^{-1}}. \tag{by the assumption $\Delta_{[1,t]}\leq \rho(t)$} \\
\label{eq: error bound 1}
\end{align}
Thus,
\begin{align*}
\sum_{\tau=1}^t \left(\widetilde{f}_\tau - R_\tau\right)
&=\sum_{\tau=1}^t \left(\widetilde{f}_\tau - a_\tau^\top \theta_\tau\right) + \sum_{\tau=1}^t \left(a_\tau^\top \theta_\tau - R_\tau\right) \\
&= \sum_{\tau=1}^t a_\tau^\top \left(\widehat{\theta}_\tau - \theta_\tau\right) + 2 \sum_{\tau=1}^t \beta\|a_\tau\|_{\Lambda^{-1}_{\tau}} + \mathcal{O}\left(\sqrt{t\log(T/\delta)}\right) \tag{by the definition of $\widetilde{f}_\tau$ and that OFUL chooses $a_\tau = \argmax_{a}\left(a^\top \widehat{\theta}_\tau + 2 \beta\|a\|_{\Lambda_{\tau}^{-1}}\right)$}\\
& = \mathcal{O}\left(\sum_{s=1}^t \beta\|a_s\|_{\Lambda^{-1}_{s}}\right) + \mathcal{O}\left(\sqrt{t\log(T/\delta)}\right) \tag{by \prettyref{eq: error bound 1}}\\
& = \mathcal{O}\left(\beta\sqrt{dt\log t}\right) = \mathcal{O}\left(t\rho(t)\right)\leq \mathcal{O}\left(t\rho(t) + t\Delta_{[1,t]}\right).
\end{align*}
This verifies \prettyref{eq: general condition 4}. Also, by \prettyref{eq: error bound 1},
\begin{align*}
\widetilde{f}_t
&= \max_{a}\left(a^\top \widehat{\theta}_t + 2 \beta\|a\|_{\Lambda_{t}^{-1}}\right) \geq \max_a a^\top \theta_t = f_t^\star \geq \min_{\tau\in[1,t]}f_\tau^\star - \Delta_{[1,t]}.
\end{align*}
This verifies \prettyref{eq: general condition 3}.
\subsection{GLM-UCB for Generalized Linear Bandits}
\begin{exampl}
\caption{GLM-UCB for generalized linear bandits}
\label{ex: glm-ucb}
\textbf{input}: $\mathcal{A}\subset \mathbb{R}^d, T, \delta, \mu$ (link function), $\lambda$. \\
\textbf{define}: $k_\mu = \sup_{x\in[0,1]} \frac{\mathrm{d}\mu(x)}{\mathrm{d}x}$,\quad $c_\mu = \inf_{x\in[0,1]} \frac{\mathrm{d}\mu(x)}{\mathrm{d}x}>0$. \\
\For{$t=1, \ldots, T$}{
Choose $a_t = \argmax_{a\in\mathcal{A}}\left( \mu(a^\top \widehat{\theta}_t) + 2\beta\|a\|_{\Lambda_t^{-1}} \right)$
\\
where
\begin{align*}
\beta = \frac{4k_\mu}{c_\mu}\left(\sqrt{d\log(c_\mu T/(\lambda\delta))} + c_\mu\sqrt{ \lambda}\right), \qquad \Lambda_t = \lambda I + \sum_{\tau=1}^{t-1}a_\tau a_\tau^\top,
\end{align*}\\
and $\widehat{\theta}_t$ is the unique solution of the following set of equations (define $g_t(x)\triangleq \lambda c_\mu x + \sum_{\tau=1}^{t-1} \mu(a_\tau^\top x)a_\tau$):
\begin{align*}
g_t(\theta_t') = \sum_{\tau=1}^{t-1}R_\tau a_\tau, \qquad
\widehat{\theta}_t = \argmin_{\theta: \|\theta\|_2\leq 1} \norm{g_t(\theta_t') - g_t(\theta)} _{\Lambda_t^{-1}}.
\end{align*}
\\
Receive $R_t$ with $\mathbb{E}[R_t] = \mu(a_t^\top \theta_t)$.
}
\end{exampl}
Generalized linear bandit is proposed by \cite{filippi2010parametric} and extended to the non-stationary case by \cite{cheung2019learning, pmlr-v108-zhao20a, russac2020algorithms, faury2021regret}. We refer the readers to these papers for the introduction of the setting. Again, we consider the special case where the action set is fixed over time, and for simplicity, we assume that the action set $\mathcal{A}$ is a subset of $\{a\in\mathbb{R}^d:~\|a\|_2\leq 1\}$ and the hidden parameter $\theta_t$ satisfies $\|\theta_t\|_2\leq 1$. The generalized linear bandit problem is accompanied with an increasing link function $\mu: \mathbb{R}\rightarrow \mathbb{R}$. It fits in our framework with $\Pi=\mathcal{A}$ and $f_t(a) = \mu(a^\top \theta_t)$.
The standard GLM-UCB is presented in
\prettyref{ex: glm-ucb}.
Below we show that GLM-UCB satisfies Assumption 1' with the following definitions:
\begin{align*}
\Delta(t) = \frac{k_\mu^2d}{c_\mu} \sqrt{\log (T/\delta)}\|\theta_t - \theta_{t+1}\|_2, \;\; \widetilde{f}_t = \max_{a\in\mathcal{A}} \left(\mu(a^\top \widehat{\theta}_t) + 2\beta \|a\|_{\Lambda_{t}^{-1}}\right), \;\;
\rho(t) = \beta\sqrt{\frac{d\log (T/\delta)}{t}},
\end{align*}
where $c_\mu$ and $k_\mu$ are the infimum and supremum of the derivative of $\mu$ (defined in \prettyref{ex: glm-ucb}).
Define $G_t\triangleq \sum_{\tau=1}^{t-1}\left[\int_{v=0}^1 \dot{\mu}\left(\inner{a_\tau, (1-v)\widehat{\theta}_t + v\theta_t} \right) \mathrm{d}v\right]a_\tau a_\tau^\top + \lambda c_\tau I \succeq c_\mu \Lambda_t$.
Under the assumption that $\Delta(t)\leq \rho(t)$, for all $a\in\mathcal{A}$,
\begin{align}
\left|\mu(a^\top \theta_t) - \mu(a^\top \widehat{\theta}_t) \right|
&\leq k_u \left|a^\top (\theta_t - \widehat{\theta}_t)\right|
\leq k_\mu \left| a^\top G_t^{-1} (g_t(\theta_t) - g_t(\widehat{\theta}_t)) \right| \nonumber \\
&\leq k_\mu \|a\|_{G_t^{-1}} \norm{g_t(\theta_t) - g_t(\widehat{\theta}_t)}_{G_t^{-1}} \nonumber \\
&\leq \frac{k_\mu}{c_\mu} \|a\|_{\Lambda_t^{-1}} \norm{g_t(\theta_t) - g_t(\widehat{\theta}_t)}_{\Lambda_t^{-1}} \nonumber \\
&\leq \frac{k_\mu}{c_\mu} \|a\|_{\Lambda_t^{-1}} \norm{g_t(\theta_t) - g_t(\theta_t')}_{\Lambda_t^{-1}} \nonumber \\
&= \frac{k_\mu}{c_\mu} \|a\|_{\Lambda_t^{-1}} \norm{ \sum_{\tau=1}^{t-1} \left(\mu(a_\tau^\top \theta_t) - \mu(a_\tau^\top \theta_\tau)\right)a_\tau + \sum_{\tau=1}^{t-1} \left(\mu(a_\tau^\top \theta_\tau) - R_\tau)\right)a_\tau + \lambda c_\mu \theta_t }_{\Lambda_t^{-1}} \nonumber \\
&\leq \frac{k_\mu}{c_\mu} \|a\|_{\Lambda_t^{-1}} \left(k_\mu\max_{\tau\leq t}\norm{\theta_t - \theta_\tau}_2 \sum_{\tau=1}^{t-1} \norm{ a_\tau }_{\Lambda_t^{-1}} + \norm{\sum_{\tau=1}^{t-1}\eta_\tau a_\tau}_{\Lambda_t^{-1}} + \sqrt{\lambda}c_\mu\right) \tag{define $\eta_\tau=\mu(a_\tau^\top \theta_\tau) - R_\tau$}\\
&\leq \frac{k_\mu}{c_\mu} \|a\|_{\Lambda_t^{-1}} \left(k_\mu\frac{c_\mu\Delta_{[1,t]}}{k_\mu^2d\sqrt{\log(T/\delta)}} \sqrt{dt} + \sqrt{d\log(c_\mu T/\delta)} + \sqrt{\lambda}c_\mu\right) \nonumber \\
&\leq \frac{k_\mu}{c_\mu} \|a\|_{\Lambda_t^{-1}} \left(\frac{c_\mu\rho(t)}{k_\mu} \sqrt{\frac{t}{d\log(T/\delta)}} + \sqrt{d\log(c_\mu T/\delta)} + \sqrt{\lambda}c_\mu\right) \tag{by the assumption $\Delta_{[1,t]}\leq \rho(t)$} \\
&\leq 2 \beta \|a\|_{\Lambda_t^{-1}}. \label{eq: glm optimistic}
\end{align}
Thus,
\begin{align*}
\sum_{\tau=1}^t \left(\widetilde{f}_\tau - R_\tau\right)
&= \sum_{\tau=1}^t \left(\widetilde{f}_\tau - \mu(a_\tau^\top \theta_\tau)\right) + \sum_{\tau=1}^t \left(\mu(a_\tau^\top \theta_\tau) - R_\tau\right) \\
&\leq \sum_{\tau=1}^t \left(\mu(a_\tau^\top \widehat{\theta}_\tau) - \mu(a_\tau^\top \theta_\tau)\right) + 2\beta \sum_{\tau=1}^t \|a_\tau\|_{\Lambda_t^{-1}} + \mathcal{O}\left(\sqrt{t\log(T/\delta)}\right) \\
&\leq \mathcal{O}\left(\beta \sum_{\tau=1}^t \|a_\tau\|_{\Lambda_t^{-1}}\right) = \mathcal{O}\left(\beta \sqrt{dt\log(T/\delta)}\right)\\
&= \mathcal{O}\left(t\rho(t)\right) = \mathcal{O}\left(t\rho(t) + t\Delta_{[1,t]}\right).
\end{align*}
This verifies \prettyref{eq: general condition 4}. Furthermore, by \prettyref{eq: glm optimistic},
\begin{align*}
\widetilde{f}_t = \max_{a\in\mathcal{A}} \left(\mu(a^\top \widehat{\theta}_t) + 2\beta \|a\|_{\Lambda_{t}^{-1}}\right) \geq \max_{a\in\mathcal{A}} \mu(a^\top \theta_t) = f_t^\star \geq \min_{\tau\in[1,t]}f_\tau^\star - \Delta_{[1,t]}.
\end{align*}
This verifies \prettyref{eq: general condition 3}.
\subsection{Q-UCB for Finite-horizon Tabular MDPs}
\begin{exampl}
\caption{Q-UCB for finite-horizon tabular MDPs}
\label{ex: qucb}
\textbf{input}: $S$ (number of states), $A$ (number of actions), $H$, $T, \delta$. \\
$Q_h(s,a)\leftarrow H, \quad N_h(s,a)\leftarrow 0$ for all $h, s, a$. \\
\For{$t=1, \ldots, T$}{
\For{$h=1, \ldots, H$}{
Choose $a^t_h \leftarrow \argmax_a Q_h(s^t_h, a)$. \\
$\tau = N_h(s^t_h, a^t_h)\leftarrow N_h(s^t_h, a^t_h) +1$, \qquad $b_\tau\leftarrow c\sqrt{H^3\log(SAT/\delta)/\tau}$. \myComment{$c$ is a universal constant}\\
$Q_h(s^t_h, a^t_h)\leftarrow (1-\alpha_\tau)Q_h(s^t_h, a^t_h) + \alpha_\tau \left[r_h^t(s_h^t, a_h^t) + V_{h+1}(s_{h+1}^t) + b_\tau\right]$. \myComment{$\alpha_\tau\triangleq \frac{H+1}{H+\tau}$} \\
$V_h(s^t_h)\leftarrow \min\left\{H, \max_a Q_h(s^t_h, a)\right\}$.
}
}
\end{exampl}
The finite-horizon tabular MDP problem fits in our framework with $\Pi$ being the set of deterministic polices on the MDP, and $f_t(\pi)$ being the expected reward of policy $\pi$ in episode $t$. Q-UCB (Hoeffding-style) is a model-free algorithm for finite-horizon tabular MDPs proposed by \cite{jin2018q}, whose pseudocode is in \prettyref{ex: qucb}. Let $H$ denote the horizon length, $s^t_h, a^t_h$ denote the state and actions visited at step $h$ of episode $t$, and $r^t_h$, $p^t_h$ denote the reward and transition functions at step $h$ of episode $t$. Without loss of generality, we assume that $s^t_1=s_1$ for all $t$ (i.e., the initial state is fixed).
It has been shown in the proof of \citep[Theorem 1]{mao2020nearoptimal} that Q-UCB satisfy Assumption 1' with the following choices:
\begin{align*}
\Delta(t)&=H\sum_{h=1}^H \max_{s,a}|r_h^{t}(s,a)-r_{h}^{t+1}(s,a)| + H^2\sum_{h=1}^H \max_{s,a}\|p_h^{t}(\cdot|s,a)-p_{h}^{t+1}(\cdot|s,a)\|_1, \\
\widetilde{f}_t &= V^t_h(s_1), \tag{$V^t_h$ is the $V_h$ in \prettyref{ex: qucb} at the beginning of episode $t$} \\
\rho(t)&=\widetilde{\mathcal{O}}\left(\sqrt{\frac{H^5SA}{t}} + \frac{H^3SA}{t}\right).
\end{align*}
The proof details are omitted here.
\subsection{LSVI-UCB for Finite-horizon Linear MDPs}
\begin{exampl}
\caption{LSVI-UCB for finite-horizon linear MDP}
\label{ex: linear mdp ucb}
\textbf{input}: $\mathcal{S}$ (state space), $\mathcal{A}$ (action space), $\phi(\cdot,\cdot): \mathcal{S}\times \mathcal{A}\rightarrow \mathbb{R}^d, H, T, \delta$. \\
\For{$t=1, \ldots, T$}{
\For{$h=H, \ldots, 1$}{
$\Lambda_h \leftarrow \sum_{\tau=1}^{t-1}\phi(s^\tau_h, a^\tau_h)\phi(s^\tau_h, a^\tau_h)^\top + I$. \\
$w_h \leftarrow \Lambda_h^{-1} \sum_{\tau=1}^{t-1}\phi(s^\tau_h, a^\tau_h)\left[r_h^\tau(x_h^\tau, a_h^\tau)+ \max_{a\in\mathcal{A}} Q_{h+1}(x^\tau_{h+1},a)\right]$. \\
$Q_h(\cdot, \cdot)\leftarrow \min\left\{w_h^\top \phi(\cdot,\cdot) + 2\beta \left(\phi(\cdot,\cdot)\Lambda_h^{-1}\phi(\cdot,\cdot)\right)^{1/2}, H\right\}$. \\
\ \myComment{$\beta = cdH\sqrt{\log(T/\delta)}$ for some universal constant $c$} \label{line: linear mdp time t}
}
\For{$h=1, \ldots, H$}{
Take action $a^t_h\leftarrow \argmax_{a\in\mathcal{A}} Q_h(s^t_h, a)$.
}
}
\end{exampl}
See \citep{zhou2020nonstationary, touati2020efficient} for the non-stationary finite-horizon linear MDP setting. We assume that the reward function and the transition function at step $h$ of episode $t$ are $r^t_h(s,a)=\phi(s,a)^\top \theta^t_h$ and $p^t_h(s'|s,a)=\phi(s,a)^\top \mu^t_h(s')$ where $\phi(\cdot,\cdot)$ is the feature function that maps a state-action pair to a $d$-dimensional feature vector.
The problem fits in our framework with $\Pi$ being the set of deterministic policies, and $f_t(\pi)$ being the expected reward of policy $\pi$ in episode $t$. The LSVI-UCB algorithm is an optimism-based algorithm proposed by \cite{jin2020provably}, whose pseudocode is shown in \prettyref{ex: linear mdp ucb}. We define $Q^t_h, w^t_h, \Lambda^t_h$ to be the $Q_h, w_h, \Lambda_h$ at \prettyref{line: linear mdp time t} of round $t$. Furthermore, define $V^t_h(s) = \max_{a\in\mathcal{A}} Q^t_h(s,a)$. Again, without loss of generality, we assume $s^t_1=s_1$ (the initial state is fixed).
We define
\begin{align*}
\Delta(t) &= dH\sqrt{\log(T/\delta)}\left(\sum_{h=1}^H \|\theta^t_{h}-\theta^{t+1}_{h}\|_2 + H\sum_{h=1}^H \|\mu^t_{h}-\mu^{t+1}_{h}\|_{F}\right), \\
\widetilde{f}_t &= V^t_1(s_1), \\
\rho(t) &= c\sqrt{\frac{d^3 H^4}{t}}\log (T/\delta) = \beta H\sqrt{\frac{d\log(T/\delta)}{t}}. \tag{$c$ and $\beta$ defined in \prettyref{ex: linear mdp ucb}}
\end{align*}
Below, we verify that LSVI-UCB satisfies Assumption 1' with the $\Delta$, $\widetilde{f}$, and $\rho$ defined above. Assume that $\Delta_{[1,t]}\leq \rho(t)$. By similar arguments as in the proof of \citep[Lemma 3]{zhou2020nonstationary}, we have
\begin{align}
&\left| \phi(s,a)^\top w^t_h - Q_{h}^\star(s,a) - \mathbb{P}^t_h(V^t_h - V^\star_h)(s,a)\right|
\leq \left(\beta + \sqrt{dt} B_{\theta, [1,t]} + H\sqrt{dt} B_{\mu,[1,t]} \right) \left\|\phi(s,a)\right\|_{\left(\Lambda^t_h\right)^{-1}} \label{eq: error linear MDP bound}
\end{align}
where $B_{\theta,[1,t]} = \sum_{\tau=1}^{t-1}\sum_{h=1}^H \|\theta^\tau_{h}-\theta^{\tau+1}_{h}\|_2$ and $B_{\mu, [1,t]} = \sum_{\tau=1}^{t-1}\sum_{h=1}^H \|\mu^\tau_{h}-\mu^{\tau+1}_{h}\|_{F}$. By the definition of $\Delta(t)$, the right-hand side of \prettyref{eq: error linear MDP bound} can be further upper bound by
\begin{align}
\left(\beta + \frac{1}{H}\sqrt{\frac{t}{d\log(T/\delta)}}\Delta_{[1,t]}\right)\left\|\phi(s,a)\right\|_{\left(\Lambda^t_h\right)^{-1}} \leq 2\beta \left\|\phi(s,a)\right\|_{\left(\Lambda^t_h\right)^{-1}}, \label{eq: final bound linear MDP}
\end{align}
where the inequality is by the assumption that $\Delta_{[1,t]}\leq \rho(t)$. Similar to the proof of \citep[Lemma 4]{zhou2020nonstationary}, we can then show that for any $t, h$,
\begin{align*}
Q_h^t(s,a) - Q_h^{\star}(s,a)
&= \phi(s,a)^\top w^t_h - Q_h^\star(s,a) + 2\beta\|\phi(s, a)\|_{\left(\Lambda^t_h\right)^{-1}} \\
&\geq \max_{s'} \left(V^t_{h+1}(s') - V^\star_{h+1}(s')\right) \tag{by \prettyref{eq: error linear MDP bound} and \prettyref{eq: final bound linear MDP}}
\end{align*}
and further using induction to show that $V^t_1(s)\geq V^\star_1(s)$. Thus, $\widetilde{f}_t=V_1^t(s_1)\geq V_1^\star(s_1)$, which verifies \prettyref{eq: general condition 3}. One can also show that $\sum_{\tau=1}^t \left(\widetilde{f}_\tau - R_\tau\right) = \mathcal{O}(t\rho(t))$ using the standard analysis of LSVI-UCB (e.g., \citep[Theorem 3.1]{jin2020provably}, \citep[Theorem 5]{zhou2020nonstationary}). This verifies \prettyref{eq: general condition 4}.
\subsection{ILOVETOCONBANDITS for Contextual Bandits}
\begin{exampl}[H]
\caption{ILOVETOCONBANDITS for contextual bandits}
\label{ex: ilove}
\textbf{input}: $\Pi$ (policy set), $\mathcal{A}$ (action set), $T$, $\delta$. \\
\For{$t=1, \ldots, T$}{
Calculate $Q_t\in\Delta_{\Pi}$ that satisfies the following constraints with some universal constant $c'>0$:
\begin{align*}
\sum_{\pi}Q(\pi)\widehat{\text{\rm Reg}}_{[1,t-1]}(\pi) &\leq 2c'A\mu_t \\
\forall \pi\in\Pi, \qquad \frac{1}{t-1}\sum_{\tau=1}^{t-1} \frac{1}{Q^{\mu_t}(\pi(x_\tau)|x_{\tau})} &\leq 2A + \frac{\widehat{\text{\rm Reg}}_{[1,t-1]}(\pi)}{c'\mu_t}
\end{align*}
where $\mu_t\triangleq \sqrt{\frac{\log(|\Pi|T/\delta)}{At}}$, $Q^{\mu}(a|x)\triangleq (1-A\mu)\sum_{\pi\in\Pi}Q(\pi)\mathbbm{1}[\pi(x)=a] + \mu$, and
\begin{align*}
\widehat{\text{\rm Reg}}_{\mathcal{I}}(\pi) \triangleq \frac{1}{|\mathcal{I}|}\max_{\pi'}\sum_{\tau\in\mathcal{I}} \left(\widehat{r}_\tau(\pi'(x_\tau)) - \widehat{r}_\tau(\pi(x_\tau))\right), \qquad \widehat{r}_\tau(a) \triangleq \frac{R_\tau\mathbbm{1}[a_\tau=a]}{p_{\tau}(a)}.
\end{align*}
Let $p_t(\cdot) = Q^{\mu_t}(\cdot|x_t)$ and sample $a_t\sim p_t$.
}
\end{exampl}
In the contextual bandit problem, in each round, the learner first sees a context $x_t\in\mathcal{X}$, and then chooses an action $a_t\in[A]$ based on it. The learner then receives the reward $r_t(a_t)\in\mathbb{R}$. We assume that $(x_t, r_t)$ is sampled from the distribution $\mathcal{D}_t$. The goal of the learner is to be comparable to the best mapping $\pi: \mathcal{X}\rightarrow [A]$ within a given set of mappings $\Pi$ (which are called \emph{policies}), i.e., the learner wants to minimize $\sum_t(r_t(\pi_t^*(x_t)) - r_t(a_t)$ where $\pi_t^*\triangleq \max_{\pi'\in\Pi} \mathbb{E}_{(x,r)\sim\mathcal{D}_t}[r(\pi'(x))]$. See \citep{agarwal2014taming} for more detailed description of the problem. This problem fits in our framework with the same $\Pi$ and $f_t(\pi) = \mathbb{E}_{(x,r)\sim \mathcal{D}_t}[r(\pi(x))]$.
The algorithm ILOVETOCONBANDITS (\prettyref{ex: ilove}) by \cite{agarwal2014taming} achieves the optimal regret bound in the i.i.d. case.
The analysis for ILOVETOCONBANDITS is more involved. Fortunately, \cite{chen2019new} already has helpful lemmas for ILOVETOCONBANDITS in the non-stationary case, and we can simply reuse them.
We show a more general result that \prettyref{assum:assump2} is satisfied no matter how large $\Delta_{[1,t]}$ is.
Let $\mathcal{R}_{\mathcal{I}}(\pi)=\frac{1}{|\mathcal{I}|}\sum_{\tau\in\mathcal{I}}\mathbb{E}_{(x,r)\sim \mathcal{D}_\tau}\left[r(\pi(x))\right]$ be the expected of policy $\pi$ in the interval $\mathcal{I}$, $\widehat{\calR}_{\mathcal{I}}(\pi)= \frac{1}{|\mathcal{I}|}\sum_{\tau\in\mathcal{I}} \widehat{r}_\tau(\pi(x_\tau))$ be an unbiased estimator of $\mathcal{R}_{\mathcal{I}}(\pi)$, with $\widehat{r}_\tau$ an unbiased estimator for the action reward constructed with inverse propensity weighting at time $\tau$. Let $\text{Reg}\xspace_{\mathcal{I}}(\pi)=\max_{\pi'}\mathcal{R}_{\mathcal{I}}(\pi') - \mathcal{R}_{\mathcal{I}}(\pi)$ and $\widehat{\text{\rm Reg}}_{\mathcal{I}}(\pi)=\max_{\pi'}\widehat{\calR}_{\mathcal{I}}(\pi') - \widehat{\calR}_{\mathcal{I}}(\pi)$. Below, we will show that ILOVETOCONBANDITS satisfies Assumption 1' with the following definitions:
\begin{align*}
\Delta(t)&\triangleq \|\mathcal{D}_t-\mathcal{D}_{t+1}\|_{\text{TV}}=\int_{r}\int_x |\mathcal{D}_t(x,r) - \mathcal{D}_{t+1}(x,r)| \mathrm{d}x\mathrm{d}r, \\
\widetilde{f}_t &\triangleq \max_{\pi}\widehat{\calR}_{[1,t-1]}(\pi) + c_2 A\mu_{t-1} \tag{for some universal constant $c_2>0$} \\
\rho(t)&\triangleq \sqrt{\frac{A\log(|\Pi|T/\delta)}{t}}.
\end{align*}
Note that $\Delta(t)$ upper bounds $|\mathbb{E}_{(x,r)\sim \mathcal{D}_{t}}[r(\pi(x))] - \mathbb{E}_{(x,r)\sim \mathcal{D}_{t+1}}[r(\pi(x))]|$.
Combining the proofs of Lemma 14 and Lemma 16 in \citep{chen2019new}, we get the following guarantee with probability at least $1-\delta$ for any policy $\pi$:
\begin{align}
\left|\widehat{\calR}_{[1,t]}(\pi) - \mathcal{R}_{[1,t]}(\pi)\right| \leq c_1 \widehat{\text{\rm Reg}}_{[1,t]}(\pi) + c_2A\mu_t + c_3 \Delta_{[1,t]}, \label{eq: CB R deviaion}
\end{align}
where $\mu_t=\sqrt{\frac{\log(|\Pi|T/\delta)}{At}}$ and $c_1, c_2, c_3$ are universal constants. To see how to get \prettyref{eq: CB R deviaion}, notice that Lemma 14 of \citep{chen2019new} gives $\left|\widehat{\calR}_{[1,t]}(\pi)-\mathcal{R}_{[1,t]}(\pi)\right| \leq \mathcal{O}\left(\frac{\mu_t}{t}\sum_{\tau=1}^t U_\tau + \frac{\log(|\Pi| T/\delta)}{t\mu_t}\right)$, and they further upper bound $U_\tau$ by $\mathcal{O}\left( \frac{\text{Reg}\xspace_{[1,t]}}{\mu_t} + A + \frac{\Delta_{[1,t]}}{\mu_t} \right)$ in the second-to-last line in their proof of Lemma 16. Combining them yields \prettyref{eq: CB R deviaion}. Notice that they have an additional $\log T$ factor which we do not suffer.
Below, let $\overline{\pi}_t=\argmax_{\pi}\mathcal{R}_{[1,t]}(\pi)$. Then we have
\begin{align}
\max_{\pi}\widehat{\calR}_{[1,t]}(\pi)
&\geq \widehat{\calR}_{[1,t]}(\overline{\pi}_t)
\geq \mathcal{R}_{[1,t]}(\overline{\pi}_t) - c_3\Delta_{[1,t]} - c_2 A\mu_t \nonumber
\\
&=\max_{\pi}\mathcal{R}_{[1,t]}(\pi) - c_3\Delta_{[1,t]} - c_2 A\mu_t, \label{eq: CB6}
\end{align}
where in the second inequality we use \prettyref{eq: CB R deviaion} with the fact that $\text{Reg}\xspace_{[1,t]}(\overline{\pi}_t)=0$.
Therefore, if we choose $\widetilde{f}_t = \max_{\pi}\widehat{\calR}_{[1,t-1]}(\pi) + c_2 A\mu_{t-1}$, then
\begin{align}
\widetilde{f}_t
&\geq \max_\pi \mathcal{R}_{[1,t-1]}(\pi) - c_3\Delta_{[1,t-1]} \tag{using \prettyref{eq: CB6} and the definition of $\widetilde{f}_t$} \\
&\geq \max_{\pi}\max_{\tau\in[1,t]} \mathcal{R}_\tau(\pi) - (c_3+1)\Delta_{[1,t]} \label{eq: ILOVE first condition}
\end{align}
which proves \prettyref{eq: general condition 3}.
Next, we show \prettyref{eq: general condition 4}:
\begin{align}
\widetilde{f}_t - \mathbb{E}_t[R_t]
&\leq \sum_\pi Q_t(\pi)\left(\widetilde{f}_t - \mathcal{R}_t(\pi)\right) + \mathcal{O}(A\mu_t) \tag{by the algorithm, which uses $\mathcal{O}(A\mu_t)$ probability to explore actions}\\
&= \sum_\pi Q_t(\pi)\left(\max_{\pi'}\widehat{\calR}_{[1,t-1]}(\pi') - \mathcal{R}_t(\pi)\right) + \mathcal{O}(A\mu_t) \nonumber \\
&\leq \sum_\pi Q_t(\pi)\left(\max_{\pi'}\widehat{\calR}_{[1,t-1]}(\pi') - \mathcal{R}_{[1,t-1]}(\pi)\right) + \mathcal{O}\left(A\mu_t+ \Delta_{[1,t]} \right) \nonumber \\
&= \sum_\pi Q_t(\pi)\left( \widehat{\text{\rm Reg}}_{[1,t-1]}(\pi) + \widehat{\calR}_{[1,t-1]}(\pi) - \mathcal{R}_{[1,t-1]}(\pi)\right) + \mathcal{O}\left(A\mu_t+ \Delta_{[1,t]} \right) \nonumber \\
&\leq \sum_\pi Q_t(\pi)\left( \widehat{\text{\rm Reg}}_{[1,t-1]}(\pi) + c_1\text{Reg}\xspace_{[1,t-1]}(\pi)\right) + \mathcal{O}\left(A\mu_t+\Delta_{[1,t]}\right) \tag{by \prettyref{eq: CB R deviaion}} \nonumber \\
&\leq (1+2c_1)\sum_{\pi}Q_t(\pi)\widehat{\text{\rm Reg}}_{[1,t-1]}(\pi) + \mathcal{O}\left(A\mu_t + \Delta_{[1,t]}\right) \label{eq: ILOVE condition 2}
\end{align}
where the last inequality is by Lemma 16 of \citep{chen2019new}, which bounds $\text{Reg}\xspace_{[1,t-1]}(\pi)$ by $2\widehat{\text{\rm Reg}}_{[1,t-1]}(\pi) + \mathcal{O}\left(A\mu_t + \Delta_{[1,t]}\right)$.
By the algorithm, $\sum_{\pi}Q_t(\pi)\widehat{\text{\rm Reg}}_{[1,t-1]}(\pi)$ is of order $\mathcal{O}\left(A\mu_t\right)$. Therefore, the last expression can further be upper bounded by $\mathcal{O}\left(A\mu_t + \Delta_{[1,t]}\right)$.
Finally, with the above calculation and Azuma's inequality, we get
\begin{align*}
\sum_{\tau=1}^t \left(\widetilde{f}_\tau - R_\tau\right)\leq \sum_{\tau=1}^t \left(\widetilde{f}_\tau - \mathbb{E}_\tau[R_\tau]\right) + \sum_{\tau=1}^t \left(\mathbb{E}_\tau[R_\tau] - R_\tau\right) \leq \mathcal{O}\left(\sqrt{At\log(|\Pi|T/\delta)} + \Delta_{[1,t]}\right)
\end{align*}
Since we choose $\rho(t)=\sqrt{\frac{A\log(|\Pi|T/\delta)}{t}}$, \prettyref{eq: general condition 4} is also satisfied.
\subsection{FALCON for Contextual Bandits}
\begin{exampl}
\caption{FALCON for realizable contextual bandits}
\label{ex: falcon}
\textbf{input}: $\Phi$ (reward function class), $\mathcal{A}$ (action sets), $T, \delta$. \\
\For{$t=1, \ldots, T$}{
Let $\gamma_t = \sqrt{At/\log(|\Phi|T/\delta)}$. \\
Compute $\widehat{\phi}_t = \argmin_{\phi\in\Phi} \sum_{\tau=1}^{t-1}(\phi(x_\tau, a_\tau) - r_t(a_t))^2$ \\
Observe context $x_t$. \\
Let $\widehat{a}_t = \argmax_{a\in\mathcal{A}} \widehat{\phi}(x_t,a)$. Define
\begin{align*}
p_t(a)\triangleq
\begin{cases}
\frac{1}{A+\gamma_t \left(\widehat{\phi}_t(x_t,\widehat{a}_t) - \widehat{\phi}_t(x_t, a)\right)}, \qquad &\text{for\ } a\neq \widehat{a}_t, \\
1- \sum_{a'\neq \widehat{a}_t} p_t(a'), &\text{for\ } a=\widehat{a}_t.
\end{cases}
\end{align*}
Sample $a_t\sim p_t$ and observe reward $R_t$.
}
\end{exampl}
FALCON is an algorithm for stationary contextual bandits. It relies on the assumption that the expected reward of action $a$ under context $x$ is given by an unknown function $\phi^\star(x,a): \mathcal{X}\times \mathcal{A}\rightarrow [0,1]$. The learner is given the function class $\Phi$ that contains $\phi^\star$. For each $\phi\in\Phi$, one can derive a policy $\pi_\phi: \mathcal{X}\rightarrow \mathcal{A}$ such that $\pi_\phi(x)=\argmax_{a\in\mathcal{A}} \phi(x,a)$. It is straightforward to see that the optimal policy is $\pi_{\phi^\star}$, and the learner's goal is to be competitive with it. This problem falls into our framework with $\Pi = \{\pi_\phi: \phi\in\Phi\}$ and $f_t(\pi) = \mathbb{E}_{x\in\mathcal{D}_t}\left[\phi^*(x, \pi(x))\right]$ where $\mathcal{D}_t$ is the distribution of context at time $t$. The algorithm FALCON is shown in \prettyref{ex: falcon}.
Below, we show that it also satisfies Assumption~1'.
At time $t$, the context $x_t$ is sampled from $\mathcal{D}_t$, and the reward is generated by $\mathbb{E}[r_t(x_t,a_t)]=\phi^\star_t(x_t, a_t)$. We slightly modify their algorithm so that at every round $t$, the algorithm call the regression oracle once and obtain $\widehat{\phi}_t = \argmin_{\phi\in\Phi} \sum_{\tau=1}^{t-1}(\phi(x_\tau, a_\tau)-R_\tau)^2$ (the original algorithm does this only when the time index doubles), and then construct a mapping from context to action distribution $p_t(\cdot|\cdot)$ as specified in their algorithm.
Analogous to their definitions, we define
\begin{align*}
\mathcal{R}_{[1,t-1]}(\pi) &= \frac{1}{t-1} \sum_{\tau=1}^{t-1} \mathbb{E}_{x\sim \mathcal{D}_\tau}\left[ \phi^\star_\tau(x, \pi(x))\right], \\
\widehat{\calR}_{[1,t-1]}(\pi) &= \frac{1}{t-1}\sum_{\tau=1}^{t-1} \mathbb{E}_{x\sim \mathcal{D}_\tau}\left[\widehat{\phi}_{t}(x, \pi(x))\right], \\
\widehat{\text{\rm Reg}}_{[1,t-1]}(\pi) &= \widehat{\calR}_{[1,t-1]}(\pi_{\widehat{\phi}_t}) - \widehat{\calR}_{[1,t-1]}(\pi), \\
\text{Reg}\xspace_{[1,t-1]}(\pi) &= \max_{\phi\in\Phi} \mathcal{R}_{[1,t-1]}(\pi_\phi) - \mathcal{R}_{[1,t-1]}(\pi), \\
V_t(p, \pi) &= \mathbb{E}_{x\sim \mathcal{D}_t}\left[\frac{1}{p(\pi(x)|x)}\right], \\
\mathcal{V}_t(\pi) &= \max_{\tau\in[1,t]} V_\tau(p_\tau, \pi).
\end{align*}
We will show that FALCON satisfies Assumption~1' with the following definitions:
\begin{align*}
\Delta(t)&=\sqrt{A}\max_{x,a} |\phi_t^\star(x,a) - \phi_{t+1}^\star(x,a)| + \int_x |\mathcal{D}_t(x) - \mathcal{D}_{t+1}(x)|\mathrm{d}x, \\
\rho(t)&=\sqrt{\frac{A\log(|\Phi|T/\delta)}{t}}
\end{align*}
By the same calculation as in Lemma 7 of \citep{simchi2020bypassing}, for any $\pi$,
\begin{align*}
&(t-1)\left|\widehat{\calR}_{[1,t-1]}(\pi) - \mathcal{R}_{[1,t-1]}(\pi)\right|^2 \\
&= \frac{1}{t-1}\left(\sum_{\tau=1}^{t-1} \mathbb{E}_{x\sim \mathcal{D}_\tau}\left[\widehat{\phi}_t(x, \pi(x)) - \phi^\star_\tau(x,\pi(x))\right]\right)^2 \\
&\leq \sum_{\tau=1}^{t-1} \left(\mathbb{E}_{x\sim \mathcal{D}_\tau}\left[\widehat{\phi}_t(x, \pi(x)) - \phi^\star_\tau(x,\pi(x))\right]\right)^2 \\
&\leq \sum_{\tau=1}^{t-1} \left(\mathbb{E}_{x\sim \mathcal{D}_\tau}\left[\sqrt{\frac{1}{p_\tau(\pi(x)|x)}p_\tau(\pi(x)|x)\left(\widehat{\phi}_t(x, \pi(x)) - \phi^\star_\tau(x,\pi(x))\right)^2}\right]\right)^2\\
&\leq \sum_{\tau=1}^{t-1} \left(\mathbb{E}_{x\sim \mathcal{D}_\tau}\left[\sqrt{\frac{1}{p_\tau(\pi(x)|x)}\mathbb{E}_{a\sim p_\tau(\cdot|x)}\left(\widehat{\phi}_t(x, a) - \phi^\star_\tau(x,a)\right)^2}\right]\right)^2 \\
&\leq \sum_{\tau=1}^{t-1} \mathbb{E}_{x\sim \mathcal{D}_\tau}\left[\frac{1}{p_\tau(\pi(x)|x)}\right] \mathbb{E}_{x\sim \mathcal{D}_\tau}\mathbb{E}_{a\sim p_\tau(\cdot|x)}\left[\left(\widehat{\phi}_t(x, a) - \phi^\star_\tau(x,a)\right)^2\right] \\
&\leq \sum_{\tau=1}^{t-1} V_\tau(p_\tau, \pi) \mathbb{E}_{a\sim p_\tau(\cdot|x)}\left[\left(\widehat{\phi}_t(x, a) - \phi^\star_\tau(x,a)\right)^2\right] \\
&\leq \mathcal{V}_{t-1}(\pi) \sum_{\tau=1}^{t-1} \mathbb{E}_{a\sim p_\tau(\cdot|x)}\left[\left(\widehat{\phi}_t(x, a) - \phi^\star_\tau(x,a)\right)^2\right].
\end{align*}
Using \prettyref{lemma: falcon lemma 1} and \prettyref{lemma: FALCON bound max V} below, when $\widehat{\Delta}_{[1,t]}\leq \mathcal{O}\left(\rho(t)\right)$, we have
\begin{align*}
\sum_{\tau=1}^{t-1} \mathbb{E}_{x\sim\mathcal{D}_\tau, a\sim p_\tau(\cdot|x)}\left[\left(\widehat{\phi}_t(x,a) - \phi^\star_\tau(x,a)\right)^2\right]\leq \mathcal{O}\left( \log(T|\Phi|/\delta) \right)
\end{align*}
and $\mathcal{V}_{t-1}(\pi) \leq \mathcal{O}(A) + \max_{\tau\in[1,t-2]}\gamma_{\tau}\widehat{\text{\rm Reg}}_{[1,\tau]}(\pi) $, where $\gamma_t=\Theta\left(\sqrt{\frac{At}{\log(|\Phi|T/\delta)}}\right)$. Note that they are actually of the same order as in the Lemma 7 of \citep{simchi2020bypassing} since the additional terms contributed by $\Delta_{[1,t]}$ are dominated by other terms. Thus, the bound we get for $\left|\mathcal{R}_{[1,t-1]} - \widehat{\calR}_{[1,t-1]}\right|$ is of the same order as their Lemma 7, which is
\begin{align}
\left|\mathcal{R}_{[1,t-1]}(\pi) - \widehat{\calR}_{[1,t-1]}(\pi)\right|
&\leq \mathcal{O}\left(\sqrt{\frac{\log(T|\Phi|/\delta)}{t}\left(A+\max_{\tau\in[1,t-2]}\gamma_{\tau} \widehat{\text{\rm Reg}}_{[1,\tau]}(\pi)\right)}\right) \\
&\leq \frac{1}{16}\max_{\tau\in[1,t-2]}\widehat{\text{\rm Reg}}_{[1,\tau]}(\pi) + \mathcal{O}\left(\rho(t)\right) \tag{by AM-GM}. \\
\label{eq: R deviation}
\end{align}
Then one can follow the derivation in their Lemma 8 using \prettyref{eq: R deviation}, and get
\begin{align*}
\text{Reg}\xspace_{[1,t]}(\pi) - \widehat{\text{\rm Reg}}_{[1,t]}(\pi)
&\leq \frac{1}{8}\max_{\tau\in[1,t-1]}\widehat{\text{\rm Reg}}_{[1,\tau]}(\pi) + \mathcal{O}(\rho(t)), \\
\widehat{\text{\rm Reg}}_{[1,t]}(\pi) - \text{Reg}\xspace_{[1,t]}(\pi)
&\leq \frac{1}{8}\max_{\tau\in[1,t-1]}\widehat{\text{\rm Reg}}_{[1,\tau]}(\pi) + \mathcal{O}(\rho(t)).
\end{align*}
Using these two inequalities, together with $\left| \text{Reg}\xspace_{[1,\tau]}(\pi) - \text{Reg}\xspace_{[1,t]}(\pi) \right| \leq \mathcal{O}(\Delta_{[1,t]}) = \mathcal{O}\left( \rho(t) \right)$,
we can also prove
\begin{align}
\text{Reg}\xspace_{[1,t]}(\pi)\leq 2\widehat{\text{\rm Reg}}_{[1,t]}(\pi) + \mathcal{O}(\rho(t)), \qquad \widehat{\text{\rm Reg}}_{[1,t]}(\pi)\leq 2\text{Reg}\xspace_{[1,t]}(\pi) + \mathcal{O}(\rho(t)) \label{eq: FALCON closeness}
\end{align}
by induction as their Lemma 8. One can see that all bounds we obtain are of the same order as in the stationary case shown in \citep{simchi2020bypassing}, thanks to the condition $\Delta_{[1,t]}=\mathcal{O}(\rho(t))$.
Then following their Lemmas 9 and 10, we obtain regret bound $\max_\phi t\mathcal{R}_{[1,t]}(\pi_\phi) - \sum_{\tau=1}^t R_\tau = \mathcal{O}\left(\sqrt{At\log(T|\Phi|/\delta)}\right)$.
Similar to the calculation in \prettyref{eq: ILOVE first condition}, by picking $\widetilde{f}_t = \widehat{\calR}_{[1,t-1]}(\pi_{\widehat{\phi}_t}) + c\sqrt{\frac{A\log(T|\Phi|/\delta)}{t}}$ with large enough $c$, we have
\begin{align*}
\widetilde{f}_t &\geq \max_\phi \mathcal{R}_{[1,t-1]}(\pi_{\phi}) + c\sqrt{\frac{A\log(T|\Phi|/\delta)}{t}} - \mathcal{O}(\rho(t)) \tag{by \prettyref{eq: R deviation}}\\
&\geq \mathcal{R}_{[1,t-1]}(\pi_{\phi_1^\star})
\geq \mathcal{R}_1(\pi_{\phi_1^\star}) - \mathcal{O}(\Delta_{[1,t]}) \geq \min_{\tau\in[1,t]}\max_{\phi} \mathcal{R}_\tau(\pi_\phi) - \mathcal{O}(\Delta_{[1,t]}),
\end{align*}
which verifies \prettyref{eq: general condition 3}. To upper bound $\sum_{\tau=1}^t \left(\widetilde{f}_\tau - R_\tau\right)$, we follow a similar calculation as \prettyref{eq: ILOVE condition 2}, and use the condition $\Delta_{[1,t]}=\mathcal{O}(\rho(t))$. This verifies \prettyref{eq: general condition 4}.
\begin{lemma}
\label{lemma: falcon lemma 1}
If $\Delta_{[1,t]}\leq \mathcal{O}\left(\rho(t)\right)$, then
\begin{align*}
\sum_{\tau=1}^{t-1} \mathbb{E}_{x\sim\mathcal{D}_\tau, a\sim p_\tau(\cdot|x)}\left[\left(\widehat{\phi}_t(x,a) - \phi^\star_\tau(x,a)\right)^2\right] = \mathcal{O}\left(\log(T|\Phi|/\delta)\right).
\end{align*}
\end{lemma}
\begin{proof}
First, we consider a speicific $\phi$. Define $Y_\tau = \left( \phi(x_\tau, a_\tau) - R_\tau \right)^2 - \left(\phi^\star_{\tau}(x_\tau, a_\tau) - R_\tau\right)^2$. Then we have $\mathbb{E}[Y_\tau]=\mathbb{E}\left[(\phi(x_\tau, a_\tau) - \phi_\tau^\star(x_\tau, a_\tau))^2\right]$ and $\mathbb{E}\left[ Y_\tau^2 \right]\leq 4\mathbb{E}\left[(\phi(x_\tau, a_\tau)-\phi_\tau^\star(x_\tau, a_\tau))^2\right]=4\mathbb{E}[Y_\tau]$. By Freedman's inequality,
\begin{align*}
\sum_{\tau=1}^{t-1} Y_\tau
&\geq \sum_{\tau=1}^{t-1} \mathbb{E}[Y_\tau] - c_1\sqrt{\sum_{\tau=1}^{t-1}\mathbb{E}\left[Y_\tau^2\right] \log(T/\delta) } - c_2 \log(T|\Phi|/\delta) \\
&\geq \sum_{\tau=1}^{t-1} \mathbb{E}[Y_\tau] - 2c_1\sqrt{\sum_{\tau=1}^{t-1}\mathbb{E}\left[Y_\tau\right] \log(T/\delta) } - c_2 \log(T|\Phi|/\delta).
\end{align*}
The above implies (by solving for $\sum_{\tau=1}^{t-1} \mathbb{E}[Y_\tau]$)
\begin{align}
\sum_{\tau=1}^{t-1} \mathbb{E}[Y_\tau] \leq 2\sum_{\tau=1}^{t-1}Y_\tau + 4(c_1^2+c_2)\log(T|\Phi|/\delta). \label{eq: CB freedman 1}
\end{align}
For the other direction, we also have
\begin{align}
\sum_{\tau=1}^{t-1} Y_\tau
&\leq 2\sum_{\tau=1}^{t-1} \mathbb{E}[Y_\tau] + \left(\frac{c_1^2}{4} + c_2\right) \log(T|\Phi|/\delta). \label{eq: CB freedman 2}
\end{align}
Then we can bound
\begin{align*}
&\sum_{\tau=1}^{t-1} \mathbb{E}_{x\sim\mathcal{D}_\tau, a\sim p_\tau(\cdot|x)}\left[\left(\widehat{\phi}_t(x,a) - \phi^\star_\tau(x,a)\right)^2\right] \tag{using \prettyref{eq: CB freedman 1}}\\
&\leq 2\sum_{\tau=1}^{t-1} \left(\widehat{\phi}_{t}(x_\tau, a_\tau) - R_\tau\right)^2 - 2\sum_{\tau=1}^{t-1} \left(\phi_{\tau}^\star(x_\tau, a_\tau) - R_\tau\right)^2 + 4(c_1^2+c_2)\log(T|\Phi|/\delta) \\
&\leq 2\sum_{\tau=1}^{t-1} \left(\phi_1^\star(x_\tau, a_\tau) - R_\tau\right)^2 - 2\sum_{\tau=1}^{t-1} \left(\phi_{\tau}^\star(x_\tau, a_\tau) - R_\tau\right)^2 + 4(c_1^2+c_2)\log(T|\Phi|/\delta) \tag{by the optimality of $\widehat{\phi}_t$} \\
&\leq 4\sum_{\tau=1}^{t-1} \mathbb{E}\left[\left(\phi_1^\star(x_\tau, a_\tau) - \phi_\tau^\star(x_\tau, a_\tau)\right)^2\right] + c_3 \log(T|\Phi|/\delta) \tag{using \prettyref{eq: CB freedman 2}}\\
&\leq \frac{4(t-1)}{A}\Delta_{[1,t]}^2 + c_3\log(T|\Phi|/\delta). \tag{by the definition of $\Delta_{[\cdot,\cdot]}$}
\end{align*}
By the condition on $\Delta_{[1,t]}$, we have $\frac{4(t-1)}{A}\Delta_{[1,t]}^2=\mathcal{O}\left(\log(T|\Phi|/\delta)\right)$, which proves the lemma.
\end{proof}
\begin{lemma} \label{lemma: FALCON bound max V}
If $\Delta_{[1,t]}\leq \mathcal{O}\left(\rho(t)\right)$, then
\begin{align*}
\mathcal{V}_{t}(\pi) \leq \mathcal{O}(A) + \max_{\tau\in[1,t-1]}\gamma_{\tau}\widehat{\text{\rm Reg}}_{[1,\tau]}(\pi)
\end{align*}
where $\gamma_t=\Theta\left(\sqrt{\frac{At}{\log(|\Phi|T/\delta)}}\right)$.
\end{lemma}
\begin{proof}
Similar to Lemma 6 of FALCON, for $\tau\in[1,t]$,
\begin{align*}
V_\tau(p_\tau, \pi)
&\leq A + \gamma_{\tau-1} \mathbb{E}_{x\sim \mathcal{D}_\tau}\left[ \widehat{\phi}_\tau(x, \pi_{\widehat{\phi}_\tau}(x)) - \widehat{\phi}_\tau(x,\pi(x)) \right] \\
&\leq A + \frac{\gamma_{\tau-1}}{\tau-1}\sum_{s=1}^{\tau-1}\mathbb{E}_{x\sim \mathcal{D}_s}\left[ \widehat{\phi}_\tau(x, \pi_{\widehat{\phi}_\tau}(x)) - \widehat{\phi}_\tau(x,\pi(x)) \right] + \gamma_{\tau-1} \Delta_{[1,t]} \\
&\leq A + \gamma_{\tau-1}\widehat{\text{\rm Reg}}_{[1,\tau-1]}(\pi) + \gamma_{\tau-1} \Delta_{[1,t]}.
\end{align*}
By the condition $\widehat{\Delta}_{[1,t]}\leq \mathcal{O}\left(\rho(t)\right)$, that last term $\gamma_{\tau-1} \Delta_{[1,t]}$ is of order $\mathcal{O}\left( A \right)$. By the definition of $\mathcal{V}_{t}(\pi)$, this finishes the proof.
\end{proof}
\section{Extension to Reinforcement Learning in Infinite-horizon Communicating MDPs}
\label{sec: average-reward}
As mentioned, applying our results to infinite-horizon RL~\citep{jaksch2010near} requires some extra care and extensions.
We refer the reader to \citep{cheung2020reinforcement} for a thorough introduction on the problem setup of infinite-horizon RL in time-varying communicating MDPs.
Here, we only highlight its difference compared to episode RL and explain how to fit it into our framework.
Specifically, in episodic RL, we have treated each episode (consisting of multiple steps in an MDP) as one round of our framework, each state-to-action mapping as a policy $\pi$, and the expected reward of executing $\pi$ in the MDP for round $t$ as $f_t(\pi)$.
In infinite-horizon RL, while the meaning of $\pi$ and $f_t$ remains the same, there is no episode any more and the learner interacts with the changing MDP from the start to the end without any reset on her state.
In this case, we treat each step (that is, each state transition) in the MDP as one round in our framework, and the meaning of the reward feedback $R_t$ has now changed from a noisy observation of the policy's reward $f_t(\pi_t)$ to just the reward of $\pi_t$ for this single step.
With this change, the dynamic regret definition remains the same.
Due to the black-box nature of our approach, if one has a base algorithm that satisfies something close to \prettyref{assum:assump2} within this setup, then it is not hard to imagine that the same idea of {\small\textsf{\textup{MASTER}}}\xspace can be applied.
In \prettyref{subsec:ACW}, we provide such a base algorithm,
and in \prettyref{subsec:MUCRL}, we combine it with appropriate multi-scale scheduling and detection to obtain our final results.
\subsection{{\normalsize\textsf{UCRL}} with Adaptive Confidence Widening}\label{subsec:ACW}
Our base algorithm, {\small\textsf{\textup{UCRL-ACW}}}\xspace, is an improvement of the standard {\small\textsf{\textup{UCRL}}}\xspace algorithm~\citep{jaksch2010near} and its variant {\small\textsf{\textup{UCRL-CW}}}\xspace~\citep{cheung2020reinforcement}.
The pseudocode is shown in \prettyref{alg: base alg UCRL} (\prettyref{app:ucrl_details}), where we highlight the differences compared to {\small\textsf{\textup{UCRL}}}\xspace and {\small\textsf{\textup{UCRL-CW}}}\xspace in blue.
The first difference is the explicit mention that the next state of the learner might sometime be {\it arbitrarily} assigned instead of following the transition of the current MDP (\prettyref{line: reassignment}).
This is necessary because of the multi-scale scheduling of {\small\textsf{\textup{MALG}}}\xspace.
Indeed, recall that in {\small\textsf{\textup{MALG}}}\xspace, an instance of the base algorithm can sometimes be paused and then resumed later.
In the infinite-horizon RL setup, this means that the instance can be resumed from an arbitrary state.
Other than making this detail explicit, however, nothing really needs to be changed in the algorithm, since this happens infrequently and only incurs small additional regret due to the communicating property of the MDPs.
The second key difference is an adaptive version of the {\it Confidence Widening} technique of~\citep{cheung2020reinforcement} (see \prettyref{line: adaptive start}--\prettyref{line: active end}).
As pointed out in~\citep{cheung2020reinforcement}, in non-stationary environments, the Extended Value Iteration (EVI) subroutine of {\small\textsf{\textup{UCRL}}}\xspace might return a bias vector ($\widetilde{h}_k$) with span much larger than $D_{\max}$, the maximum diameters of all the MDPs.
To address this issue, their confidence widening technique adds a constant $\eta$, tuned based on $\Delta$, to the confidence level of the confidence set $\mathcal{P}_k$,
which eventually leads to sub-optimal regret $\Delta^{\nicefrac{1}{4}}T^{\nicefrac{3}{4}}$.
Our \emph{adaptive confidence widening}, on the other hand, adaptively selects the value of $\eta$ in a doubling manner, so that in a relatively stationary environment we only widen the confidence set slightly, while in a more non-stationary environment the widening is more significant.
To avoid incurring too much additional regret in the latter case,
we also monitor the cumulative widening amount and terminate the algorithm if it exceeds a certain threshold (\prettyref{line: accumulate}--\prettyref{line: return terminate}), because this implies that the environment is highly non-stationary.
(This termination will also be a restart signal for {\small\textsf{\textup{MASTER}}}\xspace.)
Finally, notice that our black-box approach requires knowing the regret bound $\rho(\cdot)$ of the base algorithm, which in this case depends on $D_{\max}$, a potentially unknown quantity.
To deal with this issue, {\small\textsf{\textup{UCRL-ACW}}}\xspace takes a guess $\overline{D}$ on the value of $D_{\max}$ as an additional input.
In the next subsection, we discuss how {\small\textsf{\textup{MASTER}}}\xspace decides the value of $\overline{D}$ when $D_{\max}$ is unknown.
With all these modifications, our base algorithm {\small\textsf{\textup{UCRL-ACW}}}\xspace indeed provides a guarantee similar to \prettyref{eq: general condition 2} of \prettyref{assum:assump2}; see \prettyref{lem: modified UCRL}.
\subsection{Multi-scale {\normalsize\textsf{UCRL-ACW}} and Its Combination with {\normalsize\textsf{MASTER}}}
\label{subsec:MUCRL}
Now, we use the same idea as in \prettyref{subsec: multiscale} to create a multi-scale version of {\small\textsf{\textup{UCRL-ACW}}}\xspace, under a fixed input $\overline{D}$. The resulted algorithm is called Multi-scale {\small\textsf{\textup{UCRL-ACW}}}\xspace or {\small\textsf{\textup{MUCRL}}}\xspace for short (see \prettyref{alg: multialg-ucrl}). {\small\textsf{\textup{MUCRL}}}\xspace is basically identical to {\small\textsf{\textup{MALG}}}\xspace with {\small\textsf{\textup{UCRL-ACW}}}\xspace as the base algorithm, except that we let {\small\textsf{\textup{MUCRL}}}\xspace terminate whenever the currently active {\small\textsf{\textup{UCRL-ACW}}}\xspace instance makes a restart signal (due to having an abnormally large cumulative widening amount).
The guarantee for {\small\textsf{\textup{MUCRL}}}\xspace is provided in \prettyref{lemma: ucrl aggregated regret}, which parallels \prettyref{lemma: multi-scale reg}.
Next, as in \prettyref{sec: final algorithm sec}, we further combine {\small\textsf{\textup{MUCRL}}}\xspace with non-stationarity tests, leading to {\small\textsf{\textup{MASTER-UCRL}}}\xspace (see \prettyref{alg: master ucrl}).
The only difference compared to {\small\textsf{\textup{MASTER}}}\xspace is an additional condition to restart (highlighted in blue) --- when {\small\textsf{\textup{MUCRL}}}\xspace terminates due to a restart signal from an {\small\textsf{\textup{UCRL-ACW}}}\xspace instance. We provide a single-block regret bound guarantee for {\small\textsf{\textup{MASTER-UCRL}}}\xspace under a fixed $\overline{D}$ in \prettyref{lemma: block lemma for RL}, which parallels \prettyref{lemma: bound term1}.
Finally, we discuss three different cases with knowledge of different parameters (if any), leading to the three results listed in \prettyref{tab:compare}.
\paragraph{Known $D_{\max}$}
When $D_{\max}$ is known, we simply set $\overline{D}=D_{\max}$.
In this case, all restarts of {\small\textsf{\textup{MASTER-UCRL}}}\xspace are due to non-stationarity, and we can bound their number in terms of $L$ or $\Delta$.
Together with the single-block regret guarantee from \prettyref{lemma: block lemma for RL},
we prove that {\small\textsf{\textup{MASTER-UCRL}}}\xspace's dynamic regret is $\widetilde{\mathcal{O}}(\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\})$;
see \prettyref{thm: known Dmax case} for the dependence on other parameters.
\paragraph{Unknown $D_{\max}$ and Known $L$ or $\Delta$} When $D_{\max}$ is unknown, we unfortunately require the knowledge of $L$ to get $\Reg_L^\star\xspace$ and the knowledge of $\Delta$ to get $\Reg_\Delta^\star\xspace$.
However, as shown in \prettyref{tab:compare},
this still significantly improves over the best existing bounds $\widetilde{\mathcal{O}}(L^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}})$ and $\widetilde{\mathcal{O}}(\Delta^{\nicefrac{1}{4}}T^{\nicefrac{3}{4}})$ when $L$ and $\Delta$ are known.
Specifically, we apply a doubling trick to set the value of $\overline{D}$ following the
strategy below, where we call the interval between two restarts an {\it epoch}:
\begin{enumerate}
\item Initialize $\overline{D}\leftarrow 1$.
\item Run {\small\textsf{\textup{MASTER-UCRL}}}\xspace with $\overline{D}$. If the number of epochs exceeds $\overline{N}$, then double $\overline{D}$ and repeat this step.
Here, $\overline{N}$ is set to $L$ if $L$ is known or $1+3(S^{-2}A^{-1}\Delta^2T)^{\nicefrac{1}{3}}$ if $\Delta$ is known.
\end{enumerate}
The rationale behind monitoring the number of epochs is that, when $\overline{D}$ is too small, {\small\textsf{\textup{UCRL-ACW}}}\xspace might have an abnormally large cumulative widening amount and signal a restart even in a fairly stationary environment.
In \prettyref{lemma: RL number epoch}, we show that if $\overline{D}\geq D_{\max}$, the number of epochs produced by {\small\textsf{\textup{MASTER-UCRL}}}\xspace is upper bounded by the value of $\overline{N}$ set above. Therefore, if it exceeds this number, we can infer $\overline{D} < D_{\max}$ and double its value.
This allows us to prove the regret bound $\Reg_L^\star\xspace$ or $\Reg_\Delta^\star\xspace$ again; see \prettyref{thm: doubling trick algo for RL} for the details.
\paragraph{No prior knowledge at all}
When nothing is known, we apply the Bandit-over-Reinforcment-Learning (BoRL) framework of~\citep{cheung2019learning, cheung2020reinforcement} to get a suboptimal bound of order $\widetilde{\mathcal{O}}(\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\} +T^{\nicefrac{3}{4}})$.
BoRL also serves as a black-box reduction to obtain parameter-free algorithms (albeit suboptimal), so applying it to our algorithm is straightforward.
We omit the details and only give the concrete bound in \prettyref{app: borl discuss}.
We leave the question of whether the optimal bound is achievable when $L$, $\Delta$, and $D_{\max}$ are all unknown as a future direction.
\section{Introduction}\label{sec: intro}
Most existing works on reinforcement learning consider a stationary environment and aim to find or be comparable to an optimal policy (known as having low \emph{static regret}).
In many applications, however, the environment is far from being stationary.
In these cases, it is much more meaningful to minimize \emph{dynamic regret}, the gap between the total reward of the optimal {\it sequence} of policies and that of the learner.
Indeed, there is a surge of studies on this topic recently~\citep{jaksch2010near, gajane2018sliding, li2019online, ortner2020variational, cheung2020reinforcement, fei2020dynamic, domingues2020kernel, mao2020nearoptimal, zhou2020nonstationary, touati2020efficient}.
One common issue of all these works, however, is that their algorithms crucially rely on having some prior knowledge on the \emph{degree of non-stationarity} of the world, such as how much or how many times the distribution changes, which is often unavailable in practice. \citet{cheung2020reinforcement} develop a Bandit-over-Reinforcement-Learning (BoRL) framework to relax this assumption, but it introduces extra overhead and leads to suboptimal regret. Indeed, as discussed in their work, there are multiple aspects (which they call \emph{endogeneity}, \emph{exogeneity}, \emph{uncertainty}, and \emph{bandit feedback}) combined in non-stationary reinforcement learning that make the problem highly challenging.
For bandit problems, the special case of reinforcement learning, the works of~\citet{auer2019adaptively} and~\citet{chen2019new} are the first to achieve near-optimal dynamic regret without any prior knowledge on the degree of non-stationarity.
The same technique has later been adopted by \cite{chen2020combinatorial} for the case of combinatorial semi-bandits.
Their algorithms maintain a distribution over arms (or policies/super-arms in the contextual/combinatorial case~\citep{chen2019new, chen2020combinatorial}) with properly controlled variance for all reward estimators. This approach is generally incompatible with standard reinforcement learning algorithms, which are usually built upon the \emph{optimism in the face of uncertainty} principle and do not maintain a distribution over policies (see also \citep{lykouris2019corruption, wang2020long} for related discussions). Another drawback is that their algorithms are very specialized to their problems, and it is highly unclear whether the ideas can be extended to other problems.
In this work, we address all these issues and make significant progress in this direction.
Specifically, we propose a {\it general} approach that is applicable to various reinforcement learning settings (including bandits, episodic MDPs, infinite-horizon MDPs, etc.) and achieves {\it optimal dynamic regret without any prior knowledge} on the degree of non-stationarity.
Our approach, called {\small\textsf{\textup{MASTER}}}\xspace, is a {\it black-box reduction} that turns any algorithm with optimal performance in a (near-)stationary environment and additionally some mild requirements into another algorithm with optimal dynamic regret in a non-stationary environment,
again, without the need of any prior knowledge.
For example, all existing UCB-based algorithms satisfy the conditions of our reduction and are readily to be plugged into our black-box.
\renewcommand{\arraystretch}{1.15}
\begin{table}[t]
\centering
\caption{A summary of our results and comparisons with the state-of-the-art.
Our algorithms are named in the form of ``{\small\textsf{\textup{MASTER}}}\xspace + $X$'' where $X$ is the base algorithm used in our reduction.
Here, $\Reg_L^\star\xspace = \sqrt{LT}$ and $\Reg_\Delta^\star\xspace = \Delta^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}} + \sqrt{T}$, where $T$ is the number of rounds and $L$ and $\Delta$ are the number and amount of changes of the world respectively (dependence on other parameters is omitted).
$D_{\max}$ is the maximum diameters of the MDPs.
}
\label{tab:compare}
\vspace*{5pt}
\begin{tabular}{|c|c|c|c|}
\Xhline{\arrayrulewidth}
Setting & Algorithm & Regret in $\widetilde{\mathcal{O}}(\cdot)$ & \makecell{Required \\ knowledge} \\
\Xhline{\arrayrulewidth}
\multirow{2}{*}{Multi-armed bandits} & \makecell{\citep{auer2019adaptively}} & $\Reg_L^\star\xspace$ & \\
\cline{2-4}
& \cellcolor{blue!25}\text{{\small\textsf{\textup{MASTER}}}\xspace + UCB1} & $\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\}$ & \\
\Xhline{\arrayrulewidth}
\multirow{3}{*}{Contextual bandits} & \makecell{\citep{chen2019new}} & \multirow{3}{*}{$\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\}$} & \\
\cline{2-2
& \cellcolor{blue!25}\text{{\small\textsf{\textup{MASTER}}}\xspace + ILTCB} & & \\
\cline{2-2}
& \cellcolor{blue!25}\text{{\small\textsf{\textup{MASTER}}}\xspace + FALCON} & & \\
\Xhline{\arrayrulewidth}
\multirow{3}{*}{Linear bandits} & \makecell{\citep{cheung2018hedging}} & $\Reg_\Delta^\star\xspace$ & $\Delta$ \\
\cline{2-4}
& \makecell{\citep{cheung2018hedging}} & $\Reg_\Delta^\star\xspace + T^{\nicefrac{3}{4}}$ & \\
\cline{2-4}
& \cellcolor{blue!25} \text{{\small\textsf{\textup{MASTER}}}\xspace + OFUL} & $\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\}$ & \\
\Xhline{\arrayrulewidth}
\multirow{3}{*}{Generalized linear bandits} & \makecell{\citep{russac2020algorithms}} & $L^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}$ & $L$ \\
\cline{2-4}
& \makecell{\citep{faury2021regret}} & $\Delta^{\nicefrac{1}{5}}T^{\nicefrac{4}{5}} + T^{\nicefrac{3}{4}}$ & \\
\cline{2-4}
& \cellcolor{blue!25} \text{{\small\textsf{\textup{MASTER}}}\xspace + GLM-UCB} & $\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\}$ & \\
\Xhline{\arrayrulewidth}
\multirow{2}{*}{\makecell{Episodic MDPs\\ (tabular case)}} & \makecell{\citep{mao2020nearoptimal}} & $\Reg_\Delta^\star\xspace$ & $\Delta$ \\
\cline{2-4}
& \cellcolor{blue!25}\text{{\small\textsf{\textup{MASTER}}}\xspace + Q-UCB} & $\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\}$ & \\
\Xhline{\arrayrulewidth}
\multirow{2}{*}{\makecell{Episodic MDPs\\ (linear case)}} & \makecell{\citep{touati2020efficient}} & $\Delta^{\nicefrac{1}{4}}T^{\nicefrac{3}{4}} + \sqrt{T}$ & $\Delta$ \\
\cline{2-4}
& \cellcolor{blue!25}\text{{\small\textsf{\textup{MASTER}}}\xspace + LSVI-UCB} & $\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\}$ & \\
\Xhline{\arrayrulewidth}
\multirow{6}{*}{\makecell{Infinite-horizon \\ communicating MDPs \\ (tabular case)}} & \makecell{ \citep{gajane2018sliding}} & $L^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}$ & $L$ \\
\cline{2-4}
& \makecell{\citep{cheung2020reinforcement}} & $\Delta^{\nicefrac{1}{4}}T^{\nicefrac{3}{4}} + \sqrt{T}$ & $\Delta$ \\
\cline{2-4}
& \makecell{\citep{cheung2020reinforcement}} & $\Delta^{\nicefrac{1}{4}}T^{\nicefrac{3}{4}} + T^{\nicefrac{3}{4}}$ & \\
\cline{2-4}
&\cellcolor{blue!25} \multirow{1}{*}{\text{{\small\textsf{\textup{MASTER}}}\xspace + UCRL}} & \multirow{1}{*}{ $\Reg_L^\star\xspace$ or $\Reg_\Delta^\star\xspace$} & $L$ or $\Delta$ \\
\cline{2-4}
& \cellcolor{blue!25}\text{{\small\textsf{\textup{MASTER}}}\xspace + UCRL} & $\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\}$ & \makecell{$D_{\max}$} \\
\cline{2-4}
&\cellcolor{blue!25}\text{{\small\textsf{\textup{MASTER}}}\xspace + UCRL + BoRL} & \makecell{$\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\} +T^{\nicefrac{3}{4}}$} & \\
\Xhline{\arrayrulewidth}
\end{tabular}
\end{table}
\paragraph{Applications and comparisons}
To showcase the versatility of our approach, we provide a list of examples by considering different settings and applying our reduction with different base algorithms.
These examples, summarized in \prettyref{tab:compare}, recover the results of \cite{auer2019adaptively} and \cite{chen2019new} for (contextual) multi-armed bandits, and more importantly, improve the best known results for (generalized) linear bandits, episodic MDPs, and infinite-horizon MDPs in various ways.
More specifically, let $L$ and $\Delta$ be the number and amount of changes of the environment respectively (see \prettyref{sec:setting} for formal definition).
For all settings except infinite-horizon MDPs, ignoring other parameters, our algorithms achieve dynamic regret $\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\}$ without knowing $L$ and $\Delta$, where $\Reg_L^\star\xspace = \sqrt{LT}$, $\Reg_\Delta^\star\xspace = \Delta^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}} + \sqrt{T}$, and $T$ is the number of rounds.
These bounds are known to be optimal {\it even when $L$ and $\Delta$ are known},
and they improve over~\citep{cheung2018hedging, cheung2019learning, russac2019weighted, kim2019randomized, pmlr-v108-zhao20a, zhao2021nonstationary} for linear bandits, \citep{russac2020algorithms, faury2021regret} for generalized linear bandits, \citep{mao2020nearoptimal} for episodic tabular MDPs,
and \citep{touati2020efficient, zhou2020nonstationary} for episodic linear MDPs.
For infinite-horizon MDPs, we achieve the same optimal regret when the maximum diameter of the MDPs is known, or when $L$ and $\Delta$ are known, improving over the best existing results by~\citep{gajane2018sliding} and~\citep{cheung2020reinforcement}.
When none of them is known, we can still adopt the BoRL technique~\citep{cheung2020reinforcement} with the price of paying extra $T^{3/4}$ regret,
which is suboptimal but still outperforms best known results.
In particular, we emphasize that achieving dynamic regret $\Reg_L^\star\xspace$ beyond (contextual) multi-armed bandits is one notable breakthrough we make.
Indeed, \emph{even when $L$ is known}, previous approaches based on restarting after a fixed period, a sliding window with a fixed size, or discounting with a fixed discount factor, all lead to a suboptimal bound of $\widetilde{\mathcal{O}}(L^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}})$ at best~\citep{gajane2018sliding}.
Since this bound is subsumed by $\Reg_\Delta^\star\xspace$, related discussions are also often omitted in previous works.
For non-stationary linear bandits, although several existing works \citep{russac2019weighted, kim2019randomized, pmlr-v108-zhao20a}
claim that their algorithms achieve the bound $\Reg_\Delta^\star\xspace$ (when $\Delta$ is known), there is in fact a technical flaw in all of them, as explained and corrected recently in~\citep{zhao2021nonstationary, touati2020efficient}. After the correction, their bounds all deteriorate to $\Delta^{\nicefrac{1}{4}}T^{\nicefrac{3}{4}}+\sqrt{T}$, which is no longer near-optimal. Recently, \cite{cheung2018hedging} sidesteps this difficulty by leveraging adversarial linear bandit algorithms, and achieves the tight bound of $\Reg_\Delta^\star\xspace$ when $\Delta$ is known.
On the other hand, our approach is based on stochastic linear bandit algorithms; however, it not only sidesteps the difficulty met in previous works, but also avoids the requirement of knowing $\Delta$. When dealing with other linear-structured problems including generalized linear bandits and linear MDPs, our bounds $\Reg_\Delta^\star\xspace$ is new \emph{even when $\Delta$ is known}. Previous results \citep{russac2020algorithms, faury2021regret, touati2020efficient, zhou2020nonstationary} cannot achieve the optimal bound due to the same technical difficulty mentioned above.
\paragraph{High-level ideas}
The high-level idea of our reduction is to schedule multiple instances of the base algorithm with different durations in a carefully-designed randomized scheme, which facilitates non-stationarity detection with little overhead.
A related and well-known approach for non-stationary environments is to maintain multiple instances of a base algorithm with different parameter tunings or different starting points and to learn the best of them via another ``expert'' algorithm,
which can be very successful when learning with full information~\citep{hazan2007adaptive, luo2015achieving, daniely2015strongly, jun2017improved} but is suboptimal and has many limitations when learning with partial information~\citep{luo2018efficient,cheung2019learning,cheung2020reinforcement}.
Our approach is different as we do not try to learn the best instance; instead, we always follow the decision suggested by the instance with the currently shortest scheduled duration, and also only update this instance after receiving feedback from the environment.
The is because base algorithms with shorter duration are responsible for detecting larger distribution changes,
and always following the shortest one ensure that it is not blocked by the longer ones and thus every scale of distribution change is detected in a timely manner.
Another related approach is \emph{regret balancing}, developed recently for model selection in bandit problems~\citep{abbasi2020regret, pacchiano2020regret}. The idea is also to run multiple base algorithms in parallel, each with a putative regret upper bound. The learner executes one of them in each round which incurs the least regret so far, and also constantly compares the performance among base algorithms, eliminating those whose putative regret bounds are violated. While our algorithm resembles regret balancing in some aspects, the way it chooses the base algorithm in each round is clearly quite different, which is also crucial for our problem.
\paragraph{Other related work}
There are also a series of works on learning MDPs with adversarial rewards and a {\it fixed} transition \citep{even2009online, neu2010online, arora2012deterministic, neu2012adversarial, dekel2013better, neu2013online, zimin2013online, dick2014online, rosenberg2019online, cai2020provably, jin2020learning, shani2020optimistic, rosenberg2020adversarial, lee2020bias, jin2020simultaneously, chen2020minimax, lancewicki2020learning}.
These models can potentially handle non-stationarity in the reward function but not the transition kernel (in fact, most of these works also only consider static regret).
\citet{lykouris2019corruption} investigate an episodic MDP setting where an adversary can corrupt both the reward and the transition for up to $L'$ episodes, and achieve dynamic regret $\widetilde{\mathcal{O}}(\min\{L'\sqrt{T}, L'/\textit{gap}\})$ without knowing $L'$, where $\textit{gap}$ is the minimal suboptimality gap and could be arbitrarily small. Since corruption of up to $L'$ episodes implies that the world changes at most $L=2L'$ times, our result improves theirs from $\widetilde{\mathcal{O}}(L'\sqrt{T})$ to $\widetilde{\mathcal{O}}(\sqrt{L'T})$ when $1/\textit{gap} > \sqrt{T}$.
On the other hand, it is possible that $L$ is much smaller than $L'$ (e.g. $L=\Theta(1)$ while $L' = \Theta(T)$), in which case our results are also significantly better.
\section{Conclusion and Future Directions}
In this work, we study reinforcement learning in non-stationary environments. We propose a general black-box approach that can convert an algorithm with near-optimal regret in a (near-)stationary environment to another algorithm with near-optimal dynamic regret in a non-stationary environment. Prior to our work, the bound of $\widetilde{\mathcal{O}}(\Delta^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}})$ is only achievable with the knowledge of $\Delta$, and no algorithm achieves the bound of $\widetilde{\mathcal{O}}(\sqrt{LT})$ even with the knowledge of $L$. Our algorithm achieves both bounds simultaneously without any prior knowledge.
It would be interesting to see whether algorithms with data-dependent bounds work with our black-box approach. Previous work in this direction \citep{wei2016tracking} achieves an improved dynamic regret bound for multi-armed bandits when the cumulative variance of the loss is small; however, their approach crucially relies on the knowledge on the degree of non-stationarity as well as the cumulative variance. On the other hand, there are some immediate difficulties in applying our black-box approach to data-dependent algorithms. For example, the monotonicity of the the average regret $\rho(\cdot)$ may not hold anymore, and it is unclear how to set the probability of initiating a new base algorithm. Therefore, the task of achieving data-dependent dynamic bounds without prior knowledge seems to be challenging and requires other innovations.
Another future direction is to study a class of contextual bandit problems where the context is adversarially generated \citep{abbasi2011improved, cheung2019learning, foster2020beyond}. In this case, the expected reward of the optimal policy changes over time even if the environment is stationary, so our current algorithm cannot be directly applied. For linear contextual bandits with adversarial contexts \citep{abbasi2011improved, cheung2019learning}, the fix is straightforward though: instead of requiring the base algorithm to generate a scalar $\widetilde{f}_t$ in each round, we let it generate a \emph{confidence set} for the hidden parameter, and check the inconsistency of the confidence set over time. However, for general contextual bandits with adversarial contexts, where algorithms do not necessarily maintain a confident set for the hidden parameter~\citep{foster2020beyond}, the extension is less clear and is left for future investigation.
Finally, we are not aware of any near-optimal \emph{convex bandit} algorithm satisfying our \prettyref{assum:assump2}, so achieving near-optimal dynamic regret bound in general convex bandits is also left open.
\paragraph{Acknowledgments}
We thank Peng Zhao for pointing out the technical flaw made in previous works on non-stationary linear bandits as well as a fix in~\citep{zhao2021nonstationary}, and thank Ruihao Zhu for informing us their non-stationary linear bandit algorithm with the $\Reg_\Delta^\star\xspace$ bound \citep{cheung2018hedging}.
We also thank anonymous reviewers for pointing out the relation between our algorithm and regret balancing \citep{abbasi2020regret, pacchiano2020regret}.
This work is supported by NSF
Award IIS-1943607 and a Google Faculty Research Award.
\bibliographystyle{plainnat}
\section{Problem Setting, Main Results, and High-level Ideas}
Throughout the paper, we fix a probability parameter $\delta$ of order $1/\text{poly}(T)$, and write $h_1(x)=\widetilde{\mathcal{O}}(h_2(x))$ or $h_2(x)=\widetilde{\Omega}(h_1(x))$ if $h_1(x)=\mathcal{O}\left(\text{poly}(\log(T/\delta)) h_2(x)\right)$. We say ``with high probability, $h_1=\widetilde{\mathcal{O}}(h_2(x))$'' if ``with probability $1-\delta$, $h_1=\widetilde{\mathcal{O}}(h_2(x))$''.
For an integer $n$, we denote the set $\{1,2,\ldots, n\}$ by $[n]$;
and for integers $s$ and $e$, we denote the set $\{s,s+1,\ldots, e\}$ by $[s,e]$.
\subsection{Problem setting}\label{sec:setting}
We consider the following general reinforcement learning (RL) framework that covers a wide range of problems.
Ahead of time, the learner is given a policy set $\Pi$,
and the environment decides $T$ reward functions $f_1, \ldots, f_T: \Pi \to [0,1]$ unknown to the learner.
Then, in each round $t = 1, \ldots, T$, the learner chooses a policy $\pi_t\in\Pi$ and receives a noisy reward $R_t \in [0,1]$ whose mean is $f_t(\pi_t)$.\footnote
The range $[0,1]$ is only for simplicity. Our results can be directly extended to the case with sub-Gaussian noise.
}
The dynamic regret of the learner is defined as
$
\textsc{D-Reg}\xspace = \sum_{t=1}^T \left(f^\star_t - R_t\right)
$,
where $f^\star_t = \max_{\pi\in\Pi} f_t(\pi)$ is the expected reward of the optimal policy for round $t$.
Many heavily-studied problems fall into this framework.
For example, in the classic multi-armed bandit problem~\citep{lai1985asymptotically},
it suffices to treat each arm as a policy;
for finite-horizon episodic RL (e.g.~\citep{jin2018q}),
each state-to-action mapping is considered as a policy, and $f_t(\pi)$ is the expected reward of executing $\pi$ in the $t$-th episode's MDP with some transition kernel and some reward function.
See more examples in \prettyref{app: verify example}.
Note that our framework ignores many details of the actual problem we are trying to solve (e.g. not even mentioning the MDPs for RL).
This is because our results only rely on certain guarantees provided by a base algorithm, making these details irrelevant to our presentation.
There is also some technicality to fit the infinite-horizon RL problem into our framework, which we will discuss in detail in \prettyref{sec: average-reward}.
\paragraph{Non-stationarity measure}
A natural way to measure the distribution drift between rounds $t$ and $t+1$ is to see how much the expected reward of any policy could change, that is, $\max_{\pi\in\Pi} |f_{t}(\pi) - f_{t+1}(\pi)|$.
However, to make our results more general, we take a sligtly more abstract way to define non-stationarity whose exact form eventually depends on what guarantees the base algorithm can provide for a concrete problem.
To this end, we define the following.
\begin{definition}\label{def: deviation}
$\Delta: [T]\to \mathbb{R}$ is a non-stationarity measure if it satisfies $\Delta(t) \geq \max_{\pi\in\Pi} |f_{t}(\pi) - f_{t+1}(\pi)|$ for all $t$.
Define for any interval $\mathcal{I} = [s,e]$, $\Delta_{\mathcal{I}}=\sum_{\tau=s}^{e-1} \Delta(\tau)$ (note $\Delta_{[s,s]} = 0$) and $L_\mathcal{I}=1+\sum_{\tau=s}^{e-1}\mathbbm{1}[\Delta(\tau)\neq 0]$.
With slight abuse of notation, we write $\Delta=\Delta_{[1,T]}$ and $L=L_{[1,T]}$.
\end{definition}
\iffalse
Protocol~\ref{proto: main protocol} covers the following settings (not exhaustive):
\begin{itemize}
\item Multi-armed bandit: $\Pi=\mathcal{A}$ is a finite set of arms, and $f_t(a)$ is the expected reward of arm $a$.
\item Linear bandit: $\Pi$ is a convex set in $\mathbb{R}^d$, and $f_t(a)=\theta_t^\top a$ is a linear reward function of $a$.
\item Convex bandit: $\Pi$ is a convex set in $\mathbb{R}^d$, and $f_t(\cdot)$ is a concave reward function.
\item Contextual bandit: $\Pi\subseteq \mathcal{A}^{\mathcal{X}}$ is a set of mappings from contexts $\mathcal{X}$ to actions $\mathcal{A}$, and $f_t(\pi)=\mathbb{E}_{(x,r)\sim\mathcal{D}_t}[r(\pi(x))]$ is the expected reward of policy $\pi$ under $\mathcal{D}_t$, a joint distribution over (context, action reward) pair. We use $x$ to denote the context, and $r: \mathcal{A}\rightarrow \mathbb{R}$ to denote a mapping from actions to action rewards. (Contextual bandit usually allows the learner to directly choose actions without going through policies, making it not totally fit to Protocol~\ref{proto: main protocol}. However, this can solved by augmenting the policy set $\Pi$ with $|\mathcal{A}|$ policies each of which chooses a fixed action regardless of the context.)
\item Episodic MDP: $\Pi$ is the set of policies for the MDP, and $f_t(\pi)$ is the expected reward of the policy at episode $t$. (In episodic MDPs, for each $t$, the learner walks through a sequence of states and chooses a sequence of actions. But in this work, we view the whole episode as \emph{one round}.)
\end{itemize}
\fi
\paragraph{Base algorithm and requirements}
As mentioned, our approach takes a base algorithm that tackles the problem when the environment is (near-)stationary, and turns it into another algorithm that can deal with non-stationary environments.
Throughout the paper, we denote the base algorithm by {\small\textsf{\textup{ALG}}}\xspace and assumes that it satisfies the following mild requirements when run alone.
\begin{assumption}
\label{assum:assump2}
{\small\textsf{\textup{ALG}}}\xspace outputs an auxiliary quantity $\widetilde{f}_t\in[0, 1]$ at the beginning of each round $t$.
There exist a non-stationarity measure $\Delta$ and a non-increasing function $\rho: [T] \to \mathbb{R}$ such that running {\small\textsf{\textup{ALG}}}\xspace satisfies the following: for all $t\in[T]$, as long as $\Delta_{[1,t]}\leq \rho(t)$, without knowing $\Delta_{[1,t]}$ {\small\textsf{\textup{ALG}}}\xspace ensures with probability at least $1-\frac{\delta}{T}$:
\begin{align}
&\widetilde{f}_t \geq \min_{\tau\in [1,t]} f_\tau^\star - \Delta_{[1,t]}
\qquad\text{and}\qquad
\frac{1}{t}\sum_{\tau=1}^t \left(\widetilde{f}_\tau - R_\tau\right) \leq \rho(t) + \Delta_{[1,t]}. \label{eq: general condition 2}
\end{align}
Furthermore, we assume that $\rho(t)\geq \frac{1}{\sqrt{t}}$ and $C(t)= t\rho(t)$ is a non-decreasing function.
\end{assumption}
We unpack the meaning of this assumption and explain why this is a mild requirement via a few remarks below, followed by examples of existing algorithms that do satisfy our assumption.
First, consider choosing $\Delta(t) = \max_{\pi\in\Pi} |f_{t}(\pi) - f_{t+1}(\pi)|$ and see what the assumption means for a stationary environment with $f_t= f$ and $\Delta(t) = 0$ for all $t$.
In this case, \prettyref{eq: general condition 2} simply becomes
$\widetilde{f}_t \geq \max_{\pi \in \Pi} f(\pi)$ and $\sum_{\tau=1}^t \left(\widetilde{f}_\tau - R_\tau\right) \leq C(t)$,
which are standard properties of \emph{Upper-Confidence-Bound (UCB)-based} algorithms, where $\widetilde{f}_t$ is an \emph{optimistic} estimator of the optimal reward and $C(t)$ is the regret bound usually of order $\sqrt{t}$.
In fact, even for non-UCB-based algorithms that do not explicitly maintain optimistic estimators,
by looking into their analysis, it is still possible to extract a quantity $\widetilde{f}_t$ satisfying these two properties (see our example for contextual bandits in \prettyref{app: verify example}).
We also note that this requirement for the special stationary case is in fact all we need to achieve our claimed regret bound $\Reg_L^\star\xspace$.
Second, to simultaneously achieve the regret bound $\Reg_\Delta^\star\xspace$ as well, we require {\it slightly more} from the base algorithm: in a {\it near-stationary} environment with $\Delta_{[1,t]}\leq \rho(t)$, the two aforementioned properties still hold approximately with degradation $\Delta_{[1,t]}$ (that is, \prettyref{eq: general condition 2}).\footnote
We use $\min_{\tau\in [1,t]} f_\tau^\star$ instead of the more natural one $f_t^\star$ since the former is weaker and the difference between these two is at most $\Delta_{[1,t]}$ anyway.
}
We call this a near-stationary environment because $\Delta_{[1,t]}$ can be of order $\Theta(t)$ in a highly non-stationary environment, while here we restrict it to be at most
$\rho(t)$, which is non-increasing in $t$ (and in fact of order $1/\sqrt{t}$ in all our examples).
To the best of our knowledge, all UCB-based algorithms satisfy \prettyref{assum:assump2} with some suitable choice of $\Delta$. The fact that we only require \prettyref{eq: general condition 2} to hold for near-stationary environments is the key to bypassing the technical difficulty of getting the optimal bound $\Reg_\Delta^\star\xspace$ met in \citep{russac2019weighted, pmlr-v108-zhao20a, russac2020algorithms, faury2021regret, touati2020efficient, zhou2020nonstationary} for linear bandits, generalized linear bandits, and linear MDPs, as mentioned in \prettyref{sec: intro}.
Finally, noting that $\rho(t)$ and $C(t)$ represent an average and an cumulative regret bound respectively, the monotonicity requirement on them is more than natural.
The requirement $\rho(t)\geq \frac{1}{\sqrt{t}}$ is also usually unavoidable without further structures in the problem.
Note that while we write $\rho$ and $C$ as a function of $t$ only, they can depend on $\log(1/\delta)$, $\log T$, the complexity of $\Pi$, and other problem-dependent parameters such as the number of states/actions of an MDP.
Following the order in \prettyref{tab:compare}, we now give a list of existing algorithms that satisfy \prettyref{assum:assump2} in different problem settings with proper non-stationarity measure $\Delta$ and regret bound $C$.
We defer the concrete form of $\widetilde{f}_t$ (which requires introducing other notations) and all the proofs to \prettyref{app: verify example}.
\begin{itemize}
\item UCB1~\citep{auer2002finite}: $C(t)=\widetilde{\mathcal{O}}(\sqrt{At} + A)$ and $\Delta(t)=\Theta(\|r_t - r_{t+1}\|_\infty)$, where $A$ is the number of arms, and $r_t$ is the expected reward vector at time $t$.
\item ILTCB~\citep[short for ILOVETOCONBANDITS]{agarwal2014taming}:
$C(t)=\widetilde{\mathcal{O}}(\sqrt{At\log|\Pi|}+A\log|\Pi|)$ and $\Delta(t)=\Theta\left( \int_r \int_x |\mathcal{D}_t(x,r) - \mathcal{D}_{t+1}(x,r)| \mathrm{d}x\mathrm{d}r \right)$, where $A$ is the number of actions and $\mathcal{D}_t$ is the joint distribution of the context-reward pair $(x,r)$ at time $t$.
\item FALCON~\citep{simchi2020bypassing}: $C(t)=\widetilde{\mathcal{O}}(\sqrt{At\log|\Phi|} + A\log|\Phi|)$ and $\Delta(t)=\Theta(\sqrt{A}\max_{x,a} |\phi_t^\star(x,a) - \phi_{t+1}^\star(x,a)| + \int_x |\mathcal{D}_t(x) - \mathcal{D}_{t+1}(x)|\mathrm{d}x)$, where $A$ is the number of actions, $\Phi$ is the set of regressors (each of which maps a context-action pair to a predicted reward), $\phi^\star_t \in \Phi$ is the true regressor at time $t$, and $\mathcal{D}_t$ is the distribution of contexts at time $t$.
\item OFUL~\citep{abbasi2011improved}: $C(t)=\widetilde{\mathcal{O}}\left(d\sqrt{t}\right)$ and $\Delta(t)=\widetilde{\Theta}(d\|\theta_t-\theta_{t+1}\|_2)$, where $d$ is the feature dimension and $\theta_t \in\mathbb{R}^d$ parameterizes the linear reward function at time $t$.
\item GLM-UCB~\citep{filippi2010parametric}: $C(t)=\widetilde{\mathcal{O}}\left(\frac{k_\mu d}{c_\mu}\sqrt{t}\right)$ and $\Delta(t)=\widetilde{\Theta}\left(\frac{k_\mu^2 d}{c_\mu}\|\theta_t-\theta_{t+1}\|_2\right)$, where $d$ is the feature dimension, $\theta_t \in\mathbb{R}^d$ parameterizes the linear reward function at time $t$, and $k_\mu, c_\mu$ are the upper and lower bounds of the gradient of the link function.
\item Q-UCB~\citep[short for Q-learning UCB-H]{jin2018q}:\footnote
For ease of comparison, here, the reward range is changed from $[0,1]$ to the more common $[0,H]$. \label{fn:scaling1}
}
$C(t)=\widetilde{\mathcal{O}}(\sqrt{H^5SAt} + H^3SA)$ and $\Delta(t)=\Theta(H\sum_{h=1}^H \max_{s,a}|r_h^{t}(s,a)-r_{h}^{t+1}(s,a)| + H^2\sum_{h=1}^H \max_{s,a}\|p_h^{t}(\cdot|s,a)-p_{h}^{t+1}(\cdot|s,a)\|_1)$, where $H$, $S$ and $A$ are the numbers of layers, states, and actions of the MDP respectively, and
$p_h^{t}$ and $r_h^{t}$ are the transition and reward functions for layer $h$ of episode $t$.
\item LSVI-UCB~\citep{jin2020provably}:\footnote
Same as \prettyref{fn:scaling1}.
} $C(t)=\widetilde{\mathcal{O}}(\sqrt{d^3 H^4 t})$ and $\Delta(t)=\widetilde{\Theta}(dH\sum_{h=1}^H \|\theta_h^{t}-\theta_{h}^{t+1}\|_2 + dH^2\sum_{h=1}^H \|\mu_h^{t}-\mu_{h}^{t+1}\|_{F})$, where $d$ is the feature dimension, $H$ is the number of layers, and $\theta_h^{t}$ and $\mu_h^{t}$ are the parameters of the linear MDP for layer $h$ of episode $t$.
\end{itemize}
\subsection{Main results}\label{sec:main_results}
Our main result is that, with an algorithm satisfying \prettyref{assum:assump2} at hand, our proposed black-box reduction, {\small\textsf{\textup{MASTER}}}\xspace (\prettyref{alg: final adaptive alg}), ensures the following dynamic regret bound.
\begin{theorem}\label{thm: regret bound}
If \prettyref{assum:assump2} holds with $C(t)= c_1t^p + c_2$ for some $p\in[\frac{1}{2}, 1)$ and $c_1, c_2 >0$, then {\small\textsf{\textup{MASTER}}}\xspace (\prettyref{alg: final adaptive alg}), without knowing $L$ and $\Delta$, guarantees with high probability:
\begin{align*}
\textsc{D-Reg}\xspace = \widetilde{\mathcal{O}}\left( \min\left\{\left(c_1 + \frac{c_2}{c_1}\right)\sqrt{LT}, \; \left(c_1^{\nicefrac{2}{3}} + c_2c_1^{-\nicefrac{4}{3}}\right)\Delta^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}} + \left(c_1 + \frac{c_2}{c_1}\right)\sqrt{T} \right\} \right)
\end{align*}
when $p=\frac{1}{2}$, and
$
\textsc{D-Reg}\xspace= \widetilde{\mathcal{O}}\left(\min\left\{c_1 L^{1-p} T^{p}, \;\left(c_1\Delta^{1-p}T\right)^{\frac{1}{2-p}} + c_1T^{p}\right\}\right)
$
when $p>\frac{1}{2}$ (omitting some lower-order terms).
\end{theorem}
For ease of presentation, in this theorem we assume that $C(\cdot)$ takes a certain form that is common in the literature and holds for all our examples with $p=\frac{1}{2}$.
Applying this theorem to all the examples discussed earlier, we achieve all the optimal $\min\{\Reg_L^\star\xspace, \Reg_\Delta^\star\xspace\}$ bounds listed in \prettyref{tab:compare} (except for infinite-horizon MDPs which will be discussed in \prettyref{sec: average-reward}).
Our definitions of $L$ are the same as in previous works,
and our definitions of $\Delta$ are sometimes larger by some problem-dependent factors (such as $d$ and $H$) in order to fit \prettyref{assum:assump2}.
More specifically, for (contextual) bandits, our {\small\textsf{\textup{MASTER}}}\xspace combined with UCB1 and ILTCB recovers the same optimal bounds (in terms of all parameters) achieved by~\citep{auer2019adaptively, chen2019new}.
{\small\textsf{\textup{MASTER}}}\xspace with FALCON obtains a similar bound as in \citep{chen2019new} but with a different definition of $\Delta$ specific to the regressor setting.
For other settings, we present our results in terms of the common definition of the non-stationarity measure (denoted by $\widehat{\Delta}$) and compare them with the state of the art:
\begin{itemize}
\item {\small\textsf{\textup{MASTER}}}\xspace + OFUL: $\textsc{D-Reg}\xspace=\widetilde{\mathcal{O}}( \min\{d\sqrt{LT}, d\widehat{\Delta}^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}+d\sqrt{T}\})$, where $\widehat{\Delta} = \sum_t \|\theta_t-\theta_{t+1}\|_2$.
This improves~\citep{cheung2019learning, russac2019weighted, kim2019randomized, pmlr-v108-zhao20a, zhao2021nonstationary} which get $\widetilde{\mathcal{O}}(d^{\nicefrac{7}{8}}\widehat{\Delta}^{\nicefrac{1}{4}}T^{\nicefrac{3}{4}} + d\sqrt{T})$ when $\widehat{\Delta}$ is known.
\item {\small\textsf{\textup{MASTER}}}\xspace + GLM-UCB: $\textsc{D-Reg}\xspace=\widetilde{\mathcal{O}}\Big( \min\Big\{\frac{k_\mu}{c_\mu}d\sqrt{LT}, \frac{k_\mu^{\nicefrac{4}{3}}}{c_\mu}d\widehat{\Delta}^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}+\frac{k_\mu}{c_\mu}d\sqrt{T}\Big\}\Big)$, where $\widehat{\Delta} = \sum_t \|\theta_t-\theta_{t+1}\|_2$.
This improves~\citep{russac2020algorithms} which gets $\widetilde{\mathcal{O}}\big(\frac{k_\mu}{c_\mu}d^{\nicefrac{2}{3}}L^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}\big)$ when $L$ is known, and \citep{faury2021regret} which gets $\widetilde{\mathcal{O}}\big(\frac{k_\mu}{c_\mu}d^{\nicefrac{9}{10}}\widehat{\Delta}^{\nicefrac{1}{5}}T^{\nicefrac{4}{5}}\big)$.
\item {\small\textsf{\textup{MASTER}}}\xspace + Q-UCB:
$\textsc{D-Reg}\xspace=\widetilde{\mathcal{O}}(\min\{\sqrt{H^5SALT}, (H^{7}SA\widehat{\Delta})^{\nicefrac{1}{3}} T^{\nicefrac{2}{3}} + \sqrt{H^5SAT}\})$, where $\widehat{\Delta} = \sum_{t,h}\max_{s,a}(|r_h^{t}(s,a)-r_{h}^{t+1}(s,a)| + \|p_h^{t}(\cdot|s,a)-p_{h}^{t+1}(\cdot|s,a)\|_1)$.\footnote
Due to the scaling mentioned in \prettyref{fn:scaling1},
here, we first scale down $C(\cdot)$ and $\Delta$ by an $H$ factor, then apply \prettyref{thm: regret bound}, and finally scale up the final bound by an $H$ factor. \label{fn:scaling2}
}
\citep[Theorem 3]{mao2020nearoptimal} gets $\widetilde{\mathcal{O}}((H^5SA\widehat{\Delta})^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}} + \sqrt{H^3SAT})$ when $\widehat{\Delta}$ is known.\footnote{The bound reported in \citep{mao2020nearoptimal} is $(H^3SA\widehat{\Delta})^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}} + \sqrt{H^2SAT}$; however, their $T$ is the total number of timesteps while our $T$ is the number of episodes, and we have performed a proper translation between notations here. Their bound has a better $H$ dependency thanks to the use of Freedman-style confidence bounds.
The same idea unfortunately does not improve our bound due to the lower-order term $c_2$ in the definition of $C(t)$. \label{fn:translation1} }
\item {\small\textsf{\textup{MASTER}}}\xspace + LSVI-UCB:
$\textsc{D-Reg}\xspace=\widetilde{\mathcal{O}}(\min\{\sqrt{d^3H^4 LT}, (d^4 H^{6}\widehat{\Delta})^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}} + \sqrt{d^3 H^4 T}\})$, where $\widehat{\Delta} = \sum_{t, h} (\|\theta_h^{t}-\theta_{h}^{t+1}\|_2 + \|\mu_h^{t}-\mu_{h}^{t+1}\|_{F})$.
This improves~\citep{zhou2020nonstationary, touati2020efficient} which get $\widetilde{\mathcal{O}}((d^5 H^8 \widehat{\Delta})^{\nicefrac{1}{4}}T^{\nicefrac{3}{4}}+ \sqrt{d^3H^4T})$ when $\widehat{\Delta}$ is known.\footnote{The same scaling as in \prettyref{fn:scaling2} and \prettyref{fn:translation1} has been performed here.}
\end{itemize}
\iffalse
\renewcommand{\arraystretch}{1.2}
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Algorithm & $C(t)$ in $\widetilde{\mathcal{O}}\left(\cdot\right)$ & $\Delta_{[t,t+1]}$ in $\Theta(\cdot)$ & Dynamic Regret in $\widetilde{\mathcal{O}}\left(\cdot\right)$\\
\hline
UCB1 & $\sqrt{Kt} + K$ & $\|r_t - r_{t+1}\|_\infty$ & $\min\{\sqrt{KLT}, (K\Delta)^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}+\sqrt{KT}\}$ \\
\hline
OFUL & $d\sqrt{t}$ & $\|\theta_t - \theta_{t+1}\|_2$ & $\min\{d\sqrt{LT}, d^{\nicefrac{2}{3}}\Delta^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}\}$ \\
\hline
ILTCB & \makecell{$\sqrt{K't} + K'$ \\ ($K'\triangleq K\log|\Pi|$)} & $\|\mathcal{D}_t-\mathcal{D}_{t+1}\|_{\text{TV}}$ & $\min\{\sqrt{K'LT}, (K'\Delta)^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}+\sqrt{K'T}\}$ \\
\hline
FALCON & \makecell{$\sqrt{K't} + K'$ \\ ($K'\triangleq K\log|\Pi|$)} & \makecell{$\|\mathcal{D}_t-\mathcal{D}_{t+1}\|_{\text{TV}}$ \\ $ + \|\phi^\star_t - \phi^\star_{t+1}\|$} & $\min\{\sqrt{K'LT}, (K'\Delta)^{\nicefrac{1}{3}}T^{\nicefrac{2}{3}}+\sqrt{K'T}\}$ \\
\hline
Tabular MDP & $S\sqrt{AT} + S^2A$ &
\end{tabular}
\caption{}
\label{tab:compare}
\end{table}
\fi
\iffalse
Below we list some algorithms that satisfy \prettyref{assum:assump2} in various problems:
\begin{itemize}
\item UCB1~\citep{auer2002finite} for multi-armed bandit: $C(t)=\widetilde{\mathcal{O}}\left(\sqrt{At} + A\right)$, $\Delta_{[t,t+1]}=\|r_t - r_{t+1}\|_\infty$, where $A$ is the number of arms, and $r_t$ is the expected reward vector at time $t$.
\item OFUL~\citep{abbasi2011improved} for linear bandit (for a fixed action set): $C(t)=\widetilde{\mathcal{O}}\left(d\sqrt{t}\right)$, $\Delta_{[t,t+1]}=\|\theta_t-\theta_{t+1}\|_2$, where $d$ is the feature dimension, and $\theta_t$ is the reward vector at time $t$.
\item ILOVETOCONBANDITS~\citep{agarwal2014taming} for contextual bandit: $C(t)=\widetilde{\mathcal{O}}\left(\sqrt{At\log|\Pi|}+A\log|\Pi|\right)$, $\Delta_{[t,t+1]}=\Theta\left( \int_r \int_x |\mathcal{D}_t(x,r) - \mathcal{D}_{t+1}(x,r)| \mathrm{d}x\mathrm{d}r \right)$, where $A$ is the number of actions, $\Pi$ is the policy set, $\mathcal{D}_t$ is the joint distribution of the context-reward pair at time $t$.
\item FALCON~\citep{simchi2020bypassing} for realizable contextual bandit: $C(t)=\widetilde{\mathcal{O}}\left(\sqrt{At\log|\Phi|} + A\log|\Phi|\right)$, $\Delta_{[t,t+1]}=\Theta\left(\sqrt{A}\max_{x,a} |\phi_t^\star(x,a) - \phi_{t+1}^\star(x,a)| + \int_x |\mathcal{D}_t(x) - \mathcal{D}_{t+1}(x)|\mathrm{d}x\right)$, where $A$ is the number of actions, $\Phi$ is the set of regressor functions, $\phi^\star_t$ is the ground-truth regressor at time $t$.
\item UCRL~\citep{jaksch2010near}, UCBVI~\citep{azar2017minimax}, UCB-Q~\citep{jin2018q} for episodic MDP: $C(t)=\text{poly}(H,S,A)\sqrt{t}$, $\Delta_{[t,t+1]}=\max_{s,a}|r_t(s,a)-r_{t+1}(s,a)| + H\max_{s,a}\|p_t(\cdot|s,a)-p_{t+1}(\cdot|s,a)\|_1$ where $H$ is the episode length, $S$ and $A$ are number of states and actions respectively, $p_t, r_t$ are the transition kernel and reward function at episode $t$ respectively.
\end{itemize}
We verify them in Appendix~\ref{app: verify example}.
\fi
|
\section{Introduction}
Adherence to medication{\color{black} s is the process by which patients take their medications as prescribed} \cite{vrijens2012}. It is well-documented that nonadherence is a major problem, resulting in over 100,000 preventable deaths and \$100 billion in preventable health care costs per year in the United States alone \cite{osterberg2005}. In fact, the World Health Organization noted that ``increasing the effectiveness of adherence interventions may have a far greater impact on the health of the population than any improvement in specific medical treatments'' \cite{who2003, haynes2002}. Nonadherence is especially prevalent and problematic in patients with chronic diseases that require long term pharmacotherapy \cite{burnier2019}. As former US surgeon general C Everett Koop famously observed, ``Drugs don't work in patients who don't take them'' \cite{lindenfeld2017}.
{\color{black} Medication adherence has been divided into the three phases of initiation, implementation, and discontinuation \cite{vrijens2012}. Initiation and discontinuation refer to a patient starting and stopping a regimen as prescribed (and the term ``persistence'' describes the time from initiation to discontinuation \cite{vrijens2012, geest2018}). In this paper, we focus on implementation, which is the extent to which a patient's actual dosing follows the prescribed dosing regimen \cite{vrijens2012}.}
Many outstanding questions surround {\color{black} the implementation phase of} adherence and how it relates to therapeutic outcomes. Adherence is often reported as the percentage of doses of medication actually taken by the patient over a specified
time \cite{osterberg2005}. How does an adherence percentage $p$ translate into treatment efficacy? How much worse is, for example, $p=70\%$ compared to $p=85\%$? How much adherence is needed for full treatment benefits? How can clinicians increase patient adherence? Are there protocols to increase treatment benefits in spite of poor adherence?
While the causes of nonadherence vary, a significant portion of nonadherence stems from patients simply forgetting to take their medication \cite{spilker1991, barfod2006}. What should a patient do if they miss a dose of medication? Although patients commonly ask this question, they often do not receive adequate instructions for what to do when a dose is missed \cite{howard1999, albassam2020, gilbert2002}.
To address these questions, we formulate and analyze a mathematical model of the drug {\color{black} concentration} in a patient with imperfect adherence. Mathematical modeling is especially well-suited to investigate these questions, given the ethics of clinical trials that force patients to miss doses of medication. To model imperfect adherence, we assume that the patient takes their medication at only a given percentage of the prescribed dosing times. Doses are missed at random, and thus the drug {\color{black} concentration} in the body is random. {\color{black} For simplicity, we assume that the patient misses each dose with a fixed probability $1-p$, independent of their prior behavior.} We find exact mathematical formulas for statistics of this model, including the average drug {\color{black} concentration}, the drug {\color{black} concentration} coefficient of variation, and how the drug {\color{black} concentration} deviates from a patient with perfect adherence. These statistics are obtained as explicit functions of the adherence percentage, the drug half-life, and the prescribed dosing interval (i.e.\ the time between scheduled doses). Furthermore, we determine how these statistics depend on how the patient handles missed doses, including the case that they skip missed doses and the case that they take double doses following missed doses.
From a mathematical standpoint, the random variables that model the drug {\color{black} concentration} in our model generalize infinite Bernoulli convolutions \cite{peres2000, solomyak1995, peres1998, escribano2003, hu2008}. The study of infinite Bernoulli convolutions has a rich history in the pure mathematics literature, dating back to Erd\H{o}s and others in the 1930s \cite{jessen1935, kershner1935, erdos1939}. Infinite Bernoulli convolutions typically have very irregular distributions, including singular distributions supported on a Cantor set \cite{kershner1935}. Infinite Bernoulli convolutions also arose in the pharmacokinetic models in \cite{levy2013, fermin2017}. Our analysis of the generalized infinite Bernoulli convolutions that arise in our model relies on the theory of random pullback attractors \cite{Crauel01, Mattingly99, Schmalfuss96, lawley15sima, lawley2019hhg}.
From the standpoint of pharmacology, there are {\color{black} several results} of our analysis. First, we provide quantitative estimates of how an adherence percentage $p$ translates into statistics of drug {\color{black} concentration}s in the body, and how these statistics depend on the drug half-life $t_{\textup{half}}$, the dosing interval ${\tau}$, and how missed doses are handled. Further, these estimates show how the effects of nonadherence can be lessened by drugs with half-lives that are long compared to the dosing interval, i.e.\ $t_{\textup{half}}\gg {\tau}$. While clinical recommendations {\color{black} require extensive validation and} should depend on drug and patient specifics, as a general principle {\color{black} our theory suggests} that the effects of nonadherence are best mitigated by taking double doses following missed doses if $t_{\textup{half}}\gg {\tau}$, whereas missed doses should be skipped if $t_{\textup{half}}\ll {\tau}$. This conclusion contradicts some existing recommendations that cite long drug half-lives as the reason to avoid a double dose after a missed dose {\color{black}(for example, see recommendations for perampanel \cite{albassam2020} and valproate \cite{gilbert2002}), as well as the general recommendation that double doses should not be taken to compensate for missed doses \cite{siwale2016}}. Since double doses are sometimes avoided due to concern that they may cause toxic drug {\color{black} concentration}s, we provide an upper bound for the highest possible drug {\color{black} concentration} in the body. We find that a patient who takes double doses after missed doses can {\color{black} have at most only a slightly higher drug concentration (and exposure) than} a perfectly adherent patient if $t_{\textup{half}}\gg {\tau}$. We also investigate other ways of handling missed doses, including taking an extra half dose following a missed dose, which we find is most appropriate when $t_{\textup{half}}\approx {\tau}$.
The rest of the paper is organized as follows. We formulate and analyze the mathematical model in the Methods section {\color{black}(details of the mathematical analysis are in the Appendix)}. In the Results section, we explore the pharmacological implications of the mathematical analysis. Since these pharmacological implications depend on rather complicated mathematics, we also provide an intuitive explanation for our results in this section. The Discussion section concludes by describing related work, model limitations, and future directions. We also discuss our results in the context of hypothyroid patients taking levothyroxine. The Appendix collects some technical points and the proofs of the theorems.
\section{Methods}
\subsection{Mathematical model}\label{model}
Our model builds on the classical pharmacokinetic model of extravascular (oral) administration in a single compartment with first order kinetics \cite{gibaldi1982, bauer2015}. In the standard model, the drug concentration, $c_{0}$, in the body at time $s>0$ satisfies the ordinary differential equation (ODE),
\begin{align}\label{code}
\frac{\text{d} c_{0}}{\text{d} s}
=k_{\textup{a}}\frac{g}{V}-k_{\textup{e}} c_{0},
\end{align}
where $k_{\textup{a}}$ and $k_{\textup{e}}$ are the respective rates of absorption and elimination, $V$ is the volume of distribution, and $g$ is the drug amount at the absorption site. The amount $g$ satisfies the ODE,
\begin{align}\label{gode}
\frac{\text{d} g}{\text{d} s}
=-k_{\textup{a}} g+I(s),
\end{align}
where $I(s)$ describes the drug input.
For most drugs administered extravascularly in conventional dosage forms, the absorption rate is much larger than the elimination rate, meaning $k_{\textup{a}}\ggk_{\textup{e}}$ (see \cite{gibaldi1982, peletier2017, ma2018, fillastre1987, leroy1990, strandgaarden1999}). In this parameter regime, the solution of \eqref{code} is well-approximated by the solution to
\begin{align}\label{simpler}
\frac{\text{d} c}{\text{d} s}
=\frac{I(s)}{V}-k_{\textup{e}} c,
\end{align}
which is the standard model for intravascular administration with first order elimination. In this paper, we assume $k_{\textup{a}}\ggk_{\textup{e}}$ and thus consider the simpler model in \eqref{simpler} rather than the system in \eqref{code}-\eqref{gode}.
\subsubsection{Perfect adherence}
Suppose a patient is instructed to take a dose of size $D>0$ at regular time intervals of length $\tau>0$ beginning at time $0$. If the patient has perfect adherence, then the drug input is
\begin{align}\label{iperf}
I^{\textup{perf}}(s)
=DF\sum_{n\ge0}\delta(t-n\tau),
\end{align}
where $F\in(0,1]$ is the bioavailability fraction and $\delta$ denotes the Dirac delta function. Solving \eqref{simpler}-\eqref{iperf} yields the following well-known formula for the drug concentration at time $s\ge0$ in the perfectly adherent patient \cite{bauer2015},
\begin{align}\label{cstart}
c^{\textup{perf}}(s)
:=
\frac{DF}{V}\sum_{n=0}^{N(s)}e^{-k_{\textup{e}}(s-n\tau)},
\end{align}
where $N(s)+1$ is the number of dosing times elapsed by time $s$,
\begin{align*}
N(s)
:=\sup\{n\ge0:n\le s/\tau\}.
\end{align*}
If
\begin{align*}
t=s-N(s)\tau\in[0,\tau)
\end{align*}
denotes the time elapsed since the most recent dosing time, then \eqref{cstart} can be written as
\begin{align}\label{cform}
c^{\textup{perf}}(s)
=
\alpha^{t/\tau}\frac{DF}{V}\sum_{n=0}^{N(s)}\alpha^{n},
\end{align}
where we have defined the dimensionless constant
\begin{align}\label{alphabeta}
\alpha:=e^{-k_{\textup{e}}{\tau}}\in(0,1),
\end{align}
which is the fraction of a dose that remains in the body after one dosing interval.
If the patient continues their perfect adherence for a long time, then it is easy to see from the form in \eqref{cform} that the drug concentration approaches the following function,
\begin{align}\label{perfper}
c^{\textup{perf}}(N\tau+t)
\to C^{\textup{perf}}(t)
:=\alpha^{t/\tau}\frac{DF}{V}A^{\textup{perf}}\quad\text{as }N\to\infty,
\end{align}
where $t\in[0,\tau)$ is the time since the last dose and
\begin{align*}
A^{\textup{perf}}:=
\sum_{n\ge0}\alpha^{n}
=\frac{1}{1-\alpha}.
\end{align*}
In pharmacokinetics, it is common to measure the drug exposure over a single dosing interval by the so-called ``area under the curve,'' which for this case of perfect adherence is
\begin{align}\label{aucperf}
\textup{AUC}^{\textup{perf}}
:=\int_{0}^{\tau}C^{\textup{perf}}(t)\,\text{d} t
=\frac{DF}{V}\frac{1}{k_{\textup{e}}}.
\end{align}
\subsubsection{Nonadherence}
To model patient nonadherence, we suppose that the patient occasionally misses a dose. Specifically, at each dosing time, the patient ``remembers'' to take their medication with probability $p\in(0,1)$, and the patient ``forgets'' with probability $1-p$. Mathematically, let $\{\xi_{n}\}_{n}$ be a sequence of independent and identically distributed (iid) Bernoulli random variables with parameter $p$, meaning
\begin{align}\label{xin}
\begin{split}
\xi_{n}
=\begin{cases}
1 & \text{with probability }p,\\
0 & \text{with probability }1-p.
\end{cases}
\end{split}
\end{align}
Hence, $\xi_{n}=1$ means that the patient takes their medication at the $n$th dosing time. {\color{black} We emphasize that $\{\xi_{n}\}_{n}$ is a sequence of independent random variables, which means that the patient misses doses independently of their prior behavior.}
If $Df_{n}\ge0$ denotes the amount taken at the $n$th dosing time, then the drug input is
\begin{align}\label{iimp}
I(s)
=DF\sum_{n\ge0}\delta(t-n\tau)f_{n},
\end{align}
and solving \eqref{simpler} with $I(s)$ in \eqref{iimp} yields the drug concentration in the patient,
\begin{align}\label{CC0}
c(s)
=
\alpha^{t/\tau}\frac{DF}{V}\sum_{n=0}^{N(s)}\alpha^{N(s)-n}f_{n}.
\end{align}
Notice that \eqref{CC0} reduces to \eqref{cform} if $f_{n}=1$ for all $n$. We take
\begin{align*}
f_{n}=0,\quad\text{if }\xi_{n}=0,
\end{align*}
which means the patient does not take any medication when they forget. However, we allow for the possibility that
\begin{align*}
f_{n}> 1,\quad \text{if }\xi_{n}=1,
\end{align*}
which means that the patient may take more than a single dose to make up for prior missed doses. In general, we allow $f_{n}$ to be a function of the history $\{\xi_{i}\}_{i\le n}$, and we refer to a choice of $f_{n}$ as a ``dosing protocol.''
The simplest dosing protocol is for the patient to merely take a single dose if they remember, which means
\begin{align}
\label{fsingle}
f_{n}^{\textup{single}}
:=
\begin{cases}
0 & \text{if }\xi_{n}=0,\\
1 & \text{if }\xi_{n}=1.
\end{cases}
\end{align}
We refer to \eqref{fsingle} as the ``single dose'' protocol. Another common dosing protocol is for the patient to take a double dose to make up for a missed dose at the prior dosing time, which means
\begin{align}
\begin{split}\label{fdouble}
f_{n}^{\textup{double}}
:=\begin{cases}
0 & \text{if }\xi_{n}=0,\\
1 & \text{if }\xi_{n}=1,\,\xi_{n-1}=1,\\
2 & \text{if }\xi_{n}=1,\,\xi_{n-1}=0.
\end{cases}
\end{split}
\end{align}
We refer to \eqref{fdouble} as the ``double dose'' protocol. Notice that in the double dose protocol, the patient never takes more than two doses at a time, even if they missed more than one previous dose. Our analysis below covers other dosing protocols, but we are primarily interested in comparing the single dose and double dose protocols in \eqref{fsingle}-\eqref{fdouble}. As a technical aside, we are ultimately interested in the large time behavior of $c(s)$ in \eqref{CC0}, and thus the values of $f_{n}$ in \eqref{CC0} for small $n$ are irrelevant. In particular, the fact that the definition of $f_{0}^{\textup{double}}$ in \eqref{fdouble} depends on $\xi_{-1}$ is immaterial.
\begin{figure}[t
\centering
\includegraphics[width=\linewidth]{Fig1.pdf}
\caption{(a) The black dotted curve depicts how the drug concentration in a perfectly adherent patient evolves in time, and the gray shaded region is the area between the peaks and troughs of the perfectly adherent patient. The red dashed curve and blue solid curve describe patients with imperfect adherence following the single dose and double dose protocols, respectively. The blue dashed line depicts the largest possible drug concentration for the double dose protocol (see {\color{black}\eqref{largeste}}). (b) The distribution of the relative drug {\color{black} concentration} $Z=\textup{AUC}/\textup{AUC}^{\textup{perf}}=C(t)/C^{\textup{perf}}(t)$ for the single dose protocol (red) and the double dose protocol (blue) obtained from stochastic simulations. The black dotted vertical line at $Z=1$ describes the perfectly adherent patient, and the blue dashed vertical line describes the largest possible drug {\color{black} concentration} for the double dose patient (see {\color{black}\eqref{largeste}}). In both plots, $p=0.8$ and $\alpha=0.9$.}
\label{figsim0}
\end{figure}
Figure~\ref{figsim0}a illustrates how the drug {\color{black} concentration} in the body evolves in time. The black dotted curve describes the perfectly adherent patient, and the red dashed curve and blue solid curve describe patients with imperfect adherence following the single dose and double dose protocols, respectively. We set the initial drug concentration equal to $C^{\textup{perf}}(0)$ in this illustration.
\subsubsection{Large-time drug {\color{black} concentration} distribution}
For the case of perfect adherence, the drug concentration at large time is described by $C^{\textup{perf}}(t)$ in \eqref{perfper}. Analogously, for the case of imperfect adherence, we prove below that the drug concentration converges in distribution at large time,
\begin{align}\label{Cearly}
c(N\tau+t)
\to_{\text{d}}
C(t)
\quad\text{as }N\to\infty,
\end{align}
where $t\in[0,\tau)$ is the time since the last dosing time and $C(t)$ is a certain random function given below. In particular, $C(t)$ describes the drug concentration in a patient who has been taking the drug for a long time with adherence $p\in(0,1)$. Furthermore, the patient's drug exposure over a dosing interval is
\begin{align}\label{auc}
\textup{AUC}
:=\int_{0}^{\tau}C(t)\,\text{d} t.
\end{align}
We emphasize that $C(t)$ and $\textup{AUC}$ are random since patient adherence is modeled by a random process.
\subsubsection{The effects of nonadherence}
We measure the effects of nonadherence by comparing the {\color{black} drug concentration} in a patient with imperfect adherence to the drug {\color{black} concentration} in a patient with perfect adherence. It is natural to quantify this in terms of the drug exposure ratio $\textup{AUC}/\textup{AUC}^{\textup{perf}}$ or the drug concentration ratio $C(t)/C^{\textup{perf}}(t)$ {\color{black} at some time $t\in[0,\tau)$ since the last scheduled dose}. It turns out that these two ratios are the same, as we prove below that
\begin{align}\label{Z}
Z:=
\frac{\textup{AUC}}{\textup{AUC}^{\textup{perf}}}
=\frac{C(t)}{C^{\textup{perf}}(t)}
,\quad\text{for all }t\in[0,\tau).
\end{align}
{\color{black} We therefore emphasize that we} study the effects of nonadherence {\color{black} in terms of the relative drug exposure and the relative drug concentration} by studying the single random variable $Z$. Figure~\ref{figsim0}b plots the distribution of $Z$ for the single dose protocol (red) and the double dose protocol (blue) obtained from stochastic simulations (the distribution is obtained from $10^{7}$ realizations of $c(N\tau+t)$ with $N=100$).
We study $Z$ primarily in terms of the following three statistics. First, we define the mean,
\begin{align}\label{mu}
\mu
:=\mathbb{E}[Z]
=\frac{\mathbb{E}[\textup{AUC}]}{{\textup{AUC}^{\textup{perf}}}}
{\color{black}=\frac{\mathbb{E}[C(t)]}{C^{\textup{perf}}(t)},}
\end{align}
which compares the average drug {\color{black} concentration} to the perfectly adherent patient. We further define the deviation,
\begin{align}\label{deviation0}
\Delta
:=\sqrt{\mathbb{E}[(Z-1)^{2}]}
=\frac{\sqrt{\mathbb{E}\big[({\textup{AUC}}-{{{\textup{AUC}}^{\textup{perf}}}})^{2}\big]}}{{{{\textup{AUC}}^{\textup{perf}}}}}
{\color{black}=\frac{\sqrt{\mathbb{E}\big[(C(t)-{{C^{\textup{perf}}(t)}})^{2}\big]}}{{{C^{\textup{perf}}(t)}}}},
\end{align}
which measures how the drug {\color{black} concentration} deviate{\color{black} s} from the perfectly adherent patient. In statistics, \eqref{deviation0} is called the relative root-mean-square deviation or relative root-mean-square error. We also compute the coefficient of variation of $Z$, but we find that $\Delta$ is a better measure of the effects of nonadherence. Finally, since dosing protocols in which the patient takes more than a single dose at a time may cause drug {\color{black} concentration}s to rise too high, another useful statistic is the largest possible drug {\color{black} concentration} compared to the perfectly adherent patient,
\begin{align}\label{lambda0}
\lambda
:=\sup_{\xi}Z
{\color{black}=\sup_{\xi}\frac{\textup{AUC}}{\textup{AUC}^{\textup{perf}}}
=\sup_{\xi}\frac{C(t)}{C^{\textup{perf}}(t)},}
\end{align}
where $\sup_{\xi}$ denotes the supremum over patterns of the patient remembering or forgetting to take their medication (i.e. $\{\xi_{n}\}_{n}$ in \eqref{xin}). {\color{black} We emphasize that \eqref{lambda0} means that $\lambda$ bounds both the relative drug exposure (i.e.\ $\textup{AUC}$ to $\textup{AUC}^{\textup{perf}}$) and the relative drug concentration at any time (i.e.\ $C(t)$ to $C^{\textup{perf}}(t)$).}
We point out that the statistics $\mu$, $\Delta$, and $\lambda$ in \eqref{mu}-\eqref{lambda0} are dimensionless, and thus they are independent of the units used to measure drug amounts, concentrations, time, etc. Furthermore, these statistics depend only on $\alpha$ in \eqref{alphabeta}, the adherence percentage $p\in(0,1)$, and the dosing protocol $f_{n}$. {\color{black} We emphasize that since $\mu$, $\Delta$, and $\lambda$ are defined relative to the perfectly adherent patient, their values} are unchanged if $\textup{AUC}$ and $\textup{AUC}^{\textup{perf}}$ are replaced by $C(t)$ and $C^{\textup{perf}}(t)$ for any $t\in[0,\tau)$ {\color{black}(as indicated by \eqref{mu}-\eqref{lambda0})}.
{\color{black}
\subsubsection{Single and double dose protocols}
In the Appendix, we analyze the mathematical model and compute statistics of the drug {\color{black} concentration}s for general dosing protocols. Here, we present the formulas for $\mu$, $\Delta$, and $\lambda$ for the single and double dose protocols.
\begin{theorem}\label{maintext}
Using superscripts to denote the dosing protocol, the relative means in \eqref{mu} are
\begin{align*}
\mu^{\textup{single}}=p,\quad
\mu^{\textup{double}}=p+p(1-p).
\end{align*}
Similarly, the deviations in \eqref{deviation0} are
\begin{align}
\Delta^{\textup{single}}
&=\sqrt{\frac{1-p}{1+\alpha}}\sqrt{1-\alpha (2 p-1)},\label{es}\\
\Delta^{\textup{double}}
&=\sqrt{\frac{1-p}{1+\alpha}}\sqrt{1+p+(1-7 p+2 p^2)\alpha+2 p^2(2-p) \alpha ^2}.\label{ed}
\end{align}
Finally, the largest relative drug {\color{black} concentration}s in \eqref{lambda0} are
\begin{align}\label{largeste}
\lambda^{\textup{single}}=1,\quad
\lambda^{\textup{double}}=\frac{2}{1+\alpha}.
\end{align}
\end{theorem}
}
\section{Results}
We now explore some pharmacological implications of the analysis above. Recall that $\alpha:=e^{-k_{\textup{e}}\tau}$, where $\tau$ is the dosing interval and $k_{\textup{e}}$ is the drug elimination rate. Since elimination rates are often expressed in terms of half-lives, we note that the drug half-life, $t_{\textup{half}}>0$, is related to the other parameters via
\begin{align}\label{alphahalf}
\alpha
&=2^{-{\tau}/t_{\textup{half}}},
\quad
t_{\textup{half}}
=\bigg(\frac{\ln\frac{1}{2}}{\ln\alpha}\bigg){\tau}
=\frac{\ln2}{k_{\textup{e}}}.
\end{align}
Hence, in the following a ``long drug half-life'' means $t_{\textup{half}}$ is long compared to ${\tau}$, and thus $\alpha$ is large (i.e.\ $\alpha$ is near 1). Similarly, a ``short drug half-life'' means $t_{\textup{half}}$ is short compared to ${\tau}$, and thus $\alpha$ is small.
\subsection{Long half-lives reduce the effects of patient nonadherence}
\begin{figure}[t
\centering
\includegraphics[width=\linewidth]{Fig2.pdf}
\caption{Deviation $\Delta$ from perfect adherence. Panel (a) plots the deviation $\Delta^{\textup{single}}$ for the single dose protocol as a function of $\alpha$ for different adherence percentages $p$. Panel (b) compares the deviations $\Delta^{\textup{single}}$ and $\Delta^{\textup{double}}$ for the single and double dose protocols with $p=0.8$.
}
\label{fige}
\end{figure}
We begin by considering the single dose protocol. In Figure~\ref{fige}a, we plot $\Delta^{\textup{single}}$ as a function of $\alpha$ for different patient adherence levels $p$. As expected, $\Delta^{\textup{single}}$ decreases as the patient adherence $p$ increases. Furthermore, $\Delta^{\textup{single}}$ decreases as $\alpha$ increases, and $\Delta^{\textup{single}}$ approaches its minimum value as $\alpha\to1$,
\begin{align*}
\Delta^{\textup{single}}\to1-p\quad\text{as }\alpha\to1.
\end{align*}
These properties can be seen from Figure~\ref{fige}a and equation~\eqref{es}.
Importantly, Figure~\ref{fige}a shows that the effect of patient nonadherence, as measured by the deviation $\Delta^{\textup{single}}$ from perfect adherence, depends critically on $\alpha$. For example, notice that the horizontal line in Figure~\ref{fige}a at $\Delta^{\textup{single}}=0.25$ intersects the $\Delta^{\textup{single}}$ curves for the three different levels of patient adherence considered (namely, $p=0.85$, $p=0.8$, and $p=0.75$). Therefore, a patient with high adherence $p$ and small $\alpha$ and a patient with low adherence $p$ and large $\alpha$ can have the same deviation from the perfectly adherent patient.
Put another way, the effects of nonadherence can be lessened by increasing the $\alpha$ value of the drug (i.e.\ increasing the half-life $t_{\textup{half}}$ or decreasing the dosing interval ${\tau}$) without changing the patient's actual adherence $p$. This result is inline with previous analysis, as it is commonly noted that drugs with long half-lives tend to be more ``forgiving'' of missed doses \cite{osterberg2010}. This analysis thus quantifies drug ``forgiveness.'' For other measures of drug forgiveness, see \cite{assawasuwannakit2015, pellock2016, morrison2017}.
\subsection{Double dose protocol mitigates patient nonadherence for drugs with long half-lives}\label{compare}
To compare the deviations from perfect adherence for the single dose and double dose protocols, in Figure~\ref{fige}b we plot $\Delta^{\textup{single}}$ and $\Delta^{\textup{double}}$ as functions of $\alpha$. This figure shows that
\begin{align}
\Delta^{\textup{single}}
&<\Delta^{\textup{double}}\quad\text{if $\alpha$ is small (i.e.\ short half-life)},\nonumber\\
\text{and}\quad\Delta^{\textup{double}}
&<\Delta^{\textup{single}}\quad\text{if $\alpha$ is large (i.e.\ long half-life)}.\label{want}
\end{align}
We set $p=0.8$ in Figure~\ref{fige}b, but other values of $p$ yield similar results. Indeed, the formulas in \eqref{es} and \eqref{ed} imply the small $\alpha$ limits,
\begin{align*}
\lim_{\alpha\to0}\Delta^{\textup{single}}
=\sqrt{1-p}
< \lim_{\alpha\to0}\Delta^{\textup{double}}
=\sqrt{1-p^{2}},
\end{align*}
and the large $\alpha$ limits,
\begin{align}\label{ll}
\lim_{\alpha\to1}\Delta^{\textup{double}}
=(1-p)^{2}
<\lim_{\alpha\to1}\Delta^{\textup{single}}
=1-p.
\end{align}
In practical terms, \eqref{ll} means that if $\alpha$ is large and the patient has adherence of $p=0.9$, then the deviation from perfect adherence is roughly 10 times smaller for the double dose protocol compared to the single dose protocol.
While we have shown \eqref{want} for large $\alpha$, it follows from \eqref{es}-\eqref{ed} that it is actually the case that
\begin{align}\label{ac}
\Delta^{\textup{double}}
<\Delta^{\textup{single}}\quad\text{if and only if }\alpha>\alpha_{\text{c}}
:=\frac{2}{5-2 p+\sqrt{12 (p-3) p+25}}.
\end{align}
It is straightforward to check that the critical value $\alpha_{\text{c}}$ always lies in the interval,
\begin{align*}
\alpha_{\text{c}}\in(0.2,0.5)\quad \text{for all }p\in(0,1).
\end{align*}
Therefore, $\alpha>0.5$ is a sufficient condition for $\Delta^{\textup{double}}<\Delta^{\textup{single}}$, and $\alpha>0.5$ is equivalent to $t_{\textup{half}}>{\tau}$.
These results imply that if $\alpha>\alpha_{\text{c}}$, then a patient {\color{black} following the double dose protocol with actual adherence ${p}$} can have the same deviation from perfect adherence as they would have by following the single dose protocol with a higher adherence $p_{+}$. {\color{black} We thus refer to $p_{+}$ as their ``effective adherence.''} To calculate $p_{+}$, suppose the patient has actual adherence ${p}$. We then find the value of $p_{+}$ which satisfies
\begin{align}\label{set}
\Delta^{\textup{single}}|_{p_{+}}
=\Delta^{\textup{double}}|_{{p}},
\end{align}
where $\Delta^{\textup{single}}|_{p_{+}}$ and $\Delta^{\textup{double}}|_{{p}}$ denote setting the adherence equal to $p_{+}$ and $p$ in the respective formulas in \eqref{es}-\eqref{ed}. Solving \eqref{set} yields the ``effective adherence'' $p_{+}\in(0,1)$ as a function of $\alpha$ and the actual adherence ${p}$. Note that \eqref{ac} implies that $p_{+}>{p}$ if and only if $\alpha>\alpha_{\text{c}}$.
\begin{figure}[t
\centering
\includegraphics[width=\linewidth]{Fig3.pdf}
\caption{Panel (a) plots the effective adherence $p_{+}$ obtained by following the double dose protocol rather than the single dose protocol as a function of the actual adherence $p$. For a patient following the double dose protocol, panel (b) shows how their drug {\color{black} concentration}s could rise above the {\color{black} concentration}s in a perfectly adherent patient.}
\label{figp}
\end{figure}
In Figure~\ref{figp}a, we plot $p_{+}$ as a function of ${p}$ for different values of $\alpha$. This figure shows that the {\color{black} increase in the effective adherence obtained} by following the double dose protocol is quite substantial, especially if $\alpha$ is close to 1. For example, if $\alpha=0.8$, then a patient with actual adherence of only ${p}=0.6$ can have effective adherence $p_{+}=0.8$, and a patient with actual adherence of ${p}=0.8$ can have effective adherence $p_{+}=0.92$. Notice that $p_{+}=p$ if $\alpha=\alpha_{\textup{c}}$.
Summarizing, this analysis suggests that (i) the single dose protocol is best when $t_{\textup{half}}\ll {\tau}$ and (ii) the double dose protocol is best when $t_{\textup{half}}\gg {\tau}$. Conclusion (ii) contradicts some common dosing recommendations. Indeed, long drug half-lives are sometimes stated as the reason to avoid the double dose protocol in favor of the single dose protocol (for example, see recommendations for perampanel \cite{albassam2020} and lamotrigine sodium valproate \cite{gilbert2002}). However, we have shown that drugs with long half-lives are precisely the drugs for which patients could benefit from taking a double dose following a missed dose.
\subsection{Double dose protocol is not toxic for drugs with long half-lives}
Taking a double dose is sometimes avoided due to concern that it may cause a toxic drug {\color{black} concentration} in the body. For a patient following the double dose protocol, {\color{black}\eqref{largeste}} provides an upper bound to how their drug {\color{black} concentration or exposure} {\color{black}($C(t)$ or }$\textup{AUC}^{\textup{double}}$) could compare to the perfectly adherent patient {\color{black}($C^{\textup{perf}}(t)$ or }$\textup{AUC}^{\textup{perf}}$). Indeed, {\color{black}\eqref{largeste}} ensures that
\begin{align}\label{bb}
\begin{split}
\textcolor{black}{\frac{C^{\textup{double}}(t)}{C^{\textup{perf}}(t)}}
&\textcolor{black}{<\lambda^{\textup{double}}
=\frac{2}{1+\alpha},}\quad\text{\textcolor{black}{for any }}\textcolor{black}{t\in[0,\tau)},\\
\frac{\textup{AUC}^{\textup{double}}}{{\textup{AUC}^{\textup{perf}}}}
&<\lambda^{\textup{double}}
=\frac{2}{1+\alpha}.
\end{split}
\end{align}
{\color{black} We emphasize that \eqref{bb} means that $\lambda^{\textup{double}}$ bounds both the relative drug concentration and the relative drug exposure.}
We plot the maximum possible ``overshoot'' $\lambda^{\textup{double}}-1$ in Figure~\ref{figp}b as a function of $\alpha$. Importantly, $\lambda^{\textup{double}}$ approaches 1 for large $\alpha$, which means that the possible overshoot from following the double dose protocol vanishes for drugs with long half-lives. In practical terms, \eqref{bb} means that if $\alpha=0.8$, then the drug {\color{black} concentration} is at most $11\%$ greater than the perfectly adherent patient, and if $\alpha=0.9$, then the drug {\color{black} concentration} is at most $5\%$ greater than the perfectly adherent patient.
Furthermore, it is extremely rare for a patient to have a drug {\color{black} concentration} near the theoretical upper bound in \eqref{bb} if $\alpha$ is large. Indeed, the upper bound in \eqref{bb} is approached only by a patient that alternates exactly between taking and missing the scheduled doses for many dosing intervals if $\alpha$ is large. A more typical overshoot occurs in the following way. If the patient has been taking their medication as prescribed for a long time, then the concentration in their body time $t\in[0,\tau)$ after a dose is roughly the same as the perfectly adherent patient, which is $C^{\textup{perf}}(t)$. Then, if they miss one dose and take a double dose at the following dosing time, then the drug concentration time $t\in[0,\tau)$ after the double dose is (compared to $C^{\textup{perf}}(t)$)
\begin{align}\label{typical}
\lambda^{\textup{typ}}
:=\frac{\alpha^{2}C^{\textup{perf}}(t)+2\alpha^{t/\tau}\frac{DF}{V}}{C^{\textup{perf}}(t)}
=1+(1-\alpha)^{2}
<\lambda^{\textup{double}},
\end{align}
which is shown in Figure~\ref{figp}b.
Summarizing, if $\alpha$ is large, then a patient following the double dose protocol cannot have much more drug in their body than the perfectly adherent patient, where the precise upper bound is in \eqref{bb}. Furthermore, it is rare for the drug {\color{black} concentration} to approach the upper bound in \eqref{bb} if $\alpha$ is large, and the more typical overshoot is in \eqref{typical}.
\subsection{Adherence thresholds should depend on drug half-life, dosing interval, and dosing protocol}
How much patient adherence is needed for the patient to obtain full treatment benefits? The adherence threshold
\begin{align}\label{standard}
p\ge0.8
\end{align}
has long been considered the definition of an ``adherent patient'' \cite{burnier2019, haynes1976}. However, our calculations show the inadequacy of defining an acceptable patient adherence rate solely in terms of the adherence $p$. To illustrate, suppose patient \#1 has adherence $p_{1}=0.85$ and $\alpha_{1}=0.2$, whereas patient \#2 has adherence $p_{2}=0.75$ and $\alpha_{2}=0.8$. Therefore, by the standard definition in \eqref{standard}, patient \#1 would be deemed ``adherent'' and patient \#2 would be deemed ``non-adherent.'' However, if both patients are following the single dose protocol, then their deviations from the perfectly adherent patient are
\begin{align*}
\Delta_{1}^{\textup{single}}
&\approx0.33\quad\text{for patient \#1},\\
\text{and}\quad\Delta_{2}^{\textup{single}}
&\approx0.29\quad\text{for patient \#2}.
\end{align*}
Hence, the supposed ``non-adherent patient'' (patient \#2) is actually closer to the perfectly adherent patient than the ``adherent patient'' (patient \#1).
In fact, the situation is more exasperated if we consider the double dose protocol. To illustrate, suppose patient \#1 and patient \#2 again have respective adherence rates of $p_{1}=0.85 $ and $p_{2}=0.75$, but now suppose they both have $\alpha=0.8$. If patient \#1 follows the single dose protocol and patient \#2 follows the double dose protocol, then their deviations from the perfectly adherent patient are
\begin{align*}
\Delta_{1}^{\textup{single}}
&\approx0.19\quad\text{for patient \#1},\\
\text{and}\quad\Delta_{2}^{\textup{double}}
&\approx0.14\quad\text{for patient \#2}.
\end{align*}
Therefore, the drug {\color{black} concentration}s in the ``non-adherent patient'' are again closer to the perfectly adherent patient than the ``adherent patient.''
\subsection{More complicated dosing protocols}
We now consider more complicated dosing protocols. Notice that in the double dose protocol, the patient never takes more than two doses at a time, even if they missed two or more consecutive prior doses. A more aggressive protocol is the ``triple dose'' protocol in \eqref{ftriple5} in which the patient takes a double dose to make up for a single missed dose and a triple dose to make up for two or more consecutive missed doses. An even more aggressive protocol is the ``all dose'' protocol in \eqref{fall5} in which the patient takes all of their missed doses. As another example, consider the ``fractional'' dosing protocol,
\begin{align}
\begin{split}\label{fractional}
f_{n}^{\textup{frac}}
:=\begin{cases}
0 & \text{if }\xi_{n}=0,\\
1 & \text{if }\xi_{n}=1,\,\xi_{n-1}=1,\\
1+\alpha & \text{if }\xi_{n}=1,\,\xi_{n-1}=0,
\end{cases}
\end{split}
\end{align}
in which the patient takes an extra large fractional dose if they missed one or more prior doses. The reasoning behind the size of this extra dose is that if the patient had taken their prior dose, then the fraction of that prior dose remaining in their body at the next dosing time would be $\alpha$. Note that \eqref{fractional} is a special case of the boost protocol in \eqref{boost} with $b=\alpha$.
These protocols may be impractical, as \eqref{ftriple5}-\eqref{fall5} require the patient to keep fairly detailed records and \eqref{fractional} requires the ability to take a fractional dose. Nevertheless, it is interesting to consider the implications of these dosing protocols. In the Appendix, we obtain exact analytical formulas for the deviation $\Delta$ for these different dosing protocols (see {\color{black} Corollary~\ref{useful}}). In Figure~\ref{figall}, we plot $\Delta$ for these protocols and for the single and double dose protocols as functions of $\alpha$ (we set $p=0.8$). There are two important points to observe from Figure~\ref{figall}.
\begin{figure}[t
\centering
\includegraphics[width=.6\linewidth]{Fig4.pdf}
\caption{Deviation $\Delta$ from perfect adherence for various dosing protocols. }
\label{figall}
\end{figure}
First, if $\alpha$ is large, then the deviation $\Delta$ is smallest for the triple dose and all dose protocols. Indeed, {\color{black} the formulas in Corollary~\ref{useful}} imply that
\begin{align*}
\lim_{\alpha\to1}\Delta^{\textup{triple}}
&=(1-p)^{3},\quad
\lim_{\alpha\to1}\Delta^{\textup{all}}
=0,
\end{align*}
where the superscript indicates the corresponding dosing protocol. Hence, one might recommend the triple dose protocol or even the all dose protocol if $\alpha$ is large. However, while $\Delta^{\textup{double}}$ is much less than $\Delta^{\textup{single}}$ for large $\alpha$, the further reductions in the deviation $\Delta$ for the triple dose and all dose protocols are comparatively much smaller. Furthermore, compared to the double dose protocol, the triple dose and all dose protocols come with the costs of (i) being more complicated and (ii) allowing for a higher possible drug concentration in the body (see {\color{black}\eqref{largeste}}).
Next, notice in Figure~\ref{figall} that the fractional dose protocol in \eqref{fractional} results in a deviation $\Delta^{\textup{frac}}$ that is near minimal for all values of $\alpha\in(0,1)$. This is perhaps not surprising, since the fractional dose protocol interpolates between the single dose and double dose protocols as $\alpha$ ranges from 0 to 1. Furthermore, it is noteworthy that a patient following the fractional dose protocol is assured to never have too much drug in their body. Indeed, {\color{black}\eqref{largeste}} implies that the drug exposure $\textup{AUC}^{\textup{frac}}$ for a patient following the fractional dose protocol is bounded above by the exposure for the perfectly adherent patient,
\begin{align*}
\textup{AUC}^{\textup{frac}}
\le{\textup{AUC}^{\textup{perf}}}.
\end{align*}
{\color{black} Similarly, the drug concentration for a patient following the fractional dose protocol is bounded above by the concentration in a perfectly adherent patient,
\begin{align*}
C^{\textup{frac}}(t)
\le C^{\textup{perf}}(t),\quad\text{for any }t\in[0,\tau).
\end{align*}
Therefore,} if a patient is able to take fractional doses, then the fractional dose protocol (i) yields a small deviation $\Delta$ and (ii) ensures that the patient cannot have more drug in their body than the perfectly adherent patient (regardless of $\alpha$ and $p$).
Of course, the fractional dose protocol is similar to the single dose protocol if $\alpha$ is small, and it is similar to the double dose protocol if $\alpha$ is large. In particular, the fractional dose protocol differs significantly from both the single and double dose protocols only in the case that $\alpha\approx0.5$, which means $t_{\textup{half}}\approx {\tau}$. Therefore, this analysis suggests that (i) the single dose protocol is best when $t_{\textup{half}}\ll {\tau}$, (ii) the double dose protocol is best when $t_{\textup{half}}\gg {\tau}$, and (iii) the ``1.5 dose'' protocol is best when $t_{\textup{half}}\approx {\tau}$, where the 1.5 dose protocol means the patient takes an extra half dose to make up for a missed dose,
\begin{align}
\begin{split}\label{half}
f_{n}^{\textup{half}}
:=\begin{cases}
0 & \text{if }\xi_{n}=0,\\
1 & \text{if }\xi_{n}=1,\,\xi_{n-1}=1,\\
1.5 & \text{if }\xi_{n}=1,\,\xi_{n-1}=0.
\end{cases}
\end{split}
\end{align}
From a practical standpoint, the 1.5 dose protocol may often be feasible to implement (if a standard dose is two pills, then the patient takes three pills if they missed their prior dose). Since \eqref{half} is a special case of \eqref{boost} with $b=0.5$, {\color{black}\eqref{largeste}} implies that if a patient follows the 1.5 dose protocol, then their drug exposure, $\textup{AUC}^{1.5}$, is bounded above by
\begin{align*}
\textup{AUC}^{1.5}
\le\begin{cases}
{\textup{AUC}^{\textup{perf}}} &\text{if }\alpha\ge0.5,\\
\frac{1.5}{1+\alpha}{\textup{AUC}^{\textup{perf}}} &\text{if }\alpha<0.5.
\end{cases}
\end{align*}
Hence, a patient following the 1.5 dose protocol will never have much more drug in their body than the perfectly adherent patient if $t_{\textup{half}}\approx\tau$.
\subsection{Intuition}\label{intuition}
We have found that the single dose protocol is best when $t_{\textup{half}}\ll {\tau}$ and the double dose protocol is best when $t_{\textup{half}}\gg {\tau}$. These results relied on rather technical mathematical analysis. The purpose of this section is to provide an intuitive explanation for these results.
\subsubsection{Stochastic simulations}
\begin{figure}[t
\centering
\includegraphics[width=\linewidth]{Fig5.pdf}
\caption{Stochastic simulations of drug concentration time courses. In panels (a) and (b), we set $\alpha=0.25$ (meaning $t_{\textup{half}}\ll\tau$), with panel (b) plotted for a long time period. Panels (c) and (d) are the same as (a) and (b), except $\alpha=0.95$ (meaning $t_{\textup{half}}\gg\tau$). The adherence is $p=0.8$ in all panels.}
\label{figsim}
\end{figure}
We begin by plotting stochastic simulations of the drug concentration in the body as a function of time. In Figure~\ref{figsim}a, we set $p=0.8$ and $\alpha=0.25$ (meaning $t_{\textup{half}}\ll\tau$) and plot the concentration under perfect adherence (black dotted curve), and for imperfect adherence for the single dose protocol (red dashed curve) and double dose protocol (blue solid curve). The shaded gray highlights the region between the peaks and troughs for perfect adherence. While this is just one particular realization of the missed doses (the patient happens to miss doses at the first and fourth dosing times), it nevertheless illustrates that the curve for the patient with perfect adherence is better approximated by the single dose protocol than the double dose protocol. Indeed, the single dose and double dose protocols both undershoot the perfect adherence case when a dose is missed, but the double dose protocol then overcompensates when the patient takes their next dose. This is further illustrated in Figure~\ref{figsim}b, which plots the same scenario but for a longer time period.
\begin{figure}[t
\centering
\includegraphics[width=\linewidth]{Fig6.pdf}
\caption{The distribution of the drug exposure for the single dose protocol (red) and the double dose protocol (blue) obtained from stochastic simulations. We take $p=0.8$ in both plots and $\alpha=0.25$ in (a) and $\alpha=0.95$ in (b). In both plots, the black dotted vertical line at $\textup{AUC}/{\textup{AUC}^{\textup{perf}}}\textcolor{black}{=C(t)/C^{\textup{perf}}(t)}=1$ describes the perfectly adherent patient, and the blue dashed vertical line describes the largest possible drug {\color{black} concentration} for the double dose patient (see {\color{black}\eqref{largeste}}).
}
\label{figsim1}
\end{figure}
In Figures~\ref{figsim}c and \ref{figsim}d, we plot the same curves as in Figures~\ref{figsim}a and \ref{figsim}b except in the case that $\alpha=0.95$ (meaning $t_{\textup{half}}\gg\tau$). In this case, the double dose protocol approximates perfect adherence much better than the single dose protocol. While the double dose protocol curve does rise above the perfect adherence curve, it is only by a few percent. In contrast, the single dose protocol curve dips far below the double dose and perfect adherence curves when doses are missed.
In Figure~\ref{figsim1}, we plot the distribution of the drug concentration for the two scenarios in Figure~\ref{figsim}, with $\alpha=0.25$ in Figure~\ref{figsim1}a and $\alpha=0.95$ in Figure~\ref{figsim1}b. The distributions are computed from $10^{7}$ realizations of {\color{black}$c(N\tau+t)$ with $N=100$}. {\color{black} The very irregular distributions in Figure~\ref{figsim1}a are typical for small values of $\alpha$. The reason for this irregularity is difficult to intuit and is in fact a rich mathematical topic. Indeed, these irregular distributions for the single dose protocol have been studied in the pure mathematics literature for many decades under the name infinite Bernoulli convolutions \cite{jessen1935, kershner1935, erdos1939, peres2000, solomyak1995, peres1998, escribano2003, hu2008}.}
It is again evident from Figure~\ref{figsim1} that the single dose protocol best approximates the perfectly adherent patient when $\alpha$ is small (i.e.\ short drug half-life), whereas the double dose protocol best approximates the perfectly adherent patient when $\alpha$ is large (i.e.\ long drug half-life). Notice also from Figure~\ref{figsim1}b that it is very rare for the double dose protocol to ever result in a drug {\color{black} concentration} much larger than the perfectly adherent patient.
\subsubsection{A simple calculation}
The phenomena seen above can be explained with a simple calculation. Suppose the patient has been taking the drug as prescribed for a long time and so the drug concentration time $t\in[0,\tau)$ after a dose is $C^{\textup{perf}}(t)$. Suppose the patient then misses one dose and remembers to take the drug at the following dosing time. Under the single dose protocol, the concentration time $t\in[0,\tau)$ after the single dose is
\begin{align*}
\rho^{\textup{single}}(t)
:=\alpha^{2}C^{\textup{perf}}(t)+\alpha^{t/\tau}\tfrac{DF}{V}
=(1-\alpha(1-\alpha))C^{\textup{perf}}(t),
\end{align*}
where we have used that $C^{\textup{perf}}(t)=\alpha^{t/\tau}\tfrac{DF}{V} A^{\textup{perf}}$ and $A^{\textup{perf}}=1/(1-\alpha)$. Alternatively, under the double dose protocol, the concentration time $t\in[0,\tau)$ after the double dose is
\begin{align*}
\rho^{\textup{double}}(t)
:=\alpha^{2}C^{\textup{perf}}(t)+2\alpha^{t/\tau}\tfrac{DF}{V}
=(1+(1-\alpha)^{2})C^{\textup{perf}}(t).
\end{align*}
For small $\alpha$, we have that
\begin{align*}
\rho^{\textup{single}}(t)
&\approx(1-\alpha){C^{\textup{perf}}(t)},\quad
\rho^{\textup{double}}(t)
\approx(2-2\alpha){C^{\textup{perf}}(t)},\quad\text{if $\alpha$ is near 0},
\end{align*}
which means the single dose protocol puts the patient slightly below the desired $C^{\textup{perf}}(t)$, but the double dose protocol puts the patient at almost twice $C^{\textup{perf}}(t)$. However, for large $\alpha$, we have that
\begin{align*}
\rho^{\textup{single}}(t)
&\approx(1-(1-\alpha)){C^{\textup{perf}}(t)},\\
\rho^{\textup{double}}(t)
&=(1+(1-\alpha)^{2}){C^{\textup{perf}}(t)},
\quad\text{if $\alpha$ is near 1},
\end{align*}
which means that while the single dose protocol puts the patient below $C^{\textup{perf}}(t)$, the double dose protocol puts the patient above $C^{\textup{perf}}(t)$ by a much smaller amount. In practical terms, if $\alpha=0.9$, then the single dose patient undershoots $C^{\textup{perf}}(t)$ by about $10\%$, whereas the double dose patient overshoots $C^{\textup{perf}}(t)$ by a mere $1\%$.
\section{Discussion}
We have formulated and analyzed a mathematical model to investigate how nonadherence to medication affects drug {\color{black} concentration}s in the body. We computed pharmacologically relevant statistics of the drug {\color{black} concentration} in the body, thus providing quantitative descriptions of the effects of nonadherence, and how these effects depend on the adherence percentage $p$, drug half-life $t_{\textup{half}}$, dosing interval ${\tau}$, and how missed doses are handled (i.e.\ the dosing protocol). In agreement with previous results \cite{osterberg2005}, we found that drug {\color{black} concentration}s are less affected by missed doses if the half-life is long compared to the dosing interval, and we quantified this effect. As a general principle, we found that nonadherence is best mitigated by taking double doses following missed doses if the drug half-life is long compared to the dosing interval (i.e.\ $t_{\textup{half}}\gg {\tau}$). Furthermore, in this scenario we found that taking double doses following missed doses cannot cause the drug {\color{black} concentration} to rise much above the desired {\color{black} concentration}. Although long drug half-lives are sometimes stated as the reason to avoid a double dose after a missed dose, we have shown that drugs with long half-lives are precisely the drugs for which patients could benefit from taking a double dose after a missed dose.
As an application of these results, consider the synthetic form of thyroxine known as levothyroxine \cite{brent2017}. Levothyroxine is the standard treatment for hypothyroidism, which is one of the most common diseases in the world and affects up to 5\% of the global population \cite{duntas2019}. Levothyroxine pills are used to replace missing thyroid hormone in hypothyroid patients and are usually taken once daily for the remainder of the patient's life \cite{duntas2019}. Hence, the dosing interval is ${\tau}=1\;\text{day}$. The half-life of levothyroxine for hypothyroid patients is between 9 and 10 days \cite{lexicomp, dynamed}, and therefore setting $t_{\textup{half}}=9\;\text{days}$ yields
\begin{align*}
\alpha=2^{-{\tau}/t_{\textup{half}}}\approx0.93.
\end{align*}
Since this $\alpha$ value is close to 1, our results imply that a hypothyroid patient taking levothyroxine with imperfect adherence can make the drug {\color{black} concentration} in their body much closer to the {\color{black} concentration} in a perfectly adherent patient by following the double dose protocol rather than the single dose protocol. That is, if the patient {\color{black} misses a dose, then it is better to take a double dose at the next dosing time than to skip the missed dose}. These results conflict with common recommendations for levothyroxine, which advise patients to skip any dose that is delayed by more than 12 hours \cite{mayoclinic, rxlist, drugs, nhs}. However, some sources recommend a double dose of levothyroxine after a missed dose (see Chapter 376 of \cite{jameson2018}), and indeed taking a double dose is recognized as safe (see Chapter 36 in \cite{miller2018}). {\color{black} In fact, the American Thyroid Association has proposed taking up to 7 doses of levothyroxine at once \cite{jonklaas2014}.}
Furthermore, although following the double dose protocol may cause the drug {\color{black} concentration} in the patient to rise above the {\color{black} concentration} in a perfectly adherent patient, the maximum possible overshoot for levothyroxine is less than $4\%$ since
\begin{align*}
\lambda^{\textup{double}}
=\frac{2}{1+\alpha}
\approx\frac{2}{1+0.93}<1.04.
\end{align*}
In addition, it would be very rare for a patient to have drug {\color{black} concentration}s near this maximum, as this maximum corresponds to a patient missing doses every other day for many days. Indeed, the typical overshoot is {\color{black} less than 1\%} for this $\alpha$ value (see \eqref{typical}).
Our model assumes that the drug absorption rate $k_{\textup{a}}$ is much faster than the drug elimination rate $k_{\textup{e}}$. This is true for most drugs administered orally in conventional dosage forms \cite{gibaldi1982, peletier2017, ma2018, fillastre1987, leroy1990, strandgaarden1999}, including levothyroxine. Indeed, for hypothyroid patients taking levothyroxine, the time to maximum concentration, $t_{\textup{max}}$, is only 3 hours \cite{benvenga1995}, whereas the elimination half-life is $t_{\textup{half}}=9\;\text{days}$ \cite{lexicomp, dynamed}. Using \eqref{alphahalf} and the relation \cite{gibaldi1982},
\begin{align*}
t_{\textup{max}}
=\frac{\ln(k_{\textup{a}}/k_{\textup{e}})}{k_{\textup{a}}-k_{\textup{e}}},
\end{align*}
implies that $k_{\textup{e}}/k_{\textup{a}}<0.0015$ for levothyroxine.
Several important prior works have used mathematical modeling to investigate the effects of medication nonadherence. Li and Nekka developed stochastic models of the effect of medication nonadherence on patient drug {\color{black} concentration}s \cite{li2007, li2009}. The models in \cite{li2007, li2009} allow the drug to be administered at irregular times, and the authors obtained analytical formulas for drug {\color{black} concentration} statistics. In a series of papers \cite{levy2013, fermin2017}, another group of authors developed a variety of stochastic pharmacokinetic models, including ones that allow for variation in dosing times, dose amounts, and elimination rates. The discrete time model proposed in \cite{levy2013} is essentially identical to the model in the present paper in the special case of the single dose protocol. These prior works did not analyze different protocols for handling missed doses. Ma \cite{ma2017} analyzed the mean first passage time for the patient's drug {\color{black} concentration} to reach a therapeutic range for various ways of handling a missed dose assuming that the patient never misses two or more consecutive doses. Numerical simulations of computational models have also been useful for understanding the effects of nonadherence for specific drugs \cite{gu2020}, especially for antiepileptic drugs \cite{garnett2003, reed2004, dutta2006, ding2012, chen2013, gidal2014, brittain2015, sunkaraneni2018} and antipsychotic drugs \cite{hard2018, elkomy2020}.
Naturally, our model neglects various pharmacological details. We have developed a simple model aimed at addressing patients remembering or forgetting to take their medication{\color{black}, and we assumed that the patient forgets their medication at each dosing time with a fixed probability, independent of their prior behavior}. However, nonadherence is a dynamic process and patients exhibit a variety of patterns of nonadherence \cite{burnier2019}, including extended ``drug holidays'' \cite{vrijens2008} and ``white-coat adherence'' \cite{burnier2013}. {\color{black} We also assumed that the patient takes a double dose only if they missed their prior dose. However, actual patients might cause harm by mistakenly taking a double dose when they did not miss their prior dose.} Furthermore, {\color{black} our model did not allow delayed doses, and a more detailed model would} allow patients to take medication at times that vary continuously. Another source of stochasticity is that pharmacokinetic parameters vary between patients, which has been modeled by analyzing a population of patients with a distribution of parameters \cite{levy2013, fermin2017}.
{\color{black} Another limitation of our analysis is that we considered only a single compartment pharmacokinetic model with linear elimination and immediate absorption. The pharmacokinetics of some drugs are better described by multicompartment models \cite{gibaldi1982}, and while most drugs can be adequately described by linear processes, there are drugs which exhibit nonlinear kinetics (see Chapter 7 in \cite{gibaldi1982}). In addition, our assumption of fast absorption does not hold for so-called extended release or sustained release drugs \cite{gidal2021, wheless2018, vadivelu2011}. Our model also did not address pharmacodynamics, and an interesting avenue for further research would be to couple the stochastic pharmacokinetic model in this work to a pharmacodynamic model.}
To conclude, medication nonadherence is a complex and multi-faceted problem, and steps toward its alleviation require contributions from a variety of disciplines. Mathematical modeling is a valuable tool in this endeavor, especially given the ethics of clinical trials that require sporadic dosing. Further, mathematical models can disentangle the effects of various factors and quickly investigate the efficacy of possible interventions. Moving forward, we anticipate that mathematical modeling and analysis will play an important role in understanding and alleviating the effects of medication nonadherence.
|
\section{\label{Introduction}Introduction}
The 2012 discovery of the Higgs boson at the Large Hadron Collider (LHC)~\cite{Aad:2012tfa,Chatrchyan:2012ufa} confirmed the correctness of the Higgs mechanism as the origin of the masses of subatomic particles. However, the hierarchy problem caused by sizeable radiative corrections to the Higgs mass term implies there should be new physics between the electroweak scale and the Planck scale. In addition, the Standard Model (SM) cannot explain the existence of dark matter (DM). However, current astronomical observations have confirmed that the universe is composed of $27\%$ DM~\cite{Ade:2015xua, Aghanim:2018eyx}. Therefore, new DM particle candidates and new physics are required beyond the SM. Establishing the structural nature of DM is one of the most fundamental open questions in cosmology and particle physics.
Among multiple proposed theories, the most widely accepted are supersymmetry (SUSY) models with R-parity conservation, i.e., the minimal supersymmetric standard model (MSSM)~\cite{Haber:1984rc,Gunion:1984yn} and the next-to-minimal supersymmetric standard model (NMSSM)~\cite{Ellwanger:2009dp,Haber:1986gz,Maniatis:2009re}, which provide elegant solutions to the hierarchy problem by introducing contributions from superpartners to the Higgs mass term. In addition, R-parity conservation ensures that the lightest neutralino is a stable neutral particle if it is the lightest supersymmetric particle (LSP), and it may be an excellent DM candidate. MSSM, the most economic realization of SUSY, exhibits several attractive features, but also includes some challenges (e.g., the ``$\mu$-problem''~\cite{Kim:1983dt} and ``little hierarchy problem''~\cite{BasteroGil:2000bw}) that have become exacerbated in recent years by the first run of the LHC experiments. This was particularly true for the uncomfortably large mass of the discovered Higgs boson $m_h \simeq 125~{\rm GeV}$~\cite{Aad:2015zhl,Hall:2011aa,Ellwanger:2011aa, Gunion:2012zd,King:2012tr, Kang:2012sy,King:2012is,Cao:2012fz,Vasquez:2012hn}. Alternatively, the NMSSM solves the $\mu$-problem by adding a singlet chiral superfield $\hat{S}$ to the MSSM. In this process, the $\mu$ parameter is replaced by a dynamic quantity $\mu_{eff} = \lambda v_s$ when $S$ develops a vacuum expectation value (VEV) $v_s$, the magnitude of which is naturally at the electroweak scale~\cite{Ellwanger:2009dp,Maniatis:2009re}. Furthermore, the SM-like Higgs squared mass can be enhanced by an additional tree-level contribution $\lambda^2 v^2 \sin^2 2 \beta$ and the singlet-doublet Higgs mixing ~\cite{Hall:2011aa,Ellwanger:2011aa,King:2012is,Cao:2012fz}, where $\tan\beta \equiv {v_u}/{v_d}$ with $v_u$ and $v_d$ representing the VEVs of the doublet Higgs fields and $v^2 \equiv v_u^2 + v_d^2$.
As a byproduct, the neutralino sector includes a fermionic partner of $S$ (singlino) in addition to neutral electroweak gauginos (bino and wino) and the neutral fermionic partner of the Higgs doublets (higgsinos). Since the coupling of a singlino field with SM particles may be very weak, this study focuses on the case of a singlino-dominated neutralino as a DM candidate~\cite{Guchait:2020wqn,Das:2012rr,Ellwanger:2014hia,Ellwanger:2016sur,Ellwanger:2018zxt,Aldufeery:2020dkb}. This scenario is feasible when the Yukawa-like couplings satisfy $\lambda \geq 2 \kappa$ and the gauginos are assumed to be heavier than the higgsinos.
In the NMSSM, the Z boson mass is related to the higgsino mass $\mu_{eff}$, and a natural prediction of $m_Z$ favors light higgsinos up to several hundred GeV~\cite{Baer:2012uy}. Generally, the NMSSM with $\mu_{eff} \lesssim 500~\rm{GeV}$ is considered to be a natural NMSSM \cite{King:2012tr,Baer:2012uy,Kang:2012sy,Cao:2014kya,Cao:2016nix}.
However, given the constraints of recent experiments (e.g., the searches for electroweakinos~\cite{Sirunyan:2018lul,Sirunyan:2018nwe,Sirunyan:2018ubx,Sirunyan:2017eie,Sirunyan:2017lae,Sirunyan:2017qaj,Aaboud:2018jiw,Aaboud:2018ngk,Aaboud:2018sua,
Aad:2019vvi,Aad:2019vnb,Aad:2019vvf,ATLAS:2013rla,Sirunyan:2018iwl,TheATLAScollaboration:2013zia,Aad:2014vma,Aad:2014nua,Aad:2015jqa,CMS:2013bda,CMS:2013dea}, the WMAP/Planck experiments~\cite{Ade:2015xua, Aghanim:2018eyx}, and the DM direct detection (DD) and indirect detection (ID) experiments~\cite{Aprile:2018dbl,Aprile:2019dbj,Wang:2020coa,Cui:2017nnn,Ackermann:2015zua}), a large portion of the parameter space in the natural NMSSM has been strongly constrained. As a result, the following isolated and narrow parameter spaces with a singlino-dominated $\tilde \chi^0_1$ are preferred~\cite{Cao:2018rix,Abdallah:2019znp,Abdallah:2020yag,Cao:2019qng}:
\begin{itemize}
\item $\lambda \simeq 2 \kappa$ with $\lambda \lesssim 0.05$, where $\tilde \chi^0_1$ mainly co-annihilates with higgsinos to achieve the measured abundance~\cite{Cao:2018rix}.
\item $\kappa \sim 0.01$, $\lambda \lesssim 0.2$, and at least one light singlet-dominated Higgs boson~\cite{Abdallah:2019znp}. Here, $\tilde \chi^0_1$ annihilates in certain funnel regions and the higgsinos decay in a complex manner to satisfy the LHC constraints.
\end{itemize}
These conclusions are applicable for $|\mu_{eff}| \lesssim 500~{\rm GeV}$, or equivalently, the fine-tuning criterion ${ {\Delta}m_Z} \lesssim 50$, where $\Delta m_Z$ defined in~\cite{Ellwanger:2011mu} parameterizes the sensitivity of $m_Z$ to the SUSY parameters at the weak scale. Given this situation and the fact that the fine-tuning criteria lacks a confirmed scientific basis and may reflect personal prejudice, we update the study by Cao et al. \cite{Cao:2018rix} to improve these conclusions. Specifically, we do not require ${\Delta}m_Z$ to be less than 50 but impose the condition that $\mu_{eff} \leq 1~{\rm TeV}$. We adopt an advanced MultiNest algorithm~\cite{Feroz:2008xx,Feroz:2013hea} to perform a sophisticated scan over the parameter space of the $Z_3$-invariant NMSSM with a singlino-dominated DM. This algorithm is much more efficient than the other algorithms (e.g., the Markov Chain method~\cite{Markov-Chain} adopted in~\cite{Cao:2018rix}) in providing comprehensive information of the space to reveal the underlying physics, although it usually involves a tremendous amount of calculation. To the best of our knowledge, few researchers have used it to study the NMSSM phenomenology~\cite{Gunion:2011hs,Kowalska:2012gs,Kim:2013uxa,Roszkowski:2014lga,Athron:2017fxj,AbdusSalam:2017uzr}. We also present a description of DM annihilation and the mechanisms used to suppress the DM nucleon scattering cross section through analytical formulas and numerical analysis. Some of the formulas are new, and some are consistent with excellent related works ~\cite{Cheung:2014lqa,Cao:2015loa,Badziak:2015exr,Baum:2017enm,Badziak:2017uto}. Evidently, such an analysis is helpful to understand the DM physics.
The remainder of this paper is organized as follows. In Section~\ref{sec:model}, we briefly introduce the basic properties of the $Z_3$-invariant NMSSM, including the Higgs and neutralino sections. We then demonstrate DM annihilation and scattering cross sections for singlino-dominated $\tilde{\chi}^0_{1}$ with nucleons using analytical formulas. In Section~\ref{sec:scan}, we provide a brief description of our scanning strategy. In Section~\ref{result13}, we investigate predictions for surviving samples and the properties of singlino-dominated DM scenarios to understand their distinctive features. Finally, Section~\ref{sec:conclusion} includes a discussion of the results and corresponding conclusions.
\section{\label{sec:model} Next-to-minimal supersymmetric standard model model}
\subsection{Fundamental NMSSM properties}
As the simplest extension of the MSSM, the NMSSM includes one additional gauge singlet Higgs field $\hat{S}$. The associated superpotential can be expressed as follows~\cite{Maniatis:2009re,Ellwanger:2009dp}:
\begin{align}
\label{eq:superpotential}
W_\mathrm{NMSSM}=W_\mathrm{MSSM} + \lambda \hat{S} \hat{H_u} \hat{H_d} + \frac{1}{3} \kappa \hat{S}^3,
\end{align}
where $W_\mathrm{MSSM}$ is the MSSM superpotential without the $\mu$ term, $\lambda$ and $\kappa$ are dimensionless parameters, and $\hat{H}_u$ and $\hat{H}_d$ are the common Higgs superfields. It is the most general R-parity-conservation superpotential satisfying a $Z_3$ discrete symmetry given the considered field content.
Assuming CP-conservation, the Higgs sector of the $Z_3$-NMSSM is determined by six parameters at the tree-level~\cite{Ellwanger:2009dp,Cheung:2010ba}:
\begin{align}
\lambda,~ \kappa,~ A_\lambda,~ A_\kappa,~ \mu_{eff},~ \tan\beta,
\label{eq:six}
\end{align}
where $A_\lambda$ and $A_\kappa$ are the soft trilinear coefficients defined in Eq. (2.5) of Ref.~\cite{Ellwanger:2009dp}.
In the base-vectors, $H_{\rm SM} \equiv \sin{\beta} Re[H_{u}^0] + \cos{\beta} Re[H_{d}^0], H_{\rm NSM} \equiv \cos{\beta} Re[H_{u}^0] - \sin{\beta} Re[H_{d}^0]$, and $H_{\rm S} \equiv Re[S]$ for CP-even fields and $A_{\rm NSM} \equiv \cos{\beta} Im[H_{u}^0] + \sin{\beta} Im[H_{d}^0]$ and $A_{\rm S}\equiv Im[S])$ for CP-odd fields\footnote{$H_u^0$, and $H_d^0$ denote the neutral component fields of the doublet scalar fields $H_u$ and $H_d$, respectively.}, the three CP-even mass eigenstates $h_i = \{ h, H, h_s \}$ and two CP-odd Higgs mass eigenstates $a_i = \{A_H, a_s\}$ are given as follows:
\begin{eqnarray}\label{eq:hi}
h_i &=& V_{h_i}^{\rm SM} H_{\rm SM}+V_{h_i}^{\rm NSM} H_{\rm NSM}+V_{h_i}^{\rm S} H_{\rm S}, \nonumber \\
a_i &=& V_{a_i}^{' \rm NSM} A_{\rm NSM} + V_{a_i}^{'\rm S} A_S,
\end{eqnarray}
where $V$ and $V'$ represent the unitary matrices to diagonalize the corresponding Higgs squared mass matrix. In this work, we denote the physical Higgs state with the largest $H_{\rm SM}$ component by the symbol $h$, which is called the SM-like Higgs boson hereafter, and we denote the physical Higgs state with the largest non-SM doublet (singlet) component $H_{\rm NSM}$ ($H_{\rm S}$) by $H$ ($h_{s}$). We also denote the CP-even Higgs bosons by $h_1$, $h_2$, and $h_3$, with $m_{h_1} < m_{h_2} < m_{h_3}$. The latter notation is primarily for convenience. To date, the LHC experiments have measured the couplings of the discovered Higgs boson with about $10\%$ uncertainty, and they revealed that the boson has roughly the same couplings as the SM Higgs boson~\cite{Aad:2019mbh,Sirunyan:2018koj}. These facts imply that $\sqrt{\left (V_{h}^{\rm NSM} \right )^2 + \left ( V_{h}^{\rm S} \right )^2} \lesssim 0.1$ and $|V_{h}^{\rm SM}| \sim 1$.
In the $Z_3$-NMSSM, mixtures of bino ($\tilde{B}^0$), wino ($\tilde{W}^0$), higgsino ($\tilde{H}_{d,u}^0$), and
singlino ($\tilde{S}^0$) fields form neutralinos. Assuming a basis of $\psi^0 = (-i \tilde{B}, - i \tilde{W}^0, \tilde{H}_{d}^0, \tilde{H}_{u}^0,\tilde{S})$ produces the following neutralino mass matrix~\cite{Ellwanger:2009dp}:
\begin{align}
{\cal M} = \left(
\begin{array}{ccccc}
M_1 & 0 & -\frac{g_1 v_d}{\sqrt{2}} & \frac{g_1 v_u}{\sqrt{2}} & 0 \\
& M_2 & \frac{g_2 v_d}{\sqrt{2}} & - \frac{g_2 v_u}{\sqrt{2}} &0 \\
& & 0 & -\mu_{eff} & -\lambda v_u \\
& & & 0 & -\lambda v_d\\
& & & & \frac{2 \kappa}{\lambda} \mu_{eff}
\end{array}
\right), \label{eq:massmatrix}
\end{align}
where $M_1$, $M_2$, and $\mu_{eff}$ denote the soft breaking masses of the bino, wino, and higgsinos, respectively. Diagonalizing the mass matrix with a unitary matrix $N$ yields five mass eigenstates (ordered by mass):
\begin{align}\label{eq:chi01all}
\tilde{\chi}_i^0 = N_{i1} \tilde B^0 + N_{i2} \tilde W_3^0 + N_{i3}\tilde{H}_d^0 + N_{i4}\tilde{H}_u^0 + N_{i5}\tilde{S}.
\end{align}
The lightest neutralino, $\tilde{\chi}_1^0$, acting as the DM candidate is the focus of this work.
In the limit that $|M_1|$ and $|M_2|$ are much larger than $|\mu_{eff}|$ and $v$, $\tilde{\chi}_1^0$ is approximated by
\begin{align}\label{eq:chi01SH}
\tilde{\chi}_1^0 \simeq N_{13}\tilde{H}_d^0 + N_{14}\tilde{H}_u^0 + N_{15}\tilde{S}.
\end{align}
If $|\kappa/\lambda| < 1$, the dominant composition of $\tilde{\chi}_1^0$ is a singlino. In this case,
$\kappa$ is related to $m_{\tilde{\chi}_1^0}$ as follows~\cite{Cheung:2014lqa,Cao:2015loa}:
\begin{eqnarray}
\kappa & = & \frac{\lambda}{2 \mu_{eff}} \left[ m_{\tilde{\chi}_1^0}
- \frac{\lambda^2 v^2 \left( m_{\tilde{\chi}_1^0} - \mu_{eff} \sin 2\beta
\right)}{m_{\tilde{\chi}_1^0}^2 - \mu_{eff}^2} \right], \label{eq:kappa}
\end{eqnarray}
and the elements of the matrix $N$ exhibit the following relationships~\cite{Cheung:2014lqa,Badziak:2015exr,Badziak:2017uto,Baum:2017enm,Cao:2015loa}:
\begin{align}
\label{eq:N13N15}
\frac{N_{13}}{N_{15}}
=
\frac{\lambda v}{\mu_{eff}}
\,
\frac{(m_{\tilde{\chi}_1^0}/\mu_{eff})\sin\beta-\cos\beta}
{1-\left(m_{\tilde{\chi}_1^0}/\mu_{eff} \right)^2}
\,,\\[4pt]
\label{eq:N14N15}
\frac{N_{14}}{N_{15}}
=
\frac{\lambda v}{\mu_{eff}}
\,
\frac{(m_{\tilde{\chi}_1^0}/\mu_{eff})\cos\beta-\sin\beta}
{1-\left(m_{\tilde{\chi}_1^0}/\mu_{eff} \right)^2}.
\end{align}
Thus,
\begin{eqnarray}
\label{eq:N15}
N_{15}^2 && \simeq \left(1+ \frac{N^2_{13}}{N^2_{15}}+\frac{N^2_{14}}{N^2_{15}}\right)^{-1} \\
&&\simeq \frac{\left[1-(m_{\tilde{\chi}_1^0}/\mu_{eff} )^2\right]^2}{\left[(m_{\tilde{\chi}_1^0}/\mu_{eff} )^2
-2(m_{\tilde{\chi}_1^0}/\mu_{eff} )\sin2\beta+1 \right]\left({\lambda v}/{\mu_{eff}}\right)^2
+\left[1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2\right]^2}. \nonumber
\end{eqnarray}
The higgsino and singlino fractions in $\tilde{\chi}_1^0$ can be defined as
$Z_h=N_{13}^2+N_{14}^2$ and $Z_s=N_{15}^2$, respectively. The ratio of $Z_h$ to $Z_s$ can then be expressed as follows:
\begin{align}
\label{eq:Higgsino/singlino}
\frac{Z_h}{Z_s}
=
\left(\frac{\lambda v}{\mu_{eff}}\right)^{\!\!2}
\frac{\left(m_{\tilde{\chi}_1^0}/\mu_{eff}\right)^2-2{(m_{\tilde{\chi}_1^0}}/{\mu_{eff}})\sin2\beta+1}
{\left[1-\left({m_{\tilde{\chi}_1^0}}/{\mu_{eff}}\right)^2\right]^2}
\,.
\end{align}
This expression implies that a small $\lambda$ can suppress the higgsino fraction in $\tilde{\chi}_1^0$.
The couplings of DM to scalar Higgs states, the Z boson, and the Goldstone boson $G^0$ are included in the calculation of DM annihilation.
They take the following form~\cite{Ellwanger:2009dp}:
\begin{eqnarray}
{\cal{L}}_{\rm NMSSM} \ni C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 Z} Z_\mu \overline{\tilde{\chi}_1^0} \gamma^\mu \gamma_5 \tilde{\chi}_1^0 + i C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 G^0} G^0 \overline{\tilde{\chi}_1^0} \gamma_5 \tilde{\chi}_1^0 + C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 h_i} h_i \overline{\tilde{\chi}_1^0} \tilde{\chi}_1^0 + i C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 a_i} a_i \overline{\tilde{\chi}_1^0} \gamma_5 \tilde{\chi}_1^0, \nonumber
\end{eqnarray}
where the coefficients are given by
\begin{align}
C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 Z}\simeq & \frac{m_Z}{\sqrt{2} v} \big( \frac{\lambda v}{\mu_{eff}} \big)^2
\,
\frac{Z_s\cos2\beta}
{1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2},
\label{eq1:zchi10chi10_S} \\
C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 G^0 }
\simeq& \frac{\sqrt{2} \mu_{eff}}{v} \left( \frac{\lambda v}{\mu_{eff}}\right)^2 \frac{Z_s(m_{\tilde{\chi}_1^0}/\mu_{eff})\cos2\beta}{1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2},
\label{eq1:G0chi10chi10_S}
\\
C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 h_{i}} \simeq &
V_{h_i}^{\rm SM} C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 H_{\rm SM} }+V_{h_i}^{\rm NSM} C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 H_{\rm NSM} }+V_{h_i}^{\rm S} C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 H_{\rm S} } \nonumber \\
\simeq &
\frac{\sqrt{2}\mu_{eff}}{v}\,\big( \frac{\lambda v}{\mu_{eff}} \big)^2\, \frac { Z_s V_{h_{i}}^{\rm SM}(m_{\tilde{\chi}_1^0}/\mu_{eff} -\sin 2 \beta)}{1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2} - \frac{\sqrt{2}\mu_{eff}}{v}\,\big( \frac{\lambda v}{\mu_{eff}} \big)^2\, \frac { Z_s V_{h_{i}}^{\rm NSM}\cos 2 \beta}{1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2} \nonumber \\
&+ \lambda \big( \frac{\lambda v}{\mu_{eff}} \big)^2 \frac{Z_s V_{h_{i}}^{\rm S} \sin2\beta}{\sqrt{2}\big[ 1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2 \big]}-\sqrt{2}\kappa Z_s V_{h_{i}}^{\rm S} \left[1+ \big( \frac{\lambda v}{\mu_{eff}} \big)^2\frac{2}{1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2} \right],
\label{eq1:hichi01chi01_S}
\\
C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 a_{i}} \simeq &
V_{a_i}^{' \rm NSM}C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 A_{\rm NSM} }+ V_{a_i}^{' \rm S}C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 A_{\rm S} } \nonumber \\
\simeq &
-\frac{\sqrt{2}\mu_{eff}}{v}\,\big( \frac{\lambda v}{\mu_{eff}} \big)^2\, \frac { Z_s V_{a_{i}}^{' \rm NSM}(m_{\tilde{\chi}_1^0}/\mu_{eff} \sin 2 \beta-1)}{1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2} \nonumber \\
&+ \lambda \big( \frac{\lambda v}{\mu_{eff}} \big)^2 \frac{Z_s V_{a_{i}}^{'\rm S} \sin2\beta}{\sqrt{2}\big[ 1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2 \big]}
-\sqrt{2}\kappa Z_s V_{a_{i}}^{'\rm S} \left[1+ \big( \frac{\lambda v}{\mu_{eff}} \big)^2\frac{2}{1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2} \right],
\label{eq1:aichi01chi01_S}
\end{align}
where the approximation in Eq. (\ref{eq:chi01SH}) was applied. In addition, the coupling of a CP-even Higgs $h_i$ to two quasi-pure singlet CP-odd Higgs $a_s$,
${\cal{L}}_{\rm NMSSM} \ni C_{h_i a_s a_s} h_i a_s a_s$, is relevant to our study. Its coefficient $C_{h_i a_s a_s}$ is given
as follows~\cite{Ellwanger:2004xm,Belanger:2005kh}:
\begin{align}
C_{h_i a_s a_s}\simeq& \sqrt{2}\lambda v\,V_{h_i}^{\rm SM} \,(\lambda+\kappa \sin2\beta)+ \sqrt{2}\lambda\kappa v\,
V_{h_i}^{\rm NSM}\cos2\beta
+\sqrt{2} V_{h_i}^{\rm S}( 2 \kappa^2 v_s -\kappa A_\kappa).
\label{eq:hiasas}
\end{align}
In most cases, $C_{h_s a_s a_s} \gg C_{h a_s a_s}$, since $v_s \gg v$.
\subsection{Dark matter relic density}
In the NMSSM, the abundance of the singlino-dominated DM candidate ($\tilde \chi^0_1$) tends to be unacceptably large, due to small coupling effects with SM particles. However, such a candidate can still achieve the measured abundance~\cite{Ade:2015xua, Aghanim:2018eyx} by a specific mechanism, e.g., via $s$-channel exchanges of gauge and Higgs bosons and $t$-channel exchanges of electroweakinos and sfermions. It can also be achieved through co-annihilation with heavier states, such as sleptons, the next-to-lightest neutralino ($\tilde{\chi}_2^0$), or the lightest chargino ($\tilde{\chi}_1^\pm$).
The thermal abundance of the DM at the freeze-out temperature $T_F = m_{\tilde{\chi}_1^0}/x_F$ is given as follows~\cite{Baum:2017enm}:
\begin{equation}\label{density}
\Omega h^2 = 0.12\left(\frac{80}{g_*}\right)^{1/2}\left(\frac{x_F}{25}\right) \left( \frac{2.3\times 10^{-26}\ \mathrm{cm^3/s}}{\langle \sigma v\rangle_{x_F}}\right)\;,
\end{equation}
with a thermally-averaged annihilation cross section ${\langle \sigma v\rangle_{x_F}} \equiv a+ \frac{3b}{x_F}$. Dominant contributions to $\langle \sigma v \rangle_{x_F}$ in acquiring the measured abundance are discussed in Section 2.2 of Baum et al.~\cite{Baum:2017enm}. The following conclusions were presented:
1) $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \to t\bar{t}$ is usually the most crucial channel for the abundance when $m_{\tilde{\chi}_1^0} > m_t$. It proceeds through the s-channel exchanges of Higgs and Z bosons. Since the top quark is massive, the contribution from the Z boson's longitudinal polarization to $\langle \sigma v \rangle_{x_F}$ is important in this process when the Higgs mediators are far off-shell. $\langle \sigma v \rangle_{x_F}$ can then be approximately expressed as follows:
\begin{equation} \label{eq:gxxGforOmega}
\langle \sigma v \rangle_{x_F}^{t\bar{t}} \sim 2 \times 10^{-26}\,\frac{{\rm cm}^3}{{\rm s}} \left( \frac{\left|C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 G^0}\right|}{0.1} \right)^2 \left( \frac{m_{\tilde{\chi}_1^0} }{300\,{\rm GeV}}\right)^{-2}.
\end{equation}
The measured abundance $\Omega h^2 \sim 0.12$~\cite{Ade:2015xua,Aghanim:2018eyx} is achieved through the coupling of DM pairs to the Goldstone boson $|C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 G^0}| \sim 0.1 $, which requires $\lambda \gtrsim 0.4$ according to Eq. (\ref{eq1:G0chi10chi10_S}).
2) $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \to \Phi_i \Phi_j$ is another crucial annihilation process for the abundance, where $\Phi_i$ denotes a scalar or pseudo-scalar Higgs mass eigenstate. Such processes occur via $s$-channel Higgs or $Z$ boson exchange and $t$-channel neutralino exchange. Since $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \to h_i h_j$ and $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \to a_i a_j$ are $p$-wave suppressed~\cite{Baum:2017enm}, and because $C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 h_i}$ with $h_i = h, H$ in Eq.~(\ref{eq1:hichi01chi01_S}) and
$C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 A_H}$ in Eq.~(\ref{eq1:aichi01chi01_S}) are usually smaller than 0.1, here we consider only the contribution from $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \to h_s a_s$ to $\langle \sigma v \rangle_{x_F}$. $\langle \sigma v \rangle_{x_F}^{h_s a_s}$ is then given as follows~\cite{Baum:2017enm,Griest:1990kh}:
\begin{eqnarray} \label{eq:sigvPhiPhi}
\left\langle \sigma v \right\rangle_{x_F}^{h_s a_s} \simeq && \frac{1}{64 \pi m_{\tilde{\chi}_1^0}^2} \left\{ \left[1-\frac{\left(m_{h_s} + m_{a_s}\right)^2}{4 m_{\tilde{\chi}_1^0}^2}\right] \left[1-\frac{\left(m_{h_s} - m_{a_s}\right)^2}{4 m_{\tilde{\chi}_1^0}^2}\right] \right\}^{1/2} | {\cal{A}}_s + {\cal{A}}_t |^2, \quad \label{hsaa-approximation}
\end{eqnarray}
where the $s$- and $t$-channel contributions are approximated as
\begin{eqnarray}
{\cal{A}}_s &\simeq & \frac{-2 m_{\tilde{\chi}_1^0} C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 a_s } C_{h_s a_s a_s}}{m_{a_s}^2 - 4 m_{\tilde{\chi}_1^0}^2}, \nonumber \\
{\cal{A}}_t &\simeq & - 2 C_{\tilde{\chi}_1^0\tilde{\chi}_1^0 h_s} \, C_{\tilde{\chi}_1^0\tilde{\chi}_1^0 a_s} \left[ 1 + \frac{ 2 m_{a_s}^2}{ 4 m_{\tilde{\chi}_1^0}^2 - \left(m_{h_s}^2 + m_{a_s}^2\right) } \right],
\end{eqnarray}
if there are no resonant contributions\footnote{Note that the $\tilde{\chi}_i^0$-mediated ($i \neq 1$) contribution to $\langle \sigma v \rangle_{x_F}^{h_s a_s}$ is less important than the $\tilde{\chi}_1^0$-mediated contribution for two reasons. One is that, if $|\kappa|$ is comparable to $\lambda$, $|C_{\tilde{\chi}_1^0 \tilde{\chi}_i^0 h_s}|$ and $|C_{\tilde{\chi}_1^0 \tilde{\chi}_i^0 a_s}|$ are significantly smaller than $|C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 h_s}|$ and $|C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 a_s}|$, respectively. The other is that, since $m_{\tilde{\chi}_i^0} > m_{\tilde{\chi}_1^0}$, the former contribution is relatively suppressed by the propagator.}. According to Eqs. (\ref{eq1:hichi01chi01_S}) and (\ref{eq1:aichi01chi01_S}), if $h_s$, $a_s$, and $\tilde{\chi}_1^0$ are pure singlet states, then $|C_{ \tilde{\chi}_1^0 \tilde{\chi}_1^0 h_i}| = |C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 a_i }| \sim \sqrt{2}|\kappa|$. The measured abundance then requires $\kappa \sim 0.15 \left( \frac{m_{\tilde{\chi}_1^0}}{300\,{\rm GeV}}\right)^{1/2} $ in the case of $|{\cal{A}}_t| \gg |{\cal{A}}_s|$ ~\cite{Baum:2017enm}.
Hence, it is evident that once the involved particle masses are fixed, the density is primarily determined by the parameter $\kappa$. However, because $\lambda > 2 |\kappa|$ to ensure a singlino-dominated $\tilde{\chi}_1^0$, the measured abundance can set a lower bound on $\lambda$.
To date, the sensitivities of the XENON-1T experiments have reached the precision of $10^{-47}~{\rm cm^2}$ for the SI cross section~\cite{Aprile:2018dbl} and $10^{-42}~{\rm cm^2}$ for the SD cross section~\cite{Aprile:2019dbj}. They have strongly restricted the $\lambda \gtrsim 0.3$ case (see the discussion about DM-nucleon scattering), which is preferred by the annihilations $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \to t \bar{t}, h_s a_s$ to account for the measured abundance. In addition, the LHC searches for new particles~\cite{Sirunyan:2018lul,Sirunyan:2018nwe,Sirunyan:2018ubx,Sirunyan:2017eie,Sirunyan:2017lae,Sirunyan:2017qaj,Aaboud:2018jiw,Aaboud:2018ngk,Aaboud:2018sua,
Aad:2019vvi,Aad:2019vnb,Aad:2019vvf,ATLAS:2013rla,Sirunyan:2018iwl,TheATLAScollaboration:2013zia,Aad:2014vma,Aad:2014nua,Aad:2015jqa,CMS:2013bda,CMS:2013dea}, and its precise measurement of the discovered scalar's properties~\cite{Aad:2019mbh,Sirunyan:2018koj} have limited the theoretical prediction on light sparticles and Higgs bosons. This has a significant impact on the DM annihilation channels since they are usually accompanied with light particles to account for the abundance (see the discussion in~\cite{Cao:2018rix}). Thus, the scenario preferred by the scan of Ref.~\cite{Cao:2018rix} involves a singlino-dominated DM with $\lambda \lesssim 0.05$. In this case, an effective mechanism to obtain the measured abundance includes co-annihilation with higgsinos. The corresponding reaction is $\tilde{\chi}_i \tilde{\chi}_j \rightarrow X X^\prime $, in which $ XX^\prime$ denotes SM particles and $\tilde{\chi}_i \tilde{\chi}_j$ may be an LSP-NLSP or NLSP-NLSP annihilation state (e.g., $\tilde{\chi}_1^0 \tilde{\chi}^0_2$, $\tilde{\chi}_1^0 \tilde{\chi}^+_1$, or $\tilde{\chi}_2^0 \tilde{\chi}^+_1$). This mechanism is distinct in that the effective annihilation rate at a temperature $T$ is very sensitive to the $\tilde{\chi}_1^0$-higgsino mass splitting~\cite{Griest:1990kh,Baker:2015qna}, and even for a small $\lambda$ and $\kappa$, it can still explain the measured abundance\footnote{Note that the co-annihilation mechanism applies under the premise that $\tilde{\chi}_1^0$ and the higgsinos remained in thermal equilibrium in the early universe~\cite{Griest:1990kh,Baker:2015qna}. In the $Z_3$-NMSSM, many processes, such as $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \leftrightarrow \tilde{\chi}_i \tilde{\chi}_j$, $\tilde{\chi}_1^0 X \leftrightarrow \tilde{\chi}_i X^\prime$, and $\tilde{\chi}_i \leftrightarrow \tilde{\chi}_1^0 X X^\prime$, could keep $\tilde{\chi}_1^0$ in chemical equilibrium with $\tilde{\chi}_i$, and the conversion rates of some of them might be enhanced if the mediator were around its mass-shell. We add that maintaining the thermal equilibrium does not necessarily require the involved couplings to be moderately large. For example, the equilibrium condition was discussed in Eqs. (4.3) and (4.4) of Ref.~\cite{Baker:2015qna} in the framework of the DM model ST11. It was found that the involved coupling may be as low as $10^{-4}$ to maintain the equilibrium.}.
\subsection{DM-nucleon cross sections}
Serving for a WIMP, $\tilde {\chi}^0_1$ might be detected by measuring the recoil of a nucleus after an elastic scattering of $\tilde {\chi}^0_1$ on a nucleus taking place. In the non-relativistic limit, only two different kinds of interactions between a neutralino and a nucleon need to be considered~\cite{Jungman:1995df}: the spin-dependent interaction (SD) where the WIMP couples to the spin of the nucleus, and the spin-independent interaction (SI) where the WIMP couples to the mass of the nucleus.
When $m_{\tilde{q}} \gtrsim 2~{\rm TeV}$, only the $t$-channel $Z$ exchange diagram contributes significantly to the spin-dependent (SD) scattering cross section at the tree level, which is approximated by~\cite{Pierce:2013rda,LDM-27}
\begin{align}
\label{eq:sigSD}
\sigma_{\tilde{\chi}_1^0-N}^{\rm SD} \simeq C_N \times \left ( \frac{C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 Z}}{0.01} \right )^2,
\end{align}
with $N=p(n)$ denoting protons (neutrons) and $C_p \simeq 2.9 \times 10^{-41}~{\rm cm^2} $ ($C_n \simeq 2.3 \times 10^{-41}~{\rm cm^2} $)~\cite{Badziak:2015exr,Badziak:2017uto}.
From the expression for $C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 Z}$ in Eq. (\ref{eq1:zchi10chi10_S}), it is evident that $\sigma_{\tilde{\chi}_1^0-N}^{\rm SD}$ is proportional to $(\lambda v /\mu_{eff})^4$. Furthermore, in the co-annihilation case, the degeneracy of $m_{\tilde{\chi}_1^0}$ and $\mu_{eff}$ leads to a minuscule denominator in Eq. (\ref{eq1:zchi10chi10_S}), which requires a small value of $\lambda v /\mu_{eff}$ to satisfy the DM-DD experimental constraints.
In contrast, the spin-independent (SI) scattering cross section in the heavy squark limit is dominated by a $t$-channel exchange of CP-even Higgs bosons $h_i$~\cite{Drees1993,Drees1992,Jungman1995,Belanger2008} and can be expressed as~\cite{cao:2021Gnmssm}
\begin{equation} \label{eq:SIDD_p}
\sigma_{\tilde {\chi}^0_1-{N}}^{\rm SI} = \frac{ m_N^2}{2\pi v^2} \left( \frac{m_N m_{\tilde{\chi}_1^0}}{m_N + m_{\tilde{\chi}_1^0}} \right)^2 {\left( \frac{1}{125~\rm GeV} \right)^4}\left\{ \sum_{h_i} \left[ F^N_u \left({a_u}\right)_{h_i} + F^N_d \left({a_d}\right)_{h_i} \right] \right\}^2,
\end{equation}
where $m_N$ is the nucleon mass, $F^{(N)}_d=f^{(N)}_d+f^{(N)}_s+\frac{2}{27}f^{(N)}_G$ and $F^{(N)}_u=f^{(N)}_u+\frac{4}{27}f^{(N)}_G$ with $f^{(N)}_q =m_N^{-1}\left<N|m_qq\bar{q}|N\right> $ ($q=u,d,s$) represent the normalized light quark contribution to the nucleon mass, and $f^{(N)}_G=1-\sum_{q=u,d,s}f^{(N)}_q$ influences other heavy quark mass fractions in nucleons~\cite{Drees1993,Drees1992}. In this study, the default settings for $f_q^{N}$ were used in the micrOMEGAs package~\cite{Belanger2008}, and they predict $F_u^{p} \simeq F_u^n \simeq 0.15$ and $F_d^{p} \simeq F_d^n \simeq 0.13$. Hence, SI cross sections for DM-proton scattering and DM-neutron scattering are approximately equal (i.e., $\sigma^{SI}_{{\tilde {\chi}^0_1}-p} \simeq \sigma^{SI}_{{\tilde {\chi}^0_1}-n}$)~\cite{DM-detecion-SI-SD}. The quantities
$\left( a_u\right)_{h_i}$ and $\left( a_d\right)_{h_i}$ are defined by
\begin{eqnarray}
\left( {a_u} \right)_{h_i} &=& \left( \frac{125~\rm GeV}{m_{h_i} }\right)^2 \left(V_{h_i}^{\rm SM}+\frac{1}{\tan\beta} V_{h_i}^{\rm NSM} \right) C_{ \tilde {\chi}^0_1 \tilde {\chi}^0_1 h_i} ~, \\
\left( {a_d}\right)_{h_i} &=& \left( \frac{125~\rm GeV}{ m_{h_i}}\right)^2 \left( V_{h_i}^{\rm SM}- \tan\beta V_{h_i}^{\rm NSM}\right) C_{ \tilde {\chi}^0_1 \tilde {\chi}^0_1 h_i} ~.
\end{eqnarray}
Currently, non-SM doublet Higgs bosons $H$ are preferred to be heavier than several hundreds of GeV in LHC experiments. In this case, the contribution from $H$ to the SI cross section will be suppressed by $(a_q)_{H}^2 \propto 1/m_{H}^4$, and it is much smaller than that from $h$ for a not exceedingly large $\tan \beta$. As such, the primary contribution to SI scattering comes from the t-channel exchange of SM-like Higgs bosons $h$ and the singlet Higgs boson $h_s$. The latter contribution may be crucial when $h_s$ is much lighter than $h$. Since the non-SM doublet components of $h$ and $h_s$ are approximately zero, $\left( {a_u} \right)_{h} + \left( {a_u} \right)_{h_{s}} \simeq \left( {a_d}\right)_{h} + \left( {a_d}\right)_{h_{s}}
\equiv \mathcal{A }$, which can be expressed as
\begin{align}\label{eq0:au-ad_hsm-hs}
\mathcal{A} \simeq \left( \frac{125 \rm GeV}{m_{h}}\right)^2 V_{h}^{\rm SM} C_{ \tilde {\chi}^0_1 \tilde {\chi}^0_1 h} + \left( \frac{125 \rm GeV}{m_{h_{s}}} \right)^2 V_{h_{s}}^{\rm SM} C_{ \tilde {\chi}^0_1 \tilde {\chi}^0_1 h_{s}}~,
\end{align}
where $C_{ \tilde {\chi}^0_1 \tilde {\chi}^0_1 h}$ and $C_{ \tilde {\chi}^0_1 \tilde {\chi}^0_1 h_{s}}$ are given by Eq. (\ref{eq1:hichi01chi01_S}).
Thus, the SI scattering cross section in Eq. (\ref{eq:SIDD_p}) can be rewritten as
\begin{align}
\sigma_{\tilde {\chi}^0_1-{N}}^{\rm SI} &\simeq \frac{m_N^2}{2 v^2\pi} \left( \frac{m_N m_{\tilde{\chi}_1^0}}{m_N + m_{\tilde{\chi}_1^0}} \right)^2 {\left( \frac{1}{125 \rm GeV} \right)^4} \left( F^N_u + F^N_d \right)^2 \mathcal{A}^2\sim 5 \times 10^{-45} {\rm cm^2}\times \left(\frac{\mathcal{A}}{0.1}\right)^2.
\label{eq:SIDD_p2}
\end{align}
In these expressions, if only the contribution from a pure SM Higgs state is considered, $\mathcal{A}$ can be simply expressed as
\begin{align}\label{eq0:au-ad_hsm}
\mathcal{A} \sim \left( \frac{125 \rm GeV}{m_{h}}\right)^2 \frac{\sqrt{2}\lambda^2 v}{\mu_{eff}} \, \frac { Z_s (m_{\tilde{\chi}_1^0}/\mu_{eff} -\sin 2 \beta)}{1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2}.
\end{align}
It is immediately evident that $\sigma_{\tilde{\chi}_1^0-N}^{\rm SI}$ will vanish for $m_{\tilde{\chi}_1^0}/\mu_{eff}=\sin2\beta$, which corresponds to a blind spot (BS) condition in~\cite{Badziak:2015exr, Badziak:2016qwg, Baum:2017enm}.
The above analytical formulas for the SD and SI scattering cross sections suggest that $\sigma^{SD} \varpropto \big( \frac{\lambda v}{\mu_{eff}} \big)^2\frac{1}{1-(m_{\tilde{\chi}_1^0}/\mu_{eff})^2}$ and (by contrast) $\sigma^{SI}$ depends on $\lambda$, $\mu_{eff}$, and $m_{\tilde{\chi}^0_1}$ in a complex way. In general, a larger $\lambda$, a smaller $\mu_{eff}$, and $m_{\tilde{\chi}_1^0}/\mu_{eff} \rightarrow 1$ will increase $\sigma^{SI}$ and thus strengthen the DM-DD experimental constraints.
\begin{table}[t]
\centering
\caption{Signal of final state for the electroweakino pair production processes considered in this work. Relevant experimental analyses were performed using a simplified model by ATLAS and CMS collaborations, and their results have been encoded in the SmodelS-1.2.3~\cite{Khosa:2020zar}.}
\label{tab:my-table}
\vspace{0.3cm}
\resizebox{1\textwidth}{!}{%
\renewcommand\arraystretch{0.9}
\begin{tabular}{lccr}
\hline\hline
\bf Name & \bf Simplified Scenario & \bf Signal of Final State & \bf Luminosity \bf($\bm {fb^{-1}}$\bf) \\ \hline
\multicolumn{4}{c}{\bf13 TeV} \\ \hline
\begin{tabular}[l]{@{}l@{}}\bf CMS-SUS-17-010~\cite{Sirunyan:2018lul}\\ (arXiv:1807.07799)\end{tabular} &\begin{tabular}[c]{@{}c@{}}$\bm{\tilde{\chi}_1^{\pm}\tilde{\chi}_1^{\mp}\rightarrow W^{\pm}\tilde{\chi}_1^0 W^{\mp}\tilde{\chi}_1^0}$\\$\bm{\tilde{\chi}_1^{\pm}\tilde{\chi}_1^{\mp}\rightarrow \nu\tilde{\ell} \ell\tilde{\nu}}$ \\ \end{tabular}&\bf2$\bm \ell$ \bf + $\bm{E_{\rm T}^{\rm miss}}$ & \bf 35.9 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf CMS-SUS-17-009~\cite{Sirunyan:2018nwe}\\ (arXiv:1806.05264)\end{tabular} &$\bm{\tilde{\ell}\tilde{\ell}}$ &\bf2$\bm \ell$ \bf + $\bm{E_{\rm T}^{\rm miss}}$ & \bf 35.9 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf CMS-SUS-17-004~\cite{Sirunyan:2018ubx}\\ (arXiv:1801.03957)\end{tabular} &$\bm{\tilde{\chi}_{2}^0\tilde{\chi}_1^{\pm}\rightarrow Wh(Z)\tilde{\chi}_1^0\tilde{\chi}_1^0$} & \bf n$\bm \ell$\bf(n\textgreater{}=0) + \bf nj(n\textgreater{}=0) + $\bm E_{\rm T}^{\rm miss}$ & \bf 35.9 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf CMS-SUS-16-045~\cite{Sirunyan:2017eie}\\ (arXiv:1709.00384)\end{tabular} &$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow W^{\pm}\tilde{\chi}_1^0h\tilde{\chi}_1^0}$& \bf 1$\bm \ell$\bf 2b + $\bm E_{\rm T}^{\rm miss}$ & \bf 35.9 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf CMS-SUSY-16-039~\cite{Sirunyan:2017lae}\\(arxiv:1709.05406) \end{tabular} &\begin{tabular}[c]{@{}c@{}c@{}c@{}c@{}} $\bm {\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow \ell\tilde{\nu}\ell\tilde{\ell}}$\\$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow\tilde{\tau}\nu\tilde{\ell}\ell}$\\$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow\tilde{\tau}\nu\tilde{\tau}\tau}$\\ $\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow WZ\tilde{\chi}_1^0\tilde{\chi}_1^0}$\\$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow WH\tilde{\chi}_1^0\tilde{\chi}_1^0}$\end{tabular} & \bf n$\bm{\ell(n\textgreater{}0)}$(\bm{$\tau}$) \bf + $\bm{E_{\rm T}^{\rm miss}}$ & \bf 35.9 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf CMS-SUS-16-034~\cite{Sirunyan:2017qaj}\\ (arXiv:1709.08908)\end{tabular}&$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow W\tilde{\chi}_1^0Z(h)\tilde{\chi}_1^0}$ & \bf n$\bm{\ell}$\bf (n\textgreater{}=2) + nj(n\textgreater{}=1) $\bm{E_{\rm T}^{\rm miss}}$ & \bf 35.9 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-1803-02762~\cite{Aaboud:2018jiw}\\ (ATLAS-SUSY-2016-24)\end{tabular} &\begin{tabular}[c]{@{}c@{}c@{}c@{}}$ \bm {\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow WZ\tilde{\chi}_1^0\tilde{\chi}_1^0}$\\$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow \nu\tilde{\ell}l\tilde{\ell}}$\\$\bm {\tilde{\chi}_1^{\pm}\tilde{\chi}_1^{\mp}\rightarrow \nu\tilde{\ell}\nu\tilde{\ell}}$\\ $\bm{ \tilde{\ell}\tilde{\ell}}$\end{tabular} & \bf n$\bm \ell$\bf (n\textgreater{}=2) + $\bm E_{\rm T}^{\rm miss}$ & \bf 36.1 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-1812-09432~\cite{Aaboud:2018ngk}\\ (ATLAS-SUSY-2017-01)\end{tabular} &$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow Wh\tilde{\chi}_1^0\tilde{\chi}_1^0}$ & \bf n$\bm \ell$\bf (n\textgreater{}=0) + nj(n\textgreater{}=0) + nb(n\textgreater{}=0) + n$\bm \gamma$\bf (n\textgreater{}=0) + $\bm E_{\rm T}^{\rm miss}$ & \bf 36.1 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-1806-02293~\cite{Aaboud:2018sua}\\ (ATLAS-SUSY-2017-03)\end{tabular} &$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow WZ\tilde{\chi}_1^0\tilde{\chi}_1^0}$ & \bf n$\bm \ell$\bf (n\textgreater{}=2) + nj(n\textgreater{}=0) + $\bm E_T^{miss}$ & \bf 36.1 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-1912-08479~\cite{Aad:2019vvi}\\ (ATLAS-SUSY-2018-06)\end{tabular} &$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow W(\rightarrow l\nu)\tilde{\chi}_1^0Z(\rightarrow\ell\ell)\tilde{\chi}_1^0}$& \bf 3$\bm \ell $ \bf + $\bm E_{\rm T}^{\rm miss}$ & \bf 139 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-1908-08215~\cite{Aad:2019vnb}\\ (ATLAS-SUSY-2018-32)\end{tabular} &\begin{tabular}[c]{@{}c@{}}$\bm{\tilde{\ell}\tilde{\ell}}$\\$\bm{\tilde{\chi}_1^{\pm}\tilde{\chi}_1^{\mp}}$ \\ \end{tabular} & \bf 2$\bm \ell$ \bf + $\bm E_{\rm T}^{\rm miss}$ & \bf 139 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-1909-09226~\cite{Aad:2019vvf}\\ (ATLAS-SUSY-2019-08)\end{tabular} & $\bm{\tilde{\chi}_{2}^0\tilde{\chi}_1^{\pm}\rightarrow Wh\tilde{\chi}_1^0\tilde{\chi}_1^0}$ & \bf 1$\bm \ell$ \bf + h\bf($\bm \rightarrow$\bf bb) + $\bm E_{\rm T}^{\rm miss}$ & \bf 139 \\ \hline
\multicolumn{4}{c}{\bf 8 TeV} \\ \hline
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-CONF-2013-035~\cite{ATLAS:2013rla}\end{tabular}&\begin{tabular}[c]{@{}c@{}}$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow Z^{(*)}\tilde{\chi}_1^0W^{(*)}\tilde{\chi}_1^0}$\\$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow
\tilde{l}(\tilde{\nu})\ell(\nu)\tilde{\ell}(\tilde{\nu})\nu(\tilde{\ell})}$ \end{tabular} & \bf n$\bm \ell $\bf (n\textgreater{}=2) + $\bm E_{\rm T}^{\rm miss}$ & \bf 20.3 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-CONF-2013-049~\cite{Sirunyan:2018iwl}\\(arxiv:1801.01846) \end{tabular} &\begin{tabular}[c]{@{}c@{}} $\bm{\tilde{\ell}\tilde{\ell}}$\\$\bm{\tilde{\chi}_1^{\pm}\tilde{\chi}_1^{\mp}\rightarrow \tilde{\ell}\nu(\tilde{\nu}\ell)}$\end{tabular} &\bf 2$\bm \ell$ \bf + $\bm E_{\rm T}^{\rm miss}$ & \bf 20.3 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-CONF-2013-093~\cite{TheATLAScollaboration:2013zia} \end{tabular} &\begin{tabular}[c]{@{}c@{}} $\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow W\tilde{\chi}_1^0h\tilde{\chi}_1^0}$\end{tabular} & \bf 1$\bm \ell$ \bf + 2$\bm b$ \bf + $\bm E_{\rm T}^{\rm miss}$ & \bf 20.3 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-1403-5294~\cite{Aad:2014vma}\\(ATLAS-SUSY-2013-11)\end{tabular}&\begin{tabular}[c]{@{}c@{}c@{}c@{}}$\bm{\tilde{\chi}_1^{\pm}\tilde{\chi}_1^{\mp}\rightarrow \tilde{l}(\tilde{\nu})\nu(l)}$\\$\bm{\tilde{\chi}_1^{\pm}\tilde{\chi}_1^{\mp}\rightarrow
W\tilde{\chi}_1^0 W\tilde{\chi}_1^0}$\\$\bm{\tilde{\chi}_1^{\pm}\tilde{\chi}_2^{0}\rightarrow W\tilde{\chi}_1^0 Z\tilde{\chi}_1^0}$\\$\bm{\tilde{\ell}\tilde{\ell}}$ \end{tabular} & \bf n$\bm \ell$\bf (n\textgreater{}=2) + $\bm{E_{\rm T}^{\rm miss}}$ & \bf 20.3 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-1402-7029~\cite{Aad:2014nua}\\ (ATLAS-SUSY-2013-12)\end{tabular} &\begin{tabular}[c]{@{}c@{}c@{}c@{}}$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow W\tilde{\chi}_1^0Z\tilde{\chi}_1^0}$\\
$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow W\tilde{\chi}_1^0h\tilde{\chi}_1^0}$\\
$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow \ell\tilde{\nu}\ell\tilde{\ell}(\tilde{\nu}\nu)}$,$\bm {\ell\tilde{\ell}\nu\tilde{\ell}(\tilde{\nu}\nu)}$\\
$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow\tilde{\tau}\nu\tilde{\tau}\tau(\tilde{\nu}\nu)}$,$\bm{\tau\tilde{\nu}\tilde{\tau}\tau(\tilde{\nu}\nu)}$\end{tabular} & \bf 3$\bm \ell(\tau)$ \bf + $\bm E_{\rm T}^{\rm miss}$ & \bf 20.3 \\ \\
\begin{tabular}[l]{@{}l@{}}\bf ATLAS-1501-07110~\cite{Aad:2015jqa}\\ (ATLAS-SUSY-2013-23)\end{tabular}&$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow W\tilde{\chi}_1^0h\tilde{\chi}_1^0}$ & \bf n$\bm\ell$
\bf (n$\bm \textgreater{}0$\bf ) + n$\bm \gamma$\bf (n\textgreater{}=0) + n$\bm b$\bf (n\textgreater{}=0) + $\bm E_{\rm T}^{\rm miss}$ & \bf 20.3 \\ \\
\begin{tabular}[l]{@{}c@{}}\bf CMS-PAS-SUSY-12-022\cite{CMS:2013bda}\\\end{tabular} &\begin{tabular}[c]{@{}c@{}c@{}c@{}c@{}}
$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow \tilde{\ell}\ell\tilde{\nu}\ell}$,$\bm{\ell\tilde{\ell}\nu\tilde{\ell}}$\\
$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow Z\tilde{\chi}_1^0 W\tilde{\chi}_1^0}$\\$\bm{\tilde{\chi}_2^0\tilde{\chi}_3^0\rightarrow Z\tilde{\chi}_1^0 Z\tilde{\chi}_1^0}$\\$\bm{\tilde{\chi}_1^{\pm}\tilde{\chi}_1^{\mp}\rightarrow \ell\tilde{\nu}\nu\tilde{\ell}}$\\$\bm{\tilde{\ell}\tilde{\ell}}$\end{tabular} &\bf n$\bm \ell$\bf (n$\bm \textgreater{}=2$\bf ) + $\bm E_{\rm T}^{\rm miss}$& \bf 9.2 \\ \\
\begin{tabular}[l]{@{}c@{}}\bf CMS-SUSY-13-006~\cite{CMS:2013dea}\\\end{tabular} &\begin{tabular}[c]{@{}c@{}c@{}c@{}}
$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow \tilde{\ell}\ell\tilde{\nu}\ell}$,$\bm{\ell\tilde{\ell}\nu\tilde{\ell}}$\\
$\bm{\tilde{\chi}_2^0\tilde{\chi}_1^{\pm}\rightarrow Z\tilde{\chi}_1^0 W\tilde{\chi}_1^0}$\\$\bm{\tilde{\chi}_1^{\pm}\tilde{\chi}_1^{\mp}\rightarrow \ell\tilde{\nu}\nu\tilde{\ell}}$\\$\bm{\tilde{\ell}\tilde{\ell}}$\end{tabular} & \bf n$\bm \ell$\bf (n\textgreater{}=2) + $\bm E_{\rm T}^{\rm miss}$& \bf 19.5 \\ \hline\hline
\end{tabular}}
\end{table}
\section{\label{sec:scan} Model scans and constraints}
The \texttt{NMSSMTools-5.4.1} package~\cite{Ellwanger:2004xm,Ellwanger:2005dv} was used to produce samples of singlino-dominated DM scenarios in the $\rm Z_3$-NMSSM and to model the corresponding features in detail\footnote{The \texttt{NMSSMTools} package includes codes to compute various observables in Higgs physics, DM physics, B physics, and sparticle physics. In this sector, we only briefly introduce the calculation of the observables we are interested in.}. A sophisticated scan was first performed over the following ranges in the parameter space:
\begin{eqnarray}\label{parameter-scan}
&& 0 < \lambda \leq 0.7,~~|\kappa| \leq 0.7,~~1 \leq \tan \beta \leq 60,~~100 ~{\rm GeV} \leq \mu_{eff} \leq 1000~{\rm GeV}, \nonumber
\\
&& |A_\kappa| < 1 ~{\rm TeV},~~0 <A_\lambda \leq 5~{\rm TeV}, ~~|A_t| \leq 5~{\rm TeV}, ~~|M_1| \leq 500~{\rm GeV}, \label{Parameters}
\end{eqnarray}
in which all parameters were defined at the scale $Q = 1~{\rm TeV}$. Upper bounds of $0.7$ were imposed on $\lambda$ and $|\kappa|$ to maintain a perturbable theory up to the grand unification scale. A lower bound of $100~{\rm GeV}$ was placed on $\mu_{eff}$ by the LEP search for electroweakinos~\cite{Tanabashi:2018oca}, and an upper bound of 1000 GeV for $\mu_{eff}$ is large enough to allow us to consider various possibilities (see the discussion presented below). In addition, noting that the LHC search for SUSY prefers massive charged sparticles, the following assumptions were made concerning unimportant SUSY parameters. The electroweak gaugino masses were set to $M_2=2~{\rm TeV}$, and the gluino masses were set to $M_3=5~{\rm TeV}$. Soft SUSY-breaking parameters in the squarks sector were fixed at $2~{\rm TeV}$, excluding trilinear couplings $A_t = A_b$ used as free parameters to adjust the Higgs mass spectrum to coincide with relevant experimental measurements at the LHC. In addition, all slepton soft parameters were set to $2~{\rm TeV}$, as we did not want to explain the muon $g$-2 anomaly. We also required $\lambda \geq 2 |\kappa|$ in the scan to achieve a singlino-dominated $\tilde\chi^0_1$.
Specifically, the MultiNest algorithm~\cite{Feroz:2008xx,Feroz:2013hea} with flat distributions for all the parameters in Eq.~(\ref{Parameters}) and {\it nlive} = 20000 were adopted during the scan to ensure that the conclusions were as complete as possible, and more than 20 thousands CPU hours were spent on the calculations\footnote{The multi-nest sampling algorithm explores a high-dimensional parameter space by determining the iso-likelihood contour in each iteration with {\it nlive} active points (the integer {\it nlive} is an input parameter of the algorithm, and it usually takes a value larger than 1000. In general, the larger value {\it nlive} adopts, the more accurate the scan result becomes.)~\cite{Feroz:2008xx,Feroz:2013hea}. It is good at dealing with the case in which the samples' posterior distribution is multi-modal or degenerate, which is frequently encountered in new physics studies. In contrast, the Markov Chain method~\cite{Markov-Chain} is highly inefficient for such a situation, and thus, it usually provides incomplete information about the distribution.}. Several constraints were imposed by constructing the following corresponding likelihood function to guide the process:
\begin{eqnarray}
\mathcal{L}=\mathcal{L}_{\rm{m_{h}}} \times \mathcal{L}_{h, {\rm coupling}} \times {\mathcal{L}_{B} } \times {\mathcal{L}_{EW} } \times {\mathcal{L}_{\Omega h^2} }\times {\mathcal{L}_{DD}},
\label{Likelihood}
\end{eqnarray}
where $\mathcal{L}_{\rm{m_{h}}}$ and $\mathcal{L}_{h, {\rm coupling}}$ are likelihood functions for the experimentally measured SM-like Higgs boson mass and couplings, respectively.
The computation of $m_h$ included leading electroweak corrections, two loop terms, and propagator corrections, as in Ref. \cite{Degrassi:2009yq}. Its experimental central value was taken as $m_h = 125.09~{\rm GeV}$~\cite{Aad:2015zhl}, and a total experimental and theoretical uncertainty of $3~{\rm GeV}$ was assumed. $\mathcal{L}_{h, {\rm coupling}}$ works in a seven-parameter $\kappa$-framework with related experimental measurements, such as the central values and uncertainties of the Higgs couplings and their correlation coefficients, taken from the ATLAS analysis, using 80${\rm fb}^{-1}$ data collected during the LHC Run-II~\cite{Aad:2019mbh}. Some knowledge about probability and statistics was used in constructing $\mathcal{L}_{h, {\rm coupling}}$ (see the introduction in Ref. \cite{Tanabashi:2018oca}). ${\mathcal{L}_{B} }$ is the likelihood function for the measurement of the branching ratio for the decays $B \to X_s \gamma$ and $B_s \to \mu^+\mu^-$. These ratios were calculated by the formulae in Refs.~\cite{Domingo:2007dx,Domingo:2015wyn}, and their experimental values were taken from Ref. \cite{Tanabashi:2018oca}. ${\mathcal{L}_{\Omega h^2} }$ and ${\mathcal{L}_{DD}}$ are likelihood functions for the measured abundance from the WMAP/Planck experiments~\cite{Ade:2015xua, Aghanim:2018eyx} and the detection of both spin-independent (SI) and spin-dependent (SD) DM-nucleon scattering in the XENON-1T experiment~\cite{Aprile:2018dbl,Aprile:2019dbj}. Relevant quantities
were calculated using the \texttt{micrOMEGAs} package~\cite{Belanger:2004yn,Belanger:2005kh,Belanger:2006is,Belanger:2013oya,Belanger:2018ccd}.
In addition, $ {\mathcal{L}_{EW} }$ denotes a likelihood function for precision electroweak observables of $\epsilon_i$ (i=1,2,3)~\cite{Altarelli:1990zd,Altarelli:1991fk,Altarelli:1994iz} or, equivalently, $S$, $T$, and $U$ parameters~\cite{Peskin:1990zt,Peskin:1991sw} calculated using the formulas from Cao and Yang~\cite{Cao:2008rc} and fitted to corresponding measurements by the procedure presented in Ref. \cite{deBlas:2016ojx}. Each of these likelihood functions was assumed to follow a Gaussian distribution, with explicit representations provided in Cao et al.~\cite{Cao:2018iyk}.
The acquired samples were further refined using the following criteria: the SM-like Higgs mass was within the range of 122--128 GeV, the observed DM relic abundance was within $\pm 10\%$ of the measured central value $\Omega h^2 = 0.1187$~\cite{Aghanim:2018eyx}, the upper bound was a $90\%$ confidence level of the XENON-1T experimental results of the SI cross section~\cite{Aprile:2018dbl} and the SD cross section~\cite{Aprile:2019dbj}, and all other constraints implemented in \texttt{NMSSMTools}, including various B-physics observables in corresponding experimentally allowed ranges, were at the $2\sigma$ level. We also required the samples to satisfy $\chi^2_{EW} \leq 7.8$ and $\chi^2_{h, {\rm coupling}} \leq 14.1$, which corresponded to $95\%$ confidence level exclusion limits for three and seven degrees of freedom, respectively. Since a more significant deviation of the electroweak precision observables (the Higgs couplings) from their measured values would enhance $\chi^2_{EW}$ ($\chi^2_{h, {\rm coupling}}$), this requirement delineated the experimentally allowed range of these observables to further limit the $Z_3$-NMSSM.
Constraints were also implemented from the LHC search for electroweakinos using the \texttt{SModelS-1.2.3} code~\cite{Ambrogi:2017neo, Kraml:2013mwa,Khosa:2020zar}. The final states listed in Table~\ref{tab:my-table} from all the electroweakino pair production processes were considered during this process\footnote{The concrete procedure to determine the limitation is as follows: we first determined the signal region (SR) with the largest expected sensitivity for a given sample, then we checked its $R$ value defined by $R \equiv S/S_{95}^{OBS}$, where $S$ stands for the number of signal events in the SR with the
statistical uncertainty considered, and $S_{95}^{OBS}$ denotes the observed limit at 95\% confidence level for the SR. Evidently, $R$ represents the capability
of the LHC in exploring a point. $R > 1$ implies that the point is excluded; otherwise, it is allowed.}. It is worth noting that a large portion of the samples satisfying these constraints were characterized by a small mass splitting between $\tilde{\chi}_1^0$ and $\tilde{\chi}_2^0$. However, the latest ATLAS analysis of the compressed mass spectra, acquired by searching soft di-lepton signals~\cite{Aad:2019qnd}, was not included in the \texttt{SModelS-1.2.3} code. As such, the analysis constraints are validated in the Appendix, and they were applied to each sample by elaborate Monte Carlo simulations. We verified that they were very effective in excluding the co-annihilation case. We illustrated this point in our recent publication~\cite{cao:2021Gnmssm}.
We add that we did not consider constraints from indirect DM detection experiments (i.e., the Fermi-LAT search for DM annihilation from dwarf spheroidal galaxies) as they become loose for $m_{\tilde \chi^0_1} > 100~\rm GeV $~\cite{Ackermann:2015zua}. In addition, in the co-annihilation case encountered in this work, the annihilation rate of singlino-dominated DM in present day is very small, which weakens the constraints.
\begin{table}[t]
\centering
\caption{Benchmark points satisfying various experimental constraints. Mass parameters are in units of GeV, and the DM-nucleon scattering cross sections are in units of ${\rm cm^2}$. The number preceding each annihilation process represents its fraction of contributions to the total DM annihilation cross section at the freeze-out temperature. Dots include the information of the decay modes with smaller branching ratios. } \label{table2}
\vspace{0.2cm}
\scalebox{0.7}{
\begin{tabular}{lcclcccccc}
\hline\hline
\multicolumn{5}{c|}{SM-like Higgs: $h_1$ } & \multicolumn{5}{c}{SM-like Higgs: $h_2$} \\ \hline
\multicolumn{1}{c}{} & \multicolumn{2}{c}{$P_1$} & \multicolumn{2}{c|}{$P_2$} & & \multicolumn{2}{c}{$P_3$} & \multicolumn{2}{c}{$P_4$} \\ \hline
\multicolumn{1}{l}{$\lambda$} & \multicolumn{2}{c}{0.696} & \multicolumn{2}{c|}{0.028} &
& \multicolumn{2}{c}{0.108} & \multicolumn{2}{c}{$9.64\times10^{-3}$} \\
\multicolumn{1}{l}{$\kappa$} & \multicolumn{2}{c}{0.208} & \multicolumn{2}{c|}{$-$0.013} &
& \multicolumn{2}{c}{0.048} & \multicolumn{2}{c}{$-$$4.2\times10^{-3}$}
\\
\multicolumn{1}{l}{$\tan\beta$} & \multicolumn{2}{c}{2.15} & \multicolumn{2}{c|}{7.2} &
& \multicolumn{2}{c}{12.8} & \multicolumn{2}{c}{9.9}
\\
\multicolumn{1}{l}{$\mu_{eff}$} & \multicolumn{2}{c}{683.7} & \multicolumn{2}{c|}{227.8} &
& \multicolumn{2}{c}{206.1} & \multicolumn{2}{c}{164.4}
\\
\multicolumn{1}{l}{$A_t$} & \multicolumn{2}{c}{$-$2734} & \multicolumn{2}{c|}{2754} &
& \multicolumn{2}{c}{$-$3394} & \multicolumn{2}{c}{$-$3510}
\\
\multicolumn{1}{l}{$A_\lambda$} & \multicolumn{2}{c}{1128} & \multicolumn{2}{c|}{3908} &
& \multicolumn{2}{c}{3125} & \multicolumn{2}{c}{1142}
\\
\multicolumn{1}{l}{$A_\kappa$} & \multicolumn{2}{c}{$-$28.3} & \multicolumn{2}{c|}{24.4} &
& \multicolumn{2}{c}{$-$323.1} & \multicolumn{2}{c}{147.3}
\\
\multicolumn{1}{l}{$M_1$} & \multicolumn{2}{c}{$-$471.3} & \multicolumn{2}{c|}{$-$418.2} &
& \multicolumn{2}{c}{$-$484.7} & \multicolumn{2}{c}{211.8}
\\
\hline
\multicolumn{1}{l}{$m_{\tilde{\chi}^0_1}$} & \multicolumn{2}{c}{422.4} & \multicolumn{2}{c|}{$-$216.5} &
& \multicolumn{2}{c}{181.9} & \multicolumn{2}{c}{$-$148.6}
\\
\multicolumn{1}{l}{$m_{\tilde {\chi}^0_2}$} & \multicolumn{2}{c}{$-$473.7} & \multicolumn{2}{c|}{$-$233.0} &
& \multicolumn{2}{c}{213.2} & \multicolumn{2}{c}{151.9}
\\
\multicolumn{1}{l}{$m_{\tilde {\chi}^0_3}$} & \multicolumn{2}{c}{702.3} & \multicolumn{2}{c|}{234.5} &
& \multicolumn{2}{c}{$-$215.1} & \multicolumn{2}{c}{$-$174.5}
\\
\multicolumn{1}{l}{$m_{\tilde {\chi}^\pm_1}$} & \multicolumn{2}{c}{694.9} & \multicolumn{2}{c|}{233.2} &
& \multicolumn{2}{c}{211.4} & \multicolumn{2}{c}{170.4}
\\
\multicolumn{1}{l}{$N_{13}$} & \multicolumn{2}{c}{$-$0.035} & \multicolumn{2}{c|}{$-$0.164} &
& \multicolumn{2}{c}{$-$0.32} & \multicolumn{2}{c}{$-$0.032}
\\
\multicolumn{1}{l}{$N_{14}$} & \multicolumn{2}{c}{0.176} & \multicolumn{2}{c|}{$-$0.171} &
& \multicolumn{2}{c}{$-$0.35} & \multicolumn{2}{c}{$-$0.038}
\\
\multicolumn{1}{l}{$N_{15}$} & \multicolumn{2}{c}{0.984} & \multicolumn{2}{c|}{0.971} &
& \multicolumn{2}{c}{0.87} & \multicolumn{2}{c}{0.999}
\\
\multicolumn{1}{l}{$Z_h$} & \multicolumn{2}{c}{0.032} & \multicolumn{2}{c|}{0.056} &
& \multicolumn{2}{c}{0.23} & \multicolumn{2}{c}{0.003}
\\
\multicolumn{1}{l}{$Z_S$} & \multicolumn{2}{c}{0.968} & \multicolumn{2}{c|}{0.943} &
& \multicolumn{2}{c}{0.76} & \multicolumn{2}{c}{0.997}
\\
\hline
\multicolumn{1}{l}{$m_{h_1}$} & \multicolumn{2}{c}{124.6} & \multicolumn{2}{c|}{125.2} &
& \multicolumn{2}{c}{64.2} & \multicolumn{2}{c}{100.7}
\\
\multicolumn{1}{l}{$m_{h_2}$} & \multicolumn{2}{c}{440.4} & \multicolumn{2}{c|}{206.5} &
& \multicolumn{2}{c}{126.7} & \multicolumn{2}{c}{125.3}
\\
\multicolumn{1}{l}{$m_{h_3}$} & \multicolumn{2}{c}{1536} & \multicolumn{2}{c|}{2532} &
& \multicolumn{2}{c}{2905} & \multicolumn{2}{c}{1311}
\\
\multicolumn{1}{l}{$m_{a_1}$} & \multicolumn{2}{c}{178.0} & \multicolumn{2}{c|}{88.1} &
& \multicolumn{2}{c}{298.0} & \multicolumn{2}{c}{178.4}
\\
\multicolumn{1}{l}{$m_{a_2}$} & \multicolumn{2}{c}{1536} & \multicolumn{2}{c|}{2532} &
& \multicolumn{2}{c}{2905} & \multicolumn{2}{c}{1311}
\\
\multicolumn{1}{l}{$V_{h_1}^{\rm NSM}, V_{h_2}^{\rm NSM}$} & \multicolumn{2}{c}{0.0, $-$0.05} & \multicolumn{2}{c|}{$-$0.0, $-$0.0} &
& \multicolumn{2}{c}{$-$0.0, $-$0.0} & \multicolumn{2}{c}{$-$0.0, $-$0.0}
\\
\multicolumn{1}{l}{$V_{h_1}^{\rm SM}, V_{h_2}^{\rm SM}$} & \multicolumn{2}{c}{0.99, 0.15} & \multicolumn{2}{c|}{0.99, $-$0.11} &
& \multicolumn{2}{c}{0.147, 0.989} & \multicolumn{2}{c}{0.04, 0.999}
\\
\multicolumn{1}{l}{$V_{h_1}^{\rm S}, V_{h_2}^{\rm S}$} & \multicolumn{2}{c}{$-$0.15, 0.99} & \multicolumn{2}{c|}{0.11, 0.99} &
& \multicolumn{2}{c}{0.989, $-$0.147} & \multicolumn{2}{c}{0.999, $-$0.04}
\\
\multicolumn{1}{l}{$V_{a_1}^{' \rm NSM}, V_{a_1}^{' \rm S}$} & \multicolumn{2}{c}{$-$0.03, 1.0} & \multicolumn{2}{c|}{$-$0.0, 1.0} &
& \multicolumn{2}{c}{$-$0.0, 1.0} & \multicolumn{2}{c}{$-$0.0, 1.0}
\\
\hline
\multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c|}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{}
\\
\multicolumn{1}{l}{$\sigma^{SI}_{\tilde\chi^0_1-p}$} & \multicolumn{2}{c}{$4.2 \times 10^{-49}$} & \multicolumn{2}{c|}{$3.1\times 10^{-47}$} &
& \multicolumn{2}{c}{$8.89 \times 10^{-47}$} & \multicolumn{2}{c}{$2.67 \times 10^{-49}$}
\\
\multicolumn{1}{l}{$\sigma^{SD}_{\tilde\chi^0_1-n}$} & \multicolumn{2}{c}{$3.0 \times 10^{-41}$} & \multicolumn{2}{c|}{$2.94 \times 10^{-43}$} &
& \multicolumn{2}{c}{$2.23 \times 10^{-41}$} & \multicolumn{2}{c}{$6.57 \times 10^{-45}$}
\\
\multicolumn{1}{l}{$\Omega h^2$} & \multicolumn{2}{c}{0.118} & \multicolumn{2}{c|}{0.109}
& & \multicolumn{2}{c}{0.12} & \multicolumn{2}{c}{0.108}
\\
\multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c|}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{}
\\
\multicolumn{1}{c}{} & \multicolumn{1}{l}{$68.4\%$} & \multicolumn{1}{l}{$\tilde\chi^0_1 \tilde\chi^0_1 \rightarrow t \bar{t}$} & \multicolumn{1}{l}{$14.0\%$} & \multicolumn{1}{l|}{$\tilde\chi^0_2 \tilde\chi^+_1 \rightarrow u \bar{d}$}
& \multicolumn{1}{l}{} & \multicolumn{1}{l}{$28.7\%$} & \multicolumn{1}{l}{$\tilde\chi^0_1 \tilde\chi^0_1 \rightarrow W^+ W^-$}& \multicolumn{1}{l}{$35.0\%$} & \multicolumn{1}{l}{$\tilde\chi^0_2 \tilde\chi^0_2 \rightarrow W^+ W^-$}
\\
\multicolumn{1}{c}{} & \multicolumn{1}{l}{$19.4\%$} & \multicolumn{1}{l}{$~~~~~~\rightarrow h_s a_s$} & \multicolumn{1}{l}{$5.4\%$} & \multicolumn{1}{l|}{$~~~~~~\rightarrow \nu_l \bar{l}$}
& \multicolumn{1}{l}{} & \multicolumn{1}{l}{$22.0\%$}
& \multicolumn{1}{l}{$~~~~~~\rightarrow Z Z$}& \multicolumn{1}{l}{$25.4\%$} & \multicolumn{1}{l}{$~~~~~~\rightarrow Z Z$}
\\
\multicolumn{1}{c}{} & \multicolumn{1}{l}{$5.0\%$} & \multicolumn{1}{l}{$~~~~~~\rightarrow h_{sm} a_s$} & \multicolumn{1}{l}{$13.0\%$} & \multicolumn{1}{l|}{$\tilde\chi^0_3 \tilde\chi^+_1 \rightarrow u \bar{d}$}
& \multicolumn{1}{l}{} & \multicolumn{1}{l}{$17.6\%$}
& \multicolumn{1}{l}{$~~~~~~\rightarrow t \bar{t}$} & \multicolumn{1}{l}{$16.2\%$} & \multicolumn{1}{l}{$\tilde\chi^0_2 \tilde\chi^+_1 \rightarrow u \bar{d}$}
\\
\multicolumn{1}{c}{} & \multicolumn{1}{l}{$2.2\%$} & \multicolumn{1}{l}{$~~~~~~ \rightarrow W^+ W^-$} & \multicolumn{1}{l}{$4.8\%$} & \multicolumn{1}{l|}{$~~~~~~\rightarrow \nu_l \bar{l}$} & \multicolumn{1}{c}{} & \multicolumn{1}{l}{$13.0\%$} & \multicolumn{1}{l}{$\tilde\chi^0_1 \tilde\chi^+_1 \rightarrow u \bar{d}$}& \multicolumn{1}{l}{$6.6\%$} & \multicolumn{1}{l}{$~~~~~~\rightarrow \nu_l \bar{l}$}
\\
\multicolumn{1}{l}{annihilation} & \multicolumn{1}{l}{...} & \multicolumn{1}{l}{~~~~~...} & \multicolumn{1}{l}{$11.4\%$} & \multicolumn{1}{l|}{$\tilde\chi^0_1 \tilde\chi^+_1 \rightarrow u \bar{d}$} & & \multicolumn{1}{c}{$4.8\%$} & \multicolumn{1}{l}{$~~~~~~\rightarrow \nu_l \bar{l}$}& \multicolumn{1}{l}{4.1\%} & \multicolumn{1}{l}{$\tilde\chi^0_2 \tilde\chi^0_3 \rightarrow q \bar{q}$}
\\
\multicolumn{1}{l}{channels} & \multicolumn{1}{c}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{4.2\%} & \multicolumn{1}{l|}{$~~~~~~\rightarrow \nu_l \bar{l}$}
& & \multicolumn{1}{l}{$3.8\%$} & \multicolumn{1}{l}{$\tilde\chi^0_1 \tilde\chi^0_3 \rightarrow q \bar{q}$}& \multicolumn{1}{l}{1.2\%} & \multicolumn{1}{l}{$~~~~~~\rightarrow \nu_l \bar{\nu}_l$}
\\
\multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{6.8\%} & \multicolumn{1}{l|}{$\tilde\chi^0_2 \tilde\chi^0_3 \rightarrow q \bar{q}$} & \multicolumn{1}{l}{}& \multicolumn{1}{l}{$...$} & \multicolumn{1}{l}{$~~~~~~...$}& \multicolumn{1}{l}{3.6\%} & \multicolumn{1}{l}{$\tilde\chi^0_1 \tilde\chi^0_2 \rightarrow q \bar{q}$}
\\
\multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{1.8\%} & \multicolumn{1}{l|}{$~~~~~~ \rightarrow \nu_l \bar{\nu}_l$}
& \multicolumn{1}{l}{}& \multicolumn{1}{l}{} & \multicolumn{1}{l}{}& \multicolumn{1}{l}{...} & \multicolumn{1}{l}{~~~~~...}
\\
\multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{7.8\%} & \multicolumn{1}{l|}{$\tilde\chi^+_1 \tilde\chi^-_1 \rightarrow q \bar{q}$} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{}
\\
\multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{2.7\%} & \multicolumn{1}{l|}{$~~~~~~\rightarrow l \bar{l}$} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{}& \multicolumn{1}{l}{} & \multicolumn{1}{l}{}
\\
\multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{2.2\%} & \multicolumn{1}{l|}{~~~~~~$\rightarrow W^+ W^-$}
& \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{}& \multicolumn{1}{c}{} & \multicolumn{1}{l}{}
\\
\multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{5.7\%} & \multicolumn{1}{l|}{$\tilde\chi^0_1 \tilde\chi^0_3\rightarrow q \bar{q}$}
& \multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{}
\\
\multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{1.5\%} & \multicolumn{1}{l|}{~~~~~~$\rightarrow \nu_l \bar{\nu}_l$}
& \multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{}
\\
\multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{~~...} & \multicolumn{1}{l|}{$~~~~~~...$}
& \multicolumn{1}{l}{} & \multicolumn{1}{c}{ } & \multicolumn{1}{l}{ } & \multicolumn{1}{l}{} & \multicolumn{1}{l}{}
\\
\multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{l|}{}
& \multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{}
\\
\hline\hline
\end{tabular}}
\end{table}
\begin{figure*}[t]
\centering
\resizebox{1.\textwidth}{!}{
\includegraphics[width=0.90\textwidth]{Fig1-a.png}
\includegraphics[width=0.90\textwidth]{Fig1-b.png}
}
\vspace{-0.6cm}
\caption{Type-I, -II, and -III samples projected on the $\sin2\beta-m_{\tilde{\chi}_1^0}/\mu_{eff}$ and $\lambda v / \mu_{eff}-Z_h$ planes. \label{fig1} }
\end{figure*}
\begin{figure*}[t]
\centering
\resizebox{1.03\textwidth}{!}{
\includegraphics[width=0.90\textwidth]{Fig2-a.png}
\includegraphics[width=0.90\textwidth]{Fig2-b.png}
}
\vspace{-0.6cm}
\caption{Type-I samples projected on the $\sigma_{\tilde{\chi}_1^0-p}^{SI}-m_{\tilde{\chi}_1^0}$ and $\sigma_{\tilde{\chi}_1^0-n}^{SD}-m_{\tilde{\chi}_1^0}$ planes. The color bar represents the values of $\lambda v /\mu_{eff}$. The solid line represents the current exclusion bound of the XENON-1T experiment on the cross section and the dashed line denotes the projected sensitivity of the near-future PandaX-4T experiment. \label{fig2} }
\end{figure*}
\section{\label{result13}Numerical Results}
A closer analysis suggested that the eventual surviving parameter space could be classified into the following three types:
\begin{itemize}
\item Type-I samples: the lightest $\rm {CP}$-even Higgs boson $h_1$ as the SM-like Higgs boson, $0.4\lesssim\lambda\lesssim0.7$, $0.13 \lesssim \kappa \lesssim 0.23$, $1.5 \lesssim \tan\beta \lesssim 6 $, $450~{\rm GeV} \lesssim \mu_{eff} \lesssim 720~{\rm GeV}$, and the Bayesian evidence $\ln Z_1 = -24.2$. Due to the limited capability of the Markov Chain algorithm, this type of samples were neglected in~\cite{Cao:2018rix}.
\item Type-II samples: the lightest $\rm {CP}$-even Higgs boson $h_1$ as the SM-like Higgs boson, $\lambda\lesssim0.08$, $-0.04 \lesssim \kappa <0$ with $2 |\kappa|/\lambda \simeq 1$, $4 \lesssim \tan\beta \lesssim 24$, $170~{\rm GeV} \lesssim \mu_{eff} \lesssim 420~{\rm GeV}$, and the Bayesian evidence $\ln Z_2 = -27.5$.
\item Type-III samples: the second lightest $\rm {CP}$-even Higgs boson $h_2$ as the SM-like Higgs boson, $\lambda \lesssim 0.15$, $|\kappa| \lesssim 0.06$ with $2 |\kappa|/\lambda \simeq 1$, $ 4.5 \lesssim \tan\beta \lesssim 32$, $135~{\rm GeV} \lesssim \mu_{eff} \lesssim 260~{\rm GeV}$, and the Bayesian evidence $\ln Z_3 = -27.0$.
\end{itemize}
It is evident that the three types of parameter spaces are extremely narrow. The Bayesian evidence $Z_i$ ($i=1,2,3$) produced the Jeffreys’ scales~\cite{Bayes,Jeffreys} $\delta_{12} \equiv \ln Z_1 - \ln Z_2 = 3.3$ and $\delta_{13} \equiv \ln Z_1 - \ln Z_3 = 2.8$. These results suggest that the considered experiments slightly preferred Type-I samples to Type-II and -III samples. However, as discussed below, Type-I samples will be examined by the near-future PandaX-4T experiment~\cite{Zhang:2018xdp}. Furthermore, Type-II and -III samples share the following features in the parameter space: both $\lambda$ and $|\kappa|$ are small, $2|\kappa|/\lambda \simeq 1$, and the higgsinos are moderately light. These facts lead to the similarities of the Type-II and -III samples in many aspects of DM physics (see the discussions below for detailed similarities and differences). Nevertheless, they are still distinguished from each other in Higgs physics~\cite{Ellwanger:2011aa, Gunion:2012zd,King:2012tr, King:2012is,Cao:2012fz,Vasquez:2012hn}.
In the following, we investigate the characteristics of the singlino-dominated DM based on numerical results. In Figure~\ref{fig1}, Type-I, -II, and -III samples are projected on the $\sin2\beta-m_{\tilde{\chi}_1^0}/\mu_{eff}$ and $\lambda v / \mu_{eff}-Z_h$ planes. In Figures~\ref{fig2}, \ref{fig3}, and \ref{fig4}, Type-I, -II, and -III samples are projected on the $\sigma_{\tilde{\chi}_1^0-p}^{SI}-m_{\tilde{\chi}_1^0}$ and $\sigma_{\tilde{\chi}_1^0-n}^{SD}-m_{\tilde{\chi}_1^0}$ planes, respectively, with different colors indicating the value of $\lambda v / \mu_{eff}$. In Table~\ref{table2}, a selection of benchmark points are shown to further clarify features in each of the three scenarios.
The following points about Type-I samples can be determined from Figures~\ref{fig1} and \ref{fig2} and the point $P_1$ in Table~\ref{table2}.
\begin{itemize}
\item Type-I samples are characterized by a relatively large $\lambda v/\mu_{eff}$ ranging from $0.14$ to $0.21$ (see right panel of Figure~\ref{fig1}). This will increase the higgsino composition in $\tilde{\chi}_1^0$ through Eq. (\ref{eq:Higgsino/singlino}) and the $\tilde{\chi}_1^0 \tilde{\chi}_1^0 Z$ coupling through Eq. (\ref{eq1:zchi10chi10_S}). As indicated in the right panel of Figure~\ref{fig2}, the SD scattering rates are thus larger than $7 \times 10^{-42}\ {\rm cm^2}$, which are exceedingly close to the near-future PandaX-4T exclusion limit.
\item As shown in the left panel of Figure~\ref{fig1}, $m_{\tilde{\chi}^0_1}/\mu_{eff}$ and $\sin2\beta$ are correlated by $m_{\tilde{\chi}^0_1}/\mu_{eff} \simeq \sin2\beta$.
In this case, $C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 H_{\rm SM} }$ in Eq. (\ref{eq1:hichi01chi01_S}) is suppressed by $m_{\tilde{\chi}^0_1}/\mu_{eff}$ and $\sin2\beta$ cancellation, favored by the stringent bound of the XENON-1T experiment on the SI cross section. We further explore its implication by focusing on the point $P_1$ in Table~\ref{table2}, which predicts the following four terms in Eq. (\ref{eq1:hichi01chi01_S}):
\begin{eqnarray}
C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 h} &\simeq& 0.2702 \times (0.6178 - 0.7647 ) + 0.0-0.0029+0.0465 \sim 0.0027,
\label{res_C_xsxshsm1}
\\
C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 h_{s}} &\simeq& -0.006-0.0+0.0183-0.312 \sim -0.2997.
\label{res_C_xsxshs1}
\end{eqnarray}
The two contributions in Eq. (\ref{eq0:au-ad_hsm-hs}) are as follows:
\begin{eqnarray}
\mathcal{A} & \simeq ~0.0027 -0.0036 \sim -0.0009~.
\end{eqnarray}
These results show that, besides the mentioned cancellation, there is a strong offsetting effect between the first and fourth terms within the $\tilde{\chi}^0_1 \tilde{\chi}^0_1 h$ coupling itself (i.e., cancellation between $C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 H_{\rm SM}}$ and $C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 H_{\rm S}}$ terms in $C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 h}$) and a strong cancellation between the two contributions to the SI cross section from $h$ and $h_{s}$. These accidental cancellations result in $\sigma^{SI}_{\tilde{\chi}_1^0-p} \sim 10^{-49}\ {\rm cm^2}$. In contrast, $\sigma^{SI}_{\tilde{\chi}_1^0-p}$ would be around $10^{-42}\ {\rm cm^2}$ without them. This feature explains the SI cross section for Type-I samples possibly being as low as $10^{-50}\ {\rm cm^2}$, as shown in the left panel of Figure~\ref{fig2}.
\item Based on $m_{\tilde{\chi}_1^0}/\mu$, $\lambda v/\mu$, $\sin 2 \beta$, and $Z_h$ in Figure~\ref{fig1}, it is usually predicted that $C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 G^0} \sim 0.1$ by
Eq. (\ref{eq1:G0chi10chi10_S}) and $\langle \sigma v \rangle_{x_F}^{t\bar{t}} \sim 10^{-26}\ {\rm cm}^3{\rm s^{-1}}$ by Eq. (\ref{eq:gxxGforOmega}). This implies that $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \to t \bar{t}$ played a significant role in determining the abundance. Concerning the point P1 in Table~\ref{table2}, we found $C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 G^0} \simeq -0.108$ and $\langle \sigma v \rangle_{x_F}^{t\bar{t}} \simeq 1.2 \times 10^{-26}\ {\rm cm}^3{\rm s^{-1}}$, indicating that the annihilation contributed to the total annihilation rate by about $52\%$. We also obtained $\langle \sigma v \rangle_{x_F}^{h_s a_s} \simeq 3.4 \times 10^{-27} \ {\rm cm}^3{\rm s^{-1}}$ by Eq. (\ref{hsaa-approximation}), which means that $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \to h_s a_s$ contributed to the total rate by $15\%$. We add that
our estimation roughly agrees with the results in Table~\ref{table2}, calculated by the \texttt{micrOMEGAs} package.
\end{itemize}
\begin{figure*}[t]
\centering
\resizebox{1.\textwidth}{!}{
\includegraphics[width=0.90\textwidth]{Fig3-a.png}
\includegraphics[width=0.90\textwidth]{Fig3-b.png}
}
\vspace{-0.6cm}
\caption{Same plots as in Fig.\ref{fig2}, but for the results of the Type-II samples. \label{fig3} }
\end{figure*}
Next we consider Type-II samples. Figures~\ref{fig1} and \ref{fig3} demonstrate following features:
\begin{itemize}
\item $\lambda v/\mu_{eff} \lesssim 0.06$ and $-0.95 \lesssim m_{\tilde{\chi}^0_1}/\mu_{eff} \lesssim -0.90$, which imply that
\begin{eqnarray}
&& |C_{\tilde{\chi}_1^0 \tilde{\chi}_1^0 Z}| \lesssim 2.4 \times (\lambda v/\mu_{eff})^2 < 0.009, \nonumber \\
&& \sigma^{SD}_{\tilde{\chi}_1^0-n}/{\rm cm^2} \lesssim 1.3 \times 10^{-36} \times (\lambda v/\mu_{eff})^4 < 1.7 \times 10^{-41}
\end{eqnarray}
by Eq. (\ref{eq1:zchi10chi10_S}) and (\ref{eq:sigSD}), respectively. This feature is shown on the right panel of Figure~\ref{fig3}.
\item As indicated in the left panel of Figure~\ref{fig1}, $m_{\tilde{\chi}^0_1}/\mu_{eff}$ and $\sin 2 \beta$ are of opposite sign, and thus,
their contributions to $C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 H_{\rm SM} }$ in Eq. (\ref{eq1:hichi01chi01_S}) do not cancel each other. This implies that both
$C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 h }$ and $\mathcal{A}$ in Eq. (\ref{eq0:au-ad_hsm-hs})
are mainly contributed by
$C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 H_{\rm SM} }$.
Consequently, $\sigma^{SI}_{\tilde{\chi}_1^0-p}$ in Eq.(\ref{eq:SIDD_p2}) is approximated by
\begin{eqnarray}
\sigma_{\tilde {\chi}^0_1-{p}}^{\rm SI} &\simeq& 5 \times 10^{-45} {\rm cm^2}\times \left(\frac{C_{\tilde{\chi}^0_1 \tilde{\chi}^0_1 H_{\rm SM}}}{0.1}\right)^2,
\end{eqnarray}
where $C_{ \tilde {\chi}^0_1 \tilde {\chi}^0_1 H_{\rm SM}} \simeq (4--10) \times \sqrt{2} \lambda (\lambda v/\mu_{eff})$\footnote{It is noticeable that the enhancement coefficient $4--10$ comes from the factor $1/\{1-(m_{\tilde{\chi}_1}/\mu_{eff})^2\}$ in Eq. (\ref{eq1:hichi01chi01_S}). This is a common characteristic for the Type-II and -III samples.}. This approximation reflects the relation $\sigma_{\tilde {\chi}^0_1-{p}}^{\rm SI} \propto \lambda^4$. For a small $\lambda$, the cross section may be as low as $10^{-48}\ {\rm cm^2}$, as shown on the left panel of Figure~\ref{fig3}.
We add that the SI cross sections in Figure~\ref{fig3} are larger than $3 \times 10^{-49}\ {\rm cm^2}$ because smaller cross sections require smaller values of $\lambda$, which are not readily available since Bayesian evidence is suppressed significantly. We also add that the characteristics of $\sigma_{\tilde {\chi}^0_1-{p}}^{\rm SI}$ for the Type-I and -II samples are different because $\sigma_{\tilde {\chi}^0_1-{p}}^{\rm SI}$ has no significant cancellation effect in the latter case.
Furthermore, we verified the approximation for $\sigma_{\tilde {\chi}^0_1-{p}}^{\rm SI}$ by considering the point $P_2$ in Table~\ref{table2}. The four terms in Eq. (\ref{eq1:hichi01chi01_S}) and the two contributions in Eq. (\ref{eq0:au-ad_hsm-hs}) were as follows:
\begin{eqnarray}
C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 h} & \simeq & -0.01 + 0.0 + 0.0 + 0.002 \sim -0.008, \nonumber
\\
\mathcal{A}
&\simeq& ~-0.008 - 0.0008\sim -0.0088.
\end{eqnarray}
\item Since $2 |\kappa|/\lambda \simeq 1$, the singlino-dominated $\tilde{\chi}_1^0$ co-annihilated with the higgsino-dominated neutralinos and charginos to provide the measured abundance. In addition, because $\lambda$ and $\kappa$ must be small to suppress $C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 G^0}$ and $C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 a_s}$, the channels $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \to t \bar{t}, h_s a_s$ never have a crucial effect on the abundance, even if they are kinematically accessible.
\end{itemize}
\begin{figure*}[t]
\centering
\resizebox{1.\textwidth}{!}{
\includegraphics[width=0.90\textwidth]{Fig4-a.png}
\includegraphics[width=0.90\textwidth]{Fig4-b.png}
}
\vspace{-0.6cm}
\caption{Same plot as in Fig. \ref{fig2}, but for the results of the Type-III samples. \label{fig4} }
\end{figure*}
Finally, we investigate the Type-III samples. Figure~\ref{fig1} indicates that some samples correspond to the same parameter space as the Type-II samples, and thus, they predict
similar DM physics. This conclusion was verified by studying some points in the two scenarios, e.g.,
points P2 and P4 in Table~\ref{table2}. In the following, we only concentrate on the samples with same sign $m_{\tilde{\chi}_1^0}/\mu$ and $\sin 2 \beta$. The following results were obtained from Figures~\ref{fig1} and \ref{fig4}.
\begin{itemize}
\item Compared with the Type-II samples, $\lambda v/\mu_{eff}$ may have a greater value of up to 0.12. The reason is there are accidental cancellations contributing to $\sigma_{\tilde {\chi}^0_1-{p}}^{\rm SI}$, similar to the Type-I samples. This will relax the XENON-1T constraint. We show this characteristic by considering the point $P_3$ in Table~\ref{table2} and finding
\begin{eqnarray}
C_{\tilde {\chi}^0_1 \tilde {\chi}^0_1 h}
&\simeq& 0.047 \times (0.88 - 0.16) + 0.0-0.0 +0.0085 \sim 0.0437, \nonumber \\
\mathcal{A} & \simeq & 0.042 - 0.0285\sim 0.0135~. \nonumber
\end{eqnarray}
The cancellation also explains why $\sigma_{\tilde {\chi}^0_1-{p}}^{\rm SI}$ was as low as $10^{-50} {\rm cm^2}$, as shown in the left panel of Figure~\ref{fig4}.
\item Due to a relatively large $\lambda v/\mu_{eff}$ and $| m_{\tilde{\chi}^0_1}/\mu_{eff} |\rightarrow 1$ in Eq. (\ref{eq1:zchi10chi10_S}), $\sigma_{\tilde {\chi}^0_1-{n}}^{\rm SD}$ may be on the border of $10^{-41}\ {\rm cm^2}$ (see right panel of Figure~\ref{fig4}). Furthermore, similar to the Type-II samples, $\tilde{\chi}_1^0$ co-annihilated with the higgsino-dominated neutralinos and charginos to achieve the measured abundance.
We add that the point $P_3$ predicts a significantly larger $\lambda$ than points $P_2$ and $P_4$. Consequently, this will be readily tested in the near-future PandaX-4T experiments for both SI and SD scattering.
\end{itemize}
\section{Conclusion}\label{sec:conclusion}
This work provided updates to previous studies of singlino-dominated DM mainly in four aspects:
\begin{itemize}
\item[(1)] We no longer require the fine-tuning measurement ${\Delta}m_Z$ as a model selection criterion, since it may only reflect personal prejudice.
\item[(2)] We adopted an advanced MultiNest algorithm to perform a sophisticated scan over the $Z_3$-NMSSM parameter space to ensure that the obtained conclusions were as complete as possible.
\item[(3)] We utilized the latest experimental results to restrict singlino-dominated DM scenarios, including the XENON-1T search for both SI and SD DM-nucleon scattering since 2018, ATLAS analyses of sparticle signals with 139-${\rm fb^{-1}}$ data, and measurements of the Higgs couplings with 80-${\rm fb^{-1}}$ data.
\item[(4)] We provided simplified analytical formulas for both the DM annihilation cross sections and the SI and SD cross sections of DM-nucleon scattering, and we numerically scrutinized each contribution in these formulas.
\end{itemize}
As a result, a new singlino-dominated DM scenario (Type-I samples) was found. More model information, such as its Bayesian evidence, was obtained, and the current and future statuses of the scenarios were presented. More importantly, this study provided clear insight into singlino-dominated DM scenarios and explained why they had been tightly limited in the $Z_3$-NMSSM.
Specifically, this study indicated that the surviving samples can be categorized into three types:
\begin{itemize}
\item For Type I samples, $0.4\lesssim\lambda\lesssim0.7$, $0.13 \lesssim \kappa \lesssim 0.23$, $1.5 \lesssim \tan\beta \lesssim 6 $, $450~{\rm GeV} \lesssim \mu_{eff} \lesssim 720~{\rm GeV}$, and the annihilation $\tilde{\chi}_1^0 \tilde{\chi}_1^0 \to t \bar{t}$ is primarily responsible for the DM abundance.
\item For Type II and III samples, $0 < \lambda \lesssim 0.15$, $\lambda \simeq 2 |\kappa|$, and the dominant annihilation involves a co-annihilation with higgsinos.
\end{itemize}
The Bayesian evidence ($Z$) for the three sample types showed that the experiments slightly preferred Type-I samples to Type-II and -III. However, Type-I samples will be examined in the
near-future PandaX-4T experiment. They will become highly disfavored if the experiment shows no signs of DM. It should be emphasized that DM-annihilation by a singlet scalar or pseudo-scalar funnel, a Z-boson funnel, and the SM-like Higgs funnel presented by Abdallah et al.~\cite{Abdallah:2019znp} were not observed in this study, due to the small Bayesian evidence. In addition, both analytical formulas and numerical results were used to summarize the theory's four primary cases in significantly suppressing the SI scattering cross section for DM-nucleons. This included 1) a small $\lambda v/\mu_{eff}$, 2) cancellation between $m_{|tilde{\chi}_1^0}/\mu_{eff}$ and $\sin 2 \beta$, 3) cancellation between two contributions from the SM doublet component ($H_{\rm SM}$) and singlet component ($H_s$) within the $\tilde{\chi}^0_1 \tilde{\chi}^0_1 h$ coupling strength itself, and 4) cancellation between two contributions from $h$ and $ h_{s}$.
In summary, the interaction of the singlino-dominated DM with nucleons in the $Z_3$-NMSSM has been tightly restricted in current DM-DD experiments, while the measured abundance favors its involvement in weak interactions. Given the theory's natural preference for electroweak symmetry breaking, it has become increasingly difficult to represent these two seemingly paradoxical features using neutralino DM, due to the limited theoretical structure. Thus, the benefits of singlino-dominated DM are waning unless one extends the $Z_3$-invariant theory. Recent studies on the MSSM and NMSSM imply that DM candidates should be gauge singlet fields or that singlet components should at least be naturally dominant over other components~\cite{Cao:2019qng}. This requirement motivates us to extend the $Z_3$-NMSSM to a general NMSSM to increase the Bayesian evidence of the scenarios significantly~\cite{cao:2021Gnmssm}. It also motivates us to augment the $Z_3$-NMSSM with a seesaw mechanism to generate neutrino masses and select the lightest sneutrino as a DM candidate~\cite{Cao:2017cjf,Cao:2018iyk,Cao:2019qng,Cao:2019aam}\footnote{The properties of the sneutrino DM in the $Z_3$-NMSSM were first studied one decade ago~\cite{Cerdeno:2008ep,Cerdeno:2009dv,Cerdeno:2011qv,Cerdeno:2013oya}. However, these studies considered the cases where the cross section of the sneutrino-nucleon scattering was much larger than current DD experimental bounds, so they obtained different conclusions from our work.}.
\section*{Acknowledgments}
This work was supported by the National Natural Science Foundation of China (NNSFC) under Grant No. 11575053 and No. 12075076.
\section*{Appendix \label{sec:appendix}}
Recently, the ATLAS collaboration aimed to limit the compressed mass spectra case. They analyzed 139 ${\rm fb^{-1}}$ of $\sqrt{s} = 13\ {\rm TeV}$ proton--proton collision data collected at the LHC, focusing on
the events with missing transverse momentum and two same-flavor, oppositely charged, low-transverse-momentum leptons, further categorizing them by the presence of hadronic activity from initial-state radiation~\cite{Aad:2019qnd}. We repeated this analysis using the simulation tools MadGraph5\[email protected]~\cite{mad-1,mad-2} to generate the parton level events, Pythia-8.2~\cite{pythia} for parton fragmentation and hadronization, Delphes-3.4.2~\cite{delphes} for fast simulation of the performance of the ATLAS detector, and CheckMATE-2.0.26~\cite{cmate-1,cmate-2,cmate-3} to implement the analysis cut selections.
Below, we validate our code for all signal regions (SRs)~\cite{Aad:2019qnd}. We considered $\tilde{l}^{+} \tilde{l}^{-}$ production in the MSSM and set the masses of all charginos and neutralinos other than the bino-like $\tilde{\chi}_1^0$ to be 2.5 TeV. Thus, the sleptons will decay by $\tilde{l}^{\pm}\rightarrow \l^{\pm} \tilde{\chi}_1^0 $. We consider the benchmark point $m_{\tilde{l}} = 150~{\rm GeV}$ and $m_{\tilde{\chi}_1^0}=140~{\rm GeV}$. As a result, the cross section at the next-leading order is $126.62~{\rm fb}$ for $\tilde{l}_{L} \tilde{l}_{L}$ production and $47.62~{\rm fb}$ for $\tilde{l}_{R} \tilde{l}_{R}$ production.
The involved cards were set as follows:
\begin{lstlisting}[backgroundcolor=\color{back},frame=trBL]
import model MSSM_SLHA2 --modelname
generate p p > sl1+ sl1-, (sl1- > e- n1), (sl1+ > e+ n1)
add process p p > sl1+ sl1- j, (sl1- > e- n1), (sl1+ > e+ n1)
generate p p > sl2+ sl2-, (sl2- > mu- n1), (sl2+ > mu+ n1)
add process p p > sl2+ sl2- j, (sl1- > mu- n1), (sl2+ > mu+ n1).
\end{lstlisting}
For the proc\_card.dat:
\begin{lstlisting}[backgroundcolor=\color{back},frame=trBL]
100000 = nevents ! Number of unweighted events requested.
0 = ickkw ! 0 no matching, 1 MLM
37.5 = ktdurham,
\end{lstlisting}
For the run\_card.dat:
\begin{lstlisting}[backgroundcolor=\color{back},frame=trBL]
Block mass
1000011 1.500000e+02 # Msl1
1000013 1.500000e+02 # Msl2
1000022 1.400000e+02 # Mneu1
1000023 2.500000e+03 # Mneu2
1000024 2.500000e+03 # Mch1
1000025 -2.50000e+03 # Mneu3
1000035 2.500000e+03 # Mneu4
1000037 2.500000e+03 # Mch2
Block selmix
1 1 1.000000e+00 # RRl1x1
2 2 1.000000e+00 # RRl2x2
4 4 1.000000e+00 # RRl4x4
5 5 1.000000e+00 # RRl5x5
\end{lstlisting}
For the param\_card.dat and the pythia8\_card.dat:
\begin{lstlisting}[backgroundcolor=\color{back},frame=trBL]
Merging:Process = pp>{sl1-,1000011}{sl1+,-1000015}{sl2-,1000013}
{sl2+,-1000013}
Merging:mayRemoveDecayProducts=on
\end{lstlisting}
In our simulation, we generated 100,000 events for the production process. The results, shown in Table~\ref{R,L}, indicate that we can reproduce the ATLAS analysis at the $20\%$ level for most cases.
\begin{table*}[t]
\centering
\caption{Cut flow for the analysis in~\cite{Aad:2019qnd}. We considered the point $m(\tilde{\ell}, \tilde{\chi}_1^0) = (150, 140)~{\rm GeV}$ in the calculations.}
\vspace{0.2cm}
\label{R,L}
\scalebox{0.75}{
\begin{tabular}{l|rrrr}
\hline
Process & \multicolumn{4}{c}{Production of $\tilde{\ell}\tilde{\ell}$} \\ \hline
Point & \multicolumn{4}{c}{$m_{\tilde{\ell}}$ = 150 GeV; $m_{\tilde{\chi}_1^{0}}$ = 140 GeV} \\ \hline
Generated Events & \multicolumn{4}{c}{100000} \\ \hline
\multirow{2}{*}{Selection} & \multicolumn{2}{c}{ATLAS} & \multicolumn{2}{c}{CheckMATE} \\
& Events & Efficiency & Events & Efficiency \\ \hline
Total events & 24069 & $\cdots$ & 24069 & $\cdots$ \\
$E_{T}^{\rm miss}$ trigger & 2355.37 & $\cdots$ & 2355.37 & $\cdots$ \\
Two leptons & 1014.55 & 43.07$\%$ & 1079.07 & 45.81$\%$ \\
veto 3 GeV $< m_{\ell\ell} < 3.2~{\rm GeV}$ & 1013.21 & 99.87$\%$ & 1077.69 & 99.87$\%$ \\
lepton author 16 veto & 1009.48 & 99.63$\%$ & 1077.69 & 100.00$\%$ \\
min$(\Delta\phi(\rm any jet, p_{\rm T}^{\rm miss})) >$ 0.4 & 970.36 & 96.12$\%$ & 1049.11 & 97.35$\%$ \\
$\Delta\phi(\rm j_1, p_{\rm T}^{\rm miss}) >$ 2.0 & 961.15 & 99.05$\%$ & 1027.05 & 97.90$\%$ \\
lepton truth matching & 958.99 & 99.78$\%$ & 1027.05 & 100.00$\%$ \\
1 $< m_{\ell\ell} <$ 60 GeV & 827.86 & 86.33$\%$ & 883.55 & 86.03$\%$ \\
$\Delta R_{\rm ee} >$ 0.3, $\Delta R_{\rm \mu\mu} >$ 0.05, $\Delta R_{\rm e\rm \mu} >$ 0.2 & 826.19 & 99.80$\%$ & 883.48 & 99.99$\%$ \\
$p_{\rm T}^{j_1} >$ 5 GeV & 823.70 & 99.70$\%$ & 880.95 & 99.71$\%$ \\
$n_{\rm jet} \geq$ 1 & 810.59 & 98.41$\%$ & 880.95 & 100.00$\%$ \\
$p_{\rm T}^{\rm j_1} >$ 100 GeV & 705.86 & 87.08$\%$ & 702.58 & 79.75$\%$ \\
$n_{b-{\rm jet}} = 0$ & 611.05 & 86.57$\%$ & 643.78 & 91.63$\%$ \\
$m_{\rm \tau\tau} < 0$ or $>$ 160 GeV & 533.29 & 87.27$\%$ & 569.78 & 88.51$\%$ \\
ee or $\rm \mu\mu$ & 532.33 & 99.82$\%$ & 569.01 & 99.86$\%$ \\ \hline
\bf SR-highMass & & & & \\
$E_{\rm T}^{\rm miss} > 200~{\rm GeV}$ & 229.81 & 43.17$\%$ & 265.83 & 46.72$\%$ \\
max(0.85, 0.98 - 0.02 $\times m_{\rm T_2}^{\rm 100}$) $< R_{\rm ISR} <$ 1.0 & 160.30 & 69.75$\%$ & 165.78 & 62.36$\%$ \\
$p_{\rm T}^{\ell_2} >$ min(20.0, 2.5 + 2.5$\times (m_{\rm T_2}^{\rm 100} - \rm 100))$ & 70.71 & 44.11$\%$ & 72.51 & 43.74$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 140 GeV & 70.71 & 100.00$\%$ & 72.51 & 100.00$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 130 GeV & 70.71 & 100.00$\%$ & 72.51 & 100.00$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 120 GeV & 70.71 & 100.00$\%$ & 72.31 & 99.73$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 110 GeV & 70.71 & 100.00$\%$ & 72.23 & 99.90$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 105 GeV & 53.72 & 75.97$\%$ & 57.10 & 79.05$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 102 GeV & 20.21 & 37.62$\%$ & 23.77 & 41.63$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 101 GeV & 9.38 & 46.41$\%$ & 9.90 & 41.62$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 100.5 GeV & 4.68 & 49.89$\%$ & 4.86 & 49.11$\%$ \\ \hline
\bf SR-lowMass & & & & \\
150 $< E_{\rm T}^{\rm miss} <$ 200 GeV & 146.36 & 27.49$\%$ & 167.63 & 29.46$\%$ \\
0.8 $< R_{\rm ISR} <$ 1.0 & 107.82 & 73.67$\%$ & 93.17 & 55.58$\%$ \\
$p_{\rm T}^{\ell_2} >$ min(15.0, 7.5 + 0.75$\times (m_{\rm T2}^{\rm 100} - \rm 100))$ & 52.74 & 48.91$\%$ & 42.29 & 45.39$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 140 GeV & 52.74 & 100.00$\%$ & 42.29 & 100.00$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 130 GeV & 52.74 & 100.00$\%$ & 42.29 & 100.00$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 120 GeV & 52.74 & 100.00$\%$ & 42.29 & 100.00$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 110 GeV & 52.64 & 99.81$\%$ & 41.65 & 98.49$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 105 GeV & 38.05 & 72.28$\%$ & 29.09 & 69.85$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 102 GeV & 16.66 & 43.78$\%$ & 11.24 & 38.62$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 101 GeV & 8.70 & 52.22$\%$ & 5.60 & 49.82$\%$ \\
$m_{\rm T2}^{\rm 100} <$ 100.5 GeV & 4.39 & 50.46$\%$ & 2.29 & 40.88$\%$ \\ \hline
\end{tabular}}
\end{table*}
|
\section{Introduction}
Sampling and optimization are fundamental tasks in mathematics, statistical physics, and various subfields of computer science such as cryptography, differential privacy, machine learning, and artificial intelligence. In continuous settings, sampling and optimization are known to be intimately connected; convex sets, and more generally log-concave distributions, are the natural domains where either task is algorithmically tractable. For a more formal treatment of this connection in continuous settings see \cite{LV06, LSV18}.
On discrete/combinatorial domains, the relationship between sampling and optimization is less clear. For example, the intersection of two matroids is easy to optimize over, but not known to be easy to sample from, and the opposite holds for determinantal point processes, which are easy to sample from \cite[see, e.g.,][]{AOR16} and hard to optimize \cite{CM10}.
The goal of this work is to establish a new connection between sampling and optimization in discrete settings. For a family of distributions $\mu$ defined on size $k$ subsets of a ground set of elements\footnote{The restriction of the domain to size $k$ subsets of a ground set should be thought of as a ``canonical form''; many other discrete domains can be naturally transformed into this form.} that is closed under external fields, we show that rapid mixing of natural local random walks implies the existence of simple approximation algorithms to find $\max \mu(\cdot)$. More specifically, we show that local search can approximately find $\max \mu(\cdot)$ within a nearly-optimal approximation factor.
We study a family of natural local search algorithms (\cref{alg:localsearch}) to find $\max \mu(\cdot)$. These algorithms start with a set $S$, and repeatedly try to increase $\mu(S)$ by swapping a constant number of elements in $S$ with elements outside of $S$ until no more improvements can be made.
More formally, suppose that the domain of the objective $\mu$ is the collection of size $k$ subsets of the ground set $[n]=\set{1,\dots,n}$, which we denote by $\binom{[n]}{k}$. Then, local search is defined with a parameter $r \geq 0$ which specifies the ``local neighborhood'' the algorithm searches over in each iteration. The $r$-neighborhood of $S\in \binom{[n]}{k}$ are all the sets that can be reached by swapping at most $r$ elements:
\[ \cN_r(S) := \set*{T \in \binom{[n]}{k} \given \card{S-T}\leq r }. \]
Each iteration of local search goes from a set $S$ to $\hat{S}\in \cN_r(S)$ which maximizes $\mu(\hat{S})$. If we reach a local optimum, i.e., $S=\hat{S}$, then $\mu(S)=\max\set{\mu(T)\given T\in \cN_r(S)}$.
We show in this work that rapid mixing of natural local random walks, the \emph{(multi-step) down-up random walks}, designed to sample from $\mu$ and related distributions, implies that \emph{local maxima} of $\mu$ are \emph{approximate global maxima}.
\begin{definition}[Down-Up Random Walks]\label{def:local-walk}
For a density $\mu:\binom{[n]}{k}\to\R_{\geq 0}$, and an integer $\l\leq k$, we define the $k\leftrightarrow\l$ down-up random walk as the sequence of random sets $S_0, S_1,\dots$ generated by the following algorithm:
\begin{Algorithm*}
\For{$t=0,1,\dots$}{
Select $T_t$ uniformly at random from subsets of size $\l$ of $S_t$.\;
Select $S_{t+1}$ with probability $\propto \mu(S_{t+1})$ from supersets of size $k$ of $T_t$.\;
}
\end{Algorithm*}
\end{definition}
This random walk is time-reversible, always has $\mu$ as its stationary distribution, and moreover has positive real eigenvalues \cite[see, e.g.,][]{ALO20}. This random walk, specially for the case of $\l=k-1$, has received a lot of attention in the literature on high-dimensional expanders \cite[see, e.g.,][]{LLP17,KO18,DK17,KM16,AL20,ALO20}. Each step of this random walk can be efficiently implemented as long as $k-\l=O(1)$ and we have oracle access to $\mu$. We remark that down-up walks generalize other well-known local random walks like the Glauber dynamics \cite[see, e.g.,][]{ALO20}. Note that the down-up random walk is \emph{local} in the sense that $S_{t+1}\in \cN_{k-\l}(S_t)$. Naturally, we tie mixing of these random walks to local search with $r=k-\l$ neighborhoods.
There has been a recent surge of interest in analyzing the mixing properties of down-up random walks due to a number of breakthrough applications to open problems in sampling and counting \cite{ALOV19,AL20,ALO20,alimohammadi2021fractionally,CLV21a,CLV21b,FGYZ21,Liu21,JPV21,BCCPSV21,CFYZ21,ALOVV21,AJKPV21}.
Key to many of these works was the notion of spectral independence. \textcite{alimohammadi2021fractionally} introduced a stronger notion called fractional log-concavity, and showed that it implies a $k^{-O(1)}$ lower bound on the spectral gap of $k \leftrightarrow (k-O(1))$-down-up random walks on $\mu.$ We remark that fractional log-concavity, unlike spectral independence, is preserved under external fields, formally defined as follows.
For a distribution $\mu$ on $\binom{[n]}{k}$ and $\lambda = (\lambda_1,\dots, \lambda_n) \in \R^{n}_{>0}$, the \emph{$\lambda$-external field} applied to $\mu$ is another distribution on $\binom{[n]}{k}$, denoted by $\lambda \ast \mu$, defined up to normalization as follows:
\[\mathbb{P}_{\lambda \ast \mu}[S] \propto \mu(S)\cdot \prod_{i \in S}\lambda_i.\]
\begin{comment}
\begin{theorem}
Consider $\mu: \binom{[n]}{k} \to \R_{\geq 0}.$
Suppose that for all external field $\lambda \in \R_{> 0}^n,$ the $\ell$-down-up random walk on $\lambda \ast \mu$ has conductance at least $\Omega(k^{-c}).$ Let $S$ be a local optima with respect to $\mu$ in $\ell$-neighborhood, then $S$ is also a $k^{O(k(\ell +c))}$-approximate global optimum i.e. $k^{O(k(\ell +c))} \mu(S) \geq \max_{S'} \mu(S').$
\end{theorem}
\end{comment}
As established in \cite{alimohammadi2021fractionally}, various distributions of interest involving determinants are fractionally log-concave. For a fractionally log-concave distribution $\mu$, the $k\leftrightarrow (k-O(1))$-down-up walk on $\lambda\ast \mu$ has inverse-polynomially large spectral gap, even when an arbitrary external field $\lambda\in \R_{\geq 0}^n$ is applied to $\mu$. We show that this property\footnote{Curiously, in continuous settings applying an external field also preserves log-concavity, the standard of algorithmic tractability for sampling; applying an external field is the same as multiplication by a log-linear function.} implies nearly optimal approximation for (multi-step) local search on $\mu$.
\begin{theorem
\label{thm:main}
Consider a distribution $\mu: \binom{[n]}{k} \to \R_{\geq 0}.$ Suppose that for some $r = O(1)$, the $k \leftrightarrow (k-r)$ down-up random walk on $\lambda \ast \mu$ has spectral gap at least $k^{-O(1)}$ for all external fields $\lambda\in \R_{\geq 0}^n$. Then any approximate local maximum, that is a set $S\in \binom{[n]}{k}$ such that
\[ \mu(S)\geq \Omega(1)\cdot \max\set*{\mu(T)\given T\in \cN_r(S)} \]
is a $k^{O(k)}$-approximate global maximum, that is
\[ \mu(S)\geq k^{-O(k)}\cdot \max\set*{\mu(T)\given T\in \binom{[n]}{k}}. \]
Moreover, such an approximate local maximum can be found efficiently given oracle access to $\mu$ and a starting point in the support of $\mu$.
\end{theorem}
In particular, combined with rapid mixing results of \cite{alimohammadi2021fractionally}, \cref{thm:main} implies that local search is an efficient $k^{O(k)}$-approximation algorithm for the optimization problem on nonsymmetric determinantal point processes (see \cref{subsec:nonsymDPP}), and on the intersection of a strongly Rayleigh distributions over $\binom{[n]}{k}$ and constantly many partition constraints (\cref{cor:strongly Rayleigh partition}). Our approximation algorithm for nonsymmetric determinantal point processes is the first unconditional multiplicative approximation algorithm for this problem.
\begin{remark}
We remark that the approximation factor of $k^{O(k)}$ is nearly optimal amongst efficient algorithms. The special case of symmetric determinantal point processes was shown to be hard to approximate within a factor of $c^k$ for some constant $c>1$ \cite{CM10}. Further, the factor of $k^{O(k)}$ is tight for local search, even in the special case of symmetric determinantal point processes \cite[see, e.g.,][]{AV20}.
\end{remark}
\subsection{MAP Inference on Nonsymmetric DPPs} \label{subsec:nonsymDPP}
Determinantal point processes (DPPs) have found many applications in machine learning, such as data summarization \cite{gong2014largemargin,LB12}, recommender systems \cite{GartPK16,Wilhelm18}, neural network compression \cite{MS15}, kernel approximation \cite{LiJS16},
multi-modal output generation \cite{Elfeki19}, etc.
Formally, a DPP on a set of $[n]$ items is a probability distribution over subsets $Y\subseteq [n]$ parameterized by a matrix $L \in \R^{n\times n}$ where $Y$ is chosen with probability proportional to the determinant of the principal submatrix $L_Y$ whose columns are rows are indexed by $Y$: \[ \P{Y}\propto \det(L_Y).\]
A related and perhaps more widely used model, is a $k$-DPP, where the size of $Y$ is restricted to be exactly $k$. In applications, usually $k$ is set to be much smaller than $n$. We study this model in this paper.
A fundamental optimization problem associated to probabilistic models, including DPPs, is to find the most likely, or the maximum a posteriori (MAP) configuration \cite{GKT12}:
\begin{equation} \label{eq:opt}
\max\set*{\P{S}\given S\in \binom{[n]}{k}}.
\end{equation}
MAP inference is particularly useful when the end application requires outputting a single set; e.g., in recommender systems, the task is to produce a fixed-size subset of items to recommend to the user.
Most prior work on DPPs requires the kernel matrix $L$ to be symmetric, but such symmetric kernels are known to be able to only encode repulsive (negatively correlated) interactions between items \cite{BBL09}. This severely limits their modeling power in practical settings. For example, a good recommender system for online shopping should model iPads and Apple Pencils as having positive interactions, since these are \emph{complementary} items and tend to be bought together.
To remedy this, recent work has considered the more
general class of nonsymmetric DPPs (NDPPs) and shown that these have additional useful modeling
power \cite{BWLGCG18,Gartrell2019LearningND}. \citet{Gartrell2019LearningND} consider NDPPs parameterized by nonsymmetric positive semi-definite (nPSD) kernel matrices $L$, i.e., those matrices where $L + L^\intercal \succcurlyeq 0$, and show efficient algorithms for learning such NDPPs.
\begin{definition}\label{def:nPSD}
A (not-necessarily-symmetric) matrix $L\in \R^{n\times n}$ is \textit{nonsymmetric positive semidefinite} (nPSD) if $L + L^\intercal \succeq 0$.
\end{definition}
Throughout, we will consider only NDPPs with nPSD kernels (nPSD-NDPPs) \citep[see][for a survey on fixed-size DPPs and their applications]{KT12}. \citet{alimohammadi2021fractionally} showed how to efficiently sample from fixed-size nPSD-NDPPs using natural Markov chains. \citet{gartrell2020scalable} proposed a new learning algorithm, and showed how to efficiently implement and analyze the natural greedy MAP inference heuristic for symmetric DPPs on nPSD-NDPPs. This greedy heuristic (\cref{alg:common-greedy}) starts from an empty set and runs for $k$ iterations, in each iteration adding the item that most increases the DPP score.
Though this greedy algorithm is guaranteed to obtain a $k^{O(k)}$-approximation for symmetric DPPs \cite{CM10}, it could not achieve even a finite approximation factor for nPSD-NDPPs. For example, on a skew-symmetric matrix $X$, i.e., $X = -X^{\intercal}$, since all odd-sized principle minors of $X$ are zero, \cref{alg:common-greedy} would necessarily resort to picking an arbitrary/random item at every other iteration, which can result in an arbitrarily bad final answer. Consider a concrete example, which helps build intuition on why greedy fails to achieve a meaningful approximation factor. This example also shows that local search greedy \cite{KD16}\footnote{This algorithm starts with the output $S$ of \cref{alg:common-greedy}, then continuously swaps out an element in $S$ with one outside $S$ to increase the DPP score, until either a local maximum is reached or $k^2 \log k$ swaps have been performed.}, another candidate MAP inference algorithm with theoretical performance guarantees for symmetric DPPs, also used by \citet{gartrell2020scalable} as a baseline to compare their greedy method, also fails to achieve a meaningful approximation factor.
\begin{example}\label{remark:greedy same as localSearch}
Consider $L$ composed of $2\times 2$ blocks $D_i = \begin{bmatrix} c_i& x_i\\ -x_i & c_i \end{bmatrix}$ where $c_i> 1$:
\[
L:=\begin{bmatrix}
D_1 & 0 & \dots & 0\\
0 & D_2 & \dots & 0\\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \dots & D_n
\end{bmatrix},
\]
We further assume that $c_1 > c_2 > \dots > c_n,$ $x_1 < x_2 < \dots < x_n$ and $x_i \gg c_j \; \forall i, j.$ It is easy to check that \cref{alg:common-greedy} (greedy) on input $k= 2t$ will select $ S = \set*{1,2 ,\dots, 2t-1, 2t}. $ Indeed, \cref{alg:common-greedy} first picks item $1$ since $\det(L_{\set*{1}}) =c_1$ is maximum among all $L_{\set*{i}},$ then picks item $2$ since $ \det(L_{\set*{1,2}} ) = c_1^2 + x_1^2 >> \det(L_{\set*{1, i}}) = c_1 c_i \forall i \neq 1,$ and so on. On the other hand, the optimal subset is $\set*{n-2t+1, \dots, n}$ by our choice of $x$, and this could be arbitrarily better than \cref{alg:common-greedy}'s solution. We may think of items $2i- 1$ and $2i$ as complementary items, say, e.g., toothpaste and toothbrush proposed in a recommender system. The conditions on $c_i$'s and $x_i$'s mean that the degree of complementarity between these pairs increases with $i$. So $2n-1$ and $2n$ are the most likely pair to appear together, but each one of $2n-1$ and $2n$ is most unlikely to appear as a singleton, and the opposite holds for item $1$ and $2$; for example, think of $2n-1$ and $2n$ as a tea cup and tea cup lid, which are almost always bought together, but $1$ and $2$ as toothpaste and toothbrush, which are sometimes purchased separately.
Furthermore, switching out any item $2 i-1 $ or $2i$ in $S$ for an item $2j-1$ or $2j$ outside of $S$ reduces the determinant by $ (c_i^2 + x_i^2) /c_i c_j > c_i >1,$ so $S $ is also maximum among its $1$-neighborhood. Thus local search greedy, or equivalently, local search initialized at $S$, will simply output $S$ itself.
\end{example}
We remark that it is easy to construct an example where \cref{alg:common-greedy} produces a subset with zero determinant, whereas the optimal subset can have arbitrarily large determinant. E.g., in \cref{remark:greedy same as localSearch}, we can make all diagonal entries except for $L_{1,1}$ zero; then, \cref{alg:common-greedy} with even $k$ will necessarily produces a zero determinant.
\begin{comment}
Though this greedy algorithm is guaranteed to obtain a $k^{O(k)}$-approximation for symmetric DPPs \cite{CM10}, it could not achieve even a finite approximation factor for nPSD-NDPPs. For example, on a skew-symmetric matrix $X$, i.e., $X = -X^{\intercal}$, since all odd-sized principle minors of $X$ are zero, \cref{alg:common-greedy} would necessarily resort to picking an arbitrary/random item at every other iteration, which can result in an arbitrarily bad final answer. Consider a concrete example, which helps build intuition on why greedy fails to achieve a meaningful approximation factor. This example also shows that local search greedy \cite{KD16}\footnote{This algorithm starts with the output $S$ of \cref{alg:common-greedy}, then continuously swaps out an element in $S$ with one outside $S$ to increase the DPP score, until either a local maximum is reached or $k^2 \log k$ swaps have been performed.}, another candidate MAP inference algorithm with theoretical performance guarantees for symmetric DPPs, also used by \citet{gartrell2020scalable} as a baseline to compare their greedy method, also fails to achieve a meaningful approximation factor.
\begin{example}
Consider the same $L$ as defined in \cref{remark:greedy same as localSearch}
It is easy to check that \cref{alg:common-greedy} (greedy) on input $k= 2t$ will select $ S = \set*{1,2 ,\dots, 2t-1, 2t}. $ Indeed, \cref{alg:common-greedy} first picks item $1$ since $\det(L_{\set*{1}}) =c_1$ is maximum among all $L_{\set*{i}},$ then picks item $2$ since $ \det(L_{\set*{1,2}} ) = c_1^2 + x_1^2 >> \det(L_{\set*{1, i}}) = c_1 c_i \forall i \neq 1,$ and so on. On the other hand, the optimal subset is $\set*{n-2t+1, \dots, n}$ by our choice of $x$, and this could be arbitrarily better than \cref{alg:common-greedy}'s solution. We may think of items $2i- 1$ and $2i$ as complementary items, say, e.g., toothpaste and toothbrush. The conditions on $c_i$'s and $x_i$'s mean that the degree of complementarity between these pairs increases with $i$. So $2n-1$ and $2n$ are the most likely pair to appear together, but each one of $2n-1$ and $2n$ is most unlikely to appear as a singleton, and the opposite holds for item $1$ and $2$; for example, think of $2n-1$ and $2n$ as a tea cup and tea cup lid, which are almost always bought together, but $1$ and $2$ as toothpaste and toothbrush, which are sometimes purchased separately. Furthermore, switching out any item $2 i-1 $ or $2i$ in $S$ for an item $2j-1$ or $2j$ outside of $S$ reduces the determinant by $ (c_i^2 + x_i^2) /c_i c_j > c_i >1,$ so $S $ is also maximum among its $1$-neighborhood. Thus local search greedy, or equivalently, local search initialized at $S$, will simply output $S$ itself.
\end{example}
We remark that it is easy to construct an example where greedy produces a subset with zero determinant, whereas the optimal subset can have arbitrarily large determinant. E.g., in \cref{remark:greedy same as localSearch}, we can make all diagonal entries except for $L_{1,1}$ zero; then, \cref{alg:common-greedy} with even $k$ will necessarily produces a zero determinant.
\end{comment}
As our main application, we show the first efficient algorithm for MAP inference on nPSD-NDPPs that gives a \emph{multiplicative} factor approximation for $\max\set{\det(L_{S})\given S\in \binom{[n]}{k}},$ without requiring any additional assumption on the kernel matrix $L$. Further, we obtain multiplicative approximation guarantees for $\det(L_{S})$, unlike prior related work \cite{gartrell2020scalable} which obtained multiplicative approximations for $\log \det(L_{S})$; this is often a stronger guarantee when $\OPT$ is sufficiently large -- roughly super-exponentially large in $k$. The assumptions behind prior work often implicitly imply that $\OPT$ is at least exponentially large in $k$, making our approximation guarantees attractive.
\begin{theorem}\label{thm:nonsym DPP main}
There is a polynomial time algorithm that on input $L\in \R^{n\times n}$ that is nPSD, outputs a set of indices $S\in \binom{[n]}{k}$ guaranteeing
\[ \det(L_{S})\geq k^{-O(k)}\cdot \max\set*{\det(L_{S})\given S\in \binom{[n]}{k}}. \]
Moreover, the algorithm runs in $O(n^4 k + n^2 k^5\log n) $ time given the entries of $L$, and $O(n^2 d^2 k + n^2 d^2 k^3 \log n)$ time given a rank-$d$ decomposition of $L$, i.e., $L = B C B^{\intercal}$ with $B\in \R^{n\times d}, C\in \R^{d\times d}.$
\end{theorem}
Our approximation factor matches that of the standard greedy heuristic on \textit{symmetric} DPPs, as well as the guarantee of other simple heuristics proposed for \textit{symmetric} DPPs \cite{CM10,KD16}. As mentioned earlier, \citet{CM10}'s greedy and \citet{KD16}'s local search algorithm do not achieve any finite approximation factor for nPSD-NDPPs. Our result is incomparable to \citet{gartrell2020scalable} as \begin{enumerate}[label=\roman*., series = tobecont, itemsep=0em, topsep=0em]
\item multiplicative
approximations for maximizing $\log \det(L_S)$ do not imply similar results for $\det(L_S)$,
\item we place no additional assumption on $L$. As demonstrated earlier, our approximation guarantees hold for matrices $L$ where \cref{alg:common-greedy} fails to achieve even a finite approximation factor.
\end{enumerate}
Our local search algorithm for nPSD-NDPPs searches over $2$ neighborhoods, unlike most prior related works which typically use $1$ neighborhoods; using $2$ neighborhoods is necessary, and is compatible with intuition from prior work of \textcite{AV20} who first studied $2$ neighborhood local search for the related problem of finding the maximum $k\times k$ subdeterminant of a rectangular matrix. Unlike \cite{AV20}, our analysis of local search is not based on algebraic identities, which we believe do not have a counterpart in the world of nPSD-DPPs, but rather mixing properties of random walks.
\begin{Algorithm}
Initialize $S \gets \emptyset$.\;
\While{$\abs{S} < k$}{
Pick $i\not\in S$ that maximizes $\det(L_{S\cup \set*{i}}),$ and update $S \gets S \cup \set*{i}$.\;
}
\caption{Standard GREEDY for DPPs} \label{alg:common-greedy}
\end{Algorithm}
\subsection{Composable Core-Sets for Strongly Rayleigh Distributions and Log-Concave Polynomials}
As further application of our methods, we extend prior work of \textcite{Mahabadi2019ComposableCF} on the construction of composable core-sets for maximizing symmetric DPPs to the more general class of distributions that satisfy the strongly Rayleigh property \cite{BBL09} or have a log-concave generating polynomial \cite{AOV18}.
Composable core-sets are a tool \cite{indyk2014composable} to handle computational problems involving large amounts of data. Roughly speaking, a core-set is a summary of a dataset that is enough to solve the computational problem at hand; a \emph{composable} core-set has the additional property that the union of summaries for multiple datasets is itself a good summary for union of all datasets. More precisely, in the context of the optimization problem on $\mu: \binom{[n]}{k}\to \R_{\geq 0}$, a function $c$ that maps any set $P\subseteq [n]$
to one of its subsets is called an $\alpha$-composable core-set (\cite{Mahabadi2019ComposableCF}) if it
satisfies the following condition: given any integer $m$ and
any collection of sets $P_1, \cdots , P_m \subseteq [n]$
\[\alpha \cdot \max\set*{\mu(S) \given S \subseteq \bigcup_{i=1}^m c(P_i)} \geq \max \set*{\mu(S) \given S \subseteq \bigcup_{i=1}^m P_i}. \]
We also say $c$ is a core-set of size $ t$ if $\card{c(P)} \leq t$ for all sets $P$. Composable core-sets are very verstaile; when a composable core-set is designed for a task, they automatically imply
efficient streaming and distributed algorithms for the same task.
One strategy for constructing composable core-sets is local search. \Textcite{Mahabadi2019ComposableCF} showed that for $k$-DPP parameterized by symmetric PSD matrix $L$, (1-step)-local seaarch (\cref{alg:localsearch} with $r=1$) gives a $k^{O(k)}$-composable core-sets of size $k.$ The approximation factor of $k^{O(k)}$ is nearly optimal.
Recall that $k$-DPP parameterized by symmetric PSD matrix $L$ belongs to the family of homogeneous \emph{strongly Rayleigh} distributions, i.e., distributions $\mu$ whose generating polynomial $g_{\mu}$ is nonvanishing on the upper half plane \cite{BBL09}. An even more general family of distributions is the family of log-concave distributions \cite{AOV18}.
We extend \cite{Mahabadi2019ComposableCF}'s result to any distribution $\mu: \binom{[n]}{k} \to \R_{\geq 0}$ that is strongly Rayleigh or has a log-concave generating polynomial.
\begin{theorem} \label{thm:core-set}
Given a distribution $\mu: \binom{[n]}{k} \to \R_{\geq 0}$, let $c$ be a map that takes $P \subseteq [n]$ to some $c(P)\in \binom{P}{k}$ that is an $\zeta$-approximate local maximum in the $1$-neighborhood with respect to $\mu,$ for some fixed constant $\zeta \in (0,1)$:
\[ \mu(c(P))\geq \zeta \cdot \max\set*{\mu(S)\given S\in \cN_1(c(P))}. \]
Then $c$
is an $\alpha$-composable core-set
of size $k$ for the MAP-inference problem on $\mu$ with $\alpha = k^{O(k)}$ for strongly Rayleigh $\mu$, and $\alpha=2^{O(k^2)}$ when $\mu$ has a log-concave generating polynomial.
\end{theorem}
\subsection{Techniques}
Our main tool for proving \cref{thm:main} is a form of (approximate) exchange inequality. Exchange inequalities have been traditionally been studied in discrete convex analysis \cite{Mur16}, but have recently been extended and used in sampling \cite{ALOVV21} and optimization \cite{AV20} problems beyond the reach of traditional discrete convex analysis. Unlike prior works, here we go in the opposite direction and show that efficient sampling implies a form of exchange inquality. To prove \cref{thm:main}, we set the external field $\lambda$ appropriately, and use the lower bound on the spectral gap of the down-up walk on $\lambda \ast \mu$ to derive our approximate exchange property \cref{lem:approx exchange FLC}.
We then show that this approximate exchange property implies the desired approximation factor for local search (\cref{cor:local search guarantee FLC}).
Since nonsymmetric DPPs are $1/4$-fractionally log-concave \cite{alimohammadi2021fractionally},
\cref{thm:main} already implies an efficient algorithm (\cref{alg:localsearch} with $r=4$) to get $k^{O(k)}$-approximation factor for the MAP inference problem on nonsymmetric DPPs. We can further improve the the local search radius $r$ to $2$, and get a faster algorithm that matches the runtime stated in \cref{thm:nonsym DPP main} by showing a stronger approximate exchange property (\cref{thm:approximate_exchange}).
To prove \cref{thm:core-set}, we use the approximate exchange property introduced by \cite{ALOVV21} that is satisfied by strongly Rayleigh and log-concave distributions. This exchange property is a quantitative version of the strong basis exchange axiom for matroids. We rename it the strong approximate basis exchange property (\cref{def:strong basis exchange}), to distinguish it from weaker exchange properties that we show in this paper. We show that the strong approximate basis exchange implies that approximate local maxima in the 1-neighborhood is a size-$k$ core-set with the desired approximation factor (\cref{lem:core-set}).
\section{Preliminaries}
We use $[n]$ to denote the set $\set{1,\dots,n}$ and $\binom{[n]}{k}$ to denote the family of size $k$ subsets of $[n]$. We use $\mathds{1}$ to denote the all $1$ vector. When $n$ is clear from context, we use $\mathds{1}_S\in \R^n$ to denote the indicator vector of the set $S\subseteq [n]$, having a coordinate of $0$ everywhere except for elements of $S$, where the coordinate is $1$. For sets $S, T$ of the same size we define their \emph{distance} to be
$ d(S, T): = \card{S\Delta T}/2 = \card{S \setminus T} = \card{T \setminus S}$.
With this notion of distance, we can define neighborhoods:
\begin{definition}
For $r \geq 0$ let the $r$-neighborhood of $S\in \binom{[n]}{k}$ be
\[ \cN_r(S) := \set*{T \in \binom{[n]}{k} \given d(S, T)\leq r }. \]
\end{definition}
For a density $\mu: 2^{[n]} \to \mathbb{R}_{\geq 0}$, the \emph{generating polynomial} of $\mu$ is defined as
$$
g_\mu(z_1, \ldots, z_n) = \sum_{S \in 2^{[n]}} \mu(S) \prod_{i \in S} z_i
$$
\subsection{Determinantal Point Processes (DPPs)}
A DPP on a set of $n$ items defines a probability distribution over subsets $Y \subseteq [n].$ It is parameterized by a matrix $L \in \R^{
n\times n}$: $\P_{L}{Y } \propto \det(L_Y ),$ where $L_Y$ denote the principle submatrix whose columns and rows are indexed by $Y.$ We call $L$ the kernel matrix.
For $Y \subseteq [n]$, if we condition the distribution $\mathbb{P}_{L}$ on the event that items in $Y$ are included in the sample, we still get a DPP; the new kernel is given by the Schur complement $L^Y = L_{\tilde{Y}} - L_{\tilde{Y}, Y} L_{Y, Y}^{-1} L_{Y, \tilde{Y}} $ where $\tilde{Y} = [n] \setminus Y.$
Given a cardinality constraint $k$, the $k$-DPP paremeterized by $L$ is a distribution over subsets of size $k$ of $Y$ defined by $\mathbb{P}_{L}^k[Y] =\frac{ \det(L_Y) } {\sum_{\abs{Y'} =k } \det (L_{Y'})} . $
To ensure that $\mathbb{P}_L$ defines a probability distribution, all principal minors of $L$ must be
non-negative: $\det(L_S ) \geq 0.$ Matrices that satisfy this property are called $P_0$-matrices \citep[Definition 1]{FANG19891}. Any nonsymmetric (or symmetric) PSD matrix is automatically $P_0$-matrix \citep[Lemma 1]{Gartrell2019LearningND}.
We say a NDPP kernel $L\in \R^{n\times n}$ has a low-rank decomposition \cite{Gartrell2019LearningND,gartrell2020scalable} if $L$ can be written as $L = B C B^{\intercal}$ for some $d\leq n$, $B\in \R^{n\times d}, C\in \R^{d\times d}.$ Clearly, $\rank(L) = d$, and we say $L = B C B^{\intercal}$ is a rank-$d$ decomposition of $L.$ We will need the following identity, which is derived from Schur complements; it has previously appeared in \cite{gartrell2020scalable}. For $S \subseteq [n],$ let $B_S$ denote the sub-matrix of $B$ consisting of rows in $S$; then $L_S = B_S C B_S^{\intercal} $ and
\begin{equation} \label{eq:condition kernel}
\begin{split}
\det(L_{Y \cup D}) &= \det(L_{Y}) \det( L_D - L_{D, Y} L_Y^{-1} L_{Y,D})\\
&= \det(L_{Y}) \det(L_D - B_D C (B_Y^{\intercal} L_Y^{-1} B_Y) C B_D^T ).
\end{split}
\end{equation}
Given $\det(L_Y)$ and $L_Y^{-1},$ we can compute $\det(L_{Y \cup D})$ in $O(\abs{D} d^2 + \abs{D}^2 d+ \abs{D}^3)$ time.
\subsection{MAP Inference}
Given a density $\mu:\binom{[n]}{k} \to \R_{\geq 0},$ the optimization with respect to $\mu$ or MAP inference on $\mu$ is to find
\[S^*: = \arg\max_{S \in \binom{[n]}{k}} \mu(S).\]
Throughout the paper, we let $\OPT:= \max_{S\in \Omega} \mu(S).$
We say an algorithm gives a factor $c$-approximation for MAP inference on $\mu$ if its output $\hat{S} \in \binom{[n]}{k}$ such that $c \dot \mu(\hat{S}) \geq \OPT.$
When $\mu$ is defined by a DPP, i.e. $\mu(S) = \det(L_{S,S})$ for a $n \times n$ matrix $L,$ MAP inference on $\mu$ is also called the determinant maximization problem \cite[e.g., see][]{Mahabadi2019ComposableCF}.
\subsection{Markov Chains}
For two measures $\mu, \nu$ defined on the same state space $\Omega$, we define their total variation distance as
\[ \dtv(\mu, \nu)=\frac{1}{2}\sum_{\omega\in \Omega}\abs{\mu(\omega)-\nu(\omega)}=\max\set{\P_\mu{S}-\P_\nu{S}\given S\subseteq \Omega}. \]
A Markov chain on a state space $\Omega$ is defined by a row-stochastic matrix $P\in \R^{\Omega\times \Omega}$. We view distributions $\mu$ on $\Omega$ as row vectors, and as such $\mu P$ would be the distribution after one transition according to $P$, if we started from a sample of $\mu$. A stationary distribution $\mu$ for the Markov chain $P$ is one that satisfies $\mu P=\mu$. Under mild assumptions on $P$ (ergodicity), stationary distributions are unique and the distribution $\nu P^t$ converges to this stationary distribution as $t\to \infty$ \cite{LP17}. We refer the reader to \cite{LP17} for a detailed treatment of Markov chain analysis.
In this paper, we will only consider reversible Markov chain. We say a Markov chain with transition matrix $P$ is reversible if \[\mu(x) P(x,y) =\mu(y) P(y,x)\forall x, y\in \Omega.\]
The conductance\footnote{also known as bottleneck ratio in \cite{LP17}} of a subset $S$ of states in a Markov chain is
\[\Phi(S) = \frac{Q(S, \Omega \setminus S)}{\mu(S)}\]
where $Q(S, \Omega\setminus S) = \sum_{x\in S, y \in \Omega\setminus S} \mu(x) P(x,y)$ is the ergodic flow between $S$ and $\Omega\setminus S,$ and $\mu(S) = \sum_{x\in S} \mu(x).$
The conductance of a Markov chain is defined as the minimum conductance
over all subsets $S$ with $\mu(S) \leq 1/2,$ i.e.
\[\Phi = \min_{S: \mu(S) \leq 1/2} \Phi(S)\]
\begin{theorem}[{\cite[see, e.g.,][Thm. 13.10]{LP17}}] \label{thm:cheeger}
Let $\lambda_2$ be the second largest eigenvalue of the transition matrix $P$, then
\[ \frac{\Phi^2}{2}\leq 1-\lambda_2 \leq 2\Phi. \]
\end{theorem}
For a Markov chain $P$, we define the mixing time from a starting distribution $\nu$ as the first time $t$ such that $\nu P^t$ gets close to the stationary distribution $\mu$.
\[ \tmix(P, \nu, \epsilon)=\min\set{t\given \dtv(\nu P^t, \mu)\leq \epsilon}. \]
We drop $P$ and $\nu$ if they are clear from context. If $\nu$ is the Dirac measure on a single point $\omega$, we write $\tmix(P, \omega, \epsilon)$ for the mixing time. When mixing time is referenced without mentioning $\epsilon$, we imagine that $\epsilon$ is set to a reasonable small constant (such as $1/4$). This is justified by the fact that the growth of the mixing time in terms of $\epsilon$ can be at most logarithmic \cite{LP17}.
We can relate the mixing time and conductance as follow.
\begin{theorem}[{\cite[see, e.g.,][Thm. 7.4]{LP17}}]
For a reversible Markov chain $P$ with conductance $\Phi,$ we have
\[ \tmix(P, 1/4) \leq 4 \Phi.\]
\end{theorem}
\subsection{The Down-Up Random Walk}
Consider a distribution $\mu: \binom{[n]}{k}\to \R_{\geq 0}.$ The down-up walk is given by the composition of two row-stochastic operators, known as the down and up operators.
\begin{definition}[Down Operator]
For a ground set $\Omega$, and $|\Omega| \geq k\geq \l$, define the down operator $D_{k\to \l}\in \R^{\binom{\Omega}{k}\times \binom{\Omega}{\l}}$ as
\[
D_{k\to \l}(S, T)=\begin{cases}
\frac{1}{\binom{k}{\l}}&\text{ if }T\subseteq S,\\
0&\text{ otherwise}.\\
\end{cases}
\]
\end{definition}
Note that $D_{k\to \l}D_{\l\to m}=D_{k\to m}$.
\begin{definition}[Up Operator]
For a ground set $\Omega$, $|\Omega|\geq k\geq \l$, and density $\mu:\binom{\Omega}{k}\to \R_{\geq 0}$, define the up operator $U_{\l \to k}\in \R^{\binom{\Omega}{\l}\times \binom{\Omega}{k}}$ as
\[
U_{\l\to k}(T, S)=\begin{cases}
\frac{\mu(S)}{\sum_{S'\supseteq T}\mu(S')}&\text{ if }T\subseteq S,\\
0&\text{ otherwise}.\\
\end{cases}
\]
\end{definition}
If we define $\mu_k=\mu$ and more generally let $\mu_\l$ be $\mu_k D_{k\to \l}$, then the down and up operators satisfy
\[ \mu_k(S)D_{k\to \l}(S, T)=\mu_\l(T)U_{\l \to k}(T, S). \]
This property ensures that the composition of the down and up operators have the appropriate $\mu$ as a stationary distribution, are reversible, and have nonnegative real eigenvalues.
\begin{proposition}[{\cite[see, e.g.,][]{KO18,AL20,ALO20}}]
The operators $D_{k\to \l}U_{\l\to k}$ and $U_{\l\to k}D_{k\to \l}$ both define Markov chains that are time-reversible and have nonnegative eigenvalues. Moreover $\mu_k$ and $\mu_\l$ are respectively their stationary distributions.
\end{proposition}
\begin{definition}[Down-Up Walk]
For a ground set $\Omega$, $|\Omega|\geq k\geq \l$, and density $\mu:\binom{\Omega}{k}\to \R_{\geq 0}$, the $k\leftrightarrow \ell$ down-up walk is defined by the row-stochastic matrix $U_{\ell \to k}D_{k\to \ell}$.
\end{definition}
\subsection{Real-Stable and Sector-Stable Polynomials}
We use $\F[z_1,\dots,z_n]$ to denote $n$-variate polynomials with coefficients from $\F$, where we usually take $\F$ to be $\R$ or $\C$. We denote the degree of a polynomial $g$ by $\deg(g)$. We call a polynomial homogeneous of degree $k$ if all nonzero terms in it are of degree $k$.
\begin{definition}[Stability]
For an open subset $U\subseteq \C^n$, we call a polynomial $g\in \C[z_1,\dots,z_n]$ $U$-stable iff
\[ (z_1,\dots,z_n)\in U\implies g(z_1,\dots,z_n)\neq 0. \]
We also call the identically $0$ polynomial $U$-stable. This ensures that limits of $U$-stable polynomials are $U$-stable. For convenience, when $n$ is clear from context, we abbreviate stability w.r.t.\ regions of the form $U\times U\times \cdots \times U$ where $U\subseteq \C$ simply as $U$-stability.
\end{definition}
Our choice of the region $U$ in this work is the product of open sectors in the complex plane.
\begin{definition}[Sectors]
We name the open sector of aperture $\alpha\pi$ centered around the positive real axis $\Gamma_\alpha$:
\[ \Gamma_\alpha:=\set{\exp(x+iy)\given x\in \R, y\in (-\alpha\pi/2,\alpha\pi/2)}. \]
\end{definition}
Note that $\Gamma_1$ is the right-half-plane, and $\Gamma_1$-stability is the same as the classically studied Hurwitz-stability \cite[see, e.g.,][]{Bra07}. Another closely related notion is that of real-stability where the region $U$ is the upper-half-plane $\set{z\given \Im(z)>0}$ \cite[see, e.g.,][]{BBL09}. Note that for \emph{homogeneous} polynomials, stability w.r.t.\ $U$ is the same as stability w.r.t.\ any rotation/scaling of $U$; so Hurwitz-stability and real-stability are the same for \emph{homogeneous} polynomials.
We use $\alpha$-sector-stable as a shorthand for $\Gamma_\alpha$-stable. Naturally, we call a distribution $\alpha$-sector-stable if its generating polynomial is $\alpha$-sector-stable.
\begin{proposition}[{\citep{alimohammadi2021fractionally}}]\label{prop: ssProperties}
The following operations preserve $S_\alpha$-sector-stability on homogeneous multi-affine polynomials:
\begin{enumerate}[itemsep=0em, topsep=0em]
\item\label{part: spec}Specialization: $g\mapsto g(a,z_2,\ldots,z_n)$, for $a\in \bar S_\alpha$.
\item \label{part: derivative} Derivative: $g \mapsto \frac{\partial}{\partial z_1} g(z_1, \cdots, z_n) $.
\item \label{part:scaling} Scaling: $g \mapsto g(\lambda_1 z_1, \ldots, \lambda_n z_n)$, for $\lambda \in \R_{\geq 0}^n.$
\end{enumerate}
\end{proposition}
We state some examples of sector stable distributions.
\begin{lemma}[{\citep{alimohammadi2021fractionally}}] \label{lem:p0Constrained}
Consider $L\in \R^{n \times n}$ that is nPSD, i.e., $L + L^T\succcurlyeq 0$, then $\mu: \binom{[n]}{k}\to \R_{\geq 0}$ defined by $\mu(S) = \det(L_{S,S})$ is $1/2$-sector-stable.
\end{lemma}
\begin{lemma}[{\citep{alimohammadi2021fractionally}}] \label{lem:strong Rayleigh partition constraint}
Given a density $\mu:\binom{[n]}{k}\to \R_{\geq 0}$ and a partition $T_1\cup T_2\cup \cdots \cup T_s=[n]$, and numbers $c_1,\dots,c_s\in \Z_{\geq 0}$, let the partition constraint density $\mu_{T, c}$ be $\mu$ restricted to sets $S\in \binom{[n]}{k}$ where $\card{S \cap T_i} = c_i.$ When $\mu$ is strongly Rayleigh, $\mu_{T,c}$ is $1/2^c$-sector-stable.
\end{lemma}
\subsection{Log-Concavity and Fractional Log-Concavity}
We now formally introduce \emph{log-concavity} for distributions over size-$k$ subsets of $n$ elements, and its direct generalization, \emph{$\alpha$-fractional-log-concavity}.
\begin{definition}
A function $f: \mathbb{R}_{\geq 0}^n \to \mathbb{R}_{\geq 0}$ is \emph{log-concave} if $\log f(z_1, \ldots, z_n)$ is concave over $\mathbb{R}_{\geq 0}^n$, i.e. for all $x, y \in \mathbb{R}_{\geq 0}^n$ and $\lambda \in [0, 1]$, we have:
$$
g(\lambda x - (1 - \lambda)y) \geq g(x)^\lambda \cdot g(y)^{1 - \lambda} \Longleftrightarrow \log g(\lambda x - (1 - \lambda)y) \geq \lambda \log g(x) + (1 - \lambda) \log g(y)
$$
We say a probability distribution $\mu: \binom{[n]}{k} \to \mathbb{R}_{\geq 0}$ is \emph{log-concave} if $\log g_\mu(z_1, \ldots, z_n)$ is concave over $\mathbb{R}_{\geq 0}$, or in other words, that its generating polynomial is a log-concave function over $\mathbb{R}_{\geq 0}$.
\end{definition}
\cite{ALOV19,BH19} shows that for homogeneous multiaffine polynomials, real-stability implies log-concavity. A similar relationship holds for sector stability and fractional log-concavity.
\begin{lemma} \label{lem:sector-stability-to-log-concavity}
\emph{(Lemma 67 from \cite{alimohammadi2021fractionally})} If a polynomial $g$ is $\alpha$-sector-stable, then it is $\frac{\alpha}{2}$-fractionally-log-concave.
\end{lemma}
We note that scaling preserves $\alpha$-log-concavity
of homogeneous distributions \cite{alimohammadi2021fractionally} i.e. if $\mu$ is $\alpha$-log-concave, then so is $\lambda \ast \mu$ for all $\lambda \in \R_{\geq 0}^n.$
\begin{theorem}[{\cite{alimohammadi2021fractionally,AJKPV21}}] \label{thm:spectral gap FLC}
Suppose $\mu: \binom{[n]}{k}$ is $\alpha$-fractional-log-concave. The $k \leftrightarrow (k-\lceil 1/\alpha \rceil)$-down-up-walk on $\mu$ has spectral gap at least $\Omega(k^{-1/\alpha}).$
\end{theorem}
\subsection{Composable Core-Set}
\begin{definition}[{\cite[Definition 2.2]{Mahabadi2019ComposableCF}}] \label{def:composable core-set}
A function $c(P)$ that maps the input set $P \subseteq \R^d$ to one of its subsets is called an $\alpha$-composable core-set for a function $f:2 ^{\R^d} \to \R$ if, for any collection of sets $P_1, \ldots, P_n\subseteq \R^d$ we have $f(C) \geq f(P)/\alpha$ where $P = \bigcup_{i\leq n} P_i$ and $C= \bigcup_{i\leq n} c(P_i)$
\end{definition}
\section{MAP Inference via Local Search} \label{sec:implement}
In this section, we show how to efficiently find a local optima\footnote{More precisely, we show how to find an approximate local optima, which is sufficient for our purpose.} of a given distribution $\mu.$ We run a two stage algorithm:
\begin{enumerate}[label = (\roman*)]
\item first, we find some ``good'' initial subset $S_0 \in \binom{[n]}{k}$, i.e., one such that the ratio $\OPT/\mu(S_0)$ is bounded by $2^{\poly(n,k)}$ (see \cref{lem:crude}),
\item \label{step: local search} then, for a suitably chosen radius $r \in \N_{\geq 1}$, we run a simple local search (\cref{alg:localsearch}) that starts with $S\leftarrow S_0$, and find better and better solutions by swapping \emph{at most $r$} elements in $S$ for elements outside of $S$
until no more improvement in term of $\mu(S)$ can be found.
\end{enumerate}
To ensure that our algorithm terminates within polynomial time, we will only take improvements that increase the determinant by at least a lower multiplicative threshold, say, by a factor of $2$.
\begin{Algorithm}
\textbf{Input:} $ \alpha \leq 1, S_0 \in \binom{[n]}{k}$ with $\mu(S_0) > 0.$\;
Initialize $S \leftarrow S_0$.\;
\While{$\mu(S) < \zeta\cdot \mu(T)$ for some $T \in \cN_r(S)$}{
Update $S \leftarrow \arg\max_{T\in \cN_r(S)} \mu(T)$.\;
}
\caption{LOCAL-SEARCH-$r$ ($\LS_r$)} \label{alg:localsearch}
\end{Algorithm}
We prove the algorithmic part of \cref{thm:main},
that with a suitable choice for $S_0$, \cref{alg:localsearch} runs in polynomial time.
\begin{proposition}\label{prop:steps-bound}
The number of steps taken by \cref{alg:localsearch} with $r = O(1)$ starting from $S_0$ is at most
\[ \log_{1/\alpha}\parens*{\OPT/\mu(S_0)}. \]
Each step can be implemented using $O((nk)^r)$ oracle access to $\mu.$
\end{proposition}
\begin{proof}
Each iteration improves $\mu(S)$ by a factor of at least $1/\alpha$. On the other hand, this value can never exceed $\OPT$, and it starts as $\mu(S_0) > 0$.
Clearly, to perform local search in the $r$-neighborhood of a set $S,$ we only need to query $\mu((S \setminus U_1) \cup U_2)$ for $U_1 \in \binom{S}{\leq r}$ and $U_2 \in \binom{[n]}{\leq r}.$ The total number of such queries is $ O((nk)^r).$
\end{proof}
\begin{definition}
For $\mu: \binom{[n]}{k}\to \R_{\geq 0}$ and $\alpha > 0,$ we say $S\in \binom{[n]}{k}$ is a $(r,\zeta)$-local maximum w.r.t. $\mu$ if
$\mu(S) \geq \zeta \mu(T)$ for all $T\in \cN_r(S)$.
\end{definition}
Clearly, when \cref{alg:localsearch} terminates, the output is a $(r,\zeta)$-local maximum.
Next, we show how to obtain a ``good'' initialization $S_0$ by a simple greedy algorithm, which we call INDUCED-GREEDY, that is based on maximizing the \textit{marginal gain} defined by the distribution on size $\leq k$ subsets. This gain is \emph{induced} by the distribution $\mu,$ as defined below.
For subset $T$ of $[n]$ of size $\leq k$, let $\mu(T) = \sum_{S \in \binom{[n]}{k}: S \supseteq T} \mu(S).$
\begin{Algorithm}
Initialize $S \leftarrow \emptyset$.\;
\While{$\abs{S} < k$}{
Pick $i\not\in S$ that maximizes $\mu(S \cup \set*{i})$ and update $S \leftarrow S \cup \set*{i}$.\;
}
\caption{INDUCED-GREEDY} \label{alg:greedy}
\end{Algorithm}
\begin{lemma}\label{lem:crude}
\cref{alg:greedy}
returns $S_0$ with
\[ O(n^k) \cdot \det(L_{S_0})\geq \OPT. \]
\end{lemma}
\begin{proof
For $j\in [k],$ let $i_j$ be the element added to $S$ at the $j$-th iteration of the while loop. Let $S_0 = \emptyset$, $S_j = S_{j-1} \cup \set*{i_j}.$ Observe that $\abs{S_j} =j$ and for each $j\geq 0$
\begin{align*}
\mu(S_j) &= \frac{1}{k - \abs{S_j} } \sum_{i\not\in S_j} \mu(S_j\cup \set*{i}) \leq \frac{n - j }{k - j } \mu(S_{j+1})
\end{align*}
thus $\binom{n}{k}\mu(S_k) \geq \mu(S_0) = \mu(\emptyset) = \sum_{S'\in \binom{[n]}{k}} \mu(S') \geq \OPT$.
\end{proof}
\begin{remark} \label{remark:greedy runtime}
In \cref{alg:greedy}, it is enough to find $i$ that approximately maximizes $\mu(S\cup {i})$ i.e. for some constant $\zeta \in (0,1)$, $\mu(S\cup {i}) \geq \zeta \mu(S \cup {j}) $ for all $j\not \in S.$ In that case, \cref{lem:crude} still holds, and \cref{alg:greedy} can be efficiently implemented given access to efficient algorithms that approximately sample from $\lambda \ast \mu$ for $\lambda \in \R_{\geq 0}^n.$ Indeed, note that $\mu(S\cup {i})/\mu(S)$ is the marginal of $\lambda \ast \mu$ where $\lambda_i=\begin{cases} \infty &\text{ for } i \in S \\1 &\text{ else} \end{cases}.$ Thus, $\mu(S\cup {i})$ can be approximate within some small constant factor.
\end{remark}
\section{From Sampling to Optimization via Local Search}
In this section, we prove \cref{thm:main}.
\begin{definition}[$r$-exchange]
For $\mu: \binom{[n]}{k} \to \R_{\geq 0}$, $r \in \N$ and $S, T \in \binom{[n]}{k}$, we let \[\cE^r(S,T) : = \set*{U \subseteq S\Delta T \given \abs{U \cap S } = \abs{U \cap T } =r} \] be the set of all $r$-exchanges between $S$ and $T.$
\end{definition}
\begin{definition}[Weak $(r,\beta)$-approximate exchange]
We say a distribution $\mu: \binom{[n]}{k}$ satisfies \emph{weak $(r,\beta)$-approximate exchange} if for any $S, T \in \binom{[n]}{k},$ there exists $s \in \set*{1,\cdots,r}$ and $U \in \cE^{s}(S,T) $ such that
\[\mu(S) \leq \beta\cdot \mu(S \Delta U) \parens*{\frac{\mu(S)}{\mu(T)}}^{s/d(S,T)}\]
\end{definition}
\begin{lemma}\label{lem:approx exchange FLC
Consider $\mu:\binom{[n]}{k} \to \R_{\geq 0}$ such that for all external field $\lambda\in \R_{\geq 0}^n,$ the conductance of the $k\leftrightarrow (k-r)$-down-up walk on $\lambda \ast \mu$ is at least $\Omega(k^{-c}).$ Then $\mu$ satisfies weak $(r, O(k^{r+c}))$-approximate exchange.
\end{lemma}
\begin{proposition} \label{cor:local search guarantee FLC}
If $\mu:\binom{[n]}{k} \to \R_{\geq 0}$ satisfies weak $(r,\beta)$-approximate exchange then any $(r,\zeta)$-local max with respect to $\mu$ is also an $O((\beta/\zeta)^k)$-approximate global max.
\end{proposition}
In particular, when $\mu$ is $\alpha$-fractionally log-concave, \cref{lem:approx exchange FLC,cor:local search guarantee FLC} hold with $r = \lceil1/\alpha \rceil$ and $ c= 1/\alpha$ and $\beta = O(k^{r+c}).$
\begin{corollary}\label{cor:strongly Rayleigh partition}
Let $\mu:\binom{[n]}{k} \to \R_{\geq 0}$ be strongly Rayleigh. Given a density $\mu:\binom{[n]}{k}\to \R_{\geq 0}$ and a partition $T_1\cup T_2\cup \cdots \cup T_s=[n]$, and numbers $c_1,\dots,c_s\in \Z_{\geq 0}$, let the partition constraint density $\mu_{T, c}$ be $\mu$ restricted to sets $S\in \binom{[n]}{k}$ where $\card{S \cap T_i} = c_i.$ When $\mu$ is strongly Rayleigh and $c=O(1)$, one can efficiently finds a $k^{O(k)}$-approximation for $\max \mu_{T,c} (\cdot)$ using \cref{alg:localsearch} with $r = 2^c.$
\end{corollary}
The local search guarantee in \cref{thm:main} follows from \cref{thm:spectral gap FLC,thm:cheeger,lem:approx exchange FLC,cor:local search guarantee FLC}, and the runtime bound follows from \cref{remark:greedy runtime,prop:steps-bound}.
\begin{proof}[Proof of \cref{lem:approx exchange FLC}]
If $d(S,T) \leq r$ then the lemma holds trivially by setting $U = S \Delta T.$ In what follows, we assume $d(S,T) \geq r.$
Wlog we can assume that $S = \set*{1, \dots, t} \cup C$ and $T = \set*{t+1, \dots, 2t} \cup C$ with $C = \set*{2t+1, \dots, t+k}$ and $t = d(S,T).$
Consider distribution $\mu' = \lambda \ast \mu$ with $\lambda_i = \begin{cases} 1 &\text{ if } 1\leq i \leq t\\
(\mu(S)/\mu(T))^{t} &\text{ if } t+1\leq i \leq 2t \\
\infty &\text{ if } 2t+ 1 \leq i \leq t+k\\
0 &\text{ else }
\end{cases}.$
Note that $\mu'$ is supported on $W\in \binom{[n]}{k}$ where $(S \cap T) = C \subseteq W \subseteq (S \cup T).$
Let $\Phi$ be the conductance of the $k\leftrightarrow (k-r)$-down-up walk on $\mu'$, then $\Phi \geq \Omega(k^{-c}).$ On the other hand, since $\mu'(S) = \mu'(T) = \mu(S)\leq \frac{\sum_{W}\mu'(W)} {2},$ we have that by definition of $\Phi$
\[ \Phi=\min_{\mu'(\mathcal{S})\leq \mu'(\Omega)/2 } \frac{Q (\mathcal{S}, \Omega \setminus \mathcal{S} ) }{\mu'(\mathcal{S}) } \leq \frac{ Q(\set*{S}, \Omega \setminus \set*{S}) }{\mu'(S) }\]
where we can rewrite $Q(\set*{S}, \Omega \setminus \set*{S})$ as
\[Q(\set*{S}, \Omega \setminus \set*{S}) = \mu'(S) \frac{1}{\binom{k}{r}} \sum_{U_1 \in\binom{S}{r}} \sum_{\substack{W \supseteq S \setminus U_1\\W \in \supp(\mu')\setminus \set*{S}}} \frac{\mu'(W)}{\mu'(S \setminus U_1)}\
where $\mu'(S \setminus U_1) = \sum_{W \in \binom{[n]}{k}, W \supseteq S \setminus U_1} \mu'(W).$
Note that \[ \set*{W \in \supp(\mu') \setminus \set*{S} \given W \supseteq S \setminus U_1 } \subseteq \set*{(S \setminus U_1 ) \cup U_2 \given U_2 \in \binom{T \cup U_1}{r} \setminus \set*{U_1}}\]
thus \[\bigcup_{U_1 \in \binom{S}{r}} \set*{W \in \supp(\mu') \setminus \set*{S} \given W \supseteq S \setminus U_1 } \subseteq \set*{S \Delta U \given U \in \bigcup_{s=1}^r \cE^s(S, T)} .\] Moreover, $\abs{\binom{S}{r}} = \binom{k}{r}$ and for each $U_1 \in\binom{S}{r},$ the cardinality of $ \set*{W \in \supp(\mu') \setminus \set*{S} \given W \supseteq S \setminus U_1 } $ is at most $ \leq \binom{k+r}{r}-1\leq k^r.$
Hence, there must exist $r \in [s]$ and $U \in \cE^s(S, T)$ such that
\[\frac{\mu'(S\Delta U)}{\mu'(S\setminus U)} \geq \frac{1}{k^r} \frac{ Q(\set*{S}, \Omega \setminus \set*{S}) }{\mu'(S) } \geq \Omega(k^{-(r+c)}) . \]
Thus
\[\mu(S) = \mu'(S) \leq \mu'(S \setminus U) \leq O(k^{r + c}) \mu'(S \Delta U) = O(k^{r + c}) \mu(S \Delta U)(\frac{\mu(S)}{\mu(T)})^{s/t}. \]
\begin{comment}
WLOG assume $S \cap T =\emptyset, S \cup T = [n].$ The other cases can be reduced to this via scaling etc.
We can further assume $ S =\set*{1, \dots, k}, T = \set*{k+1, \dots, 2k}$ and that $k \geq \ell,$ otherwise we are done. Let $\lambda_{k+1} = \dots = \lambda_{2k} = (\mu(S)/\mu(T))^{1/k}.$ Consider distribution $\mu'$ generated by
\[g(z_1, \dots, z_k, \lambda_{k+1} z_{k+1}, \dots, \lambda_{2k} z_{2k}) =\sum_{W} \mu(W) (\frac{\mu(S)}{\mu(T)})^{\abs{W \setminus S}/k} z^W =\mu(S) z^S + \dots + \mu(S) z^T\]
The $k\leftrightarrow (k-\ell)$ walk on $\mu'$ still mixes fast. Let $\Phi$ be the conductance of this walk, then $\Phi \geq \frac{1}{k^{O(\ell)}}.$ On the other hand, since $\mu'(S) = \mu'(T) = \mu(S)\leq \frac{\sum_{W}\mu'(W)} {2},$ we have that by definition of $\Phi$
\[ \Phi=\min_{\mu'(\mathcal{S})\leq \mu'(\Omega)/2 } \frac{Q (\mathcal{S}, \Omega \setminus \mathcal{S} ) }{\mu'(\mathcal{S}) } \leq \frac{ Q(\set*{S}, \Omega \setminus \set*{S}) }{\mu'(S) }\]
where $Q(\set*{S}, \Omega \setminus \set*{S}) = \mu'(S)\sum_{W\in \cN_{\ell}(S)\setminus \set*{S} } \frac{\mu'(W)}{\mu'(S\cap W) }.$ Since $ \abs{\cN(\ell)(S) } = O(k^{\ell}), $ there must exists $W \in \cN_{\ell}(S)\setminus \set*{S} $ such that
\[\frac{1}{k^{O(\ell)}} \leq \frac{\mu'(W)}{\mu'(S\cap W) } \]
thus \[\mu(S) = \mu'(S) \leq \mu'(S \cap W) \leq k^{O(\ell)} \mu'(W) = k^{O(\ell)} \mu(W) (\frac{\mu(S)}{\mu(T)})^{\abs{W\setminus S}/k} \leq k^{O(\ell)} \mu(S) (\frac{\mu(S)}{\mu(T)})^{\abs{W\setminus S}/k}\]
where the last inequality follows from the fact that $S$ is local max. Since $W$ is not $S$, we must have that $\abs{W\setminus S} \geq 1.$ Simplifying the above equation, we get
\[\mu(T) \leq (k^{O(\ell)})^{k/\abs{W \setminus S}} \mu(S) \leq k^{O(k\ell)} \mu(S)\]
\end{comment}
\end{proof}
\begin{proof}[Proof of \cref{cor:local search guarantee FLC}]
Apply \cref{lem:approx exchange FLC} for $S$ being a $(r,
\zeta)$-local max and $T: = \arg\max \mu(W).$ Let $t = d(S,T).$ For some $s \in [r]$ and $U \in \cE^s(S,T)$
\[\mu(S) \leq O(k^{r+c}) \mu(S \Delta U) (\frac{\mu(S)}{\mu(T)})^{s/t} \leq O(k^{r+c}/\zeta) \mu(S) (\frac{\mu(S)}{\mu(T)})^{s/t} \]
where the inequality follows from definition of $(r,\zeta)$-local max. Divide both sides by $\mu(S) > 0,$ we get
\[\mu(T) \leq O(k^{r+c}/\zeta)^{t/s} \mu(S) \leq O(k^{r+c}/\zeta)^k \mu(S). \]
where we use the fact that $t/s \leq k.$
\end{proof}
\section{Improved Local Search for Sector-Stable Distributions}
By \cref{lem:sector-stability-to-log-concavity,lem:approx exchange FLC}, for any $\alpha$-sector-stable distribution $\mu:\binom{[n]}{k},$ \cref{alg:localsearch} with $r = \lceil \frac{2}{\alpha} \rceil$ finds a $k^{O(k/\alpha)}$-approximation of $\OPT.$ In this section, we show how to improve the local search radius $r$ to $\lceil \frac{1}{\alpha} \rceil$ for $\alpha \in [1/2,1].$ As an application, we prove \cref{thm:nonsym DPP main}.
When $\alpha =1,$ the distribution $\mu$ is real stable, thus log-concave, and \cref{cor:local search guarantee FLC} already shows $\LS_1$ gives a $k^{O(k)}$-approximation for MAP inference. Clearly, for $\alpha \in [1/2,1),$ any $\alpha$-sector-stable $\mu$ is also $1/2$-sector-stable, and $\lceil 1/\alpha \rceil = 2,$ so we only need to consider the case $\alpha =1/2.$
\begin{definition}[$(r, \beta)$-approximate exchange]\label{def:approx_exchange}
For $r\in \N_{\geq 1}$ and $\beta > 0$, we say $\mu: \binom{[n]}{k} \to \R_{\geq 0} $ satisfies $(r, \beta)$-approximate exchange if for any $S, T \in \binom{[n]}{k}$
\[\mu(S) \mu(T) \leq \max_{i=1}^r \set*{\beta^i M^i(S \to T) M^i(T \to S) }\]
where $M^i (S \to T) : = \max_{U \in \cE^i (S, T)}\mu(S \Delta U)$.
\end{definition}
\begin{theorem}\label{thm:approximate_exchange}
Suppose $\mu: \binom{[n]}{k} \to \R_{\geq 0}$ is $1/2$-sector stable. For any $S, T \in \binom{[n]}{k}$
\begin{align}\label{ineq:exchange}
\mu(S) \mu(T) &\leq \max_{i=1}^2 \set*{\parens*{\sum_{U \in \cE^i(S,T)} \mu(S \Delta U)}\parens*{\sum_{U \in \cE^i(S,T)} \mu(T \Delta U)}}\nonumber \\
&\leq \max_{i=1}^2 \set*{k^{4i} M^i(S \to T) M^i(T \to S) }.
\end{align}
Consequently, $\mu$ satisfies $(2, k^4)$-approximate exchange.
\end{theorem}
We prove the approximate exchange property by relying on the following theorem.
\begin{theorem}[\cite{asner70}] \label{thm:hurwitz minor}
Consider a univariate $1$-sector-stable (Hurwitz-stable) polynomial $f(z) = \sum_{i=0}^n a_i z^i$ with $a_i \geq 0 \forall i.$ Its Hurwitz matrix $H = (h_{ij}) \in \R^{n\times n}$ is defined by $h_{ij} = a_{2j-i}$ when $0 \leq 2j-i \leq n$, otherwise $h_{ij}=0$. $H$ is totally nonnegative, in the sense that all its minors are nonnegative.
\end{theorem}
As an immediate consequence, we obtain the following lemma about coefficients of univariate Hurwitz stable polynomial.
\begin{lemma} \label{lemma:hurwitzCoeff}
If $f(z) = a_n z^n + \cdots + a_1 z + a_0$ with $a_i\geq 0\forall i$ is $1$-sector stable,then $a_n a_0 \leq \max \set*{a_1 a_{n-1}, a_2 a_{n-2}}$
\end{lemma}
\begin{proof}
If $n\leq 2$ then the claim is trivially true. Below, we assume $n\geq 3.$
We consider two cases, when $n$ is odd and when $n$ is even.
Suppose $n = 2t -1$ for $t\in \N.$
By \cref{thm:hurwitz minor}, all minors of $H$ are non-negative, hence
\begin{align*}
\det\begin{bmatrix}
h_{1,1} & h_{1, t}\\
h_{2,1} & h_{2, t}
\end{bmatrix} &= \det \begin{bmatrix}
a_1 & a_{2t-1}\\
a_0 & a_{2t-2}
\end{bmatrix} = a_1 a_{2t-2} - a_0 a_{2t-1} \\
&= a_1 a_{n-1} - a_0 a_n \geq 0.
\end{align*}
Suppose $n = 2t$ for $t\in \N.$ Again, \cref{thm:hurwitz minor} implies \begin{align*}
\det\begin{bmatrix}
h_{2,2} & h_{2, t+1}\\
h_{4,2} & h_{4, t+1}
\end{bmatrix} &= \det \begin{bmatrix}
a_2 & a_{2t}\\
a_0 & a_{2t-2}
\end{bmatrix} = a_2 a_{2t-2} - a_0 a_{2t} \\
&= a_2 a_{n-2} -a_0 a_n \geq 0.\qedhere
\end{align*}
\end{proof}
\cref{lemma:hurwitzCoeff}, in turn implies the following fact about coefficients of $1/2$-sector-stable univariate polynomials that only have even-degree terms.
\begin{corollary} \label{cor:S[1/2]Coeff}
If $f(z) = \sum_{i=0}^t a_{2i}z^{2i}$ with $a_{2i}\geq 0 \forall i$ is $1/2$-sector stable, then $a_0 a_{2t} \leq \max \set*{a_2 a_{2t-2}, a_4 a_{2t-4}} $
\end{corollary}
\begin{proof}
Let $g(z) = f(z^{1/2}) = \sum_{i=0}^t a_{2i} z^i$ then $g(z)$ is $1$-sector stable, and the claim follows from \cref{lemma:hurwitzCoeff}.
\end{proof}
\begin{proof}[Proof of \cref{thm:approximate_exchange}]
Let $f (z_1, \cdots, z_n) = \sum_{W \in \binom{[n]}{k}}\mu(W) z_W$ be the generating polynomial of $\mu.$ We deal with the case where $S\cap T = \emptyset$ and $[n] = S \cup T.$ Other cases can be reduced to this scenario by setting $z_i$ to $0$ for $i\not \in S \cup T$, and taking derivative(s) of $f$ with respect to $i\in S\cap T.$ Recall that setting variables to $0$ and taking derivative(s) preserve $1/2$-sector-stability and homogeneity of polynomials (see \cref{prop: ssProperties}).
W.l.o.g., assume $S = [t]$ and $T = \set*{t+1, \cdots, 2t}.$ We can rewrite $f$ as
$f(z_1, \cdots, z_{2t}) = \sum_{W \in \binom{[2t]}{t}} \mu(W) z_W$.
In $f$, set $z_i = z$ if $i\in S$ and $z_i = z^{-1}$ if $i\in T.$ We obtain a single variate $1/2$-stable polynomial \[\tilde{f}(z) = z^t f (z, \cdots, z, z^{-1}, \cdots, z^{-1}) = \sum_{i=0}^{2t} b_i z^i = \sum_{i=0}^t b_{2i} z^{2i} \]
Note that a term $\mu(W) z^W$ contribute to $b_i$ if and only if $i = \abs{W \cap S} + (t - \abs{W\cap T}) = 2 \abs{W\cap S} .$ In particular, $b_{2i+1} = 0$ for all $i\in \N$ and
\begin{align*}
b_{2i} &= \sum_{W: \abs{W \cap S} = i}\mu(W) = \sum_{U \in \cE^i(S,T)} \mu(T \Delta U)
\\
&= \sum_{U \in \cE^{t-i}(S,T)} \mu(S \Delta U).
\end{align*}
In particular, $b_{2t} = \mu(S)$ and $b_0 = \mu(T).$
The first line of \cref{ineq:exchange} follows by applying \cref{cor:S[1/2]Coeff} to $\tilde{f}$, and the second line follows by observing that $ \sum_{U \in \cE^i(S,T)} \mu(T \Delta U) \leq \binom{\abs{S \Delta T} /2}{i}^2 \max_{U \in \cE^i(S,T)} \mu(T \Delta U) $.
\vskip -0.25in
\end{proof}
\begin{lemma}
\label{lem:local-to-global_gen}
Suppose $\mu: \binom{[n]}{k} \to \R_{\geq 0}$ satisfies $(r, \beta)$-exchange and $S$ is a $(r,\zeta)$ local maximum with $\alpha \leq 1$ and $\mu(S) > 0$. Then $S$ is a $(\beta/\zeta)^{ k}$-approximate global optimum:
\[ (\beta/\zeta)^{k}\mu(S)\geq \max_{T\in \binom{[n]}{k} } \mu(T).\]
\end{lemma}
\begin{proof}
Let $S\in \binom{[n]}{k}$ be a $(r,\alpha)$-local maximum with $\mu(S) > 0$ and let $S^* : =\arg\max_{T \in \binom{[n]}{k}} \mu(T)$. We first prove the following claim.
\begin{claim} \label{clm:reduceDistance}
For any $T \in \binom{[n]}{k}$ where $T \neq S$,
there exists $i\in [r]$ and $W \in \binom{[n]}{k}$ such that $d(S, W) = d(S, T) -i$ and
$
\mu(T) \leq \frac{\beta^i}{\zeta} \cdot \mu(W).
$
\end{claim}
\begin{proof}[Proof of \cref{clm:reduceDistance}]
By \cref{def:approx_exchange}, for some $i\in [r]$, there exists $U_1, U_2 \in \cE^i(S, T)$ such that
\begin{align*}
\mu(S) \mu(T) \leq \beta^{i} \mu(S \Delta U_1) \mu(T\Delta U_2) \leq \beta^i \frac{\mu(S)}{\zeta} \cdot \mu(T \Delta U_2)
\end{align*}
where the last inequality follows from the definition of $(r,\zeta)$-local maximum.
Note that $d(S, T\Delta U_2) = d(S, T) - i.$ Setting $W = T\Delta U_2 $ and dividing both sides by $\mu(S)> 0$ gives the desired inequality.
\end{proof}
Note that initially $d(S, S^*) \leq k$. We can iteratively apply \cref{clm:reduceDistance} for up to $k$ times to obtain the desired inequality. Indeed, let $T_0 = S^*$, and for $j\geq 1$ let $i_j\in [r]$ and $T_j\in \binom{[n]}{k}$ be such that $\mu(T_{j-1}) \leq\frac{\beta^{i_j}}{\zeta}\cdot T_j $ and $d(S, T_j) = d(S, T_{j-1}) - i_j.$ \cref{clm:reduceDistance} guarantees the existence of such $i_j$ and $T_j,$ as long as $T_{j-1} \neq S.$ Let $s$ be the minimum index such that $d(S,T_s) = 0.$ Note that $s \leq k$ and $T_s = S.$ We have
\begin{align*}
\mu(S^*) = \mu(T_0) \leq \frac{\beta^{i_1}}{\zeta}\cdot T_1 &\leq \frac{\beta^{i_1}}{\zeta}\cdot \frac{\beta^{i_2}}{\zeta}\cdot T_2 \leq \cdots\\ \leq
&\prod_{j=1}^s \frac{\beta^{i_j}} {\zeta} \cdot \mu(T_s) \leq \frac{\beta^k}{\zeta^k} \mu(S)
\end{align*}
where the last inequality follows from the facts that $\sum_{j=1}^s i_j = d(S,T_0) - d(S,T_s) \leq k$ and $(\frac{1}{\zeta})^s \leq (\frac{1}{\zeta})^k.$
\end{proof}
Now, we are ready to prove \cref{thm:nonsym DPP main}.
\begin{proof}[Proof of \cref{thm:nonsym DPP main}]
We let $\mu(S) = \det(L_S)$ and run the two stage algorithm in \cref{sec:implement} with $r=2.$ The approximation guarantee is a direct consequence of \cref{lem:p0Constrained,thm:approximate_exchange,lem:local-to-global_gen}.
Suppose we are given access to the entries of $L.$ Each iteration of \cref{alg:localsearch} clearly runs in $O(n^2 k^5)$ time, since $\cN_2(S)$ has at most $O(k^2 n^2)$ elements and computing the determinant of $k\times k$ matrices costs $O(k^3)$ time. The cost of $\LS_2$ can be reduced to $O(n^2 k^4)$ time using Schur complements to compute all $\det(L_{Y \cup D})$ for each fixed $Y$ and all $D$ of size $\leq r $ in $O(k^3 + n^2 k^2)$ time.
\citep[see \cref{eq:condition kernel} or ][for example]{gartrell2020scalable}.
If we are only given $B, C$, then each of these submatrices and their determinant can be computed in $O(d^2)$ time, so that each iteration takes $O(n^2 d^2 k^2)$ time.
Now, we bound the runtime of \cref{alg:greedy}.
To implement each iteration of \cref{alg:greedy}, we need to compute $\mu(Y) = \sum_{S \in\binom{[n]}{k}: S\supseteq Y} \det(L_Y) $ , which is the coefficient of $\lambda^{n-k}$ in $g(\lambda) = \det(L + \lambda\cdot \diag{\mathds{1}_{\tilde{Y}} })$ where $\tilde{Y} = [n] \setminus Y.$
There are several ways to compute $\mu(Y).$
To compute the coefficients of polynomial $g(\lambda)$ of degree $\leq n$, we can evaluate $g$ at $n+1$ distinct points $\lambda$ and use polynomial interpolation, i.e., solve a linear system of equations involving the the Vandermonde matrix.
A more efficient way, which costs $O(n^3)$ per computation of $\mu(T)$, for a total runtime of $O(n^4 k)$, is as follow:
\begin{enumerate}[label = (\roman*)]
\item Let $D=\diag{\mathds{1}_{\tilde{Y}} }. $ We use the QZ decomposition algorithm \citep[Section 7.7, p. 313]{QZDecomp} to compute unitary matrices $Q$, $Z$ such that
\[ L = Q \tilde{A} Z^*, D = Q \tilde{D} Z^* \]
where $\tilde{A}, \tilde{D}$ are both upper triangular.
Note that $\deg(g) \leq n - \abs{T}.$
Compute the roots of $g(\lambda) =\det(L + \lambda D),$ which are exactly the generalized eigenvalues $ \lambda_1, \dots,\lambda_{\deg(g)} $ defined by $\lambda_i = \frac{\tilde{A}_{i,i} }{ \tilde{B_{i,i}}}$ where we may assume w.l.o.g. that $\tilde{D}_{i,i} \neq 0$ for $i = 1, \dots, \deg(g)$, and is zero otherwise. Let $c : = \prod_{i \in [n]: B_{i,i} \neq 0 } \tilde{D}_{i,i} \prod_{i \in [n]: D_{i,i} = 0 } \tilde{A}_{i,i} $. Then
\[g (\lambda) = c \prod_{i \in [\deg(g)]} (\lambda - \lambda_i) \]
\item We then compute the $(k-n+n')^{th}$-symmetric polynomial of $ \lambda_1, \dots,\lambda_{n - \abs{T}} $ where
$e_t = \sum_{W \in\binom{[k - \abs{T}] }{ t} } \prod_{j\in W} \lambda_j$ using the recursion \cite{KT12}
\[ te_t = e_{t-1}p_1 - e_{t-2}p_2 + e_{t-3}p_3 -\dots \pm p_k \]
with $p_t = \sum \lambda_j^t,$ and output $ \mu(Y) = \abs{c e_{k-(n - \deg(g) )}} $.
\end{enumerate}
Given the low-rank decomposition $L = B C B^{\intercal},$ we can further optimize by reducing the cost of step (i) to $O(nd^2)$. Then the total runtime will be $O(n^2 k d^2 ).$
Let $L^Y$ be the kernel of $\mathbb{P}_{L}$ conditioned on the inclusion of items in $Y.$
The eigenvalues of $L^Y$ are exactly the roots of $g(\lambda).$ By \cref{eq:condition kernel}, $L^Y $ can be rewritten as product of two matrices of rank $\leq d$, thus the nonzero eigenvalues of $L^Y$ can be computed in $O(d^3)$ time. Indeed, let $D_Y : = B_{Y}^{\intercal} ( B_Y C B_Y^{\intercal})^{-1} B_{Y}$ then $L^Y = B_{\tilde{Y}} ( C- C D_Y C)B_{\tilde{Y}}^{\intercal} $ and $\rank(D_Y)\leq k$ and $D_Y$ can be computed in $O(k d^2)$ time (see \cref{eq:condition kernel}).
The matrix $F_Y:= \left((C - C D_Y C) B_{\tilde{Y}}^{\intercal}\right) B_{\tilde{Y}} $ has the same characteristic polynomial and nonzero eigenvalues as $L^Y.$ Clearly, $\rank(F_Y ) \leq \rank(B) \leq d$, so $F_Y$ and its eigenvalues can be computed in $O(nd^2)$ time.
\end{proof}
\section{Composable Core-Sets via Local Search}
Here we prove that local search yields composable core-sets for distributions that satisfy a strong form of exchange.
\begin{definition}[$\beta$-strong approximate basis exchange] \label{def:strong basis exchange}
For $\beta \geq 1$, we say $\mu: \binom{[n]}{k} \to \R_{\geq 0}$ satisfy $\beta$-strong approximate basis exchange if, for $ S\in \binom{[n]}{k}$ and $j\not \in S$,
\begin{equation}\label{ineq:rsExchange2}
\mu(S)\mu(T)\leq \beta \mu(S-i+j)\mu(T+i-j)
\end{equation}
\end{definition}
\begin{lemma}\label{lem:core-set}
Suppose $\mu:\binom{[n]}{k} \to \R_{\geq 0}$ satisfies $\beta$-strong approximate basis exchange,
then the Local Search algorithm achieves an $O(\beta)^{k}$-composable core-set
of size $k$ for the MAP-inference problem for $\mu.$
\end{lemma}
\begin{proof
Consider a partition $P_1\cup \dots \cup P_r$ of $[n]$, and let $C_i \in \binom{P_i}{k}$ be a $\alpha$-local optimum in $ P_i$ with $\mu(C_i) > 0.$ We want to show
\[(\frac{\beta}{\alpha})^k \OPT(C) \geq \OPT(\bigcup_{i=1}^r P_i) \]
where $C := \bigcup_{i=1}^r C_i.$
Let $S^*$ be such that $\mu(S^*)= \OPT(\bigcup_{i=1}^r P_i).$ We need the following fact.
\begin{claim}\label{claim:ls core-set}
For any $W \in \binom{[n]}{k}$ with non-empty $(W\setminus C_i),$ there exists $W' \in \binom{[n]}{k}$ s.t. $\abs{W' \setminus C_i} = \abs{W \setminus C_i} -1$ and $ \beta \mu(W')\geq\mu(w). $
\end{claim}
\begin{proof}[Proof of \cref{claim:ls core-set}]
Take an arbitrary $j \in (W \cap P_i) \setminus C_i.$ There exists $e \in C_i \setminus W$ s.t.
\begin{align*}
\mu(C_i) \mu(W) \leq \beta \mu(C_i - e +j) \mu(W +e -j) \leq \frac{\beta}{\alpha} \mu(C_i) \mu(W+e-j
\end{align*}
Setting $W' = W+e-j$ and dividing both sides by $\mu(C_i)> 0$ gives the desired inequality, since $\abs{W'\setminus C_i} = \abs{W\setminus C_i} - 1.$
\end{proof}
We can iteratively apply \cref{claim:ls core-set} for up to $k$ times to obtain the desired inequality. Indeed, let $W_0 : = S^*$, and for $j\geq 1$ let $i_j\in [r]$ and $W_j\in \binom{[n]}{k}$ be such that $\mu(W_{j-1}) \leq\frac{\beta}{\alpha}\cdot \mu(W_j) $ and $\abs{W_j\setminus C_{i_j} } = \abs{W_{j-1}\setminus C_{i_j}} - 1.$ \cref{claim:ls core-set} guarantees the existence of such $i_j$ and $W_j,$ as long as $W_{j-1} \not\subseteq C.$ Let $s$ be the minimum index such that $W_s \subseteq C.$ Note that $s \leq k$ and $\mu(W_s) \leq \OPT(C).$ We have
\begin{align*}
\OPT([n]) = \mu(W_0) \leq \frac{\beta}{\alpha}\cdot \mu(W_1) &\leq (\frac{\beta}{\alpha})^2\cdot \mu(W_2) \leq \cdots\\ \leq
&(\frac{\beta}{\alpha})^s \mu(W_s) \leq (\frac{\beta}{\alpha})^k \OPT(C)
\end{align*}
\end{proof}
\cref{thm:core-set} is a direct consequence of \cref{lem:core-set} and the fact that strongly Rayleigh (log concave resp.) distributions satisfy $k^{O(k)}$-strong approximate basis exchange ($2^{O(k^2)}$-strong approximate basis exchange resp.) \cite{ALOVV21}.
\PrintBibliography
\end{document}
|
\section{Introduction}
Numerous emerging smart applications (e.g. IoT, wearables, drones, etc.) demand on-chip continuous learning, compelling the development of application-specific memories and architectures. More often these applications demand the implementation of learning algorithms for large network models in an energy-efficient manner. Conventional digital memory solutions based on SRAM or DRAM cannot address the required density/energy requirements due to large area and restrictive off-chip memory access costs. High-end expensive graphical processing units (GPUs) have been a default choice to perform DNN training. The energy and time requirement of training the state-of-the-art DNN architectures on GPUs is high \cite{gpu_cost}. This necessitates the development of more energy-/area-efficient custom hardware accelerators for performing deep learning training workloads. A few ASIC processors have been recently reported for DNN training \cite{asic1,asic2,Yin_SSCL2020}, but based on conventional SRAM for on-chip storage, which requires a large amount of memory access with associated density and leakage power constraints.
The multi-level storage offered by non-volatile memories such as phase-change memory (PCM) \cite{PCM}, resistive RAM (RRAM) \cite{RRAM} when combined with in-memory computing forms a basis of analog DNN training accelerators \cite{Y2018sebastianJAP,Y2017burrAPX,Joshi2020,isaac,NSI2020,prime,pipe_layer}. In-memory computing with PCM devices has shown promise in training deep neural networks (DNNs) on several cognitive tasks \cite{Y2017burrAPX,NSI2020,rpu}. However, in these solutions weight gradients are stored on off-chip storage which again inherits the inherent density/power constraints.
In \cite{Luo_ToC_2020} a hybrid precision synapse for DNN training is proposed. In our implementation, unlike in \cite{Luo_ToC_2020}, the higher significant bits are programmed only if there is an overflow event on the lower significant bits. Furthermore, our implementation ensures that the conductance decay due to PCM device drift (similar to capacitor leakage in \cite{Luo_ToC_2020}) does not affect the network training accuracy.
Here, we propose a hybrid in-memory computing (HIC) architecture for training of DNNs on hardware accelerators and memory-efficient inference. Specifically, we propose to map a higher significant bits of the weight values using PCM devices with multi-level storage capability and the lower significant bits using PCM devices with binary-level storage capability. In the following sections, we show that this scheme can outperform baseline network trained, in floating-point $32$-bit (FP$32$) precision, by leveraging appropriate network width multiplier and requires about $50\,$\% less inference model size to achieve similar accuracy in comparison to the baseline.
\section{DNN training using Hybrid In-memory Computing} \label{sec:hic}
In this section, we will describe the weight representation strategy on two memory arrays used in the HIC architecture. We also discuss, a set of hardware aware operations that are an essential basis of the HIC architecture.
\subsection{Weight representation strategy}
An overview of the weight representation strategy in the HIC architecture is shown in Fig. \ref{FIG:HIC_ARCH}. In the HIC architecture the MSB part of the weight values is stored on an array of multi-level PCM cells denoted as MSB array. The LSB part of the weight values is stored on a memory array of binary-valued PCM devices denoted as LSB array. This hybrid design is motivated by the fact that only MSBs of weights are needed for forward/backward propagation of DNN training that targets low-precision inference, while typically weight updates are small values that mostly modify only LSBs of weights. The MSB values of weights are programmed and read using statistically accurate PCM models proposed in \cite{NSIJAP}. The LSB part of the weights is mapped using their binary representation on multiple binary-level PCM devices. Notably, the write operation on a memory location in the LSB array is performed by simply reading and flipping the binary state of the appropriate PCM device. This results in reduced write operation cost on the LSB array but also implies that some PCM devices shall drift more compared to their neighbors in the same memory location (as device drift is based on the last programming time). However, we will empirically show that this does not adversely affect the network training accuracy even for extended periods of time (year).
The MSB array is a crossbar array formed by a differential pair of multi-level PCM devices at each cross point. We perform our experiments using the PCM models proposed for DNN training in \cite{NSIJAP}. In \cite{NSIJAP}, PCM models were developed by gathering the read/write statistics from 10K PCM devices. The proposed PCM model has a strong statistical match with the obtained data. Hence, the use of this PCM model for this study is an accurate choice. The proposed PCM model consists of 4 different non-ideal components, (1) stochastic write, (2) stochastic read, (3) temporal variation in device conductance, and (4) nonlinearity of the programming curve. For the LSB array, we use a binary-level PCM model with stochastic write of high conductance state, and the read operation consists of the conductance drift coupled with stochastic read as proposed in \cite{NSIJAP}. For a binary PCM device, stochastic write is simulated by adding a zero mean fixed standard deviation Gaussian noise to the expected high state conductance value. A multi-level PCM can also be used as binary-level storage by utilizing the lowest conductance state and desired high conductance state.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{FIGURES/1_HIC_ARCH/HIC_ARCH}
\caption{Hybrid in-memory computing (HIC) architecture for DNN training - the MSB part of the synaptic weight is stored on a multi-level PCM cell that offers the equivalent of 4-bit precision while the LSB part is stored on a 7-bit memory formed by seven binary PCM devices. This choice of precision in MSB and LSB parts is found to be optimal for DNNs studied in this paper.}
\label{FIG:HIC_ARCH}
\end{figure}
\subsection{Implementation of hybrid in-memory computing}
The HIC architecture performs vector-matrix multiplication (VMM) operation using an analog crossbar array of PCM devices. We assume that all other operations required in DNN training are performed in digital CMOS circuits. A digital to analog converter (DAC) is required to apply input voltage to read/program the analog crossbar array, and an analog to digital converter (ADC) is required to read the crossbar array output. A convolution operation is essentially a matrix-matrix multiplication \cite{cnnXbarMapping}. Hence it can be implemented on a crossbar as a vector-matrix multiplication. A recurrent layer usually performs vector-matrix multiplication followed by some elementwise non-linear activation functions \cite{lstm}. Hence, the operation layout, shown in Fig. \ref{FIG:OP_LAYOUT}, is valid for three types of commonly used DNN layers, namely fully-connected, convolution, and recurrent layer.
Fig. \ref{FIG:OP_LAYOUT} shows a typical operation layout of a DNN layer in the HIC-based training. A transposable crossbar array performs VMM required in forward and backpropagation phases of DNN training. A digital to analog converter (DAC) applies an input (activation or error gradient) to the crossbar array. An analog to digital converter reads the output current of the crossbar array. A network-specific normalization function such as batch normalization \cite{batchnorm} or group normalization \cite{groupnorm} is computed on ADC output. An activation function such as ReLU, Sigmoid, or Tanh is computed on the normalization layer output. The gradients are backpropagated by applying voltages proportional to the error gradients on the columns of the transposable crossbar array. During the weight update phase, an outer product is computed on the input $X$ and output error gradients $\Delta Y_{A}$ to obtain weight gradients $\Delta W$. The weight gradients are quantized and used to update the LSB array. The values in the LSB array are updated by simply flipping the binary states of the devices if required. In case there is an overflow on the LSB array, the MSB array is programmed with a corresponding update value. There are no other specific programming events on the MSB array.
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{FIGURES/2_OPERATION_LAYOUT/OPERATION_LAYOUT}
\caption{The overview of DNN layer operation during HIC-based training - the MSB array is used to perform vector-matrix multiplication operations involved in the forward and backpropagation phases of DNN training, while the weight updates are quantized and accumulated in the LSB array. An overflow event in the LSB array triggers the programming of corresponding weight elements in the MSB array.}
\label{FIG:OP_LAYOUT}
\end{figure}
\section{Numerical validation} \label{sec:numerical_validation}
In this section, we discuss the experimental setup used for our simulations and the results from four different studies.
\subsection{Experimental setup and network details}
We have chosen ResNet-$32$ network for all our evaluations as they are representative of large networks used in the machine learning research \cite{resnet}. The ResNet-$32$ network has a total of $33$ convolution layers and $1$ fully-connected layer for classification. It has about $470\,$K trainable parameters and all the weights and updates are stored on PCM-based memory arrays. All the $34$ layers of the ResNet-$32$ network are trained using HIC architecture. The network is trained to classify the images from the CIFAR-$10$ dataset \cite{cifar}.
The scheme used to program PCM devices on the MSB array can only increment the device conductance. Application of several programming pulses to the devices in a differential pair can result in saturation of their conductance level. After every $10$ batches of training, we perform a refresh operation on the PCM devices in the MSB array to avoid saturation of device conductance \cite{Boybat2018}. A differential pair of PCM devices used in the MSB array offers an equivalent precision of approximately $4$-bits \cite{pcm4-bit}. In the LSB array, we use seven PCM devices for $7$-bit signed fixed-point representation. All the DACs and ADCs have $8$-bit precision\textcolor{red}, as they have been reported to be design points with good trade-off between precision and energy consumption \cite{Rekhi2019}. We use the same hyperparameter setting for the baseline network and preprocessing of the CIFAR-10 images as given in \cite{resnet}. The hyperparameter setting in HIC implementation is the same as \cite{resnet} except that the learning rate is 0.05 with a decay factor of 0.45 and a batch size of 100. All our simulations are performed using TensorFlow \cite{TF}.
\subsection{Effect of individual non-idealities}
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{FIGURES/3_PCM_SIMULATIONS/MODEL_ABLATION_STUDY}
\caption{The impact of different non-ideal aspects of the PCM on HIC-based training accuracy. PCM device non-linearity, as well as stochasticity associated with write and read operations results in a notable drop in the training accuracy while the temporal drift in the device conductance results in accuracy improvement as it has the same effect on DNN training as that of the weight decay regularization. The results are based on the average of five distinct training runs.}
\label{FIG:ABLATION}
\end{figure}
Our simulations are based on the PCM model proposed in \cite{NSIJAP}. This PCM model is composed of four non-ideal components, (1) stochastic write, (2) stochastic read, (3) temporal variation in the device conductance, and (4) nonlinear programming curve. We study the effect of each non-ideal component on the network training accuracy by performing the ablation of the PCM model. As shown in Fig. \ref{FIG:ABLATION}, light blue bars indicate network trained with linear PCM model and at most one other non-ideal component (see the text on bars). The colorless bar indicates a network trained with the nonlinear PCM model and at most two other non-ideal components. The right-most colored bar (denoted as "Full-model") indicates network trained with a PCM model including all the non-ideal components.
We observe that the nonlinearity of the programming curve causes a notable drop in the network training accuracy compared to the linear PCM model. This is attributed to the fact that the expected conductance increment in the PCM device is an inverse function of the number of applied programming pulses \cite{NSIJAP}. The stochastic read and write of the PCM model have a stronger negative effect on the network training accuracy due to the the large amount of write noise present in the PCM programming process. Interestingly, the implementation that incorporates device conductance drift achieves higher accuracy in comparison to other nonideal components. This accuracy improvement can be attributed to the fact that the device drift is similar to the weight decay regularization technique commonly used in deep learning \cite{wd1,wd2}. Like weight decay regularization, PCM conductance drift causes higher decay in the weight values that do not contribute to the network training and weights that are regularly updated (or trained) are subjected to smaller drift in their conductance value.
Although the network training accuracy in the full-model case including all the non-ideal components is about 4.4\% lower than the baseline network trained in FP32 precision, note that HIC is utilizing only 9-PCM devices.
We show in the next section that HIC-based training is more memory-efficient and it achieves better accuracy compared to the baseline.
\subsection{Effect of the model size on network accuracy}
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{FIGURES/4_MODEL_SIZE_STUDY/MODEL_SIZE_STUDY}
\caption{With network width multiplication, HIC-based training of ResNet-32 network outperforms the FP32 software baseline for comparable inference model size and also shows better accuracy improvement with increase in model size. Note that HIC achieves comparable accuracy to the baseline with about $50\,$\% less inference model size. The network width multiplier (values near markers) compensates for the loss in accuracy due to non-idealities of the PCM devices. The results are based on the average of five distinct training runs.}
\label{FIG:MODEL_SIZE}
\end{figure}
Now, we discuss the network training accuracy as a function of the model size required for inference. We increase the network size by increasing the number of neurons in each layer by a desired network width multiplier (see values near markers in Fig. \ref{FIG:MODEL_SIZE}) as proposed in \cite{mobilenet}. We observe that the increase in network width compensates for the loss in training accuracy due to non-ideal PCM devices. Fig. \ref{FIG:MODEL_SIZE} shows the network training accuracy as a function of the inference model size. The inference model size is the amount of memory required to store weights during the inference phase. The HIC-based implementation requires approximately $4$-bits for storing weights and $32$-bits for the baseline. Like markers indicate the same network architecture, in other words, the same number of neurons in each layer. The model size for HIC is lower because of the low precision representation of weights.
We observe that the accuracy improvement in HIC-based training per additional neurons/model size is better compared to the baseline. Furthermore, for a comparable inference model size, HIC-based training achieves at least 1\% better accuracy than the baseline. Notably, for a comparable network accuracy, HIC-based training requires about $50\,$\% less inference model size. This suggests that HIC-based training results in memory-efficient inference computation on the hardware without compromising on accuracy compared to the baseline.
\subsection{Effect of Device Drift on Post-training Inference Accuracy}
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{FIGURES/5_ACCURACY_RETENTION/ACCURACY_RETENTION}
\caption{Post-training inference accuracy estimate of ResNet-$32$ network caused by the temporal drift in PCM devices - for a period of $10^{6}\, s$ (11.7\,days) there is no observable drop in the inference accuracy. The AdaBS compensation technique \cite{Joshi2020} is shown to be effective in compensating for weight degradation due to PCM drift, resulting in maintaining the inference accuracy close to the baseline for over a year. The plots shown are averaged over 10 distinct training runs and 10 distinct inference runs per training (total of 100 runs).}
\label{FIG:ACCURACY_RETENTION}
\end{figure}
Now, we discuss the effect of PCM conductance drift on the inference accuracy computed as a function time after network training. As shown in Fig. \ref{FIG:ACCURACY_RETENTION}, we compute the inference on a version of the HIC network that uses a width multiplier of $1.7$.
In these simulations, we initially train the network for $205$ epochs and incorporate the drift in the PCM devices which causes the weights to degrade throughout training. We perform the inference on the trained network from $100\,$s to $4\times10^{7}\,$s causing further degradation in the weight value.
In \cite{Joshi2020}, AdaBS technique was proposed to compensate for the weight decay caused by the PCM conductance drift. The AdaBS technique infrequently performs a calibration phase that recomputes the global mean and variance of all the batch normalization layers in the network. The calibration phase requires about $5$\% of the images in the training set.
The inference accuracy of HIC is unaffected for the time duration of about $10^{6}\,$s when no compensation is applied to the weights. After $10^{6}\,$s, there is more adverse impact of the drift on the inference accuracy. The AdaBS compensation technique helps to recover the degradation in the network accuracy. For a year long simulation the inference accuracy of HIC drops by just $0.12\,$\% compared to the training accuracy at $100\,$s when AdaBS compensation is applied on the weights. The drop in the inference accuracy is $9.37\,$\% when no compensation is applied and AdaBS helps to recover such a significant drop in the network accuracy. Note that using AdaBS compensation no significant gain in the inference accuracy is observed till $10^6\,$s but afterwards AdaBS plays a crucial role in maintaining the inference accuracy close to the baseline.
\subsection{Write-Erase Cycle Estimation}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{FIGURES/6_ENDURANCE_N_LATENCY/WRITE_ERASE_CYCLES_LS}
\caption{The number of write-erase cycles applied to the PCM devices on the MSB or LSB array during the training of ResNet-$32$ network is well within the reported endurance limit of PCM devices ($10^{8}$).}
\label{FIG:ENDURANCE}
\end{figure}
During our training simulations we also tracked the number of write-erase cycles that were applied to the devices on the MSB or LSB array throughout the training.
Following the definition in \cite{pcm_endurance}, we define a write-erase cycle as a sequence of at most $10$ SET pulses followed by a RESET pulse. Fig. \ref{FIG:ENDURANCE} shows the distribution of write-erase cycles applied on all devices during one full training of the ResNet-$32$ network. The number of write-erase cycles applied to a PCM device in the HIC-based training of the ResNet-$32$ network is less than $150$ for the MSB array and less than $20\,$K for the LSB array. A PCM device endurance is of the order of $10^{8}$ \cite{pcm_endurance} and the number of write-erase cycles seen by any device in HIC implementation is a small fraction of PCM endurance.
\section{Conclusion}
\label{SEC:Conclusion}
We proposed hybrid in-memory computing (HIC) architecture for memory-efficient training of deep neural networks on hardware accelerators. Based on the simulations demonstrating training of ResNet-32 network on CIFAR-10 dataset, the HIC architecture shows promising memory savings and higher training accuracy compared to the baseline network trained in floating-point 32-bit precision. Specifically, the HIC implementation outperformed baseline software accuracy by at least $1\,$\% while still using a comparable amount of inference model size by leveraging a network width multiplier. Interestingly, the HIC implementation achieved similar accuracy to that of baseline but with $50\,$\% less inference model size by leveraging a network width multiplier. With a suitable PCM device drift compensation technique, we showed that post-training inference accuracy suffers a negligible drop. Finally, the HIC-based training incurred write-erase cycles that are a small fraction of PCM endurance demonstrating the usefulness of this architecture for the memory-constrained deep learning applications such as edge computing, IoT, wearable technology.
\section*{Acknowledgment}
This work was in part supported by NSF grants 1652866
and 1715443, SRC AIHW program, and C-BRIC,
one of six centers in JUMP, a SRC program sponsored by
DARPA.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:intro}
Deployment of deep neural networks (DNNs) in inference mode is challenging for applications with limited resources such as in edge devices~\cite{lin2019toward, salehinejad2019ising}. Pruning is one of the major approaches for compressing a DNN by permanently dropping a subset of network parameters. Pruning methods are divided into unstructured and structured approaches. Unstructured pruning does not follow a specific geometry and removes any subset of the weights~\cite{anwar2017structured}. Structured pruning typically follows a geometric structure and happens at channel, kernel, and intra-kernel levels~\cite{anwar2017structured, cheng2017survey}. One of the early attempts for pruning neural networks was to use second derivative information to minimize a cost function that reduces network complexity by removing excess number of trainable parameters and further training the remaining of the network to increase inference accuracy~\cite{lecun1990optimal}. \textit{Deep Compression} is a popular pruning method which has three stages that are pruning, quantization, and Huffman coding~\cite{han2015deep}. This method works by pruning all connections with weights below a threshold followed by retraining the sparsified network.
We have proposed an Ising energy model in~\cite{salehinejad2019ising} for pruning hidden units in multi-layer perceptron (MLP) networks, \cite{salehinejad2019ising2}. In this paper, we propose IPruning, which targets pruning convolutional kernels, including all corresponding input/output connections, and hidden units based on modeling a DNN as a graph and quantifying interdependencies among trainable variables using the Ising energy model.
A DNN is modeled as a graph where the nodes represent kernels/hidden units, and edges represent the relationship between nodes. This relationship is modeled using entropy of feature maps between convolutional layers and relative entropy (Kullback–Leibler (KL) divergence) between convolutional kernels in a layer. These relationships are represented as weights in an Ising energy model, which targets dropping kernels with low activity and eliminating redundant kernels. We initiate a set of candidate pruning state vectors which correspond to different subgraphs of the original graph. The objective is to search for the state vector that minimizes the Ising energy of the graph. Each step of the optimization procedure happens within a training iteration of the network, where only the kernels identified by the best pruning state vector are trained using backpropagation. This is indeed similar to training with dropout~\cite{labach2019survey}, where the original network is partially trained. However, after a number of iterations the set of candidate state vectors can converge to a best pruning state vector, which represents the pruned network\footnote{The codes and more details of experiments setup is available at: \textit{https://github.com/sparsifai/ipruning}}.
\section{Proposed Ising Pruning Method}
The weights of a DCNN with $L$ layers are defined as the set $\bm{\Theta}=\{\Theta^{[1]},...,\Theta^{[L]}\}$, where $\Theta^{[l]}\in\bm{\Theta}$ is the set of weights in layer $l$, $\Theta_{i}^{[l]}$ is the weight kernel $i$ of size $N_{i}^{[l]}$, and $N^{[l]}$ is the number of weight kernels in the convolution layer $l$. Similarly, in a dense layer $N^{[l]}$ is the number of weights from layer $l$ to the next layer $l+1$. Generally, a feature map is constructed using the convolution operation defined as
\begin{equation}
\begin{split}
\mathbf{F}^{[l]}_{i}&= \sigma(\Theta^{[l]}_{i}\star\mathbf{F}^{[l-1]}),
\end{split}
\end{equation}
where $\mathbf{F}^{[l]}_{i}$ is the feature map $i$ in layer $l$, $\mathbf{F}^{[l-1]}$ is the set of feature maps from the previous layer, $\sigma(\cdot)$ is the activation function, and $\star$ is the convolution operation.
The key questions is ``\textit{How to detect redundant and inactive kernels in a DCNN?}''. To answer this question, we suggest to quantitatively evaluate activity and redundancy of the kernels using entropy and KL divergence as follows.
\subsection{Measuring Kernels Activity}
Feature maps are the activation values of a convolutional layer, representing activation of weight kernel for a given input.
We use feature maps of a kernel as a means of evaluating its activation. Assuming $\sigma(x)=max(0,x)$, a feature map value $f_{i,j}\in \mathbb{R}_{\geq 0}$, where $j$ is the $j^{th}$ element of $\mathbf{F}_{i}$, is generally a real number in a continuous domain. We quantify a feature map in a discrete domain $\Lambda$ by mapping the feature map values as $\mathbb{R}_{\geq 0} \xrightarrow{\mathcal{Q}(\cdot)}\Lambda$ where $\Lambda=\{0,1,...,255\}$ is an 8-bit discrete state space and
\begin{equation}
\mathcal{Q}(f_{i,j})=\lfloor 255\cdot \frac{f_{i,j}}{max(\mathbf{F}_{i})}\rceil,
\end{equation}
where $\lfloor\cdot\rceil$ is the round to the nearest integer function. Let us define the random variable $F =\mathcal{Q}(f_{i,j})$ with possible outcome $\lambda\in\Lambda$. Then, the probability of $\lambda$ is
\begin{equation}
p_{F}(\lambda)=n_{\lambda}/|\mathbf{F}_{i}| \quad \forall \quad \lambda\in\Lambda,
\label{eq:pmf_fms}
\end{equation}
where $n_{\lambda}$ is the number of times $\lambda$ occurs and $|\cdot|$ is the cardinality of the feature map. The entropy of the feature map $\mathbf{F}_{i}$ is then defined as
\begin{equation}
\mathcal{H}_{i}=-\sum_{\lambda\in\Lambda}p_{F}(\lambda)log_{2}p_{F}(\lambda).
\label{eq:entropy}
\end{equation}
\subsection{Measuring Kernels Redundancy}
\label{sec:kl}
The weights in neural networks are generally initialized from a normal distribution.
A DCNN can have redundancy between kernels in a layer. Removing the redundant kernels prunes the network while may slightly drop the classification accuracy. A kernel $\Theta$ is generally a three-dimensional tensor of size $K_{1}\times K_{2}\times N$ where $K=K_{1}\times K_{2}$ is the size of a filter and $N$ is the number of filters, corresponding to the number of input channels. Therefore, we can represent the weights in a kernel with $K$ sets which are ${W_{1},...,W_{K}}$
where ${W_{k}=\{\theta_{k,1},...,\theta_{k,N}\}}$ and $k\in\{1,...,K\}$. Let us assume the weights $W_{k}$ have a normal distribution. Hence, for the kernel $i$ we have a multivariate normal distribution with means ${\bm{\mu}_{i}=(\mu_{i,1},...,\mu_{i,K})}$ and the ${K\times K}$ covariance matrix ${\bm{\Sigma}}_{i}$. The distributions $\mathcal{N}_{i}(\bm{\mu}_{i},\bm{\Sigma}_{i})$ and $\mathcal{N}_{j}(\bm{\mu}_{j},\bm{\Sigma}_{j})$ of two given kernels $i$ and $j$, respectively, have the same dimension. Hence, we can compute the KL divergence between the two kernels $i$ and $j$ as
\begin{equation}
\begin{split}
\mathcal{D}_{KL}(\mathcal{N}_{i}||\mathcal{N}_{j})=\\&
\frac{1}{2}\Big(
tr(\bm{\Sigma}_{j}^{-1}\bm{\Sigma}_{i})
+ (\bm{\mu}_{j}-\bm{\mu}_{i})^\top \bm{\Sigma}_{j}^{-1} (\bm{\mu}_{j}-\bm{\mu}_{i}))\\
&-K
+ln\Big(\frac{|\bm{\Sigma}_{j}|}{|\bm{\Sigma}_{i}|}\Big)
\Big),
\end{split}
\label{eq:kldivergence}
\end{equation}
where $tr(\cdot)$ is the trace and $|\cdot|$ is the determinant.
\subsection{Ising Energy Formulation}
A neural network $\mathcal{F}$ has the set of layers $A=\{A_{1}\cup A_{2}\}$, where $A_{1}$ and $A_{2}$ are the set of convolutional and dense layers, respectively. Obviously, the sets $A_{1}$ and $A_{2}$ are disjoint (i.e. $A_{1}\cap A_{2}=\emptyset$). Hereafter we refer to a hidden unit or a convolutional kernel a unit for simplicity. A binary state vector $\mathbf{s}$ with length $D$ represents the state of the units, where $s_{d}\in\{0,1\}\:\forall\:d\in\{1,...,D\}$. If $s_{d}=0$ unit $d$ is inactive and if $s_{d}=1$ the unit participates in training and inference. Therefore, the state vector $\mathbf{s}$ represents a subnetwork of the original network. The unit $d$ belongs to a layer $l\in A$.
Let us represent the network $\mathcal{F}$ as a graph $G = (\mathcal{D},\Gamma)$, where $\mathcal{D}$ is the set of vertices
(nodes) with cardinality $D$ and $\Gamma$ is the set of edges (connections) with weight $\gamma_{d,d'}$ between vertices $d$ and $d'$. The graph has two types of connections, where the connection between vertices of a layer is bidirectional and the connection between layers is unidirectional. In dense layers, unidirectional connections exist between nodes of two layers where each node has a state $s_{d}\in\{0,1\}$, except in the last layer (logits layer), where $s_{d}=1$. We are interested in pruning the vertices and all corresponding edges.
We model the dependencies between vertices in the graph $G$ using the Ising energy model as
\begin{equation}
\mathcal{E}=-\sum_{d\in \mathcal{D}}\sum_{d'\in \mathcal{D}}\gamma_{d,d'}s_{d}s_{d'} - b\sum_{d\in \mathcal{D}}s_{d},
\label{eq:ising}
\end{equation}
where $b$ is the bias coefficient and $\gamma_{d,d'}$ is the weight between the vertices $d$ and $d'$ defined as
\begin{equation}
\gamma_{d,d'}=\begin{cases}
\mathcal{D}_{KL}(\mathcal{N}_{d}||\mathcal{N}_{d'})-1\:$if$\: d,d'\in l \: \& \:l\in A_{1} \\
\mathcal{H}_{d}-1\:$if$\: d\in l,d'\in l+1 \: \& \:l,l+1\in A_{1}\\
\mathcal{A}_{d}-1\:$if$\: d\in l,d'\in l+1 \: \& \:l,l+1\in A_{2}\\
0 \:\:\:\:$otherwise$\\
\end{cases},
\label{eq:cases}
\end{equation}
where $\mathcal{H}_{d}$ is calculated using~(\ref{eq:entropy}). Similar to the approach we have proposed in~\cite{salehinejad2019ising} for hidden units in dense layer, we have
\begin{equation}
\mathcal{A}_{i}=tanh(a_{i}),
\label{eq:denseent}
\end{equation}
which maps the activation value $a_{i}$ of the unit $i$, generated by the ReLU activation function such that a dead unit has the lowest $\mathcal{A}_{i}$ and a highly activated unit has a high $\mathcal{A}_{i}$.
In~(\ref{eq:cases}) a high weight is allocated to the unidirectional connections of a unit with high activation value and a high weight is allocated to the bidirectional connections with high KL divergence, and vice-versa. From another perspective, the first case allocates small weight to low-active units and the latter case allocates small weight to redundant units.
Assuming all the states are active (i.e. $s_{d}=1\;\forall\;d\in\mathcal{D}$), the bias coefficient is defined to balance the interaction term and the bias term by setting $\mathcal{E}=0$. Hence,
\begin{equation}
\begin{split}
b &= -\frac{\sum_{d\in \mathcal{D}}\sum_{d'\in \mathcal{D}}\gamma_{d,d'}s_{d}s_{d'}}{\sum_{d\in \mathcal{D}}s_{d}}\\
&=-\frac{|\bm{\gamma}|}{D},
\end{split}
\end{equation}
where $|\bm{\gamma}|$ is the sum of weights $\bm{\gamma}$ and $\sum_{d\in \mathcal{D}}s_{d}=D$. Minimizing (\ref{eq:ising}) is equivalent to finding a state vector which represents a sub-network of $\mathcal{F}$ with a smaller number of redundant kernels and inactive units.
\begin{algorithm}[t]
\small
\begin{algorithmic}
\State Set $t$ = 0 // Optimization counter
\State Initiate the neural network $\mathcal{F}$
\State Set $\mathbf{S}^{(0)}\sim Bernoulli(P=0.5)$ // States initialization
\State Set $\Delta\mathbf{s}\neq 0$ // Early state threshold
\For{ $ \mathit{i_{epoch}} = 1 \rightarrow \mathit{N_{epoch}}$} // Epoch counter
\For{ $ \mathit{i_{batch}} = 1 \rightarrow \mathit{N_{batch}}$} // Batch counter
\State $t = t+1$
\If{$\Delta\mathbf{s}\neq 0$}
\If{$i_{epoch}=1 \; \& \; i_{batch}=1$}
\State Compute energy of $\mathbf{S}^{(0)}$ using~(\ref{eq:ising})
\EndIf
\For {$i=1\rightarrow S$} // States counter
\State Generate mutually different $i_{1},i_{2},i_{3}\in \{1,...,S\}$
\For {$d=1\rightarrow D$} // State dimension counter
\State Generate a random number $r_{d}\in[0,1]$
\State Compute mutation vector $v_{i,d}$ using (\ref{eq:mutation})
\State Compute candidate state $\tilde{s}^{(t)}$ using (\ref{eq:crossover})
\EndFor
\EndFor
\State Compute energy loss of $\tilde{\mathbf{S}}^{(t)}$ using~(\ref{eq:ising})
\State Select $\mathbf{S}^{(t)}$ and corresponding energy using (\ref{eq:selection})
\State Select the state with the lowest energy from $\mathbf{S}^{(t)}$ as $\mathbf{s}^{(t)}_{b}$
\Else
\State $\mathbf{s}^{(t)}_{b}=\mathbf{s}^{(t-1)}_{b}$
\EndIf
\State Temporarily drop weights of $\mathcal{F}$ according to $\mathbf{s}^{(t)}_{b}$
\State Compute cross-entropy loss of the sparsified network
\State Perform backpropagation to update active weights
\EndFor
\State Update $\Delta\mathbf{s}$ for early state convergence using (\ref{eq:stateconvergence})
\EndFor
\end{algorithmic}
\small
\caption{IPruning}
\label{alg:IsingEnergy-basedDropout}
\end{algorithm}
\subsection{Optimization of Ising Energy}
Algorithm~\ref{alg:IsingEnergy-basedDropout} shows different steps of IPruning. The process of searching for the pruning state vector with lowest energy is incorporated into the typical training of the neural network $\mathcal{F}$ with backpropagation. First, a population of candidate state vectors is initiated and then the Ising energy loss is computed for each vector. Then, the population of vectors is evolved on the optimization landscape of states with respect to the Ising energy and the state with lowest energy is selected. Dropout is performed according to the selected state vector and only active weights are updated with backpropagation. The population is then evolved and the same procedure is repeated until the population of states converges to a best state solution or a predefined number of iterations is reached.
Let us initialize a population of candidate states $\mathbf{S}^{(t)}\in\mathbb{Z}_{2}^{S\times D}$ such that $\mathbf{s}^{(t)}_{i}\in\mathbf{S}^{(t)}$, where $t$ is the iteration and $s_{i,d}^{(0)}\sim Bernoulli(P=0.5)$ for ${i\in\{1,...,S\}}$ and ${d\in\{1,...,D\}}$. A state vector $\mathbf{s}^{(t)}_{j}\in\mathbf{S}^{(t)}$ selects a subset of the graph $G$.
The optimization procedure has three phases which are mutation, crossover, and selection. Given the population of states $\mathbf{S}^{(t-1)}$, a mutation vector is defined for each candidate state $\mathbf{s}_{i}^{(t-1)}\in\mathbf{S}^{(t-1)}$ as
\begin{equation}
v_{i,d}=\begin{cases}
1-s_{i_{1},d}^{(t-1)} \:\:\:\:$if$\:\:\:s_{i_{2},d}^{(t-1)}\neq s_{i_{3},d} ^{(t-1)}\;$\&$\; r_{d}<F\\
s_{i_{1},d}^{(t-1)} \:\:\:\:\:\:\:\:\:\:\:$ otherwise $
\end{cases},
\label{eq:mutation}
\end{equation}
for $d\in\{1,..,D\}$ where $i_{1},i_{2},i_{3}\in \{1,...,S\}$ are mutually different, $F$ is the mutation factor~\cite{salehinejad2017micro}, and $r_{d}\in[0,1]$ is a random number.
The next step is to crossover the mutation vectors to generate new candidate state vectors as
\begin{equation}
\tilde{s}^{(t)}_{i,d}=\begin{cases}
v_{i,d} \:\:\:\:\:\:\:\:\:\:\:\:$if$\:\:\: r'_{d}\in[0,1] \leq C\\
s_{i,d}^{(t-1)} \:\:\:\:\:\:\:\:\:\:\:$ otherwise $
\end{cases},
\label{eq:crossover}
\end{equation}
where $C=0.5$ is the crossover coefficient~\cite{salehinejad2017micro}. The parameters $C$ and $F$ control exploration and exploitation of the optimization landscape. Each generated state $ \tilde{\mathbf{s}}_{i}^{(t)}$ is then compared with its corresponding parent with respect to its energy value $\tilde{\mathcal{E}}^{(t)}_{i}$ and the state with smaller energy is selected as
\begin{equation}
\mathbf{s}_{i}^{(t)}=\begin{cases}
\tilde{\mathbf{s}}_{i}^{(t)} \:\:\:\:\:\:\:\:\:\:$if$\:\:\: \tilde{\mathcal{E}}^{(t)}_{i}\leq \mathcal{E}^{(t-1)}_{i} \\
\mathbf{s}_{i}^{(t-1)} \:\:\:\:\:$ otherwise $
\end{cases}\:\forall\:i\in\{1,...,S\}.
\label{eq:selection}
\end{equation}
The state with minimum energy $\mathcal{E}_{b}^{(t)}=min\{\mathcal{E}_{1}^{(t)},...,\mathcal{E}_{S}^{(t)}\}$ is selected as the best state $\mathbf{s}_{b}$, which represents the sub-network for next training batch. This optimization strategy is simple and feasible to implement in parallel for a large $S$.
After a number of iterations, depending on the capacity of the neural network and complexity of the dataset, all the states in $\mathbf{S}^{(t)}$ may converge to the best state vector $\mathbf{s}_{b}\in\mathbf{S}^{(t)}$ with the Ising energy $\mathcal{E}_{b}^{(t)}$. Hence, we can define
\begin{equation}
\Delta\mathbf{s} =
\mathcal{E}_{b}^{(t)} -
\frac{1}{S}\sum\limits_{j=1}^{S}\mathcal{E}_{j}^{(t)},
\label{eq:stateconvergence}
\end{equation}
such that if $\Delta\mathbf{s}=0$, we can call for an {early state convergence} and continue training by fine-tuning the sub-network identified by the state vector $\mathbf{s}_{b}$.
\begin{table}[!ht]
\captionsetup{font=footnotesize}
\caption{Classification performance on the test datasets. $R$ is kept trainable parameters and $\#p$ is approximate number of trainable parameters. All the values except loss and $\#p$ are in percentage. (F) refers to full network used for inference and (P) refers to pruned network using \textit{IPruning}.}
\begin{subtable}{0.99\linewidth}
\centering
\captionsetup{font=footnotesize}
\caption{ \textbf{CIFAR-10} }
\begin{adjustbox}{width=1\textwidth}
\begin{tabular}{lcccccc}
\hline
\multicolumn{1}{c}{Model} & Loss & Top-1 & Top-3 & Top-5 & $R$ & $\#p$ \\ \hline \hline
\multicolumn{1}{l}{ResNet-18}
& \multicolumn{1}{l}{0.3181}
& \multicolumn{1}{l}{92.81}
& \multicolumn{1}{l}{98.78}
& \multicolumn{1}{l}{99.49}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{11.2M}\\
\multicolumn{1}{l}{ResNet-18+DeepCompression}
& \multicolumn{1}{l}{0.6893}
& \multicolumn{1}{l}{76.18}
& \multicolumn{1}{l}{94.21}
& \multicolumn{1}{l}{98.63}
& \multicolumn{1}{l}{49.19}
& \multicolumn{1}{l}{5.5M} \\
\multicolumn{1}{l}{ResNet-18+IPruning(F)}
& \multicolumn{1}{l}{0.5167}
& \multicolumn{1}{l}{84.12}
& \multicolumn{1}{l}{96.74}
& \multicolumn{1}{l}{99.24}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{11.2M}\\
\multicolumn{1}{l}{ResNet-18+IPruning(P)}
& \multicolumn{1}{l}{0.5254}
& \multicolumn{1}{l}{84.09}
& \multicolumn{1}{l}{96.77}
& \multicolumn{1}{l}{99.33}
& \multicolumn{1}{l}{49.19}
& \multicolumn{1}{l}{5.5M}\\ \hline\hline
\multicolumn{1}{l}{ResNet-34}
& \multicolumn{1}{l}{0.3684}
& \multicolumn{1}{l}{92.80}
& \multicolumn{1}{l}{98.85}
& \multicolumn{1}{l}{99.71}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{21.3M} \\
\multicolumn{1}{l}{ResNet-34+DeepCompression}
& \multicolumn{1}{l}{0.8423}
& \multicolumn{1}{l}{71.45}
& \multicolumn{1}{l}{93.28}
& \multicolumn{1}{l}{98.39}
& \multicolumn{1}{l}{49.61}
& \multicolumn{1}{l}{10.5M}\\
\multicolumn{1}{l}{ResNet-34+IPruning(F)}
& \multicolumn{1}{l}{0.6352}
& \multicolumn{1}{l}{88.78}
& \multicolumn{1}{l}{98.14}
& \multicolumn{1}{l}{99.41}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{21.3M} \\
\multicolumn{1}{l}{ResNet-34+IPruning(P)}
& \multicolumn{1}{l}{0.6401}
& \multicolumn{1}{l}{88.72}
& \multicolumn{1}{l}{97.93}
& \multicolumn{1}{l}{99.42}
& \multicolumn{1}{l}{49.61}
& \multicolumn{1}{l}{10.5M} \\ \hline\hline
\multicolumn{1}{l}{ResNet-50}
& \multicolumn{1}{l}{0.3761}
& \multicolumn{1}{l}{92.21}
& \multicolumn{1}{l}{98.70}
& \multicolumn{1}{l}{99.51}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{23.5M} \\
\multicolumn{1}{l}{ResNet-50+DeepCompression}
& \multicolumn{1}{l}{1.0355}
& \multicolumn{1}{l}{67.47}
& \multicolumn{1}{l}{90.45}
& \multicolumn{1}{l}{97.26}
& \multicolumn{1}{l}{43.46}
& \multicolumn{1}{l}{10.2M} \\
\multicolumn{1}{l}{ResNet-50+IPruning(F)}
& \multicolumn{1}{l}{0.8200}
& \multicolumn{1}{l}{82.32}
& \multicolumn{1}{l}{95.92}
& \multicolumn{1}{l}{97.37}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{23.5M} \\
\multicolumn{1}{l}{ResNet-50+IPruning(P)}
& \multicolumn{1}{l}{0.8374}
& \multicolumn{1}{l}{82.45}
& \multicolumn{1}{l}{95.32}
& \multicolumn{1}{l}{97.27}
& \multicolumn{1}{l}{43.46}
& \multicolumn{1}{l}{10.2M} \\ \hline\hline
\multicolumn{1}{l}{ResNet-101}
& \multicolumn{1}{l}{0.3680}
& \multicolumn{1}{l}{92.66}
& \multicolumn{1}{l}{98.69}
& \multicolumn{1}{l}{99.65}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{42.5M} \\
\multicolumn{1}{l}{ResNet-101+DeepCompression}
& \multicolumn{1}{l}{1.083}
& \multicolumn{1}{l}{66.63}
& \multicolumn{1}{l}{92.03}
& \multicolumn{1}{l}{97.97}
& \multicolumn{1}{l}{42.41}
& \multicolumn{1}{l}{18.0M} \\
\multicolumn{1}{l}{ResNet-101+IPruning(F)}
& \multicolumn{1}{l}{0.8233}
& \multicolumn{1}{l}{84.47}
& \multicolumn{1}{l}{97.42}
& \multicolumn{1}{l}{98.47}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{42.5M} \\
\multicolumn{1}{l}{ResNet-101+IPruning(P)}
& \multicolumn{1}{l}{0.8372}
& \multicolumn{1}{l}{84.38}
& \multicolumn{1}{l}{97.03}
& \multicolumn{1}{l}{98.37}
& \multicolumn{1}{l}{42.41}
& \multicolumn{1}{l}{18.0M} \\ \hline\hline
\multicolumn{1}{l}{AlexNet}
& \multicolumn{1}{l}{0.9727}
& \multicolumn{1}{l}{84.32}
& \multicolumn{1}{l}{96.58}
& \multicolumn{1}{l}{99.08}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{57.4M} \\
\multicolumn{1}{l}{AlexNet+IPruning(F)}
& \multicolumn{1}{l}{0.8842}
& \multicolumn{1}{l}{74.02}
& \multicolumn{1}{l}{92.79}
& \multicolumn{1}{l}{97.63}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{57.4M} \\
\multicolumn{1}{l}{AlexNet+IPruning(P)}
& \multicolumn{1}{l}{0.8830}
& \multicolumn{1}{l}{73.62}
& \multicolumn{1}{l}{92.35}
& \multicolumn{1}{l}{97.03}
& \multicolumn{1}{l}{62.84}
& \multicolumn{1}{l}{36.0M} \\
\multicolumn{1}{l}{SqueezeNet}
& \multicolumn{1}{l}{0.5585}
& \multicolumn{1}{l}{81.49}
& \multicolumn{1}{l}{96.31}
& \multicolumn{1}{l}{99.01}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{0.73M} \\
\multicolumn{1}{l}{SqueezeNet+IPruning(F)}
& \multicolumn{1}{l}{0.6894}
& \multicolumn{1}{l}{76.74}
& \multicolumn{1}{l}{95.53}
& \multicolumn{1}{l}{98.54}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{0.73M} \\
\multicolumn{1}{l}{SqueezeNet+IPruning(P)}
& \multicolumn{1}{l}{0.6989}
& \multicolumn{1}{l}{76.35}
& \multicolumn{1}{l}{95.13}
& \multicolumn{1}{l}{98.34}
& \multicolumn{1}{l}{51.26}
& \multicolumn{1}{l}{0.37M} \\ \hline
\end{tabular}
\end{adjustbox}
\label{T:results_cifar10_IPruning}
\end{subtable}
\begin{subtable}{0.99\linewidth}
\centering
\captionsetup{font=footnotesize}
\caption{\textbf{CIFAR-100}}
\begin{adjustbox}{width=1\textwidth}
\begin{tabular}{lcccccc}
\hline
\multicolumn{1}{c}{Model} & Loss & Top-1 & Top-3 & Top-5 & $R$ & $\#p$ \\ \hline \hline
\multicolumn{1}{l}{ResNet-18}
& \multicolumn{1}{l}{1.3830}
& \multicolumn{1}{l}{69.03}
& \multicolumn{1}{l}{84.44}
& \multicolumn{1}{l}{88.90}
& \multicolumn{1}{c}{100}
& \multicolumn{1}{l}{11.2M}\\
\multicolumn{1}{l}{ResNet-18+DeepCompression}
& \multicolumn{1}{l}{2.2130}
& \multicolumn{1}{l}{40.15}
& \multicolumn{1}{l}{61.92}
& \multicolumn{1}{l}{71.84}
& \multicolumn{1}{l}{47.95}
& \multicolumn{1}{l}{5.3M} \\
\multicolumn{1}{l}{ResNet-18+IPruning(F)}
& \multicolumn{1}{l}{1.8431}
& \multicolumn{1}{l}{55.43}
& \multicolumn{1}{l}{74.94}
& \multicolumn{1}{l}{82.60}
& \multicolumn{1}{c}{100}
& \multicolumn{1}{l}{11.2M}\\
\multicolumn{1}{l}{ResNet-18+IPruning(P)}
& \multicolumn{1}{l}{1.8696}
& \multicolumn{1}{l}{56.43}
& \multicolumn{1}{l}{75.37}
& \multicolumn{1}{l}{82.43}
& \multicolumn{1}{c}{47.95}
& \multicolumn{1}{l}{5.3M}\\ \hline\hline
\multicolumn{1}{l}{ResNet-34}
& \multicolumn{1}{l}{1.3931}
& \multicolumn{1}{l}{69.96}
& \multicolumn{1}{l}{85.65}
& \multicolumn{1}{l}{90.10}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{21.3M} \\
\multicolumn{1}{l}{ResNet-34+DeepCompression}
& \multicolumn{1}{l}{2.1778}
& \multicolumn{1}{l}{42.09}
& \multicolumn{1}{l}{65.01}
& \multicolumn{1}{l}{74.31}
& \multicolumn{1}{l}{49.41}
& \multicolumn{1}{l}{10.5M} \\
\multicolumn{1}{l}{ResNet-34+IPruning(F)}
& \multicolumn{1}{l}{2.3789}
& \multicolumn{1}{l}{60.73}
& \multicolumn{1}{l}{79.26}
& \multicolumn{1}{l}{85.48}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{21.3M} \\
\multicolumn{1}{l}{ResNet-34+IPruning(P)}
& \multicolumn{1}{l}{2.3794}
& \multicolumn{1}{l}{61.13}
& \multicolumn{1}{l}{79.23}
& \multicolumn{1}{l}{85.30}
& \multicolumn{1}{c}{49.41}
& \multicolumn{1}{l}{10.5M} \\ \hline\hline
\multicolumn{1}{l}{ResNet-50}
& \multicolumn{1}{l}{1.3068}
& \multicolumn{1}{l}{71.22}
& \multicolumn{1}{l}{86.47}
& \multicolumn{1}{l}{90.74}
& \multicolumn{1}{c}{100}
& \multicolumn{1}{l}{23.7M} \\
\multicolumn{1}{l}{ResNet-50+DeepCompression}
& \multicolumn{1}{l}{2.4927}
& \multicolumn{1}{l}{43.72}
& \multicolumn{1}{l}{66.93}
& \multicolumn{1}{l}{76.15}
& \multicolumn{1}{l}{44.63}
& \multicolumn{1}{l}{10.8M} \\
\multicolumn{1}{l}{ResNet-50+IPruning(F)}
& \multicolumn{1}{l}{1.8750}
& \multicolumn{1}{l}{60.44}
& \multicolumn{1}{l}{79.25}
& \multicolumn{1}{l}{86.24}
& \multicolumn{1}{c}{100}
& \multicolumn{1}{l}{23.7M} \\
\multicolumn{1}{l}{ResNet-50+IPruning(P)}
& \multicolumn{1}{l}{2.1462}
& \multicolumn{1}{l}{60.05}
& \multicolumn{1}{l}{78.83}
& \multicolumn{1}{l}{85.78}
& \multicolumn{1}{c}{44.63}
& \multicolumn{1}{l}{10.8M} \\ \hline\hline
\multicolumn{1}{l}{ResNet-101}
& \multicolumn{1}{l}{1.3574}
& \multicolumn{1}{l}{71.19}
& \multicolumn{1}{l}{85.54}
& \multicolumn{1}{l}{90.00}
& \multicolumn{1}{c}{100}
& \multicolumn{1}{c}{42.6M} \\
\multicolumn{1}{l}{ResNet-101+DeepCompression}
& \multicolumn{1}{l}{2.6232}
& \multicolumn{1}{l}{36.58}
& \multicolumn{1}{l}{57.82}
& \multicolumn{1}{l}{68.36}
& \multicolumn{1}{l}{41.36}
& \multicolumn{1}{c}{17.6M} \\
\multicolumn{1}{l}{ResNet-101+IPruning(F)}
& \multicolumn{1}{l}{2.1338}
& \multicolumn{1}{l}{60.52}
& \multicolumn{1}{l}{79.91}
& \multicolumn{1}{l}{83.22}
& \multicolumn{1}{c}{100}
& \multicolumn{1}{c}{42.6M} \\
\multicolumn{1}{l}{ResNet-101+IPruning(P)}
& \multicolumn{1}{l}{2.2952}
& \multicolumn{1}{l}{60.35}
& \multicolumn{1}{l}{78.99}
& \multicolumn{1}{l}{83.01}
& \multicolumn{1}{c}{41.36}
& \multicolumn{1}{c}{17.6M} \\ \hline\hline
\multicolumn{1}{l}{AlexNet}
& \multicolumn{1}{l}{2.8113}
& \multicolumn{1}{l}{60.12}
& \multicolumn{1}{l}{79.18}
& \multicolumn{1}{l}{83.31}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{57.4M} \\
\multicolumn{1}{l}{AlexNet+IPruning(F)}
& \multicolumn{1}{l}{2.7420}
& \multicolumn{1}{l}{53.52}
& \multicolumn{1}{l}{72.42}
& \multicolumn{1}{l}{79.70}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{57.4M} \\
\multicolumn{1}{l}{AlexNet+IPruning(P)}
& \multicolumn{1}{l}{2.7396}
& \multicolumn{1}{l}{53.05}
& \multicolumn{1}{l}{72.28}
& \multicolumn{1}{l}{79.69}
& \multicolumn{1}{l}{65.35}
& \multicolumn{1}{l}{37.5M} \\
\multicolumn{1}{l}{SqueezeNet}
& \multicolumn{1}{l}{1.4150}
& \multicolumn{1}{l}{67.85}
& \multicolumn{1}{l}{85.81}
& \multicolumn{1}{l}{89.69}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{0.77M} \\
\multicolumn{1}{l}{SqueezeNet+IPruning(F)}
& \multicolumn{1}{l}{1.9285}
& \multicolumn{1}{l}{61.93}
& \multicolumn{1}{l}{80.74}
& \multicolumn{1}{l}{86.92}
& \multicolumn{1}{l}{100}
& \multicolumn{1}{l}{0.77M} \\
\multicolumn{1}{l}{SqueezeNet+IPruning(P)}
& \multicolumn{1}{l}{1.9437}
& \multicolumn{1}{l}{61.46}
& \multicolumn{1}{l}{80.45}
& \multicolumn{1}{l}{85.81}
& \multicolumn{1}{l}{53.20}
& \multicolumn{1}{l}{0.41M} \\ \hline
\end{tabular}
\end{adjustbox}
\label{T:results_cifar100_IPruning}
\end{subtable}
\label{T:results_IPruning}
\vspace{-6mm}
\end{table}
\section{Experiments}
\label{sec:experiemnts}
The experiments were conducted on the CIFAR-10~and
CIFAR-100~\cite{krizhevsky2009learning} datasets using ResNets (18, 34, 50, and 101 layers)~\cite{he2016deep}, AlexNet~\cite{krizhevsky2012imagenet}, SqueezeNet~\cite{iandola2016squeezenet}, and Deep Compression~\cite{han2015deep}. Horizontal flip and Cutout~\cite{devries2017improved} augmentation methods were used.
The results are averaged over five independent runs. The Adadelta optimizer with Step adaptive learning rate (step: every 50 epoch at gamma rate of 0.1) and weight decay of $10e^{-6}$ is used. The number of epochs is 200 and the batch size is 128. Random dropout rate is set to 0.5 where applicable, except for the proposed model. The early state convergence in~(\ref{eq:stateconvergence}) is used with a threshold of 100.
As Table~\ref{T:results_IPruning} shows, IPruning on average has removed more than $50\%$ of the trainable weights and the Top-1 performance has dropped less than $10\%$ compared to the original model. We used the pruning rate achieved by IPruning to prune the original network using Deep Compression \cite{han2015deep}. Since this method is tailored to pruning certain layers, we have modified it to prune every layer, similar to IPruning. We also have evaluated inference results of IPruning in full and pruned modes. The former refers to training the network with IPruning but performing inference using the full model, and the latter refers to training the network with IPruning and performing inference with the pruned network. The results show that the full network has slightly better performance than the pruned network. It shows that we are able to achieve very competitive performance using the pruned network compared with the full network, which has a larger capacity, trained with IPruning.
\begin{figure}[!t]
\centering
\centering
\captionsetup{font=footnotesize}
\begin{subfigure}[t]{0.4\textwidth}
\captionsetup{font=footnotesize}
\centering
\includegraphics[width=1\textwidth]{energycifarsr18.eps}
\caption{Average energy and best energy of states population.}
\label{fig:}
\end{subfigure
\begin{subfigure}[t]{0.4\textwidth}
\captionsetup{font=footnotesize}
\centering
\includegraphics[width=1\textwidth]{keptratecifarsr18.eps}
\caption{Rate of kept trainable parameters.}
\label{fig:}
\end{subfigure
\caption{Energy loss and kept rate for ResNet-18 over 1,000 training iterations.}
\vspace{-6mm}
\label{fig:visualization_cifar_res18_idropout}
\end{figure}
Figure~\ref{fig:visualization_cifar_res18_idropout} shows the energy loss and corresponding pruning rate over 1,000 training iterations of IPruning for ResNet-18. Since CIFAR-100 is more complicated than CIFAR-10, it converges slower. Results show that the pruning rates generally converge to a value close to $50\%$, regardless of the initial distribution of the population. This might be due to the behavior of optimizer in very high dimensional space and its limited capability of reaching all possible states during the evolution.
\section{Conclusions}
\label{sec:conclusion}
We propose an Ising energy-based framework, called IPruning, for structured pruning of neural networks. Unlike most other methods, IPruning considers every trainable weight in any layer of a given network for pruning. From an implementation perspective, most pruning methods require manual modification of network architecture to apply the pruning mask while IPruning can automatically detect trainable weights and construct a pruning graph for a given network.
\section{Acknowledgment}
The authors acknowledge financial support of Fujitsu Laboratories Ltd. and Fujitsu Consulting (Canada) Inc
\bibliographystyle{IEEEbib}
|
\section{Introduction}
The development of deep learning models has contributed to the development of solutions for challenges previously considered unsolvable.
However, a concerning vulnerability in those models was discovered:
An imperceptible perturbation to a legitimate input sample creates an adversarial example that causes the model to output an incorrect prediction with high confidence \citep{szegedy2013intriguing}.
Although adversarial examples were initially observed in the digital space \citep{goodfellow2014explaining, papernot2016limitations}, they were later demonstrated in the real world \citep{kurakin2016adversarial}; making the threat even greater.
While several studies have presented various use cases for real-world adversarial perturbations in the image domain \citep{sharif2016accessorize, evtimov2017robust, lee2019physical}, they all utilized a similar methodology:
First, one or more 2D photos of the target scene are used to craft an adversarial perturbation that can be digitally added to the entire image or applied as a patch to a portion of an image.
Then, the perturbation is recreated in the real world (i.e., printed) and placed in the scene.
Finally, photos of the scene with the adversarial perturbation are fed to the neural network for evaluation.
Although that methodology has shown promising results, such 2D image-based methods do not accurately represent the 3D real world (as shown in \fig{challenges}).
The main challenge stems from the fact that while both the photos of the scene and the adversarial perturbation are flat, a real-world scene is not.
A flat patch must be placed on a flat surface and must always face the camera; otherwise, parts of the patch will be hidden, and the attack may fail.
Furthermore, crafting the adversarial perturbation based on photos of the scene limits the attacker's ability to model real-world properties as part of the attack.
Changes to the environmental settings, such as lighting, must be manually added to the real-world scene, and only then can the attacker use them as part of the attack.
In addition, such attacks can only be implemented when the attacker fully controls the target scene, which not a realistic assumption.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{images/challenges.PNG}
\caption{
Some of the challenges that should be addressed when crafting an adversarial patch for a real-world scene:
1) The patch is placed on a curved surface.
2) The patch is partially hidden by the object it is placed on.
3) The patch should have the same lighting as the rest of the scene.
4) The patch is hidden by an object that is placed in front of it.
5) Objects casting shadows on one another.
6) The scene includes an object that affects multiple objects (e.g., an object casts a shadow on more than one object).
7) The scene may have different environmental conditions (i.e., point yellow light).
}
\label{fig:challenges}
\end{figure}
Several studies focused on crafting adversarial perturbations for 3D objects.
A prominent example \citep{athalye2018synthesizing} presents the expectation over transformation (EOT) framework for crafting adversarial perturbations that are robust to random transformations (e.g., rotation, translation); the framework relies on the attacker’s ability to model those transformations as part of the attack process.
\citet{athalye2018synthesizing} used EOT to perturb the texture of a digital 3D object, which was later printed using a color 3D printer.
Although successful, the attack targeted a single object, as opposed to the complex scenes that are more common in the real world.
As \fig{challenges} shows, a realistic scene contains environmental characteristics (e.g., ambient lighting) and different objects that can affect one another (e.g., hide, shadow).
Failing to consider those attributes can impair the attack’s performance.
Additionally, EOT relies on a specific implementation (limiting the ability to use external modeling and rendering tools) and a 3D color printer, a piece of equipment that is not accessible to many people.
Both issues complicate the attack and make it less feasible.
Another limitation of prior work in this area is the inability to perform the experiment in the real world multiple times and obtain similar results.
In some studies, performing the same real-world experiment more than once can be challenging thus limiting the ability to evaluate the attack.
For instance, it is impossible to replicate and validate an experiment in which a set of photos of the scene was obtained from random positions if the authors do not document the exact positions of the camera.
In other studies, different researchers cannot replicate the results because the target scene or evaluation process cannot be (exactly) recreated by subsequent researchers.
Therefore, in cases like these, the scientific community cannot properly assess methods proposed in prior research on real-world adversarial perturbations.
In this paper, we present a framework that utilizes 3D computer graphic methods to craft adversarial patches that can be added to an existing real-world scene.
Our approach allows the attacker to target complex realistic scenes, with multiple objects, environmental characteristics, etc.
The framework uses a digital 3D replica of the target scene to simulate the real world, thus allowing the attacker to assess the limitations of the patch and improve it without risking detection.
Since the attacker controls the digital replica, he/she can add, change, and remove different elements and as a result, can create an adversarial patch that is more robust to real-world transformations.
The framework is designed to allow the integration of external modeling and rendering tools, which gives the attacker the flexibility to implement the attack using the tools he/she prefers.
We implement the framework using open-source tools, and thus the attack is also accessible to an attacker with a limited budget.
Additionally, we propose an evaluation process that is specifically designed so that the experiment can be performed multiple times and replicated in future studies.
The two-step evaluation process is used in several experiments to evaluate our attack in both the digital space and the real world.
We also examine whether the use of a digital replica of the scene to create and improve an adversarial patch can simulate the patch's performance in the real world.
To do so, we use the framework to create adversarial patches of multiple target classes and use the evaluation process to compare their performance in the digital space and the real world.
Our results show that using the digital replica to evaluate the adversarial patch can expose useful information about the patch's performance in the real world.
Furthermore, in realistic settings, unexpected changes may occur in the target scene.
Therefore, we examine the patch’s robustness to changes in the scene that were not modeled as part of the scene's replica.
To do so, we evaluate how changing or adding new objects to the scene affects the patch's ability to fool the target neural network.
Finally, by publishing our evaluation setup and code \citep{author_repository}, researchers can reproduce our results and improve upon them.
The main contributions of this study are:
\begin{itemize}
\item We present a framework for crafting and improving adversarial patches for an existing real-world scene in the risk-free environment of the digital space.
\item We demonstrate how the framework can be used to craft a low-budget adversarial patches using free, open-source, and common 3D modeling tools.
\item We present an evaluation process that enables reproducible experiments in the digital space and the real world.
\end{itemize}
\section{Background}
When discovered by \citet{szegedy2013intriguing}, adversarial perturbations were considered a minor bug, but that changed with the development of advanced adversarial attacks and the growing number of failed attempts to defend against them \citep{goodfellow2014explaining, papernot2016limitations, carlini2017towards}.
Adversarial perturbations' success at fooling neural networks led to an interest in implementing attacks in the real world \citep{kurakin2016adversarial}.
While the first real-world studies performed were unable to reproduce the results obtained in the digital space, these studies set the stage for research on novel forms of adversarial perturbations.
One example is adversarial patches, which are small shapes that, when added to a specific part of an image, could fool object classifiers \citep{evtimov2017robust, brown2017adversarial} or object detectors \citep{lee2019physical, liu2018dpatch}.
Adversarial patches can also be printed and placed in a real-world scene, but their performance is limited.
In all of these cases, the adversarial patches were crafted using a set of 2D photos of the target scene but were used in a 3D space; as a result, the flat patch produced would be unable to address the real-world challenges presented in \fig{challenges}.
Such challenges pushed to the development of methods to improve the robustness of adversarial perturbations in real world settings.
The use of the EOT framework, presented by \citet{athalye2018synthesizing}, to craft perturbations that are robust to specific transformations offered a solution.
This framework builds a set of images by transforming the original sample, using parameters that were randomly sampled from the transformation function's distribution, and use the set to craft a robust adversarial perturbation.
For example, to craft an adversarial example that is robust to rotation, the training set includes samples of the original image rotated at different angles.
To demonstrate the framework's abilities, EOT was used to perturb the texture of a digital 3D object, which was later printed in the real world using a color 3D printer.
The study focused on perturbing the texture of a single digital object and thus did not consider the challenges of a complex realistic scene (presented in \fig{challenges}).
The EOT framework was demonstrated on a premodeled digital 3D object that was later printed in the real world, however an attacker is more likely to target an existing real-world scene.
Moreover, the implementation of the rendering process and use of a color 3D printer complicate the attack, making it less accessible to inexperienced attackers.
While some studies examined how the representation of 3D data affects the model's robustness to adversarial perturbations \citep{su2018deeper}, others suggested methods for perturbing the structure (mesh) or texture of 3D objects \citep{Xiang_2019_CVPR, xiao2019meshadv}.
However, the methods focused on manipulating a single 3D object, thus failing to consider the unique characteristics of a complex real-world scene.
\citet{zeng2019adversarial} showed a more realistic approach which perturb different elements of a digital 3D scene to gain more insights about attacks in the real world.
Although using 3D objects to create realistic adversarial examples seems promising, the studies mentioned above used premade digital objects, and did not target an existing complex real-world scene or examine the attack outside of the digital space.
Our study aims to use the techniques mentioned above to craft adversarial patches that are robust in real-world settings.
We also improve upon previous research by allowing the attacker to work in a flexible environment that simulates the real world.
By creating a digital replica of the target scene, we give the attacker control of every element in the scene.
Then, the attacker can effectively utilize EOT by transforming the digital scene to improve the adversarial patch's robustness to the same transformations in the real world.
\section{Suggested Approach}
\subsection{Assumptions and Threat Model}
We assume an attacker that wants to add a sticker to an existing real-world scene to hide an object from an object detection system that uses a deep learning model.
By doing so, all photos of the scene will result in a false prediction by the neural network.
We assume the attacker has complete knowledge of the target neural network (e.g., parameters, architecture) but not the other system components (e.g., camera).
Moreover, the attacker can examine the real-world scene to create a 3D digital replica of it, such that the neural network classifies the rendered images of the replica as the original class.
Finally, we assume that the attacker has physical access to the actual scene to add the patch in the real world.
While this study demonstrates an attack under white-box settings, our framework can use black-box attack methods to craft an adversarial patch under more restrictive settings.
\subsection{Framework Overview}
We suggest the following framework for crafting an adversarial patch for an existing real-world scene.
\fig{framework} presents the six steps of the framework: model, render, combine, craft, evaluate, and apply.
The first five steps are performed in the digital space, while in the sixth step, the adversarial patch is transferred from the digital space to the real world.
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{images/framework.png}
\caption{
The framework’s steps:
1) Model a digital replication for the real-world scene.
2) Render 2D images with realistic transformations.
3) Combine the output of the rendering process into differentiable images.
4) Craft the adversarial patch.
5) Examine the adversarial patch in the digital space and improve the attack if needed.
6) Apply the patch to the physical scene.
}
\label{fig:framework}
\end{figure*}
First, the attacker uses 3D modeling techniques to create a 3D digital replication of the real-world scene and adds an empty digital patch object to the replica; in the following steps the texture of the empty object will be perturbed to craft the adversarial patch.
Second, the digital replica is rendered into 2D images, such that each image is a rendering of the scene under real-world transformations (e.g., rotation, a change in the lighting).
The rendering process results in two outputs: a background image of the scene (without the patch) and the properties of the patch (all pixels that include the patch).
Third, the attacker uses a differentiable method to combine the background and patch properties into one image to create a set of the scene’s images.
This way, the attacker can use a wide variety of external rendering tools without implementing a differentiable rendering process or approximating the gradients of a non-differentiable rendering.
Fourth, the images are then used to perturb the patch’s texture, using any method of crafting an adversarial perturbation.
Fifth, the attacker adds the adversarial patch to the scene’s digital replica, renders images of the scene, and feeds them to the neural network.
As a result, the attacker can examine the patch’s effect on the scene and improve it if needed.
Finally, the attacker prints the final patch on a sticker and adds it to the real-world scene.
\subsubsection{Modeling a Digital Replica of the Real-World Scene}
The first step is to create a 3D digital replica that approximates the real-world target scene.
To build this replica, the attacker can choose any tool (e.g., Blender, Maya) or resources (e.g., use free 3D objects, buy 3D designs, use a 3D scanner).
Then the attacker adds an empty 3D object to the digital scene to serve as the adversarial patch in the replica.
In this step, the attacker should consider the following elements: the objects in the scene, the ambient characteristics, and the adversarial patch.
First, the replica is built from digital 3D objects that represent the objects in the real-world scene.
Then, the scene’s ambient characteristics (e.g., light sources, smoke) are added to the replica to improve the replica's similarity to the real world.
Finally, the digital patch object is added to the scene according to its expected location in real life.
Decisions regarding the location, shape, and size of the patch should be made based on the attacker’s goals and the camera’s expected location.
The modeling stage is affected by the expertise and resources of the attacker.
Our findings show that a successful attack can be launched by roughly approximating the target scene, however when more realistic replications are modeled, the success rate will likely increase.
This finding demonstrates a trade-off between the attacker’s effort and the success rate, which can be used to balance the goals and capabilities of the attacker.
\subsubsection{Rendering 2D Images with Realistic Transformations}
The EOT process uses a collection of images with different transformations (“views”) to craft a single perturbation that fools the target neural network for all views.
Unlike past studies that could only use EOT with a limited number of transformation functions, our framework's design allows the attacker to transform any of the scene’s properties, including the scene’s objects (one or more), ambient characteristics, camera view, and more.
As presented by \citet{athalye2018synthesizing}, flexibility in choosing the transformations results in a better adversarial perturbation.
Therefore, in this step, the attacker aims to create a collection of views, such that each view represents the digital replica under a different set of transformations.
We define $T = \{ T_1,...,T_k \}$ such that each $T_i$ $(1 \leq i \leq k )$ is a distribution of transformation functions $t_i$ on the digital replica (i.e., the scene).
For example, a transformation function that rotates the scene through an angle $\theta$ around the x-axis can be sampled from $U(30,90)$; hence, $\theta \in [30,90]$.
Similarly, we define $C$ as the distribution for the transformation functions on the digital replica's camera (e.g., the camera's position).
For each set of transformation functions sampled from $T$ and $C$, the rendering process $R$ applies the transformations to the 3D digital replica $S$ with patch texture $P$ and outputs a 2D image of the 3D scene with the $t$ transformations from viewpoint $c$ (a view).
In this step, the attacker samples transformation functions multiple times to build a collection of views $X$:
\begin{equation}
\nonumber
X = \mathbb{E}_{t \sim T, c \sim C} [R (S,P,t,c)]
\end{equation}
\subsubsection{Combining the Rendering Output Buffers}
In the next step (i.e., crafting), the set of views is used to perturb the patch's texture $P$ to create the adversarial patch.
As shown in previous studies, crafting the adversarial perturbation is usually done by solving an optimization problem by calculating or approximating the attack’s loss gradient concerning the patch.
Because our framework allows the use of external rendering tools, an attacker can use a non-differentiable rendering process to craft $X$.
As a result, the attacker cannot calculate $\frac{\partial X}{\partial P}$, those preventing him/her from using gradient-based methods for crafting adversarial perturbations.
Past studies overcome this issue by implementing the rendering process as a differentiable part of the attack \citep{athalye2018synthesizing, Xiang_2019_CVPR} or by approximating the loss gradient \citep{xiao2019meshadv, zeng2019adversarial}.
However, these solutions limit the attacker from using external tools and increase the knowledge required to implement an attack.
Therefore, we suggest modifying the output of the rendering process, which is a simple configuration change that can be done in almost any rendering tool.
Then, instead of outputting a single image, the result of the rendering process is multiple buffers that can be split into two types: the background and the patch’s properties.
The background is an image of the scene without the patch, and the patch’s properties form a set of buffers with information that allows the patch to be added to the background in a realistic manner; combining the two creates an image of the scene.
Hence, the rendering stage results in a set $X=\{(b_i,p_i)\}_{i=1}^n$, where for sample $x_i=(b_i,p_i) \in X$, $b_i$ is the scene background, and $p_i$ are the patch properties.
Additionally, for a sample $x_i$, let $B(P,b_i,p_i)$ be a differentiable method that combines $P$, $b_i$, and $p_i$ into an image of the rendered scene.
Given the result of the rendering step $X$, the attacker builds a set of views $\tilde{X}$ that are differentiable by $P$:
\begin{equation}
\nonumber
\tilde{X} = \{ B(P,b_i,p_i): \forall (b_i,p_i) \in X \}
\end{equation}
An example of combining buffers into a scene image is presented in \fig{combining}.
The background buffer $b_i$ is an image of the scene without the patch, and the two patch properties buffers $p_i$ are the patch's texture map and lighting.
Each pixel in the texture map contains the coordinates of a pixel in the patch's texture $P$, thus allowing the framework to build an image of the patch's colors by sampling pixels from $P$ (shown in \fig{combining-sample}).
Then, the patch's color and lighting and the scene background are combined into one complete image of the complete scene (shown in \fig{combining-buffers}).
Since the image is built using simple operators, like sampling $P$, addition, and multiplication, the image $B(P,b_i,p_i)$ is differentiable by $P$.
\begin{figure*}
\centering
\null\hfill
\subfloat[Sampling the texture]{\label{fig:combining-sample}
\includegraphics[width=0.40\linewidth]{images/resample.PNG}}\hfill
\subfloat[Combining the buffers]{\label{fig:combining-buffers}
\includegraphics[width=0.40\linewidth]{images/buffers.PNG}}
\hfill\null
\caption{
Creating the images of the scene from the patch properties and background buffers.
(a) Sampling the texture of the patch to create an image of the patch's colors.
(b) Combining three buffers to create an image of the scene: the patch's colors, the shading of the patch, and an image of the background without the patch.
}
\label{fig:combining}
\end{figure*}
\subsubsection{Crafting the Adversarial Patch}
After combining the buffers into a set of images $\tilde{X}$, the attacker can use them to craft the adversarial patch.
Crafting an adversarial perturbation under white-box settings is usually done by feeding a benign input sample to the target learning model, calculating both the attack loss based on the model’s output and the loss gradient with respect to the input sample, and perturbing the original input sample.
The attack loss is used to ensure that the adversarial perturbation meets the attack's requirements, such as fooling the target model to output a specific result, limiting the perturbation size, etc.
In most attacks, the original input sample is gradually changed by an iterative optimization process that adds a small perturbation to the sample in each iteration.
While most attacks rely on a similar methodology, they differ by parameters and configuration (e.g., the attack loss, optimization process, number of iterations).
Given our framework's design, most methods for crafting adversarial perturbations can be used in this step to perturb the texture of the empty patch digital object into an adversarial patch.
Since combining the rendering output buffers results in a set of images $\tilde{X}$ that are differentiable by the patch texture $P$, the attacker can use any gradient-based method to craft an adversarial perturbation to $P$.
Most methods perturb the patch texture $P$ to fool a neural network $M$ by optimizing an objective function with a customized attack loss $\mathcal{L}$, e.g., finding $P$ such that $\argmin_P(\mathcal{L}(M(\tilde{X})))$.
The attacker can construct $\mathcal{L}$ to create a targeted attack, apply constraints, etc.
Then, since it is easy to calculate $\frac{\partial \tilde{X}}{\partial P}$, the attacker can also calculate $\frac{\partial \mathcal{L}(M(\tilde{X}))}{\partial P}$ to solve the optimization problem with a gradient-based optimizer (e.g., gradient descent).
As done in previous studies, the objective function and optimization method should be selected according to the attack's goal and target neural network.
While in this study we demonstrate our framework with a white-box attack, future work can replace the attack method we used in this step with a black-box attack (such as \citep{chen2017zoo, chen2020hopskipjumpattack}) to craft an adversarial patch under more restrictive settings.
\subsubsection{Examining the Patch in the Digital Space}
Next, by adding the patch to the digital replica, the attacker can simulate the adversarial patch’s effect on the neural network’s prediction concerning the real-world scene.
Examining the adversarial patch in the real world requires both the attacker's presence at the original scene and the performance of actions that could be considered abnormal thereby exposing the attacker to the risk of detection.
Using the digital replica to simulate the real-world scene allows the attacker to identify potential problems and improve the patch in a controlled and safe environment.
Since the attacker controls the digital replica, he/she can simulate events that are challenging to control in real life (e.g., waning daylight, the presence of smoke).
Moreover, the attacker can use the digital replica to compare different adversarial patches, identify the most effective one, and improve the attack's success in the real world.
Based on the findings in this step, the attacker might choose to change the attack process, requiring modifications to the scene’s replica, the addition of new transformations to the rendering step, or changes to the attack’s optimization function.
Since the examination process is performed in the digital space, the attacker can improve and evaluate the patch as long as the attacker wishes.
\subsubsection{Applying the Patch to the Physical Scene}
Finally, the attacker creates the adversarial patch in the real world and adds it to the scene.
For instance, the patch can be printed on a sticker or a piece of paper using a home printer.
\section{Experimental Setup}
In this study, we implement the framework using free and open-source software to create an adversarial patch for a typical office scene in which a standard white mug is placed on a desk, as seen in \fig{mugs-rw}.
A webcam (Microsoft LifeCam VX-700) films the scene, and then we divide the video stream into photos, crop them into $299 \times 299$ pixel color images, and feed them to a state-of-the-art object classifier, InceptionV3 \citep{xia2017inception}.
This classifier was trained on the ImageNet dataset \citep{deng2009imagenet}, achieving 1-top accuracy of 78\% and 5-top accuracy of 93.9\% for valid input.
We validated that the scene is classified as the original class (i.e., Coffee Mug) for 100\% of the images rendered from the digital replica without the patch.
Additional information on the experiment setup, code, and parameters is available in \citep{author_repository}.
\begin{figure*}
\centering
\null\hfill
\subfloat[Real world]{\label{fig:mugs-rw}
\includegraphics[width=0.40\linewidth]{images/mug_rw.jpg}}\hfill
\subfloat[Digital replica]{\label{fig:mugs-digital}
\includegraphics[width=0.40\linewidth]{images/mug_digital.png}}
\hfill\null
\caption{
The target scene: a white mug placed on an office desk; (a) is a photo of the original real-world scene; and (b) is a rendering of our digital replica with the empty adversarial patch (blue strip) that was modeled using Blender.
}
\label{fig:mugs}
\end{figure*}
\subsection{Crafting the Adversarial Patch}
\subsubsection{Modeling}
To model the replica of the target scene, we use Blender \citep{blender}, a free creation software.
We start by approximating the objects in the real-world scene:
For the desk and the walls of the room, we use default cube shapes in Blender.
However, the other objects are more complex, and a professional expertise in 3D modeling is required to create them from scratch.
Therefore, we use a free, premade mug object \citep{coffeemugmodel} and add it to the digital replica.
Afterward, we add a yellow point light source that simulates the light bulb in the original office.
Then, to create the empty placeholder for the adversarial patch, we crop and edit a part of the mug's model \citep{coffeemugmodel} and the result around the mug's 3D object in the digital scene.
For each 3D object, we choose standard configurations for the materials and use textures made from photos of the real-world scene.
We note that the replica is modeled using online tutorials for beginners to produce a simple approximation of the real-world scene.
Therefore, while our digital scene lacks some realistic elements (\fig{mugs-digital}), it can be easily replicated by attackers and researchers with limited experience in 3D modeling.
\subsubsection{Rendering}
The rendering process is implemented with ModernGL \citep{moderngl}, based on the examples in the library’s repository, and uses the configurations and materials to portray the ambient elements in the rendered images (e.g., shadows).
Due to the attacker's assumed lack of knowledge about the camera, we estimate the camera's configurations (e.g., field of view).
Then, we render a set of views $X$, i.e., images of the scene under different transformations:
For the scene transformations $T$, we choose translation and rotation in the $x, y,$ and $z$ axes, and changes in the light color, and for the camera transformation $C$, we choose changes in the camera's position.
To determine the ranges that define each distribution of the different transformation functions, we examine both the digital replica and the real-world scene.
Based on our findings, we define the ranges according to possible changes in the real-world scene while ensuring that the mug and patch are visible in all views.
Additional information about the rendering step (including parameters and code) is available in \citep{author_repository}.
In this study, we explore two methods for sampling the parameters for the transformations: \textit{random sampling} and \textit{systematic sampling}.
Random sampling is commonly used in the EOT framework, in which a transformation function $t_i$ is randomly sampled from a uniform distribution $T_i \sim U(\alpha_i,\beta_i)$; hence, the parameter that defines $t_i$ is randomly sampled from the range $[\alpha_i,\beta_i]$.
Additionally, we examine a deterministic approach, which we refer to as systematic sampling, where the transformation functions $t_i$ are predefined by systematically sampling the function's parameter in constant even steps across $[\alpha_i,\beta_i]$.
For example, we want to create $l$ scene rotation functions $\{ t_{i,1}, ...,t_{i,l}\} \in T_i$ that are systematically sampled from the range $[\alpha_i,\beta_i]$.
Therefore, we predefine $\{ t_{i,1}, ...,t_{i,l}\}$ such that for each $1 \leq j \leq l$, the rotation function $t_{i,j}$ rotates the scene through an angle of $\theta_j = \alpha_i + \frac{j \cdot (\beta_i - \alpha_i)}{l}$.
After selecting the transformation functions, we build the set of views by rendering an image of the scene using every combination of those parameters for the different transformation distributions.
\subsubsection{Combining}
The rendering step outputs a set of views $X$ that is determined by the sampling method; hence, each view is defined by a set of scene and camera transformations.
Additionally, for each view, the rendering process outputs four buffers that can be split into two types: one buffer with background image and three buffers with the patch properties.
The patch information buffers include the texture mapping, lighting, and a mask that defines the parts of the patch object visible in the rendered image.
To build a differentiable image from the four buffers, we follow a similar process to the one presented in \fig{combining}:
We use the texture mapping to sample the colors from $P$, merge the result with the patch's lighting, and finally, use the mask to combine the background image with the patch.
The result is an image of the complete scene with the patch, similar to the one shown in \fig{mugs-digital}.
The building process is implemented using TensorFlow \citep{abadi2016tensorflow} and uses simple operations such as sample, add, and multiply tensors; as a result, the output image is differentiable by $P$.
We use this process for each view in $X$ to build a set of differentiable images $\tilde{X}$.
\subsubsection{Crafting}
To craft the adversarial patch, we follow previous studies and define an objective function with a customized attack loss:
\begin{equation}
\nonumber
\mathcal{L}(\tilde{X},P) = CE(\tilde{X}, y_{tg}) -\kappa \cdot CE(\tilde{X}, y_{og}) +\lambda \cdot TV(P)
\end{equation}
where $\kappa$ and $\lambda$ are tuning parameters, $y_{og}$ and $y_{tg}$ are the original and attack target class respectively, $CE$ is the cross-entropy loss, and $TV$ is the total variation.
Then, to perturb $P$, we use the Adam optimizer to solve the following optimization problem:
\begin{equation}
\nonumber
\argmin_P \{ \mathcal{L}(\tilde{X},P) \}
\end{equation}
The attack loss $\mathcal{L}$ uses the cross-entropy loss to cause the scene with the patch to be classified as $y_{tg}$ and not as $y_{og}$, while maintaining the smoothness of the patch by minimizing the total variation distance.
Examples of two adversarial patches are presented in \fig{patch}.
\begin{figure}[t]
\centering
\includegraphics[width=0.85\linewidth]{images/digital_patch.png}
\caption{
The systematic (upper) and random (bottom) adversarial patches for the Platypus target class.
}
\label{fig:patch}
\end{figure}
\subsubsection{Examining}
To examine the patch, we use the digital evaluation process, which is described in Section~\ref{sec:evaluation}.
Based on our findings, we perform several changes in the transformation ranges and updates to the 3D replica to improve it (e.g., creating realistic lighting by modeling the room as a box instead of the three visible walls).
The attack was initially designed for the Armadillo target class; later, we change the seed and initialization parameters and use the attack to craft patches for the Armadillo and nine additional target classes.
We use the same attack to perform a non-biased comparison between the patches for the different target classes.
However, it is more realistic to build a tailored attack for each target class, an approach we will explore in future research.
\subsubsection{Applying}
We print the patch on an A4 piece of paper using a Xerox WorkCentre 6605, manually crop it, and apply it to the mug using transparent adhesive tape.
\subsection{Evaluation Process}\label{sec:evaluation}
The ability to perform the same experiment multiple times under the same conditions and obtain similar results is essential for our study’s integrity and to enable future research to reproduce our work.
Therefore, we present a replicable two-step evaluation process, the first step of which takes place in the digital space; the second step takes place in the real world.
It is important to note that the evaluation process was designed for research purposes and is not part of our framework.
\subsubsection{Digital Space}
In the first step, the digital replica is used to simulate the evaluation in the real world, similarly to our framework's examination step (step 5 in \fig{framework}).
Here, we add the adversarial patch to the scene, render a set of images under the expected real-world settings and transformations (e.g., camera's position), send the images to the neural network, and analyze the predictions.
\subsubsection{Real-World}
To ensure that the real-world evaluation process is reproducible, we suggest using the evaluation setup that is presented in \fig{rw-eval}.
The setup includes a camera slider that can be placed at different distances from the scene and allows the camera to film the scene from predefined positions.
As shown in \fig{setup-closeup-platf}, the camera is placed on a spinning platform, which is located on top of a cart that moves the camera from side to side in front of the scene.
The position of the camera can be changed by spinning the platform to a specific angle, which is defined by marks on the gradations at the bottom of the platform.
Then, a screw is used to secure the camera in place to ensure that the camera's angle does not change during the experiment.
A motor spins a screw rod, which moves the cart across two metal rods at a constant speed.
We suggest defining small ranges on the slider, as shown in \fig{setup-closeup-cart}, from which the scene is filmed at different angles.
For example, after choosing a camera position, we identify the range on the slider in which the mug is visible to the camera; during the experiment, the cart moves only in the defined range to avoid filming irrelevant parts of the scene.
Each range should be defined based on the required observation area (from the center, left, or right sides of the slider), along with the corresponding position of the camera, and can be set by physically limiting the cart's movement or by configuring the motor's behavior.
Additionally, as shown in \fig{setup-closeup-slider}, the slider can be moved forward and backward across a grooved base to film the scene from different distances.
Finally, since the real-world scene might change, the location and position of each object, including the adversarial patch, must be marked, thus ensuring that the same scene can be re-evaluated.
By physically marking each configuration, the same actions can be performed in future experiments and thus, achieve similar results.
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{images/rw-eval-setup.jpg}
\caption{
Our real-world evaluation setup:
The position of the camera (1) is controlled by spinning a platform with gradations on its base (2).
The platform is located on a cart (3), which moves within predefined ranges (4) on a spinning screw rod (5).
The slider can be moved forward and backward into one of three predefined positions (6).
We can rebuild the scene using the markings for the mug's location and position (7) and the patch's location on the mug (8).
}
\label{fig:rw-eval}
\end{figure*}
\begin{figure*}
\centering
\null\hfill
\subfloat[Platform]{\label{fig:setup-closeup-platf}
\includegraphics[width=0.32\linewidth]{images/setup_closeup_platform.png}}\hfill
\subfloat[Cart and ranges]{\label{fig:setup-closeup-cart}
\includegraphics[width=0.32\linewidth]{images/setup_closeup_cart.png}}\hfill
\subfloat[Slider]{\label{fig:setup-closeup-slider}
\includegraphics[width=0.32\linewidth]{images/setup_closeup_slider.png}}
\hfill\null
\caption{
A closeup of the three main components of the real-world evaluation setup:
(a) the spinning platform for the camera, (b) the cart that moves the camera in front of the scene, and (c) the slider which can be placed at different distances from the scene.
}
\label{fig:setup-closeup}
\end{figure*}
\subsubsection{Our Evaluation Setup}
In this study, we use the the evaluation process described above, in both the digital space and real world, to evaluate the performance of the patches that we create using our framework.
\textbf{Digital space:}
We determine 3,360 different positions and locations for the camera around the object, based on the patch’s visibility in the digital replica; for each of them, we render an image and send it to the object classifier.
\textbf{Real world:}
We build the structure, as shown in \fig{rw-eval} and \fig{setup-closeup} and described above, with the following configuration:
The slider can be placed at one of three predefined distances from the scene (close, middle, and far).
We defined three ranges (left, center, and right) for each distance and mark each of the nine ranges using metal eye straps.
The location of the mug, the position of the mug's handle, and the place of the patch on the mug are marked, thus ensuring that the same scene is used throughout the various experiments.
Sketches and instructions for replicating our real-world evaluation setup are available in \citep{author_repository}.
\section{Results}
For each target class, we craft four patches: \textit{random, systematic, google,} and \textit{imagenet}.
The \textit{random} and \textit{systematic} patches are adversarial patches crafted by using our framework with the random or systematic sampling of the transformations during the rendering step.
The \textit{google} and \textit{imagenet} patches are made out of images of the target class that were obtained from a Google search or the ImageNet dataset respectively, and are used to ensure that the results are non-biased.
We note that the images used for the \textit{imagenet} patches were collected from the dataset that was used to train the classifier, Inception V3.
We also examine the classification results for a mug without a patch (\textit{clean}) and a patch with random pixel values (\textit{noise}).
For each target class and its four patches, we evaluate the percentage of images classified as the original class (Coffee Mug), target class, or other classes (out of the ImageNet dataset).
\subsection{Results for Evaluation in the Digital Space}
After changing the seed and initialization parameters, we craft \textit{random, systematic, google,} and \textit{imagenet} patches for the following classes: Armadillo, Hat, (bottle) Nipple, Platypus, Mask, Pencil Box, Syringe, Screw, Mousetrap, and Ladybug.
Then, we perform the evaluation process in the digital space for \textit{clean}, with the \textit{noise} patch, and each of the patches described.
All of the 3,360 images of the rendered replica for \textit{clean} and with the \textit{noise} patch are classified as the original class, Coffee Mug.
Table \ref{table:digital-results} summarizes the results for the remaining patches.
\begin{table*}[t]
\caption{
The classification results (percentage) in the digital space for the original (Og), target (Tg), and other (Ot) classes.
All of the images (100\%) of \textit{clean} and \textit{noise} are classified as the original class.
}
\begin{center}\small
\begin{tabular}{|l|ccc|ccc|ccc|ccc|}\hline
\multicolumn{1}{|l|}{} & \multicolumn{3}{c|}{Systematic} & \multicolumn{3}{c|}{Random} & \multicolumn{3}{c|}{Google} & \multicolumn{3}{c|}{Imagenet} \\ \hline
\makecell{Target\\Class} & Og & Tg & Ot & Og & Tg & Ot & Og & Tg & Ot & Og & Tg & Ot \\ \hline\hline
Armadillo & 0.3 & 99.5 & 0.2 & 0.8 & 98.5 & 0.7 & 98.1 & 0 & 1.9 & 98 & 0 & 2 \\ \hline
Hat & 0.3 & 99.6 & 0.1 & 0.1 & 99.5 & 0.4 & 99.5 & 0 & 0.5 & 99.1 & 0 & 0.9 \\ \hline
Nipple & 0.1 & 97.6 & 2.3 & 0.4 & 93.1 & 6.5 & 99.6 & 0 & 0.4 & 100 & 0 & 0 \\ \hline
Platypus & 1 & 97.5 & 1.5 & 1.3 & 98 & 0.7 & 98.1 & 0 & 1.9 & 96.4 & 0 & 3.6 \\ \hline
Mask & 1 & 99 & 0 & 6.9 & 92.9 & 0.2 & 99.7 & 0 & 0.3 & 99.7 & 0 & 0.3 \\ \hline
Pencil Box & 0 & 99.9 & 0.1 & 0.7 & 98.7 & 0.6 & 99.7 & 0 & 0.3 & 99.9 & 0 & 0.1 \\ \hline
Syringe & 0 & 98.9 & 1.1 & 0.8 & 95 & 4.2 & 99.2 & 0 & 0.8 & 97.6 & 0 & 2.4 \\ \hline
Screw & 0.1 & 99.8 & 0.1 & 0.6 & 97.6 & 1.8 & 100 & 0 & 0 & 97.9 & 0 & 2.1 \\ \hline
Mousetrap & 0 & 100 & 0 & 0.1 & 99.1 & 0 & 100 & 0 & 0 & 97.9 & 0 & 2.1 \\ \hline
Ladybug & 0 & 100 & 0 & 0 & 100 & 0 & 99.7 & 0 & 0.3 & 100 & 0 & 0 \\ \hline
\end{tabular}
\end{center}
\label{table:digital-results}
\end{table*}
We expected that the \textit{imagenet} patches, which were taken from the dataset that was used to train Inception V3, would act like an adversarial patch, thus causing the scene to be classified as the target class.
Although the scene is simple, the \textit{google} and \textit{imagenet} patches, with clear images of the target class, do not affect the classifier.
On average, 99\% of the rendered images of the replica with the \textit{google} and \textit{imagenet} patches are classified as the original class, and none of them are classified as the target class.
In contrast, more than 99\% and 97\% (on average) of the images are classified as the target class for the \textit{systematic} and \textit{random} patches respectively.
When comparing the two types of adversarial patches, the systematic sampling approach is significantly better at causing the scene to be classified as the target class than the random sampling approach (the p-value is 0.04 for a paired sample t-test).
\subsection{Results for Evaluation in the Real World}
For the real-world evaluation, we use the same patches we created in the digital space for the following target classes: Armadillo, Hat, Nipple, and Platypus.
Each of these patches, including the \textit{noise} patch, is printed on a piece of paper, cropped, and placed on the designated location on the coffee mug using transparent adhesive tape.
Then, we follow the setup presented in \fig{rw-eval} to take approximately 700 photos, from all nine ranges, of the real-world scene with each patch and \textit{clean}.
Similarly to the evaluation in the digital space, 100\% of the photos of the real-world scene for \textit{clean} and with the \textit{noise} patch are classified as the original class.
Table \ref{table:rw-results} summarize the results for the remaining patches.
\begin{table*}[t]
\caption{
The classification results (percentage) in the real world for the original (Og), target (Tg), and other (Ot) classes.
All of the photos (100\%) of \textit{clean} and \textit{noise} are classified as the original class.
}
\begin{center}\small
\begin{tabular}{|l|ccc|ccc|ccc|ccc|}\hline
\multicolumn{1}{|l|}{} & \multicolumn{3}{c|}{Systematic} & \multicolumn{3}{c|}{Random} & \multicolumn{3}{c|}{Google} & \multicolumn{3}{c|}{Imagenet} \\ \hline
\makecell{Target\\Class} & Og & Tg & Ot & Og & Tg & Ot & Og & Tg & Ot & Og & Tg & Ot \\ \hline\hline
Armadillo & 1.5 & 92.8 & 5.7 & 2.1 & 95.3 & 2.6 & 95.8 & 0 & 4.2 & 99.7 & 0 & 0.3 \\ \hline
Hat & 0 & 99.1 & 0.9 & 0.2 & 98.9 & 0.9 & 99.3 & 0 & 0.7 & 95.9 & 0 & 4.1 \\ \hline
Nipple & 0 & 98.7 & 1.3 & 0.7 & 92.7 & 6.6 & 100 & 0 & 0 & 99.1 & 0 & 0.9 \\ \hline
Platypus & 0.3 & 92.6 & 7.1 & 6.2 & 87 & 6.8 & 99.2 & 0 & 0.8 & 99.1 & 0 & 0.9 \\ \hline
\end{tabular}
\end{center}
\label{table:rw-results}
\end{table*}
The results for the real-world evaluation are similar to the results obtained in the digital space.
For all target classes, the scene with the \textit{google} and \textit{imagenet} patches is mainly classified as the original class (for 98.5\% of the photos on average) and is never classified as the target class.
Similarly, the scene with the \textit{systematic} and \textit{random} adversarial patches is mainly classified as the target class;
The average difference between the digital and real-world results is 5\%, and the actual difference never exceeds 7\%.
We further examine the patches with the greatest difference between the digital space and real-world results.
The \textit{google} patch for the Armadillo target class and the \textit{imagenet} patch for Hat are classified as the original class in 98.1\% and 99.1\% of the images respectively, but only in 95.8\% and 95.9\% of the photos (respectively) of the real-world scene.
Both patches are classified as the Candle class from the ImageNet dataset (other) for photos taken from positions in which the mug's handle is less visible; therefore, the misclassification might stem from the cylindrical shape of the mug without the handle.
We also observe that the \textit{random} patch for the Platypus class is less effective when the camera is located far from the scene, in both the digital space and the real world.
Since the digital space evaluation reveals its weaknesses, the attacker can utilize this information to improve the patch by rendering more views in which the camera is located far from the scene.
This is an example of how the attacker can identify the patch's flaws in advance, learn how to improve the patch, and increase the chances of a successful attack.
\section{Resilience to Unexpected Transformations}
In realistic scenarios, the attacker cannot control the real-world scene, which means that by the time the attacker returns to the scene with the patch, the scene may have changed.
As we discussed, the attacker can use our framework to improve the patch's robustness to predictable changes, yet this is not the case for unexpected major transformations to the scene (e.g., the sudden removal of an object).
If the patch is only effective in the modeled scene and under expected transformations, then our framework is less feasible in real-world settings.
Therefore, we want to examine how unexpected transformations to the real-world scene affect the ability of adversarial patches created using our framework to fool the target neural network.
To do so, we use seven new transformations to change the real-world scene.
The adversarial patches are not expected to be robust to those transformations, since they are not part of the digital replica.
As shown in \fig{mugs-rw}, the original scene contains a white coffee mug with a handle on the right side and a patch placed in the middle of the mug, which is placed on an office desk.
We defined the seven unexpected transformations as \textit{up, down, red, wood, color, shape}, and \textit{flipped}.
\textit{Up} and \textit{down} are transformations to the patch's location on the mug: the patch is placed on the top and bottom of the mug, respectively.
We also change the surface on which the mug is placed to a red circle (\textit{red}) and a wooden mat (\textit{wood}).
For \textit{color} and \textit{shape}, we replace the original white mug:
for \textit{color}, we use a mug of the same shape but with a different color (a dark background with colorful illustrations), while for \textit{shape}, we use a mug of a similar color (light gray instead of white) which has a different shape.
The \textit{shape} mug is shorter and cone-shaped, and has a smaller handle which is located at a higher position on the mug.
Finally, for \textit{flipped}, we rotate the mug by $180^\circ$.
In this experiment, for each transformation (some of the transformations are shown in \fig{transform}), we change the scene and perform the real-world evaluation process with the two adversarial patches (\textit{systematic} and \textit{random}) used for the Nipple target class.
Table \ref{table:transformations} presents how each unexpected transformation affects the classification results of the real-world scene with the adversarial patches.
\begin{figure*}
\centering
\null\hfill
\subfloat[Different mugs]{\label{fig:transform-mugs}
\includegraphics[width=0.40\linewidth]{images/transformations_mugs.jpg}}\hfill
\subfloat[Red mat]{\label{fig:transform-red}
\includegraphics[width=0.40\linewidth]{images/transformations_red.jpg}}
\hfill\null
\caption{
Unexpected transformations to the real-world scene; (a) the three coffee mugs on the wooden mat used for \textit{wood} (from left to right: \textit{shape}, the original mug, and \textit{color}), and (b) \textit{red}: the original coffee mug with an adversarial patch on a red mat.
}
\label{fig:transform}
\end{figure*}
\begin{table*}[t]
\caption{
The results of a real-world evaluation for the Nipple patches with seven unexpected transformations added to the scene.
The results indicate the percentage of photos classified as the original, target, or other classes.
}
\begin{center}
\begin{tabular}{|l||c|c|c||c|c|c|}\hline
\multicolumn{1}{|c||}{} & \multicolumn{3}{c||}{Systematic} & \multicolumn{3}{c|}{Random} \\ \hline
Transformation & Original & Target & Other & Original & Target & Other \\ \hline\hline
Up & 0.6 & 88.9 & 10.5 & 2.2 & 77.9 & 19.9 \\ \hline
Down & 2.7 & 83.7 & 13.6 & 2.7 & 87.8 & 9.5 \\ \hline
Red & 1.9 & 84.6 & 13.5 & 2.7 & 76.5 & 20.8 \\ \hline
Wood & 5.8 & 68.9 & 25.3 & 2.4 & 64.7 & 32.9 \\ \hline
Color & 6.5 & 39.1 & 54.4 & 5.5 & 28.1 & 66.4 \\ \hline
Shape & 0 & 87.1 & 12.9 & 1.2 & 71.7 & 27.1 \\ \hline
Flipped & 2.7 & 85.8 & 11.5 & 4.3 & 55.1 & 40.6 \\ \hline
\end{tabular}
\end{center}
\label{table:transformations}
\end{table*}
The unexpected transformations cause the scene to be classified as the target class less often and more often as other classes.
Additionally, the scene is classified as the original class in only 2.9\% of the photos (on average) and in no more than 6.5\% of the photos.
This shows that the neural network fails to identify the mug even when there are major unexpected changes in the scene.
It seems that the patches are robust to changes in the patch's location on the mug; in \textit{up} and \textit{down}, the scene is classified as the original class for no more than 2.7\% of the photos.
However, changes in the object that the mug is placed on show mixed results; while the patches are more robust to \textit{red}, \textit{wood} is classified as the target class in less than 70\% of the photos, and in just up to 5.8\% of the photos as the original class.
We believe that the difference may stem from the shape of the mats: unlike the wooden mat, the red mat and the desk in the original scene have a solid shape.
Similarly, replacing the original mug also shows mixed results; both patches perform the worst on \textit{color} but perform well on \textit{shape}.
For \textit{color}, the scene is classified as the target class in less than 40\% of the photos and as the original class in up to 6.5\% of the photos.
We assume that the use of a mug with an inconsistent design (i.e., many colorful illustrations) reduces the performance of the patches.
Although the patches are the least robust to \textit{color}, the neural network still fails to identify the scene in more than 90\% of the photos.
In contrast, the scene with \textit{shape} is classified as the target scene in 71.7\% and 87.1\% of the photos for the \textit{random} and \textit{systematic} patches respectively, and the \textit{systematic} patch is never classified as the original class.
Additionally, we find that \textit{flipped} affects the adversarial patches differently: \textit{systematic} performs well, with 85.8\% of the photos classified as the target class, but for \textit{random}, only 55.1\% of the photos are classified as the target class.
Finally, in most cases, the \textit{systematic} patch performs better than the \textit{random} patch.
This experiment was designed to examine whether using a digital replica limits the attack to a specific model of the scene, thus causing the adversarial patch to become ineffective when unexpected changes occur in the real world.
To do so, we performed noticeable, yet realistic, transformations to the original real-world scene: changes in an object's location and position, the replacement of an object, and adding a new object to the scene.
The results show that although the transformations reduce the attacker's ability to control the neural network's prediction, the patches can still fool the neural network so that it misclassifies the scene.
On average, the scene is not classified as the original label in more than 97\% of the photos, which supports the feasibility of our framework.
Therefore, the attacker can improve the attack's robustness to real-world transformations by creating a 3D replica of the scene (as we proposed), suffering just a minor reduction in the patch's fooling ability if unexpected changes occur in the target scene.
\section{Conclusions}
In this work, we demonstrated how 3D modeling techniques and tools can be used to craft inexpensive adversarial patches that are robust to real-world transformations.
By creating a digital replica of the target scene, our method gives the attacker control of every aspect of the scene, including the objects, lighting, and more.
The replica simulates the real-world scene, thus allowing the attacker to test and improve the attack without the risk of detection.
We also demonstrate that such approach can improve the patch’s robustness to both expected and unexpected changes in the real-world scene.
Additionally, we present an evaluation process that enables other researchers to reproduce our experiments and validate our results.
We believe that such an evaluation process can be used in future studies and contribute to replicating, examining, and improving other real-world attacks.
In future work, we plan to improve the suggested framework by adding 3D elements (e.g., reflection and normal maps) to support more complex scenes.
Then, we will use it to tailor an attack for other domains, like the facial recognition domain, where our framework can improve individuals' privacy concerning such systems.
We also aim to create imperceptible perturbations that attract less attention.
Finally, our results suggest that using the EOT framework with systematic sampling might be better than random sampling, and we plan to perform additional experiments to examine this further.
\section*{Acknowledgment}
We gratefully acknowledge Matan Yesharim for his major contribution to the development of both the attack and evaluation process.
We also thank Boris Zadov for his professional expertise.
Finally, we extend a special thank you to Mathov Designs for designing and building the real-world evaluation setup and for allowing us to publish their designs which will assist the research community worldwide.
\bibliographystyle{elsarticle-num-names}
|
\section{Introduction}
In order to study the uniqueness and existence problems of the K\"ahler-Einstein metrics
on a Fano manifold $X$, Mabuchi (\cite{Ma}, \cite{BM}) introduced a useful energy functional ${\mathcal M}$,
called the $K$-energy or Mabuchi functional, on the space ${\mathcal H}$ of all the smooth K\"ahler potentials
in a given cohomology class $[\omega]$.
In other words, we can write the space as
$$ {\mathcal H}: = \{ \vp\in C^{\infty}(X); \ \ \omega+ dd^c\vp >0 \}. $$
It is observed that this functional ${\mathcal M}$ is convex along a smooth geodesic ${\mathcal G}$,
connecting with arbitrary two points in ${\mathcal H}$.
Moreover, if the Mabuchi functional is affine along a smooth geodesic ${\mathcal G}$,
then ${\mathcal G}$ must be generated by a holomorphic vector field.
In this case, we say that the Mabuchi functional is strictly convex along the geodesic ${\mathcal G}$.
It turns out that the Mabuchi functional ${\mathcal M}$ has also played an important role in the study of the
constant scalar curvature K\"ahler(cscK) metrics.
In particular, the convexity of ${\mathcal M}$ is crucial in the proof of the uniqueness of the cscK metrics (\cite{BB}, \cite{CLP}, \cite{CPZ}).
Moreover, the asymptotic behavior of this convex function ${\mathcal M}(\vp_t)$ along a geodesic ray is an invariant on a K\"ahler manifold $X$,
and it is proved (\cite{CC1}, \cite{CC2}, \cite{CC3}) that the manifold $X$ admits a cscK metric if and only if this invariant is non-negative.
However, one can not expect that there always exists a smooth geodesic connecting
two points in ${\mathcal H}$, due to the example in Darvas-Lempert (\cite{LD}).
What we can rely on in realty is the so called ${\mathcal C}^{1, \bar 1}$-geodesic (as the solution of certain homogenous complex Monge-Amp\`ere equation),
whose existence was guaranteed in the work of Chen (\cite{C00}).
The difficulty to deal with a ${\mathcal C}^{1, \bar 1}$-geodesic ${\mathcal G}$
is the lack of the regularities and the possible degeneracy on the geodesic ${\mathcal G}$.
Nevertheless, the Mabuchi functional ${\mathcal M}$ was proved to be
convex and continuous along a ${\mathcal C}^{1,\bar 1}$-geodesic,
by the work of Berman-Berndtsson (\cite{BB}) and also Chen-Li-P\u aun (\cite{CLP}).
Then people try to ask the question if the Mabuchi functional is also strictly convex
along such a geodesic.
If we allow that the boundary of ${\mathcal G}$ has merely ${\mathcal C}^{1,\bar 1}$-regularities,
then the answer is negative due to the example of Berman (\cite{Ber}).
On the other hand, the answer is affirmative (\cite{LL}), when
${\mathcal G}$ is connecting with two non-degenerate energy minimizers of ${\mathcal M}$.
In fact, the boundary of ${\mathcal G}$ must be smooth in the later case by the work of He-Zeng (\cite{HZ}).
Therefore, a geodesic ${\mathcal G}$ will always be assumed to have ${\mathcal C}^{1,\bar 1}$-regularities and
its boundary belongs to the space ${\mathcal H}$ through out this paper.
In order to circumvent the difficulties arising from a ${\mathcal C}^{1,\bar 1}$-geodesic ${\mathcal G}$,
we are appealing to using the so called $\ep$-geodesic ${\mathcal G}_{\ep}$ instead (\cite{C00}).
The $\ep$-geodesic ${\mathcal G}_{\ep}$ is a sequence of smooth approximation of ${\mathcal G}$,
satisfying some non-homogenous complex Monge-Amp\`ere equation.
This is easier to handle in the practical computation,
and we have proved in (\cite{CLP}) that the Mabuchi functional is \emph{almost convex}
along the $\ep$-geodesic.
However, the new difficulty is that the convergence ${\mathcal G}_{\ep} \rightarrow {\mathcal G}$
is merely weakly $L^p$ for all $p >1$.
In other words, it is even not clear to us if the
Mabuchi functional ${\mathcal M}(\vp_{\ep})$ along ${\mathcal G}_{\ep}$
converges to the Mabuchi functional ${\mathcal M}(\vp)$ along ${\mathcal G}$.
Therefore, we can not conclude the convexity of ${\mathcal M}$
by using the $\ep$-geodesic ${\mathcal G}_{\ep}$ in \cite{CLP}.
The new observation is that the convergence ${\mathcal M}(\vp_\ep) \rightarrow {\mathcal M}(\vp)$
is true, provided that the Mabuchi function ${\mathcal M}$ is affine along ${\mathcal G}$.
This first leads us to the following $L^2$-convergence of the fiberwise volume element of ${\mathcal G}_{\ep}$.
Write $\omega_{\ep}: = {\mathcal G}_{\ep}|_{X_t}$ and $\omega_{\vp}: = {\mathcal G}|_{X_t} $ on a fiber $X_t: = \{t\}\times X $
for any $t\in [0,1]$.
\begin{theorem}[Theorem (\ref{thm-l2-001})]
\label{thm-intr-001}
Suppose the Mabuchi functional ${\mathcal M}$ is affine along a ${\mathcal C}^{1,\bar 1}$-geodesic ${\mathcal G}$.
Then the fiber-wise volume element of the $\ep$-geodesic converges to the
volume element of the geodesic in the strong $L^2$ sense.
In other words, we have on each fiber $X_t$
\begin{equation}
\label{intr-000}
\frac{\omega_{\ep}^n}{\omega^n} \rightarrow \frac{\omega_{\vp}^n}{\omega^n}, \ \ \ \ep\rightarrow 0,
\end{equation}
under the $L^2$-norm, possibly after passing to a subsequence.
\end{theorem}
We emphasis that the $L^2$-convergence of the volume element (equation (\ref{intr-000})) may not be true in general.
Next, a slightly stronger condition than the affine Mabuchi functional will be introduced,
aiming to resolve the possible degeneracy on the geodesic ${\mathcal G}$.
Write the restriction of the Mabuchi functional along a $\ep$-geodesic ${\mathcal G}_{\ep}$ as
$${\mathcal K}_{\ep} (t) : = {\mathcal M} (\vp_{\ep} (t))$$
for all $t\in[0,1]$.
Then we say that the Mabuchi functional is \emph{$\ep$-affine} along the geodesic ${\mathcal G}$ if it satisfies
\begin{equation}
\label{intr-001}
\frac{d {\mathcal K}_{\ep}}{dt}|_{t=1} - \frac{d {\mathcal K}_{\ep}}{dt}|_{t=0} = O(\ep),
\end{equation}
for all $\ep>0$ small.
We note that ${\mathcal G}$ and ${\mathcal G}_{\ep}$ are uniquely determined if the boundary of ${\mathcal G}$ is given.
Thanks to a result in (\cite{BB}),
the Mabuchi functional must be
affine along the ${\mathcal C}^{1,\bar 1}$-geodesic ${\mathcal G}$ if it is $\ep$-affine.
Moreover, we proved the following result.
\begin{theorem}[Theorem \ref{thm-ch-001}]
\label{thm-intr-002}
Suppose the Mabuchi functional ${\mathcal M}$ is $\ep$-affine along a ${\mathcal C}^{1,\bar 1}$-geodesic ${\mathcal G}$.
Then ${\mathcal G}$ is uniformly fiberwise non-degenerate,
namely, there exists a uniform constant ${\kappa}_1 >0$ such that
$$ {\mathcal G}|_{X_t} > \kappa_1 \omega,$$
for almost everywhere $t\in [0,1]$.
\end{theorem}
There are three main \text{steps} of the proof for the above Theorem.
\textbf{Step (1)} is to figure out the so called \emph{gap phenomenon} of the geodesic ${\mathcal G}$,
which is first observed in our previous work (\cite{LL}).
\textbf{Step (2)} is to establish a kind of $W^{1,2}$-estimate for the volume element $\omega_{\ep}^n$,
which is provided from the $\ep$-affine condition.
\textbf{Step (3)} (see Proposition (\ref{prop-nd-001})) is to prove
that a non-negative function must have a positive lower bound
if it has the gap phenomenon and satisfies a certain partial $W^{1,2}$-estimate.
As an application of our Theorem (\ref{thm-intr-001}) and (\ref{thm-intr-002}),
we can prove (see Theorem (\ref{thm-app-001})) that
${\mathcal M}(\vp_{\ep})$ is not only converging to ${\mathcal M}(\vp)$
in the pointwise sense, but also in its complex Hessian, if the later is $\ep$-affine.
That is to say, we actually have
\begin{equation}
\label{intr-003}
\frac{d^2 {\mathcal K}_{\ep}}{dt^2}|_t \rightarrow 0,
\end{equation}
for almost everywhere $t\in [0,1]$.
Another application is that we can recover the strict convexity result
in (\cite{LL}), as stated before.
Moreover, if the manifold $X$ satisfies $c_1(X) = 0$
or $c_1(X) <0$, then we can utilize Chen's argument (\cite{C00}) to conclude the strict convexity of the Mabuchi functional,
provided the $\ep$-affine condition (see Theorem (\ref{thm-app-003})).
Therefore, we conjecture that the Mabuchi functional is strictly convex
along a ${\mathcal C}^{1,\bar 1}$-geodesic, if it is $\ep$-affine.
In fact, the convergence (equation (\ref{intr-003})) further implies
an $L^2$-estimate on $\bar\partial v_{\ep}$ for a sequence of smooth vector fields $v_{\ep}$ (see Theorem (\ref{thm-app-001})).
Unfortunately, there is still some difficulties to conclude the holomorphicity of $v_{\infty}$ as the limit of this sequence $v_{\ep}$.
Finally, it might be worthy to point out that
a geodesic ${\mathcal G}$ possibly possess more regularities than ${\mathcal C}^{1,\bar 1}$,
when the Mabuchi functional is $\ep$-affine along it.
Hopefully, we will see more examples about this fact, and
the regularity problem will be considered in our following projects.
\bigskip
\textbf{Acknowledgment: }
The author is very grateful to Prof. Chen and Prof. P\u aun who introduced this problem,
and have given continuous encouragement.
He also wants to thank Prof. Chengjian Yao, Dr. Jingchen Hu and Prof. Wei Sun for lots of useful discussion.
\bigskip
\section{Preliminary}
Suppose $\Sigma$ is an annulus in ${\mathbb C}$ with boundary, and $\pi$
is the holomorphic projection from the product space $Y: = X\times \Sigma$ to $X$.
Therefore, $Y$ is a compact complex K\"ahler manifold with boundary.
Let $\Phi$ be a quasi-plurisubharmonic function on $Y$ continuous up to the boundary.
Denote ${\mathcal G}$ by the closed positive $(1,1)$ current
$$ \pi^*\omega + dd^c\Phi $$
on $Y$.
We say that ${\mathcal G}$ is a geodesic in the space of K\"ahler potential,
if it is $S^1$-invariant in the argument direction of $\Sigma$, and satisfies the following
\emph{Homogeneous complex Monge-Amp\`ere} (HCMA) equation
\begin{equation}
\label{pre-0000}
{\mathcal G}^{n+1} = ( \pi^*\omega + dd^c\Phi )^{n+1} = 0,
\end{equation}
in a suitable sense on $Y$.
The boundary value of $\Phi$ is required to be in the space ${\mathcal H}$ of the smooth K\"ahler potentials.
We say that ${\mathcal G}$ is a geodesic connecting two points $\varphi_0, \varphi_1\in {\mathcal H}$
if
$$\Phi|_{X\times \{0 \}} = \varphi_0; \ \ \ \Phi|_{X\times\{1\}} = \varphi_1,$$
where we identify the annulus $\Sigma$ by a cylinder $[0,1]\times S^1$
via the standard diffeomorphism.
It is proved by Chen (\cite{C00}) that such a geodesic is unique with fixed boundary value,
and has the so called ${\mathcal C}^{1,\bar 1}$-regularities,
namely, writing ${\mathcal G}$ locally as
$$ g_{{\tau}\bar{\tau}} d{\tau}\wedge d\bar{\tau} + \sum_{{\alpha},\beta =1}^n ( g_{{\tau}\bar\beta} d{\tau}\wedge d\bar z^{\beta} + g_{{\alpha}\bar{\tau}}dz^{{\alpha}}\wedge d\bar{\tau} + g_{{\alpha}\bar\beta} dz^{{\alpha}}\wedge d\bar z^{\beta} ),$$
we have
$$ || g_{{\tau}\bar{\tau}} ||_{L^{\infty}} + \sum_{{\alpha},\beta =1}^n ( ||g_{{\tau}\bar\beta}||_{L^{\infty}} + ||g_{{\alpha}\bar{\tau}}||_{L^{\infty}} + ||g_{{\alpha}\bar\beta}||_{L^{\infty}} ) < +\infty. $$
In other words, there exist a uniform constant $C>0$ such that we have
$$ 0 \leq {\mathcal G} \leq C (\pi^* \omega + i d{\tau} \wedge d\bar{\tau} )$$
on $Y$.
Therefore the quasi-plurisubharmonic function $\Phi$ is of class $C^{1,{\alpha}}$ for any ${\alpha}\in (0,1)$, and
the wedge product ${\mathcal G}^{n+1}$ can be interpreted in the sense of Bedford and Talyor (\cite{BT}).
There is another way to describe the domain of ${\mathcal G}$, which will be useful in our later consideration on energy functionals.
Let $$ \Gamma: = \{ z\in {\mathbb C}; \ \ 0\leq Re z \leq 1 \}$$ be a strip domain in ${\mathbb C}$.
Then there is a holomorphic map from $\Gamma$ to $\Sigma$ as
$$ {\tau}(z): = e^z. $$
This is a branched cover of $\Sigma$,
and the inverse map is multi-valued in general.
Fortunately, ${\mathcal G}$ is $S^1$-invariant in the argument direction of $\Sigma$.
Therefore, we can select one of the branches as the inverse map of $\tau$.
Then the pull back of ${\mathcal G}$ under $\tau$ is the unique solution of the HCMA equation on $\Gamma\times X$,
which is independent of the imaginary part of $z$.
For this reason, the $z$-variable on $\Gamma$ can be taken as the complex coordinate of the cylinder
$$\mathcal{R}:= [0,1]\times S^1,$$
and then we can view that the solution ${\mathcal G}$ is actually defined on $\mathcal{R}\times X$.
Furthermore,
we can identify the complex variable
$z: = t + is $
with its real part $t$
for the same reason.
\subsection{The Mabuchi functional }
It was in introduced in (\cite{BM}) by Mabuchi the following functional
on the space ${\mathcal H}$
$${\mathcal M}: = \underline R {\mathcal E} - {\mathcal E}^{Ric\omega} + H, $$
where the constant $\underline R$ is the average of the scalar curvature
$$ \underline R = \frac{n c_1(X)\cdot [\omega]^{n-1}}{[\omega]^n}. $$
The energy functional ${\mathcal E}$ is defined for any $\vp\in{\mathcal H}$ as
$$ {\mathcal E} (\vp): = \frac{1}{n+1} \sum_{i=0}^n \int_X \varphi \omega^i \wedge \omega_{\vp}^{n-i}.$$
The twisted energy functional ${\mathcal E}^{{\alpha}}$ (by a closed smooth $(1,1)$ form ${\alpha}$)
is defined as
$$ {\mathcal E}^{{\alpha}} (\vp): = \sum_{i=0}^{n-1} \int_X \vp \omega^{i} \wedge \omega_{\vp}^{n-i-1}\wedge {\alpha}.$$
Finally, the entropy functional is
$$ H(\vp): = \int_X \left( \log\frac{\omega_{\vp}^n}{\omega^n} \right) \omega_{\vp}^n. $$
Suppose $\Phi$ is a $\pi^*\omega$-plurisubharmonic function on $Y$, which corresponds to a geodesic ${\mathcal G}$.
Then its restriction $\vp_{{\tau}}: = \Phi|_{X_{\tau}}$ on a fiber is actually a $\omega$-plurisubharmonic function on $X_{{\tau}}$ and
has the ${\mathcal C}^{1,\bar 1}$-regularities. It is observed that
${\mathcal M}$ can be defined on such functions.
Therefore, we can write the Mabuchi functional along the geodesic ${\mathcal G}$ as
$$ {\mathcal K}({\tau}): = {\mathcal M} (\vp_{\tau}), \ \ {\tau}\in\Sigma.$$
Next, we introduce the following modified versions of the Mabuchi functional.
Suppose $\Psi (\tau, \cdot) = \psi_{{\tau}}(\cdot)$ is a locally bounded singular metric on the relative canonical bundle $K_{Y/\Sigma}$,
and then $-\psi_{{\tau}}$ is a metric on the anti-canonical line bundle $-K_{X_{\tau}}: = \bigwedge^n TX_{{\tau}}$.
Therefore, the following is a measure on $X$
$$ \mu: = e^{\psi_{{\tau}}}.$$
We note that $\mu$ is absolutely continuous with respect to the Lebesgue measure.
Now define the following energy function on $\Sigma$ as
$$ {\mathcal K}^{\Psi}({\tau}) : = \underline R {\mathcal E} (\vp_{{\tau}}) - {\mathcal E}^{Ric\omega} (\vp_{{\tau}}) + \int_{X} \log\left(\frac{e^{\psi_{{\tau}}}}{\omega^n} \right) \omega_{\vp_{\tau}}^n. $$
This energy function will equal to ${\mathcal K}({\tau})$, if $\Psi$ is the (unbounded) metric defined by $\omega_{\vp_{{\tau}}}^n$.
For any large constant $A$,
Berman-Berndtsson (\cite{BB}) introduced the
following \emph{A-truncated Mabuchi functional} along the geodesic ${\mathcal G}$.
$$ {\mathcal K}^{\Psi_A}({\tau}):= \underline R {\mathcal E} (\vp_{\tau}) - {\mathcal E}^{Ric\omega} (\vp_{\tau}) + \int_{X} \log\left( \max\left\{ \frac{\omega^n_{\vp_{\tau}}}{\omega^n} , \frac{h_A}{\omega^n} \right\} \right) \omega_{\vp_{\tau}}^n, $$
where $h_A: = e^{\chi -A}$,
and $\chi$ is a fixed continuous metric on $K_{Y/\Sigma}$ satisfying
$$ dd^c\chi \geq k_0 ( \pi^*\omega + dd^c \Phi), $$
for some positive integer $k_0$.
The values of these energy functions ${\mathcal K}, {\mathcal K}^{\Psi}, {\mathcal K}^{\Psi_A}$
do not depend on the argument part of $\Sigma$,
and hence we can view that they are actually defined
on the cylinder ${\mathcal R} = [0,1]\times S^1$ by our previous discussion.
Moreover, it is proved (\cite{BB}) that ${\mathcal K}^{\Psi_A}$ is a convex function,
and it converges to ${\mathcal K}$ as $A\rightarrow \infty$ by the dominated convergence theorem.
Eventually, we conclude the following convexity result (\cite{BB}, \cite{CLP}).
\begin{theorem}
\label{thm-pre-0000}
The Mabuchi functional ${\mathcal M}$ is convex and continuous along a ${\mathcal C}^{1,\bar 1}$-geodesic ${\mathcal G}$.
That is to say, ${\mathcal K}$ is an $S^1$-invariant, convex and continuous function on ${\mathcal R}$.
\end{theorem}
\section{Energies on the $\ep$-geodesics}
Suppose we have two points $\vp_0, \vp_1\in {\mathcal H}$.
For each $\ep>0$ small enough,
there exists a smooth strictly $\pi^*\omega$-plurisubharmonic function $\Phi_{\ep}$ on $\Gamma\times X$ satisfying
\begin{equation}
\label{int-001}
{\mathcal G}_{\ep}^{n+1} = ( \pi^*\omega + dd^c \Phi_{\ep})^{n+1} = \ep \sqrt{-1}dt\wedge d\bar t \wedge \omega^n,
\end{equation}
with boundary conditions
$$\Phi_{\ep}(0, \cdot) = \vp_0 (\cdot);\ \ \ \Phi_{\ep}(1,\cdot) = \vp_1(\cdot).$$
Then we say that ${\mathcal G}_{\ep}: = \pi^*\omega + dd^c \Phi_{\ep} $ is the $\ep$-geodesic
connecting with $\vp_0$ and $\vp_1$.
It is proved (\cite{C00}) that the $\ep$-geodesic ${\mathcal G}_{\ep}$
is uniformly bounded in the ${\mathcal C}^{1,\bar 1}$-norm.
Moreover, we have known that $\Phi_{\ep} \rightarrow \Phi $ in $C^{1,{\alpha}}$-norm for any ${\alpha}\in (0,1)$,
and also in weakly $W^{2,p}$-norm for all $1 < p < \infty$.
In general, we write the restriction of the geodesic potential on each fiber $X_t: = \{ t\}\times X$ as
$$ \vp(t, \cdot): = \Phi|_{X_t}; \ \ \ \omega_{\vp}: = {\mathcal G}|_{X_t} = \omega+ dd^c \vp. $$
Similarly, we have for the $\ep$-geodesic potential
$$ \vp_{\ep}(t, \cdot): = \Phi_{\ep}|_{X_t}; \ \ \ \omega_{\ep}: = {\mathcal G}_{\ep}|_{X_t} = \omega + dd_X^c \vp_{\ep}.$$
Then a standard computation shows the following equation
\begin{equation}
\label{int-002}
( \pi^*\omega + dd^c \Phi_{\ep})^{n+1} = \rho_{\ep} \sqrt{-1}dt\wedge d\bar t \wedge \omega_{\ep}^n,
\end{equation}
where $\rho_{\ep}: = \rho_{\ep}(\vp_{\ep}) = g_{t\bar t} - g^{\bar\beta{\alpha}} g_{{\alpha} \bar t} g_{t\bar\beta} $,
and hence the $\ep$-geodesic equation (\ref{int-001}) can be re-written as
\begin{equation}
\label{int-003}
\rho_{\ep}(\vp_{\ep}) = \ep \frac{\omega^n}{ \omega_{\ep}^n}.
\end{equation}
\subsection{Energy and entropy}
Denote ${\mathcal K}_{\ep}: \Gamma \rightarrow \mathbb{R}$ by the restriction of the Mabuchi functional ${\mathcal M}$ to the $\ep$-geodesic as
$$ {\mathcal K}_{\ep} (t): = \underline{R}{\mathcal E} (\vp_{\ep}) - {\mathcal E}^{Ric\omega} (\vp_{\ep}) + \int_X \log\frac{\omega^n_{\ep}}{\omega^n} \omega^n_{\ep}.$$
Then the $A$-truncated Mabuchi functional along the $\ep$-geodesic can also be introduced as
$$ {\mathcal K}_{\ep, A} (t): = \underline{R}{\mathcal E} (\vp_{\ep}) - {\mathcal E}^{Ric\omega} (\vp_{\ep}) + \int_X \log\max \left\{ \frac{\omega_{\ep}^n}{\omega^n}, \frac{h_A}{\omega^n} \right\} \omega_{\ep}^n,$$
where $ h_A : = h_{\ep, A} = e^{\chi_{\ep} -A} $ is a smooth volume element whose associated curvature is greater than $-C{\mathcal G}_{\ep}$ for some fixed positive constant $C$.
More precisely, we construct this auxiliary element as follows.
Let $\chi_0$ be a smooth metric on the line bundle $K_{X}$,
and $k_0$ be a positive number such that $dd^c \chi_0 + k_0\omega > 0$.
Then we set $\chi_{\ep}: = \pi^*\chi_0 - k_0 \Phi_{\ep}$, and hence
$$ dd^c\chi_{\ep} = \pi^* \chi_0 + k_0 \pi^* \omega - k_0 (\pi^*\omega+ dd^c \Phi_{\ep}) \geq -k_0 {\mathcal G}_{\ep}. $$
We emphasis that the sub-index $\ep$ in the notation $h_{A}$ is omitted, since
$\chi_{\ep}$ is uniformly bounded in its $C^{1, \bar 1}$-norm, and converges uniformly to
$\chi: = \pi^*\chi_0 - k_0 \Phi $
in $C^{1,{\alpha}}$-norm for any ${\alpha}\in(0,1)$.
Therefore, this omission will be harmless for our later argument.
We note that ${\mathcal K}_{\ep, A}$ is continuous in $t$ by the construction.
Moreover, its complex Hessian can be computed in sense of local currents.
Suppose $v$ is a locally compact supported smooth test function on $\Gamma$
\begin{eqnarray}
\label{int-004}
\langle dd^c {\mathcal K}_{\ep, A}, v \rangle &=& \underline{R}\int_{\Gamma\times X} v (\pi^*\omega + dd^c\Phi_{\ep})^{n+1}
\nonumber\\
&-& \int_{\Gamma\times X} v (\pi^*\omega + dd^c\Phi_{\ep})^{n}\wedge \pi^* Ric(\omega)
\nonumber\\
&+& \int_{\Gamma\times X} v dd^c \left( \max \left\{ \log\frac{\omega_{\ep}^n}{\omega^n}, \log\frac{h_A}{\omega^n} \right\} \right) \wedge (\pi^*\omega + dd^c\Phi_{\ep})^{n}.
\end{eqnarray}
In other words, if we take the fiberwise integral $\int_{X_t}$ as the push forward operator acting on the currents from $\Gamma\times X$ to $\Gamma$,
we obtain the following equation.
\begin{eqnarray}
\label{int-005}
dd^c {\mathcal K}_{\ep, A}(t) &=& \frac{\underline{R}}{n+1}\int_{X_t} {\mathcal G}_{\ep}^{n+1} - \int_{X_t} Ric(\omega) \wedge {\mathcal G}_{\ep}^n
\nonumber\\
&+& \int_{X_t} dd^c \left( \max \left\{ \log\frac{\omega_{\ep}^n}{\omega^n}, \log\frac{h_A}{\omega^n} \right\} \right) \wedge {\mathcal G}_{\ep}^{n}.
\end{eqnarray}
On the one hand, we have seen locally
\begin{equation}
\label{int-006}
\frac{1}{n+1}\left(\Delta_{{\mathcal G}_{\ep}} \log{h_A} \right) {\mathcal G}_{\ep}^{n+1} = dd^c \log{h_A} \wedge{\mathcal G}_{\ep}^{n} \geq - k_0 {\mathcal G}_{\ep}^{n+1}.
\end{equation}
Next,
we denote the function $f_{\ep}: \Gamma\times X \rightarrow \mathbb{R}$ by the equality
$$ \left. \frac{\omega^n_{\ep}}{\omega^n} \right|_{X_t} = e^{f_{\ep} (t, \cdot)},$$
and introduce the following sub-level set
$$ P_{\ep, A}: = \left\{ (t,z)\in \Gamma\times X; \ \ \ f_{\ep}(t,z) > \log \frac{h_{ A}}{\omega^n} (t,z) \right\}. $$
It is proved in (\cite{CLP}) that there exists a positive number $c_A$ such that the following inequality is satisfied in $P_{\ep, A}$ (see more details in the following sections).
\begin{equation}
\label{int-007}
\left( dd^c f_{\ep} - Ric(\omega) \right) \wedge {\mathcal G}_{\ep}^n \geq -c_A {\mathcal G}_{\ep}^{n+1}.
\end{equation}
We emphasis that $c_A$ only depends on the constant $A$, the background metric $\omega$ and the uniform upper bound of $\omega_{\ep}$,
and it can be assumed to be increasing in $A$.
Hence we have the following estimate locally in $P_{\ep, A+1}$.
\begin{equation}
\label{int-008}
\left( \Delta_{{\mathcal G}_{\ep}} \log\omega_{\ep}^n \right) {\mathcal G}_{\ep}^{n+1} \geq -c_{A+1} {\mathcal G}_{\ep}^{n+1}.
\end{equation}
By a Theorem proved by Greene-Wu (see Lemma (5.2), \cite{CLP}),
we can establish the following inequality in an open neighbourhood of each point on $\Gamma\times X$.
\begin{equation}
\label{int-009}
\left( \Delta_{{\mathcal G}_{\ep}} \max\{ \log\omega_{\ep}^n, \log h_A \} \right) {\mathcal G}_{\ep}^{n+1} \geq -c'_{A} {\mathcal G}_{\ep}^{n+1},
\end{equation}
where $c'_A: = \max\{ c_{A+1}, (n+1)k_0 \}$.
Thus we infer the inequality
\begin{equation}
\label{int-010}
\left( dd^c \max\left\{ \log\frac{\omega_{\ep}^n}{\omega^n}, \log\frac{h_A}{\omega^n} \right\} - Ric \omega \right) \wedge {\mathcal G}_{\ep}^n \geq - c'_A {\mathcal G}_{\ep}^{n+1},
\end{equation}
globally on $\Gamma\times X$.
Combing this estimate with equation (\ref{int-005}), we have the following result.
\begin{theorem}[Chen-Li-P\u aun]
\label{int-thm-001}
For each positive number $A$, there is a uniform constant $C_A > 0$ such that the function
\begin{equation}
\label{int-0091}
\widetilde{{\mathcal K}}_{\ep, A}: = {\mathcal K}_{\ep, A}(t) - \ep C_A t(1-t)
\end{equation}
is convex and continuous on $[0,1]$ for each $\ep>0$ small.
\end{theorem}
One may expect that ${\mathcal K}_{\ep, A}$ converges to the energy ${\mathcal K}$.
If so, then we can directly conclude the convexity of ${\mathcal M}$ by Theorem (\ref{int-thm-001}).
The convergence of the energy parts of ${\mathcal M}$, i.e. ${\mathcal E}(\vp_{\ep}) \rightarrow {\mathcal E}(\vp)$ and ${\mathcal E}^{{\alpha}}(\vp_{\ep}) \rightarrow {\mathcal E}^{{\alpha}}(\vp)$
as $\ep\rightarrow 0$, is indeed true (\cite{BB}, \cite{LL}).
However, there is a rather severe difficulty:
the convergence of the fiber-wise volume element
$$ \frac{\omega_{\ep}^n}{\omega^n} \rightharpoonup \frac{\omega_{\vp}^n}{\omega^n} $$
is only known to be in the weakly $L^p$ sense for any $1<p< \infty$.
Unfortunately, this is not enough to conclude the convergence of the entropy functional $H(\vp_{\ep})$.
Nevertheless, we have the following estimate on the entropy functionals.
Recall that the entropy functional along a geodesic ${\mathcal G}$ is defined as
$$ H(\vp): = \int_X {\mathfrak{f}}(\vp) \log {\mathfrak{f}}(\vp) \cdot \omega^n, $$
where ${\mathfrak{f}}(\vp): = \omega_{\vp}^n / \omega^n$,
and its truncated version is
$$ H_{A}(\vp): = \int_X {\mathfrak{f}} (\vp) \log {\mathfrak{f}}_{A} (\vp) \cdot \omega^n, $$
where
$$ {\mathfrak{f}}_A (\vp): = \max \left\{ \frac{\omega_{\vp}^n}{\omega^n}, \frac{h_A}{\omega^n} \right\}. $$
As before, we omit the sub-index $\ep$ in the definition of $f_A$ as in $h_A$'s, and hope that this will be clear from the context.
Then the energy function ${\mathcal K}_{\ep, A}$ can be re-written as
\begin{equation}
\label{int-012}
{\mathcal K}_{\ep, A} (t) = \underline{R}{\mathcal E} (\vp_{\ep}) - {\mathcal E}^{Ric\omega} (\vp_{\ep}) + H_A(\vp_{\ep}).
\end{equation}
Finally we state the following lower semi-continuity type property (see Lemma (4.8), \cite{CLP}) for the truncated entropy functionals.
\begin{lemma}
\label{lem-int-001}
There exists a function $\eta: \mathbb{R}^+ \rightarrow \mathbb{R}^+$
with $\eta(x) \rightarrow 0$ as $x\rightarrow +\infty$ such that we have
\begin{equation}
\label{int-013}
\liminf_{\ep\rightarrow 0} H_A (\vp_{\ep}) \geq H_A (\vp) - \eta(A),
\end{equation}
for all $A$ large enough.
\end{lemma}
We emphasis that $\eta(A)$ is independent of $\ep$ and $t$.
\subsection{The affine energy}
From now on, we assume that the Mabuchi functional ${\mathcal M}$ is affine along ${\mathcal G}$, namely,
we have ${\mathcal K}(t)$ is a linear function on $[0,1]$.
Denote ${\mathcal K}_{A}$ by the following limit for each $t\in [0,1]$ and $A$ large
$$ {\mathcal K}_A(t): = \limsup_{\ep\rightarrow 0} {\mathcal K}_{\ep, A}(t) = \limsup_{\ep\rightarrow 0} \widetilde{{\mathcal K}}_{\ep, A} (t),$$
where $\widetilde{{\mathcal K}}_{\ep, A}$ is defined in equation (\ref{int-0091}).
Then ${\mathcal K}_A$ is a convex function on the unit interval $[0,1]$,
since the $\limsup$ of a sequence of convex functions which are locally bounded above is still convex.
Eventually, we will see that this convex function ${\mathcal K}_{A}$ obtained from taking the limit of ${\mathcal K}_{\ep, A}$
is exactly equal to ${\mathcal K}^{\Psi_A}$ (see Corollary (\ref{cor-ae-001})).
First, we note that ${\mathcal K}_A$ is actually a decreasing sequence in $A$.
This is because ${\mathcal K}_{\ep, A}$ is a decreasing sequence in $A$ for each $\ep$ fixed.
In fact, we have
$$H_{A'}(\vp_{\ep}) \leq H_A(\vp_\ep), $$
for each $t\in [0,1]$ and any $A' \geq A$, since $f_{A'}(\vp_\ep) \leq f_A (\vp_\ep)$ and $f_{\vp} \geq 0$ in this case.
Bearing this in mind, we conclude the following result.
\begin{theorem}
\label{thm-ae-001}
Suppose the Mabuchi functional ${\mathcal M}$ is affine along a geodesic ${\mathcal G}$.
Then there is a positive number $A_0$ such that we have
$$ {\mathcal K}_A (t) = {\mathcal K} (t), $$
for all $t\in[0,1]$ and $A \geq A_0$.
\end{theorem}
\begin{proof}
Up to a linear function on $[0,1]$, we can assume that the Mabuchi functional
${\mathcal M}$ is identically zero along ${\mathcal G}$, namely we have on $[0,1]$
$${\mathcal K}(t)\equiv 0. $$
The first observation is that we have
\begin{equation}
\label{ae-000}
{\mathcal K}_A(0) = 0; \ \ \ {\mathcal K}_A (1) = 0,
\end{equation}
since the $\ep$-geodesic potential $\Phi_{\ep}$ coincides with the geodesic potential $\Phi$ for each $\ep$ on the boundary of ${\mathcal R} \times X$.
Then it is easy to see that ${\mathcal K}_{\ep, A} (0) = {\mathcal K}_{\ep, A} (1) = 0$ for all $A$ large enough.
As a convex function on $[0,1]$, ${\mathcal K}_A$ is upper semi-continuous near the boundaries, and then we have
$$ \limsup_{t\rightarrow 0, 1} {\mathcal K}_A (t) \leq 0. $$
Thanks to the convexity again, ${\mathcal K}_A$ must be below the line segment joining its two boundaries.
Therefore, it is non-positive under our assumption and we have
\begin{equation}
\label{ae-001}
{\mathcal K}_A(t) \leq 0,
\end{equation}
for all $t\in[0,1]$.
On the other hand ,
we have for each $t\in (0,1)$
\begin{equation}
\label{ae-002}
{\mathcal K}_A (t) \geq {\mathcal K}(t) - \eta(A),
\end{equation}
This directly follows from Lemma (\ref{lem-int-001}), where we have proved
$$ \limsup_{\ep\rightarrow 0} H_A (\vp_{\ep}) \geq H_A (\vp) - \eta(A).$$
Next define a new function on $[0,1]$ as
$$ \widetilde{{\mathcal K}}(t): = \limsup_{A \rightarrow +\infty} {\mathcal K}_A (t). $$
For the same reason, $\widetilde{{\mathcal K}}$ is a convex function which verifies the boundary condition
$\widetilde{{\mathcal K}}(0)= \widetilde{{\mathcal K}}(1) = 0$ by equation (\ref{ae-000}).
Moreover, inequality (\ref{ae-002}) implies that
$\widetilde{{\mathcal K}}(t) \geq 0$ for each $t\in (0,1) $,
since $\eta(A)\rightarrow 0$ as $A\rightarrow +\infty$ by Lemma (\ref{lem-int-001}).
Therefore, we conclude that $\widetilde{{\mathcal K}}$ must be identically equal to zero for all $t\in [0,1]$.
However, as mentioned before, ${\mathcal K}_A$ is actually a decreasing sequence in $A$,
namely, we have ${\mathcal K}_A(t) \searrow \widetilde{{\mathcal K}}(t)$ for each $t\in [0,1]$.
Therefore, the following inequality follows
$${\mathcal K}_A(t) \geq 0, $$
for each $t\in [0,1]$.
Combining this with equation (\ref{ae-001}), we have for all $A$ large enough
$${\mathcal K}_A(t) \equiv 0 $$
on $[0,1]$, and our result follows.
\end{proof}
In other words, the limit of the $A$-truncated Mabuchi functional on the $\ep$-geodesics
will coincide with ${\mathcal M}$ for all $A$ large enough, provided the linearity of the Mabuchi functional along ${\mathcal G}$.
\subsection{Gap phenomenon }
When the Mabuchi functional is affine along ${\mathcal G}$, we have seen from (\cite{LL}) that the $A$-truncated Mabuchi functional ${\mathcal K}^{\Psi_A}$
will also coincide with ${\mathcal K}$ for all $A$ large.
This implies
\begin{equation}
\label{ae-0025}
H_A(\vp) = H(\vp),
\end{equation}
for all $A$ large enough, and then we have the following corollary.
\begin{corollary}
\label{cor-ae-001}
Suppose the Mabuchi functional ${\mathcal M}$ is affine along a geodesic ${\mathcal G}$.
Then for all $A$ large enough, the following limit exists for each $t\in [0,1]$ and satisfies
\begin{equation}
\label{ae-0026}
\lim_{\ep\rightarrow 0} {\mathcal K}_{\ep, A} (t) = {\mathcal K}^{\Psi_A}(t) ={\mathcal K} (t).
\end{equation}
\end{corollary}
\begin{proof}
For each $t\in [0,1]$ fixed, we define
$$ \widetilde{H}_A (t) : = \liminf_{\ep\rightarrow 0} H_A(\vp_\ep). $$
Thanks to Lemma (\ref{lem-int-001}), we have
\begin{equation}
\label{ae-003}
\lim_{A\rightarrow +\infty} \widetilde{H}_A(t) \geq H(\vp),
\end{equation}
by equation (\ref{ae-0025}).
However, $ \widetilde{H}_A$ is actually a decreasing sequence in $A$.
This follows from its construction and the fact that $H_{A}(\vp_{\ep})$ is decreasing for each $\ep$ fixed.
Hence, we conclude
\begin{equation}
\label{ae-004}
\widetilde{H}_A(t) \geq H(\vp).
\end{equation}
On the other hand,
Theorem (\ref{thm-ae-001}) implies for all $A$ large
\begin{equation}
\label{ae-005}
\limsup_{\ep\rightarrow 0} H_A(\vp_\ep) = H(\vp) = H_A(\vp).
\end{equation}
Combined the two inequalities above together, we have
\begin{equation}
\label{ae-005}
H(\vp) = \limsup_{\ep\rightarrow 0} H_A(\vp_\ep) \geq \liminf_{\ep\rightarrow 0} H_A(\vp_\ep) \geq H (\vp),
\end{equation}
at each $t\in[0,1]$, and then our result follows.
\end{proof}
Moreover, equation (\ref{ae-0025}) implies the so called ``gap phenomenon" (\cite{LL})
for the fiber-wise volume element of the geodesic ${\mathcal G}$.
Denote $P$ by the following measurable subset
$$ P: = \left\{ (t,z)\in \Gamma\times X;\ \ \ {\mathfrak{f}} (\vp) > 0 \right\}, $$
and $P_A$ by the following sub-level set
$$ P_A:= \left\{ (t,z)\in \Gamma\times X;\ \ \ {\mathfrak{f}} (\vp) > \frac{e^{\chi-A}}{\omega^n} \right\}. $$
For each $t\in[0,1]$, we introduce
their fiber-wise restrictions as
$$ P_t: = P \bigcap X_t; \ \ \ P_{A,t}: = P_A \bigcap X_t. $$
Then there is a large constant $A_0 $ such that for all $A\geq A_0$, and each $t\in[0,1]$,
the following equality holds
\begin{equation}
\label{ae-0055}
X_t = P_{A,t} \bigcup P_t^c
\end{equation}
up to a set of measure zero.
In other words,
there exists a uniform constant ${\kappa}_0 > 0$ such that either we have
on each fiber $X_t$
\begin{equation}
\label{ae-006}
{\mathfrak{f}}(\vp) > {\kappa}_0 \ \ \emph{or}\ \ \ \ {\mathfrak{f}} (\vp)= 0
\end{equation}
almost everywhere.
We emphasis that ${\kappa}_0$ does not depend on $t$ or $x$.
This is called the gap phenomenon for the fiberwise volume element of the geodesic ${\mathcal G}$.
\section{Convergence of the volume element}
Recall that the $\ep$-geodesic potential $\Phi_{\ep}$ converges to the geodesic potential $\Phi$
in the weak $C^{1,\bar 1}$-norm.
Fixing a fiber $X_{\tau}$,
we can pick up a convergent subsequence of the volume elements on this fiber
$$ f_{\ell} (x) : = {\mathfrak{f}}(\vp_{\ep_{{\ell}}})(x); \ \ \ f(x): = {\mathfrak{f}}(\vp)(x),$$
satisfying
$$ f_{\ell} \rightarrow f $$
weakly in $L^p$ for all $p\geq 1$ on $X_{\tau}$.
It is interesting to know whether we have strong convergence in $L^p$ for this sequence or not,
and it turns out that this is indeed the case if the Mabuchi functional is affine.
\begin{theorem}
\label{thm-l2-001}
Suppose the Mabuchi functional ${\mathcal M}$ is affine along a geodesic ${\mathcal G}$.
Then the fiber-wise volume element of the $\ep$-geodesic converges to the
volume element of the geodesic in the strong $L^2$ sense as $\ep\rightarrow 0$.
More precisely, we have
$$ \lim_{{\ell}\rightarrow +\infty} || f_{\ell} - f ||_{L^2} = 0, $$
on each fiber $X_{\tau}$.
\end{theorem}
In the following, we will provide two different proves of Theorem (\ref{thm-l2-001}).
The first one is more complicated, but it will
give an accurate estimate for the $L^2$-norm of the difference $f_{\ell} - f$
directly from the convergence of the truncated entropies.
We expect that this estimate will be useful for some independent interests.
For the beginning, denote ${\kappa}$ by the auxiliary function on the fiber $X_{\tau}$
\begin{equation}
\label{kappa}
{\kappa}(x, A): = \frac{e^{\chi -A}}{\omega^n} (x).
\end{equation}
As before, we omit the sub-index $\ep_{{\ell}}$ in $\chi$ due to the uniform control on their ${\mathcal C}^{1,\bar 1}$-norms.
Then the following result holds.
\begin{lemma}
\label{lem-l2-0010}
Suppose we have
$$ \lim_{{\ell}\rightarrow \infty} \int_X {\mathfrak{f}}(\vp_{\ep_{\ell}}) \log{\mathfrak{f}}_A (\vp_{\ep_{\ell}}) = \int_X {\mathfrak{f}}(\vp)\log{\mathfrak{f}}_A(\vp).$$
Then there exists a uniform constant $C$, only depending on the upper bound of $f_{\ell}$ and $f$, satisfying
\begin{equation}
\label{l2-0001}
\lim_{{\ell}\rightarrow \infty} \int_X (f_{\ell} - f)^2 \omega^n \leq \left( 2+ \frac{8C}{{\kappa}_0} \right) \max_{X_{{\tau}}} {\kappa},
\end{equation}
for all ${\kappa}$ small enough (or $A$ large enough).
\end{lemma}
Here ${\kappa}_0$ is the gap of the volume element ${\mathfrak{f}}(\vp)$ (equation (\ref{ae-006})), which is a fixed constant.
Therefore, Theorem (\ref{thm-l2-001}) directly follows from Lemma (\ref{lem-l2-0010}) if we take $A\rightarrow \infty$.
\subsection{The maximum function}
In order to prove this lemma,
the first step is to investigate the following maximum function.
Define a function $F: [0, +\infty) \rightarrow \mathbb{R}$ as
$$F(x): = x\log x, $$
and $F(0) = 0$. Then $F$ is a convex continuous function on its domain.
In fact, it is smooth in $\mathbb{R}^+$, and
its first and second derivatives are
$$ F'(x) = \log x +1;\ \ \ F''(x) = x^{-1} $$
Moreover, we can truncate $F$ by a small number ${\kappa}>0$ and introduce the following maximum function
$$ h_{{\kappa}}(x): = \max\{ x\log x, x\log{\kappa} \}. $$
This is also a convex and continuous function on $[0, +\infty)$,
and it is piecewise smooth in this domain.
Its first derivative exists everywhere on $\mathbb{R}^+$ except at the point $x={\kappa}$, and we have
\begin{equation}
\label{l2-001}
h'_{{\kappa}}(x) = \left\{ \begin{array}{rcl}
\log{\kappa},
& \mbox{for} &
x< {\kappa} \\
1+ \log x,
& \mbox{for} & x > {\kappa}. \\
\end{array}\right.
\end{equation}
Furthermore, we can also compute its second derivative on $\mathbb{R}^+$ as
\begin{equation}
\label{l2-002}
h''_{{\kappa}}(x) = \left\{ \begin{array}{rcl}
0,
& \mbox{for} &
x< {\kappa} \\
\delta({\kappa}),
&\mbox{for}& x={\kappa} \\
\frac{1}{x},
& \mbox{for} & x > {\kappa}, \\
\end{array}\right.
\end{equation}
where $\delta({\kappa})$ is the \emph{Dirac-delta} function at the point $x={\kappa}$.
We note that the Fundamental Theorem of Calculus is still satisfied for $h_{{\kappa}}, h'_{{\kappa}}, h''_{{\kappa}}$
on the interval $[0,1]$, since we can take the differentiation in the sense of the generalized derivatives.
Fixing a point $x$ on the fiber, we introduce another variable $t\in[0,1]$ and take
$$u_t: = t f_{\ell} + (1-t) f = at +b, $$
where $a: = (f_{\ell} - f)(x)$ and $b: = f(x)$.
We note that $u_t$ is strictly positive and has a uniform upper bound for all $t$, ${\ell}$ and $x$.
Define a new composition function
$$ F_{{\ell}, {\kappa}}(t): = h_{{\kappa}} (u_t),$$
and then its derivatives can be written as
\begin{equation}
\label{l2-003}
F'_{{\ell}, {\kappa}}(t) = \left\{ \begin{array}{rcl}
a \log {\kappa},
& \mbox{for} &
at +b < {\kappa} \\
a\log(at +b) +a ,
&\mbox{for}& at+b > {\kappa}; \\
\end{array}\right.
\end{equation}
\begin{equation}
\label{l2-004}
F''_{{\ell}, {\kappa}}(t) = \left\{ \begin{array}{rcl}
0,
& \mbox{for} &
at +b < {\kappa} \\
a \delta(t_0),
&\mbox{for}& at +b ={\kappa} \\
\frac{a^2}{at+b},
& \mbox{for} & at+b > {\kappa}, \\
\end{array}\right.
\end{equation}
where $t_0$ is determined by the equation
$$at_0 + b = {\kappa}. $$
In particular, we have $F'_{{\ell}, {\kappa}}(0) = a\log{\kappa}$ if $f(x) < {\kappa}$,
and $F'_{{\ell}, {\kappa}}(0) = a\log b + a$ if $f(x) > {\kappa}$.
Then the following convergence holds.
\begin{lemma}
\label{lem-l2-001}
For all $A$ large enough, we have
$$ \lim_{{\ell}\rightarrow + \infty} \int_X F'_{{\ell}, {\kappa}} (0) \omega^n = 0.$$
\end{lemma}
\begin{proof}
When the constant $A$ is large enough, we can assume ${\kappa} < {\kappa}_0/2$ on the fiber $X_{{\tau}}$,
where ${\kappa}_0$ is the gap of the fiber-wise volume element of ${\mathcal G}$ defined in equation (\ref{ae-006}).
Then the fiber can be completely decomposed into two parts as in equation (\ref{ae-0055})
$$ X_{{\tau}} = P_{A,{\tau}} \bigcup P_{{\tau}}^c, $$
up to a set of measure zero.
Recall that the two sets can be re-written as follows
$$ P_{A, {\tau}} = \{ x\in X_{{\tau}}; \ \ \ f(x) > {\kappa}(x) \}; \ \ \ P^c_{{\tau}}: = \{ x\in X_{\tau}; \ \ \ f(x) =0 \}. $$
Then we have
\begin{eqnarray}
\label{l2-005}
\int_X F'_{{\ell}, {\kappa}} (0) &= &\int_{P_{A, {\tau}}} F'_{{\ell}, {\kappa}} (0)+ \int_{P_{{\tau}}^c} F'_{{\ell}, {\kappa}} (0)
\nonumber\\
&=& \int_{P_{A, {\tau}}} (f_{{\ell}} - f) \log f + \int_{P_{A, {\tau}}} (f_{{\ell}} - f) + \int_{P_{\tau}^c} f_{{\ell}}\log{\kappa}
\nonumber\\
&= & \int_{P_{A, {\tau}}} (f_{{\ell}} - f) \log f + \int_{P_{{\tau}}^c} (f_{\ell} -f)\log{\kappa} + \int_{P_{{\tau}}^c} f \log{\kappa}
\nonumber\\
&+& \int_X (f_{\ell} - f) - \int_{P^c_{\tau}} (f_{\ell} - f)
\nonumber\\
&=& \int_X (f_{\ell} - f) \log\max\{ f, {\kappa}\} + \int_X (f_{\ell} - f) - \int_{P^c_{{\tau}}} f_{\ell}.
\end{eqnarray}
The three terms on the RHS of equation (\ref{l2-005}) will all converge to zero as ${\ell}\rightarrow +\infty$,
since $|| f ||_{L^{\infty}}, || f_{\ell} ||_{L^{\infty}}$ are uniformly bounded and $f_{\ell} \rightarrow f$ weakly in $L^p$ for any $p\geq 1$,
and then our result follows.
\end{proof}
\subsection{The four cases}
Next we will apply the Fundamental Theorem of Calculus on the function $F_{{\ell}, {\kappa}}(t)$ and its first derivative,
namely, we have
\begin{eqnarray}
\label{l2-006}
F_{{\ell}, {\kappa}}(1) - F_{{\ell}, {\kappa}}(0) &=& h_{{\kappa}}(f_{\ell}) - h_{{\kappa}}(f)
\nonumber\\
&=& \int_0^{t_0} F'_{{\ell}, {\kappa}}(t)dt + \int_{t_0}^1 F'_{{\ell}, {\kappa}}(t)dt.
\end{eqnarray}
A first observation is that the point $t_0$ may not be in the integration domain above.
Suppose the point $x$ is in the subset $P_{{\tau}}^c$,
and then we have $b =f(x) = 0$, $a = f_{{\ell}}(x) > 0$.
Therefore, the point $t_0$ belongs to the interval $(0,1)$
if and only if $ 0< {\kappa} < a$.
Otherwise, we have $t_0 \geq 1$ if $ f_{\ell} (x) \leq {\kappa} $, but $t_0 \leq 0$ is not possible in this case
since it means ${\kappa}\leq 0$.
On the other hand, suppose the point $x$ is in the subset $P_{A,{\tau}}$.
Then we have $b = f(x) > {\kappa}$ and $a = f_{\ell} (x)- f(x)$.
Hence $t_0 \in (0,1)$ if and only if
$a< 0$ and $b > {\kappa} > a+b = f_{{\ell}}(x)$.
Otherwise, when $a\geq 0$, we have $f_l(x) \geq f(x) > {\kappa} $ for $t_0 \leq 0$;
or when $a<0$, we have ${\kappa} \leq f_{\ell} (x)$ for $t_0 \geq 1$.
In conclusion, we distinguish all situations into the following four cases:
\begin{enumerate}
\item[(i)]
$x\in P_{A,{\tau}}$, $f_{\ell} (x) \geq f(x) > {\kappa} $ or $ f(x) > f_{\ell}(x) \geq {\kappa} $;
\smallskip\\
\item[(ii)]
$x\in P^c_{{\tau}}$, $f(x) = 0$ and $f_{\ell} (x) \leq {\kappa}$;
\smallskip\\
\item[(iii)]
$x\in P^c_{{\tau}}$, $f(x)=0$ and $f_{\ell} (x) > {\kappa}$;
\smallskip\\
\item[(iv)]
$x\in P_{A, {\tau}}$, $f_{\ell} (x) < {\kappa} < f(x)$.
\end{enumerate}
\bigskip
We note that these four cases are disjoint from each other.
Then we will discuss case by case.
For $\textbf{Case (i)}$, we note that
$u_t > {\kappa}$ and then $h_{{\kappa}}(u_t) = u_t \log u_t$ for all $t\in [0,1]$.
Therefore, we can further compute as follows.
\begin{eqnarray}
\label{l2-007}
F_{{\ell}, {\kappa}}(1) - F_{{\ell}, {\kappa}}(0)
&=& F'_{{\ell},{\kappa}}(0) + \int_0^1 \int_0^t F_{{\ell},{\kappa}}''(s) ds dt
\nonumber\\
&=& F'_{{\ell},{\kappa}}(0) + \int_0^1 \int_0^t \frac{a^2 ds}{as +b} dt
\nonumber\\
&\geq& F'_{{\ell},{\kappa}}(0) + \frac{a^2}{2C},
\end{eqnarray}
where the constant $C$ is the uniform upper bound of $u_t$.
For $\mathbf{Case (ii)}$, we note $u_t\leq {\kappa}$ and then $h_{{\kappa}}(u_t) = u_t \log {\kappa}$.
Hence we have
\begin{eqnarray}
\label{l2-008}
F_{{\ell}, {\kappa}}(1) - F_{{\ell}, {\kappa}}(0)
&=& F'_{{\ell},{\kappa}}(0) + \int_0^1 \int_0^t F_{{\ell},{\kappa}}''(s) ds dt
\nonumber\\
&=& F'_{{\ell},{\kappa}}(0).
\end{eqnarray}
The two cases above are the easy ones.
For the remaining cases, we will utilise equations (\ref{l2-003}) and (\ref{l2-004}) in the computation.
In $\mathbf{Case (iii)}$, we note that $h_{{\kappa}}(u_t) = u_t \log{\kappa} $ for $t\leq t_0$ and $h_{{\kappa}}(u_t) = u_t \log u_t$ for $t> t_0$.
Recall that $t_0 = {\kappa} /a$ in this case, and hence the computation follows.
\begin{eqnarray}
\label{l2-009}
F_{{\ell}, {\kappa}}(1) - F_{{\ell}, {\kappa}}(0)
&=& \int_0^{{\kappa}/a} F'_{{\ell}, {\kappa}} (t) dt + \int_{{\kappa}/a}^1 F'_{{\ell},{\kappa}}(t) dt
\nonumber\\
&=& \frac{{\kappa}}{a} F'_{{\ell}, {\kappa}} (0) + \int_0^{{\kappa}/a} \int_0^t F''_{{\ell}, {\tau}} (s) ds dt + \int_{{\kappa}/a}^1 F_{{\ell},{\kappa}}'(t) dt
\nonumber\\
&=& F'_{{\ell}, {\kappa}} (0) + \int_{{\kappa}/a}^1 \left\{ a + \int_{{\kappa}/a}^t F''_{{\ell}, {\kappa}}(s) ds \right\} dt
\nonumber\\
&=& F'_{{\ell}, {\kappa}} (0) + (a - {\kappa}) + \int_{{\kappa}/a}^1 \int_{{\kappa}/a}^t \frac{a^2 ds}{as +b } dt
\nonumber\\
&\geq & F'_{{\ell}, {\kappa}} (0) + (a - {\kappa}) + \frac{a^2}{2C} (1- {\kappa}/a)^2
\nonumber\\
&\geq & F'_{{\ell}, {\kappa}} (0) + \frac{a^2}{2C} + a(1- {\kappa}/C) - {\kappa}.
\end{eqnarray}
Recall that $a > 0$ in this case, and then we have
\begin{equation}
\label{l2-010}
F_{{\ell}, {\kappa}}(1) - F_{{\ell}, {\kappa}}(0) \geq F'_{{\ell}, {\kappa}} (0) + \frac{a^2}{2C} - {\kappa},
\end{equation}
for all ${\kappa}$ small enough.
Finally, the most difficult one is $\mathbf{Case (iv)}$.
As before, we first note that $h_{{\kappa}}(u_t) = u_t\log u_t$ for
$t\leq t_0$ and $h_{{\kappa}}(u_t)$ and $h_{{\kappa}} (u_t) = u_t\log{\kappa}$
for $t > t_0$. Then we compute in a similar way.
\begin{eqnarray}
\label{l2-011}
F_{{\ell}, {\kappa}}(1) - F_{{\ell}, {\kappa}}(0)
&=& \int_0^{\frac{{\kappa} - b}{a}} F'_{{\ell}, {\kappa}} (t) dt + \int_{\frac{{\kappa}-b}{a}}^1 F'_{{\ell},{\kappa}}(t) dt
\nonumber\\
&=& \frac{{\kappa}-b}{a} F'_{{\ell}, {\kappa}} (0) + \int_0^{\frac{{\kappa}-b}{a}} \int_0^t F''_{{\ell}, {\tau}} (s) ds dt + \int_{\frac{{\kappa}-b}{a}}^1 F_{{\ell}, {\kappa}}'(t) dt
\nonumber\\
&=& F'_{{\ell}, {\kappa}} (0) + \int_0^{\frac{{\kappa}-b}{a}} \int_0^t \frac{a^2 ds}{as +b} dt + \int^1_{\frac{{\kappa}-b}{a}} \left\{ \int_0^{\frac{{\kappa}-b}{a}} F''_{{\ell},{\kappa}}(s)ds + a \right\}
\nonumber\\
&\geq & F'_{{\ell}, {\kappa}} (0) + a +b + \frac{({\kappa}-b)^2}{2C} + \frac{ ({\kappa} -b)(a+b -{\kappa})}{C} -{\kappa}.
\end{eqnarray}
Recall that we have $a+b = f_{\ell}(x) >0$, ${\kappa}-b = {\kappa} - f(x) < 0$ and $a+b - {\kappa} = f_{\ell}(x) - {\kappa} <0$.
Hence the following estimate holds.
\begin{eqnarray}
\label{l2-012}
F_{{\ell}, {\kappa}}(1) - F_{{\ell}, {\kappa}}(0) &\geq& F'_{{\ell}, {\kappa}} (0) + \frac{({\kappa} -b)^2 }{2C} - {\kappa}
\nonumber\\
&\geq & F'_{{\ell}, {\kappa}} (0) + \frac{ {\kappa}_0^2 }{8C} - {\kappa}.
\end{eqnarray}
The last inequality in equation (\ref{l2-012}) is because that
we have picked up ${\kappa} < {\kappa}_0 /2$, and $P_{A, {\tau}}$ is actually the set where $f(x) > 0$ on the fiber which is equal to
$$ \{ x\in X_{{\tau}}; \ \ \ f (x) > {\kappa}_0 \}$$
up to a set of measure zero by the gap phenomenon.
Combining with equations (\ref{l2-007}) - (\ref{l2-012}) above, we conclude the following inequality after taking the integral on $X_{{\tau}}$.
\begin{eqnarray}
\label{l2-013}
\int_X ( h_{{\kappa}} (f_{\ell}) - h_{{\kappa}} (f) ) \omega^n &\geq& \int_X F'_{{\ell}, {\kappa}}(0) \omega^n - \max_{X_{{\tau}}} {\kappa}
\nonumber\\
&+& \frac{1}{2C}\int_{\left( P^c_{{\tau}} \bigcap \{ f_{\ell} \leq {\kappa} \} \right)^c \bigcap \left( P_{A,{\tau}} \bigcap \{ f_l < {\kappa} < f \} \right)^c } (f_{\ell} - f)^2 \omega^n
\nonumber\\
&+& \frac{1}{8C} {\kappa}_0^2 \mu\left( P_{A, {\tau}} \bigcap \{ f_l < {\kappa} < f \} \right).
\end{eqnarray}
Then we are ready to prove the main theorem in this section.
\begin{proof}
[Proof of Lemma (\ref{lem-l2-0010})]
By our choice on the function ${\kappa}$ and the maximum function $h_{{\kappa}}$,
it follows
\begin{equation}
\label{l2-014}
H_A (\vp_{\ep_{\ell}}) - H_A (\vp) = \int_X ( h_{{\kappa}} (f_{\ell}) - h_{{\kappa}} (f) ) \omega^n.
\end{equation}
Thanks to Corollary (\ref{cor-ae-001}), the LHS of equation (\ref{l2-013}) converges to zero as ${\ell} \rightarrow +\infty$.
Meanwhile, our Lemma (\ref{lem-l2-001}) implies the first term on the RHS of equation (\ref{l2-013}) also converges to zero.
Therefore, it implies
\begin{equation}
\label{l2-015}
\frac{8C}{{\kappa}_0^2} ( \max_{X_{\tau}} {\kappa} ) \geq \lim_{{\ell}\rightarrow +\infty} \mu\left( P_{A, {\tau}} \bigcap \{ f_l < {\kappa} < f \} \right).
\end{equation}
Furthermore, we note that the two subsets $P_{A,{\tau}} \bigcap \{ f_l < {\kappa} < f \} $ and $P^c_{{\tau}} \bigcap \{ f_{\ell} \leq {\kappa} \} $ are mutually disjoint.
Then the third term on the RHS of equation (\ref{l2-013}) can be decomposed into the following three parts.
$$ \int_X (f_{\ell} -f)^2 - \int_{P^c_{{\tau}} \bigcap \{ f_{\ell} \leq {\kappa} \} }(f_{\ell} -f)^2 - \int_{P_{A,{\tau}} \bigcap \{ f_l < {\kappa} < f \} } (f_{\ell} -f)^2. $$
The first negative term in the equation above can be estimated as
\begin{equation}
\label{l2-016}
\int_{P^c_{{\tau}} \bigcap \{ f_{\ell} \leq {\kappa} \} }(f_{\ell} -f)^2 \leq \max_{X_{\tau}} {\kappa}^2,
\end{equation}
and the second negative term can be estimated by equation (\ref{l2-015}) as
\begin{equation}
\label{l2-017}
\int_{P_{A,{\tau}} \bigcap \{ f_l < {\kappa} < f \} } (f_{\ell} -f)^2 \leq C^2 \mu\left( P_{A, {\tau}} \bigcap \{ f_l < {\kappa} < f \} \right).
\end{equation}
Combing with equations (\ref{l2-014}) - (\ref{l2-017}) and take the limit in ${\ell}$, we eventually conclude the following estimate
\begin{equation}
\label{l2-017}
\left( 2+ \frac{8C^3}{{\kappa}_0} \right) \max_{X_{\tau}} {\kappa} \geq \lim_{{\ell} \rightarrow +\infty} \int_X (f_{\ell} - f)^2\omega^n,
\end{equation}
and then our result follows.
\end{proof}
\subsection{Yet another proof}
From the previous argument,
we can see that the $L^2$-convergence of the difference $(f_{\ell} - f)$
actually follows from the convergence of the entropy as
\begin{equation}
\label{l2-018}
\lim_{{\ell}\rightarrow \infty} \int_X {\mathfrak{f}}(\vp_{\ep_{\ell}}) \log {\mathfrak{f}}(\vp_{\ep_{\ell}}) = \int_X {\mathfrak{f}}(\vp) \log {\mathfrak{f}} (\vp),
\end{equation}
but this is a direct consequence of our Corollary (\ref{cor-ae-001}) as follows.
\begin{corollary}
\label{cor-l2-001}
Suppose the Mabuchi functional is affine along a geodesic ${\mathcal G}$.
Then on each fiber $X_{{\tau}}, {\tau}\in [0,1]$, we have
$$ \lim_{\ep\rightarrow 0} H(\vp_{\ep}) = H(\vp). $$
\end{corollary}
\begin{proof}
During the proof of Corollary (\ref{cor-ae-001}),
we have seen that $H_A(\vp_{\ep})$ is decreasing in $A$ for each $\ep$ fixed.
Moreover, Theorem (\ref{thm-ae-001}) and the gap phenomenon imply that
$$ \limsup_{\ep\rightarrow 0} H_A(\vp_\ep) = H_A(\vp) = H(\vp), $$
for all $A$ large enough.
Therefore, we conclude
the following inequality
\begin{equation}
\label{l2-019}
\limsup_{\ep\rightarrow 0} H(\vp_\ep) \leq \limsup_{\ep\rightarrow 0} H_A(\vp_\ep) = H(\vp).
\end{equation}
On the other hand, by the lower semi-continuity property of the entropy functional, we have
\begin{equation}
\label{l2-020}
\liminf_{\ep\rightarrow 0} H(\vp_\ep) \geq H(\vp),
\end{equation}
and then our result follows.
\end{proof}
In particular, the above corollary implies the convergence of the Mabuchi functional along the $\ep$-geodesic,
namely, we have
$$ \lim_{\ep\rightarrow 0} {\mathcal K}_{\ep} (t) = {\mathcal K}(t), $$
for each $t\in [0,1]$.
As a simple application of Theorem (\ref{thm-l2-001}),
the fiber-wise volume element $f_{\ell}$ of the $\ep$-geodesic converges to
the volume element $f$ of the geodesic in measure.
Moreover, thanks to the Riesz-Lebesgue Theorem, we have
$$f_{\ell} \rightarrow f $$
almost everywhere on each fiber, possibly after passing to a subsequence.
\section{The $\ep$-affine energy and non-degneracy }
Suppose the Mabuchi functional ${\mathcal M}$ is affine along a geodesic ${\mathcal G}$.
For the convergence $\widetilde{{\mathcal K}}_{\ep, A} \rightarrow {\mathcal K}_A$ of a sequence of convex functions, the first derivative $\widetilde{{\mathcal K}}'_{\ep, A}(t)$
also converges uniformly to the slope $k: = {\mathcal K}'_A(t)$ on the closed interval $[\delta, 1 - \delta ]$ for any $\delta >0$ small.
However, the boundary behaviour of $\widetilde{{\mathcal K}}'_{\ep, A}(t)$ is unclear in general.
Therefore, we impose the following condition on the boundary of ${\mathcal G}$.
\begin{defn}
\label{def-ch-001}
The Mabuchi functional ${\mathcal M}$ is essentially affine along a geodesic ${\mathcal G}$, if
${\mathcal K}(t)$ is a linear function on $[0,1]$ with slope $k$ and we have
\begin{equation}
\label{ch-001}
- \int_X \dot{\vp}_t ( R_{\vp_t} - \underline{R}) \omega_{\vp_t}^n = k,
\end{equation}
at $t=0, 1$.
\end{defn}
It is proved in Berman-Berndtsson (\cite{BB}) that the one side inequality of equation (\ref{ch-001})
always holds at the two boundaries.
$$ - \int_X \dot{\vp}_0 ( R_{\vp_0} - \underline{R}) \omega_{\vp_0}^n \leq k, $$
and
$$ - \int_X \dot{\vp}_1 ( R_{\vp_1} - \underline{R}) \omega_{\vp_1}^n \geq k. $$
As we have seen before, the potential $\vp_\ep(t, \cdot)$ of the $\ep$-geodesic converges to
the potential $\vp(t,\cdot)$ of the geodesic ${\mathcal G}$ in $C^{1,{\alpha}}$-norm for each ${\alpha}\in (0,1)$ on $\Gamma\times X$.
Therefore, we have
\begin{equation}
\label{eaf-001}
\lim_{\ep\rightarrow 0} \int_X \dot{\vp}_{\ep} ( R_{\vp_\ep} - \underline{R}) \omega_{\vp_\ep}^n = \int_X \dot{\vp} ( R_{\vp} - \underline{R}) \omega_{\vp}^n,
\end{equation}
at the boundaries $t=0,1$, since $\omega_{\ep}$ coincides with $\omega_{\vp}$ at these two boundaries.
Moreover, we note that the volume element $\omega_{\ep}^n$ varies smoothly near the boundaries,
and then ${\mathfrak{f}}_A(\vp_\ep)$ will keep to be the same as ${\mathfrak{f}}(\vp_{\ep})$ for all $A$ large in
a small neighbourhood near the boundaries.
Therefore,
the $A$-truncated Mabuchi functional along ${\mathcal G}_{\ep}$
is equal to the Mabuchi functional along ${\mathcal G}_{\ep}$
in a small neighbourhood of the two boundaries.
That is to say, there exist a small number $\delta>0$,
possibly depends on $\ep$ and $A$, such that we have
\begin{equation}
\label{eaf-002}
{\mathcal K}_{\ep, A} (t) = {\mathcal K}_{\ep}(t),
\end{equation}
for all $t\in [0,\delta) \bigcup ( 1-\delta, 1]$.
Therefore, we have for all $t$ in this small interval
\begin{equation}
\label{eaf-003}
{\mathcal K}'_{\ep, A} (t) = {\mathcal K}'_{\ep}(t) = -\int_X \dot{\vp}_{\ep} ( R_{\vp_\ep} - \underline{R}) \omega_{\vp_\ep}^n.
\end{equation}
In particular, we conclude the following result.
\begin{lemma}
\label{lem-eaf-001}
Suppose the Mabuchi functional ${\mathcal M}$ is essentially affine along a geodesic ${\mathcal G}$ with slope $k$.
Then ${\mathcal K}'_{\ep, A}$ converges uniformly to $k$ on $[0,1]$ for all $A$ large enough.
\end{lemma}
\begin{proof}
Combing with equation (\ref{eaf-001}) and (\ref{eaf-003}), we have for each $A$ large
\begin{equation}
\label{eaf-004}
\lim_{\ep\rightarrow 0} \widetilde{\mathcal K}'_{\ep, A} (t) = \lim_{\ep\rightarrow 0} {\mathcal K}'_{\ep, A} (t) = k,
\end{equation}
for $t = 0,1$.
Moreover, the function $\widetilde{\mathcal K}_{\ep, A} (t) $ is convex on the interval $[0,1]$,
and then we have the control
$$ \widetilde{\mathcal K}_{\ep, A}' (0) \leq \widetilde{\mathcal K}'_{\ep, A} (t) \leq \widetilde{\mathcal K}'_{\ep, A} (1),$$
for all $t\in (0,1)$.
Therefore, $ \widetilde{\mathcal K}'_{\ep, A} (t) \rightarrow k$
uniformly as $\ep\rightarrow 0$.
\end{proof}
However, this condition is still too difficult to handle in the application.
Therefore, we will introduce an even stronger one based on the $\ep$-geodesic as follows.
\begin{defn}
\label{def-ch-002}
The Mabuchi functional ${\mathcal M}$ is $\ep$-affine along a geodesic ${\mathcal G}$,
if ${\mathcal K}(t)$ is a linear function on $[0,1]$ and
we have
$$ {\mathcal K}_{\ep}'(1) - {\mathcal K}_{\ep}'(0) = O(\ep), $$
where ${\mathcal K}_{\ep}$ is the Mabuchi functional along the $\ep$-geodesic ${\mathcal G}_{\ep}$.
\end{defn}
For the Mabuchi functional ${\mathcal M}$ along ${\mathcal G}$,
it is easy to see that the $\ep$-affine is a stronger condition than the essentially affine.
Therefore, Lemma (\ref{lem-eaf-001}) implies that ${\mathcal K}'_{\ep, A}$ converges to the constant slope $k$
uniformly on $[0,1]$ if ${\mathcal M}$ is $\ep$-affine.
Moreover, we have the following result under this assumption.
\begin{theorem}
\label{thm-ch-001}
Suppose the Mabuchi functional ${\mathcal M}$ is $\ep$-affine along a geodesic ${\mathcal G}$.
Then ${\mathcal G}$ is uniformly fiberwise non-degenerate,
namely, there exists a uniform constant ${\kappa}_1 >0$ such that
$$ {\mathcal G}|_{X_t} > {\kappa}_1\omega,$$
for almost everywhere $t\in [0,1]$.
\end{theorem}
We emphasis that the constant ${\kappa}_1$ does not depend on $t$.
Before moving on, we need to recall and improve some computations in \cite{CLP}.
\subsection{Computations}
We will take a closer look at equation (\ref{int-007}),
and try to evaluate the lower bound of the following $(n+1, n+1)$ form on $P_{\ep, A}$
$$ (dd^c f_{\ep} - Ric( \omega) )\wedge {\mathcal G}_{\ep}^n. $$
Locally near a point $p\in \Gamma\times X$, we write the $\ep$-geodesic as follows
$$ {\mathcal G}_{\ep}: = g_{t\bar t} dt\wedge d\bar t + \sum_{{\alpha},\beta =1}^n \left( g_{t\bar\beta} dt\wedge d\bar z^{\beta} + g_{{\alpha}\bar t} dz^{{\alpha}}\wedge d\bar t + g_{{\alpha}\bar\beta} dz^{{\alpha}}\wedge d\bar z^{\beta} \right).$$
This is a K\"ahler metric on $\Gamma\times X$,
and its restriction on the fiber $X_t: = \{ t\} \times X$ can be written as
$$ \omega_{\ep}(t, \cdot): = {\mathcal G}_{\ep}|_{X_t} = \sum_{{\alpha},\beta =1 }^n g_{{\alpha}\bar\beta} dz^{{\alpha}}\wedge d\bar z^{\beta}. $$
Up to a change of holomorphic coordinates on $X_t$, we can assume
\begin{equation}
\label{com-001}
g_{{\alpha}\bar\beta} = \delta_{{\alpha}\beta}; \ \ \ dg =0
\end{equation}
at this particular point $p$.
Near this point, the $\ep$-geodesic equation can be re-written as
\begin{equation}
\label{com-002}
\rho_\ep: = g_{t\bar t} - g^{\bar\beta{\alpha}} g_{{\alpha} \bar t} g_{t\bar\beta} = \ep e^{-f_{\ep}}.
\end{equation}
Then we introduce another $(1,1)$ form $\chi_{\ep}$ defined by the following equation
$$ \chi_{\ep}:= {\mathcal G}_{\ep} - \rho_{\ep} \sqrt{-1}dt\wedge d\bar t. $$
This $(1,1)$-form may not be closed anymore, but it is still positive definite on each fiber $X_t$,
and satisfies
$$ \chi_{\ep}^{n+1} = 0. $$
Hence we have
\begin{eqnarray}
\label{com-0020}
{\mathcal G}_{\ep}^n &=&
\chi_{\ep}^n + n \rho_{\ep} \sqrt{-1} dt \wedge d\bar t \wedge \chi_{\ep}^{n-1}
\nonumber\\
&=& \chi_{\ep}^n + n \rho_{\ep} \sqrt{-1} dt \wedge d\bar t \wedge {\mathcal G}_{\ep}^{n-1}
\end{eqnarray}
The first factor
$$ n\rho_{\ep} (dd^c f_{\ep} - Ric(\omega) ) \wedge \sqrt{-1} dt\wedge d\bar t \wedge {\mathcal G}_{\ep}^{n-1} $$
can be computed as follows.
\begin{equation}
\label{com-0021}
n\rho_{\ep} dd^c f_{\ep} \wedge \sqrt{-1} dt\wedge d\bar t \wedge {\mathcal G}_{\ep}^{n-1} = \ep (\Delta_{\ep} f_{\ep}) \sqrt{-1} dt\wedge d\bar t \wedge \omega^{n},
\end{equation}
and
\begin{equation}
\label{com-0022}
- n\rho_{\ep} Ric (\omega) \wedge \sqrt{-1} dt\wedge d\bar t \wedge {\mathcal G}_{\ep}^{n-1} = - \ep ( \mathop{\rm tr}\nolimits_{\omega_{\ep}} Ric\omega ) \sqrt{-1} dt\wedge d\bar t \wedge \omega^{n}.
\end{equation}
We note that
there exist a constant $c_A$ such that we have
$$ \mathop{\rm tr}\nolimits_{\omega_{\ep}} Ric\omega < c_A,$$
on $P_{\ep, A}$, since the eigenvalues of $\omega_{\ep}$ are bounded from below (and above) by a uniform constant on this set.
Therefore, we conclude the inequality
\begin{eqnarray}
\label{com-0023}
&& n\rho_{\ep} (dd^c f_{\ep} - Ric (\omega))\wedge \sqrt{-1} dt\wedge d\bar t \wedge {\mathcal G}_{\ep}^{n-1}
\nonumber\\
&\geq& ( \Delta_{\ep} f_{\ep}) {\mathcal G}_{\ep}^{n+1} - c_A {\mathcal G}_{\ep}^{n+1}.
\end{eqnarray}
Next, we compute the second factor as follows.
Introduce the following vector field on $\Gamma\times X$ as
$$ v: = \frac{\partial}{\partial t} - g^{\bar\beta {\alpha}} g_{t\bar\beta} \frac{\partial}{\partial z^{{\alpha}}}. $$
Then one observes that this vector field $v$ generates
the kernel of the $(1,1)$-form $\chi_{\ep}$.
Hence we have
\begin{equation}
\label{com-0024}
(dd^c f_{\ep} - Ric (\omega)) \wedge \chi_{\ep}^n = (dd^c f_{\ep} - Ric (\omega)) (v, \bar v) \sqrt{-1} dt\wedge d\bar t \wedge \chi_{\ep}^n
\end{equation}
Then the goal is to compute the lower bound of the following term
$$ \partial\bar\partial \log\mathop{\rm det}\nolimits (g_{{\alpha}\bar\beta}) (v, \bar v). $$
At this point $p$,
a standard computation shows the following equation (here we are using the Einstein summation convention)
\begin{eqnarray}
\label{com-003}
\partial\bar\partial \log\mathop{\rm det}\nolimits (g_{{\alpha}\bar\beta}) (v, \bar v) &=& g_{t\bar t, {\alpha}\bar{\alpha}} - \sum_{{\alpha},\beta} | g_{t\bar\beta, {\alpha}}|^2
\nonumber\\
&-& g_{t\bar\gamma} g_{\gamma \bar t, {\alpha}\bar{\alpha}} - g_{\gamma\bar t} g_{t\bar\gamma, {\alpha}\bar{\alpha}}
\nonumber\\
&+ & R_{{\alpha}\bar\gamma} g_{t\bar{\alpha}} g_{\gamma\bar t}
\end{eqnarray}
Taking the Laplacian $\Delta_{\ep}$ with respect to the metric $\omega_{\ep}$ on the both sides of equation (\ref{com-002}), we have
\begin{eqnarray}
\label{com-004}
g_{t\bar t, {\alpha}\bar{\alpha}} - \ep \Delta_{\ep} (e^{-f_{\ep}}) &=& \sum_{{\alpha}, p} |g_{p\bar t, \bar{\alpha}}|^2 + \sum_{\beta, q} |g_{t\bar q, \bar\beta}|^2
\nonumber\\
& -& R_{q\bar p} g_{p\bar t} g_{t\bar q}
\nonumber\\
& + & g_{t\bar p} g_{p\bar t, {\alpha}\bar{\alpha}} + g_{p\bar t} g_{t \bar p, {\alpha}\bar {\alpha}}.
\end{eqnarray}
Combing with the two equations above, it turns out that
\begin{equation}
\label{com-005}
\partial\bar\partial \log\mathop{\rm det}\nolimits (g_{{\alpha}\bar\beta}) (v, \bar v) = ||\bar\partial_X v ||^2_{\ep} + \ep e^{-f_{\ep}} |\nabla_{\ep} f_{\ep}|^2 - \ep e^{-f_{\ep}} ( \Delta_{\ep} f_{\ep} ).
\end{equation}
Furthermore, we can improve the above equality as follows.
First, we observe
$$ \bar\partial_X v = - g^{\bar\beta {\alpha}} g_{t\bar\beta,\bar\lambda} d\bar z^{\lambda} \otimes\frac{\partial}{\partial z^{{\alpha}}}. $$
In other words, it can be written in tensors as
\begin{equation}
\label{com-006}
\bar\partial_{\lambda} v^{{\alpha}} = - g_{t\bar{\alpha},\bar\lambda}; \ \ \ \bar\partial_{\lambda} v^{t}=0,
\end{equation}
at the point $p$.
Moreover, we have in the time direction
\begin{equation}
\label{com-007}
\bar\partial_{t} v^{\gamma} = - g_{t\bar t, \bar\gamma} + g_{t\bar\beta}g_{\beta\bar\gamma, \bar t} ; \ \ \ \bar\partial_{t} v^{t}=0.
\end{equation}
By differentiating equation (\ref{com-002}) once we get
\begin{equation}
\label{com-008}
\ep e^{-f_{\ep}} (\partial_{{\alpha}} f_{\ep}) = g_{t\bar t,{\alpha}} - g_{t\bar p} g_{p\bar t, {\alpha}} - g_{p\bar t} g_{t\bar p,{\alpha}},
\end{equation}
and similarly
\begin{equation}
\label{com-009}
\ep e^{-f_{\ep}} (\partial_{\bar\beta} f_{\ep}) = g_{t\bar t,\bar\beta} - g_{t\bar q} g_{q\bar t, \bar\beta} - g_{q\bar t} g_{t\bar q,\bar\beta}.
\end{equation}
Hence we have
\begin{eqnarray}
\label{com-010}
( \ep e^{-f_{\ep}} )^2 |\nabla_{\ep} f_{\ep}|^2
&=&g^{\bar\beta{\alpha}} \left\{ ( g_{t\bar t,{\alpha}} - g_{p\bar t} g_{{\alpha}\bar p,t} ) - g_{t\bar p} g_{p\bar t, {\alpha}} \right\}
\nonumber\\
&& \cdot \left\{ ( g_{t\bar t,\bar\beta} - g_{t\bar q} g_{q\bar\beta, \bar t} ) - g_{q\bar t} g_{t\bar q,\bar\beta} \right\}
\nonumber\\
&=& g^{\bar\beta{\alpha}} ( \partial_t v^{\bar{\alpha}} - g_{t\bar p} \partial_p v^{\bar{\alpha}} ) (\bar\partial_t v^{\beta} - g_{q\bar t} \bar\partial_q v^{\beta} )
\nonumber\\
&=& \sum_{\beta} |\bar\partial_t v^{\beta} |^2 + g_{t\bar\mu} g_{\lambda\bar t} \partial_{\mu} v^{\bar\beta} \bar\partial_{\lambda} v^{\beta}
\nonumber\\
&-& g_{\mu\bar t} \partial_t v^{\bar\beta} \bar\partial_{\mu} v^{\beta} - g_{t\bar\lambda} \bar\partial_t v^{\beta} \partial_{\lambda} v^{\bar\beta}.
\end{eqnarray}
On the other side, we can compute the inverse matrix of ${\mathcal G}_{\ep}$ at the point $p$.
First notice that
$$\mathop{\rm det}\nolimits {\mathcal G}_{\ep} (p) = \ep e^{-f_{\ep}(p)},$$
since $g_{{\alpha}\bar\beta} = \delta_{{\alpha}\beta}$ at this point.
Then a standard calculation shows the following equations.
$$ {\mathcal G}_{\ep}^{\bar t t} = (\ep e^{-f_{\ep}})^{-1}; \ \ \ {\mathcal G}_{\ep}^{\bar t p} = - (\ep e^{-f_{\ep}})^{-1} g_{t\bar p}, $$
and also
\begin{equation}
\label{com-011}
{\mathcal G}_{\ep}^{\bar p q} = \delta_{p q} + (\ep e^{-f_{\ep}})^{-1} g_{p\bar t} g_{t\bar q}
\end{equation}
Therefore, the four terms on the RHS of equation (\ref{com-010}) can be re-written as
\begin{equation}
\label{com-012}
(\ep e^{-f_{\ep}})^{-1} \sum_{\beta} |\bar\partial_t v^{\beta} |^2 = {\mathcal G}_{\ep}^{\bar t t} g_{{\alpha}\bar\beta} \bar\partial_{t} v^{{\alpha}} \partial_{t} v^{\bar\beta};
\end{equation}
\begin{equation}
\label{com-013}
- (\ep e^{-f_{\ep}})^{-1} g_{\mu\bar t} \partial_t v^{\bar\beta} \bar\partial_{\mu} v^{\beta}= {\mathcal G}_{\ep}^{\bar\mu t} g_{{\alpha}\bar\beta} \bar\partial_{\mu} v^{{\alpha}} \partial_t v^{\bar\beta};
\end{equation}
\begin{equation}
\label{com-014}
(\ep e^{-f_{\ep}})^{-1} g_{t\bar\mu} g_{\lambda\bar t} \partial_{\mu} v^{\bar\beta} \bar\partial_{\lambda} v^{\beta} = {\mathcal G}_{\ep}^{\bar\lambda\mu} g_{{\alpha}\bar\beta} \bar\partial_{\lambda} v^{{\alpha}} \partial_{\mu} v^{\bar\beta}
- g_{{\alpha}\bar\beta} \bar\partial_{\lambda} v^{{\alpha}} \partial_{\lambda} v^{\bar\beta}.
\end{equation}
Combining with equations (\ref{com-005}) - (\ref{com-014}) above, we eventually derive the following equality.
\begin{lemma}
\label{lem-com-001}
We have
\begin{eqnarray}
\label{com-015}
\partial\bar\partial \log\mathop{\rm det}\nolimits (g_{{\alpha}\bar\beta}) (v, \bar v) &=& ||\bar\partial_X v ||^2_{\omega_{\ep}} + \ep e^{-f_{\ep}} |\nabla_{\ep} f_{\ep}|^2 - \ep e^{-f_{\ep}} ( \Delta_{\ep} f_{\ep} )
\nonumber\\
&=& || \bar\partial v ||^2_{{\mathcal G}_{\ep}} - \ep e^{-f_{\ep}} ( \Delta_{\ep} f_{\ep} ).
\end{eqnarray}
\end{lemma}
Combing with Lemma (\ref{lem-com-001}) and equations (\ref{com-0023}), (\ref{com-0024}),
we infer as equation (\ref{int-010}) the following
\begin{eqnarray}
\label{com-0150}
&& \left( dd^c \max\left\{ \log\frac{\omega_{\ep}^n }{ \omega^n }, \log\frac{h_A}{\omega^n} \right\} - Ric\omega \right) \wedge {\mathcal G}_{\ep}^n + c'_A {\mathcal G}_{\ep}^{n+1}
\nonumber\\
&\geq& \chi_{P_{\ep, A}} \left( ||\bar\partial_X v ||^2_{\omega_{\ep}} + \ep e^{-f_{\ep}} |\nabla_{\ep} f_{\ep}|^2 \right) i dt\wedge d\bar t \wedge \omega_{\ep}^n,
\end{eqnarray}
where $\chi_{P_{\ep, A}}$ is the characteristic function of the set $P_{\ep, A}$.
Finally, by using the modified energy $\widetilde{\mathcal K}_{\ep, A}$ (defined in the Theorem (\ref{int-thm-001})),
we conclude the following integral estimate.
\begin{eqnarray}
\label{com-016}
\widetilde{\mathcal K}'_{\ep, A} (1) - \widetilde{\mathcal K}'_{\ep, A} (0) &\geq & \int_{P_{\ep, A}} || \bar\partial v ||^2_{{\mathcal G}_{\ep}} i dt\wedge d\bar t\wedge \omega_{\ep}^n
\nonumber\\
&\geq& \int_{P_{\ep, A}}||\bar\partial_X v ||^2_{\omega_{\ep}} i dt\wedge d\bar t\wedge \omega_{\ep}^n + \ep \int_{P_{\ep, A}} |\nabla_{\ep} f_{\ep}|^2 i dt\wedge d\bar t\wedge \omega^n.
\end{eqnarray}
If we assume that the Mabuchi functional is essentially affine along ${\mathcal G}$,
then the RHS of equation (\ref{com-016}) converges to zero as $\ep\rightarrow 0$.
Moreover,
suppose the Mabuchi functional is $\ep$-affine along the geodesic.
Then we have
\begin{eqnarray}
\label{com-017}
\widetilde{\mathcal K}'_{\ep, A} (1) - \widetilde{\mathcal K}'_{\ep, A} (0) &= & {\mathcal K}'_{\ep}(1) - {\mathcal K}'_{\ep}(0) + 2\ep C_A
\nonumber\\
&\leq & \ep C'_A,
\end{eqnarray}
for some constant $C'_A$.
Therefore, we conclude the following estimate from equations (\ref{com-016}) and (\ref{com-017}).
\begin{eqnarray}
\label{com-018}
\int_{P_{\ep, A}} |\nabla {\mathfrak{f}} (\vp_{\ep}) |^2 i dt\wedge d\bar t\wedge \omega^n &\leq & C^2 \int_{P_{\ep, A}} \frac{ |\nabla {\mathfrak{f}} (\vp_{\ep}) |^2}{{\mathfrak{f}}^2(\vp_\ep)} i dt\wedge d\bar t\wedge \omega^n
\nonumber\\
& \leq & C^2 \int_{P_{\ep, A}} |\nabla \log {\mathfrak{f}}(\vp_\ep) |^2 i dt\wedge d\bar t\wedge \omega^n
\nonumber\\
& \leq & C' \int_{P_{\ep, A}} |\nabla_{\ep} \log {\mathfrak{f}}(\vp_{\ep}) |^2 i dt\wedge d\bar t\wedge \omega^n \leq C''_A,
\end{eqnarray}
where we used the fact $\omega_{\ep} \leq c \omega$ for a uniform constant $c$.
Moreover, if we take
$$F_{\ep,A}(t): = \int_{P_{\ep, A} \bigcap X_{t}} |\nabla {\mathfrak{f}} (\vp_{\ep}) |^2 \omega^n, $$
then Fatou's lemma implies
\begin{equation}
\label{com-019}
\int_0^1 \liminf_{\ep} F_{\ep, A} (t) dt \leq \liminf_{\ep} \int_0^1 F_{\ep,A} (t) dt \leq C''_A.
\end{equation}
Therefore, for almost everywhere $t\in [0,1]$, there exist a constant $C_{t}$
and a subsequence $ F_{\ep_{\ell}, A} $ such that
$$\lim_{{\ell}\rightarrow +\infty} F_{\ep_{\ell}, A} (t)\leq C_{t}. $$
Hence we have the following result.
\begin{lemma}
\label{lem-com-002}
Suppose the Mabuchi functional is $\ep$-affine along the geodesic.
Then there exists a constant $C$ (possibly depending on $t$ and $A$),
and a subsequence of volume elements ${\mathfrak{f}}(\vp_{\ep_{\ell}}) $
satisfying
\begin{equation}
\label{com-020}
\int_{P_{\ep_{\ell}, A} \bigcap X_{t}} |\nabla {\mathfrak{f}} (\vp_{\ep_{\ell}}) |^2 \omega^n \leq C,
\end{equation}
for almost everywhere $t\in [0,1]$ and any ${\ell}$ large enough.
\end{lemma}
Thank to Theorem (\ref{thm-l2-001}), we can further assume
that ${\mathfrak{f}}(\vp_{\ep_{\ell}}) \rightarrow {\mathfrak{f}}(\vp)$ in $L^2$ on $X_t$,
possibly after passing to a subsequence.
\subsection{ Positive lower bound }
To deal with the non-degeneracy of the fiber-wise volume element of ${\mathcal G}$,
we would like to utilise
the partial $W^{1,2}$-estimate obtained in equation (\ref{com-020}) and the $L^2$ convergence of the volume elements.
However, the difficulty is that the integral on the LHS of this equation is not taken on the whole manifold $X$,
and the integration domain varies with respect to $\ep$ and $A$.
In order to overcome this difficulty, we first investigate a local model as follows.
Suppose $f_{{\ell}}$ is a sequence of positive smooth functions on the domain $D: = [0,1]^m \subset \mathbb{R}^m$
with uniformly bounded $L^{\infty}$-norm,
and $f$ is an $L^{\infty}$ non-negative function on $D$ such that
$f_{\ell}$ converges to $f$ in $L^2$-norm.
We further assume that
the function $f$ satisfies the gap phenomenon, namely,
there exists a constant ${\kappa}_0 >0$ such that we have
$$ \{ f> {\kappa}_0\} \bigcup \{ f =0 \} = D, $$
up to a set of measure zero.
In the following, we set
$$ P: = \{ f > {\kappa}_0 \}; \ \ \ P^c: = \{ f =0 \},$$
and assume $\mu(P)>0$.
Let ${\kappa}$ denote a continuous function on $D$ and we set
$$ P_{{\ell}, {\kappa}}: = \{ x\in D; \ \ \ f_{\ell}(x) > {\kappa} \}. $$
Then the following result is crucial.
\begin{prop}
\label{prop-nd-001}
Assume $\max {\kappa} < {\kappa}_0 /4$ on $D$.
Suppose there exists a constant $C >0$, such that the following estimate holds
\begin{equation}
\label{nd-001}
\int_{P_{{\ell}, {\kappa}}} |\nabla f_{\ell} |^2 < C,
\end{equation}
for a fixed ${\kappa}$ and all ${\ell}$ large enough.
Then $f > {\kappa}_0$ almost everywhere on $D$.
\end{prop}
First we will prove that Proposition (\ref{prop-nd-001})
holds in $\mathbb{R}$, namely, we assume $D = [0,1]$,
and then the following fact is clear by H\"older's inequality.
\begin{lemma}
\label{lem-nd-001}
Suppose $u$ is a smooth positive function on $[0,1]$, and we assume
that $ u(a) > 2k $ and $u(b) < k $ for some $0\leq a < b \leq 1$ and a constant $k >0$.
Then we have
$$ |b -a | > \frac{k^2}{ \int_a^b ( f'(t) )^2 dt }. $$
\end{lemma}
Suppose $E_1, E_2 $ are two subsets of $D$,
and we denote $d(E_1, E_2)$ by the distance between them
$$d(E_1, E_2): = \inf_{x_1\in E_1, x_2\in E_2} d (x_1, x_2). $$
Then the following observation is crucial.
\begin{lemma}
\label{lem-nd-0010}
Suppose $P, Q$ are two non-empty disjoint subsets of the interval $[0,1]$.
Assume that the union of $P, Q$ is the whole interval $[0,1]$ up to a set with measure zero.
Then we have $d(P, Q)=0$.
\end{lemma}
\begin{proof}
We will prove by contradiction.
Suppose the distance between the two sets is positive as
$$ d(P, Q) > \delta >0. $$
\bigskip
Take a large number $m\in \mathbb{Z}$ satisfying $ \frac{1}{m} < \frac{\delta}{100} $.
For each point $p\in P$, we define the following open interval as
$$ \mathcal{I}_{p,m}: = ( p - \frac{1}{m}, p+ \frac{1}{m}), $$
and similarly for each point $q\in Q$
$$ \mathcal{I}_{q, m} := (q- \frac{1}{m}, q+ \frac{1}{m}). $$
We note that $\mathcal{I}_{p,m}$ is disjoint from the set $Q$ for each $p\in P$,
and $\mathcal{I}_{q,m}$ is disjoint from the set $P$ for each $q\in Q$ by our assumptions.
Therefore, the following two unions
$$ \mathcal{U}: = \bigcup_{p\in P} \mathcal{I}_{p,m},\ \ \ \mathcal{V}: = \bigcup_{q\in Q} \mathcal{I}_{q,m}$$
are mutually disjoint open subsets of the interval $[0,1]$.
Moreover, there is a subset $E$ of $[0,1]$ with measure zero satisfying
$$P\bigcup Q \bigcup E = [0,1]. $$
We claim that $E$ is contained in the union $\mathcal{U}\bigcup \mathcal{V}$.
Otherwise, there is a point $a\in E$ such that we have
$$ d(\{a\}, P) > \frac{1}{2m}, \ \ \ d(\{a\}, Q) > \frac{1}{2m}. $$
Then the open interval $(a- \frac{1}{2m}, a + \frac{1}{2m})$ must be contained in $E$,
but this is impossible since $E$ has measure zero.
Therefore, we have proved that the union $\mathcal{U}\bigcup \mathcal{V}$ is exactly the interval $[0,1]$,
which is impossible since $[0,1]$ is a connected set.
\end{proof}
After passing to a subsequence, we can assume that $f_{\ell} $ converges to $f$ almost everywhere
on $[0,1]$ due to the $L^2$ convergence.
Therefore, there is a subset $E\subset [0,1]$ with measure zero such that
$f_{\ell} \rightarrow f$ in the poinwise sense outside of $E$.
Then we are ready to prove the $1$-dimensional case as follows.
\begin{lemma}
\label{lem-nd-002}
Proposition (\ref{prop-nd-001})
holds in one dimension.
\end{lemma}
\begin{proof}
Thanks to Lemma (\ref{lem-nd-0010}), everything boils down to prove
$$ d(P- E, P^c - E) > 0, $$
if $P^c$ is non-empty.
We will prove by contradiction again.
If not, then there exists a sequence of pairs $(a_j, b_j)\in ( P-E ) \times (P^c -E) $
such that $|a_j - b_j| \rightarrow 0 $ as $j\rightarrow \infty$.
Without loss of generality, we assume $a_j < b_j$ in the following.
Fix one interval $ [a_j, b_j] $.
For all ${\ell}$ large enough, we have
$$f_{\ell} (a_j) > \frac{3{\kappa}_0}{4}; \ \ \ \ f_{{\ell}} (b_j) < \frac{{\kappa}_0}{4}.$$
Therefore, there exists a point $c_j\in [a_j, b_j]$ such that $f_{\ell} (c_j) = \frac{{\kappa}_0}{2}$ by the continuity of $f_{{\ell}}$.
Moreover, the point $c_j$ can be chosen close enough to $a_j$ such that we have
\begin{equation}
\label{nd-0010}
f_{\ell}|_{ [ a_j, c_j ]} > \frac{{\kappa}_0}{4},
\end{equation}
since the first derivative $f'_{\ell}$ is bounded.
Hence we have $ [a_j, c_j]\subset P_{{\ell}, {\kappa}}$ as $\max{\kappa} < {\kappa}_0 /4$.
Then the following inequality holds
by Lemma (\ref{lem-nd-001}) and equation (\ref{nd-001})
\begin{equation}
\label{nd-002}
|b_j - a_j | > |c_j - a_j | \geq \frac{{\kappa}^2_0}{16C}.
\end{equation}
However, this contradicts to the fact that $d(a_j, b_j)\rightarrow 0$ as $j\rightarrow \infty$,
and then our result follows.
\end{proof}
\begin{proof}
[Proof of Proposition (\ref{prop-nd-001})]
We will use induction on the dimension $m$.
Lemma (\ref{lem-nd-002}) implies that the result is true for $m=1$,
and we assume that Proposition (\ref{prop-nd-001}) holds on $\mathbb{R}^{m-1}$ for some integer $m\geq 2$.
Write the coordinate $x\in [0,1]^m$ as
$ (x_1,\cdots, x_{m-1}, x_m) = (x', y ) $, where $x' = (x_1, \cdots, x_{m-1})$ and $y = x_m$.
Denote $X_y$ the $(m-1)$-dimensional slice by
$$ X_y: = \{ x \in [0,1]^m; \ \ \ x = (x', y) \}.$$
Consider the $L^2$ difference of $f_{\ell}$ and $f$ on each slice as
$$ F_{{\ell}}(y): = \int_{X_y} |f_{\ell} - f|^2 dx' $$
Then the $L^2$-convergence of $f_{\ell} -f$ on $[0,1]^m$
implies that $F_{{\ell}}$ converges to zero
as $L^1$ functions on $[0,1]$,
and then it converges in measure.
By the Riesz-Lebesgue Theorem, it follows that
there exists a subsequence $F_{{\ell}_k}$ such that
$$ \lim_{{\ell}_k\rightarrow \infty } F_{{\ell}_k} (y) \rightarrow 0, $$
for almost everywhere $y\in [0,1]$.
We emphasis that this subsequence does not depend on $y$.
Therefore, after re-writing the sub-index,
we can assume
\begin{equation}
\label{nd-0025}
|| f_{\ell} - f ||_{L^2(X_y)} \rightarrow 0, \ \ \ \ {\ell} \rightarrow \infty
\end{equation}
for almost everywhere $y\in [0,1]$.
Moreover,
equation (\ref{nd-001}) can be written as
\begin{equation}
\label{nd-003}
\int_0^1 \left( \int_{X_y\bigcap P_{{\ell}, {\kappa}}} |\nabla f_{\ell} |^2 dx' \right) dy < C
\end{equation}
Then by Fatou's lemma (as we argued in equation (\ref{com-019})),
there exists a constant $C_y > 0 $ and a subsequence $f_{\ell}$, possibly depending on $y$, satisfying
\begin{equation}
\label{nd-003}
\int_{X_y\bigcap P_{{\ell}, {\kappa}}} |\nabla_{x'} f_{\ell} |^2 \leq \int_{X_y\bigcap P_{{\ell}, {\kappa}}} |\nabla f_{\ell} |^2 dx' < C_y,
\end{equation}
for almost everywhere $y\in [0,1]$ and all ${\ell}$ large.
Furthermore, the gap phenomenon (either we have $f> {\kappa}_0$ or $f =0$ almost everywhere )
must be satisfied on $X_y$ for almost everywhere $y \in [0,1]$.
Up to this point, all conditions in Proposition (\ref{prop-nd-001}), possibly except $\mu(P\bigcap X_y)>0$, are satisfied
on such a slice $X_y$ and a subsequence $f_{\ell}$.
Then for almost everywhere $y\in [0,1]$, from our induction hypothesis,
we conclude that either one of the following two cases happens:
\begin{enumerate}
\item[(1)]
$f> {\kappa}_0$
almost everywhere on $X_y$;
\smallskip
\item[(2)]
$f = 0$
almost everywhere on $X_y$.
\end{enumerate}
\bigskip
In fact, we claim that the $\mathbf{Case\ (1)}$ occurs for almost everywhere $y\in [0,1]$,
and then our result follows.
Otherwise, the following set
$$ S: = \{ y\in [0,1];\ \ \ f = 0 \ \ \emph{a.e. on $X_y$} \}, $$
will have positive measure on $[0,1]$.
Therefore, the set
$$ \mathcal{S}: = \bigcup_{y \in S} X_{y} $$
has positive measure on $[0,1]^m$.
On the other hand,
switch the direction and take another slicing as
$$ x: = ( x_1, x_2, \cdots, x_m) = ( y', x'' ),$$
where $x'' = (x_2, \cdots, x_m)$.
Repeating our previous argument on this new slicing,
we can also conclude that for almost everywhere $y' \in [0,1]$,
either $f> {\kappa}_0$ or $f = 0$ as an $L^{\infty}$ function on the slice $X_{y'}$.
However, the measure of
the following set on each slice $X_{y'} \approx [0,1]^{m-1}$
$$ \mathcal{S}_{y'}: = \bigcup_{y \in S}\left( X_{y'} \bigcap X_y \right) $$
is the same for different $y'$,
and then $\mu(\mathcal{S}_{y'})$ must be positive by Fubini's Theorem.
In other words, the set $ P^c \bigcap X_{y'} $ has positive measure for each $y' \in [0,1]$.
Therefore, we conclude that $f = 0$ on $X_{y'}$ for almost everywhere $y' \in [0,1]$.
This contradicts to the fact that $\mu(P) > 0$, and our claim follows.
\end{proof}
Now we are going to prove the main theorem in this section.
\begin{proof}
[Proof of Theorem (\ref{thm-ch-001})]
For almost everywhere $t\in [0,1]$,
we fix a fiber $X_t$ such that the estimate in Lemma (\ref{lem-com-002}) holds for a sequence
$ {\mathfrak{f}}(\vp_{\ep_{\ell}})$.
Then we will prove that the volume element $ {\mathfrak{f}} (\vp) = \omega_{\vp}^n / \omega^n$
is bounded below by the gap ${\kappa}_0$ on $X_t$ as an $L^{\infty}$ function.
Hence the restriction of the geodesic ${\mathcal G}|_{X_t}$,
as a metric on the fiber, must have a lower bound
determined by ${\kappa}_0$ and its uniform upper bound.
We will prove by contradiction.
Suppose the set $ P^c = \{ {\mathfrak{f}}(\vp) =0 \}$ has a positive measure.
As before, we denote $P$ by the set
$$ P = \{ {\mathfrak{f}} (\vp) > {\kappa}_0 \},$$
and it also has a positive measure.
Recall that $P_{\ep, A}$ is the subset of $\Gamma\times X$ where $f_{\ep}$ is larger than the auxiliary function $\log\frac{e^{\chi -A }}{\omega^n}$,
or equivalently,
$$ P_{\ep, A}\bigcap X_t = \left\{ x\in X_t;\ \ \ {\mathfrak{f}}(\vp_{\ep} (t, x)) > {\kappa} \right\}, $$
where ${\kappa} = e^{\chi - A} / \omega^n $ is the auxiliary function (equation (\ref{kappa})).
When the constant $A$ is large enough, we can assume $\max_{X_t}{\kappa} < {\kappa}_0 /4$ for all $t\in [0,1]$.
As a compact connected K\"ahler manifold, the fiber $X_t$ has an open covering by holomorphic coordinate charts,
namely, we have
$$ X_t = \bigcup_{ j =1}^N U_j, $$
and the local trivialisation map $\pi_j: B_j \rightarrow U_j$,
where $B_j \subset \mathbb{C}^n $ is an open ball.
Without loss of generality, we assume that each ball $B_j$ is centred at the origin of $\mathbb{C}^n$
and has radius larger than $2$.
In fact, we can further assume that the manifold $X_t$ is covered by the union of $V_j \subset U_j$,
where $V_j $ denotes the open set
$$ V_j: = \pi_j (B), $$
for the unit ball $B\subset \mathbb{C}^n$.
Therefore, there exists at least one $j$, such that we have on the coordinate $U_{j}$
\begin{equation}
\label{nd-004}
\mu\left( \{{\mathfrak{f}} (\vp)=0 \} \bigcap V_{j} \right) >0.
\end{equation}
Then we claim that among all these $j$'s, there exists at least one $j'$
such that the set $P^c$ does not cover the whole unit ball, namely,
we have on $U_{j'}$
\begin{equation}
\label{nd-005}
a_{2n} > \mu\left( \{{\mathfrak{f}} (\vp)=0 \} \bigcap V_{j'} \right) >0.
\end{equation}
If not, then we can find a non-empty subset $\Lambda\subsetneqq \{1,\cdots, N \}$
such that we have ${\mathfrak{f}} = 0$ pointwise a.e. on $V_j$ for each $j\in \Lambda$
and ${\mathfrak{f}} > {\kappa}_0$ pointwise a.e. on $V_k$ for each $k\in \Lambda^c$.
Take two open sets as
$$ \mathcal{U}: = \bigcup_{j\in \Lambda} V_j; \ \ \ \ \mathcal{V}: = \bigcup_{k\in \Lambda^c} V_k.$$
Then we note that ${\mathfrak{f}} = 0$ pointwise a.e. on ${\mathcal U}$ and ${\mathfrak{f}} > {\kappa}_0$ pointwise a.e. on ${\mathcal V}$.
This implies that
the intersection of the two open sets ${\mathcal U}\bigcap {\mathcal V} $ is empty.
Otherwise, the intersection is an open set with positive measure,
and we both have ${\mathfrak{f}} = 0$ and ${\mathfrak{f}} > {\kappa}_0$ pointwise a.e. on ${\mathcal U}\bigcap {\mathcal V}$,
which is not possible. Therefore, we have
\begin{equation}
\label{nd-006}
{\mathcal U}\bigcap {\mathcal V} = \emptyset; \ \ \ {\mathcal U} \bigcup {\mathcal V} = X_t,
\end{equation}
but this is also impossible, since $X_t$ is a compact connected manifold
and ${\mathcal U}, {\mathcal V}$ are both non-empty finite union of coordinate charts.
Hence our claim follows.
Pick up such a $j'$ that equation (\ref{nd-005}) holds, and then we take
$$ f_{\ell}: = {\mathfrak{f}} (\vp_{\ep_{\ell}})|_{U_{j'}}, \ \ \ f : = {\mathfrak{f}}(\vp)|_{U_{j'},} $$
and consider the $m^{th}$-cube $D: = [-1,1]^m $ in $U_{j'}$.
Combined with equation (\ref{com-020}) and (\ref{nd-005}),
all the conditions in Proposition (\ref{prop-nd-001})
are satisfied on this domain $D$ for the sequence $f_{\ell}$.
Therefore, we conclude that $f > {\kappa}_0$ almost everywhere on $D$.
However, this contradicts to the fact that $\{ f =0 \}$ has positive measure on $V_{j'} \subset D$,
and our main result follows.
\end{proof}
\section{Applications}
As one application of the main result (Theorem (\ref{thm-ch-001})),
we will prove that the Mabuchi functional ${\mathcal K}_{\ep}: = {\mathcal M}(\vp_{\ep})$ along the $\ep$-geodesic
converges to the Mabuchi functional ${\mathcal M}(\vp)$ along the geodesic,
not only pointwise on each fiber (Corollary (\ref{cor-ae-001})), but also in its complex Hessian.
For the first step, we will take a closer look at equations (\ref{com-002}) as follows.
Recall that the $\ep$-geodesic ${\mathcal G}_{\ep}$ is uniformly ${\mathcal C}^{1,\bar 1}$ on $\Gamma\times X$.
Therefore, we have
\begin{equation}
\label{app-001}
C_2 > g_{t\bar t} - g^{\bar\beta{\alpha}} g_{{\alpha}\bar t} g_{t\bar\beta} = \frac{\ep\omega^n}{ \omega_{\ep}^n },
\end{equation}
for some uniform constant $C_2$.
For the same reason,
the eigenvalues of $\omega_{\ep}$ are bounded from above by a uniform constant $C_0 >0$.
Moreover, at a particular point $p\in\Gamma\times X$, we can write
$$ \omega = \sum_{ j=1}^n dz^j \wedge d\bar z^j; \ \ \ \omega_{\ep} = \sum_{j=1}^n \lambda_j dz^j \wedge d\bar z^j. $$
where $0 < \lambda_1 \leq \cdots \leq \lambda_n$
are the $n$ eigenvalues of the metric $\omega_{\ep}$ at this point.
Hence we have the following inequality.
\begin{eqnarray}
\label{app-002}
\mathop{\rm tr}\nolimits_{\omega_{\ep}} \omega &=& \sum_{k=1}^n \frac{1}{\lambda_i}
\nonumber\\
&\leq& \frac{n \lambda_2\cdots \lambda_n}{\lambda_1 \cdots \lambda_n}
\nonumber\\
&\leq & \frac{n C_0^{n-1}}{\mathop{\rm det}\nolimits(g_{{\alpha}\bar\beta})} \leq \frac{nC_2 C_0^{n-1}}{\ep}.
\end{eqnarray}
Equipped equation (\ref{app-002}) into the RHS of equation (\ref{com-0022}), we have
\begin{eqnarray}
\label{app-003}
- \ep \int_{X_t} ( \mathop{\rm tr}\nolimits_{\omega_{\ep}}Ric\omega ) \omega^n &\geq& - \ep C \int_{X_t} ( \mathop{\rm tr}\nolimits_{\omega_{\ep} }\omega )\omega^n
\nonumber\\
& \geq& -\ep C \int_{P_{\ep, A_1}\bigcap X_t } ( \mathop{\rm tr}\nolimits_{\omega_{\ep} }\omega )\omega^n - C_4 \mu \left( \{ {\mathfrak{f}}(\vp_{\ep}) \leq {\kappa}_1 \} \right)
\nonumber\\
&\geq & -\ep C_{A_1} - C_4 \mu \left( \{ {\mathfrak{f}}(\vp_{\ep}) \leq {\kappa}_1 \} \right),
\end{eqnarray}
where we fix a constant $A_1$ large enough such that
$$\max_{X_t} {\kappa}_1 < {\kappa}_0/2. $$
Then the first factor on the RHS of equation (\ref{app-003}) converges to zero.
Moreover, Theorem (\ref{thm-ch-001}) implies that the measure $ \mu( \{ {\mathfrak{f}}(\vp_{\ep}) \leq {\kappa}_1 \} )$
converges to zero as $\ep\rightarrow 0$, possibly after passing to a subsequence.
Therefore, we conclude the following result.
\begin{theorem}
\label{thm-app-001}
Suppose the Mabuchi functional is $\ep$-affine along a geodesic.
Then for almost everywhere $t\in[0,1]$, we have on the fiber $X_t$
$$ dd^c{\mathcal K}_{\ep}|_{t} \rightarrow 0, $$
as $\ep\rightarrow 0$, possible after passing to a subsequence.
In particular, we have the following convergence of the $L^2$-norms
\begin{equation}
\label{app-005}
\int_{X_t} || \bar\partial_X v ||_{\omega_{\ep}}^2 \omega_{\ep}^n, \int_{X_t} || \bar\partial v ||_{{\mathcal G}_{\ep}}^2 \omega_{\ep}^n \rightarrow 0,
\end{equation}
as $\ep\rightarrow 0$ for this subsequence.
\end{theorem}
\begin{proof}
The complex Hessian of ${\mathcal K}_{\ep}$ can be computed as
\begin{equation}
\label{app-004}
dd^c {\mathcal K}_{\ep}(t) = \frac{\underline{R}}{n+1}\int_{X_t} {\mathcal G}_{\ep}^{n+1}
+ \int_{X_t} ( dd^c f_{\ep} - Ric\omega) \wedge {\mathcal G}_{\ep}^{n}.
\end{equation}
Repeat the previous computation on the factor
$$ ( dd^c f_{\ep} - Ric\omega) \wedge {\mathcal G}_{\ep}^{n}. $$
Equipped the estimate (\ref{app-003}) into (\ref{com-0022}), and then we have .
\begin{eqnarray}
\label{com-0150}
&& \left( dd^c \log\frac{\omega_{\ep}^n }{ \omega^n } -Ric\omega \right)\wedge {\mathcal G}_{\ep}^n + C'_{A_1} {\mathcal G}_{\ep}^{n+1}
\nonumber\\
&\geq& \left( ||\bar\partial_X v ||^2_{\omega_{\ep}} + \ep e^{-f_{\ep}} |\nabla_{\ep} f_{\ep}|^2 \right) i dt\wedge d\bar t \wedge \omega_{\ep}^n,
\end{eqnarray}
where $$C'_{A_1}: = C_{A_1} + \chi_{ \{ {\mathfrak{f}}(\vp_{\ep}) \leq {\kappa}_1\} } \ep^{-1} C_4 $$ is a function on $\Gamma\times X$.
Then our previous argument shows that
the following energy
$$ {\mathcal K}_{\ep} (t) - \ep \left( \int_{X_t} C'_{A_1} \right) t(1-t) $$
is convex on the unit interval,
and it converges to an affine function point-wise as $\ep\rightarrow 0$.
Therefore, its second derivative
converges to zero for almost everywhere $t\in [0,1]$.
Moreover, we have seen
$$ \mu \{{\mathfrak{f}}(\vp_{\ep}) \leq {\kappa}_1\} \rightarrow 0, \ \ \ \ep\rightarrow 0, $$
possibly after passing to a subsequence on the fiber $X_t$, and then our result follows.
\end{proof}
By utilizing the estimate in the above Theorem, i.e. $\bar\partial v_{\ep} \rightarrow 0$ in the $L^2$ sense, one may expect that
$v_{\ep}$ would converge to a vector field $v_{\infty}$, and this limit $v_{\infty}$ is holomorphic on $\Gamma\times X$.
However, this is still unclear to us since the fiberwise volume element $\omega_{\ep}^n$ may not converge uniformly to the
volume element $\omega^n_{\vp}$ of the geodesic ${\mathcal G}$.
Up to this stage, we can only conclude the holomorphicity of $v_{\infty}$ under some special cases.
\subsection{Special cases}
Suppose the two boundaries $\vp_0, \vp_1$ of ${\mathcal G}$ are both non-degenerate energy minimizers of ${\mathcal M}$.
Then ${\mathcal M}(\vp_t)$ keeps to be a constant along this geodesic, and it satisfies the $\ep$-affine condition automatically.
Therefore, our Theorem (\ref{thm-ch-001}) implies that the geodesic ${\mathcal G}$ is fiberwise uniformly non-degenerate,
and then the regularities of ${\mathcal G}$ can be improved by the work of He-Zeng(\cite{HZ}).
Then we recover one of our result in \cite{LL}.
\begin{theorem}[L.]
\label{thm-app-002}
Suppose the two boundaries of a $C^{1,\bar 1}$-geodesic ${\mathcal G}$ are both non-degenerate energy minimizers of ${\mathcal M}$.
Then the geodesic ${\mathcal G}$ is generated by a holomorphic vector field.
\end{theorem}
On the other hand, we can assume that the K\"ahler manifold $X$ satisfies $c_1(X)< 0$ or $c_1(X) = 0$.
Then our computation in Section (4) would recover Chen's estimates in \cite{C00}, under the $\ep$-affine condition.
In conclusion, we can infer the following result by a similar argument as in Section (6), \cite{C00}.
\begin{theorem}
\label{thm-app-003}
Suppose the manifold $X$ satisfies $c_1(X)< 0$ or $c_1(X) = 0$.
Assume that the Mabuchi functional ${\mathcal M}$ is $\ep$-affine along a $C^{1,\bar 1}$-geodesic ${\mathcal G}$.
Then ${\mathcal G}$ is generated by a holomorphic vector field.
\end{theorem}
Finally, we would like to emphasis that the boundaries of the geodesic ${\mathcal G}$ are assumed to be smooth and non-degenerate in our set up.
Therefore, one possible way to utilize the $L^2$-convergence of $\bar\partial v_{\ep}$ is to consider their behavior close enough to the boundary.
Hence we will end up with the following observation, which may be useful in our later consideration.
\begin{prop}
\label{prop-app-001}
Suppose ${\mathcal G}$ is a ${\mathcal C}^{1,\bar 1}$-geodesic connecting two K\"ahler potentials $\vp_0, \vp_1\in {\mathcal H}$.
Then its fiberwise volume element has the following convergence near the boundaries
$$ {\mathfrak{f}}(\vp_t) \rightarrow {\mathfrak{f}}(\vp_0); \ \ \ {\mathfrak{f}}(\vp_s) \rightarrow {\mathfrak{f}}(\vp_1), $$
as $t\rightarrow 0$ and $s\rightarrow 1$ in the $L^2$-sense.
\end{prop}
\begin{proof}
By a Theorem proved in Chen-Tian (Theorem 7.1.1, \cite{CT}),
it is enough to show
$$ \lim_{t\rightarrow 0} \int_X {\mathfrak{f}}(\vp_t) \log {\mathfrak{f}}(\vp_t) \rightarrow \int_X {\mathfrak{f}}(\vp_0) \log {\mathfrak{f}}(\vp_0), $$
and
$$ \lim_{s\rightarrow 1} \int_X {\mathfrak{f}}(\vp_s) \log {\mathfrak{f}}(\vp_s) \rightarrow \int_X {\mathfrak{f}}(\vp_1) \log {\mathfrak{f}}(\vp_1). $$
In other words, the entropy $H(\vp_t) (H(\vp_s))$ converges to $H(\vp_0) (H(\vp_1))$ as $t\rightarrow 0$ and $s\rightarrow 1$.
This is true because the Mabuchi functional ${\mathcal M} $ is convex and continuous up to the boundaries of ${\mathcal G}$.
\end{proof}
\begin{bibdiv}
\begin{biblist}
\bib{Ma}{article}{
author={ Mabuchi, T. },
title={A functional integrating Futaki's invariant},
journal={Proc. Japan. Acad},
volume={61 Ser. A},
date={1985},
number={},
pages={119-120},
}
\bib{BM}{article}{
author={ Bando, S.H. },
author={ Mabuchi, T. },
title={Uniqueness of Einstein K\"ahler metrics modulo connected group actions},
journal={Algebraic geometry, Sendai, 1985, 11-40, Adv. Stud. Pure Math.},
volume={10},
date={1987},
number={},
pages={11-40},
}
\bib{BT}{article}{
author={Bedford, E.}
author={Talyor, A.},
title={A new capacity for plurisubharmonic functions},
journal={Acta Math.},
volume={149},
date={1982},
number={},
pages={1-41},
}
\bib{Ber}{article}{
author={Berman, R.},
title={On the strict convexity of the K-energy},
journal={arXiv:1710.09075}
}
\bib{BB}{article}{
author={Berman, R.}
author={Berndtsson, B.},
title={Convexity of the K-energy on the space of K\"ahler metrics and uniqueness of extremal metrics},
journal={JAMS},
volume={30},
date={2017},
number={4},
pages={1165--1196},
}
\bib{Bo}{article}{
author={Berndtsson, Bo},
title={Subharmonicity properties of the Bergman kernel and some other functions associated to pseudoconvex domains},
journal={Ann. Inst. Fourier, Grenoble},
volume={56}
date={2006}
number={6}
page={1633-1662}
}
\bib{Bo11}{article}{
author={Berndtsson, Bo},
title={A Brunn-Minkowvski type inequality for Fano manifolds and some uniqueness theorems in K\"ahler geometry},
journal={Invent. math.},
volume={200}
date={2015}
page={149-200}
}
\bib{C00}{article}{
author={Chen, Xiuxiong},
title={The space of K\"ahler metrics},
journal={J. differential geometry},
volume={56}
date={2000}
page={189-234}
}
\bib{CC1}{article}{
author={Cheng, Jingrui}
author={Chen, Xiuxiong},
title={On the constant scalar curvature metrics, I: apriori estimates},
journal={arXiv: 1712.06697 },
volume={}
date={}
page={}
}
\bib{CC2}{article}{
author={Cheng, Jingrui}
author={Chen, Xiuxiong},
title={On the constant scalar curvature metrics, II: existence results},
journal={arXiv: 1801.00656},
volume={}
date={}
page={}
}
\bib{CC3}{article}{
author={Cheng, Jingrui}
author={Chen, Xiuxiong},
title={On the constant scalar curvature metrics, III: general automorphism group},
journal={arXiv: 1801.05907},
volume={}
date={}
page={}
}
\bib{CFH}{article}{
author={Chen, Xiuxiong},
author={ M. Feldman},
author={J. Hu},
title={ Geodesically convexity of small neighbourhood in space of K\"ahler metrics},
journal={arXiv:1805.02373},
}
\bib{CLP}{article}{
author={Chen, Xiuxiong},
author={Li, Long}
author={P\u aun, Mihai},
title={Approximation of weak geodesics and subharmonicity of Mabuchi energy},
journal={ Annales de la faculte des sciences de Toulouse Ser. 6},
volume={25},
date={2016},
number={5},
page={935-957}
}
\bib{CPZ}{article}{
author={Chen, Xiuxiong},
author={P\u aun, Mihai},
author={Yu, Zeng}
title={On deformation of extremal metrics },
journal={arXiv:1506.01290},
}
\bib{CT}{article}{
author={Chen, Xiuxiong},
author={Tian, Gang}
title={Geometry of K\"ahler metrics and foliations by holomorphic disks},
journal={Publ. Math. Inst. Hautes \'Etudes Sci. },
number={107},
date={2008},
page={1-107}
}
\bib{LD}{article}{
author={Darvas, T.},
author={Lempert, L. },
title={Weak geodesics in the space of K\"ahler metrics},
journal={Math. Res. Lett.},
volume={19},
date={2012}
number={5},
pages={1127-1135}
}
\bib{Dem}{article}{
author={J.P. Demailly},
title={Regularization of closed positive currents and Intersection Theory},
journal={J. ALG. GEOM},
volume={1},
date={1992},
page={361-409}
}
\bib{GT}{article}{
author={Gilbarg, D.},
author={Trudinger, N.S.}
title={Elliptic partial differential equations of second order},
journal={Springer},
date={2001}
}
\bib{HZ}{article}{
author={He, Weiyong },
author={Zeng, Yu },
title={Constant scalar curvature equation and the regularity of its weak solution},
journal={Communications on Pure and Applied mathematics},
volume={72},
date={2019},
number={2},
pages={422-448},
}
\bib{LL}{article}{
author={Li, Long},
title={The strict convexity of the Mabuchi functional for energy minimizers },
journal={to appear in Annales de la faculte des sciences de Toulouse.}
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
Locomotion of microscopic organisms such as bacteria or sperms is governed by laws which are different from those governing the world tangible by humans. Due to the dominance of viscous drag over inertia and the ensuing time-independence of the Stokes equations, a successful swimming strategy has to break the time-reversal symmetry \cite{Purcell1977}. Several theoretical models \cite{Blake1971, Felderhof2006, EarlPooleyRyder2007, NajafiGolestanian2004, AvronKennethOaknin2005, DowntonStark2009, PandeSmith2015, ZieglerHubertVandewalle2019, Wang2019, DaddiLisickiHoell2018, DaddiKurzthalerHoell2019, RizviFarutinMisbah2018}, experimental realizations \cite{ DreyfusBaudryRoper2005, AhmedLuNourhani2015, GrosjeanLagubeauDarras2015, GrosjeanHubertLagubeau2016, GrosjeanHubertCollard2018, ZhengDaiWang2017, HamiltonPetrovWinlove2017, BryanShelleyParish2017, HamiltonGilbertPetrov2018, Collard2020} and simulations \cite{BeckerKoehlerStone2003, ZottlStark2012, PicklGoetzIglberger2012, SukhovZieglerXie2019, PicklPandeKoestler2017} have been employed to scrutinize the details of locomotion under these laws. Overall, due to the absence of inertia, the force to self-propel exerted by a microswimmer on its surrounding is balanced by friction, and force monopoles are not present. Consequently, a self-propelled swimmer typically induces a dipolar flow field, allowing to classify swimmers as pushers or pullers. In addition to it, higher order flow fields are induced \cite{SpagnolieLauga2012}, which, nonetheless, induce hydrodynamic interactions between distant swimmers. Interestingly, living microswimmers are capable of not only sensing these flows, but also reacting to them, which is associated with important physiological functions \cite{Shen2012, Guasto2011, Uppaluri2012, Mathijssen2019}.
Hydrodynamic interactions have been investigated for a number of model microswimmers \cite{PooleyAlexanderYeomans2007, AlexanderPooleyYeomans2008, AlexanderPooleyYeomans2009, FarzinRonasiNajafi2012, KurodaYasudaKomura2019, IshikawaSimmondsPedley2006, MirzakhanlooJalaliAlam2018}, leading to a consensus that, in a first approximation, a swimmer can be considered as a passive entity subject to the average flow produced by nearby swimmers. With this approximation as a starting point, suspensions of many swimmers have been investigated using simulations \cite{Bardfalvy2020, EvansIshikawaYamaguchi2011} and theoretical approaches \cite{Hoell2018, Reinken2018}.
Additionally, at the single swimmer level an active component to the interactions, depending on the time-resolved stroke of both swimmers, has been reported throughout the literature \cite{PooleyAlexanderYeomans2007, AlexanderPooleyYeomans2009, FarzinRonasiNajafi2012, KurodaYasudaKomura2019}. In opposition to the expectation that passive interactions are sufficient, it has been reported that the active component actually dominates for swimmer separations smaller than a threshold value depending on the details of the stroke \cite{PooleyAlexanderYeomans2007}.
However, several competing models for the active interactions are currently discussed, providing different predictions for the scaling and the sign of the interaction effects \cite{PooleyAlexanderYeomans2007, AlexanderPooleyYeomans2009, FarzinRonasiNajafi2012}. Given that most models rely on the same basic assumptions, the origins of these differences have not been clearly established thus far.
Most of these models consider interacting bead-based microswimmers subject to a prescribed swimming stroke. Yet, there is a second family of bead-based microswimmers where the swimmer arms are replaced with elastic springs, and oscillating external driving forces induce the swimming stroke \cite{Felderhof2006, PandeSmith2015, ZieglerHubertVandewalle2019}. These so-called bead-spring swimmers can adapt their swimming stroke and velocity to the surrounding fluid \cite{PandeMerchantKrueger2017a} and also to the presence of other swimmers, an effect not taken into account in previous works. For a single swimmer a mapping from prescribed forces to the corresponding stroke exists, which reproduces the same swimming velocity and flow field in both approaches. Consequently, a natural question to ask is if the same correspondence applies for the interaction of two swimmers.
In this paper, we answer this question by calculating the interaction of two linear three-bead microswimmers in both the force-based (FB) and stroke-based (SB) models using a recently developed perturbative approach \cite{ZieglerHubertVandewalle2019}.
While passive interactions are equivalent in the stroke-based and force-based models, the additional active component to the interactions, depending on the swimming stroke of both species, is in general not equivalent in both approaches. Interestingly, the interactions between force-based swimmers becomes equivalent to the interactions between stroke-based swimmers if the driving frequency in the force-based model is small as compared to the inverse viscous time of the system. Hence, the interactions of stroke-based swimmers should actually be considered as special case of the interactions of two force-based swimmers. These findings are systematically discussed in the context of the existing literature \cite{PooleyAlexanderYeomans2007,AlexanderPooleyYeomans2009, FarzinRonasiNajafi2012, KurodaYasudaKomura2019}.
The analysis of the scaling laws is complemented by an in-depth discussion of the cooperative effects between swimmers. Unlike previously observed, we show that two collinear or side-by-side swimmers with parallel swimming direction typically benefit mutually from each other due to the time-averaged flow fields they produce. In simple words, a swimmer propagates faster in a pair compared to when it is alone, even if it is leading. This result, which is a consequence of vanishing Reynolds numbers, is independent of the details of the swimming stroke due to a close relation of the leading order swimming velocity and flow fields produced by each swimmer.
The remainder of the article is structured as follows: We first introduce both the force-based and stroke-based swimmer models, elaborate on the perturbative approaches and establish the mapping between both models. We then proceed in section \ref{sec:results} to calculate the average flow fields of our swimmers. Focusing on a pair of side-by-side or trailing swimmers, we then investigate the passive interactions and, consequently, the active translations and rotations, comparing the obtained results to findings in the literature. Finally, we expand our analysis to arbitrary positioning of the swimmers, and estimate the long-term behavior of a pair. The most important findings are summarized in section \ref{sec:conclusion}, which concludes the paper. Details on the perturbative calculations for the force-based and stroke-based models as well as for the proportionality of the leading order swimming velocity and flow fields are given in the Appendices A-D.
\section{Model}
\label{sec:Model}
\begin{figure}
\begin{center}
\includegraphics[scale=0.75]{./Figures/201910_Scheel_figure1_v2.eps}
\caption{Sketch of two interacting in-plane linear three-sphere swimmers.}
\label{fig_1}
\end{center}
\end{figure}
\subsection{Force-based model of the three-bead swimmer}
\label{sec:forcebasedmodel}
We consider two linear three-bead swimmers identified by subscripts $s, p \in \{I, I\!I \}$. We restrict our analysis to a two-dimensional configuration space ($d = 2$), in our case the $x$-$y$-plane, as depicted in figure \ref{fig_1}. Each swimmer consists of three identical spherical beads of radius $a$, denoted by subscripts $i, j, m \in \{1,2,3\}$. The beads are connected by identical linear harmonic springs of spring constant $k$ and equilibrium length $L$, such that the spring force between two connected beads $si$ and $sj$, acting on the latter, is given by
\begin{equation}
\vec{g}(\vec{x}) := - k (|\vec{x}| - L) \cdot \frac{\vec{x}}{|\vec{x}|},
\end{equation}
with $\vec{x} = \vec{R}_{sj} - \vec{R}_{si}$ the vector connecting both beads. We employ a convenient double index notation to identify the beads, where the first index corresponds to the swimmer and the second index to the bead number within this swimmer.
In this work we only consider swimmers with a rigid joint connecting both swimmer arms, i.e. the two swimmer arms cannot bend with respect to each other. This allows to systematically compare force-based to linear stroke-based swimmers \cite{NajafiGolestanian2004, PooleyAlexanderYeomans2007, AlexanderPooleyYeomans2009, FarzinRonasiNajafi2012}, to which this constraint is inherent.
As shown in figure \ref{fig_1}, swimmer $I$ is oriented along $\vec{n}_I = (\sin \theta_I, -\cos \theta_I)$, swimmer $I\!I$ along $\vec{n}_{I\!I} = (\sin \theta_{I\!I}, -\cos \theta_{I\!I})$ and the middle beads of both swimmers are separated by $\vec{r} = r (\sin (\varphi + \theta_I), -\cos (\varphi + \theta_I))$. Here, $\varphi$ denotes the angle enclosed between $\vec{n}_I$ and $\vec{r}$. The swimmers are immersed in a viscous fluid of viscosity $\eta$, where the Reynolds number of a single bead is assumed to be zero. The interaction between two spherical beads at sufficiently large distance is then described by the Oseen tensor
\begin{equation}
\hat{T}(\vec{x}) := \frac{1}{8 \pi \eta |\vec{x}|} \left( \hat{1} + \frac{\vec{x} \otimes \vec{x} }{|\vec{x}|^2} \right),
\end{equation}
with $\hat{1}$ the unit matrix and $\otimes$ the tensor product.
Each swimmer $s$ is driven by sinusoidal forces of frequency $\omega$ acting on the two outer beads (bead s1 and s3) along the respective adjacent arm with relative phase shift $\alpha_s$. The driving forces on swimmer $I\!I$ precede those on swimmer $I$ by a phase shift of $\gamma$:
\begin{eqnarray}
\vec{E}_{I1} (t) = A_{I1} \sin(\omega t) \vec{n}_I, \ \vec{E}_{I3} (t) = A_{I3} \sin(\omega t + \alpha_I) \vec{n}_I, \nonumber \\ \vec{E}_{\II1} (t) = A_{\II1} \sin(\omega t + \gamma) \vec{n}_{I\!I}, \ \vec{E}_{\II3} (t) = A_{\II3} \sin(\omega t + \gamma + \alpha_{I\!I}) \vec{n}_{I\!I},
\label{eq:drivingForces}
\end{eqnarray}
with $\vec{E}_{si}$ the force on the bead $si$, $A_{si}$ the respective force amplitude and $t$ the time.
We introduce a second set of independent variables for the driving forces which is required in the perturbative calculation, $A := A_{I1}, B_I := A_{I2}/A_{I1}, C := A_{I\!I 1}/A_{I1}, B_{I\!I} := A_{I\!I 2}/ A_{I\!I 1}$, with $A$ the overall strength of the driving forces and $B_I, B_{I\!I}, C$ dimensionless relative parameters.
The force on each middle bead $s2$ is determined by requesting the sum of all three forces on the beads of swimmer $s$ to vanish \cite{PandeSmith2015}, $\vec{E}_{s2}(t) := -\vec{E}_{s1} (t) - \vec{E}_{s3} (t)$, as a prerequisite for self-propulsion.
We then devise an equation of motion for all of the $n = 6$ beads,
\begin{eqnarray}
\fl
\frac{d}{dt} \vec{R}_{si} (t) = & \mu \left( \vec{E}_{si} (t) + \sum_{j \in \mathrm{NN}(i)} \vec{g}(\vec{R}_{sj}(t) - \vec{R}_{si}(t)) \right) + \nonumber \\
\fl
& \sum_{(pj) \neq (si)} \hat{T}(\vec{R}_{si}(t) - \vec{R}_{pj}(t)) \cdot \left( \vec{E}_{pj} (t) + \sum_{m \in \mathrm{NN}(j)} \vec{g}(\vec{R}_{pm}(t) - \vec{R}_{pj}(t)) \right),
\label{eq:EOM}
\end{eqnarray}
with $\vec{R}_{si}$ the position of the respective bead and $\mu = (6 \pi \eta a)^{-1}$ the Stokes mobility. $\mathrm{NN}(j)$ denotes all beads connected to bead $j$ via a spring within the geometry of the linear swimmer, and in the remaining summation at the beginning of the second line we sum over all beads $pj$ different from the bead $si$.
To obtain the swimmer behavior analytically, we employ a recently developed perturbative approach to bead-spring microswimmers \cite{ZieglerHubertVandewalle2019} (details in \ref{ch:appendixB}). Also, we solve \eqref{eq:EOM} numerically using the NDSolve function from Mathematica \cite{Mathematica2017}, with a superimposed angle spring potential ensuring the rigidity of the joint at the middle bead of each swimmer (details in \ref{ch:appendixA}).
A system of bead-spring swimmers as described above can be rescaled using the bead radius $a$ as the length unit and the viscous time $t_V := 6 \pi \eta a/k$ as the time unit, in order to identify the effective parameters \cite{ZieglerHubertVandewalle2019}. We define the dimensionless parameters $q:= a/r$ and $\nu := a/L$ encoding together with $\varphi, \theta_1$ and $\theta_2$ the swimmer geometry, and the rescaled driving frequency $\Gamma := \omega t_V$, comparing the two time scales set by the driving forces and the viscous time.
In the subsequent analysis, we will illustrate the results for the swimmer interactions using a swimmer of puller- and a swimmer of pusher-type with distinct parameters. For the puller, we choose
\begin{equation}
\frac{a}{L} = \frac{1}{6}, \ A_{s1} = A_{s3} = \frac{5}{4} ka, \ \beta_s = \frac{\pi}{2}, \ \omega t_V = 0.7402
\end{equation}
and for the pusher-type swimmer
\begin{equation}
\frac{a}{L} = \frac{1}{6}, \ A_{s1} = \frac{5}{3} k a, \ A_{s3} = \frac{5}{6} ka, \ \beta_s = \frac{\pi}{2}, \ \omega t_V = 0.7402.
\end{equation}
\subsection{Stroke-based model of the three-bead swimmer}
\label{ch:strokeBasedModel}
In the stroke-based model, we consider two swimmers with identical geometry as in the force-based model. In contrast to the force-based swimmer, the length of each swimmer arm is directly prescribed as a function of time
\begin{eqnarray}
L_{I1} (t) &= L + \xi_{I1} \sin( \omega t), \ L_{I2} (t) = L + \xi_{I2} \sin( \omega t + \beta_I), \nonumber \\ L_{\II1}(t) &= L + \xi_{\II1} \sin(\omega t + \delta), \ L_{\II2} = L + \xi_{\II2} \sin(\omega t + \delta + \beta_{I\!I}),
\end{eqnarray}
with $L$ the average arm length, $\xi_{sb}$ the corresponding arm oscillation amplitude (indices for the swimmer arms $b,c \in \{1, 2\}$) and $\beta_I$, $\beta_{I\!I}$ and $\delta$ the phase shifts within each swimmer and between both, respectively.
Similarly to the force-based model, we introduce a second set of parameters for the arm oscillation amplitudes required in the perturbative analysis, $\xi := \xi_{I1}, D_I := \xi_{I2}/\xi_{I1}, F := \xi_{I\!I 1}/\xi_{I1}, D_{I\!I} := \xi_{I\!I 2}/\xi_{I\!I 1}$, with $\xi$ the overall amplitude of the arm oscillations.
With both swimmers constrained to linear shape and to the prescribed arm lengths, the system contains six undetermined degrees of freedom, which we choose to be the positions of both middle beads, $\vec{R}_{I2}$ and $\vec{R}_{I\!I 2}$, as well as the orientations of both swimmers, $\theta_I$ and $\theta_{I\!I}$. The positions of all other beads are then given by the prescribed arm lengths and the constraint on the linear swimmer shape.
The relation of the bead velocities to the forces on the beads is given similarly as in the force-based model by
\begin{equation}
\frac{d}{dt} \vec{R}_{si} (t) = \mu \, \vec{F}_{si} (t) + \sum_{(pj) \neq (si)} \hat{T} \left(\vec{R}_{si}(t) - \vec{R}_{pj}(t)\right) \cdot \vec{F}_{pj}(t),
\label{eq:StokesLaw}
\end{equation}
where $\vec{F}_{si} (t)$ denotes the forces which are acting on the beads to enforce the prescribed strokes.
The remaining six degrees of freedom are then determined by the condition that the total vectorial force on each swimmer vanishes,
\begin{equation}
\sum_{i = 1}^3 \vec{F}_{si}(t) = 0, \ s \in \{I, I\!I\},
\label{eq:forceFreeCond}
\end{equation}
as well as the total torque on each swimmer, which is a scalar quantity in the 2D framework employed here:
\begin{equation}
(- L_{s1} (t) \vec{F}_{s1}(t) + L_{s2} (t) \vec{F}_{s3}(t)) \cdot \vec{n}^\perp_s = 0, \ s \in \{I, I\!I\}.
\label{eq:torqueFreeCond}
\end{equation}
Using the compact notation $\bm{R} = (\vec{R}_{I1}, \vec{R}_{I2}, \vec{R}_{I3}, \vec{R}_{I\!I 1}, \vec{R}_{I\!I 2}, \vec{R}_{I\!I 3})$ and $\bm{F} = (\vec{F}_{I1}, ..., \vec{F}_{I\!I 3})$, we can re-express \eqref{eq:StokesLaw} as
\begin{equation}
\frac{d}{dt} \bm{R} = \underline{\mu}(\bm{R}) \, \bm{F}.
\label{eq:StokesLawCompact}
\end{equation}
Here, $\underline{\mu} (\bm{R})$ denotes the $(n \cdot d) \times (n \cdot d)$-dimensional mobility matrix \cite{Dhont1996, ZieglerHubertVandewalle2019} and we neglect the time-dependence of $\bm{R}$ and $\bm{F}$ in our notation for the sake of brevity. Vectors in the $(n \cdot d)$-dimensional configuration space of all bead positions are denoted by bold symbols and higher order tensors on this space by underlined symbols.
In order to perturbatively calculate the swimmer behavior, we expand the undetermined variables $\vec{R}_{I2}, \vec{R}_{I\!I 2}, \theta_I$ and $\theta_{I\!I}$ as power series in $\xi$ and $q$, and from this calculate an expansion of $d \bm{R} /dt$ and $\underline{\mu}(\bm{R})$. By also expanding \eqref{eq:StokesLawCompact}, we are able to express the components of $\bm{F}$ associated to different powers of $\xi$ and $q$ in terms of the middle bead positions and swimmer orientations. The force-free and torque-free conditions, \eqref{eq:forceFreeCond} and \eqref{eq:torqueFreeCond}, close the equations obtained and by solving them we obtain the full swimmer behavior as an expansion in $\xi$ and $q$. The perturbative calculation is explained in more detail in \ref{ch:appendixC2}.
\subsection{Mapping from force-based to stroke-based model}
For a single force-based swimmer, the swimming velocity can be either calculated directly in the force-based framework \cite{ZieglerHubertVandewalle2019}, or one can equivalently extract the time-dependent arm lengths and insert them in the respective expression for the swimming velocity in the stroke-based model \cite{GolestanianAjdari2008}. Since the single swimmer behavior is uniquely defined by the time-dependent arm lengths together with the force-free condition \cite{GolestanianAjdari2008}, both ways must yield the same result. This means that it exists a mapping from the parameter space of the force-based system to the parameter space of the stroke-based system which preserves the overall swimmer dynamics. For the sake of simplicity, we restrict the mapping to the first order in $A$ and $\xi$ in the subsequent calculations, where the average arm lengths in the force-based model are still equal to the spring lengths in the mechanical equilibrium, $L$ \cite{ZieglerHubertVandewalle2019}. Hence, we assume the geometry of the force-based swimmer in its mechanical equilibrium for the time-averaged shape of the stroke-based swimmer, i.e. we assume average arm lengths $L$, and only require to map the swimming stroke, i.e. the arm oscillation amplitudes. This reproduces consistent second order swimming velocities and flow fields in both models and is hence sufficient for the subsequent comparison.
\begin{figure*}
\centering
\includegraphics[scale=0.85]{./Figures/201910_Scheel_figure2.eps}
\caption{Flow fields of a puller- and pusher-type swimmer oriented along the $y$-axis. Intermediate field, far field and flow field amplitude on the y-axis for the puller-like swimmer ((a), (b) and (c), respectively) and for the pusher-like swimmer ((e), (f) and (d), respectively). In both cases, the swimmer propagates in positive $y$-direction. In (c) and (d), the flow amplitude in the front of the swimmer ($y > 0$) is given by orange pluses and in the back of the swimmer ($y < 0$) by blue triangles, both obtained numerically. Solid lines in the respective colors denote the corresponding analytical results. The swimmer parameters are chosen as defined in section \ref{sec:forcebasedmodel}.}
\label{fig_2}
\end{figure*}
The inverse map from the stroke-based to the force-based system is not unique, since the force-based model comprises one additional parameter, namely the spring constant $k$. It gives rise to the viscous time scale $t_V$ and the rescaled driving frequency $\Gamma$, with respect to which a force-based swimmer attains the maximum driving speed at $\Gamma \approx 1$ \cite{SukhovZieglerXie2019}, and is unable to self-propel for $\Gamma \gg 1$ and $\Gamma \ll 1$ \cite{PandeMerchantKrueger2017a}. In contrast, $\omega$ has no impact on the propulsion of a stroke-based swimmer measured over one stroke \cite{GolestanianAjdari2008} and drops out in a suitable rescaling. This is because with a prescribed stroke, calculating the swimmer behavior reduces to a purely geometric problem due to the time-independence of the Stokes equations \cite{GolestanianAjdari2008}.
The natural question arising then is: Is also the interaction of two or more swimmers equivalent in the force- and stroke-based models under the above mapping, or does the interaction in the force-based model change with $\Gamma$, when the driving forces on each swimmer are such that the stroke amplitude and phase of each swimmer alone would be constant?
\section{Results and discussion}
\label{sec:results}
\subsection{Flow field of a single swimmer}
A single linear three-sphere swimmer with equal bead radii typically produces a flow field which is dominantly dipolar in the very far field and quadrupolar at intermediate distances (figure \ref{fig_2}) \cite{PooleyAlexanderYeomans2007}. The transition between both fields depends on the swimmer parameters, and appears in our case at the order of $100 \, L$.
The time-averaged dipolar flow field scales to leading order as $A^4$ or $\xi^4$, depending on the model considered, and is given in terms of the swimming stroke as
\begin{eqnarray}
\fl
\vec{u}^\mathrm{dip}(\vec{r}) = \frac{\omega a^2}{2 r^2} \frac{\xi_1 \xi_2 \left(16947 a^3-24924 a^2 L+11664 a L^2-1856 L^3\right) \left(\xi_1^2-\xi_2^2\right)}{ (3 a-4 L)^3 (7 a-4 L)^3} \times \nonumber \\ \sin (\beta) \frac{1}{2} \left[3 \frac{(\vec{r} \cdot \vec{n})^2}{r^2} - 1 \right] \frac{\vec{r}}{r},
\label{eq:dipolarFlowField}
\end{eqnarray}
with $r := |\vec{r}|$ and $\vec{n}$ the unit vector along the swimmer axis (details on the calculation in \ref{ch:appendixC1}). We have omitted the swimmer index in $\xi_1$ and $\xi_2$ since this result holds for a single swimmer.
In contrast, the average quadrupolar flow field scales to leading order as $A^2$ or $\xi^2$ and reads in terms of the stroke
\begin{eqnarray}
\fl
\vec{u}^\mathrm{quad}(\vec{r}) = -\frac{\omega a^2}{r^3}\frac{\xi_1 \xi_2 L^2 (147 a-68 L) \sin (\beta )}{(3 a-4 L) (7 a-4 L)^2} \times \nonumber \\ \frac{1}{4} \left[ 3 \frac{(\vec{r} \cdot \vec{n})}{r} \left( 5 \frac{(\vec{r} \cdot \vec{n})^2}{r^2} - 3 \right) \frac{\vec{r}}{r} - \left( 3 \frac{(\vec{r} \cdot \vec{n})^2}{r^2} - 1 \right) \vec{n} \right].
\label{eq:quadrupolarFlowField}
\end{eqnarray}
Using the mapping from the parameters of the force-based to the stroke-based model at sufficient order in $\xi$ and $A$, one obtains similar expressions for the force-based swimmer, which we do not include as they would become too large.
To understand the predominance of the quadrupolar regime in the average flow field up to distances of hundreds of swimmer lengths, we first consider the special case of a swimmer with $\xi_{1} = \xi_{2}$, i.e. equal arm oscillation amplitudes. In this case, the swimmer becomes invariant under a combined time-reversal ($t \to -t$) and parity ($\vec{r} \to -\vec{r}$) transformation (TP transformation) and the dipolar regime in the time-averaged flow field is lost \cite{PooleyAlexanderYeomans2007}. Indeed, flow fields which decay with an even order in the distance $r$ from the swimmer (dipolar, octopolar, ...) are not consistent with TP invariance. Thus, the far field of a TP invariant and self-propelled swimmer is quadrupolar, decaying as $r^{-3}$, and the dipolar regime in figure \ref{fig_2} would disappear. However, even a TP-invariant swimmer still produces a non-zero dipolar flow field that oscillates with the frequency of the swimming stroke.
For $\xi_{1} \neq \xi_{2}$, which is the case for both the puller- and pusher-type swimmer defined in section \ref{sec:Model}, the TP invariance is broken and a non-zero time-averaged dipolar flow field arises. Notably, the leading order, $\xi^2$ contribution to all even-order time-averaged flow fields still vanishes, and the first non-zero contribution arises at order $\xi^4$. The reasoning to show this extends an argument made originally by Golestanian and Ajdari \cite{GolestanianAjdari2008}. Similarly to the time-averaged swimming velocity, the amplitude of each component (dipolar, quadrupolar, ...) of the average $\xi^2$ flow field must be given by a geometric prefactor times the area in the configuration space enclosed by the swimmer's stroke (\ref{ch:appendixD}). First, this leads to the important conclusion that, at order $\xi^2$, average flow field and swimming velocity are proportional and linked by a factor depending on the average swimmer geometry only. Accordingly, knowing the swimmer's geometry, one can, at leading order, directly infer the average flow field from the swimming velocity. Second, we conclude that the proportionality coefficient needs to vanish for all even-order components (dipolar, octopolar, ...) of the average flow field, since it vanishes for the special case $\xi_{1} = \xi_{2}$ and is generic for linear three-bead swimmers with equal radii and equal average arm lengths.
In the case of sinusoidal driving forces with a single frequency, also the $\xi^3$ time-averaged flow field vanishes since it is associated to odd products of the sinusoidal stroke, which averages to zero over a stroke cycle. This explains why passive dipolar interactions therefore arise firstly at fourth order in the stroke amplitude ($\sim \xi^4/r^2$) \eqref{eq:dipolarFlowField}.
In \cite{PooleyAlexanderYeomans2007}, a stroke comprising stepwise arm contractions at constant velocity \cite{NajafiGolestanian2004} has been employed, associated to infinitely many harmonics of the base frequency $\omega$ in a Fourier decomposition. Therefore, a non-zero average dipolar flow field has already been observed at third order in $\xi$, making the quadrupolar regime less dominant than in the case of sinusoidal driving with a single pure frequency.
\begin{figure*}[h]
\centering
\includegraphics[scale=0.9]{./Figures/201910_Scheel_figure3_v4.eps}
\caption{Behavior of two linear three-bead swimmers interacting collinearly. (a) Sketch. In (b), the dependence of the boost on the phase shift $\gamma$ between both swimmers at distance $r = 500a$ is shown for two pullers. Panels (c) - (f) show the increase/decrease in the swimming velocity for all four combinations of pullers and pushers with the phase shift between both swimmers fixed to $\gamma = 0$. The headlines of each plot indicate the swimmer types with first the trailing and second the leading swimmer. Numerical results are shown by orange pluses for the leading and by blue triangles for the trailing swimmer. The solid lines in respective color represent the analytical results.}
\label{fig_3}
\end{figure*}
\subsection{Passive interaction}
While a single swimmer alone would propagate with velocity $v_0$ along its axis, in the presence of another swimmer it experiences both additional translation and rotation due to the hydrodynamic interactions. We split the swimmer's velocity into a component $v_0 + \Delta v_\mathrm{par}$ parallel to its axis $\vec{n}_s$ and into a component $v_\mathrm{ort}$ along $\vec{n}^\perp_s$ (see figure \ref{fig_1}). Additionally, the swimmer rotates with an angular velocity $\Omega$ which is defined as positive when the swimmer rotates counterclockwise. We restrict our subsequent analysis to the velocities time-averaged over one swimming cycle.
We distinguish between active and passive interaction effects by their dependence on the phase shift between both swimmers, $\gamma$ in the force-based model and $\delta$ in the stroke-based model. Considering the velocity of say swimmer $I$, passive terms are constant with respect to $\gamma$ and $\delta$. In contrast, active interactions explicitly depend on the phase shift between both swimmers.
In agreement with previous work we find that all passive interactions, in particular both translation and rotation, are solely due to the average flow field that swimmer $I\!I$ would produce if it was alone in an otherwise empty and unperturbed fluid \cite{PooleyAlexanderYeomans2007, AlexanderPooleyYeomans2009}. This result holds for both the force-based and the stroke-based model. Note that for the passive interactions the single-swimmer flow field of swimmer $I\!I$ is relevant, and not the average flow field produced by swimmer $I\!I$ in the presence of swimmer $I$. Since the difference between the two situations arises from the swimming activity of swimmer $I$, this effect gives rise to active interactions. Thus, passive interactions are, via their definition, equivalent in the force-based and stroke-based models, i.e. they agree if the force protocol and stroke are chosen such that they produce the same single swimmer behavior.
The passive translation of swimmer $I$ equals the local average flow field $\vec{u}(\vec{x})$ produced by swimmer $I\!I$. The calculation of passive rotation is more involved,
\begin{equation}
\Omega^\mathrm{pass.}_I = \dot{\theta}^\mathrm{pass.}_{I} = (\vec{n}_s \cdot \nabla) \left[\vec{n}^\perp_s \cdot \vec{u}(\vec{x}) \right]|_{\vec{x} = \vec{R}_{I 2}},
\end{equation}
as it results from the interplay of the swimmer shape and the local gradient of the average flow field.
Mediated by the passive interactions, the transition from predominantly quadrupolar to dipolar with increasing distance is also found in the interaction of two collinear swimmers, i.e. swimmers with a common axis ($\theta_1 = \theta_2 = 0, \varphi = 0$) (figure \ref{fig_3}). For side-by-side swimmers with parallel swimming direction, defined by $\theta_1 = \theta_2 = 0$ and $\varphi = \pi/2$, passive dipolar interaction effects are found only in $v^\mathrm{ort}$ and passive quadrupolar interactions only in $v^\mathrm{par}$ (figure \ref{fig_4}). This is due to the shapes of the respective flow fields (figure \ref{fig_2}). In the general case of two interacting swimmers, also passive quadrupolar rotation would be observed, resulting from the gradient of the dipolar time-averaged flow field. However, in the two most simple swimmer configurations chosen here for illustration, this term is absent due to the symmetries of the dipolar flow field.
The proportionality of the $\xi^2$ or $A^2$ time-averaged flow field and the swimming velocity has important implications for the swimmer interaction. Namely, for swimmers of given geometry, the passive interactions between the swimmers depend only on their swimming velocities and their relative positioning, but not on the details of the strokes. The factor of proportionality for the quadrupolar flow field of a single swimmer, measured in front of the swimmer along the swimmer axis $\vec{n}_s$, relative to the swimming velocity is given by
\begin{equation}
(\vec{u}^\mathrm{quad} (r \vec{n}_s) \cdot \vec{n}_s)/v_0 = \frac{3 a L^3 (68 L - 147 a)}{r^3 \left(56 L^2 -198 a L + 189 a^2\right)} \geq 0.
\end{equation}
This ratio is positive for values $L \geq 6a$ where the Oseen approximation can be expected to be sufficiently good.
Therefore, two linear swimmers in the collinear or side-by-side configuration with the same swimming direction always mutually benefit from passive interactions in the quadrupolar regime (figure \ref{fig_3}, \ref{fig_4}).
\begin{figure*}
\centering
\includegraphics[scale=0.9]{./Figures/201910_Scheel_figure4_v4.eps}
\caption{Behavior of two linear three-bead swimmers in the side-by-side configuration. (a) Sketch. In (b), the dependence of all three velocities on the phase shift $\gamma$ between both swimmers is shown for two pullers at distance $r = 500a$.
Panels (c) - (f) show the boost/slowdown along the swimmer axis which scales as $r^{-3}$ at large $r$ ($\mathrm{v^{par}}$, orange triangles), the induced velocity orthogonal to it which scales as $r^{-2}$ at large $r$ and shows a transition to $r^{-4}$ scaling in (d), (e) marked by blue circles ($\mathrm{v^{ort}}$, blue pluses) and the angular velocity $\Omega$ which scales $r^{-4}$ at large $r$ (black stars) for all four combinations of pullers and pushers. The phase shift between both swimmers is fixed to $\gamma = 0$. For equal swimmer types and $\gamma = 0$, the swimmers behave symmetrically and the values plotted hold for both swimmers (attr.: attractive), for different swimmer types the values are plotted for the swimmer highlighted in bold in the headline (CW: clockwise, CCW: counterclockwise). }
\label{fig_4}
\end{figure*}
While in some of the previous works passive interaction has been found to be equal to the average local flow field \cite{PooleyAlexanderYeomans2007, AlexanderPooleyYeomans2009} as we do, other authors have reported results for the passive interaction different from the time-averaged flow field. In particular, a quadrupolar passive interaction scaling to leading order as $L^1$ for $L\rightarrow\infty$ has been reported \cite{FarzinRonasiNajafi2012, KurodaYasudaKomura2019}. This is in contrast to the average quadrupolar flow field of a linear-three sphere swimmer which scales as $L^0$ in the same limit \cite{PooleyAlexanderYeomans2007}.
We have been able to reproduce the linear scaling in $L$ of the passive interaction in a slightly altered perturbative scheme, where the distance between the two swimmers is assumed to be a constantly $r$ throughout one swimming stroke, instead of treating it as an unknown variable to be determined. In this work however, we impose no constraints on the distance between both swimmers, only assuming that the initial distance at $t = 0$ is $r$. Consequently, we obtain the exact correspondence of passive interaction and average unperturbed flow field.
\subsection{Active translation}
Beyond the approximation of the swimmer as a passive object drifting in the local time-averaged flow field around it, the swimmer deforms periodically in order to self-propel, introducing interaction effects which depend on the swimming stroke of both interacting swimmers as well as their relative phase. This effect has been termed active interactions \cite{PooleyAlexanderYeomans2007}.
In both the force-based and the stroke-based model, active translational interactions arise firstly at the quadrupolar order ($\sim 1/r^3$) and order $A^2$ or $\xi^2$, depending on the choice of model. In both models, this term has only a component parallel to the swimmer axis for arbitrary swimmer positioning. The analytical result for the change in swimming velocity of swimmer $I$ in interaction with a second swimmer $I\!I$ side-by-side with it reads in the stroke-based model
\begin{eqnarray}
\fl
\Delta v_I^\mathrm{par, SB} = - \frac{a^2 L^2 \omega }{4 r^3 \left(21 a^2-40 a L+16 L^2\right)^2} \times \nonumber \\ \fl \left(12 \left(63 a^2-64 a L+16 L^2\right) (\xi_{I2} \xi_{I\!I 2} \sin (\beta_I-\beta_{I\!I}-\delta )+\xi_{I2} \xi_{I\!I 1} \sin (\beta_I-\delta )+ \right. \label{eq:velQuadrStrokeBased} \\ \fl \left. \xi_{I1} \xi_{I\!I 2} \sin (\beta_{I\!I}+\delta )+\xi_{I1} \xi_{I\!I 1} \sin (\delta ))+\xi_{I\!I 1} \xi_{I\!I 2} \left(441 a^2-792 a L+272 L^2\right) \sin (\beta_{I\!I})\right), \nonumber
\end{eqnarray}
including both active and passive terms (details of the stroke-based perturbative calculation are presented in \ref{ch:appendixC2}).
For the force-based model, we restrict the analytical result to two pullers in the same configuration, as otherwise the expression would become too large,
\begin{eqnarray}
\fl
\Delta v_I^\mathrm{par, FB} = \frac{3 a^2 A_{I 1}^2 \omega }{4 k^2 (7 \nu -4) r^3 \left(16 \Gamma ^2+9 (4-7 \nu )^2\right) \left(16 \Gamma ^2+(4-3 \nu )^2\right)^2} \times \nonumber \\ \fl \left(-32 \Gamma \left(16 \Gamma ^2 \left(189 \nu ^2-162 \nu +40\right)+3 \left(7371 \nu ^4-21222 \nu ^3+21144 \nu ^2-8992 \nu + \right. \right. \right. \nonumber \\ \fl \left. \left. \left. 1408\right)\right) \sin (\gamma)+8 \left(512 \Gamma ^4-16 \Gamma ^2 \left(2835 \nu ^3-3870 \nu ^2+1920 \nu -352\right)+9 (9 \nu -4) \right. \right. \label{eq:velQuadrForceBased} \\\fl \left. \left. \left(21 \nu ^2-40 \nu +16\right)^2\right) \cos(\gamma)+(147 \nu -68) \left(256 \Gamma ^4+128 \Gamma^2 \left(9 \nu ^2-18 \nu +8\right)+ \right. \right. \nonumber \\ \fl \left. \left. 3 (3 \nu -4)^3 (7 \nu -4)\right)\right).\nonumber
\end{eqnarray}
For the details of the force-based calculation we refer to \ref{ch:appendixB}.
Both results, \eqref{eq:velQuadrStrokeBased} and \eqref{eq:velQuadrForceBased}, scale to leading order constant in $L$ in an expansion around $L = \infty$, similarly to the passive effect \eqref{eq:quadrupolarFlowField}. For moderate stroke amplitudes of both swimmers ($A \approx k a, \xi \approx a$), we therefore observe that passive and active quadrupolar interaction are typically of the same order of magnitude. In the collinear and side-by-side configuration, passive interaction always enhances the self-propulsion of both swimmers, whereas the sign of the active interaction depends on the phase shift between both swimmers via a sine or cosine function, as shown in \eqref{eq:velQuadrStrokeBased}, \eqref{eq:velQuadrForceBased}. Thus, for a large part of the parameter space, two swimmers in such configurations overall benefit in their propulsion (figure \ref{fig_3} (b), \ref{fig_4} (b)), suggesting that a swarm of linear swimmers should propagate faster if the swimmers arrange behind each other or all side-by-side \cite{MirzakhanlooJalaliAlam2018}. However, this result is sensitive to the relative swimmer positioning and orientation. In particular, if the swimmers are arranged on a line or side-by-side, but with opposing swimming directions, the swimmers will predominantly hinder each other's propulsion due to passive effects.
\begin{figure}
\centering
\includegraphics[scale=0.8]{./Figures/201910_Scheel_figure5.eps}
\caption{Amplitude of the $\xi^2/r^3$ or $A^2/r^3$ translational interaction along the swimmer axis of puller $I$ ($\Delta v^\mathrm{par}_I$), in interaction with puller $I\!I$ side-by-side to it, in the stroke-based (SB) and the force-based (FB) model for different values of $\Gamma$ in dependence of the phase shift $\delta$ between both swimmers. The horizontal blue line denotes the passive component, which is independent of $\Gamma$ and equal for the SB and FB models. In the force-based model, the driving forces of both swimmers are chosen such that the resulting stroke of a single swimmer equals the stroke employed for the stroke-based swimmers.}
\label{fig_5}
\end{figure}
Comparing the active interactions using the map from force parameters to stroke parameters, we find that the results in both models generally differ, but agree for arbitrary swimmer positioning in the limit $\Gamma = 6 \pi \eta a \omega / k \to 0$, i.e. when the driving frequency in the force-based model $\omega$ becomes small compared to the inverse viscous time $t_V^{-1}$.
In figure \ref{fig_5}, we plot the total quadrupolar interaction along the swimmer axis to second order in $A$ and $\xi$, respectively, for two side-by-side swimmers in dependence of the phase shift $\gamma$, while varying the rescaled driving frequency $\Gamma$.
It becomes apparent that for $\Gamma$ approaching zero, the respective force-based curve approaches smoothly the stroke-based curve until they coincide in the limit $\Gamma \to 0$. In contrast, for large values of $\Gamma$, the interaction effect in the force-based model is out of phase by approximately $\pi$ compared to the stroke-based active interaction, which means that while active effects in one model boost the swimmer, they slow it down in the other model. This shows that for frequencies with $\Gamma \approx 1$ or larger, the stroke-based results are insufficient to describe the interaction of elastic swimmers and a theory accounting for the altered swimming stroke is inevitable. While figure \ref{fig_5} corresponds to the interaction of two pullers, qualitatively very similar results are found for the different other combinations of pushers and pullers. Figure \ref{fig_5} is hence representative for interacting linear swimmers, independently of the dipolar swimmer types.
To compare both models in figure \ref{fig_5}, a certain stroke amplitude and phase between the two swimmer arms had to be assumed for each swimmer in the stroke-based model. The driving forces for the force-based swimmers were then adjusted for each choice of frequency such that at the single swimmer level the resulting stroke amplitude and phase would coincide with the before fixed stroke. We highlight that when comparing the interactions within two-swimmer systems in both approaches, the driving forces on each swimmer have to be adjusted such that they produce the SB model stroke when the swimmer is alone. If we, hypothetically, would adjust the driving forces such that the resulting swimmer strokes would coincide with the SB strokes while the swimmers interact, we would simply recover the stroke-based system since fixed strokes together with the force-free and torque-free conditions already determine the whole system dynamics.
Analyzing the terms responsible for the active interactions allows to get an intuitive understanding for why they are different in both models. Active quadrupolar translation of swimmer $I$, to second order in $A$ and $\xi$, results from the time-dependent oscillating dipolar flow field, scaling as $A^1$ or $\xi^1$ respectively, produced by swimmer $I\!I$, interacting with the stroke of swimmer $I$.
This flow field can be decomposed in a Taylor series around the position of swimmer $I$. The zeroth order, spatially constant term is associated with the same hydrodynamic force on all three beads, an effect averaging out over one stroke cycle due to the purely oscillatory nature of the flow field.
At the next order in the expansion, the gradient of this instantaneous dipolar flow field, which swimmer $I$ senses along its axis of symmetry, is associated with hydrodynamic forces which expand and compress the swimmer arms and is responsible for the active interaction.
In the force-based model, the interplay of those forces with the springs and the viscous friction alters the swimming stroke by an additional term $\sim A^1/r^3$. Since the swimmer's velocity is effectively quadratic in its stroke, cross terms of this additional term and the swimmer's original stroke $\sim A^1 r^0$ yield the leading order active interactions with the correct scalings.
In the stroke-based model, the arm lengths are prescribed a priori such that the gradient of the time-dependent dipolar flow field evokes additional external forces necessary to maintain the prescribed stroke rather than an altered stroke.
The active interactions then arise as a consequence of the additional forces \cite{FarzinRonasiNajafi2012}.
From this, we can understand why both the stroke-based and force-based models in general differ in the active interactions, but become equivalent when in the force-based model $\Gamma = 6 \pi \eta a \omega / k \to 0$. This limit can also be interpreted as the limit of $k \to \infty$ when fixing $\omega$ and $\eta$, i.e. as the limit of very stiff springs. Consequently, to maintain a constant swimming stroke amplitude of a single swimmer, it is then necessary to scale up the driving force amplitude accordingly. A such force-based swimmer with stiff arms will barely alter its stroke in response to compressing or expanding hydrodynamic forces produced by another nearby swimmer. Instead, the stiff springs will exert forces along the arms counterbalancing the hydrodynamic forces - and the mechanism of the active interactions becomes identical to the one for the stroke-based swimmers.
Comparing to the literature, we find that the results \eqref{eq:velQuadrStrokeBased} and \eqref{eq:velQuadrForceBased} have not been reported yet. Other authors have obtained results for the quadrupolar active interaction scaling to leading order linear in $L$ \cite{FarzinRonasiNajafi2012, KurodaYasudaKomura2019}. Again, we have reproduced such linear scaling in a calculation employing the altered perturbative scheme, which assumes the distance between the middle beads of both swimmers to be fixed to $r$ throughout the swimming stroke.
Conversely, we find in our calculation that the active quadrupolar interaction scales to leading order constant in $L$, which is in agreement with our numerical calculations (figure \ref{fig_2}, \ref{fig_3}, \ref{fig_4}) and also backed up by the explanation of the active interaction in terms of time-dependent dipolar flow field, which yields a similar scaling. We are able to confirm that the second order quadrupolar translational interaction acts along the swimmer axis only, a finding which has been reported previously \cite{FarzinRonasiNajafi2012}.
In contrast to the quadrupolar translation, the octopolar ($\sim 1/r^4$) translational interaction has in general both a component along the swimmer axis, which typically becomes important only at close swimmer separations since it decays quicker than quadrupolar effects, as well as a component acting orthogonal to the swimmer axis. As the quadrupolar translational interaction has no orthogonal component, a transition from the passive dipolar far field interaction to active octopolar interaction at intermediate separations can be observed for $v^\mathrm{ort}$ in the side-by-side swimmer configuration (\ref{fig_4} (d), (e)) at a few tenths of swimmer lengths. For two equal swimmers and $\gamma = 0$, the active orthogonal octopolar component vanishes, therefore the transition is not observed in figure \ref{fig_4} (c) and (f).
In the stroke-based calculation we obtain for the $\xi^2$ octopolar interaction of two side-by-side swimmers orthogonal to the swimmer axis
\begin{eqnarray}
\fl
v_I^\mathrm{ort, SB} = \frac{9 a L^3 \omega \left(21 a^2-54 a L+32 L^2\right)}{r^4 (7 a-8 L)^2 (3 a-4 L)} \left( \xi_{I 2} \xi_{I\!I 2} \sin (\beta_I-\beta_{I\!I}-\delta )+ \nonumber \right. \\ \left. \xi_{I 2} \xi_{I\!I 1} \sin (\beta_I-\delta )-\xi_{I 1} (\xi_{I\!I 2} \sin (\beta_{I\!I} +\delta )+\xi_{I\!I 1} \sin (\delta )) \right).
\label{eq:swimVelOctoOrth}
\end{eqnarray}
Since all summands in \eqref{eq:swimVelOctoOrth} are dependent on $\delta$, this term is purely active, consistent with the absence of even order $\xi^2$ flow fields and passive interaction for the linear swimmer geometry. This result is, to leading order in $1/L$, in agreement with the corresponding formula from \cite{AlexanderPooleyYeomans2009}.
Comparing to the corresponding result for the force-based interaction,
\begin{eqnarray}
\fl
v_I^{\mathrm{ort, FB}} = - \frac{9 a^3 A^2 C \omega \left(63 \nu ^3-246 \nu ^2+312 \nu -128\right) }{r^4 k^2 (8-7 \nu )^2 \nu ^2 \left(16 \Gamma ^2+(4-3 \nu )^2\right)} \left( -B_I B_{I\!I} \sin (\alpha_I-\alpha_{I\!I}-\gamma)+ \nonumber \right. \\ \left. B_I \sin(\alpha_I-\gamma)-B_{I\!I} \sin(\alpha_{I\!I}+\gamma)+\sin(\gamma) \right),
\end{eqnarray}
we find that both results are indeed equivalent with respect to the mapping from the force-based to the stroke-based model. This means, this interaction effect is independent of the spring constant $k$ in the force-based model and equal to the interaction effect for two stroke-based swimmers, assuming that the driving forces one each swimmer correspond to the same single swimmer stroke. This result holds for arbitrary swimmer configurations. In contrast, we find that the component of the octopolar translation along the swimmer axis is not equivalent in both models.
\subsection{Active rotation}
Active rotation is firstly observed at octopolar order ($\sim 1/r^4$) and order $A^2$ and $\xi^2$. In the stroke-based model, we find for two side-by-side swimmers
\begin{eqnarray}
\fl
\Omega_I^\mathrm{SB} = -\frac{9 a L^2 \omega}{8 r^4 (7 a-8 L) (3a-4 L) (7 a-4 L)^2} \times \nonumber \\ \left(2 a \xi_{I\!I 1} \xi_{I\!I 2} \left(1029 a^2-1652 a L+544 L^2\right) \sin (\beta_{I\!I}) - \left(1617 a^3- \right. \right. \nonumber \\ \left. \left. 3136 a^2 L+1936 a L^2- 384 L^3\right) (\xi_{I2} \xi_{I\!I 2} \sin(\beta_I-\beta_{I\!I}-\delta )+ \right. \\ \left. \xi_{I2} \xi_{I\!I 1} \sin (\beta_I-\delta ) +\xi_{I1} \xi_{I\!I 2} \sin (\beta_{I\!I}+\delta )+ \xi_{I1} \xi_{I\!I 1} \sin (\delta ))\right), \nonumber
\label{eq:rotationStrokeBased}
\end{eqnarray}
where both the active and passive contribution is included.
In the force-based model, we again restrict the analytical result to two pullers in the same configuration, finding
\begin{eqnarray}
\fl
\Omega_I^\mathrm{FB} = - \frac{27 a^2 A_{I1}^2 \omega }{4 r^4 k^2 \nu (7 \nu -8) (7 \nu -4) \left(16 \Gamma ^2+9 (4-7 \nu )^2\right) \left(16 \Gamma ^2+(4-3 \nu )^2\right)} \times \nonumber \\ \left(\left(1617 \nu ^3-3136 \nu ^2+1936 \nu -384\right) \left(16 \Gamma ^2+63 \nu ^2-120 \nu + \right. \right. \nonumber \\ \left. \left. 48\right) \cos (\gamma)- \nu \left(1029 \nu ^2- 1652 \nu +544\right) \left(16 \Gamma ^2+63 \nu ^2-120 \nu + \right. \right. \\ \left. \left. 48\right)- 8 \Gamma \left(14553 \nu ^4-34692 \nu ^3+ 29968 \nu ^2- 11200 \nu + 1536\right) \sin (\gamma)\right). \nonumber
\label{eq:rotationForceBased}
\end{eqnarray}
Employing the mapping from force-based to stroke-based parameters reveals that also this octopolar rotation is equivalent in both models.
The active octopolar rotation scales to leading order as $L^1$ for $L \rightarrow \infty$, consistent with the results reported in the literature \cite{PooleyAlexanderYeomans2007, AlexanderPooleyYeomans2009, FarzinRonasiNajafi2012}. In particular, we find that our result agrees algebraically in the leading order in $1/L$ with the result reported in \cite{AlexanderPooleyYeomans2009}.
Unlike the active contribution, the passive contribution to octopolar rotation scales to leading order as $L^0$, as can be seen from \eqref{eq:rotationStrokeBased} and \eqref{eq:rotationForceBased}. Therefore, the active component of the rotation is typically dominating over the passive one, and clockwise (CW) as well as counterclockwise (CCW) rotation can both be observed when varying $\gamma$ or $\delta$ (figure \ref{fig_4} (b)) due to the sinusoidal dependence of active interaction on the phase shift.
Investigating the origin of the terms responsible for the active octopolar rotation at this order in the perturbative scheme, we find that three different terms contribute to it.
First, swimmer $I$ rotates with magnitude $\sim A^1/r^4$ in oscillatory fashion due to the curl associated to the time-dependent quadrupolar flow field produced by swimmer $I\!I$. From the interplay of this time-dependent rotation with the swimmer's own self-propulsion, swimmer $I$ experiences an active overall rotation $\sim A^2/r^4$.
Second, swimmer $I$ moves within the instantaneous flow field produced by the other swimmer due to its $A^1 r^0$ original swimming stroke, experiencing the curl of this flow field at different positions. The inference of both the oscillating motion and the time-dependent curl of the flow field contributes to the active interactions as well.
Third, swimmer $I$ experiences hydrodynamic forces which would bend the swimmer, resulting from the instantaneous dipolar flow field produced by swimmer $I\!I$. The counteracting forces exerted by the rigid joint, although themselves being torque-free, induce fluid flows, which, in interplay with the $A^1 r^0$ swimming stroke of swimmer $I$, give also rise to active rotation.
All three of these mechanisms are independent of whether the swimmer arms are stiff or elastic, explaining why the active octopolar rotation is equivalent in both models.
\subsection{Swimmer behavior for arbitrary positioning and long-term behavior}
We now extend our considerations to two arbitrarily positioned force-based swimmers and the temporal evolution of their relative positions. For the latter, different types of long-term trajectories for two swimmers have already been classified in the stroke-based model in dependence on the initial relative positioning \cite{PooleyAlexanderYeomans2007, FarzinRonasiNajafi2012}. Both authors investigating the long-term behavior have obtained broadly similar results for the long-term trajectories, although different driving protocols were employed.
We extend these studies to force-based swimmers with sinusoidal driving and show that the long term behavior is qualitatively the same as for stroke-based swimmers \cite{PooleyAlexanderYeomans2007, FarzinRonasiNajafi2012} and is also robust with respect to varying the swimmer characteristics between puller- and pusher-type.
This can be understood from the fact that the long-term behavior is predominantly a result of the self-propulsion and rotational interaction, which are to leading order equivalent in both approaches, and is only weakly influenced by translational interaction \cite{PooleyAlexanderYeomans2007}. In contrast to previous works, we infer the long-term behavior from the instantaneous interactions evaluated in the parameter space of all possible relative swimmer configurations, allowing for a simple graphical representation of the system's dynamics.
The relative positioning of both swimmers is parametrized by their distance $r$, the orientation of the connection between the middle beads of both swimmers relative to the axis of swimmer $I$, $\varphi$, and the difference between the two swimmer orientations, $\theta_{I\!I} - \theta_I$ (figure \ref{fig_1}).
The equations governing the momentary evolution of a system of two swimmers, obtained by elementary geometry, are given by
\begin{align}
&\frac{d}{dt} \varphi = \frac{v_{0, I} + \Delta v_{I}^\mathrm{par}}{r} \sin \varphi - \frac{v_I^\mathrm{ort}}{r} \cos \varphi - \frac{v_{0, I\!I} + \Delta v_{I\!I}^\mathrm{par}}{r} \sin (\varphi^{\prime}) - \frac{v_{I\!I}^\mathrm{ort}}{r} \cos(\varphi^{\prime}) - \Omega_I, \label{eq:phiDot} \\
&\frac{d}{dt} (\theta_{I\!I} - \theta_I) = \Omega_{I\!I} - \Omega_I, \label{eq:thetasDot} \\
&\frac{d}{dt} r =- (v_{0, I} + \Delta v_{I}^\mathrm{par}) \cos \varphi - v_I^\mathrm{ort} \sin \varphi - (v_{0, I\!I} + \Delta v_{I\!I}^\mathrm{par}) \cos (\varphi^{\prime}) + v_{I\!I}^\mathrm{ort} \sin(\varphi^{\prime}),
\end{align}
with $\varphi^\prime = \pi - \varphi + \theta_{I\!I} - \theta_I$. $\Delta v_s^\mathrm{par}$, $v_{s}^\mathrm{ort}$ and $\Omega_s$ denote the time-averaged interaction effect parallel to the swimmer axis, orthogonal to the swimmer axis and the rotational interaction effect on swimmer $s$. $v_{0, s}$ denotes the corresponding time-averaged single swimmer speed.
The first four terms on the right hand side of \eqref{eq:phiDot} correspond to how $\varphi$ changes due to the swimmer translation, while the last term accounts for the effect due to the rotation of swimmer $I$. To illustrate the behavior of both swimmers, we display the difference in the angular velocities of both swimmers, $\Omega_{I\!I} - \Omega_I$, time-averaged over a complete stroke cycle in dependence of $\varphi$ and $\theta_{I\!I} - \theta_I$ for a fixed distance $r$ by the color in figure \ref{fig_6} (a) and (b). The momentary time evolution of the system is illustrated by arrows corresponding to the vector field
\begin{equation}
X_t = \left(\frac{d}{dt} \varphi, \frac{d}{dt} (\theta_{I\!I} - \theta_I) \right).
\label{eq:temporalVectorField}
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.805]{./Figures/201910_Scheel_figure6_v2.eps}
\caption{Difference of the angular velocities of swimmer $I\!I$ and $I$ in dependence of angles encoding the relative swimmer positions, $\varphi$ and $\theta_2 - \theta_1$, for two force-based pullers with (a) phase shift $\gamma = 0$ and (b) $\gamma = \pi$ for distance $r = 30 a$. The small white arrows indicate the time evolution that the system undergoes. The length of the arrows has been normalized for better visibility. The large black and white arrows illustrate the different types of long-term behavior observed, the different colors are to improve the contrast of the figure in print.
In (c), typical typical trajectories of the two swimmers moving from left to right are depicted. Point of reference is the central bead in each swimmer.}
\label{fig_6}
\end{figure}
An arrow pointing to the left or right indicates that $\varphi$ increases or decreases, and an arrow pointing upwards or downwards corresponds to $\theta_{I\!I} - \theta_I$ increasing or decreasing. When both $\varphi$ and $\theta_{I\!I} - \theta_I$ change with time, the arrow is tilted such that the horizontal and the vertical component represent the corresponding rate of change. For better visibility the lengths of all vectors have been normalized. We obtain qualitatively similar plots for two interacting pushers or for combinations of a pusher and a puller. Figure \ref{fig_6} can therefore be considered representative for the interaction of linear swimmers at intermediate distances up hundred swimmer lengths, independently of the dipolar characteristics.
This effectively results from the fact that the linear swimmer with equal bead radii produces only a weak average dipolar flow field, as discussed before, and interaction at the distance of $r = 30 a$ considered here is dominated by quadrupolar effects. It is in contrast to swimmer models with dominant dipolar average flow fields, which often exhibit strong differences in the collective behavior of pullers and pushers \cite{Hoell2018, Yoshinaga2017, Bardfalvy2020}.
The time evolution for two in-phase swimmers ($\gamma = 0$, figure \ref{fig_6} (a)) and two swimmers out-of-phase by $\gamma = \pi$ (figure \ref{fig_6} (b)) is dominated for a major part of the parameter space spanned by $\varphi$ and $(\theta_{I\!I} - \theta_I)$ by the self-propulsion terms in \eqref{eq:phiDot}, which are linear in $v_{0, s}$. In these parts of the parameter space, we have $|d/dt \, \varphi| \gg |d/dt \, (\theta_{I\!I} - \theta_I)|$, associated with horizontal arrows. Exceptions from this are first the regions around the green dashed lines, corresponding to $\theta_{I\!I} - \theta_I = 2 \varphi \pm \pi$, and second the region around the $\varphi$-axis, i.e. $\theta_{I\!I} - \theta_I = 0$, corresponding to swimmers with parallel swimming direction. It is straight forward to verify that the first case corresponds to both swimmers being symmetric with respect to some axis in the $x$-$y$-plane, as sketched in the inset in the first panel of figure \ref{fig_6} (c). In both of the two cases, the self-propulsion terms proportional to $v_{0, s}$ in \eqref{eq:phiDot} cancel. Therefore, the swimmer behavior in these regions is dominated by rotational interactions. In the symmetric configuration (green dashed lines), in-phase pullers rotate towards each other (which can be seen from the vectors pointing down at $\varphi = \pi/2, \theta_{I\!I} - \theta_{I} = 0$ in \ref{fig_6} (a)), but rotate away from each other when out-of-phase, consistently with figure \ref{fig_4} (b).
During swimmer interaction, not only the values of $\varphi, \theta_{I}$ and $\theta_{I\!I}$ change, but also the distance $r$ between both swimmers does.
Varying the value of $r$, the angular velocities $\Omega_I$ and $\Omega_{I\!I}$ scale as $r^{-4}$ as long as $r$ does not leave the regime dominated by the quadrupolar flow fields. Hence, the color structure in figure \ref{fig_6} stays constant while only the value corresponding to each color is rescaled.
Conversely, the arrows align with the horizontal direction when $r$ is increased, since the horizontal component \eqref{eq:phiDot} scales to leading order as $r^{-1}$ due to $v_{0, s}$ being independent of $r$. In contrast, the vertical component \eqref{eq:thetasDot} scales as $r^{-4}$.
Although figure \ref{fig_6} (a) and (b) correspond to constant $r$ and cannot account for the swimmers coming closer or separating, we are able to reproduce the different types of long-term trajectories (figure \ref{fig_6} (c)), which have been similarly reported for two interacting linear stroke-based swimmers with initially parallel swimming direction (figure 4 (a) and (b) in \cite{PooleyAlexanderYeomans2007} and figure 3, upper left panel in \cite{FarzinRonasiNajafi2012}). In figure \ref{fig_6} (c), the middle bead of each swimmer is taken as reference for the swimmer position. Since the translational interaction is small compared to self-propulsion, both swimmers are in general oriented tangential to their trajectories, swimming from left to right.
We first consider two swimmers with zero phase shift $\gamma$ (figure \ref{fig_6} (a)) and classify the different types of long-term trajectories observed when the swimmers have initially parallel swimming direction ($\theta_{I\!I} - \theta_I = 0$) but are positioned arbitrarily otherwise, corresponding to varying the initial value of $\varphi$.
Two approximately collinear swimmers, i.e. $\varphi \approx 0$, swim on oscillatory (O) trajectories. Mapping this behavior to the phase space of $\varphi$ and $\theta_{I\!I} - \theta_I$, it corresponds to the curl around $\varphi = 0, \theta_{I\!I} - \theta_I = 0$ found in figure \ref{fig_6} (a).
With increasing $\varphi$, we next enter the repulsive regime (R) in which the swimmers rotate away from each other and thus separate such that rotational interaction dies out and the swimmers self-propel on straight trajectories away from each other. Increasing $\varphi$ further, still keeping the swimmers initially parallel, the swimmers typically move on parallel trajectories (P), associated to the curl close to $\varphi \approx \pi/3, \theta_{I\!I} - \theta_I = 0$. Since this curl is associated to positive values of $\varphi$ only, the swimmer trajectories do not cross, distinguishing it from the oscillatory case.
For $\varphi$ close to $\pi/2$, the swimmers rotate towards each other, effectively attracting each other (A), until the beads come so close that the Oseen approximation is not valid anymore. This behavior corresponds to the area around the green dashed lines in figure \ref{fig_6} (a).
Similarly, also the types of trajectories of two out-of-phase swimmers are accounted for in the respective plot of the $\varphi$-$(\theta_{I\!I} - \theta_I)$ space (figure \ref{fig_6} (b)). Besides the types of trajectories already mentioned, we also observe trajectories where the swimmers first rotate towards each other, their trajectories cross and the swimmers subsequently repel ($\mathrm{O}_\mathrm{R}$). Also for this behavior, the qualitative trajectory in terms of $\varphi$ and $\theta_{I\!I} - \theta_I$ can be depicted in figure \ref{fig_6} (b).
\section{Summary and Conclusions}
\label{sec:conclusion}
We have calculated the scaling function, the strength, and the direction of the hydrodynamic interactions between two linear three-bead swimmers. Both the force-based and stroke-based models were considered within perturbative approaches, and the results were compared to direct numerical results with excellent agreement. This framework is applicable to arbitrary positioning of the two interacting swimmers, however, the most simple cases of collinear and side-by-side swimmers have been used to illustrate the main results of the paper.
In both the force-based and the stroke-based model, the swimmers experience passive hydrodynamic interactions given by the time-averaged flow field produced by all nearby swimmers. Comparing the interaction of force-based and stroke-based swimmers under the constraint that the individual devices display identical actuation, passive interactions are equivalent in both models.
In terms of active hydrodynamic interactions, which are a result of the interference of the time-dependent swimming strokes of both swimmers, active rotation and translational interaction orthogonal to the swimmer axis are equivalent to the leading order in the stroke- and the force-based model. However, the leading order of the active translation along the swimmer axis is not. This shows that the interaction is in general not conserved with respect to the mapping between stroke-based and force-based swimmers and that the deformation of one swimmer in the presence of others significantly alters the change in propulsion speed that a swimmer experiences, depending on the design. However, we have shown that both models are equivalent when the driving frequency is small compared to the inverse viscous time, which is meaningful only in the force-based model. Consequently, the stroke-based interaction should be regarded as the special case of the force-based interaction in the limit of small frequencies.
At the leading order in the swimmer actuation, the time-averaged flow field produced by a three-sphere swimmer is proportional to its velocity, with a proportionality coefficient depending solely on the swimmer geometry. This novel result, valid for force-based as well as stroke-based swimmers, allows us to infer the leading order passive interactions directly from the geometry of a system of two interacting swimmers. Applied to the three-bead swimmers this has interesting consequences. Namely, compared to the velocity of a single swimmer, passive interactions enhance the swimming velocity if swimmers are one behind another, or side-by-side in the same direction. This cooperative effect for pairs of swimmers is independent of the details of the driving and is found similarly for pullers and pushers. It relies on the passive quadrupolar interactions, which dominate in the major part of the parameter space.
In the long-time limit, the behavior of a force-based two-swimmer system is independent of the pusher/puller characteristics of both swimmers up to separations of hundreds of swimmer lengths. We find that the long-term behaviors for force-based and stroke-based swimmers agree qualitatively, which is a result of the equivalence of the leading order rotational interactions in both models. Following a numerical analysis, we have also shown that the different types of long-term trajectories can be qualitatively inferred directly from the instantaneous behavior of both swimmers evaluated in the space of all relative swimmer configurations.
The insights presented here may help to better understand the collective motion of swarms. In particular, we have thoroughly studied active interactions that rely on resolving the flow fields produced by each swimmer on time scales below the single stroke. This feature is typically not taken into account in studies of swimmer suspensions, and given that active contribution may be important, accounting for these effects may give rise to new insights in the collective behavior. Our analytical approach may be a particularly suitable tool for the task, as it can be applied explicitly to an arbitrary number of swimmers, the limiting factor being only the tractability. Furthermore, the approach can be easily tuned in terms of the precision of the hydrodynamic interactions at smaller bead separations. This can be done by using the Rotne-Prager approximation or some higher order theory instead of the Oseen approximation, a feature which becomes necessary when the density of swimmers increases. Given that hydrodynamic interactions have been shown to be important even in dense suspensions of bacteria \cite{IshikawaHota2006}, the emerging results may find application well beyond the physical context presented herein.
\section*{Acknowledgements}
We thank the Priority Programme 'Microswimmers - From Single Particle Motion to Collective behavior' (SPP 1726) of the Germany Research Foundation for funding, and acknowledge the support of the Excellence Cluster 'Engineering of Advanced Materials' at the Friedrich-Alexander-University Erlangen-Nuremberg. We benefited from insightful discussions with Alexander Sukhov, Oleg Trosman and Nicolas Vandewalle.
|
\section{Introduction}
Fast and ultrasensitive electrometers have been instrumental to the advancement of basic science. They have been used to detect in real time the tunneling of electrons in a quantum dot \cite{Lu2003}, determine the tunneling rates of quasiparticles in superconducting devices \cite{Naaman2006}, and search for signatures of Majorana zero modes in nanowires \cite{Zanten2020}. In addition, the rapid detection of single electrons is crucial for the readout of quantum-dot-based qubits \cite{Petta2005}, for which operating at lower photon numbers reduces measurement backaction \cite{D'Anjou2019}. In this same vein, ultrasensitive electrometers are at the heart of many schemes for sensing the displacement of charged mechanical resonators \cite{Knobel2003, LaHaye2004, Naik2006}, as well as for coherently coupling mechanical resonators to microwave cavities \cite{Rimberg2014, Heikkila2014, Pirkkalainen2015}. To observe and take advantage of quantum effects in such hybrid systems it is often essential that their coupling be strong at the single-photon level, a regime that has been achieved for quantum dots \cite{Mi2016, Mi2018} but not yet for mechanical resonators despite significant effort \cite{Zoepfl2020, Schmidt2020, Kounalakis2020, Bera2021}. Reaching the single-photon strong optomechanical coupling regime, where a single cavity photon causes sufficient radiation pressure to displace the mechanical resonator by more than its zero-point uncertainty, would enable the generation of nonclassical states of both light and motion \cite{Nunnenkamp2011, Rabl2011}, as well as provide a rich platform for studying the quantum-to-classical transition and other fundamental physics \cite{Aspelmeyer2014}.
Electrometers based on the single electron transistor (SET) are among the fastest and most sensitive reported in the literature to date. Radio-frequency single electron transistors (rf-SETs) are the best known of these devices, having achieved sensitivities below $1$~$\mu e/\sqrt{\mrm{Hz}}$ \cite{Brenning2006} and bandwidths greater than $100$ MHz \cite{Schoelkopf1998}. The rf-SET encodes the charge gating the SET in the power dissipated by the SET, which is embedded in a tank circuit to enable RF readout of this dissipation. This dissipative detection typically requires picowatts of power, corresponding to hundreds of thousands of photons in the tank circuit, rendering the rf-SET unsuitable for some of the aforementioned applications and making it impossible to integrate the rf-SET with modern near-quantum-limited amplifiers \cite{Castellanos-Beltran2007, Macklin2015, Sivak2019} (which typically saturate well below the picowatt scale). Dispersive electrometers based on the SET have also been developed, which encode the gate charge in the resonant frequency of a tank circuit. Such electrometers have been operated using femtowatts of power \cite{Sillanpaa2004,Naaman2006}, corresponding to tens or hundreds of photons, and have achieved sensitivities as low as $30$~$\mu e/\sqrt{\mrm{Hz}}$ \cite{Sillanpaa2005}. More recently, dispersive gate-based sensors have been developed \cite{Gonzalez-Zalba2015} that have surpassed the performance of SET-based electrometers. These devices have achieved sensitivities as low as $0.25$~$\mu e/\sqrt{\mrm{Hz}}$ with bandwidths approaching $1$ MHz using $100$ attowatts of power, corresponding to hundreds of photons \cite{Schaal2020}.
In this letter we demonstrate ultrasensitive dispersive charge detection with a cavity-embedded Cooper pair transistor (cCPT) \cite{Brock2021_characterization,Kanhirathingal2021}. Using $16$ attowatts of power, corresponding to the single-photon level of the cavity, we measure a minimum charge sensitivity of $14$~$\mu e/\sqrt{\mathrm{Hz}}$. We find that the cCPT operates within a factor of $5$ of its theoretical quantum-limited sensitivity, this discrepancy being due to frequency noise, amplifier noise, and the nonlinearity of the device. Another limitation of the present device is quasiparticle poisoning \cite{Aumentado2004}, which prevents us from studying the cCPT at its theoretically-optimal operating point. Based on these results we expect an optimized sample could achieve a sensitivity as low as $0.4$~$\mu e/\sqrt{\mathrm{Hz}}$, rivaling that of the best gate-based sensor \cite{Schaal2020}. Due to its ability to operate at the single-photon-level, the cCPT has been proposed as a platform for reaching the single-photon strong coupling regime of optomechanics \cite{Rimberg2014}. Our results support the feasibility of this proposal and represent an important step toward its realization.
Here we study the same device characterized experimentally in Ref. \cite{Brock2021_characterization}. The most important parameters of this realization of the cCPT are shown in Table \ref{tab:ccpt_parameters}. For more information on this device, including sample images, fabrication methods, and characterization techniques, see Ref. \cite{Brock2021_characterization}.
The cCPT, depicted schematically in Fig. \ref{fig:schematics_and_noise}(a), has two components: a quarter-wavelength ($\lambda/4$) coplanar waveguide cavity and a Cooper pair transistor (CPT). The CPT consists of two Josephson junctions (JJs) with an island between them that can be gated via the capacitance $C_{g}$. The CPT is connected between the voltage antinode of the cavity and the ground plane, such that the two form a SQUID loop. Embedded in this way, the CPT behaves as a nonlinear Josephson inductance $L_{J}$ in parallel with the cavity that can be tuned by both the number of electrons $n_{g}$ gating the island and the flux $\Phi_{\mathrm{ext}}$ threading the SQUID loop. The gate charge $n_{g}$ is thus encoded in the resonant frequency $\omega_{0}$ of the cavity, which can then be detected via microwave reflectometry. The theoretical charge sensitivity of the cCPT in this mode of operation is derived from first principles in Ref. \cite{Kanhirathingal2021}. This device can be operated at much lower powers than comparable SET-based dispersive electrometers \cite{Naaman2006, Sillanpaa2004, Sillanpaa2005} for two key reasons. First, we use a distributed superconducting microwave cavity rather than a lumped-element LC circuit, yielding much lower dissipation. Second, we can tune the CPT band structure via the external flux $\Phi_{\mrm{ext}}$, which provides us greater flexibility in biasing the device to an optimally-sensitive point.
\begin{table}[t]
\begin{tabular}{|c|c|} \hline
Josephson energy & $E_{J}/h = 14.8$ GHz \\ \hline
Charging energy & $E_{C}/h = 54.1$ GHz \\ \hline
Gate capacitance & $C_{g} = 6.3$ aF \\ \hline
Coupling capacitance & $C_{c} = 7.1$ fF \\ \hline
Bare cavity frequency & $\omega_{\lambda/4}/2\pi \approx 5.76$ GHz \\ \hline
Cavity linewidth & $\kappa_{\mrm{tot}}/2\pi \approx 1.4$ MHz \\ \hline
Cavity length & $\ell = 5135 \mu$m \\ \hline
Characteristic impedance & $Z_{0} = 50$ $\Omega$ \\ \hline
\end{tabular}
\caption{Parameters of the cCPT \cite{Brock2021_characterization}.}
\label{tab:ccpt_parameters}
\end{table}
\begin{figure}[!t]
\includegraphics{combined_schematics_and_noise.pdf}
\caption{(a) Schematic of the cCPT. (b) Schematic of the measurement circuitry. The cavity behaves as a parallel RLC circuit when driven near its fundamental frequency \cite{text_pozar}, and the CPT behaves as an inductance $L_{J}$ in parallel with the cavity. (c) System noise referred to the sample plane (solid black line). Shaded areas show the contribution of each noise source. The dashed white line is the quantum limit.}
\label{fig:schematics_and_noise}
\end{figure}
To measure the charge sensitivity of the cCPT we drive the cavity with a resonant carrier signal while modulating the gate about a dc bias point $n_{g}$ such that $n_{g}(t) = n_{g} + \sqrt{2}(q_{\mrm{rms}}/e)\cos(\omega_{g}t)$, which in turn modulates the resonant frequency according to $\omega_{0}(t) = \omega_{0} + \sqrt{2}(\partial\omega_{0}/\partial n_{g})(q_{\mathrm{rms}}/e)\cos(\omega_{g}t)$. As a result, the reflected carrier signal is phase-modulated leading to output power $P_{\mrm{out}}$ proportional to $q_{\mrm{rms}}^{2}$ at the sideband frequencies $\omega_{0}\pm\omega_{g}$. Thus, given the rms charge modulation amplitude $q_{\mrm{rms}}$, we can use a spectrum analyzer with resolution bandwidth $B$ to measure the sidebands and thereby extract the charge sensitivity $\delta q$ from
\begin{equation}\label{eq:empirical_sensitivity}
\delta q = \frac{q_{\mrm{rms}}}{\sqrt{2B}\times 10^{\mrm{SNR}/20}},
\end{equation}
where $\mrm{SNR}$ is the single-sideband signal to noise ratio expressed in decibels \cite{Brenning2006, Aassime2001}. Here we consider the total power at the two sidebands to be the signal of interest, since it is possible to combine them via homodyne mixing, leading to the factor of $1/\sqrt{2}$ above. Theoretically, the output sideband power $P_{\mrm{out}}(\omega_{0}\pm\omega_{g})$ can be expressed
\begin{equation}\label{eq:output_sideband_power}
P_{\mrm{out}}(\omega_{0}\pm\omega_{g}) = \frac{2\kappa_{\mrm{ext}}^{2}}{\kappa_{\mrm{tot}}^{2}(\omega_{g}^{2} + \kappa_{\mrm{tot}}^{2}/4)}\abs{\frac{q_{\mrm{rms}}}{e}\frac{\partial\omega_{0}}{\partial n_{g}}}^{2}P_{\mrm{in}},
\end{equation}
where $\kappa_{\mrm{ext}}$ and $\kappa_{\mrm{tot}}$ are the external and total damping rates of the cavity, respectively, and $P_{\mathrm{in}}$ is the input carrier power at the plane of the sample \cite{Brock2021_characterization}. The theoretical charge sensitivity can therefore be expressed as \cite{Kanhirathingal2021}
\begin{equation}\label{eq:theoretical_charge_sensitivity}
\delta q = \frac{\kappa_{\mrm{tot}}}{2\kappa_{\mrm{ext}}}\sqrt{\frac{S_{\mrm{noise}}}{P_{\mrm{in}}}\left(\omega_{g}^{2} + \frac{\kappa_{\mrm{tot}}^{2}}{4}\right)}\abs{\frac{\partial\omega_{0}}{\partial n_{g}}}^{-1}e.
\end{equation}
To evaluate this expression we use the sample-referred $S_{\mrm{noise}}$ and $P_{\mrm{in}}$ (discussed below), as well as the values of $\kappa_{\mrm{ext}}$, $\kappa_{\mrm{tot}}$, and $\omega_{0}(n_{g}, \Phi_{\mrm{ext}})$ determined from a detailed characterization of the device \cite{Brock2021_characterization}. The damping rates are approximately $\kappa_{\mrm{ext}}/2\pi\approx 1.2$ MHz and $\kappa_{\mrm{tot}}/2\pi\approx 1.4$ MHz, though these depend on $\omega_{0}(n_{g},\Phi_{\mrm{ext}})$ and can vary by $10\%-20\%$. The corresponding quantum-limited sensitivity of the device is obtained by evaluating Eq. \eqref{eq:theoretical_charge_sensitivity} at the quantum limit of system noise for our measurement scheme, $S_{\mrm{noise}}^{\mrm{QL}}=\hbar\omega$, as discussed below.
Importantly, both Eqs. \eqref{eq:empirical_sensitivity} and \eqref{eq:theoretical_charge_sensitivity} are only valid when $q_{\mrm{rms}}/e\ll \omega_{g}/(\partial\omega_{0}/\partial n_{g})$, which ensures that the amplitude of the resulting frequency modulation is small compared to $\omega_{g}$ and that $P_{\mrm{out}}(\omega_{0}\pm\omega_{g}) \propto q_{\mrm{rms}}^{2}$. In all of our measurements we use sufficiently small $q_{\mrm{rms}}$ to satisfy this constraint. Furthermore, Eq. \eqref{eq:theoretical_charge_sensitivity} is most accurate in the linear response regime for which $n\ll\kappa_{\mrm{tot}}/|K|$, where $n=4\kappa_{\mrm{ext}}P_{\mrm{in}}/\hbar\omega_{0}\kappa_{\mrm{tot}}^{2}$ is the average number of intracavity photons and $K$ is the Kerr nonlinearity of the cCPT \cite{Brock2021_characterization}. Experimentally, we find that for $n\ll\kappa_{\mrm{tot}}/|K|$ the output sideband power grows linearly with $P_{\mrm{in}}$ as expected from Eq. \eqref{eq:output_sideband_power}, but as $n$ approaches $\kappa_{\mrm{tot}}/|K|$ this trend becomes sub-linear. Near this threshold, $P_{\mrm{out}}(\omega_{0}\pm\omega_{g})$ begins to decrease with increasing $P_{\mrm{in}}$. For the present device this threshold corresponds to the single-photon-level \cite{Brock2021_characterization}, so we perform all of our measurements with $n\lesssim 1$. In this sense the single-photon-level operation of the cCPT can be viewed as both an enabling feature (for the reasons described earlier) and a constraint, but this constraint could be avoided in future devices by changing $E_{J}$, $E_{C}$, and $\kappa_{\mrm{tot}}$.
The detection bandwidth of the present device, which determines the maximum rate at which the cavity can respond to changes in $n_{g}$, is set by $\kappa_{\mrm{tot}}$ and is on the order of $1$ MHz. The bandwidth can be improved by increasing the coupling capacitance $C_{c}$, thereby increasing $\kappa_{\mrm{ext}}$, but this also affects the single-photon-level charge sensitivity. Setting $n=1$ and assuming negligible internal loss such that $\kappa_{\mrm{tot}}\approx \kappa_{\mrm{ext}}$, Eq. \eqref{eq:theoretical_charge_sensitivity} predicts $\delta q \propto \sqrt{\kappa_{\mrm{tot}}}$ for $\omega_{g}\ll\kappa_{\mrm{tot}}$. However, if we restrict ourselves to the linear-response regime rather than the single-photon level we can operate with $n \sim \kappa_{\mrm{tot}}/\abs{K}$, in which case $\delta q$ is independent of $\kappa_{\mrm{tot}}$. Lastly, it is worth noting that if one increases $\kappa_{\mrm{ext}}$, one also increases the charge noise coupling to the cCPT via the input-output transmission line \cite{Kanhirathingal2021}, which is negligible in the present device.
The cCPT is housed in a dilution refrigerator with a base temperature of $T\lesssim 30$ mK and measured using the circuitry depicted schematically in Fig. \ref{fig:schematics_and_noise}(b), which is nearly identical to that used in Ref. \cite{Brock2021_characterization}. The one difference here is that we use a near quantum-limited TWPA \cite{Macklin2015} as a first-stage amplifier. We use the techniques described in Ref. \cite{Brock2021_characterization} to refer all input and output powers, as well as the system noise $S_{\mrm{noise}}(\omega)$, to the plane of the sample. The measured system noise, shown in Fig. \ref{fig:schematics_and_noise}(c), is due to the half-photon of vacuum noise $S_{\mrm{vac}}=\hbar\omega/2$ in the input/output transmission line \cite{Clerk2010} and the added noise of our amplifier chain $S_{\mrm{amp}}$, such that $S_{\mrm{noise}} = S_{\mrm{vac}}+S_{\mrm{amp}}$. For all of the charge sensitivity measurements we report, the noise floor near the sideband frequencies is dominated by this system noise, which is why we use the same notation for these two quantities. At sufficiently low gate modulation frequencies, however, the noise floor is dominated by $1/f$ charge noise \cite{Paladino2014}. This regime occurs below about $1$ kHz in our case \cite{Brock2021_characterization}. We determine the noise added by the TWPA and HEMT independently by measuring the gain of the amplifier chain and total system noise twice: once with the TWPA pump on and once with it off. Over the operating range of the cCPT (between $5.68$ GHz and $5.82$ GHz), the TWPA contributes $1.2$ photons of noise ($50\%$ of total) while the HEMT contributes $0.7$ photons ($30\%$ of total) on average. The room temperature amplifier contributes negligibly to the sample-referred system noise $S_{\mrm{noise}}$. The quantum limit of noise in this system is one photon, such that $S_{\mrm{noise}}^{\mrm{QL}} = \hbar\omega$, since phase-insensitive amplifiers must add at least a half-photon of noise \cite{Caves1982}. Thus, our average system noise is only a factor of $2.4$ greater than the quantum limit for this measurement scheme, such that the theoretical sensitivity (Eq. \ref{eq:theoretical_charge_sensitivity}) is only a factor of $\sqrt{2.4}$ greater than the quantum-limited sensitivity.
\begin{figure}[!t]
\includegraphics{combined_charge_sensitivity.pdf}
\caption{(a) Measured and theoretical charge sensitivities, obtained using Eqs. \eqref{eq:empirical_sensitivity} and \eqref{eq:theoretical_charge_sensitivity} respectively, as a function of gate and flux. Data is omitted where the sidebands could not be resolved from the noise floor. (b) Sample-referred spectrum analyzer trace of the optimal charge sensitivity measurement, corresponding to $\delta q = 14$~$\mu e/\sqrt{\mrm{Hz}}$. The carrier frequency is $\omega_{0}/2\pi = 5.806$ GHz, the gate modulation frequency is $\omega_{g}/2\pi = 350$ kHz, the span of each segment is $1$ kHz, and the resolution bandwidth is $B=10$ Hz. The noise floor near the carrier is due to $1/f$ charge noise \cite{Brock2021_characterization,Paladino2014}.}
\label{fig:combined_charge_sensitivity}
\end{figure}
In order to compare the cCPT's charge sensitivity with its theoretical performance, given by Eq. \eqref{eq:theoretical_charge_sensitivity}, we first measure $\delta q$ as a function of both the gate charge $n_{g}$ and external flux $\Phi_{\mrm{ext}}$. Although we can access a full period of $\Phi_{\mrm{ext}}$ (from $0$ to the magnetic flux quantum $\Phi_{0}$), we can only access the gate range $-0.65<n_{g}<0.65$ due to quasiparticle poisoning \cite{Brock2021_characterization}. We perform this measurement using an input power $P_{\mrm{in}}= -141$ dBm $\approx 8$ aW and gate modulation amplitude $q_{\mrm{rms}} = 10^{-3}e$. Ideally we would set $\omega_{g}$ to be significantly less than $\kappa_{\mrm{tot}}/2\approx 2\pi\times 700$ kHz to minimize Eq. \eqref{eq:theoretical_charge_sensitivity}, but in our experiments we observe cross-talk between our gate and flux lines at frequencies below about $650$ kHz. We therefore use $\omega_{g}/2\pi = 800$ kHz, such that the gate modulation does not also induce a flux modulation. To measure the reflected power and noise floor at $\omega_{0}\pm\omega_{g}$ we use a resolution bandwidth $B=1$ Hz.
The results of this measurement are shown in Fig. \ref{fig:combined_charge_sensitivity}(a). We find that the variation of $\delta q$ with $n_{g}$ and $\Phi_{\mrm{ext}}$ is in good agreement with theory, but our measured sensitivities are about $3$ times worse than theory. We attribute this discrepancy to two factors. First and foremost, the resonant frequency fluctuates due to $1/f$ charge and flux noise \cite{Brock2021_characterization,Brock2020} over the course of each measurement, which means our carrier is not always on resonance. On average, this reduces the output sideband power yielding worse charge sensitivity than expected. Second, we used a sufficiently high input power that $P_{\mrm{out}}(\omega_{0}\pm\omega_{g})$ scales sublinearly with $P_{\mrm{in}}$ due to the Kerr nonlinearity. Although this improves the sensitivity overall and was necessary to resolve the sidebands over a large area of the gate/flux parameter space, it causes the measured sensitivity to diverge from theory since the latter assumes $P_{\mrm{in}} \propto P_{\mrm{out}}(\omega_{0}\pm\omega_{g})$. Finally, since $S_{\mrm{noise}}/S_{\mrm{noise}}^{\mrm{QL}} \approx 2.4$, the factor of $3$ discrepancy between theory and experiment means our measured sensitivities are within a factor of $5$ of the quantum limit. In this measurement we find a minimum charge sensitivity of $24$~$\mu e/\sqrt{\mrm{Hz}}$ at $(n_{g},\Phi_{\mrm{ext}}) = (0.63, 0.0)$, whereas our predicted theoretical and quantum-limited sensitivities at this point are $9$~$\mu e/\sqrt{\mrm{Hz}}$ and $6$~$\mu e/\sqrt{\mrm{Hz}}$, respectively.
In order to optimize $\delta q$ we narrow our search to the gate range $0.6\leq |n_{g}|\leq 0.65$ and the flux points $\Phi_{\mrm{ext}}=0,\Phi_{0}/2$. At these flux points the resonant frequency of the cCPT is insensitive to flux, so we can reduce our gate modulation frequency to $\omega_{g}/2\pi = 350$ kHz without the gate/flux cross-talk interfering with our results. To maintain a small frequency modulation amplitude relative to $\omega_{g}$, we also reduce $q_{\mrm{rms}}$ to $5\times 10^{-4}e$. For this measurement we use a resolution bandwidth $B=10$ Hz.
\begin{table}[t]
\begin{tabular}{|c|c|c|c|} \hline
\textbf{Electrometer} & \begin{tabular}{@{}c@{}}$\delta q$ ($\mu e/\sqrt{\mrm{Hz}}$) \end{tabular} & \begin{tabular}{@{}c@{}} $P_{\mrm{in}}$ (aW) \end{tabular} & \begin{tabular}{@{}c@{}} $n$ photons \end{tabular} \\ \hline
cCPT* & $14$ & $16$ & $1$ \\ \hline
Best gate sensor\cite{Schaal2020}* & $0.25$ & $100$ & $190$ \\ \hline
Best rf-SET\cite{Brenning2006} & $0.9$ & $6\times 10^{6}$ & $2\times 10^{5}$ \\ \hline
Andresen et al.\cite{Andresen2008} & $2.3$ & $3\times 10^{8}$ & $2\times 10^{6}$ \\ \hline
L-SET\cite{Sillanpaa2005}* & $30$ & $1\times 10^{4}$ & $70$ \\ \hline
Naaman et al.\cite{Naaman2006}* & $52$ & $2\times 10^{3}$ & $150$ \\ \hline
Bell et al.\cite{Bell2012}* & $70$ & $3\times 10^{7}$ & $2\times 10^{5}$ \\ \hline
rf-QPC\cite{Cassidy2007} & $200$ & $1\times 10^{9}$ & $7\times 10^{7}$ \\ \hline
\end{tabular}
\caption{Comparison of the cCPT with a representative set of fast and ultrasensitive electrometers. Asterisks indicate dispersive electrometers.}
\label{tab:comparing_electrometers}
\end{table}
We find a minimum charge sensitivity of $14$~$\mu e/\sqrt{\mrm{Hz}}$ at $(n_{g},\Phi_{\mrm{ext}}) = (0.625,0.0)$ using an input power $P_{\mrm{in}} = -138$ dBm $\approx 16$ aW. Under these conditions our predicted theoretical and quantum-limited sensitivities are $5$~$\mu e/\sqrt{\mrm{Hz}}$ and $3$~$\mu e/\sqrt{\mrm{Hz}}$, respectively. The spectrum analyzer trace of this optimal measurement is shown in Fig. \ref{fig:combined_charge_sensitivity}(b). At this bias point the resonant frequency is $\omega_{0}/2\pi = 5.806$~GHz, the external damping is $\kappa_{\mrm{ext}}/2\pi = 1.24$~MHz, and the total damping is $\kappa_{\mrm{tot}}/2\pi = 1.62$~MHz, such that the number of intracavity photons is $n = 4\kappa_{\mrm{ext}}P_{\mrm{in}}/\hbar\omega_{0}\kappa_{\mrm{tot}}^{2} \approx 1$. This single-photon-level sensitivity is rivaled only by gate-based sensors \cite{Schaal2020}, rf-SETs \cite{Brenning2006}, and carbon nanotube-based rf-SETs \cite{Andresen2008}, all of which operate with orders of magnitude more photons. In Table \ref{tab:comparing_electrometers} we compare the performance of the cCPT to a representative set of fast (detection bandwidth $\gtrsim 1$ MHz) and ultrasensitive ($\delta q < 10^{-3} e/\sqrt{\mrm{Hz}}$) electrometers. Clearly, the cCPT is unparalleled in its ability to operate at low powers and photon numbers. As discussed earlier, this makes it ideal for mediating an optomechanical interaction that reaches the single-photon strong coupling regime \cite{Rimberg2014}.
There remains significant room for improving the sensitivity of the cCPT, with two distinct approaches for doing so. The most promising approach is to reduce quasiparticle poisoning (QP) \cite{Aumentado2004}, which prevents us from operating at gate biases above $|n_{g}|\approx 0.65$ \cite{Brock2021_characterization}. If we were able to operate the present device at $(n_{g},\Phi_{\mrm{ext}}) = (0.9,\Phi_{0}/2)$ we would expect to attain a charge sensitivity of $\delta q \approx 0.4$~$\mu e/\sqrt{\mrm{Hz}}$, assuming the same factor of $3$ discrepancy with theory as we observe experimentally. The present device was designed with a $9$ nm thick CPT island \cite{Brock2021_characterization} to suppress QP \cite{Yamamoto2006}, but other fabrication techniques could be employed to reduce it further. These include oxygen-doping the CPT island \cite{Aumentado2004} and embedding quasiparticle traps near the CPT \cite{Rajauria2012}. The other approach is to mitigate the discrepancy between our measured sensitivities and the quantum limit. One such improvement would be to use a truly quantum-limited amplifier, which would improve our sensitivities by a factor of $\sqrt{S_{\mrm{noise}}/\hbar\omega}~\approx~\sqrt{2.4}$. Another such improvement would be to stabilize the resonant frequency against $1/f$ noise using a Pound-locking loop \cite{Lindstroem2011}. It may also be possible to improve the sensitivity of the cCPT by exploiting the nonlinearity of the cCPT \cite{Laflamme2011,Tosi2019} or incorporating a parametric drive near $2\omega_{0}$ \cite{Krantz2016}.
Several important applications exist for single-photon-level charge sensing with the cCPT. First and foremost, the cCPT can be used to dispersively sense any quantity that can be tied to electrical charge, two notable examples being the spin state of quantum-dot-based qubits and the position of a charged nanomechanical resonator. For quantum-dot-based qubits, the spin states can be encoded in charge states via spin-to-charge conversion \cite{Petta2005,Hanson2007}. For a charged nanomechanical resonator, the position of the resonator can be encoded in the charge on a capacitor \cite{Knobel2003,LaHaye2004,Naik2006}. In both cases, the measurement backaction on the relevant degree of freedom is proportional to the number of photons in the cavity, such that single-photon-level operation is preferable \cite{D'Anjou2019,Clerk2010}. Second, the cCPT can be readily integrated with near-quantum-limited amplifiers \cite{Castellanos-Beltran2007, Macklin2015, Sivak2019}, which typically saturate well below the level of power required by rf-SETs \cite{Brenning2006,Schoelkopf1998} and rf-QPCs \cite{Cassidy2007}. Finally, the cCPT has been proposed as a platform for mediating an optomechanical interaction that reaches the single-photon strong coupling regime \cite{Rimberg2014}. Our demonstration of single-photon-level electrometry with the cCPT supports the feasibility of this proposal and represents an important step toward its realization.
\begin{acknowledgments}
We thank W. F. Braasch for helpful discussions and W. Oliver for providing the TWPA used in these measurements. The sample was fabricated at Dartmouth College and the Harvard Center for Nanoscale Systems. B.L.B., S.K., and A.J.R. were supported by the National Science Foundation under Grant No. DMR-1807785. J.L. was supported by the Army Research Office under Grant No. W911NF-13-1-0377. M.P.B. was supported by the National Science Foundation under Grant No. DMR-1507383.
\end{acknowledgments}
|
\section{Numerically stable Multinomial Diffusion in log space}
\label{appendix:numerically_stable_diffusion}
In this section we explain how Multinomial Diffusion models can be implemented in a numerically safe manner in log-space. Note that in addition to this appendix with pseudo-code, the actual source code will also be released. First we define a few helper functions:
\begin{lstlisting}
def log_add_exp(a, b):
maximum = max(a, b)
return maximum + log(exp(a - maximum) + exp(b - maximum))
def log_sum_exp(x):
maximum = max(x, dim=1, keepdim=True)
return maximum + log(exp(x - maximum).sum(dim=1))
def index_to_log_onehot(x, num_classes):
# Assume that onehot axis is inserted at dimension 1
x_onehot = one_hot(x, num_classes)
# Compute in log-space, extreme low values are later
# filtered out by log sum exp calls.
log_x = log(x_onehot.clamp(min=1e-40))
return log_x
def log_onehot_to_index(log_x):
return log_x.argmax(1)
def log_1_min_a(a):
return log(1 - a.exp() + 1e-40)
\end{lstlisting}
Then we can initialize the variables we are planning to utilize for the multinomial diffusion model. This is done with float64 variables to limit the precision loss in the \texttt{log\_1\_min\_a} computation. Since these are precomputed and later converted to float32, there is no meaningful increase in computation time.
\begin{lstlisting}
alphas = init_alphas()
log_alpha = np.log(alphas)
log_cumprod_alpha = np.cumsum(log_alpha)
log_1_min_alpha = log_1_min_a(log_alpha)
log_1_min_cumprod_alpha = log_1_min_a(log_cumprod_alpha)
\end{lstlisting}
Then we can define the functions that we utilize to compute the log probabilities of the categorical distributions of the forward process. The functions below compute the probability vectors for $q({\bm{x}}_t | {\bm{x}}_{t-1})$, $q({\bm{x}}_t | {\bm{x}}_{0})$ and $q({\bm{x}}_{t-1} | {\bm{x}}_t, {\bm{x}}_{0})$.
\begin{lstlisting}
def q_pred_one_timestep(log_x_t, t):
# Computing alpha_t * E[xt] + (1 - alpha_t) 1 / K
log_probs = log_add_exp(
log_x_t + log_alpha[t],
log_1_min_alpha[t] - log(num_classes)
)
return log_probs
def q_pred(log_x0, t):
log_probs = log_add_exp(
log_x0 + log_cumprod_alpha[t],
log_1_min_cumprod_alpha[t] - log(num_classes)
)
return log_probs
def q_posterior(log_x0, log_x_t, t):
# Kronecker delta peak for q(x0 | x1, x0).
if t == 0:
log_probs_xtmin = log_x0
else:
log_probs_xtmin = q_pred(log_x0, t - 1)
# Note log_x_t is used not x_tmin, subtle and not straightforward
# why this is true. Corresponds to Algorithm 1.
unnormed_logprobs = log_probs_xtmin + q_pred_one_timestep(log_x_t, t)
log_probs_posterior = unnormed_logprobs - log_sum_exp(unnormed_logprobs)
return log_probs_posterior
\end{lstlisting}
Some magic is happening in \texttt{q\_pred\_one\_timestep}. Recall that at some point we need to compute $\mathcal{C}({\bm{x}}_t | (1 - \beta_t) {\bm{x}}_{t-1} + \beta_t / K )$ for different values of ${\bm{x}}_t$, which when treated as a function outputs $(1 - \beta_t) + \beta_t / K$ if ${\bm{x}}_t = {\bm{x}}_{t-1}$ and $\beta_t / K$ otherwise. This function is symmetric, meaning that $\mathcal{C}({\bm{x}}_t | (1 - \beta_t) {\bm{x}}_{t-1} + \beta_t / K ) = \mathcal{C}({\bm{x}}_{t-1} | (1 - \beta_t) {\bm{x}}_{t} + \beta_t / K )$. This is why we can switch the conditioning and immediately return the different probability vectors for ${\bm{x}}_t$. This also corresponds to Equation~\ref{eq:q_posterior}.
Then using the \texttt{q\_posterior} function as parametrization we predict the probability vector for $p({\bm{x}}_{t-1} | {\bm{x}}_t)$ using a neural network.
\begin{lstlisting}
def p_pred(log_x_t, t):
x_t = log_onehot_to_index(log_x_t)
log_x_recon = logsoftmax(neuralnet(x_t, t))
log_model_pred = q_posterior(log_x_recon, log_x_t, t)
return log_model_pred
\end{lstlisting}
And then finally we can compute the loss term $L_t$ using the KL divergence for categorical distributions:
\begin{lstlisting}
def categorical_kl(log_prob_a, log_prob_b):
kl = (log_prob_a.exp() * (log_prob_a - log_prob_b)).sum(dim=1)
return kl
def compute_Lt(log_x0, log_x_t, t):
log_true_prob = q_posterior(log_x0, log_x_t, t)
log_model_prob = p_pred(log_x_t, t)
kl = categorical_kl(log_true_prob, log_model_prob)
loss = sum_except_batch(kl)
return loss
\end{lstlisting}
Coincidentally this code even works for $L_0$ because ${\bm{x}}_0$ is onehot and then:
$$-\log \mathcal{C}({\bm{x}}_0 | \hat{{\bm{x}}}_{0}) - \sum_k {\bm{x}}_{0,k} \log \hat{{\bm{x}}}_{0,k} = \sum_k {\bm{x}}_{0,k} [\underbrace{\log {\bm{x}}_{0,k}}_{0 \text{ or } \log 0} - \log \hat{{\bm{x}}}_{0,k}] = \mathrm{KL}(\mathcal{C}({\bm{x}}_{0})||\mathcal{C}(\hat{{\bm{x}}}_{0})),$$ where in the last term ${\bm{x}}_{0}$ and $\hat{{\bm{x}}}_{0}$ are probability vectors and $0 \log 0$ is defined to be $0$.
\section{Additional experiments}
\label{sec:additional_experiments}
A comparison of the performance for Cartesian products with different bases is shown in Table \ref{tab:cartesian_products}. Note that this experiment was performed using a somewhat smaller architecture then in the main text. As can be seen, the performance difference between different Cartesian products is relatively small. The performance does decreases slightly over larger base numbers, indicating that it is better to choose a small base that results in fewer overall dimensions.
\begin{table}[H]
\centering
\caption{Cartesian Products with different base numbers trained using a slightly smaller version of the Argmax AR Flow on \texttt{text8}.}
\label{tab:cartesian_products}
\scalebox{.95}{
\begin{tabular}{l c }
\toprule
Model & text8 (bpc) \\ \midrule
$d_m=1, M=27$ & 1.45 \\
$d_m=2, M=6$ & 1.44 \\
$d_m=3, M=3$ & 1.44 \\
$d_m=5, M=2$ & 1.44 \\
\bottomrule
\end{tabular}}
\vspace{-.5cm}
\end{table}
A comparison of sampling time speeds are shown in Table~\ref{tab:time_experiment}. A couple of orders in magnitude difference can be seen comparing autoregressive versus non-autoregressive models. This highlights the importance of researching generative models that can be built from non-autoregressive components. The main source of difference between our coupling approach and IAF/SCF is that we utilize mixture of discretized logistics \citep{ho2019flow++} as coupling transformation, which requires a iterative process to invert over 1 dimension. The multinomial diffusion takes in-between the time of autoregressive and coupling models. Also reducing steps reduces the required sampling time, as is expected.
\begin{table}[H]
\centering
\vspace{-.25cm}
\caption{Comparison of different methods in terms of sample time. Sample time is measured by generating a single text sample of length 256 averaged over $10$ runs, unless specified otherwise. }
\label{tab:time_experiment}
\begin{tabular}{l l l c c c}
\toprule
Model type & & Model & Sample time (s) \\ \midrule
\multirow{1}{*}{ARM}
& & 64 Layer Transformer {\footnotesize \citep{alrfou2019transformer}} & 35.5$^\dagger$ \\ \midrule
\multirow{2}{*}{\small VAE} & & AF/AF$^\star$ (AR) {\footnotesize \citep{ziegler2019latent}} & 156 {\small$\pm 1.8$} \\
& & IAF / SCF$^\star$ {\footnotesize \citep{ziegler2019latent}} & 0.04 {\footnotesize $\pm0.004$} \\ \midrule
\multirow{3}{*}{Generative Flow}
& & Argmax Flow, AR (ours) & 115 {\small$\pm 0.03$} \\
& & Argmax Coupling Flow (ours) & 0.40 {\footnotesize $\pm0.03$} \\
& & Discrete Flow {\footnotesize \citep{discrete2019tran}} & 0.16$^\dagger$\hspace{.75cm} \\ \midrule
\multirow{2}{*}{Diffusion} & & Multinomial Text Diffusion (ours) & 26.6 {\small $\pm2.2$}$^{\ddagger}$ \\
& & Multinomial Text Diffusion, 100 steps (ours) & 2.4 {\small $\pm0.16$} \\ \bottomrule \end{tabular}
\begin{flushleft}
\scriptsize{$\dagger$ Computed on a 288-length sequence instead of 256-length, taken from \citep{discrete2019tran}.} \\
\scriptsize{$\ddagger$ This result is for the complete 1000 timesteps chain, improvements are possible by skipping steps.}
\vspace{-.5cm}
\end{flushleft}
\end{table}
Due to the computational cost of running normalizing flows, it is not possible for us to run every model many times. However, generally single-run results suffice, as the performance variance of these models is relatively small. In Table~\ref{tab:stdev} the standard deviation and average performance for a selection of models is shown, taken over $3$ runs. Observe that these standard deviations are small compared to the reported differences between the models. Notice that standard deviations for coupling models are larger, but the performance difference between those types of models is also larger.
\begin{table}[H]
\centering
\vspace{-.25cm}
\caption{Average and standard deviations of several models.}
\label{tab:stdev}
\scalebox{.9}{
\begin{tabular}{l c c c c c}
\toprule
Dequantization & Flow type & Dataset & average & stdev \\ \midrule
Argmax Flow (ours) & AR & text8 & 1.38 & $0.001$ \\
Argmax Flow (ours) & AR & enwik8 & 1.42 & $0.008$ \\
Argmax Flow (ours) & Coupling & text8 & 1.82 & $0.017$ \\
Argmax Flow (ours) & Coupling & enwik8 & 1.93 & $0.012$ \\\bottomrule
\end{tabular}}
\vspace{-.3cm}
\end{table}
Finally, we also compare argmax flows to a situation where its density model exactly matches the density model in \citep{lippe2020categorical} on \texttt{text8}. In this experiment Argmax Flows (1.43 bpc) outperform CategoricalNF (1.45 bpc) in an equal setting.
\section{Experimental details}
\label{app:experimental_details}
This section gives details on experimental setup, architectures and optimization hyperparameters. In addition, the code to reproduce experiments will be released publicly.
\paragraph{Diffusion settings}
For diffusion we use the cosine schedule for $\{\alpha_t\}$ from \cite{nichol2021improved} with the difference that what was previously $\sqrt{\bar{\alpha}_t}$ is now $\bar{\alpha}_t$, so that their factor $\sqrt{\bar{\alpha}_t}$ for the Gaussian mean is equal to our factor $\bar{\alpha}_t$ for categorical parameters. Specifically, our $\bar{\alpha}_t$ are defined using:
\begin{equation*}
\bar{\alpha}_t =\frac{f(t)}{f(0)} \quad f(t) = \cos \left( \frac{t/T + s}{1 + s} \cdot \frac{\pi}{2} \right), \quad s = 0.008,
\end{equation*}
where $T$ is the total number of diffusion steps. \citet{nichol2021improved} show that instead of sampling $t$ uniformly, variance is reduced when $t$ is importance-sampled with $q(t) \propto \sqrt{\mathbb{E}[L_t^2]}$, which is estimated using training statistics, and we use their approach. The objective can be summarized as:
\small
\begin{equation}
\log P({\bm{x}}_0) \geq \mathbb{E}_{t \sim q(t), {\bm{x}}_t \sim q({\bm{x}}_t | {\bm{x}}_0)}
\left[ - \frac{1}{q(t)} \mathrm{KL} \big{(}q({\bm{x}}_{t-1} | {\bm{x}}_t, {\bm{x}}_0) | p({\bm{x}}_{t-1} | {\bm{x}}_t) \right].
\end{equation}
\normalsize
\paragraph{Gumbel properties}
In Table~\ref{tab:gumbel} a useful overview of Gumbel properties are given. These equations can be used to sample and compute the likelihood of the (truncated) Gumbel distributions. For a more extensive treatment see \citep{Maddison2014Astarsampling,kool2019}.
\begin{table}[H]
\centering
\caption{Summary of Gumbel properties.}
\label{tab:gumbel}
\scalebox{.67}{
\begin{tabular}{l p{3.95cm} p{4.1cm}}
\toprule
Description & $\log p$ & Sample \\ \midrule
\multirow{2}{*}{$\mathrm{Gumbel}(g | \phi)$} & \multirow{2}{*}{$\phi - g - \exp(\phi - g)$} & $g = -\log (-\log(u)) + \phi$ \newline $u \sim \mathcal{U}(0, 1)$
\\ \midrule
\multirow{2}{*}{$\max_i \mathrm{Gumbel}(g_i | \phi)$} & $\log \mathrm{Gumbel}(g_{\max} | \phi_{\max})$ \newline $\phi_{\max} = \log \sum_i \exp \phi_i$ & $g_{\max} \sim \mathrm{Gumbel}(\phi_{\max})$ \newline $\phi_{\max} = \log \sum_i \exp \phi_i$
\\ \midrule
\multirow{2}{*}{\small$\mathrm{TruncGumbel}(g| \phi, T)$} & \small $\phi - g - \exp(\phi - g) + \exp(\phi - T)$ \newline \text{ if } $g < T$ \text{ else } $-\infty$ & \small $g = \phi - \log(\exp(\phi - T) - \log u)$ \newline $u \sim \mathcal{U}(0, 1)$ \\ \bottomrule
\end{tabular}}
\end{table}
\subsection{Language Modelling}
For the language modelling experiments we utilize the standard \texttt{text8} dataset with sequence length $256$ and \texttt{enwik8} dataset with sequence length $320$. The train/val/test splits are 90000000/5000000/5000000 for both \texttt{text8} and \texttt{enwik8}, as is standard in literature. The Multinomial Text Diffusion models are trained for $300$ epochs, whereas the Argmax Flows are trained for $40$ epochs, with the exception of the Argmax Coupling Flow on enwik8 which only needs to be trained for $20$ epochs. Further details are presented in Tables~\ref{tab:details_text_models} and \ref{tab:architecture_text_models}. In addition, the code to reproduce results will be publicly available. There are no known ethics issues with these datasets at the time of writing.
\begin{table}[H]
\centering
\caption{Optimization details for text models.}
\label{tab:details_text_models}
\scalebox{.85}{
\begin{tabular}{l l l l l l l l l }
\toprule
Model & batch size & lr & lr decay & optimizer & dropout \\ \midrule
Multinomial Text Diffusion (text8) & 32 & 0.0001 & 0.99 & Adam & 0 \\
Multinomial Text Diffusion (enwik8) & 32 & 0.0001 & 0.99 & Adam & 0 \\
Argmax AR Flow (text8) & 64 & 0.001 & 0.995 & Adam & 0.25 \\
Argmax AR Flow (enwik8) & 64 & 0.001 & 0.995 & Adam & 0.25 \\
Argmax Coupling Flow (text8) & 16 & 0.001 & 0.995 & Adamax & 0.05 \\
Argmax Coupling Flow (enwik8) & 32 & 0.001 & 0.995 & Adamax & 0.1 \\ \bottomrule
\end{tabular}}
\end{table}
\begin{table}[H]
\centering
\caption{Architecture description for text models.}
\label{tab:architecture_text_models}
\scalebox{.85}{
\begin{tabular}{l l l l l l l l l }
\toprule
Model & Architecture description \\ \midrule
Multinomial Text Diffusion (text8) & 12-layer transformer 8 global, 8 local heads / 1000 diffusion steps \\
Multinomial Text Diffusion (enwik8) & 12-layer transformer 8 global, 8 local heads / 4000 diffusion steps \\
Argmax AR Flow (text8) & 2-layer LSTM, 2048 hidden units \\
Argmax AR Flow (enwik8) & 2-layer LSTM, 2048 hidden units \\
Argmax Coupling Flow (text8) & 2-layer bi-directional LSTM, 512 hidden units \\
Argmax Coupling Flow (enwik8) & 2-layer bi-directional LSTM, 768 hidden units \\ \bottomrule
\end{tabular}}
\end{table}
\subsection{Cityscapes}
\paragraph{Preprocessing}
The Cityscapes \citep{Cordts2016cityscapes} segmentation maps are re-sampled to a $32$ by $64$ pixel image using nearest neighbour interpolation. The original segmentation maps are downloaded from \url{https://www.cityscapes-dataset.com/downloads/} where all files are contained in \texttt{gtFine\_trainvaltest.zip}. Note that we train on a $8$-class problem since we only consider what is called the \texttt{category\_id} field in torchvision. We re-purpose the validation set as test set, containing $500$ maps. The original train set containing $2975$ maps is split into $2500$ maps for training and $475$ maps for validation. The original test set is not utilized. To aid reproducibility we will publish source code that includes the preprocessing and the dataloaders. There are no known ethics issues with the segmentation maps at the time of writing. License is located at \url{https://www.cityscapes-dataset.com/license/}.
\paragraph{Architectures} For Cityscapes all models utilize the same architectures, although they represent a different part for their respective model designs. The density model $p({\bm{v}})$ consist of $4$ levels with $10$ subflows each, separated by squeeze layers, where each subflow consists of a $1$ $\times$ $1$ convolution and an affine coupling layer. The coupling layers are parametrized by DenseNets \citep{huang2017densely}. The same model is used for the latent distribution in the VAE (usually referred to as $p({\bm{z}})$ in literature). The probabilistic inverse $q({\bm{v}} | {\bm{x}})$ is modelled by a single level flow that has $8$ subflows, again consisting of affine coupling layers and $1$ $\times$ $1$ convolutions. To condition on ${\bm{x}}$ it is processed by a DenseNet which outputs a representation for the coupling layers that is concatenated to the original input. The same model is utilized to parametrize the VAE encoder (commonly referred to as $q({\bm{z}} | {\bm{x}})$). The VAE additionally has a model for the decoder $p({\bm{x}} | {\bm{z}})$ which is parametrized by a DenseNet which outputs the parameters for a categorical distribution. The models are optimized using the same settings, and no hyperparameter search was performed. Specifically, the models are optimized with minibatch size $64$ for $2000$ epochs with the Adamax optimizer with learning rate $0.001$ and a linear learning rate warmup of $10$ epochs and a decay factor of $0.995$.
\subsection{Range of considered hyperparameters}
For Multinomial Text Diffusion we experimented with the depth of transformers $\{1, 2, 4, 8, 12, 16, 20\}$ and the hidden size $\{128, 256, 512, 1024\}$. We found that models with depth $12$ and $512$ could be trained in a reasonable amount of time while giving good performance. For the cityscapes experiments no hyperparameter search was performed.
\subsection{Details on latent normalizing flows for text8}
We utilize the official code repository from \cite{ziegler2019latent} in here\footnote{\url{https://github.com/harvardnlp/TextFlow}}. The original code utilizes $10$ ELBO samples, which is relatively expensive. For that reason we instead opt for $1$ ELBO sample and find it gives similar results. The batch size is increased from $16$ to $32$. Additionally we reduce the KL scheduling from $4$ initial $10^{-5}$ epochs to only $2$ initial $10^{-5}$ epoch and we anneal linearly over the next $4$ epochs instead of over the next $10$ epochs. In total the models are optimized for $30$ epochs. We verify that the resulting models still achieve similar performance on the Penn Tree Bank experiment compared to the original paper in terms of ELBO values: Our hyperparameter setup for AF/AF achieves slightly better performance with 1.46 versus 1.47 bpc and for IAF/SCF achieves slightly worse 1.78 versus 1.76 bpc.
\subsection{Computing infrastructure}
Experiments where run on NVIDIA-GTX 1080Ti GPUs, CUDA 10.1 with Python version 3.7.6 in Pytorch 1.5.1 or 1.7.1.
\section{Samples from the text models}
Samples from our proposed models are presented in Table \ref{tab:samples_text_appendix} and a Multinomial Text Diffusion train is shown in Figure \ref{fig:samples_text_chain_appendix}, these results were not cherry-picked.
\begin{table}[H]
\centering
\caption{Samples from models trained on text8.}
\label{tab:samples_text_appendix}
\scalebox{.6}{
\begin{tabular}{l l p{20cm}}
\toprule
Model & Nr & Text \\ \midrule
\multirow{12}{*}{\rotatebox[origin=c]{90}{ \textit{Multinomial Diffusion}}}
& 1 & \texttt{ that the role of tellings not be required also action characters passed on constitution ahmad a nobilitis first be closest to the cope and dhur and nophosons she criticized itm specifically on august one three movement and a renouncing local party of exte} \\
& 2 & \texttt{nt is in this meant the replicat today through the understanding element thinks the sometimes seven five his final form of contair you are lotur and me es to ultimately this work on the future all all machine the silon words thereis greatly usaged up not t} \\
& 3 & \texttt{arity island louis has convinced privatist provinces the restrained marriage of his income ted guilds which in gulick performed in one nine six seven then sponly onward the bambat loving in separate including tichatta westell s doubled a bound of his futur} \\
& 4 & \texttt{same early duration without education as a golden core power to the pirit of spain arriving wise speech art and r t plain firman q one five six the same as part of herald h rogenszers a art poetic of literature at shaft bressen three five three five eight } \\ \midrule
\multirow{12}{*}{\rotatebox[origin=c]{90}{\textit{AR Argmax Flow}}}
& 1 & \texttt{ heartedness frege thematically infered by the famous existence of a function f from the laplace definition we can analyze a definition of binary operations with additional size so their functionality cannot be reviewed here there is no change because its } \\
& 2 & \texttt{otal cost of learning objects from language to platonic linguistics examines why animate to indicate wild amphibious substances animal and marine life constituents of animals and bird sciences medieval biology biology and central medicine full discovery re} \\
& 3 & \texttt{o use language combined with any of its subsets evolved into the group containing the primary concepts of a daily line on off the road and the material emulation of welcomes and prospects of pleasure and exercise have been committed projects in the economy} \\
& 4 & \texttt{en that are beginning to forge since october one nine five zero the mandate was planted at k nigsberg during the car horizon at first please refer to a small government situated as well as in all these countries finally giving birth to a band here he was a} \\ \midrule
\multirow{12}{*}{\rotatebox[origin=c]{90}{ \textit{Coupling Argmax Flow}}}
& 1 & \texttt{ns fergenur d alpha and le heigu man notabhe leglon lm n two six a gg
opa movement as sympathetic dutch the term bilirubhah acquired the bava
rian cheeh segt thmamouinaire vhvinus lihnos ineoneartis or medical iod
ine the rave wesp published harsy varb hhgh} \\
& 2 & \texttt{and inequalities syllee mike jean demet in standard rather than fmxed liga and a piare nut is gruncionde aodadneveshiopyhabally uchc one viredtlty three ben yi agricultariis the only mefamantia or nuil and mid satio for kigou wore not on the war rits af} \\
& 3 & \texttt{e g chain within the sale of cooperative oppine p nge tyae yarot bouatta real frequency one mbj or rorbepetam iw by someone c langt b kindoms is the single yenta valve nor eosed collagen surkeys in the goubark cuisine of animum and two trantual measurement} \\
& 4 & \texttt{hilepuin the king pete was added to or who cefralded to kiark n and panhpur not souhhvestern bat batas mudtlu for this creatures chew palenque lii lasron gentla tzanemi derived from oo four issais nivissos with the name convertinus magaa named wes orieanr} \\ \midrule
\end{tabular}}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=.99\textwidth]{images/samples/text_chain.pdf}
\caption{Intermediate steps of the generation chain of the Multinomial Text Diffusion model trained on \texttt{text8}.}
\label{fig:samples_text_chain_appendix}
\end{figure}
\section{Reproducing Discrete Flows}
\label{sec:reproducing_discrete_flows}
In this section we detail our efforts to reproduce the results from discrete flows \citep{discrete2019tran}. Specifically, we are interested in the discrete flows models that map to \textit{factorized} distributions, for instance the discrete bipartite (coupling) flow. We avoid situations where an autoregressive base distribution is used, it may be difficult to identify how much the flow is actually learning versus the ARM as base. For this paper an official implementation was released at \url{https://github.com/google/edward2/blob/master/edward2/tensorflow/layers/} in the files \texttt{discrete\_flows.py} and \texttt{utils.py}. However, this codebase contains only the high-level modules and code for the toy example, it does not contain the specific code related to the language experiments. These high-level modules and the toy problem were ported to PyTorch here: \url{https://github.com/TrentBrick/PyTorchDiscreteFlows}. Using this codebase, we were able to compare on the quantized eight Gaussians toy dataset, as depicted in Figure~\ref{fig:8gaussiansexp}. In this experiment we clearly see that argmax flows outperform discrete flows both numerically (6.32 versus 7.0 nats) and visually by comparing the samples or probability mass function.
\begin{figure}[h]
\centering
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=\textwidth]{images/quantizedgaussian/original_df.pdf}
\caption{Samples from Discrete Flow using a single layer, taken from \citep{discrete2019tran}.}
\end{subfigure}
\hspace{1cm}
\begin{subfigure}{.4\textwidth}
\includegraphics[width=\textwidth]{images/quantizedgaussian/data_pmf.pdf}
\caption{Samples from the quantized 8 Gaussians data distribution.}
\end{subfigure}
\begin{subfigure}{.4\textwidth}
\includegraphics[width=\textwidth]{images/quantizedgaussian/pytorch_df_pmf.pdf}
\caption{Samples from the Discrete Flows PyTorch re-implementation, achieving 7.0 nats.}
\end{subfigure}
\hspace{1cm}
\begin{subfigure}{.4\textwidth}
\includegraphics[width=\textwidth]{images/quantizedgaussian/argmax_pmf.pdf}
\caption{Probability mass of our Argmax Flow using a single layer, achieving 6.32 nats.}
\end{subfigure}
\caption{Reproduction of the quantized eight Gaussians experiment. Plots show either the probability mass function or weighted number of samples (which will tend towards the pmf).}
\label{fig:8gaussiansexp}
\end{figure}
Subsequent efforts by others to reproduce the language experiments failed (see \url{https://github.com/TrentBrick/PyTorchDiscreteFlows/issues/1}). In another work, \citet{lippe2020categorical} also noticed the difficulty of getting discrete flows to succesfully optimize, as detailed in the set shuffling/summation experiment corresponding to Table 5 in the paper.
For this paper we also tried to reproduce the language experiments. After verifying the correctness of the \texttt{one\_hot\_argmax}, \texttt{one\_hot\_minus} and \texttt{one\_hot\_add} functions in \url{https://github.com/TrentBrick/PyTorchDiscreteFlows}, we implemented an autoregressive discrete flow layer with an expressive network, in an effort to limit the accumulated gradient bias. Recall that an autoregressive layer is more expressive than a coupling layer as it has more dependencies between dimensions. As can be seen in Table~\ref{tab:discrete_flows_reimplementation} our re-implementation also performed considerably worse, matching the experience of the others described above.
\begin{table}[H]
\centering
\caption{Discrete Flows on \texttt{text8}. Note that AR is more expressive than coupling.}
\label{tab:discrete_flows_reimplementation}
\scalebox{.95}{
\begin{tabular}{l c }
\toprule
Model & text8 (bpc) \\ \midrule
Discrete Flows from paper (coupling, factorized base, without scale) & 1.29 \\
Discrete Flows from paper (coupling, factorized base, with scale) & 1.23 \\
Discrete Flows reimplementation (AR, factorized base, without scale) & 4.13 \\ \midrule
Argmax Flow, AR (ours) & 1.38 \\
Argmax Coupling Flow (ours) & 1.80 \\
\bottomrule
\end{tabular}}
\end{table}
\paragraph{Final remarks}
We have had extensive contact with the authors of \citep{discrete2019tran} to resolve this issue over the course of several months. Unfortunately it is not possible for them to share the code for the language flows due to internal dependencies. Also, we have not been able to find any implementation of discrete flows online that achieves the reported performance on text. The authors generously offered to look at our reimplementation, which we have shared with them. At the time of writing we have not yet heard anything back on the code. For the reasons described in this appendix, we currently assume that the language experiments in discrete flows are not reproducible.
\newpage
\section{Background}
\label{sec:prelim}
\paragraph{Normalizing Flows}
Given ${\mathcal{V}} = \mathbb{R}^d$ and ${\mathcal{Z}} = \mathbb{R}^d$ with densities $p_{V}$ and $p_{Z}$ respectively, normalizing flows \citep{rezende2015norm} learn a bijective and differentiable transformation $g : {\mathcal{Z}} \to {\mathcal{V}}$ such that the change-of-variables formula gives the density at any point ${\bm{v}} \in {\mathcal{V}}$:
\small
\begin{equation}
\label{eq:cov}
p_{V}({\bm{v}}) = p_{Z}({\bm{z}})\cdot \left|\det \frac{\mathrm{d}{\bm{z}}}{\mathrm{d}{\bm{v}}} \right|, \qquad {\bm{v}} = g({\bm{z}}),
\end{equation}
\normalsize
where $p_{Z}$ can be any density (usually chosen as a standard Gaussian). Thus, normalizing flows provide a powerful framework to learn \emph{exact} density functions. However, \Cref{eq:cov} is restricted to continuous densities.
To learn densities on ordinal discrete data (such as natural images), typically dequantization noise is added \citep{uria2013rnade, theis2016note, ho2019flow++}. \citet{nielsen2007survae} reinterpreted dequantization as a surjective flow layer ${\bm{v}} \mapsto {\bm{x}}$ that is deterministic in one direction (${\bm{x}} = \mathsf{round}({\bm{v}})$) and stochastic in the other (${\bm{v}} = {\bm{x}} + {\bm{u}}$ where ${\bm{u}} \sim q({\bm{u}}|{\bm{x}})$). Using this interpretation, dequantization can be seen as a probabilistic right-inverse for the rounding operation in the latent variable model given by:
\begin{equation*}
\small
{P({\bm{x}}) =\! \int \! P({\bm{x}}|{\bm{v}}) p({\bm{v}})\rmd{\bm{v}}, \quad P({\bm{x}}|{\bm{v}}) \!=\! \delta\big({\bm{x}} \!=\! \mathsf{round}({\bm{v}})\big)},
\end{equation*}
where $\mathsf{round}$ is applied elementwise. In this case, the density model $p({\bm{v}})$ is modeled using a normalizing flow. Learning proceeds by introducing the variational distribution $q({\bm{v}}|{\bm{x}})$ that models the probabilistic right-inverse for the rounding surjection and optimizing the evidence lower bound (ELBO):
\small
\begin{align}
\log P({\bm{x}}) &\geq \mathbb{E}_{{\bm{v}} \sim q({\bm{v}}|{\bm{x}})} \left[ \log P({\bm{x}} | {\bm{v}}) + \log p({\bm{v}}) - \log q({\bm{v}}|{\bm{x}}) \right] = \mathbb{E}_{{\bm{v}} \sim q({\bm{v}}|{\bm{x}})} \left[ \log p({\bm{v}}) - \log q({\bm{v}}|{\bm{x}}) \right].
\label{eq:elbo_objective}
\end{align}
\normalsize
The last equality holds under the constraint that the support of $q({\bm{v}}|{\bm{x}})$ is enforced to be only over the region $\mathcal{S} = \{ {\bm{v}} \in \mathds{R}^d : {\bm{x}} = \mathsf{round}({\bm{v}}) \}$ which ensures that $P({\bm{x}}|{\bm{v}}) = 1$.
\paragraph{Diffusion Models}
Given data ${\bm{x}}_0$, a diffusion model \citep{sohl2015diffusion} consists of predefined variational distributions $q({\bm{x}}_t | {\bm{x}}_{t-1})$ that gradually add noise over time steps $t \in \{1, \ldots, T\}$. The diffusion trajectory is defined such that $q({\bm{x}}_{t} | {\bm{x}}_{t-1})$ adds a small amount of noise around ${\bm{x}}_{t-1}$. This way, information is gradually destroyed such that at the final time step, ${\bm{x}}_T$ carries almost no information about ${\bm{x}}_0$. Their generative counterparts consists of learnable distributions $p({\bm{x}}_{t-1} | {\bm{x}}_{t})$ that learn to denoise the data.
When the diffusion process adds sufficiently small amounts of noise, it suffices to define the denoising trajectory using distributions that are factorized (without correlation) over the dimension axis. The distribution $p({\bm{x}}_T)$ is chosen to be similar to the distribution that the diffusion trajectory approaches. Diffusion models can be optimized using variational inference:
\begin{equation*}
\log P({\bm{x}}_0) \geq \mathbb{E}_{x_1, \ldots x_T \sim q}
\Big{[} \log p({\bm{x}}_T) + \sum_{t=1}^T \log \frac{p({\bm{x}}_{t-1} | {\bm{x}}_t)}{q({\bm{x}}_{t} | {\bm{x}}_{t-1})} \Big{]}.
\end{equation*}
An important insight in diffusion is that by conditioning on ${\bm{x}}_0$, the posterior probability $q({\bm{x}}_{t-1} | {\bm{x}}_{t}, {\bm{x}}_0) = q({\bm{x}}_{t} | {\bm{x}}_{t-1})q({\bm{x}}_{t-1} | {\bm{x}}_0) / q({\bm{x}}_t | {\bm{x}}_0)$ is tractable and straightforward to compute, permitting a reformulation in terms of KL divergences that has lower variance \citep{sohl2015diffusion}. Note that $\mathrm{KL} \big{(} q({\bm{x}}_T | {\bm{x}}_0) | p({\bm{x}}_T)\big{)} \approx 0$ if the diffusion trajectory $q$ is defined well:
\small
\begin{align}
\begin{split}
&\log P({\bm{x}}_0) \geq \mathbb{E}_{q}
\Big{[}\log p({\bm{x}}_0 | {\bm{x}}_1) - \mathrm{KL} \big{(} q({\bm{x}}_T | {\bm{x}}_0) | p({\bm{x}}_T) \big{)} - \sum_{t=2}^T \mathrm{KL} \big{(}q({\bm{x}}_{t-1} | {\bm{x}}_t, {\bm{x}}_0) | p({\bm{x}}_{t-1} | {\bm{x}}_t)\big{)} \Big{]}
\end{split}
\label{eq:diffusion_final_objective}
\end{align}\normalsize
\section{Social Impact and Conclusion}
\label{sec:conclusion}
\paragraph{Social Impact} The methods described in this paper can be used to learn categorical distributions. For that reason, they can potentially be used to generate high-dimensional categorical data, such as text or image segmentation maps, faster than iterative approaches. Possibly negative influences are the generation of fake media in the form of text, or very unhelpful automated chat bots for customer service. Our work could positively influence new methods for text generation, or improved segmentation for self-driving cars. In addition, our work may also be used for outlier detection to flag fake content. Also, we believe the method in its current form is still distant from direct applications as the ones mentioned above.
\paragraph{Conclusion} In this paper we propose two extensions for Normalizing Flows and Diffusion models to learn categorical data: Argmax Flows and Multinomial Diffusion. Our experiments show that our methods outperform comparable models in terms of negative log-likelihood. In addition, our experiments highlight distinct performance gaps in the field: Between standard ARMs, continuous autoregressive models and non-autoregressive continuous models. This indicates that future work could focus on two sources of decreased performance: 1) when discrete variables are lifted to a continuous space and further 2) when removing autoregressive components.
\textbf{Funding Disclosure}\\
There are no additional sources of funding to disclose, beyond the affiliations of the authors.
\section{Related Work}
Deep generative models broadly fall into the categories autoregressive models ARMs \citep{germain2015made}, Variational Autoencoders (VAEs) \citep{kingma2014auto,rezende2014stochasticvariationalinference}, Adversarial Network (GANs) \citep{goodfellow2014generative}, Normalizing Flows \citep{rezende2015norm}, Energy-Based Models (EBMs) and Diffusion Models \citep{sohl2015diffusion}.
Normalizing Flows typically learn a continuous distribution and dequantization is required to train these methods on ordinal data such as images. A large body of work is dedicated to building more expressive continuous normalizing flows \citep{dinh2016density,germain2015made,kingma2016improved,papamakarios2017masked,chen2018neural,song2019mintnet,perugachi2020idensenets}.
To learn ordinal discrete distributions with normalizing flows, adding uniform noise in-between ordinal classes was proposed in \citep{uria2013rnade} and later theoretically justified in \citep{theis2016note}. An extension for more powerful dequantization based on variational inference was proposed in \citep{ho2019flow++}, and connected to autoregressive models in \citep{nielsen2020closingdequantizationgap}. Dequantization for binary variables was proposed in \citep{winkler2019learning}. \citet{discrete2019tran} propose invertible transformations for categorical variables directly. However, these methods can be difficult to train because of gradient bias and results on images have thus far not been demonstrated. In addition flows for ordinal discrete data (integers) have been explored in \citep{hoogeboom2019integer,berg2020idfpp}. In other works, VAEs have been adapted to learn a normalizing flow for the latent space \citep{ziegler2019latent, lippe2020categorical}. However, these approaches typically still utilize an argmax heuristic to sample, even though this is not the distribution specified during training.
Diffusion models were first introduced in \cite{sohl2015diffusion}, who developed diffusion for Gaussian and Bernoulli distributions. Recently, Denoising Diffusion models \cite{ho2020denoising} have been shown capable of generating high-dimensional images by architectural improvements and reparametrization of the predictions. Diffusion models are relatively fast to train, but slow to sample from as they require iterations over the many timesteps in the chain. \citet{song2020denoisingimplicit,nichol2021improved} showed that in practice samples can be generated using significantly fewer steps. \citet{nichol2021improved} demonstrated that importance-weighting the objective components greatly improves log-likelihood performance. In \cite{song2020scorebasedSDEs} a continuous-time extension of denoising diffusion models was proposed. After initial release of this paper we discovered that \citet{song2020denoisingimplicit} concurrently also describe a framework for discrete diffusion, but without empirical evaluation.
\section{Introduction}
\begin{wrapfigure}{r}{.5\textwidth}
\vspace{-.5cm}
\centering
\begin{subfigure}{.499\textwidth}
\includegraphics[width=.99\linewidth]{images/overviews/overview_a.pdf}
\caption{Argmax Flow: Composition of a flow $p({\bm{v}})$ and argmax transformation which gives the model $P({\bm{x}})$. The flow maps from a base distribution $p({\bm{z}})$ using a bijection $g$.}
\label{fig:overview_argmax}
\end{subfigure}
\begin{subfigure}{.499\textwidth}
\includegraphics[width=.99\textwidth]{images/overviews/overview_b.pdf}
\caption{Multinomial Diffusion: Each step $p({\bm{x}}_{t-1} | {\bm{x}}_{t})$ denoises the signal starting from a uniform categorical base distribution which gives the model $p({\bm{x}}_0)$.}
\end{subfigure}
\vspace{-.25cm}
\caption{Overview of generative models.}
\label{fig:overviews}
\vspace{-.5cm}
\end{wrapfigure}
Many sources of high-dimensional data are categorical, for example language and image segmentation. Although natural images have been studied to a large extent with generative flows and diffusion models, categorical data has not had the same extensive treatment. Currently they are primarily modelled by autoregressive models, which are expensive to sample from \citep{Cooijmans2017recurrentbatch, dai2019transformerxl}.
Normalizing flows are attractive because they can be designed to be fast both in the evaluation and sampling direction. Typically, normalizing flows model continuous distributions. As a result, directly optimizing a flow on discrete data may lead to arbitrarily high likelihoods. In literature this problem is resolved for ordinal data by adding noise in a unit interval around the discrete value \citep{uria2013rnade,theis2016note,ho2019flow++}. However, because these methods have been designed for ordinal data, they do not work well on categorical data.
Other attractive generative models are diffusion models \citep{sohl2015diffusion}, which are fast to train due to an objective that decomposes over time steps \citep{ho2020denoising}. Diffusion models typically have a fixed diffusion process that gradually adds noise. This process is complemented by a learnable generative process that denoises the signal. \citet{song2020denoisingimplicit,nichol2021improved} have shown that diffusion models can also be designed for fast sampling. Thus far, diffusion models have been primarily trained to learn ordinal data distributions, such as natural images.
\begin{table*}[t]
\setlength{\tabcolsep}{10pt}
\renewcommand{\arraystretch}{1.0}
\caption{Surjective flow layers for applying continuous flow models to discrete data. The layers are deterministic in the generative direction, but stochastic in the inference direction.
Rounding corresponds to the commonly-used dequantization for ordinal data.}
\centering
\scalebox{.9}{
\begin{tabular}{lccl}
\toprule
\textbf{Layer} & \textbf{Generation} & \textbf{Inference} & \textbf{Applications} \\ \midrule
\multirow{2}{*}{Rounding} & \multirow{2}{*}{${\bm{x}} = \lfloor {\bm{v}} \rfloor$} & ${\bm{v}} \sim q({\bm{v}}|{\bm{x}})$ with support & Ordinal Data \\
& & $\mathcal{S}({\bm{x}}) = \{{\bm{v}} | {\bm{x}} = \lfloor {\bm{v}} \rfloor \}$ & e.g. images, audio \\
\hline
\multirow{2}{*}{Argmax} & \multirow{2}{*}{${\bm{x}} = \argmax {\bm{v}}$} & ${\bm{v}} \sim q({\bm{v}}|{\bm{x}})$ with support & Categorical Data \\
& & $\mathcal{S}({\bm{x}}) = \{{\bm{v}} | {\bm{x}} = \argmax {\bm{v}} \}$ & e.g. text, segmentation \\
\bottomrule
\end{tabular}}
\label{tab:surjective}
\end{table*}
Therefore, in this paper we introduce extensions of flows and diffusion models for categorical variables (depicted in Figure~\ref{fig:overviews}): \textit{i)} Argmax Flows bridge the gap between categorical data and continuous normalizing flows using an argmax transformation and a corresponding family of probabilistic inverses for the argmax. In addition \textit{ii)} we introduce Multinomial Diffusion, which is a diffusion model directly defined on categorical variables. Opposed to normalizing flows, defining diffusion for discrete variables directly does not require gradient approximations, because the diffusion trajectory is fixed. As a result of our work, generative normalizing flows and diffusion models can directly learn categorical data.
\section*{Checklist}
\begin{enumerate}
\item For all authors...
\begin{enumerate}
\item Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
\answerYes{}
\item Did you describe the limitations of your work?
\answerYes{Section~\ref{sec:experiments}}
\item Did you discuss any potential negative societal impacts of your work?
\answerYes{Section~\ref{sec:conclusion}}
\item Have you read the ethics review guidelines and ensured that your paper conforms to them?
\answerYes{}
\end{enumerate}
\item If you are including theoretical results...
\begin{enumerate}
\item Did you state the full set of assumptions of all theoretical results?
\answerYes{}
\item Did you include complete proofs of all theoretical results?
\answerYes{}
\end{enumerate}
\item If you ran experiments...
\begin{enumerate}
\item Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?
\answerYes{}
\item Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
\answerYes{ in Appendix~\ref{app:experimental_details}.}
\item Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
\answerYes{ for a selection of runs in the interest of computational budget, in Appendix~\ref{app:experimental_details}.}
\item Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
\answerYes{in Appendix~\ref{app:experimental_details}.}
\end{enumerate}
\item If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
\begin{enumerate}
\item If your work uses existing assets, did you cite the creators?
\answerYes{}
\item Did you mention the license of the assets?
\answerYes{if available.}
\item Did you include any new assets either in the supplemental material or as a URL?
\answerYes{code.}
\item Did you discuss whether and how consent was obtained from people whose data you're using/curating?
\answerYes{in Appendix~\ref{app:experimental_details}}
\item Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
\answerYes{in Appendix~\ref{app:experimental_details}.}
\end{enumerate}
\item If you used crowdsourcing or conducted research with human subjects...
\begin{enumerate}
\item Did you include the full text of instructions given to participants and screenshots, if applicable?
\answerNA{}
\item Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
\answerNA{}
\item Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
\answerNA{}
\end{enumerate}
\end{enumerate}
\section{Experiments}
\label{sec:experiments}
In our experiments we compare the performance of our methods on language modelling tasks and learning image segmentation maps unconditionally.
\begin{table}
\centering
\caption{Comparison of a coupling and autoregressive generative flows with uniform \citep{uria2013rnade} and variational \citep{ho2019flow++} dequantization and our proposed Argmax flows.}
\label{tab:performance_generative_flow_text}
\scalebox{.9}{
\begin{tabular}{l l c c c c}
\toprule
Dequantization & Flow type & text8 (bpc) & enwik8 (bits per raw byte) \\ \midrule
Uniform dequantization & \multirow{3}{*}{Autoregressive} & 1.90 & 2.14 \\
Variational dequantization & & 1.43 & 1.44 \\
Argmax Flow (ours) & & \textbf{1.38} & \textbf{1.42} \\ \midrule
Uniform dequantization & \multirow{3}{*}{Coupling} & 2.01 & 2.33 \\
Variational dequantization & & 2.08 & 2.28 \\
Argmax Flow (ours) & & \textbf{1.82} & \textbf{1.93} \\ \midrule
\end{tabular}}
\vspace{-.3cm}
\end{table}
\begin{table}[b]
\centering
\caption{Comparison of different methods on \texttt{text8} and \texttt{enwik8}. Results are reported in negative log-likelihood with units bits per character (bpc) for \texttt{text8} and bits per raw byte (bpb) for \texttt{enwik8}.}
\label{tab:performance_text}
\scalebox{.9}{
\begin{tabular}{l l l c c c}
\toprule
Model type & & Model & text8 (bpc) & enwik8 (bpb) \\ \midrule
\multirow{2}{*}{ARM}
& & 64 Layer Transformer {\footnotesize \citep{alrfou2019transformer}} & 1.13 & 1.06 \\
& & TransformerXL {\footnotesize \citep{dai2019transformerxl}} & 1.08 & 0.99 \\ \midrule \midrule
\multirow{3}{*}{\small VAE} & & AF/AF$^\star$ (AR) {\footnotesize \citep{ziegler2019latent}} & 1.62 & 1.72\\
& & IAF / SCF$^\star$ {\footnotesize \citep{ziegler2019latent}} & 1.88 & 2.03 \\
& & CategoricalNF (AR) {\footnotesize \citep{lippe2020categorical}} & 1.45 & - \\ \midrule
\multirow{2}{*}{Generative Flow}
& & Argmax Flow, AR (ours) & 1.39 & 1.42 \\
& & Argmax Coupling Flow (ours) & 1.82 & 1.93 \\
\midrule
Diffusion & & Multinomial Text Diffusion (ours) & 1.72 & 1.75 \\ \bottomrule
\end{tabular}}
\begin{flushleft}
\scriptsize{$\star$ Results obtained by running code from the official repository for the \texttt{text8} and \texttt{enwik8} datasets.} \\
\vspace{-.5cm}
\end{flushleft}
\end{table}
\subsection{Language data}
In this section we compare our methods on two language datasets, \texttt{text8} and \texttt{enwik8}. \texttt{text8} contains 27 categories (`a' through `z' and ` ') and for \texttt{enwik8} the bytes are directly modelled which results in 256 categories.
\paragraph{Model description}
Two versions of generative argmax flows are tested: using an autoregressive (AR) flow and a coupling-based flow for $p({\bm{v}})$. In these experiments the probabilistic inverse is based on the thresholding approach. Specifically, a conditional diagonal Gaussian $q({\bm{u}} | {\bm{x}})$ is trained and thresholded which gives the distribution $q({\bm{v}} | {\bm{x}})$.
The argmax flow is defined on binary Cartesian products. This means that for $K=27$, a $5$-dimensional binary space is used and for $K=256$ an $8$-dimensional binary space. The argmax flow is compared to the current standard of training generative flows directly on discrete data: dequantization. We compare to both uniform and variational dequantization, where noise on a (0, 1) interval is added to the onehot representation of the categorical data. The autoregressive density model is based on the model proposed in \citep{lippe2020categorical}. The coupling density model consists of 8 flow layers where each layer consists of a $1$ $\times$ $1$ convolution and mixture of logistics transformations \cite{ho2019flow++}. In the multinomial text diffusion model, the $\mu$ network is modeled by a 12-layer Transformer. For more extensive details about the experiment setup see Appendix \ref{app:experimental_details}.
\paragraph{Comparison with Generative Flows}
Firstly we compare the performance of generative flows directly trained on language data (Table~\ref{tab:performance_generative_flow_text}). These experiments are using the same underlying normalizing flow: either a coupling-based flow or an autoregressive flow. Note that Argmax Flows consistently outperform both uniform and variational dequantization. This indicates that it is easier for a generative flow to learn the lifted continuous distribution using an argmax flow. An advantage of Argmax flows that may explain this difference is that they lift the variables into the entire Euclidean space, whereas traditional dequantization only introduce probability density on $(0, 1)$ intervals, leaving gaps with no probability density. The performance improvements of Argmax flows are even more pronounced when comparing coupling-based approaches. Also note that coupling flows have worse performance than autoregressive flows, with a difference that is generally smaller for images. This indicates that designing more expressive coupling layers for text is an interesting future research direction.
\paragraph{Comparison with other generative models} The performance compared to models in literature is presented in Table~\ref{tab:performance_text} alongside the performance of our Argmax Flows and Multinomial Diffusion. The latent variable approaches containing autoregressive components are marked using (AR). Although autoregressive flows still have the same disadvantages as ARMs, they provide perspective on where performance deficiencies are coming from. We find that our autoregressive Argmax Flows achieve better performance than the VAE approaches, they outperform AF/AF \citep{ziegler2019latent} and CategoricalNF \citep{lippe2020categorical}.
\begin{figure}
\begin{minipage}[t]{.48\textwidth}
\begin{figure}[H]
\centering
\begin{subfigure}{.99\textwidth}
\includegraphics[width=\textwidth]{images/samples/text_sample_mult.pdf}
\caption{Samples from Multinomial Text Diffusion.}
\end{subfigure}
\begin{subfigure}{.99\textwidth}
\includegraphics[width=\textwidth]{images/samples/text_sample_AR.pdf}
\caption{Samples from Argmax AR Flow.}
\end{subfigure}
\begin{subfigure}{.99\textwidth}
\includegraphics[width=\textwidth]{images/samples/text_sample_coupling.pdf}
\caption{Samples from Argmax Coupling Flow.}
\end{subfigure}
\caption{Samples from models, \texttt{text8}.}
\label{fig:samples_text}
\end{figure}
\end{minipage}
\hfill
\begin{minipage}[t]{.48\textwidth}
\begin{figure}[H]
\centering
\begin{subfigure}{.99\textwidth}
\includegraphics[width=\textwidth]{images/small/samples_small_city.png}
\caption{Samples from the Argmax Flow.}
\end{subfigure}
\begin{subfigure}{.99\textwidth}
\includegraphics[width=\textwidth]{images/cityscapes/sample_ep2000_s0.png}
\caption{Samples from the Multinomial Diffusion model.}
\end{subfigure}
\begin{subfigure}{.99\textwidth}
\includegraphics[width=\textwidth]{images/small/cityscapes_data_small.png}
\caption{Cityscapes data.}
\end{subfigure}
\caption{Samples from models, cityscapes.}
\label{fig:experiments_cityscapes}
\vspace{-.5cm}
\end{figure}
\end{minipage}
\end{figure}
When comparing non-autoregressive models, Argmax Flows also outperforms the method that lifts the categorical space to a continuous space: IAF / SCF \citep{ziegler2019latent}. Interestingly, the multinomial text diffusion is a non-autoregressive model that performs even better than the argmax coupling flow, but performs worse than the autoregressive version. For this model it is possible that different diffusion trajectories for $q$ would result in even better performance, because in the current form the denoising model has to be very robust to input noise. These experiments also highlight that there is still a distinct performance gap between standard ARMs and (autoregressive) continuous density model on text, possibly related to the dequantization gap \citep{nielsen2020closingdequantizationgap}. Samples from different models trained on \texttt{text8} are depicted in Figure~\ref{fig:samples_text}. Because of difficulties in reproducing results from Discrete Flows, a comparison and analysis of discrete flows are left out of this section. Instead they are extensively discussed in Appendix~\ref{sec:reproducing_discrete_flows}. For additional experiments regarding Cartesian products and sampling time see Appendix~\ref{sec:additional_experiments}.
\begin{wrapfigure}{r}{.5\textwidth}
\begin{minipage}[t]{.485\textwidth}
\vspace{-.2cm}
\begin{figure}[H]
\centering
\begin{subfigure}{.99\textwidth}
\includegraphics[width=\textwidth]{images/samples/spellcheck_groundtruth.pdf}
\caption{Ground truth sequence from \texttt{text8}.}
\end{subfigure}
\begin{subfigure}{.99\textwidth}
\includegraphics[width=\textwidth]{images/samples/spellcheck_corrupted.pdf}
\caption{Corrupted sentence.}
\end{subfigure}
\begin{subfigure}{.99\textwidth}
\includegraphics[width=\textwidth]{images/samples/spellcheck_suggestion.pdf}
\caption{Suggested, prediction by the model.}
\end{subfigure}
\caption{Spell checking with Multinomial Text Diffusion.}
\label{fig:spellcheck}
\end{figure}
\end{minipage}
\vspace{-.3cm}
\end{wrapfigure}
\paragraph{Unsupervised spell-checking}
An interesting by-product of the text diffusion model is that it can be used to spell-check text using a single forward pass. To demonstrate this, a sentence taken from the test data is corrupted by changing a few characters. This corrupted sequence is given as ${\bm{x}}_1$ to the generative denoising model, which is close to the data at step $0$. Then the denoising model predicts $p({\bm{x}}_0 | {\bm{x}}_1)$ and the most-likely ${\bm{x}}_0$ can be suggested. Note that this model only works for character-level corruption, not insertions. An example is depicted in Figure~\ref{fig:spellcheck}. Since the model chooses the most-likely matching word, larger corruptions will at some point lead to word changes.
\subsection{Segmentation maps}
For image-type data, we introduce a categorical image dataset: the cityscapes dataset is repurposed for \textit{unconditional} image segmentation learning. In contrast with the standard setting, the distribution over the segmentation targets needs to be learned \textit{without} conditioning on the photograph. To reduce computational cost, we rescale the segmentation maps from cityscapes to $32 \times 64$ images using nearest neighbour interpolation. We utilize the global categories as prediction targets which results in an 8-class problem.
\begin{wrapfigure}{r}{.5\textwidth}
\begin{minipage}[t]{.49\textwidth}
\vspace{-1.1cm}
\begin{table}[H]
\centering
\caption{Performance of different dequantization methods on squares and cityscapes dataset, in bits per pixel, lower is better.}
\label{tab:results_cityscapes}
\scalebox{.89}{
\begin{tabular}{l l r r r r}
\toprule
Cityscapes & ELBO & IWBO \\ \midrule
Round / Unif. \citep{uria2013rnade} & 1.010 & 0.930 \\
Round / Var. \citep{ho2019flow++} & 0.334 & 0.315 \\ \midrule
Argmax / Softplus thres. (ours) & \textbf{0.303} & \textbf{0.290} \\
Argmax / Gumbel dist. (ours) & 0.365 & 0.341 \\
Argmax / Gumbel thres. (ours) & \textbf{0.307} & \textbf{0.287} \\ \midrule
Multinomial Diffusion (ours) & \multicolumn{2}{c}{0.305} \\ \bottomrule
\end{tabular}}
\vspace{-.1cm}
\end{table}
\end{minipage}
\vspace{-.3cm}
\end{wrapfigure}
\vspace{-.15cm}
\paragraph{Model description}
The Argmax Flows are defined directly on the $K=8$ categorical space. The density model $p({\bm{v}})$ is defined using affine coupling layers parametrized by DenseNets \citep{huang2017densely}. For the probabilistic inverse we learn a conditional flow $q({\bm{u}} | {\bm{x}})$ which is also based on the affine coupling structure. Depending on the method, either softplus or Gumbel thresholding is applied to obtain ${\bm{v}}$. Recall that for our first Gumbel approach it is equivalent to set $q({\bm{u}} | {\bm{x}})$ to the unit uniform distribution, whereas $q({\bm{u}} | {\bm{x}})$ is learned for Gumbel thresholding. We compare to existing dequantization strategies in literature: uniform \citep{uria2013rnade} and variational dequantization \citep{ho2019flow++} which are applied on the onehot representation. All models utilize the same underlying flow architectures and thus the number of parameters is roughly the same. The exception are uniform dequantization and the Gumbel distribution, since no additional variational flow distribution is needed. For more extensive details see Appendix \ref{app:experimental_details}.
\vspace{-.15cm}
\paragraph{Comparison}
The results of this experiment are shown in Table~\ref{tab:results_cityscapes} in terms of ELBO and if available the IWBO (importance weighted bound) \citep{burda2016} with $1000$ samples measured in bits per pixel. Consistent with the language experiments, the traditional dequantization approaches (uniform / variational) are outperformed by Argmax Flows. Interestingly, although argmax flows with softplus thresholding achieves the best ELBO, the argmax flow with Gumbel thresholding approach achieves a better IWBO. The Multinomial Diffusion model performs somewhat worse with 0.37 bpp on test whereas it scored 0.33 bpp on train. Interestingly, this the only model where overfitting was an issue and data augmentation was required, which may explain this portion of the performance difference. For all other models training performance was comparable to test and validation performance. Samples from the different models trained on cityscapes are depicted in Figure~\ref{fig:experiments_cityscapes}. Another interesting point is that coupling flows had difficulty producing coherent text samples (Figure~\ref{fig:samples_text}) but do not suffer from this problem on the cityscapes data which is more image-like. As coupling layers where initially designed for images \citep{dinh2014nice}, they may require adjustments to increase their expressiveness on text.
\section{Multinomial Diffusion}
In this section we introduce an alternative likelihood-based model for categorical data: Multinomial Diffusion. In contrast with previous sections, ${\bm{x}}_t$ will be represented in one-hot encoded format ${\bm{x}}_t \in \{0, 1\}^K$. Specifically, for category $k$, $x_{k} = 1$ and $x_{j} = 0$ for $j \not= k$. Note that again the dimension axis is omitted for clarity as all distributions are independent over the dimension axis. We define the multinomial diffusion process using a categorical distribution that has a $\beta_t$ chance of resampling a category uniformly:
\begin{equation}
q({\bm{x}}_t | {\bm{x}}_{t-1}) = \mathcal{C}({\bm{x}}_t | (1 - \beta_t) {\bm{x}}_{t-1} + \beta_t / K ),
\label{eq:diff_categorical_forward}
\end{equation}
where $\mathcal{C}$ denotes a categorical distribution with probability parameters after $|$. Further addition (and subtraction) between scalars and vectors is done elementwise. This convention kept throughout this section.
Since these distributions form a Markov chain, we can express the probability of any ${\bm{x}}_t$ given ${\bm{x}}_0$ as:
\begin{equation}
q({\bm{x}}_t | {\bm{x}}_{0}) = \mathcal{C}({\bm{x}}_t | \bar{\alpha}_t {\bm{x}}_{0} + (1 - \bar{\alpha}_t) / K )
\label{eq:diff_categorical_forward_x0}
\end{equation}
where $\alpha_t = 1 - \beta_t$ and $\bar{\alpha}_t = \prod_{\tau=1}^t \alpha_\tau$. Intu\"{i}tively, for each next timestep, a little amount of uniform noise $\beta_t$ over the $K$ classes is introduced, and with a large probability $(1 - \beta_t)$ the previous value ${\bm{x}}_{t-1}$ is sampled. Using Equation \ref{eq:diff_categorical_forward} and \ref{eq:diff_categorical_forward_x0} the categorical posterior $q({\bm{x}}_{t-1} | {\bm{x}}_{t}, {\bm{x}}_0)$ can be computed in closed-form:
\begin{align}
\begin{split}
q({\bm{x}}_{t-1} | {\bm{x}}_{t}, {\bm{x}}_0) &= \mathcal{C}({\bm{x}}_{t-1} | {\bm{\theta}}_{\mathrm{post}}({\bm{x}}_t, {\bm{x}}_0)), \,\, \text{ where } \,\, {\bm{\theta}}_{\mathrm{post}}({\bm{x}}_t, {\bm{x}}_0) = \tilde{{\bm{\theta}}} / \sum_{k=1}^K \tilde{\theta}_k
\\
\text{and } \,\tilde{{\bm{\theta}}} &= [\alpha_t {\bm{x}}_t + (1 - \alpha_t) / K] \odot [\bar{\alpha}_{t-1} \boldsymbol{x}_0 + (1 - \bar{\alpha}_{t-1}) / K ].
\label{eq:q_posterior}
\end{split}
\end{align}
\begin{figure}
\centering
\includegraphics[width=.99\linewidth]{images/overviews/overview_multdiff.pdf}
\caption{Overview of multinomial diffusion. A generative model $p({\bm{x}}_{t-1} | {\bm{x}}_{t})$ learns to gradually denoise a signal from left to right. An inference diffusion process $q({\bm{x}}_{t} | {\bm{x}}_{t-1})$ gradually adds noise form right to left.}
\label{fig:overview_mult_diffusion}
\end{figure}
One of the innovations in \cite{ho2020denoising} was the insight to not predict the parameters for the generative trajectory directly, but rather to predict the noise using the posterior equation for $q$. Although predicting the noise is difficult for discrete data, we predict a probability vector for $\hat{{\bm{x}}}_0$ from ${\bm{x}}_t$ and subsequently parametrize $p({\bm{x}}_{t-1} | {\bm{x}}_{t})$ using the probability vector from $q({\bm{x}}_{t-1} | {\bm{x}}_t, \hat{{\bm{x}}}_0)$, where ${\bm{x}}_0$ is approximated using a neural network $\hat{{\bm{x}}}_0 = \mu({\bm{x}}_t, t)$. Equation~\ref{eq:q_posterior} will produce valid probability vectors that are non-negative and sums to one under the condition that the prediction $\hat{{\bm{x}}}_0$ is non-negative and sums to one, which is ensured with a softmax function in $\mu$. To summarize:
\begin{align}
\begin{split}
p({\bm{x}}_{0} | {\bm{x}}_{1}) = \mathcal{C}({\bm{x}}_{0} | \hat{{\bm{x}}}_0) \, \text{ and } \,
p({\bm{x}}_{t-1} | {\bm{x}}_{t}) = \mathcal{C}({\bm{x}}_{t-1} | \boldsymbol{\theta}_{\mathrm{post}}({\bm{x}}_t,
\hat{{\bm{x}}}_0)) \,
\text{ where }\, \hat{{\bm{x}}}_0 = \mu({\bm{x}}_t, t)
\end{split}
\label{eq:p_parametrization}
\end{align}
The KL terms in Equation \ref{eq:diffusion_final_objective} can be simply computed by enumerating the probabilities in Equation \ref{eq:q_posterior} and \ref{eq:p_parametrization} and computing the KL divergence for discrete distributions in $L_{t-1}$ with $t \geq 2$:
\small\begin{align}
\begin{split}
\mathrm{KL} \big{(}q({\bm{x}}_{t-1} | {\bm{x}}_t, {\bm{x}}_0) | p({\bm{x}}_{t-1} | {\bm{x}}_t)\big{)} &= \mathrm{KL} \big{(} \mathcal{C}(\boldsymbol{\theta}_{\mathrm{post}}({\bm{x}}_t, {\bm{x}}_0)) | \mathcal{C}(\boldsymbol{\theta}_{\mathrm{post}}({\bm{x}}_t,
\hat{{\bm{x}}}_0))\big{)},
\end{split}
\end{align}\normalsize
which can be computed using $\sum_k \boldsymbol{\theta}_{\mathrm{post}}({\bm{x}}_t, {\bm{x}}_0))_k \cdot \log \frac{\boldsymbol{\theta}_{\mathrm{post}}({\bm{x}}_t, {\bm{x}}_0))_k}{\boldsymbol{\theta}_{\mathrm{post}}({\bm{x}}_t,
\hat{{\bm{x}}}_0))_k}$. Furtermore, to compute $\log p({\bm{x}}_0 | {\bm{x}}_1)$ use that ${\bm{x}}_0$ is onehot:
\small\begin{equation}
\log p({\bm{x}}_0 | {\bm{x}}_1) = \sum_k {\bm{x}}_{0,k} \log \hat{{\bm{x}}}_{0,k}
\end{equation}\normalsize
\subsection{Probabilistic Inverse}
The argmax layer may be viewed as a surjective flow layer \citep{nielsen2007survae}. With this view, the variational distribution $q({\bm{v}}|{\bm{x}})$ specifies a distribution over the possible right-inverses of the argmax function, also known as a \emph{stochastic inverse} or \emph{probabilistic inverse}.
Recall that the commonly-used dequantization layer for ordinal data corresponds to the probabilistic inverse of a rounding operation. As summarized in Table \ref{tab:surjective}, this layer may thus be viewed as analogous to the argmax layer, where the round is for ordinal data while the argmax is for categorical data.
We are free to specify any variational distribution $q({\bm{v}}|{\bm{x}})$ that satisfies the argmax constraint. In the next paragraphs we outline three possible approaches. Since operations are performed independently across dimensions, we omit the dimension axis and let ${\bm{v}} \in \mathbb{R}^K$ and $x \in \{1, \ldots, K\}$.
\paragraph{Thresholding (Alg.~\ref{alg:thresholding}).}
A straightforward method to construct a distribution $q({\bm{v}}|x)$ satisfying the argmax constraint is to use thresholding. That is, we first sample an unbounded variable ${\bm{u}} \in \mathbb{R}^{K}$ from $q({\bm{u}} | x)$, which can be for example a conditional Gaussian or normalizing flow. Next, we map ${\bm{u}}$ to ${\bm{v}}$ such that element $x$ is the largest:
\begin{equation}
v_x = u_x \quad \text{and} \quad {\bm{v}}_{-x} = \mathrm{threshold}_T({\bm{u}}_{-x})
\end{equation}
where the thresholding is applied elementwise with threshold value $T=v_x$. This ensures that element $v_x$ is the largest, and consequently that $q({\bm{v}}|x)$ satisfies the argmax constraint.
Note that we require the threshold function to be bijective, $\mathrm{threshold}_T: \mathbb{R} \rightarrow (-\infty,T)$, so that we can use the change-of-variables formula to compute $\log q({\bm{v}}|x)$.
In our implementation, thresholding is implemented using a softplus such that all values are mapped below a limit $T$:
\begin{equation}
\small
v = \mathrm{threshold}_T(u) = T-\mathrm{softplus}(T-u),
\end{equation}
where $\mathrm{softplus}(z) = \log(1 + e^z)$ and for which it is guaranteed that $v \in (-\infty, T)$.
\begin{table}[]
\begin{minipage}[t]{.485\textwidth}
\begin{algorithm}[H]
\caption{Thresholding-based $q({\bm{v}}|{\bm{x}})$}
\label{alg:thresholding}
\begin{algorithmic}
\STATE {\bfseries Input:} ${\bm{x}}$, $q({\bm{u}}|{\bm{x}})$
\STATE {\bfseries Output:} ${\bm{v}}$, $\log q({\bm{v}}|{\bm{x}})$
\STATE ${\bm{u}} \sim q({\bm{u}}|{\bm{x}})$
\STATE ${\bm{v}}_{{\bm{x}}} = {\bm{u}}_{{\bm{x}}}$
\STATE ${\bm{v}}_{-{\bm{x}}} = \mathrm{threshold}({\bm{u}}_{-{\bm{x}}}, {\bm{x}})$
\STATE $\log q({\bm{v}}|{\bm{x}}) = \log q({\bm{u}}|{\bm{x}}) - \log |\det \rmd{\bm{v}} / \rmd{\bm{u}}|$
\end{algorithmic}
\end{algorithm}
\end{minipage}
\hfill
\begin{minipage}[t]{.485\textwidth}
\begin{algorithm}[H]
\caption{Gumbel-based $q({\bm{v}}|{\bm{x}})$}
\label{alg:gumbel}
\begin{algorithmic}
\STATE {\bfseries Input:} ${\bm{x}}$, $\boldsymbol{\phi}$
\STATE {\bfseries Output:} ${\bm{v}}$, $\log q({\bm{v}}|{\bm{x}})$
\STATE $\phi_{\max} = \log \sum_i \exp \phi_i$
\STATE ${\bm{v}}_{{\bm{x}}} \sim \mathrm{Gumbel}(\phi_{\max})$
\STATE ${\bm{v}}_{-{\bm{x}}} \sim \mathrm{TruncGumbel}(\boldsymbol{\phi}_{-{\bm{x}}}, {\bm{v}}_{{\bm{x}}})$
\STATE $\log q({\bm{v}}|{\bm{x}}) = \log \mathrm{Gumbel}({\bm{v}}_{\bm{x}} | \phi_{\max})$
\STATE \hspace{1.0cm} $+ \log \mathrm{TruncGumbel}({\bm{v}}_{-{\bm{x}}} | \boldsymbol{\phi}_{-{\bm{x}}}, {\bm{v}}_{\bm{x}})$
\end{algorithmic}
\end{algorithm}
\end{minipage}
\end{table}
\paragraph{Gumbel (Alg.~\ref{alg:gumbel}).}
An alternative approach is to let $q({\bm{v}}|x) = \mathrm{Gumbel}({\bm{v}}|\boldsymbol{\phi})$ \textit{restricted} to $\argmax {\bm{v}} = x$, where the location parameters $\boldsymbol{\phi} \leftarrow \mathrm{NN}(x)$ are predicted using a neural network $\mathrm{NN}$.
The Gumbel distribution has favourable properties: The $\argmax$ and $\max$ are independent and the $\max$ is also distributed as a Gumbel:
\begin{equation}
\small
\label{eq:gumbel_max}
\max_i v_i \sim \mathrm{Gumbel}(\phi_{\max}),
\end{equation}
where $\phi_{\max} = \log \sum_i \exp \phi_i$.
For a more extensive introduction see \citep{Maddison2014Astarsampling,kool2019}. To sample ${\bm{v}} \sim q({\bm{v}}|x)$, we thus first sample the maximum $v_x$ according to Eq.~\ref{eq:gumbel_max}. Next, given the sample $v_x$, the remaining values can be sampled using \textit{truncated} Gumbel distributions:
\begin{equation}
\small
v_{i} \sim \mathrm{TruncGumbel}(\phi_i; T) \text{ where } i \not= x
\label{eq:gumbel_remaining}
\end{equation}
where the truncation value $T$ is given by $v_x$ which ensures that the argmax constraint $v_x > v_i$ for $i\not=x$ is satisfied. Recall that to optimize Eq.~\ref{eq:elbo_objective}, $\log q({\bm{v}} | {\bm{x}})$ is also required, which can be computed using the closed-form expressions for the log density functions (see Table \ref{tab:gumbel}). Another property of Gumbel distributions is that
\begin{equation}
\small
P(\argmax {\bm{v}} = i) = \exp \phi_i / \sum_i \exp \phi_i,
\end{equation}
which we use to initialize the location parameters $\boldsymbol{\phi}$ to match the empirical distribution of the first minibatch of the data.
\paragraph{Gumbel Thresholding.} This method unifies the methods from the previous two sections: Gumbel distributions and thresholding. The key insight is that the Gumbel sampling procedures as defined above can be seen as a reparametrization of a uniform noise distribution $\mathcal{U}(0, 1)^K$ which is put through the inverse CDF of the Gumbel distributions (see Table \ref{tab:gumbel}). From the perspective of change-of-variables, the log likelihood denotes the log volume change of this transformation. To increase expressitivity the uniform distribution can be replaced by a normalizing flow $q({\bm{u}} | x)$ that has support on the interval $(0, 1)^K$, which can be enforced using a sigmoid transformation. This section shows that a large collection of thresholding functions can be found by studying (truncated) inverse CDFs. In practice we find that performance is reasonably similar as long as the underlying noise ${\bm{u}}$ is learned.
\paragraph{Behavior of the Variational Posterior}
Although several methods to learn $q$ have been proposed, it is unclear what expressitivity is required. In the following, the interactions between $q({\bm{v}} | {\bm{x}})$ and the density model $p({\bm{v}})$ are discussed. Recall that the variational bound that is optimized under expectation of a data distribution $\mathcal{D}$ can be seen as minimizing the $\operatorname{KL}$ distance between the aggregated posterior $q({\bm{v}}) = \mathbb{E}_{{\bm{x}} \sim \mathcal{D}}q({\bm{v}} | {\bm{x}})$ and the density model $p({\bm{v}})$, so $\operatorname{KL}(q({\bm{v}})|p({\bm{v}}))$. There are two distinct reasons which can cause this distance to be large: Firstly, the density model $p({\bm{v}})$ may not have the right probability mass in each argmax region. These desired probabilities solely depend on the data distribution $\mathcal{D}$. Secondly, the variational posterior $q({\bm{v}} | {\bm{x}})$ may not have the correct shape compared to $p({\bm{v}})$, \textit{within} an argmax region. At initialization, the thresholding within $q$ can create low density regions at argmax boundaries.
In theory, if $p({\bm{v}})$ is a universal density approximator, then the model can be fitted for any well-behaved $q({\bm{v}}|{\bm{x}})$. Then $p({\bm{v}})$ can even fit the low density regions in the boundaries. This argument is trivial, as one can simply set $p({\bm{v}})$ to $q({\bm{v}}) = \mathbb{E}_{{\bm{x}} \sim \mathcal{D}}q({\bm{v}} | {\bm{x}})$. In practice, over training steps we find that $q$ does smooth out these boundary artifacts, and counteracts the thresholding so that the aggregated posterior becomes smoother.
\subsection{Cartesian Products of Argmax Flows}
In the current description, Argmax Flows require the same number of dimensions in ${\bm{v}}$ as there are classes in ${\bm{x}}$. To alleviate this constraint we introduce Cartesian products of Argmax Flows. To illustrate our method, consider a 256 class problem. One class can be represented using a single number in $\{1, \ldots, 256\}$, but also using two hexadecimal numbers $\{1, \ldots, 16\}^2$ or alternatively using eight binary numbers. Specifically, any base $K$ variable ${\bm{x}}^{(K)} \in \{1, \ldots, K\}^D$ can be converted to a base $M$ variable ${\bm{x}}^{(M)} \in \{1, \ldots, M\}^{d_m \times D}$ where $d_m = \lceil \log_M K \rceil$.
Then the variable ${\bm{x}}^{(M)}$ with dimensionality $M \cdot d_m \cdot D$ represents the variable ${\bm{x}}^{(K)}$ with dimensionality $K \cdot D$, trading off symmetry for dimensionality. Even though this may lead to some unused additional classes, the ELBO objective in Equation \ref{eq:elbo_objective} can still be optimized using an $M$-categorical Argmax Flow. Finally, note that Cartesian products of binary spaces are a special case where the variable can be encoded symmetrically into a single dimension to the positive and negative part using binary dequantization \citep{winkler2019learning}. In this case, by trading-off symmetry the dimensionality increases only proportional to $\log_2 K$ .
\section{Argmax Flows}
Argmax flows define discrete distributions using 1) a density model $p({\bm{v}})$, such as a normalizing flow, and 2) an argmax layer that maps the continuous ${\bm{v}} \in \mathbb{R}^{D \times K}$ to a discrete ${\bm{x}} \in \{1,2,...,K\}^D$ using
\begin{equation}
\small
{\bm{x}} = \argmax {\bm{v}} \quad \text{ where } \quad x_d = \argmax_k v_{dk}.
\end{equation}
This is a natural choice to model categorical variables, because it divides the entire continuous space of ${\bm{v}}$ into symmetric partitions corresponding to categories in ${\bm{x}}$. To sample from an argmax flow sample ${\bm{v}} \sim p({\bm{v}})$ and compute ${\bm{x}} = \argmax {\bm{v}}$ (Algorithm~\ref{alg:sample_argmax}). To generate reasonable samples, it is up to the density model $p({\bm{v}})$ to capture any complicated dependencies between the different dimensions. While sampling from an argmax flow is straightforward, the main difficulty lies in \textit{optimizing} this generative model. To compute the likelihood of a datapoint ${\bm{x}}$, we have to compute
\begin{equation}
\small
P({\bm{x}}) =\! \int P({\bm{x}} | {\bm{v}}) p({\bm{v}})d{\bm{v}}, ~~P({\bm{x}}|{\bm{v}}) \!=\! \delta\big({\bm{x}} \!=\! \argmax({\bm{v}})\big),
\end{equation}
which is intractable. Consequently, we resort to variational inference and specify a variational distribution $q({\bm{v}}|{\bm{x}})$. We note that na\"{i}vely choosing any variational distribution may lead to samples ${\bm{v}} \sim q({\bm{v}} | {\bm{x}})$ where $\delta({\bm{x}} = \argmax {\bm{v}}) = 0$, which yields an ELBO of negative infinity. To avoid this, we need a variational distribution $q({\bm{v}}|{\bm{x}})$ that satisfies what we term the \emph{argmax constraint}:
\begin{equation*}
\small
{\bm{x}} = \argmax {\bm{v}} \quad \text{for all} \quad {\bm{v}} \sim q({\bm{v}}|{\bm{x}}).
\end{equation*}
That is, the variational distribution $q({\bm{v}}|{\bm{x}})$ should have support limited to $\mathcal{S}({\bm{x}}) = \{{\bm{v}} \in \mathbb{R}^{D\times K} : {\bm{x}} = \argmax {\bm{v}}\}$. Recall that under this condition, the ELBO simplifies to $\mathbb{E}_{{\bm{v}} \sim q({\bm{v}}|{\bm{x}})} \left[ \log p({\bm{v}}) - \log q({\bm{v}}|{\bm{x}}) \right]$, as shown in Algorithm~\ref{alg:optimize_argmax}. For an illustration of the method see Figure \ref{fig:overview_argmax}.
\begin{table}
\begin{minipage}[t]{.47\textwidth}
\begin{algorithm}[H]
\caption{Sampling from Argmax Flows}
\label{alg:sample_argmax}
\begin{algorithmic}
\STATE {\bfseries Input:} $p({\bm{v}})$
\STATE {\bfseries Output:} Sample ${\bm{x}}$
\STATE Sample ${\bm{v}} \sim p({\bm{v}})$
\STATE Compute ${\bm{x}} = \argmax {\bm{v}}$
\end{algorithmic}
\end{algorithm}
\end{minipage}
\hfill
\begin{minipage}[t]{.47\textwidth}
\begin{algorithm}[H]
\caption{Optimizing Argmax Flows}
\label{alg:optimize_argmax}
\begin{algorithmic}
\STATE {\bfseries Input:} ${\bm{x}}$, $p({\bm{v}})$, $q({\bm{v}}|{\bm{x}})$
\STATE {\bfseries Output:} ELBO $\mathcal{L}$
\STATE Sample ${\bm{v}} \sim q({\bm{v}}|{\bm{x}})$
\STATE Compute $\mathcal{L} = \log p({\bm{v}}) - \log q({\bm{v}}|{\bm{x}})$
\end{algorithmic}
\end{algorithm}
\end{minipage}
\end{table}
\input{sections/method/dequantization2}
\input{sections/method/multinomial_diffusion}
|
\section{Introduction}
\label{sec:intro}
In recent years, the growing body of literature claims a strong relationship between culture as defined and measured by \cite{Hofstede2010} and institutions as systematically described i.a.\ by \cite{North1990}, \cite{Olson1996, Olson1998} and a large variety of socio-economic phenomena. Culture and economic institutions are both independently important for economic development but their impact is stronger when combined \citep{Williamson2011}. This combined effect is at least of the same magnitude as variables of the standard production function \citep{Evan2021}. The relationship between culture and institutions remains, however, notoriously complex and thus difficult to measure. In this paper we have decided to discern the role of a cultural background in governance of a country. Specifically, we have two goals:
\begin{enumerate}
\item To examine the strength of the relationship between cultural characteristics of countries and the quality of their institutions.
\item To determine how well individual countries transform their cultural characteristics into institutions.
\end{enumerate}
In line with \cite{Williamson2011}, we consider that the establishment of institutions is a process originating in a nation's culture. Holders of a culture conducive to development may choose to formalize informal institutions into formal ones thus promoting this development. To this end, we employ stochastic frontier analysis of \cite{Aigner1977} and \cite{Meeusen1977}, which studies how efficiently a producer can transform inputs into outputs. As inputs, we use Hofstede's Cultural Dimensions \citep{Hofstede2010} measuring the characteristics of a culture. As outputs, we use Worldwide Governance Indicators \citep{Kaufmann2011} measuring the quality of institutions. Furthermore, to capture the economic environment, we use the gross domestic product (GDP) per capita. This framework allows us to achieve both of our research goals.
The rest of the paper is structured as follows. In Section \ref{sec:theory}, we lay out the theoretical foundations and review the related economic literature. In Section \ref{sec:ind}, we describe the cultural dimensions and the governance indicators. In Section \ref{sec:model}, we specify the stochastic frontier model we use. In Section \ref{sec:res}, we discuss the results of the efficiency analysis. We conclude the paper in Section \ref{sec:con}.
\section{Culture, Institutions, and Economic Development}
\label{sec:theory}
The idea of a significant cultural influence over institutions and governance is over a hundred years old \citep{Weber1905} but was strongly criticised by both Marxists \citep{Grossman2006} and by mainstream economists and remains a hotly debated issue to this day \citep{Blum2001, Becker2009, Cantoni2015, Kersting2020}. Mainstream economics has not readily even accepted institutions as a source of economic growth and development, innovation or quality of government, as the criticisms of the separate school of New Institutional Economics attests (\citealp{Coase1960, Coase1998, North1968, North1990, Olson1998, Hall1999}, i.a.). The idea has thus only very slowly made its appearance in economics before growing exponentially in recent years. There are papers using statistical analysis linking culture to innovation \citep{Zien1997, Tellis2009, Williams2010}, population growth \citep{Shennan2001, Yacout2015, Kumar2019}, environmental issues \citep{Peng2009, Nagy2018, Dangelico2020}, tax systems and collection \citep{Alm2006, Koenig2012, Cabelkova2013}, corruption \citep{Huber2001, Yeganeh2014}, software piracy \citep{Husted2000, Simmons2002}, and terrorism \citep{Meierrieks2013, Kluch2017} as well as a vast array of other systems or institutions. The first meta-analyses have appeared \citep{Taras2010, Buschgens2013} as there is enough data to attempt more complex studies. It seems the pendulum has swung from the cautious Boettke: “we cannot assume away cultural influences as economists have often done” \citep[p.\ 436]{Boettke2009} to the more radical Landes: “Max Weber had it right. If we learn anything from the history of economic development, it is that culture makes almost all the difference” \citep[p.\ 2]{Landes2000}.
The idea that culture has a significant impact on the reality surrounding us either directly or indirectly through institutions in a variety of situations can be accepted. There is agreement on the primacy of human marketable capital, or culture as defined by \cite{Olson1996} for both the improvement of institutions (e.g.\ democratization) and economic growth. There is also overwhelming evidence for a causal link between particular economic institutions, most notably property rights and economic freedom, and economic development (for a recent literature review and summary of evidence, please consult i.e.\ \citealp{Feld2008, Acemoglu2010, Czegledi2014, Wanjuu2017}). To establish a definite causal link between institutions and economic growth has proven difficult. It means among other things understanding the technology of the transmission of institutional quality to economic growth and development. This includes three challenges as described by \cite{Docquier2014}, i.e.\ (i) disentangling the causal effects and reversing the causal effects, (ii) accounting for unobserved shocks affecting both institutions and growth, and (iii) capturing the lag structure of the relationship. \cite{Acemoglu2005} claims political institutions as a fundamental cause of long-run economic growth. According to \cite{Acemoglu2005}, the knowledge of political institutions and the distribution of resources, or de facto political power, are sufficient to determine all the other variables in the system.
Some authors suggested that current measurement strategies have conceptual flaws and Human capital rather than political institutions have a causal effect on economic growth \citep{Djankov2003, Glaeser2004} also hinting at potentially significant reverse causal effects: “institutional outcomes also get better as the society grows richer, because institutional opportunities improve” \citep[p.\ 298]{Glaeser2004}. The standard institutionalists' approach that civil rights and democratic institutions cause development has been rather weakened in recent years as several studies using Granger regressions have found no evidence of this causality \citep{Paldam2012, Murtin2014}. Some authors returned to Lipset's hypothesis \citep{Lipset1959, Barro1997} reversing that causation and suggesting that development leads to democracy and civil rights and the issue remains unresolved \citep{Acemoglu2014, Jung2014, Czegledi2015}.
While we do not fully understand the relationship between institutions and economic growth or development it is more difficult still to determine the relationship between culture and institutions, or which one is more powerful in explaining the development in societies. A point in case is “Long-Term Persistence”, a paper by \cite{Guiso2016} in which they argue that Italian cities experiencing self-governance in the Middle Ages had a higher level of civic capital than other Italian cities. They offer three hypotheses with different causalities between culture and institutions (participation in public life in communes teaches people to cooperate and it is not forgotten; past democratic institutions change levels of trust and fairness in societies; historical events lead to changes in socialisation).
Apart from complexity and long time periods for change, the issue of complementarities arises between culture and institutions hindering definitions and identification of channels of causality \citep{Alesina2015}\footnote{Please also refer to \cite{Alesina2015} for a literature review of culture and institutions.}. The complementarity hypothesis suggests the strength of the impact of culture and institutions lies in their combination. The impact of a high trust culture complemented by a high level of rule of law enforcement on the business environment would be a good example. Both culture and institutions would be significant in regression analysis. \cite{Williamson2011} shed some light on the issue by testing it. They claim it is the substitution effect between culture and institutions which is more important. They describe this important mechanism: “a culture conducive to economic growth may choose to formalize the informal institutions into institutions associated with economic freedom” (p.\ 316). Once the formal rules (institutions) are credible “the informal norms and mechanisms once relied upon for economic interaction and exchange, such as trust networks, may be rendered much less important” (p.\ 316). Culture is important when institutions promoting economic freedom and thus growth are absent but diminishing in significance when those institutions are established. This implies the different relative importance of culture and institutions depending on the level of economic development of particular countries. Less developed countries should ceteris paribus have a stronger influence of culture on institutions than more developed ones.
\section{Cultural Dimensions and Governance Indicators}
\label{sec:ind}
The term culture in its broad sense includes both civic culture (formal, institutions) and personal culture (informal) (\citealp{North1990, Olson1996}, i.a.). In this text, we use the narrow definition referring only to personal culture as we want to discern the difference between the two. To measure culture, we use \emph{Hofstede's Cultural Dimensions}. In our efficiency analysis, these dimensions act as input variables. \cite{Hofstede2010} define the six cultural dimensions in the following way:
\begin{enumerate}
\item \emph{Power Distance (PDI)} is the extent to which the less powerful members of institutions and organizations within a country expect and accept that power is distributed unequally.
\item \emph{Individualism (IDV)} pertains to societies in which the ties between individuals are loose: everyone is expected to look after him- or herself and his or her immediate family. Collectivism as its opposite pertains to societies in which people from birth onward are integrated into strong, cohesive in-groups, which throughout people’s lifetime continue to protect them in exchange for unquestioning loyalty.
\item \emph{Masculinity (MAS)} refers to society where emotional gender roles are clearly distinct: men are supposed to be assertive, tough, and focused on material success, whereas women are supposed to be more modest, tender, and concerned with the quality of life. A society is called feminine when emotional gender roles overlap: both men and women are supposed to be modest, tender, and concerned with the quality of life.
\item \emph{Uncertainty Avoidance (UAI)} is the extent to which the members of a culture feel threatened by ambiguous or unknown situations.
\item \emph{Long-Term Orientation (LTO)} stands for the fostering of virtues oriented toward future rewards -- in particular, perseverance and thrift. Its opposite pole, short-term orientation, stands for the fostering of virtues related to the past and present -- in particular, respect for tradition, preservation of “face”, and fulfilling social obligations.
\item \emph{Indulgence (IVR)} stands for a tendency to allow relatively free gratification of basic and natural human desires related to enjoying life and having fun. Its opposite pole, restraint, reflects a conviction that such gratification needs to be curbed and regulated by strict social norms.
\end{enumerate}
We standardize the variables to lie between 0 and 1. Note that as culture is viewed as unchangeable, the variables are static. The source of the data is \cite{Hofstede2020} with the methodology described in \cite{Hofstede2010}. Alternatively, the personal culture could be measured by Schwartz's Theory of Basic Human Values and Tabellini's Indicators of Individual Values and Beliefs. The theory of Hofstede's cultural dimensions, however, remains the most popular tool. For a survey of the literature based on Hofstede's Cultural Dimensions, we refer to \cite{Kirkman2006} and \cite{Beugelsdijk2017}.
To define civic culture or institutions we use the \emph{Worldwide Governance Indicators} consisting of six indicators for over 200 countries and territories over the period 1996--2019. In our efficiency analysis, these indicators play the role of output variables. \cite{Kaufmann2011} define the six governance indicators in the following way:
\begin{enumerate}
\item \emph{Voice and Accountability (VA)} captures perceptions of the extent to which a country's citizens are able to participate in selecting their government, as well as freedom of expression, freedom of association, and a free media.
\item \emph{Political Stability and Absence of Violence/Terrorism (PV)} captures perceptions of the likelihood that the government will be destabilized or overthrown by unconstitutional or violent means, including politically‐motivated violence and terrorism.
\item \emph{Government Effectiveness (GE)} captures perceptions of the quality of public services, the quality of the civil service and the degree of its independence from political pressures, the quality of policy formulation and implementation, and the credibility of the government's commitment to such policies.
\item \emph{Regulatory Quality (RQ)} captures perceptions of the ability of the government to formulate and implement sound policies and regulations that permit and promote private sector development.
\item \emph{Rule of Law (RL)} captures perceptions of the extent to which agents have confidence in and abide by the rules of society, and in particular the quality of contract enforcement, property rights, the police, and the courts, as well as the likelihood of crime and violence.
\item \emph{Control of Corruption (CC)} captures perceptions of the extent to which public power is exercised for private gain, including both petty and grand forms of corruption, as well as "capture" of the state by elites and private interests.
\end{enumerate}
The variables are standardized to have zero mean and unit standard deviation. Higher values correspond to better governance. The source of the data is \cite{WGI2020} with the methodology described in \cite{Kaufmann2011}.
Finally, we consider the gross domestic product (GDP) per capita to be a control variable representing the economic environment. For the importance of inclusion of the operating environment in an efficiency analysis, see e.g.\ \cite{Holy2020a}. To remove the trend in time but to keep differences in levels between countries, we define our \emph{GDP Level} variable as the difference between the logarithm of GDP per capita in USD current prices and the logarithm of the mean GDP per capita in each year. Note that suitable specification of the GDP variable is of great importance as it can influence results of an efficiency analysis (see e.g.\ \citealp{Holy2018e}). The source of the data is \cite{EconomicOutlook2020}.
\section{Stochastic Frontier Model}
\label{sec:model}
To capture dependence between the inputs (Hofstede's Cultural Dimensions) and the outputs (the Worldwide Governance Indicators) and assess efficiency of the individual countries, we utilize the framework of stochastic frontier analysis of \cite{Aigner1977} and \cite{Meeusen1977}. We also include GDP Level as a control variable. We consider a separate stochastic frontier model for each output variable but employ the same structure for all six models. For a textbook treatment of stochastic frontier analysis, we refer to \cite{Kumbhakar2000}, \cite{Coelli2005}, \cite{Fried2008}. For a recent survey of the efficiency literature, we refer to \cite{Daraio2020}.
First, we build a standard linear regression model, estimate it by the ordinary least squares method and verify whether there is inefficiency present in the data. Technical inefficiency manifests as negatively skewed residuals. All output variables exhibit negative skewness ranging from -0.993 (RQ) to -0.027 (CC). It is therefore suitable to include the inefficiency variable into the model and utilize stochastic frontier analysis. Evenmore, the inefficient component in stochastic frontier models is dominant for all six output variables as its share of the total variance ranges from 0.804 (PV) to 0.922 (VA).
Second, we specify whether efficiencies are static or time-varying. As our main explanatory variables are static, we consider the static model to be more meaningful in our application. To quantify this, we start with the time-varying model of \cite{Battese1992} and find that the trend in efficiencies is quite negligible as its associated coefficient ranges from -0.006 (PV) to 0.003 (GE). We therefore resort to the time-invariant specification in the fashion of \cite{Pitt1981}.
Third, we specify the distribution of the technical inefficiency variable. As proposed by \cite{Stevenson1980}, we start with the two-parameter truncated normal distribution and find that the mode $\mu$ ranges from 0.896 (GE) to 1.089 (CC). Clearly, the parameter $\mu$ is significantly different from 0 and the distribution does not reduce to the half-normal distribution with zero mode. We therefore stick with the truncated normal distribution.
Our final model is described as follows. Let $N$ be the number of countries, $T$ the number of years, $M$ the number of output variables, $K$ the number of input variables, and $L$ the number of control variables. Dependent variable $Y_{itj}$; $i=1,\ldots,N$; $t=1,\ldots,T$; $j=1,\ldots,M$ then follows
$$
Y_{itj} = \alpha_{j} + \sum_{k=1}^K \beta_{jk} X_{itk} + \sum_{l=1}^L \gamma_{jl} Z_{itl} - U_{ij} + V_{itj},
$$
where $X_{itk}$ are input variables, $Z_{itl}$ are control variables, $U_{ij}$ are non-negative random variables capturing technical inefficiency, and $V_{itj}$ are random variables representing the error term. We assume that $U_{ij}$ are i.i.d.\ with the normal distribution $\mathrm{N} (\mu_j , \theta_j \sigma_j^2 )$ truncated at zero, $V_{itj}$ are i.i.d.\ with the normal distribution $\mathrm{N} ( 0, (1 - \theta_j) \sigma_j^2 )$, and $U_{ij}$ are independent of $V_{itj}$. Note that we utilize the parametrization of \cite{Battese1977} for $U_{ij}$ and $V_{itj}$. The model includes the constant parameter $\alpha_{j}$, the parameters for the input variables $\beta_{jk}$, the parameters for the control variables $\gamma_{jl}$, the variance of the random component $\sigma_j^2$, the ratio between the variance of the inefficiency variable and the error term $\theta_j$, and the mode of the inefficiency variable $\mu_j$. Using the abbreviations from Section \ref{sec:ind}, we can rewrite the model more specifically as
$$
\begin{aligned}
Y_{itj} &= \alpha_{j} + \beta_{j1} PDI_{it} + \beta_{j2} IDV_{it} + \beta_{j3} MAS_{it} + \beta_{j4} UAI_{it} \\
& \quad + \beta_{j5} LTO_{it} + \beta_{j6} IVR_{it} + \gamma_{j} GDP_{it} - U_{ij} + V_{itj}. \\
\end{aligned}
$$
We estimate the model by the maximum likelihood method.
\section{Results}
\label{sec:res}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=13cm]{figCoef.pdf}
\caption{Estimated coefficients of the linear regression model and the stochastic frontier model.}
\label{fig:coef}
\end{center}
\end{figure}
\begin{table}[ht]
\begin{center}
\caption{Estimated coefficients with standard errors of the stochastic frontier model.}
\label{tab:coef}
\footnotesize
\begin{tabular}{@{\extracolsep{5pt}}lcccccc}
\hline
& VA & PV & GE & RQ & RL & CC \\
\hline
Constant & 0.882$^{*}$ & 1.366$^{**}$ & 1.632$^{***}$ & 2.008$^{***}$ & 2.007$^{***}$ & 2.574$^{***}$ \\
& (0.485) & (0.582) & (0.277) & (0.263) & (0.314) & (0.355) \\
& & & & & & \\
Power Distance Index & $-$1.559$^{***}$ & $-$0.173 & $-$1.168$^{***}$ & $-$1.511$^{***}$ & $-$1.717$^{***}$ & $-$1.967$^{***}$ \\
& (0.393) & (0.487) & (0.191) & (0.400) & (0.505) & (0.382) \\
& & & & & & \\
Individualism & 1.261$^{***}$ & 0.207 & 0.557$^{*}$ & 0.017 & 0.621 & 0.544 \\
& (0.458) & (0.502) & (0.287) & (0.374) & (0.531) & (0.347) \\
& & & & & & \\
Masculinity & $-$0.130 & $-$0.912$^{***}$ & $-$0.240 & $-$0.179 & $-$0.587$^{*}$ & $-$0.784$^{***}$ \\
& (0.251) & (0.223) & (0.229) & (0.338) & (0.324) & (0.241) \\
& & & & & & \\
Uncertainty Avoidance & 0.484$^{**}$ & $-$0.585$^{***}$ & $-$0.701$^{***}$ & $-$0.372$^{*}$ & $-$0.436$^{*}$ & $-$0.753$^{***}$ \\
& (0.229) & (0.219) & (0.108) & (0.200) & (0.245) & (0.124) \\
& & & & & & \\
Long-Term Orientation & 0.686$^{**}$ & 0.566$^{*}$ & 0.879$^{***}$ & 0.780$^{***}$ & 0.770$^{***}$ & 0.815$^{***}$ \\
& (0.290) & (0.298) & (0.176) & (0.251) & (0.222) & (0.219) \\
& & & & & & \\
Indulgence & 0.895$^{***}$ & 0.215 & 0.482$^{***}$ & 0.326 & 0.415 & 0.568$^{***}$ \\
& (0.262) & (0.251) & (0.144) & (0.211) & (0.270) & (0.180) \\
& & & & & & \\
GDP Level & 0.076$^{***}$ & 0.397$^{***}$ & 0.318$^{***}$ & 0.346$^{***}$ & 0.281$^{***}$ & 0.264$^{***}$ \\
& (0.016) & (0.026) & (0.017) & (0.018) & (0.016) & (0.016) \\
& & & & & & \\
Parameter $\sigma^2$ & 0.371$^{***}$ & 0.488$^{***}$ & 0.233$^{***}$ & 0.276$^{***}$ & 0.258$^{***}$ & 0.329$^{***}$ \\
& (0.063) & (0.111) & (0.032) & (0.050) & (0.066) & (0.035) \\
& & & & & & \\
Parameter $\theta$ & 0.922$^{***}$ & 0.804$^{***}$ & 0.862$^{***}$ & 0.862$^{***}$ & 0.887$^{***}$ & 0.901$^{***}$ \\
& (0.013) & (0.044) & (0.017) & (0.023) & (0.027) & (0.009) \\
& & & & & & \\
Parameter $\mu$ & 1.065$^{***}$ & 0.912$^{***}$ & 0.896$^{***}$ & 0.976$^{***}$ & 0.958$^{***}$ & 1.089$^{***}$ \\
& (0.115) & (0.213) & (0.101) & (0.090) & (0.319) & (0.074) \\
\hline
\multicolumn{7}{r}{\textit{Note:} $^{***}p < 0.001$; $^{**}p < 0.01$; $^{*}p < 0.05$}
\end{tabular}
\end{center}
\end{table}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=13cm]{figWorld.pdf}
\caption{Mean efficiency scores obtained from the stochastic frontier model.}
\label{fig:world}
\end{center}
\end{figure}
Our data sample consists of $N=94$ countries observed over $T=21$ time periods from 1996 to 2019 with years 1997, 1999 and 2001 missing. Furthermore, there are additional 13 observations missing. We therefore have 1961 observations in total. The range of countries in our data sample is limited primarily due to the availability of the Hofstede's Cultural Dimensions data, while the range of years is limited due to the availability of the Worldwide Governance Indicators data.
We estimate both the linear regression model and the stochastic frontier model. The estimated coefficients are reported in Figure \ref{fig:coef} and Table \ref{tab:coef}. The models are able to capture a large portion of variability in the output variables. Specifically, the R$^2$ statistic in the regression model ranges from 0.545 (PV) to 0.813 (GE). Both models have very similar values of the coefficients suggesting the robustness of our approach. The significance of the coefficients, however, differs. In the stochastic frontier models, there are much fewer significant variables. As the stochastic frontier model is based on a more general distribution, it is more reliable and we focus solely on it from now on.
We find that the direction of the effect of Hofstede's Cultural Dimensions is consistent across all six Worldwide Governance Indicators. Individualism, Long-Term Orientation, and Indulgence have a positive effect on governance while Power Distance, Masculinity, and Uncertainty Avoidance have a negative effect. The only exception is the effect of Uncertainty Avoidance on Voice and Accountibility, which is positive. Power Distance and Long-Term Orientation play a significant role in five out of six indicators suggesting their universal impact. Masculinity, on the other hand, is found significant only for Political Stability and Control of Corruption; Individualism only for Voice and Accountability. The GDP control variable has a significant positive effect for all indicators. When omitted from the model, however, the results do not distinctly change.
Finally, we examine efficiency of the individual countries. In general, the efficiency score lies in the interval from 0 to 1 with lower values indicating inefficiency. The efficiency averaged over countries ranges from 0.357 (CC) to 0.423 (GE). The efficiency averaged over the output variables is shown in Figure \ref{fig:world} for the individual countries.
The countries that did well in our efficiency analysis are Chile and Uruguay in Latin America, Mozambique, Burkina Faso and Ivory Coast in Africa, Slovakia and Poland and Portugal in Europe as well as Malaysia in Asia. All of these countries have high values in Hofstede's Cultural Dimensions which we identified as a roadblock to good governance and development. All of the above mentioned countries have at least the score of 60 in Power Distance, they are Short-Term Oriented (below 40, with the exception of Slovakia), and generally have a high Uncertainty Avoidance (above 50, with the exception of Malaysia and Mozambique). These countries seem to have good governance relative to their culture.
The other end is represented by countries whose culture would allow for relatively good governance, yet, it is not in place. If we omit countries in civil war (Venezuela and to a lesser extent Argentina in Latin America), these are Angola, Nigeria, and Algeria in Africa, Iran in Asia, and Russia and Belarus in Europe. The rest of the countries fall in the middle of the sample as they seem to have as good governance as their culture would predict.
When sorted by Worldwide Governance Indicators, countries occupied an entire range of possible outcomes in efficiency. A group of three countries namely Portugal, Slovakia, and the former Portuguese colony of Cape Verde ranked highest in the Voice and Accountability indicator, all above 0.9, while four countries namely Iran, Saudi Arabia, Libya and China ranked below 0.1. In the Political Stability and Absence of Violence indicator, São Tomé \& Príncipe, Mozambique, and Slovakia, respectively were the only countries reaching the 0.9 mark, while Iraq, Turkey, and Nigeria, followed closely by Russia were the worst. Again, with the Government Effectiveness indicator only three countries achieved the value of 0.9. These are Malaysia, Portugal, and Chile. Libya, Venezuela, and Angola\footnote{It is of some interest that Portugal and former Portuguese colonies did very well in our results as they have cultures with high Power Distance but decent results in governance indicators. Angola is the only exeption where the Worldwide Governance Indicators are at alarming levels.} were the worst and the only countries of the sample below 0.2 level. Chile and Hong Kong both above 0.9, followed by Slovakia (0.777) were the three best countries in the Regulatory Quality indicator and Libya, Iran, and Venezuela the three worst. Chile, Portugal, and Hong Kong also were the best in Rule of Law while Venezuela, Libya, and Angola were the three worst. Finally, the Control of Corruption indicator saw Singapore, Uruguay, and Chile at the top and Angola, Latvia, and Lithuania at the bottom. The unexpected inclusion of the latter two countries suggests we might see an anti-corruption movement in those countries as the culture of those countries (Power Distance just 44) seems to be hostile to the current levels of corruption.
The overall results in mean efficiency score are as follows: Portugal (0.822), Chile (0.821), Slovakia (0.758), Hong Kong (0.743) and Ghana (0.686) are the top five countries and Libya (0.126), Venezuela (0.143), Iran (0.144), Angola (0.156) and Russia (0.186) are the five bottom ones.
\section{Conclusion}
\label{sec:con}
We have found an albeit rudimentary measurement of the strength of the relationship between cultural characteristics of countries and the quality of their institutions in 94 countries over 1996--2019. We have also determined efficiency scores for the transformation of cultural characteristics into institutions in each of those individual countries.
The results of our study pose several policy implications and avenues for further research. After more detailed analysis, policies on foreign aid levels can be improved as it might be possible to discern which countries' governance can be relatively successfully improved and/or which countries' governments are more effective in reforming national institutions for example in situations of adverse cultural patterns. The migration policies of countries deciding on levels and source countries of economic migration can also be improved as the difference between the influence of culture and governance in each country becomes clearer. There are also significant implications for international security studies and related fields.
There are several possible directions for the development of our model. In the current paper, we have considered GDP per capita as a control variable but did not further investigate its role in countries' culture and governance. As this is a hot topic in the literature, it will certainly be interesting to study this relationship in more detail. Migration and heterogeneity of the culture within a country could also be included in the model. Finally, we could improve our model by adding a spatial structure to capture dependency between countries.
\section*{Funding}
\label{sec:fund}
The work of Vladimír Holý was supported by the Czech Science Foundation under project 19-08985S.
|
\section{Introduction}
Stochastic linear bandits have a broad range of applications in practice, including online recommendations, job assignments in crowdsourcing, and clinical trials in healthcare. Most existing studies on stochastic linear bandits formulated them as unconstrained online optimization problems, limiting their application to problems with operational constraints such as safety, fairness, and budget constraints.
In this paper, we consider a stochastic linear bandit with general constraints.
As in a standard stochastic linear bandit, at the beginning of each round $t\in [T],$ the learner is given a context $c(t)$ that is randomly sampled from the context set $\mathcal C$ (a countable set), and takes an action $A(t)\in[J].$ The learner then receives a reward $R(c(t), A(t))=r(c(t),A(t)) + \eta(t),$ where $r(c, j) = \langle \theta_{*}, \phi(c, j) \rangle,$ $\phi(c,j)\in \mathbb R^d$ is a $d$-dimensional feature vector for (context, action) pair $(c,j),$ $\theta_{*} \in \mathbb R^d$ is an unknown underlying vector to be learned, and $\eta(t)$ is a zero-mean random variable.
For constrained stochastic linear bandits, we further assume when action $A(t)$ is taken on context $c(t),$ it incurs $K$ different types of costs, denoted by $W^{(k)}(c(t), A(t)).$ We assume $W^{(k)}(c,j)$ is a random variable with mean $w^{(k)}(c,j)$ that is unknown to the learner.
This paper considers general cost functions and does {\em not} require $w^{(k)}(c,j)$ to have a linear form like $r(c,j)$.
Denote the action taken by policy $\pi$ in round $t$ by $A^{\pi}(t)$. The learner's objective is to learn a policy $\pi$ that maximizes the cumulative rewards over horizon $T$ subject to {\em anytime cumulative constraints}:
\begin{align}
&\max_{\pi} \mathbb E\left[\sum_{t=1}^{T} R(c(t), A^\pi(t)) \right]
\label{obj-intro}\\
\hbox{subject to: }& \mathbb E\left[\sum_{t=1}^{\tau} W^{(k)}\left(c(t), A^\pi(t)\right) \right] \leq
0, ~\forall\ \tau\in[T], k\in[K]. \label{eq:cons-intro}
\end{align}
The constraint \eqref{eq:cons-intro} above may represent different operational constraints including safety, fairness, and budget constraints.
\subsection*{Anytime cumulative constraints}
In the literature,
constraints in stochastic bandits have been formulated differently. There are two popular formulations. The first one is a cumulative constraint over horizon $T,$ including knapsack bandits \cite{BadLanSli_14,BadAshKle_18,AgrDev_16,AgrDev_14,AgrDevLi_16,FerSimWan_18,CayErySri_20} where the process terminates when the total budget has been consumed; fair bandits where the number of times an action can be taken must exceed a threshold at the end of the horizon \cite{CheCueLuo_20}; and contextual bandits with a cumulative budget constraint \cite{WuSriLiu_15,ComJiaSri_15}. In these settings, the feasible action set in each round depends on the history. In general, the learner has more flexibility in the earlier rounds, close to that in the unconstrained setting. Another formulation is {\em anytime} constraints, which either require the expected cost of the action taken in each round to be lower than a threshold \cite{AmaAliThr_19,MorAmaAli_19} or the expected cost of the policy in each round is lower than a threshold \cite{PacGhaBar_20}. We call them {\em anytime} action constraints and {\em anytime} policy constraints, respectively.
\begin{wrapfigure}{r}{0.45\textwidth}
\centering
\includegraphics[width=0.4\textwidth]{figures/constraints.pdf}
\caption{\small{A conceptual description of feasible policy sets under different constraint formulations.}}
\label{fig:constraints}
\end{wrapfigure}
Our constraint in the form of \eqref{eq:cons-intro} is an {\em anytime cumulative constraint}, i.e., it imposes a cumulative constraint in {\em every} round. This anytime cumulative constraint is most similar to the anytime policy constraint in \cite{PacGhaBar_20} because the average cost of a policy is close to its mean after the policy has been applied for many rounds and the process converges, so it can be viewed as a cumulative constraint on actions over many rounds (like ours). Furthermore, when our anytime cumulative constraint \eqref{eq:cons-intro} is satisfied, it is guaranteed that the time-average cost is below a threshold in every round.
In summary, our anytime cumulative constraint is stricter than a cumulative constraint over fixed horizon $T$ but is less restrictive than anytime action constraint in \cite{AmaAliThr_19,MorAmaAli_19}. Figure~\ref{fig:constraints} provides a conceptual description of the relationship between these different forms of constraints.
\subsection*{Main Contributions}
\label{sec:contribution}
This paper presents a pessimistic-optimistic algorithm based on the primal-dual approach in optimization for the problem defined in \eqref{obj-intro}-\eqref{eq:cons-intro}. The algorithm is efficient in two aspects. First, the algorithm yields $\tilde{\cal O}\left(\left(\frac{K^{0.75}}{\delta}+d\right)\sqrt{\tau}\right)$ regret in round $\tau\leq T$ and achieves {\em zero} constraint violation in any round $\tau>\tau'$ for a constant $\tau'$ {independent} of horizon $T$. Second, the algorithm is computationally efficient.
For computational efficiency, the design of our algorithm is based on the primal-dual approach in optimization. The computation of the primary component is similar to unconstrained stochastic linear bandits \cite{DanHayTho_08,RusTsi_10,LiChuLan_10,AbbPalSze_11,ChuLiRey_11}.
The dual component includes a set of Lagrangian multipliers that are updated in a simple manner to keep track of the levels of constraint violations so far in each round; the update depends on the number of constraints, but it is independent of the sizes of the contextual space, the action space, and the feature space. Thus, the overall computational complexity of our algorithm is similar to that of LinUCB in the unconstrained setting. This results in a much more efficient calculation comparing to OPLB proposed in \cite{PacGhaBar_20}. OPLB needs to construct a safe policy set in each round, hence, its computational complexity is prohibitively high as the authors acknowledged.
For constraint violation, our algorithm guarantees that for any $\tau>\tau',$ the constraint holds with probability one. In other words, after a constant number of rounds, the constraint is always satisfied. This is in contrast to prior works \cite{PacGhaBar_20,AmaAliThr_19}, where anytime constraints are proven to hold over horizon $T$ with probability $1-\chi$ for a constant $\chi$. In other words, the anytime constraints may be violated with probability $\chi,$ and it is not clear how often they are violated when it happens. Furthermore, beyond mean cost constraints considered in \eqref{eq:cons-intro} and in \cite{PacGhaBar_20,AmaAliThr_19}, we prove that a sample-path version of constraint \eqref{eq:cons-intro} holds with probability $1-O\left(e^{-\frac{\delta\sqrt{\tau}}{50K^{2.5}}}\right)$ in round $\tau$ under our algorithm.
To summarize, our algorithm is computationally efficient and provides strong guarantees on both regret and constraint violations. Additionally, our cost function is in a general form and does not need to be linear as those in \cite{PacGhaBar_20,AmaAliThr_19}.
We discuss more related work in the following.
\subsection*{Related Work}
\label{sec:related}
Stochastic linear bandits \cite{AbeLon_99, Auer_03} are a special class of contextual bandits \cite{Mic_79, LanZha_08}, which generalize multi-armed bandits \cite{LaiRob_85}.
Besides \cite{PacGhaBar_20}, \cite{CheCueLuo_20} considered an adversarial contextual bandit with anytime policy constraint representing fairness.
The proposed algorithm has $\tilde{\cal O}(\sqrt{|\mathcal C|JT})$ regret when the context distribution is known to the learner; otherwise it has $\tilde{\cal O}(\sqrt{|\mathcal C|JT})$ regret and $\tilde{\cal O}(\sqrt{|\mathcal C|T})$ constraint violation. \cite{LiLiuJi_19} studied a combinatorial sleeping bandits problem under cumulative fairness
constraints and proposed an algorithm based on UCB which they {\em conjectured} to have $\tilde{\cal O}(\sqrt{T})$ regret and $\tilde{\cal O}(\sqrt{T})$ constraint violation.
Recent work studied unconstrained structured bandits and proposed primal-dual approach based on asymptotically lower bound problem in bandits \cite{KirLatVerSze_21, TirPirResLaz_20, DegShaKoo_20}.
However, our algorithm is different from them in three aspects. Our primal component is a greedy algorithm instead of a (sub-)gradient algorithm (as in \cite{TirPirResLaz_20}). Our dual component does not solve a best response problem, which is a constrained optimization problem as in \cite{KirLatVerSze_21, DegShaKoo_20}. Our analysis is based on the Lyapunov-drift analysis for queueing systems, e.g., we establish a bound on the exponential moment of the dual variable, which is not present in \cite{KirLatVerSze_21, TirPirResLaz_20, DegShaKoo_20}. It is also worth mentioning that \cite{KiaEil_20, AbbMohYad_17,AhmChrMah_20} studied ``conservative'' bandits which require that the reward or the cumulative reward exceeds a threshold at each step. Another line of related work is
online convex optimization with constraints, studied in \cite{ShiJohYu_09,MahJinYan_12,YuNee_20,YuNeeWei_17,WeiYuNee_20,UsmKraKam_19}, where online primal-dual with proximal regularized algorithms have been proposed to achieve $O(\sqrt{T})$ regret and $O(1)$ violation for static constraints in \cite{YuNee_20} and $O(\sqrt{T})$ violation for stochastic constraints in \cite{YuNeeWei_17} .
\noindent{\bf Notation.}
$f(n) = \tilde{\mathcal O}(g(n))$ denotes $f(n) = O(g(n){\log}^k n)$ with $k>0;$ $[N]$ denotes the set $\{1,2,\cdots, N\};$ $\langle \cdot, \cdot \rangle$ denotes the inner product; $(\cdot)^\dag$ denotes the transpose of a vector or a matrix; $||\cdot||=||\cdot||_2,$ and $||\mathbf x||_{\Sigma} = \sqrt{\mathbf x^{\dag} \Sigma \mathbf x}.$ We add subscript $t$ to a variable when it is a time-varying sequence of constants (e.g., $c_t$), and add $(t)$ when they are random variables or decision variables (e.g., $c(t)$). We summarize our notation in Appendix \ref{app:notation}.
\section{A Pessimistic-Optimistic Algorithm}
\label{sec:alg}
We consider a stochastic linear bandit over horizon $T$ as described in the introduction. The learner's objective is to maximize the cumulative reward over time horizon $T$ subject to $K$ anytime cumulative constraints as defined in \eqref{obj-intro}-\eqref{eq:cons-intro}.
To address the challenges on the \emph{unknown} reward and cost in constraint, as well as the anytime cumulative constraints, we develop a pessimistic-optimistic algorithm based on the primal-dual approach for constrained optimization. We first give out the intuition of the algorithm and then provide the formal statement of the algorithm.
To start, we consider a baseline, deterministic problem that replaces all the random variables with their expectations. Different from the conventional setup, we introduce a ``tightness'' constant $\epsilon>0$:
\begin{align}
\max_{\mathbf x} & ~ \sum_{c\in\mathcal C,j\in[J]} p_c r(c,j) x_{c,j}
\label{obj-fluid-tightened}\\
\text{s.t.} & ~ \sum_{j\in[J]} x_{c,j} = 1, ~x_{c,j} \geq 0, \forall c \in \mathcal C, \label{arrival-fluid-tightened} \\
& ~ \sum_{c\in\mathcal C,j\in[J]}p_cw^{(k)}(c,j) x_{c,j} +\epsilon\leq 0, ~\forall k \in [K], \label{resource limit-fluid-tightened}
\end{align}
where $x_{c,j}$ can be viewed as the probability of taking action $j$ on context $c,$ and $p_c$ is the probability that context $c$ is selected in each round. We will discuss in further details the importance of the tightness constant $\epsilon$ in Section~\ref{sec:main proof}.
The Lagrangian of the problem above is
\begin{align}
\max_{{\mathbf x}: \sum_{j} x_{c,j} = 1, x_{c,j} \geq 0} \sum_{c,j} p_c r(c,j) x_{c,j}-\sum_{k}\lambda^{(k)}\left(\sum_{c\in\mathcal C,j\in[J]}p_cw^{(k)}(c,j) x_{c,j} +\epsilon \right),
\end{align} where $\lambda^{(k)}$ is the Lagrange multiplier associated with the $k$th constraint in \eqref{resource limit-fluid-tightened}. Fixing the values of the Lagrange multipliers, solving the optimization problem is equivalent to solving $|\mathcal C|$ separate subproblems \eqref{eq:subprob}, one for each context $c,$ because the optimization variables $\bf x$ are coupled through $j$ only:
\begin{align}
\max_{{\mathbf x}: \sum_{j} x_{c,j} = 1, x_{c,j} \geq 0} p_c\left(\sum_{j} r(c,j) x_{c,j}-\sum_{k}\lambda^{(k)}\left(\sum_{j}w^{(k)}(c,j) x_{c,j} \right)\right).\label{eq:subprob}
\end{align}
Since the problem above is a linear programming, one of the optimal solutions is $x_{c,j}=1$ for $j=j^*$ and $x_{c,j}=0$ otherwise, where
\begin{equation}
j^*\in\arg\max_j r(c,j) -\sum_{k}\lambda^{(k)} w^{(k)}(c,j)
\label{eq:maxvalue}
\end{equation} and a tie can be broken arbitrarily. If we call $r(c,j) -\sum_{k}\lambda^{(k)} w^{(k)}(c,j)$ the action-value of context $c,$ then the solution for fixed values of Lagrange multipliers is to choose an action with the highest action-value. We may view the action value here plays a similar role as the Q-value (also called action-value function) in Q-learning \cite{WatDay_92}.
Now the challenges to find a solution according to \eqref{eq:maxvalue} include: (i) both $r(c,j)$ and $w^{(k)}(c,j)$ are unknown, and (ii) the optimal Lagrange multipliers $\lambda^{(k)}$ are also unknown. To overcome these challenges, we develop a pessimistic-optimistic algorithm that
\begin{itemize}[leftmargin=*]
\item Uses LinUCB to estimate $r(c,j)$ based on its linear structure.
\item Uses observed $W^{(k)}(c(t),j)$ to replace $w^{(k)}(c(t),j)$ at each round $t.$
\item Uses the following function to dynamically approximate the Lagrange multipliers:
\begin{align*}
Q^{(k)}(t+1) =& \left[Q^{(k)}(t) + \sum_{j\in [J]}W^{(k)}(c(t),j) X_{j}(t) + \epsilon_t \right]^{+}, \forall k.
\end{align*} In other words, we increase its value when the current cost exceeds the current ``budget,'' and decrease it otherwise. Therefore, $Q^{(k)}(t)$ keeps track of the cumulative constraint violation by round $t$.
\item We further add a scaling parameter $1/V_t$ to $Q^{(k)}(t),$ i.e. $\frac{Q^{(k)}(t)}{V_t},$ to approximate $\lambda^{(k)}.$ With a carefully designed $V_t,$ we can control the tradeoff between maximizing reward and minimizing constraint violation in the policy and achieve the regret and constraint violation bounds to be presented in the main theorem.
\end{itemize}
\noindent Next, we formally state our algorithm. This algorithm
takes the following information as input at the beginning of each round $t$: (i) historical observations $${\cal F}_{t-1}=\{c(s), A(s), R(c(s),A(s)), W^{(k)}(c(s), A(s))\}_{s\in[t-1], k\in[K]},$$ (ii) current observations $c(s)$ and $\{W^{(k)}(c(s), j)\}_{k\in[K],j\in[J]},$ and (iii) system parameters: the feature map $\{\phi(c,j)\}_{c\in\mathcal C,j\in[J]},$ time horizon $T,$ and a pre-set constant $\delta$. In the analysis of our algorithm, we will reveal the connection of this constant $\delta$ with Slater's condition.
The algorithm outputs the action in each round, observes reward $R(c(t), A(t))$, makes updates, and then moves to the next round $t+1$.
\vspace{0.1in}
\hrule
\vspace{0.1in}
\noindent{\bf A Pessimistic-Optimistic Algorithm}
\vspace{0.1in}
\hrule
\vspace{0.1in}
\noindent {\bf Initialization:} $Q^{(k)}(1)=0,$ $\mathcal B_1 = \{\theta | ||\theta||_{\Sigma_0} \leq \sqrt{\beta_1} \},
\Sigma_0 = \mathbf I~\text{and}~\sqrt{\beta_1} = m + \sqrt{2\log T}.$
\noindent For $t=1,\cdots, T,$
\begin{itemize}[leftmargin=*]
\item \noindent {\bf Set:}
$V_t=\delta K^{0.25}\sqrt{\frac{2t}{3}}$ and $\epsilon_t=K^{0.75}\sqrt{\frac{6}{t}}.$
\item {\bf LinUCB (Optimistic):} Use LinUCB to estimate $r(c(t),j)$ for all $j:$ $$\hat r(c(t),j) = \min\{1,\tilde r(c(t),j)\} ~~\text{with}~~ \tilde r(c(t),j) = \max_{\theta \in \mathcal B_t} \langle \theta, \phi(c(t),j)\rangle.$$
\item {\bf MaxValue:} Compute {\em pseudo-action-value} of context $c(t)$ for all action $j,$ and take the action $j^*$ with the highest pseudo-action-value, breaking a tie arbitrarily
\begin{align*}
j^*\in \arg\max_j \underbrace{\hat r(c(t),j) - \frac{1}{V_t}\sum_k W^{(k)}(c(t),j)Q^{(k)}(t)}_{\text{pseudo action value of $(c(t),j)$}}.
\end{align*}
\item {\bf Dual Update (Pessimistic):} Update the estimates of dual variables $Q^{(k)}(t)$ as follows:
\begin{align}
Q^{(k)}(t+1) =& \left[Q^{(k)}(t) + \sum_jW^{(k)}(c(t),j) X_{j}(t) + \epsilon_t\right]^{+}, \forall k.\label{eq:dual-update}
\end{align}
\item {\bf Confidence Set Update:} Update $\Sigma_t,$ $\hat \theta_t,$ $\beta_{t+1}$ and $\mathcal B_{t+1}$ according to the received reward $R(c(t), j^*):$
\begin{align*}
\Sigma_t = \Sigma_{t-1} + \phi(c(t),j^*) \phi^{\dag}(c(t),j^*), ~~~~
\hat \theta_t = \Sigma^{-1}_t\sum_{s=1}^t \phi(c(s),A(s)) R(c(s), A(s)), \\
\sqrt{\beta_{t+1}} = m + \sqrt{2 \log T + d\log \left(\frac{d+t}{d}\right)}, ~~~ \mathcal B_{t+1} = \{\theta ~|~ ||\theta - \hat \theta_{t} ||_{\Sigma_{t}} \leq \sqrt{\beta_{t+1}} \}.
\end{align*}
\end{itemize}
\vspace{0.1in}
\hrule
\vspace{0.1in}
The complexity of our algorithm is similar to LinUCB. The additional complexity is proportional to the number of constraints (for updating $Q^{(k)}$), and it is much lower than OPLB in \cite{PacGhaBar_20}, where the construction of a safe policy set in each round is a major computational hurdle. Our algorithm is computationally efficient. Additionally, our algorithm does not estimate $w^{(k)}(c,j)$, hence, we do not need to make any specific assumption on $w^{(k)}(c,j).$
\section{Main Results: Regret and Constraint Violation Bounds}
\label{sec:main-results}
To understand the performance of a given policy $\pi,$ we will analyze both the regret and the constraint violation. For that, we first define the baselines and state the assumptions made for the performance analysis. Then, we present our main results on the regret bound and constraint violations -- for the latter, we present both results on expected violation and an additional high probability bound for pathwise constraint violation.
\subsection{Baselines and Assumptions}
\noindent{\bf Regret baseline:} We consider the following optimization problem:
\begin{align}
\max_{\mathbf x} & ~ \sum_{c\in\mathcal C,j\in[J]} p_c r(c,j) x_{c,j}
\label{obj-fluid}\\
\text{s.t.} & ~ \sum_{j\in[J]} x_{c,j} = 1, ~x_{c,j} \geq 0, \forall c \in \mathcal C, \label{arrival-fluid} \\
& ~ \sum_{c\in\mathcal C,j\in[J]}p_cw^{(k)}(c,j) x_{c,j} \leq 0, ~\forall k \in [K]. \label{resource limit-fluid}
\end{align}
\noindent{\bf Constraint violation baseline:} We choose zero (no violation) as our baseline.
It worth noting the baseline we use in the regret analysis is derived from relaxed cumulative constraints instead of anytime cumulative constraints in the original problem. Since the cumulative constraint is the least restrictive constraint, a learner obtains the highest cumulative rewards in such a setting. In other words, our regret analysis is with respect to the best (the most relaxed) baseline.
We make the following assumptions for all the results present in this paper.
\begin{assumption}
The context $c(t)$ are i.i.d. across rounds. The mean reward $r(c, j)=\langle \theta_*, \phi(c,j) \rangle\in [0, 1]$ with $||\phi(c,j)||\leq 1, ||\theta_*|| \leq m$ for any $c \in \mathcal C,\ j \in [J].$ $\eta(t)$ is zero-mean $1$-subGaussian conditioned on $\{\mathcal F_{t-1}, A(t)\}$. \label{assumption:reward}
\end{assumption}
\begin{assumption}
The costs in the constraints
satisfy $|W^{(k)}(c,j)|\leq 1.$
Furthermore, we assume $\{W^{(k)}(c, j)\}_{t=1}^{T}$ are i.i.d. samples for given $c$ and $j$. \label{assumption:constraints}
\end{assumption}
\begin{assumption}[Slater's condition]
There exists $\delta > 0$ such that there exists a feasible solution $\bf x$ to optimization problem \eqref{obj-fluid}-\eqref{resource limit-fluid} that guarantees $\sum_{c\in\mathcal C,j\in[J]} p_cw^{(k)}(c,j) x_{c,j} \leq -\delta, \forall k \in [K].$ We assume $\delta\leq 1$ because if the condition holds for $\delta>1,$ it also holds for $\delta=1$. \label{assumption:slater}
\end{assumption}
We call $\delta$ Slater's constant because it comes from Slater's condition in optimization -- this is the constant used as an input our algorithm. This constant plays a similar role as the cost of a safe action in \cite{AmaAliThr_19,PacGhaBar_20}. In fact, a safe action guarantees the existence of a Slater's constant, and we can estimate the constant by running the safe action for a period of time. However, the existence of a Slater's constant does not require the existence of a safe action. It is also a more relaxed quantity than the safety gap in \cite{AmaAliThr_19}, which is defined under the optimal policy. Slater's constant can be from a feasible solution that is not necessarily optimal.
The next lemma shows that the optimal value of \eqref{obj-fluid}-\eqref{resource limit-fluid} is an upper bound on that of \eqref{obj-intro}-\eqref{eq:cons-intro}. The proof of this lemma can be found in Appendix \ref{app: baseline}.
\begin{lemma} \label{lemma: fluid upper bound}
Assume $\{c(t)\}$ are i.i.d. across rounds, and $\{R(c, j)\}$ and $\{W^{(k)}(c, j)\}$ are i.i.d. samples given action $j$ and context $c.$ Let $\pi^*$ be the optimal policy to problem \eqref{obj-intro}-\eqref{eq:cons-intro} and $\mathbf x^*$ be the solution to \eqref{obj-fluid}-\eqref{resource limit-fluid} with entries $\{x_{c,j}^*\}_{c\in\mathcal C, j \in [J]}$. We have
$$\mathbb E\left[\sum_{t=1}^{T}\sum_{j\in[J]} R(c(t), j)X^{\pi^*}_{j} (t)\right] \leq T\sum_{c\in\mathcal C,j\in[J]} p_c r(c,j) x^*_{c,j}.$$
\end{lemma}
\noindent The baseline problem \eqref{obj-fluid}-\eqref{resource limit-fluid} is the same as the one presented in Section~\ref{sec:alg} except that the tightness constant $\epsilon = 0$ here. Any feasible solution for the tightened problem in Section~\ref{sec:alg} is a feasible solution to the baseline problem. Under Slater's condition, when $\epsilon<\delta,$ the tightened problem also has feasible solutions.
\subsection{Regret and Constraint Violation Bounds}
Given the baselines above, we now define regret and constraint violation.
\noindent{\bf Regret:} Given policy $\pi,$ we define the (pseudo)-regret of the policy to be
\begin{align}
\mathcal R(\tau) =& \tau \sum_{c\in\mathcal C,j\in[J]} p_c r(c,j) x^*_{c,j} - \mathbb E\left[\sum_{t=1}^{\tau} \sum_{j\in [J]}R(c(t), j) X^{\pi}_{j}(t)\right]. \label{def:regret}
\end{align}
\noindent{\bf Constraint violation:} The constraint violation in round $\tau$ is defined to be
\begin{align}
\mathcal V(\tau) = \sum_{k\in[K]} \left( \mathbb E\left[\sum_{t=1}^{\tau}\sum_{j\in[J]}W^{(k)}(c(t),j)X^{\pi}_{j}(t)\right] \right)^+. \label{def:cv}
\end{align}
Note that the operator $(\cdot)^+=\max(\cdot, 0)$ is imposed so that different types of constraint violations will not be canceled out.
\begin{theorem}
Under Assumptions \ref{assumption:reward}-\ref{assumption:slater}, the pessimistic-optimistic algorithm
achieves the following regret and constraint violations bounds for any $\tau \in [T]:$
\begin{align*}
\mathcal R(\tau) \leq& \frac{60K^{3}}{\delta^3}+ \frac{4\sqrt{6}K^{0.75}\sqrt{\tau}}{\delta} + 2 +\sqrt{8d\tau \beta_{\tau}(T^{-1}) \log \left(\frac{ d + \tau}{ d}\right)},\\
\mathcal V(\tau) \leq& K^{1.5}\left(\frac{48K^2}{\delta}\log\left(\frac{16}{\delta}\right) + \frac{24K^{1.5}}{\delta^2}+\frac{30K^{1.5}}{\delta}+ 8K - \sqrt{\tau} \right)^{+}.
\end{align*}
where $\sqrt{\beta_{\tau}(T^{-1})} = m+\sqrt{2\log T + d \log \left(1+\tau/d\right)}.$
\label{thm:formal-UCB}
\end{theorem}
We make a few important observations from our theoretical results. First, for the \textit{reward regret}, we observe
$${\cal R}(\tau)=\tilde{\cal O}\left(\left(\frac{K^{0.75}}{\delta}+d\right)\sqrt{\tau}\right).$$
So the regret is independent of the number of contexts, action space $[J]$ and the dimension of cost functions $W^{(k)}(\cdot,\cdot).$ It grows sub-linearly in $\tau$ and the number of constraints $K,$ and linearly in the dimension of reward feature $d$ and the inverse of Slater's constant $\delta.$
Second, for the \textit{constraint violation}, we observe
\begin{equation}
{\cal V}(\tau)=\begin{cases}
\tilde{\cal O}\left( \frac{K^{3.5}}{\delta} + \frac{K^{3}}{\delta^2} \right)&\tau\leq \tau'=\tilde{\cal O}\left(\frac{K^4}{\delta^2}+\frac{K^3}{\delta^4}\right)\\
0&\hbox{otherwise}
\end{cases}.
\end{equation}
That is, the constraint violation is {\em independent} of horizon $T$ and becomes {\em zero} when $\tau > \tau'.$ The constraint violation, however, has a strong dependence on $K$ and $\delta$ when $\tau\leq \tau'.$ This is not surprising because $K$ defines the number of constraints, and $\delta$ represents the tightness of the constraints (or size of the feasible set).
\noindent\textbf{Dependence on Slater's constant. } Both the regret and constraint violation increase in $\delta$. To see the intuition, note that $\delta$ determines the size of the feasible set for the optimization problem. A larger $\delta$ implies a larger feasible set, so it is easier to find a feasible solution, vice versa. Therefore, both regret and constraint violation increase as $\delta$ decreases because the problem becomes harder and requires more accurate learning.
\noindent\textbf{Sharpness of the bound.}
In terms of horizon $T,$ the bounds in Theorem \ref{thm:formal-UCB} are sharp because the regret bound $\mathcal R(T)$ matches the instance-independent regret $\Omega(\sqrt{T})$ in multi-armed bandit problems without constraints \cite{AueBiaFre_95, GerLat_16} up to logarithmic factors. Furthermore, zero constraint violation is the best possible. Therefore, the bounds are sharp up to logarithmic factors in terms of horizon $T$. It is not clear whether these bounds are sharp in terms of $K$, $d$, and $\delta$, which are interesting open questions.
\subsection{A High Probability Bound on Constraint Violation}
The constraint \eqref{eq:cons-intro} defined in the original problem and the constraint violation measure defined in \eqref{def:cv} are both in terms of expectation. An interesting, related question is what the probability is for a {\em sample-path version} of the constraints to be satisfied. It turns out that our algorithm provides a high probability guarantee on that as well. The proof can be found in Appendix \ref{app: tail prob}.
\begin{corollary}\label{cor: tail prob}
The pessimistic-optimistic algorithm guarantees that for any $\tau \geq \frac{\kappa K^5}{\delta^2}\left(\log\left(\frac{K}{\delta}\right)\right)^2$, where $\kappa$ is a positive constant independent of $\tau,$ $K,$ $\delta$ and $d$,
$$\mathbb P\left(\sum_{t=1}^{\tau}\sum_{j\in[J]}W^{(k)}(c(t),j)X_{j}(t)>0\right)=O\left(e^{-\frac{\delta\sqrt{\tau}}{50K^{2}}}\right).$$
\end{corollary}
\section{ Proof of Theorem \ref{thm:formal-UCB}}\label{sec:main proof}
We first explain the intuition behind the main result. Recall that the algorithm selects action $j^*$ such that
\begin{align*}
j^* \in \arg\max_j \left(\hat r(c(\tau),j) - \frac{1}{V_\tau}\sum_k W^{(k)}(c(\tau),j)Q^{(k)}(\tau)\right),
\end{align*} and $V_\tau=O(\sqrt{\tau}).$ Therefore, when $Q^{(k)}(\tau)=o(\sqrt{\tau}),$ the reward term dominates the cost term, and our algorithm uses LinUCB to maximize the reward. When $Q^{(k)}(\tau)=\omega(\sqrt{\tau}),$ the cost term dominates the reward term and our algorithm focuses on reducing $Q^{(k)}$. Slater's condition implies that there exists a policy that can reduce $Q^{(k)}$ by a constant (related to $\delta$) in each round. Therefore, the algorithm takes $\tilde{\cal O}(\sqrt{\tau})$ rounds to reduce $Q^{(k)}$ to $o(\sqrt{\tau}),$ which may add $\tilde{\cal O}(\sqrt{\tau})$ to the regret during this period. The argument above also implies that $Q^{(k)}(\tau)=O(\sqrt{\tau}).$ Then, because
\begin{align}
\mathbb E\left[\sum_{t=1}^\tau \sum_{j\in [J]}W^{(k)}(c(t),j) X_{j}(t)\right]\leq \mathbb E\left[Q^{(k)}(\tau+1)\right]-\sum_{t=1}^\tau \epsilon_t.\nonumber
\end{align}
we can further bound the constraint violation at time $\tau$ to be a constant or even zero via the bound on $\mathbb E[Q^{(k)}(\tau+1)]$ and a proper choice of $\epsilon_t.$
\subsection{Regret Bound}\label{sec: regret}
Now consider the regret defined in \eqref{def:regret} and define $\mathbf x^{\epsilon_t,*}$ to be the optimal solution to the tightened problem \eqref{obj-fluid-tightened}-\eqref{resource limit-fluid-tightened} with $\epsilon=\epsilon_t.$ We obtain the following decomposition by adding and subtracting corresponding terms:
\begin{align}
{\cal R}(\tau)=&\tau \sum_{c,j}p_cr(c,j)x_{c,j}^* - \mathbb E\left[\sum_{t=1}^{\tau} \sum_{j}R(c(t), j)X_{j}(t)\right] \nonumber\\
\stackrel{(a)}{=}&\tau \sum_{c,j}p_cr(c,j)x_{c,j}^* - \mathbb E\left[\sum_{t=1}^{\tau} \sum_{j}r(c(t), j)X_{j}(t)\right] \nonumber\\
=&
\underbrace{\sum_{t=1}^\tau \sum_{c,j}p_cr(c,j)\left(x_{c,j}^*-x_{c,j}^{\epsilon_t,*}\right)}_{\text{$\epsilon_t$-tight}}
+ \underbrace{\mathbb E\left[\sum_{t=1}^{\tau} \sum_{j} \left(r(c(t),j)-\hat r(c(t),j)\right)x_{c(t),j}^{\epsilon_t,*}\right]}_{\text{reward mismatch}}
\nonumber\\
&+\underbrace{\mathbb E\left[ \sum_{t=1}^{\tau} \sum_{j}\left(\hat r(c(t),j)x_{c(t),j}^{\epsilon_t,*} - \hat r(c(t),j)X_{j}(t)\right)\right] - \sum_{t=1}^{\tau} \frac{K(1+\epsilon_t^2)}{V_t}}_{\text{Lyapunov drift}} \label{eq:driftterm}\\
&+\underbrace{\sum_{t=1}^{\tau} \frac{K(1+\epsilon_t^2)}{V_t}}_{\text{accumulated tightness}}+ \underbrace{ \mathbb E\left[\sum_{t=1}^{\tau}\sum_{j} \left(\hat r(c(t),j)-r(c(t),j)\right)X_{j}(t)\right]}_{\text{reward mismatch}}, \label{intuition}
\end{align} where $(a)$ holds because the random reward is revealed after action $A(t)$ is taken so the noise is independent of the action.
We next present a sequence of lemmas that bounds the terms above. The proofs of these lemmas are presented in Appendices \ref{app: drif in regret}, \ref{app: epsilon gap}, and \ref{app: linucb-r}. The key to the proof, particularly the proof of Lemma~\ref{lemma: drift in regret}, relies on the Lyapunov-drift analysis; a comprehensive introduction of the method can be found in \cite{Nee_10,SriYin_14}.
\begin{lemma}\label{lemma: drift in regret}
Under the Pessimistic-Optimistic Algorithm, we have
\begin{align}
\mathbb E\left[\sum_{t=1}^{\tau}\sum_{j\in [J]} \hat r(c(t),j) \left(x^{\epsilon_t}_{c(t),j}-X_{j}(t)\right) \Big| \mathbf H(t)=\mathbf h\right] - \sum_{t=1}^{\tau}\frac{K(1+\epsilon_t^2)}{V_t} \leq \sum_{t=1}^{\tau}\frac{ Kt(\epsilon_t+\epsilon_t^2) \mathbb I(\epsilon_t > \delta)}{V_t}. \nonumber
\end{align}
\end{lemma}
\begin{lemma}\label{lemma: elsilon gap}
Under Assumptions \ref{assumption:reward}-\ref{assumption:slater}, we can bound the difference between the baseline optimization problem and its tightened version:
$$\sum_{t=1}^\tau \sum_{c,j}p_cr(c,j)\left(x_{c,j}^*-x_{c,j}^{\epsilon_t,*}\right) \leq \sum_{t=1}^{\tau}\frac{\epsilon_t}{\delta}.$$
\end{lemma}
\begin{lemma}\label{lemma: bandits}
Under the Pessimistic-Optimistic Algorithm, LinUCB guarantees that
\begin{align*}
\mathbb E\left[\sum_{t=1}^{\tau}\sum_{j}(\hat r(c(t),j)-r(c(t),j))X_{j}(t)\right] \leq& 1 + \sqrt{8d\tau \beta_{\tau}(T^{-1}) \log \left(\frac{ d + \tau}{d}\right)},\\
\mathbb E\left[ \sum_{t=1}^{\tau} \sum_{j}(r(c(t),j) - \hat r(c(t),j))x^{\epsilon_t,*}_{c(t),j}\right] \leq& 1.
\end{align*}
\end{lemma}
Based on Lemmas \ref{lemma: drift in regret}, \ref{lemma: elsilon gap}, and \ref{lemma: bandits}, we conclude that
\begin{align*}
{\cal R}(\tau)\leq & \sum_{t=1}^{\tau} \frac{Kt(\epsilon_t+\epsilon_t^2)\mathbb I(\epsilon_t > \delta)}{V_t} + \sum_{t=1}^{\tau}\frac{\epsilon_t}{\delta} + \sum_{t=1}^{\tau}\frac{K(1+\epsilon_t^2)}{V_t}
+ 2 + \sqrt{8d\tau \beta_{\tau}(T^{-1}) \log \left(\frac{ d + \tau}{ d}\right)}.
\end{align*}
By choosing $\epsilon_t=K^{0.75}\sqrt{\frac{6}{t}}$ and $V_t=\delta K^{0.25}\sqrt{\frac{2t}{3}},$ we have
$$\sum_{t=1}^\tau \epsilon_t \leq 2K^{0.75}\sqrt{6\tau},~ \sum_{t=1}^\tau1/V_t\leq \frac{\sqrt{6\tau}}{K^{0.25}},~\text{and}~\sum_{t=1}^{\tau} \frac{Kt(\epsilon_t+\epsilon_t^2) \mathbb I(\epsilon_t > \delta)}{V_t}\leq \frac{60K^{3}}{\delta^3},$$ which yields the regret bound
\begin{align*}
\mathcal R(\tau) \leq& \frac{60K^{3}}{\delta^3}+ \frac{4\sqrt{6}K^{0.75}\sqrt{\tau}}{\delta} + 2 +\sqrt{8d\tau \beta_{\tau}(T^{-1}) \log \left(\frac{ d + \tau}{ d}\right)}.
\end{align*}
\subsection{Constraint Violation Bound}\label{sec: cv}
According to the dynamic defined in \eqref{eq:dual-update}, we have
\begin{align*}
Q^{(k)}(\tau+1) \geq \sum_{t=1}^\tau\sum_jW^{(k)}(c(t),j) X_{j}(t) + \sum_{t=1}^\tau \epsilon_t,
\end{align*} where we used the fact $Q^{(k)}(0)=0.$ This it implies the constraint violation can be bounded as follows:
\begin{align}
\mathcal V(\tau) \leq& \sum_{k}\left(\mathbb E[Q^{(k)}(\tau+1)] - \sum_{t=1}^{\tau} \epsilon_t\right)^{+}. \label{eq:cv-q}
\end{align}
Next, we introduce a lemma on the upper bound of $Q^{k}(\tau).$ Define $\tau'$ the first time such that $\epsilon_{\tau'} \leq \delta/2,$ that is, $\epsilon_\tau > \delta/2, \forall 1\leq \tau < \tau'.$ Note that $Q^{(k)}(\tau') \leq \sum_{t=1}^{\tau'}(1+\epsilon_t)$ because $Q^{(k)}$ can increase by at most $(1+\epsilon_t)$ during each round.
\begin{lemma}\label{lemma: Q bound}
For any time $\tau\in [T]$ such that $\tau \geq \tau',$ i.e., $\epsilon_\tau \leq \delta/2,$ we have
\begin{align*}
\mathbb E\left[Q^{(k)}(\tau)\right]\leq \sqrt{K}\left(\frac{48K^2}{\delta}\log\left(\frac{16K}{\delta}\right)+2K+\frac{4(V_\tau + K(1+\epsilon_\tau^2))}{\delta} + \tau' + \sqrt{K}\sum_{t=1}^{\tau'}\epsilon_t\right).
\end{align*}
\end{lemma}
Based on our choices of $\epsilon_t=K^{0.75}\sqrt{\frac{6}{t}}$ and $V_t=\delta K^{0.25}\sqrt{\frac{2t}{3}},$ we obtain
$$\tau' = \frac{24K^{1.5}}{\delta^2} ~\text{and}~\sum_{t=1}^\tau \frac{\epsilon_t}{\sqrt{K}} - \frac{4V_{\tau+1}}{\delta}\geq \sqrt{\tau} - 6K.$$
Note $\mathcal V(\tau) \leq K\tau'$ for any $\tau \leq \tau'.$ Combine Lemma \ref{lemma: Q bound} into \eqref{eq:cv-q}, we conclude
\begin{align*}
\mathcal V(\tau)
\leq& K^{1.5}\left(\frac{48K^2}{\delta}\log\left(\frac{16}{\delta}\right) + \frac{24K^{1.5}}{\delta^2}+\frac{30K^{1.5}}{\delta}+ 8K - \sqrt{\tau} \right)^{+}.
\end{align*}
\section{Numerical Evaluations}
In this section, we present numerical evaluations of the proposed algorithm, including 1) the constrained multi-armed bandit (MAB) example studied in \cite{PacGhaBar_20}; 2) a constrained linear bandit example based on a healthcare dataset on inpatient flow routing.
\noindent{\bf The Constrained MAB Example in \cite{PacGhaBar_20}:} As acknowledged in \cite{PacGhaBar_20}, the proposed algorithm OPA in \cite{PacGhaBar_20} suffers from high computational complexity for linear bandits so they evaluated the performance of their algorithm with classical multi-armed bandits (MAB), for which a computationally efficient algorithm, called OPA, is proposed (Lemma 5 in \cite{PacGhaBar_20}). Therefore, we compared our algorithm with OPA by considering the MAB example in \cite{PacGhaBar_20} with $K = 4$-arms where the reward and cost distributions are Bernoulli with means $\bar r = (0.1, 0.2, 0.4, 0.7)$ and $\bar c = (0, 0.4, 0.5, 0.2)$ and the total cost in each round should not exceed $0.5.$ In particular, we set parameters of our algorithms with $V_t=\sqrt{t}, \epsilon_t = 6.0/\sqrt{t}$ and UCB bonus terms $\sqrt{\log t/N(k)}$ for arm-$k$ at time $t.$
The results are presented in Figure \ref{fig:com with OPA}, where we can see that our algorithm has significant lower regret than that under OPA while the cost constraints are satisfied under both algorithms.
\begin{figure}[H]
\centering
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/compare_ucb_factor_onepfive_alpha_zeropfive_regret_trajectory_tau_zeropfive.png}
\caption{Regret}
\label{fig:first}
\end{subfigure}
\hfill
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/compare_ucb_factor_onepfive_alpha_zeropfive_reward_trajectory_tau_zeropfive.png}
\caption{Reward}
\label{fig:second}
\end{subfigure}
\hfill
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/compare_ucb_factor_onepfive_alpha_zeropfive_cost_trajectory_tau_zeropfive.png}
\caption{Cost}
\label{fig:third}
\end{subfigure}
\caption{Our Algorithm versus OPA in \cite{PacGhaBar_20}.}
\label{fig:com with OPA}
\end{figure}
\noindent{\bf Constrained Linear Bandits for Inpatient Flow Routing:} We also evaluated our algorithm for inpatient flow routing on a real-world dataset, where incoming patients have different features (context) such as age, gender, medical history, etc, and incur different amounts of ``rewards'' when being assigned to different wards (actions). The rewards are different because the levels of care provided by different wards match with the patients' needs differently. In this evaluation, we measured the reward via the avoided 30-day readmission penalty, i.e., a reward is collected if the patient is not readmitted to the hospital within 30 days since being discharged. Calibrating from the data we have,
we considered three types of constraints: capacity, fairness, and resource.
After normalizing, the capacity constraint for each ward is $[0.2, 0.2, 0.175, 0.175, 0.175, 0.175],$ the fairness requirement is $[0.175,0.175,0.15,0.15,0.125,0.125],$ and the nursing resource constraint is $[0.1875,0.1875,0.1875,0.1875,0.1875,0.1875]$, where each patient consumes ``one" unit of resources after being assigned to a ward. We note that these constraints are strict since the hospital capacity is highly constrained.
In the experiment, we set scaling parameter $V_t = 4\sqrt{t}, \epsilon_t = 1/\sqrt{t}$ in our algorithm, and we ran various learning horizons $T = [50^2, 100^2, 150^2, 200^2, 250^2].$
The regrets and constraint violations at the end of the horizon are plotted in Figure \ref{fig:healthcare-regret} and \ref{fig:healthcare-violation}, which show that our algorithm achieves a low regret and zero violation.
To further evaluate the anytime constraint violation, we plotted a representative trajectory with $T=10,000$ in Figure \ref{fig:healthcare-trace} to see how the violations evolve and if zero constraint violation can be achieved after a constant number of steps. The results show that the constraint violations decrease to zero quickly and $\tau'$ for capacity, fairness and resource constraints are $(32,50,54),$ respectively, which confirms our theoretical results on zero constraint violation.
\begin{figure}[H]
\centering
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/neurips_regret_comparison_lin_and_classical.png}
\caption{Regret}
\label{fig:healthcare-regret}
\end{subfigure}
\hfill
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/neurips_violation_comparison_lin_and_classical.png}
\caption{Violation}
\label{fig:healthcare-violation}
\end{subfigure}
\hfill
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/neurips_resource_violation_tv_and_fixed.png}
\caption{Trajectory of Violation}
\label{fig:healthcare-trace}
\end{subfigure}
\caption{Regret and Violation in Inpatient Flow Routing.}
\label{fig:healthcare}
\end{figure}
\section{Conclusions and Extensions}
\label{sec:conclusion}
In this paper, we study stochastic linear bandits with general anytime cumulative constraints. We develop a pessimistic-optimistic algorithm that is computationally efficient and has strong guarantees on both regret and constraint violations. We conclude this paper by mentioning an extension on the case where the cost signals $W^{(k)}(c(t),A(t))$ are revealed after action $A(t)$ is taken. However, we assume the costs can be linearly parameterized as in \cite{PacGhaBar_20}.
\noindent{\bf Linear Cost Functions:} In this case, we assume that the costs $,W^{(k)}(c(t),A(t)),$ are not available before the action is taken but the costs are linear as in \cite{PacGhaBar_20}. The learner observes cost \emph{after} taking action $A(t),$ $$W(c(t), A(t))= \langle \mu_{*}, \psi(c, j) \rangle + \xi(t),$$ where $\psi(c,j)\in \mathbb R^d$ is a $d$-dimensional feature vector for (context, action) pair $(c,j),$ $\mu_{*} \in \mathbb R^d$ is an unknown underlying vector to be learned, and $\xi(t)$ is a zero-mean random variable. In this case, we also obtain an estimate $\check W(c(t), j)$ of $W(c(t), j)$ with LinUCB and replace $W(c(t), j)$ with $\check W(c(t), j)$ in the steps of MaxValue and Dual Update in the Pessimisitic-Optimistic Algorithm. This variation of Pessimisitic-Optimistic Algorithm has a similar computational complexity as our main algorithm, and it can provide similar regret and constraint violation guarantees:
\begin{theorem}[Informal]
With linear costs as in \cite{PacGhaBar_20}, a variation of our algorithm
achieves ${\cal R}(\tau)=\tilde{\cal O}\left(\frac{d}{\delta}\sqrt{\tau}+\frac{d^4}{\delta^3}\right)$ for any $\tau\in [T]$ and ${\cal V}(\tau)=0$ for $\tau\geq \tau''=O\left(\frac{d^2}{\delta^4}\log^2 T\right).$\label{thm:linearcost}
\end{theorem}
The formal statement and the detailed analysis are in Appendix \ref{app:linearcost}.
\newpage
\bibliographystyle{plain}
|
\section{Introduction}
Given an abstract (and possibly very complicated) manifold, a
natural question is whether it is possible to see
it as a subset of a simpler space. In the real category, a fundamental
theorem by Nash states that this is always possible, and in a very strong
sense: every Riemannian manifold can be isometrically embedded in some $\mathbb{R}^N$.
When moving to the complex category, we can then ask
the following natural question: is it possible to embed any complex
manifold in some $\mathbb{C}^N$, by means of a holomorphic map?
We call \emph{Stein} a manifold for which the above holds true. This
time, the rigidity of holomorphic functions readily provides negative
examples: for instance, the maximum principle implies that any holomorphic
map on a compact complex manifold must be constant, and thus the manifold
cannot be Stein. A central question is then to understand
when
a given complex manifold is
Stein.
More specifically, given a dimension $n$,
one would like to
understand the \emph{obstructions} for an $n$-dimensional manifold to be Stein.
A major advance in this direction was provided
by
Grauert \cite{Gra}:
a complex manifold is Stein if and only if it admits a
$\mathcal{C}^2$
strictly plurisubharmonic
(psh for short) exhaustion function.
The $\mathcal{C}^2$ assumption was relaxed to $\mathcal{C}^0$ by
Narasimhan \cites{Na1,Na2}.
In view of these results,
it is natural to tackle the question
by studying the positive cone
${\rm Psh}_e^0(X)$
of all continuous
psh exhaustion functions on
$X$ (or more generally the cone
${\rm Psh}_e^k (X) := {\rm Psh}_e^0(X) \cap \mathcal{C}^k$ for some
$k \in [0,\infty]$,
and in particular to find obstructions for them to be strictly psh.
As a rough idea, such obstructions must correspond to the presence of some sets in $X$ along
which all continuous
psh functions must necessarily be pluriharmonic. As a prototypical example,
the blow-up of a point and its corresponding exceptional divisor give precisely this kind of obstruction.
A precise study of this kind of phenomena was started
by Slodkowski
and Tomassini
in \cite{ST} in the setting of \emph{weakly complete complex manifolds}, i.e., manifolds admitting a
continuous
psh exhaustion function.
A crucial definition
is the following:
for $k\in [0,\infty]$
the \emph{minimal kernel} of a manifold $X$
(with respect to ${\rm Psh}^k_e$) is
\begin{equation}\label{eq:sigmaX}
\Sigma_X^k :=\{x\in X\ :\ i\partial\oli{\de} u\textrm{ is degenerate at }x\, \forall u\in {\rm Psh}^k_e(X)\},
\end{equation}
i.e., the subset of $X$ where
no element of ${\rm Psh}^k_e$
can be strictly psh. A key
result of \cite{ST} is that,
whenever ${\rm Psh}_e^k (X)$ is not empty,
there actually
exists a function $\phi_0\in {\rm Psh}^k_e (X)$
(called \emph{minimal})
which fails to be strictly psh \emph{precisely}
on
the minimal kernel
$\Sigma_X^k$.
Moreover, the minimal kernels are
\emph{local maximum sets} (see Definition \ref{defi_localmax}).
Some finer properties are also established (some requiring
at least the $\mathcal{C}^2$ regularity, see for instance \cite{ST}*{Theorem 3.9}).
Observe that the $\Sigma_X^k$'s are increasing in $k$, but it is not known whether
equalities should occur
in general,
see for instance
\cite[Section 5.10]{Sl_pseudo}.
In \cites{mst,mst2}, the second author, Slodkowski, and Tomassini
showed that, if $X$ has complex dimension $2$
and ${\rm Psh}_e^\infty$
contains at least one real analytic function,
the minimal kernel is either a union of
countably many compact (and negative) curves
or equal to the whole manifold, by
giving a full classification of the possible structures that such a manifold can present.
An important point here is that, although in general the
minimal kernel does not have a priori an analytic structure,
however its intersection with any level of a psh exhaustion function
does (at least in dimension $2$).
In \cite{Sib}, Sibony introduced the notion of \emph{Levi current} (see Definition \ref{def_Levic}),
which is related to the (non-)existence of strictly psh functions on a complex manifold and thus to
the problem of determining whether a given
manifold is Stein, see also
\cite{OS, Sib_pf, Sib_pseudo}.
Extremal Levi-currents are supported
on sets where all
continuous psh
functions are constant.
In the case of infinitesimally homogeneous manifolds, a
foliation is constructed and linked to the obstructions to Steinness.
Our goal here is
to
compare these two approaches,
and in particular to
use the notion of Levi current on $X$
to study the analytic structure of
the minimal kernels $\Sigma_X^k$.
In order to do this, let
us denote by ${\rm Psh}^k$, for $0\leq k\leq \infty$,
the cone of $\mathcal{C}^k$ psh function on $X$ and
define the \emph{distribution}
$\mathcal E^k$ in $TX$ as
\begin{equation}\label{eq:e}
\mathcal{E}^{k}:=\{(x,v)\in TX\ :\ (d\varphi)_x(v)=0\ \forall\varphi \in {\rm Psh}^{k} (X) \}.
\end{equation}
A \emph{distribution} is a subset of $TX$ whose intersection with
$T_xX$ is a (real) vector subspace of the latter for every $x\in X$.
In general, $\mathcal{E}^k$ will not be a subbundle of $TX$, as $\dim\mathcal{E}^k_x$ is not constant;
however it is a closed subset of $TX$, hence the function $x\mapsto \dim \mathcal{E}^k_x$ is upper semicontinuous.
The following is our main result.
\begin{teorema}\label{t:main}
Let $X$ be a weakly complete complex
manifold
and $T$ a Levi current on $X$.
Denote by $\Sigma_X^k$ the minimal kernels of $X$ and by $F$ the union
of the supports of all Levi currents on $X$.
Then $F \subseteq \Sigma_X^k$ for all $k\geq 0$
and
\begin{enumerate}
\item\label{item_t_locmax} if $T$ has compact support $K_T$, then $K_T$ is a local maximum set;
\item\label{item_t_KY} if
$K$
is a local maximum set, then
there exists a Levi current
supported on
$K$. In particular, $K \cap F \neq \emptyset$.
\end{enumerate}
Moreover, if $X$ is a surface,
$\phi \in \mathcal{C}^k$
a psh exhaustion function, and
$Y$
a regular connected component of a level set $\{\phi=c\}$,
\begin{enumerate}
\setcounter{enumi}{2}
\item\label{item_t_curves} if $4\leq k\leq \infty$ and $U\subseteq \Sigma_X^k$
is an open set in $Y$ and there exists $x\in U$
such that $\dim\mathcal{E}^k_x=2$,
then $X$
is a union of compact complex curves. In particular, $F=\Sigma_X^k =X$;
\item\label{item_t_Y} if
$2 \leq k \leq \infty$ and $Y \subseteq \Sigma_X^k$, then there exists $c'<c$
such that
the connected component of $\phi^{-1} ([c',c])$ containing $Y$ is contained in $F$.
In particular, $Y \subseteq F$.
\end{enumerate}
\end{teorema}
Moreover, given \emph{any}
psh function $u\in {\rm Psh}^0 (X)$,
any Levi current $T$ can be naturally disintegrated as $T= \int T_c \, d\mu$, where
$\mu$ is a positive measure on ${\mathbb R}$ and $T_c$ is a Levi current supported on $\{u=c\}$, see Corollary
\ref{cor_disint}.
This in particular gives examples of Levi currents for which Item \eqref{item_t_locmax} applies.
Notice that, whenever two level sets of $u,v\in {\rm Psh}^0(X)$
do not coincide, this allows to further refine the description of the extremal Levi current. This
motivates the definition \eqref{eq:e} of the
distributions $\mathcal E^k$.
It follows from \cite{mst,ST} that, when $k\geq 2$,
for any level set $Y$ of an exhaustion function $\phi\in {\rm Psh}_e^0$,
the set
$\Sigma_X^k\cap Y$ is a local maximum set (or empty),
see Lemma \ref{l:intersection}. Hence Item \eqref{item_t_KY} applies for instance to such sets.
Finally, the manifold $X= {\mathbb C} \times {\mathbb P}^1$ provides an example where the Items \eqref{item_t_curves} and \eqref{item_t_Y} apply.
\begin{rem}
It would
be interesting to know if the equality holds in Item \eqref{item_t_KY} (and in particular for the
intersections between levels sets of a psh exhaustion function and the minimal kernel). Namely if, for \emph{any}
point in a local maximum set $K$, there exists a Levi current $T$ such that $x \in {\rm spt} T$.
\end{rem}
\medskip
The paper is organized as follows.
In Section \ref{s:levi} we recall the definition of Levi currents and
the properties that we will need in the sequel.
In Section \ref{section_locmax}
we prove
Items \eqref{item_t_locmax}
and \eqref{item_t_KY}
of Theorem \ref{t:main}. The first item
is established for
$\Sigma_X^k \cap \{\phi_0=c_0\}$
(where $c_0$ is an attained value for a
continuous
minimal function $\phi_0$ and $k\geq 2$)
in \cite{ST}*{Theorem 3.6},
and
is actually a consequence of
\cite{Sib_pf}*{Theorem 3.1},
where it is proved through an integration by parts, see also
\cite{Sib}*{Proposition 4.2}
for an analogous statement
for $F$.
We give here a different proof by means of
a characterization of the
local maximum property
due to Slodkowski
\cite{Sl}
which allows to
bypass the use of Brebermann functions and
Jensen measures
as in \cite{ST}.
In Section \ref{s:kernels} we study the relation between
the minimal kernels $\Sigma_X^k$ and
distributions in
the tangent bundle $TX$ given by
directions satisfying some degeneracy
condition. This leads to the proof
of Item \eqref{item_t_curves}.
The proof
of Theorem \ref{t:main} is completed in Section \ref{s:proof_main}, where
we establish Item \eqref{item_t_Y}.
In Section \ref{s:analytic} we
consider
the case where
$X$ is a surface and the exhaustion
function in Theorem \ref{t:main} can be chosen analytic. By
exploting the main result in \cite{mst},
we deduce
a classification of Levi currents in this case.
\subsection*{Acknowledgements}
The authors would like to thank Nessim Sibony
for the references
and for very useful
comments that also helped to improve Item \eqref{item_t_KY} in the main theorem,
and Zbigniew Slodkowski
for
very
helpful remarks on a preliminary version of this paper.
They also would like to thank Vi\^{e}t-Anh Nguy\^{e}n and Giuseppe
Tomassini
for useful observations and discussions.
This work was supported by the
Research in Pairs 2019 program of the CIRM (Centro Internazionale di Ricerca Matematica), Trento
and the FBK (Fondazione Bruno Kessler).
The authors would like to warmly thank CIRM-FBK for their support,
hospitality, and
excellent work conditions.
This project also
received funding from the
I-SITE ULNE (ANR-16-IDEX-0004 ULNE),
the
LabEx CEMPI (ANR-11-LABX-0007-01)
and from the CNRS
program PEPS JCJC 2019.
\section{Levi currents on complex manifolds}\label{s:levi}
In this section
we recall the definition of Levi current and give the properties that we need in the sequel.
These results are essentially contained in \cite[Section 4]{Sib} and \cite[Section 3]{Sib_pseudo},
we sketch here the proofs for completeness.
We let
$X$ be any complex manifold and we denote by ${\rm Psh}^0(X)$ the space
of continuous plurisubharmonic functions on $X$.
\begin{defin}[Sibony \cite{Sib}]
\label{def_Levic}A current $T$ on $X$ is called \emph{Levi current} if
\begin{enumerate}
\item $T$ is non zero;
\item $T$ is of bidimension $(1,1)$;
\item $T$ is positive;
\item $i\partial\oli{\de} T=0$;
\item $T\wedge i\partial\oli{\de} u=0$ for all $u\in {\rm Psh}^0(X)$.
\end{enumerate}
\end{defin}
A Levi current $T$ is \emph{extremal}
if $T=T_1 = T_2$ whenever $T=(T_1+T_2)/2$ for $T_1,T_2$ Levi currents.
\begin{lemma}\label{lemma_vanish} Take $u\in{\rm Psh}^0 (X)$ and
let
$T$
be
a Levi current. The currents
$$T\wedge \partial u\;,\quad \quad T\wedge\oli{\de} u\;,\quad \mbox{and} \quad \quad T\wedge \partial u\wedge \oli{\de} u$$
are all well defined and
vanish identically on $X$.\end{lemma}
\begin{proof}
The currents in the statement are well defined when $u$ is smooth, and the arguments from
\cite{DS}*{Section 2} and
\cite{Sib}*{Section 4} prove the good definition for $u \in {\rm Psh}^0(X)$.
If $u\in {\rm Psh}^0(X)$, then also $\exp(u)\in{\rm Psh}^0(X)$.
So, by Definition
\ref{def_Levic}
of Levi current,
$T\wedge i\partial\oli{\de} \exp(u)$ vanishes identically. Hence, we have
$$0=\exp(u)T\wedge i\partial\oli{\de} u
+i\exp(u)T\wedge\partial u\wedge\oli{\de} u\;.$$
Given that $T\wedge i \partial\oli{\de} u=0$, we conclude that
$T\wedge \partial u \wedge\oli{\de} u=0$.
This gives the last identity.
We prove now the first one, the proof for the second one is similar.
Since $T$ is positive,
for any $(0,1)$-form $\alpha$
by Cauchy-Schwarz's
inequality we get
\[
|\langle T, \partial u \wedge \alpha\rangle|^2 \leq
c
\langle T, \partial u \wedge \bar \partial u\rangle
\cdot
\langle T, \alpha \wedge \bar \alpha\rangle
\]
where $c$ is a constant independent of $T,u$, and $\alpha$. Since the
first factor in the RHS is zero by the first part of the proof, the assertion follows.
\end{proof}
By a standard disintegration procedure, we obtain the following consequence.
\begin{corol}\label{cor_disint}Suppose $T$ is a Levi current and
$u\in {\rm Psh}^0(X)$; then there exists a measure $\mu$ on ${\mathbb R}$ and a collection of currents $T_c$, $c\in{\mathbb R}$ such that
\begin{itemize}
\item $T_c$ is supported on $Y_c=\{x\in X\ :\ u(x)=c\}$ for all $c\in{\mathbb R}$;
\item $T_c$ is non zero for $\mu$-almost every $c\in{\mathbb R}$;
\item whenever $T_c\neq 0$, $T_c$ is a Levi current;
\item for every $2$-form $\alpha$ on $X$ we have
$$\langle T,\alpha\rangle=\int_{{\mathbb R}}\langle T_c,\alpha\rangle d\mu(c)\;.$$
\end{itemize}
Moreover, if $u\in{\rm Psh}^1(X)$ and $c$ is a regular value for $u$,
then $T_c=j_*S_c$, where
$j$ is the inclusion of $Y_c$ in $X$ and $S_c$ a current on the real manifold $Y_c$.
\end{corol}
Notice that $T_c$ needs not be extremal,
as it is easily seen considering $X={\mathbb C} \times {\mathbb P}^1$.
\begin{rem}
Suppose now that $X$ is weakly complete and let
$\phi$ be a psh exhaustion function.
By Corollary \ref{cor_disint}, every Levi current is
obtained by averaging Levi currents which are supported on the level sets
of $\phi$; as the latter is an exhaustion of $X$, its level sets are compact, so
every Levi current on $X$ is an integral average of compactly supported Levi
currents, i.e., positive currents of bidimension $(1,1)$ which
are $\partial\oli{\de}$-closed and compactly supported.\end{rem}
\begin{corol}\label{corol_spsh}
If $T$ is a Levi current
and $u\in {\rm Psh}^1(X)$, then the vector field associated
to $T$ is tangent to the kernel of $\partial u\wedge\oli{\de} u$,
whenever the latter is non-zero
(and the former is defined).
Moreover, if there exists $v\in{\rm Psh}^0 (X)$
which is strictly plurisubharmonic at a point $x\in X$, then $x\not\in{\rm spt} T$
for any Levi current $T$.\end{corol}
\begin{proof}
The first statement is equivalent to $T\wedge\partial u\wedge\oli{\de} u=0$, hence follows
from Lemma \ref{lemma_vanish}.
Suppose now that
we have $v\in{\rm Psh}^0 (X)$ which is strictly psh at $x$
and a Levi current $T$.
First, by Richberg \cite[Satz 4.3]{Ri}
we can assume that $v$ is $\mathcal{C}^\infty$ and strictly psh near $x$.
Then, if $\rho\in\mathcal{C}^\infty_0(X)$ is supported in a neighbourhood
of $x$ where $v$ is strictly psh and $\|\rho\|_{\mathcal{C}^2}$ is small enough,
then also $v+\rho$ is
psh.
In particular, we can choose $\rho_i, 1\leq i \leq 2n$ such
that the
$\ker(\partial(v+\rho_i)\wedge\oli{\de}(v+\rho_i))_x$
are independent (over $\mathbb R$).
This property holds true in a neighbourhood of $x$.
Hence, as the vector field associated to $T$
(on a full measure subset of the support $T$ for the mass measure)
should belong to all
these subspaces, the only possibility
is that $x\not\in{\rm spt} T$. This concludes the proof.\end{proof}
\begin{lemma}\label{lmm_spsh}
Let $T$ be a Levi current such that $K={\rm spt} T$ is compact. If $u$
is defined and plurisubharmonic in an open neighbourhood $V$ of $K$
and strictly plurisubharmonic at $x\in V$,
then $x\not\in{\rm spt} T$.\end{lemma}
\begin{proof}
Let
$V'\Subset V$ be an open neighbourhood of $K$ containing $x$.
Let $\chi\in\mathcal{C}^\infty_0(V)$ be such that $\chi\vert_{V'}\equiv 1$, then $\chi u$ is
defined on $X$ and
psh on $V'$.
As ${\rm spt} T\subseteq V'$,
also ${\rm spt}(T\wedge i\partial\oli{\de} u)\subseteq V'$, so $T\wedge i\partial\oli{\de} u$
is positive; moreover, as $T$ is a Levi current,
we have $i \partial\oli{\de} T=0$, hence
$$0=\langle i\partial\oli{\de} T, u\rangle=\langle T, i\partial\oli{\de} u\rangle=\langle T\wedge i \partial\oli{\de} u, 1\rangle.$$
Therefore, $T\wedge i\partial\oli{\de} u=0$ as a (positive) measure.
Since $(i\partial\oli{\de} u)_x>0$,
this happens in a neighbourhood of $x$, so $T\wedge i\partial\oli{\de} u$ is strictly
positive in a neighbourhood of $x$ unless $T$ is zero there. This gives
$x\not\in{\rm spt} T$ and concludes the proof.
\end{proof}
\begin{lemma}\label{lmm_cptspp}
Suppose that a current $T$ satisfies requests $1-4$ of Definition \ref{def_Levic}
and $T$ has compact support.
Then $T$ is a Levi current.\end{lemma}
\begin{proof}
Given that $T$ is compactly supported, so are
$uT$, $T\wedge \partial u$, $T\wedge \oli{\de} u$,
and $T\wedge i \partial\oli{\de} u$
for all $u\in {\rm Psh}^0(X)$. Moreover, as $T$ is positive and $u$ is psh,
$T\wedge i \partial\oli{\de} u$ is a positive measure on $X$; therefore, it is zero if and only if
$\langle T\wedge i \partial\oli{\de} u, 1\rangle=0$.
Notice that, by Stokes' theorem, we have $\langle i \partial\oli{\de} (uT), 1=0\rangle$, hence
$$0=-\langle i\partial\oli{\de} u \wedge T, 1\rangle + \langle ui\partial\oli{\de} T, 1\rangle +\langle i\partial(\oli{\de} u\wedge T), 1\rangle-\langle i\oli{\de}(\partial u\wedge T), 1\rangle.$$
We have $i\partial\oli{\de} T=0$ by hypothesis, while $\langle i\partial(\oli{\de} u\wedge T), 1\rangle$ and $\langle i\oli{\de}(\partial u\wedge T), 1\rangle$ vanish by another application of Stokes' theorem. Therefore
$T\wedge i \partial\oli{\de} u=0$,
that is, $T$ is a Levi current.
\end{proof}
\section{Local maximum sets}\label{section_locmax}
We establish here Items
\eqref{item_t_locmax}
and \eqref{item_t_KY}
of Theorem \ref{t:main}.
We recall the following definition, see also \cite{ST}*{Section 2}
and \cite{R}.
\begin{defin}\label{defi_localmax}
Let $X$ be a complex manifold and
$K \subset X$ be compact. We say that
\emph{$K$ is a local maximum set} if
every $x\in K$ has a neighbourhood $U$
with the following property: for every compact set $K' \subset U$ and every
function $\psi$ which is strictly psh in a neighbourhood of $K'$, we have
\[
\max_{K \cap K'}\psi = \max_{K \cap bK'} \psi.
\]
\end{defin}
\begin{propos}\label{lemma:levi_local_max}
Suppose that $X$ is weakly complete. If
$T$ is a Levi current with compact support, then ${\rm spt} T$ is a local maximum set.
\end{propos}
\begin{proof}
Suppose that $K:={\rm spt} T$ is not a local maximum set.
By \cite{Sl}*{Proposition 2.3}
there exist
$x\in K$,
a neighbourhood $B$ of $x$, with local
coordinates $z$ with origin in $x$,
such that $B\equiv\{z\in{\mathbb C}^n\ :\ \|z\|<1\}$, and
$\psi\colon B \to {\mathbb R}$
strictly psh
with $\psi(x)=0$ and $\psi(y)\leq -\epsilon\|z(y)\|^2$ for all $y\in K\cap B$.
Up to replacing $\psi$ by an element of a continuous approximating sequence, we can directly
assume that $\psi$ is continuous.
By taking a possibly smaller ball a $x$, we can also assume that
$-\epsilon\|z(y)\|^2 -\epsilon/8 \leq \psi(y)$ on $K$.
Set
\[A:= \{ y \in B \colon \|z(y)\|^2 > 3/4\}
\quad \mbox{ and } \quad
V=\{y\in B\ :\ |\psi(y)+\epsilon\|z(y)\|^2|<\epsilon/4\}.
\]
By the continuity of $\psi$ and the bounds above,
$V$ is an open subset of $B$ containing $K$.
Consider $u\in{\rm Psh}^0(X)$
such that $u(x)=-\epsilon/4$ and $\sup_B|u|< \epsilon/2$
(this function exists because of the assumption on $X$).
Since $K\cap B\subset V$, there also exists
$\chi\in\mathcal{C}^\infty_0(X)$ be
such that $\chi\vert_K\equiv 1$
and ${\rm spt}\chi\cap B\subset V$.
Define the function
$v\colon X \to {\mathbb R}$ as
\[
v=\begin{cases}
\chi \max\{u,\psi\} & \mbox{ on } B, \\
\chi u & \mbox{ on } X\setminus B.
\end{cases}
\]
We claim that $v= \chi u$ on $A$.
Indeed,
for every $p\in {\rm spt} \chi \cap A$, we have
that $p \in V$, and so
$\psi(p)<-3\epsilon/4+\epsilon/4=-\epsilon/2$. Hence,
$\psi(p)< u(p)$ and $v(p)= ( \chi u)(p)$.
By construction, $v$ coincides with
$\psi$ in a neighbourhood of $x$.
It follows that $v$
is psh in a neighbourhood of $K$ and
strictly psh in $x$. Therefore, we have $x\not\in{\rm spt} T$
by Lemma \ref{lmm_spsh}. This gives a contradiction with the choice of $x\in K$ and
completes the proof.
\end{proof}
\begin{propos}\label{propos-locmax-supp-new}
Let $K\subset X$ be a local maximum set.
There exists a Levi current $T$ such that ${\rm spt} T \subseteq K$.
\end{propos}
\begin{proof}
By \cite[Theorem 3.1]{Sib_pseudo} (see also \cite[Section 4]{Sib})
and Lemma \ref{lmm_cptspp}, if there are no Levi currents supported on $K$, there exists a
smooth strictly
psh function $u$
on some open neighbourhood $U$ of $K$. By slightly perturbing $u$,
for every $x_0 \in K$ we can construct
$2n$ continuous strictly psh functions on some neighbourhood
$K \subset U' \subseteq U$
such that $du_1, \dots, du_{2n}$ are linearly independent at $x_0$. This implies that,
in a neighbourhood of $x_0$, we have
\[
\{u_1 = u_1 (x_0) \} \cap \dots \{u_{2n} = u_{2n}(x_0)\}= \{x_0\}.
\]
By \cite[Corollary 1.11]{Sl_pseudo} and \cite[Theorem 4.2]{Sl},
for every family
of continuous psh functions on $U'$
there
exists a local maximum set $K'\subseteq K$ with the property that all functions of the family
are constant on $K'$. Choosing a point $x_0 \in K'$, the previous paragraph
gives that $x_0$ is isolated in $K'$.
This is a contradiction, and the proof is complete.
\end{proof}
We conclude the section with the following result, that we will need to prove
Item
\eqref{item_t_curves}
of Theorem \ref{t:main}.
\begin{lemma}\label{l:intersection}
Let $X$ be a weakly complete complex surface and $Y$ a regular level for a
$\mathcal{C}^0$
exhaustion function $\phi$. Then,
for all $k\geq 2$,
$\Sigma_X^k \cap Y$ is a local maximum set and, for all
local maximum sets $K\subseteq \Sigma_X^k$,
$K$
is
foliated by holomorphic discs, i.e., it is
locally a union of disjoint holomorphic discs.
\end{lemma}
\begin{proof}
The intersection $\Sigma_X^k \cap Y$ is a local maximum set
by \cite{mst}*{Theorem 3.2}. As observed in
\cite{mst}*{Proposition 3.5}, the proof of the
lemma is then essentially
given in
\cite{ST}*{Lemma 4.1}, see
in particular
the Assertion in the proof of that lemma.
\end{proof}
We point out that \cite[Lemma 4.1]{ST}
relies on a result by Shcherbina \cite{Sh},
which holds true only in dimension 2; this is the reason for
restricting ourselves to the case of surfaces. it would be interesting
to prove (or disprove) a similar statement in higher dimension.
\section{Kernels and tangent directions}
\label{s:kernels}
In this section we let $X$ be a
weakly complete complex manifold of
dimension $n$ and
assume that
${\rm Psh}^k(X)$ contains at least one exhaustion function $\phi:X\to{\mathbb R}$ for some $2\leq k\leq \infty$.
Recall that
the minimal kernel $\Sigma_X^k$
of $X$
is
defined as in \eqref{eq:sigmaX}
and the distribution $\mathcal E^k$ of $TX$
as in \eqref{eq:e}.
We consider further the distribution
$\mathcal{S}^k$ of $TX$ given by
\begin{equation} \label{eq_s}
\mathcal{S}^k:=\{(x,\xi)\in TX\ :\ \xi\in T_xX,\ (i\partial\oli{\de} u)_x(\xi,\xi)=0\, \forall u\in {\rm Psh}^k (X)\}\; .
\end{equation}
Similar objects have already appeared in relation to the study of the Levi problem, see for instance
\cite{Hir} in the case of homogeneous manifolds and \cite{ST,HST}.
We also
set
\[
E^k_{\ell} := \{x\in X \colon\dim \mathcal{E}^k_x \geq \ell \}
\quad \mbox{ and } \quad
S^k_1 := \{x\in X \colon\dim \mathcal{S}^k_x \geq 1\}.
\]
By definition,
$E^k_{\ell} \subseteq E^k_{\ell-1}$ and $E^k_{\ell}$ is closed in $E^k_{\ell-1}$ for all $\ell \geq 1$.
Observe moreover that $\mathcal S^k$ is a \emph{complex} distribution.
\begin{rem}
Let $T$ be a Levi current. Then, for almost every point of the support
of $T$ (with respect to the mass measure), the
vector field associated to $T$ at $x$ belongs to the fibre
$\mathcal S^k_x$ of $\mathcal S^k$ at $x$.
\end{rem}
\begin{propos}\label{prop_es}
We have $\mathcal{S}^k \subseteq \mathcal {E}^k$, and
$S^k_1 = E^k_2 = E^k_1 = \Sigma_X^k$.
\end{propos}
\begin{proof}
It follows
from the definition
of $\mathcal E^k$ that $E^k_1 = E^k_2$. Moreover,
if $(x,v)\not\in\mathcal{E}^k$,
there exists $\psi\in{\rm Psh}^{k}(X)$
such that $(d\psi)_x(v)\neq 0$; then
$$i\partial\oli{\de} \exp(\psi)=\exp(\psi)i\partial\oli{\de}\psi+i \exp(\psi)\partial\psi\wedge\oli{\de}\psi\geq 0,$$
which implies that
$(i\partial\oli{\de}\exp(\psi))_x(v,v)\geq \exp(\psi(x))|\partial\psi_x(v)|^2>0$ and so
$(x,v)\not\in \mathcal{S}^k$. It follows that $\mathcal S^k \subseteq \mathcal E^k$.
We now prove that $E^k_1 = \Sigma_X^k$.
If $x\not\in\Sigma_X^k$, then there exists
$\psi \in {\rm Psh}^k_e (X)$
which is strictly psh around $x$; therefore, given any
$\rho:X\to{\mathbb R}$ smooth with compact support near $x$, there exists $\epsilon>0$
such
that $\psi+\epsilon\rho$ is still psh.
So, we can construct
psh functions of class $\mathcal{C}^k$
whose
differentials span the tangent space at $x$, which implies that these differentials
do not
have any nontrivial common kernel in $T_xX$. So $\mathcal{E}_x^k=\{0\}$, hence
$E_1^k \subseteq \Sigma_X^k$.
On the other hand, if $\mathcal{E}^k_x=\{0\}$, given $v_1,\ldots,v_{2n}\in T_xX$
linearly independent, we can choose
psh functions
$\psi_{ij}$, $i,j=1,\ldots, 2n$
of class $\mathcal{C}^k$ and
such that
$(d\psi_{ij})_x(v_i+v_j)\neq 0$. Therefore, the function
$\psi:=\sum_{i,j=1}^{2n} \psi_{ij}^2$
has positive defined Levi form at $x$.
Adding to the exhaustion
function $\phi$ suitable multiples of $\psi$, we see that
$x\not\in\Sigma_X^k$.
This
gives $E_1^k \supseteq \Sigma_X^k$, hence $E_1^k=\Sigma_X^k$.
In order to
conclude, we need to prove that
$S_1^k\supseteq \Sigma_X^k$.
Take $x \in \Sigma_X^k$ and
suppose by contradiction
that, for every $v\in T_xX$ there is $\varphi_v:X\to{\mathbb R}$
which is
$\mathcal{C}^k$, psh, and such that $(dd^c\varphi_v)_x(v,v)>0$.
Then, as above,
we
can construct a $\mathcal{C}^k$ function $\psi$
which is strictly psh
at $x$. This gives the desired contradiction
and completes the proof.
\end{proof}
The following result
gives Item \eqref{item_t_curves} of our main Theorem
\ref{t:main}.
\begin{propos}\label{t:e2}
Let $X$ be a weakly complete complex surface, $\phi$ a
$\mathcal{C}^k$, $4\leq k\leq \infty$,
exhaustion psh function and
$Y$ a regular connected component of a level set
$\{\phi=c\}$
of $\phi$.
Suppose that $U\subseteq Y$
is an open set in $Y$ and $U\subseteq\Sigma_X^k$. If there exists $x\in U$
such that $\dim\mathcal{E}_x^k=2$,
then $X$
is a union of compact complex curves.
In particular, $\Sigma_X^k=F=X=E_2^k$,
and
$E_3^k$
is contained in a (possibly empty)
analytic subset of the
singular levels for $\phi$.
\end{propos}
We will need the
following theorem by Nishino, see \cite{Ni}*{Proposition 9 and Théorème II}
and \cite{MT}*{Section 2.2.1}.
\begin{teorema}[Nishino]
\label{t_nishino}
Let $X$ be a weakly complete or compact surface
that contains an uncountable family $\mathcal F$
of disjoint connected compact complex curves.
Then there exist a Riemann surface $R$ and a meromorphic map $h \colon X\to R$ with compact fibers.
\end{teorema}
\begin{proof}[Proof of Proposition \ref{t:e2}]
By Theorem \ref{t_nishino}, to prove the first assertion it is enough to show
that $X$ contains uncountably many disjoint compact complex curves.
Since $x\in U$ is such that $\dim\mathcal{E}_x^k=2$,
there exists $\psi \in {\rm Psh}_e^k (X)$
such that $(d\phi)_x$ and $(d\psi)_x$
are linearly independent;
hence,
the map $\psi\vert_Y:Y\to{\mathbb R}$
is not constant.
Since $k\geq 4$, by Sard's theorem
we can find regular values $b$ for $\psi$ arbitrarily close
to $b_0:=\psi(x)$, therefore the sets $C_b=\{y\in Y\ :\ \psi(y)=b\}$
intersect the open set $U\subseteq \Sigma_X^k$.
For any $y\in C_b\cap \Sigma_X^k$, by Proposition \ref{prop_es}
we have $T_yC_b=\mathcal{E}^k_y =\mathcal S^k_y$.
Therefore, $C_b\cap\Sigma_X^k$ is a complex curve, being a real, smooth
$2$-dimensional manifold with complex tangent space.
On the other hand, the set $C_b\setminus\Sigma_X^k$ is open in $C_b$. Let $z\in C_b$
be a boundary point (with respect to $C_b$);
as $z\in\Sigma_X^k$, by Lemma \ref{l:intersection}
there is a holomorphic disc $f:\mathbb{D}\to X$
such that $f(\mathbb{D})\subset Y$ and $f(0)=z$.
If $\zeta\in\mathbb{D}$ is close enough to $0$,
then, setting $w=f(\zeta)$, we have $w\in\Sigma_X^k$, and
$(d\phi)_w$ and $(d\psi)_w$ independent. This gives
$w\in E_2^k\setminus E_3^k$, which in turn implies that
$\mathcal{E}_w^k =T_wC_{\psi(w)}$. Therefore $f(\mathbb{D})$ coincides locally with a leaf $C_{b'}$. Hence $C_b$
is contained in $\Sigma_X^k$, so it is a compact complex curve.
As $b$ was taken arbitrarily among the regular values close enough to $b_0$, we find uncountably many
disjoint (since they correspond to distinct values)
compact complex curves in $X$,
as desired.
In order to conclude, we need to prove the final assertion on $E_3^k$.
We proved above
that there exists a meromorphic map $h\colon X\to R$ with compact fibres, where
$R$ is Riemann surface. It is enough to prove that
$E_3^k \subseteq \{h'=0\}$.
Let $x\in X$ be such that $h'(x)\neq 0$.
Consider a strictly psh
exhaustion function $\psi$ for $R$
(which we can assume to be $\mathcal{C}^\infty$ near $x$ by \cite{Ri})
and the family of functions
$\mathcal F :=\{ \psi + \epsilon \rho\}$, where
$\rho$ is a smooth function compactly supported near $h(x)$. For every such $\rho$,
$\psi + \epsilon \rho$ is still strictly psh for $\epsilon$ sufficiently small. Thus,
we can obtain
a set of generators
for the tangent space given by
differentials at $h(x)$ of psh functions in $\mathcal F$.
Pre-composing the corresponding functions with $h$, we obtain that the space
of differentials at $x$
of psh
functions on $X$ has dimension at least $2$. Hence, $x \notin E_3^k$, and the
proof is complete.
\end{proof}
\begin{rem}
Suppose that $X$ is a surface
and $Y$ a regular level for an
exhaustion function $\phi\in {\rm Psh}^0_e(X)$.
Let
$K\subseteq Y\cap\Sigma_X$ be a local maximum set.
By Lemma \ref{l:intersection},
$Y$ and $K$ are foliated by holomorphic discs.
For every such disk, its tangent bundle is exactly the restriction of
$\mathcal S$. By \cite{BrSib}*{Theorem 1.4}, there exists
a $\partial\oli{\de}$-closed
positive current of bidimension $(1,1)$, directed by $\mathcal S$, supported in $K$.
By Lemma \ref{lmm_cptspp}, such current is a Levi current. This gives a different proof
of Item \eqref{item_t_KY} when $\dim X=2$.
\end{rem}
\section{End of the proof of Theorem \ref{t:main}}\label{s:proof_main}
It follows
from Corollary \ref{corol_spsh} (or Lemma \ref{lmm_spsh})
that
${\rm spt} T\subseteq \Sigma_X^k$ for every Levi current $T$
and all $k\geq 0$.
Thus, we have $F \subseteq \Sigma_X^k$ for all $k\geq 0$.
Moreover,
Items \eqref{item_t_locmax}, \eqref{item_t_KY}, and
\eqref{item_t_curves}
follow from Propositions
\ref{lemma:levi_local_max}, \ref{propos-locmax-supp-new}, and
\ref{t:e2}, respectively.
Let now $Y$ be a regular
connected component of a level set for an
exhaustion psh function $\phi\in {\rm Psh}^k_e(X)$ for some $k\geq 2$.
The remaining item
follows
from the next proposition.
\begin{propos}\label{prop-levels-K}
If $k\geq 2$ and
$Y\subseteq \Sigma_X^k$,
there exists $c'<c$ such that
the connected component of $\phi^{-1}([c',c])$
containing $Y$ is contained in $F$.
\end{propos}
\begin{proof}
We
assume for simplicity that the level $\{\phi=c\}$ is regular and connected, the argument is similar
otherwise.
Since $k\geq 2$, by \cite{ST}*{Theorem 3.9} there is $c'<c$ such that, setting
$$K=\{x\in X\ :\ c'\leq \phi(x)\leq c\},$$
the
form
$(dd^c\phi)^2$
vanishes
on the interior of
$K$, hence on $K$.
So, we have $K \subseteq \Sigma_X^k$.
Consider the current $T$ given by
\[
T:= i\partial\phi\wedge\oli{\de}\phi.
\]
It is clear that $T$ is
a current of bidimension $(1,1)$,
positive and
directed by the complex subspace of the tangent
of the levels of $\phi$. Moreover, $i\partial\oli{\de} T$ is induced by
the form
\[i\partial\oli{\de}
(i\partial\phi\wedge\oli{\de}\phi)
=
-(\partial\oli{\de}\phi)^2.
\]
So, $i\partial\oli{\de} T$
vanishes
where $\phi$ is not strictly psh, hence on $\Sigma_X^k$.
Let $B$ be
the interior of $K$, then the restriction of $T$ to $B$ is a current of bidimension $(1,1)$, positive,
$\partial\oli{\de}$-closed (in $B$); moreover, given $u\in {\rm Psh}^0(X)$, we have that
$T\wedge\partial\oli{\de} u=0$ on $\Sigma_X^k$, so $T$ is a Levi current.
By construction and Lemma \ref{lemma_vanish}
we have $T\wedge \partial\phi=0$, so we
can disintegrate $T$ along the levels of $\phi$, see Corollary \ref{cor_disint}:
there exist
currents $T_s$ with $s\in (c',c)$, such that, for $\alpha$ a $2$-form with ${\rm spt} \alpha\subset B$,
$$\langle T,\alpha\rangle=\int_{c'}^c\langle T_s,\alpha\rangle\, d\mu(s)
\quad
\mbox{ for some measure } \mu \mbox{ on } (c',c).$$
Since
$\phi\in \mathcal{C}^2$,
the measure $\mu$ is absolutely
continuous with respect to the Lebesgue measure on $(c',c)$.
As $T$ is $\partial\oli{\de}$-closed in $B$, so is $\mu$-almost every $T_s$ in $B$;
therefore, for a dense open set of $s\in (c',c)$, $T_s$ is
a positive, $\partial\oli{\de}$-closed current of bidimension $(1,1)$ and
$${\rm spt} T_s=\{x\in X\ :\ \phi(x)=s\}.$$
The set in the RHS is compact since $\phi$ is an exhaustion function.
By Lemma \ref{lmm_cptspp}, $T_s$
is a Levi current.
In conclusion, the level set $\{x\in X\ :\ \phi(x)=s\}$
is contained in $F$ for almost all $s\in (c',c)$, so $\phi^{-1}([c',c])\subseteq F$,
as $F$ is closed. In particular, $Y\subseteq F$.
\end{proof}
The proof of Theorem \ref{t:main} is complete.
\section{Real analytic exhaustion function}\label{s:analytic}
A classification of
those
weakly complete complex surfaces
$X$
admitting
an analytic exhaustion function is given in \cite{mst}. As a direct consequence, we can get an analogous
complete classification of the possible Levi currents in this setting.
First notice that each exceptional divisor $V$ in $X$
corresponds to an extremal Levi current given by the current of integration $[V]$.
Without loss of generality, to simplify our next statement, we
can thus assume that $X$ has no such divisors on the regular levels of $\alpha$.
The statement for a general $X$ is then a direct consequence.
\begin{teorema}
Let $X$ be a weakly complete complex surface admitting an
analytic exhaustion function $\alpha$. Assume that $X$ has no exceptional divisors on the regular levels of $\alpha$.
Then one of the following possibilities hold:
\begin{enumerate}
\item $X$ is Stein (and so, admits no Levi currents);
\item $F=\Sigma_X^{\infty} = X= \cup V_i$, where all the $V_i$ are (disjoint) connected
compact curves, and all extremal Levi currents are of the form $\lambda [V_i']$ for some positive $\lambda$, with $V_i'$ an irreducible component of some $V_i$;
\item $F=\Sigma_X^{\infty} =X$, every regular level $Y_c$ of $\alpha$ is foliated by curves $U_i$, and
the support of any extremal Levi currents on $Y_c$ is equal to
(a connected component of) $Y_c$.
\end{enumerate}
\end{teorema}
Observe also that, although a priori we would only have $\Sigma_X^k \subseteq \Sigma_X^\infty$ for all $k\geq 0$,
the above geometric description implies that
$\Sigma_X^k = \Sigma_X^\infty$
for all $k\geq 0$.
\begin{proof}
It follows from \cite{mst}*{Theorem 1.1} that
one of the following possibilities holds:
\begin{enumerate}
\item $X$ is a Stein space;
\item $X$ is proper over a (possibly singular) complex curve;
\item the connected components of the regular levels of $\alpha$
are foliated with dense complex curves.
\end{enumerate}
In the first and second cases, the assertion
follows from the characterization of Levi currents given in Section \ref{s:levi}.
In the third case, a Levi current can be constructed, for instance, by means of \cite{BrSib}*{Theorem 1.4}.
By proposition \ref{lemma:levi_local_max}, the support of any Levi current
is a local maximum set.
By \cite{MZ}*{Lemma 3.3}, a local maximum set contained in a Levi-flat hypersurface must be a union of leaves of the Levi foliation.
Hence, in the third case,
any Levi current on a regular level set of the exhaustion function
is supported on the whole level set, as all the leaves of the Levi foliation are dense.
This in particular applies to extremal Levi currents. The proof is complete.
\end{proof}
\begin{bibdiv}
\begin{biblist}
\bib{BrSib}{article}{
author={Berndtsson, Bo},
author={Sibony, Nessim},
title={The $\overline\partial$-equation on a positive current},
journal={Inventiones Mathematicae},
volume={147},
date={2002},
number={2},
pages={371--428},
issn={0020-9910},
doi={10.1007/s002220100178},
}
\bib{DS}{article}{
title={Pull-back of currents by holomorphic maps},
author={Dinh, Tien-Cuong},
author={Sibony, Nessim},
journal={Manuscripta Mathematica},
volume={123},
number={3},
pages={357--371},
year={2007},
publisher={Springer},
doi={10.1007/s00229-007-0103-5},
}
\bib{Gra}{article}{
author={Grauert, Hans},
title={On Levi's problem and the imbedding of real-analytic manifolds},
journal={Annals of Mathematics},
volume={68},
number={2},
pages={460--472},
year={1958},
doi={10.2307/1970257},
}
\bib{HST}{article}{
title={On defining functions for unbounded pseudoconvex domains},
author={Harz, Tobias}
author={Shcherbina, Nikolay},
author={Tomassini, Giuseppe},
journal={Mathematische Zeitschrift},
volume={286},
pages={987--1002},
year={2017},
doi={10.1007/s00209-016-1792-9},
}
\bib{Hir}{article}{
title={Le probleme de L{\'e}vi pour les espaces homogenes},
author={Hirschowitz, Andr{\'e}},
journal={Bulletin de la Soci{\'e}t{\'e} Math{\'e}matique de France},
volume={103},
pages={191--201},
year={1975},
doi={10.24033/bsmf.1801},
}
\bib{M}{article}{
author={Mongodi, S.},
title={Weakly complete domains in Grauert type surfaces},
journal={Annali di Matematica Pura ed Applicata (1923 -)},
volume={198},
number={4},
pages={1185--1189},
year={2019},
doi={10.1007/s10231-018-0814-0},
}
\bib{MZ}{article}{
author={Mongodi, S.},
author={Slodkowski, Z.},
title={Domains with a continuous exhaustion in weakly complete surfaces},
journal={Mathematische Zeitschrift},
volume={296},
pages={1011–1019},
year={2020},
doi={10.1007/s00209-020-02466-z},
}
\bib{crass}{article}{
author={Mongodi, Samuele},
author={Slodkowski, Zbigniew},
author={Tomassini, Giuseppe},
title = {On weakly complete surfaces},
journal = {Comptes Rendus Mathematique},
volume = {353},
number = {11},
pages = {969 -- 972},
year = {2015},
doi = {10.1016/j.crma.2015.08.009},
}
\bib{mst}{article}{
author = {Mongodi, Samuele},
author = {Slodkowski, Zbigniew},
author = {Tomassini, Giuseppe},
title = {Weakly complete complex surfaces},
journal = {Indiana University Mathematics Journal},
volume = {67},
year = {2018},
number = {2},
pages = {899 -- 935},
doi={10.1512/iumj.2018.67.6306},
}
\bib{mst2}{article}{
author = {Mongodi, Samuele},
author = {Slodkowski, Zbigniew},
author = {Tomassini, Giuseppe},
title = {Some properties of Grauert type surfaces},
journal = {International Journal of Mathematics},
volume = {28},
number = {8},
pages = {1750063 (16 pages)},
year = {2017},
doi = {10.1142/S0129167X1750063X},
}
\bib{MT}{incollection}{
title={Minimal kernels and compact analytic objects in complex surfaces},
author={Mongodi, Samuele},
author={Tomassini, Giuseppe},
booktitle={Advancements in Complex Analysis},
pages={329--362},
year={2020},
doi={10-1007/978-3-030-40120-7{\_}9},
}
\bib{Na1}{article}{
title={The Levi problem for complex spaces},
author={Narasimhan, Raghavan},
journal={Mathematische Annalen},
volume={142},
number={4},
pages={355--365},
year={1961},
doi={10.1007/BF01451029},
}
\bib{Na2}{article}{
title={The Levi problem for complex spaces II},
author={Narasimhan, Raghavan},
journal={Mathematische Annalen},
volume={146},
number={3},
pages={195--216},
year={1962},
doi={10.1007/BF01470950},
}
\bib{Ni}{article}{
author={Nishino, Toshio},
title={L'existence d'une fonction analytique sur une vari\'{e}t\'{e} analytique
complexe \`a deux dimensions},
language={French},
journal={Publications of the Research Institute for Mathematical Sciences},
volume={18},
date={1982},
number={1},
pages={387--419},
issn={0034-5318},
doi={10.2977/prims/1195184029},
}
\bib{OS}{article}{
title={Bounded psh functions and pseudoconvexity in K{\"a}hler manifold},
author={Ohsawa, Takeo},
author={Sibony, Nessim},
journal={Nagoya Mathematical Journal},
volume={149},
pages={1--8},
year={1998},
doi={10.1017/S0027763000006516},
}
\bib{Ri}{article}{
title={Stetige streng pseudokonvexe Funktionen},
author={Richberg, Rolf},
journal={Mathematische Annalen},
volume={175},
number={4},
pages={257--286},
year={1967},
doi={10.1007/BF02063212},
}
\bib{R}{article}{
title={The local maximum modulus principle},
author={Rossi, Hugo},
journal={Annals of Mathematics},
volume={72},
number={1},
pages={1--11},
year={1960},
doi={10.2307/1970145},
}
\bib{Sh}{article}{
title={On the polynomial hull of a graph},
author={Shcherbina, Nikolay},
journal={Indiana University Mathematics Journal},
volume={42},
number={2},
pages={477--503},
year={1993},
doi={10.1512/iumj.1993.42.42022},
}
\bib{Sib_pf}{article}{
title={Pfaff systems, currents and hulls},
author={Sibony, Nessim},
journal={Mathematische Zeitschrift},
volume={285},
number={3-4},
pages={1107--1123},
year={2017},
publisher={Springer},
doi={10.1007/s00209-016-1740-8},
}
\bib{Sib}{article}{
title={Levi problem in complex manifolds},
author={Sibony, Nessim},
journal={Mathematische Annalen},
date={2018},
volume={371},
pages={1047--1067},
doi={10.1007/s00208-017-1539-x},
}
\bib{Sib_pseudo}{article}{
title={Pseudoconvex domains with smooth boundary in projective spaces},
author={Sibony, Nessim},
journal={Mathematische Zeitschrift},
year={2020},
publisher={Springer},
doi={10.1007/s00209-020-02613-6},
}
\bib{Sl}{article}{
title={Local maximum property and q-plurisubharmonic functions in uniform algebras},
author={Slodkowski, Zbigniew},
journal={Journal of mathematical analysis and applications},
volume={115},
number={1},
pages={105--130},
year={1986},
doi={10.1016/0022-247X(86)90027-2},
}
\bib{Sl_pseudo}{incollection}{
title={Pseudoconcave decompositions in complex manifolds},
author={Slodkowski, Zbigniew},
booktitle={Contemporary Mathematics, Advances in Complex geometry}
volume={31},
pages={239--259},
year={2019},
doi={10.1090/conm/735/14829},
}
\bib{ST}{article}{
author={Slodkowski, Zibgniew},
author={Tomassini, Giuseppe},
title={Minimal kernels of weakly complete spaces},
journal={Journal of Functional Analysis},
volume={210},
date={2004},
number={1},
pages={125--147},
doi={10.1016/S0022-1236(03)00182-4},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
\label{sec:chap4_intro}
Low-mass X-ray binaries (LMXBs) are systems containing either a neutron star (NS) or a black hole (BH) that accretes mass from a companion star. Accretion happens via an accretion disc, which produces radiation with a characteristic spectrum peaking in the X-rays \citep{Shakura1973}.
NS LMXBs have historically been classified as either {\it Atolls} or {\it Z-sources}, based on the patterns they trace in their Colour-Colour diagrams or Hardness-Intensity diagrams (CCDs and HIDs, \citealt{Hasinger1989}), and later connected with the average accretion rates typically observed in either class: very high (often super-Eddington) for Z-sources, and relatively low (below 50\% Eddington) for Atolls (\citealt{Homan2007}).
\cite{Munoz2014} showed that the same state/transition scheme typically used for BH systems (via the HID, \citealt{Homan2001}, and the rms-intensity diagrams, RIDs, \citealt{Munoz2011}) is evident in NS systems as well, which show hard, intermediate and soft states very similar to those in BH systems.
The X-ray power density spectra (PDS) of both Atolls and Z-sources evolve along the HID track showing different types of narrow features superposed, usually, on broad-band noise, called quasi-periodic oscillations (QPOs). NS LMXB QPOs have been divided into high- and low-frequency QPOs. Low-frequency QPOs (LF QPOs) have centroid frequencies ranging between $\sim 0.1$ Hz and $\sim 60$ Hz. For Z-sources, LF QPOs have been historically divided into normal-branch oscillations (NBOs, \citealt{middleditch1986}), horizontal-branch oscillations (HBOs, \citealt{vanderklis1985}) and flaring-branch oscillations (FBOs, \citealt{Klis1989}) based on where they are detected along the CCD.
Similar QPOs have been found in Atoll sources (see e.g. \citealt{DiSalvo2003}), which were divided into HBO-like and FBO-like QPOs \citep{Motta2017} in analogy with the oscillations typical of Z-sources (note that there are no NBO-like QPOs in Atoll sources).
High-frequency QPOs (HF QPOs, \citealt{Stroh2001, belloni2012}) are called kHz QPOs in NS systems. These often appear in pairs, and thus are divided into upper and lower kHz QPOs, both with centroid frequencies spanning the range between a few hundred hertz and over a thousand hertz \citep{vanderklis1996}.
X-ray QPOs originate in the innermost regions of the accretion flow, and are believed to be related to the geometry and dynamics of the accretion flow \citep{bookvanderklis}. However, despite being known for decades now \citep{patterson1977, Ingram2020}, QPOs remain poorly understood, and their physical origin is still largely debated.
There are different groups of suggested QPO mechanisms in the literature, all somewhat involving either the characteristic motion of matter in a strong field regime, or the oscillation of different parts of the accretion flow (see \citealt{Ingram2020} for a recent review). The relativistic precession model (RPM), originally proposed by \cite{Stella1998}, considers the motion of matter with elliptical orbits slightly tilted with respect to the spin of a compact object, and associates such motions to specific types of QPOs visible in the PDS. In particular, the nodal precession ascribed to the motion of matter related to the frame-dragging occurring around spinning compact objects - known as the Lense-Thirring (LT) effect \citep{Bardeen1975} - has been associated to the HBO and HBO-like QPOs observed in NS LMXBs. The precession of the elliptical orbit's semi-major axis - the periastron precession - and the orbital frequency are instead associated to the lower and upper kHz QPOs, respectively.
The RPM has been used to interpret QPOs in both NS \citep{stella1999, done1, duBuisson2020} and BH systems \citep{m2014, Motta2014b}. A remarkable exception is however represented by a NS LMXB located in the globular cluster Terzan 5, the $11$ Hz accreting pulsar IGR J17480-2446, which seems to show QPOs that cannot be explained in terms of LT precession (\citealt{terzan5}, ALT2012 from here onwards). ALT2012 analysed the fast-time variability of the source using data from the Rossi X-ray Timing Explorer (RXTE)\footnote{RXTE: \burl{https://heasarc.gsfc.nasa.gov/docs/xte/}}, and found \mbox{LF QPOs} in the range $\sim 35-50$ Hz in $6$ observations, and kHz QPOs in the range $\sim 800-920$ Hz, which in some cases appeared simultaneously with the LF QPOs. These authors classified the LF QPOs as HBOs, and showed that given the slow spin rate of the NS in the system, they could not be interpreted as the effect of LT precession, which should instead result in QPOs at frequencies strictly below $0.82$ Hz, which were however not detected. These findings thus cast doubt on the LT interpretation of HBOs in NS LMXBs and, by extension, on the RPM.
In this paper we present the results of a new analysis of the RXTE observations of \mbox{IGR J17480-2446} during its 2010 outburst. We systematically searched for QPOs at very low frequencies in the data from the Proportional Counter Array (PCA, \citealt{Jahoda2006}) in order to reassess the presence or absence of HBOs. We also searched for QPOs at up to a few thousand hertz in order to reproduce the findings by ALT2012 and allow a direct comparison to be made.
The paper is structured as follows: Section \ref{sec:chap4_rpm} summarises the RPM while Section \ref{sec:chap4_data} details our sample selection and data analysis. In \mbox{Section \ref{sec:chap4_results}} we describe our results, in Section \ref{sec:chap4_disc} we discuss them, and Section \ref{sec:chap4_conc} summarises and concludes the study.
\section{The Relativistic precession model}
\label{sec:chap4_rpm}
The RPM was originally proposed by \cite{Stella1998}, and later revisited by the same authors as well as by others (\citealt{stella1999}, \citealt{Stella1999a}, \citealt{Merloni1999}, \citealt{Ingram2009}, \citealt{m2014}, \citealt{IngramRPM2014}). According to the RPM, the characteristic motions of a test-particle following an elliptical orbit slightly tilted out of the equatorial plane of a spinning, compact object can be associated with specific types of QPOs, which are detected in the light curves and PDS of accreting BHs and NSs.
The nodal precession of the orbits - driven by the LT mechanism \citep{Bardeen1975} - has a characteristic frequency given by $\nu_\textrm{nod} = \nu_\phi - \nu_\theta$, where $\nu_\phi$ is the orbital frequency and $\nu_\theta$ the vertical frequency. The precession of the elliptical orbit's semi-major axis (periastron precession) occurs instead at a frequency $\nu_\textrm{per} = \nu_\phi - \nu_\textrm{r}$, where $\nu_\textrm{r}$ is the radial epicyclic frequency. For the explicit forms of the above frequencies, the reader should refer to \cite{Merloni1999} and \cite{m2014}.
The RPM associates $\nu_\textrm{nod}$, $\nu_\textrm{per}$ and $\nu_\phi$ each with a respective QPO. In NS LMXBs, in particular, $\nu_\textrm{nod}$, $\nu_\textrm{per}$ and $\nu_\phi$ are associated with the HBO, and the lower and upper kHz QPO, respectively (\citealt{Stella1998}).
\cite{m2014} showed that under the assumption that when observed simultaneously, the three QPOs relevant for the RPM are generated at the same emission radius, the three equations of the RPM form a system of three unknown parameters: the mass $M$ and spin $a$ of the compact object, and the emission radius $r$. This system can be solved analytically if all three QPOs are present, or can be used to put constraints on any of the unknown parameters if the frequency of one of the QPOs is unknown \citep{IngramRPM2014}. This approach has been followed by \cite{Franchini2017} to set constraints on the spin parameters of a sample of accreting LMXBs.
\section{Observations and data analysis}
\label{sec:chap4_data}
We analysed all publicly available archival RXTE observations of the 11 Hz accreting pulsar IGR J17480-2446 during its 2010 outburst, consisting of 47 observations. All observations had source count rates above 10 cts/s/PCU\footnote{PCU: Proportional Counter Unit}, ensuring adequately high signal-to-noise ratios (S/N) for the subsequent analysis.
For each observation we considered \textsc{Binned}, \textsc{Event}, \textsc{Single Bit} and \textsc{Good Xenon} PCA data modes \citep{jahoda1996, Bradt1993} and calculated the PDS using a custom software under IDL\footnote{GHATS: \hspace{1mm} \burl{http://www.brera.inaf.it/utenti/belloni/GHATS_Package/Home.html}}. We used a maximum time resolution of \mbox{$1/8192$ s} ($\sim 122 \hspace{1mm} \mu s$), and divided each observation into segments of both 16 and 512 seconds, respectively, for average PDS production (see \textit{Case 1} and \textit{Case 2} below). We excluded from the analyses in \textit{Case 1} and \textit{Case 2} short observations which contained fewer than five segments, and averaged the Leahy-normalised PDS created from each segment to produce one averaged PDS per observation with a Nyquist frequency of \mbox{$4096$ Hz}, or two or more average PDS in the case of observations longer than $12000$s. We did not subtract the contribution of the Poisson noise \textit{a priori}, but fitted it when modeling the source PDS.
We note that observation 95437-01-01-00 contains a lunar eclipse which was cut from the observation before any further analysis (\citealt{Motta2011}, \citealt{Riggio2012}). In cases of sudden drops in an observation's count rate (usually towards the beginning or end of an observation) that can be ascribed to the re-pointing of the satellite, the beginning/end of an observation is clipped away to prevent the inclusion of low-quality data. It should be noted that all data gaps are removed prior to any FFT being carried out. Our method of PDS production differed depending on the type of QPO we were attempting to detect, as described in \textit{Case 1} and \textit{Case 2} below.
We also computed the HID for the 47 observations of \mbox{IGR J17480-2446} we consider here (see Figure \ref{fig:chap4_hid}). The count rates necessary for the computation of the HID were obtained using energy spectra extracted from \textsc{Standard 2} data, and only using PCU \mbox{unit 2}.
For each observation, the source intensity was measured in the \mbox{$2-16$ keV} energy band, while the hardness was calculated as the ratio of counts in two energy bands as $H_{\textrm{HID}} = A/B$, where $A$ stretches between \mbox{$6-10$ keV} and $B$ stretches between $4-6$ keV (\textsc{Standard 2} channels 6-9 and 11-19, respectively).
\subsection{Case 1: searching for very low frequency QPOs}
\label{sec:chap4_case1}
Following the prescriptions of the RPM, ALT2012 calculated that the HBO QPO of IGR J17480-2446 should fall strictly below \mbox{$0.82$ Hz}. In order to detect such very low frequency (VLF) QPOs, it is necessary to calculate PDS with a frequency resolution of \mbox{$0.08$ Hz} or better. We therefore divided our observations into segments of 512 seconds, from which we calculated PDS with a frequency resolution of \mbox{$1/512$ Hz $\sim 0.002$ Hz}, sufficient for the task at hand.
The observations of \mbox{IGR J17480-2446} contain, however, a large number of Type-I X-ray bursts (see \citealt{Motta2011}). These bursts have a soft, thermal spectrum, but their short recurrence times (down to about 200s) introduces quasi-periodic variability in the light curve that can take the form of a QPO in the PDS (in this case generated from the NS surface) that is not easily distinguishable from other (accretion-driven) types of QPOs, thereby obstructing the process of finding VLF QPOs.
The short Type-I X-ray burst recurrence time implies that - in most cases - cutting the bursts out of our observations leaves too little data for proper analysis. In other words, apart from for a few cases, it was impossible to recover data stretches long enough to reach the frequency resolution needed for our analysis.
We therefore instead performed an energy selection on our data, considering only photons in higher energy bands. In order to select bands that sufficiently removed Type-I X-ray bursts, we applied the following cuts to the entire dataset and inspected the resulting light curves:
\begin{itemize}
\item $0-120$ keV \,\,\,\, (absolute PCA channels $0-249$)
\item $8-120$ keV \,\,\,\, (absolute PCA channels $20-249$)
\item $10-120$ keV \, (absolute PCA channels $25-249$)
\item $12-120$ keV \, (absolute PCA channels $30-249$)
\item $15-120$ keV \, (absolute PCA channels $35-249$)
\item $17-120$ keV \, (absolute PCA channels $40-249$).
\end{itemize}
\noindent An example of the effect of these cuts on an observation's light curve can be seen in Appendix \ref{AppA}. From this assessment, it was found that cuts associated to the $15-120$ keV and $17-120$ keV bands adequately removed the Type-I X-ray bursts. We finally decided to use the $\sim 15-120$ keV band (absolute PCA channels 35 to 249) to derive our PDS, as this still allowed for a high enough S/N in the PDS for the subsequent analysis. We also note that the accretion-driven aperiodic and quasi-periodic variability tend to have a hard spectrum (e.g., \citealt{Sobolewska2006}), meaning our strategy effectively reduces the number of soft, non-variable photons, thus emphasising the remaining variability. Following the described strategy, we effectively minimise the contribution of Type-I X-ray bursts in our PDS while retaining possible HBO QPOs.
The recurrence time of the Type-I X-ray bursts in our observations vary between $200$s to over $500$s, depending on the observation. Their presence could therefore generate peaks in PDS at frequencies smaller than $0.005$ Hz. We thus conservatively exclude all significant QPO-like features falling below $0.009$ Hz from our analysis (which could still be due to residual X-ray burst contributions) in order to avoid any possible contamination of our results.
\subsection{Case 2: searching for low frequency QPOs and kHz QPOs}
\label{sec:chap4_case2}
ALT2012 found LF QPOs in the $\sim 35-50$ Hz range, and \mbox{kHz QPOs} between $\sim 800$ Hz and $920$ Hz. In order to find these features in the observations considered here, we divided each of our observations into intervals of 16 seconds for the calculation of PDS, resulting in a frequency resolution of $0.0625$ Hz. This is preferable to the 512s intervals used for \textit{Case 1}, as the large number of PDS produced in this way is averaged, significantly increasing the S/N. However, it can happen that the averaged PDS from long observations contain broadened features due to the movement of QPOs in frequency as time progresses. For the cases where an observation was longer than $12000$s, we split the observation into shorter segments of approximately $3500$s in length, and calculated an averaged PDS for each of them, which we fitted individually.
We then investigated three different methods of PDS production to determine which one resulted in the optimal detection of \mbox{LF QPOs} and kHz QPOs. First, we simply derived the PDS using the energy band $\sim 2 - 120$ keV (absolute PCA channels 0 to 249).
Next we used the same energy band, but also cut the Type-I X-ray bursts out of observations. The short time intervals used to calculate PDS allowed us to generate a good S/N for the average PDS using the time intervals between consecutive Type-I X-ray bursts.
Finally, we produced PDS in the energy band \mbox{$\sim 15-120$ keV} (absolute PCA channels 35 to 249) without any burst cuts.
The first method presented us with the largest number of significant QPOs in our data, and it was therefore employed to search for LF QPOs and kHz QPOs.
\subsection{Power spectral fitting}
\label{subsec_chap4:selection}
To find the QPOs present in our dataset, we preselected the PDS of observations that visually contained these features for each of the cases above. The features of each power spectrum were fit with a combination of Lorentzians and a power-law component (to account for the Poisson noise) by means of the XSPEC package (\citealt{Arnaud1996}) by using a one-to-one energy-frequency conversion for our PDS and a unity response matrix. We excluded all non-significant features from the analysis: for very low frequencies (where flat-top noise or red noise were present), this meant excluding features below a significance\footnote{Significance is calculated as the integral of the power of the Lorentzian used for the fitting of the feature divided by the negative $1\sigma$ error on this integral.} of $2\sigma$ (in order to account for the lower relative frequency resolution); at all other frequencies features had to be significant at or above $3\sigma$. QPOs were identified by requiring that a given feature is detected at a significance of $3\sigma$ or above, and has a quality factor $Q \geq 2$ (taking uncertainties into account). Here $Q = v_c / \Delta \nu$, where $v_c$ is the centre frequency and $\Delta \nu$ the FWHM\footnote{FWHM: full width at half maximum} of the Lorentzian.
Due to the X-ray pulsar nature of the NS in \mbox{IGR J17480-2446}, a very narrow peak corresponding to the 11 Hz pulsation is visible in our PDS. We do not cut this peak from the PDS - instead, we fit it along with the rest of the features.
\section{Results}
\label{sec:chap4_results}
We report in Table \ref{tab:chap4_qpos} the VLF QPOs found by carrying out the analysis described in \mbox{Case 1} and fitting the resulting PDS (top section), as well as the LF and \mbox{kHz QPOs} found through the analysis described in Case 2 (bottom two sections). For more information on the fitting of the PDS of the observations containing VLF QPOs, see Appendix \ref{AppB}.
We found $7$ VLF QPOs having centre frequencies in the range $\sim 0.01 - 0.03$ Hz (for illustration, three of these are shown in Figure \ref{fig:vlfqpos}), $3$ \mbox{LF QPOs} in the range $\sim 44 - 50$ Hz and $3$ kHz QPOs in the range $\sim 840 - 870$ Hz. The frequency ranges of our LF and kHz QPOs fall within those reported by ALT2012. We note, however, that the analysis performed by these authors differed slightly from ours, and a list of their six observations considered is not given. It is therefore not trivial to determine if and when our detections correspond exactly to those reported in ALT2012. We further note that for the observations in which VLF QPOs were found, the recurrence times for Type-I X-ray bursts (before they were removed by our energy cuts) were $>670$s - this would translate to a peak (if any) in the PDS at $<0.0015$ Hz. Our conservative exclusion of all significant QPO-like features falling below $0.009$ Hz from our analysis (see \mbox{Section \ref{sec:chap4_case1}}) is therefore justified.
\begin{table}
\renewcommand{\arraystretch}{1.3}
\caption[QPOs]{A list of QPOs detected in our analysis. VLF QPOs were found following \textit{Case 1} analysis (see Section \ref{sec:chap4_case1}). VLF QPO observations with a star next to their Obs IDs indicate observations for which second harmonics of the VLF QPOs were also found. These are not listed in the table as they do not form part of our analysis, but more detail regarding fit parameters for observations containing VLF QPOs can be found in Appendix \ref{AppB}. LF and kHz QPOs were found through \textit{Case 2} analysis (see Section \ref{sec:chap4_case2}). In the cases where observations were cut into two or more independent segments, an extra digit attached at the end of their observation IDs indicate which segment was used. See the text for more details. All errors reported in this table are $1 \, \sigma$ errors.}
\begin{tabular}{lccc}
\toprule
\textbf{Obs ID} & \textbf{Frequency (Hz)} & \textbf{Q factor} & \textbf{Significance} \\ \hline \hline
\multicolumn{4}{c}{\textbf{VLF QPOs}} \\
\hline \hline
95437-01-02-01 & $0.0101^{+0.0011}_{-0.0008}$ & $1.9 \pm 0.7$ & $4.3 \, \sigma$ \\
95437-01-10-05 & $0.010^{+0.002}_{-0.001}$ & $1.6 \pm 0.8$ & $6.2 \, \sigma$\\
95437-01-11-03 & $0.016 \pm 0.002$ & $2 \pm 1$ & $3.2 \, \sigma$ \\
95437-01-11-06* & $0.020 \pm 0.001$ & $4 \pm 1$ & $3.9 \, \sigma$ \\
95437-01-12-04* & $0.0175 \pm 0.0006$ & $5 \pm 2$ & $4.2 \, \sigma$ \\
95437-01-13-04 & $0.029 \pm 0.002$ & $3 \pm 1$ & $3.8 \, \sigma$ \\
95437-01-14-00 & $0.014 \pm 0.002$ & $2 \pm 1$ & $3.2 \, \sigma$ \\
\hline \hline
\multicolumn{4}{c}{\textbf{LF QPOs}} \\
\hline \hline
95437-01-07-00 & $47.5 \pm 0.5$ & $3.2 \pm 0.4$ & $11.9 \, \sigma$ \\
95437-01-08-00-1 & $49.2^{+0.8}_{-0.9}$ & $5 \pm 2$ & $4.8 \, \sigma$\\
95437-01-09-00 & $44.7 \pm 0.6$ & $5 \pm 1$ & $6.9 \, \sigma$ \\
\hline \hline
\multicolumn{4}{c}{\textbf{kHz QPOs}} \\
\hline \hline
95437-01-07-00 & $840^{+13}_{-11}$ & $12 \pm 5$ & $3.3 \, \sigma$ \\
95437-01-09-00 & $854 \pm 4$ & $22 \pm 8$ & $4.9 \, \sigma$ \\
95437-01-10-01 & $870^{+6}_{-7}$ & $21 \pm 10$ & $3.7 \, \sigma$ \\
\toprule
\end{tabular}
\label{tab:chap4_qpos}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{vlf_qpos_paper.png}
\caption{The PDS of three observations in which we found VLF QPOs (see also Table \ref{tab:chap4_qpos}). Top: Obs ID 95437-01-02-01. Middle: Obs ID \mbox{95437-01-10-05.} Bottom: Obs ID 95437-01-11-03.}
\label{fig:vlfqpos}
\end{figure}
In order to compare the quasi-periodic features we detected with the predictions of the RPM, we used the RPM equations given in \citealt{m2014} (see also Section \ref{sec:chap4_rpm}) to plot the theoretical estimates of $\nu_{\textrm{nod}}$, $\nu_{\textrm{per}}$ and $\nu_{\phi}$ as a function of the emission radius $r$.
We used the measured spin frequency $\nu = 11$ Hz of the source, and calculated the minimum and maximum dimensionless spins ($a_{min}$ and $a_{max}$) by assuming a minimum and maximum moment of inertia for the NS ($I_{min}$ and $I_{max}$), respectively. To do so, we followed \cite{mukherjee2018}, and in particular their Figure 4, which shows the moment of inertia as a function of different NS masses inferred for a number of realistic candidate NS equations of state. We found $I_{min} = 0.75 \times 10^{45}$ $\textrm{g} \hspace{0.5mm} \textrm{cm}^2$ for a NS mass of $M_{min} = 1.0$ $\textrm{M}_{\odot}$, and $I_{max} = 5 \times 10^{45}$ $\textrm{g} \hspace{0.5mm} \textrm{cm}^2$ for a mass of $M_{max} = 2.7$ $\textrm{M}_{\odot}$. We then calculated the minimum and maximum dimensionless spin parameters $a_{min}$ and $a_{max}$, to find $0.0054 \leq a \leq 0.0059$. Next, we used Equations 9, 6 and 1 in \cite{IngramRPM2014} to determine the minimum and maximum theoretical estimates of $\nu_{\textrm{nod}}$, $\nu_{\textrm{per}}$ and $\nu_{\phi}$ as a function of the emission radius $r$, assuming first $a_{min}$ and $M_{min}$, and then $a_{max}$ and $M_{max}$. Our result is displayed in Figure \ref{fig:chap4_boundaries}, where we also marked the frequency ranges where we detected QPOs in our data.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{v_r.png}
\caption{The minimum and maximum theoretical values of $\nu_{\textrm{nod}}$, $\nu_{\textrm{per}}$ and $\nu_{\phi}$ as a function of the emission radius $r$ for IGR J17480-2446, inferred using the RPM. Coloured lines indicate the maximum and minimum predicted values, respectively. Straight horizontal dashed lines mark the frequency ranges in which QPOs were found.}
\label{fig:chap4_boundaries}
\end{figure}
The HID for our data is shown in Figure \ref{fig:chap4_hid}, showing the source count rate and hardness of each observation in our analysis. Consecutive observations are connected via thin lines, with the earliest and latest observations circled and numbered. Red data points indicate observations containing VLF QPOs, light blue observations contain LF QPOs and black observations contain kHz QPOs (note that in two cases LF QPOs and kHz QPOs are detected simultaneously). VLF QPOs are detected both at the beginning and at the end of the outburst, in a hard state, while LF QPOs and kHz QPOs are found in a relatively soft state, close to the peak of the outburst which, according to \cite{Motta2011}, reached and possibly exceeded the Eddington limit.
\section{Discussion}
\label{sec:chap4_disc}
Based on the assumption that all types of QPOs, both in NS and in BH systems, should show at least a mild dependence on the spin of the compact object, we hypothesised that given the very low spin of IGR J17480-2446 ($11$ Hz, i.e. over an order of magnitude smaller than the average spin of accreting NSs, see \citealt{vanD2017}), HBOs should be detected at very low frequencies in this system (i.e. significantly below 1 Hz), a hypothesis made by ALT2012 themselves. We therefore repeated the analysis of the RXTE observations of IGR J17480-2446, following a method specifically aimed at finding PDS features with very low centroid frequencies. We also searched for QPOs up to a few thousand hertz in order to reproduce the results of ALT2012 and to allow for a direct comparison to be made. We found $7$ VLF QPOs with centre frequencies in the range $\sim 0.01-0.03$ Hz (see Table \ref{tab:chap4_qpos}), consistent with the values predicted by ALT2012 for HBOs in a slow-spinning NS. We also found $3$ LF QPOs and $3$ kHz QPOs (see Table \ref{tab:chap4_qpos}), with centroid frequencies consistent with those reported by the same authors.
We estimated the theoretical values of $\nu_{\textrm{nod}}$, $\nu_{\textrm{per}}$ and $\nu_{\phi}$ as a function of the radius $r$, obtained assuming a NS with mass $M$ between 1 and 2.7 $M_{\odot}$, and adopting a dimensionless spin parameter $0.0054 \leq a \leq 0.0059$.
By comparing our estimates with the QPOs we detected we observe the following:
\begin{enumerate}
\item The VLF QPOs we detected are consistent with being HBOs - QPOs generated through LT precession of the accretion flow - at a radius larger than approximately $15 \hspace{1mm} R_g$ ($\approx 35$ km) from the NS centre. This is supported by the fact that these features all appear in a relatively hard state, as is clear from the HID, where the inner disc radius is believed to be truncated far from the NS surface (see e.g. \citealt{Done2007}).
\item The kHz QPOs that we detected are consistent with being either the upper or lower kHz QPOs, thus associated to the orbital or periastron precession frequency of matter at a radius lower than approximately $12 \hspace{1mm} R_g$ ($\approx 25$ km) from the NS centre. The lack of a simultaneous detection of two kHz QPOs prevents any further classification.
\item As already noted by ALT2012, the LF QPOs detected are not consistent with the nodal precession frequency around a NS spinning at $11$ Hz. They are, in principle, consistent with the periastron precession frequency of matter orbiting at a distance of $15-25 \hspace{0.5mm} R_g$ from the NS. However, we note that two of our \mbox{LF QPOs} are detected together with kHz QPOs at $\sim 840$ Hz (see \mbox{Table \ref{tab:chap4_qpos}}, observation 95437-01-07-00 and 95437-01-09-00). According to Figure \ref{fig:chap4_boundaries}, this would imply the simultaneous detection of QPOs at two very different radii, which would constitute a violation of one of the key assumptions of the RPM.
\end{enumerate}
Concerning point (iii) we also tested for the non-simultaneity of the two QPOs in observations 95437-01-07-00 and \mbox{95437-01-09-00} by splitting each into several smaller segments, but both features seem to be present during a large fraction of these two observations. Assuming the correctness of the RPM, the simultaneity of the kHz QPOs and LF QPOs point to two main scenarios: either that the QPO associated to the periastron precession frequency (that at $\sim 47.5$ Hz) is generated at a radius larger than that at which the orbital frequency QPO (that at $\sim 840$ Hz) is generated, thus effectively invalidating a key assumption of the RPM; or that the \mbox{LF QPOs} we detected are not associated to any of the three motions relevant for the RPM, having already stated in (iii) that they cannot be explained by LT precession.
The first scenario pushes us toward the same conclusion drawn by ALT2012, i.e. that the RPM cannot explain the frequency properties of at least some of the QPOs, at least in its simplest form.
The second scenario, instead, further stresses the question of the real nature of the LF QPOs observed between $35$ and $50$ Hz in \mbox{IGR J17480-2446.} ALT2012 classified the QPOs at $35-50$ Hz based on their frequency (comparable with the values typically seen in other NS systems), and based on the position these features occupied in the so-called Wijnands -- van der Klis correlation (\citealt{Wijnands1999a}), formed by the frequency of a broad PDS component (called $L_b$) and the frequency of the HBO. The small number of QPOs detected by both ALT2012 and by ourselves does not really allow for the establishing of a correlation, and the fact that most of the components identifiable in a NS PDS correlate with one another (see \citealt{Psaltis1999}) suggests that ALT2012's classification of the LF QPOs at $35-50$ Hz might not be correct. Unfortunately, given the very low frequency of the VLF QPOs that we report in this work, it is not possible to determine whether these do fit into the Wijnands \& van der Klis correlation (the $L_b$ component's centroid frequency would be visible below $10$ mHz, which is lower than our frequency resolution, assuming HBO frequencies lower than $0.1$ Hz). Only a new outburst form this source and new data will therefore allow us to reach more conclusive results.
Is there an alternative explanation for these LF QPOs? PDS from NS systems are notoriously more structured and feature-rich than those from BH systems. Among the still-largely unknown features typical of NS systems are hectohertz QPOs (see e.g. \citealt{Altamirano2008}). These QPOs have been detected in a number of Atoll sources - 4U0614+09, 4U1608-52, 4U1728-34 and \mbox{4U1636-53} - around $\sim 100$ Hz. Interestingly, all these sources contain NSs spinning at fairly high frequencies ($415$, $620$, $363$ and $581$ Hz, respectively). While there is no known correlation between the NS spin and the frequency of hectohertz QPOs, it seems plausible that in a slowly spinning NS such as the one in \mbox{IGR J17480-2446} hectohertz QPOs can appear at frequencies lower than $\sim 100$ Hz. We therefore speculate that the LF QPOs detected in IGR J17480-2446 might be relatively low-frequency hectohertz QPOs.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{HID.png}
\caption{The HID of IGR J17480-2446, where the source count rate is plotted against the spectral hardness. Each data point represents an observation, and consecutive observations are connected. The first and last observation considered are marked on the plot and denoted by `first' and `last', respectively. Small red dots are observations containing VLF QPOs, larger light blue dots are observations containing LF QPOs, and black crosses are observations containing kHz QPOs (see also \mbox{Table \ref{tab:chap4_qpos}}). Light blue dots overlaid by black crosses contain both LF QPOs and kHz QPOs. A colour version of this figure is available online.}
\label{fig:chap4_hid}
\end{figure}
\section{Summary and conclusions}
\label{sec:chap4_conc}
We have examined all $47$ RXTE PCA observations of the $11$ Hz accreting pulsar IGR J17480-2446 located in the globular cluster Terzan 5 during its 2010 outburst. We searched for QPO features located between $0.01$ Hz and $\sim 4000$ Hz.
We found $7$ VLF QPOs with centre frequencies in the range $\sim 0.01-0.03$ Hz, $3$ LF QPOs in the range $\sim 44-50$ Hz, and $3$ kHz QPOs in the range $\sim 840-870$ Hz. We compared the theoretical values of the nodal frequency $\nu_{\textrm{nod}}$, periastron precession frequency $\nu_{\textrm{per}}$ and orbital frequency $\nu_{\phi}$ as a function of the emission radius as predicted by the RPM to our findings. We find that the centroid frequencies of our detected VLF QPOs are consistent with the predicted nodal frequencies if generated at radii larger than $15 \hspace{1mm} R_g$. We also find that our LF QPOs detected at $40-50$ Hz could be consistent with the periastron precession frequency of material orbiting at approximately $15 \hspace{1mm} R_g$ from the NS centre. The presence of kHz QPOs simultaneous to these LF QPOs, however, either disproves this hypothesis or invalidates one of the main assumptions of the RPM.
We have shown that VLF QPOs at frequencies consistent with those expected for LT driven modulations are present in the data, even though more data are required to confirm the classification of such features. While not conclusive, our results cast (even more) doubt on the nature of the $35-50$ Hz QPOs detected in \mbox{IGR J17480-2446}. These LF QPOs could either be associated with the periastron precession frequency (though in this case a key assumption of the RPM needs to be relaxed significantly), or with a type of QPO known as hectohertz QPOs, observed here at smaller frequencies possibly due to the low NS spin. It is also, of course, entirely possible that these LF QPOs are simply a new type of QPO, possibly peculiar to this unique system, or perhaps typical of slowly spinning accreting NS LMXBs, of which IGR J17480-2446 is currently the only example.
\section*{Acknowledgments}
\label{sec:chap4_acknowledgments}
LdB acknowledges support from the Rhodes Trust and Christ Church College. SEM \mbox{acknowledges} the Science and Technology Facilities Council (STFC) for financial support, and the Oxford Centre for Astrophysical Surveys, which is funded through generous support from the Hintze Family Charitable Foundation. The authors thank Diego Altamirano for valuable discussions.
\section*{Data availability}
\label{sec:data_avail}
The data underlying this article are publicly available from the RXTE Archive: \url{https://heasarc.gsfc.nasa.gov/docs/xte/archive.html}.
\bibliographystyle{mnras}
|
\section{introduction}
For a real number $x$, the nearest integer of $x$ denoted by $||x||$ and it is defined as
$$
||x||:=\mbox{min}\{|x-m|:m\in\mathbb{Z}\}.
$$
In 1957, Mahler \cite{mahler} used Ridout's theorem, which is a $p$-adic extension of Roth's theorem to prove the following: {\it let $\alpha\in \mathbb{Q}\backslash\mathbb{Z}$ and $\varepsilon$ be a positive real number . Then there are only many finitely many $n\in \mathbb{N}$ satisfying $||\alpha^n||<2^{-\varepsilon n}$}. Mahler also asked for which algebraic numbers $\alpha$ the above conclusion remains hold.
\smallskip
In 2004, by ingenious applications of the Subspace Theorem, Corvaja and Zannier \cite{corv} proved a `Thue-Roth' type inequality with `moving targets'. Then as an application of this result, they answered the question of Mahler and proved the following: {\it let $\alpha>1$ be a real algebraic number and let $\varepsilon$ be a positive real number. Suppose that $|| \alpha^n ||< 2^{- \varepsilon n}$ for infinitely many $n$.
Then, there is some integer $d\geq 1$ such that the number $\alpha^d$ is a Pisot number. In particular $\alpha$ is an algebraic integer. }
\bigskip
In this paper, the main motivation is to prove an inhomogeneous extension of Thue-Roth's type inequality with moving targets in the same spirit as the result of Corvaja and Zannier in \cite{corv}. Indeed, we prove the following.
\begin{theorem}\label{maintheorem}
Let $\Gamma\subset \overline{\mathbb Q}^{\times}$ be a finitely generated multiplicative group of algebraic numbers. Let $\delta$ be a non-zero algebraic number, $\beta\in(0,1)$ be an algebraic irrational, and $\varepsilon>0$ be a fixed real number. Then there exist only finitely many triples $(u, q, p)\in\Gamma\times\mathbb{Z}^2$ with $d=[\mathbb{Q}(u):\mathbb{Q}]$ such that $|\delta q u|>1$ and
\begin{equation*}\label{eq1.1}
\tag{1.1}
0<|\delta qu+\beta-p|<\frac{1}{H^\varepsilon(u)q^{d+\varepsilon}}.
\end{equation*}
\end{theorem}
Recently in 2019, Kulkarni, Mavraki and Nguyen \cite{kul} generalize Mahler problem to an arbitrary linear recurrence sequence of the form $\{Q_1(n)\alpha_1^n+\cdots+Q_k(n)\alpha^n_k:n\in\mathbb{N}\}$, where $\alpha_i$'s are non-zero algebraic number and $Q_i(x)\in\overline{Q}[x]\backslash \{0\}$. In particular case, they proved the following inhomogeneous extension of the problem of Mahler: {\it let $\alpha>1$ and $\beta$ be real algebraic numbers and let
$\varepsilon$ be a positive real number.
Supose that $||\alpha^n+\beta|| < 2^{- \varepsilon n}$ for infinitely many $n$.
Then there is some integer $d\geq 1$ such that the number $\alpha^d$ is a Pisot number.
In particular $\alpha$ is an algebraic integer. We recall that a real algebraic number $\alpha > 1$ is called {\it a Pisot number}, if the modulus value of all the other conjugates of $\alpha$ is $< 1$. }
\bigskip
In the above result, if $\beta$ is an integer and $\alpha$ is an algebraic number such that $\alpha^d$ is a Pisot number for some integer $d\geq 1$, then clearly there are infinitely many natural numbers $n$ satisfying $||\alpha^{dn}+\beta||<2^{-\eps n}$ for some $\eps>0$. Thus, we can conclude that the above assertion is best possible, if $\beta$ is a rational number.
Surprisingly, when $\beta$ is an algebraic irrational, we prove a contrast result as an application of our main theorem as follows.
\begin{theorem}\label{maintheorem2}
Let $\alpha>1$ be a real number. Let $\beta$ be an algebraic irrational and $\lambda$ be a non-zero real algebraic number. For a given real number $\varepsilon >0$, if there are infinitely many natural numbers $n$ for which $||\lambda\alpha^n+\beta|| < 2^{- \varepsilon n}$ holds true, then $\alpha$ is transcendental.
\end{theorem}
Note that Theorem \ref{maintheorem2} strengthen the main result of Wagner and Ziegler \cite{wagner}.
\section{Preliminaries}
Let $K\subset \mathbb{C}$ be a Galois extension over $\mathbb{Q}$. Let $M_K$ be the set of all inequivalent places of $K$ and $M_\infty$ be the set of all archimedian places of $K$. For each place $v\in M_K$, we denote $|\cdot |_v$ the absolute value corresponding to $v$, normalized with respect to $K$. Indeed if $v\in M_\infty$, then there exists an automorphism $\sigma\in\mbox{Gal}(K/\mathbb{Q})$ of $K$ such that for all $x\in K$,
\begin{equation*}\label{eq2.1}
\tag{2.1}
|x|_v=|\sigma(x)|^{d(\sigma)/[K:\mathbb{Q}]},
\end{equation*}
where $d(\sigma) =1$ if $\sigma(K) = K\subset \mathbb{R}$ and $d(\sigma) = 2$ otherwise. Non-archimedian absolute values are normalized accordingly so that the product formula $\displaystyle\prod_{\omega\in M_K}|x|_\omega=1$ holds for any $x\in K^\times$. For each automorphism $\rho\in \mbox{Gal}(K/\mathbb{Q}),$ one defines an archimedian valuation on $K$ by the formula
\begin{equation*}\label{eq2.2}
\tag{2.2}
|x|_\rho:= |\rho^{-1}(x)|^{d(\rho)/[K:\mathbb{Q}]},
\end{equation*}
where $|\cdot |$ denotes the complex absolute value. Note that two distinct automorphisms, say, $\rho_1$ and $\rho_2$ defines the same absolute value if and only if $\rho_1^{-1}\circ \rho_2$ is a complex conjugation. Thus, for each $v\in M_\infty$, let $\rho_v$ be an automorphism defining the valuation $v$ according to \eqref{eq2.2}. Then the set $\{\rho_v : v\in M_\infty\}$ represents the left cosets of the subgroup generated by the complex conjugation in $\mbox{Gal}(K/\mathbb{Q})$.
The absolute Weil height $H(x)$ is defined as
$$
H(x):=\prod_{\omega\in M_K}\mbox{max}\{1,|x|_\omega\} \mbox{ for all } x\in K.
$$
For a vector $\mathbf{x}=(x_1,\ldots,x_n)\in K^n$ and for a place $\omega\in M_K$, the $\omega$-norm for $\mathbf{x}$ denote by $||\mathbf{x}||_\omega$ and given by
$$
||\mathbf{x}||_\omega:=\mbox{max}\{|x_1|_\omega,\ldots,|x_n|_\omega\}
$$
and the projective height, $H(\mathbf{x})$, is defined by
$$
H(\mathbf{x})=\prod_{\omega\in M_K}||\mathbf{x}||_\omega.
$$
Now we are ready to present a more general version of the Schmidt Subspace Theorem, which was formulated by Schlickewei and Evertse. For the reference, see (\cite[ Chapter 7]{bomb}, \cite[ Chapter V, Theorem 1D$^\prime$]{schmidt} and \cite[Page 16, Theorem II.2]{zannier})
\smallskip
\begin{theorem} (Schlickewei) \label{schli}
Let $K$ be an algebraic number field and $m \geq 2$ an integer. Let $S$ be a finite set of places on $K$ containing all the archimedian places. For each $v \in S$, let $L_{1,v}, \ldots, L_{m,v}$ be linearly independent linear forms in the variables $X_1,\ldots,X_m$ with coefficients in $K$. For any $\varepsilon>0$, the set of solutions $\textbf{x} \in K^m\backslash\{0\}$ to the inequality
\begin{equation*}
\prod_{v\in S}\prod_{i=1}^{m} \frac{|L_{i,v}(\textbf{x})|_v}{\|\textbf{x}\|_v} \leq \frac{1}{H(\textbf{x})^{m+\varepsilon}}
\end{equation*}
contained in finitely many proper subspaces of $K^m$.
\end{theorem}
The following lemma, established in \cite{corv}, is used at several places in the proof of the main result
of \cite{corv}.
\begin{lemma}\label{lemCZ1}
Let $K$ be a Galois extension over $\mathbb{Q}$ and $S$ be a finite subset of places, containing all the archimedean places. Let $\sigma_1,\ldots,\sigma_n$ be distinct automorphism of $K$ and let
$\lambda_1,\ldots,\lambda_n$ be non-zero elements of $K$.
Let $\varepsilon>0$ be a positive real number and $\omega\in S$ be a distinguished place.
Let $c > 0$.
Let $\mathfrak{E}\subset \mathcal{O}_S^\times$ be the set of solutions $u\in\mathcal{O}_S^\times$
of the inequality
\begin{equation*}
0< |\lambda_1 \sigma_1(u)+\cdots+\lambda_n \sigma_n(u)|_\omega<
c \max\{|\sigma_1(u)|_\omega,\ldots,|\sigma_n(u)|_\omega\} H^{-\varepsilon}(u).
\end{equation*}
If $\mathfrak{E}$ is infinite subset of $\mathcal{O}_S^\times$,
then there exists a non-trivial linear relation of the form
$$
a_1 \sigma_1(u)+\cdots+a_n \sigma_n(u)=0,\quad \mbox{with } a_i\in K
$$
which holds for infinitely many elements of $u$ in $\mathcal{O}_S^\times$.
\end{lemma}
The following application of Theorem \ref{schli} is a slight modification of Lemma \ref{lemCZ1}.
\begin{lemma}\label{lem2.1}
Let $K$ be a Galois extension over $\mathbb{Q}$ and $S$ be a finite subset of places, containing all the archimedean places. Let $\sigma_1,\ldots,\sigma_n$ be distinct automorphism of $K$ and let
$\lambda_0, \lambda_1,\ldots,\lambda_n$ be non-zero elements of $K$.
Let $\varepsilon>0$ be a positive real number and $\omega\in S$ be a distinguished place. Let $\mathfrak{E}\subset \mathcal{O}_S^\times\times\mathbb{Z}$ be the subset defined as
\begin{equation*}
\mathfrak{E} := \left\{ (u,q)\in \mathcal{O}_S^\times \times\mathbb{Z}^\times
\ : 0<\ |\lambda_0+\lambda_1 q\sigma_1(u)+\cdots+\lambda_nq\sigma_n(u)|_\omega<\frac{\max\{|q\sigma_1(u)|_\omega,\ldots,|q\sigma_n(u)|_\omega\}}{|q|^{n+\varepsilon}H^{\varepsilon}(u)}\right\}.
\end{equation*}
If $\mathfrak{E}$ is infinite subset of $\mathcal{O}_S^\times\times\mathbb{Z}$, then there exists a non-trivial linear relation of the form
$$
a_1 \sigma_1(u)+\cdots+a_n \sigma_n(u)=0,\quad \mbox{with } a_i\in K
$$
which holds for infinitely many elements of $u$ in $\mathcal{O}_S^\times$.
\end{lemma}
\begin{proof}
In order to prove this lemma, we shall apply Theorem \ref{schli} as in the proof \cite[Lemma 1]{corv}. Without loss of generality, we can assume that
\begin{equation*}
|q\sigma_1(u)|_\omega=\max\{|q\sigma_1(u)|_\omega,\ldots,|q\sigma_n(u)|_\omega\}
\end{equation*}
for all $(u,q)\in \mathfrak{E}$. For $\nu\in S$, let us define $n+1$ linear forms $L_{0,\nu},\ldots,L_{n,\nu}$ in $n+1$ variables ${\bf X}=(X_0,X_1,\ldots,X_n)$ as follows: Put
\begin{align*}
L_{\omega,0}(X_0,X_1,\ldots,X_n)&=X_0\\
L_{\omega,1}(X_0,X_1,\ldots,X_n)&=\lambda_0 X_0+\lambda_1 X_1+\cdots+\lambda_n X_n
\end{align*}
for $2\leq i\leq n$, we define
$$
L_{\omega,i}(X_0,X_1,\ldots,X_n)=X_i,
$$
and for each $\mathit{v}\neq \omega\in S$, $0\leq j\leq n $, we let
$$
L_{\mathit{v},j}(X_0,X_1,\ldots,X_n)=X_j.
$$
Now we put
$$
{\bf X}=(1, q\sigma_1(u),\ldots,q\sigma_n(u))\in K^{n+1}
$$
and consider the product
\begin{equation*}\label{eq2.4}
\tag{2.3}
\prod_{\mathit{v}\in S}\prod_{i=0}^n\frac{|L_{\mathit{v},i}(\bf X)|_\mathit{v}}{||\bf X||_\mathit{v}}.
\end{equation*}
Using the fact that $L_{v,j}({\bf X}) =q \sigma_j(u)$ for $2\leq j\leq n $, we obtain
$$
\prod_{\mathit{v\in S}}\prod_{j=2}^{n}|L_{\mathit{v},j}(\mathbf{X})|_\mathit{v}
= \prod_{\mathit{v}\in S}\prod_{j=2}^{n}|q|_\mathit{v}\prod_{j=2}^{n}\prod_{\mathit{v}\in S}|\sigma_j(u)|_\mathit{v}.
$$
Since $\sigma_j(u)$ are $S$-units, by the product formula, we obtain
$$
\prod_{\mathit{v}\in S}|\sigma_j(u)|_\mathit{v}=\prod_{\mathit{v}\in M_K}|\sigma_j(u)|_\mathit{v}=1.
$$
Consequently, the above inequality gives
\begin{equation*}\label{eq2.5}
\tag{2.4}
\prod_{\mathit{v\in S}}\prod_{j=2}^{n}|L_{\mathit{v},j}(\mathbf{X})|_\mathit{v}=\prod_{\mathit{v}\in S}\prod_{j=2}^{n}|q|_\mathit{v}\leq \prod_{v\in M_\infty}\prod_{j=2}^{n}|q|_\mathit{v}= |q|^{n-1}.
\end{equation*}
Now we estimate $\prod_{\nu\in S}\prod_{i=0}^n||\bf X||_\nu$:
\begin{equation*}\label{eq2.6}
\tag{2.5}
\prod_{\mathit{v}\in S}\prod_{i=0}^n||{\bf X}||_\mathit{v}=\prod_{i=0}^n\left(\prod_{\mathit{v}\in S}||{\bf X}||_\mathit{v}\right)\geq (H({\bf X}))^{n+1}=H^{n+1}(1,q\sigma_1(u),\ldots,q\sigma_n(u))
\end{equation*}
since $||{\bf X}||_\mathit{v}\leq 1$ for all $\mathit{v}$ not in $S$.
\smallskip
By re-writing \eqref{eq2.4}, we have
\begin{align*}
\prod_{\mathit{v}\in S}\prod_{i=0}^n\frac{|L_{\mathit{v},i}(\bf X)|_\mathit{v}}{||\bf X||_\mathit{v}}&= |L_{\omega,0}({\bf X})|_\omega |L_{\omega, 1}({\bf X})|_\omega\left(\prod_{\mathit{v}\neq \omega\in S}\prod_{i=0}^1
|L_{\mathit{v},i}(\bf X)|_\nu\right) \left(\prod_{\mathit{v}\in S}\prod_{i=2}^n|L_{\mathit{v},i}({\bf X})|_\mathit{v}\right)\left(\prod_{\mathit{v}\in S}\prod_{i=0}^n\frac{1}{||\bf X||_\mathit{v}}\right)
\end{align*}
Then by using the fact that $L_{\mathit{v},0}({\bf X})=1$ for $\mathit{v}$ in $S$ and the product formula
$$
\prod_{\mathit{v}\neq \omega\in S}|q\sigma_1(u)|_\mathit{v}=\left(\prod_{\mathit{v}\neq \omega\in S}|q|\right)(|\sigma_1(u)|_\omega)^{-1},
$$
we get
\begin{align*}
\prod_{\mathit{v}\in S}\prod_{i=0}^n\frac{|L_{\mathit{v},i}(\bf X)|_\mathit{v}}{||\bf X||_\mathit{v}}&=\frac{|L_{\omega, 1}({\bf X})|_\omega}{|\sigma_1(u)|_\omega}
\prod_{\mathit{v}\neq \omega\in S} |q |_\mathit{v}
\left(\prod_{\mathit{v}\in S}\prod_{i=2}^n|L_{\mathit{v},i}({\bf X})|_\mathit{v}\right)\left(\prod_{\mathit{v}\in S}\prod_{i=0}^n\frac{1}{||\bf X||_\mathit{v}}\right)\\
&\leq\frac{|L_{\omega, 1}({\bf X})|_\omega}{|\sigma_1(u)|_\omega}\frac{|q|}{|q|_\omega} \left(\prod_{\mathit{v}\in S}\prod_{i=2}^n|L_{\mathit{v},i}({\bf X})|_\mathit{v}\right)\left(\prod_{\mathit{v}\in S}\prod_{i=0}^n\frac{1}{||\bf X||_\mathit{v}}\right),
\end{align*}
since $|q|_\mathit{v}\leq 1$ for all $\mathit{v}$-non-archimedian absolute value.
\smallskip
Thus from the assumption and \eqref{eq2.5}, \eqref{eq2.6}, we obtain
\begin{align*}
\prod_{\mathit{v}\in S}\prod_{i=0}^n\frac{|L_{\mathit{v},i}(\bf X)|_\mathit{v}}{||\bf X||_\mathit{v}}&\leq \frac{|\lambda_0+\lambda_1 q\sigma_1(u)+\cdots+\lambda_n q\sigma_n(u)|_\omega |q|^n}{|q\sigma_1(u)|_\omega}\frac{1}{H^{n+1}({\bf X})}\\
&\leq\frac{\max\{|q\sigma_1(u)|_\omega,\ldots,|q\sigma_n(u)|_\omega\}}{|q\sigma_1(u)|_\omega |(|q|H(u))^\varepsilon}\frac{1}{H^{n+1}({\bf X})},
\end{align*}
since $(u,q)\in\mathfrak{E}$. Using that
$$
|q\sigma_1(u)|_\omega=\max\{|q\sigma_1(u)|_\omega,\ldots,|q\sigma_n(u)|_\omega\},
$$
we obtain
$$
\prod_{\mathit{v}\in S}\prod_{i=0}^n\frac{|L_{\mathit{v},i}(\bf X)|_\mathit{v}}{||\bf X||_\mathit{v}}\leq \frac{1}{H^{n+1}({\bf X})}\frac{1}{(|q|H(u))^\varepsilon}.
$$
The height of the vector ${\bf X}=(1,q\sigma_1(u),\ldots,q\sigma_n(u))$ satisfies the following inequality:
$$
H({\bf X})\leq |q|H^{K:\mathbb{Q}}(u)=|q|H^n(u).
$$
Hence the above estimate becomes
$$
\prod_{\mathit{v}\in S}\prod_{i=0}^n\frac{|L_{i,\mathit{v}}(\bf X)|_\mathit{v}}{||\bf X||_\mathit{v}}\leq \frac{1}{H^{n+1}({\bf X})}\frac{1}{H^{\varepsilon/[K:\mathbb{Q}]}({\bf X})}=\frac{1}{H^{n+1+\varepsilon/[K:\mathbb{Q}]}({\bf X})}.
$$
Then by Theorem \ref{schli}, there exists a non-trivial relation of the form
\begin{equation*}\label{eq2.7}
\tag{2.6}
a_0+a_1 q\sigma_1(u)+\cdots+a_n q\sigma_n(u)=0
\end{equation*}
satisfied by infinitely many elements of $(u,q)\in \mathfrak{E}$.
Our next goal is to prove the following claim.
\bigskip
\noindent{\bf CLAIM.~} There exist a non-trivial relation as \eqref{eq2.7} with vanishing coefficients $a_0$.
\bigskip
Assume that $a_0\neq 0$. Then by rewriting the relation \eqref{eq2.7}, we obtain
\begin{equation*}
a_0=-a_1 q\sigma_1(u)-\cdots-a_n q\sigma_n(u).
\end{equation*}
Then from \eqref{eq2.3}, we get
$$
0<\left|-\lambda_0\left(\frac{a_1}{a_0}q\sigma_1(u)+\cdots+\frac{a_n}{a_0} q\sigma_n(u)\right)+\lambda_1 q\sigma_1(u)+\cdots+\lambda_nq\sigma_n(u)\right|_\omega<\frac{\mbox{max}\{|q\sigma_1(u)|_\omega,\ldots,|q\sigma_n(u)|_\omega\}}{|q|^{n+\varepsilon}H^{\varepsilon}(u)}.
$$
holds for infinitely many pairs $(u,q)\in\mathfrak{E}$. This is equivalent to
$$
0<\left|\left(\lambda_1-\frac{\lambda_0 a_1}{a_0}\right)\sigma_1(u)+\cdots+\left(\lambda_n-\frac{\lambda_0 a_n}{a_0}\right)\sigma_n(u)\right|_\omega<\frac{\mbox{max}\{|\sigma_1(u)|_\omega,\ldots,|\sigma_n(u)|_\omega\}}{|q|^{n+\varepsilon}H^{\varepsilon}(u)}.
$$
holds for infinitely many pairs $(u,q)\in\mathfrak{E}$. This inequality implies that $\beta_i=\lambda_i-\frac{\lambda_0 a_i}{a_0}$ are in $K$ and not all are zero.
In particular, there are infinitely many $u$ in $\mathcal{O}_S^\times$ such that
$$
0<\left|\left(\lambda_1-\frac{\lambda_0 a_1}{a_0}\right)\sigma_1(u)+\cdots+\left(\lambda_n-\frac{\lambda_0 a_n}{a_0}\right)\sigma_n(u)\right|_\omega<\frac{\mbox{max}\{|\sigma_1(u)|_\omega,\ldots,|\sigma_n(u)|_\omega\}}{H^{\varepsilon}(u)}.
$$
Then by Lemma \ref{lemCZ1}, there exists a non-trivial relation of the form
$$
a_1 \sigma_1(u)+\cdots+a_n\sigma_n(u)=0
$$
which holds for infinitely many values of $u$ in $\mathcal{O}_S^\times$.
This proves the claim and hence the lemma.
\end{proof}
\section{Key lemma for the proof of Theorem \ref{maintheorem}}
The following lemma is key to the proof of Theorem \ref{maintheorem} and its proof is based on the Subspace Theorem along with the idea in \cite{corv}, with various modifications.
\begin{lemma}\label{lem3.1}
Let $K$ be a Galois extension over $\mathbb{Q}$ of degree $n$ and $k\subset K$ be a subfield of degree $d$ over $\mathbb{Q}$. Let $\delta, \beta$ be two non-zero elements of $K$.
Let $S$ be a finite set of places on $K$ containing all the archimedean places
and let $\varepsilon>0$ be a given real number. Let
\begin{equation*}\label{eq3.1}
\tag{3.1}
\mathcal{B} = \left\{(u, q, p)\in (\mathcal{O}_S^\times\cap k)\times\mathbb{Z}^2 \ : \
0<|\delta qu+\beta-p|<\frac{1}{H^\varepsilon(u)q^{d+\varepsilon}}\right\}
\end{equation*}
such that for each triple $(u,q,p)\in\mathcal{B}$, $|\delta q u|>1$. If $\mathcal{B}$ is infinite, then there exist a proper subfield $k'\subset k$, a non-zero element $\delta'$ in $k$ and an infinite subset $\mathcal{B}'\subset \mathcal{B}$ such that for all triples $(u, q, p)\in\mathcal{B}'$
we have $u/\delta'\in k'.$
\end{lemma}
\bigskip
\noindent{\bf Remark.} Once hypothesis of the Lemma \ref{lem3.1} is satisfied, by the Subspace Theorem given in the form Theorem \ref{schli}, we can conclude that the infinite sequence $u$ satisfies \eqref{eq3.1} can not be fixed; that is, $H(u)\rightarrow\infty$ as the triple varies $(u,q, p) \in \mathcal{B}$.
\begin{proof}
Since $\mathcal{B}$ is an infinite set of solutions of \eqref{eq3.1}, we first observe that
we may assume that $H(u)\rightarrow\infty$.
Suppose that $H(u)$ is bounded. Then there exists an infinite subset $\mathcal{A}$ of $\mathcal{B}$ such that the number $u$ is constant for all elements in $\mathcal{A}$, let say $u_0$ for all triples $(u,q,p)\in\mathcal{A}$ and $q$ is unbounded along the set $\mathcal{A}$. Now we apply Theorem \ref{schli}
to the field $\mathbb{Q}$
with the input $S=\{\infty\}$,
linear forms
$L_{1,\infty}(X_1, X_2,X_3)=\alpha u_0 X_1+\beta X_2-X_3$, $L_{i,\infty}(X_1,X_2,X_3)=X_i$
for $2\leq i\leq 3$
and the points $(q,1,p)$. Then from \eqref{eq3.1}, we see that there is $\delta > 0$ such that the inequality
$$
\prod_{i=1}^3|L_{j,\infty}(q,1,p)|_\infty\leq \frac{1}{(\max\{|q|,1,|p|\})^{\delta}}
$$
holds for infinitely many triples $(q,1,p)\in\mathbb{Z}^3$. Thus by Theorem \ref{schli}, there exists a proper subspace of $\mathbb{Q}^3$ containing infinitely many triples $(q,1,p)$. That is, we have a non-trivial relation of the form
$$
a_0+a_1p+a_2q=0
$$
satisfied by infinitely many triples of the form $(q,1,p)$. Since $a_i$ is in $\mathbb{Z}$
and $q\rightarrow\infty$ along the set $\mathcal{A}$, we conclude that $a_1\neq 0$. Then by substituting the value of $p$ from this above equality into the inequality \eqref{eq3.1} along the set $\mathcal{A}$, we get
$$
0<\left|\alpha q u_0+\beta+\left(\frac{a_0}{a_1}+\frac{a_2}{a_1}q\right)\right|\leq \frac{1}{H^{\varepsilon}(u_0)}\frac{1}{q^{d+\varepsilon}}.
$$
This is equivalent to
$$
0<\left|\left(\alpha u_0+\frac{a_2}{a_1}\right)q+\beta+\frac{a_0}{a_1}\right|\leq \frac{1}{H^{\varepsilon}(u_0)}\frac{1}{q^{d+\varepsilon}},
$$
which cannot hold if $|q|$ is large enough.
Therefore, we conclude that $H(u)\to\infty$ along the set $\mathcal{B}$.
\bigskip
Let $\mathcal{G} = \mbox{Gal}(K/\mathbb{Q})$ be the Galois group of $K$ over $\mathbb{Q}$. Since $K$ over $k$ is Galois, we let $\mathcal{H}:=\mbox{Gal}(K/k)\subset \mathcal{G}$ be the subgroup fixing $k$. Hence, $|\mathcal{G}/\mathcal{H}| = [k:\mathbb Q] = d$. Therefore, among the $n$ embedding of $K$, there are exactly $d$ embeddings, say, $\sigma_1,\ldots,\sigma_d$ are a representatives for the left cosets of $\mathcal{H}$ in $\mathcal{G}$ with $\sigma_1$ being identity and more precisely, we have
$$
\mathcal{G}/\mathcal{H}:=\{\mathcal{H}, \sigma_2 \mathcal{H},\ldots,\sigma_d \mathcal{H}\}.
$$
For each $j = 1, 2, \ldots, d$, let
$$
S_j = \left\{v\in M_\infty\ : \ \rho_v\vert_k = \sigma_j: k\rightarrow \mathbb{C}\right\}
$$
and hence $S_1\cup\ldots\cup S_d=M_\infty$.
We keep this notation throughout the paper.
Then for each $\rho\in\mbox{Gal}(K/\mathbb{Q})$, by \eqref{eq2.2}, we have
\begin{equation*}\label{eq3.2}
\tag{3.2}
|\delta qu+\beta-p|^{d(\rho)/[K:\mathbb{Q}]}=|\rho(\delta)\rho(qu)+\rho(\beta)-\rho(p)|_\rho=|\rho(\delta)q\rho(u)+\rho(\beta)-p|_\rho
\end{equation*}
and hence
\begin{equation*}\label{eq3.3}
\tag{3.3}
\prod_{\mathit{v}\in M_\infty}|\rho_\mathit{v}(\delta)\rho_{\mathit{v}}(qu)+\rho_\mathit{v}(\beta)-p|_\mathit{v}=\prod_{j=1}^d\prod_{\mathit{v}\in S_j}|\rho_{\mathit{v}}(\delta)\sigma_j(qu)+\rho_{\mathit{v}}(\beta)-p|_v.
\end{equation*}
By \eqref{eq3.2}, we see that
$$
\prod_{\mathit{v}\in M_\infty}|\rho_\mathit{v}(\delta)\rho_{\mathit{v}}(qu)+\rho_\mathit{v}(\beta)-p|_\mathit{v}=\prod_{\mathit{v}\in M_\infty}|\delta q u+\beta-p|^{d(\rho_\mathit{v})/[K:\mathbb{Q}]}=|\delta q u+\beta-p|^{{\sum_{\mathit{v}\in M_\infty}}d(\rho_\mathit{v})/[K:\mathbb{Q}]}.
$$
Then, from \eqref{eq3.3} and the well-known formula $\displaystyle\sum_{\mathit{v}\in M_\infty}d(\rho_\mathit{v})=[K:\mathbb{Q}]$, it follows that
\begin{equation*}\label{eq3.4}
\tag{3.4}
\prod_{j=1}^d\prod_{\mathit{v}\in S_j}|\rho_{\mathit{v}}(\delta)\sigma_j(qu)+\rho_{\mathit{v}}(\beta)-p|_v=|\delta q u+\beta-p|.
\end{equation*}
Now, for each $\mathit{v}\in S$, we define $d+2$ linearly independent linear forms in $d+2$ variables as follows: For $j = 1, 2, \ldots, d$ and for $\mathit{v}\in S_j$, we let
\begin{eqnarray*}
L_{\mathit{v},0}(x_0,x_1,\ldots,x_{d+1})&=& \rho_{\mathit{v}}(\beta) x_0 - x_1 + \rho_{\mathit{v}}(\delta) x_{j+1}\\
L_{\mathit{v},1}(x_0,x_1,\ldots,x_{d+1})&=& x_0,
\end{eqnarray*}
and for $2\leq i\leq d+1$, we define
$$
L_{v, i}(x_1,\ldots,x_{d+1})=x_i,
$$
and for $\mathit{v}\in S\backslash{M_\infty}$ and for $0\leq i \leq d+1$, we let
$$
L_{v,i}(x_1,\ldots,x_{d+1})=x_i.
$$
Let $\mathbf{X}$ be the point in $K^{d+2}$ which is of the form
$$
\mathbf{X}=(1,p,q\sigma_1(u),\ldots,q\sigma_d(u)) \in K^{d+2}.
$$
In order to apply Theorem \ref{schli}, we need to calculate the following quantity
\begin{equation*}\label{eq3.5}
\tag{3.5}
\prod_{\mathit{v\in S}}\prod_{i=0}^{d+1}\frac{|L_{\mathit{v},i}(\mathbf{X})|_\mathit{v}}{||\mathbf{X}||_\mathit{v}}.
\end{equation*}
Using the fact that $L_{\mathit{v},i}(\mathbf{X})=q\sigma_i(u)$, for $2\leq i\leq d+1$, we obtain
\begin{eqnarray*}
\prod_{\mathit{v\in S}}\prod_{i=2}^{d+1}|L_{\mathit{v},i}(\mathbf{X})|_\mathit{v}&=&\prod_{\mathit{v}\in M_\infty}\prod_{i=2}^{d+1}|q\sigma_i(u)|\prod_{\mathit{v}\in S\backslash{M_\infty}}\prod_{i=2}^{d+1}|q\sigma_i(u)|_\mathit{v}\\
&=&\prod_{\mathit{v}\in S}\prod_{i=2}^{d+1}|q|_\mathit{v}\prod_{i=2}^{d+1}\prod_{\mathit{v}\in S}|\sigma_i(u)|_\mathit{v}.
\end{eqnarray*}
Since $\sigma_j(u)$ are $S$-units, by the product formula, we obtain
$$
\prod_{\mathit{v}\in S}|\sigma_i(u)|_\mathit{v}=\prod_{\mathit{v}\in M_K}|\sigma_i(u)|_\mathit{v}=1.
$$
Consequently, the above inequality gives
$$
\prod_{\mathit{v\in S}}\prod_{i=2}^{d+1}|L_{\mathit{v},i}(\mathbf{X})|_\mathit{v}=\prod_{\mathit{v}\in S}\prod_{i=2}^{d+1}|q|_\mathit{v}\leq \prod_{v\in M_\infty}\prod_{i=2}^{d+1}|q|_\mathit{v}=\prod_{i=2}^{d+1}|q|^{\sum_{\mathit{v}\in M_\infty}d(\rho_\mathit{v})/[K:\mathbb{Q}]}
$$
Then, from the formula $\sum_{\mathit{v}\in M_\infty}d(\rho_\mathit{v})=[K:\mathbb{Q}]$, we get
\begin{equation*}\label{eq3.6}
\tag{3.6}
\prod_{\mathit{v\in S}}\prod_{i=2}^{d+1}|L_{\mathit{v},i}(\mathbf{X})|_\mathit{v}\leq \prod_{v\in M_\infty}\prod_{i=2}^{d+1}|q|_\mathit{v}=\prod_{i=2}^{d+1}|q|^{\sum_{\mathit{v}\in M_\infty}d(\rho_\mathit{v})/[K:\mathbb{Q}]}=|q|^d.
\end{equation*}
Now we estimate the product of the denominators in \eqref{eq3.5} as follows: consider
$$
\prod_{\mathit{v\in S}}\prod_{i=0}^{d+1}||\mathbf{X}||_\mathit{v}\geq \prod_{\mathit{v\in M_K}}\prod_{i=0}^{d+1}||\mathbf{X}||_\mathit{v}=\prod_{i=0}^{d+1}\left(\prod_{\mathit{v\in M_K}}||\mathbf{X}||_\mathit{v}\right) = \prod_{i=0}^{d+1}H(\mathbf{X}),
$$
since $||\mathbf{X}||_\mathit{v}\leq 1$ for all $\mathit{v}\not\in S$. Thus, we get,
\begin{equation*}\label{eq3.7}
\tag{3.7}
\prod_{\mathit{v\in S}}\prod_{i=0}^{d+1}||\mathbf{X}||_\mathit{v}\geq H(\mathbf{X})^{d+2}.
\end{equation*}
By \eqref{eq3.5}, \eqref{eq3.6} and \eqref{eq3.7}, it follows that
$$
\prod_{\mathit{v\in S}}\prod_{i=0}^{d+1}\frac{|L_{\mathit{v},i}(\mathbf{X})|_\mathit{v}}{||\mathbf{X}||_\mathit{v}}\leq \frac{1}{H^{d+2}(\mathbf{X})}|q|^d|\alpha q u+\beta-p|.
$$
Thus, from \eqref{eq3.1}, we have
$$
\prod_{\mathit{v\in S}}\prod_{i=0}^{d+1}\frac{|L_{\mathit{v},i}(\mathbf{X})|_\mathit{v}}{||\mathbf{X}||_\mathit{v}}\leq \frac{1}{H^{d+2}(\mathbf{X})}|q|^d\frac{1}{H^{\varepsilon}(u)}\frac{1}{|q|^{d+\varepsilon}}=\frac{1}{H^{d+2}(\mathbf{X})}\frac{1}{(|q|H(u))^{\varepsilon}}.
$$
Notice that
\begin{eqnarray*}
H(\mathbf{X})&=&\prod_{\mathit{v}\in M_K}\mbox{max}\{1, |p|_\mathit{v},|q\sigma_1(u)|_\mathit{v},\ldots,|q\sigma_d(u)|_\mathit{v}\}\\
&\leq&\prod_{\mathit{v}\in S}\mbox{max}\{1, |p|_\mathit{v},|q\sigma_1(u)|_\mathit{v},\ldots,|q\sigma_d(u)|_\mathit{v}\}\\
&\leq& \prod_{\mathit{v}\in S}\mbox{max}\{1,|p|_\mathit{v},|q|_\mathit{v} \}\prod_{\mathit{v}\in S}\mbox{max}\{1,|\sigma_1(u)|_\mathit{v},\ldots,|\sigma_d(u)|_\mathit{v}\}\\
&\leq& \max\{|p|, |q|\}\left(\prod_{\mathit{v}\in S}\mbox{max}\{1,|\sigma_1(u)|_\mathit{v}\}\right)\cdots \left(\prod_{\mathit{v}\in S}\mbox{max}\{1,|\sigma_d(u)|_\mathit{v}\}\right)\\
&=&\max\{|p|, |q|\}H^d(u).
\end{eqnarray*}
By using the inequality
$
||x|-|y||\leq |x-y|
$
and the fact that $H(u)\to \infty$ for $(u, q, p)\in \mathcal{B}$, from \eqref{eq3.1}, we conclude that $|p|\leq |\delta q u+\beta|+1.$ Since $|u|^{\frac{1}{d}}\leq H(u)$, and by using the fact $H(u)\to\infty$, we get that
$$
|p|\leq |\delta q u+\beta|+1\leq |q||\delta+\beta|H^d(u)+1\leq |q|H^{2d}(u)
$$
for all but finitely many triples $(u, q, p)\in \mathcal{B}$. By combining both these above inequalities, we obtain $H(\mathbf{X})\leq |q| H(u)^{3d}$, and hence we get $H(\mathbf{X})^{1/3d} \leq |q|H(u)$. Therefore, we get
$$
\prod_{\mathit{v\in S}}\prod_{i=0}^{d+1}\frac{|L_{\mathit{v},i}(\mathbf{X})|_\mathit{v}}{||\mathbf{X}||_\mathit{v}}\leq \frac{1}{H^{d+2}(\mathbf{X})}\frac{1}{(|q|H(u))^\varepsilon}\leq \frac{1}{H(\mathbf{X})^{d+2+(\varepsilon)/3d}} = \frac{1}{H(\mathbf{X})^{d+2+ \varepsilon'}},
$$
for some $\varepsilon' >0$ holds for infinitely many tuples $(1,p,q\sigma_1(u),\ldots,q\sigma_d(u))$ along the triples$(u,q,p)\in\mathcal{B}$. Then by Theorem \ref{schli}, there exists a proper subspace of $K^{d+2}$ containing infinitely many $\mathbf{X}=(1,p,q\sigma_1(u),\ldots,q\sigma_d(u))$ along the triples $(u,q,p) \in \mathcal{B}$.
That is, we have a non-trivial linear relation of the form
\begin{equation*}\label{eq3.8}
\tag{3.8}
a_0+a_1 p+b_1q\sigma_1(u)+\cdots+b_d q \sigma_d(u)=0,\quad a_i, b_j\in K,
\end{equation*}
satisfied by all the triples $(u, q, p)\in\mathcal{B}_1$ for an infinite subset of $\mathcal{B}_1$.
\bigskip
Under the hypotheses of the Main Theorem of \cite{corv}, the authors establish
the existence of such a non-trivial linear relation with $a_0 = 0$.
The present situation is slightly more complicated.
As in \cite{corv}, we will establish that there is
a non-trivial linear relation as above with $a_0 = a_1 = 0$, and then we will conclude exactly as in \cite{corv}.
\smallskip
\noindent{\bf Claim 1.} At least one of the $b_j$'s is non-zero in the relation \eqref{eq3.8}.
\bigskip
If not, suppose $b_i=0$ for all $1\leq i\leq d$. Then from \eqref{eq3.8}, we have
\begin{equation*}\label{eq3.9}
\tag{3.9}
0\neq p=\frac{-a_0}{a_1}\in K.
\end{equation*}
We deduce from \eqref{eq3.1} and \eqref{eq3.9} that
$$
0<|\delta q u+\beta+\frac{a_0}{a_1}|<\frac{1}{H^\varepsilon(u)q^{d+\varepsilon}}
$$
holds for infinitely many pairs $(u,q)$ along the set $\mathcal{B}_1$. Now, we apply Theorem \ref{schli} with
$S$ being the finite set composed of the archimedean places on $K$, the linear
forms $L_{\mathit{v},1}(X_1, X_2)=(\beta+\frac{a_0}{a_1})X_1+\delta X_2$, $L_{\mathit{v},2}(X_1, X_2)=X_1$~~ for $\mathit{v}\in S$,
and the pairs $(1,qu)\in K^2$. Then by Theorem \ref{schli}, we get a non-trivial relation of the form
$$
c_0+c_1 qu=0
$$
which holds for infinitely many pairs $(u,q)$ along the set $\mathcal{B}_1$.
Since $|qu|\to\infty$ as the pairs $(q,u)$ runs through the set $\mathcal{B}_1$.
Therefore, we conclude that $c_0=c_1=0$, which is a contradiction.
Consequently, at least one of the $b_j$'s is non-zero in the relation \eqref{eq3.8}.
\bigskip
\noindent{\bf Claim 2.} There exists a non-trivial relation as \eqref{eq3.8} with $a_0=a_1=0$.
\bigskip
Assume that $a_0\neq 0$. Then by rewriting the relation \eqref{eq3.8}, we obtain
\begin{equation*}\label{eq3.10}
\tag{3.10}
\beta=-\beta\left(\frac{a_1}{a_0}p+\frac{b_1}{a_0}q\sigma_1(u)+\cdots+\frac{b_d}{a_0}q\sigma_d(u)\right).
\end{equation*}
Substituting the value of $\beta$ from \eqref{eq3.10} in \eqref{eq3.1}, we get
$$
0<\left|\delta q u-\beta\left(\frac{a_1}{a_0}p+\frac{b_1}{a_0}q\sigma_1(u)+\cdots+\frac{b_d}{a_0}q\sigma_d(u)\right)-p\right|<\frac{1}{H^\varepsilon(u)}\frac{1}{q^{d+\varepsilon}}.
$$
This is equivalent to write
\begin{equation*}\label{eq3.11}
\tag{3.11}
0<\left|\delta q u-(\beta a_1/a_0+1)p-\beta\left(\frac{b_1}{a_0}q\sigma_1(u)+\cdots+\frac{b_d}{a_0}q\sigma_d(u)\right)\right|<\frac{1}{H^\varepsilon(u)}\frac{1}{q^{d+\varepsilon}}.
\end{equation*}
\vspace{.1cm}
Now we divide rest of the proof in two cases, according as $\beta a_1/a_0+1$ is $0$ or not.
\bigskip
\noindent{\bf Case 1.} The quantity $\beta a_1/a_0+1$ vanishes.
\vspace{.3cm}
Under this case the relation \eqref{eq3.10} can be view as
\begin{equation*}\label{eq3.12}
\tag{3.12}
\beta-p=-\left(\frac{b_1}{a_1}q \sigma_1(u)+\cdots+\frac{b_d}{a_1}q\sigma_d(u)\right).
\end{equation*}
Since $K$ over $\mathbb{Q}$ is Galois and $\beta$ is an algebraic irrational, then there exists an automorphism $\rho_0\in\mbox{Gal}(K/\mathbb{Q})$ such that $\rho_0(\beta)\neq \beta$.
\smallskip
By applying the automorphism $\rho_0$ on both sides of the equality \eqref{eq3.12}, we get
$$
\rho_0(\beta)-p=-\left(\rho_0\left(\frac{b_1}{a_1}\right)q \rho_0\circ\sigma_1(u)+\cdots+\rho_0\left(\frac{b_d}{a_1}\right)q\rho_0\circ\sigma_d(u)\right).
$$
Since the restriction of $\rho_0$ on $k$ is in $\{\sigma_1,\ldots,\sigma_d\}$, we can write this above relation as
$$
\rho_0(\beta)-p=-\left(\rho_0\left(\frac{b_1}{a_1}\right)q \sigma_{1,0}(u)+\cdots+\rho_0\left(\frac{b_d}{a_1}\right)q\sigma_{d,0}(u)\right),
$$
where $\sigma_{i,0}\in\{\sigma_1,\ldots,\sigma_d\}$ for $1\leq i\leq d$.
\vspace{.2cm}
Now by subtracting this equality from \eqref{eq3.12}, we obtain a relation of the form
$$
0\neq \rho_0(\beta)-\beta=c_1 q \sigma_1(u)+\cdots+c_d q\sigma_d(u),\quad c_i\in K
$$
for all the pairs $(q,u)$ along the triples $(u,q,p)\in\mathcal{B}$. This is equivalent to write
\begin{equation*}
\gamma-(c_1 q \sigma_1(u)+\cdots+c_d q\sigma_d(u))=0
\end{equation*}
holds for all the pairs $(q,u)$ along the set $\mathcal{B}$, where we have set
$$
\gamma=\rho_0(\beta)-\beta.
$$
We can easily see that in this relation at least one of $c_i$'s is non-zero.
Dividing this equality by $\gamma$, we get the non-trivial relation of the kind
\begin{equation*}
\beta=\beta\left(\frac{c_1}{\gamma}q \sigma_1(u)+\cdots+\frac{c_d}{\gamma}q \sigma_d(u)\right).
\end{equation*}
Substituting value of $\beta$ from this equality in \eqref{eq3.1}, we get
$$
0<\left|\delta q u+\beta\left(\frac{c_1}{\gamma}q \sigma_1(u)+\cdots+\frac{c_d}{\gamma}q \sigma_d(u)\right)-p\right|<\frac{1}{H^\varepsilon(u)}\frac{1}{q^{d+\varepsilon}}.
$$
This is equivalent to write
\begin{equation*}\label{eq3.16}
\tag{3.13}
0<\left|\left(\delta+\beta\frac{c_1}{\gamma}\right)q \sigma_1(u)+\beta\left(\frac{c_2}{\gamma}q \sigma_2(u)+\cdots+\frac{c_d}{\gamma}q \sigma_d(u)\right)-p\right|<\frac{1}{H^\varepsilon(u)}\frac{1}{q^{d+\varepsilon}}.
\end{equation*}
Recall that, for $j = 1, 2, \ldots, d$, we have set
$$
S_j = \left\{v\in M_\infty\ : \ \rho_v\vert_k = \sigma_j: k\rightarrow \mathbb{C}\right\}
$$
and hence $S_1\cup\ldots\cup S_d=M_\infty$, where $M_\infty$ denotes the set of all archimedean places on $K$. Then for each $\rho\in\mbox{Gal}(K/\mathbb{Q})$, by \eqref{eq2.2}, we have
\begin{align*}
&\left|\left(\delta+\beta\frac{c_1}{\gamma}\right)q \sigma_1(u)+\beta\left(\frac{c_2}{\gamma}q \sigma_2(u)+\cdots+\frac{c_d}{\gamma}q \sigma_d(u)\right)-p\right|^{d(\rho)/[K:\mathbb{Q}]}\\ &=\left|\rho\left(\delta+\beta\frac{c_1}{\gamma}\right)q \rho\circ\sigma_1(u)+\left(\rho\left(\beta\frac{c_2}{\gamma}\right)q \rho\circ\sigma_2(u)+\cdots+\rho\left(\beta\frac{c_d}{\gamma}\right)q \rho\circ\sigma_d(u)\right)-p\right|_\rho
\end{align*}
and hence
\begin{align*}
&\prod_{\mathit{v}\in M_\infty}\left|\rho_{\mathit{v}}\left(\delta+\beta\frac{c_1}{\gamma}\right)q \rho_\mathit{v}\circ\sigma_1(u)+\left(\rho_\mathit{v}\left(\beta\frac{c_2}{\gamma}\right)q \rho_\mathit{v}\circ\sigma_2(u)+\cdots+\rho_\mathit{v}\left(\beta\frac{c_d}{\gamma}\right)q \rho_\mathit{v}\circ\sigma_d(u)\right)-p\right|_\mathit{v}\\ &=\prod_{j=1}^d\prod_{\mathit{v}\in S_j}\left|\rho_{\mathit{v}}\left(\delta+\beta\frac{c_1}{\gamma}\right)q \rho_\mathit{v}\circ\sigma_1(u)+\left(\rho_\mathit{v}\left(\beta\frac{c_2}{\gamma}\right)q \rho_\mathit{v}\circ\sigma_2(u)+\cdots+\rho_\mathit{v}\left(\beta\frac{c_d}{\gamma}\right)q \rho_\mathit{v}\circ\sigma_d(u)\right)-p\right|_\mathit{v}.
\end{align*}
For each $\mathit{v}\in M_\infty$ and $j=1,2,\ldots,d$, we define $\mathit{v}(j)$ such that $\rho_\mathit{v}\circ\sigma_j=\sigma_{\mathit{v}(j)}$ on the field $k$, where $\{\mathit{v}(1), \ldots,\mathit{v}(d)\}$ is a permutation of $\{1,\ldots,d\}$. Hence the above relation can be written as
\begin{align*}
&\prod_{j=1}^d\prod_{\mathit{v}\in S_j}\left|\rho_{\mathit{v}}\left(\delta+\beta\frac{c_1}{\gamma}\right)q \rho_\mathit{v}\circ\sigma_1(u)+\left(\rho_\mathit{v}\left(\beta\frac{c_2}{\gamma}\right)q \rho_\mathit{v}\circ\sigma_2(u)+\cdots+\rho_\mathit{v}\left(\beta\frac{c_d}{\gamma}\right)q \rho_\mathit{v}\circ\sigma_d(u)\right)-p\right|_\mathit{v}\\&=\prod_{j=1}^d\prod_{\mathit{v}\in S_j}\left|\rho_{\mathit{v}}\left(\delta+\beta\frac{c_1}{\gamma}\right)q \sigma_{\mathit{v}(1)}(u)+\left(\rho_\mathit{v}\left(\beta\frac{c_2}{\gamma}\right)q \sigma_{\mathit{v}(2)}(u)+\cdots+\rho_\mathit{v}\left(\beta\frac{c_d}{\gamma}\right)q \sigma_{\mathit{v}(d)}(u)\right)-p\right|_\mathit{v}.
\end{align*}
By \eqref{eq3.2}, we see that
\begin{align*}
&\prod_{j=1}^d\prod_{\mathit{v}\in S_j}\left|\rho_{\mathit{v}}\left(\delta+\beta\frac{c_1}{\gamma}\right)q \sigma_{\mathit{v}(1)}(u)+\left(\rho_\mathit{v}\left(\beta\frac{c_2}{\gamma}\right)q \sigma_{\mathit{v}(2)}(u)+\cdots+\rho_\mathit{v}\left(\beta\frac{c_d}{\gamma}\right)q \sigma_{\mathit{v}(d)}(u)\right)-p\right|_\mathit{v}\\ &=\left|\left(\delta+\beta\frac{c_1}{\gamma}\right)q \sigma_1(u)+\beta\left(\frac{c_2}{\gamma}q \sigma_2(u)+\cdots+\frac{c_d}{\gamma}q \sigma_d(u)\right)-p\right|^{{\sum_{\mathit{v}\in M_\infty}}d(\rho_\mathit{v})/[K:\mathbb{Q}]}.
\end{align*}
Then, from \eqref{eq3.16} and the well-known formula $\displaystyle\sum_{\mathit{v}\in M_\infty}d(\rho_\mathit{v})=[K:\mathbb{Q}]$, it follows that
\begin{align*}
&\prod_{j=1}^d\prod_{\mathit{v}\in S_j}\left|\rho_{\mathit{v}}\left(\delta+\beta\frac{c_1}{\gamma}\right)q \sigma_{\mathit{v}(1)}(u)+\left(\rho_\mathit{v}\left(\beta\frac{c_2}{\gamma}\right)q \sigma_{\mathit{v}(2)}(u)+\cdots+\rho_\mathit{v}\left(\beta\frac{c_d}{\gamma}\right)q \sigma_{\mathit{v}(d)}(u)\right)-p\right|_\mathit{v}\\ &=\left|\left(\delta+\beta\frac{c_1}{\gamma}\right)q \sigma_1(u)+\beta\left(\frac{c_2}{\gamma}q \sigma_2(u)+\cdots+\frac{c_d}{\gamma}q \sigma_d(u)\right)-p\right|<\frac{1}{H^\varepsilon(u)}\frac{1}{q^{d+\varepsilon}}.
\end{align*}
Now for each $\mathit{v}\in S$, we define $d+1$ linearly independent linear forms in $d+1$ variables as follows: For $j = 1, 2, \ldots, d$ and for $\mathit{v}\in S_j$, we let
\begin{align*}
L_{\mathit{v},0}(x_0, x_1,\ldots,x_d)&=-x_0+\rho_{\mathit{v}}\left(\delta+\beta\frac{c_1}{\gamma}\right)x_{\mathit{v}(1)}+\rho_\mathit{v}\left(\beta\frac{c_2}{\gamma}\right)x_{\mathit{v}(2)}+\cdots+\rho_\mathit{v}\left(\beta\frac{c_d}{\gamma}\right)x_{\mathit{v}(d)}
\end{align*}
and for $1\leq i\leq d+1$, we define
$$
L_{v, i}(x_1,\ldots,x_{d+1})=x_i,
$$
and for $\mathit{v}\in S\backslash{M_\infty}$ and for $0\leq i \leq d+1$, we let
$$
L_{v,i}(x_1,\ldots,x_{d+1})=x_i.
$$
Clearly, the linear forms $L_{\mathit{v},0},\ldots,L_{\mathit{v},d}$ are linearly independent for each $\mathit{v}\in S$. Finally, let ${}\mathbf{X}$ be the point in $K^{d+1}$, which is of the form
$$
{\mathbf{X}}=(p, q\sigma_1(u),\ldots,q\sigma_d(u)).
$$
Then by using Theorem \ref{schli} similar to the first part of this lemma, we get a non-trivial relation of the form
$$
a_1 p+b_1 q \sigma_1(u)+\cdots+b_d q \sigma_d(u)=0,\quad a_1, b_i\in K.
$$
Now we prove that there exists a relation with $a_1=0$. In order to prove this, we follow the similar procedure as in \cite[Lemma 3, Claim]{corv} along with our Lemma \ref{lem2.1}. If $a_1\neq 0$, then we have
\begin{equation*}\label{eq3.17}
\tag{3.14}
p=-\frac{b_1}{a_1} q \sigma_1(u)-\cdots-\frac{b_d}{a_1} q\sigma_d(u).
\end{equation*}
First we consider the case when $\displaystyle\sigma_j\left(\frac{b_1}{a_1}\right) \ne \frac{b_j}{a_1}$ for some $j$ with $2\leq j\leq d$.
\smallskip
By applying the automorphism $\sigma_j$ on both sides of \eqref{eq3.16}, we get
$$
p=-\sigma_j\left(\frac{b_1}{a_1}\right)q\sigma_j\circ\sigma_1(u)-\cdots-\sigma_j\left(\frac{b_d}{a_1}\right)q\sigma_j\circ\sigma_d(u).
$$
By subtracting this relation from \eqref{eq3.17}, we get a relation of the form
\begin{equation*}
b_1 q \sigma_1(u)+\cdots+b_d q \sigma_d(u)=0,\quad \mbox{with~~} b_i\in K.
\end{equation*}
Now we assume that $\displaystyle\frac{b_j}{a_1}= \sigma_j\left(\frac{b_1}{a_1}\right)$ for all $2\leq j\leq d$.
\smallskip
Note that $b_1 \ne 0$. If not, then $0 =\sigma_j(b_1/a_1) = b_j/a_1$ for every $j$. Hence $b_i=0$ for all $i$, which contradicts Claim 1. By putting $\lambda = -b_1/a_1$, we rewrite \eqref{eq3.17} as
\begin{equation*}\label{eq3.18}
\tag{3.15}
p=-q(\sigma_1(\lambda)\sigma_1(u)+\cdots+\sigma_d(\lambda)\sigma_d(u)).
\end{equation*}
Since $b_j \in K$, it may happen that $\lambda$ does not belong to $k$. If so, then there exists an automorphism $\tau\in \mathcal{H}$ with $\tau(\lambda)\neq \lambda$. By applying the automorphism $\tau$ on both sides of \eqref{eq3.18} to eliminate $p$, we obtain the linear relation
$$
(\lambda-\tau(\lambda))q\sigma_1(u)+q\sum_{i=2}^d(\sigma_i(\lambda)\sigma_i(u)-\tau \circ\sigma_i(u))=0.
$$
Note that $\tau \circ\sigma_j$ coincides on $k$ with some $\sigma_i$ and since $\tau\in \mathcal{H}$ and $\sigma_2,\ldots,\sigma_d \not\in \mathcal{H}$, none of the $\tau \circ \sigma_j$ with $j\geq 2$ belongs in $\mathcal{H}$. Hence the above relation can be view as a linear combination of the $\sigma_i(u)$ with the property that the coefficients of $\sigma_1(u)$ will remain $\lambda-\tau(\lambda)$. By our assumption that $\lambda\notin k$ and hence we have $\lambda-\tau(\lambda)\neq 0$. Therefore, we obtain a non-trivial relation among $\sigma_i(u)$ of the form
$$
b_1 \sigma_1(u)+\cdots+b_d\sigma_d(u)=0,\quad \mbox{with~} b_i\in K.
$$
\smallskip
Therefore we can assume that $\lambda\in k$ and write \eqref{eq3.18} in the simpler form
\begin{equation*}\label{eq3.19}
\tag{3.16}
p=q\mbox{Tr}_{k/\mathbb{Q}}(\lambda u)=\mbox{Tr}_{k/\mathbb{Q}}(q\lambda u).
\end{equation*}
By adding $-\delta qu-\beta$ on both sides of the equality \eqref{eq3.19}, we get
$$
0<|p-\delta qu -\beta|=\left|-\beta+(\lambda-\delta)q\sigma_1(u)+q\sigma_2(\lambda)\sigma_2(u)+\cdots+q\sigma_d(\lambda)\sigma_d(u)\right|.
$$
Then from \eqref{eq3.1}, we get
\begin{equation*}\label{eq3.20}
\tag{3.17}
0<\left|-\beta+(\lambda-\delta)q\sigma_1(u)+q\sigma_2(\lambda)\sigma_2(u)+\cdots+q\sigma_d(\lambda)\sigma_d(u)\right|<\frac{1}{H^\varepsilon(u)q^{d+\varepsilon}}
\end{equation*}
holds for infinitely many pairs $(u,q)$ along the triples $(u,q,p)\in\mathcal{B}$.
\smallskip
Now we consider the case when $\lambda=\delta$. Then from \eqref{eq3.20}, we have
\begin{equation*}\label{eq3.21}
\tag{3.18}
0<\left|-\beta+q\sigma_2(\lambda)\sigma_2(u)+\cdots+q\sigma_d(\lambda)\sigma_d(u)\right|<\frac{1}{H^\varepsilon(u)q^{d+\varepsilon}}.
\end{equation*}
First we note that
$$
\max\{|\sigma_2(q\lambda u)|,\ldots,|\sigma_d(q\lambda u)|\}\geq \frac{|\beta|}{2d}
$$
holds for all but finitely many pairs $(u,q)$ satisfied \eqref{eq3.21}. If not, we assume that
$$
\max\{|\sigma_2(q\lambda u)|,\ldots,|\sigma_d(q\lambda u)|\}< \frac{|\beta|}{2d}
$$
for all pairs $(q,u)$ satisfied \eqref{eq3.21}. Then, we get
$$
\left|-\beta+q\sigma_2(\lambda)\sigma_2(u)+\cdots+q\sigma_d(\lambda)\sigma_d(u)\right|\geq \frac{|\beta|}{2}.
$$
Thus from \eqref{eq3.21}, we have
\begin{equation*}\label{eq3.23}
\tag{3.19}
\frac{|\beta|}{2}\leq \left|-\beta+q\sigma_2(\lambda)\sigma_2(u)+\cdots+q\sigma_d(\lambda)\sigma_d(u)\right|<\frac{1}{H^\varepsilon(u)q^{d+\varepsilon}}.
\end{equation*}
Since $H(u)\to\infty$ along infinitely many pairs $(u,q)$ satisfied \eqref{eq3.21} and $\beta$ is non-zero, we see that the inequality \eqref{eq3.23} can have only finitely many solutions in the pairs $(q,u)$. Therefore we conclude that
$$
\max\{|\sigma_2(q\lambda u)|,\ldots,|\sigma_d(q\lambda u)|\}\geq \frac{|\beta|}{2d}
$$
holds for all but finitely many pairs $(q,u)$ satisfied \eqref{eq3.21}.
This can be re-written as
$$
\max\{|q\sigma_2(u)|,\ldots,|q\sigma_d(u)|\}\geq \frac{\min\{1,|\beta|\}}{2d(\max\{|\sigma_2(\lambda)|,\ldots,|\sigma_d(\lambda)|\})}.
$$
Then from \eqref{eq3.21}, we conclude that
$$
0<\left|-\beta+q\sigma_2(\lambda)\sigma_2(u)+\cdots+q\sigma_d(\lambda)\sigma_d(u)\right|<\frac{1}{H^\varepsilon(u)q^{d+\varepsilon}}<\frac{C\max\{|q\sigma_2(u)|,\ldots,|q\sigma_d(u)|\} }{H^\varepsilon(u)q^{d+\varepsilon}},
$$
where $C=\frac{2d(\max\{|\sigma_2(\lambda)|,\ldots,|\sigma_d(\lambda)|\})}{\min\{1,|\beta|\}}$. Hence by Lemma \ref{lem2.1}, we get a non-trivial relation as we claimed.
\smallskip
Now we assume that $\lambda\neq \delta$. In this case the term $(\lambda-\delta)\sigma_1(u)$ does appear in \eqref{eq3.20}. Since we supposed that $|\lambda q u|>1$, we have
$$
\max\mbox\{|q\sigma_1(u)|,\ldots,|q\sigma_d(u)|\}\geq |qu|>|\lambda|^{-1} ,
$$
for infinitely many pairs $(u,q)$ along the triples $(u,q,p)\in\mathcal{B}$ . Then from \eqref{eq3.20}, we obtain
$$
0<\left|-\beta+(\lambda-\delta)q\sigma_1(u)+q\sigma_2(\lambda)\sigma_2(u)+\cdots+q\sigma_d(\lambda)\sigma_d(u)\right|<\frac{|\lambda|\max{|q\sigma_1(u)|,\ldots,|q\sigma_d(u)|}}{H^\varepsilon(u)q^{d+\varepsilon}}.
$$
Thus by applying Lemma \ref{lem2.1} with the inputs $\omega$ is an archimidean places corresponding to the embedding of $K$ defined by $\alpha\hookrightarrow {\alpha}$, $n=d$, $\lambda_0=-\beta, \lambda_1=\lambda-\delta$, $\lambda_i=\sigma_{i}(\lambda)$ for $2\leq i\leq d$, we get a non-trivial relation of the form
\begin{equation*}
b_1 \sigma_1(u)+\cdots+b_d\sigma_d(u)=0
\end{equation*}
valid for all triples $(u, q,p)\in \mathcal{B}_1$ for an infinite subset $\mathcal{B}_1\subset\mathcal{B}$.
\bigskip
\noindent{\bf Case 2.~} $\beta a_1/a_0+1\neq 0$.
\vspace{.2cm}
First we recall \eqref{eq3.11},
$$
0<\left|\delta q u-\left(\frac{\beta a_1}{a_0}+1\right)p-\beta\left(\frac{b_1}{a_0}q\sigma_1(u)+\cdots+\frac{b_d}{a_0}q\sigma_d(u)\right)\right|<\frac{1}{H^\varepsilon(u)}\frac{1}{q^{d+\varepsilon}}.
$$
The crucial point in Case 1 was the non-vanishing of the coefficient of $p$ in the inequality \eqref{eq3.16}. Since in this case we also have the similar situation as the coefficient $(\beta a_1/a_0+1)$ of $p$ is non-zero in this inequality, we follow the similar procedure to this inequality as we have done for the inequality \eqref{eq3.16} in Case 1.
\smallskip
For each $\rho\in\mbox{Gal}(K/\mathbb{Q})$, by \eqref{eq2.2}, we have
\begin{align*}
&\left|\left(\delta-\beta\frac{b_1}{a_0}\right)q\sigma_1(u) -\left(\frac{\beta a_1}{a_0}+1\right)p-\beta\left(\frac{b_2}{a_0}q\sigma_2(u)+\cdots+\frac{b_d}{a_0}q\sigma_d(u)\right)\right|^{d(\rho)/[K:\mathbb{Q}]}\\&=\left|\rho\left(\delta-\beta\frac{b_1}{a_0}\right)q\rho\circ\sigma_1(u)-\rho\left(\frac{\beta a_1}{a_0}+1\right)p-\rho(\beta b_2/a_0q\sigma_1(u))-\cdots-\rho(\beta b_d/a_0 q\sigma_d(u))\right|_\rho
\end{align*}
and hence
\begin{align*}
&\prod_{\mathit{v}\in M_\infty}\left|\rho_\mathit{v}\left(\delta-\beta\frac{b_1}{a_0}\right)q\rho_\mathit{v}\circ\sigma_1(u)-\rho_\mathit{v}(\beta a_1/a_0+1)p-\rho_\mathit{v}\left(\frac{\beta b_2}{a_0}\right)q\rho_\mathit{v}\circ\sigma_2(u)-\cdots-\rho_\mathit{v}\left(\frac{b_d}{a_0}\right) q\rho_\mathit{v}\circ\sigma_d(u)\right|_\mathit{v}\\&=\prod_{j=1}^d\prod_{\mathit{v}\in S_j}\left|\rho_\mathit{v}\left(\delta-\beta\frac{b_1}{a_0}\right)q\rho_\mathit{v}\circ\sigma_1(u)-\rho_\mathit{v}\left(\frac{\beta a_1}{a_0}+1\right)p-\cdots-\rho_\mathit{v}\left(\frac{\beta b_d}{a_0}\right) q\rho_\mathit{v}\circ\sigma_d(u)\right|_\mathit{v}.
\end{align*}
For each $\mathit{v}\in M_\infty$ and $j=1,2,\ldots,d$, we define $\mathit{v}(j)$ such that $\rho_\mathit{v}\circ\sigma_j=\sigma_{\mathit{v}(j)}$ on the field $k$, where $\{\mathit{v}(1), \ldots,\mathit{v}(d)\}$ is a permutation of $\{1,\ldots,d\}$.
By \eqref{eq3.2}, we see that
\begin{align*}
&\prod_{j=1}^d\prod_{\mathit{v}\in S_j}\left|\rho_\mathit{v}\left(\delta-\beta\frac{b_1}{a_0}\right)q\sigma_{\mathit{v}(1)}(u)-\rho_\mathit{v}\left(\frac{\beta a_1}{a_0}+1\right)p-\rho_\mathit{v}\left(\frac{\beta b_2}{a_0}\right)q\sigma_{\mathit{v}(2)}(u)-\cdots-\rho_\mathit{v}\left(\frac{\beta b_d}{a_0}\right) q\sigma_{\mathit{v}(d)}(u))\right|_\mathit{v}\\&=\left|\left(\delta-\beta\frac{b_1}{a_0}\right) q \sigma_1(u) -\left(\frac{\beta a_1}{a_0}+1\right)p-\beta\left(\frac{b_2}{a_0}q\sigma_1(u)+\cdots+\frac{b_d}{a_0}q\sigma_d(u)\right)\right|^{{\sum_{\mathit{v}\in M_\infty}}d(\rho_\mathit{v})/[K:\mathbb{Q}]}.
\end{align*}
Then, from \eqref{eq3.11} and the well-known formula $\displaystyle\sum_{\mathit{v}\in M_\infty}d(\rho_\mathit{v})=[K:\mathbb{Q}]$, it follows that
\begin{align*}
&\prod_{j=1}^d\prod_{\mathit{v}\in S_j}\left|\rho_\mathit{v}\left(\delta-\beta\frac{b_1}{a_0}\right)q\sigma_{\mathit{v}(1)}(u)-\rho_\mathit{v}\left(\frac{\beta a_1}{a_0}+1\right)p-\rho_\mathit{v}\left(\frac{\beta b_2}{a_0}\right)q\sigma_{\mathit{v}(2)}(u)-\cdots-\rho_\mathit{v}\left(\frac{\beta b_d}{a_0}\right) q\sigma_{\mathit{v}(d)}(u))\right|_\mathit{v}\\&=\left|\left(\delta-\beta \frac{b_1}{a_0}\right) q \sigma_1(u) -(\beta a_1/a_0+1)p-\beta\left(\frac{b_2}{a_0}q\sigma_2(u)+\cdots+\frac{b_d}{a_0}q\sigma_d(u)\right)\right|<\frac{1}{H^\varepsilon(u)}\frac{1}{q^{d+\varepsilon}}.
\end{align*}
Now for each $\mathit{v}\in S$, we define $d+1$ linearly independent linear forms in $d+1$ variables as follows: for $j=1,\ldots,d$ and for each $\mathit{v}\in S_j$, we define
\begin{align*}
L_{\mathit{v},0}(x_0, x_1,\ldots,x_d)&=-\rho_{\mathit{v}}\left(\frac{\beta a_1}{a_0}+1\right)x_0+\rho_\mathit{v}\left(\delta-\beta\frac{b_1}{a_0}\right)x_{\mathit{v}(1)}-\rho_{\mathit{v}}\left(\frac{\beta b_2}{a_0}\right)x_{\mathit{v}(2)}-\cdots-\rho_{\mathit{v}}\left(\frac{\beta b_d}{a_0}\right)x_{\mathit{v}(d)}
\end{align*}
and for $1\leq i\leq d+1$, we define
$$
L_{v, i}(x_1,\ldots,x_{d+1})=x_i,
$$
and for $\mathit{v}\in S\backslash{M_\infty}$ and for $0\leq i \leq d+1$, we let
$$
L_{v,i}(x_1,\ldots,x_{d+1})=x_i.
$$
Since in this case $\beta\frac{a_1}{a_0}+1$ is non-zero, we see that the linear forms $L_{\mathit{v},0},\ldots,L_{\mathit{v},d}$ are linearly independent for each $\mathit{v}\in S$. Finally, let $\mathbf{X}$ be the point in $K^{d+1}$, which is of the form
$$
{\mathbf{X}}=(p, q\sigma_1(u),\ldots,q\sigma_d(u)).
$$
Then by using Theorem \ref{schli} similar to Case 1 of this lemma, we get a non-trivial relation of the form
$$
a_1 p+b_1 q \sigma_1(u)+\cdots+b_d q \sigma_d(u)=0.
$$
Hence by follow the similar procedure as we have done in Case 1, we obtain a non-trivial relation of the form
$$
b_1 \sigma_1(u)+\cdots+b_d\sigma_d(u)=0,\quad \mbox{with~} b_i\in K,
$$
valid for all triples $(u, q,p)\in \mathcal{B}_1$ for an infinite subset $\mathcal{B}_1\subset\mathcal{B}$.
\bigskip
Thus by combining all the Cases, we obtain a non-trivial relation of the form
$$
b_1 \sigma_1(u)+\cdots+b_d \sigma_d(u)=0, \quad b_i\in K
$$
holds for infinitely many $u$ along the triples $(u,q,p)\in\mathcal{B}$. This proves our Claim 2.
We then conclude exactly as in \cite[Lemma 3]{corv}.
\end{proof}
\section{Proofs}
\noindent{\bf Proof of Theorem \ref{maintheorem}.~} Since $\Gamma$ is finitely generated multiplicative subgroup of $\overline{\mathbb{Q}}^\times$, we can reduce to the situation where $\Gamma\subset\overline{\mathbb{Q}}^\times$ is the subgroup of $S$-units, namely,
$$
\Gamma\subset\mathcal{O}^\times_S=\{u\in K:\prod_{\mathit{v}\in S}|u|_\mathit{v}=1\}
$$
of a suitable Galois extension $K$ over $\mathbb{Q}$ containing $\delta,\beta$ and for a suitable finite set $S$ of places of $K$ containing all the archimedean places. Also, $S$ is stable under Galois conjugation.
\smallskip
Suppose that the conclusion of Theorem \ref{maintheorem} is not true. Then there exist infinitely many triples $(u,q,p)\in \mathcal{B} \subset\Gamma\times \mathbb{Z}^2$ satisfying \eqref{eq1.1}. Then by inductively, we construct a sequence $\{\delta_i\}_{i=0}^\infty$ elements of $K$, an infinite decreasing chain $\mathcal{B}_i$ of an infinite subset of $\mathcal{B}$ and an infinite strictly decreasing chain $k_i$ of subfields of $K$ with the following properties:
\bigskip
{\it For each integer $n\geq 0$, $\mathcal{B}_n\subset (k_n\times\mathbb{Z})\cap\mathcal{B}_{n-1}$, $k_n\subset k_{n-1}$, $k_n\neq k_{n-1}$ and for all but finitely many triples $(u, q,p)\in\mathcal{B}_n$ satisfying
\begin{equation*}\label{eq4.1}
\tag{4.1}
|\delta_0\cdots\delta_n q u+\beta-p|<\frac{1}{H^{\varepsilon/(n+1)}(u)q^{d+\varepsilon}}.
\end{equation*}}
If such a sequence exists, then we eventually get a contradiction from the fact that the number field $K$ does not admit an infinite strictly decreasing chain of subfields. Thus in order to complete the proof of the theorem, it is enough to construct such a sequence.
\smallskip
We proceed our construction by applying induction on $n$: for $n=0$, put $\delta_0=\delta$, $k_0=K$ and $\mathcal{B}_0=\mathcal{B}$, and we are done in this case, since by our supposition the inequality
$$
|\delta_0 q u+\beta-p|<\frac{1}{H^\epsilon(u) q^{d+\eps}}
$$
has infinitely many solutions in triples $(u, q, p)$.
Then by the induction hypothesis, we assume that $\delta_n$, $k_n$ and $\mathcal{B}_n$ exist for an integer $n\geq 0$ such that \eqref{eq4.1} holds.
Then by Lemma \ref{lem3.1} to the choices $\delta=\delta_0\delta_1\cdots\delta_n$ and $k=k_n$, we obtain an element $\delta_{n+1}\in k_n$, a proper subfield $k_{n+1}$ of $k_n$ and an infinite set $\mathcal{B}_{n+1}\subset\mathcal{B}_n$ such that all triples $(u, q, p)\in\mathcal{B}_{n+1}$ satisfy $u=\delta_{n+1}u'$ with $u'\in k_{n+1}$.
Now, since $u'\in K$, $H(\delta_{n+1}u')\geq H(\delta)^{-1} H(u')$, we have in particular that for almost for all $u'\in K$, $H(\delta_{n+1} u')\geq H^{\frac{n+1}{n+2}}(u')$. Therefore by replacing $u$ by $\delta_{n+1}u'$, for all but finitely many triples $(u',q,p)\in\mathcal{B}_{n+1}$, we have the following inequality
$$
|\delta_0\delta_1\cdots\delta_n q \delta_{n+1}u'+\beta-p|<\frac{1}{H(u')^{\varepsilon/(n+2)}q^{d+\varepsilon}}
$$
This fulfil the induction and hence the theorem. $\hfill\Box$
\bigskip
\noindent{\bf Proof of Theorem \ref{maintheorem2}.~} Suppose the conclusion of Theorem \ref{maintheorem2} is not true. Then we may assume that $\alpha$ is an algebraic number.
\smallskip
Since $|\alpha|>1$, we have $|\lambda\alpha^n|>1$ for $n$ large enough. Now we choose $\epsilon'>0$ small enough so that $\varepsilon'<\varepsilon\log 2/\log H(\alpha)$. Then we get that for infinitely many $n$,
\begin{equation*}\label{eq4.2}
\tag{4.2}
0<\Vert \lambda\alpha^n+\beta\Vert<H(\alpha^n)^{-\varepsilon'}.
\end{equation*}
On the other hand by taking $\Gamma$ is the subgroup generated by $\alpha$ and $q=1$, $\delta=\lambda$ and $u=\alpha^n$, we see that the hypothesis of Theorem \ref{maintheorem} is satisfied. Thus from Theorem \ref{maintheorem}, the inequality \eqref{eq4.2} can have only finitely many solutions in $n$. This contradiction proves that $\alpha$ must be a transcendental number and hence the theorem.
\bigskip
\noindent{\bf Acknowledgements.} I would like to express my deep gratitude to Professor Yann Bugeaud for his valuable suggestions and comments in the earlier draft of this article. Also, I would like to acknowledge the Department of Atomic Energy, Govt. of India, for providing the research grant.
|
\section{Introduction}
Emerging markets have been commonly acclaimed for providing robust growth potential and offering investors a higher expected return compared to developed markets. Indeed, due to the possibility of higher profits and the low level of global equity markets integration, EMs have been considered as an investment opportunity for investors, whose aim to build an internationally diversified portfolio. EMs liquidity and transparency have continuously enhanced [\cite{mcguire2006common}; \cite{bunda2009correlations}]. Moreover, the reputation of EMs, in the framework of portfolio diversification, has received the attention of international investors, especially after the financial crisis that affected mostly developed markets.
However, the investment’s benefits of EMs come with additional risks, which are usually not as prominent in developed markets. In fact, EMs are exposed to additional economic, political, and currency risks. Further, the EMs’ economies fast growth and, as consequence, the quick evolution of their structure result in market information being rapidly outdated. Therefore, the existing more traditional methods of risk evaluation may be misleading in EMs, especially in short and medium investment horizons. Indeed, the existing risk measure methods are not suited to provide the up to date point of view representing current market structures, if they not supplemented with the latest market information. Hence, an efficient systemic risk measure is needed for EM.
For that purpose, it is crucial to understand and measure the spillover risk across EMs financial system network, which is important for financial risk measurement and portfolio diversification;\\
From the perspective of financial risk measurement, the interdependence among FIs becomes more important, especially during periods of distress, when losses spread through institutions, rendering the global financial system more vulnerable. In this regard, a systemic crisis that disturbs the financial system stability can have serious effects and lead to high losses for the entire economy and society.
From the perspectives of risk management and portfolio diversification, the contagion risk across the FIs from the same market, and across the worldwide markets, leads to a decrease in diversification potential. Hence, understanding the network structure of interdependence among FIs is crucial to risk managers and portfolio investors, as this can help them design investment strategies to reduce dependence risk and thereby increase diversification. Investors are also interested in recognizing the FIs that contribute the most (least) risk to their portfolio so that they are considered with caution in their portfolio design, especially during financial market turmoil.
To understand the FIs co-movements in EMs, our study examines the effect of Macro factors on the FIs in BRIMST (Brazil, Russia, India, Mexico, South Africa, Turkey) EMs, using daily interval data for the period between 2000–2020, with particular focus on the last two years. It is well known that the economics of the mentioned markets are strongly linked to the US economy [\cite{ozatay2009emerging}], so it is crucial to analyze the effects of US Macro factors on volatility in EMs financial system. It is worth to note that our dataset covers several global crisis periods, allowing us to examine how the EMs Financial systems respond to the different crisis. After determining the interdependencies among FIs and macroeconomic factors, our research aims also to build a robust strategy based on portfolio diversification in EMs.
In order to achieve the mentioned goals, our paper seeks to answer the following questions: Can the US and EMs Macro factors explain BRIMST financial equity indices? Are some categories of Macro factors more important than others? What FIs from EMs are the largest (smallest) spillover transmitters (receivers)? What FIs contribute the most (least) risk to total portfolio risk? What FIs offer greater diversification benefits? And lastly, how the tail spillover effect and portfolio weights change over time, and how they react to the different tail risk levels?
Answering all these questions is crucial, as international investors interested in understanding the forces behind the interdependence among macroeconomic factors and FIs, to identify potential risks and rewards and benefit from global diversification. Economic policymakers and regulators in BRIMST are interested in forces behind the co-movement between these markets to further establish market resilience in EMs.
We tackle these questions by employing the Financial Risk Meter (FRM) technology in EMs. The FRM is based on Lasso quantile regression designed to examine tail event co-movements financial securites. The objective is to understand the FIs interconnectedness and represent them in a network topology. Moreover, the FRM indices summarize the systematic risk at a given area and identify risk factors. Briefly, FRM represents tail event behavior in a financial risk factors network, which allows us (i) to identify the “stress emitters” and “stress receivers” companies, (ii) to measure the tail dependencies among the FIs and the Macro factors (iii) analyze the risk level in EMs over time. Concerning risk management, (iv) the FRM network is adopted in the portfolio selection process. More precisely, by interpreting the correlation coefficients of FIs equity indices in the adjacency matrix of the FRM network, an overlapping region between the portfolio optimisation strategies and FRM network is developed, (v) The FRM adjacency matrix is also adopted to lift the Hierarchical Risk Parity (HRP) approach [propsed by \cite{de2016building}] to the quantile level.
The remainder of this research is organized as follows. Section 2 presents a brief review of risk measurement and risk management methods. The econometric approach is discussed in Section 3, as well as a discussion on centrality measures. Section 4 illustrates the proposed portfolio optimization strategies. The FRM results are analysed in section 5. Section 6 develops the portfolio construction, while section 7 provides policy recommendations before concluding the general scope.\\
The codes are published on \url{www.quantlet.de} indicated by \includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg} with keyword FRM. The current level of FRM for Emerging markets as well as other channels can be found at \url{http://hu.berlin/frm}.
\section{Literature Review}
The large-scale breakdown of FIs after the global Lehman brother’s crisis in 2008 had caused serious social and economic losses. Therefore, controlling and maintaining financial system stability is one of the principals and the mutual responsibilities of central banks and financial market regulators around the world. Previous researches have been conducted in the framework of FIs interconnectedness. In this context, empirical studies commonly focus on resolving this issue in European countries; \cite{betz2016systemic}, \cite{abbassi2017credit}, and \cite{aldasoro2018multiplex}. Some studies have studied the FIs networks of developed American countries. Among others, \cite{cai2018syndication}; \cite{kreis2018systemic}, and \cite{tonzer2015cross}. Relatively few researchers have investigated bank networks of emerging American countries; \cite{silva2016network}; and \cite{berndsen2018financial}. All of the above studies focused on data from only one country and did not consider networks between FIs from several countries across the globe.
In order to bridge this gap, our study investigates the FIs networks from six among the biggest emerging countries worldwide, as globalization leads to closely linked economic activity among these countries in terms of trade and finance.
It is well known that FIs are interconnected within networks of several types of financial connections and contracts. The complex links among these institutions, which can be considered as risk factors, can cause systemic risk and results in spillover-effects that deteriorate the network stability and its functioning. These observations on the joint network dynamics motivated practitioners and researchers to insert tail events into risk management.
The Value-at-Risk (VaR) approach \cite{franke2019statistics} is frequently used to measure market risk, by computing the monetary loss of an institution for a given confidence level [\cite{slim2017value}]. However, the VaR measures a tail event probability hosting only one single node, which does not reflect its connection to overall systemic risk.
Recently, \cite{adrian2011covar} \& (\citeyear{tobias2016covar}) developed CoVaR to measure the systemic risk [e.g.\cite{fang2018identifying}, \cite{souhir2019price}]. The CoVaR measures spillover effects across financial markets by providing the VaR of one market under the condition that the other market is in financial distress as given by its VaR. Consequently, the CoVaR approach can gauge the size of the financial risk spillover. However, it can only capture the extent of risk spillovers for a simple bivariate system and cannot simultaneously measure the risk spillover effects across multiple financial markets.
\cite{hardle2016tenet} developed the Tail Event NETwork (TENET) risk approach by generalizing the CoVaR to be able to accommodate all system nodes as risk factors. TENET applies the quantile regression method on a set of network nodes stock market information and macroeconomic variables in a rolling window approach. The innovation of TENET model is to employ dimensional reduction (in a semi-parametric setting) by using the Lasso in a quantile regression framework \cite{tibshirani1996regression}. Also, \cite{chen2019tail} propose a Tail Event driven Network Quantile Regression (TENQR) to adress the interdependence, dynamics and riskiness of financial institutions. The TENQR captures risk dynamics within a panel quantile autoregression in a network topology, quantified through a time-varying adjacency matrix. TENQR technique is evaluated using the SIFIs (systemically important financial institutions). To extend the TENET further, \cite{mihoci2020frm} developed an improved systemic risk measure that summarizes the high-dimensional tail stress into a single real value indicator, called the Financial Risk Meter (FRM). The FRM is computed as an average of each node and at each time window selected penalization terms. The FRM level contains fundamental information about the active set of influential neighboring nodes and about the contributors to systemic risk. \\
\cite{yu2019ai} compared the proposed FRM to other measures for systemic risk, such as Google Trends, SRISK, and VIX. They found a Granger causality between the FRM and these measures, which confirmed the validity of the FRM as an efficient systemic risk measure. \cite{mihoci2020frm} applied the FRM to measure the dependencies between FIs and Macro factors exploiting tail event information. They built two FRM indices namely FRM@Americas and FRM@Europe, but also on bond yield and credit default swap spread co-movements. \\
So far, the FRM applications took into consideration financial securities from one economic region and its domestic Macro factors. In our study, we extend the application of FRM for EMs, and we will consider the domestic Macro factors from the BRIMST as well as those from the US (proposed by \cite{tobias2016covar} to represent both the domestic and foreign state of the economy.
In many portfolio construction approaches, the correlation between financial assets represents the basis for portfolio selection. To exemplify, \cite{markowitz1959portfolio} developed the modern portfolio theory (MPT). He found that when correlation between assets is not perfect, a diversified portfolio can be constructed. Therefore, to reach efficient diversification, investors should select anti-correlated assets and verify and ensure the satisfaction of this condition over time. But the correlation structure among assets changes over time and evolves especially during crises periods. For that reason, the Markowitz theory is usually oriented to select the most stable assets such as the industrial assets, and hence, the Markowitz optimal portfolio is often composed of a limited and invariant set of assets.
The second weakness of Markowitz's MPT [\cite{markowitz1959portfolio}] is related to the large estimation errors of the expected returns vector [\cite{merton1980estimating}], as well as of the covariance matrices [\cite{jobson1980estimation}]. Hence, a robust method for the modelling of the dynamic interconnectedness of assets is needed to support the MPT and guide investors in building an efficient and well diversified portfolio.
In this regard, recent researchers designed financial markets in networks (FMN), where assets are represented by nodes and the correlation of returns are represented by links that relate these nodes [\cite{chi2010network}; \cite{diebold2014network}; \cite{peralta2016network}]. Despite the originality and interesting results provided by this network approach, the majority of its applications are descriptive and lack concrete applications in the portfolio optimization procedure.
Recently, \cite{pozzi2013spread} extracted the dependency structure of financial equities from the network approach to build a diversified portfolio to reduce investment risk. This procedure visualized portfolio selection directly over the FMN design. \cite{peralta2016network} used the FMN as a powerful device to enhance the portfolio optimization procedure by selecting a set of assets according to their network centrality.
However, the adopted spillover and financial network methods focused on estimating the risk spillover based on the return distributions' first two conditional moments, thereby ignoring higher moments of the distribution (i.e. right and left tails). Indeed, the existing spillover measures concentrate only on the mean and variance of the distributions, which may underestimate the real spillover effects among FIs in tail-events, since they do not take into consideration the extreme risk spillover across financial markets. More specifically, previous researchers investigated mean-to-mean effect, assuming that investors are mean–variance optimizers. But ideally, a portfolio selection decision should be based on the entire return distribution estimation, since investors are more risk averse to the extreme downside risk related to the tails of the distribution.
To overcome this limitation, we adopt the FRM to measure the joint tail events. In fact, the FRM has the advantage to represent instantaneously the co-movements and the dynamics of high-dimension networks. Furthermore, the FRM can display the hidden interdependency structures between the financial network's nodes.
Given that the functions of complex FMN are reflected in the risk evaluation and portfolio optimization [\cite{haluszczynski2017linear}; \cite{WANG20181}], we contribute to this line of research by investigating the extent to which the underlying structure of this financial market tail event network can be used as an effective tool in enhancing the portfolio selection process. For that purpose, we establish a bridge between the FRM network and portfolio optimisation strategies. More precisely, we study the relationship between optimal portfolio weights and the FIs’ centrality in the FRM network. \vspace{\baselineskip}
The HRP is a second approach that aims to overcome the shortcomings of MPT. The Hierarchical risk parity (HRP) is firstly proposed by \cite{LopezdePrado59} as a risk parity allocation algorithm. The HRP applies machine learning and graph theory to extrapolate a hierarchical implementation of an inverse-variance allocation with weights computed among the formed correlated asset return clusters. By substituting the covariance structure with a hierarchical structure of clusters, HRP fully benefits from the covariance matrix information and improves the stability of the weights.
\cite{LopezdePrado59} shows that the HRP achieves higher risk-adjusted return and lower out-of-sample volatility than inverse-variance allocations. Despite its appealing features, the application of HRP is still rare, with increased interest \cite{lohre2020hierarchical}, \cite{risks7030074}, and \cite{Raffinot89} test the performance of HRP in a multi-asset allocation. Furthermore, \cite{Raffinot89} and \cite{Alipour2016} evaluate the performance of variant HRP.
However, the HRP focusses on the covariance matrix in computing the weights, consequently, it offers only a conditional mean view of the assets’ connectedness, which may underestimate the real spillover effects among assets in tail events. Taking into consideration tail spillover effects using quantile regression methods seeks to broaden this view, by providing a complete description of the stochastic relationship between assets and offering more robust and more efficient estimation \cite{Lingjie2008}. In this context, our paper contributes to the growing literature by employing an uplifted HRP based on FRM as a quantile regression method.
More precisely, our research introduces two modifications to the HRP algorithm of \cite{LopezdePrado59}: (i) Replace the covariance matrix with the FRM-adjacency matrix. Indeed, the HRP assign portfolio weights without the necessity to invert the covariance matrix. This propriety of HRP makes the replacement of the symmetric covariance matrix with the non-symmetric adjacency matrix possible. (ii) The HRP uses the inverse variance of each asset to calculate the optimal weights, and we replace the variance with the FRM's Lasso penalization parameter ($\lambda$). By introducing these two modifications, we build the uplifted HRP based on the FRM technology.
The HRP portfolio is then benchmarked against the classical HRP, the Minimum variance (MinVar) and inverse variance portfolio (IVP) optimization methods. \vspace{\baselineskip}
Our research investigates a comprehensive set of risk measurement and portfolio optimization in EMs, contributing a new dimension to the existing literature as follows:
(1) The FRM based Lasso quantile regression yields novel insights into the co-movement among FIs and US and EM domestic macroeconomic risk factors. The risk of spillover and its direction are also quantified and visualized, as well as how spillover risk evolves during the financial crisis. Therefore, our research paper is among a few investigating the risk spillover across BRIMST EMs.
To the best of our knowledge, there is a considerable gap in financial literature on the subject of tail risk spillover among US Macro factors and FIs in EMs. This paper is an attempt to bridge this gap by building our research on the above subject. For that purpose, we cover a sample of the biggest six emerging countries from different areas around the world which offers a reasonable basis for comparisons at country and regional market levels. Most of the earlier studies focus on fewer countries, mostly in one region. In addition, we adopt a sample that contains the largest FIs from each country as well as the domestic and US Macro factors to build our FRM network. Hence, our study investigates a comprehensive set of emerging equity markets, contributing a new dimension to the literature on international equity market co-movement that has traditionally focused on developed markets.
(2) Examine the existence of portfolio diversification benefits for foreigners investing in EMs.
Several studies have been conducted to examine the existence of portfolio diversification benefits in less correlated financial markets; \cite{jrfm7020045}; and \cite{SAITI2014196}. However, few other studies investigate this subject from the perspective of an EMs framework [\cite{ARREOLAHERNANDEZ2020101219}].
(3) This paper sheds light on the connection between the portfolio optimisation approaches and the financial tail event networks. Our strategy aims to simplify the portfolio selection procedure by targeting a set of assets within a certain range of network centrality. As far as we are aware, \cite{ARREOLAHERNANDEZ2020101219} is the only paper that attempts to take advantage of the topology of the tail event financial market network for investment purposes. They applied a directional spillover index, the tail-event driven network (TENET), and nonlinear portfolio optimization methods on the bank returns from emerging and developed America.
(4) While most previous studies focus on portfolio optimization based on mean variance estimation, there is a lack of empirical literature on quantile estimation (both tails of distribution). This study attempts to fill this gap by extending the classical HRP approach to build an optimal portfolio based on tail information provided by the FRM quantile regression.
(5) Finally, it is worth mentioning that our results are robust to different estimation time windows, market situations (pre-crisis, crisis, and post crisis), and tail risk levels. We argue that our proposed FRM based portfolio optimisation approach benefits from the tail event co-movement and asset clustering making more useful use of fundamental tail-event information, resulting in an efficient risk portfolio selection strategy. This practice is unique to this research and remains an important contribution to the literature on risk measurement and international portfolio diversification.
\section{Econometric Methodology}
\subsection{FRM Systemic Risk Measure Framework}
This section describes the genesis and framework of FRM. For that purpose, we lay down the fundamental structure and the background of our systemic risk analysis. More precisely, we will state the systemic risk measure models that have been evaluated to lead to an augmented risk measure, the FRM.
\subsubsection{FRM Genesis}
The Value at Risk (VaR) and the expected shortfall (ES) are traditional risk measures. They compute the risk of a given financial institution based either on company characteristics or by introducing macroeconomic variables as a proxy for the state of economy. To exemplify, the VaR of a financial institution $i$ at a quantile level $\tau$ is given by the following equation:
\begin{equation}
P(X_{i,t}\leq VaR_{i,t,\tau})\overset{def}{=}\tau, \quad\tau \in(0, 1)
\label{eq1}
\end{equation}
Where $X_{i,t}$ represent the log return of financial institution $i$ at time $t$.
\cite{adrian2011covar} proposed the Conditional Value at Risk (CoVaR), the CoVaR considers the spillover effects and the macro economy state. The CoVaR of $FI_j$ given $X_i$ at quantile level $\tau$ at time $t$ is given as follow:
\begin{equation}
P(X_{j,t}\leq CoVaR_{j|i,t,\tau}|R_{i,t})\overset{def}{=}\tau,
\label{eq2}
\end{equation}
Where $M_{t-1}$ denotes the vector of macro state variables, and
$R_{i,t}$ indicates the information set that involves the event of $X_{i,t}=VaR_{i,t,\tau}$ and $M_{t-1}$.
The CoVaR is estimated in two steps of linear quantile regression assuming the following equations:
\begin{equation}
X_{i,t}=\alpha_{i}+\gamma_{i}^{\top}M_{t-1}+\varepsilon_{i,t},
\label{eq3}
\end{equation}
\begin{equation}
X_{j,t}=\alpha_{j|i}+\beta_{j|i}X_{i,t}+\gamma_{j|i}^{\top}M_{t-1}+\varepsilon_{j,t},
\label{eq4}
\end{equation}
\begin{equation}
F^{-1}_{\varepsilon_{i,t}}(\tau|M_{t-1})=0\,\, \mbox{and}\,\, F^{-1}_{\varepsilon_{j,t}}(\tau|M_{t-1},X_{i,t})=0
\label{eq5}
\end{equation}
Where $\beta_{j|i}$ in Equation(\ref{eq4}) defines the sensitivity of log return of a company $j$ to variation in tail event log return of a company $i$.
\textbf{Step 1: Estimate the $VaR$ of an institution $i$ ($VaR_i$)}\\
To estimate the $VaR_i$, \cite{adrian2011covar} apply the quantile regression of log return of company $i$ on macro state variables. The estimated equation is defined as follow:
\begin{equation}
\widehat{VaR}^{\tau}_{i,t}=\widehat\alpha_{i}+\widehat\gamma^{\top}_{i}M_{t-1},
\label{eq6}
\end{equation}
\textbf{Step 2: Estimate the $CoVaR$}\\
The CoVaR is computed by integrating the $VaR_{i,t}^{\tau}$ in Equation (\ref{eq6}) into the CoVaR equation as follow:
\begin{equation}
\widehat{CoVaR}^{\tau}_{j|i,t}=\widehat\alpha_{j|i}+\widehat\beta_{j|i}\widehat{VaR}^{\tau}_{i,t}+\widehat\gamma^{\top}_{j|i}M_{t-1}.
\label{eq7}
\end{equation}
Where the coefficient $\beta{j|i}$ of Equation (\ref{eq7}) indicates the degree of interconnectedness.
Therefore, the risk of a financial company $j$ is computed through a $VaR$ and macro state of company $i$.
By setting $j$ equal to the return of a system, and $i$ to be the return of a financial company $i$, we obtain the contribution CoVaR that illustrates how a company $i$ effects the rest of the financial system. Or, by setting $j$ to be a financial company and $i$ to be a financial system, we obtain exposure $CoVaR$, which characterizes how a single institution is exposed to the overall risk of a system.
Recently, \cite{hardle2016tenet} developed the tail-event driven network (TENET). The TENET is a risk approach generalizes CoVaR by joining systemic interconnectedness between FIs and accommodates all system nodes as risk factors. It is estimated based on quantile regressions and illustrated by three steps.
\paragraph{Step 1: Estimate the $VaR$ of each FI}\
The $VaR$ for the studied FIs is estimated using the linear quantile regression based on Equations (\ref{eq3} and \ref{eq6}).
\paragraph{Step 2: The network analysis}\
The TENET estimates the non-linear relationship among FIs, and takes more institutions into consideration to compute the tail-driven risk interdependencies. Therefore, we have:
\begin{equation}
X_{j,t}=\mbox{g}(\beta_{j|R_{j}}^{\top}R_{j,t})+\varepsilon_{j,t},
\label{eq8}
\end{equation}
\begin{equation}
\widehat{CoVaR}^{TENET}_{j|\tilde{R}_{j,t,\tau}}\overset{def}{=}\widehat{\mbox{g}}(\widehat{\beta}_{j|\tilde{R}_{j}}^{\top}\tilde{R}_{j,t}),
\label{eq9}
\end{equation}
\begin{equation}
\widehat{D}_{j|\tilde{R}_{j}}\overset{def}{=}\frac{\partial \widehat{\mbox{g}}(\widehat{\beta}_{j|R_{j}}^{\top}R_{j,t})}{\partial R_{j,t}}|_{R_{j,t}=\tilde{R}_{j,t}}=\widehat{\mbox{g}}^{\prime}(\widehat{\beta}_{j|\tilde{R}_{j}}^{\top}\tilde{R}_{j,t})\widehat{\beta}_{j|\tilde{R}_{j}}
\label{eq10}
\end{equation}
where $R_{j,t}\overset{def}{=}\{{X_{\_j,t}},\,M_{t-1},\,B_{j,t-1}\}$ denotes the set of information. Note that $X_{\_j,t}\overset{def}{=}\{X_{1,t}\,X_{2,t}\,...,X_{1,t}\}$ is a set of $\{k-1\}$ independent explanatory variables, such as the log returns of all the FIs expected the $FI_j$, and $k$ here denotes the number of FIs. The term $B_{j,t-1}$ characterizes the internal factors of the institution $j$.The parameters are defined through the term $\beta_{j|R_{j}}\overset{def}{=}\{\beta_{j|\_{j}},\,\beta_{j|M},\,\beta_{j|B_{j}}\}^{\top}$.
The $\widehat{CoVaR}^{TENET}$ stands for tail-event driven network risk by means of a single-index model (SIM) model and is estimated by plugging in $VaR$ of institution $i$ at level $j|Rj,t$, defined in Equation (\ref{eq6}) into Equation (\ref{eq8}).\\
Note that $\beta_{j|R_{j}}\overset{def}{=}\{\beta_{j|\_{j}},\,\beta_{j|M},\,\beta_{j|B_{j}}\}^{\top}$ and $R_{j,t}\overset{def}{=}\{\widehat{VaR}^{\tau}_{i,t},\,M_{t-1},\,B_{j,t-1}\}.\widehat{\mbox{g}}(\bullet)$ characterize the non-linear relationship among them.
The parameter $D_{j|\tilde{R_j}}$ represents the gradient evaluating the marginal effect of covariates measured at $R_{j,t}=\tilde{R}_{j,t}$. The component wise expression is given by:\\ $\widehat{D}_{j|\tilde{R_j}}\equiv\{\widehat{D}_{j|\_{j}},\,\widehat{D}_{j|M},\,\widehat{D}_{j|B_{j}}\}^{\top}$.
Specifically, $\widehat{D}_{j|\_j}$ permits to quantity spillover effects through the FIs and to illustrate their evolution as a network system.
The TENET network represents a set of FIs links. The estimation results of this interconnectedness can be summarized in weighted form of an adjacency matrix.
Note $\widehat{D}^s_{j|i}$ an element in $\widehat{D}^s_{j|\_j}$ at estimation window $s$ for the $FI_j$ given another $FI_i$ (where $i$ is an element in the other $FI_j$).
Therefore, the adjacency matrix includes absolute values
of $\widehat{D}^s_{j|i}$ (in upper triangular matrix) and $\widehat{D}^s_{i|j}$ (in lower triangular matrix), where $\widehat{D}^s_{j|i}$ denotes the influence from a $FI_i$ to a $FI_j$ and $\widehat{D}^s_{i|j}$ is the influence from $FI_j$ to $FI_i$. The adjacency matrix $A_s=(k*k)$ is represented in the following equation, where for each time windows $s$ only one adjacency matrix is estimated.
\begin{equation}
A_{s} =
\begin{pNiceMatrix}[first-row,first-col,nullify-dots]
&I_1 & I_2 & I_3 & \cdots & I_k\\
I_1 & 0 & \lvert \widehat{D}^s_{1|2}\rvert & \lvert\widehat{D}^s_{1|3}\rvert &\cdots & \lvert\widehat{D}^s_{1|k}\rvert \\
I_2 & \lvert \widehat{D}^s_{2|1}\rvert & 0 &\lvert \widehat{D}^s_{2|3}\rvert & \cdots&\lvert \widehat{D}^s_{2|k}\rvert \\
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots \\
I_k & \lvert \widehat{D}^s_{k|1}\rvert & \lvert \widehat{D}^s_{k|2}\rvert & \lvert \widehat{D}^s_{k|3}\rvert & \cdots & 0
\end{pNiceMatrix}
\label{eq11}
\end{equation}
The matrix $A_s$ summarizes the overall connectedness between variables at time window $s$, and $I_i$ denotes the name of $FI_i$.
\paragraph{Step 3: Systemic risk contributions}\
The objective here is to measure the systemic risk relevance of a specific FI by its total in- and out-connections, weighted by market capitalization.
Hence, we define the Systemic Risk Receiver Index $(SRR)$ for a $FI_j$ at time windows $s$ as follow:
\begin{equation}
SSR_{j,s}\overset{def}{=} MC_{j,s}\Biggl\{\sum_{i\in {k_s^{IN}}}(\lvert \widehat{D}^s_{j|i}\rvert.MC_{i,s})\Biggl\}
\label{eq12}
\end{equation}
The Systemic Risk Emitter Index $(SRE)$ for a $FI_j$ at time windows $s$ is given by the following equation:
\begin{equation}
SSE_{j,s}\overset{def}{=} MC_{j,s}\Biggl\{\sum_{i\in {k_s^{OUT}}}(\lvert \widehat{D}^s_{i|j}\rvert.MC_{i,s})\Biggl\}
\label{eq13}
\end{equation}
Here $k_s^{IN}$ and $k_s^{OUT}$ denotes the set of FIs linked with the $FI_j$ at time windows $s$ by incoming and outgoing links respectively. $MC_{i,s}$ is the market capitalisation of FI $i$ at the starting point of time windows $s$. $\lvert \widehat{D}^s_{j|i}\rvert$ and $\lvert \widehat{D}^s_{i|j}\rvert$ are absolute partial derivatives derived from Equation (\ref{eq10}) which represents row (incoming) and column (outgoing) direction connection of $FI_j$. Thus both $SR{R_j,s}$ and $SRE_{j,s}$ take into consideration the $FI_j$ and its linked FI’ market capitalization as well as its connectedness within our network.
For more details see \cite{hardle2016tenet}.
\subsubsection{FRM Financial Risk Meter}
The FRM is a systemic risk measure based on the penalization parameter$(\lambda)$ of a linear quantile Lasso regression using moving-window approach. In this section we present the methodology and algorithm that constitutes the FRM technology. Since the penalization parameters $(\lambda)$ are computed based on an $L_1$-norm (Lasso) quantile linear regression, this method is introduced first.
\paragraph{Linear Quantile Lasso Regression Model}\
The FRM aims to simultaneously capture all interdependencies in one single number based on the log return of FIs and a set of macroeconomic variables that illustrate the state of the economy. Consider $J$ the number of FI under consideration, where $j\in\{1,...,J\}$. Therefore $p=J+M-1$ denotes the number of covariates. Let $t=\{1,...,T\}$ be the time point, where $T$ denotes the total number of daily observations. Suppose $s$ is the index of time windows, where $s\in\{1,...(T-(n-1))\}$ and $n$ is the length of windows size.
The Linear quantile Lasso regression for return series $X$ is defined as follow:
\begin{equation}
X^{s}_{j,t}=\alpha^{s}_{j,t}+A^{s\top}_{j,t}\beta^{s}_{j}+\varepsilon^{s}_{j,t}, \qquad
\label{eq14}
\end{equation}
Where $A^{s\top}_{j,t}\overset{def}{=}\begin{bmatrix}M^{s}_{t-1}\\
X^{s}_{-j,t}\end{bmatrix}$, $M^s_{t-1}$ is the $M$ dimensional vector of macro variables, $X^{s}_{-j,t}$ is the $p-M$ dimensional vector of log returns of all other FIs except $FI_j$ at time $t$ and in moving window $s$, $\beta^s_{j}$ is a $p\times 1$ vector defined for moving window $s$ and $\alpha^s_{j}$ is a constant term.
The regression is performed using an $L_1$-norm penalisation given a parameter $\lambda_j$, defined by \cite{tibshirani1996regression} as the least absolute shrinkage and selection operater (Lasso). According to \cite{bassett1978asymptotic}, the current company’s $\lambda_j$ are estimated by a modification of Lasso in a quantile regression [see \cite{belloni2011} and \cite{li20081} for more details), where the optimization is solved as follow:
\begin{equation}
\operatorname{\min_{\alpha^{s}_{j},\beta^{s}_{j}}}\Big\{n^{-1}\sum_{t=s}^{s+(n-1)}\rho_{\tau}\,(X_{j,t}^{s}-\alpha^{s}_{j}-A_{j,t}^{s\top}\beta_{j}^{s})+\lambda_{j}^s\parallel\beta_{j}^{s}\parallel_1\Big\}
\label{eq15}
\end{equation}
with the following check function:
\begin{equation}
\rho_\tau\,(u) = |u|^c \,|\tau - \mathrm{I}_{\{u<0\}}|
\label{eq16}
\end{equation}
where $c=1$ corresponds to quantile regression employed here and $c=2$ corresponds to expectile regression.
\paragraph{Penalization Parameter $\lambda$}\
The formula of Lasso’s penalization parameter $\lambda$ can be estimated in a linear regression context, following the work of \cite{osborne2000lasso}
Treating $\lambda$ as a fixed value in the objective function of the penalized regression:
\begin{equation}
f(\beta,\lambda)=\Biggl\{\frac{1}{2}\sum_{i=1}^n (y_{i}-X_{i}^{\top}\beta)^2+\lambda\sum_{i=1}^p|\beta_{j}|\Biggl\}
\label{eq17}
\end{equation}
Here $ f(\beta,\lambda)$ is convex in the parameter $\lambda$. In addition, with diverging $\beta$ we note that $f(\beta,\lambda)\to\infty$. Consequently, the function $f(.,\lambda)$ admits at least one minimum, which is attained in $\widehat{\beta}(\lambda)$ [\cite{osborne1985finite}] if and only if the null vector $0\in R^p$ is an element of sub-differential:
\begin{equation}
\frac{\partial f(\beta,\lambda)}{\partial \beta}=-X^{\top}(Y-X\beta)+\lambda u(\beta)
\label{eq18}
\end{equation}
Were:
\begin{equation*}
u(\beta)=(u_1(\beta),...,u_p(\beta))^{\top} \text{is defined as}
\begin{cases}
u_j(\beta)=1\; \; if\; \;\beta_j>0\\
u_j(\beta)=-1\; \; if\; \;\beta_j<0\\
u_j(\beta)\in[-1,\;1]; \; if\; \;\beta_j=0\\
\end{cases}
\end{equation*}
Therefore, considering that $ f(\beta,\lambda)$ admit a minimum in $\widehat{\beta}$, the following equation has to be satisfied:
\begin{equation}
0=-X^{\top}\{Y-X\widehat{\beta}(\lambda)\}+\lambda u(\widehat{\beta}(\lambda)
\label{eq19}
\end{equation}
The estimator of the vector of parameters $\beta$ here is a function of the penalization parameter $\lambda$. This dependency follows from the formulation of the penalized regression method and its objective function Equation (\ref{eq17}). Following this method we select first the penalization parameter $\lambda$ and then search for $\beta_{\lambda}$ that minimizes Equation (\ref{eq17}). Given that $u(\beta))^{\top} \beta=\sum_{j=1}^p\lvert\beta_j\rvert=\left \Vert \beta\right \|_1$, where $\left \Vert . \right \|_1$ represents here $L1$-norm of a $p$-dimensional vector, Equation (\ref{eq19}) can be rewritten in the following formula:
\begin{equation}
\lambda = \displaystyle \frac{\{ Y - X\beta(\lambda)\}^{\top}X\beta\left(\lambda\right)}{\left\|\beta\right\|_{1}}
\label{eq20}
\end{equation}
Looking to formula Equation (\ref{eq20}), we can define the elements that influence the value of $\lambda$ and its dynamics when treated in a time- varying framework. The following three elements affect the size of $\lambda$:\vspace{\baselineskip}
1. The size of residuals of the model;\vspace{\baselineskip}
2. The absolute size of the coefficients of the model, $\left \Vert\\ \beta\right \|_1$; \vspace{\baselineskip}
3. The singularity of a matrix $X^{\top}X$. \vspace{\baselineskip}
The formulae for the penalization parameter $\lambda$ in a quantile regression problem Equations (\ref{eq15} and \ref{eq16})can be derived following \cite{li20081}:
\begin{equation}
\lambda = \displaystyle\frac{ \left( \theta \right)^{\top}X\widehat{\beta}\left(\lambda\right)}{\left\|\widehat{\beta}(\lambda)\right\|_{1}}
\label{eq21}
\end{equation}
Where $\theta=(\theta_1,...,\theta_n)^{\top}$ satisfies the following conditions:
\begin{equation*}
\theta_i=\begin{cases}
\tau\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;if\;\;\;\;Y_i-X_i^{\top}\widehat\beta(\lambda)>0\\
-(1-\tau)\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;if\;\;\;\;Y_i-X_i^{\top}\widehat\beta(\lambda)<0\\
\in(-(1-\:\tau)\;,\tau)\;\;\;\;if\;\;\;\;Y_i-X_i^{\top}\widehat\beta(\lambda)=0\\
\end{cases}
\end{equation*}
Since Equation (\ref{eq15}) has an $L_1$ loss function and an $L_1$-norm penalty term, the optimization problem is an $L_1$-norm quantile regression estimation problem. The selection of the penalization parameter $\lambda^s_j$ is fundamental. There are different possibilities to choose $\lambda^s_j$, for example with the Generalized Approximate Cross- Validation criterion (GACV) or the Bayesian Information Criterion $(BIC)$. In this regards, \cite{yuan2006gacv} conducted simulations and concluded that GACV outperforms BIC in terms of statistical efficiency. Hence, we estimate $\lambda^s_j$ with the GACV criterion in the FRM model and set $\lambda^s_j$ as the solution of the following minimization problem:
\begin{equation}
\operatorname{min}GACV(\lambda_{j}^{s})=\operatorname{min}\frac{\sum^{s+(n-1)}_{t=s}\rho_{\tau}(X_{j,t}^{s}-\alpha^{s}_{j}-A_{j,t}^{s,\top}\beta_{j}^{s})}{n-df}
\label{eq22}
\end{equation}
where $df$ stands for the estimated effective dimension of the fitted model. The advantage of GACV is that it can be also adopted for $p > n$, which can be crucial for the FRM if the moving window size is small. For further details see \cite{mihoci2020frm}.
\vspace{\baselineskip}
\paragraph{Financial Risk Meter (FRM) }\
The FRM is estimated using a regression analysis as explained above and select the $\lambda_j$ for each FI $j$ using GACV. The distribution of the$\lambda^s_j$ in a moving window gives important information on the network dependencies among the financial nodes. The standard FRM is defined as the average of $\lambda_j$ over the set of $J$ FIs for all windows. It is formally written as follows:
\begin{equation}
FRM=J^{-1}\sum_{j=1}^J\lambda_j
\label{eq23}
\end{equation}
Note that the distribution of $\lambda_j$ provides details regarding the overall market movement and gives information to macro-prudential decision makers about the network dynamics. Therefore, the FRM is adopted to identify the high joint tail event risks resulting from each FI. More precisely, the FI with high $\lambda_j$ exhibits common high stress levels as the FI at the origin of the crisis. Therefore, this FI is considered as having high ”co-stress”.
\vspace{\baselineskip}
\subsection {FRM Network and node centrality measures}
The FRM network illustrates the tail event interaction between the selected FIs based on the adjacency matrix $A_s$ (11). To describe the topology of the FRM networks, we focus on node centrality measures, specifically, eigenvalue centrality, degree centrality, indegree, outdegree, betweenness and closeness. The concept of centrality aims to measure the impact and the importance of a given node in the network. \vspace{\baselineskip}
\paragraph{Eigenvector centrality}\
\cite{bonacich1972factoring} proposed the so-called eigenvector centrality, which evolved to be a standard measure in network analysis.
Consider a network $G = \{N, \omega\}$, constituted by a set of links $\omega$ that connecting pairs of nodes and a set of nodes $N = {1, 2, ... , N}$. If there is a connection between two nodes $i$ and $j$, we denote it as $(i,j)\in\omega$. The network connections are defined by a simplified version of (11): the $(N\times N)$ adjacency matrix $A_{i,j}^s=[\beta_{ij}^s]$ whose element $\beta_{ij}\neq 0$ whenever $(i,j)\in \omega$, $s$ is the rolling window.
\begin{equation}
A^s =
\begin{pmatrix}
0 & \beta_{1,2} & \cdots & \beta_{1,n} \\
\beta_{2,1} & 0 & \cdots & \beta_{2,n} \\
\vdots & \vdots & \ddots & \vdots \\
\beta_{n,1} & \beta_{n,2} & \cdots & 0
\end{pmatrix}
\label{eq24}
\end{equation}
Based on the \cite{bonacich1987power}; \citeyearpar{bonacich1972factoring} definition, the eigenvector centrality of node $i$, $v_i$, is expressed as the proportional sum of its neighbors' centralities. This concept has been extended by \cite{bonanno2004networks} suggesting that for the weighted networks, $v_i$ is relative to the weighted sum of neighbors’ centralities of node $i$ with the adjencency matrix $A_{i,j}^s$ that illustrates the corresponding weighting factors. Therefore, the eigenvector centrality of node $v_i$ is calculated as follows:
\begin{equation}
v_i=\delta^{-1}\sum_j\beta_{ij}^s v_j
\label{eq25}
\end{equation}
where $\delta$ here denotes the eigenvalue. A large value of $v_i$ means that the node $i$ is highly central, implying that node $i$ is linked either to several other nodes or is linked to a few highly central nodes.
The Equation (\ref{eq25}) can be rewritten in matrix terms. We have $\delta v=A v$ specifying that the centrality vector $v$ is defined by the eigenvector of $A^s$ corresponding to the largest eigenvalue $\delta$.\vspace{\baselineskip}
In general, each eigenvector of $A^s$ is a solution to Equation (\ref{eq25}). Nevertheless, the centrality vector matching to the largest network component is specified using the eigenvector corresponding to the largest eigenvalue \cite{bonacich1972factoring} and \cite{peralta2016network}.
\paragraph{Closeness}\
Closeness is a centrality measure proposed by \cite{freeman1978centrality}. It highlights the distance of a given vertex to the rest of vertices in the network. It can be considered as duration of the information spread from one vertex to another. Closeness of a given vertex $j$ is defined as follow:
\begin{equation}
Closeness_j=\sum{_{i=1}^N \frac{1}{d_{(i,j)}}}
\end{equation}
Where $d_{(i,j)}$ measure the distance between a vertix $i$ and another vertix $j$ in the network.
\paragraph{Betweenness}\
Betweenness is a centrality measure, which computes the number of times a vertex lies on the shortest path between other vertices in the network [\cite{vyrost2019network}].\\
Suppose we need to compute this measure for a vertex $j\in\omega$, for any two distinct vertices other than $j$, named $l$ and $k$ $\in\omega$, the number of shortest paths between $l$ and $k$ is $n_{l,j,k}$ $\in N$ and the total number of shortest paths between $l$ and $k$ is $n_{l,k}\in N$, the betweenness for $j$ is then computed as follow:
\begin{equation}
Betweenness_j=\sum_{\substack{l\neq j\neq k{} \\ l,j,k\in \omega}} \frac{n_{l,j,k}}{n_{l,k}}
\label{eq26}
\end{equation}
\paragraph{Degree centrality}\
Degree centrality captures total connectedness in the network, it is given by the following equation:
\begin{equation}
D=\sum_{j=1}^N \sum_{i=1}^N \textbf{1} (\beta_{j,i}^s)
\label{eq27}
\end{equation}
where
\begin{equation*}
\textbf{1} (\beta_{j,i}^s)=\begin{cases}
1\; \; if\; \;\beta_{j,i}^s\neq 0\\
0\; \; if\; \;\beta_{j,i}^s =0\\
\end{cases}
\end{equation*}
\paragraph{Indegree:} Indegree computes the number of inflows, or in this paper the number of other FIs influencing one node. Indegree of $FI_j$ is defined as:
\begin{equation}
Ind_j=\sum_{i=1}^N\textbf{1} (\beta_{j,i}^s)
\label{eq28}
\end{equation}
Here the $FI_j$ can be considered as a risk receiver.
\paragraph{Outdegree:} Outdegree computes the number of out-going links, or in this paper the number of other FIs influenced by the node. Outdgree of $FI_i$ is defined as:
\begin{equation}
Outd_i=\sum_{j=1}^N\textbf{1} (\beta_{j,i}^s)
\label{eq29}
\end{equation}
Here the $FI_i$ can be considered as a "risk emitter".
\section{Portfolio optimisation}
\subsection{The minimum-variance}\
Consider $N$ risky assets with a vector of expected returns $\mu$, and covariance matrix, $\Sigma=[\sigma_{ii}]$. The problem is to define the optimal portfolio weights vector $w$, which minimizes the portfolio variance subject to $w^{\top}\textbf{1}=\textbf{1}$ where $\textbf{1}$ represents a column vector whose components are equivalent to one \cite{markowitz1959portfolio}. This approach is usually defined as minimum-variance or shortly $MinVar$. Formally the problem is stated as follow:
\begin{equation}
min_w\sigma^2_p=w^{\top}\Sigma{w};\;\;\;\;\;\text{subject to}\;\;\;\;w^{\top}\textbf{1}=\textbf{1}
\label{minvareq}
\end{equation}
The solution of Equation (\ref{minvareq}) is:
\begin{equation}
\widehat{w}^*_{minv}=\frac{\textbf{1}}{\textbf{1}^{\top}\Sigma^{-1}1}\Sigma^{-1}\textbf{1}
\label{minvarweights}
\end{equation}
\subsection{A machine learning-based Hierarchical Risk Parity HRP}
\subsubsection{Classical HRP approach}
The HRP is a risk-based portfolio optimization approach that diversifies portfolios without imposing a positive-definite return covariance matrix [\cite{LopezdePrado59}]. The algorithm employs machine learning methods and graph theory to classify the underlying hierarchical correlation structure of the assets, which allow clusters of similar assets to compete for capital allocation. The algorithm of the HRP approach can be broken down into three main steps: tree clustering, quasi-diagonalization, and recursive bisection. In the following, we explain each step in detail. \vspace{\baselineskip}
\paragraph{Step 1- Hierarchical Tree Clustering}\
This step involves breaking down the assets of the considering portfolio into different hierarchical clusters exploiting a Hierarchical Tree Clustering algorithm. The clusters are formed as follows:\\
1. For $N$ stock returns compute the correlation matrix [see \cite{hardle2019mva}(p 431-442)], which gives an $N\times N$ matrix $\Omega$ of these correlations $\rho$, \\
$\rho= \{ \rho_{i,j}\}_{i,j=1,...,N}$\\
Where $\rho_{i,j}$ is the correlation coefficient between a pair of assets $\{i,j\}$ $\rho[X_{i},X_{j}]$\\
2. Transform the correlation matrix to a correlation-distance matrix $D$, where for\\
$d:(X_{i},X_{j}) \subset B \rightarrow {R} \in [0,1];$
\begin{equation}
d_{i,j} = d[X_{i},X_{j}]=\sqrt{\frac{1}{2}(1-\rho_{i,j})}
\label{eq32}
\end{equation}
3. Compute a new distance matrix $\tilde{d}$ where, by taking the Euclidean distance among columns in a pair-wise manner, the augmented distance matrix is given as follow:
\begin{equation}
\tilde{d}= \tilde{d}[d_{i},d_{j}] = \sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^2}
\label{eq33}
\end{equation}
where $\tilde{d}_{i,j}: (d_{i},d_{j}) \subset B \rightarrow {R} \in [0, \sqrt{N}]$.\\
Note that for two assets $i$ and $j$, $ D_{i,j}$ represents the distance between them, however $\tilde{d}_{i,j}$ represents the closeness in similarity of $\{i,j\}$ with the rest of the portfolio. More precisely, a lower $\tilde{d}_{i,j}$ indicates that the assets $i$ and $j$ are similarly correlated to the rest of stocks in the portfolio.\\
4. Form the assets clusters in a recursive manner based in Equation(\ref{eq33}). The set of clusters is donated by $U$ and the first formed cluster$(i^*,j^*)$ is define as:
\begin{equation}
U[1]=(i^{*},j^{*}) =\text{argmin}(i,j)_{i \neq j}\{\tilde{d}_{i,j}\}
\label{eq34}
\end{equation}
5. Update the distance matrix $d$ by computing the distances of other assets from the newly formed cluster $U(1)$ using single linkage clustering. Therefore, for any asset $i$ outside of $U(1)$, the distance from $U(1)$ is updated as follows:
\begin{equation}
d_{i,U[1]}=\min\bigl[\{\tilde{d}_{i,j}\}_{j \in U[1]}\bigl]
\label{eq35}
\end{equation}
Thereby, the algorithm recursively forms assets clusters and updates the distance matrix until we are left with one giant cluster of all stocks.
Finally, the clusters are visualised in a dendrogram. See \cite{hardle2019mva} (p. 363-393) for more details. \vspace{\baselineskip}
\paragraph{Step 2- Quasi Diagonalisation or Matrix Seriation}\
The Quasi-Diagonalisation of the covariance matrix or Matrix seriation is adopted to rearrange the data to clearly represent the inherent clusters. Using the hierarchical clusters from the previous step, the columns and rows of the covariance matrix are reorganized so that similar assets are placed together and dissimilar assets are placed far apart. More precisely, the larger covariances are positioned along the diagonal and smaller ones around this diagonal and since the off-diagonal elements are not completely zero. This is named a quasi-diagonal covariance matrix.\vspace{\baselineskip}
\paragraph{Step 3- Recursive Bisection}\
The final recursive bisection step implicates assigning actual portfolio weights to the assets in the portfolio.\\
1. Assign a unit weight to all assets,
\begin{equation*}
W_i=1,\;\;\;\forall\;i=1,...N
\end{equation*}
2. Bisect each cluster into two sub-clusters in a top-down manner, it means by starting with the topmost cluster, so, for each cluster, we obtain a left and right sub-cluster.\\
3. Calculate the variance for each of these sub-cluster
\begin{equation}
V_{1,2}=w^{\top}\Sigma w\;\;\;\;; \Sigma \;\;\text{is the covariance matrix}
\label{eq36}
\end{equation}
where,
\begin{equation}
w= \frac{\mbox{diag}[\Sigma]^{-1}}{\mbox{tr[diag}[\Sigma]^{-1}]}
\label{IVPweights}
\end{equation}
Since we are dealing with a quasi-diagonal matrix, the algorithm uses the property of the portfolio that the inverse-variance allocation is optimal for a diagonal covariance matrix. Hence, we adopt the inverse-variance allocation weights when computing the variance for sub-clusters.\\
4. Calculate the weighting factor based on the quasi-diagonalised covariance matrix
\begin{equation}
\alpha_{1}=1-\frac{V_{1}}{V_{1}+V_{2}}, \text{so \ that} \ 0 \leq \alpha_{i} \leq 1\; \; ;\alpha_2=1-\alpha_1
\label{eq38}
\end{equation}
5- Update the weights $w_1$ and $w_2$ for both sub-clusters:
\begin{equation}
w'_1=\alpha_1\cdot w_1;\;\;\;\
w'_2=\alpha_2\cdot w_2
\label{eq39}
\end{equation}
6- Execute recursively steps 2-5. The algorithm stops when we have a single asset for each cluster and then the weights are assigned to all assets in the portfolio.\\
Since the weights are assigned in a top-down manner, only assets within each cluster compete for allocation for the final portfolio, rather than competing with all the assets in the portfolio, see \cite{vyrost2019network}.\vspace{\baselineskip}
\subsubsection{An uplifted HRP approach based on FRM}
The classical HRP approach focused on calculating the optimal portfolio weights based on the variance and the covariance matrix, in other words it is focused on estimating the spillover risk on the mean-variance levels, ignoring analysis of other quantile levels. To overcome this limitation, we extend the classical HRP based on FRM in order to take into consideration the hidden interdependency structures among the FIs tail quantile level when optimizing the EMs portfolio.
The basic idea is to use the FRM output in order to compute the portfolio weights based on quantile level analysis. In the classical approach, these weights are computed using the variance (which is a measure of individual risk) and the covariance matrix that measures the relationship between a pair of assets in the mean-variance level. To improve on this, the FRM provides the penalty parameters $(\lambda_{j})$ as a measure of individual risk of each company or FI at the quantile level. In addition, the FRM provides also the adjacency matrix to estimate the interconnectedness between each pair of the studied FIs $(\beta_{i,j})$. By analogy, for the uplifted HRP approach based FRM, we replace the covariance matrix with the adjacency matrix. Recall that the diagonal elements of the adjacency matrix are zero (Equation (\ref{eq24})), which allow us to introduce the vector of penalty parameters $(\lambda_{j})$ in the diagonal of the adjacency matrix to replace the variance in the classical approach.
Following this reasoning, the uplifted HRP steps will be as follows:\vspace{\baselineskip}
\paragraph{Step 1- Hierarchical Tree Clustering}:\\
1. Calculate the the FRM adjacency matrix $A^s$ (Equation (\ref{eq24}))
Where $\beta_{i,j}$ is the degree of connectedness between a pair of assets $\{i,j\}$ \\
2. Introduce the vector of $(\lambda_{j})$ as diagonal elements in the adjacency matrix $A^s$, so we obtain the new adjacency matrix $\tilde{A}$
\begin{equation}
\tilde{A^s} =
\begin{pmatrix}
\lambda_1 & \beta_{1,2} & \cdots & \beta_{1,n} \\
\beta_{2,1} & \lambda_2 & \cdots & \beta_{2,n} \\
\vdots & \vdots & \ddots & \vdots \\
\beta_{n,1} & \beta_{n,2} & \cdots & \lambda_n
\end{pmatrix}
\label{eq40}
\end{equation}
3. Replace the covariance matrix in the HRP algorithm with the adjacency matrix $\tilde{A^s}$.\\
4. Transform $\tilde{A^s}$ to an adjacency-distance matrix $D$,
\begin{equation}
D_{i,j} = D[X_i, X_j]=
\begin{cases}
\sqrt{\frac{1}{2}(1-\beta_{i,j})}\; \;\; \; if \; \;\; \; \vert \beta_{i,j}\vert <1\\
\; \; \\
\mbox{max}D[X_i, X_j]\; \;\; \;\; \;\; \;\; \; otherwise.
\end{cases}
\label{eq41}
\end{equation}
5. Compute a new distance matrix $\tilde{D}$
\begin{equation}
\tilde{D}= \tilde{D}[D_{i},D_{j}] = \sqrt{\sum_{n=1}^{N}(D_{n,i}-D_{n,j})^2}
\label{eq42}
\end{equation}
6. Form the assets clusters in a recursive manner based in Equation(\ref{eq42}). The set of clusters is donated by $\tilde{U}$ and the first formed cluster$(i^*,j^*)$ is define as:
\begin{equation}
\tilde{U}[1]=(i^{*},j^{*}) =\text{argmin}(i,j)_{i \neq j}\{\tilde{D}_{i,j}\}
\label{eq43}
\end{equation}
7. Update the distance matrix $D$ by computing the distances of other assets from the newly formed cluster $\tilde{U}(1)$ as follows:
\begin{equation}
D_{i,\tilde{U}[1]}=\min\bigl[\{\tilde{D}_{i,j}\}_{j \in U[1]}\bigl]
\label{eq44}
\end{equation}
\paragraph{Step 2- Quasi Diagonalisation or Matrix Seriation}:\\
The Quasi-Diagonalisation of the adjacency matrix or Matrix seriation is adopted to rearrange the data to represent clearly the inherent clusters (as explained in the classical approach).
\paragraph{Step 3- Recursive Bisection}:\\
1. Assign a unit weight to all assets,
\begin{equation*}
W_i=1,\;\;\;\forall\;i=1,...N
\end{equation*}
2. Bisect each cluster into two sub-clusters.\\
3. Calculate the variance for each sub-cluster
\begin{equation}
\tilde{V}_{1,2}=\tilde{w}^{\top}\tilde{A} \tilde{w}\;\;\;\;; \tilde{A} \;\;\text{is the adjacency matrix}
\label{eq45}
\end{equation}
where,
\begin{equation}
\tilde{w}= \frac{\mbox{diag}[\tilde{A}]^{-1}}{\mbox{tr[diag}[\tilde{A}]^{-1}]}
\label{eq46}
\end{equation}
4. Calculate the weighting factor based on the quasi-diagonalised adjacency matrix.
\begin{equation}
\tilde{\alpha}_{1}=1-\frac{\tilde{V}_{1}}{\tilde{V}_{1}+\tilde{V}_{2}}, \text{so \ that} \ 0 \leq \tilde{\alpha}_{i} \leq 1\; \; ;\tilde{\alpha}_2=1-\tilde{\alpha}_1
\label{eq47}
\end{equation}
5- Update the weights $\tilde{w}_1$ and $\tilde{w}_2$ for both sub-clusters:
\begin{equation}
\tilde{w}'_1=\tilde{\alpha}_1\cdot w_1;\;\;\;\
\tilde{w}'_2=\tilde{w}{\alpha}_2\cdot w_2
\label{eq48}
\end{equation}
6- Execute recursively steps 2-5, the algorithm stops when we have a single asset for each cluster and then the weights are assigned to all assets in the portfolio.\vspace{\baselineskip}
\section{Financial Risk Meter in Emerging Markets}
\label{FRMEMinterp}
\subsection{FRM@EM data-set description}
In this paper we study the largest 25 Emerging Market financial institutions (FIs) by market capitalisation at any given point in time, with focus on the BRIMST FIs. We compile a database of daily price levels in as well as market capitalisations in U.S. Dollars from Bloomberg, and select the biggest financial institutions on a daily basis from the most liquid local EM equity market indices. On any given trading day in consideration, we take the price returns of those biggest $j=25$ FIs over an estimation window $s=63$ business days.
As for the macroeconomic data, we follow \cite{tobias2016covar} concerning the developed market specific risk factors, to repeat, returns in US REITs, S\&P 500 index, U.S. three months treasury bill rates, the spread between 3 months and 10 year U.S. treasury rates, the spread between BAA rates corporate bonds by the rating agency Moody's to U.S. treasury bonds, and the implied volatility index VIX based on outstanding options on the S\&P 500 equity index. We add the following EM specific macroeconomic risk variables: The J.P. Morgan Emerging Bond Index Global Sovereign Spread index tracking the Emerging Bond Index yields over the benchmark U.S. Treasury bonds and thereby representing the risk compensation demanded from investors when investing in EM sovereign bonds, as well as the respective countries currency versus the U.S. dollar cross.
\subsection{FRM@EM Interpretation and Network Analysis}
Figure \ref{FRMEMtimeseriestotal} depicts the FRM@EM from April 2000 to June 2020. Clearly observable are the periods of distress in the global financial system around 2008, 2012, more recently in 2020 but also during the EM specific market distress periods such as 2002 (Argentina) and 2013 (following the Federal Reserve Board's Open Market Committee forward guidance). \\
Figure \ref{FRMEMboxplot} introduces one of the FRM's tool: FRM is not just the mean but actually a distribution of $\lambda^s_j$ as well as the adjacency matrix containing all (Lasso penalised) $\beta^{s}_{j}$ between FIs, and also between FIs and macroeconomic risk variables. As an example of the information contained, we mark some of the more extreme maxima in Figure \ref{FRMEMboxplot}. For example, into the crisis, Standard Bank Group (SBK SJ) had a very high $\lambda^s_j$ reading, indicating the bank was a "risk receiver", thus at risk to be impacted by spill over effects. \\
But we can also have a more detailed look at the adjacency matrix containing the $\beta^{s}_{j}$ between FIs and macroeconomic risk variables, with two examples shown for 20200429 in Figure \ref{ADJMAT20200429005} as well as post crisis 20200630 in Figure \ref{ADJMAT20200630005}, both estimated at $\tau=0.05$, and colour scaled in blue (low/negative) to high (red/positive). Observable are the high dependencies between countries of the same region (\textbf{B}S for Brazil, \textbf{R}M for Russia, \textbf{I}S for India, \textbf{M}F for Mexico, \textbf{S}J for South Africa and \textbf{T}I for Turkey), and often negative relationships (adjusted for co-movements with macroeconomic risk variables) in between regions. However, there are also detectable co-dependencies, which necessitate closer inspection from both investors as well as regulators. For example in Figure \ref{ADJMAT20200429005} South Africa's Sanlam Ltd. (SLM SJ)'s returns are explained to a significant extend by Bajaj Finance Ltd. of India (BAF IS), as both financial services companies, providing an assortment of financial services. Similarly, Banco BTG Pactual (BPAC11 BS)'s returns are explained not only by other Brazilian FIs, but also significantly by Russia's VTB Bank (VTBR RM), both operating in more banking and investment banking related markets. Clearly, sub-sector dependencies across EM FIs are to be considered to prevent risk clusters. \\
Another component to consider is the impact from macroeconomic risk variables' changes on FIs returns. As can be seen on both adjacency matrices, the "classic" macroeconomic risk variables do have an impact, however, to a large extend, the Emerging Market Sovereign Yield Spread to U.S. Treasuries (JPEGSOSD) is the main influence. In fact, this is true across from $\tau=0.05$ to $\tau=0.50$. In Figure \ref{fig:macro} we show the smoothed (rolling seven day mean) of the share of FIs impacted by the respective macroeconomic risk variable. The Emerging Market Yield spread is the dominant driver, followed by more general market risk measures such as the VIX, Moody's Baa corporate yield spreads, and the shape of the U.S. Treasury yield curve. Emerging market currencies as a cross versus the U.S. Dollar have a lesser impact. It is mostly one or two EM currencies having an overall impact, and not only on domestic banks. In Figure \ref{ADJMAT20200630005} for example, the Brazilian Real (USDBRL) has marginal negative return contribution to Bajaj Finance Ltd in India (BAF IS), and positive return contribution to Brazilian Itausa SA (ITSA4 BS) and Mexican Grupo Elektra SAB (ELEKTRA MF). \\
In Figure \ref{FRMcentmeasures}, we show the time series of FRM against various centrality measures. We observe that Betweenness, Eigenvector have similar trends as the FRM EM series, especially during the crisis period of March to May 2020. On the other hand, Closeness centrality drops sharply into the crisis period.
When the FRM rises, the number of $\beta^{s}_{ij}$ equal to zero increase, increasing with it the distance between the vertices. The average length of one node (one FI in our case) and all other FIs increases, thereby sharply reducing Closeness centrality.
Betweeness as a centrality measure of a vertex within a network rises when the FRM rises, as information flow has a very high probability to pass between some central FIs, indicating a concentration of risk around certain FIs which we call "risk emitters".
Similarly, Eigenvector centrality is a measure of an FIs influence in an observed financial system network. Central FIs Eigencentrality rises sharply around a crisis period, as the FRM increases in value.
In- and Out-degree centrality drop when FRM rises, since the edges or connections between FIs have reduced sharply, to mostly the network's risk emitters.
We can conclude therefore, that a close inspection if the distribution of $\lambda^s_j$ as well as the detailed information within the adjacency matrix across a range of $\tau$ are particularly important. In Section \ref{portfolioconstruction} we aim to make use of this richness of information for the construction of more robust, tail-event network behaviour attentive portfolios. As an indication, Figure \ref{FRMEMnetwork} shows a network graph with edges between the 25 largest FIs, estimated at $\tau=0.05$ on 20200429. We highlight one exemplary bank, and its edges stemming from the adjacency matrix. Condensing such information into clusters of risk as outlined above is the focus of the following portfolio construction discussion.
In Figure \ref{fig:macro} we depict the macroeconomic risk variables influence over time. Clearly, the EM Sovereign Yield Spread to U.S. Treasury Bonds is the prominent macroeconomic risk variable not only during tail events but also at "normal" times estimated at $\tau=0.50$. We can conclude that most EM risk is rapidly priced into yield spreads, and then consequently impacts financial institutions. This link between the sovereign and banks needs to be considered for investors and policy makers as well. Other strongly influencing risk variables are more expected such as the VIX, the US yield curve shape, and Corporate Bond yield spreads. However, EM currency fluctuations only have a marginal effect in tail events, on some of the FIs. With the increase of debt issuance in local currency denominated debt, the risk of mismatches versus developed market currencies has diminished overall, and has lower influence on EM FIs.
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{Graphics/FRM-EM.png}
\caption{FRM@EMs Time series}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{FRMEMtimeseriestotal}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{Graphics/FRM_EM_Boxplot.png}
\caption{FRM@EM Boxplot, with \textcolor{blue}{mean} and \textcolor{red}{maximum} at $\tau=0.05$}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{FRMEMboxplot}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Macro_005.png}
\caption{$\tau=0.05$}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Macro_010.png}
\caption{$\tau=0.1$}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Macro_025.png}
\caption{$\tau=0.25$}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Macro_050.png}
\caption{$\tau=0.5$}
\end{subfigure}
\caption{Macro variables' most frequent marginal return contribution across time with 7-day moving average: \textcolor{red}{EM Sovereign Spread}, \textcolor{pink}{VIX}, \textcolor{violet}{U.S. 3mth to 10yr yield spread}, \textcolor{brown}{Moody's BAA Corporate Yield Spread}, \textcolor{green}{S\&P 500 Index}}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{fig:macro}
\end{figure}
\begin{sidewaysfigure}
\centering
\includegraphics[width=\textwidth, scale=2]{Graphics/ADJMAT_20200429_005.pdf}
\caption{Adjacency matrix estimated at $\tau=0.05$ on 20200429}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{ADJMAT20200429005}
\end{sidewaysfigure}
\begin{sidewaysfigure}
\centering
\includegraphics[width=\textwidth, scale=2]{Graphics/ADJMAT_20200630_005.pdf}
\caption{Adjacency matrix estimated at $\tau=0.05$ on 20200630}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{ADJMAT20200630005}
\end{sidewaysfigure}
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{Graphics/FRM_EM_Network_20200429_005.png}
\caption{FRM@EM Network with one \textcolor{orange}{example node} and its \textcolor{blue}{in-degree} and \textcolor{orange}{out-degree} edges}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{FRMEMnetwork}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_betweenness_avg.png}
\caption{{\color{blue} FRM} and \color{red}{Betweenness}}
\label{FRMbetweenness}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_closeness_avg.png}
\caption{{\color{blue} FRM} and \color{red}{Closeness}}
\label{FRMclosness}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_eigenvector_avg.png}
\caption{{\color{blue} FRM} and {\color{red} Eigenvector}}
\label{FRMeigenvector}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_indegree_avg.png}
\caption{{\color{blue} FRM} and {\color{red} In-degree}}
\label{FRMindegree}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_outdegree_avg.png}
\caption{{\color{blue} FRM} and {\color{red} Out-degree}}
\label{FRMoutdgree}
\end{subfigure}
\caption{{\color{blue} FRM} at $\tau=0.05$ and {\color{red} Centrality measures}}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{FRMcentmeasures}
\end{figure}
\section{Portfolio Construction}
\label{portfolioconstruction}
The portfolio is constructed using daily FIs prices for the period between 20200101 and 20200630 (which yields 130 observations of the 25 biggest EM FIs).\\
The performance of the uplifted portfolio approaches (Inv$\lambda$ and upHRP) are benchmarked against the minimum-variance portfolio (MinVar), the inverse variance (IVP), and the classical HRP approaches. The MinVar weights are simply computed based on Equation (\ref{minvarweights}), and often lead to concentration in low-volatility assets. The IVP strategy can be considered as a naive risk parity strategy (Equation (\ref{IVPweights})), since it is agnostic with respect to asset correlations. Thereafter, an overlapping region between portfolio optimisation startegies and FRM network centrality is developed.\\
As shown in the FRM network (Figure \ref{FRMEMnetwork}), all FIs are treated as potential substitutes without specifying any hierarchical structure among them. Therefore, tree structures that integrate hierarchical relationships are needed. For that purpose, the upHRP algorithm aims to build and make use of a clustered adjacency matrix.
\paragraph{\textbf{Matrix seriation}}\
For the classical approach, the steps of the HRP are applied using the price time series of the 25 biggest FIs selected during the studied period. Figure \ref{fig:Classical HRP: Matrix Seriation} presents the estimation results:
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Corr_Heatmap_EM_Cla.HRP_unsorted.png}
\caption{Unclustered Correlations}
\label{UnclusteredCorrelations}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Corr_Heatmap_EM_Cla.HRP_sorted.png}
\caption{clustered Correlations}
\label{ClusteredCorrelations}
\end{subfigure}
\caption{Classical HRP: Matrix Seriation}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{fig:Classical HRP: Matrix Seriation}
\end{figure}
Figure \ref{UnclusteredCorrelations} shows the original or the unclustered correlation matrix. Figure \ref{ClusteredCorrelations} illustrates the correlation matrix after reordering in clusters using the hierarchical tree clustering, also called clustered correlation matrix. This matrix then serves as the input for the asset allocation procedure. As shown in Figure \ref{ClusteredCorrelations} the reorganizing results group similar FIs together and the dissimilar further away in a new correlation matrix, which helps to construct more meaningful asset allocation decisions and building more risk diversified portfolios. More precisely, from Figure \ref{ClusteredCorrelations} the lighter-colored squares (indicating a higher correlation coefficient) are all concentrated around the diagonal matrix.\vspace{\baselineskip}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Corr_Heatmap_EM_m.HRP_unsorted_0505.png}
\caption{Unclustered Adjacency matrix}
\label{UnclusteredAdjmatx}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Corr_Heatmap_EM_m.HRP_sorted_0505.png}
\caption{clustered Adjacency matrix}
\label{ClusteredAdjmatx}
\end{subfigure}
\caption{Uplifted HRP: Matrix Seriation}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{fig:Uplifted HRP: Matrix Seriation}
\end{figure}
For the uplifted approach, the quasi-diagonalization step is applied to the adjacency matrix dated on 20200630 with $\tau=0.05$. In this stage, the algorithm aims at reordering the adjacency matrix by placing similar assets together. As we can notice from Figure \ref{ClusteredAdjmatx} the closer assets (similar given ($\beta$)) are placed together forming assets clusters.\vspace{\baselineskip}
\textbf{Tree clustering or Dendrogram}\vspace{\baselineskip}
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{Graphics/Dendrogram_EM_Cla.HRP.png}
\caption{Classical HRP: EM@Dendrogram}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{ClassicalHRPEM@Dendrogram}
\end{figure}
The clusters are visualised in the form of a cluster diagram called a dendrogram. Figure \ref{ClassicalHRPEM@Dendrogram} illustrates the hierarchical clusters for our FIs data, where the x-axis indicates the name of the FIs in the studied portfolio and the y-axis measures the distance between the two merging FIs.
The key to interpreting a dendrogram is to focus on the height at which any two FIs are joined together. In Figure \ref{ClassicalHRPEM@Dendrogram}, we can see that Bajaj Finance (BAF IS) and Bajaj Finserv Ltd (BJFIN IS) (plotted with green) are expectedly most similar, as the height of the link that joins them together is the lowest. Note that both of them are same ultimate parent Indian financial services companies focused on insurance. The dendrogram also detects that the most similar FIs most often belong to the same market (FirstRand Ltd (FSR SJ) and Standard Bank Group Ltd (SBK SJ) plotted with red), or a similar financial sector. Finally, the highest cluster represents the giant cluster that joins all the FIs formed clusters together.
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{Graphics/Dendrogram_EM_M.HRP_20200630.png}
\caption{Uplifted HRP: EM@Dendrogram}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{upHRPEM@Dendrogram}
\end{figure}
Based on the FRM clustered adjacency matrix, Figure \ref{upHRPEM@Dendrogram} indicates that the tree clustering architecture differs from the previous one (Figure \ref{ClassicalHRPEM@Dendrogram}). Indeed, here State Bank of India (SBIN IS) is a perturbation of BJFIN IS, and consequently the two assets present the first formed cluster since they are the most similar FIs. Also, Grupo Financiero Banorte (GFNORTEO MF) and VTR Bank (VTRB RM) are perturbations of Sberbank of Russia (SBER RM), hence these three FIs are clustered together, forming a cluster of two different markets, contrary to the previous dendrogram, Figure \ref{ClassicalHRPEM@Dendrogram}, where the most similar FIs belong the same market. \vspace{\baselineskip}
\paragraph{\textbf{Calculate allocation through recursive bisection}}\
Table \ref{WeightallocationsofFIsClassicalvsUplifted} and Figure \ref{figWeights allocationsFIs} specify further explanation regarding the weight’s allocation of the studied FIs according to the adopted strategies.
\begin{table}[H]
\begin{center}
\caption{Weight allocations of FIs: Classical approach vs {\color{blue} uplifted approach}}
\label{WeightallocationsofFIsClassicalvsUplifted}
\begin{tabular}{lccccc}
\hline\\
\multicolumn{1}{c}{\textbf{EM@ FIs}} & \textbf{MinVar} & \textbf{IVP} & \textbf{HRP} & \textbf{{\color{blue}Inv$\lambda$}} & \textbf{{\color{blue}upHRP}} \\
\hline
{\color[HTML]{212121} HDFCB.IS.EQUITY} & {\color[HTML]{212121} 0.002} & {\color[HTML]{212121} 0.045} & {\color[HTML]{212121} 0.017} & {\color[HTML]{212121} 0.039} & {\color[HTML]{212121} 0.051} \\ \hline
{\color[HTML]{212121} SBER.RM.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.054} & {\color[HTML]{212121} 0.020} & {\color[HTML]{212121} 0.021} & {\color[HTML]{212121} 0.018} \\ \hline
{\color[HTML]{212121} ITUB4.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.019} & {\color[HTML]{212121} 0.074} \\ \hline
{\color[HTML]{212121} HDFC.IS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.032} & {\color[HTML]{212121} 0.007} & {\color[HTML]{212121} 0.065} & {\color[HTML]{212121} 0.023} \\ \hline
{\color[HTML]{212121} KMB.IS.EQUITY} & {\color[HTML]{212121} 0.062} & {\color[HTML]{212121} 0.038} & {\color[HTML]{212121} 0.011} & {\color[HTML]{212121} 0.036} & {\color[HTML]{212121} 0.003} \\ \hline
{\color[HTML]{212121} BBDC3.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.055} & {\color[HTML]{212121} 0.074} \\ \hline
{\color[HTML]{212121} BBDC4.BS.EQUITY} & {\color[HTML]{212121} 0.002} & {\color[HTML]{212121} 0.005} & {\color[HTML]{212121} 0.002} & {\color[HTML]{212121} 0.017} & {\color[HTML]{212121} 0.030} \\ \hline
{\color[HTML]{212121} ICICIBC.IS.EQUITY} & {\color[HTML]{212121} 0.002} & {\color[HTML]{212121} 0.027} & {\color[HTML]{212121} 0.005} & {\color[HTML]{212121} 0.022} & {\color[HTML]{212121} 0.072} \\ \hline
{\color[HTML]{212121} QNBFB.TI.EQUITY} & {\color[HTML]{212121} 0.008} & {\color[HTML]{212121} 0.009} & {\color[HTML]{212121} 0.006} & {\color[HTML]{212121} 0.084} & {\color[HTML]{212121} 0.001} \\ \hline
{\color[HTML]{212121} SBIN.IS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.035} & {\color[HTML]{212121} 0.010} & {\color[HTML]{212121} 0.030} & {\color[HTML]{212121} 0.044} \\ \hline
{\color[HTML]{212121} B3SA3.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.040} & {\color[HTML]{212121} 0.053} \\ \hline
{\color[HTML]{212121} SANB11.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.024} & {\color[HTML]{212121} 0.007} & {\color[HTML]{212121} 0.029} & {\color[HTML]{212121} 0.107} \\ \hline
{\color[HTML]{212121} BBAS3.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.016} & {\color[HTML]{212121} 0.025} & {\color[HTML]{212121} 0.032} & {\color[HTML]{212121} 0.012} \\ \hline
{\color[HTML]{212121} BAF.IS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.021} & {\color[HTML]{212121} 0.014} & {\color[HTML]{212121} 0.056} & {\color[HTML]{212121} 0.012} \\ \hline
{\color[HTML]{212121} ITSA4.BS.EQUITY} & {\color[HTML]{212121} 0.001} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.018} & {\color[HTML]{212121} 0.024} \\ \hline
{\color[HTML]{212121} AXSB.IS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.018} & {\color[HTML]{212121} 0.009} & {\color[HTML]{212121} 0.041} & {\color[HTML]{212121} 0.056} \\ \hline
{\color[HTML]{212121} ELEKTRA.MF.Equity} & {\color[HTML]{212121} 0.866} & {\color[HTML]{212121} 0.503} & {\color[HTML]{212121} 0.787} & {\color[HTML]{212121} 0.032} & {\color[HTML]{212121} 0.073} \\ \hline
{\color[HTML]{212121} FSR.SJ.EQUITY} & {\color[HTML]{212121} 0.001} & {\color[HTML]{212121} 0.033} & {\color[HTML]{212121} 0.018} & {\color[HTML]{212121} 0.062} & {\color[HTML]{212121} 0.016} \\ \hline
{\color[HTML]{212121} SBK.SJ.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.029} & {\color[HTML]{212121} 0.011} & {\color[HTML]{212121} 0.074} & {\color[HTML]{212121} 0.034} \\ \hline
{\color[HTML]{212121} BBSE3.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.052} & {\color[HTML]{212121} 0.041} \\ \hline
{\color[HTML]{212121} BJFIN.IS.EQUITY} & {\color[HTML]{212121} 0.004} & {\color[HTML]{212121} 0.026} & {\color[HTML]{212121} 0.012} & {\color[HTML]{212121} 0.025} & {\color[HTML]{212121} 0.035} \\ \hline
{\color[HTML]{212121} GFNORTEO.MF.Equity} & {\color[HTML]{212121} 0.044} & {\color[HTML]{212121} 0.035} & {\color[HTML]{212121} 0.017} & {\color[HTML]{212121} 0.037} & {\color[HTML]{212121} 0.004} \\ \hline
{\color[HTML]{212121} BPAC11.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.010} & {\color[HTML]{212121} 0.003} & {\color[HTML]{212121} 0.032} & {\color[HTML]{212121} 0.016} \\ \hline
{\color[HTML]{212121} SLM.SJ.EQUITY} & {\color[HTML]{212121} 0.007} & {\color[HTML]{212121} 0.039} & {\color[HTML]{212121} 0.018} & {\color[HTML]{212121} 0.045} & {\color[HTML]{212121} 0.084} \\ \hline
{\color[HTML]{212121} VTBR.RM.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.037} & {\color[HTML]{212121} 0.043} \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{1\linewidth}
\includegraphics[width=\linewidth]{Graphics/Comparison_of_weights_for_HRP_IVP_and_MinVar.pdf}
\caption{Optimal weights {\color{cyan} MinVar}, {\color{yellow} IVP}, {\color{green} HRP}}
\label{classicalWeightsallocations}
\end{subfigure}
\begin{subfigure}[b]{1\linewidth}
\includegraphics[width=\linewidth]{Graphics/A_comparison_of_weights_for_upHRP_and_InvLambda_based_FRM.pdf}
\caption{Optimal weights {\color{orange} Inv$\lambda$}, {\color{blue} upHRP}
\bigskip
}
\label{UpliftedWeightsallocations}
\end{subfigure}
\caption{Weights allocations of FIs: classical approach vs uplifted approach}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{figWeights allocationsFIs}
\end{figure}
\
Table \ref{WeightallocationsofFIsClassicalvsUplifted} and Figure \ref{figWeights allocationsFIs} specify further explanation regarding the weight’s allocation of the studied FIs according to the adopted strategies.
Several points are worth noting. First, for both classical and uplifted approaches, IVP, HRP, Inv$\lambda$ and upHRP and similarly allocate their assets. This result can be explained by the fact that they are based on the inverse variance method to compute their weights. However, while IVP and Inv$\lambda$ show more stable weight distributions, HRP and upHRP adjust their weight allocation more frequently, since it also takes into consideration the correlation (for HRP)/ tail connectedness (for upHRP) and the asset clusters when assigning weights. Second, results also show that the MinVar allocations have the most concentrated weight allocation. While it accorded zero allocation to certain FIs (i.e SBIN IS, B3SA3 BS, etc.), it significantly overweights others (ELEKTRA MF).
Third, the MinVar, IVP and HRP classical approaches over-weighted the Mexican ELEKTRA MF (0.866, 0.503, 0.787, respectively). In this case, any distress situation affecting this FI will have a great impact on those concentrated portfolios. Contrary, these extreme weight concentrations disappear with the Inv$\lambda$ and upHRP uplifted approaches, where the weights are wider distributed among FIs, providing a well-diversified portfolio.
\begin{table}[H]
\begin{center}
\caption{ Weight's allocation of EMs: Classical approach vs {\color{blue}Uplifted approach}}
\label{WeightsallocationEMs}
\begin{tabular}{lrrrrrr}
\hline
{\color[HTML]{212121} \textbf{EM@ FIs}} & \multicolumn{1}{l}{{\color[HTML]{212121} \textbf{Nbr of FIs}}} & \multicolumn{1}{l}{{\color[HTML]{212121} \textbf{MinVar}}} & \multicolumn{1}{l}{{\color[HTML]{212121} \textbf{IVP}}} & \multicolumn{1}{l}{{\color[HTML]{330001} \textbf{HRP}}} & \multicolumn{1}{l}{{\color[HTML]{212121} \textbf{{\color{blue}Inv$\lambda$}}}} & \multicolumn{1}{l}{{\color[HTML]{212121} \textbf{{\color{blue}upHRP}}}} \\ \hline
{\color[HTML]{212121} \textbf{India (IS)}} & {\color[HTML]{212121} 8} & {\color[HTML]{212121} 7 \%} & {\color[HTML]{212121} 24.2\%} & {\color[HTML]{212121} 8.5\%} & {\color[HTML]{212121} 29.2\%} & {\color[HTML]{212121} 22.4\%} \\ \hline
{\color[HTML]{212121} \textbf{Brasil (BS)}} & {\color[HTML]{212121} 9} & {\color[HTML]{212121} 0.3\%} & {\color[HTML]{212121} 5.5\%} & {\color[HTML]{212121} 3.7\%} & {\color[HTML]{212121} 29.4\%} & {\color[HTML]{212121} 43.1\%} \\ \hline
{\color[HTML]{212121} \textbf{Mexico (MF)}} & {\color[HTML]{212121} 2} & {\color[HTML]{212121} 91 \%} & {\color[HTML]{212121} 53.8\%} & {\color[HTML]{212121} 80.4\%} & {\color[HTML]{212121} 6.9\%} & {\color[HTML]{212121} 7.7\%} \\ \hline
{\color[HTML]{212121} \textbf{Russia (RM)}} & {\color[HTML]{212121} 2} & {\color[HTML]{212121} 0 \%} & {\color[HTML]{212121} 5.4\%} & {\color[HTML]{212121} 2 \%} & {\color[HTML]{212121} 5.8\%} & {\color[HTML]{212121} 6.1\%} \\ \hline
{\color[HTML]{212121} \textbf{Turkey (TI)}} & {\color[HTML]{212121} 1} & {\color[HTML]{212121} 0.8\%} & {\color[HTML]{212121} 0.1\%} & {\color[HTML]{212121} 0.6\%} & {\color[HTML]{212121} 8.6\%} & {\color[HTML]{212121} 0.1\%} \\ \hline
{\color[HTML]{212121} \textbf{S.Africa (SJ)}} & {\color[HTML]{212121} 3} & {\color[HTML]{212121} 0.8\%} & {\color[HTML]{212121} 10.1\%} & {\color[HTML]{212121} 4.7\%} & {\color[HTML]{212121} 18.1\%} & {\color[HTML]{212121} 13.7\%} \\ \hline
\end{tabular}
\end{center}
\end{table}
Table \ref{WeightsallocationEMs} reports the percentage of weight allocation for the different approaches, giving further intuition on how the portfolio optimization approaches allocate their weights through EMs. From these allocation results, we can understand a few stylized features: First, the MinVar concentrates 91\% of the allocation on the two existing Mexican FIs and it assigns zero weights for the Russian FIs and only 0.3\% for the 9 Brazilian FIs. Similarly, the IVP and the HRP concentrate allocation on the same market with 53.8\% and 80.4\% weights respectively. However, they provide some diversification across the other markets compared to MinVar strategy. The classical approaches concentrate their weights on the Mexican market that contributed only by two FIs to the basket of the 25 biggest FIs in the EMs, which poses significant idiosyncratic risk.
Nevertheless, for the uplifted portfolio approaches, the weights are well distributed across the EMs. For example, the Inv$\lambda$ allocates 29\% of weights to the Brazilian and Indian FIs, while, the upHRP allocates 29.4\% to Indian FIs and 43.1\% to the Brazilian FIs since this last market has the highest number of FIs in the selection of the 25 biggest EM@FIs. Moreover, the upHRP allocates only 0.1\% for the Turkish FIs, since this market contributes only by one FI.
To recapitulate, the upHRP appears to find a compromise between classical concentrated weight approaches and the Inv$\lambda$ strategy. The classical strategies can concentrate weights on a few FIs, leading to vulnerabilities. The Inv$\lambda$ evenly assigns weights across all FIs, ignoring the correlation structure. This makes it exposed to systemic shocks. However, the upHRP finds a compromise between diversifying across all FIs and diversifying across clusters, which makes it more resistant against both types of shocks.\vspace{\baselineskip}
\paragraph{\textbf{Backtesting results}}
\begin{table}[H]
\begin{center}
\caption{Backtesting on EM FIs data: Classical approach vs {\color{blue} uplifted approach}}
\label{BacktestingEMFIsdata}
\begin{tabular}{ccccc}
\hline
\multicolumn{1}{l}{} & {\color[HTML]{212121} \textbf{Mean}} & {\color[HTML]{212121} \textbf{Std}} & {\color[HTML]{212121} \textbf{Sharpe ratio}} & {\color[HTML]{212121} \textbf{Effective n}} \\ \hline
{\color[HTML]{212121} \textbf{MinVar}} & {\color[HTML]{212121} 0.0043} & {\color[HTML]{212121} 0.0818} & {\color[HTML]{212121} 0.0528} & {\color[HTML]{212121} 1.3225} \\ \hline
{\color[HTML]{212121} \textbf{IVP}} & {\color[HTML]{212121} 0.0039} & {\color[HTML]{212121} 0.0712} & {\color[HTML]{212121} 0.0560} & {\color[HTML]{212121} 3.7121} \\ \hline
{\color[HTML]{212121} \textbf{HRP}} & {\color[HTML]{212121} 0.0035} & {\color[HTML]{212121} 0.0490} & {\color[HTML]{212121} 0.0721} & {\color[HTML]{212121} 1.6063} \\ \hline
\multicolumn{5}{l}{$\tau=5\%$} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}Inv$\lambda$}}} & {\color[HTML]{212121} 0.0767} & {\color[HTML]{212121} 0.1859} & {\color[HTML]{212121} 0.4128} & {\color[HTML]{212121} 20.986} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}upHRP}}} & {\color[HTML]{212121} 0.0522} & {\color[HTML]{212121} 0.1886} & {\color[HTML]{212121} 0.2768} & {\color[HTML]{212121} 16.801} \\ \hline
\multicolumn{5}{l}{$\tau=10\%$} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}Inv$\lambda$}}} & {\color[HTML]{212121} 0.0894} & {\color[HTML]{212121} 0.1868} & {\color[HTML]{212121} 0.4788} & {\color[HTML]{212121} 21.643} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}upHRP}}} & {\color[HTML]{212121} 0.0807} & {\color[HTML]{212121} 0.1967} & {\color[HTML]{212121} 0.4104} & {\color[HTML]{212121} 16.296} \\ \hline
\multicolumn{5}{l}{$\tau=25\%$} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}Inv$\lambda$}}} & {\color[HTML]{212121} 0.0894} & {\color[HTML]{212121} 0.2195} & {\color[HTML]{212121} 0.4073} & {\color[HTML]{212121} 23.429} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}upHRP}}} & {\color[HTML]{212121} 0.0792} & {\color[HTML]{212121} 0.2280} & {\color[HTML]{212121} 0.3476} & {\color[HTML]{212121} 20.199} \\ \hline
\multicolumn{5}{l}{$\tau=50\%$} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}Inv$\lambda$}}} & {\color[HTML]{212121} 0.0883} & {\color[HTML]{212121} 0.2093} & {\color[HTML]{212121} 0.4219} & {\color[HTML]{212121} 21.117} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}upHRP}}} & {\color[HTML]{212121} 0.0766} & {\color[HTML]{212121} 0.2145} & {\color[HTML]{212121} 0.3571} & {\color[HTML]{212121} 21.508} \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=0.8\textwidth]{Graphics/Portfolio_returns_30_days_Rebalacing.png}
\caption{Backtesting on EM FIs data: classical approaches
({\color{blue}MinVar},{\color{red}IVP}, {\color{brown}HRP}) vs uplifted approaches ({\color{purple}Inv$\lambda$}, {\color{green}upHRP})}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{figBacktestingEMFIsdataclassicalvsUplifted}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Backtesting_uplifted_approach_with_different_tau.png}
\caption{Backtesting on EM asset data: uplifted approach ({\color{purple}Inv$\lambda$ 5\%}, {\color{green}upHRP 5\%}, {\color{cyan}Inv$\lambda$ 10\%}, {\color{magenta}upHRP 10\%}, {\color{orange}Inv$\lambda$ 25\%}, {\color{lime}upHRP 25\%}, {\color{blue}Inv$\lambda$ 50\%}, {\color{red}upHRP 50\%})}
\label{BacktestingUpliftedapproachwithdifferenttau}
\end{subfigure}
\begin{subfigure}[b]{0.5\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_2020.png}
\caption{Financial Risk Meter (FRM)
\bigskip
\bigskip
\smallskip
}
\label{FinancialRiskMeter2020}
\end{subfigure}
\begin{subfigure}[b]{0.8\linewidth}
\includegraphics[width=\linewidth]{Graphics/lamda_weights_20200630-4.png}
\caption{Penalisation parameter ($\lambda$) vs Optimal weights ({\color{blue}Penalisation parameter ($\lambda$}, {\color{purple}Inv$\lambda$ 5\% weights}, {\color{green}upHRP 5\% weights})}
\label{Penalisationparametervsoptimalweights}
\end{subfigure}
\caption{Impact of Penalisation parameter ($\lambda$) and FRM on optimal weights and portfolio returns}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{figImpactofPenalisationparameterandFRMonoptimalweightsand portfolio returns}
\end{figure}
For the backtesting, we adopt a 30 days rebalancing period. Table \ref{BacktestingEMFIsdata} recapitulates the out-of-sample performance of the studied approaches.\\
For the classical approaches, the mean and the volatility of the HRP are 0.003 and 0.049, respectively. While the MinVar delivers the highest return value (0.004) and the highest volatility (0.081), the IVP provides approximately the same portfolio return (0.003) with higher volatility (0.0712). Therefore, the HRP balances out both return and volatility most efficiently providing the best approach in terms of Sharpe ratio (0.072 compared to 0.052 and 0.056 of the MinVar and IVP portfolios, respectively).\\
For the uplifted approach (with $\tau=5\%$), the return and volatility of the upHRP approach are 0.052 and 0.188, respectively. While the Inv$\lambda$ offers a higher return (0.076), and approximately the same volatility, the Inv$\lambda$ provides a risk-return balance registering the best Sharpe ratio value (0.412 compared to 0.276 of the upHRP portfolio). \\
Besides the sharp ratio, we measured the portfolio diversification effects using the Effective $N$ $Effect(N)$ measurement, proposed by \cite{strongin2000beating} as one of allocation concentration and defined as follow:
\begin{equation}
Effect(N)=\frac{1}{\sum_{j=1}^N \widehat{\omega}_{j,t}^2}
\label{eq49}
\end{equation}
The portfolio has high concentration risk if $Effect(N)$ is close to one.\vspace{\baselineskip}
Comparing the classical and the uplifted approaches(with $\tau=5\%$), it seems that the upHRP improves the Sharpe ratio of the HRP approach (0.072 and 0.276, respectively). Moreover, the HRP has discarded five EMs in favor of one single market, with ($Effect(N)=1.6)$, see Table \ref{WeightsallocationEMs}). Therefore, the HRP’s portfolio is deceitfully diversified, since any distress situation affecting this market will have a greater negative impact on HRP’s portfolio than the upHRP, which allocates only 6.9\% to the Mexican market and providing more diversification across EMs ($Effect(N)=16.8$, see Table \ref{BacktestingEMFIsdata}). \vspace{\baselineskip}
In fact, the main innovation of upHRP strategy is to apply the HRP algorithm based on tail dependence clustering instead of the standard correlation-based clustering. First, this strategy is motivated by the fact that the correlation coefficients can change drastically during financial crises due to contagion effects, and such a crisis can spill over quickly. Consequently, diversifying a portfolio based on correlation clusters may be a failing strategy without attention to tail events. Second, while the correlation matrix illustrates the relationship based on the mean-variance of the distribution, the FRM measures the co-movements in extreme events based on both tail of the distribution. To exemplify, lower tail dependence is associated with the capacity to diversify during crises, which can improve the tail risk management of a given strategy. More precisely, the main idea is to cluster the studied FIs illustrating a high probability to experience extremely negative events contemporaneously. Finally, empirical results show that the proposed uplifted approach has the potential to compete with the classical portfolio optimization approach by providing desirable diversification properties, especially if this hierarchical risk parity strategy is based on the tail dependence coefficient (provided by the FRM adjacency matrix), which is a benefit to tail risk management.\vspace{\baselineskip}
Besides, for the uplifted approach, we take into consideration four tail risk levels (5\%, 10\%, 25\%, 50\%). The results in Table \ref{BacktestingEMFIsdata} indicate that:
The $Effect N$ is an increasing function of the tail risk level, therefore for a higher tail risk level, more diversification is needed to guarantee stable portfolio returns.
Comparing the Inv$\lambda$ strategy with the upHRP, we noticed that the Inv$\lambda$ strategy needs to include a higher number of assets compared to the upHRP (i.e for $\tau=10\%$, $Effect(N)= 21.64$ and 16.24, respectively) to provide the same portfolio performance level ( approximately the same sharp ratio=0.4). On the one hand, this result can be explained by the fact that for the upHRP, only FIs within the same cluster compete for portfolio allocation rather than competing with all the FIs in the portfolio, which avoids the redundancy in the FIs included in the optimal portfolio. On the other hand, the larger number of FIs increases estimated parameters, which also can increase the risk of estimation error and thus biased results. For that purpose, for relatively similar values of sharp ratio (portfolio efficiency), the upHRP is preferred over the Inv$\lambda$.\vspace{\baselineskip}
Figure \ref{figBacktestingEMFIsdataclassicalvsUplifted} illustrates the backtesting on EM’s FIs with 30 days rebalancing period. It plots the portfolio turnover for the first half of 2020 according to the classical strategies (MinVar, IVP, and HRP) and the uplifted strategies (Inv$\lambda$ and upHRP), where the classical approaches show a relative stability in returns compared to the uplifted approaches since, the classical approaches concentrate their weights on one Mexican FI (see Table \ref{WeightsallocationEMs}), which limited their profit (increase in portfolio returns in mid-February and in June) and also limit their loss (end of March ) during the COVID 19 crisis. However, this is a trade-off between historical stability and very high concentration risk. Any idiosyncratic risk impact would lead to substantial portfolio losses.
Figure \ref{BacktestingUpliftedapproachwithdifferenttau} plots the backtesting on EM’s FIs with 30 day rebalancing period for the uplifted strategies, taking into consideration the adjacency matrices with different tail risk levels (5\%, 10\%, 25\%, and 50\%). The Figure indicates that all portfolio returns follow the same trend, despite the difference in portfolio composition, since the biggest 25 FIs that compose the adjacency matrices vary over time and the risk levels.\\
Indeed, by comparing this portfolio return’s trend with the FRM plot (Figure \ref{FinancialRiskMeter2020}) during the same period, we noticed a negative relationship, where the high-risk period plotted by the FRM is translated by a decrease in the portfolio returns (March). Moreover, Figure \ref{Penalisationparametervsoptimalweights} indicates that the uplifted approach underweights the high risky FIs (with high penalization parameter $\lambda$) and overweight the less risky FIs, specially for the Inv$\lambda$, since the upHRP takes into consideration the tail co-movement between the FIs.
\subsection{Bridging optimal portfolio weights and network centrality}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/network_20200630_005.png}
\caption{$\tau=0.05$}
\label{tau005}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Network_20200630_010.png}
\caption{$\tau=0.1$}
\label{tau01}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Network_20200630_025.png}
\caption{$\tau=0.25$}
\label{tau025}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Network_20200625_050.png}
\caption{$\tau=0.5$}
\label{tau05}
\end{subfigure}
\caption{Network graphs with the size of nodes representing total degree centrality on 20200630}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{fig:NETWORKtau}
\end{figure}
Figure \ref{fig:NETWORKtau} plots the FRM Network with the size of nodes representing total degree centrality (indegree and outdegree). The network information is provided by the adjacency matrices estimated at $\tau$ equal to 0.05, 0.10, 0.25, and 0.50, with dates 20200630. The Figure shows that the total degree centrality (both in and out) increase and become more visible with the increasing of tail risk level $\tau$.
By interpreting the adjacency matrix of the network, an overlapping region between the portfolio theory and network theory can be established.\\
Indeed, according to Figure \ref{tau005}, Itau Unibanco (ITUB4 BS) is a highly central FI followed by Banco Santander Brasil (SANB11 BS) and Banco Bradesco SA (BBDC4 BS). Therefore, the Brazilian market illustrates high-CoStress compared to the other markets (see also the boxplot, Figure \ref{FRMEMboxplot}. Contrary to that, QNB Finansbank (QNBF TI) and Housing Development Finance Corp (HDFC IS) are low-central FIs.\\
Regarding ELEKTRA MF, the networks (Figure \ref{fig:NETWORKtau}) indicate that this FI is only a risk emitter (high out-degrees with zero in-degree), which makes this FI insensitive to the other FIs shocks. Moreover, the classical portfolio approaches overweight ELEKTRA MF (see Table \ref{WeightallocationsofFIsClassicalvsUplifted}). Therefore, at least historically, ELEKTRA MF had more stable return patterns, but as a risk emitter, any change in risk perception concerning this FI would rapidly spill over into the entire market.\\
Moreover, referring to Table \ref{Optimalweightsfordifftau}, the results of the optimal weight indicate that the Inv$\lambda$ portfolio is insensitive to the centrality degree. It often allocates similar weights for the high and low central FIs (i.e. BBDC4 BS and GFNORTEO MF, see Figure (\ref{tau05}). This result is due to the fact that this portfolio approach allocates weights based only on the individual tail risk level $\Lambda$, varying by tail risk levels $\tau$ but is insensitive to the network centrality degree.\\
Referring to Table \ref{Optimalweightsfordifftau} and Figure \ref{fig:NETWORKtau}, results from different $\tau$ levels indicate that the upHRP approach underweights the high central FIs (i.e. BBDC4 BS (for $\tau$=5\% and $\tau$=50\%), ICICIBC IS (for $\tau$=10\%), ITUB BS and HDFC IS (for $\tau$=25\%)). Contrary to that, the upHRP strategy overweights the low central FIs (i.e. ELEKTRA MF and SLM SJ (for $\tau$=5\%), ELEKTRA MF (for $\tau$=10\%), GFNORTEO MF and B3SA3 BS (for $\tau$=25\% and $\tau$=50\%). These findings are in line with previous research, studying the relationship between the optimal portfolio weights and network centrality in the mean-variance level. In fact, our findings are consistent with \cite{peralta2016network} and \cite{pozzi2013spread} in establishing that "optimal portfolio strategies should overweigh low-central assets and underweight high-central ones". Nonetheless, they contradict the findings of \cite{vyrost2019network}, who argue that “asset weights must be ordered in the same way as the reciprocal of asset centrality in a given network”. \\
Indeed, the sensibility of the Adjacency matrix to different tail risk levels $\tau$ results in centrality degree variation as well as FIs clustering (see Figures \ref{upHRPEM@Dendrogram} and \ref{figdendrogramofdifftau}), which leads to upHRP optimal weights variation through different $\tau$ for all FIs.\\
In fact, the low concentration of weights for the uplifted strategies (Inv$\Lambda$ and upHRP) as well as the sensitivity of weights to different tail risk levels, don’t allow us to conclude about the best FI that offers greater diversification benefits in EMs (contrary to the classical approach that concentrates weights on ELEKTRA MF). However, uplifted strategies overweight the Brasilian FIs (for different $\tau$). Moreover, by contributing with the highest number of FIs to the basket of the 25 biggest FIs in EMs (see Table \ref{WeightsallocationEMs}), the Brazilian market seems more interesting to international investors, besides the need of diversification advantages provided by the other EMs.
\section{Policy recommendation}
Besides the outlined portfolio construction in Section \ref{portfolioconstruction}, we want to derive some recommendations for both policy makers as well as investors. \\
In Section \ref{FRMEMinterp} we have shown that particular EM FIs are not only influenced by FIs and macroeconomic risk variables from the same region, but also have tail-event co-movement dependencies to other EM geographic regions and currencies, the latter all the more important at lower levels of $\tau=0.10$ or $\tau=0.05$. Secondly, a dominant driver of EM FIs returns is the fiscal and economical stance of EM Sovereign bond issuers. If the risk perception for sovereign issuers increases, the return of EM FIs is impacted across the board and across the multivariate distribution. Lastly, co-movements can be detected within similar economic sub-sectors across geographical regions, which is also apparent in the portfolio construction exercises in Section \ref{portfolioconstruction}. We recommend investors to analyse concentrations across economic sub-sectors across regions, as well as to analyse dependencies between equity investments and bond investments in EM, given the clear linkages between the sovereign and the domestic EM FIs. Lastly, currency fluctuations have marginal return contributions in tail-events especially. In so far as EM FI investments are unhedged into developed market currencies, there is risk of a compound effect on returns. \\
As for the portfolio construction, the "classical" approaches show a relative stability in portfolio returns compared to the uplifted approaches, which limited their profit and also limit their loss during the COVID 19 crisis. But, this often comes at the cost of high concentration risks, exposing the portfolio to idiosyncratic tail events. The uplifted portfolio approaches show more volatile turnover with higher loss and profit, but are much better diversified, preventing sizeable risk clusters. Even though a risk averse investor might opt for the classical approach at first, closer inspection of tail risk behaviour and concentration risk should let the investor prefer the uplifted portfolio.\\
EM policy makers can derive important recommendations from our analysis. Firstly, coordination between EM regulatory bodies is of importance in order to mute EM FI fluctuations. Here, particular attention should be put on linkages in same sub-sector operating FIs across regions, which will be increasingly more important as globalisation continues. Secondly, fluctuations in risk perception of the sovereign issuer has an immediate impact on EM FI returns. Regulatory bodies are therefore advised to preemptively verify sound capitalisation of their domestic banks, even if the sovereign issuers distress is stemming from another geographic region. This is further amplified by a tail-event leading to a weakening of an EM's currency versus for example the U.S. Dollar. Global financial linkages, we show, between EM FIs lead to spill-over effects. Overall, in order to protect the domestic as well as global EM economies and their FIs, EM regulatory bodies should continuously work towards closer coordination between Emerging Market economies. This will help increase robustness versus developed market economies distress, of which more is likely to come, as developed markets fight with difficult fiscal situations and low growth patterns. But also, closer coordination between countries will prevent spill-over effects from one geographical region's FIs onto others, thereby increasing the attractiveness of the asset class further. Our approach via dendograms rapidly indicates any such risk clusters, and can be updated at ease and frequently during financial crises.
\section{Conclusion}
In this study, we examine the co-movements of EM FIs across six geographical regions, with aim to analyse within EM country co-dependencies in tail events but importantly also between regions as well as FIs of the same sub-sector across regions. We also analyse the important macro-economical risk variables impacting EM FIs and conclude that in addition to the developed market variables suggested by \cite{tobias2016covar}, EM specific macroeconomic risk variables have significant explanatory power. This can be used to construct more robust total asset class portfolio allocation and supplies EM regulatory bodies with detailed information on co-dependencies for better and faster stabilisation measures during periods of distress.\\
We also propose a novel asset allocation method – Hierarchical Risk Parity based the tail event information from the FRM technology, allowing us to extend this approach to the quantile level and replace the covariance matrix with the rich information contained in the FRM adjacency matrix. We applied this proposed approach to a portfolio of the biggest 25 FIs in EM, and our results show that uplifted strategies provide appropriate diversification properties. In comparison, the Inv$\lambda$ portfolios tend to be too static and the classical approaches result in too concentrated portfolios. Bridging optimal portfolio weights and network centrality, we conclude that the Inv$\lambda$ insensitive to the network centrality degree. However the upHRP portfolio underweight high-central FIs and overweight low-central ones, therefore the Inv$\lambda$ is less at risk of spill-over effects across EM regions, FIs, and financial sub-sectors.
\section{Introduction}
Emerging markets have been commonly acclaimed for providing robust growth potential and offering investors a higher expected return compared to developed markets. Indeed, due to the possibility of higher profits and the low level of global equity markets integration, EMs have been considered as an investment opportunity for investors, whose aim to build an internationally diversified portfolio. EMs liquidity and transparency have continuously enhanced [\cite{mcguire2006common}; \cite{bunda2009correlations}]. Moreover, the reputation of EMs, in the framework of portfolio diversification, has received the attention of international investors, especially after the financial crisis that affected mostly developed markets.
However, the investment’s benefits of EMs come with additional risks, which are usually not as prominent in developed markets. In fact, EMs are exposed to additional economic, political, and currency risks. Further, the EMs’ economies fast growth and, as consequence, the quick evolution of their structure result in market information being rapidly outdated. Therefore, the existing more traditional methods of risk evaluation may be misleading in EMs, especially in short and medium investment horizons. Indeed, the existing risk measure methods are not suited to provide the up to date point of view representing current market structures, if they not supplemented with the latest market information. Hence, an efficient systemic risk measure is needed for EM.
For that purpose, it is crucial to understand and measure the spillover risk across EMs financial system network, which is important for financial risk measurement and portfolio diversification;\\
From the perspective of financial risk measurement, the interdependence among FIs becomes more important, especially during periods of distress, when losses spread through institutions, rendering the global financial system more vulnerable. In this regard, a systemic crisis that disturbs the financial system stability can have serious effects and lead to high losses for the entire economy and society.
From the perspectives of risk management and portfolio diversification, the contagion risk across the FIs from the same market, and across the worldwide markets, leads to a decrease in diversification potential. Hence, understanding the network structure of interdependence among FIs is crucial to risk managers and portfolio investors, as this can help them design investment strategies to reduce dependence risk and thereby increase diversification. Investors are also interested in recognizing the FIs that contribute the most (least) risk to their portfolio so that they are considered with caution in their portfolio design, especially during financial market turmoil.
To understand the FIs co-movements in EMs, our study examines the effect of Macro factors on the FIs in BRIMST (Brazil, Russia, India, Mexico, South Africa, Turkey) EMs, using daily interval data for the period between 2000–2020, with particular focus on the last two years. It is well known that the economics of the mentioned markets are strongly linked to the US economy [\cite{ozatay2009emerging}], so it is crucial to analyze the effects of US Macro factors on volatility in EMs financial system. It is worth to note that our dataset covers several global crisis periods, allowing us to examine how the EMs Financial systems respond to the different crisis. After determining the interdependencies among FIs and macroeconomic factors, our research aims also to build a robust strategy based on portfolio diversification in EMs.
In order to achieve the mentioned goals, our paper seeks to answer the following questions: Can the US and EMs Macro factors explain BRIMST financial equity indices? Are some categories of Macro factors more important than others? What FIs from EMs are the largest (smallest) spillover transmitters (receivers)? What FIs contribute the most (least) risk to total portfolio risk? What FIs offer greater diversification benefits? And lastly, how the tail spillover effect and portfolio weights change over time, and how they react to the different tail risk levels?
Answering all these questions is crucial, as international investors interested in understanding the forces behind the interdependence among macroeconomic factors and FIs, to identify potential risks and rewards and benefit from global diversification. Economic policymakers and regulators in BRIMST are interested in forces behind the co-movement between these markets to further establish market resilience in EMs.
We tackle these questions by employing the Financial Risk Meter (FRM) technology in EMs. The FRM is based on Lasso quantile regression designed to examine tail event co-movements financial securites. The objective is to understand the FIs interconnectedness and represent them in a network topology. Moreover, the FRM indices summarize the systematic risk at a given area and identify risk factors. Briefly, FRM represents tail event behavior in a financial risk factors network, which allows us (i) to identify the “stress emitters” and “stress receivers” companies, (ii) to measure the tail dependencies among the FIs and the Macro factors (iii) analyze the risk level in EMs over time. Concerning risk management, (iv) the FRM network is adopted in the portfolio selection process. More precisely, by interpreting the correlation coefficients of FIs equity indices in the adjacency matrix of the FRM network, an overlapping region between the portfolio optimisation strategies and FRM network is developed, (v) The FRM adjacency matrix is also adopted to lift the Hierarchical Risk Parity (HRP) approach [propsed by \cite{de2016building}] to the quantile level.
The remainder of this research is organized as follows. Section 2 presents a brief review of risk measurement and risk management methods. The econometric approach is discussed in Section 3, as well as a discussion on centrality measures. Section 4 illustrates the proposed portfolio optimization strategies. The FRM results are analysed in section 5. Section 6 develops the portfolio construction, while section 7 provides policy recommendations before concluding the general scope.\\
The codes are published on \url{www.quantlet.de} indicated by \includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg} with keyword FRM. The current level of FRM for Emerging markets as well as other channels can be found at \url{http://hu.berlin/frm}.
\section{Literature Review}
The large-scale breakdown of FIs after the global Lehman brother’s crisis in 2008 had caused serious social and economic losses. Therefore, controlling and maintaining financial system stability is one of the principals and the mutual responsibilities of central banks and financial market regulators around the world. Previous researches have been conducted in the framework of FIs interconnectedness. In this context, empirical studies commonly focus on resolving this issue in European countries; \cite{betz2016systemic}, \cite{abbassi2017credit}, and \cite{aldasoro2018multiplex}. Some studies have studied the FIs networks of developed American countries. Among others, \cite{cai2018syndication}; \cite{kreis2018systemic}, and \cite{tonzer2015cross}. Relatively few researchers have investigated bank networks of emerging American countries; \cite{silva2016network}; and \cite{berndsen2018financial}. All of the above studies focused on data from only one country and did not consider networks between FIs from several countries across the globe.
In order to bridge this gap, our study investigates the FIs networks from six among the biggest emerging countries worldwide, as globalization leads to closely linked economic activity among these countries in terms of trade and finance.
It is well known that FIs are interconnected within networks of several types of financial connections and contracts. The complex links among these institutions, which can be considered as risk factors, can cause systemic risk and results in spillover-effects that deteriorate the network stability and its functioning. These observations on the joint network dynamics motivated practitioners and researchers to insert tail events into risk management.
The Value-at-Risk (VaR) approach \cite{franke2019statistics} is frequently used to measure market risk, by computing the monetary loss of an institution for a given confidence level [\cite{slim2017value}]. However, the VaR measures a tail event probability hosting only one single node, which does not reflect its connection to overall systemic risk.
Recently, \cite{adrian2011covar} \& (\citeyear{tobias2016covar}) developed CoVaR to measure the systemic risk [e.g.\cite{fang2018identifying}, \cite{souhir2019price}]. The CoVaR measures spillover effects across financial markets by providing the VaR of one market under the condition that the other market is in financial distress as given by its VaR. Consequently, the CoVaR approach can gauge the size of the financial risk spillover. However, it can only capture the extent of risk spillovers for a simple bivariate system and cannot simultaneously measure the risk spillover effects across multiple financial markets.
\cite{hardle2016tenet} developed the Tail Event NETwork (TENET) risk approach by generalizing the CoVaR to be able to accommodate all system nodes as risk factors. TENET applies the quantile regression method on a set of network nodes stock market information and macroeconomic variables in a rolling window approach. The innovation of TENET model is to employ dimensional reduction (in a semi-parametric setting) by using the Lasso in a quantile regression framework \cite{tibshirani1996regression}. Also, \cite{chen2019tail} propose a Tail Event driven Network Quantile Regression (TENQR) to adress the interdependence, dynamics and riskiness of financial institutions. The TENQR captures risk dynamics within a panel quantile autoregression in a network topology, quantified through a time-varying adjacency matrix. TENQR technique is evaluated using the SIFIs (systemically important financial institutions). To extend the TENET further, \cite{mihoci2020frm} developed an improved systemic risk measure that summarizes the high-dimensional tail stress into a single real value indicator, called the Financial Risk Meter (FRM). The FRM is computed as an average of each node and at each time window selected penalization terms. The FRM level contains fundamental information about the active set of influential neighboring nodes and about the contributors to systemic risk. \\
\cite{yu2019ai} compared the proposed FRM to other measures for systemic risk, such as Google Trends, SRISK, and VIX. They found a Granger causality between the FRM and these measures, which confirmed the validity of the FRM as an efficient systemic risk measure. \cite{mihoci2020frm} applied the FRM to measure the dependencies between FIs and Macro factors exploiting tail event information. They built two FRM indices namely FRM@Americas and FRM@Europe, but also on bond yield and credit default swap spread co-movements. \\
So far, the FRM applications took into consideration financial securities from one economic region and its domestic Macro factors. In our study, we extend the application of FRM for EMs, and we will consider the domestic Macro factors from the BRIMST as well as those from the US (proposed by \cite{tobias2016covar} to represent both the domestic and foreign state of the economy.
In many portfolio construction approaches, the correlation between financial assets represents the basis for portfolio selection. To exemplify, \cite{markowitz1959portfolio} developed the modern portfolio theory (MPT). He found that when correlation between assets is not perfect, a diversified portfolio can be constructed. Therefore, to reach efficient diversification, investors should select anti-correlated assets and verify and ensure the satisfaction of this condition over time. But the correlation structure among assets changes over time and evolves especially during crises periods. For that reason, the Markowitz theory is usually oriented to select the most stable assets such as the industrial assets, and hence, the Markowitz optimal portfolio is often composed of a limited and invariant set of assets.
The second weakness of Markowitz's MPT [\cite{markowitz1959portfolio}] is related to the large estimation errors of the expected returns vector [\cite{merton1980estimating}], as well as of the covariance matrices [\cite{jobson1980estimation}]. Hence, a robust method for the modelling of the dynamic interconnectedness of assets is needed to support the MPT and guide investors in building an efficient and well diversified portfolio.
In this regard, recent researchers designed financial markets in networks (FMN), where assets are represented by nodes and the correlation of returns are represented by links that relate these nodes [\cite{chi2010network}; \cite{diebold2014network}; \cite{peralta2016network}]. Despite the originality and interesting results provided by this network approach, the majority of its applications are descriptive and lack concrete applications in the portfolio optimization procedure.
Recently, \cite{pozzi2013spread} extracted the dependency structure of financial equities from the network approach to build a diversified portfolio to reduce investment risk. This procedure visualized portfolio selection directly over the FMN design. \cite{peralta2016network} used the FMN as a powerful device to enhance the portfolio optimization procedure by selecting a set of assets according to their network centrality.
However, the adopted spillover and financial network methods focused on estimating the risk spillover based on the return distributions' first two conditional moments, thereby ignoring higher moments of the distribution (i.e. right and left tails). Indeed, the existing spillover measures concentrate only on the mean and variance of the distributions, which may underestimate the real spillover effects among FIs in tail-events, since they do not take into consideration the extreme risk spillover across financial markets. More specifically, previous researchers investigated mean-to-mean effect, assuming that investors are mean–variance optimizers. But ideally, a portfolio selection decision should be based on the entire return distribution estimation, since investors are more risk averse to the extreme downside risk related to the tails of the distribution.
To overcome this limitation, we adopt the FRM to measure the joint tail events. In fact, the FRM has the advantage to represent instantaneously the co-movements and the dynamics of high-dimension networks. Furthermore, the FRM can display the hidden interdependency structures between the financial network's nodes.
Given that the functions of complex FMN are reflected in the risk evaluation and portfolio optimization [\cite{haluszczynski2017linear}; \cite{WANG20181}], we contribute to this line of research by investigating the extent to which the underlying structure of this financial market tail event network can be used as an effective tool in enhancing the portfolio selection process. For that purpose, we establish a bridge between the FRM network and portfolio optimisation strategies. More precisely, we study the relationship between optimal portfolio weights and the FIs’ centrality in the FRM network. \vspace{\baselineskip}
The HRP is a second approach that aims to overcome the shortcomings of MPT. The Hierarchical risk parity (HRP) is firstly proposed by \cite{LopezdePrado59} as a risk parity allocation algorithm. The HRP applies machine learning and graph theory to extrapolate a hierarchical implementation of an inverse-variance allocation with weights computed among the formed correlated asset return clusters. By substituting the covariance structure with a hierarchical structure of clusters, HRP fully benefits from the covariance matrix information and improves the stability of the weights.
\cite{LopezdePrado59} shows that the HRP achieves higher risk-adjusted return and lower out-of-sample volatility than inverse-variance allocations. Despite its appealing features, the application of HRP is still rare, with increased interest \cite{lohre2020hierarchical}, \cite{risks7030074}, and \cite{Raffinot89} test the performance of HRP in a multi-asset allocation. Furthermore, \cite{Raffinot89} and \cite{Alipour2016} evaluate the performance of variant HRP.
However, the HRP focusses on the covariance matrix in computing the weights, consequently, it offers only a conditional mean view of the assets’ connectedness, which may underestimate the real spillover effects among assets in tail events. Taking into consideration tail spillover effects using quantile regression methods seeks to broaden this view, by providing a complete description of the stochastic relationship between assets and offering more robust and more efficient estimation \cite{Lingjie2008}. In this context, our paper contributes to the growing literature by employing an uplifted HRP based on FRM as a quantile regression method.
More precisely, our research introduces two modifications to the HRP algorithm of \cite{LopezdePrado59}: (i) Replace the covariance matrix with the FRM-adjacency matrix. Indeed, the HRP assign portfolio weights without the necessity to invert the covariance matrix. This propriety of HRP makes the replacement of the symmetric covariance matrix with the non-symmetric adjacency matrix possible. (ii) The HRP uses the inverse variance of each asset to calculate the optimal weights, and we replace the variance with the FRM's Lasso penalization parameter ($\lambda$). By introducing these two modifications, we build the uplifted HRP based on the FRM technology.
The HRP portfolio is then benchmarked against the classical HRP, the Minimum variance (MinVar) and inverse variance portfolio (IVP) optimization methods. \vspace{\baselineskip}
Our research investigates a comprehensive set of risk measurement and portfolio optimization in EMs, contributing a new dimension to the existing literature as follows:
(1) The FRM based Lasso quantile regression yields novel insights into the co-movement among FIs and US and EM domestic macroeconomic risk factors. The risk of spillover and its direction are also quantified and visualized, as well as how spillover risk evolves during the financial crisis. Therefore, our research paper is among a few investigating the risk spillover across BRIMST EMs.
To the best of our knowledge, there is a considerable gap in financial literature on the subject of tail risk spillover among US Macro factors and FIs in EMs. This paper is an attempt to bridge this gap by building our research on the above subject. For that purpose, we cover a sample of the biggest six emerging countries from different areas around the world which offers a reasonable basis for comparisons at country and regional market levels. Most of the earlier studies focus on fewer countries, mostly in one region. In addition, we adopt a sample that contains the largest FIs from each country as well as the domestic and US Macro factors to build our FRM network. Hence, our study investigates a comprehensive set of emerging equity markets, contributing a new dimension to the literature on international equity market co-movement that has traditionally focused on developed markets.
(2) Examine the existence of portfolio diversification benefits for foreigners investing in EMs.
Several studies have been conducted to examine the existence of portfolio diversification benefits in less correlated financial markets; \cite{jrfm7020045}; and \cite{SAITI2014196}. However, few other studies investigate this subject from the perspective of an EMs framework [\cite{ARREOLAHERNANDEZ2020101219}].
(3) This paper sheds light on the connection between the portfolio optimisation approaches and the financial tail event networks. Our strategy aims to simplify the portfolio selection procedure by targeting a set of assets within a certain range of network centrality. As far as we are aware, \cite{ARREOLAHERNANDEZ2020101219} is the only paper that attempts to take advantage of the topology of the tail event financial market network for investment purposes. They applied a directional spillover index, the tail-event driven network (TENET), and nonlinear portfolio optimization methods on the bank returns from emerging and developed America.
(4) While most previous studies focus on portfolio optimization based on mean variance estimation, there is a lack of empirical literature on quantile estimation (both tails of distribution). This study attempts to fill this gap by extending the classical HRP approach to build an optimal portfolio based on tail information provided by the FRM quantile regression.
(5) Finally, it is worth mentioning that our results are robust to different estimation time windows, market situations (pre-crisis, crisis, and post crisis), and tail risk levels. We argue that our proposed FRM based portfolio optimisation approach benefits from the tail event co-movement and asset clustering making more useful use of fundamental tail-event information, resulting in an efficient risk portfolio selection strategy. This practice is unique to this research and remains an important contribution to the literature on risk measurement and international portfolio diversification.
\section{Econometric Methodology}
\subsection{FRM Systemic Risk Measure Framework}
This section describes the genesis and framework of FRM. For that purpose, we lay down the fundamental structure and the background of our systemic risk analysis. More precisely, we will state the systemic risk measure models that have been evaluated to lead to an augmented risk measure, the FRM.
\subsubsection{FRM Genesis}
The Value at Risk (VaR) and the expected shortfall (ES) are traditional risk measures. They compute the risk of a given financial institution based either on company characteristics or by introducing macroeconomic variables as a proxy for the state of economy. To exemplify, the VaR of a financial institution $i$ at a quantile level $\tau$ is given by the following equation:
\begin{equation}
P(X_{i,t}\leq VaR_{i,t,\tau})\overset{def}{=}\tau, \quad\tau \in(0, 1)
\label{eq1}
\end{equation}
Where $X_{i,t}$ represent the log return of financial institution $i$ at time $t$.
\cite{adrian2011covar} proposed the Conditional Value at Risk (CoVaR), the CoVaR considers the spillover effects and the macro economy state. The CoVaR of $FI_j$ given $X_i$ at quantile level $\tau$ at time $t$ is given as follow:
\begin{equation}
P(X_{j,t}\leq CoVaR_{j|i,t,\tau}|R_{i,t})\overset{def}{=}\tau,
\label{eq2}
\end{equation}
Where $M_{t-1}$ denotes the vector of macro state variables, and
$R_{i,t}$ indicates the information set that involves the event of $X_{i,t}=VaR_{i,t,\tau}$ and $M_{t-1}$.
The CoVaR is estimated in two steps of linear quantile regression assuming the following equations:
\begin{equation}
X_{i,t}=\alpha_{i}+\gamma_{i}^{\top}M_{t-1}+\varepsilon_{i,t},
\label{eq3}
\end{equation}
\begin{equation}
X_{j,t}=\alpha_{j|i}+\beta_{j|i}X_{i,t}+\gamma_{j|i}^{\top}M_{t-1}+\varepsilon_{j,t},
\label{eq4}
\end{equation}
\begin{equation}
F^{-1}_{\varepsilon_{i,t}}(\tau|M_{t-1})=0\,\, \mbox{and}\,\, F^{-1}_{\varepsilon_{j,t}}(\tau|M_{t-1},X_{i,t})=0
\label{eq5}
\end{equation}
Where $\beta_{j|i}$ in Equation(\ref{eq4}) defines the sensitivity of log return of a company $j$ to variation in tail event log return of a company $i$.
\textbf{Step 1: Estimate the $VaR$ of an institution $i$ ($VaR_i$)}\\
To estimate the $VaR_i$, \cite{adrian2011covar} apply the quantile regression of log return of company $i$ on macro state variables. The estimated equation is defined as follow:
\begin{equation}
\widehat{VaR}^{\tau}_{i,t}=\widehat\alpha_{i}+\widehat\gamma^{\top}_{i}M_{t-1},
\label{eq6}
\end{equation}
\textbf{Step 2: Estimate the $CoVaR$}\\
The CoVaR is computed by integrating the $VaR_{i,t}^{\tau}$ in Equation (\ref{eq6}) into the CoVaR equation as follow:
\begin{equation}
\widehat{CoVaR}^{\tau}_{j|i,t}=\widehat\alpha_{j|i}+\widehat\beta_{j|i}\widehat{VaR}^{\tau}_{i,t}+\widehat\gamma^{\top}_{j|i}M_{t-1}.
\label{eq7}
\end{equation}
Where the coefficient $\beta{j|i}$ of Equation (\ref{eq7}) indicates the degree of interconnectedness.
Therefore, the risk of a financial company $j$ is computed through a $VaR$ and macro state of company $i$.
By setting $j$ equal to the return of a system, and $i$ to be the return of a financial company $i$, we obtain the contribution CoVaR that illustrates how a company $i$ effects the rest of the financial system. Or, by setting $j$ to be a financial company and $i$ to be a financial system, we obtain exposure $CoVaR$, which characterizes how a single institution is exposed to the overall risk of a system.
Recently, \cite{hardle2016tenet} developed the tail-event driven network (TENET). The TENET is a risk approach generalizes CoVaR by joining systemic interconnectedness between FIs and accommodates all system nodes as risk factors. It is estimated based on quantile regressions and illustrated by three steps.
\paragraph{Step 1: Estimate the $VaR$ of each FI}\
The $VaR$ for the studied FIs is estimated using the linear quantile regression based on Equations (\ref{eq3} and \ref{eq6}).
\paragraph{Step 2: The network analysis}\
The TENET estimates the non-linear relationship among FIs, and takes more institutions into consideration to compute the tail-driven risk interdependencies. Therefore, we have:
\begin{equation}
X_{j,t}=\mbox{g}(\beta_{j|R_{j}}^{\top}R_{j,t})+\varepsilon_{j,t},
\label{eq8}
\end{equation}
\begin{equation}
\widehat{CoVaR}^{TENET}_{j|\tilde{R}_{j,t,\tau}}\overset{def}{=}\widehat{\mbox{g}}(\widehat{\beta}_{j|\tilde{R}_{j}}^{\top}\tilde{R}_{j,t}),
\label{eq9}
\end{equation}
\begin{equation}
\widehat{D}_{j|\tilde{R}_{j}}\overset{def}{=}\frac{\partial \widehat{\mbox{g}}(\widehat{\beta}_{j|R_{j}}^{\top}R_{j,t})}{\partial R_{j,t}}|_{R_{j,t}=\tilde{R}_{j,t}}=\widehat{\mbox{g}}^{\prime}(\widehat{\beta}_{j|\tilde{R}_{j}}^{\top}\tilde{R}_{j,t})\widehat{\beta}_{j|\tilde{R}_{j}}
\label{eq10}
\end{equation}
where $R_{j,t}\overset{def}{=}\{{X_{\_j,t}},\,M_{t-1},\,B_{j,t-1}\}$ denotes the set of information. Note that $X_{\_j,t}\overset{def}{=}\{X_{1,t}\,X_{2,t}\,...,X_{1,t}\}$ is a set of $\{k-1\}$ independent explanatory variables, such as the log returns of all the FIs expected the $FI_j$, and $k$ here denotes the number of FIs. The term $B_{j,t-1}$ characterizes the internal factors of the institution $j$.The parameters are defined through the term $\beta_{j|R_{j}}\overset{def}{=}\{\beta_{j|\_{j}},\,\beta_{j|M},\,\beta_{j|B_{j}}\}^{\top}$.
The $\widehat{CoVaR}^{TENET}$ stands for tail-event driven network risk by means of a single-index model (SIM) model and is estimated by plugging in $VaR$ of institution $i$ at level $j|Rj,t$, defined in Equation (\ref{eq6}) into Equation (\ref{eq8}).\\
Note that $\beta_{j|R_{j}}\overset{def}{=}\{\beta_{j|\_{j}},\,\beta_{j|M},\,\beta_{j|B_{j}}\}^{\top}$ and $R_{j,t}\overset{def}{=}\{\widehat{VaR}^{\tau}_{i,t},\,M_{t-1},\,B_{j,t-1}\}.\widehat{\mbox{g}}(\bullet)$ characterize the non-linear relationship among them.
The parameter $D_{j|\tilde{R_j}}$ represents the gradient evaluating the marginal effect of covariates measured at $R_{j,t}=\tilde{R}_{j,t}$. The component wise expression is given by:\\ $\widehat{D}_{j|\tilde{R_j}}\equiv\{\widehat{D}_{j|\_{j}},\,\widehat{D}_{j|M},\,\widehat{D}_{j|B_{j}}\}^{\top}$.
Specifically, $\widehat{D}_{j|\_j}$ permits to quantity spillover effects through the FIs and to illustrate their evolution as a network system.
The TENET network represents a set of FIs links. The estimation results of this interconnectedness can be summarized in weighted form of an adjacency matrix.
Note $\widehat{D}^s_{j|i}$ an element in $\widehat{D}^s_{j|\_j}$ at estimation window $s$ for the $FI_j$ given another $FI_i$ (where $i$ is an element in the other $FI_j$).
Therefore, the adjacency matrix includes absolute values
of $\widehat{D}^s_{j|i}$ (in upper triangular matrix) and $\widehat{D}^s_{i|j}$ (in lower triangular matrix), where $\widehat{D}^s_{j|i}$ denotes the influence from a $FI_i$ to a $FI_j$ and $\widehat{D}^s_{i|j}$ is the influence from $FI_j$ to $FI_i$. The adjacency matrix $A_s=(k*k)$ is represented in the following equation, where for each time windows $s$ only one adjacency matrix is estimated.
\begin{equation}
A_{s} =
\begin{pNiceMatrix}[first-row,first-col,nullify-dots]
&I_1 & I_2 & I_3 & \cdots & I_k\\
I_1 & 0 & \lvert \widehat{D}^s_{1|2}\rvert & \lvert\widehat{D}^s_{1|3}\rvert &\cdots & \lvert\widehat{D}^s_{1|k}\rvert \\
I_2 & \lvert \widehat{D}^s_{2|1}\rvert & 0 &\lvert \widehat{D}^s_{2|3}\rvert & \cdots&\lvert \widehat{D}^s_{2|k}\rvert \\
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots \\
I_k & \lvert \widehat{D}^s_{k|1}\rvert & \lvert \widehat{D}^s_{k|2}\rvert & \lvert \widehat{D}^s_{k|3}\rvert & \cdots & 0
\end{pNiceMatrix}
\label{eq11}
\end{equation}
The matrix $A_s$ summarizes the overall connectedness between variables at time window $s$, and $I_i$ denotes the name of $FI_i$.
\paragraph{Step 3: Systemic risk contributions}\
The objective here is to measure the systemic risk relevance of a specific FI by its total in- and out-connections, weighted by market capitalization.
Hence, we define the Systemic Risk Receiver Index $(SRR)$ for a $FI_j$ at time windows $s$ as follow:
\begin{equation}
SSR_{j,s}\overset{def}{=} MC_{j,s}\Biggl\{\sum_{i\in {k_s^{IN}}}(\lvert \widehat{D}^s_{j|i}\rvert.MC_{i,s})\Biggl\}
\label{eq12}
\end{equation}
The Systemic Risk Emitter Index $(SRE)$ for a $FI_j$ at time windows $s$ is given by the following equation:
\begin{equation}
SSE_{j,s}\overset{def}{=} MC_{j,s}\Biggl\{\sum_{i\in {k_s^{OUT}}}(\lvert \widehat{D}^s_{i|j}\rvert.MC_{i,s})\Biggl\}
\label{eq13}
\end{equation}
Here $k_s^{IN}$ and $k_s^{OUT}$ denotes the set of FIs linked with the $FI_j$ at time windows $s$ by incoming and outgoing links respectively. $MC_{i,s}$ is the market capitalisation of FI $i$ at the starting point of time windows $s$. $\lvert \widehat{D}^s_{j|i}\rvert$ and $\lvert \widehat{D}^s_{i|j}\rvert$ are absolute partial derivatives derived from Equation (\ref{eq10}) which represents row (incoming) and column (outgoing) direction connection of $FI_j$. Thus both $SR{R_j,s}$ and $SRE_{j,s}$ take into consideration the $FI_j$ and its linked FI’ market capitalization as well as its connectedness within our network.
For more details see \cite{hardle2016tenet}.
\subsubsection{FRM Financial Risk Meter}
The FRM is a systemic risk measure based on the penalization parameter$(\lambda)$ of a linear quantile Lasso regression using moving-window approach. In this section we present the methodology and algorithm that constitutes the FRM technology. Since the penalization parameters $(\lambda)$ are computed based on an $L_1$-norm (Lasso) quantile linear regression, this method is introduced first.
\paragraph{Linear Quantile Lasso Regression Model}\
The FRM aims to simultaneously capture all interdependencies in one single number based on the log return of FIs and a set of macroeconomic variables that illustrate the state of the economy. Consider $J$ the number of FI under consideration, where $j\in\{1,...,J\}$. Therefore $p=J+M-1$ denotes the number of covariates. Let $t=\{1,...,T\}$ be the time point, where $T$ denotes the total number of daily observations. Suppose $s$ is the index of time windows, where $s\in\{1,...(T-(n-1))\}$ and $n$ is the length of windows size.
The Linear quantile Lasso regression for return series $X$ is defined as follow:
\begin{equation}
X^{s}_{j,t}=\alpha^{s}_{j,t}+A^{s\top}_{j,t}\beta^{s}_{j}+\varepsilon^{s}_{j,t}, \qquad
\label{eq14}
\end{equation}
Where $A^{s\top}_{j,t}\overset{def}{=}\begin{bmatrix}M^{s}_{t-1}\\
X^{s}_{-j,t}\end{bmatrix}$, $M^s_{t-1}$ is the $M$ dimensional vector of macro variables, $X^{s}_{-j,t}$ is the $p-M$ dimensional vector of log returns of all other FIs except $FI_j$ at time $t$ and in moving window $s$, $\beta^s_{j}$ is a $p\times 1$ vector defined for moving window $s$ and $\alpha^s_{j}$ is a constant term.
The regression is performed using an $L_1$-norm penalisation given a parameter $\lambda_j$, defined by \cite{tibshirani1996regression} as the least absolute shrinkage and selection operater (Lasso). According to \cite{bassett1978asymptotic}, the current company’s $\lambda_j$ are estimated by a modification of Lasso in a quantile regression [see \cite{belloni2011} and \cite{li20081} for more details), where the optimization is solved as follow:
\begin{equation}
\operatorname{\min_{\alpha^{s}_{j},\beta^{s}_{j}}}\Big\{n^{-1}\sum_{t=s}^{s+(n-1)}\rho_{\tau}\,(X_{j,t}^{s}-\alpha^{s}_{j}-A_{j,t}^{s\top}\beta_{j}^{s})+\lambda_{j}^s\parallel\beta_{j}^{s}\parallel_1\Big\}
\label{eq15}
\end{equation}
with the following check function:
\begin{equation}
\rho_\tau\,(u) = |u|^c \,|\tau - \mathrm{I}_{\{u<0\}}|
\label{eq16}
\end{equation}
where $c=1$ corresponds to quantile regression employed here and $c=2$ corresponds to expectile regression.
\paragraph{Penalization Parameter $\lambda$}\
The formula of Lasso’s penalization parameter $\lambda$ can be estimated in a linear regression context, following the work of \cite{osborne2000lasso}
Treating $\lambda$ as a fixed value in the objective function of the penalized regression:
\begin{equation}
f(\beta,\lambda)=\Biggl\{\frac{1}{2}\sum_{i=1}^n (y_{i}-X_{i}^{\top}\beta)^2+\lambda\sum_{i=1}^p|\beta_{j}|\Biggl\}
\label{eq17}
\end{equation}
Here $ f(\beta,\lambda)$ is convex in the parameter $\lambda$. In addition, with diverging $\beta$ we note that $f(\beta,\lambda)\to\infty$. Consequently, the function $f(.,\lambda)$ admits at least one minimum, which is attained in $\widehat{\beta}(\lambda)$ [\cite{osborne1985finite}] if and only if the null vector $0\in R^p$ is an element of sub-differential:
\begin{equation}
\frac{\partial f(\beta,\lambda)}{\partial \beta}=-X^{\top}(Y-X\beta)+\lambda u(\beta)
\label{eq18}
\end{equation}
Were:
\begin{equation*}
u(\beta)=(u_1(\beta),...,u_p(\beta))^{\top} \text{is defined as}
\begin{cases}
u_j(\beta)=1\; \; if\; \;\beta_j>0\\
u_j(\beta)=-1\; \; if\; \;\beta_j<0\\
u_j(\beta)\in[-1,\;1]; \; if\; \;\beta_j=0\\
\end{cases}
\end{equation*}
Therefore, considering that $ f(\beta,\lambda)$ admit a minimum in $\widehat{\beta}$, the following equation has to be satisfied:
\begin{equation}
0=-X^{\top}\{Y-X\widehat{\beta}(\lambda)\}+\lambda u(\widehat{\beta}(\lambda)
\label{eq19}
\end{equation}
The estimator of the vector of parameters $\beta$ here is a function of the penalization parameter $\lambda$. This dependency follows from the formulation of the penalized regression method and its objective function Equation (\ref{eq17}). Following this method we select first the penalization parameter $\lambda$ and then search for $\beta_{\lambda}$ that minimizes Equation (\ref{eq17}). Given that $u(\beta))^{\top} \beta=\sum_{j=1}^p\lvert\beta_j\rvert=\left \Vert \beta\right \|_1$, where $\left \Vert . \right \|_1$ represents here $L1$-norm of a $p$-dimensional vector, Equation (\ref{eq19}) can be rewritten in the following formula:
\begin{equation}
\lambda = \displaystyle \frac{\{ Y - X\beta(\lambda)\}^{\top}X\beta\left(\lambda\right)}{\left\|\beta\right\|_{1}}
\label{eq20}
\end{equation}
Looking to formula Equation (\ref{eq20}), we can define the elements that influence the value of $\lambda$ and its dynamics when treated in a time- varying framework. The following three elements affect the size of $\lambda$:\vspace{\baselineskip}
1. The size of residuals of the model;\vspace{\baselineskip}
2. The absolute size of the coefficients of the model, $\left \Vert\\ \beta\right \|_1$; \vspace{\baselineskip}
3. The singularity of a matrix $X^{\top}X$. \vspace{\baselineskip}
The formulae for the penalization parameter $\lambda$ in a quantile regression problem Equations (\ref{eq15} and \ref{eq16})can be derived following \cite{li20081}:
\begin{equation}
\lambda = \displaystyle\frac{ \left( \theta \right)^{\top}X\widehat{\beta}\left(\lambda\right)}{\left\|\widehat{\beta}(\lambda)\right\|_{1}}
\label{eq21}
\end{equation}
Where $\theta=(\theta_1,...,\theta_n)^{\top}$ satisfies the following conditions:
\begin{equation*}
\theta_i=\begin{cases}
\tau\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;if\;\;\;\;Y_i-X_i^{\top}\widehat\beta(\lambda)>0\\
-(1-\tau)\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;if\;\;\;\;Y_i-X_i^{\top}\widehat\beta(\lambda)<0\\
\in(-(1-\:\tau)\;,\tau)\;\;\;\;if\;\;\;\;Y_i-X_i^{\top}\widehat\beta(\lambda)=0\\
\end{cases}
\end{equation*}
Since Equation (\ref{eq15}) has an $L_1$ loss function and an $L_1$-norm penalty term, the optimization problem is an $L_1$-norm quantile regression estimation problem. The selection of the penalization parameter $\lambda^s_j$ is fundamental. There are different possibilities to choose $\lambda^s_j$, for example with the Generalized Approximate Cross- Validation criterion (GACV) or the Bayesian Information Criterion $(BIC)$. In this regards, \cite{yuan2006gacv} conducted simulations and concluded that GACV outperforms BIC in terms of statistical efficiency. Hence, we estimate $\lambda^s_j$ with the GACV criterion in the FRM model and set $\lambda^s_j$ as the solution of the following minimization problem:
\begin{equation}
\operatorname{min}GACV(\lambda_{j}^{s})=\operatorname{min}\frac{\sum^{s+(n-1)}_{t=s}\rho_{\tau}(X_{j,t}^{s}-\alpha^{s}_{j}-A_{j,t}^{s,\top}\beta_{j}^{s})}{n-df}
\label{eq22}
\end{equation}
where $df$ stands for the estimated effective dimension of the fitted model. The advantage of GACV is that it can be also adopted for $p > n$, which can be crucial for the FRM if the moving window size is small. For further details see \cite{mihoci2020frm}.
\vspace{\baselineskip}
\paragraph{Financial Risk Meter (FRM) }\
The FRM is estimated using a regression analysis as explained above and select the $\lambda_j$ for each FI $j$ using GACV. The distribution of the$\lambda^s_j$ in a moving window gives important information on the network dependencies among the financial nodes. The standard FRM is defined as the average of $\lambda_j$ over the set of $J$ FIs for all windows. It is formally written as follows:
\begin{equation}
FRM=J^{-1}\sum_{j=1}^J\lambda_j
\label{eq23}
\end{equation}
Note that the distribution of $\lambda_j$ provides details regarding the overall market movement and gives information to macro-prudential decision makers about the network dynamics. Therefore, the FRM is adopted to identify the high joint tail event risks resulting from each FI. More precisely, the FI with high $\lambda_j$ exhibits common high stress levels as the FI at the origin of the crisis. Therefore, this FI is considered as having high ”co-stress”.
\vspace{\baselineskip}
\subsection {FRM Network and node centrality measures}
The FRM network illustrates the tail event interaction between the selected FIs based on the adjacency matrix $A_s$ (11). To describe the topology of the FRM networks, we focus on node centrality measures, specifically, eigenvalue centrality, degree centrality, indegree, outdegree, betweenness and closeness. The concept of centrality aims to measure the impact and the importance of a given node in the network. \vspace{\baselineskip}
\paragraph{Eigenvector centrality}\
\cite{bonacich1972factoring} proposed the so-called eigenvector centrality, which evolved to be a standard measure in network analysis.
Consider a network $G = \{N, \omega\}$, constituted by a set of links $\omega$ that connecting pairs of nodes and a set of nodes $N = {1, 2, ... , N}$. If there is a connection between two nodes $i$ and $j$, we denote it as $(i,j)\in\omega$. The network connections are defined by a simplified version of (11): the $(N\times N)$ adjacency matrix $A_{i,j}^s=[\beta_{ij}^s]$ whose element $\beta_{ij}\neq 0$ whenever $(i,j)\in \omega$, $s$ is the rolling window.
\begin{equation}
A^s =
\begin{pmatrix}
0 & \beta_{1,2} & \cdots & \beta_{1,n} \\
\beta_{2,1} & 0 & \cdots & \beta_{2,n} \\
\vdots & \vdots & \ddots & \vdots \\
\beta_{n,1} & \beta_{n,2} & \cdots & 0
\end{pmatrix}
\label{eq24}
\end{equation}
Based on the \cite{bonacich1987power}; \citeyearpar{bonacich1972factoring} definition, the eigenvector centrality of node $i$, $v_i$, is expressed as the proportional sum of its neighbors' centralities. This concept has been extended by \cite{bonanno2004networks} suggesting that for the weighted networks, $v_i$ is relative to the weighted sum of neighbors’ centralities of node $i$ with the adjencency matrix $A_{i,j}^s$ that illustrates the corresponding weighting factors. Therefore, the eigenvector centrality of node $v_i$ is calculated as follows:
\begin{equation}
v_i=\delta^{-1}\sum_j\beta_{ij}^s v_j
\label{eq25}
\end{equation}
where $\delta$ here denotes the eigenvalue. A large value of $v_i$ means that the node $i$ is highly central, implying that node $i$ is linked either to several other nodes or is linked to a few highly central nodes.
The Equation (\ref{eq25}) can be rewritten in matrix terms. We have $\delta v=A v$ specifying that the centrality vector $v$ is defined by the eigenvector of $A^s$ corresponding to the largest eigenvalue $\delta$.\vspace{\baselineskip}
In general, each eigenvector of $A^s$ is a solution to Equation (\ref{eq25}). Nevertheless, the centrality vector matching to the largest network component is specified using the eigenvector corresponding to the largest eigenvalue \cite{bonacich1972factoring} and \cite{peralta2016network}.
\paragraph{Closeness}\
Closeness is a centrality measure proposed by \cite{freeman1978centrality}. It highlights the distance of a given vertex to the rest of vertices in the network. It can be considered as duration of the information spread from one vertex to another. Closeness of a given vertex $j$ is defined as follow:
\begin{equation}
Closeness_j=\sum{_{i=1}^N \frac{1}{d_{(i,j)}}}
\end{equation}
Where $d_{(i,j)}$ measure the distance between a vertix $i$ and another vertix $j$ in the network.
\paragraph{Betweenness}\
Betweenness is a centrality measure, which computes the number of times a vertex lies on the shortest path between other vertices in the network [\cite{vyrost2019network}].\\
Suppose we need to compute this measure for a vertex $j\in\omega$, for any two distinct vertices other than $j$, named $l$ and $k$ $\in\omega$, the number of shortest paths between $l$ and $k$ is $n_{l,j,k}$ $\in N$ and the total number of shortest paths between $l$ and $k$ is $n_{l,k}\in N$, the betweenness for $j$ is then computed as follow:
\begin{equation}
Betweenness_j=\sum_{\substack{l\neq j\neq k{} \\ l,j,k\in \omega}} \frac{n_{l,j,k}}{n_{l,k}}
\label{eq26}
\end{equation}
\paragraph{Degree centrality}\
Degree centrality captures total connectedness in the network, it is given by the following equation:
\begin{equation}
D=\sum_{j=1}^N \sum_{i=1}^N \textbf{1} (\beta_{j,i}^s)
\label{eq27}
\end{equation}
where
\begin{equation*}
\textbf{1} (\beta_{j,i}^s)=\begin{cases}
1\; \; if\; \;\beta_{j,i}^s\neq 0\\
0\; \; if\; \;\beta_{j,i}^s =0\\
\end{cases}
\end{equation*}
\paragraph{Indegree:} Indegree computes the number of inflows, or in this paper the number of other FIs influencing one node. Indegree of $FI_j$ is defined as:
\begin{equation}
Ind_j=\sum_{i=1}^N\textbf{1} (\beta_{j,i}^s)
\label{eq28}
\end{equation}
Here the $FI_j$ can be considered as a risk receiver.
\paragraph{Outdegree:} Outdegree computes the number of out-going links, or in this paper the number of other FIs influenced by the node. Outdgree of $FI_i$ is defined as:
\begin{equation}
Outd_i=\sum_{j=1}^N\textbf{1} (\beta_{j,i}^s)
\label{eq29}
\end{equation}
Here the $FI_i$ can be considered as a "risk emitter".
\section{Portfolio optimisation}
\subsection{The minimum-variance}\
Consider $N$ risky assets with a vector of expected returns $\mu$, and covariance matrix, $\Sigma=[\sigma_{ii}]$. The problem is to define the optimal portfolio weights vector $w$, which minimizes the portfolio variance subject to $w^{\top}\textbf{1}=\textbf{1}$ where $\textbf{1}$ represents a column vector whose components are equivalent to one \cite{markowitz1959portfolio}. This approach is usually defined as minimum-variance or shortly $MinVar$. Formally the problem is stated as follow:
\begin{equation}
min_w\sigma^2_p=w^{\top}\Sigma{w};\;\;\;\;\;\text{subject to}\;\;\;\;w^{\top}\textbf{1}=\textbf{1}
\label{minvareq}
\end{equation}
The solution of Equation (\ref{minvareq}) is:
\begin{equation}
\widehat{w}^*_{minv}=\frac{\textbf{1}}{\textbf{1}^{\top}\Sigma^{-1}1}\Sigma^{-1}\textbf{1}
\label{minvarweights}
\end{equation}
\subsection{A machine learning-based Hierarchical Risk Parity HRP}
\subsubsection{Classical HRP approach}
The HRP is a risk-based portfolio optimization approach that diversifies portfolios without imposing a positive-definite return covariance matrix [\cite{LopezdePrado59}]. The algorithm employs machine learning methods and graph theory to classify the underlying hierarchical correlation structure of the assets, which allow clusters of similar assets to compete for capital allocation. The algorithm of the HRP approach can be broken down into three main steps: tree clustering, quasi-diagonalization, and recursive bisection. In the following, we explain each step in detail. \vspace{\baselineskip}
\paragraph{Step 1- Hierarchical Tree Clustering}\
This step involves breaking down the assets of the considering portfolio into different hierarchical clusters exploiting a Hierarchical Tree Clustering algorithm. The clusters are formed as follows:\\
1. For $N$ stock returns compute the correlation matrix [see \cite{hardle2019mva}(p 431-442)], which gives an $N\times N$ matrix $\Omega$ of these correlations $\rho$, \\
$\rho= \{ \rho_{i,j}\}_{i,j=1,...,N}$\\
Where $\rho_{i,j}$ is the correlation coefficient between a pair of assets $\{i,j\}$ $\rho[X_{i},X_{j}]$\\
2. Transform the correlation matrix to a correlation-distance matrix $D$, where for\\
$d:(X_{i},X_{j}) \subset B \rightarrow {R} \in [0,1];$
\begin{equation}
d_{i,j} = d[X_{i},X_{j}]=\sqrt{\frac{1}{2}(1-\rho_{i,j})}
\label{eq32}
\end{equation}
3. Compute a new distance matrix $\tilde{d}$ where, by taking the Euclidean distance among columns in a pair-wise manner, the augmented distance matrix is given as follow:
\begin{equation}
\tilde{d}= \tilde{d}[d_{i},d_{j}] = \sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^2}
\label{eq33}
\end{equation}
where $\tilde{d}_{i,j}: (d_{i},d_{j}) \subset B \rightarrow {R} \in [0, \sqrt{N}]$.\\
Note that for two assets $i$ and $j$, $ D_{i,j}$ represents the distance between them, however $\tilde{d}_{i,j}$ represents the closeness in similarity of $\{i,j\}$ with the rest of the portfolio. More precisely, a lower $\tilde{d}_{i,j}$ indicates that the assets $i$ and $j$ are similarly correlated to the rest of stocks in the portfolio.\\
4. Form the assets clusters in a recursive manner based in Equation(\ref{eq33}). The set of clusters is donated by $U$ and the first formed cluster$(i^*,j^*)$ is define as:
\begin{equation}
U[1]=(i^{*},j^{*}) =\text{argmin}(i,j)_{i \neq j}\{\tilde{d}_{i,j}\}
\label{eq34}
\end{equation}
5. Update the distance matrix $d$ by computing the distances of other assets from the newly formed cluster $U(1)$ using single linkage clustering. Therefore, for any asset $i$ outside of $U(1)$, the distance from $U(1)$ is updated as follows:
\begin{equation}
d_{i,U[1]}=\min\bigl[\{\tilde{d}_{i,j}\}_{j \in U[1]}\bigl]
\label{eq35}
\end{equation}
Thereby, the algorithm recursively forms assets clusters and updates the distance matrix until we are left with one giant cluster of all stocks.
Finally, the clusters are visualised in a dendrogram. See \cite{hardle2019mva} (p. 363-393) for more details. \vspace{\baselineskip}
\paragraph{Step 2- Quasi Diagonalisation or Matrix Seriation}\
The Quasi-Diagonalisation of the covariance matrix or Matrix seriation is adopted to rearrange the data to clearly represent the inherent clusters. Using the hierarchical clusters from the previous step, the columns and rows of the covariance matrix are reorganized so that similar assets are placed together and dissimilar assets are placed far apart. More precisely, the larger covariances are positioned along the diagonal and smaller ones around this diagonal and since the off-diagonal elements are not completely zero. This is named a quasi-diagonal covariance matrix.\vspace{\baselineskip}
\paragraph{Step 3- Recursive Bisection}\
The final recursive bisection step implicates assigning actual portfolio weights to the assets in the portfolio.\\
1. Assign a unit weight to all assets,
\begin{equation*}
W_i=1,\;\;\;\forall\;i=1,...N
\end{equation*}
2. Bisect each cluster into two sub-clusters in a top-down manner, it means by starting with the topmost cluster, so, for each cluster, we obtain a left and right sub-cluster.\\
3. Calculate the variance for each of these sub-cluster
\begin{equation}
V_{1,2}=w^{\top}\Sigma w\;\;\;\;; \Sigma \;\;\text{is the covariance matrix}
\label{eq36}
\end{equation}
where,
\begin{equation}
w= \frac{\mbox{diag}[\Sigma]^{-1}}{\mbox{tr[diag}[\Sigma]^{-1}]}
\label{IVPweights}
\end{equation}
Since we are dealing with a quasi-diagonal matrix, the algorithm uses the property of the portfolio that the inverse-variance allocation is optimal for a diagonal covariance matrix. Hence, we adopt the inverse-variance allocation weights when computing the variance for sub-clusters.\\
4. Calculate the weighting factor based on the quasi-diagonalised covariance matrix
\begin{equation}
\alpha_{1}=1-\frac{V_{1}}{V_{1}+V_{2}}, \text{so \ that} \ 0 \leq \alpha_{i} \leq 1\; \; ;\alpha_2=1-\alpha_1
\label{eq38}
\end{equation}
5- Update the weights $w_1$ and $w_2$ for both sub-clusters:
\begin{equation}
w'_1=\alpha_1\cdot w_1;\;\;\;\
w'_2=\alpha_2\cdot w_2
\label{eq39}
\end{equation}
6- Execute recursively steps 2-5. The algorithm stops when we have a single asset for each cluster and then the weights are assigned to all assets in the portfolio.\\
Since the weights are assigned in a top-down manner, only assets within each cluster compete for allocation for the final portfolio, rather than competing with all the assets in the portfolio, see \cite{vyrost2019network}.\vspace{\baselineskip}
\subsubsection{An uplifted HRP approach based on FRM}
The classical HRP approach focused on calculating the optimal portfolio weights based on the variance and the covariance matrix, in other words it is focused on estimating the spillover risk on the mean-variance levels, ignoring analysis of other quantile levels. To overcome this limitation, we extend the classical HRP based on FRM in order to take into consideration the hidden interdependency structures among the FIs tail quantile level when optimizing the EMs portfolio.
The basic idea is to use the FRM output in order to compute the portfolio weights based on quantile level analysis. In the classical approach, these weights are computed using the variance (which is a measure of individual risk) and the covariance matrix that measures the relationship between a pair of assets in the mean-variance level. To improve on this, the FRM provides the penalty parameters $(\lambda_{j})$ as a measure of individual risk of each company or FI at the quantile level. In addition, the FRM provides also the adjacency matrix to estimate the interconnectedness between each pair of the studied FIs $(\beta_{i,j})$. By analogy, for the uplifted HRP approach based FRM, we replace the covariance matrix with the adjacency matrix. Recall that the diagonal elements of the adjacency matrix are zero (Equation (\ref{eq24})), which allow us to introduce the vector of penalty parameters $(\lambda_{j})$ in the diagonal of the adjacency matrix to replace the variance in the classical approach.
Following this reasoning, the uplifted HRP steps will be as follows:\vspace{\baselineskip}
\paragraph{Step 1- Hierarchical Tree Clustering}:\\
1. Calculate the the FRM adjacency matrix $A^s$ (Equation (\ref{eq24}))
Where $\beta_{i,j}$ is the degree of connectedness between a pair of assets $\{i,j\}$ \\
2. Introduce the vector of $(\lambda_{j})$ as diagonal elements in the adjacency matrix $A^s$, so we obtain the new adjacency matrix $\tilde{A}$
\begin{equation}
\tilde{A^s} =
\begin{pmatrix}
\lambda_1 & \beta_{1,2} & \cdots & \beta_{1,n} \\
\beta_{2,1} & \lambda_2 & \cdots & \beta_{2,n} \\
\vdots & \vdots & \ddots & \vdots \\
\beta_{n,1} & \beta_{n,2} & \cdots & \lambda_n
\end{pmatrix}
\label{eq40}
\end{equation}
3. Replace the covariance matrix in the HRP algorithm with the adjacency matrix $\tilde{A^s}$.\\
4. Transform $\tilde{A^s}$ to an adjacency-distance matrix $D$,
\begin{equation}
D_{i,j} = D[X_i, X_j]=
\begin{cases}
\sqrt{\frac{1}{2}(1-\beta_{i,j})}\; \;\; \; if \; \;\; \; \vert \beta_{i,j}\vert <1\\
\; \; \\
\mbox{max}D[X_i, X_j]\; \;\; \;\; \;\; \;\; \; otherwise.
\end{cases}
\label{eq41}
\end{equation}
5. Compute a new distance matrix $\tilde{D}$
\begin{equation}
\tilde{D}= \tilde{D}[D_{i},D_{j}] = \sqrt{\sum_{n=1}^{N}(D_{n,i}-D_{n,j})^2}
\label{eq42}
\end{equation}
6. Form the assets clusters in a recursive manner based in Equation(\ref{eq42}). The set of clusters is donated by $\tilde{U}$ and the first formed cluster$(i^*,j^*)$ is define as:
\begin{equation}
\tilde{U}[1]=(i^{*},j^{*}) =\text{argmin}(i,j)_{i \neq j}\{\tilde{D}_{i,j}\}
\label{eq43}
\end{equation}
7. Update the distance matrix $D$ by computing the distances of other assets from the newly formed cluster $\tilde{U}(1)$ as follows:
\begin{equation}
D_{i,\tilde{U}[1]}=\min\bigl[\{\tilde{D}_{i,j}\}_{j \in U[1]}\bigl]
\label{eq44}
\end{equation}
\paragraph{Step 2- Quasi Diagonalisation or Matrix Seriation}:\\
The Quasi-Diagonalisation of the adjacency matrix or Matrix seriation is adopted to rearrange the data to represent clearly the inherent clusters (as explained in the classical approach).
\paragraph{Step 3- Recursive Bisection}:\\
1. Assign a unit weight to all assets,
\begin{equation*}
W_i=1,\;\;\;\forall\;i=1,...N
\end{equation*}
2. Bisect each cluster into two sub-clusters.\\
3. Calculate the variance for each sub-cluster
\begin{equation}
\tilde{V}_{1,2}=\tilde{w}^{\top}\tilde{A} \tilde{w}\;\;\;\;; \tilde{A} \;\;\text{is the adjacency matrix}
\label{eq45}
\end{equation}
where,
\begin{equation}
\tilde{w}= \frac{\mbox{diag}[\tilde{A}]^{-1}}{\mbox{tr[diag}[\tilde{A}]^{-1}]}
\label{eq46}
\end{equation}
4. Calculate the weighting factor based on the quasi-diagonalised adjacency matrix.
\begin{equation}
\tilde{\alpha}_{1}=1-\frac{\tilde{V}_{1}}{\tilde{V}_{1}+\tilde{V}_{2}}, \text{so \ that} \ 0 \leq \tilde{\alpha}_{i} \leq 1\; \; ;\tilde{\alpha}_2=1-\tilde{\alpha}_1
\label{eq47}
\end{equation}
5- Update the weights $\tilde{w}_1$ and $\tilde{w}_2$ for both sub-clusters:
\begin{equation}
\tilde{w}'_1=\tilde{\alpha}_1\cdot w_1;\;\;\;\
\tilde{w}'_2=\tilde{w}{\alpha}_2\cdot w_2
\label{eq48}
\end{equation}
6- Execute recursively steps 2-5, the algorithm stops when we have a single asset for each cluster and then the weights are assigned to all assets in the portfolio.\vspace{\baselineskip}
\section{Financial Risk Meter in Emerging Markets}
\label{FRMEMinterp}
\subsection{FRM@EM data-set description}
In this paper we study the largest 25 Emerging Market financial institutions (FIs) by market capitalisation at any given point in time, with focus on the BRIMST FIs. We compile a database of daily price levels in as well as market capitalisations in U.S. Dollars from Bloomberg, and select the biggest financial institutions on a daily basis from the most liquid local EM equity market indices. On any given trading day in consideration, we take the price returns of those biggest $j=25$ FIs over an estimation window $s=63$ business days.
As for the macroeconomic data, we follow \cite{tobias2016covar} concerning the developed market specific risk factors, to repeat, returns in US REITs, S\&P 500 index, U.S. three months treasury bill rates, the spread between 3 months and 10 year U.S. treasury rates, the spread between BAA rates corporate bonds by the rating agency Moody's to U.S. treasury bonds, and the implied volatility index VIX based on outstanding options on the S\&P 500 equity index. We add the following EM specific macroeconomic risk variables: The J.P. Morgan Emerging Bond Index Global Sovereign Spread index tracking the Emerging Bond Index yields over the benchmark U.S. Treasury bonds and thereby representing the risk compensation demanded from investors when investing in EM sovereign bonds, as well as the respective countries currency versus the U.S. dollar cross.
\subsection{FRM@EM Interpretation and Network Analysis}
Figure \ref{FRMEMtimeseriestotal} depicts the FRM@EM from April 2000 to June 2020. Clearly observable are the periods of distress in the global financial system around 2008, 2012, more recently in 2020 but also during the EM specific market distress periods such as 2002 (Argentina) and 2013 (following the Federal Reserve Board's Open Market Committee forward guidance). \\
Figure \ref{FRMEMboxplot} introduces one of the FRM's tool: FRM is not just the mean but actually a distribution of $\lambda^s_j$ as well as the adjacency matrix containing all (Lasso penalised) $\beta^{s}_{j}$ between FIs, and also between FIs and macroeconomic risk variables. As an example of the information contained, we mark some of the more extreme maxima in Figure \ref{FRMEMboxplot}. For example, into the crisis, Standard Bank Group (SBK SJ) had a very high $\lambda^s_j$ reading, indicating the bank was a "risk receiver", thus at risk to be impacted by spill over effects. \\
But we can also have a more detailed look at the adjacency matrix containing the $\beta^{s}_{j}$ between FIs and macroeconomic risk variables, with two examples shown for 20200429 in Figure \ref{ADJMAT20200429005} as well as post crisis 20200630 in Figure \ref{ADJMAT20200630005}, both estimated at $\tau=0.05$, and colour scaled in blue (low/negative) to high (red/positive). Observable are the high dependencies between countries of the same region (\textbf{B}S for Brazil, \textbf{R}M for Russia, \textbf{I}S for India, \textbf{M}F for Mexico, \textbf{S}J for South Africa and \textbf{T}I for Turkey), and often negative relationships (adjusted for co-movements with macroeconomic risk variables) in between regions. However, there are also detectable co-dependencies, which necessitate closer inspection from both investors as well as regulators. For example in Figure \ref{ADJMAT20200429005} South Africa's Sanlam Ltd. (SLM SJ)'s returns are explained to a significant extend by Bajaj Finance Ltd. of India (BAF IS), as both financial services companies, providing an assortment of financial services. Similarly, Banco BTG Pactual (BPAC11 BS)'s returns are explained not only by other Brazilian FIs, but also significantly by Russia's VTB Bank (VTBR RM), both operating in more banking and investment banking related markets. Clearly, sub-sector dependencies across EM FIs are to be considered to prevent risk clusters. \\
Another component to consider is the impact from macroeconomic risk variables' changes on FIs returns. As can be seen on both adjacency matrices, the "classic" macroeconomic risk variables do have an impact, however, to a large extend, the Emerging Market Sovereign Yield Spread to U.S. Treasuries (JPEGSOSD) is the main influence. In fact, this is true across from $\tau=0.05$ to $\tau=0.50$. In Figure \ref{fig:macro} we show the smoothed (rolling seven day mean) of the share of FIs impacted by the respective macroeconomic risk variable. The Emerging Market Yield spread is the dominant driver, followed by more general market risk measures such as the VIX, Moody's Baa corporate yield spreads, and the shape of the U.S. Treasury yield curve. Emerging market currencies as a cross versus the U.S. Dollar have a lesser impact. It is mostly one or two EM currencies having an overall impact, and not only on domestic banks. In Figure \ref{ADJMAT20200630005} for example, the Brazilian Real (USDBRL) has marginal negative return contribution to Bajaj Finance Ltd in India (BAF IS), and positive return contribution to Brazilian Itausa SA (ITSA4 BS) and Mexican Grupo Elektra SAB (ELEKTRA MF). \\
In Figure \ref{FRMcentmeasures}, we show the time series of FRM against various centrality measures. We observe that Betweenness, Eigenvector have similar trends as the FRM EM series, especially during the crisis period of March to May 2020. On the other hand, Closeness centrality drops sharply into the crisis period.
When the FRM rises, the number of $\beta^{s}_{ij}$ equal to zero increase, increasing with it the distance between the vertices. The average length of one node (one FI in our case) and all other FIs increases, thereby sharply reducing Closeness centrality.
Betweeness as a centrality measure of a vertex within a network rises when the FRM rises, as information flow has a very high probability to pass between some central FIs, indicating a concentration of risk around certain FIs which we call "risk emitters".
Similarly, Eigenvector centrality is a measure of an FIs influence in an observed financial system network. Central FIs Eigencentrality rises sharply around a crisis period, as the FRM increases in value.
In- and Out-degree centrality drop when FRM rises, since the edges or connections between FIs have reduced sharply, to mostly the network's risk emitters.
We can conclude therefore, that a close inspection if the distribution of $\lambda^s_j$ as well as the detailed information within the adjacency matrix across a range of $\tau$ are particularly important. In Section \ref{portfolioconstruction} we aim to make use of this richness of information for the construction of more robust, tail-event network behaviour attentive portfolios. As an indication, Figure \ref{FRMEMnetwork} shows a network graph with edges between the 25 largest FIs, estimated at $\tau=0.05$ on 20200429. We highlight one exemplary bank, and its edges stemming from the adjacency matrix. Condensing such information into clusters of risk as outlined above is the focus of the following portfolio construction discussion.
In Figure \ref{fig:macro} we depict the macroeconomic risk variables influence over time. Clearly, the EM Sovereign Yield Spread to U.S. Treasury Bonds is the prominent macroeconomic risk variable not only during tail events but also at "normal" times estimated at $\tau=0.50$. We can conclude that most EM risk is rapidly priced into yield spreads, and then consequently impacts financial institutions. This link between the sovereign and banks needs to be considered for investors and policy makers as well. Other strongly influencing risk variables are more expected such as the VIX, the US yield curve shape, and Corporate Bond yield spreads. However, EM currency fluctuations only have a marginal effect in tail events, on some of the FIs. With the increase of debt issuance in local currency denominated debt, the risk of mismatches versus developed market currencies has diminished overall, and has lower influence on EM FIs.
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{Graphics/FRM-EM.png}
\caption{FRM@EMs Time series}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{FRMEMtimeseriestotal}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{Graphics/FRM_EM_Boxplot.png}
\caption{FRM@EM Boxplot, with \textcolor{blue}{mean} and \textcolor{red}{maximum} at $\tau=0.05$}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{FRMEMboxplot}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Macro_005.png}
\caption{$\tau=0.05$}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Macro_010.png}
\caption{$\tau=0.1$}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Macro_025.png}
\caption{$\tau=0.25$}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Macro_050.png}
\caption{$\tau=0.5$}
\end{subfigure}
\caption{Macro variables' most frequent marginal return contribution across time with 7-day moving average: \textcolor{red}{EM Sovereign Spread}, \textcolor{pink}{VIX}, \textcolor{violet}{U.S. 3mth to 10yr yield spread}, \textcolor{brown}{Moody's BAA Corporate Yield Spread}, \textcolor{green}{S\&P 500 Index}}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{fig:macro}
\end{figure}
\begin{sidewaysfigure}
\centering
\includegraphics[width=\textwidth, scale=2]{Graphics/ADJMAT_20200429_005.pdf}
\caption{Adjacency matrix estimated at $\tau=0.05$ on 20200429}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{ADJMAT20200429005}
\end{sidewaysfigure}
\begin{sidewaysfigure}
\centering
\includegraphics[width=\textwidth, scale=2]{Graphics/ADJMAT_20200630_005.pdf}
\caption{Adjacency matrix estimated at $\tau=0.05$ on 20200630}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{ADJMAT20200630005}
\end{sidewaysfigure}
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{Graphics/FRM_EM_Network_20200429_005.png}
\caption{FRM@EM Network with one \textcolor{orange}{example node} and its \textcolor{blue}{in-degree} and \textcolor{orange}{out-degree} edges}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{FRMEMnetwork}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_betweenness_avg.png}
\caption{{\color{blue} FRM} and \color{red}{Betweenness}}
\label{FRMbetweenness}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_closeness_avg.png}
\caption{{\color{blue} FRM} and \color{red}{Closeness}}
\label{FRMclosness}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_eigenvector_avg.png}
\caption{{\color{blue} FRM} and {\color{red} Eigenvector}}
\label{FRMeigenvector}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_indegree_avg.png}
\caption{{\color{blue} FRM} and {\color{red} In-degree}}
\label{FRMindegree}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_outdegree_avg.png}
\caption{{\color{blue} FRM} and {\color{red} Out-degree}}
\label{FRMoutdgree}
\end{subfigure}
\caption{{\color{blue} FRM} at $\tau=0.05$ and {\color{red} Centrality measures}}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{FRMcentmeasures}
\end{figure}
\section{Portfolio Construction}
\label{portfolioconstruction}
The portfolio is constructed using daily FIs prices for the period between 20200101 and 20200630 (which yields 130 observations of the 25 biggest EM FIs).\\
The performance of the uplifted portfolio approaches (Inv$\lambda$ and upHRP) are benchmarked against the minimum-variance portfolio (MinVar), the inverse variance (IVP), and the classical HRP approaches. The MinVar weights are simply computed based on Equation (\ref{minvarweights}), and often lead to concentration in low-volatility assets. The IVP strategy can be considered as a naive risk parity strategy (Equation (\ref{IVPweights})), since it is agnostic with respect to asset correlations. Thereafter, an overlapping region between portfolio optimisation startegies and FRM network centrality is developed.\\
As shown in the FRM network (Figure \ref{FRMEMnetwork}), all FIs are treated as potential substitutes without specifying any hierarchical structure among them. Therefore, tree structures that integrate hierarchical relationships are needed. For that purpose, the upHRP algorithm aims to build and make use of a clustered adjacency matrix.
\paragraph{\textbf{Matrix seriation}}\
For the classical approach, the steps of the HRP are applied using the price time series of the 25 biggest FIs selected during the studied period. Figure \ref{fig:Classical HRP: Matrix Seriation} presents the estimation results:
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Corr_Heatmap_EM_Cla.HRP_unsorted.png}
\caption{Unclustered Correlations}
\label{UnclusteredCorrelations}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Corr_Heatmap_EM_Cla.HRP_sorted.png}
\caption{clustered Correlations}
\label{ClusteredCorrelations}
\end{subfigure}
\caption{Classical HRP: Matrix Seriation}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{fig:Classical HRP: Matrix Seriation}
\end{figure}
Figure \ref{UnclusteredCorrelations} shows the original or the unclustered correlation matrix. Figure \ref{ClusteredCorrelations} illustrates the correlation matrix after reordering in clusters using the hierarchical tree clustering, also called clustered correlation matrix. This matrix then serves as the input for the asset allocation procedure. As shown in Figure \ref{ClusteredCorrelations} the reorganizing results group similar FIs together and the dissimilar further away in a new correlation matrix, which helps to construct more meaningful asset allocation decisions and building more risk diversified portfolios. More precisely, from Figure \ref{ClusteredCorrelations} the lighter-colored squares (indicating a higher correlation coefficient) are all concentrated around the diagonal matrix.\vspace{\baselineskip}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Corr_Heatmap_EM_m.HRP_unsorted_0505.png}
\caption{Unclustered Adjacency matrix}
\label{UnclusteredAdjmatx}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Corr_Heatmap_EM_m.HRP_sorted_0505.png}
\caption{clustered Adjacency matrix}
\label{ClusteredAdjmatx}
\end{subfigure}
\caption{Uplifted HRP: Matrix Seriation}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{fig:Uplifted HRP: Matrix Seriation}
\end{figure}
For the uplifted approach, the quasi-diagonalization step is applied to the adjacency matrix dated on 20200630 with $\tau=0.05$. In this stage, the algorithm aims at reordering the adjacency matrix by placing similar assets together. As we can notice from Figure \ref{ClusteredAdjmatx} the closer assets (similar given ($\beta$)) are placed together forming assets clusters.\vspace{\baselineskip}
\textbf{Tree clustering or Dendrogram}\vspace{\baselineskip}
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{Graphics/Dendrogram_EM_Cla.HRP.png}
\caption{Classical HRP: EM@Dendrogram}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{ClassicalHRPEM@Dendrogram}
\end{figure}
The clusters are visualised in the form of a cluster diagram called a dendrogram. Figure \ref{ClassicalHRPEM@Dendrogram} illustrates the hierarchical clusters for our FIs data, where the x-axis indicates the name of the FIs in the studied portfolio and the y-axis measures the distance between the two merging FIs.
The key to interpreting a dendrogram is to focus on the height at which any two FIs are joined together. In Figure \ref{ClassicalHRPEM@Dendrogram}, we can see that Bajaj Finance (BAF IS) and Bajaj Finserv Ltd (BJFIN IS) (plotted with green) are expectedly most similar, as the height of the link that joins them together is the lowest. Note that both of them are same ultimate parent Indian financial services companies focused on insurance. The dendrogram also detects that the most similar FIs most often belong to the same market (FirstRand Ltd (FSR SJ) and Standard Bank Group Ltd (SBK SJ) plotted with red), or a similar financial sector. Finally, the highest cluster represents the giant cluster that joins all the FIs formed clusters together.
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{Graphics/Dendrogram_EM_M.HRP_20200630.png}
\caption{Uplifted HRP: EM@Dendrogram}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{upHRPEM@Dendrogram}
\end{figure}
Based on the FRM clustered adjacency matrix, Figure \ref{upHRPEM@Dendrogram} indicates that the tree clustering architecture differs from the previous one (Figure \ref{ClassicalHRPEM@Dendrogram}). Indeed, here State Bank of India (SBIN IS) is a perturbation of BJFIN IS, and consequently the two assets present the first formed cluster since they are the most similar FIs. Also, Grupo Financiero Banorte (GFNORTEO MF) and VTR Bank (VTRB RM) are perturbations of Sberbank of Russia (SBER RM), hence these three FIs are clustered together, forming a cluster of two different markets, contrary to the previous dendrogram, Figure \ref{ClassicalHRPEM@Dendrogram}, where the most similar FIs belong the same market. \vspace{\baselineskip}
\paragraph{\textbf{Calculate allocation through recursive bisection}}\
Table \ref{WeightallocationsofFIsClassicalvsUplifted} and Figure \ref{figWeights allocationsFIs} specify further explanation regarding the weight’s allocation of the studied FIs according to the adopted strategies.
\begin{table}[H]
\begin{center}
\caption{Weight allocations of FIs: Classical approach vs {\color{blue} uplifted approach}}
\label{WeightallocationsofFIsClassicalvsUplifted}
\begin{tabular}{lccccc}
\hline\\
\multicolumn{1}{c}{\textbf{EM@ FIs}} & \textbf{MinVar} & \textbf{IVP} & \textbf{HRP} & \textbf{{\color{blue}Inv$\lambda$}} & \textbf{{\color{blue}upHRP}} \\
\hline
{\color[HTML]{212121} HDFCB.IS.EQUITY} & {\color[HTML]{212121} 0.002} & {\color[HTML]{212121} 0.045} & {\color[HTML]{212121} 0.017} & {\color[HTML]{212121} 0.039} & {\color[HTML]{212121} 0.051} \\ \hline
{\color[HTML]{212121} SBER.RM.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.054} & {\color[HTML]{212121} 0.020} & {\color[HTML]{212121} 0.021} & {\color[HTML]{212121} 0.018} \\ \hline
{\color[HTML]{212121} ITUB4.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.019} & {\color[HTML]{212121} 0.074} \\ \hline
{\color[HTML]{212121} HDFC.IS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.032} & {\color[HTML]{212121} 0.007} & {\color[HTML]{212121} 0.065} & {\color[HTML]{212121} 0.023} \\ \hline
{\color[HTML]{212121} KMB.IS.EQUITY} & {\color[HTML]{212121} 0.062} & {\color[HTML]{212121} 0.038} & {\color[HTML]{212121} 0.011} & {\color[HTML]{212121} 0.036} & {\color[HTML]{212121} 0.003} \\ \hline
{\color[HTML]{212121} BBDC3.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.055} & {\color[HTML]{212121} 0.074} \\ \hline
{\color[HTML]{212121} BBDC4.BS.EQUITY} & {\color[HTML]{212121} 0.002} & {\color[HTML]{212121} 0.005} & {\color[HTML]{212121} 0.002} & {\color[HTML]{212121} 0.017} & {\color[HTML]{212121} 0.030} \\ \hline
{\color[HTML]{212121} ICICIBC.IS.EQUITY} & {\color[HTML]{212121} 0.002} & {\color[HTML]{212121} 0.027} & {\color[HTML]{212121} 0.005} & {\color[HTML]{212121} 0.022} & {\color[HTML]{212121} 0.072} \\ \hline
{\color[HTML]{212121} QNBFB.TI.EQUITY} & {\color[HTML]{212121} 0.008} & {\color[HTML]{212121} 0.009} & {\color[HTML]{212121} 0.006} & {\color[HTML]{212121} 0.084} & {\color[HTML]{212121} 0.001} \\ \hline
{\color[HTML]{212121} SBIN.IS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.035} & {\color[HTML]{212121} 0.010} & {\color[HTML]{212121} 0.030} & {\color[HTML]{212121} 0.044} \\ \hline
{\color[HTML]{212121} B3SA3.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.040} & {\color[HTML]{212121} 0.053} \\ \hline
{\color[HTML]{212121} SANB11.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.024} & {\color[HTML]{212121} 0.007} & {\color[HTML]{212121} 0.029} & {\color[HTML]{212121} 0.107} \\ \hline
{\color[HTML]{212121} BBAS3.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.016} & {\color[HTML]{212121} 0.025} & {\color[HTML]{212121} 0.032} & {\color[HTML]{212121} 0.012} \\ \hline
{\color[HTML]{212121} BAF.IS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.021} & {\color[HTML]{212121} 0.014} & {\color[HTML]{212121} 0.056} & {\color[HTML]{212121} 0.012} \\ \hline
{\color[HTML]{212121} ITSA4.BS.EQUITY} & {\color[HTML]{212121} 0.001} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.018} & {\color[HTML]{212121} 0.024} \\ \hline
{\color[HTML]{212121} AXSB.IS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.018} & {\color[HTML]{212121} 0.009} & {\color[HTML]{212121} 0.041} & {\color[HTML]{212121} 0.056} \\ \hline
{\color[HTML]{212121} ELEKTRA.MF.Equity} & {\color[HTML]{212121} 0.866} & {\color[HTML]{212121} 0.503} & {\color[HTML]{212121} 0.787} & {\color[HTML]{212121} 0.032} & {\color[HTML]{212121} 0.073} \\ \hline
{\color[HTML]{212121} FSR.SJ.EQUITY} & {\color[HTML]{212121} 0.001} & {\color[HTML]{212121} 0.033} & {\color[HTML]{212121} 0.018} & {\color[HTML]{212121} 0.062} & {\color[HTML]{212121} 0.016} \\ \hline
{\color[HTML]{212121} SBK.SJ.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.029} & {\color[HTML]{212121} 0.011} & {\color[HTML]{212121} 0.074} & {\color[HTML]{212121} 0.034} \\ \hline
{\color[HTML]{212121} BBSE3.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.052} & {\color[HTML]{212121} 0.041} \\ \hline
{\color[HTML]{212121} BJFIN.IS.EQUITY} & {\color[HTML]{212121} 0.004} & {\color[HTML]{212121} 0.026} & {\color[HTML]{212121} 0.012} & {\color[HTML]{212121} 0.025} & {\color[HTML]{212121} 0.035} \\ \hline
{\color[HTML]{212121} GFNORTEO.MF.Equity} & {\color[HTML]{212121} 0.044} & {\color[HTML]{212121} 0.035} & {\color[HTML]{212121} 0.017} & {\color[HTML]{212121} 0.037} & {\color[HTML]{212121} 0.004} \\ \hline
{\color[HTML]{212121} BPAC11.BS.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.010} & {\color[HTML]{212121} 0.003} & {\color[HTML]{212121} 0.032} & {\color[HTML]{212121} 0.016} \\ \hline
{\color[HTML]{212121} SLM.SJ.EQUITY} & {\color[HTML]{212121} 0.007} & {\color[HTML]{212121} 0.039} & {\color[HTML]{212121} 0.018} & {\color[HTML]{212121} 0.045} & {\color[HTML]{212121} 0.084} \\ \hline
{\color[HTML]{212121} VTBR.RM.EQUITY} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0} & {\color[HTML]{212121} 0.037} & {\color[HTML]{212121} 0.043} \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{1\linewidth}
\includegraphics[width=\linewidth]{Graphics/Comparison_of_weights_for_HRP_IVP_and_MinVar.pdf}
\caption{Optimal weights {\color{cyan} MinVar}, {\color{yellow} IVP}, {\color{green} HRP}}
\label{classicalWeightsallocations}
\end{subfigure}
\begin{subfigure}[b]{1\linewidth}
\includegraphics[width=\linewidth]{Graphics/A_comparison_of_weights_for_upHRP_and_InvLambda_based_FRM.pdf}
\caption{Optimal weights {\color{orange} Inv$\lambda$}, {\color{blue} upHRP}
\bigskip
}
\label{UpliftedWeightsallocations}
\end{subfigure}
\caption{Weights allocations of FIs: classical approach vs uplifted approach}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{figWeights allocationsFIs}
\end{figure}
\
Table \ref{WeightallocationsofFIsClassicalvsUplifted} and Figure \ref{figWeights allocationsFIs} specify further explanation regarding the weight’s allocation of the studied FIs according to the adopted strategies.
Several points are worth noting. First, for both classical and uplifted approaches, IVP, HRP, Inv$\lambda$ and upHRP and similarly allocate their assets. This result can be explained by the fact that they are based on the inverse variance method to compute their weights. However, while IVP and Inv$\lambda$ show more stable weight distributions, HRP and upHRP adjust their weight allocation more frequently, since it also takes into consideration the correlation (for HRP)/ tail connectedness (for upHRP) and the asset clusters when assigning weights. Second, results also show that the MinVar allocations have the most concentrated weight allocation. While it accorded zero allocation to certain FIs (i.e SBIN IS, B3SA3 BS, etc.), it significantly overweights others (ELEKTRA MF).
Third, the MinVar, IVP and HRP classical approaches over-weighted the Mexican ELEKTRA MF (0.866, 0.503, 0.787, respectively). In this case, any distress situation affecting this FI will have a great impact on those concentrated portfolios. Contrary, these extreme weight concentrations disappear with the Inv$\lambda$ and upHRP uplifted approaches, where the weights are wider distributed among FIs, providing a well-diversified portfolio.
\begin{table}[H]
\begin{center}
\caption{ Weight's allocation of EMs: Classical approach vs {\color{blue}Uplifted approach}}
\label{WeightsallocationEMs}
\begin{tabular}{lrrrrrr}
\hline
{\color[HTML]{212121} \textbf{EM@ FIs}} & \multicolumn{1}{l}{{\color[HTML]{212121} \textbf{Nbr of FIs}}} & \multicolumn{1}{l}{{\color[HTML]{212121} \textbf{MinVar}}} & \multicolumn{1}{l}{{\color[HTML]{212121} \textbf{IVP}}} & \multicolumn{1}{l}{{\color[HTML]{330001} \textbf{HRP}}} & \multicolumn{1}{l}{{\color[HTML]{212121} \textbf{{\color{blue}Inv$\lambda$}}}} & \multicolumn{1}{l}{{\color[HTML]{212121} \textbf{{\color{blue}upHRP}}}} \\ \hline
{\color[HTML]{212121} \textbf{India (IS)}} & {\color[HTML]{212121} 8} & {\color[HTML]{212121} 7 \%} & {\color[HTML]{212121} 24.2\%} & {\color[HTML]{212121} 8.5\%} & {\color[HTML]{212121} 29.2\%} & {\color[HTML]{212121} 22.4\%} \\ \hline
{\color[HTML]{212121} \textbf{Brasil (BS)}} & {\color[HTML]{212121} 9} & {\color[HTML]{212121} 0.3\%} & {\color[HTML]{212121} 5.5\%} & {\color[HTML]{212121} 3.7\%} & {\color[HTML]{212121} 29.4\%} & {\color[HTML]{212121} 43.1\%} \\ \hline
{\color[HTML]{212121} \textbf{Mexico (MF)}} & {\color[HTML]{212121} 2} & {\color[HTML]{212121} 91 \%} & {\color[HTML]{212121} 53.8\%} & {\color[HTML]{212121} 80.4\%} & {\color[HTML]{212121} 6.9\%} & {\color[HTML]{212121} 7.7\%} \\ \hline
{\color[HTML]{212121} \textbf{Russia (RM)}} & {\color[HTML]{212121} 2} & {\color[HTML]{212121} 0 \%} & {\color[HTML]{212121} 5.4\%} & {\color[HTML]{212121} 2 \%} & {\color[HTML]{212121} 5.8\%} & {\color[HTML]{212121} 6.1\%} \\ \hline
{\color[HTML]{212121} \textbf{Turkey (TI)}} & {\color[HTML]{212121} 1} & {\color[HTML]{212121} 0.8\%} & {\color[HTML]{212121} 0.1\%} & {\color[HTML]{212121} 0.6\%} & {\color[HTML]{212121} 8.6\%} & {\color[HTML]{212121} 0.1\%} \\ \hline
{\color[HTML]{212121} \textbf{S.Africa (SJ)}} & {\color[HTML]{212121} 3} & {\color[HTML]{212121} 0.8\%} & {\color[HTML]{212121} 10.1\%} & {\color[HTML]{212121} 4.7\%} & {\color[HTML]{212121} 18.1\%} & {\color[HTML]{212121} 13.7\%} \\ \hline
\end{tabular}
\end{center}
\end{table}
Table \ref{WeightsallocationEMs} reports the percentage of weight allocation for the different approaches, giving further intuition on how the portfolio optimization approaches allocate their weights through EMs. From these allocation results, we can understand a few stylized features: First, the MinVar concentrates 91\% of the allocation on the two existing Mexican FIs and it assigns zero weights for the Russian FIs and only 0.3\% for the 9 Brazilian FIs. Similarly, the IVP and the HRP concentrate allocation on the same market with 53.8\% and 80.4\% weights respectively. However, they provide some diversification across the other markets compared to MinVar strategy. The classical approaches concentrate their weights on the Mexican market that contributed only by two FIs to the basket of the 25 biggest FIs in the EMs, which poses significant idiosyncratic risk.
Nevertheless, for the uplifted portfolio approaches, the weights are well distributed across the EMs. For example, the Inv$\lambda$ allocates 29\% of weights to the Brazilian and Indian FIs, while, the upHRP allocates 29.4\% to Indian FIs and 43.1\% to the Brazilian FIs since this last market has the highest number of FIs in the selection of the 25 biggest EM@FIs. Moreover, the upHRP allocates only 0.1\% for the Turkish FIs, since this market contributes only by one FI.
To recapitulate, the upHRP appears to find a compromise between classical concentrated weight approaches and the Inv$\lambda$ strategy. The classical strategies can concentrate weights on a few FIs, leading to vulnerabilities. The Inv$\lambda$ evenly assigns weights across all FIs, ignoring the correlation structure. This makes it exposed to systemic shocks. However, the upHRP finds a compromise between diversifying across all FIs and diversifying across clusters, which makes it more resistant against both types of shocks.\vspace{\baselineskip}
\paragraph{\textbf{Backtesting results}}
\begin{table}[H]
\begin{center}
\caption{Backtesting on EM FIs data: Classical approach vs {\color{blue} uplifted approach}}
\label{BacktestingEMFIsdata}
\begin{tabular}{ccccc}
\hline
\multicolumn{1}{l}{} & {\color[HTML]{212121} \textbf{Mean}} & {\color[HTML]{212121} \textbf{Std}} & {\color[HTML]{212121} \textbf{Sharpe ratio}} & {\color[HTML]{212121} \textbf{Effective n}} \\ \hline
{\color[HTML]{212121} \textbf{MinVar}} & {\color[HTML]{212121} 0.0043} & {\color[HTML]{212121} 0.0818} & {\color[HTML]{212121} 0.0528} & {\color[HTML]{212121} 1.3225} \\ \hline
{\color[HTML]{212121} \textbf{IVP}} & {\color[HTML]{212121} 0.0039} & {\color[HTML]{212121} 0.0712} & {\color[HTML]{212121} 0.0560} & {\color[HTML]{212121} 3.7121} \\ \hline
{\color[HTML]{212121} \textbf{HRP}} & {\color[HTML]{212121} 0.0035} & {\color[HTML]{212121} 0.0490} & {\color[HTML]{212121} 0.0721} & {\color[HTML]{212121} 1.6063} \\ \hline
\multicolumn{5}{l}{$\tau=5\%$} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}Inv$\lambda$}}} & {\color[HTML]{212121} 0.0767} & {\color[HTML]{212121} 0.1859} & {\color[HTML]{212121} 0.4128} & {\color[HTML]{212121} 20.986} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}upHRP}}} & {\color[HTML]{212121} 0.0522} & {\color[HTML]{212121} 0.1886} & {\color[HTML]{212121} 0.2768} & {\color[HTML]{212121} 16.801} \\ \hline
\multicolumn{5}{l}{$\tau=10\%$} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}Inv$\lambda$}}} & {\color[HTML]{212121} 0.0894} & {\color[HTML]{212121} 0.1868} & {\color[HTML]{212121} 0.4788} & {\color[HTML]{212121} 21.643} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}upHRP}}} & {\color[HTML]{212121} 0.0807} & {\color[HTML]{212121} 0.1967} & {\color[HTML]{212121} 0.4104} & {\color[HTML]{212121} 16.296} \\ \hline
\multicolumn{5}{l}{$\tau=25\%$} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}Inv$\lambda$}}} & {\color[HTML]{212121} 0.0894} & {\color[HTML]{212121} 0.2195} & {\color[HTML]{212121} 0.4073} & {\color[HTML]{212121} 23.429} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}upHRP}}} & {\color[HTML]{212121} 0.0792} & {\color[HTML]{212121} 0.2280} & {\color[HTML]{212121} 0.3476} & {\color[HTML]{212121} 20.199} \\ \hline
\multicolumn{5}{l}{$\tau=50\%$} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}Inv$\lambda$}}} & {\color[HTML]{212121} 0.0883} & {\color[HTML]{212121} 0.2093} & {\color[HTML]{212121} 0.4219} & {\color[HTML]{212121} 21.117} \\ \hline
{\color[HTML]{212121} \textbf{{\color{blue}upHRP}}} & {\color[HTML]{212121} 0.0766} & {\color[HTML]{212121} 0.2145} & {\color[HTML]{212121} 0.3571} & {\color[HTML]{212121} 21.508} \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=0.8\textwidth]{Graphics/Portfolio_returns_30_days_Rebalacing.png}
\caption{Backtesting on EM FIs data: classical approaches
({\color{blue}MinVar},{\color{red}IVP}, {\color{brown}HRP}) vs uplifted approaches ({\color{purple}Inv$\lambda$}, {\color{green}upHRP})}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{figBacktestingEMFIsdataclassicalvsUplifted}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Backtesting_uplifted_approach_with_different_tau.png}
\caption{Backtesting on EM asset data: uplifted approach ({\color{purple}Inv$\lambda$ 5\%}, {\color{green}upHRP 5\%}, {\color{cyan}Inv$\lambda$ 10\%}, {\color{magenta}upHRP 10\%}, {\color{orange}Inv$\lambda$ 25\%}, {\color{lime}upHRP 25\%}, {\color{blue}Inv$\lambda$ 50\%}, {\color{red}upHRP 50\%})}
\label{BacktestingUpliftedapproachwithdifferenttau}
\end{subfigure}
\begin{subfigure}[b]{0.5\linewidth}
\includegraphics[width=\linewidth]{Graphics/FRM_2020.png}
\caption{Financial Risk Meter (FRM)
\bigskip
\bigskip
\smallskip
}
\label{FinancialRiskMeter2020}
\end{subfigure}
\begin{subfigure}[b]{0.8\linewidth}
\includegraphics[width=\linewidth]{Graphics/lamda_weights_20200630-4.png}
\caption{Penalisation parameter ($\lambda$) vs Optimal weights ({\color{blue}Penalisation parameter ($\lambda$}, {\color{purple}Inv$\lambda$ 5\% weights}, {\color{green}upHRP 5\% weights})}
\label{Penalisationparametervsoptimalweights}
\end{subfigure}
\caption{Impact of Penalisation parameter ($\lambda$) and FRM on optimal weights and portfolio returns}
\hspace*{\fill} \href{https://github.com/QuantLet/FRM-EM-paper}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{figImpactofPenalisationparameterandFRMonoptimalweightsand portfolio returns}
\end{figure}
For the backtesting, we adopt a 30 days rebalancing period. Table \ref{BacktestingEMFIsdata} recapitulates the out-of-sample performance of the studied approaches.\\
For the classical approaches, the mean and the volatility of the HRP are 0.003 and 0.049, respectively. While the MinVar delivers the highest return value (0.004) and the highest volatility (0.081), the IVP provides approximately the same portfolio return (0.003) with higher volatility (0.0712). Therefore, the HRP balances out both return and volatility most efficiently providing the best approach in terms of Sharpe ratio (0.072 compared to 0.052 and 0.056 of the MinVar and IVP portfolios, respectively).\\
For the uplifted approach (with $\tau=5\%$), the return and volatility of the upHRP approach are 0.052 and 0.188, respectively. While the Inv$\lambda$ offers a higher return (0.076), and approximately the same volatility, the Inv$\lambda$ provides a risk-return balance registering the best Sharpe ratio value (0.412 compared to 0.276 of the upHRP portfolio). \\
Besides the sharp ratio, we measured the portfolio diversification effects using the Effective $N$ $Effect(N)$ measurement, proposed by \cite{strongin2000beating} as one of allocation concentration and defined as follow:
\begin{equation}
Effect(N)=\frac{1}{\sum_{j=1}^N \widehat{\omega}_{j,t}^2}
\label{eq49}
\end{equation}
The portfolio has high concentration risk if $Effect(N)$ is close to one.\vspace{\baselineskip}
Comparing the classical and the uplifted approaches(with $\tau=5\%$), it seems that the upHRP improves the Sharpe ratio of the HRP approach (0.072 and 0.276, respectively). Moreover, the HRP has discarded five EMs in favor of one single market, with ($Effect(N)=1.6)$, see Table \ref{WeightsallocationEMs}). Therefore, the HRP’s portfolio is deceitfully diversified, since any distress situation affecting this market will have a greater negative impact on HRP’s portfolio than the upHRP, which allocates only 6.9\% to the Mexican market and providing more diversification across EMs ($Effect(N)=16.8$, see Table \ref{BacktestingEMFIsdata}). \vspace{\baselineskip}
In fact, the main innovation of upHRP strategy is to apply the HRP algorithm based on tail dependence clustering instead of the standard correlation-based clustering. First, this strategy is motivated by the fact that the correlation coefficients can change drastically during financial crises due to contagion effects, and such a crisis can spill over quickly. Consequently, diversifying a portfolio based on correlation clusters may be a failing strategy without attention to tail events. Second, while the correlation matrix illustrates the relationship based on the mean-variance of the distribution, the FRM measures the co-movements in extreme events based on both tail of the distribution. To exemplify, lower tail dependence is associated with the capacity to diversify during crises, which can improve the tail risk management of a given strategy. More precisely, the main idea is to cluster the studied FIs illustrating a high probability to experience extremely negative events contemporaneously. Finally, empirical results show that the proposed uplifted approach has the potential to compete with the classical portfolio optimization approach by providing desirable diversification properties, especially if this hierarchical risk parity strategy is based on the tail dependence coefficient (provided by the FRM adjacency matrix), which is a benefit to tail risk management.\vspace{\baselineskip}
Besides, for the uplifted approach, we take into consideration four tail risk levels (5\%, 10\%, 25\%, 50\%). The results in Table \ref{BacktestingEMFIsdata} indicate that:
The $Effect N$ is an increasing function of the tail risk level, therefore for a higher tail risk level, more diversification is needed to guarantee stable portfolio returns.
Comparing the Inv$\lambda$ strategy with the upHRP, we noticed that the Inv$\lambda$ strategy needs to include a higher number of assets compared to the upHRP (i.e for $\tau=10\%$, $Effect(N)= 21.64$ and 16.24, respectively) to provide the same portfolio performance level ( approximately the same sharp ratio=0.4). On the one hand, this result can be explained by the fact that for the upHRP, only FIs within the same cluster compete for portfolio allocation rather than competing with all the FIs in the portfolio, which avoids the redundancy in the FIs included in the optimal portfolio. On the other hand, the larger number of FIs increases estimated parameters, which also can increase the risk of estimation error and thus biased results. For that purpose, for relatively similar values of sharp ratio (portfolio efficiency), the upHRP is preferred over the Inv$\lambda$.\vspace{\baselineskip}
Figure \ref{figBacktestingEMFIsdataclassicalvsUplifted} illustrates the backtesting on EM’s FIs with 30 days rebalancing period. It plots the portfolio turnover for the first half of 2020 according to the classical strategies (MinVar, IVP, and HRP) and the uplifted strategies (Inv$\lambda$ and upHRP), where the classical approaches show a relative stability in returns compared to the uplifted approaches since, the classical approaches concentrate their weights on one Mexican FI (see Table \ref{WeightsallocationEMs}), which limited their profit (increase in portfolio returns in mid-February and in June) and also limit their loss (end of March ) during the COVID 19 crisis. However, this is a trade-off between historical stability and very high concentration risk. Any idiosyncratic risk impact would lead to substantial portfolio losses.
Figure \ref{BacktestingUpliftedapproachwithdifferenttau} plots the backtesting on EM’s FIs with 30 day rebalancing period for the uplifted strategies, taking into consideration the adjacency matrices with different tail risk levels (5\%, 10\%, 25\%, and 50\%). The Figure indicates that all portfolio returns follow the same trend, despite the difference in portfolio composition, since the biggest 25 FIs that compose the adjacency matrices vary over time and the risk levels.\\
Indeed, by comparing this portfolio return’s trend with the FRM plot (Figure \ref{FinancialRiskMeter2020}) during the same period, we noticed a negative relationship, where the high-risk period plotted by the FRM is translated by a decrease in the portfolio returns (March). Moreover, Figure \ref{Penalisationparametervsoptimalweights} indicates that the uplifted approach underweights the high risky FIs (with high penalization parameter $\lambda$) and overweight the less risky FIs, specially for the Inv$\lambda$, since the upHRP takes into consideration the tail co-movement between the FIs.
\subsection{Bridging optimal portfolio weights and network centrality}
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/network_20200630_005.png}
\caption{$\tau=0.05$}
\label{tau005}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Network_20200630_010.png}
\caption{$\tau=0.1$}
\label{tau01}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Network_20200630_025.png}
\caption{$\tau=0.25$}
\label{tau025}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{Graphics/Network_20200625_050.png}
\caption{$\tau=0.5$}
\label{tau05}
\end{subfigure}
\caption{Network graphs with the size of nodes representing total degree centrality on 20200630}
\hspace*{\fill} \raisebox{-1pt}{\includegraphics[scale=0.008]{Graphics/Quantlets_Logo_Ring.jpeg}}
\label{fig:NETWORKtau}
\end{figure}
Figure \ref{fig:NETWORKtau} plots the FRM Network with the size of nodes representing total degree centrality (indegree and outdegree). The network information is provided by the adjacency matrices estimated at $\tau$ equal to 0.05, 0.10, 0.25, and 0.50, with dates 20200630. The Figure shows that the total degree centrality (both in and out) increase and become more visible with the increasing of tail risk level $\tau$.
By interpreting the adjacency matrix of the network, an overlapping region between the portfolio theory and network theory can be established.\\
Indeed, according to Figure \ref{tau005}, Itau Unibanco (ITUB4 BS) is a highly central FI followed by Banco Santander Brasil (SANB11 BS) and Banco Bradesco SA (BBDC4 BS). Therefore, the Brazilian market illustrates high-CoStress compared to the other markets (see also the boxplot, Figure \ref{FRMEMboxplot}. Contrary to that, QNB Finansbank (QNBF TI) and Housing Development Finance Corp (HDFC IS) are low-central FIs.\\
Regarding ELEKTRA MF, the networks (Figure \ref{fig:NETWORKtau}) indicate that this FI is only a risk emitter (high out-degrees with zero in-degree), which makes this FI insensitive to the other FIs shocks. Moreover, the classical portfolio approaches overweight ELEKTRA MF (see Table \ref{WeightallocationsofFIsClassicalvsUplifted}). Therefore, at least historically, ELEKTRA MF had more stable return patterns, but as a risk emitter, any change in risk perception concerning this FI would rapidly spill over into the entire market.\\
Moreover, referring to Table \ref{Optimalweightsfordifftau}, the results of the optimal weight indicate that the Inv$\lambda$ portfolio is insensitive to the centrality degree. It often allocates similar weights for the high and low central FIs (i.e. BBDC4 BS and GFNORTEO MF, see Figure (\ref{tau05}). This result is due to the fact that this portfolio approach allocates weights based only on the individual tail risk level $\Lambda$, varying by tail risk levels $\tau$ but is insensitive to the network centrality degree.\\
Referring to Table \ref{Optimalweightsfordifftau} and Figure \ref{fig:NETWORKtau}, results from different $\tau$ levels indicate that the upHRP approach underweights the high central FIs (i.e. BBDC4 BS (for $\tau$=5\% and $\tau$=50\%), ICICIBC IS (for $\tau$=10\%), ITUB BS and HDFC IS (for $\tau$=25\%)). Contrary to that, the upHRP strategy overweights the low central FIs (i.e. ELEKTRA MF and SLM SJ (for $\tau$=5\%), ELEKTRA MF (for $\tau$=10\%), GFNORTEO MF and B3SA3 BS (for $\tau$=25\% and $\tau$=50\%). These findings are in line with previous research, studying the relationship between the optimal portfolio weights and network centrality in the mean-variance level. In fact, our findings are consistent with \cite{peralta2016network} and \cite{pozzi2013spread} in establishing that "optimal portfolio strategies should overweigh low-central assets and underweight high-central ones". Nonetheless, they contradict the findings of \cite{vyrost2019network}, who argue that “asset weights must be ordered in the same way as the reciprocal of asset centrality in a given network”. \\
Indeed, the sensibility of the Adjacency matrix to different tail risk levels $\tau$ results in centrality degree variation as well as FIs clustering (see Figures \ref{upHRPEM@Dendrogram} and \ref{figdendrogramofdifftau}), which leads to upHRP optimal weights variation through different $\tau$ for all FIs.\\
In fact, the low concentration of weights for the uplifted strategies (Inv$\Lambda$ and upHRP) as well as the sensitivity of weights to different tail risk levels, don’t allow us to conclude about the best FI that offers greater diversification benefits in EMs (contrary to the classical approach that concentrates weights on ELEKTRA MF). However, uplifted strategies overweight the Brasilian FIs (for different $\tau$). Moreover, by contributing with the highest number of FIs to the basket of the 25 biggest FIs in EMs (see Table \ref{WeightsallocationEMs}), the Brazilian market seems more interesting to international investors, besides the need of diversification advantages provided by the other EMs.
\section{Policy recommendation}
Besides the outlined portfolio construction in Section \ref{portfolioconstruction}, we want to derive some recommendations for both policy makers as well as investors. \\
In Section \ref{FRMEMinterp} we have shown that particular EM FIs are not only influenced by FIs and macroeconomic risk variables from the same region, but also have tail-event co-movement dependencies to other EM geographic regions and currencies, the latter all the more important at lower levels of $\tau=0.10$ or $\tau=0.05$. Secondly, a dominant driver of EM FIs returns is the fiscal and economical stance of EM Sovereign bond issuers. If the risk perception for sovereign issuers increases, the return of EM FIs is impacted across the board and across the multivariate distribution. Lastly, co-movements can be detected within similar economic sub-sectors across geographical regions, which is also apparent in the portfolio construction exercises in Section \ref{portfolioconstruction}. We recommend investors to analyse concentrations across economic sub-sectors across regions, as well as to analyse dependencies between equity investments and bond investments in EM, given the clear linkages between the sovereign and the domestic EM FIs. Lastly, currency fluctuations have marginal return contributions in tail-events especially. In so far as EM FI investments are unhedged into developed market currencies, there is risk of a compound effect on returns. \\
As for the portfolio construction, the "classical" approaches show a relative stability in portfolio returns compared to the uplifted approaches, which limited their profit and also limit their loss during the COVID 19 crisis. But, this often comes at the cost of high concentration risks, exposing the portfolio to idiosyncratic tail events. The uplifted portfolio approaches show more volatile turnover with higher loss and profit, but are much better diversified, preventing sizeable risk clusters. Even though a risk averse investor might opt for the classical approach at first, closer inspection of tail risk behaviour and concentration risk should let the investor prefer the uplifted portfolio.\\
EM policy makers can derive important recommendations from our analysis. Firstly, coordination between EM regulatory bodies is of importance in order to mute EM FI fluctuations. Here, particular attention should be put on linkages in same sub-sector operating FIs across regions, which will be increasingly more important as globalisation continues. Secondly, fluctuations in risk perception of the sovereign issuer has an immediate impact on EM FI returns. Regulatory bodies are therefore advised to preemptively verify sound capitalisation of their domestic banks, even if the sovereign issuers distress is stemming from another geographic region. This is further amplified by a tail-event leading to a weakening of an EM's currency versus for example the U.S. Dollar. Global financial linkages, we show, between EM FIs lead to spill-over effects. Overall, in order to protect the domestic as well as global EM economies and their FIs, EM regulatory bodies should continuously work towards closer coordination between Emerging Market economies. This will help increase robustness versus developed market economies distress, of which more is likely to come, as developed markets fight with difficult fiscal situations and low growth patterns. But also, closer coordination between countries will prevent spill-over effects from one geographical region's FIs onto others, thereby increasing the attractiveness of the asset class further. Our approach via dendograms rapidly indicates any such risk clusters, and can be updated at ease and frequently during financial crises.
\section{Conclusion}
In this study, we examine the co-movements of EM FIs across six geographical regions, with aim to analyse within EM country co-dependencies in tail events but importantly also between regions as well as FIs of the same sub-sector across regions. We also analyse the important macro-economical risk variables impacting EM FIs and conclude that in addition to the developed market variables suggested by \cite{tobias2016covar}, EM specific macroeconomic risk variables have significant explanatory power. This can be used to construct more robust total asset class portfolio allocation and supplies EM regulatory bodies with detailed information on co-dependencies for better and faster stabilisation measures during periods of distress.\\
We also propose a novel asset allocation method – Hierarchical Risk Parity based the tail event information from the FRM technology, allowing us to extend this approach to the quantile level and replace the covariance matrix with the rich information contained in the FRM adjacency matrix. We applied this proposed approach to a portfolio of the biggest 25 FIs in EM, and our results show that uplifted strategies provide appropriate diversification properties. In comparison, the Inv$\lambda$ portfolios tend to be too static and the classical approaches result in too concentrated portfolios. Bridging optimal portfolio weights and network centrality, we conclude that the Inv$\lambda$ insensitive to the network centrality degree. However the upHRP portfolio underweight high-central FIs and overweight low-central ones, therefore the Inv$\lambda$ is less at risk of spill-over effects across EM regions, FIs, and financial sub-sectors.
|
\section{Introduction}
Set theory is a well-established vehicle for formal modeling, specification,
analysis and verification of software systems. Formal notations such as B
\cite{Abrial00} and Z \cite{Spivey00} and tools such as ProB \cite{Leuschel00},
Atelier-B \cite{atelierb} and Z/EVES \cite{Saaltink01} are good examples of
that claim. Hence, it is important to extend the capabilities of existing tools
and develop new ones for set theory as applied in the context of verification.
Besides, when these methods and tools are used for formal verification and
analysis, it is necessary to discharge a number of verification conditions or
proof obligations. Then, tools capable of automating such proofs are essential
to render the development process cost-effective. Decision procedures play a
key role in proof automation. Indeed, if a decision procedure exists for a
fragment of set theory, then it would be possible to automate the proofs of
verification conditions lying in this fragment.
$\{log\}$\xspace (read `setlog') \cite{DBLP:journals/jlp/DovierOPR96,setlog} is a Constraint Logic
Programming (CLP) language and satisfiability solver implemented in Prolog
providing: \emph{i)} a decision procedure for the algebra of \emph{hereditarily
finite sets}, i.e., finitely nested sets that are finite at each level of
nesting \cite{Dovier00}; \emph{ii)} a decision
procedure for a very expressive fragment of the class of finite set relation
algebras (\citeANP{DBLP:journals/jar/CristiaR20} \citeyearNP{DBLP:journals/jar/CristiaR20,DBLP:conf/RelMiCS/CristiaR18}); and
\emph{iii)} a decision procedure for restricted intensional sets (RIS) (\citeANP{DBLP:conf/cade/CristiaR17} \citeyearNP{DBLP:journals/jar/CristiaR21a,DBLP:conf/cade/CristiaR17}). Several
in-depth empirical evaluations provide evidence that $\{log\}$\xspace is able to solve
non-trivial problems
(\citeANP{DBLP:journals/jar/CristiaR20} \citeyearNP{DBLP:journals/jar/CristiaR21a,DBLP:journals/jar/CristiaR20,DBLP:conf/RelMiCS/CristiaR18,DBLP:conf/cade/CristiaR17,CristiaRossiSEFM13}),
in particular as an automated verifier of security properties
(\citeANP{DBLP:journals/jar/CristiaR21} \citeyearNP{DBLP:journals/jar/CristiaR21,Cristia2021}). All of these decision
procedures are based on the notion of \emph{set unification} \cite{Dovier2006}.
In this paper we add to $\{log\}$\xspace a decision procedure for the algebra of finite
sets extended with cardinality constraints. This extension is important in
terms of formal software verification because there are situations where we
need to reason about the size of a given data structure and not only about what
its elements are. For example, within the algebra of finite sets one can
partition a given set into two disjoint subsets, $C = A \cup B \land A \cap B =
\emptyset$, but there is no way to state that $A$ and $B$ must be of the same
cardinality. In practice these constraints might appear, for instance, when
part of a data container must be put into a cache---a simple $\{log\}$\xspace program is
shown in \ref{app:datacontainer}. Specifically, cardinality constraints appear
in the verification of some distributed algorithms
\cite{DBLP:conf/cav/BerkovitsLLPS19,Alberti2017} and are at the base of the
notions of integer interval, array and list.
At an abstract level, the new decision procedure combines the decision
procedure for the algebra of finite sets already existing in $\{log\}$\xspace with a
decision procedure for sets with cardinality constraints proposed by
\citeN{DBLP:conf/frocos/Zarba02}. Zarba proves that a theory of finite sets
equipped with the classic set theoretic operators, including cardinality,
combined with linear integer constraints is decidable. In his work, Zarba is
interested in proving a decidability result; as far as we know Zarba's
algorithm has never been implemented before. In fact, the new decision
procedure first uses all the power of $\{log\}$\xspace to produce a simplified,
equivalent formula that can be passed to Zarba's algorithm which makes a final
judgment about its satisfiability, in case it contains cardinality constraints.
In this way, $\{log\}$\xspace performs as well as before on the class of formulas it was able to
deal with previously.
As a consequence of the fact that the new decision procedure is still based on
set unification, it can deal with sets of sets nested at any depth. For
example, the decision procedure is able to give all possible
solutions for a goal such as $\card{\{\{x\},\{y,z\}\}} = n$, where $x$, $y$,
$z$ and $n$ are variables.
Zarba's algorithm is implemented by integrating the Prolog Boolean SAT solver
developed by \citeN{DBLP:journals/tcs/HoweK12} with SWI-Prolog's implementation
of the CLP(Q) system \cite{holzbaur1995ofai}. As a result the implementation
integrates three Prolog-based systems: Howe and King's SAT solver, CLP(Q) and
$\{log\}$\xspace.
Solving formulas over a theory of sets and cardinality is not new
\cite{DBLP:conf/cade/FerroOS80,Gervet01}. However, our proposal clearly
distinguishes itself from all previous works in some aspects that constitute
the main contributions of this paper: \emph{a)} our implementation is deeply
rooted in the CLP framework and thus inherits all its properties; in
particular, $\{log\}$\xspace preserves its features as a CLP language and as a
satisfiability solver; \emph{b}) our CLP system produces a finite
representation of all possible solutions of any satisfiable formula of its
input language; \emph{c}) as the decision procedure is based on set unification
it handles set elements of any kind including nested sets; and \emph{d)} this
is the first implementation of Zarba's algorithm and it is shown to perform
better than some other systems.
\paragraph{Structure of the paper.}
Section \ref{language} presents the syntax and semantics of the constraint
language for finite sets with cardinality constraints. The overall structure of
the constraint solver for that language is introduced in Section \ref{satcard}.
The main routine dealing with cardinality constraints is presented in Section
\ref{sizesolver}, where we also include a description of Zarba's algorithm. In
Section \ref{decproc} we prove that the resulting solver is indeed a decision
procedure for our language. Besides deciding the satisfiability of cardinality
formulas, the solver is able to find a particular form of their solutions, as
we explain in Section \ref{minimal}. Section \ref{impl} shows how $\{log\}$\xspace works
with cardinality constraints, in particular in the context of formal
verification (Section \ref{formver}); an empirical evaluation is also reported
(Section \ref{empirical}). We compare our approach with others in Section
\ref{related}. Some concluding remarks are provided in Section \ref{concl}.
\section{\label{language}$\mathit{L}_{\CARD}$: a language for finite sets and cardinality}
In this section we describe the syntax and semantics of our set-based language
$\mathit{L}_{\CARD}$ (read `l-card'). This is a quantifier-free first-order predicate
language with three distinct sorts: the sort $\mathsf{Set}$ of all terms denoting
sets, the sort $\mathsf{Int}$ of terms denoting integer numbers, and the sort $\mathsf{Ur}$
of all other terms. Terms of each sort are allowed to enter in the formation of
set terms (in this sense, the designated sets are hybrid), no nesting
restrictions being enforced (in particular, membership chains of any finite
length can be modeled). A handful of reserved predicate symbols endowed with a
pre-designated set-theoretic meaning is available. The usual linear integer
arithmetic operators are available as well. Formulas are built in the usual way
by using conjunction and disjunction. A few more complex operators (in the form
of predicates) are defined as $\mathit{L}_{\CARD}$ formulas, thus making it simpler for the
user to write complex formulas.
\subsection{Syntax}\label{syntax}
The syntax of the language is defined primarily by giving the signature upon
which terms and formulas are built.
\begin{definition}[Signature]\label{signature}
The signature $\Sigma_{\card{\cdot}}$ of $\mathit{L}_{\CARD}$ is a triple $\langle
\mathcal{F},\Pi,\mathcal{V}\rangle$ where:
\begin{itemize}
\item $\mathcal{F}$ is the set of constants and function symbols along with their sorts, partitioned as
$\mathcal{F} \defs \mathcal{F}_\Set \uplus \mathcal{F}_\Int \uplus \mathcal{F}_\Ur$, where
$\mathcal{F}_\Set \defs \{\emptyset,\plus\}$, $\mathcal{F}_\Int = \{0,-1,1,-2,2,\dots\} \cup \{+,-,*\}$
and $\mathcal{F}_\Ur$ is a set of uninterpreted constant and function symbols.
\item $\Pi$ is the set of predicate symbols along with their sorts, partitioned as
$\Pi \defs \Pi_{=} \cup \Pi_\mathsf{S} \cup \Pi_{size} \cup \Pi_\Int$, where $\Pi_{=}
\defs \{=,\neq\}$, $\Pi_\mathsf{S} \defs \{\in,\notin,\Cup,\parallel\}$, $\Pi_{size}
\defs \{size\}$, and $\Pi_\Int \defs \{\leq\}$.
\item $\mathcal{V}$ is a denumerable set of variables partitioned as
$\mathcal{V} \defs \mathcal{V}_\mathsf{S} \cup \mathcal{V}_\mathsf{Z} \cup \mathcal{V}_\mathsf{U}$.
\hfill$\Box$
\end{itemize}
\end{definition}
Intuitively, $\emptyset$ represents the empty set; $\{x \plus A\}$ represents the
set\footnote{$\plus$ is akin to Prolog's list constructor
`$\mid$'.}\textsuperscript{-}\footnote{In $\{log\}$\xspace, $\emptyset$ is written as \{\}
and $\plus$ as /, see Section \ref{impl}.} $\{x\} \cup A$; and $\mathcal{V}_\mathsf{S}$,
$\mathcal{V}_\mathsf{Z}$ and $\mathcal{V}_\mathsf{U}$ represent sets of variables ranging over sets,
integers and ur-elements\footnote{Ur-elements (also known as atoms or
individuals) are objects which have no elements but are distinct from the empty
set.}, respectively.
Sorts of function and predicate symbols are specified as follows: if $f$
(resp., $\pi$) is a function (resp., a predicate) symbol of arity $n$,
then its sort is an $n+1$-tuple $\langle s_1, \ldots ,s_{n+1} \rangle$ (resp.,
an $n$-tuple $\langle s_1, \ldots ,s_n \rangle$) of non-empty subsets of the
set $\{ \mathsf{Set} , \mathsf{Int}, \mathsf{Ur} \}$ of sorts. This notion is denoted by $f:\langle
s_1, \ldots ,s_{n+1}\rangle$ (resp., by $\pi:\langle s_1, \ldots ,s_n\rangle
$). Specifically, the sorts of the elements of $\mathcal{F}$ and $\mathcal{V}$ are the
following.
\begin{definition}[Sorts of function symbols and variables]\label{d:sorts}
The sorts of the symbols in $\mathcal{F}$ are as follows:
\begin{flalign*}
\quad\quad & \emptyset: \langle \{\mathsf{Set} \} \rangle & \\
& \mathsf{\{\cdot \plus \cdot\}: \langle \{\mathsf{Set} , \mathsf{Int}, \mathsf{Ur}\}, \{ \mathsf{Set} \} , \{ \mathsf{Set}\}\rangle } & \\
& c: \langle \{\mathsf{Int} \} \rangle \text{, for any $c \in \{0,-1,1,-2,2,\dots\}$} & \\
& \cdot + \cdot, \cdot - \cdot, \cdot * \cdot:
\langle \{\mathsf{Int}\}, \{\mathsf{Int}\} , \{\mathsf{Int}\}\rangle & \\
& f: \langle \underbrace{\{\mathsf{Set},\mathsf{Int},\mathsf{Ur}\}, \ldots,
\{\mathsf{Set},\mathsf{Int},\mathsf{Ur}\}}_n, \{{\sf
\mathsf{Ur}}\}\rangle\text{, if $f \in \mathcal{F}_\Ur$ is of arity $n \ge 0$}. &
\end{flalign*}
The sorts of variables are as follows:
\begin{flalign*}
\quad\quad & v: \langle \{\mathsf{Set} \} \rangle \text{, if $v \in \mathcal{V}_\mathsf{S}$} & \\
& v: \langle \{\mathsf{Int} \} \rangle \text{, if $v \in \mathcal{V}_\mathsf{Z}$} & \\
& v: \langle \{\mathsf{Ur} \} \rangle \text{, if $v \in \mathcal{V}_\mathsf{U}$} & \tag*{\hfill$\Box$}
\end{flalign*}
\end{definition}
\begin{definition}[Sorts of predicate symbols]\label{d:sorts_pred}
The sorts of the predicate symbols in $\Pi$ are as follows
(symbols $\Cup$ and $size$ are prefix; all other symbols in $\Pi$ are infix):
\begin{flalign*}
\quad\quad & =,\neq: \langle \{\mathsf{Set} , \mathsf{Int}, \mathsf{Ur} \}, \{ \mathsf{Set} , \mathsf{Int}, \mathsf{Ur} \} \rangle & \\
& \in,\notin: \langle \{\mathsf{Set}, \mathsf{Int}, \mathsf{Ur} \} , \{\mathsf{Set} \} \rangle & \\
& \Cup: \langle \{\mathsf{Set} \} , \{\mathsf{Set} \}, \{\mathsf{Set} \} \rangle & \\
& \parallel: \langle \{\mathsf{Set} \} , \{\mathsf{Set} \} \rangle & \\
& size: \langle \{\mathsf{Set} \} , \{\mathsf{Int} \} \rangle & \\
& \leq: \langle \{\mathsf{Int} \} , \{\mathsf{Int} \} \rangle & \tag*{\hfill$\Box$}
\end{flalign*}
\end{definition}
Note that arguments of $=$ and $\neq$ can be of any of the three considered
sorts. We do not have distinct symbols for different sorts, but the
interpretation of $=$ and $\neq$ (see Section \ref{semantics}) depends on the
sorts of their arguments.
The set of admissible (i.e., well-sorted) $\mathit{L}_{\CARD}$ terms is defined as follows.
\begin{definition}[$\card{\cdot}$-terms]\label{LCARD-terms}
The set of \emph{$\card{\cdot}$-terms}, denoted by $\mathcal{T}_{\CARD}$, is the minimal subset of
the set of $\Sigma_{\card{\cdot}}$-terms generated by the following grammar complying
with the sorts as given in Definition \ref{d:sorts}:
\begin{flalign*}
\quad\quad C &::= 0 \hspace{2pt}|\hspace{2pt}
{-1} \hspace{2pt}|\hspace{2pt}
1 \hspace{2pt}|\hspace{2pt}
{-2} \hspace{2pt}|\hspace{2pt}
2 \hspace{2pt}|\hspace{2pt}
\dots & \\
\mathcal{T}_\Int &::= C \hspace{2pt}|\hspace{2pt}
\mathcal{V}_\mathsf{Z} \hspace{2pt}|\hspace{2pt}
C * \mathcal{V}_\mathsf{Z} \hspace{2pt}|\hspace{2pt}
\mathcal{V}_\mathsf{Z} * C \hspace{2pt}|\hspace{2pt}
\mathcal{T}_\Int + \mathcal{T}_\Int \hspace{2pt}|\hspace{2pt}
\mathcal{T}_\Int - \mathcal{T}_\Int & \\
\mathcal{T}_{\CARD} & ::=
\mathcal{T}_\Int \hspace{2pt}|\hspace{2pt}
\mathcal{T}_\Ur \hspace{2pt}|\hspace{2pt} \mathcal{V}_\mathsf{U} \hspace{2pt}|\hspace{2pt}
\mathit{Set} & \\
\mathit{Set} & ::=
\text{\normalfont\'{}}\emptyset\text{\normalfont\'{}}\hspace{5pt}
\hspace{2pt}|\hspace{2pt}
\mathcal{V}_S
\hspace{2pt}|\hspace{2pt}
\text{\normalfont\'{}}\{\text{\normalfont\'{}}\hspace{5pt} \mathcal{T}_{\CARD}
\hspace{5pt}\text{\normalfont\'{}}\hspace{-2pt}\plus\hspace{-2pt}\text{\normalfont\'{}}\hspace{5pt} \mathit{Set} \hspace{5pt}\text{\normalfont\'{}}\}\text{\normalfont\'{}} &
\end{flalign*}
where $\mathcal{T}_\Int$ (resp., $\mathcal{T}_\Ur$) represents any non-variable $\mathcal{F}_\Int$-term
(resp., $\mathcal{F}_\Ur$-term). \hfill$\Box$
\end{definition}
As can be seen, through rules $C$ and $\mathcal{T}_\Int$, the grammar allows only integer
linear terms.
If $t$ is a term $f(t_1,\dots,t_n)$, $f \in \mathcal{F}, n \ge 0$, and $\langle
s_1, \ldots ,s_{n+1} \rangle$ is the sort of $f$, then we say that $t$ is of
sort $\langle s_{n+1} \rangle$. The sort of any $\card{\cdot}$-term $t$ is always
$\langle \{\mathsf{Set}\}\rangle$ or $\langle \{\mathsf{Int}\} \rangle$ or $\langle \{\mathsf{Ur}\}
\rangle$. For the sake of simplicity, we simply say that $t$ is of sort $\mathsf{Set}$
or $\mathsf{Int}$ or $\mathsf{Ur}$, respectively. In particular, we say that a $\card{\cdot}$-term
of sort $\mathsf{Set}$ is a {\em set term}, and that set terms of the form $\{t_1
\plus t_2\}$ are \emph{extensional} set terms. The first parameter of an
extensional set term is called \emph{element part} and the second is called
\emph{set part}. Observe that one can write terms representing sets which are
nested at any level.
Hereafter, we will use the following notation for extensional set terms:
$\{t_1,t_2,\dots,t_n \plus t\}$, $n \ge 1$, is a shorthand for $\{t_1
\plus \{t_2 \,\plus\, \cdots \{ t_n \plus t\}\cdots\}\}$, while
$\{t_1,t_2,\dots,t_n\}$ is a shorthand for $\{t_1,t_2,\dots,t_n \plus \emptyset\}$.
Moreover, we will use the following naming conventions: $A, B, C, D$
stand for terms of sort $\mathsf{Set}$; $i, j, k, m$
stand for terms of sort $\mathsf{Int}$; $a, b, c, d$ stand for terms of sort $\mathsf{Ur}$; and $x, y, z$
stand for terms of any of the three sorts.
\begin{example}[Set terms]
The following $\Sigma_{\card{\cdot}}$-terms are set terms:
\begin{flalign*}
\quad\quad & \emptyset &\\
& \{x \plus A\} & \\
& \{4+k,f(a,b)\}, \text{ i.e., } \{4+k \plus \{f(a,b) \plus \emptyset \}\},
\text{ where $f$ is a (uninterpreted) symbol in $\mathcal{F}_\Ur$.} &
\end{flalign*}
On the opposite, $\{x \plus 17\}$ is not a set term. \hfill$\Box$
\end{example}
The sets of well-sorted $\mathit{L}_{\CARD}$ constraints and
formulas are defined as follows.
\begin{definition}[$\card{\cdot}$-constraints]\label{primitive-constraint}
If $\pi \in \Pi$ is a predicate symbol of sort $\langle s_1, \ldots , s_n
\rangle$, and for each $i=1,\ldots , n$, $t_i$ is a $\card{\cdot}$-term of sort
$\langle s'_i \rangle$ with $s'_i \subseteq s_i$, then $\pi (t_1,\ldots ,t_n)$
is a \emph{$\card{\cdot}$-constraint}. The set of $\card{\cdot}$-constraints is denoted by
$\mathcal{C}_{\CARD}$. \hfill$\Box$
\end{definition}
$\card{\cdot}$-constraints whose arguments are of sort $\mathsf{Set}$ (including $size$
constraints) will be called \emph{set constraints}; $\card{\cdot}$-constraints whose
arguments are of sort $\mathsf{Int}$ will be called \emph{integer constraints}.
\newcommand{\Phi_{\CARD}}{\Phi_{\card{\cdot}}}
\begin{definition}[$\card{\cdot}$-formulas]\label{formula}
The set of $\card{\cdot}$-formulas, denoted by $\Phi_{\CARD}$, is given
by the following grammar:
\begin{flalign*}
\quad\quad & \Phi_{\CARD} ::=
\mathit{true} \mid false \mid \mathcal{C}_{\CARD} \mid \Phi_{\CARD} \land \Phi_{\CARD} \mid \Phi_{\CARD} \lor
\Phi_{\CARD} &
\end{flalign*}
where $\mathcal{C}_{\CARD}$ represents any element belonging to the set of
$\card{\cdot}$-constraints. \hfill$\Box$
\end{definition}
\begin{example}[$\card{\cdot}$-formulas]\label{ex:formulas}
The following are $\card{\cdot}$-formulas:
\begin{flalign*}
\quad\quad & a \in A \land a \notin B \land \Cup(A,B,C) \land C = \{x \plus D\} & \\
& \Cup(A,B,C) \land n+k > 5 \land size(C,n) \land B \neq \emptyset & \\
& x \in A \land B \in A \land size(A,x) \land size(B,y) \land x<y &
\end{flalign*}
On the contrary, $\Cup(A,B,23)$ is not a $\card{\cdot}$-formula because $\Cup(A,B,23)$
is not a $\card{\cdot}$-constraint ($23$ is not of sort $\mathsf{Set}$ as required by the
sort of $\Cup$). \hfill$\Box$
\end{example}
As we will show in Section \ref{expressiveness}, the language does not need a primitive negation connective, thanks to the presence of negative constraints.
\subsection{\label{semantics}Semantics}
Sorts and symbols in $\Sigma_{\card{\cdot}}$ are interpreted according to the
interpretation structure $\mathcal{R} \defs \langle D,\iF{\cdot}\rangle$, where $D$ and
$\iF{\cdot}$ are defined as follows.
\begin{definition} [Interpretation domain] \label{def:int_dom}
The interpretation domain $D$ is partitioned as $D \defs D_\mathsf{Set} \cup D_\mathsf{Int} \cup D_\mathsf{Ur}$
where:
\begin{itemize}
\item $D_\mathsf{Set}$ is the set of all hereditarily finite hybrid
sets built from elements in $D$. Hereditarily finite sets are those sets that
admit (hereditarily finite) sets as their elements, that is sets of sets.
\item $D_\mathsf{Int}$ is the set of integer numbers, $\mathbb{Z}$.
\item $D_\mathsf{Ur}$ is a collection of other objects. \hfill$\Box$
\end{itemize}
\end{definition}
\begin{definition} [Interpretation function] \label{app:def:int_funct}
The interpretation function $\iF{\cdot}$ is defined as follows:
\begin{itemize}
\item Each sort $\mathsf{X} \in \{\mathsf{Set},\mathsf{Int},\mathsf{Ur}\}$ is mapped to
the domain $D_\mathsf{X}$.
\item For each sort $\mathsf{X}$, each variable $x$ of sort $\mathsf{X}$ is mapped to
an element $x^\mathcal{R}$ in $D_\mathsf{X}$.
\item The constant and function symbols in $\mathcal{F}_\mathsf{S}$ are
interpreted as follows:
\begin{itemize}
\item $\emptyset$ is interpreted as the empty set, namely $\emptyset^\mathcal{R} = \emptyset$
\item $\{ x \plus A \}$ is interpreted as the set $\{x^\mathcal{R}\} \cup A^\mathcal{R}$.
\end{itemize}
\item The constant and function symbols in $\mathcal{F}_\mathsf{Z}$ are
interpreted as follows:
\begin{itemize}
\item Each element of \{0,-1,1,-2,2,\dots\} is interpreted as the corresponding integer number
\item $i + j$ is interpreted as $i^\mathcal{R} + j^\mathcal{R}$
\item $i - j$ is interpreted as $i^\mathcal{R} - j^\mathcal{R}$
\item $i * j$ is interpreted as $i^\mathcal{R} * j^\mathcal{R}$
\end{itemize}
\item The predicate symbols in $\Pi$ are interpreted as follows:
\begin{itemize}
\item $x = y$, where $x$ and $y$ have the same sort $\mathsf{X}$, is interpreted as the
identity between $x^\mathcal{R}$ and $y^\mathcal{R}$ in $D_\mathsf{X}$;
otherwise, $x = y$ is interpreted as being $false$
\item $x \in A$ is interpreted as $x^\mathcal{R} \in A^\mathcal{R}$
\item $\Cup(A,B,C)$ is interpreted as $C^\mathcal{R} = A^\mathcal{R} \cup B^\mathcal{R}$
\item $A \parallel B$ is interpreted as $A^\mathcal{R} \cap B^\mathcal{R} = \emptyset$
\item $size(A,k)$ is interpreted as $\card{A^\mathcal{R}} = k^\mathcal{R}$
\item $i \leq j$ is interpreted as $i^\mathcal{R} \leq j^\mathcal{R}$
\item $x \neq y$ and $x \notin A$ are
interpreted as $\lnot x = y$ and $\lnot x \in A$, respectively.
\hfill$\Box$
\end{itemize}
\end{itemize}
\end{definition}
It is worth noting that $size(A,k)$ is interpreted as the
\emph{cardinality}, i.e., the number of elements, of the set denoted by $A$,
and it is not to be confused with the term size, i.e., the number of function
symbols appearing in the term $A$.
The interpretation structure $\mathcal{R}$ is used to evaluate each $\card{\cdot}$-formula
$\Phi$ into a truth value $\Phi^\mathcal{R} = \{\mathit{true},false\}$ in the following way:
set constraints (resp., integer constraints) are evaluated by $\iF{\cdot}$
according to the meaning of the corresponding predicates in set theory (resp.,
in number theory) as defined above; $\card{\cdot}$-formulas are evaluated by
$\iF{\cdot}$ according to the rules of propositional logic. A $\mathit{L}_{\CARD}$-formula
$\Phi$ is \emph{satisfiable} iff there exists an assignment $\sigma$ of values
from ${\cal D}$ to the variables of $\Phi$, respecting the sorts of the
variables, such that $\Phi[\sigma]$ is true in $\mathcal{R}$, i.e., $\mathcal{R} \models
\Phi[\sigma]$. In this case, we say that $\sigma$ is a \emph{successful
valuation} (or, simply, a \emph{solution}) of $\Phi$.
In particular, observe that equality between two set terms is interpreted as
the equality in $D_\mathsf{Set}$; that is, as set equality between hereditarily finite
hybrid sets. Such equality is regulated by the standard \emph{extensionality
axiom}, which has been proved to be equivalent, for hereditarily finite sets,
to the following equational axioms \cite{Dovier00}:
\begin{gather}
\{x, x \plus A\} = \{x \plus A\} \tag{$Ab$} \label{Ab} \\
\{x, y \plus A\} = \{y, x \plus A\} \tag{$C\ell$} \label{Cl}
\end{gather}
Axiom \eqref{Ab} states that duplicates in a set term do not matter
(\emph{Absorption property}). Axiom \eqref{Cl} states that the order of
elements in a set term is irrelevant (\emph{Commutativity on the left}). These
two properties capture the intuitive idea that, for instance, the set terms
$\{1,2\}$, $\{2,1\}$, and $\{1,2,1\}$ all denote the same set.
\subsection{\label{expressiveness}{Derived Constraints}}
$\mathit{L}_{\CARD}$ can be extended to support other set and integer operators definable
by means of suitable $\mathit{L}_{\CARD}$ formulas.
\citeN{Dovier00} proved that the collection of predicate symbols
in $\Pi_{=} \cup \Pi_\mathsf{S}$ is
sufficient to define constraints implementing the set operators $\cap$,
$\subseteq$ and $\setminus$. For example, $A \subseteq B$ can be defined by the
$\mathit{L}_{\CARD}$ formula $\Cup(A,B,B)$. Likewise, $\{=,\neq\} \cup \Pi_\Int$
is sufficient to define $<$, $>$ and $\geq$. With a slight abuse of
terminology, we say that the set and integer predicates that are specified by
$\card{\cdot}$-formulas are \emph{derived constraints}.
Whenever a formula contains a derived constraint, the constraint is replaced by
its definition turning the given formula into an $\mathit{L}_{\CARD}$ formula. Precisely, if
formula $\Phi$ is the definition of constraint $c$, then $c$ is replaced by
$\Phi$ and the solver checks satisfiability of $\Phi$ to determine
satisfiability of $c$. Thus, we can completely ignore the presence of derived
constraints in the subsequent discussion about constraint solving and formal
properties of our solver.
The negated versions of set and integer operators can be introduced as derived
constraints, as well. The derived constraint for $\lnot\cup$ and
$\lnot\parallel$ (called $nun$ and $\not\disj$, respectively) are shown in
\cite{Dovier00}.
For example, $\lnot(A \cup B = C)$ is introduced as:
\begin{equation}\label{e:nun}
nun(A,B,C) \defs
(n \in C \land n \notin A \land n \notin B)
\lor (n \in A \land n \notin C)
\lor (n \in B \land n \notin C)
\end{equation}
With a
little abuse of terminology, we will refer to these predicates as
\emph{negative constraints}.
Thanks to the availability of negative constraints, (general) logical negation
is not strictly necessary in $\mathit{L}_{\CARD}$.
Now that we have derived and negative constraints it is easy to see that
$\mathit{L}_{\CARD}$ expresses the Boolean algebra of sets with cardinality.
\begin{remark}[CLP($\mathit{SET}$)\xspace]
$\{log\}$\xspace
provides an implementation of the CLP instance CLP($\mathit{SET}$)\xspace \cite{Dovier00}. In
turn, CLP($\mathit{SET}$)\xspace is based on a constraint language including
$\mathcal{F}_\mathsf{S}$ and $\Pi_\mathsf{S}$, with the same sorts; formulas in CLP($\mathit{SET}$)\xspace are
built as in $\mathit{L}_{\CARD}$. Hence, $\mathit{L}_{\CARD}$ effectively extends CLP($\mathit{SET}$)\xspace by
introducing $size$ constraints and integer arithmetic. An $\mathit{L}_{\CARD}$ formula not
including $size$ constraints nor integer constraints is a CLP($\mathit{SET}$)\xspace formula.
Hereafter, we will simply use the name CLP($\mathit{SET}$)\xspace to refer to the
constraint language offered by $\{log\}$\xspace.
\hfill$\Box$
\end{remark}
\section{\label{satcard}$\mathit{SAT}_{\CARD}$: a constraint solving procedure for $\mathit{L}_{\CARD}$}
A complete solver for CLP($\mathit{SET}$)\xspace is proposed in \cite{Dovier00}. In this section,
we show how that solver can be combined with Zarba's decision procedure
\cite{DBLP:conf/frocos/Zarba02}---hereafter simply called $\mathit{SAT}_{Za}$---to support
cardinality constraints. The resulting constraint solving procedure, called
$\mathit{SAT}_{\CARD}$ (read `sat-card'), is a decision procedure for $\mathit{L}_{\CARD}$ formulas.
Furthermore, it produces a finite representation of all possible solutions of
any satisifiable $\mathit{L}_{\CARD}$ formula (see Section \ref{decproc}).
\subsection{The solver}
The overall organization of $\mathit{SAT}_{\CARD}$ is shown in Algorithm
\ref{glob}. Basically, $\mathit{SAT}_{\CARD}$ uses four routines:
\textsf{gen\_size\_leq}, $\mathsf{STEP_S}$, \textsf{remove\_neq} and
$\mathsf{solve\_size}$. $\mathsf{solve\_size}$, which is crucial for the integration of cardinality
constraints into CLP($\mathit{SET}$)\xspace, will be presented separately in Section
\ref{sizesolver}.
\begin{algorithm
\begin{algorithmic}[0]
\State $\Phi \gets \textsf{gen\_size\_leq}(\Phi)$;
\Repeat
\State $\Phi' \gets \Phi$;
\Repeat
\State $\Phi'' \gets \Phi$;
\State $\Phi \gets \mathsf{STEP_S}(\Phi)$
\hfill{\footnotesize[$\mathsf{STEP_S}$ returns $false$ when $\Phi$ is unsat]}
\Until{$\Phi = \Phi''$}
\State $\Phi \gets \textsf{remove\_neq}(\Phi)$
\Until{$\Phi = \Phi'$}
\hfill{\footnotesize[end of main loop]}
\State \textbf{let} $\Phi$ \textbf{be} $\Phi_1 \land \Phi_2$
\hfill{\footnotesize [$\Phi_1$ contains $size$ relevant constraints, see Section \ref{integrating}]}
\State $\Phi_1 \gets \mathsf{solve\_size}(\Phi_1)$
\hfill{\footnotesize[$\mathsf{solve\_size}$ returns $false$ when $\Phi_1$ is unsat]}
\State\Return{$\Phi_1 \land \Phi_2$}
\hfill{\footnotesize[returns $false$ (unsat); or a disjunction of formulas representing all solutions]}
\end{algorithmic}
\caption{The solver $\mathit{SAT}_{\CARD}$. $\Phi$ is the input formula.} \label{glob}
\end{algorithm}
\textsf{gen\_size\_leq} simply adds integer constraints to the input formula
$\Phi$ to force the second argument of each $size$ constraint in $\Phi$ to be a
non-negative integer. $\mathsf{STEP_S}$ includes the constraint solving
procedure for the CLP($\mathit{SET}$)\xspace fragment as well as the constraint solving procedures
for cardinality constraints (see Section \ref{steps}). $\mathsf{STEP_S}$
applies specialized rewriting procedures to the current formula $\Phi$ and
returns either $false$ or the modified formula. Each rewriting procedure
applies a few non-deterministic rewrite rules which reduce the syntactic
complexity of $\card{\cdot}$-constraints of one kind. \textsf{remove\_neq} deals with
the elimination of $\neq$ constraints involving set variables. Its purpose and
definition is made evident in \ref{inequalities}.
The execution of $\mathsf{STEP_S}$ and \textsf{remove\_neq} is iterated until a
fixpoint is reached, i.e., the formula is irreducible. These routines return
$false$ whenever (at least) one of the involved procedures rewrites $\Phi$ to
$false$. In this case, a fixpoint is immediately detected.
As we will show in Section \ref{decproc}, when all the non-deterministic
computations of $\mathit{SAT}_{\CARD}(\Phi)$ return $false$, then we can conclude that
$\Phi$ is unsatisfiable; otherwise, we can conclude that $\Phi$ is satisfiable
and each solution of the formulas returned by $\mathit{SAT}_{\CARD}$ is a solution of
$\Phi$, and vice versa.
The rewrite rules used by $\mathit{SAT}_{\CARD}$ are defined as follows.
\begin{definition}[Rewrite rules]\label{d:rw_rules}
If $\pi$ is a symbol in $\Pi$ and $\phi$ is a $\card{\cdot}$-constraint based on
$\pi$, then a \emph{rewrite rule for $\pi$-constraints} is a rule of the form
$\phi \longrightarrow \Phi_1 \lor \dots \lor \Phi_n$, where $\Phi_i$, $i \geq 1$, are
$\card{\cdot}$-formulas. Each $\Sigma_{\card{\cdot}}$-predicate matching $\phi$ is
non-deterministically rewritten to one of the $\Phi_i$ s. Variables appearing in
the right-hand side but not in the left-hand side are assumed to be fresh
variables, implicitly existentially quantified over each $\Phi_i$. \hfill$\Box$
\end{definition}
A \emph{rewriting procedure} for $\pi$-constraints consists of the collection
of all the rewrite rules for $\pi$-constraints. For each rewriting procedure,
$\mathsf{STEP_S}$ checks rules in the order they are listed
in the figures below. The first rule whose left-hand side matches the input
$\pi$-constraint is used to rewrite it.
Constraints that no rule rewrites are called \emph{irreducible}. Irreducible
constraints are part of the final answer of $\mathsf{STEP_S}$ (see Definition
\ref{def:solved}).
The following conventions are used throughout the rules. $\dot x$, for any name
$x$, is a shorthand for $x \in \mathcal{V}$, i.e., $\dot x$ represents a variable. In
particular, variable names $\dot n$, $\dot n_i$, $\dot N$ and $\dot{N_i}$
denote fresh variables of sort $\mathsf{Int}$ and $\mathsf{Set}$, respectively. Moreover,
conjunctions occurring at the right-hand side of any given rule have higher
precedence than disjunctions.
\subsection{\label{steps}Set solving ($\mathsf{STEP_S}$)}
$\mathsf{STEP_S}$ can be divided into two collections of rewriting
procedures: those given as part of the CLP($\mathit{SET}$)\xspace system and those concerning
$size$ constraints.
The rewriting procedures of CLP($\mathit{SET}$)\xspace cover constraints based on $=$ when
arguments are either of sort $\mathsf{Set}$ or $\mathsf{Ur}$, $\in$, $\Cup$, and $\parallel$.
Figure \ref{f:clpset} lists some representative rewrite rules of CLP($\mathit{SET}$)\xspace
which, informally, work as follows:
\begin{itemize}
\item Rule \eqref{eq:ext} is the main rule of set unification. It states when two
non-empty, non-variable sets are equal by non-deterministically and recursively
computing four cases. These cases implement the \eqref{Ab} and \eqref{Cl}
axioms shown in Section \ref{semantics}. As an example, by applying rule
\eqref{eq:ext} to $\{1\} = \{1,1\}$ we get: ($1 = 1 \land \emptyset = \{1\}) \lor (1 =
1 \land \{1\} = \{1\}) \lor (1 = 1 \land \emptyset = \{1,1\}) \lor (\emptyset = \{1 \plus
\dot N\} \land \{1 \plus \dot N\} = \{1\})$, which turns out to be true
(due to the second disjunct).
\item Rule \eqref{in:var} rewrites a set membership constraint into an equality
constraint. This means that a formula such as $x \in \dot{A} \land y \in
\dot{A}$ will eventually be transformed into $\{x \plus \dot{N_1}\} = \{y \plus
\dot{N_2}\}$ which will be processed by rule \eqref{eq:ext}.
\item Rule \eqref{notin:ext}
deals with not membership constraints. When
the r.h.s. of a $\notin$ constraint is an extensional set term, rule
\eqref{notin:ext} operates recursively to check that $x$ is not an element of
the set. Conversely, when the r.h.s. is a variable, $\notin$ constraint
are left unchanged (see Definition \ref{def:solved}).
\item Rule \eqref{un:ext1} is one of the main rules for $\Cup$ constraints.
Observe that this rule is based on set unification. It computes two cases: $x$
does not belong to $A$ and $x$ belongs to $A$ (in which case $A$ is of the form
$\{x \plus \dot{N_2}\}$ for some set $\dot{N_2}$). In the latter case $x
\notin \dot{N_2}$ prevents Algorithm \ref{glob} from generating infinite terms
denoting the same set.
\item Finally, rule \eqref{disj:id} deals with a particular form of a disjointness
constraint.
\end{itemize}
The rest of the rewrite rules of CLP($\mathit{SET}$)\xspace can be found in \cite{Dovier00} and
online \cite{calculusBR}.
\begin{figure}
\hrule\vspace{3mm}
\begin{flalign}
\quad\quad & \{x \plus{} A\} = \{y \plus B\} \longrightarrow & \notag \\
& \qquad x = y \land A = B & \notag \\
& \qquad \lor x = y \land \{x \plus A\} = B & \label{eq:ext} \\
& \qquad \lor x = y \land A = \{y \plus B\} & \notag \\
& \qquad \lor A = \{y \plus \dot N\} \land \{x \plus \dot N\} = B & \notag \\[2mm]
& x \in \dot{A} \longrightarrow \dot{A} = \{x \plus \dot N\} & \label{in:var} \\[2mm]
& x \notin \{y \plus A\} \longrightarrow x \neq y \land x \notin A & \label{notin:ext} \\[2mm]
& \Cup(\{x \plus C\}, A, \dot{B}) \rightarrow & \notag \\
& \qquad \{x \plus C\} = \{x \plus \dot{N_1}\}
\land x \notin \dot{N_1} \land \dot{B} = \{x \plus \dot N\} & \label{un:ext1} \\
& \qquad \land (x \notin A \land \Cup(\dot{N_1}, A, \dot N) & \notag \\
& \qquad {}\qquad\lor A = \{x \plus \dot{N_2}\}
\land x \notin \dot{N_2} \land \Cup(\dot{N_1}, \dot{N_2}, \dot
N)) & \notag \\[2mm]
& \dot{X} \parallel \dot{X} \rightarrow \dot{X} = \emptyset & \label{disj:id}
\end{flalign}
\hrule
\caption{\label{f:clpset}Some rewrite rules of CLP($\mathit{SET}$)\xspace}
\end{figure}
The rewrite rules concerning $size$ constraints implemented in
$\mathsf{STEP_S}$ are listed in Figure \ref{f:card1}. Rules
\eqref{size:empty}-\eqref{size:expr} are straightforward. Rule \eqref{size:ext}
computes the size of any extensional set by counting the elements that belong
to it while taking care of avoiding duplicates. This means that, for instance,
the first non-deterministic choice for a formula such as
$size(\{1,2,3,1,4\},m)$ will be:
\[
1 \notin \{2,3,1,4\}
\land m = 1 + \dot{n} \land size(\{2,3,1,4\},\dot{n}) \land 0 \leq \dot{n}
\]
which will eventually lead to a failure due to the presence of $1 \notin
\{2,3,1,4\}$ and rule \eqref{notin:ext}. This implies that $1$ will be counted
in its second occurrence. Besides, the second choice becomes
$size(\{2,3,1,4\},m)$ which is correct given that $\card{\{1,2,3,1,4\}} =
\card{\{2,3,1,4\}}$.
\begin{figure}
\hrule\vspace{3mm}
\begin{flalign}
\quad\quad
& size(\emptyset,m) \longrightarrow m = 0 \label{size:empty} & \\[2mm]
& size(A,0) \longrightarrow A = \emptyset \label{size:zero} & \\[2mm]
& \text{If $e$ is a compound arithmetic expression:} \notag & \\
& \qquad size(A,e) \longrightarrow size(A,\dot{n}) \land \dot{n} = e \land 0 \leq \dot{n} \label{size:expr} & \\[2mm]
& size(\{x \plus A\},m) \longrightarrow & \notag \\
& \qquad x \notin A \land m = 1 + \dot n \land size(A,\dot n) \land 0 \leq \dot{n} & \label{size:ext} \\
& \qquad \lor A = \{x \plus \dot N\} \land x \notin \dot N \land size(\dot N,m)
& \notag
\end{flalign}
\hrule \caption{\label{f:card1}Rewrite rules for the $size$ constraint}
\end{figure}
Integer constraints, i.e., atomic constraints whose arguments are of sort
$\mathsf{Int}$ (including those based on $=$ and $\neq$),
are simply dealt with as irreducible by $\mathsf{STEP_S}$; hence, they are
passed ahead to be checked by the routine $\mathsf{solve\_size}$ after the main loop of
$\mathit{SAT}_{\CARD}$ terminates successfully.
\subsection{\label{irreducible}Irreducible constraints}
When no rewrite rule is applicable to the current $\card{\cdot}$-formula $\Phi$ and $\Phi$
is not $false$, the main loop of $\mathit{SAT}_{\CARD}$ terminates returning $\Phi$ as its
result. This formula can be seen, without loss of generality, as $\Phi_\mathsf{S}
\land \Phi_\mathsf{Z}$, where $\Phi_\mathsf{Z}$ contains all (and only) integer constraints
and $\Phi_\mathsf{S}$ contains all other constraints occurring in $\Phi$.
The following definition precisely characterizes the form of atomic constraints
in $\Phi_\mathsf{S}$.
\begin{definition}[Irreducible formula]\label{def:solved}
Let $\Phi$ be a $\card{\cdot}$-formula, $A$ and $A_i$ $\card{\cdot}$-terms of sort $\mathsf{Set}$, $t$
and $\dot{X}$ $\card{\cdot}$-terms of sort $\langle \{\mathsf{Set},\mathsf{Ur}\} \rangle$, $x$ a
$\card{\cdot}$-term of any sort, and $c$ a variable or a constant integer number. A
$\card{\cdot}$-constraint $\phi$ occurring in $\Phi$ is \emph{irreducible} if it has
one of the following forms:
\begin{enumerate}[label=(\roman*), leftmargin=*, widest=viii]
\item \label{i:icfirst} $\dot{X} = t$, and neither $t$ nor $\Phi \setminus \{\phi\}$
contains $\dot{X}$;
\item $\dot{X} \neq t$, and $\dot{X}$ does not occur either in $t$ or
as an argument of any constraint $\pi(\dots)$, $\pi \in \{\Cup,size\}$, in
$\Phi$;
\item $x \notin \dot{A}$, and $\dot{A}$ does not occur in $x$;
\item $\Cup(\dot{A}_1,\dot{A}_2,\dot{A}_3)$, where $\dot{A}_1$ and $\dot{A}_2$
are distinct variables;
\item $\dot{A}_1 \parallel \dot{A}_2$, where $\dot{A}_1$ and $\dot{A}_2$ are distinct variables;
\item $size(\dot{A}, c)$, $c \neq 0$.
\end{enumerate}
A $\card{\cdot}$-formula $\Phi$ is irreducible if it is $\mathit{true}$ or if all of its
$\card{\cdot}$-constraints are irreducible. \hfill$\Box$
\end{definition}
$\Phi_\mathsf{S}$, as returned by $\mathit{SAT}_{\CARD}$ once it finishes its main loop, is an
irreducible formula. This fact can be checked by inspecting the rewrite rules
presented in \cite{Dovier00} and those for the $size$ constraints given in
Figure \ref{f:card1}. This inspection is straightforward as there are no
rewrite rules dealing with irreducible constraints and all non-irreducible form
constraints are dealt with by some rule.
Putting $size$ constraints aside, $\Phi_\mathsf{S}$ is basically the formula
returned by the CLP($\mathit{SET}$)\xspace solver. \citeN[Theorem 9.4]{Dovier00} show
that such formula is always satisfiable, unless the result is $false$.
It is important to observe that the atomic constraints occurring in $\Phi_\mathsf{S}$
are indeed quite simple. In particular, all non-variable set terms occurring in the input formula have been removed,
except those occurring as right-hand sides of $=$ and $\neq$ constraints. Thus,
all (possibly complex) equalities and inequalities between set terms have been
solved. Furthermore, all arguments of $\Cup$ and $\parallel$ constraints are
necessarily simple variables.
\section{\label{sizesolver}Cardinality solving ($\mathsf{solve\_size}$)}
Due to the presence of $size$ and integer constraints, a non-$false$ formula
returned by $\mathsf{STEP_S}$ and $\mathsf{remove\_neq}$ is not always
satisfiable.
\begin{example}
Assuming all the arguments to be variables, the following formula cannot be
processed any further by $\mathsf{STEP_S}$ but is unsatisfiable:
\begin{equation*}
\Cup(A,B,C) \land size(A,m_a) \land size(B,m_b) \land size(C,m_c)
\land m_a + m_b < m_c
\end{equation*}
as it states that $\card{A} + \card{B} < \card{A \cup B}$.
\hfill$\Box$
\end{example}
Therefore, Algorithm \ref{glob} includes a new step, called $\mathsf{solve\_size}$, whose
purpose is to check satisfiability of the formula returned at the end of the
main loop of $\mathit{SAT}_{\CARD}$.
Basically, $\mathsf{solve\_size}$ encodes an adaptation of the $\mathit{SAT}_{Za}$ algorithm to our
CLP system. In order to explain how we adapted $\mathit{SAT}_{Za}$ we first introduce it
briefly; some technical details are omitted to simplify the presentation.
\subsection{\label{zarba}An algorithm for deciding set formulas with cardinality}
The language considered by Zarba--hereafter simply called $\mathit{L}_{Za}$---includes the
following function symbols: $\emptyset$, $\cup$, $\cap$, $\setminus$, $+$, $-$ and
$\card{\cdot}$; the usual predicate symbols: $=$, $\in$, $<$, $>$; and
variables and integer constants as usual. All symbols have standard sorts and
semantics; in particular, sets are finite. The language also includes the
singleton set symbol $\{\cdot\}$ to form extensional sets. Note that although
$\mathit{L}_{Za}$ does not include an integer product symbol, it still allows the
representation of expressions of the form $c*x$, with either $c$ or $x$ a
constant. Formulas in $\mathit{L}_{Za}$ are built in the usual way.
$\mathit{SAT}_{Za}$ is divided into four phases and takes as input a conjunction of
$\mathit{L}_{Za}$ literals. However, we will present the last two phases as a single one.
\begin{enumerate}
\item
\textsc{First phase}. The input formula, $\Psi$, is
transformed and divided into two subformulas, $\Psi'$ and $\Psi''$. $\Psi''$
contains only literals of the form $v = \card{x}$ where $v$ and $x$ are integer
and set variables, respectively. $\Psi'$ contains the integer constraints
present in $\Psi$ plus a transformation of the set constraints in $\Psi$. This
transformation guarantees that all set constraints are of the following forms:
$x = y$, $x \neq y$, $x = \{u\}$, $x = y \cup z$, $x = y \cap z$ and $x = y
\setminus z$, where $x$, $y$ and $z$ are set variables and $u$ is a
ur-variable.
\begin{example}
A constraint such as $y \in x$ is transformed into $x = \{y\} \cup x$ and then
into $w = \{y\} \land x = w \cup x$, where $w$ is a new variable.
A constraint such as $\{u\} \cup x = h \cap w$ is transformed into $v = \{u\}
\land v \cup x = h \cap w$ and then into $v = \{u\} \land t = v \cup x \land t
= h \cap w$, where $v$ and $t$ are new variables.
A constraint such as $\card{x} + m < k$ is transformed into $v = \card{x}
\land v + m < k$, where $v$ is a new variable. In this way, $v = \card{x}$
becomes part of $\Psi''$. \hfill$\Box$
\end{example}
\item
\textsc{Second phase}. $\Psi'$ is divided into three subformulas: $\Psi_\mathsf{U}$,
containing literals of the form $x = \{u\}$, where $u$ is a ur-element;
$\Psi_\mathsf{Z}$, containing the integer literals; and $\Psi_\mathsf{S}$, containing the
set literals. So now the input formula has been transformed and divided into
four subformulas: $\Psi_\mathsf{U}$, $\Psi_\mathsf{Z}$, $\Psi_\mathsf{S}$ and $\Psi''$. In the
next phase, $\Psi
\defs \Psi_\mathsf{U} \land \Psi_\mathsf{Z} \land \Psi_\mathsf{S} \land \Psi''$.
\item
\textsc{Third phase}. This phase consists in executing the following three
steps for each \emph{arrangement} of $\Psi$. Whenever there are no more
arrangements the input formula is unsatisfiable.
An arrangement of $\Psi$ is a tuple $\langle R, \Pi, at\rangle$ where: $R
\subseteq \mathcal{V}_\mathsf{U}^\Psi \times \mathcal{V}_\mathsf{U}^\Psi$ is an equivalence relation where
$\mathcal{V}_\mathsf{U}^\Psi$ is the collection of ur-variables in $\Psi_\mathsf{U}$; $\Pi$ is a
finite collection of non-$false$ Boolean functions $\pi: \mathcal{V}_\mathsf{S}^\Psi \fun
\{0,1\}$ where $\mathcal{V}_\mathsf{S}^\Psi$ is the collection of set variables in
$\Psi_\mathsf{S} \land \Psi''$; and $at: \mathcal{V}_\mathsf{U}^\Psi \fun \Pi$. $\pi$ is a
non-$false$ Boolean function if $1 \in \ran \pi$.
From now on $\rho = \langle R, \Pi, at\rangle$ denotes the current arrangement.
\begin{enumerate}[ref=\alph*]
%
\item\label{i:checkarr}
In this step the algorithm checks whether or not $\rho$ verifies seven
conditions. If $\rho$ does not verify these conditions the next arrangement
is chosen; if it does the next step is executed. We show only the
conditions that are used in our implementation.
\begin{enumerate}
\renewcommand{\labelenumii}{\roman{enumii}}
\item\label{i:un}
If $x = y \cup z$ is in $\Psi_\mathsf{S}$ then $\pi(x) = 1$ if and only if
$\pi(y) = 1$ or $\pi(z) = 1$, for each $\pi \in \Pi$.
%
\item\label{i:disj}
If $\emptyset = y \cap z$ is in $\Psi_\mathsf{S}$ then $\pi(y) = 0$ or $\pi(z) = 0$.
\end{enumerate}
The remaining conditions are not used because $\mathit{SAT}_{Za}$ is called after
$\mathsf{STEP_S}$; see Section \ref{integrating} for more details.
%
\item\label{i:checkint}
In this step the algorithm checks whether or not $\Psi_\mathsf{Z} \land
\mathit{res}_\mathsf{Z}(\rho)$ is satisfiable, where:
\begin{equation}\label{eq:res}
\begin{split}
\mathit{res}_\mathsf{Z}&(\rho) \defs \\
& \bigwedge_{\pi \in \Pi} 0 < v_\pi
\bigwedge_{\pi \in \ran at} v_\pi = 1
\bigwedge_{v = \card{x} \in \Psi''}
v = \sum_{\pi \in \Pi} \pi(x) * v_\pi
\end{split}
\end{equation}
If $\Psi_\mathsf{Z} \land \mathit{res}_\mathsf{Z}(\rho)$ is unsatisfiable the
next arrangement is chosen and step \eqref{i:checkarr} is executed.
%
\item
In this last step the algorithm checks whether or not there are enough
ur-elements as to satisfy $\Psi_\mathsf{U}$ considering the equivalence relation $R$
of $\rho$ and the minimum of $\sum_{\pi \in \Pi} v_\pi$ subject to $\Psi_\mathsf{Z}
\land \mathit{res}_\mathsf{Z}(\rho)$. If this is satisfiable, the input formula is
satisfiable; if not, the next arrangement is chosen and step \eqref{i:checkarr}
is executed.
\end{enumerate}
Informally, in this phase the algorithm assigns a positive cardinality ($v_\pi$)
to each non-empty Venn region involved in the formula and tries, one after the
other, all possible combinations of these assignments---each combination is
encoded in each arrangement. With each combination it builds formula
\eqref{eq:res} and checks whether the cardinality constrains are
satisfiable or not.
\end{enumerate}
\subsection{Integrating $\mathit{SAT}_{Za}$ into $\mathit{SAT}_{\CARD}$}\label{integrating}
The repeated execution of $\mathsf{STEP_S}$ and \textsf{remove\_neq} in
$\mathit{SAT}_{\CARD}$ implements up to the second phase of $\mathit{SAT}_{Za}$. The third phase
of $\mathit{SAT}_{Za}$ is implemented by $\mathsf{solve\_size}$. Formulas returned at the end of the
main loop of $\mathit{SAT}_{\CARD}$ (i.e., $\card{\cdot}$-formulas in irreducible form) can be
easily transformed into the formulas obtained after executing the second phase
of $\mathit{SAT}_{Za}$. A detailed definition of a mapping of such formulas into the
corresponding $\mathit{L}_{Za}$ formulas is given in \ref{mapping}. Hereafter, we provide
an intuitive description of which formulas are passed to
$\mathsf{solve\_size}$.
Let $\Phi \defs \Phi_1 \land \Phi_2$ be the formula in irreducible form right
after the main loop of Algorithm \ref{glob}, where $\Phi_1$ contains all
integer constraints and all of the $\Cup$, $\parallel$ and $size$ constraints, and
$\Phi_2$ is the rest of $\Phi$ (i.e., $\notin$ constraints, and $=$ and $\neq$
constraints not involving integer terms). Hence, $\mathsf{solve\_size}$ is called on
$\Phi_1$ as follows:
\begin{itemize}
\item
All integer constraints are passed basically unaltered to
$\mathsf{solve\_size}$.
\item
$\card{\cdot}$-constraints of the form $\Cup(A,B,C)$, $A \parallel B$, $size(A,m)$, where
$A, B, C$ are variables and $m$ is either a variable or an integer constant,
are mapped to literals of the form $C = A \cup B$, $A \cap B = \emptyset$, $\card{A} =
m$, respectively, in $\mathit{L}_{Za}$.
\end{itemize}
On the other hand, constraints in $\Phi_2$ are
not passed to $\mathsf{solve\_size}$:
\begin{itemize}
\item equality constraints are ignored because these
variables do not appear in the rest of $\Phi$.
\item
$\neq$ constraints not involving integer terms and $\notin$ constraints are
ignored because they do not affect the cardinality of the set variables
involved in the formula. Indeed, in $\mathit{L}_{\CARD}$ we assume that the universe of
objects which can be used as set elements is infinite---as it includes integer
numbers and (nested) sets. Hence, constraints of the form $X \neq t$ and $t
\notin X$ (with $X$ variable and $t$ any term) do not forbid any value of the
cardinality of $X$. For instance, if $\Phi$ contains $1 \notin S \land 2 \notin
S \land ... \land 20 \notin S \land size(S,m)$, with $S$ variable, then we can
find anyway $m$ constants different from $1,\dots,20$ to fill the set $S$.
\end{itemize}
Note that non-variable set terms occur only in those
constraints of $\Phi_2$ that are not passed to $\mathsf{solve\_size}$. Thus, the
translation function $\mathcal{Z}$ shown in \ref{mapping}, which only deals with
variables, is indeed capable of translating any $\mathit{L}_{\CARD}$ formula that is passed
to it.
$\mathsf{solve\_size}$ implements the first two steps of the third phase by casting step
\eqref{i:checkarr} in terms of a Boolean satisfiability problem and step
\eqref{i:checkint} in terms of an integer linear programming (ILP) problem
\cite{10.5555/1611284}. All the solutions returned by solving the Boolean
formula are collected in a set $S$ and then all possible
arrangements are the elements of $2^{S}$.
A description of a concrete implementation of these two steps is given in the
next subsection.
The last step of the third phase is not implemented again because of
the assumption about the infinity of the universe of objects which can be used
as set elements in $\mathit{L}_{\CARD}$.
It is worth noting that, in the integrated system, unsatisfiability
caused by set constraints, excluding $size$, can be caught directly by
$\mathsf{STEP_S}$ and \textsf{remove\_neq}, without executing
$\mathsf{solve\_size}$.
\begin{example}
Consider the following formula:
\begin{equation*}
\Cup(A,B,C) \land A \parallel C \land A \neq \emptyset \land size(C,k) \land k < 2.
\end{equation*}
where $A$, $B$, $C$, and $k$ are variables. The subformula $\Cup(A,B,C) \land A
\parallel C \land A \neq \emptyset$ is not in irreducible form and it is further processed
first by \textsf{remove\_neq} and then by $\mathsf{STEP_S}$, that finally
rewrites it to false. That is, the input formula is found to be unsatisfiable
disregarding the cardinality and integer constraints occurring in it. \hfill$\Box$
\end{example}
On the other hand, the presence of $\mathsf{solve\_size}$ in $\mathit{SAT}_{\CARD}$ allows us to solve
linear integer constraints even if the given formula does not contain any
$size$ constraint. For example, a formula such as $x > y \land x < y +
1$ is found to be false by exploiting the integer constraint solver included
in $\mathsf{solve\_size}$.
\subsection{A concrete implementation of $\mathsf{solve\_size}$}\label{concrete}
In this section we briefly outline a concrete Prolog implementation of
$\mathsf{solve\_size}$. This implementation is obtained by integrating into the
$\mathsf{solve\_size}$ procedure described above a Prolog Boolean SAT solver, namely the
very concise solver developed by
\citeN{DBLP:journals/tcs/HoweK12}, and the implementation of the CLP(Q) system
of SWI-Prolog \cite{holzbaur1995ofai}.
CLP(Q) implements a solver for linear equations, a Simplex algorithm to decide
linear inequalities and a branch and bound method to provide a decision
algorithm for ILP. This library provides
$\mathtt{bb\_inf}(Vars,Expr,Min,Vert)$, which finds the vertex ($Vert$) of the
minimum ($Min$) of the expression $Expr$ subjected to the integer constraints
present in the constraint store and assuming all the variables in $Vars$ take
integers values. In its way to find the minimum value, $\mathtt{bb\_inf}$ first
determines whether or not the constraints are satisfiable (in $\num$).
$\mathtt{bb\_inf}$ is complete provided all integer constraints are linear.
With respect to the completeness of $\mathtt{bb\_inf}$, observe that: \emph{a)}
$\mathit{L}_{\CARD}$ restricts integer constraints to be linear (Definition
\ref{LCARD-terms}); and \emph{b)} the integer constraints generated by any rule
for $size$ are linear.
Consider a formula $\Phi$ received by $\mathsf{solve\_size}$. Now consider the subformula
of $\Phi$ that is a conjunction of constraints of the following forms:
$\Cup(A,B,C)$ and $A \parallel B$, with $A, B$ and $C$ variables. As $\mathit{SAT}_{Za}$
must find all the non-$false$ Boolean functions $\pi:\mathcal{V}_\mathsf{S}^\Phi \fun
\{0,1\}$ verifying some Boolean conditions (see Section \ref{sizesolver} for
some examples and \cite[conditions (C1)-(C7) in
3.4]{DBLP:conf/frocos/Zarba02}), we encode the conjunction of these constraints
as a Boolean formula as follows:
\begin{itemize}
\item $\Cup(A,B,C) \longrightarrow (\lnot C \lor B \lor A)
\land (\lnot A \lor C) \land (C \lor \lnot A)$,
due to condition \ref{i:un}.
\item $A \parallel B \longrightarrow \lnot A \lor \lnot B$,
due to condition \ref{i:disj}.
\end{itemize}
Next, we call Howe and King's SAT solver on the resulting Boolean formula and
collect in a set $S$ all the Boolean solutions where at least one variable is
bound to $\mathit{true}$. Hence, $S$ contains all possible non-$false$ Boolean
functions $\pi:\mathcal{V}_\mathsf{S}^\Phi \fun \{0,1\}$ satisfying $\mathit{SAT}_{Za}$'s conditions \ref{i:un}
and \ref{i:disj}.
If $\{\pi_1,\dots,\pi_n\}$ verifies the above condition, then we use it to
execute the second step of the third phase. Then we build formula
\eqref{eq:res} as a conjunction of CLP(Q) constraints, which is easy to
implement. All the integer constraints present in $\Phi$ and all those in
\eqref{eq:res} are passed in to the CLP(Q) constraint store. Finally, we call
CLP(Q)'s \verb+bb_inf/4+ predicate\footnote{\texttt{bb\_inf/4}:
\url{https://www.swi-prolog.org/pldoc/doc_for?object=bb_inf/4}} as follows:
\begin{equation}\label{eq:bbinf}
\mathtt{bb\_inf}(\mathcal{V}_\mathsf{Z},\sum_{i=1}^k m_i,\_,Vertex)
\end{equation}
where $m_1,\dots,m_k$ are the second arguments of the $size$ constraints in
$\Phi$. That is, we ask CLP(Q) to check the satisfiability of its constraint
store assuming that all the variables there are integers, and if so, to compute
the vertex ($Vertex$) of the minimum of the sum of the cardinalities of the
sets in $\Phi$. If this call does not fail we know $\Phi$ is satisfiable and
$\mathsf{solve\_size}$ terminates; if not, we pick the next subset of $S$. If $\mathsf{solve\_size}$
fails for all subsets of $S$ it returns $false$.
\section{\label{decproc}$\mathit{SAT}_{\CARD}$ is a decision procedure for $\mathit{L}_{\CARD}$}
In this section we analyze the soundness, completeness and termination
properties of $\mathit{SAT}_{\CARD}$.
The following theorem ensures that, after termination, the
rewriting process implemented by $\mathit{SAT}_{\CARD}$ preserves the set of solutions of
the input formula.
\begin{theorem}[Equisatisfiability]\label{equisatisfiable}
Let $\Phi$ be a $\card{\cdot}$-formula and $\Phi^1, \Phi^2,\dots,\Phi^n$ be the
collection of $\card{\cdot}$-formulas returned by $\mathit{SAT}_{\CARD}(\Phi)$. Then $\Phi^1 \lor
\Phi^2 \lor \dots \lor \Phi^n$ is equisatisfiable to $\Phi$, that is, every
possible solution\footnote{More precisely, each solution of $\Phi$ expanded to
the variables occurring in $\Phi^i$ but not in $\Phi$, so as to account for the
possible fresh variables introduced into $\Phi^i$.} of $\Phi$ is a solution of
one of the $\Phi^i$s and, vice versa, every solution of one of these formulas is
a solution for $\Phi$.
\end{theorem}
\begin{proof}
According to Definition \ref{irreducible}, each formula
$\Phi_i$ returned at the end of $\mathit{SAT}_{\CARD}$'s main loop is of the form
$\Phi^i_\mathsf{S} \land \Phi^i_\mathsf{Z}$, where $\Phi^i_\mathsf{S}$ is a $\card{\cdot}$-formula
in irreducible form and $\Phi^i_\mathsf{Z}$ contains all integer constraints
encountered during the processing of the input formula.
As concerns constraints in $\Phi_i^\mathsf{S}$, the proof is based on showing that
for each rewrite rule the set of solutions of left and right-hand sides is the
same. For those rules dealing with constraints different from $size$ the
proofs can be found in \cite{Dovier00}. The proofs of equisatisfiability for
the rules for $size$ can be found in \ref{proofs}.
As concerns $\Phi^i_\mathsf{Z}$, no rewriting is actually performed on the
constraints occurring in it. Thus the set of solutions is trivially preserved.
Considering also the last step of $\mathit{SAT}_{\CARD}$, i.e., calling $\mathsf{solve\_size}$, we
observe that this step is just a check which either returns $false$ or has no
influence on its input formula.
\end{proof}
\begin{theorem}[Satisfiability of the output formula]\label{satisf}
Any $\card{\cdot}$-formula different from $false$ returned by $\mathit{SAT}_{\CARD}$ is
satisfiable w.r.t. the underlying interpretation structure $\mathcal{R}$.
\end{theorem}
\begin{proof}
Basically, the proof of this theorem relies on the fact that $\mathsf{solve\_size}$
implements $\mathit{SAT}_{Za}$. Let $\Phi$ be the input formula and $\Phi'$ its
irreducible form right before $\mathsf{solve\_size}$. Consider that $\Phi'$ is divided as
$\Phi'_1 \land \Phi'_2$ where $\Phi'_1$ contains the integer constraints and
the $\Cup$, $\parallel$ and $size$ constraints; and $\Phi'_2$ all the other
constraints. Then, $\Phi'_1$ can be easily mapped to formulas which are
accepted by $\mathit{SAT}_{Za}$ (see \ref{mapping}). As observed in Section
\ref{integrating}, $\Phi'_2$ is not passed to $\mathit{SAT}_{Za}$
because is irrelevant as regards the satisfiability of $\Phi'_1$. Then, the
satisfiability of $\Phi$ depends only on the satisfibility of $\Phi'_1$. Hence,
if $\mathsf{solve\_size}$ decides that $\Phi'_1$ is satisfiable, we can conclude that
$\Phi$ is satisfiable. In this case $\mathit{SAT}_{\CARD}$ returns $\Phi$.
\end{proof}
Thanks to Theorems \ref{equisatisfiable} and \ref{satisf} we can conclude that,
given a $\card{\cdot}$-formula $\Phi$, then $\Phi$ is
satisfiable with respect to the intended interpretation structure $\mathcal{R}$ if and
only if there is a non-deterministic choice in $\mathit{SAT}_{\CARD}(\Phi)$ that returns a
$\card{\cdot}$-formula different from $false$. Conversely, if all the
non-deterministic computations of $\mathit{SAT}_{\CARD}(\Phi)$ terminate with $false$,
then $\Phi$ is surely unsatisfiable.
The following is an example of a formula that $\mathit{SAT}_{\CARD}$ is able to detect to
be unsatisfiable.
\begin{example}
The formula
\begin{equation*}
\Cup(A,B,C) \land size(A,m_1) \land size(B,m_2) \land size(B,m_3) \land m_3
> m_1 + m_2
\end{equation*}
where all arguments are variables, is rewritten by $\mathit{SAT}_{\CARD}$ to $false$;
hence, the formula is unsatisfiable. \hfill$\Box$
\end{example}
Note that many of the rewriting procedures given in the previous section will
stop even when returning relatively complex formulas.
\begin{example}\label{ex:solution}
Assuming all the arguments are variables, the formula:
\begin{equation*}
\Cup(A,B,C) \land size(A,m_1) \land size(B,m_2) \land size(B,m_3) \land m_3
\leq m_1 + m_2
\end{equation*}
is returned unchanged by $\mathit{SAT}_{\CARD}$ because there is no rewrite rule for
constraints such as $\Cup(A,B,C)$ and $size(A,m)$ when all arguments of sort
$\mathsf{Set}$ are variables. Actually, this formula is proved to be satisfiable by applying
$\mathsf{solve\_size}$. \hfill$\Box$
\end{example}
Finally, we can state the termination property for $\mathit{SAT}_{\CARD}$.
\begin{theorem}[Termination] \label{termination-glob}
The $\mathit{SAT}_{\CARD}$ procedure can be implemented as to ensure termination
for every input $\mathit{L}_{\CARD}$ formula.
\end{theorem}
\begin{proof}
Termination of the $\mathit{SAT}_{\CARD}$ is a consequence of the termination
proved in Theorem 10.10 in \cite{Dovier00} and Zarba's
algorithm \cite[Theorem 3]{DBLP:conf/frocos/Zarba02}. The only new observations
to be done concern the treatment of $size$ constraints. Looking at the rewrite
rules for this kind of constraints shown in Figure \ref{f:card1}, we can
observe that: they generate equality and inequality constraints (in fact,
$\not\in$ constraints are rewritten to $\neq$ constraints), which in turn do
not generate any new $size$ constraint; besides, they generate new $size$
constraints which, however, are in irreducible form, since their first argument
is a (fresh) variable. Therefore, the processing of $size$ constraints cannot
trigger any infinite loop.
\end{proof}
\section{\label{minimal}Minimal solutions}
The formulas $\Phi_1,\dots,\Phi_n, n \ge 1,$ returned by $\mathit{SAT}_{\CARD}$ represent
all the concrete (or ground) \emph{solutions} of the input formula $\Phi$. If
these formulas do not contain any $size$ or integer constraints, then it is
quite easy to get concrete solutions from them. Indeed, a successful assignment
of values to variables (i.e., a concrete solution) for such formulas is
obtained by substituting each set variable occurring in them by the empty set,
with the exception of the variables $X$ in atoms of the form $X = t$.
Unfortunately, when it comes to the $size$ and integer constraints, providing
concrete solutions for certain $\mathit{L}_{\CARD}$-formulas may be difficult.
\begin{example}\label{ex:size_slvfrm}
If $\mathit{SAT}_{\CARD}$ is called on the following formula:
\begin{equation*}
size(A,m) \land 1 \leq m \land B \subseteq A
\land size(B,n) \land 5 \leq n
\end{equation*}
it will return the same formula meaning that it is satisfiable. However, a
solution is not evident. \hfill$\Box$
\end{example}
For some applications such as model-based testing \cite{CristiaRossiSEFM13}
determining the satisfiability of a formula is not enough. More explicit
solutions are needed. For this reason we provide a way in which $\mathit{SAT}_{\CARD}$
returns formulas for which finding a solution is always easy. We call such a
solution \emph{minimal} because no cardinality of a set assigned to a variable
appearing in a $size$ constraint can be lowered without making the formula
false. However, in this case we cannot get a finite representation of the set
of all possible solutions.
Let $\Phi$ be a satisfiable input formula and let $\Phi'$ the corresponding
formula right before $\mathsf{solve\_size}$ is called. Let $size(A_1,m_1), \dots,
size(A_k,m_k)$ be all the $size$ constraints in $\Phi'$. If $\mathit{SAT}_{\CARD}$ is
required to compute the minimal solution, once Algorithm \ref{glob} finishes,
it is called again with the following formula:
\begin{equation}\label{eq:minimal}
\Phi' \land \bigwedge_{i=1}^{k} m_i = V_i
\end{equation}
where $\langle V_1,\dots,V_k \rangle$ is the $Vertex$ computed in
\eqref{eq:bbinf}. In this way all sets $A_i$ of the $size$ constraints in
$\Phi'$ are bound to bounded sets of least possible cardinality so as to
satisfy $\Phi$. Note that, necessarily, $0 \leq V_i$, for $i \in [1,k]$.
Besides, when $\mathit{SAT}_{\CARD}$ runs in this mode it will not call $\mathsf{solve\_size}$ to
solve \eqref{eq:minimal}. In fact, $\bigwedge_{i=1}^{k} m_i = V_i$ turns all
$size$ constraints in $\Phi'$ into atoms of the form $size(\dot{A},c)$ with
$c$ a constant. Then, the following rewrite rule is activated:
\begin{equation}
size(\dot A,c) \text{, $c$ an integer constant} \longrightarrow
\dot A = \{\dot{n}_1,\dots,\dot{n}_c\}
\land ad(\dot{n}_1,\dots,\dot{n}_c) \label{size:const3}
\end{equation}
where $ad(y_1,\dots,y_c)$ is a shorthand for $\bigwedge_{i=1}^{c-1}
\bigwedge_{j=i+1}^c y_i \neq y_j$.
\begin{example}\label{ex:minsol}
If $\mathit{SAT}_{\CARD}$ is called on the formula of Example \ref{ex:size_slvfrm} but
requiring that all minimal solutions be computed, then the formula returned at
the end of the computation is:
\begin{flalign*}
\quad\quad & A = \{n_5,n_4,n_3,n_2,n_1\}, m = 5,
B = \{n_5,n_4,n_3,n_2,n_1\}, n = 5,
ad(n_5,n_4,n_3,n_2,n_1) &
\end{flalign*}
This formula is a finite representation of a subset of the
possible solutions for the input formula
from which it is trivial to get concrete solutions. \hfill$\Box$
\end{example}
\section{\label{impl} The Implementation and its Empirical Evaluation}
$\mathit{SAT}_{\CARD}$ is implemented by extending the solver provided by the publicly
available tool $\{log\}$\xspace \cite{setlog}. $\{log\}$\xspace is a Prolog program that can be
used as a constraint solver, as a satisfiability solver and as a constraint
logic programming language. It also provides some programming facilities not
described in this paper. In this section we describe and empirically evaluate
this implementation.
The main syntactic differences between the abstract syntax used in previous
sections and the concrete syntax used in $\{log\}$\xspace are made evident by the
following examples.
\begin{example}\label{ex:setlogformulas}
The formulas of Example \ref{ex:formulas} are written in $\{log\}$\xspace as follows:
\begin{verbatim}
a in A & a nin B & un(A,B,C) & C = {X / D}.
un(A,B,C) & N + K > 5 & size(C,N) & B neq {}.
\end{verbatim}
where names beginning with a capital letter represent variables, and all others
represent constants and function symbols. This is why we renamed some variables
w.r.t. the formulas in Example \ref{ex:formulas}. Note that \verb+{_/_}+ is the
concrete syntax for the set term $\{\_\plus\_\}$.
If $\{log\}$\xspace is asked to solve the second formula it returns the following:
\begin{verbatim}
B = {_N3/_N2}, C = {_N3/_N1}
Constraint: un(A,_N2,_N1), N + K > 5, _N3 nin _N1,
size(_N1,_N4), _N4 >= 0, N >= 1, _N4 is N - 1
\end{verbatim}
as the first solution (more can be obtained interactively). That is, $\{log\}$\xspace
binds values to $B$ and $C$ and gives a list of constraints in irreducible form
(which is guaranteed to be satisfiable). Any concrete solution must bind values
to the remaining variables in such a way as to verify the constraints.
Variables beginning with the underscore symbol (\verb+_+) represent new
variables.
\hfill$\Box$
\end{example}
The implementation in $\{log\}$\xspace of $\mathsf{STEP_S}$ consists in adding to
$\{log\}$\xspace the rewrite rules of Figure \ref{f:card1}. Due to the design of
$\{log\}$\xspace, adding new constraints and their rewrite rules is easy, and it does
not deserve to be further commented here. On the other hand, the implementation
in $\{log\}$\xspace of $\mathsf{solve\_size}$ is basically that described in Section
\ref{concrete}.
Observe
that the fact that $\{log\}$\xspace is based on set unification automatically provides
cardinality over sets of sets---nested at any level. For instance, running
\verb+size({{X},{Y}},N)+ produces two solutions:
\begin{verbatim}
N = 2, X neq Y;
Y = X, N = 1
\end{verbatim}
Concerning formulas with $size$ constraints, by default $\{log\}$\xspace decides their
satisfiability as described in Section \ref{sizesolver}. That is, if the
formula of Example \ref{ex:size_slvfrm} is executed, $\{log\}$\xspace will find it
satisfiable and will return it unchanged. If users want more concrete
solutions, as described in Section \ref{minimal}, they must execute command
\verb+fix_size+ to activate the algorithm that computes minimal solutions.
In this case, after solving the formula of Example \ref{ex:size_slvfrm},
$\{log\}$\xspace would return exactly the solution shown in Example \ref{ex:minsol}. As
another example, when solving the second formula of Example \ref{ex:setlogformulas}
in \texttt{fix\_size} mode, $\{log\}$\xspace will return (as its first
solution):
\begin{verbatim}
A = {}, B = {_N1}, C = {_N1}, N = 1
Constraint: 1 + K > 5
\end{verbatim}
which is indeed a more concrete solution for the given formula.
\subsection{\label{formver}Applications to formal verification}
We now present a simple example showing how $\{log\}$\xspace can be used as a
verification tool of problems involving cardinality constraints. In doing so we
will show how our approach differs from other tools that can deal with similar
problems---see Section \ref{related} for a detailed account. More than 250
real-world examples have been used in the empirical evaluation presented in
Section \ref{empirical} and another example is developed in
\ref{app:datacontainer}. The example is taken from \citeN{Kuncak2006}. Figure
\ref{f:insert} shows the $\mathsf{insert}$ procedure which inserts an element
$e$ into the set $content$. Besides, the procedure maintains the cardinality of
$content$ in variable $size$. In this context an \textsf{element} is an object
represented as a set of cardinality one. The procedure is annotated with its
preconditions (i.e., \textsc{requires}), its postconditions (i.e.,
\textsc{ensures}) and the invariant it preserves (i.e., \textsc{maintains}).
Kuncak then proposes a verification condition for the \textsf{insert}
procedure.
\begin{figure}
\raggedright
\textsf{var} $content$:\textsf{set}; $size$:\textsf{integer}; \\[1mm]
\textsf{procedure insert}($e$:\textsf{element}) \{
\hfill {\footnotesize
[\textsc{requires:}
$\card{e} = 1 \land \card{e \cap content} = 0$]}\\
\hspace{3mm}$content := content \cup e$;
\hfill {\footnotesize
[\textsc{maintains:} $size = \card{content}$] \\
\hspace{3mm}$size := size + 1$}; \\
\} \hfill {\footnotesize
[\textsc{ensures:} $size' > 0$]}
\caption{\label{f:insert}Procedure $\mathsf{insert}$ inserts $e$ into set
$contents$ and updates its cardinality in $size$}
\end{figure}
The $\{log\}$\xspace representation of \textsf{insert} is the following:
\begin{Verbatim}[commandchars=\\\$\$]
sl_insert(Content,Size,E,Content_,Size_) :-
un(Content,E,Content_) & \hfill\small\rm [\(content := content \cup e\)]
Size_ is Size + 1. \hfill\small\rm [\(size := size + 1\)]
\end{Verbatim}
where \verb+Content+ and \verb+Size+ are the initial values and
\verb+Content_+ and \verb+Size_+ the final ones. In this way,
\verb+sl_insert+ becomes a $\{log\}$\xspace \emph{program} and thus it
can be executed as any other program and can be part of a larger
Prolog+$\{log\}$\xspace program. For example the query:
\begin{verbatim}
sl_insert({},0,{hellow},C1,S1).
\end{verbatim}
returns:
\begin{verbatim}
C1 = {hellow}, S1 = 1
\end{verbatim}
and the following one:
\begin{verbatim}
sl_insert({},0,{hellow},C1,S1) & sl_insert(C1,S1,{world},C2,S2).
\end{verbatim}
returns:
\begin{verbatim}
C1 = {hellow}, S1 = 1, C2 = {hellow,world}, S2 = 2
\end{verbatim}
Furthermore, \verb+sl_insert+ is also a \emph{formula}. Indeed, we can
discharge the verification condition indicated by Kuncak \emph{using the same
representation} of \textsf{insert} by simply encoding the negation of the
verification condition as a $\{log\}$\xspace query:
\begin{Verbatim}[commandchars=\\\$\$]
size(E,1) & inters(E,Content,M1) & size(M1,0) & \hfill\small \rm[precondition]
size(Content,Size) & \hfill\small\rm[invariant@before state]
sl_insert(Content,Size,E,Content_,Size_) & \hfill\small \rm[\sf insert\rm is executed]
(Size_ =< 0 \hfill\small \rm[negation of postcondition@after state]
or
size(Content_,M2) & M2 neq Size_ \hfill\small \rm[negation of invariant@after state]
).
\end{Verbatim}
If the answer is \verb+no+ it means the query is unsatisfiable for all values
of the variables, and so the verification condition is a theorem. $\{log\}$\xspace runs
this query in 0.016 seconds.
As the example shows, the $\{log\}$\xspace representation of \textsf{insert} is both a
formula (or executable specification) \emph{and} a program (or prototype,
because of its lack of efficiency). Or put it in another way, $\{log\}$\xspace is the
very same tool that executes \textsf{insert} and \emph{automatically} proves
its correctness. We think this is a rare characteristic in verification tools
dealing with cardinality constraints. $\{log\}$\xspace has been used in the same fashion
on real-world problems (\citeANP{DBLP:journals/jar/CristiaR21}
\citeyearNP{DBLP:journals/jar/CristiaR21,Cristia2021}).
\subsection{Improvements}
In this section we present some improvements recently made to
$\{log\}$\xspace to render it a more usable tool.
\paragraph{Derived constraints.}
As shown in Section \ref{expressiveness}, many set operators in
$\{log\}$\xspace are defined as derived constraints, i.e., as $\card{\cdot}$-formulas built out
of the primitive constraints that $\mathit{L}_{\CARD}$ offers. For example, the predicate
$\Cap(A,B,C)$, which is true when $C$ is the intersection between sets $A$ and
$B$, can be defined as a derived constraint as follows:
\[
\Cap(A,B,C) \defs
\Cup(C,N_1,A) \land \Cup(C,N_2,B) \land N_1 \parallel N_2
\]
This approach is good from a theoretical perspective because it keeps the
language, proofs and implementation to a minimum. However, it pays the price of
reduced efficiency which, in the end, makes the tool less interesting from a
practical perspective. Therefore, we move some key set constraints from derived
constraints to \emph{built-in} constraints by defining and implementing
possibly recursive rewriting procedures for them. Specifically, we select
$\Cap$, $\subseteq$, and $diff$ (for set difference) to be implemented as
built-in constraints. The main new rewrite rules for these constraints can be
found in an on-line document \cite{calculusBR}).
\paragraph{Inference rules.}
In order to further improve the efficiency of our solver we introduce
special rewrite rules---hereafter simply called \emph{inference rules}---that
allow new $size$ and integer constraints to be inferred from the irreducible
constraints. The presence of these additional constraints will allow the solver
to detect more efficiently certain classes of unsatisfiable formulas.
Some significant inference rules are shown in Figure \ref{fig:infrule}.
\begin{figure}
\hrule\vspace{3mm}
\raggedright
\quad\quad If $X$ is any of $\dot{A}_i$; $m$ is any of $\dot{m}_i$; $\dot{m}_i$ is the cardinality of
$\dot{A}_i$; then:
\begin{flalign}
\quad\quad \quad\quad & \Cup(\dot{A}_1,\dot{A}_2,\dot{A}_3) \land size(X,\dot{m}) \longrightarrow
\Cup(\dot{A}_1,\dot{A}_2,\dot{A}_3) \land \dot{m}_3 \leq
\dot{m}_1 + \dot{m}_2 \bigwedge_{i=1,2,3}size(\dot{A}_i,\dot{m}_i) & \label{rule:infer_un} \\[2mm]
\quad\quad \quad\quad & \Cap(\dot{A}_1,\dot{A}_2,\dot{A}_3) \land size(X,\dot{m}) \longrightarrow
\Cap(\dot{A}_1,\dot{A}_2,\dot{A}_3) \bigwedge_{i=1,2,3}size(\dot{A}_i,\dot{m}_i) \bigwedge_{i=1,2} \dot{m}_3 \leq \dot{m}_i & \label{rule:infer_inters}
\end{flalign}
\hrule
\caption{Rule scheme for $size$ inference rules}
\label{fig:infrule}
\end{figure}
\begin{comment}
\begin{figure}
\hrule\vspace{3mm}
\raggedright
\quad\quad If $X$ is any of $\dot{A}$, $\dot{B}$ and
$\dot{C}$; $\dot{m}_a$, $\dot{m}_b$ and $\dot{m}_c$ are the cardinalities of
$\dot{A}$, $\dot{B}$ and $\dot{C}$, respectively, then:
\begin{flalign}
\quad\quad \quad\quad & \Cup(\dot{A},\dot{B},\dot{C}) \land size(X,\dot{m}) \longrightarrow
\Cup(\dot{A},\dot{B},\dot{C}) \land size(X,\dot{n}) \land \dot{m}_c \leq
\dot{m}_a + \dot{m}_b & \label{rule:infer_un} \\[2mm]
\quad\quad \quad\quad & \Cap(\dot{A},\dot{B},\dot{C}) \land size(X,\dot{m}) \longrightarrow
\Cap(\dot{A},\dot{B},\dot{C}) \land size(X,\dot{n}) \land \dot{m}_c \leq
\dot{m}_a \land \dot{m}_c \leq \dot{m}_b & \label{rule:infer_inters}
\end{flalign}
\hrule
\caption{Rule scheme for $size$ inference rules}
\label{fig:infrule}
\end{figure}
\end{comment}
\begin{example}
Proving a formula such as $ B = A_1 \cup \dots \cup A_{20} \land
\sum_{i=1}^{20} \card{A_i} < \card{B} $ which can be easily written in $\{log\}$\xspace
by using $\Cup$, $size$, $=$ and $<$ constraints, would cause an exponential
explosion in $\mathsf{solve\_size}$. Instead, by implementing the first inference rule
shown in Figure \ref{fig:infrule} the unsatisfiability is found in a few
milliseconds. In fact, the introduction of this rule eliminates the exponential
explosion for this class of formulas. \hfill$\Box$
\end{example}
Hence, we extend Algorithm \ref{glob} by properly adding new calls to the
inference rules inside $\mathsf{solve\_size}$, just before starting the third phase of
$\mathit{SAT}_{Za}$. If $\Phi_1$ is the formula received by $\mathsf{solve\_size}$ and ${\Phi_1}'$
the one obtained from $\Phi_1$ after applying the inference rules, then CLP(Q)
is called on the integer subformula of ${\Phi_1}'$. If CLP(Q) fails, then the
whole computation fails and the input formula is unsatisfiable; if not, the
third phase of $\mathit{SAT}_{Za}$ is started with $\Phi_1$.
\subsection{\label{empirical}Empirical evaluation}
In this section we present the results of the empirical evaluation we conducted
in order to evaluate how well the implementation of $\mathit{SAT}_{\CARD}$ in $\{log\}$\xspace
performs in practice. In previous papers, we have evaluated other aspects of
$\{log\}$\xspace such as its efficiency in producing model-based test cases
\cite{CristiaRossiSEFM13}; how well it deals with relational constraints
\cite{DBLP:journals/jar/CristiaR20} and restricted intensional sets
(\citeANP{DBLP:conf/cade/CristiaR17}
\citeyearNP{DBLP:conf/cade/CristiaR17,DBLP:journals/jar/CristiaR21a}); and we
have applied it to industrial-strength case studies such as the Bell-LaPadula
security model \cite{DBLP:journals/jar/CristiaR21} and the Tokeneer project \cite{Cristia2021}.
The empirical evaluation consists of two experiments where $\{log\}$\xspace is asked to
determine the satisfiability of a collection of $\mathit{L}_{\CARD}$ formulas. We measure
how many of those formulas are solved and the time spent in doing so. In both
experiments we use a 2 s timeout and the computing times are those of the
solved problems. The data set to reproduce these experiments can be downloaded
from \url{http://people.dmi.unipr.it/gianfranco.rossi/SETLOG/size.zip} (the
technical details can be found in \ref{app:experiments}). These experiments do
not use nested sets.
As shown in Table \ref{t:eval}, the first experiment is performed over a
collection of 468 $\mathit{L}_{\CARD}$ formulas. These formulas are taken from different
sources:
\begin{itemize}
\item
\textsc{Tests}. These are simple cardinality formulas of our own.
%
\item
\textsc{Properties}. These are formulas related to typical cardinality
properties such as $\card{A \cup B} \leq \card{A} + \card{B}$.
%
\item
\textsc{CVC4}. These are problems used by \citeN{Bansal2018} as a benchmark for
the implementation of cardinality constraints in the CVC4 SMT solver plus
problems derived from these.
%
\item
\textsc{Kuncak}. These are the five examples of program verification used by
\citeN{Kuncak2006} to show their algorithm that solves \textsf{BAPA} formulas.
\textsf{BAPA} is discussed in Section \ref{related}.
%
\item
\textsc{ssl-reachability}. This is the collection of problems used by
\citeN{Piskac2020} to evaluate their method based on a $\mathrm{LIA}^*$
encoding.
$\mathrm{LIA}^*$ is briefly discussed in Section \ref{related}.
\end{itemize}
\begin{table}
\caption{\label{t:eval}Results of the first experiment}
\begin{tabular}{lrrrrrrr}
\hline \multicolumn{1}{c}{\textsc{Collection}} &
\multicolumn{1}{c}{\textsc{\#}} & \multicolumn{2}{c}{\textsc{Satisfiable}} &
\multicolumn{2}{c}{\textsc{Unsatisfiable}} & \multicolumn{1}{c}{\textsc{\%}} &
\multicolumn{1}{c}{\textsc{Time}} \\
&
&
\textsc{Slvd} &
\textsc{Uslvd} &
\textsc{Slvd} &
\textsc{Uslvd} &
& \\\hline
\textsc{Tests} & 150 & 98 & 0 & 52 & 0 & 100 & 0.5 s \\
\textsc{Properties} & 53 & 14 & 0 & 36 & 3 & 94 & 3.8 s \\
\textsc{CVC4} & 20 & 8 & 0 & 12 & 0 & 100 & 2.5 s \\
\textsc{Kuncak} & 5 & 0 & 0 & 5 & 0 & 100 & 0.0 s \\
\textsc{ssl-reachability} & 240 & 130 & 13 & 90 & 7 & 92 & 19.1 s \\\hline
\textsc{Total} & 468 & 250 & 13 & 195 & 10 & 95 & 25.9 s \\\hline
\end{tabular}
\end{table}
As can be seen, $\{log\}$\xspace solves 95\% of the problems in 25.9 s, meaning an
average of 0.06 s per problem. Even if the first collection is not considered,
$\{log\}$\xspace solves 93\% of the resulting 318 problems in 25.4 s, thus making 0.09 s
per problem. In particular, $\{log\}$\xspace solves all the problems in the
\textsc{CVC4} and \textsc{Kuncak} collections. It also solves 92\% of the
\textsc{ssl-reachability} collection in 19.1 s (0.09 s on average) whereas
Piscak et al. manage to solve 76\% of them in 59 s (0.3 s in
average)\footnote{Piscak et al. run their evaluation on a 2018 MacBook Pro
running OS X Mojave 10.14.5 with a 2.9 GHz Intel Core i9 processor and 32GB of
RAM. Our hardware platform is older and less powerful, see below.} \cite[Table
1]{Piskac2020}. If the timeout is set to 50 s, as done by Piscak, $\{log\}$\xspace
manages to solve 11 more problems thus solving 96\% of them (although it needs
considerably more time as some problems are solved only after several seconds).
The second experiment concerns the evaluation of $\mathit{SAT}_{\CARD}$ when computing
minimal solutions---cf. Section \ref{minimal} and command \verb+fix_size+ given
in Section \ref{impl}. Then, we run $\{log\}$\xspace on the 250 satisfiable problems of
Table \ref{t:eval} that the tool is able to solve. The results are given in
Table \ref{t:evalfixsize}. This experiment sheds some light on the efficiency
of $\{log\}$\xspace in constructing more concrete solutions of satisfiable problems. As
can be seen, $\{log\}$\xspace is able to produce a more concrete solution to 99\% of the
satisfiable problems in 0.07 s on average. Note that the tool is not able to
find a concrete solution for three formulas whose satisfiability, nonetheless,
it was able to ascertain.
Even if the first collection of problems is removed from this experiment,
$\{log\}$\xspace solves 99\% of the problems in 0.1 s on average.
\begin{table}
\caption{\label{t:evalfixsize}Results of the second experiment}
\begin{minipage}{.65\textwidth}
\begin{tabular}{lrrrrrrr}
\hline \multicolumn{1}{c}{\textsc{Collection}} &
\multicolumn{1}{c}{\textsc{\#}} & \multicolumn{2}{c}{\textsc{Satisfiable}} &
\multicolumn{1}{c}{\textsc{\%}} &
\multicolumn{1}{c}{\textsc{Time}} \\
&
&
\textsc{Slvd} &
\textsc{Uslvd} &
& \\\hline
\textsc{Tests} & 98 & 97 & 1 & 99 & 0.4 s \\
\textsc{Properties} & 14 & 14 & 0 & 100 & 0.1 s \\
\textsc{CVC4} & 8 & 8 & 0 & 100 & 0.7 s \\
\textsc{ssl-reachability} & 130 & 128 & 2 & 98 & 15.3 s \\\hline \textsc{Total} & 250 & 247 & 3 & 99 & 16.5 s \\\hline
\end{tabular}
\end{minipage}
\end{table}
\subsection{\label{discussion}Discussion}
In spite of initial theoretical concerns, the empirical evaluation presented in
Section \ref{empirical} shows that, in practice, $\{log\}$\xspace's implementation of
$\mathit{SAT}_{Za}$ performs no worse than other approaches and better than special
purpose algorithms such as those by Kuncak and Piscak. It is true, however,
that in the worst case the exponential complexity of the algorithm makes it
unfit for certain problems. We can see that in the unsolved problems (23 out of 468) of Table
\ref{t:eval}.
Broadly speaking, $\{log\}$\xspace's implementation of $\mathit{SAT}_{\CARD}$ goes through three
phases: \emph{a)} solve the formula with minimal concern about cardinality;
\emph{b)} compute the set of solutions of a Boolean formula derived from the
irreducible form (cf. Definition \ref{def:solved}); and \emph{c)} solve an
integer linear programming problem \emph{for each} subset of the Boolean
solutions, which presupposes the powerset of the set of Boolean solutions being
computed. Each phase of $\mathit{SAT}_{\CARD}$ is inherently exponential, at least, in the
worse case.
However, according to our experiments, the worst of these three problems is
\emph{c)}. Its most demanding part is not the computation of the powerset
itself but solving the integer problem for each of its elements. In fact,
$\{log\}$\xspace uses backtracking in such a way as to avoid computing the powerset
explicitly. This problem bears some relationship with the number of set
variables of the input formula, but this is neither evident nor direct. For
example a formula such as $A_1 \cup \dots \cup A_{50} = \emptyset \land
\card{A_{43}} > 2*k + 5$ is solved in virtually no time, while a formula with
fewer variables but where $\cup$ is substituted by $\cap$ will take an
exponential time. As we have noted, the real problem is the number of solutions
returned by step \emph{b)} which determines the size of the powerset.
Unfortunately, the relationship between the input formula and the number of
solutions of the Boolean problem is complex. For example, $A_1 \cup \dots \cup
A_{50} = B$ will generate many more Boolean solutions than $A_1 \cup \dots \cup
A_{50} = B \land \bigwedge_{i=1}^{49} A_i \parallel A_{i+1}$. To worsen things, if
the number of set variables is large, the integer problem to be solved for each
element of the powerset becomes increasingly more complex, consuming a non
negligible time. On the other hand, a palliative to deal with \emph{c)} is the
fact that the problem is inherently parallelizable.
The introduction of inference
rules proved to be a good method to avoid many of the exponential problems we
have discussed above. As long as the application of inference rules remains
polynomial in the size of the formula received by $\mathsf{solve\_size}$, it will be, on
average, better to add them than not. It remains as an open problem whether or
not there is a set of inference rules applicable in polynomial time
constituting a decision procedure for $\mathit{L}_{\CARD}$. We believe the answer is no.
\section{\label{related}Related work}
Computable Set Theory (CST) has studied the problem of deciding the
satisfiability of set formulas involving cardinality constraints since a long
time ago \cite{DBLP:conf/cade/FerroOS80}\cite[Chapter
11]{DBLP:series/mcs/CantoneOP01}. In these works cardinality formulas are
encoded as additive arithmetic formulas over the natural numbers.
\citeN{hibti1995} proves the decidability of a similar problem by encoding it
as a propositional consistency problem.
Zarba's work is rooted in CST and thus relies on the notion of \emph{place} as
a way to represent Venn regions. This notion is used only inside $\mathsf{solve\_size}$.
Zarba also proves that a theory of multisets, without the cardinality operator,
is decidable \cite{DBLP:conf/cade/Zarba02}. Later on, Zarba proved that a
theory of (not necessarily finite) sets, including the cardinality operator,
combined with a theory of cardinal numbers is decidable
\cite{DBLP:journals/jar/Zarba05}.
In the field of Constraint Logic Programming a number of proposals have been
put forward introducing \emph{set constraints}, possibly including cardinality
\cite{DBLP:journals/constraints/Azevedo07,DBLP:journals/constraints/Gervet97,DBLP:journals/jair/HawkinsLS05}.
In these proposals, constraint (set) variables have a \emph{finite domain}
attached to them, which is exploited by the solver to efficiently compute
simplified forms of the original constraints or to detect failures. The same
approach is adopted in the constraint modeling language MiniZinc
\cite{minizinchandbook}. While the availability of finite domains for
constraint variables allows efficient handling of set constraints, it actually
prevents the user from using the solver as a general theorem prover. On the contrary, this is
feasible in $\{log\}$\xspace where constraint variables do not require finite domains.
For example, proving the property $\forall A,B,n: A \subseteq B \land
\card{A}=n \land \card{B}=n \implies A = B$, can be done in $\{log\}$\xspace by checking
that the formula
\verb!subset(A,B) & size(A,N) & size(B,N) & A neq B!
is unsatisfiable. The same general result cannot be achieved for instance in
MiniZinc, since set variables $A$ and $B$ (declared as ``decision variables''
in MiniZinc) must have a fixed domain attached to them---e.g.,
\verb!var set of 0..100: A!.
Thus, we can write the formula in MiniZinc but what we prove is not as general
as in $\{log\}$\xspace: if we get an \verb!UNSATISFIABLE! answer from MiniZinc it does
not mean we have proved the (general) property, while in $\{log\}$\xspace it does.
Furthermore, set elements in $\{log\}$\xspace can be of any type, including unbounded
constraint variables and other sets, which are not allowed in MiniZinc and in
other related proposals for set constraints.
V. Kuncak and his colleagues have worked on the decidability of the first-order
multisorted theory \textsf{BAPA} and its applications to program verification
\cite{Kuncak2006}. \textsf{BAPA} extends the combination of the theory of
Boolean algebras of sets (BA) and Presburguer arithmetic (PA). In this way
\textsf{BAPA} can deal with formulas where the cardinality of a set is treated
as an integer variable subjected to PA constraints. Kuncak's algorithm reduces
a \textsf{BAPA} sentence to an equivalent \textsf{PA} sentence. In this way,
the algorithm enjoys several nice properties (e.g., its complexity is no worse
than an optimal algorithm for deciding \textsf{PA}). This implies that the
complexity of Kuncak's algorithm is identical to the complexity of PA. Besides,
the algorithm can eliminate quantifiers from a \textsf{BAPA} formula thus
turning this into a quantifier-free \textsf{BAPA} formula---called
\textsf{QFBAPA}. The algorithm depends upon \textsf{MAXC}, an integer constant
denoting the size of the finite universe. Our method does not depend on any
constant denoting the size of the universe. Kuncak and his colleagues have
implemented this algorithm in the Jahob system, used to check the consistency
of data structures in the Java language. Kuncak shows a few problems related to
program verification that can be solved with his algorithm. All the problems
proposed by Kuncak can also be efficiently solved by $\{log\}$\xspace as is shown in
Section \ref{empirical}.
In a further development, \citeN{Piskac2008} give a decision
procedure for multisets with cardinality constraints by using a similar method
(i.e., encoding input formulas as quantifier-free PA formulas); more recently
a more efficient method based on a LIA$\star$ encoding has been proposed
\cite{Piskac2020,Levatich2020}. These algorithms have been implemented in the
MUNCH \cite{Piskac2010} and ssl-reachability \cite{Piskac2020} tools which use
existing solvers to solve the various problems involved in this approach,
e.g., linear integer arithmetic. The empirical evaluation used to evaluate the
ssl-reachability tool is included in the evaluation of the implementation of
our algorithm in $\{log\}$\xspace (cf. Section \ref{empirical}).
\citeN{Suter2011} have extended the Z3 SMT solver to
solve problems of the \textsf{QFBAPA} logic which, as said above, can be used
to encode set problems combined with PA problems through the cardinality
operator. \citeN{Bansal2018} also approach the problem of deciding
the satisfiability of finite set formulas with cardinality in the context of
SMT solvers. They propose and implement in CVC4 a calculus describing a
combination of a procedure for reasoning about membership with a procedure for
reasoning about cardinality. Their method is based on a different strategy
w.r.t. to Suter's work but it draws the concept of \emph{place} from CST
although used in an incremental way. According to Bansal and his colleagues,
Suter's method cannot scale well when the formula has set membership
constraints because these are encoded as cardinality constraints (i.e., $x \in
A \iff \{x\} \subseteq A$ and $\{x\}$ is actually a set whose cardinality is
1). Instead, they propose to avoid dealing with set membership constraints in
terms of \emph{places} or Venn regions, but to reason directly about
membership. This is aligned with how our method deals with set membership,
although we do it in terms of set unification \cite{Dovier2006}. In fact, in our
method a formula such as $x \in B \cup C$ is written as $\Cup(B,C,A) \land x
\in A$ which in turn is rewritten as $A = \{x \plus N\} \land \Cup(B,C,\{x
\plus N\})$, where $N$ is a new variable (implicitly existentially quantified)
and $\{x \plus N\}$ is a set constructor interpreted as $\{x\} \cup N$. No Venn
regions are computed when this formula is solved. Bansal et al. empirically
evaluate their method on 25 problems on program verification. The first 15 of
these problems are drawn from the evaluations performed by Kuncak and Suter on
their tools. CVC4 shows a comparable performance w.r.t. those other tools.
These 15 problems are included in the empirical evaluation of our method
reported in Section \ref{empirical}; $\{log\}$\xspace also shows a comparable
performance. Bansal et al. also compare their method with Suter's on the
constraint $x \in A_1 \cup \dots \cup A_{21}$. As expected, Suter's method runs
out of memory after some time while CVC4 solves the formula immediately.
$\{log\}$\xspace also solves the formula quickly and is able to return a finite
representation of all possible solutions which, as far as we know, no other
tool can do. $\{log\}$\xspace also supports nested sets which is apparently not the case
of CVC4.
\citeN{DBLP:conf/vmcai/YessenovPK10} prove the
decidability of a theory of sets including functions, $n$-ary relations and
some operators for the algebra of relations (e.g., relational image). Then, they
show that the cardinality operator can be added to the theory preserving its
decidability.
\citeN{DBLP:journals/constraints/Azevedo07} describes the
\textit{Cardinal} system which is part of the ECLiPSe Prolog library.
\textit{Cardinal} is based on constraint propagation on set cardinality and set
interval reasoning. Methods of this kind are, in general, restricted to
formulas where the
cardinality of each set is constrained to range over a closed integer interval.
Azevedo applies his method to some problems on digital circuits.
A proposal for extending $\{log\}$\xspace with integers and cardinality constraints had
already been put forward in a previous work \cite{Palu:2003:IFD:888251.888272}.
In that case, however, the extension is based on the integration of CLP(FD)
into $\{log\}$\xspace. Consequently, completeness of the solver is obtained only if
finite domains are provided for all integer variables and labelling is
performed over them. This in fact implies an upper limit for set cardinalities.
Furthermore, the presence of labeling can easily lead to unacceptable
performance.
\citeN{Alberti2017} extend linear integer arithmetic with free function symbols
and cardinality constraints for interpreted sets. Interpreted sets are sets of
the form $\{x \in [0,N)| \varphi\}$, for some $0 < N \in \nat$, and $\varphi$
is an arithmetic formula. Free unary function symbols are used to represent
array ID's. Thus, the language offers terms of the form $a(y)$ where $a$ is an
array ID and $y$ is a variable. Formulas such as $a(y) < 1$ are allowed to
occur in interpreted sets where $y$ is the bound variable. Then, the language
\emph{only} allows one to indicate the cardinality of interpreted sets, e.g.,
$\card{\{y \in [0,N) | a(y) < 1\}} = 0$. These authors prove that some
fragments of this logic are both decidable and expressive enough as to model
and reason about problems of fault-tolerant distributed systems. The
decidability results are obtained by mapping those fragments into Presburger
arithmetic enriched with unary counting quantifiers. One of the decidable
fragments has been implemented in a tool that uses the Z3 SMT solver as a
back-end solver for quantifier-free linear arithmetic. Alberti's logic does not
include classic set theoretic operators such as union. Hence, it is difficult
to compare the expressiveness of Alberti's logic with other logics analyzed in
this section and with ours. Although $\{log\}$\xspace's intensional sets
\cite{DBLP:journals/jar/CristiaR21a} could be used to encode Alberti's
interpreted sets, it is still necessary to extend that theory as to compute the
cardinality of intensional sets. This is a line of future research.
\citeN{Bender2017} extend some of the previous
results to theories where cardinalities are replaced by the more general notion
of measures. In this case a key aspect of the previous approaches is no longer
valid, namely the fact that only the empty set has cardinality equal to 0, as
there are non-empty sets with measure 0. The theories analyzed by these authors
are important in, for example, duration calculus.
Also the Artificial Intelligence community has studied the problem of reasoning
about the size of sets, e.g.,
\cite{ding_harrison-trainor_holliday_2020,DBLP:conf/aaai/KisbyBKM20}. We want
to remark the work by \citeN{DBLP:conf/aaai/KisbyBKM20} because
they propose two logics, combining sets with cardinality, whose decidability
can be solved in polynomial time. As expected, the gain in complexity is at the
cost of expressiveness. Nonetheless, the result may deserve being studied in
terms of software verification as it might give clues about what are the
simplest specifications and proof obligations involving sets and cardinality.
From there, compositional methods might be drawn in order to tame the
complexity constantly faced in automated program verification.
\section{\label{concl}Concluding Remarks}
In this paper we have presented a decision
procedure for the algebra of hereditarily finite hybrid sets extended with
cardinality constraints. The proposed
procedure is implemented within $\{log\}$\xspace, a CLP system able to deal with a few
decidable fragments of set theory. The empirical evaluation carried out on the
implementation proves that $\{log\}$\xspace is able to deal efficiently with formal
verification problems involving cardinality constraints.
As a future work, we plan to use this decision procedure as the base for a
decision procedure for the algebra of finite sets extended with integer
intervals. Indeed, the following identity:
\begin{equation*}
A = [m,n] \iff A \subseteq [m,n] \land \card{A} = n-m+1
\end{equation*}
becomes the key for a set unification algorithm including integer intervals
with \emph{variable} limits. In fact, it would suffice to be able to deal with
constraints of the form $A \subseteq [m,n]$ in a decidable framework to have a
decision procedure for integer intervals. In turn, integer intervals are a key
component in the definition of arrays as sets. In fact, if $array(A,n)$ is a
predicate stating that $A$ is an array of length $n$ whose components take
values on some universe $\mathcal{U}$, then it can be defined as follows:
\begin{equation*}
array(A,n) \iff A: [1,n] \fun \mathcal{U}
\end{equation*}
$\{log\}$\xspace already supports a broad class of set relation algebras
(\citeANP{DBLP:journals/jar/CristiaR20}
\citeyearNP{DBLP:journals/jar/CristiaR20,DBLP:conf/RelMiCS/CristiaR18}),
including partial functions and the domain operator. Hence, it would be
possible to use $\{log\}$\xspace to automatically reason about broad classes of programs
with arrays from a set theoretic perspective which would be different from
existing approaches
\cite{DBLP:conf/lics/StumpBDL01,DBLP:conf/vmcai/BradleyMS06}.
\bigskip
\noindent\textit{Competing interests: The authors declare none}
\bigskip
\bibliographystyle{acmtrans}
|
\section{Introduction}
\label{sec:introduction}
Money laundering is a serious financial crime that consists of the illegal process of obtaining money from criminal activities, such as drug or human trafficking, and making it appear legitimate. Cryptocurrencies, such as Bitcoin \cite{nakamoto2008peer}, are particularly susceptible to money laundering schemes due to their pseudo-anonymity and the relative lack of control by authorities. Preventing money laundering is an international effort and Anti-Money Laundering (AML) laws have been trying to cope with the new threats posed by criminals using cryptocurrencies \cite{union2018a, network2019a}.
In 2019, \citet{weber2019antimoney} released the Elliptic data set. It contains anonymized labeled Bitcoin transactions and enables researchers to study illicit behaviour in cryptocurrencies. The data set consists of a time-series graph with 200K labeled bitcoin transaction nodes and tabular data with 166 anonymized features describing each transaction. \citet{weber2019antimoney} assesses the performance of several supervised learning algorithms on the task of detecting nodes associated with illicit activities.
To improve existing supervised learning results found in the literature, we propose a new set of features that leverage the structure of the network and the existence of hubs or pockets of illicit transactions. We extract these new features with \emph{GuiltyWalker}. This random walker traverses a given network starting from a seed node and computes features related to the distance of the seed node to other nodes known to be illicit.
GuiltyWalker consists of two main components:
\begin{itemize}[leftmargin=0.5cm]
\item \textbf{Random walker}: Given a transaction graph, a set of seed nodes, and the number of desired random walks for each of the seeds, GuiltyWalker samples random walks for each seed node. Due to the temporal nature of the graph, the walker only travels to past nodes (i.e., transactions) and stops at the first illicit node found or when there are no more valid nodes to visit.
\item \textbf{Feature extractor}: Given a set of random walks for each of the seeds, GuiltyWalker computes aggregated features that summarize these walks, e.g., the average number of steps needed to reach one illicit node or the total number of different illicit nodes found.
\end{itemize}
In our experiments on the Elliptic data set, we observe that adding the features computed by GuiltyWalker improves the performance of machine learning methods; namely, we achieve a 5pp increased performance in F1-score, when compared against machine learning methods that use only the original anonymized features from \citet{weber2019antimoney}. Furthermore, the gains in performance are more pronounced during a black market shutdown, where the original performance by \citet{weber2019antimoney} dropped significantly.
This paper is organized as follows. Section 2 details the implementation of GuiltyWalker and how we generate new features from its output to enrich the data and convey additional information to supervised learning methods. Section 3 describes the experimental setup, and Section 4 the results consequently obtained. Section 5 presents the related work. Finally, we set out the main conclusions in Section 6.
\section{GuiltyWalker}
\label{sec:guiltywalker}
GuiltyWalker consists of a random walker that traverses a given transaction network from a seed node and extracts features based on the distance of that node to known illicit nodes. It includes two main components -- a random walker and a feature extractor, explained in the following subsections.
\subsection{Random Walker}
\label{subsec:random_walker}
The random walker receives as input the original transaction graph $G$, a list of seed nodes, $S \subset V(G)$, and the number of \emph{successful} random walks desired, $k$. Successful random walks are explained later in this section. GuiltyWalker's output for each seed node $s \in S$ is the list of sampled random walks $\mathcal{X}^s = \{X_{1}^s, X_{2}^s, ..., X_{k}^s\}$.
A random walk $X_i^s$ consists of a sequence of nodes
\begin{equation}
X_i^s = (x_1, x_2,...,x_n)\;,
\end{equation} such that $x_1 = s$. Due to the temporal nature of the transaction graph, the random walker can only walk backward in time. That is, it is only valid to go from node $x_n$ to node $x_m$ if $x_m$ represents a transaction older than $x_n$. This transaction network is represented as a directed graph connecting older nodes to newer nodes by an outgoing edge. Then, to address the former condition, during the random walk process, GuiltyWalker chooses a node uniformly at random from the incoming neighbors of the current node. When GuiltyWalker is in a given state of a random walk $X^s_i = (s, x_2,..., x_n)$, the process stops and returns $X^s_i$ as the final random walk if at least one of the following criteria is met:
\begin{itemize}[leftmargin=0.5cm]
\item $x_n$ is a known illicit node/transaction.
\item The set of eligible nodes to pick from is empty. This scenario happens when a given node has no incoming neighbors and, consequently, the random walker has no possible moves.
\end{itemize}
Otherwise, GuiltyWalker randomly picks the next node to add to $X^s_i$, and the process continues.
Note that since the edges only connect older transactions to newer transactions, there is always an end node in any random walk. In other words, the properties of our transaction graph guarantee that GuiltyWalker will not be stuck in an endless loop.
The number of \emph{successful} random walks, given by the user as input, intends to set the desired number of random walks ending in an illicit node from each seed node $s \in S$. However, as discussed before, the random walker may find a node with no incoming neighbors. In this case, a random walker finishes traversing the graph without reaching an illicit node. To ensure the number of desired successful random walks, GuiltyWalker performs as many random walks as needed, and only those are used for feature extraction. As we mention in Section~\ref{subsec:features_computation}, one of the features extracted from GuiltyWalker is the fraction of \emph{successful} random walks from the total number of random walks needed to perform to ensure the number of successful ones. This is a way of also considering the number of unsuccessful walks made from each node, which may be informative.
It is important to note that some nodes in a transaction graph might have no paths to any illicit node. Thus, it is impossible to obtain successful random walks (as per our definition) for those nodes. To avoid this problem, GuiltyWalker first determines which nodes actually can reach an illicit node.
To do that, we first invert the direction of the graph. Then, we use the descendants algorithm for directed acyclic graphs from NetworkX \cite{SciPyProceedings_11} \cite{descendants}. It returns all nodes reachable from a source node $s$ in the graph $G$. Afterwards, we inspect if at least one of them is illicit. This procedure is made for all nodes in the transaction graph, and only those that can reach an illicit node are given as input for the random walker.
\subsection{Features Computation}
\label{subsec:features_computation}
The second component of GuiltyWalker receives the list of random walks from each seed node and returns a data frame of features corresponding to each transaction, summarizing the random walks. In particular, GuiltyWalker obtains the following features:
\begin{itemize}[leftmargin=0.5cm]
\item Minimum size of the random walks (min);
\item Maximum size of the random walks (max);
\item Mean size of the random walks (mean);
\item Standard deviation of the random walks sizes (std);
\item Median size of the random walks (median);
\item First quartile of the random walks sizes (q25);
\item Third quartile of the random walks sizes (q75);
\item Fraction of \emph{successful} random walks from all the random walks performed by Random Walker (hit rate);
\item Number of distinct illicit nodes in the random walks (illicit).
\end{itemize}
We also add information about the transaction nodes with no possible paths to fraudulent nodes to the data frame of features, with all features set accordingly (see Section~\ref{subsec:methodology}), due to the lack of information regarding the distance to an illicit node.
\section{Experimental Setup}
\label{sec:experimental_setup}
\subsection{Elliptic Data Set}
\label{subsec:elliptic_data_set}
In this work, we use the Elliptic Data Set, a graph network of Bitcoin transactions\footnote[2]{\footnotesize Available at \url{https://www.kaggle.com/ellipticco/elliptic-data-set}}. Elliptic, a company focused on combating financial crime in cryptocurrencies, released this data set.
The data set includes 203,769 node transactions and 234,355 directed edges, representing the flow of Bitcoin currency (BTC) going from one transaction to the next. Each transaction can be categorized into three classes: "licit", "illicit" or "unknown", based on the category of the entity that created it. Licit categories include exchanges, wallet providers, miners, and financial service providers. Illicit categories include scams, malware, terrorist organizations, and Ponzi schemes. From the total number of transactions, 21\% (42,019) are labeled as licit, 2\% (4,545) as illicit, and the remaining 77\% (157,205) are unknown.
Besides the graph structure, the data set has 166 anonymized features associated with each transaction. The first 94 relate to information about the transaction itself, such as the time step, number of inputs/outputs, and transaction fee. The remaining features relate to aggregated information about the direct neighbors of the transaction, giving the maximum, minimum, standard deviation, and correlation coefficients of each transaction.
Besides, a time step from 1 to 49 is associated with each node. It represents an estimate of when the Bitcoin network confirmed the transaction. The time steps are evenly spaced with an interval of about two weeks and each one contains a single connected component of transactions that appeared on the blockchain within less than three hours between each other. Therefore, it can be considered that this data set includes 49 directed acyclic graphs associated with different sequential moments in time. Figure \ref{fig:structuredataset} provides an idea of the structure of this data set.
\subsection{Methodology}
\label{subsec:methodology}
This section gives an overview of the models used in our experiments and discusses our experimental setup. Following \citet{weber2019antimoney}, we perform a 70/30 temporal split of training and test data, respectively, for all experiments. Therefore, the train set includes all labeled samples up to the 34\textsuperscript{th} time step, and the test set includes all labeled samples from the last 15 time steps, up to the 49\textsuperscript{th}.
We use random forest for licit versus illicit prediction. First, we train the model on the train set using all 166 features and evaluate it on the entire test set. We use the scikit-learn \cite{pedregosa2011scikit} implementation of random forest, with 50 estimators, corresponding to the number of trees in the forest, and 50 max features, defined as the maximum number of features each tree can have. By doing so, we mimic the method in \citet{weber2019antimoney} enabling a fair comparison of the results. We also set the random state to 0 for the purpose of results reproducibility.
Then, we train a random forest model (using the same parameters as before) using (i) only the new set of features obtained by GuiltyWalker and (ii) both the features obtained by GuiltyWalker and the original 166 features. We extract the GuiltyWalker features after performing 100 \emph{successful} random walks. Missing values for the transaction nodes that cannot reach an illicit node are filled with -1 values, except feature \textit{hit}, which is filled with 0 values, as it represents the fraction of random walks ending in a fraudulent node. We see that the utilization of some of these alternative sets of features improves performance in Section \ref{sec:results}.
To further improve the results, we filter the set of features obtained by GuiltyWalker to keep just the most important ones. This method characterizes every single feature's importance as the decrease in the performance score after randomly shuffling its position in the set, and is called Permutation feature importance \cite{featureimportance}. After applying this method and assessing every feature's importance, the new features kept for further classification purposes are \textit{hit}, \textit{std}, \textit{illicit}, \textit{max} and \textit{mean}. We also analyse the model's performance with these features together with the 166 original ones in Section \ref{sec:results}.
Similarly to \citet{weber2019antimoney}, we evaluate the random forest classifier's performance with each set of features using the F1-score for the illicit class, hereafter referred to as illicit F1-score. This score is the weighted average of precision and recall. Moreover, it is suitable for imbalanced tasks, which is the case of our dataset (91\% of licit nodes and 9\% of illicit ones). We also use the ROC curve (and AUC value) and precision and recall measures to evaluate the models' performance.
\begin{figure}[t!]
\center
\includegraphics[width=0.45\textwidth]{graph.png}
\caption{Structure of the data set (taken from \cite{bellei_2019})}
\label{fig:structuredataset}
\end{figure}
\section{Results}
\label{sec:results}
In this section, we present the results obtained by using the standard model, random forest, with the 166 baseline features (referred to as AF), as well as only the new features extracted from GuiltyWalker (referred to as GWF) and the former ones together with the latter (referred to as AF+GWF). Furthermore, we show the results obtained using the 166 features in conjunction with the new ones obtained after performing feature reduction
(referred to as AF+GWF*).
\begin{table}[ht]
\caption{Illicit classification results using random forest, for different features. {\normalfont \textit{AF}} refers to the original all features, {\normalfont \textit{GWF}} refers features extracted from the GuiltyWalker, and, {\normalfont \textit{GWF*}} refers to GuiltyWalker features after feature selection.}
\label{tab1:f1score}
\vspace{5mm}
\resizebox{\columnwidth}{!}{%
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{lcccc}
\toprule
\multicolumn{1}{c}{} & & Illicit & \multicolumn{1}{c}{} & MicroAVG \\ \cline{2-4}
\multicolumn{1}{c}{Method} & Precision & Recall & \multicolumn{1}{c}{F1} & F1 \\ \midrule
AF & 0.91 & 0.72 & 0.80 & 0.977 \\
GWF & 0.93 & 0.11 & 0.20 & 0.942 \\
AF + GWF & 0.93 & 0.76 & 0.84 & 0.981 \\
AF + GWF* & 0.97 & 0.77 & 0.85 & 0.983 \\
\bottomrule
\end{tabular}
}
\end{table}
Table \ref{tab1:f1score} shows the testing results in terms of precision, recall and F1-score concerning the illicit class. For the sake of completeness, we also show the micro-averaged F1 score.
An important thing to note from Table~\ref{tab1:f1score} is that the GuiltyWalker features alone are not informative enough. The F1-score value obtained using only these features is very low (0.20). We can also observe higher precision, recall, and F1-score when using GuiltyWalker's additional features, suggesting the importance of the graph structure. Using GuiltyWalker features, we improved precision, recall, and F1-score by 2 percentage points (pp), 4pp, and 4pp, respectively.
In order to understand the importance of each one of the features created, we performed feature importance, using the method described in the previous section. We kept only the most important features to train together with the original ones. Results show that by filtering GuiltyWalker features and keeping only the most important ones (\textit{hit}, \textit{std}, \textit{illicit}, \textit{max} and \textit{mean}), the performance of the model slightly improves (we improved F1-score by 1pp, comparing with the model AF+GWF).
To give additional insights about the performance of the new model AF + GWF* compared against the original model, we plot the ROC curve of both models.
Note that the ROC curve shows the trade-off between sensitivity/ recall and specificity. Moreover, the area under the curve (AUC) can be seen as a measure of separability. In other words, it represents how much a model is capable of distinguishing between classes. Therefore, from the observation of Figure \ref{fig:roc}, we can infer that both models are quite good at predicting illicit nodes as illicit and licit ones as licit. However, AF + GWF* is slightly better (improves AUC value by 1pp). In particular, for very low false positive rates, our method seems to be significantly better.
In a real scenario, we would be more interested in low false-positive regions of the ROC-curve since raising too many alerts is not practical. With this in mind, we compare the recall at specific low false positive rates, namely 1\%, 5\% and 10\%, and AF+GWF* shows considerable gains when compared against AF: recall@1\% increases from 73\% to 78\% (5pp), recall@5\% increases from 75\% to 80\% (5pp), and recall@10\% increases from 76\% to 82\% (6pp). Therefore, while the gain of using GuiltyWalker's features is only 1pp in the full region, in the region of interest the gain is considerably higher.
\begin{figure}[t!]
\center
\includegraphics[scale=0.17]{roc.png}
\caption{ROC curve associated with AF + GWF* and AF models. Random baseline represented in red. }
\label{fig:roc}
\end{figure}
\begin{figure}[ht]
\center
\includegraphics[scale=0.17]{fsgwf.png}
\caption{Illicit F1-score obtained with Random Forest, for the standard 166 features and the new GuiltyWalker features before and after feature selection.}
\label{fig:fsgwf}
\end{figure}
As noted by \citet{weber2019antimoney}, a sudden dark market shutdown occurring at time step 43 extremely affects the model performance. In particular, the random forest model trained on the 166 features, from that time step forward, cannot achieve an Illicit F1-score value above 0.25. The introduction of the new set of features extracted from GuiltyWalker improves F1 results in the entire test set (i.e., time steps 35-49). However, this improvement is more notorious after this dark market shutdown (from time step 43 to 49). In fact, from time step 43 to time step 49, we observe, on average, a F1 score improvement of about 10pp and 16pp with AF + GWF and AF + GWF* models, respectively. Note that for time steps 48 and 49, both of these models still perform poorly.
As we can see in Figure \ref{fig:fsgwf}, both AF + GWF and AF + GWF* models are able to reliably capture new illicit transactions after the dark market shutdown, in comparison with the original model. To understand the additional information those models are capturing, we compute the confusion matrices of the models AF, AF+GWF and AF+GWF*. We obtain 784, 828 and 831 true positives (referred to as TP), respectively. We also determine the new TP found and the ones lost, using AF + GWF and AF + GWF* models, in comparison to the ones found training RF with the original set of features. By doing so, we can verify that with the new sets of features, AF+GWF and AF+GWF*, we found 48 and 50 new TP and lost 4 and 3 TP that the original model could find, respectively.
Concerning the AF + GWF* model, we observe that, for almost all new TP found, the features extracted from GuiltyWalker (\textit{max}, \textit{mean}, \textit{std}, \textit{illicit} and \textit{hit}) have positive values. Only 2 of the 50 elements have -1 values regarding \textit{max}, \textit{mean}, \textit{std} and \textit{illicit} and 0 with respect to \textit{hit}.
Recall that these values indicate that the associated transaction nodes have no possible paths to known illicit nodes.
This interesting fact lets us infer that the new set of features adds some new information based on the graph's structure, which allows the model to make better predictions. However, we have to notice that the fact that a given node has a path to an illicit transaction does not necessarily imply that it is also illicit and vice-versa. This information alone is not enough to make good predictions concerning the labels of the transaction nodes, as we verified from the results obtained for the GWF model. Nonetheless, it provides extra information to complement the original features in a way that boosts performance of the overall model.
\section{Related Work}
Besides the work of \citet{weber2019antimoney}, which was the baseline for our study, more recently, \citet{lorenz2020machine} proposed active learning techniques to study the minimum number of labels necessary to achieve high detection of illicit activity in cryptocurrencies and tested them also on the Elliptic data set. Thus, even though using a different approach to the same problem, the authors did not aim to achieve better results than the baseline. Moreover, \citet{alarab2020comparative} proposed an ensemble learning method, using a combination of the given supervised learning models, and applied it on the Elliptic data set, improving the baseline results. Although they improved upon existing results, our results, using the new set of features, are better. While \citet{alarab2020comparative} achieves higher precision than us (97.38\% versus 96.5\%), we achieve higher recall (76.7\% versus 72.2\%), higher F1-score (85.47\% versus 82.92\%), and higher accuracy (98.3\% versus 98.06\%).
As far as we know, previous work on the application of graph-related features and, in particular,
random walks, in a supervised learning setting are scarce. \citet{Hu2019CharacterizingAD} worked with Bitcoin transaction graphs and used various graph characteristics to differentiate money laundering transactions from regular transactions. They actually found that the main difference between them lies in their output values and neighbourhood
information. The authors also evaluated a set of classifiers based on different types of extracted features, namely immediate neighbours, curated features, deepwalk embeddings \citep{perozzi2014deepwalk}, and node2vec embeddings \citep{grover2016node2vec} to classify money laundering and regular transactions. This approach differs from ours as we are not trying to embed the graph or a particular node's neighbourhood but instead to describe distances to a specific target (i.e., malicious activity). Nonetheless, the descriptive power of random walks in networks is still recognized.
\citet{socialnetwork} studied methods based on the iterative application of traditional classifiers using graph information as features, and methods that propagate the existing labels via random walks. Moreover, concerning the application of random walks in the context of classification problems, \citet{textclassification} proposed a new approach for estimating term weights in a document based on a random walk model. They showed that the new random walk based approach outperforms the traditional term frequency approach of feature weighting. Therefore, with this work, we extend the existing knowledge regarding random walks to improve classifiers' performance in graph datasets.
\section{Conclusion}
\label{sec:conclusion}
In this study, we set out to improve the performance of supervised models in an anti-money laundering classification task. Given a transaction network, we propose a method called GuiltyWalker that extracts information from the structure of the network and the existence of past labels to create new features for a supervised model. It consists of a random walker that traverses the transaction network starting from a seed node and a feature extractor that computes features related to the distance of the seed node to other nodes known to be illicit.
We test our method on a public dataset of Bitcoin transactions published by \citet{weber2019antimoney}. Using a supervised setting similar to the original authors as our baseline, we showed that by training the same classifier considering the original 166 features and the new ones extracted from GuiltyWalker, we could obtain better results. In particular, by filtering the features extracted from GuiltyWalker and considering only the most important ones, the results were even better. The performance differences were more notorious for time steps associated with a black market shutdown, where the baseline model performed poorly. Moreover, we observed that the models that considered GuiltyWalker features could reliably capture new illicit transactions that were not captured by the model from \citet{weber2019antimoney}.
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
\label{sec:intro}
Several researchers have conducted Internet measurement studies to understand how the COVID-19 pandemic affected the Internet and user behaviors~\cite{Favale_2020,c2020impact,IMC20-covid-isp,IMC20-covid-mobile,IMC20-facebook}.
Favale et al. and Feldmann et al.~\cite{Favale_2020,IMC20-covid-isp} explored the changes in Internet traffic, Lutu et al.~\cite{IMC20-covid-mobile} explored the changes in traffic and its impact on user mobility in mobile operators, Candela et al.~\cite{c2020impact} analyzed the impact of Internet traffic changes on network latency, and Boettger et al.~\cite{IMC20-facebook} analyzed the changes in social media access patterns and the implications.
The details of these studies will be discussed in Section~\ref{sec:related}.
To the best of our knowledge, there has been no academic study that has analyzed the impact of COVID-19 in terms of registered domain names.
This work takes a first look at domain names related to COVID-19 (Cov19doms in short), using a large-scale set of registered domain names.
We note that the only literature we have been able to find on this subject is a blog article~\cite{cyberthreatcoalition2}, which analyzed the domain names associated with COVID-19.
The article reported that the number of COVID-19 domain name registrations has spiked in mid-March 2020, with some days seeing the registration of more than 5,000 Cov19doms.
However, we found that the data used in the article contained many false positives due to the naive string match heuristics.
Also, this data is no longer updated since May 2020, so we cannot perform a longer-term analysis using the data.
In this study, we attempt to extract Cov19doms accurately and analyze how it changes over a long period of time.
With so many of us keeping an eye on COVID-19 and spending more and more of our time online, it is crucial to understand the origins and implications of Cov19doms.
Given these backgrounds in mind, we attempt to answer the following research questions:
\begin{mybullet}
\item[{\bf RQ1}:] {\em Is the number of Cov19doms registrations correlated with the COVID-19 outbreaks?}
\item[{\bf RQ2}:] {\em For what purpose do people register Cov19doms?}
\end{mybullet}
To address the research questions, we compiled an exhaustive list of Cov19doms using a large-scale registered Internet domain name database~\cite{domainlistsio}, which accounted for 260M of distinct domain names registered for the 1.6K of top-level domains.
Using the dataset, we found that at least 167K of distinct Cov19doms containing strings such as ``covid'' or ``corona'' have been registered from the end of December 2019 to the end of September 2020.
We attempt to study how domain name registration behavior changed with the emergence of COVID-19; i.e., we examine whether or not the time-series of COVID-19 infections is correlated with the time series of domain name registrations.
Next, from the 167K of Cov19doms, we extracted active websites that used Cov19doms by checking DNS A record and HTTP/HTTPS response.
We then randomly sampled 10,000 of the Cov19doms websites to study how Cov19doms are used in the wild.
By applying cluster analysis to the screenshots, we systematically classified 10K websites.
For the remaining general websites, we performed manual inspection with the aid of three evaluators.
We also leveraged online virus-testing services to check whether some Cov19doms were used for malicious activities.
Our chief findings are as follows:
\begin{mybullet}
\item Similar to the global COVID-19 pandemic observed around April 2020, the number of Cov19doms registrations also experienced drastic growth, which, surprisingly, preceded the COVID-19 pandemic by about a month.
\item 70 \% of active Cov19doms websites with visible content provided useful information such as health, tools, or product sales related to COVID-19.
\item Non-negligible number (roughly 4\%) of registered Cov19doms have been used for malicious purposes such as phishing or malware distribution.
\end{mybullet}
These findings imply that it has become more challenging to distinguish between domain names registered for legitimate purposes and those that are not.
It was also indicated that it is necessary for researchers who analyze domain names, and even operators and blacklisters who take security measures based on domain names to pay close attention to how Cov19doms currently parked or in preparation will be used/misused in the future.
\section{Data}
\label{sec:data}
\subsection{Collecting Cov19doms}
To collect registered Cov19doms, we used a large-scale commercial domain name database, domainlist.io~\cite{domainlistsio}.
This database contains snapshots of approximately 260M domain names taken from 1.6K of different TLDs, and we continued to retrieve data daily from 27 December 2019 to 20 September 2020.
Of the $98,940,555$ domain names that have been newly registered since December 27, 2019,
we first extracted the domain names that contained ``covid'' or ``corona'' as a substring.
As a result, we obtained a total of $170,846$ Cov19doms.
We note that this approach could include false positives such as ``covideo.co.uk'', for instance.
However, we can safely ignore the effect of false positives in the following analysis, as our manual inspection of the randomly sampled data showed that the occurrence of such false positives was extremely rare as these words are.
We believe that these words, especially in the COVID-19 era, are mostly used in the context of a specific purpose, i.e., ``severe acute respiratory syndrome coronavirus 2,'' resulting in fewer false positives.
To study the characteristics of the Cov19doms, it is essential that we can get information about the creation date of the domain names.
Therefore, we used the WHOIS information for the extracted Cov19doms to obtain information on the date and time the domain name was created.
If the creation date of a domain name was older than December 27, 2019, those domain names were excluded from the following analysis.
This resulted in a total of 166,825 Cov19doms, as shown in Table~\ref{tab:data_stats2}. To ensure that domains registered before December 27, 2019 were not related to COVID-19, we manually checked on them and found it be correct. In fact, most of them were related to Coronado city in California, U.S.
We investigated where the specific words related to COVID-19, i.e., ``covid'' and ``corona'', are located in
the left-most labels of Cov19doms (e.g., ``covidcare'' in \texttt{covidcare[.]example}) and confirmed that (a) 59.6\% are at the beginning, (b) 24.2\% are at the end, and (c) 16.2\% are in the middle of them.
The patterns (a) and (b) mean that the left-most labels of Cov19doms were generated by concatenating any character at the beginning or end of the COVID-19-related words such as ``covid''.
We believe that patterns (a) and (b) are less likely to cause false positives than pattern (c).
We further investigated the extent to which similar COVID-19-related words, ``covid'', ``covid19'', and ``covid-19'', are included in Cov19doms and found that they are 41,718, 32,671, and 10,120 Cov19doms, respectively. These numbers do not overlap, because we checked Cov19doms that contain ``covid19'' and ``covid-19'' earlier.
It is interesting that ``covid19'' is more common in Cov19doms than its formal name of the desease, ``covid-19''.
Among these, ``covid'' was most frequently included in Cov19doms, and as far as we manually checked, the majority of cases (about 40\%) were used in the context of the COVID-19.
One of the reasons why ``covid'' is included in Cov19doms in large numbers is that there are cases where various numbers are added to the end of ``covid'' (e.g., covid-2019, covid-2020, and covid-2021). We expect those domain names to have been acquired for speculative purposes.
We looked into what country registered Cov19doms firstest by usinig WHOIS registrant information. Of the 165,185 Cov19doms we extracted, 153,243 domains had valid WHOIS registrant country information. Among the countries, United States was the first to register Cov19doms. The top-5 countries registered Cov19doms were United States (85,970), Canada (17,229), Panama (6,781), Germany (4,533) and United Kingdom (4,237).
\subsection{Collecting Active Websites Using Cov19doms}
With the aim of studying the usage of Cov19doms, we extract the active websites that are operating using Cov19doms.
To extract active websites, we first check the DNS A record to determine if an IP address is assigned to the extracted Cov19doms.
We then send an HTTP/HTTPS request to the domain name where the DNS A record exists, and record the response.
Specifically, we check if a connection can be established to Port 80 and Port 443 of each host that had a Cov19dom. Next, if a connection with either port can be established, we made an HTTP/HTTPS request to those hosts and checked whether the content could be retrieved from them.
This step removes websites that caused connection timeouts and/or TLS errors such as invalid certificate.
These steps resulted in a total of 77,333 of active websites that use Cov19doms, as shown in Table~\ref{tab:data_stats2}.
\begin{table}[tbp]
\centering
\normalsize
\caption{Statistics of extracted Cov19doms data.}
\label{tab:data_stats2}
\begin{tabular}{l|r}
\hline
& \# of domain names \\
\hline\hline
Orig. Cov19doms & 170,846 \\
WHOIS check & 166,825 \\
DNS check & 144,522 \\
HTTP/HTTPS check & 77,333 \\
\hline
\end{tabular}
\end{table}
\section{Measurement Study}
\label{sec:measurement}
Figure~\ref{fig:measurement_overview} presents an overview of the measurement processes.
We first study the correlation between the number of COVID-19 infections and the number of Cov19doms registrations (Sec~\ref{sec:correlation}).
For this analysis, we used the statistics on the number of COVID-19 infections by country, provided by WHO~\cite{who-covid19}.
We then study how Cov19doms are used for various websites (Sec~\ref{sec:categorization}).
The classification of active websites operated using Cov19doms was manually performed by three evaluators.
Due to the large number of websites to be analyzed, we conducted a random sampling study.
Finally, we report the analysis of Cov19doms that have been used for malicious activities (Sec~\ref{sec:malicious}).
We used VirusTotal~\cite{virustotal} to investigate the presence of malicious sites using Cov19doms.
\begin{figure*}[!t]
\centering
\includegraphics[width=0.8\linewidth]{figs/measurement-overview-new.pdf}
\caption{Overview of the measurement processes.}
\label{fig:measurement_overview}
\end{figure*}
\subsection{Number of new infections and Cov19doms registrations}
\label{sec:correlation}
We analyze the online behavior of people around the world in response to the unprecedented event of COVID-19 through the lens of DNS.
Specifically, we examine whether or not the time series of COVID-19 infections is correlated with the time series of domain name registrations.
\begin{figure*}[tbp]
\centering
\includegraphics[width=0.49\linewidth]{figs/gtld_large.pdf}
\includegraphics[width=0.49\linewidth]{figs/cctld_large.pdf}
\caption{Number of COVID-19 infections and Cov19doms registrations over time. Cases for gTLD (left) and ccTLD (right).}
\label{fig:time-all}
\end{figure*}
\begin{figure*}[tbp]
\centering
\includegraphics[width=0.49\linewidth]{figs/GB_large_period.pdf}
\includegraphics[width=0.49\linewidth]{figs/SE_large.pdf}
\caption{Number of COVID-19 infections and Cov19doms registrations over time. Cases for UK (left) and SE (right). In Sweden, no lockdown enforcement was taken.}
\label{fig:time-se-uk}
\end{figure*}
First, we investigate the time series of new registrations of Cov19doms and the number of new COVID-19 infections worldwide.
We take all Cov19doms and split them into groups of gTLDs (e.g., \texttt{.com}) and ccTLDs (e.g., \texttt{.uk}).
We obtained information on the number of COVID-19 infections from the official WHO website~\cite{who-covid19}.
Figure~\ref{fig:time-all} shows the time series of the number of new COVID-19 infections and the number of new registrations in Cov19doms (gTLD and ccTLD).
These figures show that similar to the situation of the COVID-19 pandemic outbreak around the world around April 2020, Cov19doms saw a significant increase in its new registrations as well.
Surprisingly, the number of new domain name registrations peaked around March 2020, about a month ahead of the peak in the number of new COVID-19 infections.
Subsequently, the number of new registrations of Cov19doms has reached a stable daily registration rate, but the number of COVID-19 infections has still increasing as of October 2020.
Second, we focus on the Cov19doms of ccTLDs and investigate the relationship between the number of new registrations of Cov19doms per ccTLD and the number of new COVID-19 infections in the country corresponding to the ccTLD over time.
Our Cov19doms data included only four ccTLDs: United Kingdom (\texttt{.uk}), Sweden (\texttt{.se}), Niue (\texttt{.nu}), and Australia (\texttt{.au}).
We excluded \texttt{.nu}, for which no information on the number of WHO infected people existed from there, and \texttt{.au}, for which we were unable to obtain the full domain name registration date from the WHOIS data, and conducted a survey of 4,766 \texttt{.uk} and 549 \texttt{.se} Cov19doms.
Figure~\ref{fig:time-se-uk} shows the time-series change in the number of new infections of Cov19doms and COVID-19 in the UK and Sweden, respectively.
Since the lockdown was implemented in the UK, the period is also shown in .uk graph.
In both cases, Cov19doms registrations tend to be more likely to be ahead of the COVID-19 infection explosion.
Furthermore, we find that registration of Cov19doms moves faster and clearer in the UK than in the Sweden case.
Our results obtained so far above indicate two things: (1) events like COVID-19 that affect so many people's lives will create a massive demand for domain names and (2) people are anticipating such demand and taking the action of registering domain names at an amazingly early stage.
In subsequent sections, we will clarify for what purpose people are registering these Cov19doms.
\subsection{Understanding the Usage of Cov19doms}
\label{sec:categorization}
In general, automatic website classification is not an easy task as the modern web is composed of rich and complex multimedia, making it difficult to automatically analyze its contents using simple data processing scheme.
Therefore, instead of fully automating the website classification process, this work adopted manual inspection to ensure the quality of the classification.
However, the number of Cov19doms we have collected is so large that it is infeasible to inspect them all manually.
Therefore, we took the approach of applying random sampling to reduce the number of domains/websites to be analyzed.
As shown in Figure~\ref{fig:measurement_overview}, we randomly sampled 10K of websites from 77K active Cov19doms websites to reduce the number of samples to be classified by human.
For the 10K of randomly sampled Cov19doms websites, we took the following two-stage approach.
In the first stage, we aim to systematically classify websites into the following categories: {\em Empty}, {\em Error}, {\em Parked}, {\em Hosted}, and Has content, where
{\em Empty} represents cases in which HTTP/HTTPS requests were responded to, but the data was empty,
{\em Error} represents the websites responded with error codes such as 404 or 501,
{\em Parked} represents the domain parking websites,
{\em Hosted} represents cases where the domain name has been purchased, but the website only shows the initial page after installation of Apache, WordPress, etc, and
{\em Has content} represents the remaining Cov19doms websites that have some content.
In the second stage, three evaluators manually classify the websites classified as ``Has content.''
In the following, we present the details of the analysis to be performed at each stage and the results obtained.
\subsubsection{Stage 1: Systematic Classification}
We classify websites into the five classes defined above based on HTTP/HTTPS response codes and screenshot information.
Among the five classes, the classification of empty and error is simple.
They can be classified by analyzing the size of the data retrieved and the response code.
For the remaining classes parking and hosted, we use cluster analysis.
For parking, we could use domain name registrar information in some cases, however, our preliminary study shows that we cannot do a comprehensive study due to the existence of so many different domain parking companies.
The key idea is that the majority of websites that are accessed for parking and hosted are similar in appearance.
Therefore, we apply cluster analysis to the screenshot images and classify the websites by determining whether each cluster is {\em Parked} or {\em Hosted} or {\em Has content}.
With this approach, we can streamline the classification.
To perform clustering of screenshot images, we need to calculate the distance between images; i.e., it is necessary to compute the similarity of images.
There are several methods for computing the similarity of images, and in this paper, we adopt the perceptual hash (pHash)~\cite{phash}, which computes close hash values for two similar images.
pHash is widely used to discover copyright infringement and is known to be effective in discovering resemblances to certain images.
We first accessed 10K of randomly sampled active websites and extract HTML, screenshots, and other metadata by navigating Google Chrome\footnote{We used the version of 81.0.4044.129.} using Selenium~\cite{selenium}.
The language was set to English, and the User-Agent was set to Windows 10 Google Chrome.
To not halm the websites set to be investigated, access to the IP address corresponding to each Cov19dom is limited twice (HTTP and HTTPS).
Next, we computed the pHash values for the 10,000 screenshots we collected, using imageHash~\cite{imagehash}.
We then grouped the corresponding Cov19doms with the same value of pHash and HTTP status code pairs into the same cluster.
Table~\ref{tab:systematic-classification} presents the classification result of the Cov19dom websites.
From the table, we can see that many of the Cov19doms websites resulted in either domain parking or errors, and that 40\% of the websites (classified as ``Has content'') requires detailed manual inspection.
We note that 60\% of the websites categorized as other than ``Has content'' do not currently provide any useful content, however, they might start providing some content in the future, so we need to pay attention to them.
In the following, we will classify the websites categorized as ``Has content.''
\begin{table}[tbp]
\centering
\normalsize
\caption{Result of Systematic Classification.}
\label{tab:systematic-classification}
{\renewcommand{\arraystretch}{0.9}
\begin{tabular}{l|rr}
\hline
Category & \# of active websites & fraction (\%) \\
\hline\hline
Empty & 609 & 6.1 \\
Error & 1,663 & 16.6 \\
Parked & 2,138 & 21.4 \\
Hosted & 1,402 & 14.0 \\
\hline
Has contents & 4,188 & 41.9 \\
\hline\hline
Total & 10,000 & 100.0 \\
\hline
\end{tabular}
}
\end{table}
\subsubsection{Stage 2: Manual Classification}
In the second stage, we will classify the Cov19doms websites marked as ``Has content'' in the Stage 1.
Since 4K of websites are too many to analyze manually, further random sampling is performed and 1,000 general websites will be carefully classified by three evaluators.
Through the Stage 1 classification, the classification categories for Stage 2 were predetermined and provided to the evaluators with detailed explanations.
Figure~\ref{fig:classify-tool} presents a screenshot image of a tool developed by the authors to help evaluators efficiently classify websites.
Although the evaluators made a classification based on screenshot image and metadata, there are cases that cannot necessarily be determined by screenshot or metadata.
For example, if the evaluators could not understand the language used in the web content, they also leveraged external resources such as a search engine.
\begin{figure*}[tbp]
\centering
\includegraphics[width=0.98\linewidth]{figs/gui_level2_new.pdf}
\caption{A screenshot of the website classification tool we developed for our analysis.}
\label{fig:classify-tool}
\end{figure*}
Three evaluators used the tool to classify 1,000 of websites marked as ``General'' taking 4.8 hour on average, resulting in 477 websites where the three evaluators agreed, 423 websites where the two evaluators agreed, and 100 websites where they all disagreed.
That is, for 90\% of the websites, at least two evaluators' classification results were consistent.
The result of calculating the Fleiss' kappa coefficient, which is a quantitative measure of inter-rater agreement, was 0.50, which can be interpreted as moderate agreement~\cite{landis1977measurement}.
The results of the interviews with the three evaluators revealed that the primary reason for the disagreement was the difference in the decisions they made when they were unsure of their classification. One of the evaluators reported that he categorized all of his confusion as ``activities for COVID-19.''
Discrepancies in judgments also arose because of the existence of websites that could be classified into multiple categories. For example, a website that displays medical products (masks, face shields) may be categorized as both Health and Sales.
Apart from such discrepancies, the classifications were generally consistent and the manual classification results can be considered reasonable.
In the final classification, a majority vote was adopted.
Websites with discrepancies between the three evaluators' classification results were marked as ``Unknown.''
Table~\ref{tab:manual-classification} presents the classification results.
About 70\% of the websites were related to COVID-19.
Those websites were medical services, selling products, providing COVID-19 information such as apps, maps and dashboards, supporting people's activities related to COVID-19, and social security.
As would be expected from the nature of COVID-19, the majority of the websites (40\%) were medical-related.
Many of these health-related websites are critical sources of information in countering COVID-19 pandemic and should never be blocked.
The remaining 30\% were completely irrelevant websites, websites with no content displayed, and ``Unknown,'' which we defined earlier.
\begin{table*}
\centering
\normalsize
\begin{threeparttable}[!t]
\caption{Results of Manual Classification.}
\label{tab:manual-classification}
{
\begin{tabular}{l|l|r}
\hline
Category & Description & \#sites \\
\hline\hline
Health & Websites providing information on health\tnote{$^\ast$} & 405\\
Sales & Websites selling products related to COVID-19 & 109\\
Tools & Websites providing apps/maps/dashboards of COVID-19 & 123\\
Activities & Websites dedicated to people's activities to address COVID-19\tnote{$^{\ast\ast}$} & 72\\
Social Security & Websites regarding to social security & 2\\
\hline
Unrelated & Websites unrelated to COVID-19 & 139\\
Login & Websites showing a login page & 26\\
Index of & Websites showing the ``Index of /'' page & 24\\
\hline
Unknown &Websites with discrepancies between the evaluators' classifications & 100\\
\hline\hline
Total &-- & 1,000\\
\hline
\end{tabular}
}
\begin{tablenotes}
\item[$^\ast$] hospitals, infection testing, sterilization, and other health-related topics.
\item[$^{\ast\ast}$] Fundraising, volunteering, business, and political movements regarding COVID-19.
\end{tablenotes}
\end{threeparttable}
\end{table*}
\subsection{Malicious Activities Using Cov19doms}
\label{sec:malicious}
Finally, we investigate whether Cov19dom websites were involved in any malicious activities.
To achieve this goal, we utilize VirusTotal, a large-scale online virus scanning service.
As shown in Figure~\ref{fig:measurement_overview}, we target 10K of active websites that had Cov19doms.
Of the 10K websites, 6,362 of the websites were detected as malicious by at least one or more scanners.
This is an alarming number, but when we analyzed the detection results, we found that one online scanner detected 6,256 websites as malicious, and that the majority of them (about 98.7\%) were classified as phishing sites.
Although we cannot determine from our data whether or not these detections were correct, the result does suggest that there may be a non-negligible number of malicious sites that use Cov19doms.
On the other hand, one of the reasons why online scanners may falsely detect Cov19doms as a phishing site is likely to be naïve detection using keyword matching.
For example, a scheme that increases the probability of detecting a website with a domain name containing the strings corona or covid-19 as a malicious site could be employed.
However, such an approach might have the risk of blocking websites that provide important information about COVID-19.
To reduce the effect of false positives from individual scanners, we examined websites that were detected as malicious by at least two online scanners. We note that this approach is consistent with the best practice used in many papers that make use of multiple engines/vendors of VirusTotal for the labeling task~\cite{IMC_VT}. As a result, we found that the number was 357, which accounted for roughly 4\% of the active Cov19doms websites.
The detection categories of those detected by two or more online scanners are summarized in Table~\ref{tab:scan}.
Note that a website may be detected as a different category (e.g., phishing site and malware site) by several online scanners . In such cases, the category is decided by majority vote, and if the category is not uniquely determined, the category is marked as ``pending''.
It can be seen that once again, phishing sites have the highest number of detections, but the number of other malicious sites is also very close.
\begin{table}[!t]
\centering
\normalsize
\caption{Breakdown of the detection results.}
\label{tab:scan}
{\renewcommand{\arraystretch}{0.9}
\begin{tabular}{l|rr}
\hline
Detected category & \# detections & fraction (\%) \\
\hline\hline
Phishing site & 117 & 32.8 \\
Malicious site & 52 & 14.6 \\
Malware site & 17 & 4.7 \\
\hline
Pending & 171 & 47.9 \\
\hline\hline
Total & 357 & 100.0 \\
\hline
\end{tabular}
}
\end{table}
\section{Discussion}
\label{sec:discussion}
\subsection{Limitations}
This study aims to understand the Cov19doms in the wild.
In order to ensure the accuracy of the results, two heuristics were applied to extract such domain names, as described in Section~\ref{sec:data}.
The first heuristic was to limit the domain search words to ``covid'' and ``corona.''
Such limitation will miss several cases where domain names contain other keywords such as ``virus'' or ``mask,'' which could bring false positives as we discussed.
We also limited our search to the e2LD part; the limitation will eliminate the cases where an FQDN contains the substrings in its hostname.
Another heuristic was to constrain the registration date for domain names.
Our analysis also excluded websites that did not include keywords in their domain name but were COVID-19-related in their website content.
Such websites existed on both malicious and benign sites.
Another limitation that we are aware of is that the URL path is not taken into account when creating a URL from an FQDN.
We only retrieved web content from the top directory on a website
in the web-crawling process.
Exploring the URL path might reduce the errors shown in the Table~\ref{tab:systematic-classification}, however we may miss web content if a website does not configure the setting of index file.
Addressing these issues is left for future study.
\subsection{Detecting malicious Cov19doms}
As we have shown in this work, simply using a list of Cov19doms as a blocklist may result in false positives, and this introduces the risk of blocking information that is useful for COVID-19 countermeasures.
In order to determine if a detected Cov19dom is malicious, we need to monitor a domain name when it is being abused and examine the content in a timely manner.
The Trademark Clearinghouse (TMCH) is a global database of trademarks and provides this information to registries and registrars during the domain name registration process to thwart unwanted domain name registrations by third-parties.
This is effectively used by trademark owners to fight against a trademark infringement using fake domain names.
Unfortunately, this countermeasure is not effective against domain names piggybacking on global crises including COVID-19, due to the fact that there is no right owner of such corresponding keywords.
Szurdi and Christin proposed the anti-bulk registration policy such as dynamic pricing to make bulk domain registrations expensive~\cite{Szurdi:WEIS18}, which is a potential countermeasure against bulk-registered COVID-19 domain names.
\subsection{Ethical considerations}
Our study analyzed publicly available DNS records and web content corresponding to the domain names without collecting personally-identifiable information.
In our web-crawling process, we sent the minimum amount of legitimate requests to websites, i.e., two requests (HTTP and HTTPS) per site, and left them and their users unharmed.
\section{Related Work}
\label{sec:related}
In this section, we present several related works and clarify how our work differs.
\noindent\textbf{Internet Measurement driven by COVID-19.}
Favale et al.~\cite{Favale_2020} analyzed the impact of the lockdown enforcement on a campus network in Italy.
Through analyzing Internet traffic statistics, they revealed that while incoming traffic
was reduced by a factor of 10 during the lockdown, outgoing traffic increased by 2.5 times, driven by more than 600 daily online classes, with around 16,000 students per day.
They concluded that the campus network infrastructure is robust enough to successfully cope with the drastic changes while maintaining the university operations.
Feldmann et al.~\cite{IMC20-covid-isp} conducted similar analysis using traffic data collected at one ISP, three IXPs, and one educational network.
They reported on changes in Internet traffic in various perspectives and concluded that the Internet infrastructure has been able to deliver the increased Internet traffic without significant impact.
Candela et al.~\cite{c2020impact} conducted a large-scale analysis of Internet latencies, which could be affected by the increased amount of online activities during the lockdown.
By leveraging the measurement data collected with the RIPE Atlas platform~\cite{ripeatlas},
they analyzed Internet latencies focusing on Italy, where people experienced more than a month of lockdown.
They reported that the increase in online activity led to an increase in the variability of Internet latencies, a trend that intensified in the evening due to the increase in the entertainment
traffic.
\noindent\textbf{Event-driven domain name registration.}
The strategy of early acquisition of domain names associated with ongoing events has been a well-known approach in the domain name business community. In fact, a patent of such a technique was filed by an Internet domain registrar~\cite{patent-2008}.
Although event-driven domain registration is a widely known best practice in the domain name business community, to the best of our knowledge, there has been little research on the topic in the research community.
One of the few available studies is that Coull et al.~\cite{10.1007/978-3-642-15257-3_7}
derived rules to describe topics, such as ongoing events, from popular Google search queries with the aim of characterizing the registration of speculative domain names and empirically evaluated the feasibility of domain acquisition based on such a method. While they attempted to extract current events using Google search, COVID-19 is a unique phenomenon, and researchers have not had an opportunity to study domain names for such a case.
Tombs et al.~\cite{ambiguity_in_covid_tld} tried to determine the level of credibility of a top-level coronavirus-related website that purport to be government websites, and find out the purpose of non-governmental entity or company register a top-level coronavirus-related domain name by analyzing data collected from 303 websites which domains related to COVID-19 between April 5 and April 6, 2020.
They found that 80\% of websites presented as government websites cannot be verified the authenticity. Additionally, about 30\% of websites collected had unverified information and nearly half were squatting domains or ``under construction."
Government websites providing critical information about coronaviruses should not be subject to ambiguous in their authenticity , and therefore should not share the top-level domain name space with non-governmental entity or company.
Their findings are important in establishing trusted communication channel between government and their citizens during this crisis.
\noindent\textbf{Malicious domain names and websites.}
Much research has been conducted on ways to observe the registration and early activity of malicious domain names~\cite{10.1145/2068816.2068842,10.1145/3196494.3196548,10.1145/3278532.3278569}.
Hao et al.~\cite{10.1145/2068816.2068842}
unveiled that DNS infrastructures and early DNS lookup patterns for a newly registered malicious domain name differ significantly from those with a legitimate domain name.
Korczynski et al.~\cite{10.1145/3196494.3196548}
collected WHOIS information, web content, and DNS records for corresponding malicious domain names provided from 11 distinct abuse feeds and observed a growing number of spam domains in new gTLDs, indicating a shift from legacy gTLDs to new gTLDs.
We conducted our measurement by referring to the ways practiced in these existing studies.
While these studies analyzed fake domain names containing strings related to brand names having specific owners, our study focuses on domain names containing strings related to generic crisis having no specific owners, which makes it be challenging to distinguish between malicious and legitimate domain names.
There are few academic studies so far on detection of malicious domain names related to COVID-19.
Ispahany and Islam developed a machine learning model using lexical features to detect malicious domain names and examined registered domain names in April 2020~\cite{ispahany2020detecting}. The purpose of our study is not to detect malicious Cov19doms, but to investigate the usage of Cov19doms.
Furthermore, our study utilized a long-term dataset obtained from the end of December 2019 to the end of September 2020.
\section{Conclusion}
\label{sec:summary}
Through the analysis of 167K of Cov19doms we collected, we found that a month before the global COVID-19 pandemic hit in April 2020, there was a flood of domain name registrations.
This phenomenon can be attributed to a variety of people registering domain names for the purpose of COVID-19 countermeasures, speculative domain name business, or to generate phishing sites, as they predicted the high impact of COVID-19.
Such a global, high-impact phenomenon is unprecedented in the past and is a remarkable event from the perspective of Internet measurement.
In conventional measures against the registration of unwanted domain names targeting brands, distinguishing between an original domain name and a fake domain name has been relatively straightforward since the brand owner has been determined.
In the case of the Cov19doms, on the other hand, there is no concept of a brand owner, and many different players have registered Cov19doms to benefit society.
Therefore, it is not feasible to apply traditional domain name analysis methods.
As this study revealed, majority of Cov19doms (about 60\%) are not active. Even if Cov19doms are uesd for active websites, many of them are parked or hosted, and it is not clear how these domain names will change in the future.
Addressing these problems is a challenge for the future.
We plan to release our dataset and tools used for our analyses at \texttt{https://github.com/cov19doms/cov19doms}
|
\section*{Data Article template}
{\fontsize{7.5pt}{9pt}\selectfont
\noindent\textbf{Specifications Table}
\begin{longtable}{|p{33mm}|p{94mm}|}
\hline
\endhead
\hline
\endfoot
Subject & Computer Science \\
\hline
Specific subject area & Auto-tuning GPU kernels using hardware performance counters\\
\hline
Type of data & Tables\newline
Python and R scripts
\\
\hline
How data were acquired &
\textbf{Raw autotuning data:} Using our autotuning framework KTT, we measured computation time and collected hardware performance counters for whole tuning spaces of five benchmark CUDA codes on four GPUs. Kernel Tuning Toolkit (KTT) is freely available in Github repository \url{https://github.com/HiPerCoRe/KTT}, five benchmarks are also there in 'examples' folder. These benchmarks cover a wide range of computational problems: computing convolution, Coulomb summation in three dimensions, matrix multiplication, matrix transposition and N-body problem. They also differ in sizes of their tuning spaces. \newline
\textbf{Prediction models:} Using our scripts, also available in KTT repository at Github, we trained models with the raw tuning data.
\\
\hline
Data format &
Raw\newline
Analyzed\newline
Scripts
\\
\hline
Parameters for
data\newline
collection &
\textbf{Raw autotuning data:} Computation time and performance counters were measured for five benchmarks (GEMM, Convolution, Matrix transposition, 3D Coulomb summation and n-body) bundled with KTT, git tag \texttt{v1.3-profile-searcher}. We ran them on four GPUs: GeForce GTX 680, GeForce GTX 750, GeForce GTX 1070 and GeForce RTX 2080. KTT was configured to perform exhaustive exploration of tuning spaces on each GPU under our test with profiling switched on. Size of input for each benchmark was chosen so that the kernel execution took 1-10 milliseconds. For the GEMM benchmark, data for several input sizes were collected on GTX GeForce 1070.
\newline
\textbf{Prediction models:} Prediction models were trained for all hardware performance counters, local and global size.
\\
\hline
Description of
data\newline
collection &
\textbf{Raw autotuning data:} KTT performed exhaustive exploration of complete tuning spaces (sets of all executable tuning configurations) of tested benchmarks for each GPU. Each tuning configuration contains information about tuning parameters (affecting how GPU kernel is created and executed), the runtime of the kernel and hardware performance counters provided by NVIDIA CUPTI library. Tuning configurations which cannot be executed on a particular GPU are not stored.
\newline
\textbf{Prediction models:} Models were created from the raw autotuning data with scripts \texttt{create\_nonlinear\_models.R} and \texttt{generate-knowledge-base.py} available with profile-based searcher in KTT.
\\
\hline
Data source location &
Institute of Computer Science, Masaryk University\newline
Brno\newline
Czech Republic\newline
49.211N, 16.598E
\\
\hline
\hypertarget{target1}
{Data accessibility} &
Repository name: Mendeley Data\newline
Data identification number: doi:10.17632/nn53dskr7z.1\newline
Direct URL to data: http://dx.doi.org/10.17632/nn53dskr7z.1\newline
\\
\hline
Related
research\newline
article &
Filipovi\v{c}, J., Hozzov\'{a}, J., Nezarat, A., O\softl{}ha, J., Petrovi\v{c}, F., Using hardware performance counters to speed up autotuning convergence on GPUs, Future Generation Computer Systems. In Press.
\end{longtable}
}
\section*{Value of the Data}
\begin{itemize}
\itemsep=0pt
\parsep=0pt
\item
Raw autotuning data contain, to the best of our knowledge, the first freely available complete tuning spaces of several CUDA kernels prepared for autotuning, alongside their computation times and hardware performance counters measurements on several GPUs. Scripts make it easier to experiment with searching tuning space in a controlled environment, so the results of searchers are comparable.
\item
These data will help those researching how to search the tuning spaces of GPU codes or those interested in mining the data related to hardware performance counters and GPU kernels' performance.
\item
With raw autotuning data, new search algorithms for navigating the tuning spaces can be easily evaluated for multiple GPUs (even those unavailable to the researchers), skipping high time demands of actually compiling, running and measuring. Moreover, the global optimum of the tuning space is known from data.
\item With scripts for simulated and real-time tuning, the results of others (with a new search method or a new prediction model for performance counters) can be consistently compared to the results of our searcher.
\item Availability of KTT autotuner and scripts for model preparation allows users to expand our dataset by measurement on their own GPUs, or their own benchmarks.
\end{itemize}
\section*{Data Description}
\subsection*{Raw Autotuning Data}
Raw autotuning data were produced by Kernel Tuning Toolkit 1.3\footnote{\url{https://github.com/HiPerCoRe/KTT/releases/tag/v1.3-profile-searcher}} running on GPUs listed in Table~\ref{tab:gpus}. For each benchmark listed in Table~\ref{tab:benchmarks} (available in KTT repository in folder \texttt{examples} as cltune-conv, coulomb\_sum\_3d, cltune-gemm, mtran and nbody), the exhaustive search of the whole tuning space was executed, measuring computation time and hardware performance counters. For details on benchmarks and their tuning spaces, see~\cite{petrovic2020benchmark}.
\begin{table}
\caption{GPU devices used to obtain our data.}
\label{tab:gpus}
\centering
\begin{tabular}{|l|l|l|l|}
\hline
Device & Architecture & Released & Abbreviation\\ \hline
GeForce GTX 680 & Kepler & 2012 & 680 \\
GeForce GTX 750 & Maxwell & 2014 & 750 \\
GeForce GTX 1070 & Pascal & 2016 & 1070 \\
GeForce RTX 2080 & Turing & 2018 & 2080 \\
\hline
\end{tabular}
\end{table}
\begin{table}
\caption{A list of the benchmarks used to obtain our data.}
\label{tab:benchmarks}
\centering
\begin{tabular}{|l|l|l|}
\hline
Benchmark & Description & Abbreviation\\
\hline
Convolution & 2D convolution kernel using $7 \times 7$ filter adopted from~\cite{nugteren2015cltune}. & conv \\
Coulomb 3D & Direct coulomb summation on 3D lattice, introduced in~\cite{filipovic2017autotuning}. & coulomb\\
GEMM & Matrix-matrix multipication adopted from~\cite{nugteren2015cltune}, tuning space & gemm-reduced\\
& reduced as in~\cite{nugteren2018clblast}. & \\
Transpose & Out-of-place matrix transposition, adopted from NVIDIA & mtran\\
& CUDA SDK 10.0. & \\
N-body & N-body kernel, adopted from NVIDIA CUDA SDK 10.0. & nbody \\
\hline
\end{tabular}
\end{table}
The raw data are available at \url{http://dx.doi.org/10.17632/nn53dskr7z.1} in directory 'raw-autotuning-data'. They are stored as CSV files with naming convention containing the abbreviation of GPU, the abbreviation of benchmark, and suffix \texttt{\_output.csv}. For example, data obtained on GeForce GTX 1070 and N-body benchmark are stored in \\\texttt{1070-nbody\_output.csv}. There are special cases for GEMM benchmark, where we obtained data on small and highly-rectangular matrices. Those benchmarks are abbreviated as \texttt{1070-gemm-128-128-128} (multiplication of $128 \times 128$ matrices), \texttt{1070-gemm-16-4096-4096} (multiplication of matrix $16 \times 4096$ with matrix $4096 \times 16$), \texttt{1070-gemm-4096-16-4096} (multiplication of matrix $4096 \times 16$ with matrix $4096 \times 4096$) and \texttt{1070-gemm-4096-4096-16} (multiplication of matrix $4096 \times 4096$ with matrix $4096 \times 16$). However, those benchmarks are measured for GeForce GTX 1070 only.
The CSV files produced by Kernel Tuning Toolkit are formatted as follows:
\begin{itemize}
\item the first line is the header containing names of columns;
\item each other line contains the profile of one tuning configuration (a combination of tuning parameters, which produces unique CUDA kernel source code and execution setting);
\item if some configuration cannot be executed on a given GPU (e.g., because of insufficient hardware resources), it is not included in CSV (therefore, the same benchmarks can produce CSV files with a different number of lines when executed on different GPUs).
\end{itemize}
Each line of the CSV file contains the following types of columns:
\begin{itemize}
\item \textit{Kernel name}: the name of the benchmarked kernel (the same for one type of benchmark);
\item \textit{Computation duration (\textmu s)}: the duration of the benchmarked kernel and unit the time is measure in;
\item \textit{Global size} and \textit{Local size}: The global and local size of the executed kernel (number of threads and block size in CUDA terminology). The size is counted as a scalar number; it reflects an overall number of threads with no respect to the grid or block dimensionality;
\item Tuning parameters: the benchmark-specific tuning parameters, named in capitals by our convention (e.g., \texttt{VECTOR\_TYPE} or \texttt{CR});
\item Hardware performance counters: performance counters measured on particular GPU (e.g., \texttt{dram\_utilization} or \texttt{inst\_fp\_32}).
\end{itemize}
Please note that not all available hardware performance counters were measured due to time demand to measure the complete tuning space. The performance counters set differs from GPU to GPU because different architectures implement different performance counters. The biggest change is with GeForce RTX 2080, where the performance counters are completely re-designed and re-named.
\begin{table}
\caption{Input sizes used for gathering raw tuning data. The shown number determines size of input matrix/matrices in both dimensions with conv, gemm-reduced and mtran benchmarks. With nbody benchmark, the shown number determines number of simulated bodies. Finally, with coulomb benchmark, the first number is size of a 3D grid (the same in all dimensions), whereas the second number determines the number of atoms.}
\label{tab:input-sizes}
\centering
\begin{tabular}{|l|rrrr|}
\hline
& 680 & 750 & 1070 & 2080\\
\hline
conv & 2048 & 4096 & 4096 & 4096 \\
coulomb & 256, 64 & 256, 64 & 256, 64 & 256, 64 \\
gemm-reduced & 1024 & 1024 & 1024 & 2048 \\
mtran & 8192 & 8192 & 8192 & 8192 \\
nbody & 16384 & 16384 & 16384 & 16384\\
\hline
\end{tabular}
\end{table}
Input size for kernels was selected so that the kernel execution took approximately 1 - 10 milliseconds. These sizes obviously differ for each benchmark and GPU, Table~\ref{tab:input-sizes} summarizes them.
\subsection*{Prediction Models for Performance Counters}
We provide pre-computed prediction models for performance counters. All predict the global size, the local size, and performance counters relevant for the GPU of the training raw autotuning data. For a detailed list of performance counters and implemented models, see \cite{Filipovic2021}.
\subsubsection*{Least-squares Nonlinear Models}
Nonlinear prediction models were produced with the script \texttt{create\_least\_squares\_models.R} bundled with KTT in 'profile-searcher/scripts-prep/'. For each raw tuning data file (i.e. for each benchmark and each GPU, and for all input sizes of gemm benchmark on GeForce GTX 1070), we ran the script, producing multiple models for each performance counter, each for a different combination of values of binary tuning parameters. Please see the section Experimental Design, Materials and Methods for details on how the models are generated. It might make it easier to understand the format of the model files. The least-squares nonlinear models are stored as CSV file, following a similar naming convention as raw autotuning data: the abbreviation of GPU, the abbreviation of benchmark and suffix \texttt{-model\_[number].csv}. The special cases for GEMM benchmark are named analogously as their raw tuning data files.
The CSV files produced by the script contain three sections. The first section includes a line for each tuning parameter describing an expression for coding this parameter, as coded values of tuning parameters are used to predict the values of performance counters). The second section includes one line called \texttt{Condition} describing a logical condition of values of binary parameters this model was trained for. Furthermore, the third section includes a line for each performance counter, describing an expression for predicting the given performance counter's value with the coded values of tuning parameters.
\subsubsection*{Decision Tree}
Decision-tree prediction models were produced with the script \texttt{generate\_decision\_tree\_model.py} bundled with KTT in 'profile-searcher/scripts-prep/'. The script takes raw tuning data as an input and creates a predictive model of performance counters. Please see the section Experimental Design, Materials and Methods for details on how the models are generated. The resulting decision tree is stored as a pickle file together with a CSV file containing a list of all performance counters predicted by the model.
\section*{Experimental Design, Materials and Methods}
\subsection*{Obtaining Raw Tuning Data via Kernel Tuning Toolkit}
The raw tuning data are obtained during an autotuning process performed by Kernel Tuning Toolkit with GPU of our interest. Note that we recommend using version tagged v1.3-profile-searcher, which couples KTT 1.3 with profile-based searcher and benchmarks listed in Table~\ref{tab:benchmarks} prepared for collecting raw data or executing tuning with the profile-based searcher. We can explore the tuning space of any benchmark bundled with KTT (in 'examples' folder) or user-provided code. To obtain tuning data with hardware performance counters, KTT has to be built with enabled profiling (e.g., by calling \texttt{premake5 --profiling=cupti gmake}, see KTT documentation for further details). Moreover, the profiling has to be switched on in the autotuned application by calling \texttt{ktt::Tuner::set\-Kernel\-Profiling()} method (the profiling in benchmarks bundled with KTT can be switched on by setting \texttt{USE\_PROFILING} macro to 1).
KTT can explore either the entire tuning space (default behaviour) or only its subset. In the latter case, it is recommended to use random searcher to randomize the observed subset (using method \texttt{ktt::Tuner::setSearcher()}). After the search of the space is complete, the tuning data are stored in CSV files by method \texttt{ktt::Tuner::Print\-Result()}). All benchmarks bundled with KTT stores resulting CSV and can execute exhaustive exploration of the tuning space by setting preprocessor macro \texttt{EXHAUSTIVE\_SEARCH} to 1. For more information about KTT methods and implementation of new autotuned codes, we refer to its documentation~\footnote{\url{https://github.com/HiPerCoRe/KTT}}.
\subsection*{Generating Prediction Models from Raw Tuning Data}
We provide scripts to generate prediction models from raw tuning data. These scripts take tuning space of the problem with collected performance counters and train a model that predicts performance counters' values when given a tuning configuration.
\subsubsection*{Generating Least-square Regression Non-linear Models}
We provide two scripts in 'profile-searcher/scripts-prep' folder in the Github KTT repository.
The main script \texttt{create\_least\_squares\_models.R} trains nonlinear models: it takes the tuning data, and for each performance counter, it generates a model that predicts its value based on values of tuning parameters. To increase the accuracy of such prediction, we divide the tuning space into subspaces based on values of binary tuning parameters, as we suspect these have a profound influence on the performance counters. Thus, we generate several models for each performance counter, each model applicable only for a given combination of values of binary tuning parameters. An example of its usage is shown in Listing~\ref{lst:create-nm}. We recommend R v3.4.4, no special Rcran libraries are necessary.
\begin{lstlisting}[label=lst:create-nm, caption=Generating nonlinear models for GEMM benchmark on GTX GeForce 1070]
Rscript ./create_least_squares_models.R 1070-gemm-reduced_output.csv \
1070-gemm-reduced 4:19 2,3,19:62
\end{lstlisting}
It takes four arguments:
\begin{itemize}
\item \texttt{[input file name]} e.g. 1070-gemm-reduced\_output.csv, must follow the formatting of raw tuning data, as described above in section Data Description;
\item \texttt{[prefix for output files names]} e.g. 1070-gemm-reduced, this will be used to name output files with models, \texttt{-model\_[number].csv} will be added;
\item \texttt{[numbers of columns with tuning parameters in input file]} in format allowing to set individual columns and interval of columns (in format 'from:to') e.g. 2,5:12 meaning columns 2 and 5 through 11, counting starts at 0;
\item \texttt{[numbers of columns with performance counters in input file]} in the same format.
\end{itemize}
After parsing the script arguments and reading the input file, we code the tuning parameters' values, i.e., scale them to the range of $<-1,1>$. This step is recommended in any regression model design, as models generally do not work well with the absolute values of the factors. Next, we select the values of tuning parameters that will determine training data. In other words, we do not choose data points (rows from the input file) for training randomly. We select a few values of non-binary tuning parameters and then include all available combinations in the training dataset. We need to moderate the number of values to prevent an exponential increase in training data size or a poor sampling of some part of the tuning space due to constraints.
The function takes two arguments: the formula and the training data. The formula includes the factors, i.e. coded tuning parameters, and arithmetic operations with them. To make the models nonlinear, we include multiplications of factors (to capture their interactions) and quadratic terms. The training data include rows from the input file with selected values of tuning parameters and corresponding values of the given profiling counter.
The output of the script are multiple files named \texttt{[output\_name]-model\_[number].csv}. The number of models corresponds to the number of combinations of values of binary parameters. If a model cannot be created for a specific combination of binary parameters (e.g. there are no data due to constraints), the closest model (i.e., the minimal number of values of binary tuning parameters differs) fills in and is printed in the output file. The format and contents of model files are described in the above section Data Description.
The script \texttt{generate\_least\_squares\_models.py} makes it quick and easy to generate models for all our raw tuning data. It requires python3, with docopt library installed. It takes one argument, \texttt{--benchmark}. The option \texttt{--benchmark GPU} generates models for benchmarks conv, coulomb, gemm-reduced, mtran and nbody for all GPUs. Option \texttt{--benchmark GEMM} generates models for different input sizes of benchmark gemm-reduced on GeForce GTX 1070. Users may need to modify the script to accommodate the names of folders with data.
\subsubsection*{Generating Decision Trees}
The script \texttt{generate\_decision\_tree\_model.py} for decision-tree model preparation is stored in 'profile-searcher/scripts-prep' folder in a KTT distribution. While generating the model, users have to supply the CSV file containing explored tuning space, columns containing tuning parameters and profiling counters with parallelism configuration ("Global size" and "Local size" columns) in the same format as with least-square regression nonlinear models, see Listing~\ref{lst:create-kb}.
\begin{lstlisting}[label=lst:create-kb, caption=How to call script for generating nonlinear model]
python3 generate_decision_tree_model.py \
-i 1070-gemm-reduced_output.csv -t 4:19 -c 2,3,19:62
\end{lstlisting}
The script builds models predicting performance counters using optimized decision trees. The performance counters prediction with decision trees is computationally more efficient than with least-squares model; therefore, we recommend them as a default choice. The decision trees have high precision in densely sampled tuning spaces, but they are poor in extrapolation. Therefore, if only a smaller part of the tuning space is sampled, we recommend testing whether the least-squares method would bring better precision and faster tuning convergence.
The resulting decision tree is stored in files with suffix \texttt{DT.sav} and a list of hardware performance counters predicted by the script has suffix \texttt{DT.sav.pc}. For example, model for the file \texttt{1070-gemm-reduced\_output.csv} is stored in \texttt{1070-gemm-reduced\_output\_DT.sav}.
\subsection*{Execution of Simulated Tuning}
Simulated tuning scripts \texttt{simulated-profiling-searcher.py} perform a search of the autotuning space on a pre-computed tuning space. It requires auxiliary files \texttt{base.py} and \texttt{mlKTTPre\-dictor.py} distributed with KTT. It also requires python3, with installed libraries docopt, numpy, pandas, pickle and sklearn.
Instead of real execution and profiling of autotuned kernels (obtaining their runtime and hardware performance counters), it reads stored raw autotuning data (i.e. just simulates their execution and profiling) and performs a search on them. The advantage of this approach is that it is performed much faster than real tuning (as no compilation, execution and profiling are performed); therefore, the simulated tuning experiments can be performed many times to get statistically relevant data. The simulated run also does not require installation of KTT and GPU we are simulating autotuning for. The convergence of search method is measured and can be compared in means of the number of search steps (equal to the number of kernel executions performed by KTT in a real environment).
Listing \ref{lst:simulated-tuning} shows three examples of running the script. We want to tune gemm-reduced benchmark. The models for predicting performance counters were trained on GeForce GTX 750, and we want to use them to guide the search on GeForce GTX 1070. These three commands differ in the model used. The first one does not predict anything, only reads the performance counters' values from the provided raw tuning data file. The second one uses a decision tree, and the third one uses least-squares nonlinear models.
\begin{lstlisting}[label=lst:simulated-tuning, caption=Example of simulated-profiling-searcher.py.]
python3 -W ignore ./simulated-profiling-searcher.py \
-o 1070-gemm-reduced_output.csv --oc 6.1 --mp 15 --co 1920 \
--cm 750-gemm-reduced_output.csv --ic 5.0 -p 1 -t 4:19 -c 2,3,19:62 \
--compute_bound -e 1000 -i 1000
python3 -W ignore ./simulated-profiling-searcher.py \
-o 1070-gemm-reduced_output.csv --oc 6.1 --mp 15 --co 1920 \
--dt 750-gemm-reduced_output_DT.sav --ic 5.0 -p 1 \
-t 4:19 -c 2,3,19:62 --compute_bound -e 1000 -i 1000
python3 -W ignore ./simulated-profiling-searcher.py \
-o 1070-gemm-reduced_output.csv --oc 6.1 --mp 15 --co 1920 \
--ls 750-gemm-reduced --ic 5.0 -p 1 -t 4:19 -c 2,3,19:62 \
--compute_bound -e 1000 -i 1000
\end{lstlisting}
The script takes multiple arguments:
\begin{itemize}
\item \texttt{-o [raw tuning data file]} the raw tuning data file following the format described in the section Data Description
\item \texttt{--oc [compute capability of GPU used to produce raw tuning data]} e.g. 6.1, if raw tuning data came from GeForce GTX 1070 \footnote{Note that values of several arguments, such as compute capability of GPUs, number of multiprocessor or CUDA cores, and column indexes for computation time, tuning parameters and profiling counters are available in script \texttt{autobench.py} for our raw tuning data. This script is described later in this section.}
\item \texttt{--mp [number of multiprocessors on that GPU]} e.g. 15, if raw tuning data came from GeForce GTX 1070
\item \texttt{--co [number of CUDA cores on that GPU]} e.g. 1920, if raw tuning data came from GeForce GTX 1070
\item one of the following
\begin{itemize}
\item \texttt{--cm [raw tuning data file]} with this option, no prediction of values of performance counters is computed, their actual values are read from the file with given raw tuning data
\item \texttt{--dt [decision tree model file]} with this option, decision tree model is employed to predict values of performance counters
\item \texttt{--ls [prefix for least squares model files]} with this option, least squares nonlinear models are employed to predict values of performance counters
\end{itemize}
\item \texttt{--ic [compute capability of the GPU of training data for model]} e.g. 5.0, if model was trained with data from GeForce GTX 750
\item \texttt{-p [column with computation time]} always 1 in provided raw tuning data
\item \texttt{-t [columns with TP]} 4:19 for gemm-reduced bechmark
\item \texttt{-c [columns with PC]} 2,3,19:62 for gemm-reduced benchmark on GeForce GTX 1070
\item \texttt{-e [number of experiments]} sets how many times the experiment is repeated to get more stable results in case of randomized searchers
\item \texttt{-i [number of iterations]} sets how many tuning iterations (i.e., search steps) are performed per experiment
\item \texttt{--compute-bound} or \texttt{--memory-bound} e.g., \texttt{--compute-bound} as that is the character of gemm-reduced problem
\end{itemize}
For details on the algorithm of profile-based search, please see \cite{Filipovic2021}.
The results of the analysis are stored as CSV files of the following format. The first line contains a header with names of the columns. Each column presents an iteration of the searcher (i.e., the exploration of the next tuning configuration, requiring its profiling). The first column contains the iteration number. The second column contains an average runtime with the standard deviation of the best kernel known in this iteration when the random searcher is utilized. The third column contains an average runtime with the standard deviation of the best kernel known in this iteration when the profile-based searcher is utilized.
The script \texttt{autobench.py} makes it easy to run simulated tuning for all our raw tuning data. It takes two arguments, \texttt{--benchmark} and \texttt{--method}. The option \texttt{--benchmark GPU} runs simulated tuning for benchmarks conv, coulomb, gemm-reduced, mtran and nbody for all GPUs. Option \texttt{--benchmark GEMM} runs simulated tuning for different input sizes of benchmark gemm-reduced on GeForce GTX 1070. The option \texttt{--method} has three possible arguments \texttt{Exact}, \texttt{DecisionTree} or \texttt{LeastSquares} to denote the used model for predicting the values of performance counters. Users may need to modify the script to accommodate the names of folders with data. Moreover, the script can be used as a source of information on possible values of several command-line arguments of \texttt{simulated-profiling-searcher.py}, such as compute capabilities of different GPUs, number of their multiprocessors and CUDA cores, and indexes of columns for computation time, tuning parameters and performance counters in raw tuning data we provide.
We used the simulated tuning to analyze the convergence speed of the profile-based searcher proposed in~\cite{Filipovic2021} and of the random search. During the analysis, the autotuning is performed in the defined number of iterations. In each iteration of the searching process, the runtime of the best kernel found is logged. The autotuning is performed multiple times, so we obtain an average speed of the best kernel for each iteration over multiple autotuning executions.
Other search methods or modifications based on our profile-based searcher might be easily added to scripts and compared consistently.
\subsection*{Execution Real-time Tuning}
Real-time tuning performs a search of the autotuning space without a pre-computed tuning space. The autotuned kernels are actually compiled, executed and profiled during the search. This is far more demanding than the simulated tuning described above, but it makes it possible to measure the actual time per tuning search iteration. The convergence of search method is measured and can be compared in means of tuning time.
Real-time tuning can be executed by running a compiled benchmark. For benchmarks bundled with KTT, the preparation includes the following steps:
\begin{itemize}
\item KTT needs to be compiled with profiling. i.e. \texttt{premake5 --profiling=cupti gmake} needs to be run before building it. In case of older GPU architectures, \\use \texttt{--profiling=cupti-legacy} instead.
\item In the code of the benchmark (in \texttt{cpp} file), a preprocessor macro \texttt{EXHAUSTIVE\_SEARCH} has to be set to 0.
\item The random search is used by default. To test profile-based searcher~\cite{Filipovic2021}, the macro \texttt{USE\_PROFILE\_SEARCHER} has to be set to 1 in the code of the benchmark (in \texttt{cpp} file).
\item The time for autotuning is restricted to a certain value set by macro \texttt{TUNE\_SEC}. This time can be altered, for example, with \texttt{TUNE\_SEC 60} performs autotuning for 60 seconds.
\item The prediction models needed for profile-based searcher have to be in 'KTT/profile-searcher\-/models' folder.
\end{itemize}
Listing \ref{lst:real-time-tuning} shows examples of running a single real-time tuning for each benchmark and saving the log file.
\begin{lstlisting}[label=lst:real-time-tuning, caption=Running a single execution of real-time tuning on every benchmark]
cd KTT/build/x86_64_Release/
./conv_cuda > conv_experiment_1.log
./coulomb_sum_3d_cuda > coulomb_experiment_1.log
./gemm_cuda > gemm_experiment_1.log
./mtran_cuda > mtran_experiment_1.log
./nbody_cuda > nbody_experiment_1.log
\end{lstlisting}
The benchmark executable has two arguments; however, both of them have default values:
\begin{itemize}
\item \texttt{[platform index]} default value 0, cannot be changed when using CUDA
\item \texttt{[device index]} default value 0, users may need to change this if multiple GPUs are available
\end{itemize}
The input size can be modified in the source code of the given benchmark, in the \texttt{main} function. Again, reasonable default values are set. In our evaluation in \cite{Filipovic2021}, we set the input size of benchmarks according to Table~\ref{tab:input-sizes}.
Proper evaluation of searcher's convergence in means of tuning time requires multiple runs of real-time tuning. This can be easily done by executing the benchmark multiple times and generating a log file for each run. For easier processing of the results in multiple log files, we provide script \texttt{histogram.py}. See an example of its usage in Listings~\ref{lst:histogram}.
\begin{lstlisting}[label=lst:histogram, caption=Processing multiple runs of real-time tuning]
python3 ./histogram.py -s gemm-experiments -t 300
\end{lstlisting}
It takes two arguments:
\begin{itemize}
\item \texttt{-s [folder with log files from real-time tuning runs]};
\item \texttt{-t [time]} in seconds that denotes the maximum running time that is analyzed.
\end{itemize}
Results of \texttt{histogram.py} script are stored as CSV files of the following format. The first line is a header containing names of the columns. Each row contains measurement in each second of the autotuning process. The missing row indicates that no new data was available in particular second of the tuning process (this may happen in the initial part of the tuning if the first profiled kernel runs a long time). The rows contain the following columns:
\begin{itemize}
\item the time (in seconds) from the beginning of the autotuning;
\item average runtime of the best kernel known in the corresponding time;
\item standard deviation of the runtime;
\item minimal runtime;
\item maximal runtime.
\end{itemize}
The data are collected from all available executions of the autotuning (all log files in folder passed by argument \texttt{-s}).
\section*{Acknowledgments}
The work was supported from European Regional Development Fund-Project "CERIT Scientific Cloud" (No. CZ.02.1.01\-/0.0/0.0/16\_013/0001802). Computational resources were supplied by the project "e-Infrastruktura CZ" (e-INFRA LM2018140) provided within the program Projects of Large Research, Development and Innovations Infrastructures.
\section*{Declaration of Competing Interest}
The authors declare that they have no known competing
financial interests or personal relationships which have, or could be
perceived to have, influenced the work reported in this article.
\bibliographystyle{model1-num-names}
|
\section{Introduction}
Gaussian beams are among the most commonly encountered optical waves in modern imaging systems. It is thus crucial to further our understanding of the effects optical aberrations can exhibit on Gaussian beams. This is especially true in the recently emerging field of active aberrations correction, colloquially known as adaptive optics. Adaptive optics (AO) is an approach that uses active optical elements for a dynamic manipulation of light amplitude and phase with fine spatial resolution in order to correct for the presence of aberration and drastically increase image quality under experimental conditions \cite{Booth:14,Ji:17}. Practical implementations of AO usually require the choice of a particular orthonormal base into which experimentally measured aberrations can be decomposed. The Zernike polynomials are a frequent and popular choice since they facilitate efficient aberration correction in practice. \cite{booth2007adaptive}.
It has been shown previously that Zernike aberrations are orthogonal only in the case of plane waves and show cross coupling in the case of Gaussian beams. The effects of this cross couplings were analysed based on the Strehl ratio approach \cite{mahajan1994zernike,mahajan1995zernike}. Here we propose an alternative formulation based on analysing the effects Zernike type aberrations in inducing power coupling into higher order Laguerre-Gauss modes \cite{Mah:19,Bond:11,Czuchnowski:20} and use it to gain intuitive understanding of various model situations in AO such as iterative improvement of cross-compensating Zernike aberrations and cascading aberrations. We also extend our approach to shed light on Zernike aberration interactions in the high aberration regimes.
\section{Coupling between Zernike aberrations and Laguerre-Gauss modes in the low aberration regime}
Laguerre-Gauss beams as eigenvectors of the wave equation are inherently orthogonal and therefore do not cross-couple between each other. However it has been shown that Zernike type aberrations can induce cross coupling between different LG-beams \cite{Bond:11}. The coupling coefficient can be expressed as:
\begin{equation}
k^{n,m}_{p,l,p',l'}=\int_S LG_{p,l}\exp(2ikZ^m_n)LG_{p',l'}^*dS
\label{eq:knmplpl}
\end{equation}
in the low aberration regime we can approximate:
\begin{equation}
\exp(2ikZ^m_n)=1+2ikZ^m_n
\end{equation}
which simplifies \textbf{Equation \ref{eq:knmplpl}} to an analytically solvable integral:
\begin{equation}
k^{n,m}_{p,l,p',l'}=\int_0^{2\pi} \int_0^R LG_{p,l}LG_{p',l'}^*(2ikZ^m_n)rdrd\phi
\end{equation}
The equation can be separated into azimuthal and radial parts which can be solved independently, starting with the azimuthal part:
\begin{equation}
\phi^{n,m}_{p,l,p',l'}=\int_0^{2\pi} e^{i\phi(l-l')}\frac{e^{im\phi}+e^{-im\phi}}{2}d\phi=\bigg[\frac{e^{i\phi(l-l'+m)}}{2i(l-l'-m)}+\frac{e^{i\phi(l-l'+m)}}{2i(l-l'-m)}\bigg]_0^{2\pi}
\end{equation}
which specifies the coupling condition:
\begin{equation}
\label{eq:phi}
\phi^{n,m}_{p,l,p',l'} =
\begin{cases}
0 & \text{if $m \neq |l-l'|$}\\
\pi & \text{if $m=|l-l'|$, \ even $Z^m_n$}\\
sgn(l-l') i\pi & \text{if $m=|l-l'|$, \ odd $Z^m_n$}\\
2\pi & \text{if $m=0$, \ $l=l'$}
\end{cases}
\end{equation}
This means that efficient coupling between modes can occur only under the condition that $m=|l-l'|$. As in most imaging applications we are mainly concerned with aberrations acting on the fundamental Gaussian mode ($LG_{00}$) this simplifies the coupling condition to $m=|l'|$. This has fundamental consequences for our analysis as it provides a mapping from particular sets of Z-modes into particular families of LG-beams (\textbf{Figure \ref{fig:1}a}).
\begin{figure}
\includegraphics[width=14cm]{Figure1.eps}
\centering
\caption{\textbf{a.} Coupling from classes of Z-modes to classes of LG-modes depending on which LG beam is assumed as the fundamental beam. \textbf{b.} Schematic representation of aberrated beams ($G_{abr}$) as vectors in the LG space showing the possibility of cross compensation by non-orthogonal aberrated beams ($G_{corr}$). \textbf{c.} Amplitude of coupling coefficients into different LG-modes induced by Zernike aberrations of the spherical aberration class. \textbf{d.} Quantification of the cross-compensation between different modes in the spherical aberration class ($Z_n^0$), expressed as the percentage of power out-coupled from the fundamental Gaussian mode that can be recovered. \textbf{e.} Iterative adaptive optics correction for two cross coupling aberrations. \textbf{f.} Geometric interpretation of iterative adaptive optics correction for two cross coupling aberrations. The iterative improvement is trapped between the two dotted lines which are the orthogonals of the aberration vectors.}
\label{fig:1}
\end{figure}
\subsection{Applications in adaptive optics}
As Zernike polynomials are frequently used as an orthonormal base for applications in adaptive optics \cite{booth2007adaptive} and aberration correction it is interesting to investigate the consequences of the cross-coupling in the context of our framework. In order to do so we first describe an AO task within our theory.
As we operate in the LG-beam space we need to define how the effects of Zernike aberrations on the beam quality can be quantified. The most intuitive metric would be to look at the power fraction coupled into higher order LG-beams. In this general sense the goal of aberration correction could be defined as redirecting the power back into the fundamental Gaussian mode. The coupling condition ($m=|l'|$) allows us to easily map all beams aberrated by a particular class of Zernike aberrations into a subspace of LG-beam vector space:
\begin{equation}
\mathbf{G^{n,m}}=\sum_p k^{n,m}_{p,\pm m}\mathbf{LG_{p,\pm m}}
\end{equation}
where $\mathbf{LG_{p,\pm m}}$ are understood as base vectors and $k^{n,m}_{p,\pm m}$ is the amplitude coupling coefficient \cite{Bond:11} (see \textbf{Figure \ref{fig:1}b, c}). The power coupled into higher order modes can then be simply calculated as the square of the vector length:
\begin{equation}
P_{n,m}=||\mathbf{G^{n,m}}||^2
\end{equation}
Additionally, in the low aberration regime interactions between different Zernike aberrations translate to simple vector additions and their effect on the outcoupled power can be expressed as follows (\textbf{Figure \ref{fig:1}b}):
\begin{equation}
P_{n*,m*}^{n,m}=||\alpha\mathbf{G^{n,m}}+\beta\mathbf{G^{n*,m*}}||^2
\end{equation}
If we assume that $\alpha$ is fixed and $\beta$ can be adjusted to attempt an AO correction the optimal reduction of outcoupled power can be calculated from geometrical relations between vectors:
\begin{equation}
min\{P_{n*,m*}^{n,m}\}=\bigg[ \alpha||\mathbf{G^{n,m}}||\sin\zeta \bigg]^2
\end{equation}
where $\zeta$ is the angle between vectors $\mathbf{G^{n,m}}$ and $\mathbf{G^{n*,m*}}$. In case $\zeta=\pi/2$ the two vectors are orthogonal and unable to cross-compensate, otherwise the vectors can cross-compensate at least residually. We investigate the cross-coupling characteristics in the spherical aberration class ($Z^0_n$) and show that there is significant cross-compensation only between neighbouring aberrations (\textbf{Figure \ref{fig:1}d}). As the angle between vectors was calculated numerically for a finite set of LG-beam it has an uncertainty from discarding the higher order beams. We propose a simple estimation for the upper bound of this uncertainty based on vector properties (\textbf{Appendix A1}) and show that the angle is significantly different from $\pi/2$ for the considered cases.
\subsection{Correcting multiple cross-compensating aberrations}
Now we can use our framework to model a typical correction scheme in indirect adaptive optics where different Zernike corrections are applied and optimized in an independent (mode-by-mode) as well as iterative fashion. This is a particularly relevant situation in applied optics where one can see the effects of aberration cross-compensation:
\begin{equation}
\mathbf{G^{n,m,n*,m*}}=(\alpha_1+\beta_1)\mathbf{G^{n,m}}+(\alpha_2+\beta_2)\mathbf{G^{n*,m*}}
\end{equation}
where $\alpha_1$ and $\alpha_2$ are fixed aberrations and $\beta_1, \beta_2$ are the correction coefficients which are updated iteratively. Aberration cross-compensations will cause the coefficients to evolve over several iterations until they converge at the proper values (\textbf{Figure \ref{fig:1}e}). This behaviour can be intuitively understood by looking at the geometrical interpretation of our theory (\textbf{Figure \ref{fig:1}f}). The situation for only a single fixed aberration is conceptually the same (\textbf{Appendix A2}). It is important to note here that odd ($m < 0$) and even ($m \geq 0$) Zernike aberrations do not cross-compensate, even though they couple into the same LG modes, because they are orthogonal in phase (as can be seen from \textbf{Equation \ref{eq:phi}}).
\begin{figure}
\includegraphics[width=15cm]{Figure2.eps}
\centering
\caption{\textbf{a.} Sequentially cascading two $Z^1_3$ aberrations (coma). \textbf{b.} Powered coupled to LG-modes for a cascading $Z^1_3$ aberration. \textbf{c.} Powered coupled to LG-modes for a cascading $Z^0_2$ aberration. \textbf{d.} Schematic showing $1^{st}$ and $2^{nd}$ order coupling into LG-modes for $Z^{\pm1}_n$-class and $Z^{0}_n$-class aberrations. \textbf{e.} Z-mode decomposition of the residual power coupled into the $LG_{p,0}$-family caused by a cascading $Z^1_3$ aberration (\textbf{panel b}) showing possible cross-compensation between Z-classes. \textbf{f.} Mean coupled mode ($\overline{p}$) quantification for the $Z^{0}_n$-class aberrations showing that higher order $Z^{0}_n$-aberrations couple power into higher order LG-modes. \textbf{g.} Mean coupled mode ($\overline{p}$) quantification for a cascading $Z^{0}_2$ aberration showing that for higher order cascades (and consequently higher order coupling coefficients) power is being transferred into higher order LG-modes resembling the effects of higher order Z-aberrations.}
\label{fig:2}
\end{figure}
\section{Cascading aberrations and the high aberrations regime}
The use of LG space in our framework allows us to look into aberrated beam evolution in a very straightforward manner as LG modes are eigenvectors of the wave equation. This allows us to address questions concerning effects of cascading aberrations where the evolving beam interacts twice with the same type of Zernike aberration (\textbf{Figure \ref{fig:2}a}). In this case we see that re-applying a $Z^1_3$ (coma) aberration to a $Z^1_3$-aberrated Gausian beam will cause coupling into LG-modes characteristic for the sperical aberration ($Z^0_n$) and trefoil ($Z^{\pm2}_n$) Z-mode clases (\textbf{Figure \ref{fig:2}b}).
Conceptually cascading aberrations show similarities to the effects of Zernike aberrations in the high aberrations regime which we will now discuss. In the high aberrations regime it becomes necessary to consider higher order coupling terms which modify the coupling conditions between Z-modes and LG-modes \cite{Bond:14}:
\begin{equation}
\exp(2ikZ^m_n)=1+2ikZ^m_n+2(ikZ^m_n)^2
\end{equation}
The resulting approximation of \textbf{Equation 1} is still solvable analytically in an analogical way \cite{Bond:14} and the resulting coupling condition from the azimutal integral is the following:
\begin{equation}
\label{eq:phi2}
\phi^{n,m,2}_{p,l,p',l'}=
\begin{cases}
0 & \text{if $2m \neq |l-l'|$, \ $l \neq l'$}\\
\frac{\pi}{2} & \text{if $2m=|l-l'|$, \ even $Z^m_n$}\\
-\frac{\pi}{2} & \text{if $2m=|l-l'|$, \ odd $Z^m_n$}\\
\pi & \text{if $l=l'$, $m \neq 0$}\\
2\pi & \text{if $l=l'$, $m=0$}\\
\end{cases}
\end{equation}
We see now that $2^{nd}$ order coupling populates the same LG-modes as a cascading aberration, which allows us to intuitively look at the Z-mode cross-compensation in the high aberration regime (\textbf{Figure \ref{fig:2}a, d}).
We observe that there are two special classes of Zernike aberrations. The $Z^0_n$-class (spherical aberrations) is special, because of the coupling from $Z^0_n$-modes is always constrained to the same class of LG-modes ($LG_{p,0}$) regardless of the strength of the aberration (\textbf{Figure \ref{fig:2}c, d}). Because of this we can define an easy metric to look at the scaling between the different orders of Z-modes and LG-modes. We show that higher order $Z^0_n$-modes couple to higher order LG-modes (\textbf{Figure \ref{fig:2}f}, see \textbf{Appendix A3} for details), and interestingly that cascading lower order aberrations ($Z^{0}_2$) over several iterations start to shift the power coupling into higher order modes (\textbf{Figure \ref{fig:2}g}). Additionally due to \textbf{Equation \ref{eq:phi2}} all Zernike aberrations will cause coupling into the $LG_{p,0}$-class if the $2^{nd}$ order is considered, facilitating possible cross-compensation between different Z-mode classes (\textbf{Figure \ref{fig:2}b, e}). Secondly, the $Z^{\pm1}_n$-class (coma) is special because it densely populates the LG-space as the aberration magnitude becomes larger, which will also facilitate possible cross-compensation. Higher order Zernike aberration classes populate the LG-space in a progressively more sparse manner thus facilitating possible cross-compensation to a lower degree.
\section{Discussion}
With the recent rise of popularity and applicability of adaptive optics in imaging demands a better understanding of the ways optical aberrations interact both with each other as well as with the active optical elements used for AO correction (such as deformable mirrors and spatial light modulators). As most of these devices are being used and calibrated in the Zernike aberration base the limitations of Zernike polynomials as optical aberration models need to be characterised and highlighted. In our work we proposed a novel framework based on analysing aberrated beams in the Laguerre-Gauss mode space as a useful tool for understanding the interactions of Zernike type aberrations with Gaussian laser beams. Our framework provides an intuitive explanation for common phenomena observed in experimental AO applications, namely the need to iterate over the same Zernike modes several times in order to achieve optimal correction results. Additionally, it describes the differences in the cross-coupling behaviour of Zernike aberrations between the low and high aberration regimes. Therefore our work might provide a basis to design superior correction strategies in AO or other applications based on aberrations described by Zernike basis sets.
\section*{Funding}
This work was supported by the European Molecular Biology Laboratory (EMBL).
\section*{Disclosures}
The authors declare no conflicts of interest.
\bibliographystyle{naturemag}
|
\section{Introduction}
Titan is the only moon of the solar system with a thick hazy atmosphere which represents approximately 20\% of its apparent
diameter. This atmosphere is mainly composed of nitrogen and methane. The photo-dissociation of these molecules by the
UV light in the upper part of the atmosphere leads to the production of a large number of other hydrocarbons and nitriles as
trace species and to photochemical haze. This haze is global and completely covers Titan. It controls
the thermal balance through its visible and thermal infrared properties \citep[\textit{e.g.}\xspace][]{Bezard2018}.
It also veils the lower atmosphere and the surface that can be perceived in a few methane windows in near infrared.
Titan's haze was first resolved in the 80's by Pioneer 11 \citep{Smith1980} and the two Voyagers
\citep{Smith1981, Smith1982, Sromovsky1981}. It had several remarkable structures: a northern (winter)
polar hood, an interhemispheric asymmetry and a thin global detached haze layer (DHL) above the main global haze layer. It was
thought that the detached haze layer had a dynamical origin \citep{Smith1981}. Photometric analyses provided a means
to derive the extinction properties of both haze layers and to evaluate the effective radius of the aerosols in the detached
haze ($\simeq$ 0.3\micron) and in the main haze ($\simeq$ 0.4\micron) layers \citep{Rages1983, Rages1983a}. Analysis of Voyager images showed
that the detached haze layer appears due to a strong depletion of aerosol extinction around 300 km, yielding a distinct
layer above the main haze layer with a maximum extinction located around 350 km \citep{Rages1983}. Its horizontal extent was
very stable in pressure and it was reported at all the southern latitudes up to \ang{45}N where it connected to the northern
polar hood. The detached haze layer was re-observed twenty years after the Voyager flybys during Cassini first flyby in 2004
\citep{Porco2005}. The main change was in its altitude location at 500 km, which was 150 km higher than in 1981.
Again, it appeared as a fairly homogeneous global shell above the main haze layer at a constant altitude and
merged with the northern polarhood. Notably, while Voyager observations were performed after the northern spring equinox,
Cassini early observations occurred during the northern winter, that is half a season earlier (Fig.~\ref{fig:titan_seasons}).
\begin{figure*}[ht!]
\plotone{Titan_seasons.pdf}
\caption{Titan orbital position as a function of the season, reported as solar longitude position ($L_s$).
The Cassini mission covered almost half a Titan year. The Pioneer and Voyager flybys are also reported
as well as the Huygens landing and ground-based stellar occultations observed on Earth.}
\label{fig:titan_seasons}
\end{figure*}
\cite{Toon1992} first attempted to explain the observation of the DHL. They used a 1D microphysical
model where an \emph{ad hoc} vertical wind maintains aloft the DHL particles at a constant altitude above the main
haze layer. Alternative scenarios were proposed to explain the DHL from purely microphysical processes. \cite{Chassefiere1995}
investigated the case of two different aerosol production layers. They proposed that the uppermost layer (500-1000 km)
produces fluffy aggregates that could be swept horizontally by winds, generating a detached haze layer. They also
proposed an alternative scenario where aerosols settle downward and interact with macromolecules from the main haze layer,
produced by the lower production zone (around 350-400 km). In the latter case, the interaction would produce by some
way an optical gap. However, they favored the scenario involving winds which would match all the constraints
known at that time. In the same vein, \cite{Lavvas2009} proposed a scenario based on a purely microphysical process. Aerosols
are produced at high altitude, as per the \cite{Chassefiere1995} hypothesis, growing as spheres down to levels around 500 km.
But there, the detached haze is produced by a sudden change in the fractal dimension of the aerosols. This produces a
sharp change in the microphysical properties, and an artificial optical gap. However, it is unclear how this model for the
production of the detached haze layer would be augmented to account for the seasonal evolution of the altitude,
disappearance, and reappearance (as described below).
Later, with a 2D-General Climate Model (GCM) accounting for the transport of haze by dynamics and the radiative
feedback, it was possible to reproduce and explain the mechanism that produces the DHL \citep{Rannou2002}. It was also
demonstrated that this feedback strongly enhances the wind speed due to the thick polar haze cap near the winter pole.
In return, this cap enhances the cooling to space during the polar night \citep{Rannou2004} and
reinforces the circulation. Due to Titan's obliquity (\ang{27}) and the slow rotation rate, seasons are well
marked and Hadley circulation cells span both hemispheres. This situation leads to the formation of a
broad ascending circulation in the summer hemisphere able to lift aerosols up to high altitudes where they remain
suspended and are transported through mid-latitudes to the winter polar region where they are transported by
subsidence (Fig.~\ref{fig:titan_atm_circulation}). In this scenario, the location of the DHL corresponds to the area
where the settling speed is compensated by upward wind and evolves with seasonal changes of illumination.
More sophisticated 3D-GCMs improved the understanding of the haze cycle, including the formation of the detached haze,
and confirmed this picture \citep{Lebonnois2012,Larson2015}. This formation mechanism implies that the DHL
is a blending of aerosols newly produced and falling from above and older and larger aerosols produced in the
stratosphere and lifted by circulation.
Although the GCM results differ in some aspects with observations, they are able to capture the main picture behind
the existence of the DHL.
\begin{figure}[ht!]
\plotone{Atmsopheric_circulation.pdf}
\caption{Synthetic representation of Titan atmospheric circulation as a function of season.}
\label{fig:titan_atm_circulation}
\end{figure}
Photometric studies performed with Cassini data taken before the equinox in 2009 provide complementary observations
of the DHL \citep{Cours2011, Koskinen2011, Seignovert2017}.
On one hand, the authors used the light intensity scattered at the limb in UV (338 nm)
at different phase angles measured by ISS. On the other hand, a single value of the tangential opacity in VUV (187 nm)
was retrieved from UVIS observations during stellar occultations.
The results show the presence of large aerosols in the DHL with an effective bulk radius $\simeq$ 0.2\micron,
producing all the UV scattering, while small nanometric aerosols are needed to explain most of the VUV
extinction \citep{Cours2011}, which is quite consistent with a DHL made of two different populations of
aerosols.
\cite{West2011} also reported a rapid collapse of the detached haze layer starting just before the equinox. The altitude of
the DHL descended by about 80 km in 200 terrestrial days and by 30 km more in about 300 terrestrial days. A simple
extrapolation of the altitude of the DHL with time indicated that it would be at the same altitude as observed by Voyager
exactly one Titan year after the Voyager epoch. \cite{West2011} concluded that such a result was coherent with a seasonal
cycle of the DHL. They compared their results with a 2D-GCM \citep{Rannou2002} and made a prediction about the reappearance of the
DHL several years later (2013-2016) at its initial altitude (around 500 km). \cite{Lebonnois2012} and \cite{Larson2015} made
similar predictions but with a reappearance of the DHL a bit later, around the next northern summer solstice ($L_S=\ang{95}$
and $\ang{70}-\ang{80}$, respectively). In practice, \cite{West2018} found that the DHL reappeared in early 2016
($L_S=\ang{73}-\ang{76}$) at 480 km, several months before the solstice (mid-2017). They followed the cycle of the DHL at the equator
and retrieved the haze extinction profile in the CL1-UV3 filter combination.
Its reappearance was much more complex than predicted. This early 2016 detached haze layer
dropped in altitude down to 470 km within a terrestrial year and vanished while a new DHL emerged again around 500 km. This
new layer appeared quite stable until the end of the Cassini mission (September 2017, $L_S=\ang{91}$). Unfortunately, no other
data exist to further probe the DHL and nothing is known about the fate of the detached haze after this date.
\medskip
In the present work, we perform a systematic latitude-altitude mapping of the detached haze layer in the range 350 to 600 km altitude.
This covers the period between July 2004 (half a season after the northern winter solstice) and the end of
the mission in September 2017 (after the summer solstice).
We used all the UV3 observations acquired by the Cassini Narrow Angle Camera (NAC) of ISS.
We used exactly the same model as \cite{West2018}, that is a ray tracing model in spherical shell geometry for the single-scattering albedo
and a correction for multiples scattering.
The outline of the article is as follows. In the next section (section~\ref{seq:observations}), we first give a global presentation of
the available data and the criteria we use to select images.
Then we describe the main principle of the retrieval model and the retrieval method.
In section~\ref{seq:lat_variations} we present the results of the photometric analysis as latitude -
altitude panels showing the spatial distribution of the DHL and the upper part of the main haze layer.
The seasonal cycle of the DHL is split in four specific periods between 2004 and 2017.
This section has then 4 subsections for each period where we explain in detail the main characteristics of the haze and
its evolution.
Section~\ref{seq:local_variations} is dedicated to the study of specific sets of observations that probe short time scales,
short term or diurnal variations. We first describe how the data were selected and then what they reveal about Titan's atmosphere.
In section~\ref{seq:comparisons}, we make comparisons between our results and results obtained at the same location and the same time with UVIS.
We also make comparisons between our results and prediction made by two Titan 3D-GCMs about the detached haze layer and its evolution.
The conclusion and the perspective of this work are given in section~\ref{seq:conclusions}.
\section{Observations and models}
\label{seq:observations}
\subsection{Selection of observations}
We conduct our survey on 138 images\footnote{See Data Availability section to get the list and the results for all the ISS images analyzed.}
taken by the Cassini Image Science Sub-System Narrow Angle Camera (ISS-NAC) with the
clear and ultra-violet filter combination CL1-UV3. At this wavelength (338 nm), ISS is sensitive to haze in the
upper part of Titan's atmosphere (300-600 km altitude). We choose the best sample among the 317 images available on the NASA Planetary Data System (PDS)
to get the highest temporal and phase coverage. For the main study we kept only images taken with at least a one-day gap.
We also kept specific sets of observations made a few hours apart to study short-term variations.
On average, the selected images are separated by 39 Earth days, \textit{i.e.}\xspace 2.5 Titan days (Fig.~\ref{fig:img_sampling}).
Although our sampling is not evenly distributed, due to orbital constraints and mission schedule, at least 90\% of the selected
images are separated by less than 120 Earth days, \textit{i.e.}\xspace 7.5 Titan days. Two main gaps of data can be observed.
The first is between 28 March 2008 and 25 January 2009 (302 Earth days/19 Titan days) and 26 November 2010 and 9 September 2011 (286 Earth days/18 Titan days).
\begin{figure*}[ht!]
\plotone{IMG_interval.pdf}
\caption{Timeline of the ISS/NAC CL1-UV3 images analyzed (vertical black ticks). The number of observations per 2 Earth months period (\textit{i.e.}\xspace 4 Titan days / $\Delta Ls \approx \ang{2}$) is reported with a green gradient scale. The gray area correspond to the gaps with no available observations.}
\label{fig:img_sampling}
\end{figure*}
The selected images are calibrated using the CISSCAL routines (v3.8) provided on the Planetary Data System. To improve
the signal to noise ratio on the limb profile, we deconvolved the images with a Poisson Maximum A Posteriori
(PMAP) method using the point spread function (PSF) measured in-flight \citep{West2010, Knowles2020}.
This deconvolution method is known to be efficient to restore fine strutures in astronomical images \citep{Hunt1996}.
In our case, the main source of light is always Titan and is resolved inside the field of view of the camera (Fig.~\ref{fig:model_uncertainties}a). Therefore, we don't expect to see a significant contribution of the stray light in this configuration \citep{West2010}.
The image pointing is initialized with the SPICE kernels \citep{Acton1996, Annex2020}. Since the surface of Titan is not visible in the UV, we improved
the location of Titan's center by fitting the limb intensity. Then, we calculated the planetocentric coordinates of each
pixel and their tangent point altitude with respect to a mean spherical body with a radius of 2575 km.
Intensity profiles are extracted every \ang{5} on both sides of the limb. Depending on the
latitude of the Sub-Cassini point on the ground, the sampling in latitude is not evenly distributed for each image.
Therefore, the polar latitudes are usually less covered than the equator.
Also, the solar illumination changes drastically during the
season between the northern mid-winter to summer, which restricts our ability to see both poles at the same time.
On average, the image pixel scale is about $10 \pm 5$~km.
\subsection{Model of scattering at the limb}
To retrieve the haze extinction profiles from the intensity observations, we model the synthetic
radiance factor ($I/F$) with a single scattering ray tracing model in a spherical shell geometry.
The effect of multiple scattering is accounted for as a correction $\varrho_k (z)$
applied to the volume scattering along line of sight.
This technique was used successfully several times before \citep[\textit{e.g.}\xspace][]{Rages1983, Rannou1997, Seignovert2017, West2018}.
In the detached haze layer, the multiple scattering is mainly produced by the light coming from the main haze below. To
evaluate $\varrho_k$, we use a representative vertical profile of the main haze that
reproduce the observed intensity of Titan in the UV. With a radiative transfer model \citep[SHDOMPP, from ][]{Evans1998}
we have access to the complete radiative source function at each level of the atmosphere as well as the optical properties (haze absorption and scattering and Rayleigh scattering). We are then able to compare the
intensity that is scattered in the direction of the observer from the direct sun only and from the direct sun and the scattered
field coming from below. $\varrho_k$ is defined as the ratio of multiple scattering to single scattering toward the observer
for a given altitude and as a function of the incident and emergent angles. This parameter is pre-computed as a function of
altitude, incident and emergent angles and saved in a look-up table \citep[see.][for details]{West2018}.
We find that the multiple scattering increases the scattered intensity at the limb of Titan in the UV by a ratio between
1.05 and 1.20, depending on the geometry of the observation.
This effect is included in our model. We also checked that significant changes in the main haze (single scattering albedo, the opacity and the vertical scale height) only affect the value of $\varrho_k$ by a few percent and can be neglected.
We discretize the atmosphere in $N = 60$ irregular layers of various thickness : $\Delta z =$ 50 km from the
ground to 200 km, $\Delta z =$ 25 km from 200 to 300 km, $\Delta z =$ 10 km from 300 to 400 km and from 550 to 700 km.
Finally, we used $\Delta z$= 5 km between 400 and 550 km. This grid allows us to take advantage of the spatial resolution
of the ISS NAC camera in the region of interest where is mainly located the DHL.
We can write the outgoing $I/F (z)$ as:
\begin{equation}
I/F (z) = \sum_{i=0}^{n_x-1} \int\limits_{x_k}^{x_{k+1}}
\frac{\left< \varpi P(\Theta) \right>_k}{4}
e^{-\left( \tau^i_k\left(z\right) + \tau^e_k\left(z\right) \right)}
\beta_k\left(z\right) \varrho_k\left(z\right) d{x}
\label{eq:west2017_sup_limb}
\end{equation}
where the summation is performed on the $n_x-1$ segments defined by the intersections of the line of sight and the spherical shell boundaries. The impact parameter $z$ (the lowest altitude reached by the line of sight) is given by the bottom of the $n^\mathrm{th}$ layer crossed. Therefore, each layer of the atmosphere is crossed twice. $x$ is the abscissa along the line of sight. $\tau^i_k$ and $\tau^e_k$ are the opacities along the incident and emergent paths.
$\left< \varpi P(\Theta)\right>_k$ is the average of the product of the single scattering and the phase function at the scattering angle of the observation $\Theta$ for the layer crossed on $x_k$. $\beta_k(z)$ is the local extinction coefficient at altitude $z$ (and the product of the haze cross-section and the local number density). Here, the altitude $z(x)$ is the local altitude at point of abscissa $x$ along the line of sight.
\begin{figure*}[!ht]
\includegraphics[width=.4\textwidth]{N1551888681_sampling.pdf}
\includegraphics[width=.57\textwidth]{Model_uncertainties.pdf}
\caption{Left: Log scale representation of the $I/F$ for the image
\textbf{N1551888681\_1} taken in March, 2007. The data are sampled at the limb by
bins of \ang{5}. The red area correspond to the bin at the photometric
equator ([\ang{45}S, \ang{50}S] and [\ang{76}W, \ang{79}W]). The pixel scale
(7.9 km) is represented in the zoomed area by the yellow rectangle.
An altitude grid is also represented where we can see the peak of intensity of the DHL
located at 500 km. Right: the red curve represents the inverted $I/F$ profile fitted
on the resampled data (orange error bars) within a \ang{5} bin in latitude.
The uncertainties on the retrieved extinction profile $\beta$ (in blue) is calculated to match the $1 \sigma$ distribution of the $I/F$ pixels. The model fits the observations down to 300 km.
The uncertainties increase rapidly when observed ($I/F$) is close
to the noise level ($3\cdot10^{-3}$).}
\label{fig:model_uncertainties}
\end{figure*}
\subsection{Retrieval method}
Based on our previous work \citep{Seignovert2017, West2018}, we make the assumption that the optical properties $\left<\varpi P(\Theta)\right>_k$ of the aerosols are constant in the upper part of Titan's atmosphere. This allows us to focus our study only on the retrieval of the extinction along the line of sight.
In our model, the $I/F (z)$ intensity profiles depend on the set haze extinction profile $\beta(z)$ and the viewing geometry of the observation (incidence, emergence and phase angles).
We assume no horizontal in-homogeneity along the line of sight \citep{Seignovert2017}. We retrieve a set of extinction values $\beta_i$ (the vector ${\beta}$), with $i$ the indices of the layers, that matches the values of the $I/F_i$ (the vector $I/F$).
From the Eq. (\ref{eq:west2017_sup_limb}), it is possible to cast the scattered intensity $I/F_i$, as a function of the extinction $\beta_j$ with $j \le i$. This forms a non-linear triangular system. To find the vector $\beta$, we have to solve the formal equation:
\begin{equation}
I/F = G(\beta)
\end{equation}
where ${G}$ is a nonlinear function which depends on $\beta_i$ and on the viewing geometry of the observation.
We solve the system by minimizing globally the difference between the modeled $I/F$ and the observations
using a Levenberg-Marquardt minimization. Therefore, we obtain simultaneously all the $\beta_i$ at once.
Moreover, the tangential opacity along a line of sight ($\tau_{los}$) is considered opaque when it reaches 3 (usually around 300 km in the UV).
Beyond this threshold, we do not retrieve the value of $\beta$.
An example of inversion is presented in the figure~\ref{fig:model_uncertainties}.
The uncertainties associated to the intensity profile ($I/F$) correspond to a $1 \sigma$ distribution of the observed $I/F$ for a given altitude within a \ang{5} bin in latitude.
The number of pixels within a \ang{5} bin varies from one image to another and depends on the viewing geometry and the altitude sampled. Usually we have more than 1500 pixels per profile between 0 and 700 km.
In the worse case scenarios (N1630432142\_1 at $6 \sigma$ of the average pixel scale) we have at least 120 pixels per profile.
This allow us to resampled our observed profile on a vertical grid with layers smaller than the image pixel scale in the 500 km region.
The uncertainties in the extinction profile ($\beta$) correspond to the minimum and maximum values required to fit the data within $1 \sigma$ of the observed $I/F$ in each layer independently \citep[contrary to][ where the errors were computed in all the layers simultaneously]{West2018}.
\section{Seasonal cycle of the haze extinction}
\label{seq:lat_variations}
In order to provide a detailed explanation of the complex latitudinal variability of the detached haze layer, we present
some of the key images that we have analyzed. Based on our previous work, focused on the evolution of the detached haze in the
equatorial region \citep{West2018}, we define four different seasonal phase characteristics of the evolution of the detached haze layer.
Between 2004 and 2008, the DHL was stable in altitude and extinction profile.
Between 2008 and 2012, it settled and disappeared in the main haze below the 300 km.
During the period 2012-2016, the DHL was not observed and only sporadic transitory layers showed up in the UV3 images.
After 2016 and up to the end of the Cassini mission, the DHL reappeared following a complex pattern.
Here we illustrate the complete time and latitude survey, covering a period of time which includes about half a Titan year.
This is valuable because it encompasses the equinoctial transition period of 2009.
For each phase, we display the altitude and latitude distribution of the instantaneous haze extinction coefficient retrieved from intensities at the illuminated limb of Titan.
The same color scale is applied to all the panels in order to keep a consistent view on the whole data-set.
It corresponds to the range where the model have the best sensitivity at the UV wavelengths. In some cases, smaller features might be harder to highlight with this color scheme, however, all the results from theses panels are available to the reader for more in depth considerations (see Data Availablity section).
Locations where no data are available are left as blank areas on the panels.
\subsection{Period 1: A very well delimited detached haze Layer during the Northern Winter (2004-2008) - $L_s=\ang{300}-\ang{340}$}
At its arrival in the Saturnian system in 2004, Cassini observed a single detached haze layer at 500 km altitude
(Fig.~\ref{fig:dhl_2004_2008}a) similar to the one observed at 350 km by Voyager 24 years before
\citep{Smith1981}. At that moment, Titan was two years after the winter solstice in the northern hemisphere, at $L_s=\ang{300}$.
With Cassini, we see that, in the southern hemisphere, the haze layer was completely detached from the main haze layer.
The haze extinction was at least one order of magnitude smaller inside the depleted zone (470 km) than in the main and
the detached haze layers (below 450 and at 500 km respectively).
Between the equator and up to about \ang{60}N it presented a local depletion in extinction
of a factor 10. There, the separation with the main haze is not as distinct as in the south, but is still sufficiently significant to
defined a detached haze layer.
The altitude of the depletion zone decreased by about 50 km between latitude \ang{30}N and \ang{60}N.
The detached haze layer merged with the polar hood beyond \ang{60}N. This description of the detached haze layer at
the beginning of Cassini mission is very consistent with the results obtained from stellar occultation in 2003 \citep{Sicardy2006}.
Throughout the period 2004-2008 the detached haze layer was quite stable in shape and altitude, with a maximum of extinction
at $500 \pm 20$ km. The top of the main haze layer was located around $450 \pm 20$ km below \ang{30}N and dropped
by 50 km between \ang{30} and \ang{60}N.
\begin{figure*}[!ht]
\plotone{Lat_beta-2004_2008.pdf}
\includegraphics[width=.5\textwidth]{Extinction_colorbar.pdf}
\caption{Latitudinal haze extinction profile ($\beta$) retrieved for 6 images taken between 2004 and 2008
($L_s=\ang{300}-\ang{340}$) showing a stable DHL at $500 \pm 20$ km altitude.
The color schema is fixed for all the figures to make direct comparison
between the different panels. The seasonal solar longitude ($L_s$) and the observation phase angle are
also provided in each panel.}
\label{fig:dhl_2004_2008}
\end{figure*}
However, there are noticeable variations in haze extinction. During several months, the
detached haze remained stable, but in December 2004 (Fig.~\ref{fig:dhl_2004_2008}b),
the detached haze layer extinction was found to be a factor of 10 lower than previously at almost all latitudes south
of \ang{30}N, and about half a decade above \ang{30}N. The polar hood and the main haze do not show a similar decrease.
In the following periods, (Fig.~\ref{fig:dhl_2004_2008}c, d and e),
the detached haze was partially restored, but not with the same amount of extinction as before. Only observations in
2007 (Fig.~\ref{fig:dhl_2004_2008}f) show extinctions in the detached haze layer comparable to those seen before
the decrease. We note that the decrease of extinction below 370 km in the polar hood above \ang{50}N
(Fig.~\ref{fig:dhl_2004_2008}e) is at the limit of sensitivity of the UV filter. The stability of the large-scale
structure of the detached haze layer is related to the steady state of the large-scale circulation during all the winter.
The observation of October 2007 (Fig.~\ref{fig:dhl_2004_2008}f) is the last view that we have of this stable state
before the seasonal turnover.
During this period, the detached haze also has a strong layering with, at some latitudes, distinct decks which are not
continuous and rather appear as foliation. This feature is more pronounced in some observations, for instance from
June 2005 to May 2006, but does not shows up in October 2007, except marginally beyond \ang{30}N. The foliated detached
haze layer has a larger geometrical thickness than before December 2004.
As we will see in the section~\ref{seq:local_variations}, the detached haze layer exhibits some longitudinal or diurnal variability
that limits our ability to interpret details of features observed in single images. The small-scale features
could depend on the local short-term dynamics such as initia-gravity waves.
\subsection{Period 2: Drop and disappearance of the main haze layer around the Vernal Equinox (2008-2012) - $L_s=\ang{340}-\ang{30}$}
A precursor sign of the drop of the detached haze can be seen in March 2008 (Fig.~\ref{fig:dhl_2008_2012}a).
The main haze starts an initial contraction around \ang{35}S. There, the depleted zone is almost
75 km thick at its maximum. In January 2009, the main haze continued to fall from 425 km down to 375 km
while the detached haze layer remained around 500 km (Fig.~\ref{fig:dhl_2008_2012}b). After the drop
of the main haze in early 2009, the detached haze starts its own descent in June 2009, just before the equinox
(Fig.~\ref{fig:dhl_2008_2012}c). This delay in collapse increased the apparent thickness of the depletion
zone between the two haze layers.
\begin{figure*}[!ht]
\plotone{Lat_beta-2008_2012.pdf}
\includegraphics[width=.5\textwidth]{Extinction_colorbar.pdf}
\caption{Same as the figure~\ref{fig:dhl_2004_2008} for 8 images taken between 2008 and 2012
($L_s=\ang{340}-\ang{30}$) showing the drop and disappearance of the DHL.
The color schema is identical to the figure~\ref{fig:dhl_2004_2008} to provide
direct comparisons. In this case, the altitude range is extended down to 300 km
(where the model is less reliable).}
\label{fig:dhl_2008_2012}
\end{figure*}
As for the main haze, the detached haze collapses first in the Southern hemisphere, from 500 km to 425 km, and
then at the equator and in the northern hemisphere (Fig.~\ref{fig:dhl_2008_2012}c).
This is associated with the circulation turnover affecting first the summer hemisphere ascending branch.
With time, the detached haze gradually settled in altitude and finally disappear below 300 km.
Later ISS observations made with the Blue and Green filters visually show that the main DHL continues
its descend below 300 km during 2011. Since our current model was only tested for UV observations, we were not able
to observe its merge with main haze.
The complete collapse of the detached haze, as it appeared in the UV3 filter, is displayed in Fig.~\ref{fig:dhl_2008_2012}c to
Fig.~\ref{fig:dhl_2008_2012}h. We note that the column extinction is smaller at equator than at
other latitudes, and this is the case during entire period of the collapse.
During the fall, a second thin detached haze layer, at planetary scale, is evident above the collapsing detached
haze layer. In January 2010 (Fig.~\ref{fig:dhl_2008_2012}e), the detached haze layer was located between 375 and 400 km.
We can still see a double deck of haze, and this time the detached haze appears higher at the equator compared to the two
hemispheres, producing an arch. The haze peak extinction has globally increased by a factor of two due to sedimentation
in denser layers.
In August 2010 (Fig.~\ref{fig:dhl_2008_2012}f), one year after equinox, the detached haze layer continued
its drop down to 375 km around \ang{40}S and 400 km at the equator. It has gained in complexity with
multiple secondary layers up to 520 km. The detached haze formed a remarkable arch with a difference of about 50 km
in altitude between the equator and the poles as previously noticed by~\cite{West2011}.
This observation and the next one correspond to the same seasonal phase of Voyager flybys ($L_s=\ang{8}$ and \ang{18}).
They can be compared directly.
We now know that this season was a time of rapid change, and that the Voyager probes observed transient situations.
Voyager also observed the detached haze higher near equator than elsewhere \citep{Rages1983, Rannou2000}.
Due to orbital constraints and mission planning, the next observation was made in September 2011
(Fig.~\ref{fig:dhl_2008_2012}g). The detached haze layer was, at that time, well below the level of the polar hoods.
Again, secondary detached layers show up as high as 470 and 520 km.
The south polar hood was not present in August 2010 (Fig.~\ref{fig:dhl_2008_2012}f) and appeared in less than 13 months.
This indicates that the circulation started to reverse around the equinox and the southward circulation sent haze to
the southern polar region and produced a polar hood. The change in haze distribution
is a very good indication of the timing of the equinoctial circulation turnover, as discussed later. We note
that the strong haze depletion at 300 km and between \ang{30}S and \ang{20}N is real (and visible in the I/F profiles) but may be exaggerated at \ang{20}N
due to the limit of the retrieval procedure. At this altitude level, Titan's atmosphere is opaque to UV radiation
(see Fig.~\ref{fig:model_uncertainties}) and does not allow us to follow the main depletion below this altitude.
The last UV3 image we have showing a detached haze layer was taken in February of 2012 (Fig.~\ref{fig:dhl_2008_2012}h). At that
time, the initial detached haze had completely disappeared and the secondary detached haze layer was still descending
and had reached 400 km altitude. The secondary detached haze layer is not well delineated by a layer strongly depleted in aerosols.
The south polar hood increased its latitudinal extent northward to \ang{50}S and became larger than the northern
polar hood which tends to retreat.
\subsection{Period 3: Absence of the DHL, with sporadic transitory layers after the Northern Spring Equinox (2012-2015) - $L_s=\ang{30}-\ang{75}$}
During this period (Fig.~\ref{fig:dhl_2012_2015}), the main haze layer has large-scale structures which slowly evolve under the influence of the
large-scale circulation. The south and north polar hoods are still visible and they evolve
with time. Superimposed on this background haze, transient structures show up and disappear from one observation
to the other. At some moments, large-scale detached hazes appear. They differ from the detached haze seen at the
beginning of the mission because they are not stable in time and in altitude.
\begin{figure*}[!ht]
\plotone{Lat_beta-2012_2015.pdf}
\includegraphics[width=.5\textwidth]{Extinction_colorbar.pdf}
\caption{Same as the figures~\ref{fig:dhl_2004_2008} and~\ref{fig:dhl_2008_2012}
for 4 images taken between 2012 and 2015 ($L_s=\ang{30}-\ang{75}$) showing sporadic
transitory layers and the absence of a stable DHL.
The color scheme is the same as in previous figures and the altitude extends down to 350 km.}
\label{fig:dhl_2012_2015}
\end{figure*}
In April 2012 (Fig.~\ref{fig:dhl_2012_2015}a), the detached haze has completely disappeared, except some residual
structures around the equator at 370 km around the equator.
These relics of the last detached haze layer are almost not perceptible in the corresponding I/F profile. At other
latitudes, we can see only the main haze with a marked south polar hood and a small increase of extinction above
\ang{65}N that could be the residual north polar hood. Sometimes, detached layers emerge from the background with large latitudinal
extent (e.g. detached haze at 500 km Fig.~\ref{fig:dhl_2012_2015}b). However, they only remain for a short time
and are not seen in the following observations.
In August 2014 (Fig.~\ref{fig:dhl_2012_2015}c) we observed a plume of aerosol between \ang{10}S and \ang{25}N,
reaching 530 km. A detached haze layer seems to spread from this plume toward the north and the south. This
detached haze is around 500 km altitude, descending to 470 km at \ang{50}S (and probably even further south). In the
north, the detached haze does not extend further than \ang{50}N and remains at 500 km. This indicates an
atmosphere circulation flowing from equator toward the south pole. These aerosols seem to originate
from the equatorial part of the main haze.
Most of the observations between 2012 and the end of 2015 are featureless as in figure~\ref{fig:dhl_2012_2015}d)
taken in October, 2015. During this period, the main haze has a uniform scale height of 45 km and
with an homogeneous extinction at the planetary scale.
\subsection{Period 4: Reappearance of a new and weaker detached haze layer around the Summer Solstice (2015-2017) - $L_s=\ang{75}-\ang{95}$}
The first occurrence of the stable detached haze in this period was seen on 3$^{rd}$ December 2015
(Fig.~\ref{fig:dhl_2015_2017}a). It
does not at first appear different from the previous sporadic detached haze layers observed in the period 2012-2015. Its signature is very weak, but, after
this observation, the detached haze was more pronounced and present in each following observation. The detached haze became stable in time,
similar to the detached haze before the equinox. Therefore, we consider this date as a limit to the beginning of the
reappearance of the detached haze ($L_s=\ang{74}$). The evolution of the haze during this period is displayed in
Fig.~\ref{fig:dhl_2015_2017}. These observations validate the long awaited reappearance of the detached haze layer,
just before the end of the Cassini mission in September 2017.
\begin{figure*}[!ht]
\plotone{Lat_beta-2015_2017.pdf}
\includegraphics[width=.5\textwidth]{Extinction_colorbar.pdf}
\caption{Same as the figures~\ref{fig:dhl_2004_2008}, \ref{fig:dhl_2008_2012}
and~\ref{fig:dhl_2012_2015} for 6 images taken between 2015 and 2017
($L_s=\ang{75}-\ang{95}$) during the reappearance of the DHL.
The image \textbf{N1884018021\_1} is one of the very last observations of Titan before
the end of the Cassini mission in September 2017.}
\label{fig:dhl_2015_2017}
\end{figure*}
The detached haze layer is not very well defined but it can be perceived at all latitudes
around 490 - 520 km (Fig.~\ref{fig:dhl_2015_2017}a). We notice a contraction
the main haze: the top of the haze dropped by 50 km in January 2016 compared to October 2015
(Fig.~\ref{fig:dhl_2012_2015}d). The main elements of the reappearance (altitude and date) confirm the
predictions made by the general circulation models \citep{Lebonnois2012,Larson2015} and the prediction
reported in \cite{West2011}. The detailed comparison between the observations and the GCM is discussed further
in section~\ref{seq:comparisons}.
With time, the detached haze layer became more distinct and the zone of depletion more pronounced, especially in the
northern hemisphere (Fig.~\ref{fig:dhl_2015_2017}b). The situation seems to be analogous to an early stage of
the structure observed in 2004 (Fig~\ref{fig:dhl_2004_2008}a, but in the opposite hemisphere).
However, although the detached haze persists in time,
it also settles and almost merges with the main haze in October, 2016 (Fig.~\ref{fig:dhl_2015_2017}c).
In the following observations (Figs.~\ref{fig:dhl_2015_2017}d and e), we are
witnessing the complex evolution of the newly-formed detached haze that merged with the main layer southward
of \ang{35}N while it remained stable around 450 km northward of \ang{35}N and seems to vanish in May, 2017 rather than settle.
This structure was still observed in the very last image of Titan taken by Cassini just before its final plunge into Saturn
(Fig.~\ref{fig:dhl_2015_2017}f) in September, 2017.
A secondary detached haze layer appeared in October, 2016 in the southern hemisphere at high altitude (around 520 km in Fig.~\ref{fig:dhl_2015_2017}c). Its northern boundary is not well defined. This new structure will be persistent with
time, at planetary scale, up to the end of the Cassini mission but it gradually descended. The results reported by \cite{West2018}
concern the detached haze at the equator only. Although they already revealed a complex behavior of the detached haze layer, the
present observations show a dichotomy between the two hemispheres. The detail of the evolution, the split in a double layer
structure, and the formation and disappearance of several structures were completely unexpected. According to the GCMs mentioned earlier, six years
after equinox, the post-equinoctial circulation was supposed to be already installed with a planetary-scale circulation cell
from the southern hemisphere to the north polar region. Apparently, this is not the case in the observations.
Cassini observations from 2004 to 2017 do not completely cover half a Titan year. The first and last observations
were taken almost at the opposite season, $L_s=\ang{297}$ and \ang{94} respectively (\emph{i.e.} \ang{157} apart).
This prevents direct comparisons of the detached haze at opposite seasonal phases,
although both 2004 and 2017 images are taken more than a season after the previous solstice.
\section{Local and short-term variability of the detached haze layer}
\label{seq:local_variations}
So far, we have considered the evolution of the detached haze layer in the frame of the seasonal change.
We then discussed the long term-evolution at the planetary scale as a function of latitude.
In this section we consider sets of observations to characterize short-term and local behavior of the detached haze.
These characterizations could only be conducted on a limited number of observations and require very specific acquisition geometries. They allow us to observe localized, secondary order variations of the detached haze layer.
First, we choose several images taken a few hours apart to
evaluate the hourly variability of the detached haze. Next, we consider
observations at low phase angle which show simultaneously the two limbs of Titan. And, finally,
we consider observations taken from a near-polar point of view which can show
longitudinal variations in a narrow range of latitudes.
\subsection{Short time or spatial variability}
\label{subseq:time_variations}
As presented before, we observed a large-amplitude variability of the
detached haze layer extinction profile at all latitudes below \ang{35}N (Fig.~\ref{fig:dhl_2004_2008}a
to \ref{fig:dhl_2004_2008}c) between December 2004 and June 2005. Fortunately, Cassini took in June, 2005 a series of 9 images of
Titan with a time-step of 80 minutes (Table.~\ref{tab:time_variability}).
\begin{deluxetable*}{ccccc}
\tablecaption{Sequence of 9 images of Titan taken June 4$^{th}$, 2005.
The longitude and local time are given for the profile at the equator
on the illuminated side of Titan.}
\label{tab:time_variability}
\tablewidth{0pt}
\tablehead{
\colhead{Image ID} & \colhead{Time (UTC)} & \colhead{Phase} &
\colhead{Longitude (Eq)} & \colhead{Local Time (Eq)}
}
\startdata
N1496548825\_1 & 03:32 & \ang{10.4} & \ang{10.7}W & 17:12 \\
N1496552665\_1 & 04:36 & \ang{10.3} & \ang{10.4}W & 17:17 \\
N1496557465\_1 & 05:56 & \ang{10.1} & \ang{10.6}W & 17:23 \\
N1496562265\_1 & 07:16 & \ang{9.9} & \ang{13.1}W & 17:17 \\
N1496567065\_1 & 08:36 & \ang{9.8} & \ang{13.4}W & 17:23 \\
N1496571865\_1 & 09:56 & \ang{9.8} & \ang{13.9}W & 17:23 \\
N1496576665\_1 & 11:16 & \ang{9.8} & \ang{14.3}W & 17:30 \\
N1496581465\_1 & 12:36 & \ang{9.9} & \ang{14.7}W & 17:30 \\
N1496586265\_1 & 13:56 & \ang{10.1} & \ang{15.2}W & 17:36 \\
\enddata
\end{deluxetable*}
With multiple observations of Titan in a short period, we are able to validate our calibration and
observe short time and local variability. Here, we analyze the sequence at three different
locations on the limb (\ang{40}S, \ang{0}, \ang{40}N). The 9 observations are made with phase
angles around \ang{10}, within an interval of \ang{0.6}. The limb longitude of the observations
varies between \ang{10}W and \ang{15}W whereas the solar local time on Titan varies between
17:12 and 17:36.
\begin{figure*}[!ht]
\plotone{Time_variability.pdf}
\caption{Extinction profiles for the series
of 9 images taken 80 minutes apart in June 2005 (cf. Table \ref{tab:time_variability}).
The global latitudinal map is shown in figure~\ref{fig:dhl_2004_2008}c.}
\label{fig:time_variability}
\end{figure*}
Figure~\ref{fig:time_variability} presents extinction profiles extracted from the analysis of these 9 images.
First, we confirm that our calibration method is reliable from one image to another and the overall
variability of the detached haze is very small. We observe the double layering at \ang{40}S and \ang{40}N reported
previously. All the locations of the maximum of the extinction peaks are located within a small altitude range (smaller
than the 8 km of the pixel scale). Then, the vertical offsets which are observed at different times are consistent
with the accuracy of the image navigation, and therefore are not significant. At the equator and at \ang{40}S, the detached
haze layer is well detached for all the profiles. When the vertical offset is accounted for, all the extinctions are found
with relative differences of about $\pm$ 10\% compared to the average value, except in the depletion zone between
the main and detached haze layers where difference can reach an order of magnitude (at \ang{0}) to several orders of magnitude
(\ang{40}N). The variations outside the depletion zone are not significant.
The variability in the depletion zone at the equator are about one order of magnitude and
consistent with uncertainties reported in the figure~\ref{fig:model_uncertainties}.
The variability observed at \ang{40}N, in the depletion zone around 460 km are significant for two reasons.
First, the differences are much larger than the expected uncertainty in this zone. Second, the sequence shows
a gradual and consistent increase of extinction with time. If these differences were due to uncertainties in
the retrieval procedure, it would have rather given a chaotic evolution of the extinction with time.
Furthermore, the corresponding extinction map (Fig.~\ref{fig:dhl_2004_2008}c) shows that in the south,
the detached haze layer is well separated from the main haze layer by a well defined depletion zone. In the north, the depletion
zone is less well defined and the detached haze and the main haze are connected vertically by a residual haze.
The variation of this residual haze is the one reported in Fig.~\ref{fig:time_variability} at \ang{40}N.
We can not strictly determine if we are witnessing time or a spatial variations since both the time and the longitude
of the observations change simultaneously during the image sequence. We note that a rotation of \ang{5} in longitude
corresponds to a maximum shift of 250 km in distance (one tenth of Titan's radius), possibly consistent with a spatial
variation of the haze extinction.
\subsection{Dawn and dusk sides}
Aside from short time and local variations, we also are interested in images showing simultaneously the two sides
of Titan. At low phase angle, the viewing geometry allows us to retrieve the haze extinction with both the
illuminated and the dark side of Titan simultaneously (Fig.~\ref{fig:dawn_dusk}).
In this case, we can compare the dawn and dusk limbs for
specific latitudes. Although Titan's day is about 16 terrestrial
days, the time spent by the haze on the night side or dayside is much shorter.
First, the atmosphere is superrotating and at altitude around 400 or 500 km the zonal wind is comparable to or
larger than the rotation speed at the ground \citep{Flasar2005, Achterberg2011, Lebonnois2012, Lellouch2019}.
This makes the actual diurnal cycle for the high altitude hazes shorter than 16 days by a factor of 2 or more.
Secondly, at high altitude, sunlight penetrating beyond the geometric terminator further shortens the time spent in darkness.
Thus, effects on the haze should be produced by processes with timescales
comparable with a terrestrial day.
\begin{figure*}[!ht]
\plotone{Dawn_dusk.pdf}
\includegraphics[width=.5\textwidth]{Extinction_colorbar.pdf}
\caption{The left and right panels show maps of the haze extinction for the dawn and dusk side of
Titan for 4 different images taken during the mission when the viewing geometry allows us to observe both dawn and dusk
limbs. The date and the seasonal solar longitude are provided for each image.}
\label{fig:dawn_dusk}
\end{figure*}
Figure~\ref{fig:dawn_dusk}a presents the haze extinction at the dawn and dusk sides as observed
in June of 2005 (Fig.~\ref{fig:dhl_2004_2008}c). The detached haze differs significantly between the two sides.
At the equator, we observe on the dawn side a double detached layer of 40 km thickness whereas it appears
as a thin layer of 10 km thickness on the dusk side. The haze extinction at the peak differs significantly,
from 7$\times 10^{-8}$ to 2.5$\times 10^{-8}$ m$^{-1}$, respectively. The depletion below the haze layer
is also less pronounced on the dawn side compared to the dusk side. Although this observation was taken
during the period of stability for the detached haze, we observed a significant asymmetry between
the dawn and dusk side. This asymmetry is also observed in all the images taken at the same moment which were
analyzed in section~\ref{subseq:time_variations}.
Two years later, another low phase angle image was recorded (Fig.~\ref{fig:dawn_dusk}b).
This time, the detached haze layer is much more symmetrical between the two sides. The peaks of haze extinction
are at the same altitude with comparable values. However, small differences can be noticed: there is a
small secondary layer above the detached haze layer at the dusk side between \ang{50}S and \ang{20}N, and may even extend
northward. In the northern hemisphere, the extinction appears slightly larger at dawn that at dusk, and the
vertical extent of the detached haze is also a bit larger. However, the overall morphologies are very similar.
At the equinox, the detached haze layer already started its drop in altitude (Fig.~\ref{fig:dawn_dusk}c).
There are significant differences between the two sides, in both hemispheres, while at the equator the two
profiles are almost identical. The haze layer is not exactly symmetrical in the southern hemisphere. The detached
haze itself is at the same altitude, but the depletion zone is at higher altitude at the dawn side, and the main
haze layer is thinner in the dusk side. In the northern hemisphere, the haze layer is more complex, and the
asymmetry is even more marked with a detached haze at different altitudes and with a different extinction. The
detached and main haze layers at the dawn side appear optically thicker than the layers at the dusk side.
This is consistent with the two previous cases.
After the equinox, fewer images were taken at low phase angle. Among them, we do not notice any
significant differences between the dawn and dusk sides. After the reappearance of the detached haze layer in
2016, we found only one image with the relevant geometry to see both sides of Titan illuminated at the same time
(Fig.~\ref{fig:dawn_dusk}d). In this case, we are close to the solstice and the sub-solar latitude is almost
at its maximal extent and does not allow us to probe the southern high latitudes near the terminator.
The main haze appears very symmetrical on both sides and almost identical
above \ang{45}N. The depletion can be followed continuously all around the North Pole. At latitudes lower than
\ang{45}N, the values of the haze extinction are similar but the altitudes of the extinction peak differ by
25 km.
The haze extinction profile and the altitude of the detached haze layer can differ between the dawn
and dusk sides. In general we notice a higher extinction in the dawn side than on the dusk side. This effect
could be due to a diurnal cycle.
\subsection{Longitudinal variability}
Due to orbital constrains, most of our observations sampled only a small range of longitudes.
However, a few observations of Titan taken from a near-polar point of view offer a unique way to study the evolution of
the detached haze with a large coverage in longitude and within a small range of latitudes. It allows us
first to check the homogeneity of the haze in longitude, and it also allows us to extend our previous
observations between the dawn/dusk sides with a local time coverage between 6:00 and 18:00.
\begin{figure*}[!ht]
\plotone{Lon_variability.pdf}
\includegraphics[width=.5\textwidth]{Extinction_colorbar.pdf}
\caption{The panels show the map of the haze extinction as a function of longitude and local
time for a set of 3 images taken in March and April, 2009 ($L_s=\ang{356}$). The latitude range covered is
also indicated for each image.}
\label{fig:lon_variability}
\end{figure*}
We analyzed a set of three images taken sequentially within a month interval. The first observation
was performed on the 29$^{th}$ of March, 2009 (Fig.~\ref{fig:lon_variability}a) during the collapse of the detached
haze layer. Two other observations were made two weeks and one month later, with the same geometry
(Figs.~\ref{fig:lon_variability}b, \ref{fig:lon_variability}c). The detached haze layer is
located at 470 km at all the longitudes. Inside the depletion region, we notice a plume of haze
between 400 and 440 km and between \ang{150}W and \ang{220}W. In mid-April, the plume is located between
\ang{160}W and \ang{240}W and between 375 and 425 km. It appears disconnected and settling from the detached
layer, which remains at 470 km. At the end of April, we see the extension around 410 km and it
is spread from \ang{180}W and \ang{250}W. This aerosol plume is almost connected with the main haze.
This feature is not correlated with the local time but remains about at the same longitude and drifts slowly
toward the West. This would correspond to a retrograde motion of about 0.6 m/s. Another solution would be
prograde motion of 6.6 m/s, in phase with the sampling of 15 terrestrial days. The vertical speed, assuming
that the aerosol cloud dropped from 400 km to 375 km in one month would be $10^{-3}$ m/s. We also identified a
modulation in the extinction and in the geometrical thickness of both the detached haze layer and the main
haze. In the last image, only the geometrical thickness is modulated and not the haze extinction.
These observations show that the haze layer is not completely homogeneous in longitude and have some
fluctuations in extinction and in geometry.
It also strengthen the idea that space and time variations, as in observations previously discussed,
can not be distinguished without additional observations (not available in this dataset).
The dawn/dusk differences and the short-term variations,
presented in the two previous subsections, could be due to longitudinal effect rather than to time variations.
Therefore, with the results of this section, we stress that the longitude inhomogeneities should be kept
in mind as a secondary effect when discussing and comparing latitudinal maps of the detached haze layer.
\section{Comparisons with UVIS occultations and GCM predictions}
\label{seq:comparisons}
\cite{West2018} already confirmed the excellent agreement between the observations made during
two Voyager flybys and the position of the detached haze layer one Titan year later.
Comparisons with Cassini Visual and Infrared Mapping Spectrometer (VIMS) and
Cassini Composite Infrared Spectrometer (CIRS) instruments, could be possible but
they are limited by the sensitivity of their detectors above 450 km where the detached haze is located.
Therefore, we only performed a comparison with two stellar occultations
made by the Cassini Ultraviolet Imaging Spectrograph (UVIS) instrument in 2009 \citep{Koskinen2011}.
The extinction profiles retrieved in the previous sections can also be compared with results
obtained with other instruments and with Global Circulation Model (GCM) predictions.
\subsection{Comparison with UVIS occultations}
\cite{Koskinen2011} derived information on the mesosphere and thermosphere of Titan using UVIS stellar
occultations. The sensitivity to haze opacity of UVIS during a stellar occultation is much better than what ISS can achieve. However,
while ISS probes the light scattered by the detached haze layer, UVIS probes the light transmitted through a tangential
path at the limb. In both cases, an extinction profile can be retrieved. ISS can retrieve the extinction of the
particles which scatter light, and under assumptions concerning the phase function and the single scattering albedo \citep{Seignovert2017, West2018}.
On the other hand, UVIS is able to retrieve the total extinction from transmission with no assumption about the haze
particles. This difference is valuable because it may give information about the change in aerosol
size with altitude. In practice, ISS sensitivity is not sufficiently sensitive to probe above
the peak of the detached haze by more than a scale height.
Two of the UVIS occultation profiles, in 2008 and 2009 (T41 and T53 flybys), can be directly compared with ISS
observations at the same location and at the same period (Fig.~\ref{fig:uvis_iss}). The UVIS profiles are scaled to
offset the spectral dependence between ISS and UVIS effective wavelengths (338 nm and 1850-1900~\AA~respectively).
This offset is due to the spectral dependence of the extinction cross-sections and to the intrinsic
differences arising from comparing the extinction retrieved from scattering properties or from occultation
\citep[see.][]{Cours2011}.
\begin{figure*}[!ht]
\plotone{UVIS_ISS.pdf}
\caption{Comparisons between ISS and UVIS extinction profiles before the equinox.
UVIS profiles are retrieved by \cite{Koskinen2011} during the T41 (2008/02/23) and T53 (2009/04/19) flybys.
ISS profiles are retrieved for the images N1585329510\_1 (2008/03/27) and
N1618568958\_1 (2009/04/16).
The UVIS profiles are scaled by a factor 0.15 to compensate the spectral dependence of the extinction
cross section and overlap ISS retrievals.}
\label{fig:uvis_iss}
\end{figure*}
The two profile can be compared in the 450 to 550 km altitude range.
In the first case (Fig.~\ref{fig:uvis_iss}a), even if the profiles don't exactly overlap, the ISS extinction profile presents a peak of extinction exactly at the same location as the one observed by UVIS. The drop, above and below the peak is more pronounced with ISS than with UVIS.
In the second case (Fig.~\ref{fig:uvis_iss}b), the two profiles presents a excellent agreement with each over in the 450 to 550 km altitude range.
Considering that UVIS and ISS profiles are not taken simultaneously and
they don't probe the same longitude, the results of the previous section demonstrate that these differences are
consistent with the natural variabilities observed in the detached haze layer.
This comparison is then a good validation of our results concerning the extinction profiles of the detached haze.
Above 575 km, UVIS extinction profiles show the presence of a secondary layer at 610 km which is not detected by ISS.
ISS is sensitive to the aerosols scattered throughout the limb, whereas UVIS observations probe the extinction along the tangential path.
Due to its stronger forward scattering peak, the large particles contribute more to the measured ISS scattering signal compared to the smaller particles, that scatter the light more isotropically. During an occultation, both small and large particles contribute to the UVIS extinction signal \citep{Cours2011}.
In theory, the difference between UVIS and ISS above 575 km could reveal a change in aerosol size distribution.
However, this layer is located at altitudes where the signal to noise is low where our model is no longer able to retrieve the extinction properly.
Therefore, it is not possible to draw a safe conclusion from the comparison between ISS and UVIS profiles above 575 km.
\subsection{Comparison with general circulation model predictions}
General circulation models are very powerful tools to understand the climate of planetary atmospheres and the
interplay between different processes at planetary scales. In the case of Titan, circulation and haze are linked
by a strong feedback loop. The large-scale structures in the haze layer are produced by the action of the
circulation. The haze layer produces a feedback effect on the circulation through the control of the stratospheric
thermal structure \citep{Rannou2004}. The detached haze is one of the noticeable features produced by the the
stratospheric circulation \citep{Rannou2002, Lebonnois2012, Larson2015}. Figures.~\ref{fig:gcm_winter}
and \ref{fig:gcm_spring} show the maps of haze extinction obtained by \cite{Lebonnois2012} and
\cite{Larson2015} at 700 nm and 525 nm respectively.
Similarly to the comparison with UVIS, the distribution of the haze from the GCMs
can be compared with the extinction map derived with ISS in the CL1-UV3 filters at 338 nm with a scaling factor.
\begin{figure*}[!ht]
\plottwo{Lebonnois_2012-winter.pdf}{Larson_2015-winter.pdf}
\plotone{N1477222048_2-lat_beta.pdf}
\caption{Zonally-averaged haze extinction at northern winter solstice
($L_s = \ang{270}$) estimated by \cite{Lebonnois2012} at the wavelength $\lambda = $ 700 nm (a)
and by \cite{Larson2015} at $\lambda = $ 525 nm (b). (c) Haze extinction map
retrieved from Cassini/ISS observation CL1-UV3 ($\lambda = $ 338 nm) in the middle of winter
(N1477222048\_2 - $L_s = \ang{300}$).}
\label{fig:gcm_winter}
\end{figure*}
At the northern winter solstice (Fig.~\ref{fig:gcm_winter}), the detached haze appears around 350 km in both models.
In \cite{Lebonnois2012}, the altitude decreases by about few tens of km from the southern latitudes to
the north polar region where it merges with the north polar hood at \ang{40}N (Fig.~\ref{fig:gcm_winter}a).
In \cite{Larson2015}, the detached haze remains at constant altitude (Fig.~\ref{fig:gcm_winter}b), appears better marked than in \cite{Lebonnois2012}, and merges with the polar hood around \ang{60}N.
In both models, the extinction increases from the south to the north by about half a magnitude.
In the ISS observations made in 2004 (Fig.~\ref{fig:gcm_winter}c), \emph{i.e.} at the middle of winter, the detached haze layer is completely developed at 500 km and covers latitudes from the south polar region to \ang{60}N where it merges with the north polar hood.
The location of the depletion zone decreases from 475 to 425 km between \ang{40}N and \ang{60}N, which is not the case in
models. It is, on the other hand, consistent with the results obtained from stellar occultation by \cite{Sicardy2006}.
The haze extinction increases from the south to the north with about the same magnitude as in the models. This
is consistent with a layer increasing in aerosol loading while the airmass is flowing from south to north. It was already
noted \citep{West2011, West2018} that the detached haze layer in models appears as a supplementary layer added to
the background aerosols while, in data, it appears detached because there is a zone strongly depleted in
aerosols. Finally, as mentioned earlier, the detached haze layer is continuous all around the South Pole, which is not
the case in the models.
\begin{figure*}[!ht]
\plottwo{Lebonnois_2012-equinox.pdf}{Larson_2015-spring.pdf}
\plottwo{N1628820904_1-lat_beta.pdf}{N1708076255_1-lat_beta.pdf}
\caption{Rescaled zonally averaged haze extinction at the northern spring equinox ($L_s = \ang{3}$)
estimated by \cite{Lebonnois2012} (a) and at 1000 days after the equinox ($L_s = \ang{30}$)
by \cite{Larson2015} (b).
(c-d) Haze extinction map retrieved from the Cassini/ISS observations
at the Northern Spring equinox (N1628820904\_1 - $L_s = \ang{0}$)
and 1000 days after the equinox (N1708076255\_1 - $L_s = \ang{30}$).}
\label{fig:gcm_spring}
\end{figure*}
At the northern spring equinox (Fig.~\ref{fig:gcm_spring}), the model from \cite{Lebonnois2012} shows a
flat main haze layer without a detached haze between \ang{60}S and \ang{60}N and with two major increases at both poles (Fig.~\ref{fig:gcm_spring}a).
The haze in the south is increasing as a consequence of the circulation reversal while the northern haze is diminishing
and will disappear later in the season. In observations (Fig.~\ref{fig:gcm_spring}c),
this thicker haze is observed only for the northern latitudes, the detached haze has not yet disappeared
and the feeding of the south polar haze has not started. To notice a major increase of extinction at the South Pole,
we need to wait until the beginning of spring at $L_s = \ang{30}$ (Fig.~\ref{fig:gcm_spring}d).
At that time, the detached haze layer almost completely collapsed into the main haze and the haze distribution is very
similar to the one predicted by \cite{Lebonnois2012} at the equinox (Fig.~\ref{fig:gcm_spring}a).
In \cite{Larson2015}, 1000 days after the equinox (Fig.~\ref{fig:gcm_spring}b), we also observed a \emph{U} shape in the meridional haze extinction distribution as in data (Fig.~\ref{fig:gcm_spring}d), but in this case, a new detached haze layer already started to grow from the South Pole in the model (Fig.~\ref{fig:gcm_spring}b), whereas in data (Fig.~\ref{fig:gcm_spring}d) the local increase seen at 380 km is the consequence of the drop of a previous secondary layer (cf. Figs.~\ref{fig:dhl_2008_2012}g and \ref{fig:dhl_2008_2012}h).
In both comparisons, this means that the timing in the circulation models is not completely in phase with the observations.
The model of \cite{Lebonnois2012} seems to be in advance by about 3 years compared to the data.
We do not have enough information to characterize the advance in phase of \cite{Larson2015} model.
Anticipated in late 2014 or early 2015 ($L_s = \ang{60}$, Fig.~\ref{fig:gcm_cycle}a) by \cite{Larson2015} or in mid-2017 ($L_s = \ang{90}$, Fig.~\ref{fig:gcm_cycle}b) by \cite{Lebonnois2012}, the detached haze layer
finally reappeared in late 2015 to early 2016 ($L_s = \ang{74}$, Fig.~\ref{fig:gcm_cycle}c).
However, in May 2017, the upper atmosphere of Titan was still evolving and did not show a polar hood in the South Pole
similar to the one observed in 2004. Moreover, the most recent observations of early-2017 (Figs.~\ref{fig:dhl_2015_2017}c and \ref{fig:dhl_2015_2017}d) seem to show that the seasonal
formation of the detached haze layer could be different from one hemisphere to the other. The double peaks at
420 and 450 km in the temperature gradient profile, a proxy for the haze extinction, obtained from the 1989
occultation \citep{Sicardy1999} supports this hypothesis.
\begin{figure*}[!ht]
\plottwo{Lebonnois_2012-time.pdf}{Larson_2015-time.pdf}
\plotone{DHL_time_eq.pdf}
\plotone{DHL_time_eq-interp.pdf}
\caption{(a) Annual variations of the zonally averaged equatorial opacity at 700 nm extracted from \cite{Lebonnois2012} and wrapped around $L_s = \ang{0}$.
(b) Seasonal evolution of the aerosol mass density (g/cm$^3$) adapted from \cite{Larson2015}.
(c) The haze extinction coefficient as a function of time and altitude at the equator (this work).
The data are binned by 2 months ($\Delta Ls \approx \ang{2} $). The gray area corresponds to the period where no UV3 observations where available (see Fig.~\ref{fig:img_sampling}).
(d) Same result as middle but linearly extrapolated in area without observations. This plot is only provided as an indicative global pattern but contains obvious interpolation artefact.
In each panel, the black vertical line corresponds to the vernal equinox ($L_s = \ang{0}$)
and the white vertical line is the date of the reappearance of the detached haze layer ($L_s = \ang{90}$, $L_s = \ang{60}$ and $L_s = \ang{74}$ respectively).}
\label{fig:gcm_cycle}
\end{figure*}
The current general circulation models do not match some of the large-scale features reported earlier. The first of them
is the shape of the vertical extinction profile which has a local depletion, producing the appearance of a detached haze layer in the
observations, while in models it appears as a high altitude haze layer superimposed to the main background haze.
The amplitude of the observed depletion reported, before and during the collapse, is sufficient to consider that the
detached haze layer is really disconnected from the main haze layer. We already stressed that the detached haze layer is
continuous around the South Pole without any visible upwelling coming from the main haze at this location. Finally, we
reported an early contraction of the main haze in 2008, just before the drop of the detached haze at the equinox. The origin
for such a contraction is likely related to the weakening of the Hadley cell when the latitudinal illumination gradient
decreases around the equinox.
Thanks to high spatial resolution of the ISS/NAC camera, we noticed some small-scale structures which could be unresolved or
erased by the temporal averaging in GCMs. During the northern winter and spring, we observed some sporadic decreases and
bursts in the extinction profiles at very short time scales. These events could have a major impact on the redistribution of
aerosols in the upper atmosphere. We also reported, in numerous cases, the existence of sub-layers above the main detached
haze layer with large latitude extend. Usually, their presence could be followed during more than one Earth year. This is
especially true during the collapse of the main detached haze layer after the equinox, when we observed several smaller
drops from 500 km down to 300 km. And finally, the double structure reported in early-2017
(Figs.~\ref{fig:dhl_2015_2017}c and \ref{fig:dhl_2015_2017}d) with two very distinct detached haze layers in the
North at 450 km, in the South at 520 km, was never mentioned before and not reported in the current GCMs. New modeling with a better vertical resolution will be need to interpret this structure.
We want to highlight facts that would help to better constrain the climate models.
GCMs are ideal tools to understand the interplay of different components of Titan's climate.
But they also have several limitations which are often technical by nature (low spatial resolution, shallow water
approximation for the dynamics, radiative transfer in plane-parallel, no link with the high mesosphere and thermosphere physics).
Limitations may also come from incorrect boundary conditions. We give some remarks:
\begin{enumerate}
\item the collapse starts in the southern hemisphere and concerns the main haze and then the detached haze. This
may indicate that aerosols of the main haze and the detached haze do not have the same fractal structure. This is
already suggested by \cite{Lavvas2009} and \cite{Larson2015}.
The weakening or the displacement of ascending branch of the circulation would first affect the more
compact aerosols of the main haze before the fluffiest aerosols of the detached haze that can remain suspended more
easily.
\item the finest details found in this work occur at length scales smaller than the GCM grid scale.
ISS observations show an instantaneous snapshot of the haze layer. GCMs generally
calculate extinction of haze averaged over several Titan days and/or are zonally averaged. Comparing GCMs instantaneous and
not zonally averaged outputs would allow to see if they are able or not to produce the same kind of variabilities as
observed. These variability may also be induced by processes occurring in higher layers not modeled by the GCMs.
\item the timing offset of the DHL reappearance in GCM is probably due to the model top levels which are generally too
low. Vertical extension of the models and of the haze layer may improve the timing of the cycle.
\item the source of the aerosols depends on photochemical processes that occur far above the DHL and which are
also are subjected to circulation in the upper mesosphere and thermosphere.
Accounting for the physics and chemistry in higher layers, for instance, a coupling with thermospheric models could
be a key to understand further the detail of the haze cycle.
\end{enumerate}
\section{Conclusions}
\label{seq:conclusions}
In this article we have analyzed observations of light scattered at the limb of Titan with ISS (UV3 filter) on the
Cassini spacecraft. We retrieved the haze extinction as a function of altitude, latitude and time between October 2004 and
September 2017. We followed, during about half Titan's year, the evolution of the DHL and the top of the
main haze. In particular, we witnessed the collapse of the DHL during the equinoctial transition of the atmospheric
circulation and its reappearance before the following solstice.
We confirmed and gave details about the collapse of the detached haze layer previously reported by \cite{West2011}.
We also tracked the small-scale variations after the equinox in order to detect the reappearance of the detached haze
layer at the end of 2015 \citep{West2018}. These two previous works focused on the DHL at equator. Here, we give
a full description of the structures of the haze layer as a function of altitude, latitude and time. We find
that the DHL has a natural variability, which can be temporal and spatial, and sometimes two distinct hazes or plumes
can be observed above the DHL. The amount of data is not large enough to distinguish between spatial or
longitudinal variability. But data taken with a polar viewing indicates that the haze is not completely
uniform in longitude.
The equinoctial collapse starts in the summer hemisphere. Its initial phase can be discerned in March, 2008.
The main haze collapses first and then the detached haze layer about one terrestrial year later. By April, 2012,
the detached haze is below 300 km and can not be seen at UV wavelengths.
The fall of the detached haze layer between 2009 and 2011 occurred at the average speed of -67 km/yr.
During the equinoctial collapse, the DHL seems to settle down at the aerosol terminal speed,
as reported by \cite{West2018}. If sedimentation does control the speed of the collapse,
it would indicate an absence of vertical wind at these two moments.
During a period of 3 years and a half (from mid-2011 to 2015), no stable detached haze layer was observed.
However, the haze layer fluctuates, and sporadic local detached layers appeared and disappeared rapidly.
The detached haze layer reappeared in December 2015. We first noticed this new detached haze when it was marginally
apparent in UV3 images.
The timing of the reappearance is offset compare to GCM predictions and it occurs with patterns more complex
that those predicted. First, it reappeared around 500 km in December, 2015 ($L_s =\ang{74}$) as a very faint
structure which became persistent and more pronounced with time.
At the equator this structure sedimented and finally disappeared in about one year, whereas it remained visible in the northern hemisphere.
A second detached haze appeared in July, 2016 around 500 km altitude, above the first DHL, and apparently started to settle down
as well. Unfortunately, the survey was interrupted in September 2017 by the end of Cassini mission. This second detached
haze layer did not cover all latitudes, was quite variable, and was present up to the final observation.
Unfortunately, no data were acquired between April 2008 and February 2009 and between December 2010 and September 2011.
We thought to use the NAC UV1 and UV2 images but these data have a very poor signal to noise ratio and
can not be included in our analysis.
However, a few images are available with the WAC camera in the VIO (Violet), BL1 (Blue) filters.
The behavior of the aerosols at these wavelength should be very similar to the UV3 filter and could fill these gaps.
It would also be interesting to perform similar analysis with data acquired through filters at larger wavelengths.
In these case, it would be possible to probe deeper layers in order to monitor the collapse of the DHL further
down and, as well, the cycle of the main haze. For instance, \cite{Rages1983} could probe as lower as 200-250 km
in clear filter ($\lambda_{eff} \simeq 0.5 \mu$m). At even longer wavelengths, we could reach levels in the low
stratosphere and, maybe probe high altitude polar clouds \citep{deKok2014,West2016}.
Comparison with General Circulation Models are fruitful. Our results reinforce, the scenario of a
detached haze cycle primarily controlled by circulation as proposed by \cite{Toon1992} with a 1D model, \cite{Rannou2002}
with a coupled 2D-GCM and \cite{Lebonnois2012, Larson2015} with coupled 3D-GCMs. Although GCMs capture the global
haze cycle, many differences remain, mainly driven by technical limitations.
\section*{Acknowledgments}
This work was supported by the French ministry of public research. The authors also thank the Programme National de Plan\'{e}tologie (PNP) for their financial support.
B.S. and R.W. thank the Cassini Mission, in particular the advocacy by Trina Ray and Mou Roy for obtaining the images studied here.
Part of this work was performed by the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration (80NM0018D0004).
P.R. and S.V. thank the French Agence Nationale de la Recherche (ANR Project APOSTIC No. 11BS56002, France).
B.S. thanks the Planetary Rings Node (SETI Institute) for their excellent PDS \href{https://tools.pds-rings.seti.org/opus}{browser interface OPUS}.
\section*{Data availability}
The list of all the ISS images and UV extinction profiles derived from this analysis, plus the source code to produce the figures, will be made
publicly available after publication on the \href{https://data.caltech.edu}{Caltech Data Archive},
doi:\href{https://doi.org/10.22002/d1.1344}{10.22002/d1.1344}.
|
\section{Introduction}
Reflected light spectroscopy of exoplanets will be important in the upcoming decades. Future space-based direct imaging missions will perform space-based optical coronagraphy, which will enable low resolution (R $\sim$ 40-200) optical spectra of directly imaged cool giant, and temperate terrestrial planets orbiting Sun-like stars. Although HST/JWST may be able to obtain transmission spectroscopy of a few number of favorable targets (e.g. HIP 41378 f ; \citet{dressing2020HIP}), reflected light offers the opportunity to directly infer the scattering properties of the atmosphere. Reflected light spectra contains key information related to the planet's scattering and chemical properties. Specifically, unlike thermal and transmission spectroscopy, clouds and hazes determine the zeroth order structure of the reflection spectra \citep{nayak17, Lupu_2016, gao2017sulfur, macdonald2018exploring,marley1999reflected}. Given the diversity of clouds and hazes expected in exoplanets \citep{morley14water,Gao20aerosol}, this presents new challenges for retrieving properties from reflected light spectra of exoplanets.
\subsection{Previous Parameterizations used in Reflected Light Retrievals}
In order to prepare for this next decade of exoplanet spectroscopy there has been a growing body of literature aimed at determining best practices in retrieving properties from reflected light observations. We discuss each modeling framework below, focusing on the methodologies for parameterizing clouds. Given our focus on determining the required cloud complexity needed for retrievals, it is important to understand the parameterization method of each of these previous works.
\subsubsection{\citet{Lupu_2016,nayak17}}
\citet{Lupu_2016} showed that the presence/absence of clouds and CH$_4$ can be inferred with high confidence from cool giants with reflection spectroscopy. The forward model used by \citet{Lupu_2016} was based on the model initially developed by \citet{mckay1989thermal, marley1999thermal,marley1999reflected} and later updated by \citet{cahoy2010exoplanet}. They used CH$_4$ molecular opacity and the collision induced opacities of CH$_4$, H$_2$ and He as gaseous opacities in their forward model as they worked only with planets where the reflection spectra is expected to be CH$_4$ dominated. \citet{Lupu_2016} used two simple retrieval models differing in cloud parameterization for retrieving on the reflected spectra, both of which contained wavelength independent clouds. Their first model included a single semi-infinite cloud layer. Their second cloud model had a second cloud deck in addition to the semi-infinite cloud deck. In this case, the semi-infinite cloud deck at the bottom is forced to be optically thick and essentially acts as a reflective surface where the asymmetry parameter for this bottom deck is not retrieved. Using these two models, \citet{Lupu_2016} retrieved on three validation test cases where the simulated data was produced using the retrieval model itself. They also tested their retrieval model on self-consistently modeled test cases of three planets. The performance of their retrieval models showed a decline on the `real' cases compared to the validation cases. For example, only lower limits on the CH$_4$ abundance could be obtained for two of the three real planet cases compared to constraints of CH$_4$ within factors of 10 of the true abundance for retrievals on validation planets. \citet{Lupu_2016} hence note the need to 1) test various other parameterizations with varying levels of cloud modeling complexities on high SNR data of self-consistent models, and 2) identify an optimal set cloud parameters that fully describes the system but minimizes the number of free parameters. Using an identical modeling framework, \citet{nayak17} expanded on this work by exploring the effect of phase and radius uncertainty on the retrieval of atmospheric properties for cool giants with CH$_4$ dominated reflection spectra. Both of these studies concluded that optical spectra with a minimum SNR of 20 is required to retrieve accurate atmospheric properties like molecular abundances and clouds from reflected light for cool giants.
\subsubsection{\citet{Feng_2018}}
\citet{Feng_2018} demonstrated the ability to ascertain atmospheric composition of earth-analog planets using higher resolution spectroscopy (R =70 or 140) as expected from mission concepts such as {\it HabEx} and {\it LUVOIR}. They considered only a single deck of H$_2$O cloud characterized by a cloud top pressure, cloud thickness and a single value of optical depth. The asymmetry parameter and single scattering albedo in their model is fixed to a constant value appropriate for H$_2$O clouds. Abundances of O$_3$, O$_2$, H$_2$O and N$_2$ are retrieved, given the focus on Earth-like planet atmospheres. In addition to these parameters, they also retrieved a parameter for the reflective surface, and a parameter to describe the patchiness of clouds, {\it f$_c$}. Although their methodology pertains to terrestrial planets, the patchy cloud concept will likely be relevant for gas giants as well.
In order to incorporate patchy cloud coverage, they create two models for each case: one with 100\% cloud coverage and a second that is cloud free. The albedo spectra of each of the runs are combined by weighting the first cloudy spectra with f$_c$ and the latter cloudless spectra with (1-{\it f$_c$}). They found that at relatively higher spectral resolutions of R$\sim$140 and SNR$\sim$20 the parameters of interest could be constrained. At R$\sim$70 and SNR$\sim$20 the presence of clouds and molecules could be detected and with low resolution (R $\sim$ 50) combined with SNR$\sim$20 one could achieve just weak detections of clouds and molecules. This retrieval model only accounts for water clouds. Terrestrial planet atmospheres might have photochemical hazes which would complicate inferring scattering properties with water only single deck cloud models. Lastly, \citet{feng2018earth} kept their single scattering albedo and asymmetry parameter fixed. This motivates additional work to determine how retrieving on these parameters for single or multiple cloud decks effects the retrieved solutions for cool giants.
\subsubsection{\citet{Hu_19, Damiano2020exorel}}
A separate model, \texttt{ExoREL} was developed by \citet{Hu_19} for modeling reflected spectra of cool giants, and was implemented in a retrieval framework by \citet{Damiano2020exorel}. This model considered the effect of H$_2$O and NH$_3$ condensation on the vapor phase mixing ratio of these molecules.
\citet{Damiano2020exorel} conducted the retrieval analysis on three test planets, including {\it 47 Uma b}, a focus of this analysis as well. Similar to \citet{lupu2016developing} and \citet{nayak17}, \citet{Damiano2020exorel} retrieved the cloud bottom pressure, cloud thickness, a well-mixed CH$_4$ mixing ratio and the gravity. But they assume only H$_2$O and NH$_3$ condensation and focused on retrieving the volume mixing ratio (VMR) of H$_2$O and NH$_3$ below the cloud bottom. They retrieve a condensation ratio which is then used to construct the depleted VMR of CH$_4$ and NH$_3$ above the cloud deck. \citet{Damiano2020exorel} retrieved on three test planet albedo spectra synthesized using \citet{Hu_19} forward model.
\subsubsection{\citet{irwin2008nemesis, barstow2014clouds}}
\texttt{NEMESIS} is a well-vetted code for retreiving properties of Solar System planets \citep{irwin2008nemesis} and exoplanets \citet{barstow2014clouds}. Most recently, it was used to demonstrate the ability of retrieving cloud scattering properties in thermal emission \citep{Taylor20howdoesthermal}. With regards to reflected light, \citet{barstow2014clouds} used \texttt{NEMESIS} to determine the atmospheric parameters from hot Jupiter {\it HD 189733 b}, observed using HST/STIS by \citet{evans13} at very low spectral resolution. Instead of a Bayesian retrieval analysis, they perform a chi-square analysis on a grid of 980 spectra with a fixed set of cloud base pressures, particle sizes and optical depths at 0.25 microns -- and a fixed set of chemistry parameters for the volume mixing ratio of Na. Similar to previous analyses of cool giants where the cloud species is assumed, \citet{barstow2014clouds} assume that the clouds are composed of MgSiO$_3$ or MnS as these are the relevant condensation species for {\it HD 189733 b}. Then the scattering properties for the cloud particles were calculated using Mie theory with a double peaked Henyey–Greenstein formulation of the phase function. Other important atmospheric parameters such as the temperature-pressure profile and volume mixing ratios of CO, CH$_4$, H$_2$O, CO$_2$ were fixed at the best-fit values derived from previously obtained emission spectra. They found that the data was consistent with a large number of cloudy cases, as well as many cloudless cases. Therefore, there was lot of degeneracy in their cloud parameter space. Given the data quality and resolution of STIS, this is consistent with the findings of \citet{lupu2016developing,nayak2017atmospheric,feng2018earth} and \citet{Damiano2020exorel} who reported the need for a SNR of 20 for proper characterization of cloud properties.
\texttt{NEMESIS} has also been used for retrievals of scattering properties of Solar System planets from reflection spectra \citep[e.g.][]{irwin2015nemesis,irwin2016storm}. Unlike in the study of exoplanets, the cloud species, location and thickness for solar system planets are generally known quantities, and can be treated as fixed parameters. Moreover, the data quality used in these Solar System studies are generally far superior in quality than the focus of this and other previous studies discussed thus far. Therefore in studies such as \citet{irwin2015nemesis}, the wavelength-dependent imaginary part of the refractive indices and the parameters of the particle size distribution for the each condensing species can be directly retrieved from the data.
\subsection{This Analysis}
Previous work has highlighted the need for retrieval studies to be conducted on self-consistent models without the assumption of pure water clouds. Therefore, in this work, we build upon the approach of \citet{Lupu_2016} and others by retrieving the atmospheric properties of cool giant planets from spectra produced with a cloud model. We test four parameterizations to retrieve the cloud properties, each with increasing levels of complexity. Ultimately, we try to compare the accuracy and precision of the retrieved chemistry and clouds of the atmospheres to the original input. We focus on three radiative properties of the clouds -- optical depth per layer, asymmetry parameter and the single scattering albedo. By retrieving directly on the cloud radiative properties, we avoid any assumptions about the condensing species in the atmosphere. For our simulated data, we model the reflected spectra for three cool giants that are priority targets for {\it Nancy Grace Roman Space Telescope}. These priority cool giant targets have high contrast ratios compared to their host stars in reflected light due to the optimal combination of their size, separation from host star and effective temperature.
Table \ref{table:0} summarizes the spectral resolutions in the optical wavelength range expected from multiple future space-based direct imaging missions \citep{luvoir2019,habex20}. We consider both the lowest spectral resolution expected from {\it Nancy Grace Roman Space Telescope} and higher spectral resolutions expected from mission concepts like HabEx and LUVOIR while producing mock observations of the reflected spectra of these three priority target exoplanets. We retrieve on these mock observation spectra in order to test and compare various methodologies for parameterizing atmospheres when retrieving properties. This exercise helps to inform the complexity of atmospheric parameterization requisite for the next decade of reflected light studies. We also account for the uncertainty in the gravity of the planets while performing the retrievals. In doing this, we aim to address the following:
\begin{enumerate}
\item Does the choice of cloud parameterization effect the retrieved cloud properties and molecular mixing ratios from the reflected spectra of cool giant planets?
\item Does the performance of our retrieval model change from one planet to another (i.e. different effective temperatures)?
\item Does the constraint on the retrieved gravity depend on the cloud parameterization?
\item How does data quality (SNR \& R) limit ability to retrieve molecular abundances and cloud properties?
\end{enumerate}
\begin{table*}
\begin{center}
\begin{tabular}{||c c c ||}
\hline
Future Mission & Wavelength Range (microns) & Spectral Resolution \\ [0.5ex]
\hline\hline
Roman Space Telescope CGI Spectroscopy & 0.675-0.785 & 47-75\footnote{\label{wfirst}\url{https://roman.ipac.caltech.edu/sims/Param_db.html}} \\
\hline
Roman Space Telescope CGI Imaging & 0.5-0.8 \textsuperscript{\ref{wfirst}} & - \\
\hline
LUVOIR-A \& B (ECLIPS) & 0.515-1.03 & 140 \footnote{\label{luvoir}\citet{luvoir2019}} \\
\hline
HabEx Coronagraph & 0.45-1.00 & 140 \footnote{\label{habex}\citet{habex20}} \\%(logscale in Pressure) \\
\hline
HabEx Starshade Instrument & 0.45-0.975 & 140 \textsuperscript{\ref{habex}}\\%(logscale in Pressure)
\hline
\end{tabular}
\end{center}
\caption{Expected Spectral resolutions from future space-based direct imaging missions.}
\label{table:0}
\end{table*}
We describe the atmospheric modeling using \texttt{PICASO} and \texttt{Virga} in \S \ref{sec:Modeling}. We briefly explain the retrieval method in \S \ref{sec:nestedsampling}. We describe the results of our analysis in \S \ref{sec:results} and finally discuss our main results in \S \ref{sec:discussion} and summarize them in \S \ref{sec:conclusion}.
\section{Modeling Reflected Spectra}\label{sec:Modeling}
We use the effective temperature vs. gravity parameter space to select a representative target population for the analysis. We calculate the equilibrium temperature ($T_{\rm eq}$) and gravity ($g$) of 23 direct imaging planet targets, most of which have radial velocity detections \citep{Butler_2006_catalog,howard16limits,fischer02second,hatzes06confirmation}. We use the planetary orbital and stellar parameters, to calculate the equilibrium temperature of these planets assuming a zero albedo. We also consider an additional internal temperature of 100 K, similar to that of Jupiter \citep{fortney2007planetary}, to get the effective temperature ($T_{\rm eff}$).
We use the $M \sin i$ and planet radius for calculating the gravity of these planets. We calculate the planet radius using the empirical mass-radius relationship for cool giants from \citet{Thorngren_2019empirical}. The $T_{\rm eff}$ vs $g$ parameter space for these planets are shown in Figure \ref{fig:fig1}. We use {\it eps Eri b}, {\it 47 Uma b} and {\it HD 62509 b} as our target planets (shown in in Figure \ref{fig:fig1}) to explore retrievals on reflected light across three different temperature regimes for directly imaged cool giants. We emphasize that the aim is not to produce highly self-consistent models of these planets. Instead, we aim to explore a range in temperature that enables a diversity in cloud formation, and chemistry scenarios. Additionally, our targets have similar gravity estimates. This allows us to isolate the effect that varying cloud and chemistry scenarios have on retrieving atmospheric parameters. In a future analysis we will explore the effect of completely unconstrained gravity.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{fig1.pdf}
\caption{The effective temperature $T_{\rm eff}$ vs gravity $g$ parameter space for 23 cool giant targets for space-based direct imaging missions. The blue, orange and green points denote the target planets used in this work {\it eps Eri b}, {\it 47 Uma b} and {\it HD 62509 b}, respectively. The upward arrows point to the effective temperature if a 100 K internal temperature is considered. {\bf Main Point}- We choose three cool giant target planets that span a wide a range of effective temperatures (135~K - 533~K) in order to explore diversity in atmospheric cloud properties.}
\label{fig:fig1}
\end{figure}
\subsection{Modeling the Planet Atmospheres}
We use \texttt{PICASO} \citep{batalha19}, which has heritage from \citet{mckay1989thermal, marley1999thermal,cahoy2010exoplanet}, to model the reflected light spectra of our target planets. \texttt{PICASO} is an open-source radiative transfer code capable of calculating transmission, reflected and/or thermal spectra of planets and brown dwarfs. \texttt{PICASO} requires the temperature-pressure ($T(P)$) profile, cloud structure and atmospheric chemistry as inputs for the radiative transfer calculation. Here, we discuss modeling each of these inputs for our reflected spectra simulation in \S\ref{sec:TP}-\S\ref{sec:cld} and discuss the basics of \texttt{PICASO} in \S\ref{sec:picaso}.
\subsubsection{The Temperature-Pressure Profile}\label{sec:TP}
We divide the planet atmosphere in to 61 plane-parallel pressure layers where the pressure rises logarithmically from $10^{-6}$ to $10^{3}$ bars. We model the temperature-pressure profile of the planet atmospheres using the empirical parameterization described in \citet{ryan18}. This empirical $T(P)$ profile parameterization is dependent on T$_{\rm eff}$, $g$ and metallicity of the planet [M/H]. The best-fit value of the coefficients in the parameterization have been determined by fitting the empirical profile to a large number of $T(P)$ profiles for cool giants produced self-consistently using the methodology of \citet{fortney08}. The $T(P)$ profile parameterization is described by,
\begin{align}
T^4(P) = {T_0}^4 + {T_{\rm deep}}^4 \left(\dfrac{P}{1000 \text{ bars}} \right)
\end{align}
where both $T_0$ and $T_{\rm deep}$ are T$_{\rm eff}$, $g$ and [M/H] dependent functions \cite[see][]{macdonald2018exploring}. We assume Jupiter metallicity of 3 $\times$ Solar metallicity \citep{wong04updated} for all the three planets. The parameterized $T(P)$ profile for the three planets are shown in Figure \ref{fig:figtp}. Although the self-consistent $T(P)$ profiles will show more structure than the parameterized profiles, as seen in \citet{macdonald2018exploring}, this will not affect the results of the analysis given the insensitivity of reflected light spectroscopy to temperature.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{fig2.pdf}
\caption{The temperature-pressure profile of {\it eps Eri b} (blue), {\it 47 Uma b} (orange) and {\it HD 62509 b} (green) generated using the empirical paramterization of \citet{ryan18} are shown. Condensation curves of the molecular condensates considered in the cloud calculation of \texttt{Virga} are shown by the colored dashed lines. {\bf Main Point}- Our target planets have different $T(P)$ profiles which causes condensation of different species at different pressures, resulting in varying cloud decks and optical properties.}
\label{fig:figtp}
\end{figure}
\subsubsection{ Atmospheric Chemistry}\label{sec:chem}
Chemical equilibrium abundances are interpolated from those computed on a grid of $(P,T)$ points as calculated using a modified version of the NASA CEA Gibbs minimization code \cite[see][]{Gordon94}. The chemistry grid is available for download at \citet{marley_18} and described fully in Marley et al. (in prep.). For these cool giants the most important gaseous absorbers in the optical are methane, ammonia, and water. The grid accounts for depletion of each chemical species above the point of condensation.
The volume mixing ratio profiles of these three species for each of our three target planets are shown in Figure \ref{fig:figchem}.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{fig3.pdf}
\caption{The volume mixing ratio as a function of pressure of CH$_4$, NH$_3$ and H$_2$O for the three planets. {\bf Main Point}- The three planets have very different chemical structure due to different T(P) profiles. }
\label{fig:figchem}
\end{figure*}
\subsubsection{Clouds with \texttt{Virga}}\label{sec:cld}
We calculate the cloud profiles for each of the targets using \texttt{Virga}. \texttt{Virga} follows the \citet{marley00} treatment of condensation in atmospheres. For each condensate species and at each atmospheric layer, the vapor pressure in excess of the saturation vapor pressure is allowed to condense. The condensation curves of all the molecular species considered by our cloud model are shown in Figure \ref{fig:figtp} in dashed lines. The condensation curves follow those in \citet{morley2012neglected} and \citet{gao2020aerosol}, and are available online \footnote{\url{https://github.com/natashabatalha/virga/blob/master/virga/pvaps.py}}.
The full cloud profiles from \texttt{Virga} are shown Figure \ref{fig:figcldopd}. The coolest planet case, {\it eps Eri b}, has H$_2$O and NH$_3$ condensation forming two separate cloud decks. The warmer case, {\it 47 Uma b}, lacks NH$_3$ clouds but is still dominated by H$_2$O clouds. The hottest case, {\it HD 62509 b}, is dominated by Na$_2$S clouds at depth (P$>1$~bar), and lacks condensation from H$_2$O or NH$_3$. As shown in Figure \ref{fig:figcldopd}, the three cases explored here probe three different cloud condensation regimes.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{fig4.pdf}
\caption{The column optical depth as a function of pressure for each condensed species for the three planets. {\it eps Eri b} has two major cloud decks of NH$_3$ and H$_2$O. The top deck for {\it 47 Uma b} is H$_2$O cloud while the bottom deck is composed primarily of Na$_2$S. {\it HD 62509 b} has a single cloud deck formed very deep in the atmosphere. Na$_2$S, ZnS and MnS are the high optical depth condensates among other smaller optical depth condensate species for this planet. {\bf Main Point}- The three planets have very different cloud structure formed out of different condensate species. }
\label{fig:figcldopd}
\end{figure*}
The vertical structure of the cloud is determined by the balance between the vertical turbulent mixing of condensates and vapor and the sedimentation of condensates described by the equation,
\begin{align}\label{eqn:am01}
-K_z\dfrac{\partial q_t}{\partial z} - f_{\rm sed}w_*q_c=0
\end{align}
where $q_c$ is the condensate mole fraction, $q_v$ is the vapor mole fraction and $q_t=q_c+q_v$ \citep{marley00}. The first term represents the vertical turbulent mixing of the condensate and vapor, where $K_z$ is the vertical eddy diffusion coefficient. The second term represents the sedimentation caused by the condensates, where $w_*$ is the convective velocity scale. $f_{\rm sed}$ is a dimensionless ratio of the sedimentation velocity to $w_*$. The cloud structure is solved by the balance of the two competing processes for each of the condensing species.
The sedimentation parameter $f_{\rm sed}$ and the vertical eddy diffusion coefficient $K_{z}$ are the two inputs which critically determine the cloud vertical extents and particle size distribution. Overall, low values of $f_{\rm sed}<1$ produce thick cloud layers with smaller particles and higher values, $f_{\rm sed}>1$, produce thinner cloud decks with larger particles. We set $f_{\rm sed} = 3$ motivated by the higher values that have been successful in modeling cool giant clouds for jupiter-like planets, in contrast to lower values of $\sim$0.1 which have been typically used for hot jupiters \citep{webber15}.
The vertical eddy diffusion coefficient ($K_{z}$) can strongly effect the cloud properties. Overall, higher $K_{z}$ values lead to the formation of larger cloud particles. We calculate the vertical eddy diffusion coefficient $K_z$ using \citep{gierasch85energy},
\begin{align}
K_z= \dfrac{H}{3}{\left(\dfrac{L}{H}\right)}^{\dfrac{4}{3}}{\left(\dfrac{RF}{\mu{\rho_a}c_p}\right)}^{\dfrac{1}{3}}
\end{align}
where $H$ is the atmospheric scale height, $L$ is turbulent mixing length, $R$ is the universal gas constant, $F$ is the thermal flux of the atmosphere (assumed to be ${\sigma}T_{\rm eff}^4$), $\mu$ is the atmospheric molecular weight, $\rho_a$ is the atmospheric density and $c_p$ is the atmospheric specific heat at constant pressure.
This formulation is based on the assumption that the vertical eddy diffusion coefficient for the vapor and the condensate of the cloud model is the same as derived for heat in free convection conditions \citep{gierasch85energy}.
This method assumes convection occurs all the way to the top of the atmosphere, which of course is not the case in reality. However, this has been used to baseline the cloud model for Jupiter in \citet{ackerman2001cloud} and, hence, is applicable for this study.
In solving Equation \ref{eqn:am01}, we compute an effective particle radius per layer per species, and assume a lognormal distribution of particles with a geometric standard deviation of 2 about that radius. The Mie scattering calculations are then computed with \texttt{PyMieScatt} \citep{sumlin18retrieving} over this distribution. This allows \texttt{Virga} to produce the altitude- and wavelength-dependent optical depth per layer -- ${\tau}$(P,${\lambda})$, single scattering albedo -- $\omega$(P,$\lambda$)) and asymmetry parameter -- $g$(P,$\lambda$) of the clouds in the atmosphere.
The single scattering albedo describes the wavelength-dependent reflectivity of the cloud particles. Higher $\omega$ leads to higher reflectivity. The asymmetry parameter captures the forward/back scattering probability of the scattering of light from the cloud particles.
The final cloud optical profiles of each planet are shown in Figure \ref{fig:Virgacld}. The optical properties can trace back to the exact cloud species. For example, the $\sim$1 bar cloud deck of eps Eri b corresponds to the highest region of single scattering and therefore, can be reasonably identified as a H$_2$O cloud. Ultimately, it is the information in these profiles that we aim to recover.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{fig5.pdf}
\caption{The cloud optical properties as outputs of the cloud modeling code \texttt{Virga} for the three planets. The optical depth per layer, asymmetry parameter and single scattering albedo are shown as function of pressure for the three planets from left to right, respectively. The cloud profiles shown here are averaged over the wavelength range of 0.3 to 1~$\mu$m. }
\label{fig:Virgacld}
\end{figure*}
\subsubsection{Reflected Spectra with \texttt{PICASO}}\label{sec:picaso}
With the $T(P)$ profile, chemical structure and the cloud optical properties as inputs, we adopt the one dimensional version of \texttt{PICASO} to calculate the reflected light spectra for the three planet cases \citep{batalha19}. \texttt{PICASO} uses the two-stream radiative transfer methodology of \citet{toon1989rapid}. We add contributions from both molecular and collision-induced absorption (CIA) opacities. Although we only focus on retrieving H$_2$O, NH$_3$, and CH$_4$, \texttt{PICASO} includes the molecular opacity from H$_2$O \citep{barber06high,tennyson2018exomol}, CH$_4$ \citep{yurchenko_2014,yurchenko13vibrational}, NH$_3$ \citep{yurchenko11vibrationally}, CO \citep{li15rovibrational}, PH$_3$ \citep{sousa14exomol}, H$_2$S \citep{azzam16exomol}, CO$_2$ \citep{HUANG2014reliable}, Na \& K \citep{Ryabchikova2015}, and others not applicable to these temperatures (e.g. TiO, VO). Among the CIA \texttt{PICASO} includes opacity from H$_2$-H$_2$ \citep{collision-inducedmartin}, H$_2$-He, H$_2$-N$_2$, H$_2$-H, H$_2$-CH$_4$, H-electron bound-free, H-electron free-free and H$_2$-electron interactions. The resultant opacity calculations are available on Zenodo \citep{batalha_zenodo_opacities}.
In order to accurately capture asymmetrical back scattering caused by Rayleigh scattering, we use the Two-Term Henyey-Greenstein (TTHG) phase function combined with Rayleigh phase function formalism (\texttt(TTHG$\_$Ray) in \texttt{PICASO}) for the direct scattering component. The One-Term Henyey-Greenstein (P$_{OTHG}$) phase function is defined as,
\begin{align}
P_{OTHG}(cos\Theta)=\dfrac{1}{2}\dfrac{1-g^2}{\sqrt{1+g^2 -2gcos{\Theta}}}
\end{align}
The TTHG phase function capturing both forward, $g_f$, and back scattering, $g_b$, is then defined as
\begin{multline}
P_{TTHG}(cos\Theta)= fP_{OTHG}(cos{\Theta},g_f) \\
+(1-f)P_{OTHG}(cos{\Theta},g_b)\label{eqn:phase}
\end{multline}
where $g_f$= $\Bar{g}$, $g_b$= -$\Bar{g}$/2 and $f$, the fraction of forward to back scattering is $f= 1- g_{b}^2$. The first term corresponds to forward scattering phase function weighted by $f$ while the second term is the back scattering phase function weighted by $(1-f)$. Our adoption of $g_b$ is arbitrary. However, it has been previously adopted in studies of exoplanet reflected light \citep{cahoy2010exoplanet, Feng_2018} due to the lack of \textit{a priori} information. The $\Bar{g}$ is the cloud asymmetry parameter weighted by the cloud fractional opacity. It is calculated using the cloud asymmetry parameter, single scattering albedo, the cloud opacity, and Rayleigh opacity:
\begin{align}
\Bar{g}=\left(\dfrac{\omega\tau_{cld}}{\omega\tau_{cld}+\tau_{Ray}}\right)g
\end{align}
The TTHG and Rayleigh scattering phase functions ($P_{Ray}$) are then combined with a weighted addition to get the final phase function, \texttt{TTHG$\_$Ray}. The weight factor for the TTHG and Rayleigh phase functions are $\tau_{cld}/\tau_{scat}$ and $\tau_{Ray}/\tau_{scat}$, respectively. P$_{TTHGRay}$ is then:
\begin{align}
P_{TTHGRay}(cos\Theta)= \dfrac{\tau_{cld}}{\tau_{scat}}P_{TTHG}(cos\Theta) + \dfrac{\tau_{Ray}}{\tau_{scat}}P_{Ray}(cos\Theta)
\end{align}
The multiple scattering phase function in \texttt{PICASO} is calculated by expanding the HG function to second order (\texttt{N=2}) and forcing the second order moment such that it reproduces Rayleigh scattering when it dominates the total scattering opacity \citep{batalha19}. We also include the effect of Raman scattering by using the \citet{POLLACK1986442} formalism. The \citet{POLLACK1986442} methodology for Raman scattering results in redshift of the photons, which dampens the overall reflectively toward the blue \citep{batalha19}. The \citet{POLLACK1986442} formulation does not model high resolution of solar emission features seen in reflected light spectra of gas giants \citep{antonija2016raman}. However, these features are far too high resolution ($R>>100$) for the next decade of direct imaging observations \citep{antonija2016raman}.
In order to understand the interplay between Rayleigh scattering, molecular, and cloud scattering, \texttt{PICASO} computes the ``photon attenuation'', which denotes the pressure level where the two-way optical depth from each component reaches $\tau=1$. Figure \ref{fig:figphot} shows the photon attenuation for the three planets cases. The ``flatness'' of the {\it eps Eri b} reflected spectra comes from the dominance of the water cloud optical properties, which are highly reflective and non-wavelength dependent at the respective particle radii from 0.3-1$\mu$m. On the other hand, the {\it HD 62509 b} spectrum is dominated by Rayleigh opacity short of 0.5~$\mu$m, by molecular absorption between 0.5-1$\mu$m. Because the cloud deck is much lower (in altitude) than the molecular opacity source, the molecular absorption dominates, causing the planet to have the lowest albedo among the three planet cases. Molecular, cloud and Rayleigh opacity contributes significantly to the {\it 47 Uma b} reflected light.
For the case of {\it eps Eri b} and {\it 47 Uma b}, which both have contribution from the cloud, the different behaviour in the spectra is a result of different cloud optical properties (optical depth, single scattering, and asymmetry profiles). As seen from Figure \ref{fig:Virgacld} and Figure \ref{fig:figcldopd}, {\it eps Eri b} has two cloud decks (an NH$_3$ cloud deck above a very optically thick H$_2$O cloud deck), whereas {\it 47 Uma b} only has a H$_2$O cloud deck at a pressure similar to the higher NH$_3$ cloud deck in {\it eps Eri b}. The H$_2$O cloud deck in {\it eps Eri b} has a relatively high optical depth ($\sim$ 200) compared to the H$_2$O cloud deck in {\it 47 Uma b}. {\it 47 Uma b} also has a second cloud deck which appears much deeper in the atmosphere and is relatively optically thick compared to the deeper ZnS cloud deck appearing in {\it eps Eri b}. The optical depth differences, in addition to the differing asymmetry and the single scattering albedo of the {\it eps Eri b} H$_2$O cloud deck, leads to a flat cloud dominated albedo spectra for {\it eps Eri b} compared to {\it 47 Uma b}.
Finally, the simulated reflected light spectra for the three planets are shown in Figure \ref{fig:figphot} lower panel. They span: 1) a case dominated by bright water clouds, 2) a case dominated by clouds, molecular opacity and Rayleigh scattering, and lastly 3) a case dominated by Rayleigh and molecular opacity. These three spectra are exemplary cases to be used in the retrieval analysis as they test the parameterizations under three different scattering regimes.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{fig6.pdf}
\caption{The photon attenuation depth maps in the top panel corresponding to an optical depth of 0.5 for our three fiducial cases: {\it eps Eri b} (left), {\it 47 Uma b} (middle) and {\it HD 62509 b} (right). The attenuation pressure levels are divided into Rayleigh, gas and cloud opacity. The lower panel shows the simulated albedo spectra for each planet. {\bf Main Point}- The three cases span three different scattering regimes. {\it eps Eri b}'s spectrum is dominated by bright water cloud reflection, {\it 47 Uma b}'s spectrum by cloud and molecular opacity, and {\it HD 62509 b}'s spectrum by molecular and Rayleigh scattering.}
\label{fig:figphot}
\end{figure*}
\subsection{The Retrieval Setup}\label{sec:retrieval}
Similar to previous works discussed, a retrieval requires parameterizations to be made in order to best capture the behavior of the physical model, described in \S2.1. For the retrieval, we replace the chemistry, cloud, and $T(P)$ profile with parameterizations that can be used in a Bayesian framework.
We use an isothermal $T(P)$ profile for our forward model with the temperature fixed at the effective temperature of the planet. This is different from \citet{Lupu_2016} where the $T(P)$ profile for retrievals was kept fixed to the profile used for modeling the simulated data. Unlike thermal and transmission spectroscopy, reflected light is only sensitive to the temperature through its contribution to the line shapes of the molecular opacity, and the scale height of the atmosphere. The opacity is not strongly temperature dependent in the parameter space probed by our three targets \citep{karkoschka1994spec, karkoschka2011haze}.
We initially assume that the atmosphere is well-mixed and start by retrieving a single value for the mixing ratios of three molecules -- CH$_4$, H$_2$O and NH$_3$ -- that dominate the opacity sources for these cooler class of planets \citep{madhusudhan16,burrows97}. The rest of the atmospheric composition, other than CH$_4$, NH$_3$ \& H$_2$O, is assumed to be composed of H$_2$ and He. The H$_2$/He fraction is taken to be f=H$_2$/He=$0.837/0.163$ \citep{lodders19solar}. Therefore the He and H$_2$ mixing ratio is given by,
\begin{align}
He &= \dfrac{1- CH_4- H_2O-NH_3}{f+1},\\
H_2 &= f*He
\end{align}
where the species' name represent its volume mixing ratios (v/v). Our initial assumption of well-mixed profiles in the retrieval is different from that of \citet{Damiano2020exorel}, who use two free parameters for H$_2$O and NH$_3$ each to describe their depleted mixing ratios.
Our chemical profiles incorporate depletion caused by condensation as is clearly evident in Figure \ref{fig:figchem}. {\it eps Eri b} shows a depletion of NH$_3$ and H$_2$O due to condensation of both the species whereas {\it 47 Uma b} shows a depletion only in the H$_2$O mixing ratio as NH$_3$ condensation is absent in {\it 47 Uma b}. Therefore, our initial assumption of well-mixed atmospheres purposely tests whether or not additional complexity is necessary. Later, we relax this assumption and discuss the results of a retrieved depleted profile in \S\ref{sec:depletedwater}.
We parameterize ${\tau}$(P), $\omega$(P) and $g$(P) in four ways with different levels of complexities. We neglect any wavelength dependence in all the three cloud optical properties for our retrieval model. The validity of this assumption, given that \texttt{Virga} calculates wavelength dependent cloud optical properties, is addressed in \S\ref{sec:discussion}. The schematic diagram for the cloud parameterizations are shown in Figure \ref{fig:fig3}. In what follows, we describe each of the four parameterizations (9-15 free parameters total), and the associated priors used in the retrieval analysis.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{fig7.png}
\caption{Schematic diagram of the four cloud structure and optics parameterizations used in our forward model for the target planets. The left column shows the optical depth per layer -- $\tau$, the middle column is for the asymmetry parameter -- {\it g} per layer and the right column shows the single scattering albedo -- $\omega$ per layer. The top row shows the parameterizations for Case 1 -- the box cloud model, the second row is for Case 2 -- the single cloud profile model, the third row shows Case 3 -- the double cloud profile model and the last row shows Case 4 -- the double cloud profile model with two valued asymmetry and single scattering albedo model. Details of the model parameters for each case can be found in \S \ref{sec:boxcld} through \S \ref{sec:case4}.
}
\label{fig:fig3}
\end{figure}
\subsubsection{The Box Cloud Model (Case 1)}\label{sec:boxcld}
The optical depth profile ($\tau(P)$) of this cloud parameterization is similar to the cloud deck used for retrieval by \citet{Feng_2018}. Unlike \citet{Feng_2018}, the asymmetry parameter and the single scattering albedo for this parameterization are also free parameters. We model the cloud structure in this case using 5 parameters. The other four parameters are the CH$_4$, NH$_3$ and H$_2$O mixing ratios and the gravity of the planet. The cloud structure for this case is parameterized according to the following equations,
\begin{align}
\tau (P,\lambda) =
\begin{cases}
0 & \text{if $P > P_0$ or $P < P_0 - dP$ ,}\\
\tau_{max} & \text{otherwise}
\end{cases}
\end{align}
\begin{align}
\omega (P,\lambda) =
\begin{cases}
0 & \text{if $P > P_0$ or $P < P_0 - dP$ ,}\\
\omega & \text{otherwise}
\end{cases}
\end{align}
\begin{align}
g (P,\lambda) =
\begin{cases}
0 & \text{if $P > P_0$ or $P < P_0 - dP$ ,}\\
g & \text{otherwise}
\end{cases}
\end{align}
\smallskip
where P$_0$, $\tau_{max}$, $\omega$, g and dP are the 5 parameters of the model. The nine parameters of this model and the priors used in retrievals are summarised in Table \ref{table:1}.
\begin{table*}
\begin{center}
\begin{tabular}{||c c c c||}
\hline
Parameter & Description & Range & Iteration type \\ [0.5ex]
\hline\hline
g & Asymmetry Parameter & 0-1 & Linear \\
\hline
$\omega$ & Single Scattering Albedo & 0-1 & Linear \\
\hline
$\tau$ & Optical Depth Per Layer & 0.1-30 & Linear \\
\hline
P$_0$ & Cloud Base Pressure Level & 10$^{-6}$-10$^2$ bars & Logscale \\%(logscale in Pressure) \\
\hline
dP & Cloud Deck Thickness & 0 to $P_0$-10$^{-6}$ bars & Logscale \\%(logscale in Pressure) \\
\hline
CH$_4$ & CH$_4$ Mixing Ratio & -6 to 0 & Logscale \\
\hline
NH$_3$ & NH$_3$ Mixing Ratio & -6 to 0 & Logscale \\
\hline
H$_2$O & H$_2$O Mixing Ratio & -6 to 0 & Logscale \\
\hline
g & Gravity & 25-65 m/s$^2$ & Linear \\ [1ex]
\hline
\end{tabular}
\end{center}
\caption{Parameters for the box cloud model.}
\label{table:1}
\end{table*}
\subsubsection{Single Cloud Profile Model (Case 2)}\label{sec:profcld}
For this case, we model the clouds with an altitude dependent optical depth profile. The asymmetry parameter and the single scattering albedo are forced to be zero beneath the base of the cloud deck and can take a value between zero and one above the base of the cloud deck till the top of the atmosphere. The cloud structure here is modeled as,
\begin{align}
\tau (P,\lambda) =
\begin{cases}
0 & \text{if $P > P_0$ ,}\\
\tau_{max} {\rm e}^{-a(lnP-lnP_0)^2} & \text{otherwise}
\end{cases}
\end{align}
\begin{align}
\omega (P,\lambda) =
\begin{cases}
0 & \text{if $P > P_0$ ,}\\
\omega & \text{otherwise}
\end{cases}
\end{align}
\begin{align}
g (P,\lambda) =
\begin{cases}
0 & \text{if $P > P_0$ ,}\\
g & \text{otherwise}
\end{cases}
\end{align}
where P$_0$, $\tau_{max}$, a, $\omega$ and g are free parameters. Hence, our forward model involves 9 free parameters consisting of 5 cloud parameters and the mixing ratios and gravity, similar to Case 1. The number of parameters are same as Case 1 and they are described in Table \ref{table:2}.
\begin{table*}
\begin{center}
\begin{tabular}{||c c c c||}
\hline
Parameter & Description & Range & Iteration type \\ [0.5ex]
\hline\hline
g & Asymmetry Parameter & 0-1 & Linear \\
\hline
$\omega$ & Single Scattering Albedo & 0-1 & Linear \\
\hline
$\tau$ & Optical Depth Per Layer & 0.1-30 & Linear \\
\hline
P$_0$ & Cloud Base Pressure & 10$^{-6}$-10$^2$ bars & Logscale \\
\hline
a & Cloud Deck Scale Height & 10$^{-4}$ to 2 & Logscale \\
\hline
CH$_4$ & CH$_4$ Mixing Ratio & -6 to 0 & Logscale \\
\hline
NH$_3$ & NH$_3$ Mixing Ratio & -6 to 0 & Logscale \\
\hline
H$_2$O & H$_2$O Mixing Ratio & -6 to 0 & Logscale \\
\hline
g & Gravity & 25-65 m/s$^2$ & Linear \\ [1ex]
\hline
\end{tabular}
\end{center}
\caption{Parameters for the single cloud profile model.}
\label{table:2}
\end{table*}
\subsubsection{Double Cloud Profile Model (Case 3)}\label{sec:case3}
This model is similar to Case 2 except a second cloud deck is allowed to form here. Additionally, the asymmetry and single scattering profiles are similar to Case 2 where they take a value between zero and one above the base of the deepest cloud deck. The twelve free parameters are described in Table \ref{table:3}. The following equations describes the parameterizations for this model,
\begin{align}
\tau (P,\lambda) =
\begin{cases}
0 & \text{if $P > P_1$ ,}\\
\tau_{max1} {\rm e}^{-a_1(lnP-lnP_1)^2} & \text{if $P < P_1$ ,} \\
\tau_{max2} {\rm e}^{-a_2(lnP-lnP_2)^2} & \text{if $P < P_2$ ,} \\
\end{cases}
\end{align}
\begin{align}
\omega (P,\lambda) =
\begin{cases}
0 & \text{if $P > P_1$ ,}\\
\omega & \text{otherwise}
\end{cases}
\end{align}
\begin{align}
g (P,\lambda) =
\begin{cases}
0 & \text{if $P > P_1$ ,}\\
g & \text{otherwise}
\end{cases}
\end{align}
\begin{table*}
\begin{center}
\begin{tabular}{||c c c c||}
\hline
Parameter & Description & Range & Iteration type \\ [0.5ex]
\hline\hline
g & Asymmetry Parameter & 0-1 & Linear \\
\hline
$\omega$ & Single Scattering Albedo & 0-1 & Linear \\
\hline
$\tau_1$ & Optical Depth Per Layer of Lower Cloud Deck & 0.1-30 & Linear \\
\hline
$\tau_2$ & Optical Depth Per Layer of Upper Cloud Deck & 0.1-30 & Linear \\
\hline
P$_1$ & Cloud Base Pressure of Lower Cloud Deck & 10$^{-6}$-10$^2$ bars & Logscale \\
\hline
P$_2$ & Cloud Base Pressure of Upper Cloud Deck & P$_1$-10$^{-6}$ bars & Logscale\\
\hline
a$_1$ & Cloud Deck Scale Height of Lower Cloud Deck & 10$^{-4}$ to 2 & Logscale \\
\hline
a$_2$ & Cloud Deck Scale Height of Upper Cloud Deck & 10$^{-4}$ to 2 & Logscale \\
\hline
CH$_4$ & CH$_4$ Mixing Ratio & -6 to 0 & Logscale \\
\hline
NH$_3$ & NH$_3$ Mixing Ratio & -6 to 0 & Logscale \\
\hline
H$_2$O & H$_2$O Mixing Ratio & -6 to 0 & Logscale \\
\hline
g & Gravity & 25-65 m/s$^2$ & Linear \\ [1ex]
\hline
\end{tabular}
\end{center}
\caption{Parameters for the double deck cloud profile model.}
\label{table:3}
\end{table*}
\subsubsection{Double Cloud Profile Model with Two Valued g0 and w0 (Case 4)}\label{sec:case4}
This fifteen parameter model has the same optical depth per layer parameterization as Case 3. The major difference for this case is that the asymmetry and single scattering are each allowed to have two-values. The asymmetry and single scattering parameterizations are,
\begin{align}
\omega (P,\lambda) =
\begin{cases}
\omega_2 & \text{if $P > P_2$ and $P < P_2-dP$ ,}\\
\omega_1 & \text{otherwise}
\end{cases}
\end{align}
\begin{align}
g (P,\lambda) =
\begin{cases}
g_2 & \text{if $P > P_2$ and $P < P_2-dP$ ,}\\
g_1 & \text{otherwise}
\end{cases}
\end{align}
Table \ref{table:4} describes the 15 parameters for Case 4.
\begin{table*}
\begin{center}
\begin{tabular}{||c c c c||}
\hline
Parameter & Description & Range & Iteration type \\ [0.5ex]
\hline\hline
$g_1$ & Asymmetry Parameter & 0-1 & Linear \\
\hline
$g_2$ & Asymmetry Parameter & 0-1 & Linear \\
\hline
$\omega_1$ & Single Scattering Albedo & 0-1 & Linear \\
\hline
$\omega_2$ & Single Scattering Albedo & 0-1 & Linear \\
\hline
$\tau_1$ & Optical Depth Per Layer of Lower Cloud Deck & 0.1-30 & Linear \\
\hline
$\tau_2$ & Optical Depth Per Layer of Upper Cloud Deck & 0.1-30 & Linear \\
\hline
P$_1$ & Cloud Base Pressure Level of Lower Cloud Deck & 10$^{-6}$-10$^2$ bars & Logscale \\
\hline
P$_2$ & Cloud Base Pressure Level of Upper Cloud Deck & P$_1$-10$^{-6}$ bars & Logscale \\
\hline
a$_1$ & Cloud Deck Scale Height of Lower Cloud Deck & 10$^{-4}$ to 2 & Logscale \\
\hline
a$_2$ & Cloud Deck Scale Height of Upper Cloud Deck & 10$^{-4}$ to 2 & Logscale \\
\hline
dP & Thickness of g0/w0 Deck & 0 to P$_2$-10$^{-6}$ bars & Logscale \\
\hline
CH$_4$ & CH$_4$ Mixing Ratio & -6 to 0 & Logscale \\
\hline
NH$_3$ & NH$_3$ Mixing Ratio & -6 to 0 & Logscale \\
\hline
H$_2$O & H$_2$O Mixing Ratio & -6 to 0 & Logscale \\
\hline
g & Gravity & 25-65 m/s$^2$ & Linear \\ [1ex]
\hline
\end{tabular}
\end{center}
\caption{Parameters for the double deck cloud profile model with two valued $g$ and $w$.}
\label{table:4}
\end{table*}
\section{Dynamic Nested Sampling}\label{sec:nestedsampling}
We use the Dynamic Nested Sampling package \texttt{Dynesty}{\footnote{\url{https://dynesty.readthedocs.io/en/latest/}}} \citep{speagle20} for our retrievals. We choose Dynamic Nested Sampling to explore non-gaussian posteriors of the parameters. The Nested Sampling \citep{skilling2006} method can efficiently and accurately determine both the evidence and the posteriors of the problem simultaneously, unlike traditional MCMC that prioritizes the estimation of the posterior. Briefly, the basic process includes: 1) drawing a large number ($\sim 50\times$ number of free parameters) of live points from the provided priors of the parameter space, and then 2) iteratively replacing the live point with the least likelihood with a new live point having a greater likelihood than the replaced point. At each iteration, the replaced points become ``dead points''. The evidence can then be estimated with a set of $N$ dead points by summing over the product of their likelihoods and prior volumes \citep{skilling2006}. This process continues until a user-defined stopping criteria is met.
We wrap the retrieval model described in \S\ref{sec:retrieval} in the \texttt{Dynesty} module and retrieve on the simulated observational spectra. In each retrieval, we assign 50 live points per free parameter, as recommended \citep{speagle20}. We use the the multi- ellipsoidal decomposition method because of its ability to efficiently capture complex, multi-modal posteriors \citep{speagle20}. We use ($\Delta ln(Z_i)$) defined as,
\begin{align}
\Delta ln(Z_i) = ln(Z_i+\Delta Z_i)-ln(Z_i)
\end{align}
as our stopping criteria. Here, $Z_i$ and $\Delta Z_i$ are the current and remaining evidence estimate, respectively. The remaining evidence can be approximated by the product of the highest likelihood among remaining live-points and the prior volume of the last dead point. The retrieval is stopped when this $\Delta ln(Z_i)$ is smaller than $N/1000$ where $N$ is the number of live points. This criteria has been optimized for evidence and posterior estimation by \citet{speagle20}. We do not use any limitation on the maximum number of iterations for the retrievals.
We use the evidences estimated from the nested sampling for calculating the Bayes factor. The Bayes factor allows us to quantitatively compare each of our models. This directly can inform us whether or not one model is favored, compared to another. The Bayes factor of model M$_0$ over model M$_1$ for a dataset $D$ is,
\begin{align}
B_{01} = \dfrac{p(D|M_0)}{p(D|M_1)}
\end{align}
where $p(D|M_0)$ and $p(D|M_1)$ are the evidence of model 0 and model 1 with the dataset $D$, respectively. Pairs of models with $\ln(B)$ less than 2.1 are said to indicate that model 0 is weakly favored over model 1 at best with a confidence of less than 2$\sigma$ \citep{Trotta_2008}. If $\ln(B)$ is greater than 5, model 1 can be strongly ruled out over model 0 with a confidence of 3.6-5$\sigma$ \citep{Trotta_2008}.
We present the retrieval results obtained using \texttt{Dynesty} in the following sections.
\section{Results}\label{sec:results}
We first bin the model spectra for {\it eps Eri b}, {\it 47 Uma b}, and {\it HD 62509 b} to a constant resolution (R) of 40. We also fix the signal-to-noise ratio (SNR) of the spectra to 20 at a wavelength of 0.35 microns. This initial choice of SNR is motivated by earlier studies \citep{Lupu_2016,nayak17,Hu_19,Feng_2018} establishing this to be the minimum SNR required for accurate retrievals of atmospheric properties. The values for the R and SNR are chosen to mimic the likely best possible data quality from near future space-based direct imaging and spectroscopy missions. We also explore the effect of degrading SNR to 5. We retrieve on the albedo spectra for each planet using our four parameterizations described in \S \ref{sec:retrieval}. Here we present the results of our analysis on a planet-by-planet basis.
\subsection{47 Uma b}
\subsubsection{Comparison of Retrieval parameterizations}\label{sec:47umab}
For the first case, we retrieve atmospheric parameters using all four cases described in \S \ref{sec:retrieval} on {\it 47 Uma b}. We estimate the effective temperature of {\it 47 Uma b} to be $\sim$ 217 K. Figure \ref{fig:fig4} shows the median retrieved solution along with 1$\sigma$ and 2$\sigma$ confidence intervals. The residuals of the retrieved median spectra from the simulated data are also shown for each case in Figure \ref{fig:fig4}.
Case 1 and Case 2 show greater residuals on the blue side of the spectra compared to Case 3 and 4. The large residuals toward the blue for the Case 1 and 2 retrievals foreshadows a potential overestimation of the molecular abundances. However, comparing the performances of Case 3 and Case 4 just with the residuals and the median spectra is not quantitatively informative. Hence, we use the evidence estimates from the nested sampling calculations in order to determine how strongly one model could be ruled out or compared to another model. Specifically, we calculate the Bayes factor described in \S \ref{sec:nestedsampling} for each pair of models. The heat map is shown in Figure \ref{fig:bayes}. As suggested by Figure \ref{fig:fig4}, we see that both Case 3 and Case 4 are favored over Case 1 and Case 2. Between the two 9 parameter models, Case 2 is favored very weakly over Case 1. The 15 parameter model Case 4 is moderately favored over the 12 parameter model Case 3. Moving forward we evaluate each of the retrieval models by comparing their retrievals of various atmospheric properties with the input properties used to simulate the mock spectra.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{fig8.pdf}
\caption{The top panel shows the model albedo spectrum for {\it 47 Uma b}. The second,third, fourth and fifth panels show the comparison of retrieved spectra by each retrieval model from Case 1 to Case 4 with the simulated observed spectra of {\it 47 Uma b} shown in black diamonds. The red line shows the median albedo spectra of the retrieved models while the dark and light blue shaded regions represent the 1$\sigma$ and 2$\sigma$, respectively. {\bf Main Point}- Case 3 and 4 provide much better fits to the spectra than Cases 1 and 2.}
\label{fig:fig4}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{fig9.pdf}
\caption{Heat map of the natural logarithm of the Bayes Factor of the ``Model 0'' on the y-axis over ``Model 1'' on the x-axis. A higher Bayes factor, as is seen for Case 3 and 4 compared to Case 1 and 2, signifies that Case 3 and 4 can be favored over Case 1 and 2. Between Case 1 and 2, Case 2 very weakly rejects Case 1. Between Case 3 and 4, Case 4 moderately rejects Case 3. {\bf Main Point}- Case 3 and 4 outperform Case 1 and 2 in retrievals. Case 4 (highest complexity) is moderately favored over Case 3.}
\label{fig:bayes}
\end{figure}
Figure \ref{fig:fig5} presents the comparison of the retrieved molecular mixing ratios compared with the input molecular mixing ratio profiles, which was used to generate the simulated observed spectra for {\it 47 Uma b}. In this first case, a single value for the mixing ratio was retrieved. The full posteriors of this single retrieved value are depicted in light blue, while the altitude-dependent profile is shown in dashed red.
H$_2$O is the most dominant molecular opacity source for {\it 47 Uma b}, among the three molecules. The H$_2$O opacity causes broadband molecular absorption features starting from 0.5-1$\mu$m. Constraints on the H$_2$O mixing ratio for Case 1 and 2 are precise but not accurate. Constraints on H$_2$O mixing ratio by Case 3 and Case 4 are relatively accurate, but the precision appears to be obscured by the altitude-dependence.
The posterior of the H$_2$O mixing ratios by Case 3 and 4 is most highly peaked at the region of highest water mixing ratio of the `true' atmosphere state (deep in the atmosphere). However, the posterior contains an additional tail that extends to the depleted value. This might indicate that there is sensitivity to the depleted abundances above the cloud deck. In \S \ref{sec:depletedwater} we report the results of adding in an additional parameter to capture the depleted water profile.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{fig10.pdf}
\caption{The retrieved molecular mixing ratios of CH$_4$, NH$_3$ and H$_2$O with the input mixing ratio profile used to generate the `observed' albedo spectra for {\it 47 Uma b}. The four rows depict the retrievals for each forward model starting from Case 1 to Case 4, respectively. The first column depicts retrieval of CH$_4$, the second column is for NH$_3$ and the last column is for H$_2$O. The red dotted line shows the input mixing ratio profile. The shaded blue region shows the posterior for the retrieval in each case. {\bf Main Point}- Case 3 and 4 retrieve molecular mixing ratios accurately, but not always precisely (i.e. NH$_3$).}
\label{fig:fig5}
\end{figure}
CH$_4$ is retrieved by both Case 3 and Case 4 within 2$\sigma$ of the `true' CH$_4$ profile, though the magnitude of the 2$\sigma$ value is relatively large (an order of magnitude in abundance), and the posterior is not Gaussian. For this planet-case, this result is not surprising. CH$_4$ contributes to some of the absorption features beyond 0.7~$\mu$m but unlike \textit{eps Eri b}, the CH$_4$ opacity contribution is smaller than H$_2$O at all wavelengths. While Case 3 retrieves CH$_4$ well within the 2$\sigma$ limit of the input CH$_4$ mixing ratio profile, Case 4 better constrains CH$_4$ within 1$\sigma$. Like H$_2$O, CH$_4$ is overestimated by Case 1 and 2 with sharply peaked posteriors.
In all cases, SNR$=$20 does not allow for the precise retrieval of an NH$_3$ abundance because, although it is relatively abundant, its opacity contribution to the spectrum is negligible until 0.8$\mu$m.
Case 1 and 2 fail to constrain NH$_3$ at all and return the produce nearly uniform posteriors (the prior). Case 3 also returns a posterior whose $1\sigma$ constraint spans $\sim$2 orders of magnitude.
The retrieved posterior distribution for NH$_3$ with Case 4 shows a bias towards higher NH$_3$ abundance, and does not yield a Gaussian posterior, which points to some degeneracy. With the CH$_4$, H$_2$O and NH$_3$ retrievals for {\it 47 Uma b} described above, it is clear that Case 1 and 2 are not robust models for retrieving abundances while Case 3 and 4 retrieve the mixing ratios of nearly all the three molecules within 2$\sigma$ of the true profiles.
Cloud retrievals for each case are shown in Figure \ref{fig:fig6}. The dashed red lines show the 0.3-1$\mu$m averaged input profiles produced by \texttt{Virga}. The left column shows the optical depth per layer retrievals for the four cases. The middle and right column shows the retrieval of the asymmetry and single scattering albedo, respectively. The median profiles (in blue) along with 1$\sigma$ and 2$\sigma$ bounds (in dark and light brown) are shown for those cases where altitude-dependence was considered. For the rest, the posteriors for the single parameter are shown in with the blue shaded curves.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{fig11.pdf}
\caption{This figure shows retrieved cloud structure and optical properties compared with the input cloud structure for {\it 47 Uma b}. The four rows depict the retrievals for each parameterization starting from Case 1 to Case 4, respectively. The first column shows the optical depth per layer, the second column is the asymmetry parameter and the last column is the single scattering albedo. The red dashed line shows the input cloud structure averaged over the wavelength range of 0.3 to 1. The dark and light brown colored patches depict the 1$\sigma$ and 2$\sigma$ confidence intervals for those cases where altitude-dependence was considered. The blue shaded curves depict the posterior distribution for a singly retrieved parameter.
}
\label{fig:fig6}
\end{figure}
The optical depth per layer profiles show that
both the single deck parameterizations -- Case 1 and 2 -- trace only the upper cloud deck. Case 1 retrieves the top H$_2$O cloud deck position more precisely compared to Case 2. Whereas Case 2 retrieves the top cloud deck position more accurately compared to Case 1. Case 3 retrieves the bottom cloud deck position and optical depth profile accurately but fails to retrieve accurate or precise parameters for the top cloud deck. Similarly, Case 4 gets the bottom deck but places the top cloud deck at much lower pressures ($<10^{-4}$bars) missing the `true' cloud deck pressure by several ($\sim$ 4) orders of magnitude.
The retrieved asymmetry parameter by Case 1 and 2 are highly degenerate (the posterior is multi-modal). Case 3 retrieval of the asymmetry parameter profile is relatively precise but it overestimates the asymmetry parameter value. Case 4 retrieves an inaccurate asymmetry parameter profile with a low ($\sim$ 0.2) asymmetry parameter across most of the atmosphere except a large value ($\sim$ 0.9) at the position of the retrieved top cloud deck.
High values ($\sim$ 1) of the single scattering albedo are retrieved by Cases 1,2 and 3. These three retrieval models hence prefer reflective cloud particles, similar to H$_2$O. This is not the case for Case 4, where the retrieved bottom deck is relatively unreflective with a single scattering albedo of 0.4, whereas the top deck is highly reflective with single scattering albedo close to 1.
None of the four retrieval models succeed in retrieving the cloud structure for {\it 47 Uma b} completely. Some aspect of the position of the cloud layers are accurately captured by Case 2, 3 and 4. For example, Case 3 and 4 retrieve the position and the optical depth profile of the deeper bottom cloud deck accurately, but incorrectly retrieve either the position or optical depth of the top layer. We discuss this further in \S \ref{sec:discussion}.
\subsubsection{47 Uma b Case 3 with a Depleted Water Profile}{\label{sec:depletedwater}}
The posterior distribution of the well-mixed H$_2$O mixing ratio retrieved using Case 3 on {\it 47 Uma b} shows a bias towards lower abundances of H$_2$O, while the retrieved median value coincides with the H$_2$O mixing ratio below the cloud deck in the `true' atmosphere. This bias, shown in Figure \ref{fig:fig5}, might be indicative of model sensitivity with respect to the depletion above the top H$_2$O cloud deck. This sensitivity was similarly highlighted in the work of \citet{Damiano2020exorel}. Here, we modify the well-mixed H$_2$O retrieval model in Case 3 to accommodate a depleted water mixing ratio above the top cloud deck. This requires two additional parameter for Case 3 since we now retrieve two H$_2$O mixing ratios -- one below (H$_2$O$^1$) and one above (H$_2$O$^2$) the cloud deck. In the retrieval model, we define the cloud top pressure, P$_{\rm top}$, as the location where the optical depth of the top cloud deck is 10$^z$ times less than the peak cloud optical depth, where $z$ is the second additional parameter. The water mixing ratio is modeled to be a straight line with negative slope in the log(pressure) vs. log(H$_2$O) space between the lower and upper cloud deck pressure.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{fig12.pdf}
\caption{Comparison of the retrieved atmospheric properties of {\it 47 Uma b} using the modified Case 3 model accounting for H$_2$O depletion above the cloud deck. The top row shows the comparison of the observed spectra in black diamonds with the median (red), 1$\sigma$ (dark blue) and 2$\sigma$(light blue) regions of the retrieved spectra. The middle row shows the retrieved mixing ratios of CH$_4$, NH$_3$ and H$_2$O from left to right with the true input profiles shown in red. The last row shows the retrieved cloud properties -- optical depth per layer, asymmetry parameter and single scattering albedo from left to right with the true input profiles shown in red. {\bf Main Point}- At this SNR=20, retrieving a depleted water profile above the cloud deck for a {\it 47 Uma b}-like planet can lead to an improvement in the constraint interval of cloud and abundance properties}
\label{fig:dep}
\end{figure}
Figure \ref{fig:dep} shows the spectral fit of the retrieved results along with the cloud and molecular abundance retrievals.
We find that this additional complexity within Case 3 doesn't improve the retrieved cloud optical depth structure significantly for {\it 47 Uma b} compared to the well-mixed Case 3 retrieval. Without the depleted water profile, Case 3 is unable to constrain the top cloud deck pressure within the lower bounds of the cloud deck pressure parameter space but with the depleted water profile, the additional complexity in this modified Case 3 model helps to put 1$\sigma$ bounds on the top cloud pressure within the lowest pressure bound of the retrieved atmospheric model. The retrieval of the deeper cloud deck remains similarly accurate/precise in both the models. This parameterization also broadens the constraints on the asymmetry parameter towards lower values but the overestimation of the asymmetry parameter as seen for Case 3 in Figure \ref{fig:fig6} still persists. Retrieval of the single scattering albedo remains similar in both the cases.
This model captures the depleted H$_2$O abundance above the cloud deck within 1$\sigma$ and can also retrieve the deep H$_2$O abundance within 1$\sigma$. This modified model does not show any improvement in the retrieval of CH$_4$ over the original Case 3 model. NH$_3$ is better constrained by the modified Case 3 model compared to Case 3 (unmodified) where NH$_3$ is not at all constrained.
Despite the apparent improvement in water abundance, a Bayes factor analysis suggests that Case 3 \textit{without} water depletion is weakly preferred over Case 3 with water depletion for {\it 47 Uma b}. The Bayes factor difference is so small that without prior knowledge of the true solution, it would be difficult to discern which scenario was correct. That is, the complexity that arises from the addition of the depleted profile parameter, is not strongly favored.
\subsubsection{Retrieval at Higher Spectral Resolution (R $=$ 140)}
Future mission concepts like LUVOIR and HabEx are expected to achieve a spectral resolution of $\sim$ 140 in optical wavelengths as shown in Table \ref{table:0}. We retrieve on a R $\sim$ 140 and SNR $\sim$ 20 spectra of {\it 47 Uma b} with
the double cloud deck model, Case 3, to investigate how a higher spectral resolution from these missions can improve upon the constraints on the molecular abundances and cloud properties obtained from lower spectral resolution data expected from {\it Roman Space Telescope}.
Figure \ref{fig:fignew} left and right panel shows the corner plots of the retrieved Case 3 parameters from the {\it 47 Uma b} reflection spectra with spectral resolution of R $\sim$ 40 and 140, respectively. Here, we opt to show full set of marginalised posterior distributions of each parameter, in order to showcase correlations between the various retrieved parameters. Overall, higher resolution spectra leads to tighter constraints on multiple parameters, relative to the R $\sim$ 40 spectra with Case 3.
Specifically, the bias in the H$_2$O retrieval seen in the lower resolution retrieval with Case 3 is no longer present in the posterior retrieved from the the higher resolution spectra. The 2$\sigma$ constraints on both H$_2$O and CH$_4$ has tightened by a factor of $\sim$2 relative to the posteriors shown in Figure \ref{fig:fignew}. CH$_4$ retrieval has improved from being within 2$\sigma$ of the `true' CH$_4$ abundance in the low resolution case to being within 1$\sigma$ at high resolution. The NH$_3$ posteriors remain unconstrained in both spectral resolutions -- though this is as expected given it's opacity contribution at these wavelengths. Lastly, although the abundances improve significantly, retrievals of the cloud parameters remain very similar (within 1$\sigma$) to that obtained from the lower resolution spectra. This further motivates the utility of low resolution spectroscopy for obtaining cloud properties.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{figadd.pdf}
\caption{ Corner Plot of the Case 3 retrieval on the {\it 47 Uma b} reflected spectra with R $\sim$ 40 (left panel) and R $\sim$ 140 (right panel). This shows the posteriors of the parameters described in Table \ref{table:3} and correlations between them. The median and the 2$\sigma$ confidence intervals on each parameter is shown with blue vertical lines in the posterior distribution panel of each parameter. The median value of each parameter along with their 2$\sigma$ errors are also reported for each parameter. \textbf{Main point-} Although higher resolution improves the accuracy and precision of molecular abundances, it does not largely improve that of the cloud parameters.}
\label{fig:fignew}
\end{figure*}
\subsection{ eps Eri b}
Out of our 3 planet sample, the spectroscopic model of {\it eps Eri b} is the most similar to Jupiter, and it is our coolest target. The reflected spectra is dominated by cloud opacity, as seen in Figure \ref{fig:figphot}. Therefore, it has a zero-sloped reflection spectra with a major CH$_4$ feature at 0.73 \& 0.9$\mu$m along with a minor NH$_3$ feature. The retrieval results are shown in \ref{fig:eps_cases}.
Overall, the accuracy and precision of the retrieved cloud and mixing ratio profiles of the atmosphere were found to be very similar for Case 2 and 3, which negates the addition of more complexity in going to Case 4. Moreover, the Bayes factor calculation suggests Case 2 is weakly favored over Case 3.
Despite the multiple cloud decks made up of different condensate species, Case 3 retrieves a single deck. The maximum optical depth of the highest pressure water cloud located at 1 bar is $\tau>10$. The NH$_3$ cloud directly above the water deck only reaches a maximum optical depth of $\tau>0.5$. Therefore, when retrieving the optical depth profile, the model favors a single larger cloud deck, which spans both the H$_2$O and the NH$_3$ cloud. Despite H$_2$O and the NH$_3$ having different optical properties, the asymmetry parameter and single scattering albedo structure for the planet is retrieved accurately and precisely within the 1$\sigma$ bounds of the input profile by both the cases. Here, we define accuracy as retrieving the true value in the region of maximum cloud opacity.
Both Case 2 and 3 overestimate the CH$_4$ abundance for {\it eps Eri b}
despite the presence of two prominent CH$_4$ features in the {\it eps Eri b} albedo spectra. This CH$_4$ overestimation was also noted in \citet{lupu2016developing} for the case of {\it HD 99492 c}, which also contained two CH$_4$ absorption features similar to the {\it eps Eri b} spectra in our case. NH$_3$ is neither precise nor accurate, which is intuitive given the opacity contribution to the spectrum. H$_2$O remains unconstrained for {\it eps Eri b} because the `true' H$_2$O mixing ratio depletes at high pressures ($>1$~bars) due to water cloud formation. Therefore, there is a limited H$_2$O contribution to the albedo spectra for {\it eps Eri b}.
We performed additional retrieval tests for {\it eps Eri b} in order to determine the source of the overestimated CH$_4$. The tests we performed included: 1) fixing the gravity to the true value, 2) increasing the resolution of the simulated spectrum to R=120, and 3) removing water as a free parameter. Fixing gravity, and removing of H$_2$O did not increase the accuracy or precision of the CH$_4$ abundance. Increasing the resolution to R=120 increased the accuracy of the retrieved CH$_4$ by a two orders of magnitude, which is a clear indication that higher resolution is needed to accurately retrieve CH$_4$. At R=120, 4 data points sample the under-saturated CH$_4$ feature at 0.7 microns, while 8 data points sample the saturated feature at 0.9 microns. This is in contrast to R=40 where only 1 data point samples the under-saturated feature at 0.7 microns and 2 data points sample the feature at 0.9 microns.
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{fig13_a.pdf}
\includegraphics[width=0.45\textwidth]{fig13_b.pdf}
\caption{Comparison of the retrieved atmospheric properties of {\it eps Eri b}. The left panel shows the retrieval results using Case 2 parameterization while the right column shows retrieval results with Case 3 parameterizations. The top row shows the comparison of the observed spectra in black diamonds with the median (red), 1$\sigma$ (dark blue) and 2$\sigma$ (light blue) regions of the retrieved spectra with each parameterization. The middle row shows the retrieved mixing ratios of CH$_4$, NH$_3$ and H$_2$O from left to right in both the columns with the true input profiles shown in red. The last row shows the retrieved cloud properties -- optical depth per layer, asymmetry parameter and single scattering albedo from left to right in both the columns with the true input profiles shown in red. {\bf Main Point}- Case 2 performs slightly better than Case 3, and is preferred according to Bayes factor.}
\label{fig:eps_cases}
\end{figure*}
\subsection{HD 62509 b}
{\it HD 62509 b } is our hottest target planet with an estimated effective temperature of $\sim$533 K. The clouds in this planet are formed much deeper in the atmosphere compared to the other two cooler targets in our consideration, as is evident in Figure \ref{fig:figtp}. As a result, most of the spectra is dominated by molecular opacity and Rayleigh scattering as shown in Figure \ref{fig:figphot}. Additionally, the planet is dim compared to the other two planets in reflected light. Similar to {\it eps Eri b}, the Bayes factor calculation suggests Case 2 is weakly preferred over Case 3 for {\it HD 62509 b }. Therefore, the accuracy and precision of both Case 2 and Case 3 in retrieving cloud properties and the molecular abundances are very similar for {\it HD 62509 b}.
H$_2$O mixing ratio is retrieved accurately within 2$\sigma$ by Case 2 \& 3 for {\it HD 62509 b} as shown in Figure \ref{fig:hd_cases}. This is because, like {\it 47 Uma b}, H$_2$O is the most significant gaseous opacity source in the atmospheres of {\it HD 62509 b}. Both Case 2 \& 3 fail to constrain the CH$_4$ and NH$_3$ abundance for {\it HD 62509 b} since the spectra is dominated by H$_2$O opacity for {\it HD 62509 b} with negligible opacity contribution from CH$_4$ and NH$_3$.
For {\it HD 62509 b}, the Case 2 \& 3 model traces the region where the lower cloud deck achieves optical depth of $\sim$10. The retrieval is able to constrain the cloud optical depth structure despite: 1) the cloud being very deep in the atmosphere and 2) the relatively small cloud opacity contribution to the reflected spectra compared to our two other target planets. Both Case 2 and 3 underestimate the single scattering albedo for {\it HD 62509 b}. This is likely because the cloud deck is at such depth in the atmosphere that molecular absorption dominates the total opacity. The retrieval models hence become less sensitive to the reflectivity of the cloud particles. The posterior of the asymmetry parameter parameter nearly spans the entire prior region. This too is because of the lack of sensitivity of the model to the cloud scattering properties due to the depth of the cloud in the atmosphere.
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{fig14_a.pdf}
\includegraphics[width=0.45\textwidth]{fig14_b.pdf}
\caption{Comparison of the retrieved atmospheric properties of {\it HD 62509 b}. The left panel shows the retrieval results using Case 2 parameterization while the right column shows retrieval results with Case 3 parameterizations. The top row shows the comparison of the observed spectra in black diamonds with the median (red), 1$\sigma$ (dark blue) and 2$\sigma$ (light blue) regions of the retrieved spectra with each parameterization. The middle row shows the retrieved mixing ratios of CH$_4$, NH$_3$ and H$_2$O from left to right in both the columns with the true input profiles shown in red. The last row shows the retrieved cloud properties -- optical depth per layer, asymmetry parameter and single scattering albedo from left to right in both the columns with the true input profiles shown in red. {\bf Main Point}- Case 2 performs slightly better than Case 3 but the retrieved results are extremely similar.}
\label{fig:hd_cases}
\end{figure*}
\section{Discussion}\label{sec:discussion}
\subsection{Required Cloud Complexity
Between the first two cases with 9 parameters, Case 1 (box model) could never accurately/precisely retrieve atmospheric or scattering properties. Therefore, we advocate against box models for future retrieval work, even in simplified studies.
Between Cases 2 and Case 3 (from one to two cloud decks), the results were less clearly defined. The reflected light spectrum of {\it eps Eri b} is dominated by cloud opacity, while the hotter {\it HD 62509 b } spectrum is dominated by molecular and Rayleigh opacity. In both the cases we find that a single deck model (e.g. Case 2) can sufficiently capture the atmospheric properties and produce similar results to the double cloud deck model (e.g. Case 3). For the case of {\it 47 Uma b}, though, double deck models (Case 3 \& 4) are significantly favored over single deck models (Case 1 \& 2). Naively, this is contrary to the intuition that might be gained from looking at Figure \ref{fig:figcldopd}, which shows a strong NH$_3$ \& CH$_4$ double cloud deck for {\it eps Eri b}.
The key factor that dictated whether an additional cloud deck was necessary was the pressure location of the highest optical depth ($\tau\sim10$) cloud deck -- relative to the region of maximum contribution of molecular opacity. The {\it eps Eri b} NH$_3$ cloud deck achieved $\tau\sim0.5$ at 0.5~bars. The H$_2$O deck achieved $\tau\sim10$ at 1~bar. The bulk of the molecular opacity (see Figure \ref{fig:figphot}) was mostly contained below this optically thick H$_2$O deck (e.g. cloud layer-cloud layer-molecular opacity). This is in contrast to {\it 47 Uma b}, where the upper cloud deck achieved $\tau\sim1$ at 0.1~bars while lower deck did not achieve $\tau\sim10$ until roughly 30~bars. The molecular opacity sat between these two regions (e.g. cloud-molecular-cloud). Therefore, the former case of {\it eps Eri b} could be accurately modeled with one larger cloud deck, where as the latter case of {\it 47 Uma b} required two separate scattering regions both below and above the region of highest molecular opacity. Of course we will not know \textit{a priori} where these cloud decks exist. However, this result can inform the interpretation of future reflected light results. Retrievals that strongly favor single deck models may not ultimately reflect the true state.
Lastly, Case 4 -- two parameters to describe asymmetry and single scattering albedo -- was only moderately favored over Case 3, according to Bayes factor analysis. However, Case 4 failed to retrieve an accurate and precise single scattering albedo and the asymmetry parameter profile for {\it 47 Uma b}.
Additionally, the accuracy of the retrieved molecular abundances with Case 4 were also similar to Case 3 (within 1$\sigma$). Even though the use of Case 4 was not strongly motivated in this work, it could be useful for cases not explored here. Specifically, the use of Case 4 would be suitable for atmospheres that have stronger vertical variation in asymmetry, or single scattering albedo. For example, a case with a water cloud deck below high-altitude photochemically-produced hazes \citep[e.g.][]{gao2017sulfur} might require at least 2-parameters in each of the scattering properties.
\subsubsection{Ability to Constrain Gravity}
Unlike for most transiting planets, there generally will only be approximate constraints on gravity for directly imaged planets in reflected light. While radial velocity and a sufficient number of images will constrain $\sin i$, planet radii will still be uncertain. Therefore it is worthwhile to determine 1) whether or not gravity can be accurately retrieved from reflected light spectroscopy alone, and 2) whether or not an imprecise gravity effects the ability to retrieve accurate atmospheric properties.
In order to determine the robustness of our results with respect to imprecise gravity measurements, we allow the gravity of our planet cases to vary by $\pm$50\% of the assumed mass. With this level of uncertainty we then can explore whether or not the knowledge of the mass can be improved with the observed 0.3-1$\mu$m reflected light spectroscopy. Figure \ref{fig:fig7} shows the posteriors for the retrieved gravity for the five parameterizations explored for {\it 47 Uma b} in \S \ref{sec:47umab}.
Case 1 and Case 2 retrieve gravity posteriors which clearly favor higher gravity values whereas Case 3 (no water depletion), Case 3 (with water depletion) and Case 4 show posteriors peaking toward lower values of gravity (beyond 2$\sigma$) compared to the true input gravity shown by the dotted line in Figure \ref{fig:fig7}.
From this we draw two conclusions: 1) none of the parameterizations here could reliably retrieve gravity with the spectral resolution and SNR of reflected spectra used in this work , 2) even with $\pm$50\% inaccuracy in gravity cloud structures and abundances of molecules can be inferred from the reflected spectra of cool giants with proper choice of cloud parameterizations. Further analysis, beyond the scope of this analysis, would need to be performed to determine if these conclusions were: 1) robust against gravity constraints that were larger than $\pm$50\%, and 2) robust against unconstrained phase angle (see further discussion on phase \S \ref{sec:phase}).
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{fig15.pdf}
\caption{Comparison of the retrieved gravity posteriors for the four cases and modified Case 3 on {\it 47 Uma b}. The black dashed line depicts the true gravity of the planet. The green shaded curve depicts the posterior for Case 1 while the blue shaded curve depicts the posterior for Case 2. The posterior for Case 3 is depicted by the red shaded curve, modified Case 3 is depicted by the black shaded curve and that for Case 4 is depicted with the blue shaded curve. {\bf Main Point}- No parameterization can accurately improve the gravity beyond the $\pm$50\% \textit{a prior} value.}
\label{fig:fig7}
\end{figure}
\subsubsection{Effect of Assumed Signal-to-Noise Ratio}
Throughout the analysis we fixed SNR. We determined that for a planet like {\it 47 Uma b} the retrieved results were highly dependent on the complexity of the used retrieval model and the overall parameterization. In order to determine the robustness of this result with respect to the assumed SNR, we degrade the SNR to see whether this complexity dependence still holds for a simulated spectra of {\it 47 Uma b} with a lower SNR of 5.
Figure \ref{fig:case2_snr5} shows the retrieval results on a spectra with SNR=5 for Case 2 and Case 3 parameterizations. At lower SNR, retrievals produced with Case 1 and Case 2 parameterizations are able to fit the observed spectra. This is an intuitive result as the extra absorption features, which were seen at SNR=20, are now buried within the systematic error bars of the simulated spectra. Similar to previous results of \citet{Lupu_2016} and \citet{Feng_2018}, we find that at such low SNR, none of the cases result in precise or accurate constraints of molecular abundances. According to the Bayes Factor, Case 2 rules out Case 3 very weakly with this quality of data.
Despite not being able to constrain molecular abundances directly, we can make inferences as to where the cloud deck is relative to the molecular and Rayleigh opacity levels based on the retrieved photon attenuation map. We demonstrate this with the retrieved photon attenuation map for Case 2 shown in Figure \ref{fig:case2_snr5_phot}. Comparing this retrieved photon attenuation map with that shown in Figure \ref{fig:figphot} shows that the cloud optical depth level can be retrieved within 2$\sigma$ of the `true' opacity levels with the Case 2 parameterization. The Rayleigh opacity levels and the gas opacity levels are also retrieved within 1$\sigma$ of their respective `true' optical depth. This estimate of the cloud base pressure level from a SNR=5 albedo spectra can roughly and indirectly inform temperature-pressure structure of the atmosphere. This is because the location of the cloud deck is predicated on the region where the temperature becomes cool enough to condense a respective species. Therefore, by combining the expected equilibrium temperature of the planet, and the retrieved cloud deck, zeroth order inferences can be made about the potential temperature regime of the atmosphere.
This analysis demonstrates that at lower SNR ($\sim$ 5) observations, single cloud deck parameterizations are preferred when performing retrievals (even when two cloud decks are present). We also verify that all the models fail to capture any of the atmospheric properties like molecular abundances directly with this data quality. This remains true even at higher resolutions of 140. However, inspection of the retrieved photon attenuation maps for lower SNR observations can be informative regarding the positions of the cloud opacity and therefore, the planet's climate.
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{fig16_a.pdf}
\includegraphics[width=0.45\textwidth]{fig16_b.pdf}
\caption{Comparison of the retrieved atmospheric properties of {\it 47 Uma b} on a spectra with SNR 5. The left panel shows the retrieval results using Case 2 parameterization while the right column shows retrieval results with Case 3 parameterizations. The top row shows the comparison of the observed spectra in black diamonds with the median (red), 1$\sigma$ (dark blue) and 2$\sigma$ (light blue) regions of the retrieved spectra with each parameterization. The middle row shows the retrieved mixing ratios of CH$_4$, NH$_3$ and H$_2$O from left to right in both the columns with the true input profiles shown in red. The last row shows the retrieved cloud properties -- optical depth per layer, asymmetry parameter and single scattering albedo from left to right in both the columns with the true input profiles shown in red. {\bf Main Point}- At low SNR of 5 neither Case 2 and Case 3 are able to accurately or precisely retrieve molecular abundances.}
\label{fig:case2_snr5}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{fig17.pdf}
\caption{Photon attenuation depth map corresponding to an optical depth of 0.5 retrieved using Case 2 from the SNR=5 spectra of {\it 47 Uma b}. The attenuation pressure levels are divided into Rayleigh, gas and cloud opacity. The dark shaded yellow, purple and green regions depict 1$\sigma$ regions around the retrieved cloud, gas and Rayleigh opacity levels, respectively. The light shaded region of the same colors depict the 2$\sigma$ intervals.}
\label{fig:case2_snr5_phot}
\end{figure}
\subsection{Validity of Wavelength Independent Cloud Properties in our Retrieval Models}
We ignore any wavelength dependence while parametrizing the cloud optical properties for all of our retrieval models. However, when modeling our simulated data we consider full wavelength dependent cloud optical properties from {\texttt{Virga}}. This begs the question of whether or not additional wavelength-dependent cloud complexity would be needed for our SNR/R and wavelength range parameter space. For the cases considered here, though, this additional complexity is not necessary because of the: 1) wavelength region explored, and 2) optical properties of the cloud species explored.
The cloud optical properties of {\it eps Eri b} and {
\it 47 Uma b} show negligible wavelength dependence. This is because the optical properties of H$_2$O and NH$_3$ are not strongly wavelength dependent within the wavelength range of our focus (0.3-1 microns). Therefore, this assumption is strongly dependent on the particular species explored. For planets that are relatively hotter than {\it eps Eri b} and {\it 47 Uma b}, condensation of sulfur-based species such as MnS, Na$_2$S and ZnS, may occur. These species show strong wavelength dependence in their scattering properties \citep[e.g.][]{querry1987optical}. Although our hottest target {\it HD 62509 b} has a cloud deck dominated by sulfur species (Na$_2$S), the overall opacity is dominated by Rayleigh and molecular contribution. That is because the cloud deck of {\it HD 62509 b} is far too low in altitude (high in pressure). Therefore the necessity of wavelength dependent properties is not warranted.
Lastly, in addition to condensates, hazes can add an additional wavelength dependence. In particular, sulfur hazes, which strongly absorb light toward 0.3$\mu$m, can create positive-sloped spectra \citep{gao2017sulfur} that would require the consideration of a wavelength-dependent cloud retrieval. A Jupiter-like, wavelength-dependent haze has also been retrieved by using a simple parameterization \citep{lacy2018wfirst}. These specific cases are beyond the scope of this analysis, but could add an additional level of complexity to the parameterizations explored here.
\subsection{Additional Uncertainty Caused By Unknown Phase Angle} \label{sec:phase}
Throughout our analysis the phase angle of our target planets have been kept to zero. Phase, however, changes the albedo spectrum of a planet significantly because the scattering properties of the atmosphere are phase dependent. Therefore, an unknown phase can lead to additional uncertainty in the retrieval analysis. This effect has been the subject of previous exploration. In particular, there exists a known degeneracy between the phase angle and the radius of the planet \citep{nayak2017atmospheric}. \citet{nayak2017atmospheric} showed that an unknown phase angle does not lead to a significant change in the accuracy of retrieved molecular abundances and cloud structure, when compared to the case of known phase angle. However, when retrieving on contrast (relative planet-to-star ratio) as opposed to albedo, the unknown phase angle does introduce significant uncertainty in the radius retrieval of the planet compared to the case where phase angle is known \citet{nayak2017atmospheric}.
To test the sensitivity to phase angle in this analysis, we performed a retrieval on the albedo spectra of {\it 47 Uma b} simulated at a phase angle of 90$^{\circ}$ with our Case 3 retrieval model. We assume a uniform prior for the phase angle between 60$^{\circ}$ and 120$^{\circ}$. Similar to \citet{nayak2017atmospheric}, we find that the precision and accuracy of the retrieved molecular abundances are unchanged relative to the case of zero phase, within 1$\sigma$. We do find significant bi-modality in the retrieved cloud solution of the 1) pressure level of the optically thick (high pressure) cloud deck, 2) the gravity, and 3) the phase angle. Instead of retrieving a single peaked posterior at 30~bars, a double peaked posterior solution of 30~bars and 0.3~bars is retrieved. Therefore, the combination of unknown gravity and phase angle will impede accurate and precise determination of cloud properties. However, given the solution is strongly bimodal, inferences could be made regarding the most likely physical scenario.
\section{Conclusions and Future Work}\label{sec:conclusion}
We have performed retrievals on reflected light albedo spectra for three high priority cool giant targets for future space-based optical high-contrast imaging and spectroscopic missions like {\it HabEx} and {\it LUVOIR}. We have chosen planets with three different estimated effective temperatures of 135 K, 217 K and 533 K. This wide range of cool giant effective temperatures helps to explore retrievals of various possible cloud structure scenarios for cool giants. Albedo spectra for these planets were calculated using the spectroscopy modeling code -- \texttt{PICASO} and robust cloud calculation model -- \texttt{Virga}. We used the modeled albedo spectra to simulate mock observation spectra with a constant spectral resolution of 40 and a SNR of 20. Here we briefly discuss the key aspects and results of our retrieval analysis.
\begin{enumerate}
\item Requisite cloud complexity is highly sensitive to the relative position of the molecular, cloud, and Rayleigh opacity. The additional complexity of a second cloud deck, for example, is only favored (according to Bayes factor analysis) when the region of highest molecular opacity contribution is between the two cloud decks. Otherwise, the atmosphere can be simply parameterized with a larger, single deck.
\item Box model parameterizations for cloud opacity result in abundance measurements that are largely over-estimated. Therefore, exponential cloud opacity parameterizations (e.g. at least Case 2) should be used instead, even at low SNR$\sim$5 observations.
\item Although single scattering and asymmetry of the cloud deck changes with altitude, a 2-valued model for these scattering properties never retrieves a more accurate solution than single-valued models for these scattering properties (i.e. Case 3 retrieves the scattering properties more accurately than Case 4). This conclusion, however, might only apply to the planets explored here (dominated by NH$_3$ and H$_2$O clouds). Planets with two cloud decks composed of condensates or hazes with drastically different optical properties might warrant additional altitude-based complexity.
\item Allowing for an altitude-dependence in the H$_2$O mixing ratio profile in order to detect H$_2$O depletion above the cloud deck due to water condensation for the case of {\it 47 Uma b}, slightly improves the precision and accuracy of the abundances. However, this seemingly ``better'' solution was weakly rejected over an identical retrieval without altitude dependence. Therefore, although the fit appears better (i.e. increased precision and accuracy with respect to the 1$\sigma$ constraint interval), the additional complexity is not statistically favored for this data quality.
\item We find that even at very low SNR=5, low R=40 (0.3-1$\mu$m), inferences can be made with respect to the position of the cloud deck without attaining accurate information regarding the abundances of molecular species. In accordance with other works \citep{Lupu_2016,nayak17,Hu_19} we are unable to attain precisely constrained molecular abundances with this data quality. However, we are able to retrieve a photon attenuation map of the expected opacity contribution of rayleigh scattering, cloud scattering, and molecular absorption. This gives a limit as to the position of the bottom of the cloud deck. This suggests that very coarse, zeroth-order, temperature information could be attained by combining the equilibrium temperature of the planet, with knowledge of condensation curves.
\item Lastly, we show that the cloud structure and molecular mixing ratios of the planets can be accurately and precisely retrieved with a $\pm$50\% uncertainty in the gravity of the planets. However, it is not possible to improve the gravity constraint beyond this value.
\end{enumerate}
Returning to our initial questions posed: 1)
Users' choice of cloud and atmospheric parameterization strongly effect the precision and accuracy of the resultant abundances and cloud structure. 2) The specific location of the cloud deck, with respect to the location of the optically thick molecular opacity, dictates whether or not accurate cloud structure information can be retrieved though this information will not be known \textit{a prior}. 3) Precise gravity information is very difficult to retrieve with the quality of simulated data used here, but atmospheric characterization with reflected light is possible even with large uncertainties in planet gravity. 4) Lastly, even low SNR=5, low R=40 reflected light spectroscopy from 0.3-1$\mu$m can give insights into the cloud deck position of the planet.
\section{Acknowledgments}
SM would like to thank the S.~N.~Bose Scholar's program by Indo-US Science and Technology Forum (IUSSTF) for funding his visit to the Department of Astronomy and Astrophysics, UC Santa Cruz through the S.~N.~Bose Scholarship. MM acknowledges the support of the Nancy Grace Roman Science Investigation Team program. The authors would like to thank the exoplanet group at UC Santa Cruz especially Jonathan Fortney for all the computational support and resources used in this work. The authors will also like to thank Ryan MacDonald for insightful discussions and the anonymous referee for their suggestions, which helped in improving the manuscript.
{\it Software:} PICASO \citep{batalha19}, Virga \citep{batalha20}, DYNESTY \citep{speagle20}, numba \citep{numba}, pandas \citep{mckinney2010data}, bokeh \citep{bokeh}, NumPy \citep{walt2011numpy}, IPython \citep{perez2007ipython}, Jupyter \citep{kluyver2016jupyter},PySynphot \citep{pysynphot2013}, sqlite3, matplotlib \citep{Hunter:2007}, PyMieScatt \citep{sumlin18retrieving}
|
\section{Introduction}
Answering a question of Whitney, {\L}ojasiewicz~\cite{lojasiewicz63}
showed that every analytic variety $f^{-1}(0)$, where $f:\mathcal{U}\subset \mathbb{R}^N\rightarrow\mathbb{R}$ is real-analytic ($\mathcal{U}\neq\emptyset$, open),
is a deformation retract of its open neighborhood. The
deformation was given by the flow of the Euclidean gradient $-\nabla (f^2)$.
The main argument of {\L}ojasiewicz was based on a famous lemma, nowadays known as
the {\L}ojasiewicz (gradient) inequality, which asserts that for some
$\vartheta\in(0,1)$ and $c>0$ we have
\begin{equation}
\Vert\nabla f(x)\Vert\geq c|f(x)-f(a)|^{\vartheta}\label{eq:1}
\end{equation}
for all $x$ sufficiently close to $a\in f^{-1}(0)$. The above inequality ensures that every bounded gradient orbit $t\mapsto \gamma (t)$
({\em i.e.}, $\dot \gamma=\nabla f(\gamma)$) has finite length and therefore converges to a singular point $\gamma_{\infty}$
with $\nabla f(\gamma_{\infty})=0$. \smallskip
Some years later, Thom conjectured that in this case,
up to a change of coordinates that identifies $\gamma_{\infty}$ to~$0$, the
spherical part of the orbit also converges. In other words, the limit of secants
\begin{equation}
\lim_{t\rightarrow+\infty}\frac{\gamma(t)-\gamma_{\infty}}{||\gamma(t)-\gamma_{\infty}
||}\text{\quad exists.}\label{eq:thom}
\end{equation}
For decades, this has been known as the (Thom) gradient
conjecture, see \cite{arnold83, thom89}. (For the more general problem of non-oscillation of trajectories, we refer to~\cite{moussu97,cms07,gs13}.)
The gradient conjecture makes sense for any gradient dynamics for which bounded orbits
converge. Partial results revealed that~\eqref{eq:thom}
should hold in the real-analytic case, see ~\cite{ichikawa92,
lin92, sanz98}, fact that was eventually published in
full generality by Kurdyka, Mostowski and Parusi\'{n}ski~\cite{kmp00} in 2000.
The proof was based on \eqref{eq:1} together with concrete analytic estimations.\smallskip
{\L}ojasiewicz showed that the gradient inequality \eqref{eq:1} remains valid also for
$\mathcal{C}^{1}$ semialgebraic (respectively, globally subabalytic)
functions, see \cite{lojasiewicz84}. In 1998, Kurdyka~\cite{kurdyka98} generalized
\eqref{eq:1} for $\mathcal{C}^{1}$ functions that are \textit{definable} in
some \textit{o-minimal structure}, an axiomatic definition due to van den
Dries~\cite{dm96,dries98} which encompasses
semialgebraic and globally subanalytic functions, but also larger
classes that include the exponential function~\cite{miller94}. More precisely, Kurdyka showed
that for every definable function $f$ and critical value
$r_{\infty}$ (which is necessarily isolated) there exists $\delta>0$
and a continuous function $\Psi:[r_{\infty},r_{\infty}+\delta)\rightarrow
\mathbb{R}$ which is $\mathcal{C}^{1}$ on $(r_{\infty},r_{\infty}+\delta)$
with $\Psi^{\prime}>0$ such that
\begin{equation}
||\nabla(\Psi\circ f)(x)||\geq1 \label{eq:2}
\end{equation}
for all $x\in\mathbb{R}^{N}$ such that $r_{\infty}<f(x)<r_{\infty}+\delta$. In
addition, Kurdyka's proof showed that the function $\Psi$ can be taken in the same o-minimal
structure as $f$. Consequently, if $f$ is semialgebraic or globally
subanalytic, then so is $\Psi$ and thanks to Puiseux's theorem we may take
$\Psi(r)=r^{1-\vartheta}$, for $\vartheta\in(0,1)$. It is then straightforward
to see that (\ref{eq:2}) actually yields (\ref{eq:1}) for $c=(1-\vartheta
)^{-1}.$\smallskip
We refer to \eqref{eq:2} as the Kurdyka-{\L}ojasiewicz (in short, K{\L}) inequality and we call
K{\L}-function any function with (upper) isolated critical values that satisfies the
K{\L}-inequality around any of them. Similarly to the gradient inequality
\eqref{eq:1}, bounded gradient orbits of a K{\L}-function have finite
length. There are well-known examples of $\mathcal{C}^{\infty}$ functions in
$\mathbb{R}^{2}$ with isolated critical values that are not K{\L}-functions
(they have bounded gradient orbits which fail to converge), see \cite{fokin81, pd82}. Bounded gradient orbits of convex functions have finite length~\cite{dddl15, mp91} and therefore converge, but there are also examples
of $\mathcal{C}^{2}$-smooth convex functions failing K{\L}-property, see \cite[\S 4.3]{bdlm10} or
\cite[\S 5.1]{bp20}. In \cite{bdlm10} we characterized the class of K{\L}-functions (among the ones with upper isolated critical values) and gave criteria for a convex function to be~K{\L}.\smallskip
In \cite{kp06}, Kurdyka and Parusinski used K{\L}-inequality together with a quasiconvex cell decomposition of o-minimal sets and concrete estimates to show that the gradient conjecture holds for $\mathcal{C}^{1}$ o-minimal functions
provided either $N=2$ (planar case) or the structure is \textit{polynomially bounded} (in particular if $f$ is semialgebraic or globally subanalytic). On the other hand, mere convexity is not sufficient to guarantee~\eqref{eq:thom}: there exist examples of convex functions whose orbits either spiral \cite[\S 7.2]{dls10} or oscillate between two secants~\cite{bp20}.\smallskip
In \cite{grandjean07}, Grandjean considered the behavior of the secants
at infinity: he showed that if $f$ is a $\mathcal{C}^{1}$
semialgebraic function and $t\mapsto \gamma(t)$ is a gradient orbit satisfying
$||\gamma(t)||\rightarrow\infty,$ as $t\rightarrow+\infty,$ then the limit of
secants at infinity
\begin{equation}
\lim_{t\rightarrow+\infty}\frac{\gamma(t)}{||\gamma(t)||}\text{\quad exists\qquad
(gradient conjecture at infinity).}\label{eq:thom-inf}
\end{equation}
The proof is based on a {\L}ojasiewicz type gradient inequality at infinity
previously obtained by the author together with D'Acunto in~\cite{dg05}.\smallskip
The behavior of secants at infinity has recently become relevant in Machine Learning. If a deep network model is unbiased and homogeneous (max-pooling, ReLu, linear and convolutional layers), then minimizing the cross-entropy or
other classification losses forces the parameters of the model to diverge in
norm to infinity \cite{ll19}. In this setting, convergence of the secants at infinity
is important. In \cite{jt20} the authors manage to establish
that for a certain type of prediction functions ($L$-homogeneous and definable in the log-exp
structure) \eqref{eq:thom-inf} holds. For the time being, no further results have been reported.
\smallskip
In a nutshell, proving the gradient conjecture (respectively, the gradient conjecture at infinity) seems to require at least the K{\L}-inequality \eqref{eq:2} together with other properties of o-minimal functions, but it is still unknown if these conjectures are true for general o-minimal functions. \smallskip
In this work we present an example of a smooth convex function in $\mathbb{R}^{2}$, which is real-analytic outside zero (its unique critical point), it satisfies the {\L}ojasiewicz inequality \eqref{eq:1} and fails the gradient conjecture both at zero and at infinity. In
particular, all gradient orbits spiral both at zero and at infinity, underlying
in this way the two failures of o-minimality of the function, despite the fact that the function is convex and satisfies the {\L}ojasiewicz gradient inequality.
\begin{theorem}[main result]
\label{thm-main} For every $k\in\mathbb{N}$, there exists a $\mathcal{C}^k$-convex function
$f:\mathbb{R}^{2}\rightarrow\mathbb{R}$ with a unique minimum at $\mathcal{O}:=(0,0)$ such that:
\begin{itemize}
\item[-] $f$ is real analytic on $\mathbb{R}^{2}\setminus\{\mathcal{O}\}$ ;
\item[-] $f$ satisfies the {\L}ojasiewicz inequality at $\mathcal{O}$ and
\item[-] every maximal gradient orbit $\gamma:(-\infty,T)\rightarrow\mathbb{R}^{2}$ of $f$ spirals infinitely many times
both when $t\rightarrow-\infty$ (around the origin $\mathcal{O}$) and $t\rightarrow T$ (at infinity). As we show in Lemma~\ref{grad-traj}, $T<+\infty$, {\it i.e.}, maximal orbits blow up in finite positive time.
\end{itemize}
\end{theorem}
\noindent Throughout the manuscript, by gradient orbits (or gradient trajectories) we refer to maximal solutions of the ordinary differential equation:
$$\gamma^{\prime}(t)=\nabla f(\gamma(t)).$$
In our example, the function $f$ will be convex, with unique critical point (global minimizer) at~$\mathcal{O}$, where we tacitly assume that $\gamma(0)\not =\mathcal{O}$ (avoiding stationary orbits). \smallskip
Let us briefly describe our strategy for the construction of this example: in Section~\ref{sec:foliation} we prescribe a family of convex sets, all being delimited by ellipses, centered at the origin, and obtained via rotations and size adjustments of a basic ellipse $E(0)$. This is done in a way that convex foliation is obtained, which can be represented by some (quasiconvex) function. \smallskip
In Section~\ref{sec:cvx-fct}, we further calibrate the parameters so that we can apply a criterium due to de Finetti~\cite{definetti49} and Crouzeix~\cite{crouzeix80} that guarantees that the aforementioned quasiconvex function is in fact convex. The construction yields that the function is real-analytic on $\mathbb{R}^2\setminus\mathcal{O}$, which of course cannot be further improved to real analycity on the whole space, due to the proof of Thom's gradient conjecture~\cite{kmp00}. Instead, we are able to show that the function can be taken $\mathcal{C}^{k}$-smooth at $\mathcal{O}$ for arbitrary large $k\in\mathbb{N}$. Still our construction fails to ensure $\mathcal{C}^{\infty}$. Finally, applying a result of \cite{bdlm10} which gives conditions for a convex function to satisfy~\eqref{eq:2}, we show that our function satisfies K{\L}-inequality and in fact even \eqref{eq:1} (the {\L}ojasiewicz inequality). \smallskip
Gradient orbits are perpendicular to the foliation and explicit calculations, conducted in Section~\ref{sec:traj}, show that the orbits turn around both at the origin and at infinity, which disproves the conjecture. An additional difficulty to establish spirality is that the evolution of the spherical part of the orbit (the rotation angle $\alpha(t)$ of $\gamma(t)$ in polar coordinates) is not monotone in time, so that the decrease rate is established in average, see Figure~\ref{dess-traj-osc} and Figure~\ref{dess:fonction-alpha-prime}. For a study of monotonic spiraling of orbits of general analytic vector fields in dimensions 2 and 3, we refer to~\cite{sanz02}.
\section{Construction of a convex real analytic foliation in $\mathbb{R}^{2}\setminus\{\mathcal{O}\}$.}
\label{sec:foliation}
Let us first consider two smooth increasing functions $a,b:\mathbb{R}\to(0,+\infty)$ for which we assume:
\begin{align}
\label{hypab}
\left\{
\begin{array}
[l]{l} \text{$\underset{t\to +\infty}\lim a(t) \, =\, \underset{t\to +\infty}\lim b(t) = +\infty$}\medskip \\
\text{$\underset{t\to -\infty}\lim a(t) \, = \, \underset{t\to -\infty}\lim b(t) = 0$ \, and} \medskip \\
\text{\phantom{olivier}$a(t)\, \geq\, b(t)$, \quad for all $t\in\mathbb{R}$.}
\end{array}
\right.
\end{align}
The exact definition of the functions $a(t)$ and $b(t)$ will be given in Lemma~\ref{lem-fct-convexe} (Section~\ref{sec:cvx-fct}).
We also consider the rotation matrix by an angle $t$ denoted by:
\begin{align}
\label{rotatM}R(t)=\left(
\begin{array}
[c]{cc}
\cos t & -\sin t\\
\sin t & \cos t
\end{array}
\right)
\end{align}
For $t\in\mathbb{R}$ and $\theta \in\mathbb{T}:=\mathbb{R}/ 2\pi\mathbb{Z}$ we set
\begin{align*}
m(t,\theta):=(x(t,\theta), \,y(t,\theta))=(a(t) \cos\theta,\, b(t)\sin\theta),
\end{align*}
and
\begin{align}
\label{Mmap}M(t,\theta):=R(t)\,m(t,\theta)\,=\,(X(t,\theta), Y(t,\theta)).
\end{align}
Therefore
\begin{align}
\label{formMXY}\left\{
\begin{array}
[c]{l}
X(t,\theta)= x(t,\theta)\cos t - y(t,\theta)\sin t\, =\, a(t) \cos t \cos\theta- b(t) \sin t \sin\theta \smallskip\\
Y(t,\theta)=\, x(t,\theta)\sin t + y(t,\theta)\cos t \,=\, a(t) \sin t \cos\theta+ b(t) \cos t \sin\theta\,.
\end{array}
\right.
\end{align}
The subset
\begin{align}
\label{ellEt}\mathcal{E}(t):=\{M(t,\theta) : \theta\in\mathbb{T}\}
\end{align}
is an ellipse with major axis of length $a(t)$ and minor axis of length $b(t)$ (see Figure~\ref{dess_ellipse-rot} for illustration). Notice that $\mathcal{E}(t)$ is the rotation by
angle $t$ of the ellipse
\begin{align*}
E(t):=\,\big\{m(t,\theta) :\,\theta\in\mathbb{T}\big\}\,\,=\,\,\left\{ \, (x,y)\in\mathbb{R}^{2}:\,\,\,
\frac{x^{2}}{a^{2}(t)}+\frac{y^{2}}{b^{2}(t)}=1\,\right\} .
\end{align*}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=14cm]{dess_ellipses_bis.pdf}
\caption{The ellipse $\mathcal{E}(t)$ and the map $(t,\theta)\mapsto M(t,\theta)$}
\label{dess_ellipse-rot}
\unitlength=1pt
\begin{picture}(0,0)
\put(-12,163){$\scriptstyle \mathcal{O}$}
\put(41,170){$\scriptstyle m(0,\theta)$}
\put(120,204){$m(t,\theta)$}
\put(75,260){$M(t,\theta)$}
\put(46,155){$\scriptstyle a(0)$}
\put(0,183){$\scriptstyle b(0)$}
\put(154,155){$\scriptstyle a(t)$}
\put(0,225){$\scriptstyle b(t)$}
\put(70,170){$\phi$}
\put(74,222){$\phi$}
\put(94,177){$t$}
\put(104,232){$t$}
\put(191,165){$x$}
\put(2,283){$y$}
\put(90,101){$E(t)$}
\put(-35,70){$\mathcal{E}(t)$}
\end{picture}
\end{center}
\end{figure}
Under an additional condition on the functions $a,b$, the family of ellipses $\{\mathcal{E}(t)\}_{t\in\mathbb{R}}$ defined in~\eqref{ellEt} is disjoint with union equal to $\mathbb{R}^2\setminus\{\mathcal{O}\}$. More precisely, denoting by $a^{\prime}$, $b^{\prime}$ the derivatives of the functions $a$, $b$ respectively, we have the following result:
\begin{lemma} [Convex foliation by ellipses]\label{lem-feuilletage}
Let $a,b:\mathbb{R}\to(0,+\infty)$ satisfy~\eqref{hypab} and assume
\begin{align}
\label{cond-foliation}4\, a(t) \,b(t)\, a^{\prime}(t)\,b^{\prime}(t)\, >\, (a(t)^2-b(t)^2)^{2}\,,\qquad\text{for all }\, t\in\mathbb{R}.
\end{align}
Then $(\mathcal{E}(t))_{t\in\mathbb{R}}$ defines an analytic convex foliation of $\mathbb{R}^{2}\setminus\{\mathcal{O}\}$.
\end{lemma}
\begin{proof
The proof is divided in three steps: \bigskip \newline
{\it Step 1.} The map $M: \mathbb{R}\times \mathbb{T}\to \mathbb{R}^2\setminus \{\mathcal{O}\}$ is a local analytic diffeomorphism. \medskip\newline
\noindent Indeed, let us first notice that the map $M$, defined by~\eqref{Mmap}--\eqref{formMXY}, is real-analytic as composition of analytic functions. Therefore, if we show that the Jacobian matrix
$\mathcal{J}M
=\left(\begin{array}{cc}\frac{\partial X}{\partial t} & \frac{\partial X}{\partial \theta}\\[1mm]\frac{\partial Y}{\partial t} & \frac{\partial Y}{\partial \theta}\end{array}\right)$
is invertible at each point $(t,\theta)\in \mathbb{R}\times \mathbb{T}$, the assertion follows from the local analytic inverse function theorem~\cite[Theorem~2.5.1]{kp02}.
To this end, we shall prove that
\begin{eqnarray}\label{eq:2.6}
{\rm det}(\mathcal{J} M)= \frac{\partial X}{\partial t}\,\frac{\partial Y}{\partial \theta} - \frac{\partial Y}{\partial t}\, \frac{\partial X}{\partial \theta}
=\big\langle \frac{\partial M}{\partial t}, n\big\rangle >0,
\end{eqnarray}
where
$n(t,\theta)=-R(\frac{\pi}{2})\,\frac{\partial M}{\partial \theta}= (\frac{\partial Y}{\partial \theta}, -\frac{\partial X}{\partial \theta})$ is the outer unit normal
to the convex set ${\rm conv}\,\mathcal{E}(t)$ (convex envelope of $\mathcal{E}(t)$) at $M(t,\theta)$. Recalling that $M(t,\theta)=R(t)\,m(t,\theta )$ (see~\eqref{Mmap}) and that
the rotation matrix~\eqref{rotatM}
satisfies $$R'(t)=R(t+\frac{\pi}{2}), \quad R(t)^{-1}=R(t)^T=R(-t) \quad \text{and} \quad R(t)\,R(s)=R(t+s),$$
we deduce
\begin{eqnarray*}
\big\langle \frac{\partial M}{\partial t}, n\big\rangle
&=&
\big\langle \frac{\partial}{\partial t}(R(t)m), -R(\frac{\pi}{2})\,\frac{\partial}{\partial \theta} (R(t)m)\,\big\rangle\\
&=&
\big\langle R'(t)m +R(t) \frac{\partial m}{\partial t}, \,-R(\frac{\pi}{2})\,R(t)\, \frac{\partial m}{\partial \theta}\big\rangle\\
&=&
\big\langle R(t+\frac{\pi}{2})\,m +R(t)\frac{\partial m}{\partial t},\, R(t-\frac{\pi}{2})\, \frac{\partial m}{\partial \theta}\big\rangle\\
&=&
\big\langle R(t-\frac{\pi}{2})^T R(t+\frac{\pi}{2})m, \,\frac{\partial m}{\partial \theta}\big\rangle\,
+\, \big\langle R(t-\frac{\pi}{2})^T R(t)\frac{\partial m}{\partial t}, \, \frac{\partial m}{\partial \theta}\big\rangle\\
&=&
- \big\langle m, \frac{\partial m}{\partial \theta}\big\rangle\, +\,\big\langle R(\frac{\pi}{2})\frac{\partial m}{\partial t},\, \frac{\partial m}{\partial \theta}\big\rangle.
\end{eqnarray*}
Plugging $$\frac{\partial m}{\partial \theta}=(-a\sin\theta , b\cos\theta)\qquad \text{ and } \qquad \frac{\partial m}{\partial t}= (a'\cos\theta, b'\sin\theta)$$
into the above equality, we end up with the expression:
\begin{eqnarray}\label{cond-normale}
{\rm det}(\mathcal{J} M)=\langle \frac{\partial M}{\partial t}, n\rangle = a' b \cos^2\theta + a b' \sin^2\theta + (a^2-b^2)\cos\theta \sin\theta.
\end{eqnarray}
This is a quadratic expression with respect to $\cos\theta$ and $\sin\theta$,
which is positive for all $\theta\in \mathbb{T}$ if and only if the discriminant
$(a^2-b^2)^2-4aa'bb'$ is negative. The result follows in view of~\eqref{cond-foliation}. \bigskip
\noindent{\it Step 2.} The map $M: \mathbb{R}\times \mathbb{T}\to \mathbb{R}^2\setminus\{\mathcal{O}\}$ is injective. \medskip\\
\noindent Fix $t\in \mathbb{R}$.
From~\eqref{eq:2.6}--\eqref{cond-normale}, using compactness of $\mathcal{E}(t)$ and smoothness
of $M$, we deduce the existence of $\delta_t, \rho_t >0$ such that, for all
$s\in [t,t+\delta_t]$, $\theta\in\mathbb{T}$,
\begin{eqnarray*}
\big\langle \,\frac{\partial M}{\partial t}(s,\theta), n(t,\theta)\, \big\rangle \geq \rho_t >0,
\end{eqnarray*}
which yields
\begin{eqnarray*}
\big\langle M(s,\theta)-M(t,\theta), n(t,\theta)\big\rangle \,\geq \,\rho_t(s-t)\, >\,0,
\quad \text{for \,$t<s\leq t+\delta_t$\,\,\text{and }\,$\theta\in\mathbb{T}$.}
\end{eqnarray*}
It follows that ${\rm conv}\,\mathcal{E}(t)\subset {\rm int}\,{\rm conv}\,\mathcal{E}(s)$
for all $s>t$. Therefore, the family $({\rm conv }\,\mathcal{E}(t))_{t\in\mathbb{R}}$ is nested
and the map $M$ is injective. \medskip\\
\noindent {\it Step 3.} The map $M: \mathbb{R}\times \mathbb{T}\to \mathbb{R}^2\setminus\{\mathcal{O}\}$ is surjective. \smallskip\\
\noindent Fix $(x,y)\in \mathbb{R}^2\setminus\{\mathcal{O}\}$ and set, for $t\in\mathbb{R}$ and
$D(t)=\left(\begin{array}{cc} a(t) & 0\\ 0 & b(t)\end{array}\right)$,
\begin{eqnarray*}
\rho(t) &:= &||D(t)^{-1}R(t)^{-1}(x,y)||^
=\frac{1}{a^2(t)}(x\cos t +y\sin t)^2 + \frac{1}{b^2(t)}(-x\sin t +y\cos t)^2.
\end{eqnarray*}
We claim that $\rho$ is a smooth decreasing function with
$\displaystyle\lim_{-\infty}\rho= +\infty$ and
$\displaystyle\lim_{+\infty}\rho= 0$. \smallskip
\noindent Indeed, since $(x,y)\not= (0,0)$, we get
$R(t)^{-1}(x,y)\not= (0,0)$ and either
$x\cos t +y\sin t\not= 0$ or $-x\sin t +y\cos t\not= 0$.
Recalling that $a(t),b(t)\to 0$ as $t\to -\infty$, we deduce
$\displaystyle\lim_{-\infty}\rho= +\infty$. We also observe that $\displaystyle\lim_{+\infty}\rho= 0$
is a direct consequence of the fact $a(t),b(t)\to +\infty$ as $t\to +\infty$.\smallskip
\noindent It remains to prove that $\rho'$ is negative. To this end, set $q(t):=x\cos t +y\sin t$ and notice that
$\rho = a^{-2} q^2 + b^{-2} (q')^2$. Using that $q''=-q$,
we infer
\begin{eqnarray*}
\rho'(t) &= &
-2a'a^{-3}q^2 + 2 a^{-2} q' q - 2b'b^{-3}(q')^2+ 2 b^{-2} q'' q' \\
&=&
-2 a^{-2} b^{-2} \left( a'a^{-1}b^2 q^2 + (a^2-b^2)qq'+ b'b^{-1}a^2 (q')^2\right).
\end{eqnarray*}
The quadratic expression $a'a^{-1}b^2 q^2 + (a^2-b^2)qq'+ b'b^{-1}a^2 (q')^2$
with respect to $q$ and $q'$
is positive if and only if its discriminant is negative, which
is equivalent, once again, to assume~\eqref{cond-foliation}.
Thus $\rho$ is strictly decreasing and the claim follows.\smallskip
\noindent Using the claim, we infer that there exists a unique $\overline{t}\in\mathbb{R}$
such that $$\rho(\overline{t})=||D(\overline{t})^{-1}R(\overline{t})^{-1}(x,y)||^2=1.$$
Therefore, there exists a unique $\overline{\theta}\in\mathbb{T}$ such that
$D(\overline{t})^{-1}R(\overline{t})^{-1}(x,y)=(\cos\overline{\theta},\sin\overline{\theta})$.
It follows that $M(\overline{t}, \overline{\theta})=(x,y)$, which proves that $M$ is onto.
\end{proof}
\smallskip
\begin{figure}[!h]
\begin{center}
\includegraphics[width=12cm]{dess_foliation.pdf}
\end{center}
\caption{The convex foliation $(\mathcal{E}(t))_{t\in\mathbb{R}}$ for $a(t)=2b(t)=2e^t$.}
\label{dess:feuilletage-ellipses}
\end{figure}
A typical instance where Lemma~\ref{lem-feuilletage} applies is to take $a=\mu
b$ for some constant $\mu>1$. Then for
$b(t)=e^{\nu t}$ with $\nu>\frac{\mu^{2}-1}{2\mu}$, it is straightforward to check that
$a,b$ satisfy~\eqref{hypab} and~\eqref{cond-foliation}. Figure~\ref{dess:feuilletage-ellipses} represents the explicit choice
$\mu=2$ and $\nu=1$ leading to $a(t)=2 e^{t}$ and $b(t)=e^{t}$.
\section{Defining the convex function and regularity properties}
\label{sec:cvx-fct}
In this section we shall show that for a more precise choice of the functions $a(t), b(t)$ we can construct a convex function whose level sets are exactly the foliation $\{ \mathcal{E}(t) \}_{t\in\mathbb{R}}$. Moreover, we shall show that this convex function is smooth, real-analytic on
$\mathbb{R}^2\setminus\{\mathcal{O}\}$ and satisfies~\eqref{eq:1}. \smallskip \\
Concretely, let us denote by $\varphi:\mathbb{R}\to\mathbb{R}$ a smooth strictly increasing function
satisfying $\varphi(-\infty):=\underset{t\rightarrow-\infty}{\lim}\varphi(t)=0$
(the concrete definition of the function $\varphi$ will be given in~\eqref{cond-a-b-varphi-nu}, see Lemma~\ref{lem-fct-convexe}) and let us set for all $M\in\mathbb{R}^{2}$
\begin{align}
\label{def-f}
f(M)=\left\{
\begin{array}
[c]{ll}
\phantom{0}0\,, & \text{if $M=(0,0)$,}\medskip \\
\varphi(t), & \text{if \,$M\in\,\mathcal{E}(t)$},\\
\end{array}
\right.
\end{align}
where $\mathcal{E}(t)$ is the ellipse given in~\eqref{ellEt}. We shall now show that we can adjust the parameters and choose $\varphi$ in a way that~\eqref{def-f} gives a well-defined convex function.
\begin{lemma} [Construction of the convex function]\label{lem-fct-convexe} Setting for $t\in\mathbb{R}$
\begin{align}
\label{cond-a-b-varphi-nu}
\begin{array}
[c]{cc}
a(t)= \sqrt{2} \,\, {\rm exp}({t}), \quad b(t)={\rm exp}(t) &
\text{in~\eqref{formMXY},} \medskip \\
\varphi(t)= {\rm exp}(t/\tau), \quad\, \tau\in(0, \frac{1}{10}), & \text{in~\eqref{def-f}},
\end{array}
\end{align}
the function $f$ defined by~\eqref{def-f} is convex, with level
sets the ellipses $\mathcal{E}(t)$ and $\mathrm{argmin}
\,f=\{\mathcal{O}\}$.
\end{lemma}
\begin{proof
Since the functions $a,b$ satisfy~\eqref{hypab} and~\eqref{cond-foliation}, we deduce by Lemma~\ref{lem-feuilletage} that
${\rm conv }(\mathcal{E}(t))_{t\in\mathbb{R}}$ is a convex foliation. In particular, the function $f$ is well defined from~\eqref{def-f} with sublevel sets
$$
[f\leq \lambda]:=\{M\in\mathbb{R}^2:\,f(M)\leq \lambda\}=\mathrm{conv }\, [\mathcal{E}(\varphi^{-1}(\lambda))]
= \mathrm{conv }\, [\mathcal{E}(\tau\log \lambda)]
$$
compact and convex. Therefore $f$ is a coercive, quasiconvex function.
\smallskip\\
\noindent We shall now use a result due to de Finetti and Crouzeix~\cite{definetti49,crouzeix80} which asserts that the quasiconvex function $f$ is convex if and only if $$\lambda\mapsto \sigma_{[f\leq \lambda]}(p) \,\,\text{is concave for every }\,p\in\mathbb{R}^2,$$
where $\displaystyle\sigma_A (p)=\max_{M\in A}\, \langle p, M\rangle$ is the support function to the subset $A$.
Without loss of generality, we may restrict to unit vectorss $p\in\mathbb{R}^2$, which results in assuming that
$p=(\cos\alpha , \sin\alpha )$, for some $\alpha\in\mathbb{T}$. Therefore, we are led to prove that the function
\begin{eqnarray*}
G_\alpha (\lambda) &:=& \sup \left\{ \Big\langle (x,y),(\cos\alpha , \sin\alpha )\Big\rangle : \,f(x,y)\leq \lambda \right \}\\
&=&
\sup \left\{ \Big\langle M(t,\theta),(\cos\alpha , \sin\alpha )\Big\rangle : \,f(M(t,\theta))=\varphi(t)\leq \lambda\right \}\\
&=&
\max \left \{ \Big\langle M(t,\theta),(\cos\alpha , \sin\alpha )\Big\rangle : \theta\in\mathbb{T}, t=t(\lambda)=\varphi^{-1}(\lambda) \right \}
\end{eqnarray*}
is concave. To this, end, after straightforward calculations we obtain
\begin{eqnarray*}
\Big\langle M(t,\theta),(\cos\alpha , \sin\alpha )\Big\rangle
&=&\Big\langle R(t)\,m(t,\theta),\,(\cos\alpha , \sin\alpha )\Big\rangle\\
&=&\Big\langle (a(t)\cos\theta,\, b(t)\sin\theta),\, R(-t)\,(\cos\alpha , \sin\alpha )\Big\rangle\\
&=& \Big\langle (\cos\theta, \sin\theta ),\,\left(a(t)\cos(\alpha -t) , b(t)\sin (\alpha-t)\right)\Big\rangle
\end{eqnarray*}
whence we deduce
\begin{eqnarray}\label{eq:G_a}
G_\alpha (\lambda) \,=\, \Big\Vert \,a(t(\lambda))\cos(\alpha -t(\lambda)) ,\, b(t(\lambda))\sin (\alpha-t(\lambda))\,\Big\Vert
\,=\, \sqrt{g_\alpha (\lambda)}
\end{eqnarray}
with
\begin{eqnarray}\label{eq:g_a}
g_\alpha (\lambda) =a(t(\lambda))^2\cos^2(t(\lambda)-\alpha)\,+\, b(t(\lambda))^2\sin^2(t(\lambda)-\alpha).
\end{eqnarray}
Calculating the second derivative of $G_{\alpha}$ in~\eqref{eq:G_a} yields
\begin{eqnarray*}
G_\alpha''=\frac{2 g_\alpha'' g_\alpha -(g_\alpha')^2}{4 g_\alpha^{3/2}}.
\end{eqnarray*}
Therefore, the functions $\{G_\alpha\}_{\alpha\in \mathbb{T}}$ are concave provided we establish:
\begin{equation}\label{eq:con}
2 g_\alpha'' g_\alpha -(g_\alpha')^2\leq 0, \quad\text{for all } \alpha\in\mathbb{T}\,.
\end{equation}
At this step, we replace in~\eqref{eq:g_a} the choice for $a$, $b$ and $\varphi$ given in~\eqref{cond-a-b-varphi-nu}:
$$ a(t)=\sqrt{2} \,e^t\,, \quad b(t)=e^t \quad\text{and }\, \lambda= \varphi(t)=e^{t/\tau},\quad\text{for all }\, t\in\mathbb{R},$$
and we seek for the values of $\tau>0 $ that ensure inequality~\eqref{eq:con}.
In particular,
$$ t:=t(\lambda)=\tau\log\lambda, \quad \text{whence }\, t'(\lambda)=\frac{\tau}{\lambda} \,\,\text{and }\, t''(\lambda)=-\frac{\tau}{\lambda^2}<0.$$
After tedious computations, we get
$$ g_\alpha= e^{2t}\left(\cos^2(t\!-\!\alpha)+1\right),\qquad
g_\alpha' = 2 \,e^{2t}\, t' \left (\cos^2(t\!-\!\alpha)+1-\cos(t\!-\!\alpha)\sin(t\!-\!\alpha)\right) $$
and
\[
g_\alpha'' = 2 e^{2t}\,\Big( (t')^2 \big( 3-4\cos(t\!-\!\alpha)\sin(t\!-\!\alpha)\big) + t'' \left( \cos^2(t\!-\!\alpha)+1 -\cos(t\!-\!\alpha)\sin(t\!-\!\alpha)\right)\Big ).
\]
Hence \medskip \\ $2 g_\alpha'' g_\alpha - (g_\alpha')^2 =$
\begin{eqnarray*}
&=& 4 e^{4t}(t')^2\Big\{ \big( \cos^2(t\!-\!\alpha)+1\big)\big(3- 4 \cos(t\!-\!\alpha)\sin(t\!-\!\alpha)\big ) -\big(\cos^2(t\!-\!\alpha)+1- \cos(t\!-\!\alpha)\sin(t\!-\!\alpha)\big)^2\medskip \\
&\phantom{ley}&\hspace{4cm} +\,\,4\, e^{4t}\,t''\, \big( \cos^2(t\!-\!\alpha)+1 \big)\big(\cos^2(t\!-\!\alpha)+1 - \cos(t\!-\!\alpha)\sin(t\!-\!\alpha) \big)\Big\} \smallskip\\
&\leq&
4 e^{4t} \Big( 5 (t')^2 +\frac{1}{2}t'' \Big)\quad\leq\quad
\frac{2\tau (10\tau -1)e^{4t}}{\lambda^2},
\end{eqnarray*}
which is negative provided we choose $\tau < 1/10$.
\end{proof}
\bigskip
\noindent We fix $M:\mathbb{R}\times\mathbb{T}\mapsto
\mathbb{R}^{2}\backslash\{\mathcal{O}\}$ under the choice made in
Lemma~\ref{lem-fct-convexe}, that is,
\begin{equation}
M(t,\theta)=(X(t,\theta),Y(t,\theta))=e^{t}\left( \sqrt{2}\cos t\cos
\theta-\sin t\sin\theta,\sqrt{2}\sin t\cos\theta+\cos t\sin\theta
\,\right) . \label{def-M}
\end{equation}
Setting
\begin{equation}
\left\{
\begin{array}
[c]{l}
\tilde{f}:\mathbb{R}\times\mathbb{T}\mapsto\mathbb{R}\medskip \\
\tilde{f}(t,\theta)=\varphi(t)=\exp(t/\tau)
\end{array}
\right.
\label{ftilde}
\end{equation}
we observe that the convex function $f$ defined in~\eqref{def-f}
satisfies:
\begin{equation}
f(x,y)=\left\{
\begin{array}
[c]{cc}
(\tilde{f}\circ M^{-1})(x,y), & \text{if \thinspace$(x,y)\neq \mathcal{O}$},\smallskip \\
\phantom{0}0\,, & \text{if $(x,y)=\mathcal{O}$.}
\end{array}
\right. \label{def-f-comp}
\end{equation}
With the next couple of lemmas we show that the function $f$, apart from being convex, enjoys several other good properties.
\begin{lemma} [Properties of the convex function]\label{lem-regularite-fct} Let
$f:\mathbb{R}^{2}\mapsto\lbrack0,+\infty)$ be the convex function
defined by~\eqref{def-M}--\eqref{def-f-comp} for $0<\tau<1/10$. Then
\begin{itemize}
\item[{\rm (i).}] $f$ is strictly positive on $\mathbb{R}^{2}\setminus\{\mathcal{O}\}$ with $f(\mathcal{O})=0\,.$
\item[{\rm (ii).}] For all $(x,y)\in\mathbb{R}^{2}$, it holds
\begin{align}
\label{ineg-f}
\left( 1/\sqrt{2}\right)^{1/\tau}\big \Vert(x,y)\big\Vert^{1/\tau} \,\leq \, f(x,y) \leq
\,\big\Vert (x,y)\big\Vert ^{1/\tau}.
\end{align}
In particular, $f$ is coercive.
\item[{\rm (iii).}] $f$ is real analytic on $\mathbb{R}^{2}\setminus\{\mathcal{O}\}$ and $f\in \mathcal{C}^{1}(\mathbb{R}^{2})$\,.
\item[{\rm (iv).}] $f$ satisfies the {\L}ojasiewicz inequality~\eqref{eq:1} with $\vartheta=1-\tau$, $c=\tau/\sqrt{2}$, $a\equiv\mathcal{O}$ and $f(\mathcal{O})=0$, that is
\begin{align}
\label{ineg-loja1}\Vert \nabla f (x,y)\Vert \, \geq\, \left(\frac{\tau}{\sqrt{2}}\right)\, f(x,y)^{1-\tau}\,,
\qquad\text{for all $(x,y) \in\mathbb{R}^{2}$.}
\end{align}
\end{itemize}
\end{lemma}
\begin{proof}
(i). It is straightforward from the definition of $f$ in~\eqref{def-f} and the choice of $\varphi$.\smallskip\\
\noindent (ii). From Lemma~\ref{lem-feuilletage}, for every $(x,y)\in \mathbb{R}^2\setminus\{\mathcal{O}\}$, there exists a unique $t\in\mathbb{R}$
such that $(x,y)\in\mathcal{E}(t)$ and we have
\begin{eqnarray*}
\frac{x^2+y^2}{a^2(t)} \,
\leq\,\,
\frac{1}{a^2(t)}(x\cos t +y\sin t)^2 + \frac{1}{b^2(t)}(-x\sin t +y\cos t)^2
\,=\, 1\,\,
\leq \,\,\frac{x^2+y^2}{b^2(t)},
\end{eqnarray*}
whence
\begin{eqnarray*}
e^t=b(t)\leq\, \Vert(x,y)\Vert\,\leq a(t)=\sqrt{2}e^t.
\end{eqnarray*}
We deduce easily that
\begin{eqnarray*}
2^{-1/(2\tau)}\,\Vert(x,y)\Vert^{1/\tau} \,\leq \, f(x,y)=\varphi(t)=e^{t/\tau}\, \leq\, \Vert(x,y)\Vert^{1/\tau}.
\end{eqnarray*}
\noindent (iii). It follows from~\eqref{def-f} that $f=\varphi\,\circ\, p_1\,\circ M^{-1}$ on $\mathbb{R}^2\setminus\{\mathcal{O}\}$, where $p_1: \mathbb{R}\times \mathbb{T} \mapsto \mathbb{R}$ with $p_1(t,\theta)= t$. By Lemma~\ref{lem-feuilletage}, the map $M:\mathbb{R}\times \mathbb{T}\mapsto \mathbb{R}^2\setminus\{\mathcal{O}\}$
given in~\eqref{def-M} is a real analytic diffeomorphism. Since $p_1$ and
$\varphi$ are analytic, the first part of the assertion follows. In particular, the function $f$ is $\mathcal{C}^\infty$-smooth on $\mathbb{R}^2\setminus\{\mathcal{O}\}$. \smallskip\\
Since $1/\tau >1$, the function $(x,y)\mapsto \Vert(x,y)\Vert^{1/\tau}$ is~$\mathcal{C}^1$ over $\mathbb{R}^2$ and~\eqref{ineg-f} yields that $f$ is differentiable at $\mathcal{O}$ with $\nabla f(\mathcal{O})=0$.
Therefore $f$ is differentiable everywhere in $\mathbb{R}^2$ and,
since it is convex, it is $\mathcal{C}^1$ (see for instance, \cite[p.~20]{phelps89}).
\bigskip\noindent(iv) Since $S:=\mathrm{argmin}\,f=\{\mathcal{O}\}$, we have
$\mathrm{dist}_{S}(M)=\Vert M\Vert$ for all $M\!=\!(x,y)\!\in
\!\mathbb{R}^{2}$. Therefore, the first inequality in~\eqref{ineg-f} can be
written
\[
f(M)\,\geq\,\mathbf{m}\!\left( \mathrm{dist}_{S}(M)\right) \quad\text{for all
$M\in\mathbb{R}^{2}$},
\]
where $\mathbf{m}(r)=2^{-1/(2\tau)}\,r^{1/\tau}$. Since
\[
\frac{\mathbf{m}^{-1}(s)}{s}=\sqrt{2}\,s^{\tau-1}\in L_{\mathrm{loc}}^{1}%
((0,+\infty)),
\]
we deduce from \cite[Theorem~30]{bdlm10} that the K{\L }-inequality
\[
\Vert\nabla(\psi\circ f)(M)\Vert\geq1,
\]
holds for all $M\in\lbrack f>0]:=\mathbb{R}^{2}\setminus\{\mathcal{O}\}$,
where
\[
\psi(s)=\int_{0}^{s}\frac{\mathbf{m}^{-1}(\sigma)}{\sigma}\,d\sigma\,=\,\frac{\sqrt{2}%
}{\tau}\,s^{\tau}.
\]
A straightforward calculation shows that~\eqref{ineg-loja1} holds.
\end{proof}
\begin{lemma}
[$\mathcal{C}^k$-smoothness of the convex function]\label{lem-Ck-fct} Let
$f$ be the convex function
defined by~\eqref{ftilde}--\eqref{def-f-comp} for $0<\tau<1/10$.
Let $k\in\mathbb{N}$ be the biggest integer such that $k<\frac{1}{\tau}$.
Then $f\in \mathcal{C}^k(\mathbb{R}^2)$ and $f\not\in \mathcal{C}^{k+1}(\mathbb{R}^2)$.
\end{lemma}
\begin{proof}
Recalling that $f$ is real analytic in $\mathbb{R}^2\setminus\{\mathcal{O}\}$ with $f(\mathcal{O})=0$ and $\nabla f(\mathcal{O})=0$, in order
to prove that $f$ is $\mathcal{C}^{k}$, it is sufficient to show
that all the partial derivatives
\begin{equation}
\frac{\partial^{l_1+l_2} f}{\partial x^{l_1}\partial y^{l_2}}, \quad l_1+l_2\leq k,
\end{equation}
which exist in $\mathbb{R}^{2}\setminus\{\mathcal{O}\}$, converge to $0$ at
$\mathcal{O}$. To this end, it is more convenient to start by computating the partial derivatives of
$\tilde{f}$ defined in~\eqref{ftilde}. We have
\begin{eqnarray*}
\tilde{f}(t,\theta):=f(M(t,\theta))=e^{t/\tau}=f(x,y) \quad\text{for $(x,y)=M(t,\theta)=(X(t,\theta),Y(t,\theta))$,}
\end{eqnarray*}
and by differentiation, we obtain
\begin{eqnarray}\label{diff1-f}
\left(\begin{array}{c}\frac{\partial\tilde{f}}{\partial t}\\[1mm]\frac{\partial\tilde{f}}{\partial \theta}\end{array}\right)
= \left(\begin{array}{c} \frac{1}{\tau}e^{t/\tau} \\0 \end{array}\right)
=
\left(\begin{array}{cc}\frac{\partial X}{\partial t} &\frac{\partial Y}{\partial t}\\[1mm]
\frac{\partial X}{\partial \theta} & \frac{\partial Y}{\partial \theta}
\end{array}\right)
\left(\begin{array}{c}\frac{\partial f}{\partial x}\\[1mm]\frac{\partial f}{\partial y}\end{array}\right).
\end{eqnarray}
We can compute explicitely the partial derivatives of $X$ and $Y$, see~\eqref{def-M}, to obtain
\begin{eqnarray*}
\frac{\partial X}{\partial t}, \frac{\partial Y}{\partial t}, \frac{\partial X}{\partial \theta}, \frac{\partial Y}{\partial \theta}
= e^t P(t,\theta),
\end{eqnarray*}
where $P(t,\theta)$ denotes generically a smooth periodic (hence bounded) function with respect to~$t$ and~$\theta$.
More generally, in what follows, $P_{n,m}(t,\theta)$ (respectively $B_{n,m}(t,\theta)$) denotes a $n\times m$ matrix,
the coefficients of which are smooth and periodic with respect to $t$ and $\theta$ (respectively bounded in $(-\infty,1]\times\mathbb R$). It follows that
\begin{eqnarray*}
\left(\begin{array}{c}\frac{\partial f}{\partial x}\\[1mm]\frac{\partial f}{\partial y}\end{array}\right)
= \frac{1}{\frac{\partial X}{\partial t}\frac{\partial Y}{\partial \theta}-\frac{\partial Y}{\partial t}\frac{\partial X}{\partial \theta}}
\left(\begin{array}{cc} \frac{\partial Y}{\partial \theta} &-\frac{\partial Y}{\partial t}\\[1mm]
-\frac{\partial X}{\partial \theta} & \frac{\partial X}{\partial t}
\end{array}\right)
\left(\begin{array}{c} \frac{1}{\tau}e^{t/\tau} \\0 \end{array}\right)
\end{eqnarray*}
Since
\begin{eqnarray*}
0<\,e^{2t}(\sqrt{2}-\frac{1}{2})\, \leq \,\,\frac{\partial X}{\partial t}\frac{\partial Y}{\partial \theta}-\frac{\partial Y}{\partial t}\frac{\partial X}{\partial \theta} \, \,=\, e^{2t}(\sqrt{2}+\cos\theta \sin\theta ) \, \leq \,e^{2t}(\sqrt{2}+\frac{1}{2}),
\end{eqnarray*}
we obtain
\begin{eqnarray}\label{form-grad-f}
\left(\begin{array}{c}\frac{\partial f}{\partial x}\\[1mm]\frac{\partial f}{\partial y}\end{array}\right)
= e^{(\frac{1}{\tau} -1)t} P_{2,1}(t,\theta),
\end{eqnarray}
from which we infer that $\frac{\partial f}{\partial x}, \frac{\partial f}{\partial y}\to 0$ as $(x,y)\to \mathcal{O}$ or equivalently as $t\to -\infty$,
since $\frac{1}{\tau} >1$.
We then recover the fact that $f$ is $\mathcal{C}^1$, with $\nabla f(\mathcal{O})=(0,0)$.\smallskip
To prove that $f$ is $\mathcal{C}^2$ (when $\frac{1}{\tau} >2$), we differentiate again~\eqref{diff1-f} to obtain
\begin{eqnarray}\label{diff-f-2}
\left(\begin{array}{c}\frac{\partial^2\tilde{f}}{\partial t^2}\\[1mm]\frac{\partial^2\tilde{f}}{\partial t\partial\theta}\\[1mm]
\frac{\partial^2\tilde{f}}{\partial \theta^2} \end{array}\right)
= \left(\begin{array}{c} \frac{1}{\tau^2}e^{t/\tau} \\0\\0 \end{array}\right)
=
e^{2t} P_{3,3}(t,\theta)
\left(\begin{array}{c}\frac{\partial^2 f}{\partial x^2}\\[1mm]\frac{\partial^2 f}{\partial x\partial y}\\[1mm]
\frac{\partial^2 f}{\partial y^2}\end{array} \right)
+ e^t P_{3,2}(t,\theta)
\left(\begin{array}{c}\frac{\partial f}{\partial x}\\[1mm]\frac{\partial f}{\partial y}\end{array}\right),
\end{eqnarray}
where the coefficients of $e^{2t} P_{3,3}(t,\theta)$ are of the form
\begin{eqnarray*}
Z_1 Z_2, \quad \text{with } Z_1,Z_2\in \mathcal{D}_1:=\Big\{ \frac{\partial X}{\partial t}, \frac{\partial Y}{\partial t}, \frac{\partial X}{\partial \theta}, \frac{\partial Y}{\partial \theta}\Big\}
\end{eqnarray*}
and the coefficients of $e^t P_{3,2}(t,\theta)$ are second derivatives of $X$, $Y$.
The matrix $P_{3,3}(t,\theta)$ is invertible since $(t,\theta)\in\mathbb R\times \mathbb T\mapsto M(t,\theta):=(x,y)\in \mathbb R^2\setminus\{\mathcal{O}\}$
is an analytic diffeomorphism.
Finally, we get
\begin{eqnarray*
\left(\begin{array}{c}\frac{\partial^2 f}{\partial x^2}\\[1mm]\frac{\partial^2 f}{\partial x\partial y}\\[1mm]
\frac{\partial^2 f}{\partial y^2}\end{array} \right)
= e^{(\frac{1}{\tau}-2)t} P_{3,1}(t,\theta) + e^{(\frac{1}{\tau}-1)t} B_{3,1}(t,\theta),
\end{eqnarray*}
which proves that the second derivatives of $f$ converge to 0 as $(x,y)\to \mathcal{O}$
if $\frac{1}{\tau} >2$. Therefore $f$ is $\mathcal{C}^2$ with $\nabla^2 f(\mathcal{O})=0_{2\times 2}$. \smallskip
Continuing along the same lines, when differentiating $l$ times, the invertible
matrix in front of the $l$-th order derivatives of $f$ has coefficients of
the form $Z_1Z_2\cdots Z_l$ with $Z_1,\cdots , Z_l\in \mathcal{D}_1$ and,
after tedious computations, we obtain
\begin{eqnarray}\label{rel-fin-l}
\left(\begin{array}{c}\frac{\partial^l f}{\partial x^l}\\[1mm]\vdots\\[1mm]
\frac{\partial^l f}{\partial x^{l-i}\partial y^{i}}\\[1mm]\vdots\\[1mm]
\frac{\partial^l f}{\partial y^l}\end{array} \right)
= e^{(\frac{1}{\tau}-l)t} P_{l+1,1}(t,\theta) + e^{(\frac{1}{\tau}-(l-1))t} B_{l+1,1}(t,\theta),
\end{eqnarray}
which converges to $0$ as $(x,y)\to \mathcal{O}$ as long as $\frac{1}{\tau} >l$.
Therefore $f$ is $\mathcal{C}^l$ and all the $l$-th order derivatives of $f$ are
zero at $\mathcal{O}$ and we conclude that $f\in \mathcal{C}^k(\mathbb R^2)$, where $k$ is the biggest integer
such that $k< \frac{1}{\tau}$. \smallskip
Let us now assume, towards a contradiction, that $f$ is $\mathcal{C}^{k+1}$. Then we can write a Taylor expansion of $f$ up to the order $k+1$ at $\mathcal{O}$. Since $\nabla^l f(\mathcal{O})=0$ for $l\leq k$, we obtain that
\begin{eqnarray}\label{allure-f-0}
f(x,y)=O(||(x,y)||^{k+1})\quad \text{in a neighborhood of $\mathcal{O}$},
\end{eqnarray}
where $O(r^{k+1})/r^{k+1}$ is bounded near $0$. If $\frac{1}{\tau}\not\in\mathbb N$, then
$k+1 > \frac{1}{\tau}$, and we obtain a straightforward contradiction with the first inequality in~\eqref{ineg-f}.
If now $k+1 = \frac{1}{\tau}\in\mathbb N$, then~\eqref{allure-f-0} is not anymore contradictory
with~\eqref{ineg-f}. But writing~\eqref{rel-fin-l} with $l=k+1$, we get
\begin{eqnarray*}
\left(\begin{array}{c}\frac{\partial^{k+1} f}{\partial x^{k+1}}\\[1mm]\vdots\\[1mm]
\frac{\partial^{k+1} f}{\partial y^{k+1}}\end{array} \right)
= P_{k+2,1}(t,\theta) + e^{t} B_{k+2,1}(t,\theta).
\end{eqnarray*}
The second term above converges to zero as $t\to -\infty$, or equivalently as $(x,y)\to\mathcal{O}$,
but $P_{k+2,1}(t,\theta)$ is a periodic nonconstant matrix with respect to $t$ and $\theta$ so
cannot converge as $t\to -\infty$, contradicting our assumption. This ends the proof.
\end{proof}
\section{Oscillating gradient trajectories}
\label{sec:traj}
Let us start by showing that maximal gradient orbits blow up in finite positive time (and converge to the unique minimum $\mathcal{O}$ of the convex function $f$ as $t\to-\infty$).
\begin{lemma}
[Gradient trajectories of the convex function]\label{grad-traj}
Let $f$ be the convex function defined in Lemma~\ref{lem-fct-convexe}.
Then the ordinary differential equation for the gradient orbits
\begin{align}
\label{edo-gradient}\left\{
\begin{array}
[c]{ll}
\gamma^{\prime}(t) = \nabla f(\gamma(t) ), & t \in\mathbb{R},\medskip \\
\gamma(0)=\gamma_0 \in\mathbb{R}^{2}\setminus\{\mathcal{O}\}. &
\end{array}
\right.
\end{align}
admits a unique maximal solution $\gamma$ defined in $(-\infty, T)$ such that
$$
\underset{t\to-\infty}{\rm lim} \gamma(t)=\mathcal{O}
$$
and $\gamma$ blows up in a finite time $$T\leq \frac{2^{1/2\tau }}{(\frac{1}{\tau}-2)\, \Vert\gamma_0\Vert^{\frac{1}{\tau}-2}}\qquad (0< \tau <\frac{1}{10}\quad\text{is introduced
in~\eqref{cond-a-b-varphi-nu}}),$$ i.e.,
$$
\underset{t\nearrow T}{\rm lim}\,\Vert \gamma (t)\Vert =+\infty.
$$
\end{lemma}
\begin{proof}
Since $f$ is $\mathcal{C}^k$ with $k\geq 2$ (Lemma~\ref{lem-Ck-fct}),
there exists a unique maximal solution of~\eqref{edo-gradient}, denoted by $\gamma\in \mathcal{C}^k((S, T))$, where $-\infty\leq S<0<T\leq +\infty$. The function $f$ being convex and
coercive with a unique minimum at $\mathcal{O}$, we infer that $S=-\infty$
and $\gamma (t)\to \mathcal{O}$ as $t\to -\infty$.
In particular, $\gamma (t)\not= \mathcal{O}$ for every $t\in (-\infty, T)$
and consequently the function $t\mapsto z(t):=\Vert \gamma(t)\Vert$ is differentiable. Using the convexity of $f$ and~\eqref{ineg-f},
we deduce:
\begin{eqnarray*}
\frac{d}{dt}\Vert \gamma(t)\Vert =\langle \gamma'(t), \frac{\gamma(t)}{\Vert \gamma(t)\Vert}\rangle
= \langle \nabla f(\gamma(t)), \frac{\gamma(t)}{\Vert \gamma(t)\Vert}\rangle
\geq \frac{f(\gamma(t))}{\Vert \gamma(t)\Vert}\geq 2^{-\frac{1}{2\tau}} \Vert \gamma(t)\Vert^{\frac{1}{\tau}-1}.
\end{eqnarray*}
It follows that
\begin{eqnarray*}
\Vert \gamma(t)\Vert \geq \frac{1}{\left( \Vert \gamma_0\Vert^{2- \frac{1}{\tau}} - 2^{-\frac{1}{2\tau}}(\frac{1}{\tau}-2)t\right)^\frac{\tau }{1-2\tau}},
\end{eqnarray*}
where the above right-hand side is the exact solution to the scalar ordinary differential equation
$z'(t)= 2^{-\frac{1}{2\tau}} z(t)^{\frac{1}{\tau}-1}$, $z(0)=\Vert \gamma_0\Vert$.
We conclude that the maximal solution $\gamma$ blows up in finite positive time.
\end{proof}
In fact, finding gradient orbits is a geometric problem. We seek the unique curve $\gamma$ passing through $\gamma_0$,
which is orthogonal to the level sets of $f$. It is convenient to parametrize $\gamma$ as
\begin{align}
\label{param-s}\gamma(s)= M(t(s),\theta(s))= (X(t(s),\theta(s)),Y(t(s),\theta
(s))), \ s\in\mathbb{R}
\end{align}
using the notations~\eqref{Mmap}--\eqref{formMXY}. Under this parametrization $\gamma(s)\!\in\!\mathcal{E}(t(s))$, for every
$s\in\mathbb{R}$ and $\gamma^{\prime}(s)$ is a normal vector at $\gamma(s)$ to the (convex) sublevel set $[f\!\leq\!f(\gamma(s))]=\mathrm{conv }\,\mathcal{E}(t(s))$. Therefore:
\begin{align}
\label{cond-orth}\gamma^{\prime}(s) \perp\partial_{\theta}M(t(s),\theta(s)),
\quad\text{for all $s\in\mathbb{R}$}.
\end{align}
We define the rotation angle $s\mapsto \alpha(s)$ as the angle between the $x$-axis and the secant $\frac{\gamma
(s)}{\Vert \gamma(s)\Vert}$ (spherical part of the orbit) varying in a continuous way. Therefore
\begin{align*}
\left\{
\begin{array}
[c]{l}
\cos\alpha(s)=\frac{X(t,\theta)}{\sqrt{X(t ,\theta)^2+Y(t ,\theta)^2}}\,,\medskip\\
\sin\alpha(s) =\frac{Y(t ,\theta)}{\sqrt{X(t ,\theta)^2+Y(t ,\theta)^2}}.
\end{array}
\right.
\end{align*}
In particular, according to the notation used in \eqref{Mmap}--\eqref{ellEt}, if $\phi(s)$ is the angle in polar coordinates of the point $m(t,\theta)$, then we have (see Figure~\ref{dess_ellipse-rot}):
$$\alpha(s)=t(s)+\phi(s),\quad\text{for all }\, s\in\mathbb{R}.$$
\begin{figure}[ht]
\begin{center}
\includegraphics[width=5cm]{X2Y0.png}\hspace*{-0.5cm}
\includegraphics[width=5cm]{X2Y0zoom.png}\hspace*{-0.5cm}
\includegraphics[width=5cm]{X2Y0zoomplus.png}
\end{center}
\caption{\small Gradient orbit $\gamma(s)$ with initial point $\gamma(0)=(2,0)$, then zoom and extra-zoom.}
\label{dess-traj-osc}
\end{figure}
\begin{lemma}
[Spiraling around the origin]\label{lem-traj-oscill} Let $f$ be the convex function defined in~\eqref{def-f} under the assumption~\eqref{cond-a-b-varphi-nu} and let $s\mapsto \gamma(s)$ be a maximal orbit of the convex foliation
$(\mathcal{E}(t))_{t\in\mathbb{R}}$. Then the rotation angle
$s\mapsto\alpha(s)$ satisfies
\begin{align}\label{rotation-alpha}
\underset{s\to
\pm\infty}{\rm lim} \alpha(s)=\pm\infty.
\end{align}
\end{lemma}
\noindent See Figure~\ref{dess-traj-osc} for a generic numerical simulation of the maximal orbit of the function $f$
associated with the convex foliation of Figure~\ref{dess:feuilletage-ellipses}.
\begin{proof}
We use the parametrization given by~\eqref{param-s}. Since
$$ \underset{s\to +\infty}{\rm lim} \Vert\gamma (s)\Vert = +\infty\quad\text{and} \quad \underset{t\to -\infty} {\rm lim}\gamma (s)= \mathcal{O},$$
we can assume that the function $s\mapsto t(s)$ satisfies
\begin{eqnarray}\label{cond-tprime}
t'(s)>0 \quad \text{and}\quad \underset{s \to \pm \infty}{\rm lim} t(s) = \pm \infty.
\end{eqnarray}
The goal is to compute $\alpha(s)$ using the orthogonality
condition~\eqref{cond-orth}, which is equivalent to
\begin{eqnarray}\label{cond-ps}
\big\langle \,\gamma' (s),\, \partial_\theta M\!\left(t(s),\theta (s)\right) \big\rangle =0, \quad \text{for all $s\in\mathbb{R}$}.
\end{eqnarray}
Using the notations of Section~\ref{sec:foliation}, we have
\begin{eqnarray*}
\gamma'(s)=\frac{d}{ds} M(t(s),\theta (s))
= t' \partial_t (Rm) + \theta' \partial_{\theta} (Rm)
= t' (R' m + R \partial_t m)+ \theta' R\partial_{\theta} m
\end{eqnarray*}
and $\partial_\theta M= \partial_\theta (Rm)= R \partial_\theta m$.
It follows
\begin{eqnarray*}
\big\langle \gamma' (s), \partial_\theta M \big\rangle
&=& t' \big\langle R' m, R \partial_\theta m \big\rangle
+ t' \big\langle R \partial_t m, R \partial_\theta m \big\rangle
+ \theta' \big\langle R \partial_\theta m, R \partial_\theta m \big\rangle\\
&=&
t' \big\langle R(\frac{\pi}{2}) m, \partial_\theta m \big\rangle
+ t' \big\langle \partial_t m, \partial_\theta m \big\rangle
+ \theta' \Vert\partial_\theta m\Vert^2\\
&=&
t'\left(ab +(bb'-aa') \cos\theta \sin\theta\right)+\theta'\left(a^2\sin^2\theta + b^2\cos^2\theta\right).
\end{eqnarray*}
By~\eqref{cond-orth}, we have $\big\langle\gamma^{\prime}(s),\partial_{\theta}M\big\rangle=0$
and after substitution $a(t)=\sqrt{2}e^t$ and $b(t)=e^t$ we get
$$
t' e^{2t}(\sqrt{2}- \cos\theta \sin\theta) +\theta' e^{2t}(1+\sin^2\theta )=0
$$
whence we deduce the following relation between $t(s)$ and $\theta(s)$:
\begin{eqnarray}\label{ttheta}
t'(s)= - \frac{1+\sin^2\theta(s)}{\sqrt{2}- \cos\theta(s) \sin\theta(s)}\theta'(s).
\end{eqnarray}
Since for every $\theta\in \mathbb{R}$ we have
\begin{eqnarray*}
0\,<\, \frac{1}{\sqrt{2}+\frac{1}{2}}\,\leq \, \frac{1+\sin^2\theta}{\sqrt{2}- \cos\theta \sin\theta}\,\leq \,\frac{2}{\sqrt{2}-\frac{1}{2}}\,,
\end{eqnarray*}
we get
\begin{eqnarray*}
-\frac{1}{\sqrt{2}+\frac{1}{2}}\,\theta'(s)\,\leq \, t'(s)\,\leq \,- \frac{2}{\sqrt{2}-\frac{1}{2}}\,\theta'(s).
\end{eqnarray*}
Therefore, from~\eqref{cond-tprime} we deduce
\begin{eqnarray}\label{cond-thetaprime}
\theta'(s)<0, \quad \theta(s)\mathop{\to}_{s\to -\infty}+\infty, \quad \theta(s)\mathop{\to}_{s\to +\infty}-\infty.
\end{eqnarray}
Next, we establish the relation between $\theta(s)$ and $\phi(s)$, see Figure~\ref{dess_ellipse-rot}.
We have
\begin{eqnarray*}
&& \cos\phi =\frac{a\cos\theta}{\sqrt{a^2\cos^2\theta + b^2\sin^2\theta}}
= \frac{\sqrt{2}\cos\theta}{\sqrt{2\cos^2\theta + \sin^2\theta}},\\
&& \sin\phi =\frac{b\sin\theta}{\sqrt{a^2\cos^2\theta + b^2\sin^2\theta}}
= \frac{\sin\theta}{\sqrt{2\cos^2\theta + \sin^2\theta}}.
\end{eqnarray*}
Differentiating $\cos\phi$ and plugging the result in the second expression,
we end up with
\begin{eqnarray}\label{phitheta}
\phi' =\frac{\sqrt{2}}{1+ \cos^2\theta} \theta'.
\end{eqnarray}
Assembling~\eqref{ttheta} and~\eqref{phitheta}, we obtain
\begin{eqnarray}\label{expr-alpha-prime}
\alpha'=t'+\phi'=\left(\frac{\sqrt{2}}{1+\cos^2\theta}-\frac{1+\sin^2\theta}{\sqrt{2}-\cos\theta\sin\theta}\right)\theta'=: h(\theta) \theta'.
\end{eqnarray}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=12cm]{dess_fct_derivee-alpha.pdf}
\end{center}
\caption{Plot of $\displaystyle h(\theta)= \frac{\sqrt{2}}{1+\cos^2\theta}-\frac{1+\sin^2\theta}{\sqrt{2}-\cos\theta\sin\theta}$.}
\label{dess:fonction-alpha-prime}
\end{figure}
The function $h$ is analytic and $2\pi$-periodic, see Figure~\ref{dess:fonction-alpha-prime}. We can expand
it in Fourier series and integrate~\eqref{expr-alpha-prime} to obtain
\begin{eqnarray}\label{val-alpha}
\alpha (s)= \frac{a_0}{2}\theta (s) + O(1),
\end{eqnarray}
where $O(1)$ is a bounded function and
\begin{eqnarray*}
a_0=\frac{1}{\pi}\int_0^{2\pi}h(\theta)d\theta \simeq -0.84 < 0.
\end{eqnarray*}
We finally conclude from~\eqref{val-alpha} and~\eqref{cond-tprime} that~\eqref{rotation-alpha}
holds.
\end{proof}
\section{Proof of Theorem~\ref{thm-main}}
\label{sec:proof-thm}
Consider the convex foliation by ellipses $\{\mathcal{E}(t)\}_{t\in\mathbb{R}}$ given by Lemma~\ref{lem-feuilletage}.
Let $k\geq 1$ be any integer and $f$ be the convex function defined by Lemma~\ref{lem-fct-convexe}
for $0<\tau < \min\{ 1/10, 1/k \}$. Then, by Lemma~\ref{lem-regularite-fct}, the function $f$ is coercive, has
its unique minimum at the origin $\mathcal{O}$, is real analytic in $\mathbb{R}^2\setminus\{\mathcal{O}\}$
and satisfies the {\L}ojasiewicz inequality~\eqref{eq:1}.
Further, Lemma~\ref{lem-Ck-fct}, ensures that $f$ is $\mathcal{C}^k$-smooth.
Finally, Lemma~\ref{lem-traj-oscill} asserts that all nontrivial gradient orbits spiral infinitely many times both near the origin (bounded part) and at infinity.\hfill$\square$
\vspace{0.5cm}
\noindent\textbf{Acknowledgement.} This work was partially supported by the
Centre Henri Lebesgue ANR-11-LABX-0020-01 and the grants CMM AFB170001,
ECOS-Sud/ANID C18E04 and FONDECYT 1211217. Major part of this work has been
done during a research visit of the first author to INSA Rennes. This author
is indebted to his hosts for hospitality.
|
\section{Introduction}\label{sec:intro}
\section{Definition of von Mises-Fisher Density}
The von Mises-Fisher (vMF for short) distribution, defined over the unit hypersphere in $\mathbb{R}^d$, has the following density function:
\begin{equation}
p({\bf x}; {\bm \mu}, \kappa) = C_d(\kappa) \cdot \exp(\kappa {\bm \mu}^\top {\bf x}), \ \ \ \ C_d(\kappa) = \frac{\kappa^{d/2-1}}{(2\pi)^{d/2} I_{d/2-1}(\kappa)},
\end{equation}
where $I_\alpha$ is the modified Bessel function of the first kind with order $\alpha$.
Here, $\kappa$ (scalar) and ${\bm \mu}\in\mathbb{R}^d$ constitute the parameters of the vMF density, and they are constrained as: $\kappa \geq 0$ and $||{\bm \mu}|| = 1$.
Obviously, the density has a single mode at ${\bf x} = {\bm \mu}$, and any hyperplane normal to ${\bm \mu}$ forms a level set, that is, the likelihood remains unchanged for $\{ {\bf x}: {\bm \mu}^\top {\bf x} = \gamma \}$ with $|\gamma| \leq 1$. Due to the symmetry of the density along the vector ${\bm \mu}$, the mean of the density is also ${\bm \mu}$. And, $\kappa$ determines how peaky the distribution is around its mode/mean ${\bm \mu}$, thus it is often called the {\em concentration} parameter (i.e., large $\kappa$ means that the density is more concentrated around ${\bm \mu}$, and vice versa).
Note also that the normalizer $C_d(\kappa)$ depends only on $\kappa$, not ${\bm \mu}$, due to the symmetry of the density function.
\section{MLE for vMF}
We derive maximum likelihood estimators (MLEs) for the vMF density function, given the training data $\{ {\bf x}_i \}_{i=1}^N$. It is known that the MLE for vMF admits a {\em near} closed-form formula, where by {\em near} one has to resort to some approximation due to the difficulty of inverting the Bessel ratio function. We will derive this in Sec.~\ref{sec:mle_vmf_batch} below by following the approximation schemes proposed in~\citep{banerjee05}, and this estimator naturally forms a {\em full-batch learner} in that we need to go through the entire data to have a model update. Alternatively we can perform the gradient ascent of the log-likelihood objective function, which is amenable to stochastic online (mini-batch) learning as discussed in Sec.~\ref{sec:mle_vmf_sgd}.
\subsection{Full-Batch Learning (Near Closed Form)}\label{sec:mle_vmf_batch}
The maximum likelihood learning can be written as the following optimization problem:
\begin{align}
&\max_{{\bm \mu}, \kappa} \ \mathcal{L}({\bm \mu}, \kappa) \ := \ \sum_{i=1}^N \log p({\bf x}_i) = N\kappa {\bm \mu}^\top \overline{{\bf x}} + N \log C_d(\kappa) \\
& \ \ \textrm{s.t.} \ \ ||\mu|| = 1 \ \ \textrm{and} \ \ \kappa \geq 0,
\end{align}
where $\overline{{\bf x}} = \frac{1}{N} \sum_{i=1}^N {\bf x}_i$.
By taking the derivatives of the objective wrt parameters and setting them to $0$, it is easy to see that the optimum $({\bm \mu}^*, \kappa^*)$ should meet:
\begin{align}
{\bm\mu}^* = \frac{\overline{{\bf x}}}{||\overline{{\bf x}}||} \ \ \ \ \textrm{and} \ \ \ \
\frac{C'_d(\kappa^*)}{C_d(\kappa^*)} = -||\overline{{\bf x}}||, \label{eq:vmf_mle_eqs}
\end{align}
where
$f'(a) = \frac{d f(a)}{d a}$.
For the second equation in (\ref{eq:vmf_mle_eqs}), by letting $\xi = (2\pi)^{d/2}$ and $s=d/2-1$, we have:
\begin{equation}
C_d(\kappa) = \frac{\kappa^s}{\xi \cdot I_s(\kappa)} \ \ \ \ \textrm{and} \ \ \ \ C'_d(\kappa) = \frac{1}{\xi} \cdot \frac{s \kappa^{s-1} I_s(\kappa) - \kappa^s I'_s(\kappa)}{I_s(\kappa)^2}.
\end{equation}
Then it follows that
\begin{equation}
\frac{C'_d(\kappa)}{C_d(\kappa)} = \frac{s}{\kappa} - \frac{I'_s(\kappa)}{I_s(\kappa)}.
\end{equation}
Using the recursive formulas for the Bessel function, more specifically,
\begin{equation}
\frac{I'_s(\kappa)}{I_s(\kappa)} = \frac{s}{\kappa} + \frac{I'_{s+1}(\kappa)}{I_s(\kappa)},
\end{equation}
the second condition in (\ref{eq:vmf_mle_eqs}) boils down to the following equation (for $\kappa$):
\begin{equation}
\frac{I_{d/2}(\kappa)}{I_{d/2-1}(\kappa)} = ||\overline{{\bf x}}||.
\label{eq:bessel_ratio_eq}
\end{equation}
Unfortunately, there is no known closed-form solution to the Bessel ratio inversion problem (\ref{eq:bessel_ratio_eq}). While there have been several approximate solutions to (\ref{eq:bessel_ratio_eq})~\citep{vmf_book,tanabe2007,sra2012}, here we focus on the approximation schemes based on the continued fraction form of the Bessel ratio function, namely
\begin{equation}
R := \frac{I_{d/2}(\kappa)}{I_{d/2-1}(\kappa)} = \frac{1}{ \frac{d}{\kappa} + \frac{1}{ \frac{d+2}{\kappa} + \cdots} } \approx
\frac{1}{\frac{d}{\kappa} + R}.
\label{eq:cont_frac}
\end{equation}
With $R = ||\overline{{\bf x}}||$, the approximation in (\ref{eq:cont_frac}) leads to $\kappa^* \approx \frac{d\cdot ||\overline{{\bf x}}||}{1-||\overline{{\bf x}}||^2}$. To reduce the approximation error further, in~\citep{banerjee05}, a correction term is added to the numerator,
\begin{equation}
\kappa^* \approx \frac{d\cdot ||\overline{{\bf x}}|| - ||\overline{{\bf x}}||^3 } {1-||\overline{{\bf x}}||^2}.
\label{eq:kappa_mle}
\end{equation}
\subsection{Stochastic Gradient Descent (SGD) Learning}\label{sec:mle_vmf_sgd}
To circumvent the inverse Bessel ratio approximation used in the previous full-batch learning, we can consider (stochastic) gradient descent of the negative log-likelihood loss function.
Assuming that we can sample a mini-batch $B$ ($\ni {\bf x}$) from the training data, the negative expected log-likelihood objective on $B$ can be written as:
\begin{equation}
NELL_B = \kappa {\bm \mu}^\top \overline{{\bf x}}_B + \log C_d(\kappa),
\end{equation}
where $\overline{{\bf x}}_B = \frac{1}{|B|} \sum_{{\bf x}\in B} {\bf x}$.
Note that the gradients of the objective can be derived as:
\begin{equation}
\frac{\partial NELL_B}{\partial {\bm \mu}} = \kappa \cdot \overline{{\bf x}}_B, \ \ \ \
\frac{\partial NELL_B}{\partial \kappa} = {\bm \mu}^\top \overline{{\bf x}}_B - \frac{I_{s+1}(\kappa)}{I_s(\kappa)} \ \ \ \ (s = d/2-1).
\end{equation}
And, to evaluate the objective itself, one needs to compute $\log I_s(\kappa)$ since
\begin{equation}
\log C_d(\kappa) = s \log \kappa - \frac{d}{2} \log 2\pi - \log I_s(\kappa).
\end{equation}
Hence, to perform the SGD training, the key quantities that we need to compute are: $\log I_s(\kappa)$ and $\frac{I_{s+1}(\kappa)}{I_s(\kappa)}$. To be more specific, in the PyTorch implementation with auto-differentiation capability, the former will be used in the \texttt{forward()} method definition of \texttt{autograd.Function}, while the latter placed in the \texttt{backward()} method.
\begin{figure
\begin{center}
\includegraphics[trim = 18mm 0mm 0mm 0mm, clip, scale=0.625
]{figs/besseli_scipy_vs_mpmath.pdf}
\end{center}
\vspace{-1.0em}
\caption{Comparison between \texttt{scipy} and \texttt{mpmath} for the Bessel function evaluation on small inputs. For the smaller input $x=0.03$, \texttt{scipy} suffers from numerical underflow.
}
\vspace{-1.0em}
\label{fig:scipy_mpmath}
\end{figure}
Although one can use the \texttt{scipy.special.iv()} for the Bessel function of the first kind, it can easily incur numerical issues when the dimesnion $d$ becomes large (esp., when $d$ is large {\em and} $\kappa$ is close to $0$)\footnote{In the S-VAE~\citep{s-vae18}, the deep latent variable model that postulates a unit hyperspherical latent space, for instance, they introduced a vMF variational posterior, and used the \texttt{scipy.special.iv()} for the Bessel. The consequence is that the results become unstable when the dimension is greater than 50.}.
To address the problem, we rather adopt \texttt{mpmath}~\citep{mpmath}, the numerical Python library that supports arbitrary precision.
The code snippet in Fig.~\ref{fig:scipy_mpmath} illustrates the basic usage of \texttt{mpmath} (e.g., $I_\alpha(x)$ can be computed by calling \texttt{mpmath.besseli($\alpha$,x)}), while we also contrast the numerical stability between \texttt{scipy} and \texttt{mpmath} for high-order and near-$0$ input cases.
Here is another useful tip. For the computational efficiency, the block operation might be required. That is, we need to apply \texttt{mpmath} functions to all elements in a tensor/array. For the \texttt{numpy}, this can be done by vectorizing the \texttt{mpmath} function using \texttt{numpy.vectorize()}. For instance, the following code snippet computes $I_{100}(x)$ for all elements $x$ in a $(3 \times 4)$ \texttt{numpy} array.
\begin{lstlisting}
>>> arr = numpy.random.rand(3,4)*mpmath.mpf(1.0)
>>> besseli = numpy.vectorize(mpmath.besseli)
>>> besseli(100, arr)
\end{lstlisting}
A clear advantage of defining the \texttt{autograd.Function} for the gradients of the log-likelihood of the vMF density wrt the vMF parameters, is that (using the chain rule) we can easily optimize a model that has tensors/parameters complexly related to the vMF parameters. For instance, we may define the vMF parameters to be the outputs of some neural networks, say ${\bm\mu}_{{\bm\nu}}({\bf y})$ and/or $\kappa_{{\bm\lambda}}({\bf y})$, with weight parameters ${\bm\nu}$ and ${\bm\lambda}$ on the input ${\bf y}$.
\subsection{Empirical Comparison between Full-Batch Learning and SGD}\label{sec:batch_vs_sgd}
For this empirical comparison, we choose a vMF density $p({\bf x}; {\bm\mu}^*, \kappa^*)$, and sample iid data $\{{\bf x}_i\}_{i=1}^N$ from it. Then the true parameters $({\bm\mu}^*, \kappa^*)$ are estimated using the two estimators discussed above, with the sampled data. To sample from a vMF, we follow the scheme in~\citep{vmf_samp}. To be concrete, sampling a $d$-dim ${\bf x} \sim p({\bf x}; {\bm\mu}, \kappa)$ can be done by the following steps:
\begin{enumerate}
\item Sample ${\bf v}$, a unit vector in $\mathbb{R}^{d-1}$ uniformly. (It can be done by first sampling $d-1$ iid $\mathcal{N}(0,1)$ numbers, forming a vector, then unit-normalizing it.)
\item Sample $v_0 \sim p(v_0) \propto e^{\kappa v_0} (1-v_0^2)^{0.5(d-3)}$, e.g., by acceptance-rejection sampling.
\item Set ${\bf v} := [ v_0; (1-v_0^2)^{\frac{1}{2}}{\bf v} ]$.
\item Compute ${\bf U} = {\bf I} - 2 \frac{ ({\bf e}_1-{\bm\mu}) ({\bf e}_1-{\bm\mu})^\top } {||{\bf e}_1-{\bm\mu}||^2}$.
\item Return ${\bf x} := {\bf U}{\bf v}$.
\end{enumerate}
The estimation errors are shown in Table~\ref{tab:mle_for_vmf}. We report the relative errors of the estimated parameters, that is, $e_{{\bm\mu}} = \frac{||{\bm\mu} - {\bm\mu}^*||} {||{\bm\mu}^*||} = ||{\bm\mu} - {\bm\mu}^*||$ for the mean parameters, and $e_{\kappa} = \frac{|\kappa-\kappa^*|}{\kappa^*}$ for the concentration parameter, where $({\bm\mu}^*,\kappa^*)$ is the parameters of the sampling distribution. To test the estimators on diverse scenarios, we vary the dimensionality $d$ from $\{5,20,100\}$ while $\kappa^*$ is chosen to be either 50 (low concentration) or 500 (high certainty).
We fix ${\bm\mu}^* = {\bf e}_1$, the unit vector on the first axis.
We generate $N=10,000$ samples. For the SGD optimization, we use the Adam optimizer~\citep{adam} where the learning rate is 0.01 which decays exponentially by the rate 0.95. The batch size 128, and we run it for 100 epochs.
\begin{table
\centering
\caption{Comparison of the full-batch and SGD estimators with vMF-sampled data.
}
\label{tab:mle_for_vmf}
\begin{small}
\begin{sc}
\centering
\begin{tabular}{c||c|c||c|c}
\toprule
\multirow{2}{*}{dim ($d$)} & \multicolumn{2}{c||}{$\kappa^*=50$} & \multicolumn{2}{c}{$\kappa^*=500$}
\\ \cline{2-5}
\Tstrut & Batch & SGD & Batch & SGD \\
\hline\hline
\Tstrut \multirow{2}{*}{$d=5$} & $e_{\bm\mu}=3.8 \times 10^{-6}$ & $e_{\bm\mu}=3.8 \times 10^{-6}$ & $e_{\bm\mu}=8.1 \times 10^{-7}$ & $e_{\bm\mu}=8.4 \times 10^{-7}$ \\
& $e_\kappa = 1.0 \times 10^{-2}$ & $e_\kappa = 6.2 \times 10^{-4}$ & $e_\kappa = 4.6 \times 10^{-3}$ & $e_\kappa = 3.7 \times 10^{-3}$
\\
\hline
\Tstrut \multirow{2}{*}{$d=20$} & $e_{\bm\mu}=7.4 \times 10^{-5}$ & $e_{\bm\mu}=7.4 \times 10^{-5}$ & $e_{\bm\mu}=3.4 \times 10^{-6}$ & $e_{\bm\mu}=3.3 \times 10^{-6}$ \\
& $e_\kappa = 5.4 \times 10^{-3}$ & $e_\kappa = 1.4 \times 10^{-3}$ & $e_\kappa = 2.8 \times 10^{-3}$ & $e_\kappa = 1.7 \times 10^{-3}$
\\
\hline
\Tstrut \multirow{2}{*}{$d=100$} & $e_{\bm\mu}=4.8 \times 10^{-4}$ & $e_{\bm\mu}=4.8 \times 10^{-4}$ & $e_{\bm\mu}=2.1 \times 10^{-5}$ & $e_{\bm\mu}=2.1 \times 10^{-5}$ \\
& $e_\kappa = 7.0 \times 10^{-3}$ & $e_\kappa = 5.9 \times 10^{-3}$ & $e_\kappa = 1.9 \times 10^{-3}$ & $e_\kappa = 1.1 \times 10^{-3}$
\\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{table}
The conclusion is as follows. The two estimators perform equally well. For the mean parameter estimation, both estimators perform nearly identically. But for the concentration parameter ($\kappa$) estimation, there are slight differences: SGD is consistently more accurate than the full-batch estimator. This may originate from the batch estimator's continued fraction approximation error. Overall, the estimation error for $\kappa$ becomes larger as the dimension $d$ increases, and as the data variability (or entropy) increases (i.e., small $\kappa^*$).
The code to reproduce the results in this section can be found in:~\url{https://github.com/minyoungkim21/vmf-lib/mle_for_vmf.py}.
\section{MLE for Mixture of vMFs}
We consider a mixture of vMF densities. A mixture with the order (the number of components) $M$ can be written down as follows:
\begin{equation}
p({\bf x}; {\bm\theta}) = \sum_{m=1}^M \alpha_m p({\bf x}; {\bm\mu}_m, \kappa_m),
\end{equation}
where $\alpha_m$'s are the mixing proportions. The parameters of the mixture are denoted by ${\bm\theta} = \{\alpha_m, {\bm\mu}_m, \kappa_m\}_{m=1}^M$.
The EM algorithm~\citep{em77} is recognized as the most popular generic algorithm for mixture estimation. However, when the size of the training data is large, each E-step can be computationally demanding since one has to pass through the entire data. It might take a long time until the model is updated by a single EM iteration.
There are broadly two workarounds to deal with large-scale training data. The first approach is to perform the stochastic gradient ascent on the log-marginal likelihood (i.e., taking the gradient step $\nabla_{{\bm\theta}} \frac{1}{|B|} \sum_{{\bf x}\in B} \log p({\bf x}; {\bm \theta})$ with the minibatch $B$). The second approach is the recent {\em stochastic EM}~\citep{sem,sem_var} that aims to solve a stochastic noisy version of the fixed point estimation through Robbins-Monro stochastic approximation method~\citep{robbins_monro}. Although deriving the stochastic EM for a vMF mixture can be done straightforwardly, in this section we instead dig out the former (direct stochastic gradient ascent), and empirically compare it with the full-batch EM learning.
\subsection{Derivation for EM }\label{sec:mixture_em}
The EM algorithm is essentially a block coordinate optimization method for the lower bound of the data log-likelihood, where the lower bound is obtained by applying the Jensen inequality to the log function. More specifically, letting $z \in \{1,\dots,M\}$ be the (hidden) component membership indicator, and $p_m({\bf x}) = p({\bf x}|z=m)$ be the component conditional distribution, it alternates the following two steps until convergence:
\begin{itemize}
\item \textbf{E-step}: With ${\bm \theta}$ fixed, evaluate: $q_i(z) = p(z|{\bf x}_i; {\bm \theta})$ for $i=1,\dots,N$.
\item \textbf{M-step}: With $q$ fixed, solve: $\max_{{\bm\theta}} \mathcal{L}({\bm\theta}) := \sum_{i=1}^N \sum_{m=1}^M q_i(z=m)
\log \big( \alpha_m p_m({\bf x}_i) \big)$.
\end{itemize}
The E-step can be done easily by
\begin{equation}
q_i(z=m) = p(z=m|{\bf x}_i)=\frac{\alpha_m p_m({\bf x}_i)}{\sum_m \alpha_m p_m({\bf x}_i)}.
\end{equation}
Note that $p_m({\bf x}) = \textrm{vMF}({\bf x}; {\bm\mu}_m, \kappa_m)$ only requires the evaluation of the normalizer $C_d(\kappa_m)$ (and hence evaluation of the Bessel function).
The M-step, with $q_i$'s fixed, can be derived for each parameter as follows, admitting (near) closed forms:
\begin{itemize}
\item ${\bm\mu}_m$: We set
$\frac{\partial \mathcal{L}}{\partial {\bm\mu}_m} =
\sum_{i=1}^N q_i(m) \kappa_m {\bf x}_i$
to $0$, and solve it with the unit-norm constraint:
\begin{equation}
{\bm\mu}_m^* = \frac{\sum_{i=1}^N q_i(m) {\bf x}_i}{||\sum_{i=1}^N q_i(m) {\bf x}_i||}.
\end{equation}
\item $\kappa_m$: Setting the derivative $\sum_{i=1}^N q_i(m) \Big( {\bm \mu}_m^\top {\bf x}_i - \frac{I_{d/2}(\kappa_m)}{I_{d/2-1}(\kappa_m)} \Big)$ to $0$, leads to:
\begin{equation}
\frac{I_{d/2}(\kappa_m^*)}{I_{d/2-1}(\kappa_m^*)} = \frac{\sum_{i=1}^N q_i(m) {\bf x}_i^\top {\bm \mu}_m^* }{\sum_{i=1}^N q_i(m)} = \frac{||\sum_i q_i(m) {\bf x}_i||}{\sum_{i=1}^N q_i(m)} \ (=: R_m).
\label{eq:em_kappa_m}
\end{equation}
To solve (\ref{eq:em_kappa_m}) for $\kappa_m$, we have to resort to an approximation scheme, and we employ the continued fraction method with the correction term, as described in Sec.~\ref{sec:mle_vmf_batch}. That is,
\begin{equation}
\kappa_m^* \approx \frac{d\cdot R_m - R_m^3 } {1-R_m^2}.
\end{equation}
\item $\alpha_m$: The maximum is attained by the empirical mean of the posteriors for the membership $m$. That is,
\begin{equation}
\alpha_m^* = \frac{\sum_{i=1}^N q_i(m)} {N}.
\end{equation}
\end{itemize}
\subsection{Derivation for Direct SGD}\label{sec:mixture_sgd}
For the large training data, we can apply the SGD by directly taking the gradient of the data log-likelihood objective over a small minibath $B$ ($\ni {\bf x}$), namely $\mathcal{L}_B := \nabla_{{\bm\theta}} \frac{1}{|B|} \sum_{{\bf x}\in B} \log p({\bf x}; {\bm \theta})$. The gradient (at a single instance ${\bf x}$) can be derived as follows:
\begin{equation}
\nabla_{{\bm\theta}} \log p({\bf x}; {\bm\theta}) =
\sum_{m=1}^M p(z=m|{\bf x}) \big( \nabla_{{\bm\theta}} \log \alpha_m + \nabla_{{\bm\theta}} \log p_m({\bf x}) \big).
\end{equation}
In particular, the gradients for individual parameters can be derived as:
\begin{align}
\frac{\partial \log p({\bf x}; {\bm\theta})}{\partial {\bm\mu}_m} & \ = \
p(z=m|{\bf x}) \kappa_m {\bf x} \\
\frac{\partial \log p({\bf x}; {\bm\theta})}{\partial \kappa_m} & \ = \
p(z=m|{\bf x}) \bigg( {\bm \mu}_m^\top {\bf x} -
\frac{I_{d/2}(\kappa_m)}{I_{d/2-1}(\kappa_m)} \bigg) \\
\frac{\partial \log p({\bf x}; {\bm\theta})}{\partial \alpha_m} & \ = \
p(z=m|{\bf x}) \frac{1}{\alpha_m}.
\end{align}
However, with the auto-differentiation feature provided in the PyTorch, one can bypass the implementation of the above steps: we simply form a computation graph of $\log p({\bf x}) = \log \sum_m \alpha_m p_m({\bf x})$ (e.g., using the PyTorch's \texttt{logsumexp()} function), and utilize the backprop.
\subsection{Empirical Study on Synthetic Data Clustering}\label{sec:mixture_synth}
To test the EM and SGD mixture learning algorithms, we choose a mixture model and sample iid data from it. The mixture order is set to 3 and the data dimension 5. The true parameters are: $\alpha_1 = 0.3$, $\alpha_2=0.4$, $\alpha_3=0.3$, ${\bm\mu}_1 = [0.0889, -0.3556, 0.6815, 0.1185, 0.6222]$, ${\bm\mu}_2 = [ 1.0000, 0.0000, 0.0000, 0.0000, 0.0000]$, ${\bm\mu}_3 = [-0.0889, 0.3556, -0.6815, -0.1185, -0.6222]$, and $\kappa_1=100.0$, $\kappa_2=50.0$, $\kappa_3=100.0$. We sample 1000 data points from the model.
For the EM algorithm we choose the maximum number of iterations 100 (but converged very quickly after a few iterations), and the tolerance of the relative log-likelihood improvement in iterations is $10^{-5}$ which serves as a stopping criterion. For the SGD learning, the batch size is set to 64, and the training goes until 100 epochs with learning rate 0.1 that decays with the rate of 0.95 every epoch. We compute the errors of the learned models, defined to be the absolute differences between true parameters and the learned parameters. To deal with the invariance of mixture components permutation in mixture models,
we consider all permutations of the mixture components of the learned models, and take the one with the smallest error. The results are summarized in Fig.~\ref{fig:mixture_em_vs_sgd}. As shown, the two learning methods work equally well, while the SGD performs slightly better (but not significantly) in terms of the L1 error in the parameter space.
\begin{figure
\begin{center}
\fbox{\includegraphics[trim = 16mm 13mm 52mm 0mm, clip, scale=0.755
]{figs/mix_vmf_results.pdf}}
\end{center}
\vspace{-0.5em}
\caption{Comparison between the (full-batch) EM and SGD learning for mixture estimation.
}
\vspace{-1.0em}
\label{fig:mixture_em_vs_sgd}
\end{figure}
The code to reproduce the results in this section can be found in:~\url{https://github.com/minyoungkim21/vmf-lib/mle_for_mix_vmf.py}.
\section{Application: Image Clustering in Embedded Unit-Hypersphere Space}\label{sec:apps}
In this section we test the vMF mixture model on the image clustering problem.
In particular, we deal with image clustering with the CIFAR-10 dataset. The idea is that we first learn the useful embeddings/features ${\bf z} = e({\bf x})$ for image ${\bf x}$, which can be done by minimizing the reconstruction error in the auto-encoding process. We impose the constraint $||{\bf z}|| = 1$ by placing a normalization layer at the end of the encoder pipeline. Once the features ${\bf z}$'s are learned, we apply the EM or SGD algorithm with the fixed ${\bf z}$ data. We choose $\dim({\bf z})=100$ for the feature dimension, and the encoder/decoder networks are built from conv/deconv layers.
As a baseline we also compare the two algorithms with the simple K-means~\citep{kmeans} on the fixed ${\bf z}$ data. For the clustering performance metrics, we use the popular: Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI), where the higher the better for both metrics.
The results are shown in Table~\ref{tab:img_clustering}. The EM and SGD attain better performance than k-means, while both perform equally well.
The code to reproduce the results in this section can be found in:~\url{https://github.com/minyoungkim21/vmf-lib/image_clustering_cifar10.py}.
\begin{table
\centering
\caption{Image clustering performance on the CIFAR-10 dataset.
}
\label{tab:img_clustering}
\vspace{-0.5em}
\begin{small}
\begin{sc}
\centering
\begin{tabular}{c||c|c}
\toprule
\Tstrut Methods & ARI & NMI \\ \hline\hline
\Tstrut k-means & 0.0487 & 0.0871 \\ \hline
\Tstrut EM & 0.0522 & 0.1068 \\ \hline
\Tstrut SGD & 0.0528 & 0.1067 \\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{table}
\small
{
\bibliographystyle{spbasic}
|
\section{Introduction}
Corrupted or missing data often raise the following questions:
If the missing values are imputed, what value is to be used for "padding"?
What is the underlying relation with the normalization scheme of the input values?
How should categorical inputs be treated?
Popular approaches include replacing the missing values with their mean, in the case of continuous variables, or the median, in the case of discrete variables.
Other approaches introduce a binary indicator of missingness that is jointly trained with the rest of the model parameters.
The simplest, although widely used, method is \emph{mean/mode impute}, where missing numerical values are replaced by their mean (0 in the Zeta Normalized Scale) and categorical values are replaced by their mode (the most frequent class of the dataset).
One disadvantage of mean/mode impute is that it can introduce ambiguity: variables that take values close to the mean or mode, might not be distinguished from variables whose missing values were imputed with the value of the mean or median.
To resolve this ambiguity, a bitwise indicator (also referred to as \emph{mask}) was introduced, that assigns 1 to missing instances and 0 to non-missing instances.
For simplicity, we refer to this method as Zero Impute (with) Mask Concat (\textbf{ZIMC}).
A straightforward extension to ZIMC is Sparsity Normalization (\textbf{SN}),
which creates a per-example (horizontal) scaling by a constant factor \cite{vsp},
based on the global missingness factor.
This method resolves the Variable Sparsity Problem, but fails to capture the dependencies in-between the features.
In more detail, no adjustment is made based on what feature is missing and the values the existing features have.
Since the problem of missing data naturally defines a context of missing information we are inspired from feature-wise transformations \cite{fwt} as methods of context-based processing. More specifically the Feature Wise Linear Modulation (\textbf{FiLM}), naturally augments the idea of SN \cite{vsp} (\textbf{SN}) since it can be seen as a conditional scaling method.
In order to efficiently capture the missing context we combine both the feature vector (i.e values) as well as the bitwise missingness indicator mask into contextual embeddings.
In this regard, our method is capable of producing independent scaling coefficients for each individual feature, surpassing the \textbf{SN} method on various datasets and thus confirms that vertical (per-feature) conditioned imputation surpasses horizontal (per-instance) methods. Furthermore, we extend the attention mechanism of AimNet \cite{aimnet} into one that uses a richer key-value embedding representation, capable of encapsulating position, missingness and value information.
On the contrary, in AimNet, only the missingness context is captured.
That enables our model to continuously improve the embedding representations with both existing and missing features.
Our approach, Multihead-Attention Imputation Networks (\textbf{MAIN}), builds on \textbf{FiLM} as it creates multiple coefficients for each feature (multiple vertical imputation), based on the attention mechanism.
In this modeling scenario, we create $N$ sets of $N$ coefficients where $N$ is the number of input features, with each set representing a conditioned modulation of a feature based on the value and existence of all features.
This learnable combination mechanism captures both global feature-value and feature existence effects.
Briefly our method extends the SN method of \citet{vsp} in two ways:
\begin{enumerate}
\item learning a different weight coefficient for each individual feature - instead of the whole feature vector, thus dismissing the concept of linear covariation as proposed in SN\cite{vsp}.
\item conditioning the coefficients not only on the existence vector of the input features but also on their observed values.
\end{enumerate}. Also our attention mechanism differs from \cite{aimnet} as we use multi-head attention and we build embeddings with both the value and the missing context of all the features.
\section{Related Work}
Imputation methods are often classified as \emph{discriminative} and \emph{generative}, based on the formulation of the missingness modeling mechanism.
In discriminative modeling, missing features are approximated directly through their conditional distributions, given the values of the existing features.
On the contrary, in generative modeling, missing features are often imputed in a two-step fashion: initially, the underlying joint distribution of all features is modeled, followed by a second step of conditionally generating the missing features from the existing information in the feature vector.
Discriminative methods depend on certain assumptions about the missingness mechanism.
Specifically, the data must conform to the Missing Completely At Random (MCAR) or - at a modest degree - Missing At Random (MAR) assumption \cite{missingness,misgan}.
For MCAR, the probability of an entry missing is independent from the values of the other features, while for MAR, that probability depends on the values of the other features.
In both cases, the probability of a feature missing is assumed to be independent from the value of that feature.
Generative models focus on the joint probability distribution of the data instead of the conditional \cite{MIWAE,HI-VAE}. In \cite{MIWAE} the authors deal with problem of missing data
with the use of deep latent variable models (DLVMs). Their approach is based on an importance-weighted autoencoder (IWAE) which maximises a potentially tight lower bound of the log-likelihood of the observed data. In \cite{HI-VAE} the authors propose a general framework implementing VAEs to fit incomplete heterogenous data.
Generative models in general are more flexible and can effectively capture multimodal distributions, while discriminative models rely on point estimates.
The extra flexibility of the generative approach allows these models to deal with strongly MAR scenarios, as well as cases where the probability of a feature missing is dependent on the value of that feature (Missing Not At Random, or MNAR). On the other hand they are difficult to train and specifically, GANs can suffer from slow convergence and mode collapse problems.
It should be pointed that, joint modeling may excel in scenarios where the underlying connections between features are of spatial and/or auto-regressive nature (e.g image completion) \cite{inpaint,misgan,ambient}, but in case of tabular data this can be equally approached by (multiple) conditional modeling, since there is not such a variety of modalities that can only be approached by manifold walks.
Furthermore, when the final objective is to build a predictive task based on specific input (e.g regression, classification) the flexibility that generative modeling provides is unfit for time-critical applications.
In a scenario of new incoming information, a generative model will act in a "two-step fashion". Firstly, it will impute the missing data producing a complete feature vector and as a second step will apply an auxiliary or different architecture/model on the downstream task. Discriminative methods on the other hand can operate both in single and two-step fashion.
The work by \citet{vsp} gave an in-depth insight to the first two questions about data imputation, and re-introduced us with the phenomenon of Variable Sparsity Problem (VSP), while making significant progress towards the ``one-step'' approach.
The VSP problem stresses an undesired phenomenon where the model's performance drops as its output significantly varies due to the rate of missingness in the given input also identified in \cite{dropblock}.
The solution proposed by \citet{vsp}, named Sparsity Normalization (SN), seems to give a significant boost in the reconstruction capabilities of existing networks, while also increasing the robustness of a model's downstream task when dealing with missing inputs.
In this work, we opt for a discriminative modeling schema, supporting the paradigm shift towards ``one-step'' methods, and study the effect of query-key-value attention performed in a multi-head fashion. We propose a custom key representation encoding, called Positional Encoded Vector (PEV). We show that even in heavily corrupted data our model incorporates the best attributes of discriminative and generative modeling, being an one-step method that is valid for both numeric and categorical inputs.
The contribution of this work can be summarized around 4 key ideas:
\begin{enumerate}
\item Due to its main application in real-time industrial scenarios, this work focuses on the performance gains of the downstream task(s), in contrast with other works that seem to sacrifice performance in favour of reconstruction of the full feature vector.
\item We show that our method has the best of both worlds between discriminative and generative methods as far as the downstream task performance is concerned.
It can handle both MCAR and MAR assumptions while not suffering from the limitation of training with complete vectors, that some discriminative and generative approaches silently imply. After testing its performance against 7 Datasets it shows consistent performance margin, as well as convergence without any instability issues.
\item We achieve a single-step, missingness-agnostic behaviour: our scheme can train on missing data directly, exploiting all the examples (including partially missing ones), and test on a different set never seen by the model, with completely untreated data (including missing features). This is not the case in autoencoder-based (i.e reconstruction) schemes like AimNet or generative methods \cite{gain}, where the model can train only on complete data where the missingness/reconstruction is artificially induced.
\item The PEV-based attention mechanism models both the missingness and the value of each individual feature in relation to the other features on the raw data distribution. AimNet \cite{aimnet} models only the missing context in a missing-induced distribution, while in VSP \cite{vsp} the feature vector is regularized horizontally, not taking into account each feature's individual variation.
\end{enumerate}
\section{Problem Formulation}
\subsection{Preliminaries}
Following the formulation of VSP \cite{vsp}, let $\rvec{x} \in \mathbb{R}^l$ and $\rvec{y} \in \mathbb{R}^d$ represent a training input/output pair instance of a model with $D$ tasks. Without loss of generality we will set $|D| = 1$ for the rest of our analysis.
To formulate missingness, we introduce $\rvec{m} \in \{0,1\}^l$, as the binary mask indicating missing values in $\rvec{x}$.
Hence, we define $\rvec{x_{\text{miss}}}=\rvec{x} \odot \rvec{m}$, where $\odot$ denotes the Hadamard product (element-wise multiplication) of two vectors, as the corrupted input the model observes after the effect of the missingness mechanism.
Finally, for the model let's assume a $N$-layer feed forward network, with convex, non-decreasing non-linearities $\sigma$ in each layer but the last.
Each layer $i$ contains $n_i$ units, and we use $\rvec{W^i} \in \mathbb{R}^{{n_i} \times {n_{i-1}}}$ to denote the weight matrix, $\rvec{b^i} \in \mathbb{R}^{n_i}$ to denote the bias, and $\rvec{h^i} \in \mathbb{R}^{n_i}$ for the post-activation output vector. The calculation executed at each layer can then be written as $\rvec{h^i} = \sigma^i (\rvec{W^i} \rvec{h^{i-1}} + \rvec{b^i})$.\footnote{Throughout the paper we use uppercase and bold notation for matrices, lower and bold notation for vectors and lower and italic notation for elements. Superscripts without parentheses as $\rvec{h}^i$ are used to denote a quantity at layer $i$, and superscripts in parentheses, as $h^{(i)}$ to denote the $i$th element of $\rvec{h}$.}
\subsection{Modeling Assumptions}
Our method works under the following two assumptions:
\begin{enumerate}
\item Each element $m^k$ of the binary mask $\rvec{m}$ is MAR, meaning it may not depend on the other mask elements or the value of the element $x^k$, but it does depend on the values of the original input vector $x_i, i \neq k$, as described in \cite{missingness}. We will denote the means of the mask vector
as $\mu_{m}$ and the means of the input vector as $\mu_{\text{missing}}$.
\item The coordinates of $\rvec{b^i}$ and the elements of $\rvec{W^i}$ are mutually independent and follow the same distribution with means $\mu^{i}_{b}$ and $\mu^{i}_{w}$ respectively, as in \cite{vsp,he,bengio}.
\end{enumerate}
\subsection{Methods}
\textbf{Theorem 1.} \textit{The expected value of the output layer of an N-Layer FFN with $l$-dimensional input and convex, non-decreasing, non-linearities under the MCAR assumption is bounded by}: $E[\rvec{h}^N] \geq f_N \circ f_{N-1} \circ \dots \circ f_1(\mu_{m}\mu_{x})$, where $f_{i}(x) = \sigma(\rvec{W}^{i} \cdot \rvec{x}^{i-1} + \rvec{b}^{i})$.
\vspace{3.5mm}
\textbf{Proof of 1.}
Proof is result of \cite{vsp}.
In Theorem 2 we extend the result of Theorem, for the more general MAR case.
\vspace{3.5mm}
\textbf{Theorem 2.} \textit{The expected value of the first layer of an N-Layer FFN of Theorem 1, under the MAR assumption, is bounded by: $E[\rvec{h}^1] \geq \sigma(n_0\mu^{1}_{w}\mu_{m}\mu_{\text{x}} + \mu^{1}_{b}) + T_1$, where $T_1 = \sigma(n_0\mu^{1}_{w}Cov(\rvec{x}^{1},\rvec{m}^{1}))$.}
\vspace{3.5mm}
\textbf{Proof of 2.}
For the output activation vector of the first layer $\rvec{h}^1$ it holds that:
\newline
\vspace{2mm}
$\rvec{h}^1 = \sigma(\rvec{W}^1 \times \rvec{x_{missing}} +\rvec{b}) = \sigma(\rvec{W}^1 \times \rvec{x} \odot \rvec{m} + \rvec{b})$.
\newline
Then, by taking the expected value of both sides:
\newline
\vspace{2mm}
$E[\rvec{h}^1] = E[\sigma(\rvec{W}^1 \times \rvec{x} \odot \rvec{m} + \rvec{b})]$,
\newline
And due to non-decreasing convexity of $\sigma$:
\newline
$E[\rvec{h}^1] \geq \sigma(E[\rvec{W}^1 \times \rvec{x} \odot \rvec{m} + \rvec{b}])$.
\vspace{2mm}
For the second part of the inequality it holds that:
\newline $\sigma(E[\rvec{W}^1 \times \rvec{x} \odot \rvec{m} + \rvec{b}])$ = $\sigma(E[\rvec{W}^1] \times E[\rvec{x} \odot \rvec{m}] + E[\rvec{b}]])$, based on assumption 2.
\vspace{2mm}
Focusing on elements of $E[\rvec{x} \odot \rvec{m}]$ we have that:
\begin{align*}
E[\rvec{x} \odot \rvec{m}] &= [E[x^{(1)} m^{(1)}], \dots, E[x^{(n_0)} m^{(n_0)}]]^T = \\
&=[\mu_m^{(1)} \mu_{\text{x}}^{(1)} + cov(x^{(1)}, m^{(1)}),\dots \\
&\dots,\mu_m^{(n_0)} \mu_{\text{x}}^{(n_0)} + cov(x^{(n_0)}, m^{(n_0)})]^T
\end{align*}
where $x^{(i)}, m^{(i)}$ refers to the $i$th component of the vector $\rvec{x}$ and
$\rvec{m}$ respectively.
For the conditional probabilities $p(x_i | m_i)$, and $p(x_i)$ we have under:
\begin{align*}
\text{MCAR: }& p(x_i | m_i, x_1, \dots x_{i-1}, x_{i+1}, \dots, x_l) = \\
& p(x_i | m_i) = p(x_i) \\
\text{MAR: }& p(x_i | x_1, \dots, x_{i-1}, x_{i+1}, \dots, x_l) = \\
&p(x_i | m_i, x_1, \dots, x_{i-1}, x_{i+1}, \dots, x_l)\\
&but\\
&p(x_i | m_i) \neq p(x_i)
\end{align*}
Recall that under the MCAR assumption $cov(x^{(i)}, m^{(i)}) = 0$, as $\rvec{x},\rvec{m}$ are independent.
Under the MAR assumption we are working with, $x^{(i)},m^{(i)}$ are conditionally independent, conditioned on the values of the rest of the features $x^{(j)}, i \neq j$, and in this case $cov(x^{(i)}, m^{(i)}) $ is not necessarily $0$.
\vspace{2mm}
In that fashion the elementwise calculations lead to:
\begin{align*}
\sigma (E[&\rvec{W}^1] \times E[\rvec{x} \odot \rvec{m}] + E[\rvec{b}]) = \sigma(E[\rvec{w}^{(1)}, \dots ,\rvec{w}^{(n_1)}] \\
& [\mu_m^{(1)} \mu_{\text{x}}^{(1)} + cov(x^{(1)}, m^{(1)}),\dots, \mu_m^{(n_0)} \mu_{\text{x}}^{(n_0)} + \\
& + cov(x^{(n_0)}, m^{(n_0)})]^T + E[b^{(1)}, \dots, b^{(n0)}])
\end{align*}
Finally for elements $i,j$ with $j = 0, \dots, n_0$ and $\rvec{w}^{(i)} \in \mathbb{R}^{n_0}$:
\begin{align*}
\sigma(E[&(\rvec{w}^{(i)})(\mu_m^{(j)} \mu_{\text{x}}^{(j)} + cov(x^{(j)}, m^{(j)})) + E[b^{(j)}])\\
&\geq \sigma(n_0\mu^{1}_{w}\mu_{m}\mu_{\text{x}} + \mu^{1}_{b}) + \sigma(n_0\mu^{1}_{w}cov(x^{(j)},m^{(j)}))
\end{align*}
Now let $T_1 = \sigma(n_0\mu^{1}_{w}Cov(\rvec{x}^{1},\rvec{m}^{1}))$, where $Cov(a,b)$ denotes the vector of element-wise covariances between items of $a$ and $b$ then:
$$E[\rvec{h}^1] \geq \sigma(n_0\mu^{1}_{w}\mu_{m}\mu_{\text{x}} + \mu^{1}_{b}) + T_1.$$
This term is propagated through all the network layers, adding bias to the model's output. (Full proof in the Appendix). A straightforward way to tackle the issue would be altering the model's input, by a subtraction debias:
$$\rvec{x_{new}} = \rvec{x} - [cov(x^{(1)}, m^{(1)}),\dots, cov(x^{(n)}, m^{(n)})]^T$$
thus alleviating the MAR-derived effect ($T1$ term). Subsequently, one could apply sparsity normalisation, as proposed by \citet{vsp}. This way, the mean of the network's output is not explicitly dependent on the the missingness rate over all data instances. However, that term is intractable since it requires the full features $x^{(i)}$ to be known.
Instead, we propose a simple mechanism that takes into account both the input feature values and their missingness pattern, and evaluates the similarity between data instances with the same, or similar, structure.
\vspace{2mm}
\textbf{Core Idea:} \textit{Latent factors conditioned on the value and missingness vector, including the covariance between each input feature and the respective missingness indicator can be approximated through transformer-like similarity scores between non-linear embedding projections under the MAR assumption.}
\vspace{2mm}
Instead of the transformation : $$\rvec{x_{new}} = (\rvec{x} - \rvec{\beta}) / \alpha $$
where $\rvec{\beta}$ is the per-feature covariance term and $\alpha$ the per-instance sparsity normalization scaling, we opt for an embedding mechanism.
The mechanism is conditioned on the fused value-existence input vectors $(\rvec{x} \odot \rvec{m}, \rvec{m})$, and creates a common representation embedding space $\textit{K} \subset \mathbb{R}^{emb}$ for the input instances.
Multihead self-attention is employed towards that goal, since, it can use the per-instance available context to retrieve the appropriate non-affine transformations from the embedding space ($\rvec{trans} \in \textit{K}$) and apply them to the provided input. In this way both the per-instance context (similar to $\alpha$), and the global context (similar to $\rvec{\beta}$) are incorporated in the final solution:
$$ lookup(\rvec{x}) = \rvec{trans} $$
$$\rvec{x_{new}} = transform(\rvec{x}, \rvec{trans})$$
where the lookup function, $\mathrm{lookup} : \mathbb{R}^{N} \to \mathbb{R}^{emb}$ is implemented with multihead self-attention, and the transform function, $\mathrm{transform} : \mathbb{R}^{(N+emb)} \to \mathbb{R}^{N}$ by non-linear feed-forward components.
\section{MAIN Algorithm}
The MAIN algorithms consists of two complementary steps: The Positional Encoded Vector (PEV) creation, and the multihead attention step with opacity gating.
Let $\rvec{x} \in \mathbb{R}^{N}$ denote an input vector of $N$ features, and $x_1, x_2, \dots , x_N$ the scalar values representing the value of each specific feature in $\rvec{x}$. PEV encoding augments the initial feature vector in order to:
\begin{enumerate}
\item Differentiate between the case of a "filled value", in case of a missing feature, i.e zero in zeta normalization scenarios, and the existing equivalent of that feature value in a feature vector - zero because the feature in that specific instance is equal to the mean of the distribution that generates it.
\item Create an orthogonal basis of feature existence. From this perspective, the original feature vector is treated as a linear combination of inter-changeable vectors whose direction is denoted by the feature position and magnitude by their respective feature's value.
\end{enumerate}
Here, for the specified input length $N$, the minimum number of bits required to describe all available positions, $bw$ is first calculated.
Then, for each training example $\rvec{x}$ the following components are calculated:
\begin{enumerate}[a)]
\item A binary mask, indicating the existence of each feature in $\rvec{x}$.
\item A binary positional encoding mask, marking the position of each feature in $\rvec{x}$.
\end{enumerate}
Finally, we concatenate a),b) feature-wise with the original input. That results to a 2D input matrix for each training example $\rvec{x}_{augmented} \in \mathbb{R}^{{N} \times {\log_{2}(N+1)}}$. In that way, we provide a simple way for the model to capture the similarity between data instances based on both their values and their missingness patterns.
The aforementioned PEV procedure is described in Algorithm 1.
\begin{algorithm}
\caption{PEV creation}\label{alg:pev}
\begin{algorithmic}[1]
\Function{PevMaskGenerator}{$x$}
\State $bw\gets \lceil{\log_2 n}\rceil$
\For{$i \gets 1$ to $n$}
\State $bin_i \gets bin_{bw}(i)$
\State $m_i \gets$ 0 if x[i] = $\emptyset$ else 1
\State $pev_i \gets [m_i|bin_i]$
\EndFor\label{euclidendwhile}
\State $pev \gets [pev_1, pev_2, ... pev_n]$
\State $m \gets [m_1, m_2, ... , m_n]$
\State \textbf{return} $pev, m$
\EndFunction
\end{algorithmic}
\end{algorithm}
\begin{figure}[H]
\centering
\captionsetup{justification=centering}
\includegraphics[width=\linewidth]{images/Pev2drawio.png}
\caption{Creation of PEV encoding Matrix.}
\label{fig:pev_creation}
\end{figure}
\begin{algorithm}
\caption{Opaque Multi Head Attention}\label{alg:main}
\begin{algorithmic}[1]
\While{$convergence$ $is$ $not$ $achieved$}
\For{$batch$ $in$ $X$}
\For{$(x,y)$ $in$ $batch$}
\State $pev, m \gets PevMaskGenerator(x)$
\State $query \gets proj_n(dq([x:pev]))$
\State $key \gets proj_k(dk([x:pev]))$
\State $value \gets proj_n(dv([x:pev]))$
\State $imputed_{x} \gets mha(query,key,value)$
\State $\hat{x} \gets \gamma \cdot imputed_{x} + (1-\gamma)\cdot query$
\For {$dsm_j$ $in$ $Model$ $Tasks$}
\State $\hat{y_j} \gets dsm_j(\hat{x})$
\EndFor
\State {$\hat{y} \gets [\hat{y_0}:\hat{y_1}:,\dots,:\hat{y_l}]$}
\State {$loss \gets e(y,\hat{y})$}
\EndFor
\State {$batch$ $loss \gets \sum loss$}
\State {$backprop(batch$ $loss)$}
\EndFor
\State \textbf{return}
\EndWhile
\end{algorithmic}
\end{algorithm}
\begin{figure}[H]
\includegraphics[width=\linewidth]{images/Query.png}
\caption{Creation of Query Projections.}
\label{fig:query_creation}
\end{figure}
\begin{figure}[H]
\includegraphics[width=\linewidth]{images/Key2.png}
\caption{Creation of Key Projections.}
\label{fig:key_creation}
\end{figure}
\begin{figure}[H]
\includegraphics[width=\linewidth]{images/Value.png}
\caption{Creation of Value Projections.}
\label{fig:value_creation}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\linewidth]{images/FullProcedure.png}
\caption{An overview of MAIN's architecture. It is comprised of three basic modules: a) PEV creation, b) Opaque Multi-head Attention, c) Downstream Tasks/Models}
\label{fig:attention}
\end{figure*}
The next step is a Luong-style \cite{luong} attention mechanism of each individual augmented feature over the total augmented feature vector, effectively serving as self-attention.
Note that the attention function is considered a mapping procedure between a Query and a set of Key-Value pairs \cite{aiayn}. In that fashion, we project the augmented input (PEV + features) into 3 embedding vector spaces, Query, Key and Value, using trainable projections. Here, the PEV augmentation will act as a mechanism capable to guide the focus of the self-attention heads into retrieving saliency from the existence-position prespective.
The example of Query projections creation is depicted in Figure \ref{fig:pev_creation}, and the Keys and Values are created similarly (Figures \ref{fig:key_creation}, \ref{fig:query_creation}). That of course, could not be possible without the projection of all features to a common embedding space. This is enabled through the use of PEV embeddings, that transform each feature from a mere scalar to a querable vector.
This is the key link with the $T_N$ bound of Theorem 2. The correlation between each feature value and the fact that it is missing, is conditioned on the values of the other features in the original measurement. That is in the end composed out of the each feature's value, and whether it exists or not in the specific instance.
As a next step, we perform multi-head attention over each augmented feature, noted as \textbf{\textit{mha}} in Algorithm \ref{alg:main}, resulting in the imputed feature vector $imputed_x$.
Finally, we employ a trainable gating mechanism called opacity gate, $\gamma(\cdot)$. The gate acts as a "trainable sigmoid knob" between the original input of the model, and the imputed input produced by the MAIN mechanism.
The gate is unconditioned on the input features, and initialized at 0.5, giving equal importance to both inputs. During training, the gate is left free to decide through backprop whether our method provides an input more "salient" towards minimizing the downstream loss (Gate $ \simeq 0 $) or if the original input is better (Gate $ \simeq 1 $). Furthermore, the gate acts in a twofold manner, providing an abstraction buffer towards the next layers. During the first iterations, the next layers will try to adapt towards an unoptimal solution to the task, based on the original inputs while the MAIN mechanism is still in early training stage. This is not permanent however, since the gate will shift towards the MAIN inputs, "fine-tuning" the rest of the model into the optimal solution.
The algorithm finishes with collecting all the opacity weighted feature vectors, $\hat{x}$ and feeding them to
an arbitrary number of downstream models, noted as \textbf{\textit{dsm}} in Algorithm \ref{alg:main}. They are jointly trained with the MAIN component to perform the required predictive tasks, and during this analysis are all considered to be 1, since no-multitask dataset was used.
\section{Experiments}
Our method was evaluated on 7 Tabular datasets, 6 of which come from the UCI Dataset Repository and 1 maritime dataset related to vessel's performance which is property of DeepSea Technologies (i.e DeepSea V9).
For all experiments, in order to be comparable with other methods that simulate missingness, binary sampling was performed per feature using the same technique as in \cite{vsp} and \cite{gain}. Our method though is not bound to this restriction and can work directly on missing data, exploiting the full capacity of the dataset. The downstream task in the UCI datasets was binary classification with imbalanced classes, while in the maritime dataset was regression. It is noteworthy to outline here that out of the 6 datasets, Breast, Credit, Spam and Heart are injected with MCAR missingness at 4 different levels: $[20\%,40\%,60\%,80\%]$, while Pima, Mammographic and DeepSea V9 have instrinsic MAR missingess at levels: $[12.2\%,3.35\%,7.28\%]$. In order for all the experiments to be compared at a common basis, the full rows of those datasets were polluted so that the total missing rate matched the $[20\%,40\%,60\%,80\%]$ scale, as a mixture of MCAR + MAR missingness.
It is important to outline that the scope of the experiments is towards to a \textbf{single-step} end-to-end training scheme. We found that the comparison with classical \textbf{two-step}, methods that first-impute then train/test formulations (i.e KNN, MICE, MCMC) is not only less relevant to the scope of this work but also unfair to them since they don't optimize for a specific task and give inferior results. For the sake of completeness however, we opt to compare with the \textbf{two-step tabular} method, GAIN \cite{gain}, since its authors shared our common interest of optimizing towards a downstream task. We also compared with SN \cite{vsp} as our method builds on the same ideas and can be seen as an extension to theirs. ZIMC (i.e zero imputation with mask concat) is serving as a baseline and FiLM \cite{fwt} although not an imputation method, since it also inspired our work can be seen as an intermediate step between SN and our work MAIN. We choose not to compare with AimNet \cite{aimnet}, as there was no official codebase, making it very difficult to repliate results. Furthermore,it was a DataBase-Oriented imputation method for the project HoloClean, not a general purpose Machine Learning Solution like the other compared methods.
The experimental design is similar to \cite{vsp} and \cite{aimnet}, reporting the test AUROC (UCI) and MSE (DeepSea V9) score of 5 runs.
In order to keep the comparison fair, the GAIN \cite{gain} method was trained and tested on different splits of the dataset unlike the original implementation where training and evaluation was in the same set.
The Breast, Spam, Credit and DeepSea V9 datasets were normalized in the range of [0,1], while Heart, Pima were normalized in the range [-1,1] as in \cite{vsp}, \cite{gain}.
We used no explicit prepossessing for categorical values in the above datasets since our scheme: a) creates automatically trainable embeddings for both scalar and categorical/ordinal variables and b) we don't reconstruct the missing feature and thus we don't have to map reconstructed logits to corresponding classes as in \cite{gain}.
Regarding train/validation/test splits, in most imputation-only approaches the RMSE reconstruction metric is usually reported on a single dataset without any splits. Since no official splits exist in any UCI datasets, instead of the 70-20-10 split used in \cite{vsp} we opt for a 70-30 split with 10-fold stratified cross-validation.
Models Setup:
In order to produce comparable results in UCI and test the performance gains of our imputation method as a base layer, we tried to keep the total number of layers (imputation + downstream) of our scheme close to what proposed in \cite{selu} as the most appropriate for the UCI and also used in \cite{vsp}; namely 4 Hidden layers @ 256 units and Adam Optimizer. The same principle applied to all the compared schemes where we had: imputation method + 4 layers dedicated to the downstream task.
Due to class imbalance, in our training scheme we pre-calculate the class weights of the target classes and use them into a weighted binary crossentropy loss, penalizing misclassification on the minority class more heavily that misclassifications on the majority class. Finally, the reported metric is the AUROC curve interpolated at 200 points with Riemann summation method.
\
\begin{table}[h]
\renewcommand{\arraystretch}{1.4}
\label{tb_breast}
\fontsize{6}{7}\selectfont
\centering
\captionsetup{singlelinecheck=off}
\caption{: UCI Breast-Wisconsin}
\begin{tabular}{l@{\hskip6pt}c@{\hskip8pt}c@{\hskip8pt}c@{\hskip8pt}c}
\hline
\textbf{MR} & \textbf{20\%} & \textbf{40\%} & \textbf{60\%} & \textbf{80\%}\\
\hline
ZIMC & 0.9501 $\pm$ 0.005 & 0.9485 $\pm$ 0.007 & 0.9182 $\pm$ 0.010 & 0.8470 $\pm$ 0.012 \\
GAIN$^{\star}$ & 0.9872 $\pm$ 0.008 & 0.9475 $\pm$ 0.011 & 0.9171 $\pm$ 0.035 & 0.8443 $\pm$ 0.017 \\
SN & 0.9683 $\pm$ 0.007 & 0.9341 $\pm$ 0.009 & 0.8593 $\pm$ 0.029 & 0.8640 $\pm$ 0.027 \\
FiLM & 0.9818 $\pm$ 0.003 & 0.9513 $\pm$ 0.005 & 0.9263 $\pm$ 0.008 & 0.8757 $\pm$ 0.015 \\
\textbf{MAIN} & \textbf{0.9821 $\pm$ 0.003} & \textbf{0.9786 $\pm$ 0.005} & \textbf{0.9693 $\pm$ 0.013} & \textbf{0.9241 $\pm$ 0.014} \\
\hline
\end{tabular}
\end{table}
\begin{table}[!htb]
\renewcommand{\arraystretch}{1.4}
\caption{: UCI Credit Dataset}
\label{tb_credit}
\centering
\captionsetup{singlelinecheck=off}
\fontsize{6}{7}\selectfont
\begin{tabular}{l@{\hskip6pt}c@{\hskip8pt}c@{\hskip8pt}c@{\hskip8pt}c}
\hline
\textbf{MR} & \textbf{20\%} & \textbf{40\%} & \textbf{60\%} & \textbf{80\%}\\
\hline
ZIMC & 0.7297 $\pm$ 0.008 & 0.7051 $\pm$ 0.007 & 0.6833 $\pm$ 0.010 & 0.6349 $\pm$ 0.013 \\
GAIN$^{\star}$ & 0.7412 $\pm$ 0.008 & 0.7173 $\pm$ 0.013 & 0.6849 $\pm$ 0.008 & 0.6019 $\pm$ 0.032 \\
SN & 0.7396 $\pm$ 0.005 & 0.7145 $\pm$ 0.007 & 0.6826 $\pm$ 0.003 & 0.6332 $\pm$ 0.012 \\
FiLM & 0.7443 $\pm$ 0.005 & 0.7187 $\pm$ 0.007 & 0.7025 $\pm$ 0.005 & 0.6528 $\pm$ 0.011 \\
\textbf{MAIN} & \textbf{0.7456 $\pm$ 0.004} & \textbf{0.7209 $\pm$ 0.006} & \textbf{0.7032 $\pm$ 0.004} & \textbf{0.6577 $\pm$ 0.009} \\
\hline
\end{tabular}
\end{table}
\begin{table}[!htb]
\renewcommand{\arraystretch}{1.4}
\centering
\captionsetup{singlelinecheck=off}
\caption{: UCI Spam Dataset}
\label{tb_spam}
\fontsize{6}{7}\selectfont
\begin{tabular}{l@{\hskip6pt}c@{\hskip8pt}c@{\hskip8pt}c@{\hskip8pt}c}
\hline
\textbf{MR} & \textbf{20\%} & \textbf{40\%} & \textbf{60\%} & \textbf{80\%}\\
\hline
ZIMC & 0.9740 $\pm$ 0.005 & 0.9519 $\pm$ 0.007 & 0.9243 $\pm$ 0.008 & 0.8675 $\pm$ 0.011 \\
GAIN$^{\star}$ & 0.9645 $\pm$ 0.004 & 0.9451 $\pm$ 0.008 & 0.9213 $\pm$ 0.001 & 0.8623 $\pm$ 0.010 \\
SN & 0.9798 $\pm$ 0.002 & 0.9568 $\pm$ 0.002 & 0.9270 $\pm$ 0.002 & 0.8707 $\pm$ 0.005 \\
FiLM & 0.9735 $\pm$ 0.005 & 0.9571 $\pm$ 0.006 & 0.9276 $\pm$ 0.007 & 0.8725 $\pm$ 0.012 \\
\textbf{MAIN} & \textbf{0.9817 $\pm$ 0.005} & \textbf{0.9589 $\pm$ 0.007} & \textbf{0.9298 $\pm$ 0.008} & \textbf{0.8732 $\pm$ 0.011} \\
\hline
\end{tabular}
\end{table}
\begin{table}[h]
\renewcommand{\arraystretch}{1.4}
\centering
\captionsetup{singlelinecheck=off}
\caption{\textsc{: UCI Pima Dataset}}
\label{tb_pima}
\fontsize{6}{7}\selectfont
\begin{tabular}{l@{\hskip6pt}c@{\hskip8pt}c@{\hskip8pt}c@{\hskip8pt}c}
\hline
\textbf{MR} & \textbf{20\%} & \textbf{40\%} & \textbf{60\%} & \textbf{80\%}\\
\hline
ZIMC & 0.8131 $\pm$ 0.016 & 0.7864 $\pm$ 0.014 & 0.7585 $\pm$ 0.015 & 0.7001 $\pm$ 0.015 \\
GAIN$^{\star}$ & 0.8074 $\pm$ 0.017 & 0.7861 $\pm$ 0.020 & 0.7503 $\pm$ 0.023 & 0.6987 $\pm$ 0.022 \\
SN & 0.8121 $\pm$ 0.013 & 0.7851 $\pm$ 0.021 & 0.7589 $\pm$ 0.017 & 0.7006 $\pm$ 0.015 \\
FiLM & 0.8017 $\pm$ 0.017 & 0.7725 $\pm$ 0.015 & 0.7475 $\pm$ 0.016 & 0.6913 $\pm$ 0.010 \\
\textbf{MAIN} & \textbf{0.8442 $\pm$ 0.009} & \textbf{0.7981 $\pm$ 0.007} & \textbf{0.7628 $\pm$ 0.011} & \textbf{0.7183 $\pm$ 0.012} \\
\hline
\end{tabular}
\end{table}
\begin{table}[!htb]
\renewcommand{\arraystretch}{1.4}
\centering
\captionsetup{singlelinecheck=off}
\caption{: UCI Heart Dataset}
\label{tb_heart}
\fontsize{6}{7}\selectfont
\begin{tabular}{l@{\hskip6pt}c@{\hskip8pt}c@{\hskip8pt}c@{\hskip8pt}c}
\hline
\textbf{MR} & \textbf{20\%} & \textbf{40\%} & \textbf{60\%} & \textbf{80\%}\\
\hline
ZIMC & 0.8256 $\pm$ 0.009 & 0.8002 $\pm$ 0.012 & 0.6752 $\pm$ 0.028 & 0.6702 $\pm$ 0.033 \\
GAIN$^{\star}$ & 0.8419 $\pm$ 0.039 & 0.7586 $\pm$ 0.065 & 0.6399 $\pm$ 0.102 & 0.6357 $\pm$ 0.094 \\
SN & 0.8533 $\pm$ 0.007 & 0.7974 $\pm$ 0.012 & 0.7570 $\pm$ 0.007 & 0.6619 $\pm$ 0.014 \\
FiLM & 0.8530 $\pm$ 0.007 & 0.7915 $\pm$ 0.014 & 0.7516 $\pm$ 0.009 & 0.6422 $\pm$ 0.017 \\
\textbf{MAIN} & \textbf{0.8702 $\pm$ 0.008} & \textbf{0.8371 $\pm$ 0.007} & \textbf{0.7851 $\pm$ 0.008} & \textbf{0.6802 $\pm$ 0.011} \\
\hline
\end{tabular}
\end{table}
\begin{table}[!htb]
\renewcommand{\arraystretch}{1.4}
\centering
\captionsetup{singlelinecheck=off}
\caption{: UCI Mammographic Dataset}
\label{tb_mamm}
\fontsize{6}{7}\selectfont
\begin{tabular}{l@{\hskip6pt}c@{\hskip8pt}c@{\hskip8pt}c@{\hskip8pt}c}
\hline
\textbf{MR} & \textbf{20\%} & \textbf{40\%} & \textbf{60\%} & \textbf{80\%}\\
\hline
ZIMC & 0.7851 $\pm$ 0.014 & 0.7433 $\pm$ 0.027 & 0.6712 $\pm$ 0.026 & 0.6049 $\pm$ 0.026 \\
GAIN$^{\star}$ & 0.8010 $\pm$ 0.024 & 0.7501 $\pm$ 0.036 & 0.7126 $\pm$ 0.028 & 0.6448 $\pm$ 0.057 \\
SN & 0.7794 $\pm$ 0.011 & 0.7367 $\pm$ 0.021 & 0.6753 $\pm$ 0.014 & 0.6063 $\pm$ 0.019 \\
FiLM & 0.8198 $\pm$ 0.012 & 0.7472 $\pm$ 0.023 & 0.6860 $\pm$ 0.042 & 0.6350 $\pm$ 0.032 \\
\textbf{MAIN} & \textbf{0.8807 $\pm$ 0.005} & \textbf{0.8535 $\pm$ 0.009} & \textbf{0.7998 $\pm$ 0.012} & \textbf{0.7682 $\pm$ 0.014} \\
\hline
\end{tabular}
\end{table}
\begin{table}[!htb]
\renewcommand{\arraystretch}{1.4}
\centering
\captionsetup{singlelinecheck=off}
\caption{: DeepSea V9}
\label{tb_ds9}
\fontsize{6}{7}\selectfont
\begin{tabular}{l@{\hskip6pt}c@{\hskip8pt}c@{\hskip8pt}c@{\hskip8pt}c}
\hline
\textbf{MR} & \textbf{20\%} & \textbf{40\%} & \textbf{60\%} & \textbf{80\%}\\
\hline
ZIMC & 0.0013 $\pm$ 0.0005 & 0.0030 $\pm$ 0.0006 & 0.0124 $\pm$ 0.0010 & 0.0280 $\pm$ 0.0013\\
GAIN$^{\star}$ & \textbf{0.0009 $\pm$ 0.0002} & \textbf{0.0027 $\pm$ 0.0002} & 0.0086 $\pm$ 0.0001 & 0.0267 $\pm$ 0.0004 \\
SN & 0.0012 $\pm$ 0.0002 & 0.0048 $\pm$ 0.0003 & 0.0116 $\pm$ 0.0005 & 0.0310 $\pm$ 0.0009 \\
FiLM & 0.0010 $\pm$ 0.0004 & 0.0035 $\pm$ 0.0005 & 0.0091 $\pm$ 0.0005 & 0.0270 $\pm$ 0.0009 \\
\textbf{MAIN} & 0.0009 $\pm$ 0.0003 & \textbf{0.0027 $\pm$ 0.0002} & \textbf{0.0085 $\pm$ 0.0002} & \textbf{0.0261 $\pm$ 0.0008} \\
\hline
\end{tabular}
\end{table}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.3\linewidth]{images/input.png}
\end{center}
\caption{An illustrated example of a training example with missing features. (feat.value/missing)}
\label{fig:ablationinput}
\end{figure*}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.90\linewidth]{images/final.png}
\end{center}
\caption{An illustrated example of the methods under test.}
\label{fig:ablation}
\end{figure*}
As Tables I-IV show, both GAIN and SN provide a far better alternative to the ZIMC baseline in every dataset and missing rate. While FiLM is of subpar performance towards both the generative GAIN and the discriminative SN in most cases, it is noteworthy, that in the case of the heavy MAR dataset (Pima), FiLM gives worse results to SN, while with none of the current state-of-the-art methods solving explicitly the MAR case.
On the contrary, the MAIN method seems to be vastly outperforming other methods, especially in the case of high missing rates, where it maintains considerably higher scores than its counterparts. In the case of the DeepSea V9 regression task, GAIN and MAIN have the least test MSE and perform equally in the low missing scenarios, while MAIN slightly outperforms GAIN in the high missing scenarios.
\section{Discussion}
In this section we will give an illustrated example of a partially missing feature vector in order to provide more insights on how our method works compared to other well-known methods. The setup is the following: in figure \ref{fig:ablationinput}, we present a z-normalized 4D feature vector that will be provided as an input to the different methods that were compared against our final solution, MAIN.
The 4D input vector consists of 3 non-missing features: [$x1, x3, x4$] with corresponding values [$0.35, 0.52, 0.00$], and one missing feature: $x2$. The respective existence bitmask of the input is depicted with a binary indicator after the "slash" symbol ('/') of each feature, where '1' means missing feature, thus the corresponding mask is $m=[0,1,0,0]$.
In figure \ref{fig:ablation} we break-down how the input information is potentially transformed (i.e Feature Transformation) by 4 different approaches, leading to a the Processed Vector on the right in the same figure. The processed vector can be seen as the result of a learnable scaling factor (for the most algorithms) which forms the input to the following-up machine learning method (e.g downstream model). The actual numbers are figurative just to illustrate the underlying mechanics of each method.
We observe that in the case of ZIMC as is the most naive method features are not explicitly affected since it is a mere zero imputation technique. It must be noted here, that the mask concatenation in ZIMC, potentially provides to the underline model with joint knowledge of missingness but that could happen in a deeper layer interactions in the case of neural nets. The even simpler zero-imputation couldn't discriminate at all, between a missing feature from a feature placed at the mode of the distribution in the z-scaled space.
In SN, the inverse L1 norm of the bitmask scales all the non-zero features, by multiplying by their number + 1 (K) which is an improvement over zero-fill. On the other hand the scaling of existing features at the mode of the distribution (i.e zero-mean like $x4$ ) are affected in the same way as the missing ones (i.e $x2$), re-scaling only the rest of them in order to alleviate the VSP problem\cite{vsp}. In FiLM, since all the features are subject to affine transformation conditioned in the missing context (feature + mask), the missing $x2$ and non-missing $x4$ zero-valued features are treated differently. Thus we claim that FiLM can be seen as a feature-wise rescaling extension of the SN which learns how to condition missing context to independent feature scalers. Following this logic and going one step further, if we explicitly define the conditioning mechanism (i.e attention) we have MAIN which offers even more expressivity to learn the underline mechanics of the missing distribution.
\section{Conclusion}
In this work, we proposed MAIN a novel method based on multi-head attention to deal with missing data in continuous or discrete datasets. Our method works in a single step by implicitly imputing missing data and is optimized directly on the downstream task, offering an end-to-end trainable system. We demonstrate that MAIN significantly outperforms state-of-the-art methods in a variety of open dataset and also in a proprietary one.
\FloatBarrier
|
\section{Introduction}
Einstein’s General Theory of Relativity is an extremely elegant and successful theory of gravitation, passing all of its experimental tests since its inception over 100 years ago \cite{Will_2014}. Yet its reconciliation with quantum theory remains elusive. While there is yet to be a proper full description of quantum gravity, a key piece of the puzzle is provided by black holes. Originally thought to be nature's ultimate repositories of matter and energy, their behaviour drastically changes once quantum physics is taken into account \cite{hawking1975particle}, leading to the well-known prediction that a black hole radiates like a thermal blackbody whose temperature is proportional to its surface gravity and whose entropy is proportional to its horizon area. In anti de Sitter spacetime they can undergo phase transitions \cite{Hawking:1982dh}, and in fact exhibit a very broad range of chemical thermodynamics \cite{Kubiznak:2016qmn}.
Another general expectation that emerges from quantum gravity is the presence of higher curvature terms that correct the Einstein-Hilbert action \cite{birrell_davies_1982,PhysRevD.16.953}. The most commonly discussed is Lovelock gravity \cite{Lovelock2,lovelock1971einstein}, which is regarded as a physically sensible generalization of Einstein gravity to higher dimensions since its field equations are second order in all metric components. These theories have the general feature that the entropy of a black hole is no longer proportional to the horizon area, and so are of particular interest in black hole thermodynamics since they provide a window into how quantum gravitational effects could modify the radiative behaviour of black holes \cite{Frassino_2014}.
So far work on black hole thermodynamics in Lovelock gravity
has been limited to black hole solutions that have a constant curvature manifold as its transverse space. However this limitation is not necessary:
recently a more general class of black hole solutions were found for Lovelock gravity in which the transverse space is a more general manifold
\cite{Ray_2015}. Such solutions take the form of a warped product of a two-dimensional space and an arbitrary transverse base manifold \cite{Dotti_2005,Dotti_2007,DOTTI_2009,Dotti_2010,Oliva_2013,Anabal_n_2011}. There is a generalization of the Birkhoff theorem that implies
this base manifold must be static. Furthermore, the field equations impose the conditions that
all the non-trivial intrinsic Lovelock tensors of the base manifold are constants that can be chosen arbitrarily. We shall refer to such objects as
`Exotic Lovelock Black Holes', or ELBHs.
We investigate here the thermodynamic properties of ELBHs in Einstein Maxwell Gauss-Bonnet Gravity, the simplest Lovelock gravity theory. ELBHs in this case depend on two parameters, and in the limit that these parameters are chosen so that when the base manifold has constant curvature
we recover thermodynamic phenomena for both neutral and charged black holes previously observed in Gauss-Bonnet gravity \cite{Frassino_2014}. We find that ELBHs exhibit new effects as they increasingly depart from this special case. For example we observe a novel triple point between thermal radiation and large and small ELBHs in $6$ dimensions. In the $d=6$ charged case we find that the particular features of the large/intermediate/small triple point depends on the horizon geometry.
We also find an interesting set of massless and negative mass black hole solutions that generalize those found previously in Einstein gravity \cite{Mann:1997jb,Aminneborg:1996iz,Smith:1997wx}. These have a more interesting structure insofar as two horizons are possible under some circumstances, and their singularity structure is more complicated than the corresponding situation in Einstein gravity.
We first begin with a review of Lovelock gravity, discussing black holes solutions whose transverse space is of constant and non-constant curvature and make a simple distinction between the two theories and how to transition between each theory, and provide a calculation of the Kretschmann scalar for an arbitrary metric. We follow that with Lovelock thermodynamics along with black hole thermodynamics. In Section 3 we examine Gauss-bonnet gravity, with first a discussion of vacumm solutions and then 5 and 6 dimensional black holes (uncharged then charged). In section 5 we give a summary of our results and future outlook.
\section{Exotic Lovelock Black Holes}
The Lagrangian for a Lovelock theory \cite{Lovelock2,lovelock1971einstein} in $d$ dimensions is
\begin{equation} \label{Lag}
\mathcal{L}=\frac{1}{16 \pi G_{N}} \sum_{k=0}^{K} \hat{\alpha}_{k} \mathcal{L}^{(k)}
\end{equation}
where $ \hat \alpha_{k}$ are the Lovelock coupling constants and $\mathcal{L}^{(k)}$ are the Euler densities
\begin{equation}
\mathcal{L}^{(k)}=\frac{1}{2^{k}} \delta_{c_{1} d_{1} \ldots c_{k} d_{k}}^{a_{1} b_{1} \ldots a_{k} b_{k}} R_{a_{1} b_{1}}^{c_{1} d_{1}} \ldots R_{a_{k} b_{k}}^{c_{k} d_{k}}.
\end{equation}
with the contraction occurring over the anti-symmetric generalized Kronecker delta. The dimension of the Euler densities is $2k$, with $\mathcal{L}^{(0)}$ beingthe cosmological constant, $\mathcal{L}^{(1)}$ the Ricci scalar, and $\mathcal{L}^{(2)}$ is the Gauss-Bonnett term. Note that we must have $d> 2K$ in order to have non-trivial field equations.
With the lagragnian \eqref{Lag} we can write our action for Lovelock theory as
\begin{equation} \label{action}
S= \int d^{d}x \sqrt{-g}\left( \frac{1}{16 \pi G_{N}}\sum_{k=0}^{K} \hat{\alpha}_{k} \mathcal{L}^{(k)} + \mathcal{L}_{matter} \right).
\end{equation}
Variation of the action with respect to the metric yields the field equations
\begin{equation} \label{vaceq}
\sum_{k}^{K} \hat{\alpha}_{k} \mathcal{G}^{(k)}_{ ab}= 8\pi G_{N} T_{ab}
\end{equation}
where $T_{ab}$ is the stress-energy tensor and
$\mathcal{G}^{(k)}_{ab}$ are the Lovelock tensors
\begin{equation}
\mathcal{G}_{ ab}^{(k)}=-\frac{1}{2^{(k+1)}} g_{z a} \delta_{b e_{1} f_{1} \ldots e_{k} f_{k}}^{z c_{1} d_{1} \ldots c_{k} d_{k}} R_{c_{1} d_{1}}^{e_{1} f_{1}} \ldots R_{c_{k} d_{k}}^{e_{k} f_{k}}.
\end{equation}
We will be examing charged black holes which means we will use the following Lagrangian $\mathcal{L}_{matter} = -4 \pi G_{N} F_{a b} F^{a b}$, which will give us our finalized field equations
\begin{equation} \label{fieldeq}
\sum_{k=0}^{K} \hat{\alpha}_{(k)} \mathcal{G}_{a b}^{(k)}=8 \pi G_{N}\left(F_{a c} F_{b}^{c}-\frac{1}{4} g_{a b} F_{c d} F^{c d}\right).
\end{equation}
We shall discuss the black hole solutions to these equations in the next section.
\subsection{Black Hole Solutions}
We will be focusing on charged AdS black hole solutions in this paper, using the ansatz
\begin{equation} \label{metric}
ds^2 =
{\textsf{g}}_{ij} {dy^i dy^j}
+ {\gamma}_{\alpha\beta}dx^\alpha dx^\beta
=-f(r)dt^2 + \frac{dr^2}{f(r)} + r^2 d\Sigma_{d-2}^2
\qquad
F = \frac{Q}{r^{d-2}}dt\wedge dr
\end{equation}
for the metric and gauge field strength,
that we require to be solutions to \eqref{fieldeq}. The coordinates $y^i = (t,r)$, with
${\textsf{g}}_{ij} = \textrm{diag}(-f(r),1/f(r))$.
The importance of this metric lies in the nature of the base manifold described by $d\Sigma_{d-2}^2$.
The most common approach is to take
$d\Sigma_{d-2}^2$ to be a $(d-2)$-dimensional compact space with constant curvature given by $(d-2)(d-3)\kappa$ with $\kappa=-1,0,+1$ corresponding to hyperbolic, flat, or spherical curvature respectively. This yields the polynomial equation
\cite{PhysRevLett.55.2656,Cai_2004,Castro_2013,Camanho_2013,Takahashi_2012}
\begin{equation} \label{constpoly}
\sum_{k=0}^{K} \alpha_{k}\left(\frac{\kappa-f(r)}{r^{2}}\right)^{k}=\frac{16 \pi G M}{(d-2) \Sigma_{d-2}^{(\kappa)} r^{d-1}}-\frac{8 \pi G_{N} Q^{2}}{(d-2)(d-3)} \frac{1}{r^{2 d-4}}
\end{equation}
from \eqref{fieldeq}.
Here M is the mass of the black hole, $\Sigma_{d-2}^{(\kappa)}$ is the volume of the compact space whose metric is
$d\Sigma_{d-2}^2$.
Q is the black hole charge given by
\begin{equation}
Q=\frac{1}{2 \Sigma_{d-2}^{(\kappa)}} \int * F
\end{equation}
The $\alpha_k$ terms are re-scaled Lovelock coupling constants
\begin{equation}
\alpha_{0}=\frac{\hat{\alpha}_{(0)}}{(d-1)(d-2)} =
\quad \alpha_{1}=\hat{\alpha}_{(1)}, \quad \alpha_{k}=\hat{\alpha}_{(k)} \prod_{n=3}^{2 k}(d-n) \quad \text { for } \quad k \geq 2
\end{equation}
where the cosmological constant
$\Lambda = -\hat{\alpha}_{(0)}/2$. In what follows we will set $\alpha_{1}=1$ to retrieve general relativity in the low energy limit.
However it is possible to
assume that
$d\Sigma_{d-2}^2$ is the metric of a more general $(d-2)$-dimensional base manifold that does not have to be of constant curvature. In this case the field equations \eqref{fieldeq} become
\begin{equation}
\mathcal{G}^{i}_{j} \equiv -\frac{ (d-2)(d-1) \delta_{j}^{i}}{2(d-1) ! r^{d-2}} \sum_{n=0}^{\bar{k}}\left\{(d-2 n-2) ! \hat{\mathcal{L}}^{(n)}\right\}\left\{\frac{d}{d r}\left(r^{d-2 n-1} A_{n}\left(\frac{-f(r)}{r^2}\right )\right)\right\}= 8 \pi G_{N} T^{i}_{j} \label{rawfield1}
\end{equation}
\begin{equation}
\mathcal{G}^{\alpha}_{\beta} \equiv \left.\frac{(d-1)(d-2)}{(d-1) ! r^{d-3}} \sum_{n=0}^{\bar{k}}\left\{(d-2 n-3) ! \hat{\mathcal{G}}^{(n)\alpha}_{\beta}\right\}\right\}\left\{\frac{d^{2}}{d r^{2}}\left(r^{d-2 n-1} A_{n}\left(\frac{-f(r)}{r^2}\right)\right)\right\} = 8 \pi G_{N} T^{\alpha}_{\beta} \label{rawfield2}
\end{equation}
with the polynomial
\begin{equation}
A_{n}\left(\frac{-f(r)}{r^{2}}\right) \equiv \sum_{k=n}^{K} \alpha_{k} \binom{k}{n} \left( \frac{-f(r)}{r^{2}}\right)^{k-n}
\end{equation}
which satisfies the recurrence relation
\begin{equation}
A_{n}^{\prime}\left(\frac{-f(r)}{r^{2}}\right)=(n+1)A_{n+1}\left(\frac{-f(r)}{r^{2}}\right).
\end{equation}
The quantities $\hat{\mathcal{L}}^{(n)}$ and $\hat{\mathcal{G}}^{(n)\alpha}_{\beta}$ are respectively the Euler Characteristic and Lovelock Tensors of the base manifold
\begin{equation}
\hat{\mathcal{L}}^{(n)}=\frac{(d-2)! b_{n}}{(d-2 n-2) !} \qquad \qquad \hat{\mathcal{G}}^{(n)\alpha}_{\beta}=-\frac{(d-3)!b_{n}}{2 (d-2n-3)!}\delta^{\alpha}_{\beta}
\end{equation}
reducing the field equations to
\begin{equation}
\mathcal{G}_{j}^{i} \equiv \frac{-(d-2) \delta_{j}^{i}}{2 r^{d-2}} \frac{d}{d r} \sum_{n=0}^{K}\left\{b_{n}\left(r^{d-2 n-1} A_{n}\left(\frac{-f(r)}{r^{2}}\right)\right)\right\}=8 \pi G_{N} T_{j}^{i} \label{finalfield1}
\end{equation}
\begin{equation}
\mathcal{G}^{\alpha}_{\beta} \equiv \frac{- \delta^{\alpha}_{\beta}}{2 r^{d-3}} \frac{d^2}{dr^2} \sum_{n=0}^{K} \left\{ b_{n}\left( r^{d-2 n-1} A_{n}\left(\frac{-f(r)}{r^{2}}\right) \right) \right\} = 8 \pi G_{N} T^{\alpha}_{\beta} \label{finalfield2}
\end{equation}
In this case the polynomial equation in $f(r)$ becomes \cite{Ray:2015ava}
\begin{equation} \label{souryapoly}
\sum_{n=0}^{K} \frac{b_{n}}{r^{2n}} \left( \sum_{k=n}^{K} \alpha_{k} \binom{k}{n}\left(\frac{-f(r)}{r^{2}}\right)^{k-n}\right )=\frac{16 \pi G_{N} M}{(d-2) \Sigma_{d-2} r^{d-1}}-\frac{8 \pi G_{N} Q^{2}}{(d-2)(d-3)} \frac{1}{r^{2 d-4}}
\end{equation}
generalizing \eqref{constpoly}.
We shall refer to the solutions that follow from solutions
to \eqref{souryapoly} in which $f(r)$ vanishes at least once for some $r>0$ as
\textit{Exotic Lovelock Black Holes}, or ELBHs, where the term $b_{n}$ is introduced and is referred to as the topological parameter. We can set $b_{0}=1$ without any loss of generality.
Before proceeding to the thermodynamics we mention a few more things with regards to the field equations \eqref{finalfield1} and \eqref{finalfield2}. Upon comparing the left-hand sides of equations \eqref{constpoly} and \eqref{souryapoly}, we obtain
\begin{equation}\label{bncc}
b_{n}=\kappa^n
\end{equation}
for black holes whose base manifolds have constant curvature.
We define the mass, using the Hamiltonian formulation, as the conserved charge corresponding to the time translational Killing vector of a background spacetime to which the black hole solutions approach in the asymptotic region. For black holes with maximally symmetric horizons, this is usually chosen as a constant curvature spacetime that solves the field equations \cite{Frassino_2014,Cai:2003kt,2013JHEP...07..164C,Camanho:2011rj,Takahashi:2011du}. In our case, we choose this background solution to have identical geometry of the base manifold as that of the black hole solution under consideration while the corresponding metric function $\bar{f}(r)$ solves the equation
\begin{equation} \label{souryapolyref}
\sum_{n=0}^{K} \frac{b_{n}}{r^{2n}} \left( \sum_{k=n}^{K} \alpha_{k} \binom{k}{n}\left(\frac{-\bar{f}(r)}{r^{2}}\right)^{k-n}\right )=0
\end{equation}
One might note an apparent discrepancy in the above equation while considering the dimension $d=2K+1$, since in this case there are $K-1$ constants $b_n$ characterizing the geometry of the base manifold. In this particular dimensionality, $b_K$ does not carry any geometric information of the base manifold but rather corresponds to the choice of the integration constant specifying the background spacetime.
This becomes evident if we rewrite the above equation as
\begin{equation} \label{souryapolyref2}
\sum_{n=0}^{K-1} \frac{b_{n}}{r^{2n}} \left( \sum_{k=n}^{K} \alpha_{k} \binom{k}{n}\left(\frac{-\bar{f}(r)}{r^{2}}\right)^{k-n}\right )=-\frac{b_K\alpha_K}{r^{2K}}
\end{equation}
For spherical base manifolds, $b_K=1$ then corresponds to the natural choice of a constant curvature spacetime as the background. We shall adopt the convention that the sum in \eqref{souryapoly} extends up to $K$ for all $d$, recognizing that if $d=2K+1$ the parameter $b_K$ corresponds to a convention for choosing $M$ and contains no geometric information.
In analyzing the structure of our solutions, we will find it useful to employ the Kretschmann scalar, which for the metric \eqref{metric} can be written as
\begin{equation}\label{Kscal}
R^{ abcd }R_{abcd }=\left(\frac{d^{2} f(r)}{d r^2}\right)^{2} + 2\frac{(d-2)}{r^2}\left( \frac{d f(r)}{dr} \right) ^2 + 2\frac{(d-2)(d-3) f(r)^2}{r^4} - 4\frac{ R[\gamma] f(r)}{r^4} +\frac{\mathcal{K}[\gamma]}{r^4}
\end{equation}
where
\begin{equation}
\mathcal{K}[\gamma]=R^{ \alpha \beta \mu \nu }R_{ \alpha \beta \mu \nu }[\gamma]
\end{equation}
is the Kretschmann scalar of the base manifold.
\subsection{Lovelock Black Hole Thermodynamics}
Lovelock black holes of mass $M$, entropy $S$, temperature $T$, and charge $Q$ obey the extended First Law and Smarr relations \cite{Jacobson_1993,Kastor_2010}
\begin{equation}
\delta M=T \delta S-\frac{1}{16 \pi G_{N}} \sum_{h} \hat{\Psi}^{(k)} \delta \hat{\alpha}_{(k)}+\Phi \delta Q
\end{equation}
\begin{equation}
(d-3) M=(d-2) T S+\sum_{k} 2(k-1) \frac{\hat{\Psi}^{(k)} \hat{\alpha}_{(k)}}{16 \pi G_{N}}+(d-3) \Phi Q
\end{equation}
where we regard the $\hat \alpha_{(k)}$ as thermodynamic parameters. The $\hat \Psi^{(k)}$ are their respective conjugate thermodynamic potentials, given by
\begin{equation}
\hat{\Psi}^{(k)}=4 \pi T \mathcal{A}^{(k)}+\mathcal{B}^{(k)}+\Theta^{(k)}
\end{equation}
with
\begin{equation}
\begin{aligned}
\mathcal{B}^{(k)} &=-\frac{16 \pi k G_{N} M(d-1) !}{b(d-2 k-1) !}\left(-\frac{1}{\ell^{2}}\right)^{k-1}, \quad b=\sum_{k} \frac{\hat{\alpha}_{k} k(d-1) !}{(d-2 k-1) !}\left(-\frac{1}{\ell^{2}}\right)^{k-1} \\
\Theta^{(k)} &=\int_{\Sigma} \sqrt{-g} \mathcal{L}^{(k)}[s]-\int_{\Sigma_{\mathrm{AdS}}} \sqrt{-g_{\mathrm{AdS}}} \mathcal{L}^{(k)}\left[s_{\mathrm{AdS}}\right]
\end{aligned}
\end{equation}
where $ \ell^2 = 1/\alpha^{(0)}$ is the 'AdS' radius. The spatial hypersurface $\Sigma$,
with timelike unit normal $n^a$ and induced metric $s_{a b}=g_{a b}+n_{a} n_{b}$,
extends from the horizon to infinity.
Black holes in Lovelock gravity no longer obey the area relation $S=\frac{A_{H}}{4}$, but instead have entropy given by \cite{Jacobson_1993}
\begin{equation}
S=\frac{1}{4 G_{N}} \sum_{k} \hat{\alpha}_{k} \mathcal{A}^{(k)}, \quad \mathcal{A}^{(k)}=k \int_{\mathcal{H}} \sqrt{\sigma} \mathcal{L}^{(k-1)}
\end{equation}
where $\sigma$ is the determinant of the induced metric on the horizon, and $\mathcal{L}^{(k-1)}$ are the corresponding Euler densities.
We will be treating the (negative) cosmological constant as the thermodynamic pressure
\begin{equation} \label{pressure}
P=-\frac{\Lambda}{8 \pi G_{N}}=\frac{\hat{\alpha}_{0}}{16 \pi G_{N}}, \quad V=-\hat{\Psi}^{(0)}
\end{equation}
with $V$ the conjugate thermodynamic volume of the black hole.
\subsubsection{ELBH Thermodynamics}
We wish to solve \eqref{metric} to obtain AdS black hole solutions, solutions for which $f(r)$ grows quadratically with $r$ for large $r$ and has $r=r_+ >0$ as its largest linear zero. Although we could explicitly solve \eqref{metric} in the $K=2$ Gauss-Bonnet case, this is not necessary as we can employ the Hamiltonian formalism
\cite{Cai_2004,Kastor_2010,RayExotic}. We can find the thermodynamic parameters of the black hole without an explicit solution of $f(r)$ for any value of $K$.
Setting $f(r_+)=0$, we find
\begin{equation}
M=\frac{\Sigma_{d-2}(d-2)}{16 \pi G_N} \sum_{k=0}^{K} \alpha_{k} b_{k} r_{+}^{d-1-2 k}+\frac{\Sigma_{d-2}}{2(d-3)} \frac{Q^{2}}{r_{+}^{d-3}}.
\end{equation}
\begin{equation} \label{temp}
T=\frac{f^{\prime}\left(r_{+}\right)}{4 \pi}=\frac{1}{4 \pi r_{+} D\left(r_{+}\right)}\left[\sum_{k=0}^{K} b_{k} \alpha_{k}(d-2 k-1)r_{+}^{-2(k-1)}-\frac{8 \pi G_N Q^{2}}{(d-2) r_{+}^{2(d-3)}}\right]
\end{equation}
where $D(r_{+})$ is
\begin{equation}
D\left(r_{+}\right)=\sum_{k=1}^{K} k \alpha_{k}b_{k-1} r_{+}^{-2(k-1)}\; .
\end{equation}
The entropy is given by
\begin{equation}
S=\frac{\Sigma_{d-2}(d-2)}{4 G_N} \sum_{k=0}^{K} \frac{k b_{k-1} \alpha_{k} r_{+}^{d-2 k}}{d-2 k}
\end{equation}
and through the first law it is easy to identify the conjugate potentials
\begin{equation}
\Psi^{(k)}=\frac{\Sigma_{d-2}(d-2)}{16 \pi G_{N}} r_{+}^{d-2 k}\left[\frac{b_{k}}{r_{+}}-\frac{4 b_{k-1} \pi k T}{d-2 k}\right].
\end{equation}
The thermodynamic volume is
\begin{equation}\label{volume}
V=-\hat{\Psi}^{(0)}=\frac{16 \pi G_{N} \Psi^{(0)}}{(d-1)(d-2)}=\frac{\Sigma_{d-2}r_{+}^{d-1}}{ {d-1}}.
\end{equation}
Using \eqref{pressure} and \eqref{temp} we obtain the equation of state
\begin{equation}\label{eos}
P=\frac{d-2}{16 \pi G_N} \sum_{k=1}^{K} \frac{\alpha_{k}}{r_{+}^{2}}r_{+}^{-2(k-1)}\left[4 \pi k b_{k-1} r_{+} T-b_{k}(d-2 k-1)\right]+\frac{Q^{2}}{2 r_{+}^{2(d-2)}}
\end{equation}
where $r_+$ is a function of the thermodynamic volume from \eqref{volume}.
We note the non-trivial dependence of the various thermodynamic parameters on the topological constants $b_k$. This implies that we might expect new phase behaviour for ELBHs as compared to their constant curvature counterparts. To investigate this we shall study the Gibbs free energy \cite{Kubiznak:2016qmn}
\begin{equation}\label{gibbs0}
G(P,T,Q) =M-TS
\end{equation}
which characterizes the canonical ensemble. A thermodynamically stable state is given by the global minimum of $G$ for any given choice of the parameters. To observe phase transitions, it is most useful to plot $G$ as a function of $T$, fixing the other parameters.
This, however, is not sufficient to determine a physically acceptable black hole thermodynamic state. We shall also require that
\begin{equation}
\sum_{k=0}^{K} \frac{k b_{k-1} \alpha_{k} r_{+}^{d-2 k}}{d-2 k} \geq 0. \label{posentropy}
\end{equation}
so that the entropy \label{entropy} is not negative. Likewise, we shall only consider $T\geq 0$ in \eqref{temp}. However we shall not require that $M>0$, since it is known that, for example, topological black holes can have negative masses that are bounded from below \cite{Mann:1997jb}.
\section{ELBHs in Gauss-Bonnet Gravity}
We now specialize our considerations to the $K=2$ Gauss-Bonnet case.
\subsection{Solutions}
\noindent Setting $K=2$, in \eqref{souryapoly} we obtain at the polynomial equation for $f(r) \equiv f$ in Gauss-Bonnet-Lovelock gravity (where we recall $b_{0}=1$ and $\alpha_{1}=1$)
\begin{equation}
\frac{\alpha_{2} f^{2}}{r^{4}}+\left(-\frac{1}{r^{2}}-\frac{2 b_{1} \alpha_{2}}{r^{4}}\right) f+\alpha_{0}+\frac{b_{1}}{r^{2}}+\frac{b_{2} \alpha_{2}}{r^{4}}=\frac{16 \pi M}{(d-2) \Sigma_{d-2} r^{d-1}}-\frac{8 \pi Q^{2}}{(d-2)(d-3) r^{2 d-4}}
\end{equation}
whose solutions are
\begin{equation}\label{elbhsol}
f= f_\pm(\textsf{m},\textsf{q}) \equiv \frac{r^{2}+2 b_{1} \alpha_{2} \pm \sqrt{\left( b_{1}^{2}- b_{2}\right)4 \alpha_{2}^{2} + r^4(1-4 \alpha_{2} \alpha_{0})
+\frac{8\textsf{m}\alpha_{2}}{r^{d-5}}
-\frac{4 \textsf{q}^2 \alpha_{2}}{r^{2(d-4)}}
}}{2 \alpha_{2}}
\end{equation}
where
\begin{equation}
\textsf{m} \equiv \frac{8 \pi M}{(d-2) \Sigma_{d-2}} \qquad
\textsf{q}^2 \equiv \frac{8 \pi Q^{2}}{(d-2)(d-3)}\; .
\end{equation}
The solution $f_{-}(\textsf{m},\textsf{q})$ has the limit
\begin{equation}\label{einlim}
\lim_{\alpha_2 \to 0} f_{-}(\textsf{m},\textsf{q}) = \alpha_0 r^2 + b_1 - \frac{2\textsf{m}}{r^{d-3}} + \frac{\textsf{q}^2}{r^{2(d-3)}}
\end{equation}
recovering the solution in Einstein gravity for $b_1 =\kappa$.
We also require that $f_\pm(\textsf{m},\textsf{q})\to r^2$ for large $r$. This implies
\begin{equation}\label{largercond}
1 - 4 \alpha_{2} \alpha_{0}\geq 0.
\end{equation}
independent of the $b_k$. Equation
\eqref{largercond} implies from \eqref{pressure} that there is a maximum pressure \cite{Frassino_2014}
\begin{equation}\label{maxP}
P \leq P_{\max}= \frac{(d-1)(d-2)}{64 \pi \alpha_2}
\end{equation}
such that if this bound is violated the spacetime is no longer asymptotically AdS.
The horizons are located at
\begin{equation}\label{horpm}
r_{\pm}^2(\textsf{m},\textsf{q}) = -\frac{1}{2\left(\alpha_0-\frac{2\textsf{m}}{r_\pm^{d-5}}
-\frac{\textsf{q}^2}{r_\pm^{2(d-4)}} \right)}\left(
b_1 \pm \sqrt{b_1^2 - 4 b_2 \alpha_2\alpha_0
+b_2 \alpha_2\left(\frac{8\textsf{m}}{r_\pm^{d-5}}
-\frac{4\textsf{q}^2}{r_\pm^{2(d-4)}} \right)
}
\right)
\end{equation}
which implicitly defines $r_\pm$. We note from this that solutions
with $\textsf{m}=\textsf{q}=0$
\begin{equation}\label{vacsol}
f_\pm(0,0) = \frac{r^{2}+2 b_{1} \alpha_{2} \pm \sqrt{\left( b_{1}^{2}- b_{2}\right)4 \alpha_{2}^{2} + r^4(1-4 \alpha_{2} \alpha_{0})}}{2 \alpha_{2}}
\end{equation}
have horizons at
\begin{equation}\label{vachor}
r_{\pm}^2(0,0) = \frac{1}{2\alpha_0}\left(-
b_1 \pm \sqrt{b_1^2 - 4 b_2 \alpha_2\alpha_0}
\right)
\end{equation}
provided either
\begin{align}\label{cond1}
\textrm{(a)}&\quad b_2<0 \quad\textrm{and}\quad b_1>0\quad
\Rightarrow r_{+}(0,0) \;\textrm{is the only horizon} \\
& \nonumber
\\
\textrm{or, if}\quad
b_2 > 0 \nonumber\\
\textrm{(b)}&\quad b_1^2 > 4 b_2 \alpha_2 \alpha_0 \quad \textrm{and}\quad 0 > b_1 >
- \sqrt{\frac{b_2 }{2(1-2\alpha_2\alpha_0)}}
\quad
\Rightarrow r_{\pm}(0,0) \;\textrm{are both horizons}
\label{cond2}
\\
\textrm{or}
\nonumber\\
\textrm{(c)}&\quad b_1^2 > 4 b_2 \alpha_2 \alpha_0 \quad \textrm{and}\quad - {\sqrt{\frac{ b_2 }{2(1-2\alpha_2\alpha_0)}}} > b_1 \quad \Rightarrow r_{+}(0,0) \;\textrm{is the only horizon}
\label{cond3}
\end{align}
where the inequalities \eqref{cond2} ensures that $r_{-}(0,0)$ is real and larger than the location of the spacetime singularity.
If any of these conditions do not hold then the solution has a naked singularity.
The solutions \eqref{vacsol} are generalizations of massless topological black holes
in Einstein gravity
\cite{Mann:1997jb,Aminneborg:1996iz,Smith:1997wx}, with $\alpha_2=0$ and $b_1=-1$, with appropriate identifications made on the transverse base space \cite{Aminneborg:1996iz,Mann:1997iz}. Here we have a richer set of possibilities insofar as two horizons are possible, as long as $b_2\alpha_2 > 0$. Negative mass solutions are likewise possible.
An evaluation of the Kretschmann Scalar \eqref{Kscal} for the
solution \eqref{vacsol} yields
\begin{align}
R^{ abcd }R_{abcd } &=\frac{1}{\alpha_2}+
{\frac {-48\, \left( {{ b_1}}^{2}-{ b_2} \right) {r}^{2}{ \alpha_0}\,
{{ \alpha_2}}^{3}+12\,{r}^{2} \left( 4{r}^{4}{{ \alpha_0}}^{2}/3+{{ b_1}
}^{2}-{ b_2} \right) {{ \alpha_2^2}}-8\,{ \alpha_0}\,{ \alpha_2}\,{r}^{6}}{{ \alpha_2}\,
\left(\left( b_{1}^{2}- b_{2}\right)4 \alpha_{2}^{2} + r^4(1-4 \alpha_{2} \alpha_{0}) \right) ^{3/2}}}
\nonumber\\
&-2\,{\frac {
\left( d-2 \right) \left( -4\,{ \alpha_0}\,{ \alpha_2}\,{r}^{2}+{r}^{2}+
\sqrt { \left( b_{1}^{2}- b_{2}\right)4 \alpha_{2}^{2} + r^4(1-4 \alpha_{2} \alpha_{0})} \right) ^{2}}{{{ \alpha_2^2}}
\left( \left( b_{1}^{2}- b_{2}\right)4 \alpha_{2}^{2} + r^4(1-4 \alpha_{2} \alpha_{0}) \right) }}\nonumber\\
&
+{\frac { \left( d-2
\right) \left( d-3 \right) \left( {r}^{2}+2\,{ b_1}\,{ \alpha_2}+
\sqrt { \left( b_{1}^{2}- b_{2}\right)4 \alpha_{2}^{2} + r^4(1-4 \alpha_{2} \alpha_{0})} \right)^{2}}{2{r}^{4}{
{ \alpha_2^2}}}}\nonumber\\
&
-2\,{\frac {R[\gamma] \left( {r}^{2}+2\,{ b_1}\,{ \alpha_2}+
\sqrt { \left( b_{1}^{2}- b_{2}\right)4 \alpha_{2}^{2} + r^4(1-4 \alpha_{2} \alpha_{0})} \right) }{{ \alpha_2}\,{r}^{4}}}+{\frac {K[\gamma]}{{r}^{
4}}}
\end{align}
which clearly has singularities at $r=0$ and at
\begin{equation}\label{sing2}
r_s=\left ( \frac{(b_2-b_{1}^{2}) 4 \alpha_{2}^{2}}{1- 4 \alpha_{2} \alpha_{0}}\right)^{\frac{1}{4}}\; .
\end{equation}
Imposing the condition \eqref{largercond},
it is straightforward to show that the inner horizon $r_{-}(0,0) > r_s$
provided \eqref{cond2} holds; otherwise $r_{+}(0,0) > r_s$
if either of \eqref{cond1} or
\eqref{cond3} hold.
If $b_2 < b_1^2$ the singularity in \eqref{sing2} is absent, but the $r=0$ singularity in general remains.
If $b_2 = b_1^2$ the spacetime is of constant curvature and all singularities are absent.
\subsection{Equation of State}
The equation of state \eqref{eos} for $K=2$ is
\begin{equation}\label{eosK2}
P=\frac{(d-2) T}{4 r_{+}}-\frac{(d-2)(d-3) b_{1}}{16 \pi r_{+}^{2}}+\frac{(d-2) \alpha_{2} b_{1} T}{2 r_{+}^{3}}-\frac{(d-2)(d-5) \alpha_{2} b_{2}}{16 \pi r_{+}^{4}}+\frac{Q^{2}}{2 r_{+}^{2(d-2)}}.
\end{equation}
and becomes
\begin{equation}\label{eosdim}
p=\frac{t}{v}-\frac{(d-2)(d-3) b_{1}}{4 \pi v^{2}}+\frac{2 b_{1} t}{v^{3}}-\frac{(d-2)(d-5) b_{2}}{4 \pi v^{4}}+\frac{q^{2}}{v^{2}(d-2)}.
\end{equation}
upon introducing the dimensionless variables $(v,t,m,q,p)$
\cite{Frassino_2014}
\begin{equation}\label{dimv}
r_{+}=v \alpha_{2}^{\frac{1}{2}}, \quad T=\frac{t \alpha_{2}^{-\frac{1}{2}}}{d-2}, \quad m=\frac{16 \pi M}{(d-2) \Sigma_{d-2}}{\alpha_{2}^{\frac{d-3}{2}}}, \quad Q=\frac{q}{\sqrt{2}} \alpha_{2}^{\frac{d-3}{2}}, \quad P=\frac{p}{4 \alpha_{2}}
\end{equation}
and recalling \eqref{maxP} it becomes
\begin{equation} \label{maxp}
p \leq p_{\max }=\frac{(d-1)(d-2)}{16 \pi }
\end{equation}.
Critical points are obtained by solving
\begin{equation}
\frac{\partial p}{\partial v}=0, \quad \frac{\partial^{2} p}{\partial v^{2}}=0 \label{criticalpointsequation}
\end{equation}
where the first equation determines critical temperature and the latter yields the critical volume. We obtain
\begin{equation}\label{critt}
t_{c}=\frac{\left(-4 v_{c}^{8-2 d} q^{2} \pi+b_{1}(d-3) v_{c}^{2}+2 b_{2}(d-5)\right)(d-2)}{2 \pi\left(v_{c}^{2}+6 b_{1}\right) v_{c}},
\end{equation}
and
\begin{equation}
\begin{split}\label{critv}
(48 v_{c}^{8-2 d} \pi d b_{1}+ & 8 v_{c}^{10-2 d} \pi d-168 v_{c}^{8-2 d} \pi b_{1}-20 v_{c}^{10-2 d} \pi) q^{2}+\\ &\left(-d b_{1}+3 b_{1}\right) v_{c}^{4}+\left(6 d b_{1}^{2}-6 b_{2} d-18 b_{1}^{2}+30 b_{2}\right) v_{c}^{2}-12 d b_{1} b_{2}+60 b_{1} b_{2}=0
\end{split}
\end{equation}
The Gibbs free energy
can also be written in dimensionless form using
\eqref{dimv} as
\begin{equation}
g=\frac{1}{\Sigma_{d-2}^{(\kappa)}} \alpha_{2}^{\frac{3-d}{2}} G
\end{equation}
yielding
\begin{equation}
\begin{split}
g&=\frac{(d-2)\left(\frac{v^{d-2}}{d-2}+\frac{2 b_{1} v^{d-4}}{d-4}\right)\left(\frac{4 \pi p v^{2}}{d-2}+(d-3) b_{1}+\frac{(d-5) b_{2}}{v^{2}}\right)}{16 \pi v\left(1+\frac{2 b_{1}}{v^{2}}\right)} \\
& +\frac{(d-2)\left(\frac{4 \pi p v^{d-1}}{(d-1)(d-2)}+b_{1} v^{d-3}+b_{2} v^{d-5}\right)}{16 \pi} \\
& +\frac{q^{2}\left((2 d-5)(d-4) v^{2}+2 b_{1}(d-2)(2 d-7)\right)}{4(d-4)(d-2)(d-3)\left(v^{2}+2 b_{1}\right) v^{d-3}}
\end{split}
\end{equation}
from \eqref{gibbs0}.
The positive entropy condition \eqref{posentropy} for $K=2$ is
\begin{equation}
\frac{r_{+}^{d-2}}{d-2}+\frac{2 \alpha_{2} b_{1} r_{+}^{d-4}}{d-4} \geq 0 \Rightarrow
\frac{v^{d-2}}{d-2}+\frac{2 b_{1} v^{d-4}}{d-4} \geq 0
\end{equation}
which is always satisfied for $b_1 > 0$. For
$b_1 < 0$
\begin{equation}
v \geq \sqrt{2\frac{|b_{1}|(d-2)}{d-4}}
\end{equation}
implying that the size of the black hole must
be sufficiently large for it to have positive entropy.
Finally, we note that
the vacuum horizon equation \eqref{vachor} becomes
\begin{equation}
v_{\pm}^{2}(0,0)=\frac{(d-1)(d-2)}{8 \pi p} \left( -b_{1} \pm \sqrt{b_{1}^{2} - \frac{16 \pi p b_{2}}{(d-1)(d-2)}} \right)
\end{equation}
using the dimensionless variables \eqref{dimv}.
\subsection{Five Dimensions}
In light of our discussion in section 2.1, in five dimensions we must obey certain conditions in order to have compatibility with the constant curvature case. The conditions are $ b_{2}=b_{1}^{2}$, where
$ b_{1} = -1,0,1$. We shall consider only $b_1=\pm 1$ in what follows.
With this, we no longer have the possibility of a singularity outside of the origin in vacuum spacetime (which we can see in \eqref{sing2} with the $b_{2}=b_{1}^{2}$ condition). Our $f(r)$ solution for vaccuum is given by
\begin{equation}
f_{\pm}(0,0)=\frac{r^{2}+2 b_{1} \alpha_{2} \pm \sqrt{r^{4}\left(1-4 \alpha_{2} \alpha_{0}\right)}}{2 \alpha_{2}}
\end{equation}
yielding
\begin{equation}
r_{\pm}^{2}(0,0)=\frac{\left (\pm \sqrt{1-4 \alpha_{2} \alpha_{0}}-1\right) b_{1}}{2 \alpha_{0}}.
\end{equation}
for the horizons.
Note that since $ 1 > 1-4 \alpha_{0} \alpha_{2}>0$, there will be two horizons provided $b_{1}<0$. For $1- 4 \alpha_{0} \alpha_{2}=0$, there are two coincident horizons
\begin{equation}
r_{+}^{2}=r_{-}^{2}=\frac{- b_{1}}{2 \alpha_{0}}
\end{equation}
which corresponds to be being at maximum AdS pressure, where
$b_{1}<0$ or else there will be no horizons.
As $\alpha_{0} \rightarrow 0$, the only horizon is
\begin{equation}
\lim_{\alpha_{0}\to\ 0} r^{2}_{+}(0,0)=- b_{1} \alpha_{2}
\end{equation}
for $b_{1}<0$, or else no horizons are present. While the former horizon corresponds to a maximum AdS pressure the later corresponds to zero AdS pressure. Which shows that as we turn off the cosmological constant we will still maintain a vacuum singularity as long as we properly choose our topological term $b_{1}$.
For nonzero $M$ and $Q$ we obtain
\eqref{elbhsol}
\begin{equation}\label{elbhsol5d}
f = \frac{r^{2}+2 b_{1} \alpha_{2} \pm \sqrt{ r^4(1-4 \alpha_{2} \alpha_{0})
+ {8\textsf{m}\alpha_{2}}
-\frac{4 \textsf{q}^2 \alpha_{2}}{r^{2}}
}}{2 \alpha_{2}}
\end{equation}
from \eqref{elbhsol}. For $Q = 0$ there is a bound on the mass
\begin{equation}
\textsf{m} \geq -\frac{b_{1}^{2}\left(1-4 \alpha_{0} \alpha_{2}\right)}{8 \alpha_{0}}
\end{equation}
which provides a lower (negative) bound for the mass, below which uncharged black hole solutions do not exist.
\subsubsection{Uncharged ELBH Thermodynamics}
The five dimensional uncharged equation of state is given by
\begin{equation}
p=\frac{t}{v}-\frac{3 b_{1}}{2 \pi v^{2}}+\frac{2 t b_{1}}{v^{3}}
\end{equation}
where
\begin{equation}
p_{max}=\frac{3}{4 \pi} \approx 0.2387324146 \label{maxp5d}
\end{equation}
is maximum dimensionless pressure.
The Gibbs free energy is
\begin{equation}
g=-\frac{3\left(\frac{1}{3} v^{3}+2 b_{1} v\right)\left(\frac{4 \pi p v^{2}}{3}+2 b_{1}\right)}{16 \pi v\left(1+\frac{2 b_{1}}{v^{2}}\right)}+\frac{3\left(\frac{1}{3} \pi p v^{4}+b_{1} v^{2}+b_{2}\right)}{16 \pi}.
\end{equation}
The critical temperature and volume are
obtained from \eqref{critt} and \eqref{critv}
\begin{equation}
t_{c}=\frac{3 b_{1} v_{c}}{\pi\left(v_{c}^{2}+6 b_{1}\right)}, \quad v_{c}^{2}-6 b_{1}=0.
\end{equation}
whose solutions are
\begin{equation}
v_{c}=\sqrt{6 b_{1}}, \quad t_{c}=\frac{\sqrt{6 b_{1}}}{4 \pi}, \quad p_{c}=\frac{1}{12 \pi} = \frac{p_{max}}{9}.
\end{equation}
We see that we must have positive values of $b_{1}$ in order to have real critical points;
this in turn ensures the positive entropy condition holds for all black hole sizes.
With the condition that $b_{1}^{2}=b_{2}$ we can first begin with the constant curvature case, $b_{1}=b_{2}=1$. We see the standard Van der Waals behaviour\cite{Frassino_2014} shown in Figure~\ref{5DUncharged1}. Seen in the center image there is an intersection between large and small black hole branches, indicating a large/small first order phase transition. However these black holes are unstable since their free energy is greater than that of $g=0$ which corresponds to AdS radiation. Instead, when the large black hole branch crosses the $g=0$ axis it will undergo a Hawking/Page transition into thermal AdS radiation \cite{Hawking:1982dh}. The phase diagram for this first order transition is displayed in the right image of Figure~\ref{5DUncharged1}.
The apparent sharp corner occurs at the critical point $p=p_c$ for the unstable black hole branch. It is actually smooth, but corresponds to a very steep rise in pressure as a function of temperature for
$p_c < p < p_{max}$ from the equation of state, as shown in Figure~\ref{pc/pmax-gt} .
\begin{figure}[H]
\centering
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{5DUncharged/pvb11.pdf}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{5DUncharged/gtb11b21.pdf}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{5DUncharged/ptfinal.pdf}
\end{subfigure}
\caption{\textbf{Phase Behaviour for $d=5$, $q=0$, $b_{1}=b_{2}=1$ Black Holes.} \textit{Left:} $p-v$ diagram with constant temperature slices of the \textit{unstable BH} showing the oscillation for $t<t_{c}$. \textit{Center}: $g-t$ diagram with constant pressure slices around $p_{c}$ again, of \textit{unstable BHs} showing swallowtail structure with intersection between large and small black holes. \textit{Right:} Phase diagram of the Black Hole / Radiation showing the termination at the maximum pressure.}
\label{5DUncharged1}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.5\textwidth]{5DUncharged/pcpmaxgt.pdf}
\caption{$g-t$ plots for two constant pressure slices: $p_{crit}$ of the unstable black hole and $p_{max}$ of the spacetime.}
\label{pc/pmax-gt}
\end{figure}
It is interesting to compare the phase behaviour of considered previously for uncharged $d=5$ Gauss-Bonnet black holes with the present case. As shown in Figure~\ref{fig:compare}, the coexistence line between the radiation/large black hole phases is very close to that of the small/large black hole phases. Approaching the diagram from the right, it is clear that as the temperature decreases, the large black hole will undergo a phase transition to radiation before that for a small black hole.
\begin{figure}[H]
\centering
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{5DUncharged/ptcomparison1.pdf}
\end{subfigure}
\centering
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{5DUncharged/ptcomparison2.pdf}
\end{subfigure}
\caption{Left: Phase diagram showing the first-order coexistence line of the unstable large/small transition (red) and the coexistence line of the Hawking-Page transition(black). Right: Close up version of the left diagram. The unstable small black hole phase is between the red and black lines. }
\label{fig:compare}
\end{figure}
\subsubsection{Charged ELBH Thermodynamics}
Including charge, the equation of state and Gibbs free energy are now
\begin{equation}
p=\frac{t}{v}-\frac{3 b_{1}}{2 \pi v^{2}}+\frac{2 t b_{1}}{v^{3}}+\frac{q^{2}}{v^{6}}
\end{equation}
\begin{equation}
g=-\frac{3\left(\frac{1}{3} v^{3}+2 b_{1} v\right)\left(\frac{4 \pi p v^{2}}{3}+2 b_{1}\right)}{16 \pi v\left(1+\frac{2 b_{1}}{v^{2}}\right)}+\frac{3\left(\frac{1}{3} \pi p v^{4}+b_{1} v^{2}+b_{2}\right)}{16 \pi}+\frac{q^{2}\left(5 v^{2}+18 b_{1}\right)}{24\left(v^{2}+2 b_{1}\right) v^{2}}
\end{equation}
and the critical temperature equation and critical volume relations \eqref{critt} and \eqref{critv} become
\begin{equation}\label{tvcrit5dq}
t_{c}=\frac{3\left(b_{1} v_{c}^{4}-2 \pi q^{2}\right)}{\pi v_{c}^{3}\left(v_{c}^{2}+6 b_{1}\right)}, \quad 3 v_{c}^{6} b_{1}-18 v_{c}^{4} b_{1}^{2}-\left(30 \pi v_{c}^{2}+108 \pi b_{1}\right) q^{2}=0.
\end{equation}
Since the latter is a cubic equation in $v_c^2$,
analytic solutions are possible for arbitrary values of $b_{1}$ and $q$. All roots of the cubic will be positive and real only if its coefficients alternate in sign, which is not possible for any values of $b_1$ or $q$. Hence there can be at most two admissible solutions for $v_c$ from \eqref{tvcrit5dq}.
Plotting in Figure~\ref{5dqvolplots} the critical volume solutions for specific choices of $q$, we see that negative values of $b_{1}$ are permitted as well as positive ones, unlike the uncharged case. For any given charge there is a negative value of $b_{1}$
below which there are no longer any real solutions for $v_c$ from \eqref{tvcrit5dq}; for $q=1$ this is approximately $b_1 = -1.5$. As we decrease the charge, the magnitude of the most negative allowed value of $b_{1}$ also decreases, as expected since only positive values of $b_{1}$ are permitted for $q=0$.
For $b_1>0$, there is always a positive root to the cubic, so all positive values of $b_1$ yield a positive critical volume.
\begin{figure}[H]
\centering
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{5DCharged/5dq01vol.pdf}
\caption{$q=0.1$}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{5DCharged/5dq05vol.pdf}
\caption{$q=0.5$}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{5DCharged/5dq1vol.pdf}
\caption{$q=1$}
\end{subfigure}
\caption{Critical volume solutions vs $b_{1}$ for varying values of $q$, the topological parameter $b_{2}$ does not play a role in these solutions.}
\label{5dqvolplots}
\end{figure}
We first begin with $q=1$, $b_{1}=b_{2}=1$, illustrating the results in Figure~\ref{pvgt5dq(1)}. In this configuration they exhibit the standard Van der Waals behaviour of a large/small black hole phase transition \cite{Frassino_2014} . There is no transition into radiation (thermal AdS) due to conservation of charge. This behaviour is qualitatively the same for all possible values of $b_1>0$.
\begin{figure}[H]
\centering
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{5DCharged/pvq1b11.pdf}
\end{subfigure}
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{5DCharged/gtq1b11b21.pdf}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{5DCharged/ptq1b11b21.pdf}
\end{subfigure}
\caption{\textbf{Phase Behaviour for $d=5$, $q=1$\,,\,$b_{1}=b_{2}=1$ black holes.} \textit{Left:} $p-v$ diagram with constant temperature values mimicking the uncharged case with Van der Waals oscillation. \textit{Bottom:} $g-t$ diagram with constant pressure slices showing large/small branch intersection. \textit{Right:} phase diagram displaying first order transition terminating at the critical point.}
\label{pvgt5dq(1)}
\end{figure}
We illustrate the situation for $b_1=-1$ (and $b_2=1$) in Figure~\ref{pvgt5db1-1}. In this case there is no Van der Waals type behaviour for $p<p_{max}$, and no interesting phase behaviour.
\begin{figure}[H]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{5DCharged/pvq1b1-1.pdf}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{5DCharged/gtq1b1-1b21.pdf}
\end{subfigure}
\caption{\textbf{Phase Behaviour for $d=5$, $q=1$, $b_{1}=-1,\: b_{2}=1$ black holes.} \textit{Left:} $p-v$ diagram for constant temperature slices. \textit{Right:} $g-t$ diagram for constant pressure slices beginning with maximum pressure and decreasing. Solid lines represent black holes with positive entropy while dotted lines correspond to negative entropy}
\label{pvgt5db1-1}
\end{figure}
\subsection{Six Dimensions}
In six dimensions we are free to choose our own values of $b_{1}$ and $b_{2}$. In this case a vacuum singularity could occur at
\begin{equation}
v=\left ( \frac{4 (b_2-b_{1}^{2}) }{1 - \frac{4 p \pi}{5} }\right)^{\frac{1}{4}}
\end{equation}
provided $b_{2}-b_{1}^2>0$,
since $p < p_{max} =\frac{5}{4\pi} = 0.3978873576$.
The vacuum horizon equation \eqref{vachor} can be rewritten as
\begin{equation}
v_{\pm}^{2}(0,0)= \frac{5}{2 \pi p} \left( -b_{1} \pm \sqrt{b_{1}^2 - \frac{4 \pi p b_{2}}{5}}\right)
\end{equation}
using the dimensionless variables \eqref{dimv}.
Now we can re-write the vacuum horizon conditions
into dimensionless form as follows:
\begin{align}
\textrm{(a)}&\quad b_2<0 \quad\textrm{and}\quad b_1>0\quad
\Rightarrow v_{+}(0,0) \;\textrm{is the only horizon} \\
& \nonumber
\\
\textrm{or, if}\quad
b_2 > 0 \nonumber\\
\textrm{(b)}&\quad b_1^2 > \frac{4 \pi p b_{2}}{5} \quad \textrm{and}\quad 0 > b_1 >
- \sqrt{\frac{b_{2}}{2 \left( 1 - \frac{2 \pi p}{5}\right) }}
\quad
\Rightarrow v_{\pm}(0,0) \;\textrm{are both horizons}
\label{cond2a}
\\
\textrm{or}
\nonumber\\
\textrm{(c)}&\quad b_1^2 > \frac{4 \pi p b_{2}}{5} \quad \textrm{and}\quad - \sqrt{\frac{b_{2}}{2 \left(1 - \frac{2 \pi p}{5} \right)}} > b_1 \quad \Rightarrow v_{+}(0,0) \;\textrm{is the only horizon}
\label{cond3a}
\end{align}
For nonzero $M$ and $Q$ we have
\begin{equation}\label{elbhsol6d}
f = \frac{r^{2}+2 b_{1} \alpha_{2} \pm \sqrt{ r^4(1-4 \alpha_{2} \alpha_{0})
+ \frac{{8\textsf{m}\alpha_{2}}}{r}
-\frac{4 \textsf{q}^2 \alpha_{2}}{r^{4}}
}}{2 \alpha_{2}}
\end{equation}
from \eqref{elbhsol}. For $Q = 0$ we obtain
\begin{equation}
\textsf{m} \geq \textsf{m}_{\pm} \equiv - \frac{\sqrt{10}\sqrt{-\alpha_{0}\left(3 b_{1} \pm \sqrt{-20 \alpha_{0} \alpha_{2} b_{2}+9 b_{1}^{2}}\right)} \left ( - 3 b_{1}^{2} \mp \sqrt{9 b_{1}^{2} - 20 \alpha_{0} \alpha_{2} b_{2}}+ 20 \alpha_{0} \alpha_{2} b_{2}\right)}{500 \alpha_{0}^{2}}
\end{equation}
as a lower (negative) bound for the mass, below which uncharged black hole solutions do not exist, with
$\textsf{m}_+$ corresponding to $b_1>0$ and
$\textsf{m}_-$ to $b_1<0$.
Note that $\textsf{m}_+ > \textsf{m}_{-}$ (solutions with
$b_1<0$ can have more negative mass) and that
$b_2 \alpha_2 \alpha_0 < 9 b_1^2/20$ for such solutions to exist. If $b_2\alpha_2 >0$ then
only $\textsf{m}_{-} < 0$.
\subsubsection{Uncharged ELBHs}
The equation of state and Gibbs free energy for $d=6$ and $q=0$ are
\begin{equation}
p=\frac{t}{v}-\frac{3 b_{1}}{\pi v^{2}}+\frac{2 t b_{1}}{v^{3}}-\frac{b_{2}}{\pi v^{4}}
\end{equation}
\begin{equation}
g= -\frac{\left(\frac{1}{4} v^{4}+b_{1} v^{2}\right)\left(\pi p v^{2}+3 b_{1}+\frac{b_{2}}{v^{2}}\right)}{4 \pi v\left(1+\frac{2 b_{1}}{v^{2}}\right)}+\frac{\frac{1}{5} \pi p v^{5}+b_{1} v^{3}+b_{2} v}{4 \pi}
\end{equation}
The critical temperature relation \eqref{critt} is now
\begin{equation}
t_{c}=\frac{2\left(3 b_{1} v_{c}^{2}+2 b_{2}\right)}{\pi v_{c}\left(v_{c}^{2}+6 b_{1}\right)}
\end{equation}
and the critical volume relation is \eqref{critv} is
\begin{equation}
6 v_{c}^{4} b_{1}+\left(-36 b_{1}^{2}+12 b_{2}\right) v_{c}^{2}+24 b_{1} b_{2}=0 \label{critv6d}
\end{equation}
whose solutions
are \begin{equation}
v^2_{c\pm} = \frac{ {\left(3 b_{1}^{2}\pm\sqrt{(9 b_{1}^{2}-b_{2})( b_{1}^2-b_{2})}-b_{2}\right)}}{b_{1}}
\end{equation}
If $b_{1}>0$ then we must have
$b_2 \leq b_1^2$ in order that
$v_{c}$ be real and positive.
If $b_2 >0$, then both
$v_{c\pm}$ are valid critical solutions. Conversely, if $b_1<0$ then we must have either
$b_2<0$ or $b_2 > 9 b^2_1$; if the latter holds both
$v_{c\pm}$ are valid critical solutions.
To examine the phase behaviour of the uncharged solutions, we first set $b_{1}=1$, with the results displayed in Figure~\ref{6dpvgt0q}. For $b_{2}=1$, corresponding to the standard horizon geometries \cite{Frassino_2014},
we get a cusp structure for the $g-t$ diagram, and a maximal pressure in the $p-v$ diagram. There is also a minimum, temperature-dependent volume for which $p=0$.
As in the 5 dimensional case, the black hole will undergo a Hawking/Page transition from a large black hole into thermal AdS.
Things become more interesting as the value of $b_2$ changes. As $b_2$ decreases, we
recover standard Van der Waals behaviour with a single oscillation in the $p-v$ diagram and the familiar swallowtail structure in the Gibbs free energy diagram.
Unlike the 5 dimensional case, this intersection of the swallowtail
occurs below the $g=0$ axis and therefore is a genuine first order large/small first order phase transition between stable black holes -- this transition is generally not observed for uncharged black holes.
As $b_2$ becomes negative, no further qualitative changes in phase behaviour are seen.
The phase diagrams are depicted in Figure~\ref{pt6duncharged}. Here we can see that when we have $b_{2}=1$ we observe a Hawking/Page transition between a large black hole and thermal AdS, whereas for $b_{2}=-1$ we have the standard Van der Waals transition from a large black hole to a small one.
We can notice something between the phase diagram displayed on the right of Figure~\ref{pt6duncharged} and that of the right of Figure~\ref{5DUncharged1}. In the 5 dimensional case there is a sharp but smooth bend in the diagram which corresponds to the critical point of the unstable black hole; however in the 6 dimensional case the change after this point is not as dramatic. This can be easily explained with reference to the $g-t$ diagram. In the 6 dimensional case there is no swallowtail behaviour for $b_{1}=b_{2}=1$; instead we only observe a cusp. The coexistence curve is correspondingly less sharp, although we still observe a steep slope as $p_{max}$ is approached. In the $g-t$ diagram, the curves for $p_{c}$ (the only value that is a solution to \eqref{criticalpointsequation}) and $p_{max}$ are close together, with only a small difference in temperature between them.
\begin{figure}[H]
\centering
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/pvb11b21.pdf}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/pvb11b20.pdf}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/pvb11b2-1.pdf}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/gtb11b21.pdf}
\caption{$b_{2}=1$}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/gtb11b20.pdf}
\caption{$b_{2}=0$}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/gtb11b2-1.pdf}
\caption{$b_{2}=-1$}
\end{subfigure}
\caption{\textbf{ Phase Behaviour for $d=6$, $q=0$, $b_{1}=1$ black holes.} \textit{Top}: Three $p-v$ diagrams for varying values of $b_{2}$ at constant temperature slices. \textit{Bottom}: Corresponding $g-t$ plots of constant pressure with $b_{2}$ values displayed below.The images on the left display critical temperature/pressure pressures for a non-stable black hole.}
\label{6dpvgt0q}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.35\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/pt6db11b2-1.pdf}
\end{subfigure}
\begin{subfigure}{0.35\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/pt6dfinal.pdf}
\end{subfigure}
\caption{\textbf{Coexistence curves for $d=6$, $q=0$,\, $b_{1}=1$ black holes.} \textit{Left:} $p-t$ phase diagram for $b_{2}=-1$ displaying first order transition between Small/Large BH, which terminates at the critical point. \textit{Right:} $p-t$ phase diagram for $b_{2}=1$ here we have a transition between Large BH and AdS radiation which terminates at the maximum pressure.
}
\label{pt6duncharged}
\end{figure}
For $1 >b_2 > 0$ swallowtail behaviour occurs and the possibility of new phenomenon emerges -- that of a new kind of black hole triple point. In Figure~\ref{6Dq0pvgtb205} we see
(for $b_{1}=1$ and $b_{2}=0.5$) that for $p<p_c$ we can have
a large black hole undergoing a first order phase transition into a small black hole that in turn undergoes a transition to thermal AdS as the temperature is lowered further. If we decrease the pressure even further, we arrive at Figure~\ref{6DNewTrip}, in which we have the large/small transition occurring on the $g=0$ axis. This implies a novel triple point where we have the coexistence large and small black holes with thermal AdS.
The phase diagram displaying this novel triple point can be found in Figure~\ref{ptnoveltrip}. For sufficiently low pressures, there are only 2 phases, thermal AdS and the large black hole. As the pressure increases, the triple point emerges where the small black hole phase coexists with the other two. At pressures above the novel triple point pressure we observe the three distinct phases as the temperature varies. There is a further critical pressure at which small and large black holes are no longer distinct phases; above this pressure we again have just a single Hawking/Page transtion between thermal AdS and a black hole.
Although within the range of $0 < b_{2} <1$ swallowtail behaviour is observed, not all values of $b_2$ in that range yield a novel-triple point. For $b_{1}=1$ we find that only the range $0<b_{2}<\frac{2}{3}$ for which the novel triple point occurs. This can be obtained by solving for the critical volume and temperature, in turn yielding the constraint \begin{equation}
\frac{\left(-3 b_{2}+12\right) \sqrt{b_{2}^{2}-10 b_{2}+9}-3 b_{2}^{2}+31 b_{2}-36}{\pi\left(\sqrt{b_{2}^{2}-10 b_{2}+9}+b_{2}-3\right)^{2}\left(b_{2}-9+\sqrt{b_{2}^{2}-10 b_{2}+9}\right)} < 0
\end{equation}
that must be satisfied in order to obtain the critical pressure; its solutions are
$0<b_{2}<\frac{2}{3}$.
If $b_{2}=2/3$, we only have one degenerate critical pressure, at which the novel triple point terminates at the critical point of the black hole, pushing the small branch out and only leaving a large/radiation transition.
\begin{figure}[H]
\centering
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/pvb11b20.5.pdf}
\end{subfigure}
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/gtb11b20.5.pdf}
\end{subfigure}
\caption{\textbf{Phase Behaviour for $d=6$, $q=0$, $b_{1}=1$, $b_{2}=0.5$ black holes.} \textit{Left:} $p-v$ diagram for constant temperature slices dislaying Van der Waals oscillations. \textit{Right:} $g-t$ diagram for constant pressure slices.}
\label{6Dq0pvgtb205}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.4\textwidth]{6DUncharged/gtnewtrip.pdf}
\caption{\textbf{Novel Triple Point for $d=6$, $b_{1}=1$,$b_{2}=0.5$}. Gibbs temperature diagram showing the two-branch intersection occuring along the $g=0$ axis.}
\label{6DNewTrip}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.5\textwidth]{6DUncharged/finalptuncharged6d.pdf}
\caption{\textbf{ Coexistence Curves for $d=6$, $b_{1}=1$ , $b_{2}=0.5$, $q=0$ black holes.} The $p-t$ phase transition diagram displays the novel triple point. The large/small BH transition terminates at the critical point while the small/radiation coexistence line extends up to the maximum pressure.
}
\label{ptnoveltrip}
\end{figure}
While the above conditions provide the information needed to know if a triple point is possible, we can also gain information on this from the $g-t$ diagram. In Figure~\ref{gtcompare} we see that moving rightward from the cusp on the lower branch, there is a discontinuity in the first derivative of $g$, indicative of a small/large critical point. If this point is above the $g=0$ axis no triple point will occur, whereas if it is below then we find the novel triple point. If this point intersects the $g=0$ axis then the novel triple point merges with the small/large critical point.
\begin{figure}[H]
\centering
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/gttripcondition1.pdf}
\caption{$b_{2}=0.7$}
\end{subfigure}
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{6DUncharged/gttripcondition2.pdf}
\caption{$b_{2}=0.4$}
\end{subfigure}
\caption{Comparing the two $g-t$ diagrams with two different values of $b_{2}$ and showing the position of the sharp "Corner" being above and below the $g=0$ axis.}
\label{gtcompare}
\end{figure}
Turning to $b_1 < 0$, we must have either $b_2<0$ or $b_2 \geq 9b_1^2$
in order to have a real and positive critical volume.
However the latter case yields a negative critical temperature, and so critical behaviour can take place only for $b_2<0$. The only phase behaviour we observe in the range $9b_1^2> b_2 > 0$ is that of a Hawking-Page transition.
We close by noting that throughout this subsection we have chosen
$b_{2}-b_{1} \leq 0$. Hence none of the black holes we consider will possess a vacuum singularity outside the origin.
\subsubsection{Charged ELBHs}
Including charge, the equation of state is now
\begin{equation}
p=\frac{t}{v}-\frac{3 b_{1}}{\pi v^{2}}+\frac{2 t b_{1}}{v^{3}}+\frac{q^{2}}{v^{8}}-\frac{b_{2}}{\pi v^{4}}
\end{equation}
with Gibbs free energy
\begin{equation}
g=-\frac{\left(\frac{1}{4} v^{4}+b_{1} v^{2}\right)\left(\pi p v^{2}+3 b_{1}+\frac{b_{2}}{v^{2}}\right)}{4 \pi v\left(1+\frac{2 b_{1}}{v^{2}}\right)}+\frac{\frac{1}{5} \pi p v^{5}+b_{1} v^{3}+b_{2} v}{4 \pi}+\frac{q^{2}\left(14 v^{2}+40 b_{1}\right)}{96\left(v^{2}+2 b_{1}\right) v^{3}}\; .
\end{equation}
The critical temperature and critical volume relations \eqref{critt} and \eqref{critv} are now
\begin{equation}\label{ctemp6}
t_{c}=\frac{6 b_{1} v_{c}^{6}+4 b_{2} v_{c}^{4}-8 \pi q^{2}}{\pi v_{c}^{5}\left(v_{c}^{2}+6 b_{1}\right)}
\end{equation}
\begin{equation}\label{cvol6}
6 v_{c}^{8} b_{1}-\left(36 b_{1}^{2}-12 b_{2}\right) v_{c}^{6}+24 v_{c}^{4} b_{1} b_{2}+\left(-56 \pi v_{c}^{2}-240 \pi b_{1}\right) q^{2}=0.
\end{equation}
This latter equation is a quartic polynomial in $v^2_c$; analytic solutions can be obtained, but they are cumbersome and so we will not display them.
If $b_{1}=b_{2}=1$, which corresponds the constant curvature case \cite{Frassino_2014}, standard Van der Waals behaviour is observed, with a large/small first order transition occurring for $q>0.1$. This was seen in the 5 dimensional case as well, so we shall not display any phase diagrams for this case. Only one critical point is present for $q \geq 0.1$.
More interesting behaviour occurs for values of $q < 0.1$. We find that more then one critical volume/temperature/pressure is possible, leading to the existence of triple points, previously observed for charged black holes in $d=6$ Lovelock gravity \cite{Frassino_2014}.
To see what happens to the triple point if the horizon curvature is not constant, it is useful to rewrite the critical volume equation \eqref{cvol6} as
\begin{equation}
w_1 \equiv \left(56 \pi v_{c}^{2}+240 \pi b_{1}\right) q^{2}=6 v_{c}^{8} b_{1}-\left(36 b_{1}^{2}-12 b_{2}\right) v_{c}^{6}+24 v_{c}^{4} b_{1} b_{2} \equiv w_2
\end{equation}
and search for the intersection points. A necessary condition for a triple point to occur is that
there are three intersection points for $v_c >0$. This will indeed occur as long as the signs of the coefficients in $w_2$ alternate -- the rule of signs then indicates there are three positive real roots for $v_c^2$ (and hence for $v_c$), in turn implying two distinct regions where $w_2>0$ for $v_c>0$, one of which has a maximum. Since $w_1$ is a quadratic in $v_c$ with coefficient $q^2$, there will be three intersection points for sufficiently small $q>0$ and $b_1>0$. An example is given in
Figure~\ref{wplotconstq} for
$b_{1}=0.8$ and $b_{2}=0.5$. The emergence of the triple point for increasing pressure is shown in Figure~\ref{mytriplepoint}, with the phase diagram given in Figure~\ref{ptmytrip}.
\begin{figure}[H]
\centering
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{6DCharged/wplotq01.pdf}
\caption{q=0.01}
\end{subfigure}
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{6DCharged/wplotq02.pdf}
\caption{q=0.02}
\end{subfigure}
\caption{Plots of $w_{1}$(red) and $w_{2}$ (blue) in 6 dimensions for constant charge, showing three intersection between the two functions, giving the possibility of a triple point. Here $b_{1}=0.8$ and $b_{2}=0.5$.
}
\label{wplotconstq}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{6DCharged/pv3plotstriple2.pdf}
\end{subfigure}
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{6DCharged/gttriple1.pdf}
\caption{p=0.06687}
\end{subfigure}
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{6DCharged/gttriple2.pdf}
\caption{p=0.06127}
\end{subfigure}
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{6DCharged/gttriple3.pdf}
\caption{p=0.05987}
\end{subfigure}
\caption{{\bf Phase Behaviour for $d=6$, $q=0.02$,$b_{1}=0.8$,$b_{2}=0.5$ black holes.} {Top:} The $p-v$ diagram for three constant temperature slices located around the tri-critical temperature. We can see in the blue line 2 oscillations, similar to the constant curvature case. {Bottom:} Three $g-t$ plots for constant pressure slices are shown. As we decrease the pressure from left to right we see the presence of two swallow tails, which eventually intersect, then separate again.}
\label{mytriplepoint}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.5\textwidth]{6DCharged/pt6dchargedfinal.pdf}
\caption{{\bf Coexistence Curves for $d=6$, $b_{1}=0.8$, $b_{2}=0.5$,$q=0.02$ black holes.} We observe the triple point at the intersection of the curves. }
\label{ptmytrip}
\end{figure}
We find that for a given value of $b_1$, the triple point in Figure~\ref{ptmytrip} moves to the right as $b_2$ increases. For sufficiently large $b_2$, the triple point merges with the large/intermediate critical point. Conversely, fixing $b_2$ and increasing $b_1$ moves the triple point to the left. This illustrates how changing the horizon geometry of exotic black holes modifies their phase behaviour.
\section{Summary \& Conclusion}
Our investigation of thermodynamic behaviour for exotic black holes has uncovered a number of interesting results.
The most interesting is that of a novel triple point
between thermal AdS (radiation), and uncharged large and small black holes in 6 dimensions. This phase behaviour was overlooked in previous studies \cite{Frassino_2014}, and arises as a consequence of the exotic geometry of the horizon. We likewise observe a range of large/intermediate/small black hole
triple point behaviour in the charged case in $d=6$ as we adjust the parameters of the horizon geometry.
Another interesting result is the generalizations \eqref{vacsol} of massless topological black holes in Einstein gravity
\cite{Mann:1997jb,Aminneborg:1996iz,Smith:1997wx}. For these exotic Gauss-Bonnet black holes two horizons are possible, yielding a richer set of possibilities warranting further study. Negative mass solutions generalizing those in Einstein gravity \cite{Mann:1997jb,Smith:1997wx} are also possible. We leave a more detailed study of these object for future investigations.
A study of 3rd order Lovelock gravity, with the possibility of finding a quadruple point, would be interesting. There are two possibilities for a quadruple point. One is that of a novel uncharged quadruple point where we have two swallowtails intersecting each other on the $g=0$ axis, giving large/intermediate/small/radiation coexistence point. Another would be that in the charged case, where four black holes of distinct size merge at a single point in the phase diagram.
More ambitious endeavours include promoting the topological parameter to a thermodynamic variable itself, generalizations to to de Sitter spacetime, and obtaining rotating solutions. Work on these areas is in progress.
\section*{Acknowledgements}
This work was supported in part by the Natural Sciences and Engineering Research Council of Canada. We are grateful to Sourya Ray for a number of helpful discussions.
\section*{Note added}
As we were completing this paper we became aware of a similar study in 3rd order Lovelock gravity \cite{Farhangkhah:2021tzq}. This study considers only uncharged black holes. We do not agree with some of their findings, particularly the failure to notice the presence of Hawking-Page transitions when relevant.
\bibliographystyle{unsrt}
\addcontentsline{toc}{section}{References}
|
\section{Introduction}
The $B_c$ meson, first observed by the CDF Collaboration \cite{Abe:1998wi}, is interesting since it has the structure of the heavy quarkonium but it decays weakly. Therefore, this meson is well suited to study both quarkonium and weak interaction features within the same hadronic system. As for weak interactions, in addition to the purely leptonic mode which proceeds through the weak annihilation of the constituent quarks, the $B_c$ decays occur through the transitions of both the charm and beauty quark. The decays induced by the charm transition represent the dominant contribution to the full width despite the smaller available phase-space \cite{Colangelo:1992cx,Beneke:1996xe,Anisimov:1998uk,Kiselev:2000pp}. In our study we focus on the exclusive semileptonic modes
$B^+_c \to B_{s,d} \bar \ell \nu_\ell$ and $B^+_c \to B^*_{s,d} \bar \ell \nu_\ell$ induced at the quark level by
$c \to (s,d) \bar \ell \nu_\ell$, with $\ell=e, \mu$ (the tauonic mode is phase-space forbidden). There are various reasons for such a choice.
The first one is the possibility of exploiting the heavy quark spin symmetry \cite{Jenkins:1992nb}, which allows us to relate the observables in the modes with final pseudoscalar and vector meson, as well as the different observables in the vector channel. The relatively small phase-space justifies the extrapolation to the full kinematical range of the spin symmetry relations, that strictly hold close to the zero-recoil point where the produced meson is at rest in the $B_c$ rest frame \cite{Colangelo:1999zn}. Invoking the heavy quark spin symmetry the relevant hadronic matrix elements can be expressed in terms of two independent functions, that can be derived from the $B_c \to B_s$ and $B_c \to B_d$ form factors (FF) precisely determined by lattice QCD \cite{Cooper:2020wnj}.
The second reason is the possibility to scrutinize the sensitivity of such processes to beyond the Standard Model (BSM) effects of the kind emerging in $B$ decays, where hints of violation of lepton flavour universality (LFU) are found \footnote{For recent overviews see \cite{Bifani:2018zmi,Gambino:2020jvv}.}. The measurement of ${\cal B}(B_c \to J/\psi \bar \tau \nu_\tau)$ is also important in this regard \cite{Aaij:2017tyk}.
Such effects can be analyzed in an effective theory framework extending the low-energy SM Hamiltonian that governs the $c \to (s,d) \bar \ell \nu_\ell$ transitions with the inclusion of the full set of semileptonic dimension-$6$ operators with lepton flavour dependent Wilson coefficients.
The impact of the new operators on the experimental $B_c$ observables can be assessed. The $D$ and $D_s$ semileptonic decay modes have been recently studied in this context, and the Wilson coefficients of the new operators in the extended Hamiltonian have been constrained using the available experimental data \cite{Fajfer:2015ixa,Fleischer:2019wlx,Fuentes-Martin:2020lea,Leng:2020fei,Becirevic:2020rzi}. The study of
the sensitivity of this class of $B_c$ decays to extensions of the Standard Model (the New Physics - NP) is timely, as these channels are accessible at the present facilities. The hadronic matrix elements of the new operators can also be given in terms of the same independent functions entering in the SM ones, invoking the heavy quark spin symmetry. Since the produced $B_s^*$ and $B_d^*$ mesons decay radiatively, we shall provide the expressions of the fully differential $B_c ^+\to B_{s,d}^*(\to B_{s,d}\, \gamma) \bar \ell \nu_\ell$ decay distribution for the extended low-energy Hamiltonian: such general expressions can also be used for different processes.
In Sec. \ref{sec:hamil} we introduce the effective semileptonic Hamiltonian comprising the full set of dimension-6 operators with left-handed neutrinos, that generalizes the SM low-energy Hamiltonian. In Sec. \ref{sec:decay} we provide the decay distributions of $B_c \to B_{s,d} \bar \ell \nu_\ell$ and $B_c \to B^*_{s,d} (\to B_{s, d} \gamma) \bar \ell \nu_\ell$ obtained from the extended Hamiltonian. In Sec. \ref{HQspin} we discuss the heavy quark spin symmetry relations connecting the SM and NP operator matrix elements.
Sec. \ref{sec:num} contains the numerical analysis in SM and a discussion of the effects of the new operators on the $B_c$ decay observables. The summary and the outlook are presented in the last section. The appendices contain the relations among the hadronic form factors obtained by the heavy quark spin symmetry (Appendix \ref{app:A}), and the coefficient functions of the full angular distribution of the four-body radiative modes $B_c \to B^*_{s,d} (\to B_{s,d} \,\gamma) \bar \ell \nu_\ell$ (Appendix \ref{app:B}).
\section{ Effective $c \to s,d$ semileptonic Hamiltonian}\label{sec:hamil}
We consider the low-energy Hamiltonian comprising the full set of dimension-$6$ semileptonic $Q \to q$ operators with left-handed neutrinos:
\begin{eqnarray}
&&H_{\rm eff}^{Q \to q \bar \ell \nu}= \frac{G_F}{\sqrt{2}} V_{CKM}\nonumber\\
&&\quad \quad \Big[(1+\epsilon_V^\ell) \left({\bar q} \gamma_\mu (1-\gamma_5) Q \right)\left( \bar \nu_{\ell } (1+\gamma_5) \gamma^\mu \ell \right)\nonumber \\
&&\quad \quad + \epsilon_R^\ell \left({\bar q} \gamma_\mu (1+\gamma_5) Q \right)\left( \bar \nu_{\ell} (1+\gamma_5)\gamma^\mu \ell \right) \nonumber \\
&&\quad \quad + \epsilon_S^\ell \, ({\bar q} Q) \left( {\bar \nu_\ell} (1+\gamma_5) \ell \right) \label{hamil} \\
&&\quad \quad + \epsilon_P^\ell \, \left({\bar q} \gamma_5 Q\right) \left({\bar \nu_\ell} (1+\gamma_5) \ell \right) \nonumber \\
&&\quad \quad + \epsilon_T^\ell \, \left({\bar q} (1+\gamma_5) \sigma_{\mu \nu} Q\right) \,\left( {\bar \nu_\ell} (1+\gamma_5) \sigma^{\mu \nu} \ell \right) \Big] , \nonumber
\end{eqnarray}
with $Q=c$, and $q$ either the $s$ or the $d$ quark. $V_{CKM}$ is the Cabibbo-Kobayashi-Maskawa (CKM) matrix element $V_{cs}$ or $V_{cd}$.
In addition to the SM operator
${\cal O}_{SM}=4 (\bar q_L \gamma^\mu Q_L) \left( {\bar \nu_{\ell L}} \gamma_\mu \ell_L\right)$
and to the operators
${\cal O}_S=\left({\bar q} Q \right)\left( {\bar \nu_\ell} (1+\gamma_5) \ell \right)$,
${\cal O}_P=\left({\bar q} \gamma_5 Q \right)\left( {\bar \nu_\ell} (1+\gamma_5) \ell \right)$ and
${\cal O}_T=\left({\bar q} (1+\gamma_5)\sigma_{\mu \nu} Q \right)\left( {\bar \nu_\ell} (1+\gamma_5) \sigma^{\mu \nu} \ell \right)$, the operator
${\cal O}_{R}=4 (\bar q_R \gamma^\mu Q_R) \left( {\bar \nu_{\ell L}} \gamma_\mu \ell_L\right)$
is included in Eq.~\eqref{hamil}. It is worth remarking that in the Standard Model Effective Field Theory the only
dimension-$6$ operator with the right-handed quark current is nonlinear in the Higgs field \cite{Buchmuller:1985jz,Cirigliano:2009wk,Aebischer:2020lsx}, and its role has been the subject of several discussions \cite{Bernard:2006gy,Crivellin:2009sd,Crivellin:2014zpa,Alioli:2017ces,Aebischer:2018iyb,Aebischer:2020lsx}.
The complex coefficients $\epsilon^\ell_{V,R,S,P,T}$ in the
low-energy Hamiltonian \eqref{hamil} are lepton-flavour dependent.
Generalized Hamiltonians as in Eq.~\eqref{hamil} have been studied for $b \to c$ transitions in connection with the anomalies in semileptonic $B \to D^{(*)} \tau \nu_\tau$ decays, obtaining information on the various operators \cite{Biancofiore:2013ki,Becirevic:2016hea,Alonso:2016oyd,Colangelo:2016ymy,Jung:2018lfu,Colangelo:2018cnj,Murgui:2019czp,Alguero:2020ukk}. Modes induced by the $b \to u$ induced transition have also been analyzed in such an effective theory approach \cite{Colangelo:2019axi}. For both classes of $b$-quark transitions, suitable observables testing the Standard Model and challenging LFU have been identified. Observables in baryon decays, in particular in inclusive modes, have also been studied
\cite{Colangelo:2020vhu}. Here we focus on the $B_c$ decays governed by the Hamiltonian \eqref{hamil}, to study the SM phenomenology and to assess the sensitivity of such channels to deviations from the SM.
\section{ Modes $B_c \to P \,\bar \ell \nu_\ell$ and $B_c \to V(\to P\gamma) \, \bar \ell \nu_\ell$ }\label{sec:decay}
The $q^2$ distribution of the $B_c \to P \bar \ell \nu_\ell$ decay, with $P$ a pseudoscalar meson, governed by the low-energy Hamiltonian (\ref{hamil}) reads:
\begin{eqnarray}
&&
\frac{d \Gamma(B_c \to P \bar \ell {\nu}_\ell)}{d q^2}=\nonumber \\
&&\frac{G_F^2 |V_{CKM}|^2 \lambda^{1/2}}{128 \, m_{B_c}^3 \pi^3 q^2 } \left( 1 - \frac{m_\ell^2}{q^2} \right)^2 \nonumber \\
&&
\Bigg\{ \left| m_\ell (1 + \epsilon_V^\ell+ \epsilon_R^\ell) + \frac{q^2 \epsilon_S^\ell}{m_Q-m_q} \right|^2 (m_{B_c}^2 - m_P^2)^2 f_0^2(q^2) \nonumber \\
&&+ \lambda
\Bigg[ \frac{1}{3} \left| m_\ell (1 + \epsilon_V^\ell+ \epsilon_R^\ell) f_+(q^2) + \frac{4 q^2}{m_{B_c}+m_P} \epsilon_T^\ell f_T(q^2) \right|^2 \nonumber \\
&&+ \frac{2 q^2}{3} \left| (1 + \epsilon_V^\ell+ \epsilon_R^\ell) f_+(q^2) +4 \frac{m_\ell}{m_{B_c}+m_P} \epsilon_T^\ell f_T(q^2) \right|^2 \Bigg] \Bigg\} . \nonumber \\ \label{GamBctoP}
\end{eqnarray}
$G_F$ is the Fermi constant, $q^2$ the squared momentum transferred to the lepton pair and $\lambda=\lambda(m_{B_c}^2,m_P^2,q^2)$ is the triangular function. The form factors $f_{+}$, $f_{0}$ and $f_{T}$ are defined in Appendix \ref{app:A}. The SM expression is recovered setting to zero all couplings $\epsilon_i^\ell$.
In the case of a final vector meson $V$ decaying to $P \gamma$, namely $B^*_{s,d}$, the four-body kinematics of $B_c \to V(\to P\gamma) \,\bar \ell \nu_\ell$ is shown in Fig.~\ref{fig:piani}.
\begin{figure}[b!]
\begin{center}
\includegraphics[width = 0.55\textwidth]{piani.pdf}
\vspace*{-1cm}
\caption{ \baselineskip 10pt \small Kinematics of the $B_c \to B_s^* (B_s \gamma) \bar \ell \nu_\ell$ decay. }\label{fig:piani}
\end{center}
\end{figure}
The fully differential decay width is expressed in terms of $q^2$ and of the angles $\theta_V$, $\theta$ and $\phi$ defined in the figure:
\begin{eqnarray}
&&\frac{d^4 \Gamma ( B_c \to V(\to P\gamma) \bar \ell \nu_\ell)}{dq^2 \,d\cos \theta_V \,d\cos \theta \,d\phi} ={\cal N_\gamma}|{\vec p}_{V}| \left(1- \frac{ m_\ell^2}{q^2}\right)^2 \nonumber \\
&&\quad \quad \Big\{I_{1s} \,\sin^2 \theta_V+I_{1c} \,(3+\cos 2\theta_V )\nonumber \\
&&\quad \quad + (I_{2s} \,\sin^2 \theta_V+I_{2c} \,(3+\cos 2\theta_V )) \cos 2\theta \nonumber \\
&&\quad \quad +I_3 \,\sin^2 \theta_V \sin^2 \theta \cos 2 \phi +I_4 \sin 2 \theta_V \sin 2\theta \cos \phi \nonumber \\
&&\quad \quad +I_5 \, \sin 2 \theta_V \sin \theta \cos \phi \label{angulargamma}\\
&&\quad \quad +(I_{6s} \, \sin^2 \theta_V+I_{6c} \,(3+\cos 2\theta_V ))\cos \theta
\nonumber \\
&&\quad \quad + I_7 \sin 2 \theta_V \sin \theta \sin \phi+I_8\sin 2 \theta_V \sin 2\theta \sin \phi \nonumber \\
&&\quad \quad +I_9 \,\sin^2 \theta_V \sin^2 \theta \sin 2 \phi \Big\}, \nonumber
\end{eqnarray}
with $|{\vec p}_{V}|=\sqrt{\lambda(m_{B_c}^2,m_V^2,q^2)}/2 m_{B_c}$.
The distribution \eqref{angulargamma} is obtained in the narrow width approximation for the meson $V$, and the factor
${\cal N}_\gamma=\displaystyle{\frac{3G_F^2 |V_{CKM}|^2 {\cal B}(V \to P \gamma)}{128(2\pi)^4m_{B_c}^2}}$
comprises the $V \to P \gamma$ branching fraction.
The angular coefficient functions $I_i(q^2)$ encode the dynamics and the SM and of NP
described by the Hamiltonian \eqref{hamil}.
We provide them for the full set of operators, generalizing the results obtained in \cite{Colangelo:2018cnj} for the tensor operator:
\begin{eqnarray}
I_i &=& |1+\epsilon_V|^2 \,I_i^{SM}+|\epsilon_R|^2I_i^{NP,R}+|\epsilon_P|^2I_i^{NP,P} \hspace*{1cm} \nonumber \\
&+&|\epsilon_T|^2I_i^{NP,T} +2 \, {\rm Re}\left[\epsilon_R(1+\epsilon_V^* )\right] I_i^{INT,R} \nonumber \\
&+&2 \, {\rm Re}\left[\epsilon_P(1+\epsilon_V^* )\right] I_i^{INT,P} \nonumber \\
&+&2 \, {\rm Re}\left[\epsilon_T(1+\epsilon_V^* )\right] I_i^{INT,T} \label{eq:Iang1} \\
&+&2 \, {\rm Re}\left[\epsilon_R \epsilon_T^* \right] I_i^{INT,RT}
+2 \, {\rm Re}\left[\epsilon_P \epsilon_T^* \right] I_i^{INT,PT}\nonumber \\
&+&2 \, {\rm Re}\left[\epsilon_P \epsilon_R^* \right] I_i^{INT,PR} \nonumber
\end{eqnarray}
for $ i=1,\dots 6$,
\begin{eqnarray}
I_7 &=&2 \, {\rm Im}\left[\epsilon_R(1+\epsilon_V^* )\right] I_7^{INT,R} \nonumber \\
&+&2 \, {\rm Im}\left[\epsilon_P(1+\epsilon_V^* )\right] I_7^{INT,P}\nonumber \\
&+&2 \, {\rm Im}\left[\epsilon_T(1+\epsilon_V^* )\right] I_7^{INT,T} \label{eq:Iang2} \\
&+&2 \, {\rm Im}\left[\epsilon_R \epsilon_T^* \right] I_7^{INT,RT}+2 \, {\rm Im}\left[\epsilon_P \epsilon_T^* \right] I_7^{INT,PT} \quad \quad \nonumber \\
&+&2 \, {\rm Im}\left[\epsilon_P \epsilon_R^* \right] I_7^{INT,PR} , \nonumber
\end{eqnarray}
and
\begin{equation}
I_i=2 \, {\rm Im}\left[\epsilon_R (1+\epsilon_V^* ) \right] I_i^{INT,R}
\label{eq:Iang3} \end{equation}
for $ i=8,\,9$.
In SM the angular coefficient functions are given in terms of the helicity amplitudes
\begin{eqnarray}
H_0 &=&\frac{1}{{2m_V(m_{B_c}+m_V) \sqrt{q^2}}} \nonumber \\
&& \Big( (m_{B_c}+m_V)^2(m_{B_c}^2-m_V^2-q^2) A_1(q^2)\nonumber \\
&-&\lambda(m_{B_c}^2,\,m_V^2,\,q^2) A_2(q^2) \Big) \label{HampV} \\
H_\pm&=& \frac{(m_{B_c}+m_V)^2 A_1(q^2)\mp\sqrt{\lambda(m_{B_c}^2,\,m_V^2,\,q^2)}V(q^2)}{m_{B_c}+m_V} \nonumber \\
H_t&=& -\frac{\sqrt{\lambda(m_{B_c}^2,\,m_V^2,\,q^2)}}{\sqrt{q^2}} \,A_0(q^2) . \,\,\, \nonumber
\end{eqnarray}
For the NP operators the following amplitudes are also introduced:
\begin{eqnarray}
H_\pm^{NP} &=&
\frac{1}{\sqrt{q^2}}\Big\{\Big(m_{B_c}^2-m_V^2 \pm \sqrt{\lambda(m_{B_c}^2,m_V^2,q^2)} \Big)(T_1+ T_2) \nonumber \\
&+&q^2(T_1- T_2)\Big\} \nonumber \\
H_L^{NP}&=& 4\Big\{
\frac{\lambda (m_{B_c}^2,m_V^2,q^2)}{m_V(m_{B_c}+m_V)^2} \, T_0+2\frac{m_{B_c}^2+m_V^2-q^2}{m_V} T_1\nonumber \\
&+&4 m_V T_2\Big\} . \label{HampNP}
\end{eqnarray}
The form factors $V$, $A_i$ and $T_i$ are defined in Appendix \ref{app:A}. The coefficient functions in Eqs.~\eqref{eq:Iang1}, \eqref{eq:Iang2} and \eqref{eq:Iang3}, expressed in terms of the amplitudes \eqref{HampV} and \eqref{HampNP}, are collected in Appendix \ref{app:B}. With such expressions the various observables can be computed by suitable integrations of the distribution in Eq.~\eqref{angulargamma}.
\section{Heavy quark spin symmetry and relations among form factors }\label{HQspin}
In the infinite heavy quark mass limit $m_Q \gg \Lambda_{\rm QCD}$ the QCD Lagrangian exhibits a heavy quark (HQ) spin symmetry, with the decoupling of the heavy quark spin from gluons \cite{Neubert:1993mb}. This produces the decoupling of the spins of the heavy quarks in $B_c$: the spin-spin interaction vanishes in this limit. Important consequences of the HQ spin symmetry are the relations among the form factors parametrizing the weak current matrix elements of $B_c$ and mesons comprising a single heavy quark ($B_s^{(*)}, B_d^{(*)}, D^{(*)},\dots$) or two heavy quarks ($\eta_c, J/\psi, \psi(2S), \dots$)
\cite{Jenkins:1992nb}.
In the semileptonic $B_c \to B_a^{(*)}$ ($a=s,d$) decays induced by the $c \to s,d$ transition, since $m_c \ll m_b$ the energy released to the final hadronic system is much smaller than $m_b$. The $b$ quark remains almost unaffected, so that the final meson keeps the same $B_c$ four-velocity $v$.
Denoting the initial and final meson four-momenta as
$p=m_{B_c} v$ and $p^\prime=m_{B_a} v^\prime =m_{B_a} v +k$ , with $k$ a small residual momentum, the four-momentum transferred to the leptons is $q=p-p^\prime=(m_{B_c}-m_{B_a})v-k$, with $v \cdot k={\cal O}(1/m_b)$.
The relations stemming from the HQ spin symmetry can be worked out using the trace formalism \cite{Falk:1990yz}. The heavy pseudoscalar and vector mesons are collected in doublets, the two components of which represent states differing only for the orientation of the heavy quark spins.
The $B_c^+$ and $B_c^{*+}$ doublet comprising the heavy $c$ and $ \bar b$ quarks is described by the effective fields
\begin{equation}
H^{c \bar b}= \frac{1+\spur{v}}{2} \left[B_c^{*\mu} \gamma_\mu - B_c \gamma_5 \right] \frac{1-\spur{v}}{2} .\label{bc-doublet}
\end{equation}
The $B_a$ and $B_a^*$ doublet ($a$ an $SU(3)_F$ index) with the single heavy antiquark $\bar b$ is described by the effective fields
\begin{equation}
H^{\bar b}= \left[B_a^{*\mu} \gamma_\mu - B_a \gamma_5 \right] \frac{1-\spur{v}}{2} . \label{qb-doublet}
\end{equation}
$B$ and $B^*$ are operators that include a factor $\sqrt{m_B}$ and $\sqrt{m_B^*}$ and have dimension $3/2$. The equations
${\spur v} H^{c \bar b}=H^{c \bar b}$, $\ H^{c \bar b} {\spur v}=-H^{c \bar b}$, ${\spur v} H^{\bar b}=H^{\bar b}$, $\ H^{ \bar b} {\spur v}=-H^{ \bar b}$ are satisfied.
Under the heavy quark spin transformations and light quark $SU(3)_F$ transformations the doublets transform as
\begin{eqnarray}
H^{c \bar b} \to S_c H^{c \bar b} S^\dagger_b \nonumber \\
H^{\bar b}_a \to (U H^{\bar b})_a S^\dagger_b . \label{eq:transf}
\end{eqnarray}
The matrix elements of the quark current ${\bar q} \Gamma Q$ between
$B_c$ and $B_a^{(*)}$, with $\Gamma$ a generic product of Dirac matrices, can be written as
\begin{eqnarray}
&&
\langle B_a^{(*)}(v,k)| {\bar q} \Gamma Q| { B_c}(v) \rangle =\nonumber \\
&&
-\sqrt{ m_{B_c} m_{B_a}} \, {\rm Tr} \left[ \overline H_a^{(\bar b)} \Omega_a(v, a_0 k) \Gamma H^{(c \bar b)}\right] , \label{omega}
\end{eqnarray}
with $\overline H_a=\gamma^0 H_a^\dagger \gamma^0$ and
are invariant under rotations of the $\bar b$ spin. The most general matrix depending on $v$ and $k$ is
\begin{equation}
\Omega_a(v, a_0 k)=\Omega_{1a}+ \spur{k} a_0\Omega_{2a} . \label{omega1}
\end{equation}
It involves two dimensionless nonperturbative functions, the form factors $\Omega_{1a}$ and $\Omega_{2a}$. The dimensionful parameter $a_0$ can be identified with the length scale of the process, typically the Bohr radius of the mesons.
At odds with the weak matrix elements of mesons comprising a single heavy quark, that are expressed in terms of a single universal function (the Isgur-Wise function \cite{Isgur:1989vq,Isgur:1989ed}) normalized to 1 at the zero-recoil point $v \cdot v^\prime =1$ due to the heavy quark flavour symmetry, no normalization is fixed for $\Omega_{1}$ and $\Omega_{2}$. Such form factors encode the QCD dynamics and must be determined by nonperturbative methods.
The SM matrix elements relevant for $B_c^+ \to B_a \ell^+ {\nu}_\ell$ involve the form factors $f_+^{B_c \to P_a}$ and $f_0^{B_c \to P_a}$ defined in (\ref{BctoP}). On the other hand, four form factors are needed in SM for each $B_c \to B_a^* \ell {\bar \nu}_\ell$ mode, $V^{B_c \to V_a}$ and $A_{1,2,0}^{B_c \to V_a}$ defined in (\ref{BctoV}). They parametrize the
hadronic matrix elements of the SM operator in the low-energy Hamiltonian (\ref{hamil}). The matrix elements of the operators with a scalar and pseudoscalar quark current in Eq.~(\ref{hamil}) do not involve new form factors: the scalar operator contributes only to $B_c \to B_a \ell {\bar \nu}_\ell$ and its hadronic matrix element is given in terms of $f_0^{B_c \to B_a}$ and of the masses of the quarks involved in the transitions. The pseudoscalar operator contributes only to $B_c \to B_a^* \ell {\bar \nu}_\ell$ and its matrix element can be expressed in terms of $A_0^{B_c \to B_a^*}$ and the quark masses (Appendix \ref{app:A}). The matrix elements of the tensor operator in (\ref{hamil}) require the form factors $f_T^{B_c \to B_a}$ for $B_c^+ \to B_a \ell^+ {\nu}_\ell$ and $T_{1,2,0}^{B_c \to B_a^*}$ for $B_c^+ \to B_a^* \ell^+ {\nu}_\ell$ defined in Appendix \ref{app:A}.
Exploiting the HQ spin symmetry all the form factors $f_+,\,f_0,\,f_T$ and $V,A_i,T_i$ can be given in terms of the functions $\Omega_{1,2}$ in \eqref{omega1}.
Such relations can be inverted to express $\Omega_1$ and $\Omega_2$ in terms of $f_+$ and $f_0$,
Eq.~\eqref{eq:om12}, and can be used once such functions are determined in a nonperturbative way. All relations are in Appendix \ref{app:A}.
The result is that $f_+$ and $f_0$, accompanied with the relations from the HQ spin symmetry, provide enough information to study the full phenomenology of the $B_c \to B^{(*)}_a$ semileptonic modes in SM and beyond.
The relations among the form factors are valid close to the zero-recoil point, at maximum momentum squared transferred to the lepton pair $q^2_{max}=(m_{B_c}-m_{B_a^{(*)}})^2$. However, since the phase space for $B_c \to B_a^{(*)}$ is small, such relations can be extrapolated to the full kinematical $q^2$ range. The assumption can be checked once other form factors are available, by a comparison with the expressions in the heavy quark limit.
\section{Numerical analysis}\label{sec:num}
We describe several observables in $B_c^+ \to B_{s,d} \ell^+ \nu_\ell$ and $B_c^+ \to B_{s,d}^*(\to B_{s,d} \gamma) \ell^+ \nu_\ell$ in the Standard Model. We also study their sensitivity
to the BSM operators in the low-energy Hamiltonian.
For the hadronic matrix elements of the various operators in Eq.~\eqref{hamil} we exploit the HQ spin symmetry and express all form factors in terms of the universal functions $\Omega_{1s(d)}$ and
$\Omega_{2s(d)}$ using the relations in Appendix \ref{app:A}. $\Omega_{1s(d)}$ and
$\Omega_{2s(d)}$ are determined from the form factors
$f_{+,0}^{B_c \to B_{s}}$ and $f_{+,0}^{B_c \to B_{d}}$ computed by lattice QCD in Ref.~\cite{Cooper:2020wnj}.
In such computation the form factors are evaluated in the full $q^2$ range, by a chain fit of the results obtained by a non-relativistic QCD treatment of the $b$ quark and by using the highly improved staggered quark method. The variable $t=q^2$, with kinematical bound $m_\ell^2 \le t \leq t_-=(m_{B_c}-m_{B_{s(d)}})^2$, is mapped into the variable
$z(t)=\frac{\sqrt{t_+-t}-\sqrt t_+}{\sqrt{t_+-t}+\sqrt t_+}$ with $t_+=(m_{B_c}+m_{B_{s(d)}})^2$ chosen to be larger than the lowest threshold for hadron production in the $t$ channel, the $D K$ and $D \pi$ threshold. To optimize the
calculation, a rescaled variable
$z_p(t)=z(t)/z(M^2_{res})$ is defined, with $M_{res}$ a suitably chosen mass parameter. Each form factor $f(t)$ is expressed (in the continuum limit of the lattice discretization) as a truncated power series of $z_p$:
\begin{equation}
f(t)=P(t) \sum_n^N A_n z_p(t)^n \,\, ,
\end{equation}
with $P(t)$ a function chosen to describe the main computed $t$-dependence. As a result, each form factor is determined by the set of coefficients $A_n$ together with their errors and error correlation matrices. The functions $\Omega_1(y)$ and $a_0 \Omega_2(y)$ obtained for the $c \to s$ and $c \to d$ transitions are depicted in Fig.~\ref{fig:omega} together with their uncertainties. They are expressed in terms of the variable $y=\displaystyle\frac{p \cdot p^\prime}{m_{B_c}m_{B_a}}=\frac{m_{B_c}^2+m_{B_a}^2-q^2}{2 m_{B_c}m_{B_a}}$ in the range $[1,y_{max}]$, with $y_{max}$ corresponding to $q^2_{min}=m_\ell^2$. The numerical values of the other parameters, taken from the Particle Data Group \cite{Zyla:2020zbs}, are listed in Table \ref{tab:par}.
\begin{widetext}
\begin{center}
\begin{figure}[t!]
\begin{center}
\includegraphics[width = 0.8\textwidth]{OmegaAll.pdf}
\caption{ \baselineskip 10pt \small Universal functions $\Omega_{1}(y)$ (top) and $a_0 \Omega_{2}(y)$ (bottom panels) obtained using Eq.~\eqref{eq:om12} and the form factors $f_+$ and $f_0$ computed in Ref.~\cite{Cooper:2020wnj} for $B_c \to B_s$ (left) and $B_c \to B_d$ matrix elements (right panels), with $\displaystyle y=\frac{p \cdot p^\prime}{m_{B_c}m_{B_a}}$. }\label{fig:omega}
\end{center}
\end{figure}
\end{center}
\end{widetext}
\begin{table}[t]
\caption{ \small Parameters, from Ref.~\cite{Zyla:2020zbs}.}\label{tab:par}
\vspace{0.3cm}
\centering
\begin{tabular}{cc}
\hline
$m_{B_c}$ & $6274.9\pm0.8$ MeV \\
$\tau_{B_c}$ & $(0.510\pm0.009) \times 10^{-12}$ s\\
$m_{B_s}$ & $5366.88\pm0.14$ MeV \\
$m_{B_s^*}$ & $5415.8\pm1.5$ MeV \\
${\cal B}(B_s^*\to B_s \gamma)$ &1 \\
$m_{B_d}$ & $5279.63\pm0.20$ MeV\\
$m_{B_d^*}$ &$5324.7\pm0.21$ MeV \\
${\cal B}(B_d^*\to B_d \gamma)$ &1 \\
$|V_{cs}|$ & $0.987\pm0.011$ \\
$|V_{cd}|$ & $0.221\pm0.004$ \\
$m_d^{\overline{MS}}(2 \, GeV)$ & $4.67^{+0.48}_{-0.17}$ MeV \\
$m_s^{\overline{MS}}(2 \, GeV)$ & $93^{+11}_{-5}$ MeV \\
$m_c$ & $1.67\pm0.07$ GeV\\
\hline
\end{tabular}
\end{table}
The analysis of the sensitivity to the BSM operators in Eq.~(\ref{hamil}) requires a set of input values for the coefficients $\epsilon_i^\ell$. There are experimental constraints, in particular from the purely leptonic $D_s$ and $D^+$ decay widths, from the semileptonic $D^{0(+)}$ decays to $K^{-(0)},K^{*-(0)}$ and $\pi^{-(0)},\rho^{-(0)}$, and from the semileptonic $D_s \to \phi$ transitions \cite{Fajfer:2015ixa,Fleischer:2019wlx,Leng:2020fei,Becirevic:2020rzi}. Ranges of values have been determined upon the assumption that all $\epsilon_i^\ell$ are real \cite{Becirevic:2020rzi}: $\epsilon_V^\mu=(1.65 \pm 2.02) \times 10^{-2} $, $\epsilon_R^\mu=(-1.35 \pm 2.02) \times 10^{-2}$, $\epsilon_S^\mu =(-1.0\pm 2.0) \times 10^{-2} $, $\epsilon_P^\mu=(0.9 \pm 1.4) \times 10^{-3}$ and $\epsilon_T^\mu=(1.2 \pm 1.8) \times 10^{-2}$
for the $c \to s$ transition, and
$\epsilon_V^\mu=(5.0 \pm 2.1) \times 10^{-2} $, $\epsilon_R^\mu=(2.0 \pm 2.0) \times 10^{-2}$, $\epsilon_S^\mu=(-9.0\pm 7.0) \times 10^{-2} $, $\epsilon_P^\mu=(-2.6 \pm 1.3) \times 10^{-3}$ and $\epsilon_T^\mu=(-2.0 \pm 1.4)\times 10^{-1}$ for the $c \to d$ transition. Interestingly, the allowed range for $\epsilon_T^\mu$ in the $c \to d$ transition is wide. We vary the couplings in these intervals with the purpose of describing the effects of the various NP operators.
Assuming a hierarchy in LFU violation, all couplings for the electron operators
$\epsilon^e_{V,R,S,P,T}$ are kept to zero, hence such modes are only described in SM.
\subsection{$B_c \to B_s \ell^- \bar \nu_\ell$ and $B_c \to B_s^*(\to B_s \gamma) \ell^- \bar \nu_\ell$ }
The semileptonic $B_c$ decays induced by the $c \to s$ transition are expected to constitute the largest fraction of semileptonic modes
\cite{Colangelo:1999zn,Ivanov:2000aj,Ebert:2003wc,Kiselev:2003mp,Ivanov:2006ni,Hernandez:2006gt,Wang:2008xt,Choi:2009ai,Barik:2009zz,Dhir:2009ub,Chang:2014jca,Shi:2016gqt}.
The prediction in SM
\begin{equation}
{\cal B}(B_c^+ \to B_s \,\mu^+ \nu_\mu)=0.0125\,(4)\, \left(\frac{|V_{cs}|}{0.987}\right)^2 \,\,\label{BrBsRis}
\end{equation}
follows from the use of form factors in \cite{Cooper:2020wnj}. The quoted error refers only to the form factor uncertainties, the errors from the CKM matrix element and from the $B_c$ lifetime in Table \ref{tab:par} can be simply added, the error from the mass parameters is small. For the electron mode the result is:
\begin{equation}
{\cal B}(B_c^+ \to B_s \, e^+ \nu_e)=0.0131\,(4)\, \left(\frac{|V_{cs}|}{0.987}\right)^2 . \label{BrBsRise}
\end{equation}
In the case of $\mu$ we describe below how the branching fraction changes due to the NP operators, studying also the correlation with other observables. We notice that the $q^2$ spectrum in Fig.~\ref{fig:BrBs} is modified with respect to the Standard Model when the additional operators in \eqref{hamil} are considered. The SM prediction including the FF uncertainty is enlarged if the NP operators are considered, varying the couplings $\epsilon_i^\mu$ in their quoted ranges. However, the shape of the spectrum is unchanged.
For $B_c^+ \to B_a^* \mu^+ \nu_\mu$ ($a=s,d$),
the SM helicity amplitudes \eqref{HampV} can be expressed in terms of $\Omega_{1a}$ and $\Omega_{2a}$:
\begin{eqnarray}
H_0&=&\sqrt{\frac{m_{B_c}}{m_{B_a^*}}}\frac{(m_{B_c}^2-m_{B_a^*}^2-q^2)}{\sqrt{q^2}}\Omega_{1a}\nonumber \\
&+&\frac{\lambda(m_{B_c}^2,m_{B_a^*}^2,q^2)}{2\sqrt{m_{B_c}m_{B_a^*}q^2}}\,a_0 \Omega_{2a} \nonumber \\
H_\pm&=&\sqrt{\frac{m_{B_a^*}}{m_{B_c}}} \bigg(2 m_{B_c} \Omega_{1a} \nonumber \\
&\mp&
\lambda^{1/2}(m_{B_c}^2,m_{B_a^*}^2,q^2) \,a_0 \Omega_{2a} \bigg) \\
H_t&=&-\frac{\lambda^{1/2}(m_{B_c}^2,m_{B_a^*}^2,q^2)}{2\sqrt{m_{B_c}m_{B_a^*}q^2}}
\bigg(2m_{B_c}\Omega_{1a}\nonumber \\
&+&(m_{B_c}^2-m_{B_a^*}^2+q^2)\,a_0 \Omega_{2a} \bigg) , \nonumber
\end{eqnarray}
while the NP amplitudes \eqref{HampNP} read:
\begin{eqnarray}
H_{\pm}^{NP}&= &2 \, \sqrt{\frac{m_{B^*_a}}{m_{B_c} q^2}} \nonumber \\
&\bigg[& \bigg(m_{B_c}^2 - m_{B^*_a}^2 + q^2 \pm \sqrt{\lambda(m_{B_c}^2,m_{B^*_a}^2,q^2)} \bigg) \, \Omega_1 \nonumber \\
&+& \bigg((m_{B_c} + m_{B^*_a}) \, \big((m_{B_c} - m_{B^*_a})^2 - q^2\big) \nonumber \\
& \pm& (m_{B_c} - m_{B^*_a}) \, \sqrt{\lambda(m_{B_c}^2,m_{B^*_a}^2,q^2)} \bigg) \, a_0 \, \Omega_2 \bigg] \nonumber \\ \\
H_{L}^{NP}&=&\frac{16}{\sqrt{m_{B_c} m_{B^*_a}}} \, \bigg[ (m_{B_c}^2 + m_{B^*_a}^2 - q^2) \, \Omega_1 \nonumber \\
&-& m_{B^*_a} \, \big((m_{B_c} - m_{B^*_a})^2 - q^2\big) \, a_0 \, \Omega_2 \bigg] \;. \nonumber
\end{eqnarray}
For $a=s$ the SM predictions
\begin{eqnarray}
{\cal B}(B_c^+ \to B_s^* \, \mu^+ \nu_\mu)&=&\,0.030 \, (1)\,\,\left(\frac{|V_{cs}|} {0.987}\right)^2 \nonumber \\
{\cal B}(B_c^+ \to B_s^* \, e^+ \nu_e)&=&\,0.032 \, (1)\,\,\left(\frac{|V_{cs}|} {0.987}\right)^2
\label{BrBsstarRis}
\end{eqnarray}
include only the error on the form factors. For $\mu$ channel,
the $q^2$ distribution in Fig.~\ref{fig:BrBs} is affected by a small FF uncertainty. In the NP extension the tensor operator has a visible effect on the spectrum.
Moreover, the spectra of longitudinally and transversely polarized $B_s^*$ in Fig.~\ref{fig:BrBsstarL}
show that NP mainly affects the longitudinal $B^*_s$ polarization in the small $q^2$ region. The ratio $\displaystyle F_T=\frac{\Gamma_T}{\Gamma_L+\Gamma_T}$, with $\Gamma_{T,L}$ the decay widths to transversely and longitudinally polarized $B_s^*$, is predicted in the SM: $F_T=0.413 \pm 0.004$, and remains smaller than $1/2$ when the NP operators are included, with the main effect due to the $T$ operator, as shown in Fig.~\ref{fig:FTBsstar}.
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.45\textwidth]{BrBs.pdf} \\
\includegraphics[width = 0.45\textwidth]{BrBsstar.pdf}
\caption{ \baselineskip 10pt \small $q^2$ spectrum of the modes $B_c^+ \to B_s \mu^+ \nu_\mu$ (top) and $B_c^+ \to B_s^* \mu^+ \nu_\mu$ (bottom). The Standard Model result (red SM band) includes the uncertainty on the form factors. The result for the full Hamiltonian Eq.~\eqref{hamil} is obtained varying the effective couplings in the quoted ranges (gray NP band). For $B_c^+ \to B_s^* \mu^+ \nu_\mu$ the spectrum obtained omitting the tensor operator $T$ is also displayed (dashed cyan lines).}\label{fig:BrBs}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\vspace*{-1.7cm}
\hspace*{-0.5cm}
\includegraphics[width = 0.55\textwidth]{BrBsstarLandT.pdf}
\vspace*{-1.7cm}
\caption{ \baselineskip 10pt \small $q^2$ distribution for longitudinally (top) and transversely polarized $B_s^*$ meson (bottom) in $B_c^+ \to B_s^* \mu^+ \nu_\mu$. The color codes are the same as in Fig.~\ref{fig:BrBs}. }\label{fig:BrBsstarL}
\end{center}
\end{figure}
\begin{figure}[b]
\begin{center}
\includegraphics[width = 0.48\textwidth]{FTBsstar.pdf}
\caption{ \baselineskip 10pt \small Fraction of transversely polarized $B_s^*$. The lines correspond to SM, to the NP operators in Eq.~\eqref{hamil} separately considered, and to the full set of NP operators.}\label{fig:FTBsstar}
\end{center}
\end{figure}
The $q^2$-dependent forward-backward (FB) lepton asymmetry
\begin{eqnarray}
&&
{\cal A_{FB}}(q^2)=\left({\displaystyle{\frac{d \Gamma}{dq^2}}} \right)^{-1}\times \label{AFB} \\
&&
\left[\int_0^1 \, d\cos \, \theta \, \displaystyle{\frac{d^2 \Gamma}{dq^2 d\cos \, \theta}} -\int_{-1}^0 \, d\cos \, \theta \, \displaystyle{\frac{d^2 \Gamma}{dq^2 d\cos \, \theta}} \right] \,\,\, \nonumber
\end{eqnarray}
is affected by a small uncertainty in the SM (Fig.~\ref{fig:AFB}). The asymmetry has a zero precisely determined at $q_0^2 \simeq 0.1905\,(5)$ GeV$^2$. This observable is particular sensitive to the tensor operator: indeed, as shown in Fig.~\ref{fig:AFB}, excluding this operator the asymmetry in NP practically coincides with SM. When all the operators in the extended Hamiltonian are considered the position of the zero is in the range $q_0^2 \in [0.149,\,0.208]$ GeV$^2$.
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.41\textwidth]{AFB.pdf}
\caption{ \baselineskip 10pt \small $q^2$-dependent forward-backward lepton asymmetry in $B_c^+ \to B_s^* \mu^+ \nu_\mu$. The red band corresponds to SM, the gray band to the full Hamiltonian \eqref{hamil}. The region obtained excluding the tensor operator $T$ is indicated by the dashed cyan lines.}\label{fig:AFB}
\end{center}
\end{figure}
The effects of the new operators can also be observed in the coefficients $c_{0,1,2}$ defined in the expression \cite{Penalva:2020xup,Penalva:2020ftd}
\begin{equation}
\frac{d {\cal B}(B_c^+ \to B_s^* \, \mu^+ \nu_\mu)}{dq^2 d \cos \theta}= c_0 +c_1 \cos \theta + c_2 \cos^2 \theta \,\, , \label{eq:c012}
\end{equation}
as shown in Fig.~\ref{fig:a0a1a2Bsstar}.
\begin{figure}[b]
\begin{center}
\includegraphics[width = 0.42\textwidth]{a0a1a2Bsstar.pdf}
\caption{ \baselineskip 10pt \small Coefficients $c_{0,1,2}$ in Eq.~\eqref{eq:c012} for $B_c \to B^*_s \mu^+ \nu_\mu$. The color codes are the same as in Fig.~\ref{fig:AFB}. }\label{fig:a0a1a2Bsstar}
\end{center}
\end{figure}
Interesting information is encoded in the correlations between the various observables in the decay modes to the pseudoscalar and vector meson. We analyze them in turn, neglecting the common FF uncertainties, considering the SM, each NP operator and all operators together. Since the scalar and pseudoscalar operators have a minor impact on the results, we do not discuss them individually.
\begin{figure}[b]
\begin{center}
\includegraphics[width = 0.45\textwidth]{BRsCorrelation.pdf}
\caption{ \baselineskip 10pt \small Correlation between the branching fractions ${\cal B}(B_c^+ \to B_s \mu^+ \nu_\mu$) and ${\cal B} (B_c^+ \to B_s^* \mu^+ \nu_\mu$) in SM (black dot) and for the NP operators in Eq.~\eqref{hamil}. The regions labeled $VR$, $V$, $R$ and $T$ are obtained varying separately the coefficients of the corresponding operators in their quoted ranges. The NP-All region refers to the full set of operators in \eqref{hamil}. }\label{fig:Brscorr}
\end{center}
\end{figure}
Fig.~\ref{fig:Brscorr} shows the correlation between the branching fractions of the pseudoscalar and vector modes ${\cal B}(B_c^+ \to B_s \, \mu^+ \nu_\mu)$ and ${\cal B}(B_c^+ \to B_s^*\, \mu^+ \nu_\mu)$. The SM point corresponds to the central values in Eqs.~(\ref{BrBsRis}) and (\ref{BrBsstarRis}). When all NP operators are considered the enlarged (pink) region is obtained. Anticorrelation between the branching fractions is found when the $R$ operator is considered. Increasing $\epsilon_V^\mu$ produces a positive correlation between the two observables. The tensor operator $T$ can allow a reduction of ${\cal B}(B_c^+ \to B_s^* \,\mu^+ \nu_\mu)$ with respect to SM.
Structured patterns are found in the correlations of the branching fractions ${\cal B}(B_c^+ \to B_s \, \mu^+ \nu_\mu)$ and ${\cal B}(B_c^+ \to B_s^* \, \mu^+ \nu_\mu)$ with the integrated FB lepton asymmetry in the $B^*_s$ mode
\begin{equation}
A_{FB}=\int_{q^2_{min}}^{q^2_{max}} dq^2 \, {\cal A_{FB}}(q^2)\,, \label{eq:AFB}
\end{equation}
as shown in Fig.~\ref{fig:BrAFBcorr}. Varying the $R$ and $V$ coefficients produces anticorrelations in case of the $B_s$ channel, same sign correlation in case of $B_s^*$.
The tensor operator results in a mild anticorrelation in the $B_s^*$ case. The combined analysis of all observables can allow to isolate the signature of the different NP operators.
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.45\textwidth]{BrAFBCorrelation.pdf} \\
\includegraphics[width = 0.45\textwidth]{BrstarAFBCorrelation.pdf}
\caption{ \baselineskip 10pt \small Correlations between the integrated forward-backward lepton asymmetry $A_{FB}$ in $B_c^+ \to B_s^* \mu^+ \nu_\mu$, defined in Eq.~\eqref{eq:AFB}, with ${\cal B}(B_c^+ \to B_s \, \mu^+ \nu_\mu)$ (top) and ${\cal B}(B_c^+ \to B_s^* \mu^+ \nu_\mu)$ (bottom panel). The color codes are the same as in Fig.~\ref{fig:Brscorr}.}\label{fig:BrAFBcorr}
\end{center}
\end{figure}
\subsection{$B_c^+ \to B_d \ell^+ \nu_\ell$ and $B_c^+ \to B_d^*(\to B_d \gamma) \ell^+ \nu_\ell$ }
The $c \to d$ semileptonic $B_c$ modes also give access to relevant information. The SM expectations
\begin{eqnarray}
{\cal B}(B_c^+ \to B_d \,\mu^+ \nu_\mu)&=&8.3 \, (5) \times 10^{-4}\, \left(\frac{|V_{cd}|}{0.221}\right)^2 \nonumber \\
{\cal B}(B_c^+ \to B_d \,e^+ \nu_e)&=&8.7 \, (5) \times 10^{-4}\, \left(\frac{|V_{cd}|}{0.221}\right)^2 \,\,\,\,\,\,\,\,\,
\,\label{BrBdRis}
\end{eqnarray}
derive from the form factors in \cite{Cooper:2020wnj}. The quoted errors are only due to the FF uncertainty.
The corresponding predictions for $B_c^+ \to B_d^* \,\bar \ell \nu_\ell$ in SM are
\begin{eqnarray}
{\cal B}(B_c^+ \to B_d^* \,\mu^+ \nu_\mu)&=&20 \,(1) \times 10^{-4}\, \left(\frac{|V_{cd}|}{0.221}\right)^2 \nonumber \\
{\cal B}(B_c^+ \to B_d^* \, e^+ \nu_e)&=&21 \,(1) \times 10^{-4}\, \left(\frac{|V_{cd}|}{0.221}\right)^2. \,\,\,\,\,\,\,\,\, \label{BrBdstarRis}
\end{eqnarray}
For the $\mu$ channel, the impact of the NP operators in the decay distributions is shown in Fig.~\ref{fig:BrBd}. The spectra in SM are affected by a small FF uncertainty. Including the NP operators sizably enlarges the spectrum of the pseudoscalar mode. The forward-backward asymmetry
Eq.~\eqref{AFB} for the pseudoscalar mode shows deviations from the SM expectation mainly due to the tensor operator, Fig.~\ref{fig:AFBctoBd}.
Large effects are allowed in $B_c^+ \to B_d^*(\to B_d \gamma) \ell^+ \nu_\ell$: this is due to the contribution of the tensor operator, that overwhelms the other ones if the coefficient $\epsilon_T^\mu$ is varied in the parameter space bound in \cite{Becirevic:2020rzi} using $D$ meson decays.
The distributions of longitudinally and transversely polarized $B_d^*$, Fig.~\ref{fig:BrBdstarL}, show that the tensor operator can sizably affect the transverse distribution. In SM the integrated width to longitudinal $B^*_d$ is larger than to the transverse one, as shown in Fig.~\ref{figFTBdstar}. The tensor operator can reverse such a hierachy.
Also the $q^2$-dependent forward-backward lepton asymmetry shows this effect, as seen in Fig.~\ref{fig:AFBd}. The inclusion of the tensor operator produces a zero for the ${\cal A_{FB}}$ distribution in the range $q_0^2 \in [0.27\, {\rm GeV}^2 ,\,q^2_{max}]$, while in the SM $q_0^2= 0.188(1)$ GeV$^2$ is expected. The position of the zero of ${\cal A_{FB}}(q^2)$ has a remarkable discriminating power of NP operators.
The effects of the new operators on the coefficients defined in \eqref{eq:c012} are shown in Fig.~\ref{fig:a0a1a2Bdstar}.
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.45\textwidth]{BrBd.pdf} \\
\includegraphics[width = 0.45\textwidth]{BrBdstar.pdf}
\caption{ \baselineskip 10pt \small $q^2$ spectrum of the modes $B_c^+ \to B_d \, \mu^+ \nu_\mu$ (top) and $B_c^+ \to B_d^* \, \mu^+ \nu_\mu$ (bottom). The Standard Model results (green SM band) include the uncertainty on the form factors. The spectra for the full Hamiltonian in Eq.~\eqref{hamil} are obtained varying the effective couplings in their quoted ranges (gray NP band). For $B_c^+ \to B_d^* \, \mu^+ \nu_\mu$ the spectrum obtained omitting the tensor operator $T$ is also shown (dashed orange lines). }\label{fig:BrBd}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.41\textwidth]{AFBBctoBd.pdf}
\caption{ \baselineskip 10pt \small $q^2$-dependent forward-backward lepton asymmetry in $B_c^+ \to B_d \, \mu^+ \nu_\mu$. The green line corresponds to SM, the gray band is obtained for the Hamiltonian \eqref{hamil}. The dashed orange lines are obtained excluding the tensor operator $T$. }\label{fig:AFBctoBd}
\end{center}
\end{figure}
\begin{figure}[b]
\begin{center}
\vspace*{-1.9cm}
\hspace*{-0.5cm}
\includegraphics[width = 0.51\textwidth]{BrBdstarLandT.pdf}
\vspace*{-1.cm}
\caption{ \baselineskip 10pt \small $q^2$ distribution of longitudinally (top) and transversely polarized $B_d^*$ (bottom) in $B_c^+ \to B_d^* \, \mu^+ \nu_\mu$. The color codes are as in Fig.~\ref{fig:BrBd}. }\label{fig:BrBdstarL}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.48\textwidth]{FTBdstar.pdf}
\caption{ \baselineskip 10pt \small Fraction of transversely polarized $B_d^*$. The lines correspond to the SM, to the NP operators in Eq.~\eqref{hamil} separately considered, and to the full set of NP operators. }\label{figFTBdstar}
\end{center}
\end{figure}
\begin{figure}[b]
\begin{center}
\includegraphics[width = 0.41\textwidth]{AFBd.pdf}
\caption{ \baselineskip 10pt \small $q^2$-dependent forward-backward lepton asymmetry in $B_c^+ \to B_d^* \, \mu^+ \nu_\mu$. The green band corresponds to SM, the gray one is obtained for the Hamiltonian \eqref{hamil}. The region obtained excluding the tensor operator $T$ (dashed orange lines) is also displayed. }\label{fig:AFBd}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.42\textwidth]{a0a1a2Bdstar.pdf}
\caption{ \baselineskip 10pt \small Coefficients $c_{0,1,2}$ in Eq.~\eqref{eq:c012} for $B_c \to B^*_d \mu^+ \nu_\mu$. The color codes are as in Fig.~\ref{fig:AFBd}. }\label{fig:a0a1a2Bdstar}
\end{center}
\end{figure}
The correlation plots in Figs.~\ref{fig:Brdcorr} and \ref{fig:BrdAFBcorr} give access to other information. The branching factions ${\cal B}(B_c^+ \to B_d^* \, \mu^+ \nu_\mu) $ and ${\cal B}(B_c^+ \to B_d \, \mu^+ \nu_\mu) $ are sizably affected by the NP contributions.
The $R$ operator anti-correlates the decay widths of the pseudoscalar and vector modes, while the $V$ contribution results in a positive correlation. In particular, ${\cal B}(B_c^+ \to B_d^* \, \mu^+ \nu_\mu) $ increases with respect to SM if $R$ is included, and decreases considering only $V$. However, the main effect is due to the tensor operator that strongly enhances ${\cal B}(B_c^+ \to B_d \, \mu^+ \nu_\mu)$ if its coefficient is varied in the range quoted in \cite{Becirevic:2020rzi}. Such a macroscopic effect on the one hand requires to further scrutinize the bounds from the $D$ meson decays, on the other hand shows the relevance of the $B_c$ modes in the search of BSM signals. This is confirmed by the correlations between the integrated forward-backward lepton asymmetry $A_{FB}$ and the branching fractions of the pseudoscalar and vector modes. As shown in Fig.~\ref{fig:BrdAFBcorr}, the integrated $A_{FB}$, that in SM is predicted to be negative, is anti-correlated with ${\cal B}(B_c^+ \to B_d \,\mu^+ \nu_\mu)$ mainly due to the tensor operator. $A_{FB}$ can become positive in the allowed range for the coefficient of such an operator, an interesting experimental signature. On the other hand, $A_{FB}$ and ${\cal B}(B_c^+ \to B_d^* \,\mu^+ \nu_\mu)$ are positively correlated, and the enhancement of the branching fraction closely follows the enhancement of $A_{FB}$ obtained varying the coefficient of the tensor operator.
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.45\textwidth]{BRdCorrelation.pdf}
\caption{ \baselineskip 10pt \small Correlation between the branching fractions ${\cal B}(B_c^+ \to B_d \mu^+ \nu_\mu$) and ${\cal B} (B_c^+ \to B_d^* \mu^+ \nu_\mu$) in SM (black dot) and considering the NP operators in Eq.~\eqref{hamil}. The regions labeled $VR$, $V$, $R$ and $T$ are obtained varying separately the coefficients of the corresponding operators in their quoted ranges. The NP-All region refers to the full set of operators in \eqref{hamil}. }\label{fig:Brdcorr}
\end{center}
\end{figure}
\begin{figure}[b]
\begin{center}
\includegraphics[width = 0.45\textwidth]{BrdAFBCorrelation.pdf} \\
\includegraphics[width = 0.45\textwidth]{BrstarAFBdCorrelation.pdf}
\caption{ \baselineskip 10pt \small Correlations between the integrated forward-backward lepton asymmetry $A_{FB}$ in $B_c^+ \to B_d^* \mu^+ \nu_\mu$, defined in Eq.~\eqref{eq:AFB}, with ${\cal B}(B_c^+ \to B_d \, \mu^+ \nu_\mu)$ (top) and ${\cal B}(B_c^+ \to B_d^* \mu^+ \nu_\mu)$ (bottom panel). The color codes are the same as in Fig.~\ref{fig:Brdcorr}. }\label{fig:BrdAFBcorr}
\end{center}
\end{figure}
\section{Conclusions}
The semileptonic $B_c$ decays induced by the $c \to s,d$ transitions play an interesting role in SM and in the search of BSM effects analogous to the ones emerging in $B$ decays. The heavy quark spin symmetry has allowed to analyze the full phenomenology of such decays using two nonperturbative form factors obtained by lattice QCD. The assessment of the role of the symmetry-breaking terms requires additional nonperturbative information, namely some other form factor in few points of the kinematical range. We have studied several significant observables in these decay modes, together with the effects and their correlations of the SM extension involving dimension-6 operators and left-handed neutrinos.
On the basis of
the available information on semileptonic $D$ decays we have found that sizable deviations from SM are allowed in $B_c^+ \to B_d^* \, \mu^+ \nu_\mu$. Of particular interest are the correlations
of the effects of the NP operators in the various observables, that can be used to pin-down the single contributions. For example, the branching fractions of the pseudoscalar and vector modes are positively or negatively correlated if the $R$ or $V$ contributions are considered. Other correlations involve the integrated FB lepton asymmetry, in particular the effect of the tensor operator in the $B_c^+ \to B_d^* \, \mu^+ \nu_\mu$ mode correlated to the branching fraction. The position of the zero in the FB lepton distribution, as well as the fraction of longitudinally vs transversely polarized final vector mesons constitute other observables worth to measure.
\section{Acknowledgements}
We thank D. Be\v{c}irevi\'c, F. Jaffredo, A. Pe\~nuelas and O. Sumensari for communications about Ref.~\cite{Becirevic:2020rzi}.
This study has been carried out within the INFN project (Iniziativa Specifica) QFT-HEP.
|
\section{Introduction}
Orbital order is a long-standing issue tracing back to the transition metal oxides~\cite{Tokura00,Maekawa04,Khomskii14}.
The precise mechanism driving orbital order remains largely unknown due to the intricate interplay among spin, orbital, charge, and lattice degrees of freedom in host crystals. In particular, a recent example of relevance is the nematic phase in iron-based superconductors, which entwines with spin Ising order, orbital order, and lattice structural distortion as dictated by symmetry~\cite{Fernandes14}. Among diverse theoretical proposals in addressing the origin of nematicity~\cite{Fradkin10,Fernandes19}, one interesting finding is that the orbital order in the nematic phase manifests its essential role in the metal-insulator transition and promotes an intermediate phase, {\it i.e.}, the orbital-selective Mott phase~\cite{Yu18}. This phase is characterized by the orbital-dependent Mott localization and interpolates the itinerant and Mott localized limits, validating the incipient Mott picture~\cite{Si09a,Si09b}. By contrast, a natural question may raise for bosonic systems: how the orbital order evolves in the superfluid-Mott insulator (SF-MI) transition.
Yet, much efforts have been denoted to the understanding of orbital order in electronic materials. While, the studies in bosonic systems are rare~\cite{Wu09,Li16}. Experimentally, artificial systems, such as ultracold atomic~\cite{Bloch07,Wirth11,Soltan12,Kock16,Zhou18} and photonic~\cite{Jacqmin14,Amo17,Amo19} systems, have been shown the exciting possibility of stimulating the crystals with $p$-orbital bosons in the first excited band. For instance, the Dirac points in the $p$-band hexagonal lattice are theoretically predicted by the early study~\cite{Wu08} and experimentally observed in photonic systems~\cite{Jacqmin14}. Later, the orbital edge state, which is extensively studied in graphene~\cite{Neto09,Novoselov11,Geim11}, is confirmed in the subsequent photonic experiment~\cite{Amo17}. More recently, the evidence of nematic superfluid (SF) phase in a hexagonal lattice, which is attributed to the orbital order, is also reported in ultracold atomic systems~\cite{Li20}.
The main purpose of our study is to give a comprehensive understanding of the orbital order in the SF-MI transition. The single-particle spectrum of the $p$-band triangular lattice exhibits a pair of Dirac points at the corners of hexagonal Brillouin zone (HBZ), resembling the low energy physics of graphene~\cite{Neto09,Novoselov11,Geim11}. The evolution of orbital order across the SF-MI transition is then studied based on the Bose-Hubbard model. In the weak-interacting limit, the $p$-band triangular lattice is frustrated due to the inability to simultaneously minimize both the kinetic and interacting energies. This weak-coupling SF phase is characterized by the intertwining of the axial $p_\pm=p_x\pm ip_y$ and in-plane $p_\theta=\cos\theta p_x+\sin\theta p_y$ orbital orders. Interestingly, the Bogoliubov excitation spectrum in this SF phase gaps the original Dirac points in the single-particle spectrum but exhibits emergent Dirac points. In the strong-interacting limit, the orbital order is also studied based on the orbital exchange model. We show that the classical ground state is of ferro-orbital type and enjoys an emergent SO$(2)$ rotational symmetry, which ensures an infinite degeneracy. The orbital fluctuation ultimately lifts the degeneracy and selects discrete quantum ground states through the order by disorder mechanism. Moreover, the phase diagram established by Gutzwiller approach interpolates these two limits. Besides these two phases, we find an intermediate SF phase with the ferro-orbital order. This intermediate phase survives in a wide range of low boson filling and gradually increase the occupation in the preferable in-plane orbital when approaching the Mott insulator (MI) with unit boson filling $n=1$. Our study provides strong clues that the ferro-orbital order in the MI $n=1$ phase likely evolves from the intermediate SF phase, facilitating the understanding on the role of orbital order in the SF-MI transition.
The reminder of this paper is organized as follow. In Sec.~\ref{sec:minimal}, we introduce the $p$-band tight-binding model in the triangular lattice as well as the Bose-Hubbard model. We establish the ground-state phase diagram by utilizing Gutzwiller approach in Sec.~\ref{sec:gutzwiller}. The orbital order in the weak-interacting limit is further studied with Bogoliubov approximation by treating the Bose-Hubbard interaction perturbatively in Sec~\ref{sec:weak}. In Sec.~\ref{sec:strong} the orbital exchange model is constructed to study the orbital order by treating the hopping processes as perturbations. Finally, we summarise and discuss the results in Sec.~\ref{sec:summary}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{./band.eps}
\caption{
(a) The structure of triangular lattice and the hexagonal Brillouin zone.
The blue lines mark the Wigner-Seitz cell and the reduced Brillouin zone due to the umklapp scattering between the band minima at $M$ points. (b) The band structure of the tight-binding model in Eq.~(\ref{eq:TB}) with the $\sigma$-bonding $t_\sigma=1$ being the energy unit. The pseudovector fields $\bm{d}\equiv\left(d_z,d_x\right)$ near the Dirac point at $K$ (c) and $K^\prime$ (d) resemble the vortex in $XY$ systems with winding number $W=1$.
}
\label{fig:band}
\end{figure}
\section{Minimal model}
\label{sec:minimal}
We begin with the tight-binding model that describes the hopping processes of bosons in the $p$-band triangular lattice depicted in Fig.~\ref{fig:band}(a). Introducing an orbital pseudospin representation, the momentum-space Hamiltonian in the basis $p_{\bm k}=\left[p_{x{\bm k}},p_{y{\bm k}}\right]^\text{T}$ reads
\begin{equation}
\mathcal{H}_{\bm k} =
d_0\left({\bm k}\right)\tau_0+
d_x\left({\bm k}\right)\tau_x+
d_z\left({\bm k}\right)\tau_z,
\label{eq:TB}
\end{equation}
where $\tau_0$ and $\bm{\tau}$ are the identity matrix and Pauli matrices respectively, and the coefficients
$d_0\left(\bm{k}\right)=\left(t_\sigma+t_\pi\right)\sum_{i}\cos k_i$, and
$\{d_x\left(\bm{k}\right),d_z\left(\bm{k}\right)\}=\left(t_\sigma-t_\pi\right)/2\times\{\sqrt{3}\left(\cos k_3-\cos k_2\right),\cos k_1+\sum_{i}\cos k_i\}$.
Here, the crystal momenta $\{k_1,k_2,k_3\}$ are measured along reciprocal lattice vectors $\{\bm{b}_1,\bm{b}_2,\bm{b}_3\equiv-\bm{b}_1-\bm{b}_2\}$, and the hopping integrals $t_\sigma$ and $t_\pi$ denote the $\sigma$ and $\pi$ bonding of $p$ orbitals, respectively. For the $\pi$ bonding, the bond vector lies in the nodal plane of $p$ orbitals. As a result, the strength of $\pi$ bonding is typically much weaker than that of $\sigma$ bonding. The band structure of the tight-binding model in Eq.~(\ref{eq:TB}) is plotted in Fig.~\ref{fig:band}(b). Notably, two bands cross at the Dirac points located at $K$ and $K^\prime$ points of HBZ. To describe the corresponding low-energy behavior around $K$ and $K^\prime$ points, we derive the effective $k\cdot p$ model
\begin{eqnarray}
\mathcal{H}_{K/K^\prime} \left({\bm q}\right) =
d_0\tau_0
+d_x\tau_x
+d_z\tau_z
+\mathcal{O}\left(q^2\right)
\label{eq:kp}
\end{eqnarray}
with the coefficients
\begin{eqnarray}
d_0 = -\frac{3}{2}\left(t_\sigma+t_\pi\right),
\{d_x,d_z\} = \pm\frac{3}{4}\sqrt{3}\left(t_\sigma-t_\pi\right)\{q_x,-q_y\}.\nonumber
\end{eqnarray}
Diagonalizing $\mathcal{H}_{K/K^\prime}\left({\bm q}\right)$ gives two non-interacting bands $E^\pm_{K/K^\prime}\left({\bm q}\right)=d_0\pm\sqrt{d_x^2+d_z^2}$, resulting in a linear dispersed Dirac point with the velocity $v=3\sqrt{3}/4(t_
\sigma-t_\pi)$. The pseudovector fields $\bm{d}\equiv\left(d_z,d_x\right)$ around $K$ and $K^\prime$ points, shown in Figs.~\ref{fig:band}(c) and \ref{fig:band}(d) respectively, have a $p$-wave symmetry. The topological charge of Dirac point is given by the winding number of pseudovector field: $W=\frac{1}{2\pi}\oint_\mathcal{C}\nabla\theta\left({\bm q}\right)\cdot d{\bm q}=1$, where $\theta\equiv\text{arctan}\left(d_x/d_z\right)$ and $\mathcal{C}$ is a contour enclosing the singular $K/K^\prime$ point, indicating that the Dirac point carries a $\pi$ Berry flux. The band minima are located at three inequivalent centres $M$ of HBZ edges, promoting a finite-momentum Bose-Einstein condensate for weakly interacting bosons. For non-interacting bosons, an infinite degenerate manifold of the single-particle ground state can be constructed by the linear superposition of the Bloch functions at these band minima. The umklapp scattering between the band minima transfers a lattice phonon which carries the momentum of multiple primitive reciprocal vectors. This process folds three $M$ points to $\Gamma$ point and underlies the reduced Brillouin zone (RBZ) and the enlarged Wigner-Seitz cell, as illustrated by the blue lines in Fig.~\ref{fig:band}(a).
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{./gw.eps}
\caption{
Gutzwiller approach.
(a) The ground-state phase diagram in $\mu/U$ vs $U/t_\sigma$ plane accommodates three distinct phases, including (1) ISF, the superfluid phase intertwining axial and in-plane orbital orders; (2) DSF, the superfluid phase with the detwinned in-plane orbital order, and (3) MI $n=1$, the Mott insulator phase with unit filling $n=1$. The transition lines separating these three phases merge at a triple point $\left(U/t_\sigma,\mu/U\right)\approx\left(12.2,0.66\right)$. (b) Evolution of amplitude of condensate order parameter $\bm{\phi}=\left(\phi_x,\phi_y\right)$, compressibility $\kappa=\partial n/\partial \mu$, and orbital pseudospin $\bm{\tau}$ at fixed $\mu/U=0.5$. The black and blue dots in (a) mark the the critical points with the vanishing order parameters $|\bm{\phi}|$ and $|\tau_y|$, respectively. In the numerical calculations, the truncation of local Fock space $N_\pm=10$ for the maximum occupation in axial orbitals $p_\pm=p_x\pm ip_y$ and $t_\pi=0$ are used.
}
\label{fig:gutzwiller}
\end{figure}
Having established the single-particle physics, we are then in a position to study the effects of many-particle interactions. The interacting Hamiltonian can be generally constructed in terms of Haldane pseudopotentials by projecting a pair of particles into relative angular momenta, respecting the quantum statistics~\cite{Duncan90,chen18}. The Bose-Hubbard interaction which is mathematically described by zero relative angular momentum takes the form
\begin{equation}
H_\text{I}=\frac{3}{2}U\sum_{i}\left[\hat{n}_{i}\left(\hat{n}_{i}-\frac{2}{3}\right)
-\frac{1}{3}\hat{L}_{zi}^2\right],
\end{equation}
where $\hat{n}_{i}=\sum_{\alpha=x,y}p_{\alpha i}^\dagger p_{\alpha i}$ is the occupation operator and $\hat{L}_{zi}=-i\sum_{\alpha,\beta=x,y}\epsilon_{z\alpha\beta}p_{\alpha i}^\dagger p_{\beta i}$ is the $z$-component orbital angular momentum at $i$-th site~\cite{Isacsson05,Liu06,Umucalilar08}. Here $\epsilon_{\alpha\beta\gamma}$ is the Levi-Civita antisymmetric tensor. The interaction can be experimentally realized through the Feshbach resonance~\cite{Chin10} and optical nonlinearities~\cite{Carusotto13} for ultracold atomic and photonic systems, respectively.
\section{Gutzwiller approach}
\label{sec:gutzwiller}
To gain an overall understanding on the ground-state phase diagram, Gutzwiller approach~\cite{Gutzwiller63,Gutzwiller65,Kotliar91,Werner92} has its advance in capturing the physics in the intermediate regime of Hubbard interaction $U$, and straddles the limits of the weakly interacting SF and strongly interacting MI phases. It has been utilized to establish the phase diagram of $p$-band Bose-Hubbard model with different lattice geometries in the early study~\cite{Isacsson05}. This approach starts from the factorized local Fock state
\begin{eqnarray}
\left|\Psi_\text{GW}\right\rangle=&&\prod_{i}\sum_{\text{F}}\eta_\text{F}^i\left|\text{F}\right\rangle_i, \nonumber\\
\left|\text{F}\right\rangle_i=\frac{1}{\sqrt{n_+^\text{F}!n_-^\text{F}!}}&&
\left(p_{+i}^{\dagger}\right)^{n_+^\text{F}} \left(p_{-i}^{\dagger}\right)^{n_-^\text{F}} \left|0\right\rangle \nonumber
\end{eqnarray}
where $n_{\pm}^\text{F}$ is the occupation of bosons in the axial orbitals $p_\pm=p_x\pm ip_y$ and $\eta_{\text{F}}^i$ is the probability weighting factor determined variationally. It takes into account that the multioccupation of bosons in the local orbitals are energetically costly. In numerical calculations, a truncation of the local Fock space is imposed and the filling of bosons is dictated by chemical potential $\mu$ for the grand canonical ensemble. We decompose the hopping terms in the tight-binding Hamiltonian as $p^\dagger_{\alpha i}p_{\beta j}\approx p^\dagger_{\alpha i}\phi_{\beta j}+\phi^*_{\alpha i}p_{\beta j}-\phi^*_{\alpha i}\phi_{\beta j}$ with the condensate order parameter $\phi_{\alpha i}=\sum_{\text{F}\text{F}^\prime} \eta_{\text{F}}^{i*}\eta_{\text{F}^\prime}^i{}_i\left\langle\text{F}|p_{\alpha i}|\text{F}^\prime\right\rangle_i$ entangling the local Fock states. The self-consistent solution of the order parameters $\phi_{\alpha i}$ requires an iterative minimization of the energy functional over the Wigner-Seitz cell in Fig.~\ref{fig:band}(a). The calculated phase diagram for $t_\pi=0$ shown in Fig.~\ref{fig:gutzwiller} (a) accommodates three different phases including two distinct SF phases and the MI phase. We have also verified that the phase diagram remains qualitatively robust against the perturbative $\pi$ bonding $t_\pi=-0.1t_\sigma$. To characterize the orbital order, we numerically evaluate the ground-state expectation of orbital pseudospin $\bm{\hat{\tau}}=\sum_{\alpha\beta} p^\dagger_{\alpha}\bm{\tau}_{\alpha\beta}p_{\beta}$. The axial orbital order $p_\pm=p_x\pm ip_y$ is characterized by the orbital polarization $\tau_y$ in $y$ axis, while the in-plane orbital order $p_\theta=\cos\theta p_x+\sin\theta p_y$ directing at angle $\theta$ with $x$ axis corresponds to the orbital polarization $\left(\tau_z,\tau_x\right)=\tau\left(\cos\left[2\theta\right],\sin\left[2\theta\right]\right)$ in $zx$ plane. Figure~\ref{fig:gutzwiller}(b) shows the detailed evolution of order parameters at fixed $\mu/U=0.5$, which determines the phases across the SF-MI transition. The stability of each phase is further checked with various sets of supercell sizes up to $8\bm{a}_1\times8\bm{a}_2$. As shown in Fig.~\ref{fig:gutzwiller}(b), two distinct SF phases share a non-vanishing uniform order parameter $|\bm{\phi}_i|$ and can be distinguished by the ground-state expectation value of orbital pseudospin $\hat{\bm{\tau}}$. Initially, the ground state at weak Hubbard interaction $U$ develops an intertwined order by entangling both the axial and in-plane orbital orders. The former is characterized by the alternating signs in adjacent rows but an identical amplitude of $\tau_y^i$ therefore suggesting the ordering of antiferro-orbital angular momentum, while the latter is indicated by the uniform $\left(\tau^i_z,\tau^i_x\right)$ implying the ferro-orbital order. Therefore, this superfluid phase intertwining axial and in-plane orbital orders is denoted as ISF. The detailed pattern of orbital orders will be further discussed in Sec.~\ref{sec:weak}. At the critical Hubbard interaction $U_{c1}$, the orbital pseudospin is completely aligned in the $zx$ plane $\left(\tau_z,\tau_x\right)$, showing a detwinned ferro-orbital order. This superfluid phase is thus denoted as DSF. Since the inter-site hopping process is treated at the mean-filed level in Gutzwiller approach, we will show in Sec.~\ref{sec:strong} that the orientation of in-plane orbital is solely determined by the quantum fluctuation due to the orbital anisotropy. With further increasing Hubbard interaction $U>U_{c2}$, the MI phase obtained within Gutzwiller approximation is simply a product of local Fock states and is thus featureless. As shown in Fig.~\ref{fig:gutzwiller}(b), the ISF-DSF-MI transition driven by he Hubbard interaction $U$ is well detected by the discontinuous jumps of the compressibility $\kappa=\partial n/\partial\mu$. Therefore, the phase transition discussed here may be experimentally probed by measuring the boson filling $n$. Below we shall justify the orbital orders above from two extrema limits.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{./bogoliubov.eps}
\caption{ Bogoliubov approach.
(a) The condensed ground state intertwines the ferro-orbital order with the in-plane orbital being parallel to the bond, and the antiferro-orbital angular momentum with the spatial pattern indicated by red and blue lines, supporting staggered fluxes. (c) The Bogoliubov excitation spectrum along the high symmetry line indicated in (b). (d) The detailed Bogoliubov excitation spectrum indicated by the pink box in (c) exhibits Dirac bosons.
(e) The pseudovector field $\bm{d}$ of Dirac boson in the lowest two excitation branches around $M$ point has winding number $W=1$. The parameters used in numerics are $\left(t_\sigma,t_\pi,nU\right)=\left(1,0,1\right)$.
}
\label{fig:bogoliubov}
\end{figure}
\section{Weak-coupling approach}
\label{sec:weak}
In the weakly interacting limit, the Hubbard interaction $U$ is treated perturbatively. The operators can be decomposed in terms of quantum fluctuations $\tilde{p}_{\alpha\bm{k}\ell}$ around the condensed ground-state wave function $\phi_{\alpha \ell}$
\begin{equation}
p_{\alpha\bm{k}\ell} = \phi_{\alpha \ell}\delta\left(\bm{k}\right)+\tilde{p}_{\alpha\bm{k}\ell}\text{, }
\alpha=x,y
\end{equation}
with $\ell$ specifying the sublattice in the Wigner-Seitz cell.
In the spirit of Bogoliubov approximation~\cite{Bogolyubov47,Abrikosov63}, the Hamiltonian is expanded in powers of quantum fluctuations and is truncated up to the quadratic order.
Detailed derivations are presented in Appendix~\ref{app:Bgl}.
The zeroth-order terms in this expansion determine the energy functional $\varepsilon\left(\bm{\phi}^*,\bm{\phi}\right)$ of the condensate at $\Gamma$ point in RBZ. The time-dependent Gross-Pitaevskii equation can be readily derived via the Euler-Lagrange equation
\begin{equation}
\frac{\partial\mathcal{L}}{\partial\phi_{\alpha\ell}^*}
-\frac{d}{dt}\left(\frac{\partial\mathcal{L}}{\partial\dot{\phi}_{\alpha\ell}^*}\right)=0,
\end{equation}
where the Lagrangian $\mathcal{L}=\sum_{\alpha\ell}i\hbar\left(\phi_{\alpha \ell}^*\dot{\phi}_{\alpha\ell}-\phi_{\alpha \ell}\dot{\phi}_{\alpha\ell}^*\right)-\varepsilon\left(\bm{\phi}^*,\bm{\phi}\right)$~\cite{Pethick08}. The ground state can be numerically solved through the imaginary-time evolution of Gross-Pitaevskii equation by propagating an initial trial state~\cite{Dalfovo99}. Mathematically, this procedure is equivalent to the minimization of the energy functional $\varepsilon\left(\bm{\phi}^*,\bm{\phi}\right)$, which causes the linear order terms in $\tilde{p}_{\alpha\bm{k}\ell}$ to vanish. The calculated ground-state condensate develops an intertwined order, confirming the results from Gutzwiller approach. The ferro-orbital order is characterized by orientating the in-plane orbital parallel to the bond, which breaks the lattice rotational symmetry. The antiferro-orbital angular momentum is characterized by the alternating sign of $\tau_y$ along the direction perpendicular to the in-plane orbital, and breaks the time-reversal symmetry as well as the lattice translational symmetry. Interestingly, the symmetry breaking of this weak-coupling phase, which has been studied in details, is shown to be universal in the strong-coupling regime with boson filling $n\ge2$~\cite{Wu06}. As schematically depicted in Fig.~\ref{fig:bogoliubov}(a), the staggered flux pattern of the ground-state condensate arising from the time-reversal symmetry breaking is characterized by the bond current $J_{ij}=-i\sum_{\alpha\beta}t_{\gamma}^{\alpha\beta}(\langle p^\dagger_{\alpha i}p_{\beta j}\rangle-\text{c.c.})\delta_{j,i\pm \bm{a}_\gamma}$ where the hopping matrix $t_{\gamma}=[(t_\sigma+t_\pi)\tau_0+(t_\sigma-t_\pi)(\cos[2\theta_\gamma]\tau_z+\sin[2\theta_\gamma]\tau_x)]/2$ and $\theta_\gamma$ is the azimuthal angle of $\bm{a}_\gamma$. The early studies find pure axial orbital orders with different lattice geometries, which support bond currents as a natural consequence~\cite{Liu06,Xu16,Liberto16}. In contrast, the intertwined orbital order in the present study, due to the inability to simultaneously minimize both kinetic and interacting energies, originates from the geometric frustration of the triangular lattice. Having settled the ground state, we then proceed with the quadratic order
\begin{equation}
\mathcal{H}^{(2)}_{\bm{k}}=
\frac{1}{2}
\left[\tilde{\bm{p}}^\dagger_{\bm{k}},\tilde{\bm{p}}_{-\bm{k}}\right]
\left[
\begin{matrix}
X_{\bm{k}} & Y \\
Y^\dagger & X_{-\bm{k}}
\end{matrix}
\right]
\left[
\begin{matrix}
\tilde{\bm{p}}_{\bm{k}} \\
\tilde{\bm{p}}^\dagger_{-\bm{k}}
\end{matrix}
\right],
\label{eq:bogoliubov}
\end{equation}
which describes the Bogoliubov excitation on top of the ground-state condensate. The diagonal terms $X_{\pm\bm{k}}$ in Eq.~(\ref{eq:bogoliubov}) receive contributions from both the hopping processes and the self-energy correction of Hubbard interaction $U$, while the off-diagonal terms $Y$ ($Y^\dagger$) describe the anomalous processes in which a pair of bosons scatter with each other into the excited states (condensates). As shown in Fig.~\ref{fig:bogoliubov}(c), the Bogoliubov excitation spectrum has a gapless Goldstone mode around $\Gamma$ point in RBZ as the signature of $U(1)$ symmetry breaking.
Interestingly, we find that the Bogoliubov spectrum along the high symmetry line $K$-$K^\prime$ gaps out the original Dirac points at $K$ and $K^\prime$ points in the band structure but exhibits emergent Dirac bosons at $M$ point. As a representative example, the Dirac point in the lowest two branches is identified by numerically calculating the winding number shown in Fig.~\ref{fig:bogoliubov}(e). It is worth mentioning that the Dirac bosons only exist on one edge of RBZ parallel to the bond that is selected by the in-plane orbital order, serving as a fingerprint of lattice rotational symmetry breaking. We have also checked that the Dirac related physics is robust for $t_\pi=-0.1t_\sigma$. While the early study focuses on the Dirac fermion in the band structure with staggered fluxes~\cite{Wang17}, the present study investigates the Dirac bosons in the elementary excitation on top of the Bose-Einstein condensate with staggered fluxes instead.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{./sw.eps}
\caption{ Strong-coupling approach.
(a) The eigenvalues of orbital interaction matrix $\Lambda_{\bm{k}}$ within the classical approximation.
(b) The emergent SO$(2)$ rotational symmetry in the easy plane $(\tau_z,\tau_x)$ of orbital pseudospin space.
(c) The zero-point energy $E_\text{ZP}\left(\theta\right)$ in Eq.~(\ref{eq:zp}) from the correction of orbital fluctuations has minima at $\theta=\pi/6+\mathbb{Z}\pi/3$. The energies in (a) and (c) are in units of $t_\sigma^2/16U$.
}
\label{fig:sw}
\end{figure}
\section{Strong-coupling approach}
\label{sec:strong}
Finally, we turn to the strong-coupling limit in which the virtual hopping processes are treated perturbatively. Since the charge excitation in MI phase is suppressed by the charge gap proportional to Hubbard interaction $U$, the orbital fluctuation is the remaining low energy degree of freedom. Following the standard second-order perturbation theory, the effective low-energy physics for the MI $n=1$ phase is captured by the following orbital exchange model
\begin{eqnarray}
H_\text{OE}=J\sum_{\langle ij \rangle\parallel\bm{a}_\gamma}\tau^i_\gamma\tau^j_\gamma
+J^\prime \sum_{\langle ij \rangle}
\left(\bm{\tau}^i\cdot\bm{\tau}^j+2\tau^i_y\tau^j_y\right)
\label{eq:orbex}
\end{eqnarray}
with
\begin{eqnarray}
\tau_\gamma=\tau_z\cos\left[2\theta_\gamma\right]+\tau_x\sin\left[2\theta_\gamma\right]. \nonumber
\end{eqnarray}
Detailed derivations are presented in Appendix~\ref{app:OEX}.
The ferro-orbital exchange $J=-(t_\sigma-t_\pi)^2/16U$ in Eq.~(\ref{eq:orbex}) is inherently anisotropic originating from the anisotropic shape of $p$ orbitals. In contrast, the exchange $J^\prime=-t_\sigma t_\pi/8U$ is antiferro-orbital due to the opposite sign of $t_\sigma$ and $t_\pi$. The first term in Eq.~(\ref{eq:orbex}) involving interacting orbital degrees of freedom is coined as the compass model~\cite{Kugel82,Nussinov15}. The ground state is first studied by treating the orbital pseudospin $\bm{\tau}$ as a classical vector. The orbital interaction can be minimized via the diagonalization of the orbital exchange Hamiltonian in momentum space $H_\text{OE} = \sum_{\bm{k}}\bm{\tau}^{-\bm{k}}\Lambda_{\bm{k}}\bm{\tau}^{\bm{k}}$. As shown in Fig.~\ref{fig:sw}(a), the lowest eigenvalue of $\Lambda_{\bm{k}}$ is found at $\Gamma$ point in RBZ and has a twofold degeneracy, suggesting that the classical ground state is ferro-orbital ordering. The degenerated eigenvalue has important implications on the structure of orbital order. A close inspection on the Hamiltonian in Eq.~(\ref{eq:orbex}) reveals that the $y$ component of orbital pseudospin $\tau_y$ is decoupled from the other two components $\tau_{z,x}$, in which the lowest degenerate eigenvalue arises. It indicates that the ordering of orbital pseudospin $\bm{\tau}$ lies in the $zx$ plane. More importantly, this degeneracy renders a continuous $\text{SO}(2)$ rotational symmetry of orbital pseudospin $(\tau_z,\tau_x)=\tau(\cos\theta,\sin\theta)$ as shown in Fig.~\ref{fig:sw}(b). Note that this symmetry restricted to the classical ground state is emergent and is not an exact symmetry of the orbital exchange model, which is only invariant under finite point group rotations. The orbital order of the classical ground state evolves in the $zx$ plane without any energy cost, which makes the system particularly susceptible to quantum fluctuations. Following Holstein-Primakoff spin wave theory~\cite{Holstein40}, the zero-point energy arises from the correction of quantum fluctuations, and is studied as a function of the rotation $\theta$ about the $y$ axis of orbital pseudospin. To the leading order, the zero-point energy takes the form
\begin{equation}
E_\text{ZP}(\theta)=\frac{1}{2N}\sum_{\bm{k}}\omega_{\bm{k}}\left(\theta\right)+6J+12J^\prime,
\label{eq:zp}
\end{equation}
where $N$ is the number of sites, the orbital excitation $\omega_{\bm{k}}(\theta)=2\sqrt{[\varphi_{\bm{k}}(\theta)+2\varphi^\prime_{\bm{k}}-6J-12J^\prime]^2-[\varphi_{\bm{k}}(\theta)+\varphi_{\bm{k}}^\prime]^2}$, and the auxiliary functions $\{\varphi_{\bm{k}}(\theta),\varphi_{\bm{k}}^\prime\}=\{2J\sum_{\gamma}\sin^2[2\theta_\gamma+\theta],4J^\prime\}\cos[\bm{k}\cdot\bm{a}_\gamma]$. Detailed derivations are presented in Appendix~\ref{app:LOW}. The numerical evaluation of zero-point energy is shown in Fig.~\ref{fig:sw}(c). The orbital fluctuation lifts the degeneracy protected by the continuous $\text{SO}(2)$ rotational symmetry, and selects the quantum ground state at $\theta=\pi/6+\mathbb{Z}\pi/3$ ($\mathbb{Z}$ is an integer). This mechanism is known as order by disorder in frustrated spin systems~\cite{Villain80,Henley89,Lacroix11,Diep13,Green18}.
\section{Conclusion and discussion}
\label{sec:summary}
To summarize, we have studied the evolution of orbital ordering across the SF-MI transition in the $p$-band triangular lattice. The ground-state phase diagram is first established by Gutzwiller approach, which interpolates continuously between two extreme limits, deep in SF phase and deep in MI phase. The orbital orders in these two limits are further examined by the perturbation approaches. With systematic analyses, we identify an intermediate SF phase with the detwined in-plane ferro-orbital order, which correctly reproduce the orbital order in the MI $n=1$ phase. It is worth remarking several directions for further studies. The quantum fluctuations, which can be partially restored with the cluster Gutzwiller approach~\cite{Luhmann13,Bai18}, may deserve to be studied for its role in selecting the orbital order in the vicinity of SF-MI transition. Alternatively, it is also interesting to investigate the details of SF-MI transition within a single unified method, {\it e.g.} the quantum Monte-Carlo simulation. Finally, we close by briefly discussing the dissipation. The experimental realization of the Bose-Hubbard model in photonic systems involves light-matter interactions, which may be better described as an open system. Therefore, another direction to generalize our work is to study the effect of dissipation.
\section*{Acknowlegdgement}
We thank Congjun Wu and W. Vincent Liu for helpful discussions.
This work is supported by the National Natural Science Foundation of China under Grants No. 11704338, No. 11534001, No. 11504008, and the National Basic Research Program of China under Grant No. 2015CB921102.
|
\section{Introduction}
Mining is the extraction of economically valuable minerals or materials from the earth. This has raised the importance of the production scheduling process due to its significant role in the profitability and efficiency of any mining operation. Mine production scheduling problem \cite{moreno2017linear} is a well-study mining engineering problem, and it has received much attention in past decades \cite{osanloo2008long} from both engineering and research. The task of the mine production schedule is to generate a mining sequence and ensure the product meets the blending resource constraints and object to maximize the net present value of the mining operation.
The mine production scheduling problem is commonly formulated as a Mixed-Integer Program (MIP) with binary variables \cite{johnson1968optimum,bley2012solving,Topal}. However, since it becomes a challenge for the MIP when the problem deals with the blending resource constraints. Lipovetzky et al.~\cite{Nir2014} introduced a combined MIP for a mine planning problem, which devises a heuristic objective function in the MIP and can improve the resulting search space for the planner. Samavati et al. \cite{samavati2017local} proposed a heuristic approach that combines local branching with a new adaptive branching scheme to tackle the production scheduling problem in open-pit mining.
Stockpiles are essential components in the supply chain of the mining industry, and play a significant role in the mine scheduling problem. Jupp et al. \cite{jupp2013role} introduced the four different reasons for stockpiling before material processing: buffering, blending, storing, and grade separation. In open-pit mine production scheduling problem, stockpiles are used for blending different grades of material from the mine or keeping low-grade ore for possible future processing \cite{moreno2017linear,rezakhah2020practical}. Rezakhah et al. \cite{rezakhah2019open} used a linear-integer model to approximate the open-pit mine production scheduling with stockpiling problem which forces the stockpile to have an average grade above a specific limit. Recently, some researchers present nonlinear-integer models to solve open pit mine production scheduling with stockpiles. Tabesh et al. \cite{tabesh2015comprehensive} proposed a nonlinear model of stockpiles to optimize a comprehensive open-pit mine plan but not give any results. Bley et al. \cite{bley2012solving} proposed a nonlinear model for mine production planning, however, they only consider one stockpile.
In this paper, we study an important component of the mine production scheduling problem, the stockpile blending problem. This problem is challenging to address in terms of blending material from stockpiles for parcels to match the demands of downstream customers. We define the stockpile blending problem as an optimization problem that aims to maximize the volume of valuable material from all parcels by finding the percentage that each stockpile provides for each parcel in the whole planning. Furthermore, the strategy has to respond to the mining schedule and the market plan where the mine schedule provides the material mining and sending to corresponding stockpiles in each period, and the market plan provides the customer requirements.
Solving the stockpile blending problem in mining optimally is critical because it is based on an uncertain supply of mineralized materials for the resource available in the mine. This uncertainty is acknowledged in the related technical literature to be the major reason for not meeting production expectations \cite{baker1998resource,asad2012optimal}. Given its substantial impact on the financial outcome of mining operations, this paper focuses on dealing with the uncertainty in metal content within a mineral deposit being mined. For the stochastic variables of the stockpile blending problem, we introduce chance-constrained programming here to tackle the uncertainty of material grades. Chance-constrained optimization problems~\cite{Charnes,Miller} whose resulting decision ensures the probability of complying with the constraints and the confidence level of being feasible to have received significant attention in the literature. Chance-constraint programming has been widely applied in different disciplines for optimization under uncertainty~\cite{Uryasev}. For example, chance-constraint programming has been applied in analog integrated circuit design~\cite{McConaghy}, mechanical engineering \cite{Mercado}, and other disciplines \cite{liu,poojari}. However, so far, chance-constraint programming has received little attention in the evolutionary computation literature~\cite{Zhang}.
It is difficult for MIP to tackle such a continuous optimization problem containing the nonlinear constraints. To address this challenge, this paper proposes two repair operators to tackles the complex constraints. Follow the paper \cite{Yue19}, we present the surrogate functions of the chance constraints by using Chebyshev's inequality. Furthermore, a well-known evolutionary algorithm, the Differential Evolution (DE) algorithm is introduced to solve the stockpile blending problem. Recently, evolutionary algorithms have received much attention in solving large-scale optimization problems and multi-dimensions problem. The DE algorithm is a simple and effective evolutionary algorithm used to solve global optimization problems in a continuous domain \cite{neri2010recent,pham2011comparative}. The DE and its variants have been successfully applied to solve numerous real-world problems from diverse domains of science and engineering \cite{das2010differential,neri2010recent}. This paper investigates the use of the DE algorithm combining the two repair operators for solving the problem. Then we compare the impactas of different chance constraints on the objective value.
The rest of the paper is organized as follows. In the next section, we present the model of the stockpile blending problem and a decision variable normalized operator for the continuous decision variables as well as a duration repair operator. After that, the chance constraints model and the surrogate functions of the chance constraints are presented in Section \ref{sec:chanceModel}. Following, we describe the approach we used to solve the problem and the fitness function of the algorithm. We set up experiments and investigate the performance of the different fitness functions in Section \ref{sec:experiment}. We conclude with Section \ref{sec:conclusion}.
\section{Deterministic model}
In this section, we present nonlinear formulations of the stockpiles blending problem with a deterministic setting. In reality, some processes such as the chemical process in the concentrate production progress are highly complex to model because it is influenced by many factors, some of which include the mineralogy of the ore, particle size of milled material, temperature, and chemical reactants available in the process. The information of these variables was not available to us, therefore within this study to recovery factors of all materials from the chemical processing stage and the copper percentage within the produced copper concentrate is assumed to be constant throughout the stockpiles blending and production schedule.
We first introduces notation as follow, and then provide the math. We use the term "material" to include ore, i.e., rock that contains sufficient minerals including metals that can be economically extracted and to include waste, and we use chemical symbol represent the corresponding material, i.e., Cu denotes Copper, Fl denotes Flerovium.
\subsection{Notation}
\label{sec:notation}
\begin{table}[th]
\centering
\scriptsize
\scalebox{1.0}{
\makebox[\linewidth][c]{
\tabcolsep=0.5cm
\begin{tabular}{ll}
\multicolumn{2}{l}{\textbf{Indices and sets:} } \\
$s\in \mathcal{S}$ & stockpiles; $1,\ldots,S$\\
$p\in \mathcal{P}$ & parcels; $1,\ldots,P$\\
$o$ & material; $\{Cu,Ag,Fe, Au,U, Fl, S\}$\\
$m \in \mathcal{M} $ & month; $1,\ldots,m$ \\
\\
\multicolumn{2}{l}{\textbf{Decision variables:} }\\
$x_{ps}$ & fraction of parcel $p$ claimed from stockpile $s$ \\
$t_p$ & produce time (duration) for parcel $p$ \\
$w_p$: & tonnage of parcel $p$\\
$\theta_{ps}$: & tonnage stores in stockpile $s$ after providing material to parcel $p$\\
$c_p$: & Cu tonne in parcel $p$\\
$g^o_p$: & grade of material $o$ in parcel $p$\\
$ \tilde{g}^o_{ps}$: & grade of material $o$ in stockpile $s$ when proving parcel $p$\\
$k_p$: & tonne concentrate of parcel $p$\\
$r^{Cu}_p$: & Cu recovery of parcel $p$\\
$r^{Fl}_p$: & Fl recovery of parcel $p$\\
\\
\multicolumn{2}{l}{ \textbf{Parameters: }} \\
$T_p^m$: & binary parameter, if $T^m_p=1$, parcel $p$\\ & is the first parcel need to prepare in month $m$, if $T^m_p=0$ otherwise\\
$\delta$: & discount factor for time period\\
$\tilde{\phi}$: & factor in chemical processing stage \\
$\phi^{Au}$: & factor of Au in chemical processing stage \\
$\phi^{U}$: & factor of U in chemical processing stage \\
$\phi^{Fe}$: & factor of Fe in chemical processing stage \\
$\phi^{Cu}$: & factor of Cu in chemical processing stage \\
$(\gamma_1,\gamma_2)$ : & factor of Cu percentage within the produced Cu concentrate \\
$\mu^{Fl}$: & factor of Fl recovery \\
$\mu^{U}$ : & factor of U recovery \\
$(\mu^{Cu}_1, \mu^{Cu}_2)$ : & factor of Cu recovery \\
$D^m$: & duration of month $m$ \\
$H^m_s$: & tonnage of material hauled to stockpile $s$ in month $m$\\
$G^{om}_s$: & grade of material $o$ that shipping to the stockpile $s$ in month $m$\\
$K_p$: & expected tonne concentrate of parcel $p$\\
$R^{Fl}_p$ : & upper threshold of Fl recovery of parcel $p$\\
${Cu}_p$ : & lower threshold of Cu grade of parcel $p$\\
$N_m$: & number of planning parcels in month $m$\\
\end{tabular}}}
\label{tab:notation}%
\end{table}%
\subsection{Model with deterministic setting}
\begin{align}
Obj: \max \sum_{p\in \mathcal{P}} c_p = max \sum_{p\in \mathcal{P}} \left(w_p g^{Cu}_p r^{Cu}_p \right)
\label{obj:function}
\end{align}
\begin{align}
s.t. \sum_{\sum_{m=1}^{m-1}N_m +1 \leq p \leq \sum_{m=1}^m N_m} t_p \leq D^m
\label{con:duration}
\end{align}
\begin{align}
\sum_{s\in\mathcal{S}} x_{ps}=1 \qquad \forall p \in \mathcal{P}
\label{con:variables}
\end{align}
\begin{align}
r^{Cu}_p = \mu^{Cu}_1 \frac{g^{Cu}_p}{g^S_p} +\mu^{Cu}_2 \qquad \forall p \in \mathcal{P}
\label{con:curecovery}
\end{align}
\begin{align}
g^o_p = \sum_{s\in \mathcal{S}} x_{ps} \tilde{g}^o_{ps} \qquad \forall p \in \mathcal{P}
\label{con:parcelgrade}
\end{align}
\begin{align}
w_p = & \delta t_p[\tilde{\phi} +(\phi^{Au}\log g^{Au}_p) + (\phi^{U}\log g^{U}_p) \nonumber\\
& - (\phi^{Fe}\log g^{Fe}_p) + (\phi^{Cu}\log g^{Cu}_p)]
\label{con:parceltonne}
\end{align}
\begin{align}
k_p= \frac{c_p}{\gamma_1 \frac{g^{Cu}_p}{g^{S}_p}+\gamma_2}
\label{con:concentrate}
\end{align}
\begin{eqnarray}
\tilde{g}^o_{ps} =
\begin{cases}
\frac{\tilde{g}^o_{(p-1)s}\cdot \theta{(p-1)s}+ G^{om}_s \cdot H^m_s}{\theta{(p-1)s}+H^m_s} & \textit{if $T_p^m=1$}
\\
\tilde{g}^o_{(p-1)s} & otherwise
\end{cases}
\label{con:stockpilegrade}
\end{eqnarray}
\begin{eqnarray}
\theta_{ps} =
\begin{cases}
\theta_{(p-1)s}+H^m_s - x_{ps} \cdot w_p & \textit{if $T_p^m=1$ } \\
\theta_{(p-1)s}- x_{ps} \cdot w_p & otherwise
\label{con:stockpilestorage}
\end{cases}
\end{eqnarray}
\begin{align}
g^{Cu}_p \geq {Cu}_p \qquad \forall p\in \mathcal{P}
\label{con:cugradebound}
\end{align}
\begin{align}
( K_p-1) \leq k_p \leq (K_p+1) \qquad \forall p\in \mathcal{P}
\label{con:concentratebound}
\end{align}
\begin{align}
\mu^{Fl} g^{Fl}_p \leq R^{Fl}_p \qquad \forall p \in \mathcal{P}
\label{con:Flrecoverybound}
\end{align}
The objective function (\ref{obj:function}) aims to maximize the sum of Cu volume of all parcels, which is obtained by the tonnage of parcels multiply the Cu grade, and multiple the Cu recovery. Constraint (\ref{con:duration}) forces the sum of duration of the parcels that planned into the same month less than the available duration of this month. Constraint (\ref{con:variables}) ensures that the sum of the decision variables for the same parcel is equal to 1. Function (\ref{con:curecovery}) denotes the simplified calculation of Cu recovery of parcels, and function (\ref{con:parcelgrade}) calculates the material grades of parcels. Function (\ref{con:parceltonne}) express the simplified calculation of parcel tonne which is a component in objection function. Function (\ref{con:concentrate}) shows the simplify version of how to calculate the tonne concentrate of parcels.
Constraint (\ref{con:stockpilegrade}) enforces material grade balance for stockpiles when providing material to parcels. Constraint (\ref{con:stockpilestorage}) enforces inventory balance when providing material to parcels. Constraint (\ref{con:cugradebound}) forces the Cu grade of parcels to less than or equal to the given lower bound of the Cu grade. Constraint (\ref{con:concentratebound}) forces the value of tonne concentrate of each parcel is no more or less than the expected tonne concentrate by one. Constraint (\ref{con:Flrecoverybound}) ensures the Fl recovery of each parcel less than the bound given in advance.
\begin{algorithm}[t]
\caption{Decision variables normalized approach}
\label{alg:variablesapproach}
\KwIn{Decision vector $X_j=\{x_{1j},x_{2j}..,x_{Ij}\}$}
$a = \sum_{i \in I }x_{ij}$\;
\For {$i=1$ to $I$}{
$x_{ij} = \frac{x_{ij}}{a}$\;
}
\Return the normalized decision variables.
\end{algorithm}
The stockpile blending problem is a non-linear optimization problem in the continuous search space. To tackle the constraint (\ref{con:variables}), we introduce a decision variables normalized approach (cf. Algorithm \ref{alg:variablesapproach}) to force solutions match the constraint. This approach first calculates the sum of the variables of each parcel separately, then each variable of the same parcel is divided by the corresponding sum. It shows the significant importance of applying this approach to a generated decision vector to meet the constraint.
Due to the complex constraint (\ref{con:concentratebound}) is too tight to construct feasible solutions, we develop a repair operator to address this problem. As shown in function (\ref{con:concentrate}), the value of tonne concentrate of parcels is related to the duration of this parcel and the material grades of the parcel. Meanwhile, referring to equation (\ref{con:parcelgrade}), material grades of the parcel are directly calculated by decision variables. Therefore, with the fixed decision variables of a parcel, the real tonne concentrate of this parcel is affected by the duration of this parcel. We present a duration repair operator (cf. Algorithm \ref{alg:durationfix}) which uses a binary search process to convert an infeasible solution into a solution without violating constraint (\ref{con:concentratebound}).
\begin{algorithm}[t]
\caption{Duration repair operator}
\label{alg:durationfix}
\KwIn{$X\in (0,1)^{I\cdot J}$, $i\in \{1,..,I\}$, $j\in\{1,..,J\}$; parameter $\zeta$; available duration $\mathit{D}$} \
\KwOut{ parcel duration: $\mathit{d}\in \{0,\mathit{D}\}$}\
initialization: $\underline{d}=0$, $\overline{d}=\mathit{D}$, $d\in\{0,\mathit{D}\}$ , $k=\zeta \cdot d$ \
\While{$d \in \{0, \mathit{D}\}$ and $k \notin \{K-1,K+1\}$ }{
\If{$k>K+1$}{
$d := (d+\underline{d})/2$\;
$k := \zeta \cdot d$\;
\If{$k>K+1$}{
$\overline{d}=d$\;
}\Else{
$\underline{d}=d$\;
}
}
\ElseIf{$k < K-1$}{
$d := (d+\overline{d})/2$}\;
$k := \zeta \cdot d$\;
\If{$k>K+1$}{
$\overline{d} :=d$\;
}\Else{$\underline{d} :=d$}
}
\Return \textit{the duration corresponding to solution $X$}
\end{algorithm}
Since the time complexity of the binary search is $\log n$ where $n$ denotes the length of the search space in the beginning. In our problem, the duration of each parcel can not exceed the total available duration of the month. The run-time of the duration repair operator for one parcel is $\log d$ in the worst case where $d$ denotes the total available duration of the current month.
\section{Model with chance constraints}
\label{sec:chanceModel}
In real-world mining engineering problems, the material grades are estimated by some tools, and in the research of mining scheduling problems, researchers treat the stochastic material grades as constant by using the expected values. In this paper, we are the first to discuss the influences of stochastic material grades on the objective value of the stockpile blending problem with chance constraints. Due to the complexity of the problem, we reformulated the constraints (\ref{con:cugradebound}) and (\ref{con:Flrecoverybound}) to chance-constrained. Chance-constrained programming is a competitive tool for solving optimization problems under uncertainty. The main feature is that the resulting decision ensures the probability of complying with constraints, i.e. the confidence of being feasible. Thus, using chance-constrained programming the relationship between profitability and reliability can be quantified.
\subsection{The formulation of the chance constraints}
First, we define additional notation as follow.
\begin{table}[h]
\centering
\scalebox{1.0}{
\makebox[\linewidth][c]{
\tabcolsep=0.5cm
\begin{tabular}{ll}
$\alpha_{Cu}$: & confidence of Cu grade chance constraint \\
$\alpha_{Fl}$: & confidence of Fl recovery chance constraint \\
\end{tabular}}}
\end{table}
Then, the new chance constraints are
\begin{align}
Pr\{g^{Cu}_p \geq Cu_P\} \geq \alpha_{Cu},
\label{chance:coppergrade}
\end{align}
and
\begin{align}
Pr\{\mu^{Fl}g^{Fl}_p \leq R_{p}^{Fl}\} \geq \alpha_{Fl}.
\label{chance:Flrecovery}
\end{align}
Constraints (\ref{chance:coppergrade}), (\ref{chance:Flrecovery}) force the confidence of ensuring the constraint are greater than or equal to the corresponding given bound.
We use Chebyshev's inequality to construct the available surrogate that translates to a guarantee on the feasibility of the chance constraint imposed by the inequalities. Firstly, we use Chebyshev's inequality to reformulate the chance constraints. The inequality has utility for being applied to any probability distribution with known expectation and variance. Therefore, we assume the stochastic material grades discussed in this paper are all estimated with given expected values and corresponding variances. Note that Chebyshev's inequality automatically yields a two-sided tail bound, there is a one-sided version of Chebyshev's inequality named Cantelli's inequality.
\begin{theorem}[Cantelli's inequality]
\label{thm:cheb}
Let $X$ be a random variable with $Var[X]>0$. Then for all $\lambda>0$,
\begin{equation}
P_r\{X\geq E[X] +\lambda\sqrt{Var[X]}\}\leq \frac{1}{1+\lambda^2}.
\label{the:Canplus}
\end{equation}
\begin{equation}
P_r\{X\leq E[X] -\lambda\sqrt{Var[X]}\}\leq \frac{1}{1+\lambda^2}.
\label{the:Canminus}
\end{equation}
\end{theorem}
We assume the material grades in the stockpiles are independent of each other, each grade corresponding expectation $a^{om}_{s}$ and variance $\sigma_s^{2om}$. Therefore, the expected material grades of stockpiles can be denoted as
\begin{eqnarray}
E(\tilde{g}^o_{ps}) =
\begin{cases}
\frac{\tilde{g}^o_{(p-1)s}\cdot \theta_{(p-1)s}+ a^{om}_s \cdot H^m_s}{\theta{(p-1)s}+H^m_s} & \textit{if $T_p^m=1$}
\\
E(\tilde{g}^o_{(p-1)s}) & otherwise.
\end{cases}
\nonumber
\end{eqnarray}
Furthermore, the variance of the material grades are
\begin{eqnarray}
Var(\tilde{g}^o_{ps})=
\begin{cases}
\left(\frac{\theta_{(p-1)s}}{\theta{(p-1)s}+H^m_s}\right)^2 Var(\tilde{g}^o_{(p-1)s}) +\left(\frac{H^m_s}{\theta{(p-1)s}+H^m_s}\right)^2 \sigma_s^{2om} & \textit{if $T_p^m=1$}
\\
Var(\tilde{g}^o_{(p-1)s}) & otherwise.
\end{cases}
\nonumber
\end{eqnarray}
Let $g^{Cu}_p = \sum_{s\in \mathcal{S}} x_{ps}\tilde{g}^{Cu}_{ps}$ be the Cu grade of parcel $p$ of a given solution $X=\{x_{p1},..,x_{ps},..,x_{p\mathbf{S}}\}$, and $$E[g^{Cu}_p ]=\sum_{s\in \mathcal{S}}x_{ps}E(\tilde{g}^{Cu}_{ps})$$ denotes the expected Cu grade of parcel $p$ of the solution derived by linearity of expectation, $$Var[g^{Cu}_p]=\sum_{s\in \mathcal{S}} (x_{ps})^2 Var(\tilde{g}^{Cu}_{ps})$$ denotes the variance of Cu grade of parcel $p$. To match the expression of the Cantelli's inequality (\ref{the:Canminus}), we set $$Cu_p=E[g^{Cu}_p]-\lambda\sqrt{Var[g^{Cu}_p]}$$ and have $$\lambda =\frac{E[g^{Cu}_p]-Cu_p}{\sqrt{Var[g^{Cu}_p]}}$$ for each parcel, then we have a formulation to calculate the upper bound of the chance constraint (\ref{chance:coppergrade}) as follows.
\begin{align}
Pr\{g^{Cu}_p \leq Cu_p\} \leq \frac{Var[g^{Cu}_p]}{Var[g^{Cu}_p]+(E[g^{Cu}_p]-Cu_p)^2} \leq (1-\alpha_{Cu})
\label{chance:cu}
\end{align}
Furthermore, let $ r^{Fl}_p=\mu^{Fl}\sum_{s\in \mathcal{S}}x_{ps}\tilde{g}^{Fl}_{ps} $
be the FL recovery of parcel $p$. Let $$E[r^{Fl}_p ]=\mu^{Fl}\sum_{s\in \mathcal{S}}x_{ps}E(\tilde{g}^{Fl}_{ps})$$
denotes the expectation of Fl recovery, and $$Var[r^{Fl}_p ]=\sum_{s\in \mathcal{S}} (\mu^{Fl} x_{ps})^2 Var(\tilde{g}^{Fl}_{ps}),$$
is the variance of FL recovery of parcel $p$ with solution $X=\{x_{p1},..,x_{ps},..,x_{p\mathcal{S}}\}.$
To match the expression of the Cantelli's inequality (\ref{the:Canplus}), we set $$R_{p}^{Fl}=\mu^{Fl}E[r_p^{Fl}]+\lambda\sqrt{Var[r^{Fl}_p]}$$ and have
$$\lambda = \frac{R_p^{Fl}-\mu^{Fl}E[g^{Fl}_p]}{\sqrt{Var[g^{Fl}_p]}}$$ for each parcel, then we have a formulation to calculate the upper bound of the chance constraint (\ref{chance:Flrecovery}) as follows.
\begin{align}
\resizebox{0.9\hsize}{!}{$Pr\{\mu^{Fl}g^{Fl}_p\geq R^{Fl}_p\} \leq \frac{Var[g^{Fl}_p]}{Var[g^{Fl}_p]+(R_p^{Fl}-\mu^{Fl}E[g^{Fl}_p])^2} \leq (1-\alpha_{Fl})$}
\label{chance:fl}
\end{align}
Now, we obtain the surrogate functions of the chance constraints. In the next section, we present the approach for solving the stockpile blending problem with chance constraints.
\section{Approaches for the stockpile blending problem with chance constraints}
\label{sec:approach}
In this section, we present the fitness functions for the differential evolution (DE) algorithm which has been proved successfully used in solving the optimization problem in continuous space.
\subsection{Fitness function for deterministic setting}
We start by designing a fitness function for the deterministic setting model that can be used in the DE algorithm. The fitness function $f$ for the approach needs to take all constraints into account. The fitness function of a solution $X$ is defined as follows.
\begin{align}
f(X) = \left(u(X),v(X),w(X),q(X),g(X), O(X) \right)
\label{fit:deter}
\end{align}
\begin{align}
& u(X)=\sum_{p\in \mathcal{P}} \max \{\left|K_p-k_p \right|,1\} \nonumber \\
& v(X)= \max\{\sum_{p\in \mathcal{P}}t_p - D^m , 0\} \nonumber \\
& w(X)=\min \{\sum_{p\in \mathcal{P}}\sum_{s\in \mathcal{S}}\theta_{ps},0\} \nonumber \\
& q(X)=\sum_{p \in \mathcal{P}} \max\{Cu_{p}- g^{Cu}_p,0\} \nonumber \\
& g(X)=\sum_{p \in \mathcal{P}} \max\{r^{Fl}_p-R^{Fl}_p,0\} \nonumber \\
& O(X)= \sum_{p\in \mathcal{P}} c_p. \nonumber
\end{align}
In this fitness function, the components $u,v,q,g$ need to be minimize while $w$ and $O$ maximized, and we optimize $f$ in lexicographic order. For the stockpile blending problem, any infeasible solution can at least violate one of the above constraints. Then, among solutions that meet all constraints, we aim to maximize the objective function. Formally, we have
\begin{align*}
&f(X) \succeq f(Y) \\
\textbf{iff} \ & u(X) <u(Y) \ \textit{or} \\
&{u(X)=u(Y)\wedge v(X)<v(Y)} \ \textit{or}\\
& \{u,v\} \text{are equal} \wedge w(X)>w(Y) \ \textit{or}\\
& \{u,v,w\} \text{are equal} \wedge q (X)<q(Y) \ \textit{or} \\
& \{u,v,w,q\} \text{are equal} \wedge g(X)<g(Y) \textit{or} \\
&\{u,v,w,q,g\} \text{ are equal} \wedge O(X)>O(Y),
\end{align*}
When comparing two solutions, the feasible solution is preferred in a comparison between an infeasible and a feasible solution. Between two infeasible solutions that violated the same constraint, the one with a lower degree of constraint violation is preferred.
\subsection{Fitness function of the problem with chance constraints}
Now, we design the fitness function for the stockpile blending problem with chance constraints. In this paper, we investigate the effectiveness of chance constraints on the objective value. We first reformulate the components $q$ and $g$ of the function (\ref{fit:deter}) with chance constraints (\ref{chance:cu} and \ref{chance:fl}) as follow,
\begin{align}
q'(X) =\sum_{p\in \mathcal{P}} max\left\{P_r\{g^{Cu}_p \leq Cu_p\}-(1-\alpha_{Cu}),0\right\}\\
g'(X) =\sum_{p\in \mathcal{P}} max\left\{P_r\{\mu^{Fl}g^{Fl}_p \geq R_p^{Fl}\}-(1-\alpha_{Fl}),0\right\}
\end{align}
where $q'$ and $g'$ need to be minimized.
To distinguish the influence of each chance constraint, we design three fitness functions where the two functions consider the chance constraints separately, and the other one uses the combination of components.
\begin{align}
f'(X) = \left(u(X),v(X),w(X),q'(X),g(X), O(X) \right)
\label{fit:Cuonly}\\
f''(X) = \left(u(X),v(X),w(X),q(X),g'(X),O(X) \right)
\label{fit:flonly}\\
f'''(X) = \left(u(X),v(X),w(X),q'(X),g'(X), O(X) \right)
\label{fit:combina}
\end{align}
\subsection{Differential evolution algorithm}
\begin{algorithm}[t]
\caption{Differential evolution algorithm}
$t \leftarrow 1$, initialize $\mathbf{P}^t =\{X_1^t,..,X_{NP}^t\}$ randomly \;
\While{stopping criterion not met}{
\For{$i\in\{1,..,NP\}$}{
$R\leftarrow$ A set of randomly selected indices from $\{1,..,NP\} \setminus \{i\}$ \;
$V^t_i \leftarrow$ mutation $(P^t,R,F)$ \;
$j_{rand} \leftarrow$ A randomly selected number from $\{1,..,n\};$ \
$U^t_i \leftarrow$ crossover$(X^t_i, V^t_i,C,j_{rand});$ \
}
\For{$i\in\{1,..,NP\}$}{
\If{$f(U^t_i) \succeq f(X^t_i)$}{
$X^{t+1}_i \leftarrow U^t_i$\;
}
\Else {
$X^{t+1}_i \leftarrow X^t_i$ \;}
}
$t\leftarrow t+1$ \;
}
\label{alg:de}
\end{algorithm}
Algorithm (\ref{alg:de}) shows the overall procedure of the basic DE algorithm. DE is usually initialized by generating a population of $NP$ individuals. For each $i\in\{1,..,NP\}$, $X^t_i$ is the $i$-th individual in the population $\mathbf{P}^t$. Each individual represents a $d$-dimensional solution of a problem. For each $j\in\{1,..,d\}$, $X_{ij}^t$ is the $j$-th element of $X^t_i$.
After the initialization of $\mathbf{P}^t$, the following steps are repeatedly performed until a termination condition is satisfied. For each $X^t_i$, the scale factor $F>0$ which controls the magnitude of the mutation, and the crossover rate $Cr\in[0,1]$ which controls the number of elements inherited from $X^t_i$ to a trail vector $U^t_i$ are constants and given in advance.
A set of parent indices $R=\{r_1,r_2\}$ are randomly selected from $\{1,..,n\}\setminus\{i\}$ such that they differ from each other. For each $X^t_i$, a mutant vector $V^t_i$ is generated by applying a mutation to $X^t_{r1},X^t_{r2}$. There are many mutation strategies that have been proposed in the literature \cite{das2010differential}. Here, we use the $DE/target-to-best/1$ strategy shown as follows, which is one of the most efficient strategies.
\begin{align}
V^t_i= X^t_i+F(X^t_{best}-X^t_i) +F(X^t_{r1}-X^t_{r2}),
\end{align}
where $X^t_{best}$ denotes the best individual in the current population.
After the mutant vector $V^t_i$ has been generated for each $X^t_i$, a trail vector $U^t_i$ is generated by applying crossover to $X^t_i$ and $V^t_i$. The scheme of the crossover can be outlined as
\begin{eqnarray}
U^t_{ij} =
\begin{cases}
V^t_{ij} & \textit{if}\ (rand_{i,j}[0,1] \leq Cr \ \textit{or}\ j=j_{rand}) \\
X^t_{ij} & otherwise
\label{DE:crossover}
\end{cases}
\end{eqnarray}
where $rand_{i,j}[0,1]$ is a uniformly distributed random number, which is called a new for each $j$-th element of the $i$-th parameter vector. $j_{rand}\in \{1,..,n\}$ is a randomly chosen index, which ensures that $U^t_i$ gets at least one element from $V^t_i$. It is instantiated once for each vector per generation.
After the trial vector, $U^t_i$ has been generated for each parent individual, the next step called selection which determines whether the target or the trailing vector survives to the next generation. The selection operation is described as
\begin{eqnarray}
X^{t+1}_{i} =
\begin{cases}
U^t_{i} & \textit{if}\ f(U^t_i)\succeq f(X^t_i) \\
X^t_{ij} & otherwise
\label{DE:selection}
\end{cases}
\end{eqnarray}
according to the fitness function.
\section{Experimental investigation}
\label{sec:experiment}
In this section, we examine the solution quality associated with different fitness functions. Due to business security, we are not able to investigate the proposed approach in real-data instances. Therefore, we first design the benchmark of the stockpile blending problem. Afterward, we compare the results obtained by using different fitness functions of the instances. Furthermore, considering the complexity of the problem with chance constraints, the instances we discussed in this section only contain one month schedule.
\subsection{Experimental Setup}
Table \ref{tab:paramRange} lists the intervals of input parameters mentioned in Section \ref{sec:notation}. The three instances we evaluated in this paper are created by randomly generated value of parameters from their intervals (see Table \ref{tab:paramRange}), we attach the parameters of these instances in the appendix. The randomly generated numbers are the expected values of material grades, and the deviation of material grades are set equal to $0.01$ multiply the expectation. Let $\alpha_{Cu}=\{0.999,0.99,0.9\}$ and $\alpha_{Fl}=\{0.999,0.99,0.9\}$. Base on this arrangement, we compare the performance of the DE algorithm with fitness functions (Eq. \ref{fit:deter}, \ref{fit:Cuonly}, \ref{fit:flonly}, \ref{fit:combina}) on the stockpile blending problem.
\begin{table}[t]
\centering
\caption{General information about the ore and processing parameters}
\scriptsize
\scalebox{1.0}{
\makebox[\linewidth][c]{
\begin{tabular}{ll}
\toprule
\text{Description} & \text{Values or Value range}\\
\midrule
Number of parcel & stockpiles; $\{3,4,5\}$\\
Number of Stockpile & $7$\\
Duration of month & ${29,30,31}$ \\
Discount factor for time period ($\delta$) & $0.98$\\
Factor in chemical processing stage ($\tilde{\phi}$) & $[1000,2000]$ \\
Factor of Au in chemical processing stage ($\phi^{Au}$) & $[200,300]$ \\
Factor of U in chemical processing stage ($\phi^{U}$) & $[300,400]$ \\
Factor of Fe in chemical processing stage ($\phi^{Fe}$) & $[560000,570000]$ \\
Factor of Cu in chemical processing stage ($\phi^{Cu}$) & $[6000000,7000000]$ \\
Factor of Cu percentage within the produced Cu concentrate $(\gamma_1,\gamma_2)$ & $([5,10],[30,40])$ \\
Factor of Fl recovery ($\mu^{Fl}$) & $[0.05,0.15]$\\
Factor of U recovery ($\mu^{U}$) & $[0.5,0.9]$\\
Factor of Cu recovery $(\mu^{Cu}_1, \mu^{Cu}_2)$ & $([1.5,3.5],[0,10])$\\
Tonnage of material hauled to stockpile & $[5000,1000000]$\\
Cu grade & $[0.05,2.5]$\\
Ag grade & $[1.0,4.0]$\\
Fe grade & $ [10.0,30.0]$\\
Au grade & $[0.3,2.0]$\\
U grade & $[30.0,400.0]$\\
Fl grade & $[1200,4500]$\\
S grade & $[0.15,1.0]$ \\
Expected tonne concentrate of parcel ($K_p$) & $[10000,\infty]$\\
Threshold of Fl recovery of parcel ( $R^{Fl}_p$) & $[1300,1500]$\\
Threshold of Cu grade of parcel (${Cu}_p$) & $[0.5,1.5]$\\
\bottomrule
\end{tabular}}}%
\label{tab:paramRange}
\end{table
We then investigate the performance of the DE algorithms with different fitness functions described in Section \ref{sec:approach} and provide the results from $30$ independent runs with $10000$ generation and $10$ population for all instances. For a closer look, we report the average, best and worst solutions obtained by the algorithm in corresponding columns. We also evaluate the algorithm by success rate which is the percentage of success for the algorithm in obtaining valid solutions out of $30$ runs.
\begin{table*}[t]
\centering
\caption{Fitness values obtained with single chance constraint}
\scalebox{0.6}{
\makebox[\linewidth][c]{
\begin{tabular}{llrrrrrrrrr}
\toprule
& & Deterministic & & \multicolumn{3}{l}{Cu Chance constraint $(\alpha_{Cu})$} & & \multicolumn{3}{l}{Fl Chance constraint $(\alpha_{Fl})$ } \\
Instance & & & & 0.999 & 0.99 & 0.9 & & 0.999 & 0.99 & 0.9 \\
\midrule
\multirow{4}{*}{1} & Mean & 103603035.94 & & 99319128.52 & 102724900.09 & 103206748.35 & & 103117715.98 & 103340755.20 & 102753876.59 \\
& Best & 110830487.20 & & 100434268.90 & 110489368.20 & 111221777.51 & & 108460913.10 & 110593860.12 & 106976825.01 \\
& Worst & 100025173.10 & & 98404158.16 & 99426947.62 & 99935646.96 & & 99979453.63 & 98951340.30 & 99444063.80 \\
& Success rate & & & 0.166666667 & 1 & 1 & & 1 & 1 & 1 \\
\midrule
\multirow{4}[0]{*}{2} &Mean & 66339866.34 & & 64280794.53 & 65346088.50 & 65440690.91 & & 65741114.11 & 66128957.43 & 64999603.10 \\
&Best & 69691652.87 & & 66062865.43 & 70504938.14 & 69307609.02 & & 70846603.30 & 69265095.80 & 67416065.20 \\
& Worst & 63401302.85 & & 62822049.78 & 61409848.38 & 62043167.43 & & 62885220.60 & 63139531.31 & 62369471.32 \\
& Success rate & & &0.3 & 1 & 1 & & 1 & 1 & 1 \\
\midrule
\multirow{4}[0]{*}{3} &Mean & 25706739.82 & & 25172345.85 & 25484058.55 & 25667396.95 & & 25414602.50 & 25487090.60 & 25737554.80 \\
&Best & 26714591.61 & & 25780112.93 & 27501228.19 & 26652385.99 & & 26542657.00 & 26440088.30 & 27420748.21 \\
&Worst & 25042412.92 & & 24939884.30 & 24517912.31 & 24338065.15 & & 24301347.20 & 24502516.90 & 24675079.20 \\
& Success rate & & & 0.166666667 & 0.733333333 & 0.8 & & 0.83333333 & 0.7 & 1 \\
\bottomrule
\end{tabular}}}%
\label{tab:singleResults}%
\end{table*}%
\begin{table*}[t]
\centering
\caption{Fitness values obtained with two chance constraints}
\scalebox{0.6}{
\makebox[\linewidth][c]{
\begin{tabular}{llrrrrrrrrrrrr}
\toprule
Instance & & &\multicolumn{11}{c}{Combine Chance constraints} \\
& & & \multicolumn{3}{c}{$\alpha_{Cu}=0.999$} & & \multicolumn{3}{c}{$\alpha_{Cu}=0.99$} & & \multicolumn{3}{c}{$\alpha_{Cu}=0.9$} \\
\midrule
$\alpha_{Fl}$ & & & 0.999 & 0.99 & 0.9 & & 0.999 & 0.99 & 0.9 & & 0.999 & 0.99 & 0.9 \\
\midrule
\multirow{4}[0]{*}{1} & Mean & & 98787701.39 & 99241024.07 & 99492631.19 & & 102600300.50 & 102276579.70 & 102682519.61 & & 103388918.00 & 102934747.10 & 102493177.23 \\
&Best & & 102031085.80 & 103755489.91 & 102599092.62 & & 106167319.20 & 107035409.13 & 109936137.61 & & 106369047.00 & 108285976.13 & 108115669.08 \\
&Worst & & 96585053.09 & 97115231.16 & 97595258.99 & & 99322445.27 & 98370524.81 & 100035474.00 & & 100191353.05 & 99131474.70 & 99124405.82 \\
&Success rate & & 0.366666667 & 0.366666667 & 0.4 & & 1 & 1 & 1 & & 1 & 1 & 1 \\
\midrule
\multirow{4}[0]{*}{2} & Mean & & 65983255.56 & 63563990.36 & 64902632.44 & & 65902068.31 & 65729892.86 & 65506695.73 & & 65678975.14 & 65559860.50 & 65651773.33 \\
&Best & & 69079010.41 & 65782360.39 & 69006540.86 & & 69991791.36 & 69754992.24 & 70596782.19 & & 69703718.21 & 68258615.00 & 70556638.81 \\
&Worst & & 62870362.84 & 61877215.08 & 61963922.28 & & 59672821.89 & 63038071.83 & 61823714.56 & & 61900565.70 & 61618562.90 & 62959270.15 \\
&Success rate & & 0.166666667 & 0.133333333 & 0.3 & & 1 & 1 & 1 & & 1 & 1 & 1 \\
\midrule
\multirow{4}[0]{*}{3}& Mean & & 25960917.39 & 25871636.05 & 25686417.52 & & 25459522.87 & 25682180.87 & 25687541.01 & & 25569238.21 & 25498602.00 & 25617280.63 \\
& Best & & 25960917.39 & 26273134.25 & 25686417.52 & & 26575050.89 & 26643851.74 & 26329733.68 & & 26457335.14 & 26589310.80 & 26326709.05 \\
& Worst & & 25960917.39 & 25473933.01 & 25686417.52 & & 24495873.72 & 24826217.28 & 24668501.30 & & 24693215.90 & 24721583.71 & 24549825.10 \\
& Success rate & & 0.033333333 & 0.133333333 & 0.033333333 & & 0.8 & 0.833333333 & 0.866666667 & & 0.8 & 0.86666667 & 0.76666667 \\
\bottomrule
\end{tabular}}}%
\label{tab:MultiResults}%
\end{table*}%
\subsection{Experimental Results}
We benchmark our approach with the combinations from the experimental setting described above. All experiments were performed using Java of version 11.0.1 and carried out on a MacBook with a 2.3GHz Intel Core i5 CPU.
\begin{figure*}[h]
\centering
\subfigure[Instance 1]{
\includegraphics[width = 0.3\textwidth]{Instance1.png}
}
\subfigure[Instance 2 ]{
\includegraphics[width = 0.3\textwidth]{Instance2.png}
}
\subfigure[Instance 3]{
\includegraphics[width = 0.3\textwidth]{Instance3.png}
}
\caption{Bar graph for DE algorithm with single chance constraint }
\label{fig:bargraph}
\end{figure*}
Table \ref{tab:singleResults} lists the results for the three instances with using fitness function (\ref{fit:deter},\ref{fit:Cuonly}) and (\ref{fit:flonly}) separately. Figure \ref{fig:bargraph} shows the how the chance-constrained bound $\alpha_{Cu}$ or $\alpha_{Fl}$ affects the quality of the solutions. The bars in the graphs are corresponding to the solutions of instances combining with the confidence of chance constraint respectively, and the three bars in each group corresponding to the threshold of confidence $\{0.999,0.99,0.9\}$. Among others, we observe that results obtained by applying the fitness function (\ref{fit:Cuonly}) are significantly affected by the value of $\alpha_{Cu}$. The results show an increasing trend as the value of $\alpha_{Cu}$ decrease. However, by observing the bars in \textit{Fl chance constraint} group, the value of $\alpha_{Fl}$ does not influence the result when using the Fl chance constraint.
As can be seen from Table \ref{tab:singleResults}, the success rate shows significantly difference between using the fitness functions (\ref{fit:Cuonly}) and (\ref{fit:flonly}) for instance $1$ and $2$. When the confidence of the chance constraint (\ref{chance:coppergrade}) is tight such as $0.999$, the DE algorithm can not generate a pure feasible population in the last generation. While the confidence of the chance constraint (\ref{chance:Flrecovery}) does not influence the success rate of the algorithm. However, for instance $3$, which has four parcels into consideration and is the most complex instance in our study, the DE algorithm fails to obtain a feasible population in the last generation when the value of $\alpha_{Fl}$ is $0.999$.
Table \ref{tab:MultiResults} lists the results obtained by considering two chance constraints together, the fitness function (\ref{fit:combina}). For each instance, we investigate different parameters setting together with the different requirement on the chance constraints determined by $\alpha_{Cu}$ and $\alpha_{Fl}$. The results list in the columns with the same $\alpha_{Cu}$ shows that there is no significant difference between the solutions obtained by applying difference $\alpha_{Fl}$. Moreover, with the same $\alpha_{Fl}$, the object value increase while the $\alpha_{Cu}$ decrease.
Now, we compare the results obtained by using single chance constraint and combined chance constraints. Comparing the solutions list in the column \textit{Cu Chance constraint} and \textit{Fl Chance constraint} in Table \ref{tab:singleResults} against that of the combined chance constraint in the same value of $\alpha_{Cu}$ and $\alpha_{Fl}$ respectively. We find that for the same instance, the results obtained by applying a single chance constraint are better than the combined chance constraints which happened in most cases. One interesting finding is that the value of $\alpha_{Fl}$ does not show significant effects on the results in the experiments for results in Table \ref{tab:singleResults} and \ref{tab:MultiResults}. A possible explanation for this might be that the parameters of the instances are not reliable or match the real-world situation, which can indicate the malfunction of the constraint. This is an important issue for feature research that develops approaches to create a benchmark that more reliable or more close to the real-world situation for the stockpile blending problem.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we consider the stockpile blending problem which is an important component in mine scheduling with the uncertainty in the geologic input data. We modeled the stockpile blending problem as a nonlinear optimization problem and introduced the chance constraints to tackle the stochastic material grades. We show how to incorporate a well-known probability tail, Chebyshev's inequality, into presenting the surrogate functions of the chance constraints. Furthermore, we designed the four fitness functions with considering different chance constraints. In our experiments, which have covered a variety of instances according to the parameters, we have observed that the confidence of the Cu chance constraint affects the results obtained by using the fitness function considering the Cu chance constraint and the fitness function with combined chance constraints. Due to the ineffectiveness of the confidence of the Fl chance constraint, for further studies, it could be interesting to deeply investigate the relationship between chance constraints. It would be also interesting to develop benchmarks for the stockpile blending problem with chance constraints as there is no available open access data-set.
\section{Acknowledgements}
This research has been supported by the SA Government through the PRIF RCP Industry Consortium
\bibliographystyle{abbrv}
|
\section{Introduction}
Many real-world data can be organized in the form of graphs, such as social relations and protein structure.
Effective graph analysis provides users a deeper understanding of what is behind the data and can help to analyze many natural phenomena and build powerful commercial applications.
However, it is not trivial to utilize the classical machine learning models to analyze relational data with a more complex structure.
These models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), are designed to extract fine-grained representation for each data sample from its own feature.
To fully utilize the rich information of graph data, a new family of machine learning (ML) models, namely graph neural networks (GNNs)~\cite{KW17,HYL17,VCCRLB18,XHLJ19}, has been introduced to address graph-related tasks in an end-to-end manner.
GNN models utilize both the feature of each sample (referred to as a node in the GNN context) and the features of the sample's neighborhood (from the graph) to represent the sample.
In this way, a GNN learns to embed the structural connections among different nodes in its training dataset.
Recent research has shown that ML models are vulnerable to privacy attacks~\cite{SSSS17,SZHBFB19,FJR15,LF20,CLEKS19,MSCS19,SS19,SS20,CYZF20,CTWJHLRBSEOR20}.
Most of the current efforts in this direction concentrate on ML models trained on sensitive data from the Euclidean space, such as images and texts.
Meanwhile, graph data, which is used to train GNNs, also contains sensitive information, such as social relations~\cite{BDK07,CBCSHK10,JWZG17} and mobility traces~\cite{CML11,BHPZ17}.
However, the potential privacy risks stemming from GNNs have been largely understudied.
\subsection{Our Contributions}
In this paper, we investigate whether a GNN model is vulnerable to membership inference attacks~\cite{SSSS17,SZHBFB19,LF20,SS19,HMDC19,CYZF20}, the major means to assess ML models' privacy risks.
Specifically, an adversary aims to infer whether a target node is used in the training dataset of a target GNN.
We concentrate on black-box membership inference, the most difficult setting for the adversary~\cite{SSSS17}.
There exists some preliminary work on privacy risks of GNNs~\cite{DBS20,ONK21,HJBGZ21}.
For instance, He et al.~\cite{HJBGZ21} propose a link stealing attack to infer the graph structure of a trained GNN model.
Also, Duddu et al.~\cite{DBS20} and Olatunji et al.~\cite{ONK21} have performed some preliminary studies on node-level membership inference attacks against GNNs.
However, the former lacks a clear attack methodology, while the latter conducts attacks in a restricted scenario, i.e., using a target node's 2-hop subgraph to query the target model to obtain the input to their attack model (see \autoref{subsection:AttackMethodology}), which falls short of providing a complete picture of GNN's membership inference risks.
As mentioned before, GNNs are designed for graph data that is not in the Euclidean space, which leads to some unique research questions for membership inference attacks in this setting.
First, an adversary needs background knowledge, such as the target GNN's architecture and a shadow dataset, to train their attack model.
State-of-the-art GNN models are normally shallow with less diverse choices of model architectures compared to CNNs and RNNs due to the fact that real-world graphs normally exhibit small-world phenomenon~\cite{EK10}.
Also, different graphs share many common properties, such as power-law degree distribution~\cite{LRU14}.
This motivates us to understand whether an adversary can have less constrained background knowledge compared to previous membership inference attacks against other types of ML models.
Second, an adversary can query a target node to a target GNN with either the node's feature alone or the node and its neighborhood's graph connections as well as their features.
This means one node can receive two different prediction outputs (posteriors) from the target GNN.
We are interested in which posteriors reveal more information of the target node's membership status and whether these two posteriors can be combined to achieve a more effective attack.
To answer these research questions, we make the following contributions in this paper.
We first systematically define the threat model of node-level membership inference attack against GNNs by categorizing an adversary's background knowledge along three dimensions, i.e., shadow dataset, shadow model, and node topology.
Specifically, we assume that an adversary may have a shadow dataset that comes from the same or different distribution of the target model's training dataset.
They can also establish a shadow model that has the same or different architecture from the target model.
Regarding node topology, we consider two situations: 1) the adversary only uses the target node's feature itself to query the target model (0-hop query) or 2) the adversary uses the target node and its 2-hop subgraph's information to query the target model (2-hop query).
Following the different threat models based on node topology, we propose three membership inference attack models, namely 0-hop attack, 2-hop attack, and combined attack.
We perform an extensive evaluation on three popular GNN models including GraphSAGE~\cite{HYL17}, Graph Attention Network (GAT)~\cite{VCCRLB18}, and Graph Isomorphism Network (GIN)~\cite{XHLJ19} with four benchmark datasets, i.e., Cora~\cite{KW17}, Citeseer~\cite{KW17}, Cora-full~\cite{BG182}, and LastFM Asia~\cite{RS20}.
Experimental results show that our attacks achieve strong performance.
For instance, our 0-hop attack achieves 0.791 accuracy on the GraphSAGE model trained on Citeseer.
More interestingly, we discover that our 0-hop attack has better performance than the 2-hop attack.
This is due to the fact a target node's 2-hop neighborhood contains a mixture of member and non-member nodes which jeopardizes the attack model's accuracy.
Also, our combined attack achieves the strongest performance by taking advantage of both the 0-hop and 2-hop attacks.
Moreover, we show that when the adversary does not know the target model's training dataset distribution or architecture, our attack is still effective.
We perform an in-depth analysis of the success behind the attack.
Our experiments reveal that a target node with higher subgraph density is more prone to membership inference.
This is due to the fact that a dense subgraph drives the node to participate more in the aggregation process of the GNN training, which amplifies the node's influence in the target GNN model.
Besides, it is easier for the adversary to mount their attack if a node shares similar features with its neighbors.
We propose two defense mechanisms to mitigate the membership inference risks of GNNs.
Empirical evaluation shows that they are able to mitigate the attack performance to a certain extent while bringing moderate utility damage.
This motivates us to further investigate advanced defenses in the future.
\subsection{Organization}
The rest of the paper is organized as the following.
In \autoref{section:gnn}, we provide some background knowledge of graph neural networks.
\autoref{section:MIAAgainstGNN} presents the threat model and attack methodology.
In \autoref{section:Evaluation}, we discuss our empirical evaluation results.
\autoref{section:RelatedWork} summarizes the related work and \autoref{section:Conclusion} concludes the paper.
\section{Graph Neural Networks}
\label{section:gnn}
In this section, we first introduce the notations used in the paper.
Then, we introduce the three representative GNN architectures we focus on.
In the end, we discuss the prediction process of GNN.
\subsection{Notations}
We define a graph dataset as $\mathcal{D}=(\mathcal{G}, \mathcal{X}, \mathcal{Y})$.
Here, $\mathcal{G}=(\mathcal{V}, \mathcal{E})$ represents a graph with $\mathcal{V}$ denoting the graph's set of nodes and $\mathcal{E}$ representing the set of edges connecting these nodes.
Each node is denoted by $v \in \mathcal{V}$ and $e_{uv} \in \mathcal{E}$ represents an edge linking two nodes $u$ and $v$.
$\mathcal{X}=\{x_1,x_2,...,x_{\vert\mathcal{V}\vert}\}$ and $\mathcal{Y}= \{y_1, y_2,...,y_{\vert\mathcal{V}\vert}\}$ represent the features and labels for all the nodes in $\mathcal{G}$, respectively.
Node $v$'s $l$-hop neighborhood is denoted by $\mathcal{N}^{l}({v})$, which contains a set of nodes at a distance less than or equal to $l$ from $v$ in $\mathcal{G}$.
For convenience, we abbreviate the $1$-hop neighborhood of $v$ as $\mathcal{N}({v})$.
The $l$-hop subgraph of node $v$, denoted by $g^{l}(v)$, contains $v$ and its $l$-hop neighborhood $\mathcal{N}^{l}({v})$, edges among these nodes, and features of these nodes.
We summarize the main notations in ~\autoref{table:notion}.
\begin{table}[!t]
\caption{List of notations.}
\label{table:notion}
\centering
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{r|l}
\toprule
Notation & Description \\
\midrule
$\mathcal{D} =(\mathcal{G},\mathcal{X},\mathcal{Y})$ & Dataset\\
$v$ & A node\\
$\mathcal{N}^{l}({v})$ & ${l}$-hop neighborhood of $v$\\
$g^{l}(v)$ & ${l}$-hop subgraph of $v$\\
$h_{v}^{(t)}$ & Representation vector of $v$ at layer $t$\\
$\mathcal{M}_{\textit{T}}$ & Target model\\
$\mathcal{D}_{\textit{Target}}$ & Target dataset \\
$\mathcal{D}_{\textit{Target}}^{\textit{Train}}$ ($\mathcal{D}_{\textit{Target}}^{\textit{Test}}$) & Target training (testing) dataset \\
$\mathcal{M}_{\textit{S}}$ & Shadow model\\
$\mathcal{D}_{\textit{Shadow}}$ & Shadow dataset \\
$\mathcal{D}_{\textit{Shadow}}^{\textit{Train}}$ ($\mathcal{D}_{\textit{Shadow}}^{\textit{Test}}$) & Shadow training (testing) dataset \\
\bottomrule
\end{tabular}
\end{table}
\subsection{GNN Architecture}
\label{subsection:gnn_architecture}
In general, there are two settings for GNNs, i.e., \textit{transductive} setting and \textit{inductive} setting.
In the transductive setting, a GNN (e.g., vanilla GCN~\cite{KW17}) is trained and tested on the same fixed graph.
It means, in the testing phase, the GNN model can only provide predictions for nodes that are in its training dataset.
Therefore, membership inference attacks against transductive GNN models are trivial.
In this paper, we focus on the inductive setting of GNNs, whereby a GNN model can classify nodes that are not from its training dataset.
Basically, a GNN contains several graph convolution layers that iteratively update a node $v$'s representation by aggregating the representation of nodes in $v$'s neighborhood.
Formally, each graph convolution layer of a GNN model can be defined as follows:
\begin{equation}
\begin{array}{l}
z_{v}^{(t)} =\mathsf{AGGREGATE}(\{h_{u}^{(t-1)}: u \in \mathcal{N}(v)\}) \\
h_{v}^{(t)}=\mathsf{UPDATE}( z_{v}^{(t)})
\end{array}
\end{equation}
where $\mathcal{N}({v})$ is the neighborhood of $v$.
$t$ represents the $t$-th layer of the GNN.
$z_{v}^{(t)}$ and $h^{(t)}_{v}$ denote the hidden state and the representation vector of node $v$ at layer $t$.
In the first step, we initialize $v$'s representation $h_{v}^{(0)}$ as its feature $x_{v}$.
$\mathsf{AGGREGATE}(\cdot)$ and $\mathsf{UPDATE}(\cdot)$ are the aggregation and update functions, respectively.
For a given node $v$, the aggregation function is used to generate the current hidden state $z_{v}^{(t)}$ using a combination of its previous representation and the aggregated representation from its neighborhood $\mathcal{N}(v)$.
The update function then conducts non-linear transformation on the current hidden state $z_{v}^{(t)}$ and produces the representation vector $h^{(t)}_{v}$.
A multilayer perceptron (MLP) is usually used as the update function's structure.
Meanwhile, different GNN models may leverage different aggregation functions.
In this paper, we focus on three representative GNN architectures, i.e., GraphSAGE~\cite{HYL17}, Graph Attention Network (GAT)~\cite{VCCRLB18}, and Graph Isomorphism Network (GIN)~\cite{XHLJ19}.
\mypara{GraphSAGE}
Hamilton et al.~\cite{HYL17} propose GraphSAGE, which first generalizes the original graph convolutional network~\cite{KW17} to the inductive setting with different aggregation functions.
In this paper, we follow the widely used mean aggregation function of GraphSAGE, which can be defined as follows:
\begin{equation}
z_{v}^{(t)} = \operatorname{\mathsf{CONCAT}}(\mathrm{h}_{v}^{(t-1)}, \frac{1}{|\mathcal{N}(v)|}\sum_{u \in \mathcal{N}(v)}\mathrm{h}_{u}^{(t-1)})
\end{equation}
where $\operatorname{\mathsf{CONCAT}}$ is the concatenation operation.
\mypara{GAT}
Inspired by the attention mechanism in deep learning~\cite{VSPUJGKP17}, Velickovic et al.~\cite{VCCRLB18} propose GAT that leverages multi-head attention to learn different attention weights and pays more attention to important neighborhoods.
GAT's aggregation function can be formulated as:
\begin{equation}
z_{v}^{(t)} = \operatorname{\mathsf{CONCAT}}_{k=1}^{K} \sigma(\sum_{u \in \mathcal{N}(v)}\alpha_{uv}^{k} \mathbf{W}^{k} \mathrm{h}_{u}^{(t-1)})
\end{equation}
where $K$ is the total number of projection heads in the attention mechanism. $\mathbf{W}^{k}$ and $\alpha_{uv}^{k}$ are the weight matrix and the attention coefficient in the $k$-th projection head, respectively.
$\sigma(\cdot)$ is the activation function.
\mypara{GIN}
Xu et al.~\cite{XHLJ19} develop GIN whose representation power is well-matched with the Weisfeiler-Lehman test for graph isomorphism.
The aggregation function of GIN can be represented as:
\begin{equation}
z_{v}^{(t)}= (1+\epsilon^{(t)}) \cdot h_{v}^{(t-1)}+\sum_{u \in \mathcal{N}(v)} \mathrm{h}_{u}^{(t-1)}
\end{equation}
where $\epsilon$ is a learnable parameter to adjust the weight of node $v$.
\subsection{GNN Prediction}
In this paper, we focus on node classification tasks.
In the training phase, an inductive GNN learns the parameters of aggregation and update functions in different layers over a training dataset.
Then, to get a precise prediction of an unseen node $v$ in a $t$-layer GNN, we can feed $v$'s $t$-hop subgraph, i.e., $g^{t}(v)$, to the GNN and obtain the prediction posteriors $p_{v}$.
Note that the $t$-hop subgraph of $v$ is not a necessary condition to acquire the posteriors $p_{v}$.
We can obtain posteriors $p_{v}$ by only querying the target node $v$'s feature to the GNN model.
Our evaluation shows that even in this case, the GNN model can achieve better performance than MLP, i.e., a model that does not consider graph structural information (see \autoref{section:Evaluation}).
\section{Node-Level Membership Inference Against GNNs}
\label{section:MIAAgainstGNN}
In this section, we first define node-level membership inference attacks against GNNs.
Then, we discuss the threat model and present the attack methodology.
\subsection{Problem Definition}
\label{subsection:ProblemDefinition}
The goal of an adversary is to determine whether a given node is used to train a target GNN model or not.
More formally, given a target node $v$, a target GNN model $\mathcal{M}_{\textit{T}}$, and the adversary's background knowledge $\mathcal{K}$, node-level membership inference attack $\mathcal{A}$ is defined as the following.
\begin{equation}
\mathcal{A}: v, \mathcal{M}_{\textit{T}}, \mathcal{K} \mapsto \{\textit{member}, \textit{non-member}\}
\end{equation}
As discussed in the previous work~\cite{SSSS17}, successful membership inference attacks can cause severe privacy risks.
In the setting of GNNs, membership threat is related to graph data, such as inferring a user being a member of a sensitive social network.
\begin{figure*}[!t]
\centering
\includegraphics[width=2\columnwidth]{fig/system_figure.png}
\caption{A schematic overview of node-level membership inference attack against GNNs.}
\label{figure:attack_pipeline}
\end{figure*}
\subsection{Threat Model}
\label{subsection:ThreatModel}
Our target model $\mathcal{M}_{\textit{T}}$ is an inductive GNN model.
First, we assume that the adversary only has black-box access to the target model, i,e, they can only query the target model and obtain the posteriors.
As mentioned by previous work~\cite{SSSS17,SZHBFB19,HJBGZ21}, black-box setting is the most challenging scenario for the adversary.
We then categorize the adversary's background knowledge $\mathcal{K}$ along three dimensions, i.e., shadow dataset, shadow model, and node topology.
\mypara{Shadow Dataset}
We assume that the adversary has a shadow dataset $\mathcal{D}_{\textit{Shadow}}$ which contains its own graph structure as well as node features and labels.
Following the previous work~\cite{SSSS17}, the shadow dataset $\mathcal{D}_{\textit{Shadow}}$ can come from the same distribution of the target model's training dataset.
However, our empirical evaluation shows that this assumption can be relaxed (see \autoref{subsection:EvaluationRelaxAssumptions}).
Note that in both cases, the shadow dataset has no node and edge intersection with the target dataset.
\mypara{Shadow Model}
With the shadow dataset, the adversary can train a shadow GNN model $\mathcal{M}_{\textit{S}}$ to mimic the behaviors of the target model $\mathcal{M}_{\textit{T}}$.
We can assume that the shadow model shares the same architecture as the target model~\cite{SSSS17,SZHBFB19}.
In this case, the adversary needs to first perform a hyperparameter stealing attack to obtain the target model's architecture~\cite{WG18}.
Also, our experimental results show that an adversary can use a different GNN architecture from the target model to establish their shadow model (see \autoref{subsection:EvaluationRelaxAssumptions}).
\mypara{Node Topology}
To get the posteriors for $v$ from $\mathcal{M}_{\textit{T}}$, we consider two cases.
In the first case, we assume that the adversary only has $v$'s feature $x_v$.
As the input to a GNN needs to be in the form of a graph, we add a self-loop for $v$~\cite{KW17} and query the target model.
We refer to this case as a node's \textit{0-hop query}.
In the second case, we assume that the adversary knows the target node $v$'s 2-hop subgraph $g^{2}(v)$, which can be directly fed to the target model.
We name this scenario as a node's \textit{2-hop query}.
Note that $g^{2}(v)$ does not need to be the complete 2-hop subgraph of $v$ as the adversary may only have a partial view of the dataset.
Besides, nodes in $g^{2}(v)$ can be a mixture of members and non-members for the target GNN.
This is more realistic as the adversary does not know any other nodes' membership status.
The goal is to infer the membership status of $v$.
In this paper, we only consider the 0-hop and 2-hop queries since they are the two extreme querying cases where 0-hop query utilizes no information from the graph structure, while 2-hop query considers the complete graph structure (2-hop subgraph).\footnote{Most of the state-of-the-art GNNs follow two-layer structure due to the fact that real-world graphs normally exhibit small-world phenomenon~\cite{EK10}, and in this case, 2-hop query is the upper bound for the query depth.
Moreover, previous empirical results~\cite{HYL17} show that deeper GNN architecture does not further improve the classification performance.}
Indeed, 1-hop subgraph is also a possible node topology to the adversary.
We leave the investigation as our future work.
\subsection{Attack Methodology}
\label{subsection:AttackMethodology}
Following the standard process of membership inference attacks against ML models~\cite{SSSS17}, our attack can be divided into three stages, i.e., shadow model training, attack model training, and membership inference.
\autoref{figure:attack_pipeline} provides a schematic overview of the attack process.
\mypara{Shadow Model Training}
Given a shadow dataset $\mathcal{D}_{\textit{Shadow}}$, the adversary first splits its node set $\mathcal{V}_{\textit{Shadow}}$ into two disjoint sets, including $\mathcal{V}_{\textit{Shadow}}^{\textit{Train}}$ and $\mathcal{V}_{\textit{Shadow}}^{\textit{Test}}$.
Then, the adversary derives their shadow training ($\mathcal{D}_{\textit{Shadow}}^{\textit{Train}}$) and testing ($\mathcal{D}_{\textit{Shadow}}^{\textit{Test}}$) datasets by involving all the features, labels, and edges within $\mathcal{V}_{\textit{Shadow}}^{\textit{Train}}$ and $\mathcal{V}_{\textit{Shadow}}^{\textit{Test}}$, respectively.
After that, $\mathcal{D}_{\textit{Shadow}}^{\textit{Train}}$ is used to train a shadow GNN model $\mathcal{M}_{\textit{S}}$.
\mypara{Attack Model Training}
The attack model is a binary machine learning classifier and its input is derived from a node's posteriors provided by a GNN.
To obtain the training dataset for the attack model, the adversary needs to query $\mathcal{M}_{\textit{S}}$ with all the nodes in $\mathcal{V}_{\textit{Shadow}}$ (both $\mathcal{V}_{\textit{Shadow}}^{\textit{Train}}$ and $\mathcal{V}_{\textit{Shadow}}^{\textit{Test}}$) and gets the corresponding prediction posteriors.
As mentioned before, depending on their knowledge of node topology, the adversary can perform 0-hop query or 2-hop query.
For a node $v$, we refer to its posteriors obtained by 0-hop query (2-hop query) as \textit{0-hop posteriors} (\textit{2-hop posteriors}).
In this paper, we consider three types of attack model input summarized from posteriors which leads to three attack models, namely 0-hop attack~$\mathcal{A}_{0}$, 2-hop attack~$\mathcal{A}_{2}$, and combined attack~$\mathcal{A}_{c}$.
\begin{itemize}
\item \textbf{0-hop Attack.}
The 0-hop attack model is essentially an MLP, which takes $v$'s largest two\footnote{Classification tasks considered in this paper have at least two classes.} values (ranked) in its 0-hop posteriors as the input.
\item \textbf{2-hop Attack.}
The 2-hop attack model is also an MLP, which takes $v$'s largest two values (ranked) in its 2-hop posteriors as the input.
\item \textbf{Combined Attack}
The combined attack model considers both the inputs for the 0-hop and the 2-hop attack by first feeding them separately to different linear layers.
Then, the attack model concatenates the two embeddings and feeds them to an MLP.
\end{itemize}
Note that if the adversary can perform 2-hop attack of a given node, they can also perform 0-hop attack.
Therefore, the combined attack requires the same background knowledge as the 2-hop attack.
In all cases, if $v\in\mathcal{V}_{\textit{Shadow}}^{\textit{Train}}$, we label it as a member, otherwise as a non-member.
In the end, the adversary constructs an attack training dataset, which they use to train their attack model.
\mypara{Membership Inference}
To determine whether a target node is used to train the target model $\mathcal{M}_{\textit{T}}$, the adversary first conducts 0-hop query or 2-hop query to the target model depending on their background knowledge.
Then, the adversary queries the attack model with the 0-hop posteriors, 2-hop posteriors, or both to get the node's membership prediction.
\section{Evaluation}
\label{section:Evaluation}
In this section, we perform a comprehensive measurement of the node-level membership privacy risks stemming from GNN models.
We first introduce the experimental setup, then present the evaluation results for the attacks in different settings.
In the end, we evaluate the performance of possible defense mechanisms.
\subsection{Experimental Setup}
\mypara{Dataset}
We conduct experiments on four public datasets, including Cora~\cite{KW17}, Citeseer~\cite{KW17}, Cora-full~\cite{BG182}, and LastFM Asia~\cite{RS20} (abbreviated as Lastfm).
Cora and Citeseer are citation graphs whose nodes represent papers and edges reflect citation relationships among papers.
Cora-full is an extended Cora dataset.
Lastfm is a social network dataset with its nodes being users and edges representing users' mutual following relationships.
All datasets contain node features and labels.
Dataset statistics are summarized in \autoref{table:dataset_statistics}.
\begin{table}[!t]
\caption{Dataset statistics.}
\label{table:dataset_statistics}
\centering
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{c|c c c c}
\toprule
Dataset & \#. Node & \#. Edge & \#. Feature & \#. Class \\
\midrule
Cora & 2,708 & 5,429 & 1,433 & 7 \\
Citeseer & 3,327 & 4,732 & 3,703 & 6\\
Cora-full & 19,793 & 65,311 & 8,710 & 70\\
Lastfm & 7,624 & 27,806 & 7,842 & 18\\
\bottomrule
\end{tabular}
\end{table}
\mypara{Dataset Configuration}
The dataset configuration process is depicted in \autoref{figure:attack_pipeline}.
For each dataset, we first randomly split its nodes by half.
The first half (including the nodes, the edges among the nodes, and the nodes' features and labels) is used to construct the target dataset, i.e., $\mathcal{D}_{\textit{Target}}$.
The other half is treated as the shadow dataset, i.e., $\mathcal{D}_{\textit{Shadow}}$.
Note that the target dataset and shadow dataset are disjoint as mentioned in \autoref{section:MIAAgainstGNN}.
For the target dataset $\mathcal{D}_{\textit{Target}}$, we further randomly split it by half creating the target training dataset $\mathcal{D}_{\textit{Target}}^{\textit{Train}}$ and the target testing dataset $\mathcal{D}_{\textit{Target}}^{\textit{Test}}$.
The target training dataset is used to train the target model, and the target testing dataset is used to test the target model's performance with respect to its original classification task.
Both $\mathcal{D}_{\textit{Target}}^{\textit{Train}}$ and $\mathcal{D}_{\textit{Target}}^{\textit{Test}}$ are used to test membership inference.
Nodes in $\mathcal{D}_{\textit{Target}}^{\textit{Train}}$ are considered as members and nodes in $\mathcal{D}_{\textit{Target}}^{\textit{Test}}$ as non-members.
As mentioned in \autoref{subsection:ThreatModel}, for the 2-hop query scenario, each node's 2-hop subgraph can contain a mixture of member and non-member nodes.
We apply the same processing procedure on the shadow dataset to generate the shadow training dataset $\mathcal{D}_{\textit{Shadow}}^{\textit{Train}}$ and the shadow testing dataset $\mathcal{D}_{\textit{Shadow}}^{\textit{Test}}$.
$\mathcal{D}_{\textit{Shadow}}^{\textit{Train}}$ is used to train the shadow model.
Both $\mathcal{D}_{\textit{Shadow}}^{\textit{Train}}$ and $\mathcal{D}_{\textit{Shadow}}^{\textit{Test}}$ are used to derive the training dataset for the attack model.
\mypara{Metric}
We use accuracy as our evaluation metric for both target model's performance and attack model's performance as it is widely used in node classification tasks~\cite{KW17,VCCRLB18,XHLJ19} as well as membership inference attacks~\cite{SSSS17,SZHBFB19}.
\mypara{Target Models}
We leverage three GNN architectures, i.e., GraphSAGE, GAT, and GIN, to construct our target models and shadow models.
For each target model, we set the number of layers to 2 and the number of neurons to 32 in the hidden layer.
Additionally, GAT models require the specification of the number of heads in the multi-head attention mechanism.
We set the number of heads for the first layer and the second layer to 2 and 1, respectively.
We also use dropout in all hidden layers to reduce overfitting, and the dropout rate is 0.5.
We adopt cross-entropy as the loss function and Adam as the optimizer.
The learning rate is set to 0.003.
The target and shadow models are both trained for 200 epochs.
\mypara{Baseline Model}
We leverage a 2-layer MLP as the baseline model to perform the same task as the target model's original task.
Each hidden layer has 32 neurons with ReLU as its activation function.
Loss function, optimizer, epochs, and learning rate are identical to those of the target GNN models.
\mypara{Attack Models}
For both 0-hop and 2-hop attacks, a 2-layer MLP is utilized as the attack model and the number of neurons in the hidden layer is set to 128.
Regarding the combined attack, the two inputs are first fed into two separated linear layers (with 64 neurons) simultaneously.
We then concatenate the two 64-dimensional embeddings and feed them to another linear layer for membership inference.
ReLU is adopted as the activation function for all the attack models.
Also, the loss function and optimizer are the same as the target model.
We set the learning rate to 0.001 and the training epochs to 500.
\mypara{Implementation}
Our code is implemented with PyTorch\footnote{\url{https://pytorch.org/}} and DGL.\footnote{\url{https://www.dgl.ai}}
The experiments are performed on an NVIDIA DGX-A100 server with Ubuntu 18.04 system.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{fig/target_task_performance.pdf}
\caption{The performance of original classification tasks when the target model's architecture is MLP or GraphSAGE (0-hop and 2-hop query).
The x-axis represents different datasets.
The y-axis represents original classification tasks' accuracy.}
\label{figure:target_model_performance}
\end{figure}
\begin{figure*}[!t]
\centering
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/tpfp_analy_Cora.pdf}
\caption{Cora}
\label{figure:tpfp_analy_Cora_log}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/tpfp_analy_Citeseer.pdf}
\caption{Citeseer}
\label{figure:tpfp_analy_Citeseer_log}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/tpfp_analy_Cora-full.pdf}
\caption{Cora-full}
\label{figure:tpfp_analy_Cora-full_log}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/tpfp_analy_Lastfm.pdf}
\caption{Lastfm}
\label{figure:tpfp_analy_Lastfm_log}
\end{subfigure}
\caption{The ratio of true positive (TP), false positive (FP), true negative (TN), and false negative (FP) of different attacks for GraphSAGE on four different datasets.
The x-axis represents different attack types.
The y-axis represents the ratio.}
\label{figure:tpfp_analysis}
\end{figure*}
\subsection{Target Model Performance}
\label{sec:target_model_performance}
We first show the performance of the target models with respect to their original classification tasks in \autoref{figure:target_model_performance}.
To get the posteriors of a given node, we consider two query scenarios for each target model, i.e., 0-hop query and 2-hop query.
For comparison, we only consider a node's feature as the input to each baseline model (i.e., a 2-layer MLP) and perform the same classification task as the target model.
Due to space limitations, we only show the results for GraphSAGE.
Other GNN models exhibit similar trends.
First of all, compared to MLP, we observe that GNN has higher performance in the original task when using 2-hop queries.
For instance, on the Cora dataset, the baseline MLP achieves 0.684 accuracy while the GraphSAGE (2-hop) achieves 0.790 accuracy.
This demonstrates the efficacy of GNN models that consider nodes' features as well as their neighborhood information jointly for classification.
Second and more interestingly, 0-hop query on GraphSAGE also achieves better performance than MLP except for Citeseer.
This indicates that the graph information used during the training phase can be generalized to boost the performance of a GNN model even when it is queried with only a node's feature (0-hop query).
\begin{figure*}[!t]
\centering
\begin{subfigure}{0.66\columnwidth}
\includegraphics[width=\columnwidth]{fig/overfit_graphsage.pdf}
\caption{GraphSAGE}
\label{figure:overfitting_graphsage}
\end{subfigure}
\begin{subfigure}{0.66\columnwidth}
\includegraphics[width=\columnwidth]{fig/overfit_gat.pdf}
\caption{GAT}
\label{figure:overfitting_gat}
\end{subfigure}
\begin{subfigure}{0.66\columnwidth}
\includegraphics[width=\columnwidth]{fig/overfit_gin.pdf}
\caption{GIN}
\label{figure:overfitting_gin}
\end{subfigure}
\caption{The performance of 0-hop and 2-hop attacks for different GNN architectures on four different datasets under different overfitting levels.
The x-axis represents different overfitting levels.
The y-axis represents membership inference attacks' accuracy.}
\label{figure:overfitting}
\end{figure*}
\subsection{0-hop and 2-hop Attacks}
\label{subsection:EvaluationBaiscAttacks}
We first show the membership inference attack performance of the 0-hop and 2-hop attacks in \autoref{table:mia_performance_0hop} and \autoref{table:mia_performance_2hop}, respectively.
We find that compared to the 2-hop attack, the 0-hop attack achieves higher membership inference accuracy.
For instance, the 0-hop attack on GraphSAGE trained on Cora achieves 0.754 accuracy while the accuracy of the corresponding 2-hop attack is only 0.671.
Such observations reveal that a node's 2-hop query to the target GNN leaks less membership information of the node.
This observation is rather interesting since we show that the 2-hop query leads to better node classification accuracy in the original task (see ~\autoref{sec:target_model_performance}).
\begin{table}[!t]
\caption{The performance of 0-hop attacks for different GNN architectures on four different datasets.}
\label{table:mia_performance_0hop}
\centering
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{c|c c c}
\toprule
& \multicolumn{3}{c}{Target Model}\\
Dataset & GraphSAGE & GIN & GAT \\
\midrule
Cora & 0.754 & 0.741 & 0.757\\
Citeseer & 0.791 & 0.797 & 0.798\\
Cora-full & 0.754 & 0.748 & 0.728\\
Lastfm & 0.686 & 0.652 & 0.658\\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[!t]
\caption{The performance of 2-hop attacks for different GNN architectures on four different datasets.}
\label{table:mia_performance_2hop}
\centering
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{c|c c c}
\toprule
& \multicolumn{3}{c}{Target Model}\\
Dataset & GraphSAGE & GIN & GAT \\
\midrule
Cora & 0.671 & 0.601 & 0.662\\
Citeseer & 0.700 & 0.647 & 0.691\\
Cora-full & 0.723 & 0.617 & 0.639\\
Lastfm & 0.637 & 0.586 & 0.602\\
\bottomrule
\end{tabular}
\end{table}
To investigate the reason behind this, we visualize the ratio of true positive (TP), false positive (FP), true negative (TN), and false negative (FN) nodes for the 0-hop and 2-hop attacks in \autoref{figure:tpfp_analysis}.
We observe that both attacks achieve a similar true positive rate.
It is reasonable since if a target node is a member, then the target model gives a relatively confident prediction for both its 0-hop and 2-hop queries, and this confident prediction is exploited by both attack models to distinguish the node from non-members.
Meanwhile, the 2-hop attack has a higher ratio of FP (misclassifying non-members as members) than the 0-hop attack.
One reason might be a non-member node's 2-hop subgraph may contain some member nodes.
When the attack model makes a prediction for the non-member node with its 2-hop query, it aggregates the information from the member nodes that might exist in its 2-hop subgraph, thus yields a less accurate prediction.
Similar to previous work~\cite{SSSS17,SZHBFB19}, we measure the relationship between overfitting and attack performance.
The overfitting level is quantified by the difference between training accuracy and testing accuracy of the target model.
In \autoref{figure:overfitting}, we observe that the attack performance is strongly correlated with the overfitting level.
Specifically, in \autoref{figure:overfitting_graphsage}, for the 2-hop attack, the overfitting level for GraphSAGE on the Lastfm dataset is 0.164 and the attack accuracy is 0.637, while a higher overfitting level (0.261) and attack accuracy (0.700) can be observed on the Citeseer dataset.
Also, compared to 0-hop query, 2-hop query has a lower overfitting level, this is due to the fact that 2-hop query achieves better testing accuracy (see \autoref{figure:target_model_performance}).
\begin{figure*}[!t]
\centering
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/degree_auc_Cora.pdf}
\caption{Cora}
\label{figure:degree_auc_Cora}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/degree_auc_Citeseer.pdf}
\caption{Citeseer}
\label{figure:degree_auc_Citeseer}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/degree_auc_Cora-full.pdf}
\caption{Cora-full}
\label{figure:degree_auc_Corafull}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/degree_auc_Lastfm.pdf}
\caption{Lastfm}
\label{figure:degree_auc_Lastfm}
\end{subfigure}
\caption{AUC for 0-hop and 2-hop attacks on different groups of nodes categorized by degree on four different datasets.
The architecture of both target and shadow model is GraphSAGE.
The x-axis represents different groups, e.g, 0-25 means the group of nodes whose degrees are in the lowest 25\% of the dataset.
The y-axis represents the AUC.}
\label{figure:micro_degree_vs_auc}
\end{figure*}
\mypara{Node Property}
We next investigate which kinds of nodes are more prone to membership inference.
To this end, we calculate three metrics for each node, i.e., degree, ego density, and feature similarity.
The first two are related to a node's graph property and the last one focuses on the node's feature.
\begin{itemize}
\item \textbf{Degree.}
For a given node $v$, the degree of the node is defined as the number of edges connected to it.
\item \textbf{Ego Density.}
Ego density measures the graph density of a node $v$'s 2-hop subgraph $g^{2}(v)$.
\item \textbf{Feature Similarity.}
Feature similarity measures how similar a node $v$'s feature to nodes' features in its 2-hop subgraph $g^{2}(v)$.
Specifically, we calculate the similarity (cosine similarity) between the feature of $v$ and the feature of each node in $g^{2}(v)$.
Then, we average all the similarity.
\end{itemize}
\begin{figure*}[!t]
\centering
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/ego_density_auc_Cora.pdf}
\caption{Cora}
\label{figure:ego_density_auc_Cora}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/ego_density_auc_Citeseer.pdf}
\caption{Citeseer}
\label{figure:ego_density_auc_Citeseer}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/ego_density_auc_Cora-full.pdf}
\caption{Cora-full}
\label{figure:ego_density_auc_Cora-full}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/ego_density_auc_Lastfm.pdf}
\caption{Lastfm}
\label{figure:ego_density_auc_Lastfm}
\end{subfigure}
\caption{AUC for 0-hop and 2-hop attacks on different groups of nodes categorized by ego density on four different datasets.
The architecture of both target and shadow model is GraphSAGE.
The x-axis represents different groups, e.g, 0-0.25 means the group of nodes whose ego density values are in the range of 0 and 0.25.
The y-axis represents the AUC.}
\label{figure:micro_density_vs_auc}
\end{figure*}
\begin{figure*}[!t]
\centering
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/feature_similarity_auc_Cora.pdf}
\caption{Cora}
\label{figure:feature_similarity_auc_Cora}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/feature_similarity_auc_Citeseer.pdf}
\caption{Citeseer}
\label{figure:feature_similarity_auc_Citeseer}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/feature_similarity_auc_Cora-full.pdf}
\caption{Cora-full}
\label{figure:feature_similarity_auc_Cora-full}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/feature_similarity_auc_Lastfm.pdf}
\caption{Lastfm}
\label{figure:feature_similarity_auc_Lastfm}
\end{subfigure}
\caption{AUC for 0-hop and 2-hop attacks on different groups of nodes categorized by feature similarity on four different datasets.
The architecture of both target and shadow models is GraphSAGE.
The x-axis represents different groups, e.g, 0-0.25 means the group of nodes whose feature similarity values are in the range of 0 and 0.25.
The y-axis represents the AUC.}
\label{figure:micro_similarity_vs_auc}
\end{figure*}
We categorize all the nodes in $\mathcal{D}_{\textit{Target}}^{\textit{Train}}$ and $\mathcal{D}_{\textit{Target}}^{\textit{Test}}$, i.e., the attack model's testing dataset, into four different groups based on their degrees, ego density, and feature similarity, respectively.
The results are summarized in \autoref{figure:micro_degree_vs_auc}, \autoref{figure:micro_density_vs_auc}, and \autoref{figure:micro_similarity_vs_auc}.
Note that the distribution of member and non-member nodes in each group is not uniform, thus we utilize AUC (area under the ROC curve) to measure the attack performance in each group as AUC is not sensitive to imbalanced classes~\cite{BHPZ17,FLJLPR14}.
In general, we find that higher degree leads to lower AUC score for both 0-hop and 2-hop attacks (see \autoref{figure:micro_degree_vs_auc}).
For instance, for GraphSAGE trained on the Cora dataset, the 0-hop attack's AUC is 0.849 on nodes in the lowest 25\% degree group while the AUC is 0.775 in the highest 25\% degree group.
Recall that during the training process, each GNN layer generates a node's representation by aggregating its neighbor nodes' representation.
With a higher degree, more neighbor nodes are involved, which may reduce the ``exposure'' of the target node itself, thus lesser membership inference risk.
In \autoref{figure:micro_density_vs_auc}, we find that larger ego density implies higher attack performance.
For instance, for GraphSAGE trained on Cora-full, the 0-hop attack achieves 0.799 AUC on nodes with less than 0.25 ego density while the AUC increases to 0.867 for nodes with larger than 0.75 ego density.
The reason behind this can be credited to the aggregation function of GNN models.
Higher density enables a node to participate more times in the aggregation process during training, which results in the model memorizing more information about the node.
Also, if the density of a node's 2-hop subgraph is high, then all the nodes in the subgraph are more likely to share similar features as the node, following social homophily theory~\cite{EK10}.
This further amplifies the influence of the node in the model.
We further measure the relation between attack performance and feature similarity (see \autoref{figure:micro_similarity_vs_auc}).
Our finding reveals that membership inference is indeed more effective when the target node has a larger feature similarity with its neighbors.
For GraphSAGE trained on Citeseer, the 2-hop attack's AUC increases from 0.657 to 0.829 when the feature similarity increase from less than 0.25 to larger than 0.75.
\subsection{Combined Attacks}
\label{subsection:EvaluationCombinedAttacks}
\begin{figure*}[!t]
\centering
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/attack_Cora.pdf}
\caption{Cora}
\label{figure:attack_Cora}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/attack_Citeseer.pdf}
\caption{Citeseer}
\label{figure:attack_Citeseer}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/attack_Cora-full.pdf}
\caption{Cora-full}
\label{figure:attack_Corafull}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/attack_lastfm.pdf}
\caption{Lastfm}
\label{figure:attack_lastfm}
\end{subfigure}
\caption{The performance of membership inference attacks against different target models' architectures on four different datasets.
The x-axis represents different target models' architectures.
The y-axis represents membership inference attacks' accuracy.}
\label{figure:attack_performance}
\end{figure*}
\begin{figure*}[!t]
\centering
\begin{subfigure}{2\columnwidth}
\includegraphics[width=\columnwidth]{fig/embedding_Citeseer.pdf}
\caption{Non-members}
\label{figure:embedding_nonmem_Citeseer}
\end{subfigure}
\begin{subfigure}{2\columnwidth}
\includegraphics[width=\columnwidth]{fig/embedding_mem_Citeseer.pdf}
\caption{Members}
\label{figure:embedding_mem_Citeseer}
\end{subfigure}
\caption{The average embeddings of non-member and member nodes obtained from the combined attack model's hidden layer.
Both the target and shadow models are GraphSAGE trained on Citeseer.
The left (right) parts are the embeddings generated by the input of the 0-hop (2-hop) attack.}
\label{figure:embedding_combine_attack}
\end{figure*}
The combined attack takes both the inputs of the 0-hop and 2-hop attack models as its input.
Its performance is summarized in \autoref{figure:attack_performance}.
In most cases, we observe that the combined attack reaches the highest membership inference accuracy compared to the 0-hop and 2-hop attacks.
For instance, when the target model is GAT trained on Citeseer, the membership inference accuracy is 0.825 for the combined attack, while only 0.798 and 0.691 for the 0-hop and 2-hop attacks, respectively.
\autoref{figure:tpfp_analysis} further shows the ratio of TP, FP, TN, and FN for the combined attack.
We also visualize the embeddings of non-member and member nodes obtained from the combined attack model's hidden layer (when the target model is GraphSAGE trained on Citeseer) in \autoref{figure:embedding_combine_attack}.
The hidden layer embedding is a 128-dimensional vector, where the first 64 dimension corresponds to the input to the 0-hop attack model (referred to as 0-hop embedding), and the second 64 dimension corresponds to the input to the 2-hop attack model (referred to as 2-hop embedding).
We use an orange dash line to separate them for better visualization.
In \autoref{figure:embedding_nonmem_Citeseer}, the first row represents the average embeddings of the non-member nodes that are correctly predicted by both the 0-hop and 2-hop attacks.
The second (third) row represents the average embeddings of the non-member nodes that are correctly (wrongly) predicted by the 0-hop attack but wrongly (correctly) by the 2-hop attacks.
We can see that the left parts of the first row and the second row (0-hop embedding) are similar, while the right parts are relatively different.
This indicates that for those non-member nodes that are classified correctly by the 0-hop attack but wrongly by the 2-hop attack, the combined attack is able to follow the prediction of the 0-hop attack.
On the other hand, when the non-member nodes are classified correctly by the 2-hop attack but wrongly by the 0-hop attack, the combined attack follows the prediction of the 2-hop attack as the right parts of the third row resemble the right parts of the first row.
We observe similar trends for member nodes (see \autoref{figure:embedding_mem_Citeseer}).
In conclusion, the combined attack has better performance since it takes the advantage of both 0-hop and 2-hop attacks.
\subsection{Relax Assumptions}
\label{subsection:EvaluationRelaxAssumptions}
\begin{table}[!t]
\caption{The performance of combined attacks when using different distribution shadow datasets to train the shadow models.
Both the target and shadow models' architecture is GraphSAGE.}
\label{table:attack_performance_different_shadow_dataset}
\centering
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{c|c c c c}
\toprule
& \multicolumn{4}{c}{Shadow Dataset}\\
Target Dataset & Cora & Citeseer & Cora-full & Lastfm \\
\midrule
Cora & 0.767 & 0.775 & 0.715 & 0.743\\
Citeseer & 0.791 & 0.801 & 0.764 & 0.773\\
Cora-full & 0.721 & 0.736 & 0.767 & 0.705\\
Lastfm & 0.696 & 0.693 & 0.643 & 0.687\\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[!t]
\caption{The performance of combined attacks when using different architectures to establish the shadow models.
The target model's architecture is GraphSAGE.
Both the target and shadow training datasets are from the same distribution.}
\label{table:attack_performance_different_shadow_model}
\centering
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{c|c c c}
\toprule
& \multicolumn{3}{c}{Shadow Model}\\
Dataset & GraphSAGE & GIN & GAT \\
\midrule
Cora & 0.767 & 0.759 & 0.742\\
Citeseer & 0.801 & 0.798 & 0.760\\
Cora-full & 0.767 & 0.753 & 0.717\\
Lastfm & 0.687 & 0.683 & 0.688\\
\bottomrule
\end{tabular}
\end{table}
We further investigate whether the two key assumptions made by our attacks (see \autoref{subsection:ThreatModel}) can be relaxed: 1) the adversary has a shadow dataset that comes from the same distribution as the target dataset, 2) the adversary has a shadow model with the same architecture as the target model.
\mypara{Different Shadow Dataset Distribution}
\autoref{table:attack_performance_different_shadow_dataset} shows the attack results when the shadow model is trained on a dataset from a different distribution.
We observe that in this case, our combined attack can still achieve similar or even better performance compared to the same distribution shadow dataset.
For instance, when the target model is GraphSAGE trained on Cora, the attack accuracy is 0.775 with Citeseer as the shadow dataset, while the accuracy is 0.767 with Cora as the shadow dataset.
This shows that even the adversary does not have the same distribution shadow dataset, they can still launch effective membership inference.
We further extract the embeddings of members and non-members from two combined attack models (one corresponds to the same distribution shadow dataset, the other corresponds to the different distribution shadow dataset), and project the embeddings into a 2-dimensional space using t-SNE~\cite{MH08}.
The results are shown in \autoref{figure:tsne_attack}.
In both cases, member and non-member nodes are easily separable.
\mypara{Different Shadow Model Architecture}
The results of using different shadow model architectures are summarized in \autoref{table:attack_performance_different_shadow_model}.
We see that a shadow model with a different architecture from the target model still yields good attack performance.
For instance, for GraphSAGE trained on Cora-full, the attack accuracy is 0.753 when the shadow model architecture is GIN while the original attack accuracy is 0.767.
We further investigate whether the different number of neurons in the shadow model affects the attack performance.
Concretely, we evaluate the case when the target model is GAT with 32 neurons in its hidden layer, and the shadow model is GraphSAGE with different numbers of neurons ranging from 16 to 128.
The results are depicted in \autoref{figure:different_shadow_neurons}.
We observe that the attack performance is relatively stable under different numbers of neurons.
\medskip
In conclusion, both the same distribution shadow dataset and same architecture shadow model assumptions can be relaxed, which further demonstrates the severe membership privacy risks of GNNs.
\begin{figure}[!t]
\centering
\begin{subfigure}{0.49\columnwidth}
\includegraphics[width=\columnwidth]{fig/tsne_Cora_same_shadow.pdf}
\caption{Cora as Shadow Dataset}
\label{figure:tsne_Cora_Cora}
\end{subfigure}
\begin{subfigure}{0.49\columnwidth}
\includegraphics[width=\columnwidth]{fig/tsne_Cora_different_shadow.pdf}
\caption{Citeseer as Shadow Dataset}
\label{figure:tsne_Cora_Citeseer}
\end{subfigure}
\caption{The embeddings of 100 randomly selected member and non-member nodes obtained from the combined attack's hidden layer.
We project them into a 2-dimensional space using t-SNE.
The target model is GraphSAGE trained on Cora.
The two shadow models are GraphSAGE trained on Cora or Citeseer.
Each point represents a node.}
\label{figure:tsne_attack}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{fig/attack_shadow_model_with_different_neurons.pdf}
\caption{The performance of membership inference attacks when the target model is GAT with 32 neurons in its hidden layer and the shadow model is GraphSAGE with different numbers of neurons in its hidden layer.
The x-axis represents the number of neurons in the shadow model's hidden layer.
The y-axis represents membership inference attacks' accuracy.}
\label{figure:different_shadow_neurons}
\end{figure}
\subsection{Possible Defenses}
\label{subsection:Defense}
\begin{figure*}[!t]
\centering
\begin{subfigure}{0.66\columnwidth}
\includegraphics[width=\columnwidth]{fig/defense_add_edge_target_performance_0-hop.pdf}
\caption{Target Model Performance (0-hop)}
\label{figure:defense_add_edge_target_performance_0hop}
\end{subfigure}
\begin{subfigure}{0.66\columnwidth}
\includegraphics[width=\columnwidth]{fig/defense_add_edge_target_performance_whole.pdf}
\caption{Target Model Performance (2-hop)}
\label{figure:defense_add_edge_target_performance_2hop}
\end{subfigure}
\begin{subfigure}{0.66\columnwidth}
\includegraphics[width=\columnwidth]{fig/defense_add_edge_attack_performance.pdf}
\caption{Attack Performance ($\mathcal{A}_{c}$)}
\label{figure:defense_add_edge_attack_performance}
\end{subfigure}
\caption{The performance of the target model's original task and membership inference attacks when applying random edge addition as the defense.
The x-axis represents different proportions of edges added.
Here, 2$\times$ means randomly adding in total 2 times more edges in the target training dataset.
The y-axis represents the accuracy of the target models' original classification tasks or membership inference attacks.
Note that we only show the results when GraphSAGE is used as the architecture for both target and shadow models.}
\label{figure:defese_add_edge}
\end{figure*}
\begin{figure*}[!t]
\centering
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/defense_label_only_attack_performance_Cora.pdf}
\caption{Cora}
\label{figure:defense_label_only_attack_performance_Cora}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/defense_label_only_attack_performance_Citeseer.pdf}
\caption{Citeseer}
\label{figure:defense_label_only_attack_performance_Citeseer}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/defense_label_only_attack_performance_Cora-full.pdf}
\caption{Cora-full}
\label{figure:defense_label_only_attack_performance_Corafull}
\end{subfigure}
\begin{subfigure}{0.50\columnwidth}
\includegraphics[width=\columnwidth]{fig/defense_label_only_attack_performance_lastfm.pdf}
\caption{Lastfm}
\label{figure:defense_label_only_attack_performance_lastfm}
\end{subfigure}
\caption{The performance of membership inference attacks when applying label-only output as the defense.
The x-axis represents different target models' architectures.
The y-axis represents the accuracy of membership inference attacks.}
\label{figure:defense_label_only_attack_performance}
\end{figure*}
To mitigate the membership inference attacks, we investigate two possible defense mechanisms, namely random edge addition and label-only output.
\mypara{Random Edge Addition}
In the first defense, we perturb the target training dataset's graph structure by randomly adding edges.
For the adversary, the shadow model is trained on the original shadow training dataset.
We evaluate the target models' performance with respect to the original classification task, i.e., utility, and the membership inference attack performance using the combined attacks.
Due to space limitations, we only show the results when both the target and shadow models use GraphSAGE as their architecture in \autoref{figure:defese_add_edge}.
Other models exhibit similar trends.
In \autoref{figure:defense_add_edge_attack_performance}, we observe that with more random edges added, the attack performance indeed drops.
For instance, the membership inference accuracy is 0.801 on the original Citeseer dataset, while the accuracy drops to 0.660 when 20 times more edges are added.
This indicates that adding random edges to the target training dataset can protect nodes' membership privacy.
As shown in \autoref{figure:micro_degree_vs_auc}, nodes with higher degree suffer less membership leakage risks, since the aggregation function of GNN during training aggregates more neighbor nodes' information and ``memorize'' less about the target node itself.
On the other hand, the target models also suffer large utility loss as shown in \autoref{figure:defense_add_edge_target_performance_0hop} and \autoref{figure:defense_add_edge_target_performance_2hop}.
For instance, the accuracy of original classification task is 0.819 on the original Lastfm dataset using 2-hop query, while the accuracy decreases to 0.733 when 20 times more edges are added (the corresponding attack accuracy drops from 0.687 to 0.584).
\mypara{Label-Only Output}
For the second defense, we let the target model only return the prediction label instead of posteriors.
In this case, we assume that the adversary knows the total number of classes of the target model.
The adversary first converts the prediction labels derived from the 0-hop and 2-hop queries into two one-hot vectors, respectively.
Then, the two vectors serve as the input to train the combined attack model $\mathcal{A}_{c}$.
The performance of membership inference attacks against different target models is shown in \autoref{figure:defense_label_only_attack_performance}.
We observe that on all the target models, membership inference attack accuracy decreases significantly.
For instance, on Cora-full, when both the target and shadow models' architectures are GraphSAGE, the membership inference accuracy of the original combined attack is 0.767, while the accuracy drops to 0.537 after applying the label-only output defense.
In addition, this defense can also limit the target model's utility as labels contain less information than posteriors.
We note that previous work~\cite{CTCP20,LZ20} investigate the label-only membership inference attack on non-GNN models.
However, its effectiveness on the proposed defense for GNN models remains unjustified and we leave it as our future work.
\medskip
In summary, our proposed defense mechanisms can reduce the risk of membership inference attack against GNN models.
However, they both limit the target model's utility.
In the future, we plan to investigate more advanced defense mechanisms.
\section{Related Work}
\label{section:RelatedWork}
\mypara{Membership Inference Attack}
Membership inference attacks aim at inferring membership of individual training samples of a target model to which an adversary has black-box access through a prediction API~\cite{SSSS17,SZHBFB19,NSH18,YGFJ18,HMDC19,NSH19,CXXLBKZ18,SS202,CLEKS19,LZ20}.
Most of the existing attacks focus on deep learning models that are trained on sensitive data from the Euclidean space, such as images and texts.
Shokri et al.~\cite{SSSS17} propose the first membership inference attack against machine learning models in the black-box setting.
The authors provide a general formulation of membership inference attack whereas the adversary trains multiple shadow models to mimic the target model's behavior with certain background knowledge of training data and leverages many attack models to conduct the attack.
Salem et al.~\cite{SZHBFB19} further relax several key assumptions from~\cite{SSSS17}, such as knowledge of the target model architecture, shadow dataset from the same distribution.
Yeom et al.~\cite{YGFJ18} discuss the relationship between overfitting and membership attacks.
Nasr et al.~\cite{NSH19} conduct a comprehensive study for membership inference attacks in both black-box and white-box settings.
To mitigate the attacks, some defense mechanisms~\cite{SSSS17,SZHBFB19,NSH18,JSBZG19} have been proposed.
Those strategies include using model stacking~\cite{SZHBFB19}, dropout~\cite{SZHBFB19}, adversarial training~\cite{NSH18}, jointly
maximize privacy and prediction accuracy~\cite{JSBZG19}, etc.
\mypara{Other Exploratory Attacks Against ML Models}
Besides membership inference, other exploratory attacks such as model inversion, attribute inference, and model stealing have been studied by many researchers.
In model inversion attacks~\cite{FLJLPR14,FJR15,ZJPWLS20}, an adversary aims to reconstruct input samples from a target ML model, i.e., model inversion enables the adversary to directly learn information about the training dataset.
Fredrikson et al.~\cite{FLJLPR14} first propose a model inversion attack in the setting of drug dose classification.
Later, Fredrikson et al.~\cite{FJR15} further extend model inversion to general ML settings relying on back-propagation.
More recently, Zhang et al.~\cite{ZJPWLS20} develop a more advanced attack based on GANs to synthesize the training dataset.
Attribute inference attacks~\cite{SS20,MSCS19} aim to infer some general properties of the training dataset.
Meils et al.~\cite{MSCS19} first show that collaborative machine learning can leak sensitive attributes about training data.
Song and Shmatikov~\cite{SS20} later demonstrate that having risks of attribute inference is an intrinsic feature of machine learning models, which is caused by overlearning.
The goal of model stealing attacks~\cite{TZJRR16,WG18,OSF19,JCBKP20} is to extract the parameters from a target model.
Tramer et al.~\cite{TZJRR16} propose the first model stealing attack, with black-box access to the target model.
Wang and Gong~\cite{WG18} propose hyperparameters stealing attacks for a variety of ML models, based on the observation that the gradient of the objective function at the value of parameters is close to 0.
More recently, Orekondy et al.~\cite{OSF19} propose a model stealing attack based on reinforcement learning, which relaxes assumptions on dataset and model architecture.
\mypara{Adversarial Attacks Against Graph Neural Networks}
Recent studies show that GNNs are susceptible to privacy and security attacks~\cite{ZAG18,BG192,DLTHWZS18,ZG19,WWTDLZ19,WG19,ZJWG20}.
Most of these attacks against GNNs are causative attacks where the adversary can manipulate the training dataset by introducing adversarial perturbations to node features, graph structure, etc.
In this direction, different adversarial attack strategies have been investigated.
Z{\"u}gner et al.~\cite{ZAG18} design adversarial examples on attributed graphs.
Bojchevski et al.~\cite{BG192} construct poisoning attacks on unsupervised node embeddings based on random walks.
Wang and Gong~\cite{WG19} perform the study on an adversarial attack for collective classification.
Z{\"u}gner and G{\"u}nnemann~\cite{ZG19} introduce training time attacks on GNNs.
The emerging GNN models~\cite{HYL17,XHLJ19, VCCRLB18} enable the adversary to launch exploratory attacks.
Unlike the causative attacks, these attacks do not intend to change the parameters of the target models.
Instead, the adversary probes the target models with carefully crafted input data and learn from the responses.
Though exploratory attacks on classical machine learning models have been extensively studied~\cite{TZJRR16,WG18,OSF19,JCBKP20,SS20,MSCS19,FLJLPR14,FJR15,ZJPWLS20}, only a few studies focus on exploratory attacks on GNNs~\cite{HJBGZ21,WYPY20,DBS20,ONK21}.
He et al.~\cite{HJBGZ21} propose the first link stealing attack to infer whether there is an edge between two nodes used to train the target GNN model.
Wu et al.~\cite{WYPY20} focus on the GCN model extraction attack given various levels of background knowledge.
Duddu et al.~\cite{DBS20} and Olatunji et al.~\cite{ONK21} have performed some preliminary studies on node-level membership inference attacks against GNNs.
However, Duddu et al.~\cite{DBS20} lack a clear attack methodology.
Olatunji et al.~\cite{ONK21} only conduct attacks in a restricted scenario, i.e., using a target node's 2-hop subgraph to query the target model to obtain the input to their attack model, which is deficient to provide a complete picture of the membership inference risks stemming from GNNs.
Also, compared to these work, we further perform an in-depth analysis of the factors that influence the attack performance, such as ego density and feature similarity, and investigate two defenses.
We refer the audience to other works~\cite{SDYWYHL18,DLTHWZS18,CLPXCXHZ20,JLXWT20,XMLDLTJ20} for comprehensive surveys of existing adversarial attacks and defenses on GNNs.
\section{Conclusion}
\label{section:Conclusion}
In this paper, we perform a comprehensive privacy risk assessment of graph neural networks through the lens of node-level membership inference attacks.
We systematically define the threat model along three dimensions, including shadow dataset, shadow model, and node topology, and propose three different attack models.
We conduct extensive experiments on three popular GNN models over four benchmark datasets.
Our evaluation results show that GNNs are indeed vulnerable to membership inference attacks even with minimal background knowledge of an adversary.
Moreover, our analysis reveals that a node's degree, ego density, and feature similarity have a large impact with respect to the attack performance.
We further show that the attacks are still effective even the adversary does not have the same distribution shadow dataset or same architecture shadow model.
To mitigate the attacks, we propose two possible defense mechanisms and discuss their trade-offs between membership privacy and model utility.
\bibliographystyle{plain}
|
\section{Introduction}
The field of teleoperation of robots and robot arms has seen a lot of activity since both collaborative robot arms and 6 degrees-of-freedom (DOF) input devices have become more affordable and more widely available \cite{lee2020}.
It has been shown that it is generally more intuitive, faster and less mentally exhausting for a human operator to operate a robot arm via head and hand tracking devices rather than via a touch interface, mouse or joystick \cite{rakita2017}.
As the kinematic structure of a robot arm is often different from that of a human arm, and because mostly only the hand position of the operator is tracked rather than every joint of the arm, it is not straightforward to find an end-effector pose that matches the positioning of the arm and the intent behind the operator's movement.
Furthermore, naïve end-effector pose matching can lead to singular or close to singular robot arm configurations which can lead to dangerously fast joint movements for small changes in the end-effector pose or can get the robot stuck in singular configurations, unable to reliably continue to match the operator's pose.
This problem gets accentuated in multi-robot settings, where not one single arm has to match an end-effector pose but instead multiple arms need to support and move a payload in a way that is either given by a predefined trajectory or by real-time human input.
Teleoperation of such a system can exhibit non-intuitive workspace constraints as well as additional singular configurations which arise from the system acting like a closed-loop kinematic chain.
In order for the operator to focus on high-level tasks it is important that the multi-robot system can reliably avoid these configurations and show a graceful degradation of pose-matching reliability, giving the operator the possibility to adapt rather than abruptly getting stuck when approaching singularities and workspace limits.
We first provide a short overview of related work in the field. We then introduce an improvement to traditional Inverse Kinematic (IK) solvers in which we relax one single rotational DOF and provide the solver with a prior configuration that locally maximises the manipulability index of each individual robot arm. This improves the system's ability to avoid dangerous discontinuities in joint configuration which can arise from near-singular configurations.
We then introduce the setup and how we are using our intuitive VR interface for payload manipulation. The interface lets the operator move freely around the virtual robots and modify the payload's pose by simply grasping and moving it. It therefore provides the operator with the ability to safely operate a group of robots remotely, or use it as a simulation tool for training or pre-defining tasks.
Finally, we show the results of applying our method to a group of 3 UR5 robots and 2 ABB YuMis respectively and provide an outlook on future developments.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{Figures/system.png}
\caption{The setup with 3 UR5 robot arms}
\label{fig:system}
\end{figure}
\section{Related Work}
In order to quantify the performance of a solution to the inverse kinematic problem, several measures have been proposed, such as the \emph{manipulability index} \cite{yoshikawa2020}, the \emph{condition number} \cite{salisbury1982} and many derivations and adaptions thereof (see e.g. \cite{patel2015} for an overview).
With the appearance of collaborative robots, the interest to operate robots in real-time and potentially close to humans has increased.
In combination with the availability of accurate and affordable 6DOF input devices, the field of teleoperation has seen a lot of exciting developments, drawing inspiration from motion capturing and animation techniques traditionally used for animated movies.
Although the single-arm case has been relatively well studied, dual- and multi-arm collaborative tasks have seen less activity, despite their vast potential for applications in construction, assembly or mobile multi-robot systems.
In the following, we review some of the closely related topics.
\subsection{Manipulability Index and Singularity Avoidance}
The capability of a robot arm to avoid singular configurations is especially important in the field of physical human-robot interaction (HRI), where a human operator, who cannot easily recognize potentially dangerous configurations, guides the robot e.g. during kinesthetic teaching.
Several frameworks have been proposed to steer the operator away from singular configurations using for example asymmetric damped least squares \cite{carmichael2017} or admittance control via virtual forces \cite{dimeas2016}.
These methods often use a \emph{manipulability} or conditioning measure to scale the generated forces depending on the proximity to a singularity and, in the asymmetric case, the sign of the \emph{gradient} of the manipulability.
Since the absolute distance to the singular configuration is often not required, scaling factors and thresholds which are tuned to work well in HRI applications are introduced.
A similar method of optimising a measure called \textit{parameter of singularity} in order to avoid singularities was used in \cite{franks2008}.
To still match the required end-effector trajectory as closely as possible, the authors consider the rotational axis of the tool that is held by the robot arm to be functionally redundant and thus have one redundant degree of freedom to optimise over.
A similar idea can be found in \cite{zimmermann2020}, where the optimal grasp position on a wooden block is chosen by parametrizing the gripper orientation and position on the block in the presence of obstacles, although the gripper position is not changed anymore once the block has been picked up.
A method that introduces the manipulability index as an optimization objective into a kinematic task was introduced in \cite{dufour2020}, where the authors showed an improved manipulability index for a predefined trajectory at the cost of reduced trajectory tracking.
They additionally combined this with obstacle avoidance to show that they are able to maintain a safe distance even with the additional objective.
\subsection{Teleoperation}
With recent advances in Virtual and Augmented Reality devices, virtual teleoperation of robots in 3D space has become more precise and much easier to implement.
Headsets such as the Microsoft Hololens 2 and Oculus Quest 2 are pushing to improve 6DOF head- and hand-tracking and several other headsets such as the HTC Vive are providing affordable and easy to use 6DOF controllers.
Since robots and robot arms are often kinematically different from their human operators, and since, with few exceptions, there is little to no haptic feedback for operators, several methods for intuitive mapping of operator motion and intent onto robotic systems have been proposed.
In \cite{lee2020}, Lee et al. introduced a method of unimanual and bimanual teleoperation of a robot arm, using Oculus Rift IR LED sensors and touch controllers
They showed that their method leads to subjective and objective improvements over traditional kinesthetic teaching methods on moderately challenging tasks.
Rakita et al. \cite{rakita2017} introduced a trade-off between IK and different goals such as obstacle or singularity avoidance.
This allows them to place more importance on avoiding bad configurations during faster and larger hand motions, while giving the operator more precise control over the end-effector pose when approaching or manipulating different objects.
They used HTC Vive controllers as the input device and show improvements over other input methods, such as a 6DOF stylus on several tasks of varying difficulty.
\subsection{Multi-arm manipulation}
While multi-arm coordination generally introduces an additional level of complexity by adding the need for re-grasping or limiting the available workspace if the robot arms are fixed in place, it also provides several benefits such as the modularity, redundancy and increased payload capabilities, while keeping each individual robot relatively simple and cost-efficient.
The growing availability of collaborative robot arms such as Universal Robots' UR5 and UR10 or the ABB YuMi promotes an increasing incentive to build teams of multiple, ready-made robots for tasks that one single robot cannot solve.
In \cite{xian2017}, the authors showed a method that enables the manipulation of large objects by considering two arms as a closed-chain system and introducing \emph{essentially mutually disconnected} components, which allow them to switch between different configurations via re-grasping operations. A method of coupling two robot arms via a virtual object is proposed in \cite{salehian2016}.
The paper focuses on the choice of a grasping position and the interaction of the system with a payload that is handed over by a human collaborator.
Another method that coordinates 4 robot arms by introducing a virtual manipulator is shown in \cite{dehio2018}. The virtual manipulator lets the authors tune the impedance of the system, making it more robust to disturbances such as added payload.
\begin{figure}[b]
\centering
\includegraphics[width=0.9\linewidth]{Figures/yumis.png}
\caption{The setup with 2 ABB YuMi robot arms}
\label{fig:systemYumi}
\end{figure}
\section{Method}
For many installation and construction tasks the robot arm's end-effector does not actually have to be in a precise location or orientation, as long as it can still guarantee the correct positioning and orientation of the payload.
Examples of this are the installation of glass panels using suction cups, or the positioning and fastening of pipes and tubes to the ceiling on construction sites.
Construction workers can adjust their grasp around the handles or pipes during installation in order to optimize leverage and accessibility.
We propose a method that optimises the manipulation index for each robot arm independently by utilizing one rotational degree of freedom for the optimisation, while limiting the maximum deviation from the fully constrained configuration in order to avoid losing grip or intersecting with the payload. Because the payload is handled by multiple robots simultaneously, its pose is always fully defined, even when opening up the rotational DOF on the robot arm.
Contrary to previous methods which consider manipulability as part of a global optimization problem (\cite{dufour2020}) or only use it to detect proximity to kinematic singularities (\cite{carmichael2017},\cite{dimeas2016}), we suggest an extremely simple algorithm that improves manipulability while not sacrificing performance and tracking fidelity.
\\
\textbf{Optimization based IK.}
\label{subsec:opt}
We begin by formulating an optimization based IK problem as follows
\begin{equation}\label{eq:IK}
\begin{aligned}
\min_{\mathbf{q}} \quad & w_{\text{err}}\|\mathcal{K}(\mathbf{q})-x\|^2 + w_{\text{reg}}\|\mathbf{q}-\mathbf{q}_0\|^2\\
\textrm{s.t.} \quad & \mathbf{q}_{\text{min}} < \mathbf{q} < \mathbf{q}_{\text{max}},
\end{aligned}
\end{equation}
where $\mathbf{q}$ are the stacked joint angles, $\mathbf{q}_{\text{min}}$ and $\mathbf{q}_{\text{max}}$ are the joints upper and lower limits, $\mathcal{K}(\mathbf{q})$ is the forward kinematics function for the pose of the robot's end effector, $\mathbf{x}$ is the target end effector pose, and $\mathbf{q}_0$ is predefined rest pose, which is used as regularization.
We solve this optimization problem using Newton's method.
Joint limits are handled as soft bound constraints using a barrier function
As an initial guess, we provide the solver with the previously computed optimal joint angles, which were based on the previous end effector target.
In the next section, we discuss a local optimization procedure that improve the manipulability of the individual arms.
\\
\textbf{Local manipulability optimization.}
Once \ref{eq:IK} is solved, we can still improve the manipulabilty of of the arms.
The manipulability index is defined by
\begin{equation}
m(\mathbf{q}) = \sqrt{\det(\mathbf{J}\mathbf{J}^T)}
\end{equation}
Our goal is to find a better orientation of the end effector w.r.t. a single rotation axis, as defined by the handles grasped by the end effector.
To this end, we could potentially solve another optimization problem using a gradient based method.
However, similarly to \cite{dimeas2016} can consider small modification to the joint angles, that, to first order, preserve the pose, except around this one rotation.
This is done via the generalized inverse of the velocity Jacobian,
\begin{equation}
\mathbf{J}^+ = \left(\mathbf{J}^T\mathbf{J}\right)^{-1}\mathbf{J}^T
\end{equation}
In order to compute the first order angle modification we use
\begin{equation}
\mathbf{q'_\pm} = \mathbf{q_0} \pm \mathbf{J^+}\mathbf{M} \Delta t
\end{equation}
where $M$ is a mask matrix (i.e. containing only zeros and ones) that selects the available DOF, and $\Delta t$ is a scaling factor.
In our case, that would be the 4th entry, which corresponds to rotation around the local $x$-axis.
In each iteration of our algorithm, we use these new joint angles $\mathbf{q'_+}$ and $\mathbf{q'_-}$ to compute the new manipulability index for each one of them.
If one of them is higher than the current manipulability we use the configuration as the new initial guess for the IK solver at the next iteration. If the increase in manipulability falls below a threshold value $\theta_m$ we discard the result in order to avoid oscillating around a locally optimal result and use the current configuration as the initial guess instead.
\section{Experimental Setup}
\subsection{Robots and Payload}
As our method does not depend on the specific structure of the robot arm, it can be applied to a wide variety of different types and groups with different numbers of robots. We validate our method on a setup of 3 6DOF UR5 robots arms and 2 7DOF ABB YuMis respectively due to their wide use in the robotics community and industry and their difference in DOFs as well as number of end effectors per robot. The UR5 setup is shown in Fig. \ref{fig:system}, while the YuMi setup is shown in Fig. \ref{fig:systemYumi}.
\subsection{Tasks}
We test our method on 2 predefined geometric tasks as well as real-time teleoperation scenarios. The geometric tasks consist of square and circular trajectories. The teleoperation tasks are carried out on a Oculus Quest VR headset, using the accompanying 6DOF controllers. In order to ensure repeatability with different configurations we prerecord a user-generated trajectory which we then stream to the controller for subsequent experiments. As proposed in \cite{rakita2018} we compare the positional errors, manipulability index and the joint velocities, accelerations and jerk over the different trajectories. \\
We compare all tasks on both the UR5 and the YuMi setups, using 3 different scenarios:
\begin{itemize}
\item [A)] IK with full constraints on position and orientation of the end effectors
\item [B)] IK with masked y and z orientations, allowing free rotation around the x axis
\item [C)] IK with masked y and z orientations, with our algorithm enabled
\end{itemize}
\subsection{Setup}
The payload is represented by a sphere of diameter 0.3m, to which one handle per robot end effector is attached at the equator of the sphere. The handles are initially facing their respective robot arms and are parallel to the floor, and the handle bar has a distance of 5cm from the payload surface. For all experiments the UR5 robot arms are placed in a circle of 1.5m diameter, facing the center of the circle. The YuMis are placed 1.2m apart, facing the center position.To ensure that the experiments are comparable, the weights for our energy function are the same for different trajectories and the different scenarios, namely $w_{\text{err}}=1000$, $w_{\text{reg}}=0.01$ and $w_{\text{lim}}=10\,000$. For our Newton minimizer we are using a maximum of 10 steps and a residual value of $10^{-5}$. For scenario B, we nullify the penalty for rotations around the x axis, as described in Section \ref{subsec:opt}. For scenario C, we additionally enable the manipulability optimization, using a $\Delta t = 0.007$ and a threshold $\theta_m = 0.0001$ for the UR5 experiments and a $\Delta t = 0.005$ and $\theta_m = 0.001$ for the YuMi experiments respectively. These values have been chosen empirically to avoid oscillations while still providing a sufficiently fast convergence towards a better manipulability value. An example of the initial setup for the 2 YuMi robots and the payload can be found in Fig. \ref{fig:yumi_manip}.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\linewidth]{Figures/yumi_comparison.png}
\caption{Comparison of the initial pose of the YuMi arms before (left) and after (right) improving manipulability via rotation around the handle x axis.}
\label{fig:yumi_manip}
\end{figure}
\begin{figure}[b]
\centering
\includegraphics[width=\linewidth]{Figures/vr_interface.png}
\caption{User interface inside the Oculus Quest. The payload pose can be modified via simple grasping with the controller.}
\label{fig:vr}
\end{figure}
\subsection{System}
All experiments were run on a Windows 10 machine with an Intel Core i7-9750H CPU @ 2.60GHz, 32GB RAM and an Nvidia Geforce RTX 2080 Max-Q GPU. \\
For the VR teleoperation we used an Oculus Quest 128GB headset and the accompanying 6DOF controller. The VR interface was built in Unity3D and a simple UDP implementation was used to wirelessly transmit manipulation data to the Windows machine. A typical view for an operator inside the headset is shown in Fig. \ref{fig:vr}.
\section{RESULTS and DISCUSSION}
\subsection{Circle trajectory}
\begin{figure}[t]
\centering
\includegraphics[width=0.705\linewidth]{Figures/manipVSerr.png}
\caption{Scenario A: Comparison of manipulability and position error for the circle trajectory on the 3 UR5 robot arms.}
\label{fig:manipVSerr}
\end{figure}
The payload is initialized in the middle between all robots. It then moves in positive x direction for 0.2m and finally moves in circles around the origin for a predefined number of steps. If they don't get stuck in a local optimum or because of a singular configuration the arms are theoretically able to meet the end effector goal pose for any point on the trajectory. For each experiment, the position error, velocity, acceleration and jerk values as well as the manipulability were measured over several iterations of the trajectory. The average and standard deviation for the different metrics are summarized in table \ref{t1}. The results indicate that both in scenario B and C the manipulability is improved substantially over scenario A. As shown in the accompanying video, the fully constrained IK solver is not able to continuously match the goal pose and exhibits discontinuities in joint configuration at several points in time. This coincides with the manipulability index approaching a zero value and leads to higher positional error for the end effectors, as can be seen in Fig. \ref{fig:manipVSerr}. While the mean manipulability and position error values as well as the standard deviations are slightly improved when using our method with the UR5 arms, acceleration and jerk are equal or even slightly worse, but generally in the same order of magnitude. Similar results can be seen on the YuMi. Although here we are able to improve the manipulability index as well, we have substantially higher positional errors in general and also when using our method, which might mostly be attributed to the difficulties of the YuMi to still reach far away points on the trajectory. In turn we can see a more substantial improvement in mean velocity, acceleration and jerk.
\subsection{Square trajectory}
Similar observations as for the circle trajectory can be made for the square trajectory. The problems of fully constraining the end effector pose are even more obvious here, as the abrupt changes in direction often lead to very high tracking errors, especially when they coincide with low manipulability. While for the UR5 the positional tracking is slightly improved when using our method, the YuMi benefits from lower velocity, acceleration and jerk values, although here improvements are not quite as substantial. This indicates that the sharp changes in velocity don't allow our algorithm to sufficiently quickly improve the manipulability. Therefore, while we achieve slightly higher values for the manipulability index, they do not translate as well to improvements in tracking as for the previous trajectory. Qualitative examples of the UR5's manipulability evolution are shown in Fig. \ref{fig:squareTraj} and the results are summarized in table \ref{t2}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{Figures/manipComparisonSquare.png}
\caption{Qualitative manipulability comparison for scenarios B and C of the square trajectory on the 3 UR5 robot arms}
\label{fig:squareTraj}
\end{figure}
\begin{table*}[ht]
\centering
\begin{tabular}{|c||c|c|c|c|c|}
\hline
& Pos. [mm] & Vel. $[\text{rad}/s]$ & Acc. $[\text{rad}/s^2]$& Jerk $[10^{-3}\text{rad}/s^3]$ & $m(\mathbf{q})$ \\
\hhline{|=#=|=|=|=|=|}
UR5 Scenario A & $0.32\pm 0.64$ & $0.12\pm 0.93$ & $0.037\pm 0.006$ & $0.94\pm 1.7$ & $0.043 \pm 0.03$ \\
\hline
UR5 Scenario B & $0.025\pm 0.015$ & $0.072\pm 0.03$ & $0.02\pm 0.003$ & $0.53\pm 0.81$ & $0.071 \pm 0.016$\\
\hline
UR5 Scenario C & $0.022\pm 0.013$ & $0.071\pm 0.02$ & $0.02\pm 0.003$ & $0.58\pm 0.89$ & $0.074 \pm 0.012$\\
\hhline{|=#=|=|=|=|=|}
YuMi Scenario A & $26.9 \pm 48$ & $0.178 \pm 0.178$ & $0.004 \pm 0.008$ & $0.72 \pm 1.4$ & $0.022 \pm 0.008$\\
\hline
YuMi Scenario B & $6.9 \pm 13.4$ & $0.143 \pm 0.065$ & $0.003 \pm 0.004$ & $0.58 \pm 0.89$ & $0.025 \pm 0.007$\\
\hline
YuMi Scenario C & $8.8 \pm 16.6$ & $0.036 \pm 0.052$ & $0.001 \pm 0.004$ & $0.10 \pm 0.52$ & $0.030 \pm 0.007$\\
\hline
\end{tabular}
\captionof{table}{Mean results for the circle trajectory}
\label{t1}
\end{table*}
\begin{table*}[ht]
\centering
\begin{tabular}{|c||c|c|c|c|c|}
\hline
& Pos. [mm] & Vel. $[\text{rad}/s]$ & Acc. $[\text{rad}/s^2]$& Jerk $[10^{-3}\text{rad}/s^3]$ & $m(\mathbf{q})$ \\
\hhline{|=#=|=|=|=|=|}
UR5 Scenario A & $2.3\pm 8.3$ & $0.12\pm 0.13$ & $0.0021\pm 0.01$ & $0.33\pm 1.9$ & $0.044 \pm 0.032$ \\
\hline
UR5 Scenario B & $0.026\pm 0.014$ & $0.077\pm 0.03$ & $0.0037\pm 0.002$ & $0.07\pm 0.42$ & $0.071 \pm 0.016$\\
\hline
UR5 Scenario C & $0.024\pm 0.012$ & $0.078\pm 0.03$ & $0.0055\pm 0.002$ & $0.13\pm 0.56$ & $0.074 \pm 0.014$\\
\hhline{|=#=|=|=|=|=|}
YuMi Scenario A & $52.5 \pm 108.7$ & $0.18 \pm 0.48$ & $0.007 \pm 0.062$ & $1.5 \pm 12$ & $0.022 \pm 0.008$\\
\hline
YuMi Scenario B & $8.0 \pm 16.0$ & $0.16 \pm 0.13$ & $0.0025 \pm 0.008$ & $0.53 \pm 1.5$ & $0.025 \pm 0.009$\\
\hline
YuMi Scenario C & $9.2 \pm 19.1$ & $0.05 \pm 0.14$ & $0.0019 \pm 0.014$ & $0.27 \pm 1.6$ & $0.028 \pm 0.008$\\
\hline
\end{tabular}
\captionof{table}{Mean results for the square trajectory}
\label{t2}
\end{table*}
\subsection{Teleoperation trajectory}
For the teleoperation trajectory, all parameters were kept the same as for the square and and circular trajectories. The trajectory was pre-recorded and then played back for the different scenarios. The results are summarized in table \ref{t3}. In this scenario, our method clearly performs best, improving the position errors for both the UR5 and the YuMi setup and leading to substantial improvements for the other metrics on the YuMi, while only exhibiting an outlier on the jerk values for the UR5. We believe that this is due to the natural movement of the user, pausing in-between different interactions with the payload, which gives our algorithm time to best adjust the manipulability index. We can also find an explanation for the outliers in Fig. \ref{fig:vrTraj}: The manipulability of arm 2 steadily declines in scenario B, whereas our method is able to make a small adjustment that improves manipulability around timestep 550 and consequently brings the arm into a better configuration. Later on, around timestep 1500, the user guides the payload in a way that makes arm 2 approach its workspace limit, sending the manipulability towards 0 very quickly. Although our method is not able to completely avoid the decline, it nevertheless minimizes the amount of time spent in the unfavorable regime, while the traditional method stays stuck until the user sufficiently adjusts the payload position. These constant adjustments are partially reflected in the mean measurements and also lead to the substantially higher standard deviations. For these cases a simple limitiation of the maximum acceleration would probably suffice to substantially improve the measured values for our method. Fig. \ref{fig:vrTrajYuMi} shows a similar situation for the teleoperation trajectory with the 2 YuMis.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{Figures/manipComparisonVR.png}
\caption{Qualitative manipulability comparison for scenarios B and C of the teleoperation trajectory on the 3 UR5 robots}
\label{fig:vrTraj}
\end{figure}
\begin{table*}[ht]
\centering
\begin{tabular}{|c||c|c|c|c|c|}
\hline
& Pos. [mm] & Vel. $[\text{rad}/s]$ & Acc. $[\text{rad}/s^2]$& Jerk $[10^{-3}\text{rad}/s^3]$ & $m(\mathbf{q})$ \\
\hhline{|=#=|=|=|=|=|}
UR5 Scenario A & $1.8\pm 9.9$ & $0.07\pm 0.12$ & $0.011\pm 0.017$ & $3.2\pm 4.8$ & $0.03 \pm 0.017$ \\
\hline
UR5 Scenario B & $0.97\pm 8.1$ & $0.05\pm 0.07$ & $0.007\pm 0.01$ & $1.8\pm 2.8$ & $0.054 \pm 0.023$ \\
\hline
UR5 Scenario C & $0.3\pm 3.8$ & $0.08\pm 0.6$ & $0.012\pm 0.12$ & $3.1\pm 30$ & $0.057 \pm 0.021$ \\
\hhline{|=#=|=|=|=|=|}
YuMi Scenario A & $27 \pm 46$ & $0.085 \pm 0.155$ & $0.008 \pm 0.015$ & $1.5 \pm 2.5$ & $0.020 \pm 0.010$\\
\hline
YuMi Scenario B & $5.4 \pm 16$ & $0.059 \pm 0.077$ & $0.007 \pm 0.009$ & $1.5 \pm 2.1$ & $0.027 \pm 0.012$\\
\hline
YuMi Scenario C & $4.9 \pm 17$ & $0.019 \pm 0.077$ & $0.001 \pm 0.009$ & $0.3 \pm 1.7$ & $0.029 \pm 0.010$\\
\hline
\end{tabular}
\captionof{table}{Mean results for the teleoperation trajectory}
\label{t3}
\end{table*}
\subsection{Computational efficiency}
We only need to calculate the generalized inverse of the Jacobian once per end effector and optimization step. Additionally, the calculations for the resulting joint configurations and the optimized manipulability are not heavy and we therefore expected little impact on the overall performance. Indeed, for the experiments presented we add $0.58$ms $\pm 0.23$ms
to a computation time of around 3ms per frame, which still allows the simulation to mostly run at 144 frames per second.
\section{CONCLUSION}
We introduced a method and a intuitive VR interface for payload manipulation in multi-arm systems. By opening up a rotational degree of freedom and using first order angle modifications to rotate around the free axis we are able to improve the manipulability index for each arm individually. The method was tested on different robots and and for different trajectories, showing that especially in teleoperation where users are not particularly aware of the system limitations we are able to better avoid singular configurations and, in most cases, improve the system behaviour. By adding further improvements such as limits on the maximum accelerations when reaching workspace limits we believe that our method can easily be further improved. The method is computationally efficient and does generally not require a trade-off against other parameters as is the case in \cite{rakita2018} or \cite{dufour2020}. We therefore believe that this is a promising first step in novel developments for future VR teleoperation for multi-arm systems.
\begin{figure}[b]
\centering
\includegraphics[width=\linewidth]{Figures/manipComparisonVRYuMi.png}
\caption{Qualitative manipulability comparison for scenarios B and C of the teleoperation trajectory on the 2 YuMi robots}
\label{fig:vrTrajYuMi}
\end{figure}
\\
\textbf{Mobile bases}: Although our setup consists of robots which are fixed to the ground this is not a limitation of our system and indeed we can already show that it works for robots on omnidirectional mobile bases by adding the DOFs of the base to the overall system and optimizing over the resulting $\mathbf{q}$ vector. Qualitative results of the experiment can be found in the accompanying video. In future iterations we would also like to extend this to directional robot bases.\\
\textbf{Collision avoidance and re-grasping}: To focus on the method in isolation, many of the details of a complete system have been neglected, e.g. collision avoidance.
Additionaly, sometimes a better configuration for the individual robot arm might exist but is not reachable without violating the end effector constraints.
Allowing re-grasping maneuvers might improve the system even further. \\
\textbf{Dynamics and task allocation}: One point of interest in multi-robot systems is that their combined payload capacity allows them to lift heavier objects, while the potential redundancy of additional robots could be used to allow for dynamic re-grasping or allocation of robots to multiple tasks or objects. We believe that this might be especially interesting for larger building sites or automation in larger factory halls.\\
\addtolength{\textheight}{-12cm}
|
\section{Introduction}
For Noetherian ring $S$ and for an ideal $J\subset S$, we have the local cohomology module $H^i_J(S)$ supported at $J$. It is a very mysterious object, even it is quite hard to know when it will vanish. Recently there have been several instances where more explicit information on local cohomology modules was obtained in special cases using combinatorial approach and one of such place is Stanley-Reisner ring.
Recall that the cohomological dimension of an ideal $J$ of a Noetherian ring $S$ is the maximum index $i \geq0$ for which the local cohomology module $H^i_J(S)$ is nonzero. In this context we mention Hartshorne-Lichtenbaum vanishing theorem or ``HLVT'' \cite{HartshorneCD}. It states that: For any complete local domain $S$ of dimension $d$, $H^d_J(S)$ vanishes if and only if $\dim(S/J)>0$. One may regard the HLVT as the ``first vanishing theorem'' for local cohomology. In \cite{RWY}, HLVT has been extended for Stanley-Reisner ring with an interpretation from combinatorial topology.
If the ring $S$ contains a field, the ``second vanishing theorem'' or ``SVT'' of local cohomology states the following: Let $S$ be a complete regular local ring of dimension $d$ with a separably closed residue field, which it contains. Let $J\subseteq S$ be an ideal such that $\dim(S/J)\geq 2$. Then $H^{d-1}_J(S)=0$ if and only if the punctured spectrum of $S/J$ is connected \cite{HartshorneCD,H-L,Ogus,P-S}. In \cite{CohDim}, the SVT has been extended to complete unramified regular local ring of mixed characteristic. In \cite{Bh}, it has been realized in the ramified case only the for extended ideals.
In this paper, we extend the result of the SVT to Stanley-Reisner ring with a combinatorial topological interpretetion, (see Theorem 3.2.):
Let $\Sigma \subset \Delta$ be simplicial complexes where $\dim k[\Delta]= d$ with $J=I_{\Sigma}$. Then $H^{d-1}_J (k[\Delta])=0$ if and
only if every $(d-1)$-face contains two vertices of $\Sigma$ and every $(d-2)$-face contains one vertex of $\Sigma$.
In this context it should be mentioned that this proof is greatly influenced by Theorem 3.5 of \cite{RWY} and goes similarly. Finally, we generalize the result of SVT (see, Theorem 4.2):
Let $\Sigma \subset \Delta$ be simplicial complexes where $\dim k[\Delta]= d$ with $J=I_{\Sigma}$. For some fixed $n\leq d-1$, $H^{d-n+1}_J (k[\Delta])=0$ if and only if for every $1\leq i \leq n$, every $(d-i)$-face of $\Delta$ contains $n-i+1$ vertices of $\Sigma$.
\section{Preliminaries}
In this section, we recall some basic results from combinatorial topology. For general references we refer \cite{Bj} and \cite{Mu}.
For $[n]= \{1,2,\ldots,n\}$, let $\Delta\subset 2^{[n]}$ be a simplicial complex i.e., if $F\in \Delta$ and $G\subset F$, then $G\in \Delta$. For $\bold{a}\in \ZZ^n$, we define the following support subsets of $[n]= \{1,2,\ldots,n\}$: $\supp_{+}\bold{a}= \{i: a_i > 0\}$, $\supp_{-}\bold{a}= \{i: a_i < 0\}$ and $\supp \bold{a}= \{i: a_i \neq 0\}$. Let $S = k[x_1, \ldots, x_n]$ be a polynomial ring over a field $k$. If $F\in \Delta\subset 2^{[n]}$, then we write square-free monomial $x^{a_1}_1\ldots x^{a_n}_n= x^F$, when $F= \supp \bold{a}$.
For a simplicial complex $\Delta$, the Stanley-Reisner ideal of $\Delta$ is the square-free monomial ideal, $I_{\Delta}= \{x^F: F\notin \Delta\}$. We define the Stanley-Reisner ring as $S/I_{\Delta}= k[\Delta]$. Let $\Sigma\subset \Delta \subset 2^{[n]}$ be two simplicial complexes. Then we have Stanley-Reisner ideals $I_{\Delta}\subset I_{\Sigma}$. In $k[\Delta]$, we denote the image of $I_{\Sigma}$ by $J$.
Given a face $F\in \Delta$, we can define three subcomplexes, called the star, deletion, and link of $F$ inside $\Delta$, as follows:
$\star_{\Delta} (F) = \{G\in \Delta: G\cup F\in \Delta\}$, $\del_{\Delta} (F) = \{G\in \Delta: G\not\subset F\}$ and $\lk_{\Delta} (F) = \{G\in \Delta: G\cap F= \phi, G\cup F\in \Delta\}$
Now to prove the result of this paper, we need the following result, Theorem 3.2 of \cite{RWY}:
\begin{mt*}
Let $\Sigma\subset \Delta$ be simplicial complexes and let $\bold{a}\in \ZZ^{n}$, $F_{+}=\supp_{+}\bold{a}$ and $F_{-}=\supp_{-}\bold{a}$. Then $H^{i}_J (k[\Delta])_{\bold{a}}= \widetilde{H}^{i-1}(||\star_{\Delta}(F_{+})||-||\Sigma||,||\del_{\star_{\Delta}(F_{+})}(F_{-})||-||\Sigma||;k)$
where $\widetilde{H}^{i}(-,-;k)$ denotes $i-th$ singular relative reduced cohomology and $||\Delta||$
denotes the geometric realization of a simplicial complex $\Delta$.
\end{mt*}
Let $\Sd (\Delta)$ denote the barycentric subdivision of the simplicial complexes $\Delta$, see Section 15 of \cite{Mu} and also section 9 of \cite{Bj} for abstract simplicial complexes. Given a subcomplex $\Sigma\subset \Delta$, $\Sd (\Delta-\Sigma)$ is the subcomplex of barycentric subdivision of $\Sd (\Delta)$ whose vertices are not the barycentre of any face of $\Sigma$.
We also need the following result, Lemma 4.7.27 of \cite{BLSWZ}.
\begin{mt**}
Let $\Delta$ be a simplicial complex and let $\Delta'$ and $\Sigma$ be two subcomplexes. Then the pair of spaces $(||\Delta||-||\Sigma||,||\Delta'||-||\Sigma||)$ is relatively homotopy equivalent to the pair $(\Sd (\Delta-\Sigma),\Sd (\Delta'-\Sigma))$
\end{mt**}
\section{main results}
Before proving our main result, we need the following Lemma. In this context it should be mentioned that this proof is greatly influenced by Theorem 3.5 of \cite{RWY} and goes similarly.
\begin{lem}
For $\Sigma \subset \Delta$ and $\dim\Delta= d-1$, if every $(d-1)$-face of $\Delta$ contains two vertices of $\Sigma$ and if every $(d-2)$-face of $\Delta$ contains one vertex of $\Sigma$, then we can remove every $(d-1)$-face and every $(d-2)$-face of $\Sd(\Delta-\Sigma)$.
\end{lem}
\begin{proof}
Denote $\Sd(\Delta-\Sigma)= X$ and set $D= d-1, d-2$. To remove a $D$-face of $X$, we adopt the way of elementary collapse \cite{Bj}, which states that, if we find a $D$-face which contains some $D-1$-face, but latter is not in any other $D$-face, we remove both of them. We claim that this process removes all $D$-faces of $\Sd(\Delta-\Sigma)$.
If not, assume that after removing few $D$-faces, when we reach the simplicial complex $X'$ and there exists one $D$-face in $X'$ which can not be removed, i.e. there is a $D$-face in $X'$ such that all of its $(D-1)$-faces lying in some other $D$-face, so that no further collapse is possible. Let $\sigma$ be one of such $D$-faces in $X'$, and let $G$ be a $D$-face of $X$ in which it lies, so that the barycenter $b_G$ of $G$ is one of its vertex.
.
Then $\lk_{X'}(b_G)$ is a subcomplex of $\lk_{\Sd (\Delta)}(b_G)$ which is the boundary of $\Sd G$ (since being a barycentre only for $F\subset \Sd G$, $\{b_G\}\cup F$ is a face of $\Sd (\Delta)$). Since $\sigma$ is in $X'$, $\dim (\lk_{X'}(b_G))= \dim (\lk_{\Sd (\Delta)}(b_G))$, and moreover any $(D-2)$-face $\tau$, $(D-1)$-face $\tau\cup \{b_G\}$ is in two $D$-faces (otherwise, we could remove the face $\sigma$). Thus we find $\lk_{X'}(b_G)= \lk_{\Sd (\Delta)}(b_G)$. But this is a contradiction, since $\lk_{\Sd (\Delta)}(b_G)$ should contain atleast one vertex of $\Sigma$, while $\lk_{X'}(b_G)$ should not contain any vertex of $\Sigma$.
\end{proof}
Now we state the main result of this paper.
\begin{thm}
Let $\Sigma \subset \Delta$ be simplicial complexes where $\dim k[\Delta]= d$ with $J=I_{\Sigma}$. Then $H^{d-1}_J (k[\Delta])=0$ if and
only if every $(d-1)$-face contains two vertices of $\Sigma$ and every $(d-2)$-face contains one vertex of $\Sigma$.
\end{thm}
\begin{proof}
For the forward direction, we assume that some $(d-2)$-face $F$ contains no vertex of $\Sigma$ and we want to prove that for some multidegree $\bold{a}$, $H^{d-1}_J (k[\Delta])_{\bold{a}}\neq 0$. We consider a multidegree $\bold{a}$ having $F_{+}=\supp_{+}\bold{a}= \phi$, $F_{-}=\supp_{-}\bold{a}= F$. From Theorem A given in Section 2, we get $H^{d-1}_J (k[\Delta])_{\bold{a}}= \widetilde{H}^{d-2}(||\Delta||-||\Sigma||,||\del_{\Delta}(F)||-||\Sigma||;k)$ where we use the fact that $\star_{\Delta}(\phi)=\Delta$. Now if we remove $||\Delta||-||\Sigma||-||F||$, using "exision" we get $\widetilde{H}^{d-2}(||\Delta||-||\Sigma||,||\del_{\Delta}(F)||-||\Sigma||;k)= \widetilde{H}^{d-2}(||F||,||\partial F||;k)$. Finally, going to the quotient space $||F||/||\partial F||$, we get $\widetilde{H}^{d-2}(||F||,||\partial F||;k)= \widetilde{H}^{d-2}(||F||/||\partial F||;k)= \widetilde{H}^{d-2}(\SS^{d-2};k)=k$. Thus $H^{d-1}_J (k[\Delta])_{\bold{a}}\neq 0$.
For the other direction, we assume that every $(d-1)$-face contains two vertices of $\Sigma$ and every $(d-2)$-face contains one vertex of $\Sigma$ and using Theorem A of Section 2, we need to show $H^{d-1}_J (k[\Delta])_{\bold{a}}= \widetilde{H}^{d-2}(||\star_{\Delta}(F_{+})||-||\Sigma||,||\del_{\star_{\Delta}(F_{+})}(F_{-})||-||\Sigma||;k)= 0$ with $F_{+} \cup F_{-}\in \Delta$. Since $\star_{\Delta}(F_{+})$ can be of dimension atmost that of $\Delta$, we can assume $F_{+}=\phi$ and above reduces that we only need to show $\widetilde{H}^{d-2}(||\Delta||-||\Sigma||,||\del_{\Delta}(F)||-||\Sigma||;k)= 0$.
Now instead of taking simplicial homology for the pair of spaces $(||\Delta||-||\Sigma||,||\del_{\Delta}(F)||-||\Sigma||)$, we can take the pair of spaces $(\Sd (\Delta-\Sigma),\Sd (\del_{\Delta}(F)-\Sigma))$, where $\Sd (\Delta-\Sigma)$ is the subcomplex of barycentric subdivision of $\Sd (\Delta)$ whose vertices are not the barycentre of any face of $\Sigma$. Using Theorem B of Section 2, we get that $\widetilde{H}^{d-2}(||\Delta||-||\Sigma||,||\del_{\Delta}(F)||-||\Sigma||;k)= \widetilde{H}^{d-2}(\Sd (\Delta-\Sigma),\Sd (\del_{\Delta}(F)-\Sigma))$.
Now from above Lemma we have shown that we can remove every $(d-1)$-face and $(d-2)$-face of $\Sd (\Delta-\Sigma)$ and this leads to the desired result.
\end{proof}
\section{generalization of the SVT}
The result of the last section can be generalized in the following
\begin{lem}
Let $\Sigma \subset \Delta$ and $\dim\Delta= d-1$. For some fixed $n\leq d-1$ and for every $1\leq i \leq n$, if every $(d-i)$-face of $\Delta$ contains $n-i+1$ vertices of $\Sigma$, then for every $i\leq n$, we can remove every $(d-i)$-face of $\Sd(\Delta-\Sigma)$.
\end{lem}
\begin{proof}
Similar to that of Lemma 3.1.
\end{proof}
Using above lemma in a similar way we can prove the generalization of Theorem 3.2.
\begin{thm}
Let $\Sigma \subset \Delta$ be simplicial complexes where $\dim k[\Delta]= d$ with $J=I_{\Sigma}$. For some fixed $n\leq d-1$, $H^{d-n+1}_J (k[\Delta])=0$ if and only if for every $1\leq i \leq n$, every $(d-i)$-face of $\Delta$ contains $n-i+1$ vertices of $\Sigma$.
\end{thm}
\begin{proof}
Similar to that of Theorem 3.2.
\end{proof}
|
\section{Introduction}
Human visual perception is acquainted with high-contrast images that are characterized by high contrast, good visibility, and minimal noise. Thus researchers have focused extensively on developing computer-vision techniques to improve the visual perception of images. Such algorithms have broad applicability, such as all-weather autonomous vehicles and illumination-invariant face detection.
Low-light image enhancement is a well-studied problem, and researchers have proposed several methods to address this problem. These methods include histogram equalization, dehazing-based approaches, and retinex theory. Although these representative state-of-the-art methods produce good results, they are limited in terms of model capacity for illumination and reflectance decomposition. Such constraints are hand-crafted and require careful hyperparameter-optimization. To mitigate this problem, researchers have used CNNs for low-level image processing. Owing to the extensive success of GANs for the problem of image-to-image translation, we build a framework that can generate visually-pleasing images through spectral guidance.
In this paper, we propose \textsc{SpecNet} which optimizes a spectral profile to achieve superior results. We first use a cycle-consistent framework to reconstruct hyperspectral images from RGB images which is further used to restore proper illumination for the given low-light or dark image. The primary GAN framework used for hyperspectral reconstruction has been carefully modified to incorporate a spectral-profile optimization framework, ultimately aimed at generating visually-pleasing images. Finally, we perform extensive set of experiments to evaluate the effectiveness of the model.
\begin{figure}[t!]
\centering
\begin{subfigure}[t]{0.23\columnwidth}
\includegraphics[width=\columnwidth,height=22mm]{Figures/748.png_low.pdf}
\caption{Dark Image}
\label{subfig:intro_hazy}
\end{subfigure}
\begin{subfigure}[t]{0.23\columnwidth}
\includegraphics[width=\columnwidth,height=22mm]{Figures/748_hsi.pdf}
\caption{Reconstructed HSI}
\label{subfig:intro_hsi}
\end{subfigure}
\begin{subfigure}[t]{0.23\columnwidth}
\includegraphics[width=\columnwidth,height=22mm]{Figures/748_generated_model4.pdf}
\caption{Enh. Output}
\label{subfig:intro_output}
\end{subfigure}
\begin{subfigure}[t]{0.23\columnwidth}
\includegraphics[width=\columnwidth,height=22mm]{Figures/748.png_high.pdf}
\caption{Ground Truth}
\label{subfig:intro_gt}
\end{subfigure}%
\caption{A sample dark image along with the reconstructed HSI and the output obtained using \textsc{SpecNet}.
}
\label{fig:intro}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth, height=35mm]{Figures/download-final.png}
\caption{Multi-channel cumulative spectral profile of sample dark image along with the reconstructed HSI and the output obtained using \textsc{SpecNet}}
\label{fig:my_label}
\end{figure}
\begin{figure*}
\centering
\begin{subfigure}[t]{0.13\textwidth}
\caption{Dark Image}
\includegraphics[width=\textwidth,height=20mm]{Figures/778.png_low.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{U-Net}
\includegraphics[width=\textwidth,height=20mm]{Figures/778_0_generated_unet.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{Pix2Pix}
\includegraphics[width=\textwidth,height=20mm]{Figures/778_0_generated_pix.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{CycleGAN}
\includegraphics[width=\textwidth,height=20mm]{Figures/778_generated_cyc.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{EnlightenGAN}
\includegraphics[width=\textwidth,height=20mm]{Figures/778_generated_enlGAN.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{\textbf{SpecNet}}
\includegraphics[width=\textwidth,height=20mm]{Figures/778_generated_model4.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{Ground Truth}
\includegraphics[width=\textwidth,height=20mm]{Figures/778.png_high.pdf}
\end{subfigure}
\caption{Qualitative comparison for different models as described in Table \ref{table:results}.}
\label{fig:ablation_comparison}
\end{figure*}
\section{Proposed Method}
To propose \textsc{SpecNet}, we hypothesize that multi-band information in the reconstructed hyperspectral images can improve the perceptual quality of images. First of all, we create a spanned 31-channel RGB image matrix to imitate the 31-channel HSI, to ease the under-constrained problem of HSI reconstruction from RGB images. The framework can be viewed as a cascaded GAN approach. The first GAN takes an unsupervised cycle-consistent approach to reconstruct HSI, which is fed into another cGAN to generate the final enhanced output image. To solve the under-constrained problem of HSI reconstruction we make use of several guiding principles such as task-aided supervision and spectral-profile optimization.
\vspace{-0.5em}
\paragraph{Aided Supervision} The lack of large-scale hyperspectral image datasets poses a problem in learning an output distribution that can imitate the underlying original hyperspectral values. A task-aided supervision addresses this distributional discrepancy. We use our original low-light enhancement task as an auxiliary optimization task to aid hyperspectral reconstruction. We modify the cycle-consistency loss as
\begin{equation}
\label{eq1}
\begin{split}
\mathcal{L}_{cyc} =& \Arrowvert y-G_h(G_x(x)) \Arrowvert_2^2 + \Arrowvert h-G_x(G_h(h)) \Arrowvert_2^2.
\end{split}
\end{equation}
where $(x, y)$ refers to dark and ehanced RGB images respectively, $G_x, G_h$ refer to dual generators used for cycle-consistency and $h$ refers to HSI.
\vspace{-0.5em}
\paragraph{Spectral-profile Optimization} As the primary task of the framework is to produce enhanced images, we incorporate a network to generate spectral-profile using multi-channel power spectrum from 2D Fourier transform \cite{durall2020watch}. The network was used to regularize the spectral distribution of reconstructed HSI. The motivation is to induce alignment in spectral distributional discrepancy in the reconstructed HSI. This is achieved by jointly optimizing the algorithm with a spectral-profile generator that discriminates between spectral profiles of reconstructed HSI and real RGB images. By minimizing the mean squared error, the algorithm encourages spectrally-enduring HSI.
\vspace{-0.5em}
\paragraph{Multi-layer Colorization Space} The multi-layer colorization space is constructed using different color models such as HSV, YCrCb, and LAB concatenated together with RGB which results in a 12-channel input image \cite{mehta2020domainaware}. This is fed into cGAN along with the reconstructed HSI.
\section{Experimental Evaluation}
The experimental results in terms of PSNR and SSIM on LOL dataset \cite{wei2018deep} are compiled in Table \ref{table:results}. \textsc{SpecNet} outperforms the existing state-of-the-art techniques in terms of PSNR and SSIM.
\begin{table}
\centering
\begin{threeparttable}
\begin{tabular}{c|l l}
\toprule
Method & \multicolumn{1}{c}{SSIM} & \multicolumn{1}{c}{PSNR} \\
\midrule
\midrule
U-Net & 0.7397 & 21.500 \\
Pix2Pix & 0.7307 & 20.483 \\
CycleGAN & 0.6850 & 20.348 \\
EnlightenGAN & 0.7694 & 23.202 \\
\textbf{\textsc{SpecNet}} & \textbf{0.8052} & \textbf{22.330} \\
\bottomrule
\end{tabular}
\end{threeparttable}
\caption{Comparative results on LOL dataset}
\label{table:results}
\end{table}
The proposed \textsc{SpecNet} consists of several components which add to performance through cumulative effort.To delineate the contributions of different components, several models were trained apart from the final model. The comparative performance is summarized in Table \ref{table:ablation_study}.
\begin{table}
\resizebox{\columnwidth}{!}{
\begin{tabular}{c|c c |c}
\toprule
& \multicolumn{2}{c|}{Components} & \\ [4pt]
\cline{2-3}
Method & \begin{tabular}[c]{@{}c@{}}Spectral Profile \\ Optimization\end{tabular} & \begin{tabular}[c]{@{}c@{}}Multi-Layer\\ Colorization Space\end{tabular} & \begin{tabular}[c]{@{}c@{}} SSIM \end{tabular} \\ \midrule \midrule
Model-1 & & & 0.6784 \\
Model-2 & \checkmark & & 0.7244 \\
\textbf{\textsc{SpecNet}} & \checkmark & \checkmark & \textbf{0.8052} \\
\bottomrule
\end{tabular}
}
\caption{Ablation Models}
\label{table:ablation_study}
\end{table}
\section{Conclusions}
This work demonstrates the use of spectral-profile optimization for low-light image enhancement using a cascaded GAN framework, referred to as \textsc{SpecNet}. It reconstructs HSI from low-light RGB images and an enhanced cGAN generates enhanced output images using reconstructed hyperspectral images. The model utilizes color spaces by concatenating a 12-channel multi-layer color space with the reconstructed HSI. Further, an ablation study is conducted which substantiates the contribution of individual components in the framework.
\begin{figure*}
\centering
\begin{subfigure}[t]{0.13\textwidth}
\caption{Dark Image\\\phantom{~\shortcite{wei2018deep}}}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/111.png_low.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{U-Net \\\shortcite{ronneberger2015u}}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/111_0_generated_unet.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{Pix2Pix\\\shortcite{isola2017image}}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/111_0_generated_pix.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{EnlightenGAN\\ \shortcite{jiang2019enlightengan}}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/111_generated_enlGAN.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{CycleGAN\\\shortcite{zhu2017unpaired}}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/111_generated_cyc.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{\textbf{SpecNet}\\ \phantom{Ours}}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/111_generated_model4.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\caption{Ground Truth\\ \phantom{\shortcite{wei2018deep}}}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/111.png_high.pdf}
\end{subfigure}
\\
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/146.png_low.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/146_0_generated_unet.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/146_0_generated_pix.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/146_generated_enlGAN.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/146_generated_cyc.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/146_generated_model4.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/146.png_high.pdf}
\end{subfigure}
\\
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/22.png_low.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/22_0_generated_unet.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/22_0_generated_pix.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/22_generated_enlGAN.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/22_generated_cyc.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/22_generated_model4.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/22.png_high.pdf}
\end{subfigure}
\\
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/23.png_low.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/23_0_generated_unet.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/23_0_generated_pix.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/23_generated_enlGAN.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/23_generated_cyc.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/23_generated_model4.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=22mm]{qual_sup/23.png_high.pdf}
\end{subfigure}
\\
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/55.png_low.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/55_0_generated_unet.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/55_0_generated_pix.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/55_generated_enlGAN.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/55_generated_cyc.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/55_generated_model4.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/55.png_high.pdf}
\end{subfigure}
\\
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/79.png_low.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/79_0_generated_unet.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/79_0_generated_pix.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/79_generated_enlGAN.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/79_generated_cyc.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/79_generated_model4.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/79.png_high.pdf}
\end{subfigure}
\\
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/665.png_low.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/665_0_generated_unet.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/665_0_generated_pix.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/665_generated_enlGAN.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/665_generated_cyc.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/665_generated_model4.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/665.png_high.pdf}
\end{subfigure}
\\
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/669.png_low.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/669_0_generated_unet.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/669_0_generated_pix.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/669_generated_enlGAN.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/669_generated_cyc.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/669_generated_model4.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/669.png_high.pdf}
\end{subfigure}
\\
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/780.png_low.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/780_0_generated_unet.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/780_0_generated_pix.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/780_generated_enlGAN.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/780_generated_cyc.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/780_generated_model4.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.13\textwidth}
\includegraphics[width=\textwidth,height=25mm]{qual_sup/780.png_high.pdf}
\end{subfigure}
\caption{Qualitative comparison with different models.}
\label{qual}
\end{figure*}
\begin{figure*}
\centering
\frame{\includegraphics[width=\linewidth]{Figures/aaai.png}}
\caption{The schematic diagram for the proposed SpecNet}
\label{fig:overview}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{Figures/ablation.pdf}
\caption{Qualitative comparison for ablation models}
\label{ablation_img}
\end{figure*}
\section{Supplementary Material}
\subsection{Architectural Details}
The proposed work adapts an unpaired cycle-consistency framework ~\cite{zhu2017unpaired} to exploit supervision at the level of sets.
The objective is to learn a mapping function $G_x: X_{31} \,\to\, Y$, where $X_{31}$ represents the stacked RGB image and $Y$ refers to the reconstructed HSI. In context to the adversarial loss, the reconstruction module can be expressed as
\begin{align}
\begin{split}
&\mathcal{L}_{GAN_x}(G_x,D_x, X_{31}, Y) \\
&= E_{x_{31}\sim p_{data(x_{31})}}[log(1-D_x(G_x(x_{31})))] \\
&+ E_{y\sim p_{data(y)}}[log(D_x(y))] \\
\end{split}
\\
\begin{split}
&\mathcal{L}_{GAN_y}(G_y,D_y, Y, X_{31}) \\
&= E_{y\sim p_{data(y)}}[log(1-D_y(G_y(y)))] \\
&+E_{x_{31}\sim p_{data(x_{31})}}[log(D_y(x_{31}))] \\
\end{split}
\\
\begin{split}
\mathcal{L}_{GAN} = \mathcal{L}_{GAN_x} + \mathcal{L}_{GAN_y}
\end{split}
\end{align}
Figure \ref{fig:overview} shows the complete schematic diagram of the proposed SpecNet.
The generators in $G_x$, $G_y$ and $G_z$ adopt a U-Net with skip connections while PatchGAN is adopted for the corresponding discriminators. We use L1 cycle consistency losses and identity losses~\cite{zhu2017unpaired} to further improve the reconstructed HSI.
The generator $G_s$ uses a ResNet-based architecture to compute the spectral profile of input image. Deriving inspiration from recent work by \citeauthor{durall2020watch}, we extend the analysis to hyperspectral images. The network aims to regularize the generated HSI with respect to spectral distribution of real images.
\section{Datasets}
To facilitate HSI reconstruction, HSCycle is trained using ICVL BGU Hyperspectral Dataset (NTIRE 2018) ~\cite{arad2016sparse, arad2018ntire} and the NTIRE 2020 dataset. The dataset is composed of 200 natural images with various indoor and outdoor scenes. The dataset provides sampled images which each having 31 spectral bands. Adjacent bands have an incremental difference of 10 nm. In addition, preprocessing like random cropping and flip is utilized to increase the total number of images upto 6000.
To train the proposed network for low light image enhancement, we use low/normal-light pairs in the LOL Dataset. The LOL Dataset consists of 500 image pairs, which is pre-divided into training and evaluation datasets.
\subsection{Additional Results}
In Figure \ref{qual}, we show additional qualitative comparison of \textsc{SpecNet} with several deep learning based models. U-Net, Pix2Pix and CycleGAN, being general computer vision models, were re-trained on the train dataset used by \textsc{SpecNet}.
In Figure \ref{ablation_img} we visually show the performance of \textsc{SpecNet} with respect to other ablated models. The red box highlights the improvement our model gets due Spectral Profile optimization and multi-layer colorization.
\section{Acknowledgements}
This work is supported by BITS Additional Competitive
Research Grant (PLN/AD/2018-19/5).
|
\section{Introduction}
In the last decade, a variety of wireless technological advances, including millimeter-wave (mmWave) communication and massive multiple-input multiple-output (mMIMO) systems, have been proposed to achieve a $ 1000 $-fold capacity increase and ubiquitous wireless connectivity among a large number of devices \cite{Boccardi2014}. Unfortunately, these technologies face practical limitations in terms of excessive energy consumption and hardware cost as well as no guaranteed quality of service (QoS) in harsh propagation environments. For instance, mmWave communications exhibit high penetration/path loss while requiring expensive and energy-consuming transceivers. Similarly, mMIMO systems manifest low performance in poor scattering conditions while the large number of active elements might render the energy usage prohibitive. Moreover, new challenging use cases will emerge with possibly similar shortcomings. As a result, future networks require radical paradigm shifts towards their energy sustainability, e.g., control to some extent over the propagation environment.
A disruptive technology, covering this gap, has emerged under the label intelligent reflecting surfaces (IRSs) or reconfigurable intelligent surfaces (RISs). An IRS consists of a meta-surface including a large number of reconfigurable passive elements that are able to function independently by inducing certain phase shifts on the impinging waves \cite{Basar2019}. The smart adjustment of the phase shifts is managed by an attached controller and allows the coherent addition of the reflected signals to boost the desired signal at the receiver.
The IRS design and applications have attracted a lot of significant research interest \cite{Basar2019,Wu2019a,Pan2020,Bjoernson2019b,Kammoun2020, Elbir2020,Guo2020,Chen2019}. For example, in \cite{Wu2019a}, the downlink of an IRS-assisted multi-user (MU) multiple-input single-output (MISO) communication system was studied by jointly optimizing the precoding and reflecting beamforming matrices (RBMs), in order to minimize the transmit power at the base station (BS) with signal-to-interference-plus-noise ratio (SINR) constraints. In addition, the sum-rate was maximized in \cite{Pan2020} subject to a transmit power constraint.
Furthermore, the outperformance of the IRS with respect to the decode-and-forward (DF) relay was presented in \cite{Bjoernson2019b}. Also, the authors in \cite{Kammoun2020} maximized the minimum user rate in the large number of antennas regime. Despite the fundamental design issues, applications regarding IRSs have started to emerge such as the maximization of the minimum secrecy rate for physical layer security \cite{Chen2019}.
In general, there are two approaches for phases optimization as the literature reveals. The first method is based on statistical channel state information (CSI) \cite{Jia2020,Zhao2020,Zhi2020,Kammoun2020,Papazafeiropoulos2021,VanChien2021,Papazafeiropoulos2021a} and the second method is based on instantaneous CSI \cite{Wu2019a,Pan2020}. According to the second method, the phases are optimized at every coherence
interval since the corresponding expressions depend on small-scale channel fading. On the contrary, the first method includes expressions that depend on the large-scale statistics, which change at every several coherence intervals. Thus, the significance of the first method is noteworthy since it reduces considerably the signal overhead which can be prohibitive in the case of a large number of reflecting elements at the IRS. Moreover,
it results in lower computational complexity. Especially, based on these observations, in high-speed scenarios with fast time-varying channels, it is more practical to design and adjust the IRS phase shifts according to the statistical
CSI while the tuning of the IRS parameters, based on instantaneous CSI, would be more challenging since they would have to
be updated more frequently. Furthermore, although the IRS does not consume ideally transmit power, its smart controller is power-consuming and its continuous overloading with operations in the case of instantaneous CSI would not be energy efficient.
Although most existing works with IRS-aided systems have relied on the knowledge of perfect CSI, this is a highly unrealistic assumption. In practice, systems have imperfect CSI. Especially, their passive elements make them energy efficient, but, contrary to conventional systems, this interesting feature does not enable them to accomplish the channel estimation (CE) task by transmission/reception of pilot symbols. Hence, it is of paramount importance to take into account the CE before arriving at realistic conclusions. Among the fundamental works \cite{Zheng2019,Mishra2019,He2019,Elbir2020,Nadeem2020}, the authors in \cite{Mishra2019} proposed an ON/OFF channel estimation scheme that obtains one-by-one least squares (LS) estimates of all IRS-assisted channels for a single-user MISO system. Moving to MUs systems, finding more applications in contemporary systems, the authors in \cite{He2019} exploited the sparsity of the channel and formulated a sparse channel matrix recovery problem for CE. In \cite{Nadeem2020}, the authors extended the model in \cite{Mishra2019} by assuming all IRS elements to be active during training while a number of sub-phases equal at least to the number of IRS elements is considered. This method provides better CE as the number of sub-phases increases, but the achievable rate worsens since the data transmission phase takes a smaller fraction of the coherence time due to excessive training overhead. Another drawback is that this method provides the estimates of the channels of the individual IRS elements while the covariance of the channel vector from all IRS elements to a specific user equipment (UE) is unknown.
On the other hand, prior literature of IRS-assisted systems has mostly assumed perfect hardware while practical applications are affected by unavoidable transceiver hardware impairments (T-HWIs) such as the in-phase/quadrature-phase (I/Q)-imbalance~\cite{Qi2010}, the quantization noise in the analog-to-digital converters (ADCs), and the oscillator phase noise (PN)~\cite{Papazafeiropoulos2016,Papazafeiropoulos2017a}. Even if mitigation/compensation algorithms exist, T-HWIs cannot be completely removed \cite{Schenk2008,Bjoernson2017}. Basically, T-HWIs are divided into two main categories being the additive and multiplicative T-HWIs.
In this work, we focus on the impact of the additive T-HWIs, while the study of multiplicative T-HWIs will be the topic for future work. In this direction, an examination of existing works with HWIs in the area of IRS-assisted systems shows that relevant studies are in their infancy \cite{Li2020,Xing2020,Qian2020,Liu2020,Shen2020,Zhou2020a}. In \cite{Li2020}, only single-antennas nodes were considered, and the phase errors were assumed known (deterministic). In addition, the phase noise, induced by an IRS and, henceforth called IRS-HWIs, has been studied in \cite{Xing2020,Qian2020} in the case of perfect CSI, but no expectation was taken over the phase noise. Note that this phase noise, coming from the finite precision configuration of the phase drifts, is irrelevant with the phase noise coming from imperfect signal generation in local oscillators in standard antenna systems. Moreover, in \cite{Liu2020}, despite its randomness, again, no averaging of the phase noise was applied. Furthermore, only a single UE communication has been considered and only upper bounds on the channel capacities have been studied, which are not also obtained in closed forms. The authors in \cite{Shen2020} provided the beamforming optimization by accounting for T-HWIs in a single-user setting, and in \cite{Zhou2020a}, the secrecy rate was derived. Notably, only a few works in IRS-aided systems have assumed correlated Rayleigh fading despite that this is normally the case in multi-antenna next-generation systems. Apart from that, most works perform RBM optimization with a high computational cost in every coherence interval.
\subsection{Contribution}
The previous observations motivate the topic of this work, which is the design/study of a general IRS-assisted MU-MISO system with imperfect CSI and HWIs at both the IRS and the transceiver while performing robust optimization. Notably, the introduction of HWIs increases the complexity/difficulty and demands substantial manipulations during the analysis of IRS-assisted systems. The main contributions are summarized as follows:
\begin{itemize}
\item Contrary to \cite{Xing2020}, we have assumed multiple antennas at the BS and multiple UEs as well as imperfect CSI. Also, compared to \cite{Liu2020}, we have considered correlated Rayleigh fading, multiple UEs, and closed-form lower bounds, which are more practical than any upper bounds. Both references have not addressed properly the impact of phase noise while, in \cite{Li2020}, only deterministic phase noise was assumed. In \cite{Shen2020} and \cite{Zhou2020a}, only the impact of T-HWIs was studied while only a single destination and perfect CSI were assumed. Notably, as far as the authors are aware, our work is the only one accounting for the randomness of the phase noise.
\item Many previous works have assumed that the optimization of the RBM should take place at every coherence interval since the corresponding expressions depend on small-scale channel fading, while our proposed results, being dependent only on large-scale statistics are suggested to be optimized at every several coherence intervals. Thus, their significance is noteworthy since they reduce considerably the signal overhead which can be prohibitive in the case of a large number of reflecting elements at the IRS.
\item We perform CE by means of linear minimum mean square error (LMMSE) while HWIs are taken into account. In parallel, we have assumed correlated Rayleigh fading. Our method provides analytical tractable expressions with low overhead compared to previous works.\footnote{Works such as \cite{He2019, Elbir2020} do not provide analytical expressions. Also,
previous CE methods with correlated fading do not allow the derivation of an optimizable achievable spectral efficiency (SE) being dependent on the RBM. In \cite{Nadeem2020}, only the estimated individual channels between each IRS element and each UE are obtained while the inter-element correlation is unknown.}
\item We derive the uplink achievable spectral efficiency (SE) (lower bound) of an IRS-assisted MU-MISO system with MRC, imperfect CSI, and HWIs in a closed-form dependent only on large-scale statistics (covariances).
\item We optimize the achievable sum SE with respect to the RBM. As mentioned, contrary to other works that depend on small-scale statistics (e.g., see \cite{Guo2020} where the stochastic successive convex approximation technique has been performed), our optimization can be performed quite efficiently by the project gradient ascent at every several coherence intervals since both the sum SE and the proposed algorithm require only the large-scale statistics. Notably, contrary to existing works such as \cite{Jia2020,Zhao2020}, based on statistical CSI, we achieve to provide the SE and the phases optimization in closed-form.
\item We shed light on the degradation of the uplink sum SE of an IRS-aided MU-MISO system due to the presence of imperfect CSI, HWIs, and correlated fading. For example, we thoroughly examine how the probability density function (PDF) of the phase noise at the IRS and the severity of the T-HWIs affect the performance of IRS-aided systems.
\end{itemize}
\subsection{Paper Outline}
The remainder of this paper is organized as follows. Section~\ref{System} presents the system model of an IRS-assisted MU-MISO system with correlated Rayleigh fading and HWIs. Section~\ref{ChannelEstimation} provides the CE. Section~\ref{PerformanceAnalysis} presents the uplink sum SE and the optimization concerning the IRS RBM.
The numerical results are placed in Section~\ref{Numerical}, and Section~\ref{Conclusion} concludes the paper.
\subsection{Notation}Vectors and matrices are denoted by boldface lower and upper case symbols, respectively. The notations $(\cdot)^{\scriptscriptstyle\mathsf{T}}$, $(\cdot)^\H$, and $\mathop{\mathrm{tr}}\nolimits\!\left( {\cdot} \right)$ represent the transpose, Hermitian transpose, and trace operators, respectively. The expectation operator is denoted by $\mathbb{E}\left[\cdot\right]$ (or $\mathbb{E}_{x}\left[\cdot\right]$ to denote expectation with respect to $ x $) while $ \diag\left({\mathbf{a}} \right) $ represents an $ n\times n $ diagonal matrix with diagonal elements being the elements of vector $ {\mathbf{a}} $. In the case of a matrix $ {\mathbf{A}} $, $\diag\left({\mathbf{A}}\right) $ denotes a diagonal matrix with elements corresponding to the diagonal elements of $ {\mathbf{A}} $. Also, $ \arg\left(\cdot\right) $ and $ \circ $ denote the argument function and the Hadamard product, respectively. Finally, ${\mathbf{b}} \sim {\cal C}{\cal N}{({\mathbf{0}},\mathbf{\Sigma})}$ represents a circularly symmetric complex Gaussian vector with {zero mean} and covariance matrix $\mathbf{\Sigma}$.
\section{System Model}\label{System}
We consider an IRS-aided MU-MISO system as depicted in Fig. \ref{Fig10}. In particular, a BS, equipped with $ M $ antennas, serves $K $ single-antenna UEs by means of one IRS consisting of $ N $ passive reflecting elements introducing phase shifts onto the incoming signal waves. The phase-shifts are adjusted by a controller exchanging information with the BS through a backhaul link. Reasonably, the IRS is deployed in the line-of-sight (LoS) of the BS by assuming that both the BS and IRS are deployed at high altitude and their locations are fixed. Moreover, the proposed model assumes direct links between the BS and the UEs, but these could be neglected in certain scenarios. For example, in mmWave transmission, suggested by 5G and beyond systems, high penetration losses and resultant signal blockages do not allow the presence of an LoS component \cite{Rappaport2019}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.9\linewidth]{SystemModel.pdf}
\caption{\footnotesize{ An IRS-assisted uplink MU-MISO communication system with $ M $ BS antennas, $ N $ IRS elements, and $ K $ UEs. }}
\label{Fig10}
\end{center}
\end{figure}
\subsection{Channel Model}\label{ChannelModel}
We assume a time-varying narrowband channel, divided into coherence blocks, where each block has a duration $\tau_{\mathrm{c}}$ channel uses.\footnote{ The extension to the wideband case could follow the lines of \cite{Zheng2019,Yang2020} and could be the topic of future work.} Note that $\tau_{\mathrm{c}}= B_{\mathrm{c}}T_{\mathrm{c}}$ with $ B_{\mathrm{c}} $ and $ T_{\mathrm{c}} $ being the coherence bandwidth and the coherence time in $\mathrm{Hz}$ and $\mathrm{s}$, respectively. Especially, we employ the standard time-division-duplex (TDD) protocol, where each block accounts for $\tau$ channel uses for the uplink training phase and $\tau_{\mathrm{u}}=\tau_{\mathrm{c}}-\tau$ channel uses for the uplink data transmission phase.
Let $ {\mathbf{h}}_{\mathrm{d},k} \in \mathbb{C}^{M \times 1} $, $ {\mathbf{H}}_{1}=[{\mathbf{h}}_{1,1}\ldots,{\mathbf{h}}_{1,N} ] \in \mathbb{C}^{M \times N}$, and $ {\mathbf{h}}_{2,k} \in \mathbb{C}^{N \times 1}$ be the direct channel between the BS and UE $ k $, the LoS channel between the BS and the IRS with $ {\mathbf{h}}_{1,i} $ for $ i=1,\ldots,N $ being its column vectors, and the channel between the IRS and UE $ k $. The subscripts $ 1$ and $ 2$ correspond to the BS-IRS and IRS-UE $ k $ links, respectively. Although the majority of existing works, e.g., \cite{Wu2019a,Pan2020}, assumed independent Rayleigh model, we account for spatial correlation,
which appears in practice and affects the performance \cite{Kammoun2020}. Hence, $ {\mathbf{h}}_{\mathrm{d},k} $ and $ {\mathbf{h}}_{2,k} $ are described in terms of correlated Rayleigh fading distributions as
\begin{align}
{\mathbf{h}}_{2,k}&=\sqrt{\beta_{2,k}}{\mathbf{R}}_{\mathrm{IRS},k}^{1/2}{\mathbf{z}}_{k},\\
{\mathbf{h}}_{\mathrm{d},k}&=\sqrt{\beta_{\mathrm{d},k}}{\mathbf{R}}_{\mathrm{BS},k}^{1/2}{\mathbf{z}}_{\mathrm{d},k},
\end{align}
where $ {\mathbf{R}}_{\mathrm{IRS},k} \in \mathbb{C}^{N \times N} $ and $ {\mathbf{R}}_{\mathrm{BS},k} \in \mathbb{C}^{M \times M} $ describe the deterministic Hermitian-symmetric positive semi-definite correlation
matrices at the IRS and the BS respectively with $ \mathop{\mathrm{tr}}\nolimits\left({\mathbf{R}}_{\mathrm{IRS},k} \right)=N $ and $ \mathop{\mathrm{tr}}\nolimits\left({\mathbf{R}}_{\mathrm{BS},k} \right)=M $. Notably, the correlation matrices $ {\mathbf{R}}_{\mathrm{IRS},k} $ and $ {\mathbf{R}}_{\mathrm{BS},k}~\forall k$ are assumed to be known by the network. They can be obtained through the existing estimation methods (see e.g., \cite{Neumann2018}). Given that correlation models for IRS using meta-surfaces are not known, we adopt the correlation model in \cite{Kammoun2015} for conventional antenna arrays.\footnote{While writing this work, the authors in \cite{Bjoernson2020} presented a more suitable correlation for IRSs. Its thorough study is the topic of ongoing research.} Also, $ \beta_{2,k} $ and $ \beta_{\mathrm{d},k} $ denote the path-loss of the IRS-UE $ k $ and BS-UE $ k $ links, respectively. Furthermore, $ {\mathbf{z}}_{k}\sim \mathcal{CN}\left({\mathbf{0}},\mat{\mathrm{I}}_{N}\right) $ and $ {\mathbf{z}}_{\mathrm{d},k} \sim \mathcal{CN}\left({\mathbf{0}},\mat{\mathrm{I}}_{N}\right) $ describe the corresponding fast-fading vectors.
By taking into account that the IRS is designed to be installed in a location providing an LoS channel with the BS, the channel matrix $ {\mathbf{H}}_{1} $ will likely have rank one, which results in performance gains only when $ K=1 $ \cite{Kammoun2020}. However, an MU scenario requires $ \mathrm{rank}\left({\mathbf{H}}_{1}\right)\ge K $. The higher rank could be achieved by placing the IRS close to the BS or by
deterministic scattering between the BS and the IRS. The high rank LoS channel $ {\mathbf{H}}_{1} $ can be obtained as
\begin{align}
[{\mathbf{H}}_{1}]_{m,n}=\sqrt{\beta_{1}} \exp\big(j \frac{2 \pi }{\lambda}\left(m-1\right)d_{\mathrm{BS}}\sin \theta_{1,n}\sin \phi_{1,n}\nonumber\\
+\left(n-1\right)d_{\mathrm{IRS}}\sin \theta_{2,m}\sin \phi_{2,m}\big)\!,
\end{align}
where $ \lambda $ is the carrier wavelength, $ \beta_{1} $ is the path-loss between the BS and IRS while $ d_{\mathrm{BS}} $ and $ d_{\mathrm{IRS}} $ are the inter-antenna separation at the BS and inter-element separation at the IRS, respectively \cite{Nadeem2020}. Also, $ \theta_{1,n} $, $ \phi_{1,n} $ describe the elevation and azimuth LoS angles of departure (AoD) at the BS with respect to IRS element $ n $, and $ \theta_{2,n} $, $ \phi_{2,n} $ describe the elevation and azimuth LoS angles of arrival (AoA) at the IRS. In practice, $ d_{\mathrm{BS}} $ and $ d_{\mathrm{IRS}} $ are known by construction while the angles, depending only on the locations, can be calculated when the locations are given according to \cite{Hu2020}. It is worthwhile to mention that the estimation of the correlation matrices could also be obtained similar to $ {\mathbf{H}}_1 $ since the dependence of their expressions on the distances and the angles is similar.
\subsection{Ideal Uplink Signal Model}\label{SignalModel}
The ideal received complex baseband signal vector by the BS is written as
\begin{align}
{\mathbf{y}}= & \sum_{k =1}^{K} \left({\mathbf{h}}_{\mathrm{d},k}+{\mathbf{H}}_{1}{\boldsymbol{\Theta}} {\mathbf{h}}_{2,k}\right)x_{k}+
{\mathbf{w}},\label{eq:Ypt1}
\end{align}
where $ {\mathbf{w}} \sim {\cal C}{\cal N}\left({\mathbf{0}},\sigma^2\mat{\mathrm{I}}_{M}\right) $ is the additive white Gaussian noise (AWGN) vector at the BS while ${\boldsymbol{\Theta}}=\mathrm{diag}\left( \alpha_{1}e^{j \theta_{1}}, \ldots, \alpha_{N}e^{j \theta_{N}} \right)\in\mathbb{C}^{N\times N}$ is the RBM being diagonal and representing the response of the $ N $ elements with $ \theta_{n} \in [0,2\pi]$ and $ \alpha_{n}\in [0,1] $ denoting the phase and amplitude coefficient for element $ n $, respectively. As commonly assumed due to recent advances towards lossless metasurfaces \cite{Epstein2016,Badloe2017}, we set $ \alpha_{n}=1~ \forall n$, i.e., we assume maximum signal reflection. For the sake of exposition, given the RBM, we denote the overall channel vector $ {\mathbf{h}}_{k}={\mathbf{h}}_{\mathrm{d},k}+ {\mathbf{H}}_{1}{\boldsymbol{\Theta}} {\mathbf{h}}_{2,k} $, distributed as $ {\mathbf{h}}_{k}\sim {\cal C}{\cal N}\left( 0, {\mathbf{R}}_{k} \right) $, where $ {\mathbf{R}}_{k}= \beta_{\mathrm{d},k}{\mathbf{R}}_{\mathrm{BS},k}+ \beta_{2,k}{\mathbf{H}}_{1} {\boldsymbol{\Theta}} {{\mathbf{R}}}_{\mathrm{IRS},k}{\boldsymbol{\Theta}}^{\H}{\mathbf{H}}_{1}^{\H}$.
\subsection{Hardware Impairments}\label{HardwareImpairments}
In this work, we consider two distinct types of HWIs in an IRS-assisted system: 1) the aggregate additive HWIs at the transceiver, and 2) the HWIs emerged from the passive elements of the IRS. Henceforth, we denote them T-HWIs and IRS-HWIs, respectively.
\subsubsection{T-HWIs}
The majority of papers in the IRS literature have relied on the unrealistic assumption of ideal transceiver hardware. Especially, next-generation antenna deployments with a large number of antennas such as a massive MIMO systems assisted by an IRS should be implemented with cheap hardware, in order to be cost-efficient as the number of antennas increases. However, cheaper hardware results in lower quality with more severe HWIs that are more power consuming and degrade further the system performance. Instead, we take into account the additive distortions at both the transmitter and the receiver being Gaussian distributed with average powers proportional to the average transmit and received signals, respectively \cite{Schenk2008}. We would like to mention that although Gaussian modeling for the HWIs can be assumed rudimentary, it is used widely because its tractability allows extracting primary conclusions, e.g., see the recent works \cite{Zhang2020,Papazafeiropoulos2021b}. The Gaussianity results by means of the aggregate contribution of many impairments. Especially, let $ p_{k}=\mathbb{E} \{|x_{k}|^{2}\}$ be the transmit power from UE $ k $ having transmit signal $ x_{k} $ and $ {\mathbf{h}}_{k} \in \mathbb{C}^{M\times 1}$ be the channel vector of this UE. The additive transceiver distortions during the uplink are described in terms of conditional distributions with respect to the channel realizations as
\begin{align}
\delta_{\mathrm{t},k}&\sim {\cal C}{\cal N}\left( 0, \Lambda_{k} \right),\label{eta_tU} \\
\hbox{\boldmath$\delta$}_{\mathrm{r}}&\sim {\cal C}{\cal N} \left( {\mathbf{0}},\bm \Upsilon \right)\label{eta_rU},
\end{align}
where $ \Lambda_{k}= \kappa_{\mathrm{UE}}p_{k}$ and
$\bm \Upsilon =\kappa_{\mathrm{BS}}\sum_{i=1}^{K}p_{i}$ $ \mathrm{diag}\left( |h_{i,1}|^{2},\ldots,|h_{i,M}|^{2} \right) $ with $ {\mathbf{h}}_{i}=\left[h_{i,1},\ldots,h_{i,M}\right]^{{\scriptscriptstyle\mathsf{T}}} $. The variance $ \bm \Upsilon $ can also be written as
$\bm \Upsilon =\kappa_{\mathrm{BS}}\sum_{i=1}^{K}p_{i} \mat{\mathrm{I}}_{M}\circ {\mathbf{h}}_{i} {\mathbf{h}}_{i}^{\H}$. The proportionality parameters $\kappa_{\mathrm{UE}}$ and $\kappa_{\mathrm{BS}}$ express the severity of the residual impairments at the transmitter and receiver side, and are met in applications in terms of the error vector magnitude (EVM)~\cite{Holma2011}. For example, the EVM at the BS is defined as
\begin{align}
\mathrm{EVM}_{\mathrm{BS}}=\sqrt{\frac{\mathbb{E}[\|\hbox{\boldmath$\delta$}_{\mathrm{r}}\|^{2}_{2}]}{\mathbb{E}[\|{\mathbf{y}}\|_{2}^{2}]}}=\sqrt{\frac{\mathrm{tr}\left(\bm \Upsilon \right)}{\mathbb{E}[\|{\mathbf{y}}\|_{2}^{2}]}}= \sqrt{\kappa_{\mathrm{BS}}},
\end{align}
where the expectations take place for a specific channel realization. Here, for the sake of simplicity, we assume that the parameter $ \kappa_{\mathrm{UE}} $ is identical for all UEs.
\subsubsection{IRS-HWIs} Taking into account that the reflection phases of the IRS passive elements cannot be configured with infinite precision, they can be modeled in terms of phase errors \cite{Badiu2019}.\footnote{We focus on the main IRS impairment, being the imperfection of phases configuration by assuming unity reflection amplitude, i.e., full signal reflection as in prior works e.g., see \cite{Basar2019,Wu2019a,Pan2020}. However, recently, it was suggested that the reflection amplitude can be phase-dependent due to hardware limitations \cite{Abeywickrama2020}, which requires a separate analysis and is left for future work.} In particular, IRS-HWIs are mathematically described by means of a random diagonal phase error matrix consisting of $ N $ random phase errors, i.e., $ \widetilde{{\boldsymbol{\Theta}}} =\diag\left( e^{j \tilde{\theta}_{1}}, \ldots, e^{j \tilde{\theta}_{N}} \right)\in\mathbb{C}^{N\times N}$, where $ \tilde{\theta}_{i}, i=1,\ldots,N $ are the random phase errors being i.i.d. randomly distributed in $ [-\pi, \pi) $ according to a certain circular distribution. Also, we assume that the PDF of $\tilde{ \theta}_{i} $ is symmetric and its mean direction is zero, i.e., $ \arg\left(\mathbb{E}[\mathrm{e}^{j \tilde{\theta}_{i}}]\right)=0 $ \cite{Badiu2019}. The most widely used PDFs, being able to describe the phase noise are the uniform and the Von Mises distributions \cite{Badiu2019}, where:
\begin{itemize}
\item the uniform distribution expresses completely lack of knowledge (random reflection) and its characteristic function (CF) denoted by $ m $ is $0 $,
\item the Von Mises distribution with a zero-mean and concentration parameter $ \kappa_{\tilde{\theta}} $, capturing the accuracy of the estimation, has a CF $ m= \frac{\mathrm{I}_{1}\!\left(\kappa_{\tilde{\theta}}\right)}{\mathrm{I}_{0}\!\left(\kappa_{\tilde{\theta}}\right)}$ with $ \mathrm{I}_{p}\!\left(\kappa_{\tilde{\theta}}\right)$ being the modified Bessel function of the first kind and order
$ p $.
\end{itemize}
\subsection{Realistic Uplink Signal Model with HWIs}\label{SignalModel1}
Overall, the realistic received signal vector by the BS after having incorporated both the T-HWIs and IRS-HWIs in \eqref{eq:Ypt1} is given by
\begin{align}
{\mathbf{y}}= & \sum_{k =1}^{K} \left({\mathbf{h}}_{\mathrm{d},k}+{\mathbf{H}}_{1}{\boldsymbol{\Theta}}\widetilde{{\boldsymbol{\Theta}}} {\mathbf{h}}_{2,k}\right)\left(x_{k} +\delta_{\mathrm{t},k}\right)+\hbox{\boldmath$\delta$}_{\mathrm{r}}+
{\mathbf{w}}\label{eq:Ypt}.
\end{align}
Now, given the RBM, the overall channel vector is $ {\mathbf{h}}_{k}={\mathbf{h}}_{\mathrm{d},k}+ {\mathbf{H}}_{1}{\boldsymbol{\Theta}}\widetilde{{\boldsymbol{\Theta}}} {\mathbf{h}}_{2,k} $, distributed as $ {\mathbf{h}}_{k}\sim {\cal C}{\cal N}\left( 0, {\mathbf{R}}_{k} \right) $, where $ {\mathbf{R}}_{k}= \beta_{\mathrm{d},k}{\mathbf{R}}_{\mathrm{BS},k}+ \beta_{2,k}{\mathbf{H}}_{1} {\boldsymbol{\Theta}} \widetilde{{\mathbf{R}}}_{\mathrm{IRS},k}{\boldsymbol{\Theta}}^{\H}{\mathbf{H}}_{1}^{\H}$ with $ \widetilde{{\mathbf{R}}}_{\mathrm{IRS},k} $ given by
\begin{align}
&\widetilde{{\mathbf{R}}}_{\mathrm{IRS},k}=\mathbb{E}[ \widetilde{{\boldsymbol{\Theta}}}{\mathbf{R}}_{\mathrm{IRS},k}\widetilde{{\boldsymbol{\Theta}}}^{\H}]\\
&=\!\!\begin{bmatrix}\!
r_{11} &\!\!\!\! r_{12} \mathbb{E}_{\tilde{\theta}}[e^{j \tilde{\theta}_{1}-j \tilde{\theta}_{2}}]&\!\!\!\! \dots&\!\!\!\!r_{1N} \mathbb{E}_{\tilde{\theta}}[e^{j \tilde{\theta}_{1}-j \tilde{\theta}_{N}}]\\
r_{21} \mathbb{E}_{\tilde{\theta}}[e^{j \tilde{\theta}_{2}-j \tilde{\theta}_{1}} ]&\!\!\!\! r_{22} & \!\!\!\!\dots&\!\!\!\!r_{2N} \mathbb{E}_{\tilde{\theta}}[e^{j \tilde{\theta}_{2}-j \tilde{\theta}_{N}}]\\
\vdots&\!\!\!\!\vdots&\!\!\!\!\ddots&\!\!\!\!\vdots\\
r_{N1} \mathbb{E}_{\tilde{\theta}}[e^{j \tilde{\theta}_{N}-j \tilde{\theta}_{1}} ]&\!\!\!\! r_{N2} \mathbb{E}_{\tilde{\theta}}[e^{j \tilde{\theta}_{N}-j \tilde{\theta}_{2}} ] &\!\!\!\! \dots&\!\!\!\!r_{NN}\!\end{bmatrix} \label{cor1}\\
&=\!\begin{bmatrix}
r_{11} & m^{2} r_{12} & \dots&m^{2} r_{1N}\\
m^{2} r_{21}& r_{22} & \dots&m^{2} r_{2N} \\
\vdots&\vdots&\ddots&\vdots\\
m^{2} r_{N1} &m^{2} r_{N2} & \dots&r_{NN}\end{bmatrix} \label{cor2}\\
&=m^{2}{\mathbf{R}}_{\mathrm{IRS},k}+\left(1-m^{2}\right)\diag\left({\mathbf{R}}_{\mathrm{IRS},k}\right)\\
&=m^{2}{\mathbf{R}}_{\mathrm{IRS},k}+\left(1-m^{2}\right)\mat{\mathrm{I}}_{N}.\label{cor3}
\end{align}
In \eqref{cor1}, $ r_{ij} $ is the $ ij $th element of the correlation matrix $ {\mathbf{R}}_{\mathrm{IRS},k} $. Also, in \eqref{cor2}, we have exploited that $ \tilde{\theta}_{i} $ are i.i.d. distributed with a symmetric PDF while we have substituted with $ m $ the corresponding CF. The next equation is written in a compact form in terms of $ {\mathbf{R}}_{\mathrm{IRS},k} $. In \eqref{cor3}, we have accounted for that in correlated Rayleigh fading, we have $ r_{ii}=1, ~\forall i $. Notably, this is a very useful equation describing the dependence on the IRS-HWIs, i.e., the phase noises from the IRS elements.
\begin{remark}\label{rem1}
In the case of the uniform distribution, where the characteristic function is zero ($ m=0 $), we obtain $ \widetilde{{\mathbf{R}}}_{\mathrm{IRS},k}=\mat{\mathrm{I}}_{N}$, i.e., there is no dependence on the phase noises. In such case, the overall covariance becomes $ {\mathbf{R}}_{k} =\beta_{\mathrm{d},k}{\mathbf{R}}_{\mathrm{BS},k}+\beta_{2,k}{\mathbf{H}}_{1}{\mathbf{H}}_{1}^{\H}$, which obviously has no dependence on the RBM and cannot be optimized. Hence, the IRS cannot be exploited. Moreover, no knowledge of $ {\mathbf{R}}_{\mathrm{IRS},k} $ is required at the BS. However, even in this case, the IRS still contributes with an additional signal to the receiver. Although it is expected to be weak, it is beneficial. Especially, when there is no direct signal. Note that these cases are very difficult to appear in practice. Specifically, regarding the independent Rayleigh assumption, it is uncommon to appear as mentioned in \cite{Bjoernson2020} while always there will be some knowledge and control in the reflection at the IRS, which means that the uniform distribution is not meaningful in practice (see Sec. II.C.2). On the contrary, any other circular PDF for the description of the phase errors allows studying the impact of these errors, and mostly, taking advantage of the IRS.
\end{remark}
\section{CE with HWIs}\label{ChannelEstimation}
In practice, a BS does not have perfect CSI but estimates its channel by a TDD operation including an uplink training phase with pilot symbols \cite{Bjoernson2017}. Differently to conventional MISO with/without relay systems, the IRS implemented by means of passive elements, is not able to send pilots to the BS for CE or process the received pilot symbols from the UEs to obtain the corresponding estimated channels. Contrary to existing works providing separately the estimated direct and cascaded channels \cite{Nadeem2020}, we provide the estimate of the overall channel. For example, compared to \cite{Mishra2019} and \cite{Nadeem2020}, we perform the CE in a single phase instead of $ N+1 $ phases. In particular, the former is known as ON/OFF channel
estimation and addresses a single UE setting while the latter assumes multiple UEs. Hence, the achievable SE in our case is higher since the pre-log factor in the SE is lower (lower training overhead). Also, we achieve to derive the covariance of the estimated cascaded channel vector while, in \cite{Nadeem2020}, it is assumed unknown.
The CE protocol assumes that the total uplink training phase has a duration of $ \tau $ sec. Let the UEs transmit orthogonal pilot sequences. Especially, we denote by ${\mathbf{x}}_{p,k}=[x_{p,k,1}, \ldots, x_{p,k,\tau}]^{{\scriptscriptstyle\mathsf{T}}}\in \mathbb{C}^{\tau\times 1} $ the pilot sequence of UE $ k $ with $ {\mathbf{x}}_{p,k}^{\H}{\mathbf{x}}_{p,l}=0~\forall k\ne l$ and $ {\mathbf{x}}_{p,k}^{\H}{\mathbf{x}}_{p,k}= \tau P$ joules, where $ P =|x_{p,k,i}|^{2} ,~\forall k,i$ is the common average transmit power per UE during the training phase.
The received signal at the BS with T-HWIs and IRS-HWIs during the training period is given by
\begin{align}
{\mathbf{Y}}^{\mathop{\mathrm{tr}}\nolimits}=\sum_{i=1}^{K}{\mathbf{h}}_{i}\!\left({\mathbf{x}}_{\mathrm{p},i}^{\H} +\hbox{\boldmath$\delta$}_{\mathrm{t},i}^{\H} \right)+{\boldsymbol{\Delta}}_{\mathrm{r}}+
{\mathbf{W}}^{\mathop{\mathrm{tr}}\nolimits},\label{train1}
\end{align}
where $ \hbox{\boldmath$\delta$}_{\mathrm{t},i} \in \mathbb{C}^{M \times 1} \sim \mathcal{CN}\left({\mathbf{0}},\kappa_{\mathrm{UE}} P\mat{\mathrm{I}}_{\tau}\right) $ is the $ \tau \times 1$ additive transmit HWI vector while $ {\boldsymbol{\Delta}}_{\mathrm{r}} \in \mathbb{C}^{M \times \tau} $ is the additive receive HWI matrix where each column is distributed as $ \hbox{\boldmath$\delta$}_{\mathrm{r}} \in \mathbb{C}^{M \times 1} \sim \mathcal{CN}\left({\mathbf{0}},\bm \Upsilon \right)$ with $ \bm \Upsilon =\kappa_{\mathrm{BS}} P\sum_{i=1}^{K} \mat{\mathrm{I}}_{M}\circ {\mathbf{h}}_{i} {\mathbf{h}}_{i}^{\H}$. Note that the phase noise is hidden inside the expression of $ {\mathbf{h}}_{k} $. In addition, $ {\mathbf{W}}^{\mathop{\mathrm{tr}}\nolimits} \in \mathbb{C}^{M \times \tau} $ is the AWGN matrix at the BS with independent columns, each one distributed as $ \mathcal{CN}\left({\mathbf{0}},\sigma^2\mat{\mathrm{I}}_{M}\right)$.
The received training signal at the BS, given by \eqref{train1}, is multiplied by the transmitted training sequence from UE $ k $ to eliminate the interference caused by other UEs, and obtain
\begin{align}
{\mathbf{r}}_{k}={\mathbf{h}}_{k}+\sum_{i=1}^{K}\frac{\tilde{\delta}_{\mathrm{t},i}}{ \tau P}{\mathbf{h}}_{i} +\frac{\tilde{\hbox{\boldmath$\delta$}}_{\mathrm{r}}+{\mathbf{w}}_{k}}{ \tau P},\label{train2}
\end{align}
where $ \tilde{\delta}_{\mathrm{t},i}=\hbox{\boldmath$\delta$}_{\mathrm{t},i}^{\H}{\mathbf{x}}_{\mathrm{p},k} $, $\tilde{\hbox{\boldmath$\delta$}}_{\mathrm{r}}={\boldsymbol{\Delta}}_{\mathrm{r}}{\mathbf{x}}_{\mathrm{p},k} $, and $ {\mathbf{w}}_{k}={\mathbf{W}}^{\mathop{\mathrm{tr}}\nolimits} {\mathbf{x}}_{\mathrm{p},k}$.
\begin{proposition}\label{PropositionDirectChannel}
The LMMSE estimate of the overall channel $ {\mathbf{h}}_{k} $ is given by
\begin{align}
\hat{{\mathbf{h}}}_{k}={\mathbf{R}}_{k}{\mathbf{Q}}_{k} {\mathbf{r}}_{k},\label{estim1}
\end{align}
where $ {\mathbf{Q}}_{k}\!=\! \left(\!{\mathbf{R}}_{k}\!+ \!\frac{\kappa_{\mathrm{UE}}}{\tau }\!\sum_{i=1}^{K} \!{\mathbf{R}}_{i}\!+\! \frac{\kappa_{\mathrm{BS}}}{\tau}\!\sum_{i=1}^{K}\!\mat{\mathrm{I}}_{M}\!\circ\! {\mathbf{R}}_{i} \!+\!\frac{\sigma^2}{ \tau P }\mat{\mathrm{I}}_{M}\!\right)^{\!-1}$, and $ {\mathbf{r}}_{k}$ is the noisy observation of the effective overall channel from UE $k$ given by \eqref{train2}.
\end{proposition}
\begin{proof}
The proof is provided in Appendix~\ref{Proposition1}.
\end{proof}
According to the property of orthogonality of LMMSE estimation, the overall perfect channel is given by
\begin{align}
{\mathbf{h}}_{k}=\hat{{\mathbf{h}}}_{k}+\tilde{{\mathbf{h}}}_{k},\label{current} \end{align}
where $\hat{{\mathbf{h}}}_{k}$ and $\tilde{{\mathbf{h}}}_{k} $ have zero mean and variances $ {\boldsymbol{\Psi}}_{k}\!=\!{\mathbf{R}}_{k}{\mathbf{Q}}_{k}{\mathbf{R}}_{k}$ and $ \tilde{{\boldsymbol{\Psi}}}_{k}={\mathbf{R}}_{k}-{\boldsymbol{\Psi}}_{k}$, respectively. Contrary to conventional estimation theory concerning independent Gaussian noise, $\hat{{\mathbf{h}}}_{k}$ and $\tilde{{\mathbf{h}}}_{k}$ are neither independent nor jointly complex Gaussian vectors because the effective distortion noises are not Gaussian, e.g., $ {\mathbf{h}}_{k} \tilde{\delta}_{\mathrm{t},k}$ is the product between two Gaussian variables. However, $\hat{{\mathbf{h}}}_{k}$ and $\tilde{{\mathbf{h}}}_{k} $ are uncorrelated and each of them has zero mean~\cite{Bjoernson2017}. In the unrealistic case of perfect HWIs, the LMMSE estimator of the overall channel vector coincides with the optimal MMSE estimator. Notably, the CE can be easily generalized to include other fading models such as independent Rayleigh fading, where $ {\mathbf{R}}_{\mathrm{IRS},k}=\mat{\mathrm{I}}_{N} $ and $ {\mathbf{R}}_{\mathrm{BS},k}=\mat{\mathrm{I}}_{M} $.
\begin{remark}\label{rem3}
A comparison with other CE schemes is difficult since the majority of works such as \cite{He2019} does not yield analytical expressions, while the proposed method is indicated for future closed-form manipulations. Compared to \cite{Nadeem2020} requiring $ N+1 $ subphases, our proposed method has a lower training overhead requiring only one phase and has achieved to obtain the estimated cascaded channel vector while, therein, only the estimated channel (scalar) concerning each element was obtained.
\end{remark}
\begin{remark}\label{rem2}
Generally, the covariances $ {\mathbf{R}}_{k} $, $ {\boldsymbol{\Psi}}_{k} $, and $ \tilde{{\boldsymbol{\Psi}}}_{k} $ depend on both the T-HWIs and the IRS-HWIs. In the special case of uniformly distributed phase errors, these covariances do not depend on these errors or the reflect phase matrix $ {\boldsymbol{\Theta}} $. Then, we can not take benefit from any IRS optimization to minimize the estimation error and achieve better estimation.
\end{remark}
The study of the NMSE is insightful. Specifically, we define
\begin{align}
\mathrm{NMSE}_{k}&=\frac{\mathop{\mathrm{tr}}\nolimits(\mathbb{E}[(\hat{{\mathbf{h}}}_{k}-{{\mathbf{h}}}_{k})(\hat{{\mathbf{h}}}_{k}-{{\mathbf{h}}}_{k})^{\H}])}{\mathop{\mathrm{tr}}\nolimits\left(\mathbb{E}[{{\mathbf{h}}}_{k}{{\mathbf{h}}}_{k}^{\H}]\right)}\\
&=1-\frac{\mathop{\mathrm{tr}}\nolimits({\boldsymbol{\Psi}}_{k})}{\mathop{\mathrm{tr}}\nolimits({\mathbf{R}}_{k})}.\label{nmse1}
\end{align}
The T-HWIs are found inside $ {\boldsymbol{\Psi}}_{k} $ in terms of $ {\mathbf{Q}}_{k} $ while the IRS-HWIs (phase noise) appear inside $ {\mathbf{R}}_{k} $. From \eqref{nmse1}, we observe that an increase of the T-HWIs results in the increase of the $ \mathrm{NMSE}_{k} $. Moreover, according to Remark \ref{rem1}, if the phase errors are uniformly distributed, the NMSE does not depend on the RBM and the NMSE can not be optimized. A similar observation takes place if uncorrelated Rayleigh fading is assumed.
%
\section{Uplink Data Transmission with HWIs}\label{PerformanceAnalysis}
In this section. we focus on the derivation of the uplink achievable sum SE of a practical IRS-aided MU-MISO setup with HWIs. The received signal by the BS can be written as
\begin{align}
{\mathbf{y}}=\sum_{i=1}^{K}{\mathbf{h}}_{i}\left(s_{i}+\delta_{\mathrm{t},i}\right) +\hbox{\boldmath$\delta$}_{\mathrm{r}}+{\mathbf{n}},\label{ULTrans}
\end{align}
where ${\mathbf{n}}\sim \mathcal{CN}\left({\mathbf{0}},\sigma^2\mat{\mathrm{I}}_{M}\right) $ is the AWGN at the BS and phase noises are found inside the expression of $ {\mathbf{h}}_{i} $, while $ \delta_{\mathrm{t},i}$ and $ \hbox{\boldmath$\delta$}_{\mathrm{r}} $ correspond to the T-HWIs.
\subsection{Achievable SE} \label{lower1}
The BS estimates $ s_{k} $ from UE $ k $ by means of~\eqref{ULTrans} in terms of linear single-user detection by applying the receive combining vector $ {\mathbf{v}}_{k} $ as $ \hat{s}_{k}= {\mathbf{v}}_{k}^{\H}{\mathbf{y}}_{}$.
Moreover, we exploit the use-and-then-forget (UatF) bound, suggested for systems with a large number of antennas ( $ M>8 $) \cite{Bjoernson2017}, in order to obtain a closed-form expression of the SE. Note that this bound can be applied with different channel estimators (not only LMMSE) and decoders. Specifically, $ \hat{s}_{k} $ can be rewritten as
\begin{align}
\hat{s}_{k}&=\underbrace{\sqrt{\rho_{k}}\mathbb{E}\left\{{\mathbf{v}}_{k}^{\H}{{\mathbf{h}}}_{k}\right\} s_{k}}_{\mathrm{DS}_{k}} + \underbrace{\sqrt{\rho_{k}}\left({\mathbf{v}}_{k}^{\H}{{\mathbf{h}}}_{k}-\mathbb{E}\left\{{\mathbf{v}}_{k}^{\H}{{\mathbf{h}}}_{k}\right\}\right) s_{k}}_{\mathrm{BU}_{k}}\nonumber\\
&+\sum_{i\ne k}^{K}\underbrace{\sqrt{\rho_{i}}{\mathbf{v}}_{k}^{\H}{{\mathbf{h}}}_{i}s_{i}}_{\mathrm{MUI}_{ik}} +\sum_{i=1}^{K}\underbrace{{\mathbf{v}}_{k}^{\H}{{\mathbf{h}}}_{i}\delta_{\mathrm{i},k}}_{\mathrm{TD}_{i}}+\underbrace{{\mathbf{v}}_{k}^{\H}\hbox{\boldmath$\delta$}_{\mathrm{r}}}_{ \mathrm{RD}_{k}}+\underbrace{{\mathbf{v}}_{k}^{\H}}_{\mathrm{RN}_{k}},
\end{align}
where
$ \mathrm{DS}_{k} $, $ \mathrm{BU}_{k} $, $ \mathrm{MUI}_{ik} $ express the desired signal (DS) part, the beamforming gain uncertainty (BU), and each term of the sum describing the MU interference (MUI). Also, $ \mathrm{TD}_{i} $, $ \mathrm{RD}_{k} $, and $ \mathrm{RN}_{k} $ express the transmit distortion, the receive distortion, and the receiver AWGN noise. Next, by applying a standard bound technique assuming worst-case uncorrelated additive noise for the inter-user interference and the distortion noise \cite{Hassibi2003}, we derive a lower bound on the uplink average SE in bps/Hz, which is known as the use-and-then-forget bound in the massive MIMO (mMIMO) literature \cite{Bjoernson2017}. In particular, the achievable SE is given by
\begin{align}
\mathrm{SE}_{k} =\frac{\tau_{\mathrm{c}}-\tau}{\tau_{\mathrm{c}}}\log_{2}\left ( 1+\gamma_{k}\right)\!,\label{LowerBound}
\end{align}
where the pre-log fraction expresses the percentage of samples per coherence block used for uplink data transmission and $ \gamma_{k}=\frac{S_{k}}{I_{k}}$ is the uplink SINR with
\begin{align}
S_{k}&=|\mathrm{DS}_{k} |^{2},\label{sig11}\\
I_{k}&=\mathbb{E}\left\{|\mathrm{BU}_{k}|^{2}\right\}+\sum_{i\ne k}^{K}\mathbb{E}\left\{\mathrm{MUI}_{ik}\right\}+\sum_{i=1}^{K} \mathbb{E}\left\{|\mathrm{TD}_{i}|^{2}\right\}\nonumber\\
& +\mathbb{E}\left\{|\mathrm{RD}_{k}|^{2}\right\}\!+\!\mathbb{E}\left\{|\mathrm{RN}_{k}|^{2}\right\}\!\label{int1}
\end{align}
describing the desired signal power and the interference plus noise power. For the sake of further convenience, we denote by $ \sigma_{\mathrm{UE}}^{2}= \sum_{i=1}^{K} \mathbb{E}\left\{|\mathrm{TD}_{i}|^{2}\right\}$ and $ \sigma_{\mathrm{BS}}^{2}= \mathbb{E}\left\{|\mathrm{RD}_{k}|^{2}\right\}$ the variances of the additive transmit and receive HWIs from all transmit UEs and at the output of the decoder.
Generally, MRC and conventional MMSE decoders are the most common linear receivers for the uplink of next-generation systems such as mMIMO \cite{Hoydis2013,Papazafeiropoulos2015a}. However, the expectations in \eqref{sig11} and \eqref{int1} cannot be derived in closed-form in the case of the optimal MMSE receiver except if the deterministic equivalent analysis is applied \cite{Hoydis2013,Papazafeiropoulos2015a,Papazafeiropoulos2016}. Also, the next step that includes the optimization with respect to reflection coefficients would be quite intractable. Hence, we focus on the derivation of a closed-form SINR by applying MRC decoding, which can be obtained even for a finite number of BS antennas. Thus, below we assume $ {\mathbf{v}}_{k}=\hat{{\mathbf{h}}}_{k}$. However, given the higher performance expected by MMSE decoding, its application in the study of HWIs with statistical CSI according to the proposed methodology is the topic of ongoing work.
\begin{Theorem}\label{theorem:ULDEMMSE}
Given the RBM $ {\boldsymbol{\Theta}} $, the uplink achievable SINR of UE $k$ with MRC decoding in an IRS-assisted MU-MISO system, accounting for imperfect CSI and HWIs, is given by
\begin{align}
\gamma_{k}\!=\! \frac{\bar{S}_{k}}{ \bar{I}_{k}},
2 \end{align}
where
\begin{align}
& \bar{S}_{k}\!=\!\rho_{k}|\!\mathop{\mathrm{tr}}\nolimits\left({\boldsymbol{\Psi}}_{k}\right)\!|^{2},\label{Num1}\\
& \bar{I}_{k}\!=\!\left(1\!+\!\kappa_{\mathrm{UE}}\right)\!\!\left(\sum_{i=1}^{K}\!\rho_{i}\! \mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k} {\mathbf{R}}_{i}\right)\!-\!\rho_{k}\!\mathop{\mathrm{tr}}\nolimits\left( {\boldsymbol{\Psi}}_{k}^{2}\right)\!\!\!\right)\!\!+\!\rho_{k}\kappa_{\mathrm{UE}}|\!\mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k}\!\right)|^{2}\nonumber\\
&+\!{\boldsymbol{\kappa}}_{\mathrm{BS}} \!\left(\!\!\rho_{k}|\! \mathop{\mathrm{tr}}\nolimits \!\left( \mat{\mathrm{I}}_{M}\!\circ\! {\boldsymbol{\Psi}}_{k}\right)\!|^{2}\! +\!\!\sum_{i=1}^{K}\!\rho_{i}\!\mathop{\mathrm{tr}}\nolimits \left(\left(\mat{\mathrm{I}}_{M}\!\circ\! {\mathbf{R}}_{i}\right) \!{\boldsymbol{\Psi}}_{k}\right)\!\!\right)\!\!+\!\sigma^2\! \mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k}\right)\!.\label{Den1}
\end{align}
\end{Theorem}
\proof The proof is provided in Appendix~\ref{theorem1}.\endproof
\begin{remark}
Theorem \ref{theorem:ULDEMMSE} provides the uplink achievable SINR with MRC under imperfect CSI in closed-form. Notably, it shows the impact of the unavoidable HWIs. Especially, it depends directly on the T-HWIs by means of $ {\boldsymbol{\kappa}}_{\mathrm{BS}} $ and $ {\boldsymbol{\kappa}}_{\mathrm{UE}} $. The impact of the phase noise appears indirectly through the covariance matrices. Moreover, the expression of $ \gamma_{k} $ depends only on slowly-varying large-scale statistics.
\end{remark}
Based on $ \gamma_{k} $, provided by Theorem \ref{theorem:ULDEMMSE}, the system (sum) SE in bps/Hz is obtained as
\begin{align}
\mathcal{R}=\frac{\tau_{\mathrm{c}}-\tau}{\tau_{\mathrm{c}}}\sum_{i=1}^{K}\log_{2}\left(1+ \gamma_{ i}\right).\label{sumse}
\end{align}
\subsection{IRS Design Problem: Formulation and Solution}\label{IRSdesign}
IRS-aided architectures require to design the corresponding RBM, found inside the covariance matrices, in order to maximize the sum SE given by \eqref{sumse}. Hence, by resorting to the common assumption of infinite resolution phase shifters, herein, we formulate and solve the RB design problem under MRC and realistic conditions accounting for imperfect CSI and HWIs as
\begin{align}\begin{split}
(\mathcal{P}1)~~~~~~~\max_{{\boldsymbol{\Theta}}} ~~~ &\mathcal{R}\\
\mathrm{s.t}~~~&|\phi_{n}|=1,~~ n=1,\dots,N,
\end{split}\label{Maximization}
\end{align}
with $ \mathcal{R} $ given by \eqref{sumse} and $ \phi_{n}= \exp\left(j \theta_{n}\right) $ are the elements of $ {\boldsymbol{\Theta}} $. Obviously, $ (\mathcal{P}1) $ is a non-convex maximization problem with respect to $ {\boldsymbol{\Theta}} $ with a unit-modulus constraint regarding $ \phi_{n} $.
\begin{remark}\label{rem5}
If the phase noise is uniformly distributed, the covariance matrices will not include the RBM according to Remark \ref{rem2}. Hence, the SINR/SE cannot be optimized, and the IRS does not serve its purpose.
\end{remark}
\begin{remark}
Given that HWIs at the transceiver and IRS degrade the performance, the use of cheaper (lower quality) hardware will have a direct impact on the QoS. In such a case, a better RB design is suggested to compensate for the loss and improve the performance.
\end{remark}
Taking the expression of $ \gamma_{k} $ into account, the optimization problem takes the form of a constrained maximization problem with a solution given by means of projected gradient ascent until converging to a stationary point as in \cite{Kammoun2020}. At every step, we project the solution onto the closest feasible point satisfying the unit-modulus constraint concerning $ \phi_{n} $. In more detail, the procedure assumes the vectors $ {\mathbf{s}}^{i} =[\phi_{1}^{i}, \ldots, \phi_{N}^{i}]^{{\scriptscriptstyle\mathsf{T}}}$ including the induced phases at step $ i $. The next iteration point, resulting in the increase of $ \mathcal{R} $ towards to its convergence, is given by
\begin{align}
\tilde{{\mathbf{s}}}^{i+1}&={\mathbf{s}}^{i}+\mu {\mathbf{q}}^{i},\label{sol1}\\
{\mathbf{s}}^{i+1}&=\exp\left(j \arg \left(\tilde{{\mathbf{s}}}^{i+1}\right)\right),\label{sol2}
\end{align}
where $ \mu $ is the step size and $ {\mathbf{q}}^{i} $ describes the adopted ascent direction at step $ i $. In particular, we have $ [{\mathbf{q}}^{i}]_{n}= \pdv{ \mathcal{R}}{\phi_{n}^{*}} $, which is obtained by Proposition \ref{Prop:optimPhase} below. The suitable step size is computed at each iteration by means of the backtracking line search \cite{Boyd2004}. The solution of the problem, described by \eqref{sol1} and \eqref{sol2}, is found based on the projection problem $ \min_{|\phi_{n} |=1, n=1,\ldots,N}\|{\mathbf{s}}-\tilde{{\mathbf{s}}}\|^{2} $ under the unit-modulus constraint. The outline of the algorithm is described by Algorithm \ref{Algoa1}.
\begin{algorithm}
\caption{Projected Gradient Ascent Algorithm for the IRS Design}
1. \textbf{Initialisation}: $ {\mathbf{s}}^{0} =\exp\left(j\pi/2\right)\mat{\mathrm{1}}_{N}$, $ {\boldsymbol{\Theta}}^{0}=\diag\left({\mathbf{s}}^{0}\right) $, $ \mathcal{R}^{0}=f\left({\boldsymbol{\Theta}}^{0}\right) $ given by \eqref{sumse}; $ \epsilon>0 $\\
2. \textbf{Iteration} $ i $: \textbf{for} $ i=0,1,\dots, $ do\\
3. $[{\mathbf{q}}^{i}]_{n}= \pdv{ \mathcal{R}}{\phi_{n}^{*}}, n=1, \ldots,N $, where $\pdv{ \mathcal{R}}{\phi_{n}^{*}} $ is given by Proposition \ref{Prop:optimPhase};\\
4. \textbf{Find} $ \mu $ by backtrack line search$( f\left({\boldsymbol{\Theta}}^{0}\right),{\mathbf{q}}^{i},{\mathbf{s}}^{i})$ \cite{Boyd2004};\\
5. $ \tilde{{\mathbf{s}}}^{i+1}={\mathbf{s}}^{i}+\mu {\mathbf{q}}^{i} $;\\
6. $ {\mathbf{s}}^{i+1}=\exp\left(j \arg \left(\tilde{{\mathbf{s}}}^{i+1}\right)\right) $; $ {\boldsymbol{\Theta}}^{i+1}=\alpha \diag\left({\mathbf{s}}^{i+1}\right) $;\\
7. $ \mathcal{R}^{i+1}=f\left({\boldsymbol{\Theta}}^{i+1}\right) $;\\
8. \textbf{Until} $ \| \mathcal{R}^{i+1}- \mathcal{R}^{i}\|^{2} <\epsilon$; \textbf{Obtain} $ {\boldsymbol{\Theta}}^{*}={\boldsymbol{\Theta}}^{i+1}$;\\
9. \textbf{end for}\label{Algoa1}
\end{algorithm}
The convergence of the proposed algorithm to a local maximum can be guaranteed because it is bounded due to the power constraint and it increases by setting $ [{\mathbf{q}}^{i}]_{n}= \pdv{ \mathcal{R}}{\phi_{n}^{*}} $, where the backtracking line search is used to find a suitable step size.
\begin{proposition}\label{Prop:optimPhase}
The derivative of $ \mathcal{R} $ with respect to $ \phi_{n} $ is provided by
\begin{align}
\pdv{ \mathcal{R}}{\phi_{n}}=\sum_{k=1}^{K} \frac{\pdv{\bar{S}_{k}}{\phi^{*}_{n}}\bar{I}_{k}-\bar{S}_{k}\pdv{\bar{I}_{k}}{\phi^{*}_{n}}}{\ln (2) \bar{I}_{k}^{2}\left(1+\frac{\bar{S}_{k}}{\bar{I}_{k}}\right)},\label{der10}
\end{align}
where $ \bar{S}_{k} $, $ \bar{I}_{k} $ follow by \eqref{Num1}, \eqref{Den1} while
\begin{align}
&\pdv{\bar{S}_{k}}{\phi^{*}_{n}} =2\rho_{k}\mathop{\mathrm{tr}}\nolimits\left({\boldsymbol{\Psi}}_{k}\right){\mathbf{L}}\! \left({\mathbf{R}}_{k},{\mathbf{R}}_{k},\mat{\mathrm{I}}_{M} \right),\\
&\pdv{\bar{I}_{k}}{\phi^{*}_{n}}=\left(1+\kappa_{\mathrm{UE}}\right)\big(\sum_{i=1}^{K}\!\rho_{i}\big({\mathbf{L}} \big({\mathbf{R}}_{i}{\mathbf{R}}_{k},{\mathbf{R}}_{i}{\mathbf{R}}_{k},{\mathbf{R}}_{i} \big) \nonumber\\&\!\!+\!\alpha\beta_{2,k}[{\mathbf{H}}_{1}^{\H}{\boldsymbol{\Psi}}_{k}{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}]_{n,n}\big)\nonumber\\
&\!\!-\!2\rho_{k}{\mathbf{L}} \big({\boldsymbol{\Psi}}_{k}{\mathbf{R}}_{k},{\boldsymbol{\Psi}}_{k}{\mathbf{R}}_{k},{\boldsymbol{\Psi}}_{k}\big)\big)\nonumber\\
&\!\!+\!{\boldsymbol{\kappa}}_{\mathrm{BS}} \big(2\rho_{k} \mathop{\mathrm{tr}}\nolimits \big( \mat{\mathrm{I}}_{M}\circ {\boldsymbol{\Psi}}_{k}\big) {\mathbf{L}} \big({\mathbf{R}}_{k},{\mathbf{R}}_{k}, \mat{\mathrm{I}}_{M}\big)\!\nonumber\\ &\!\!+\!\sum_{i=1}^{K}\!\rho_{i}\big(\alpha\beta_{2,k}[{\mathbf{H}}_{1}^{\H}{\boldsymbol{\Psi}}_{k}{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}]_{n,n}\nonumber\\
& \!\!+\!{\mathbf{L}} \big(\big(\mat{\mathrm{I}}_{M}\circ {\mathbf{R}}_{i}\big){\mathbf{R}}_{k},\big(\mat{\mathrm{I}}_{M}\circ {\mathbf{R}}_{i}\big){\mathbf{R}}_{k},\mat{\mathrm{I}}_{M}\circ {\mathbf{R}}_{i}\big)\big)\big)\nonumber\\
&\!\!+\!2\rho_{k}\kappa_{\mathrm{UE}}\mathop{\mathrm{tr}}\nolimits\big({\boldsymbol{\Psi}}_{k}\big){\mathbf{L}} \big({\mathbf{R}}_{k},{\mathbf{R}}_{k},\mat{\mathrm{I}}_{M} \big)\!+\!\sigma^2 {\mathbf{L}} \big({\mathbf{R}}_{k},{\mathbf{R}}_{k},\mat{\mathrm{I}}_{M} \big)
\end{align}
with $ {\mathbf{L}}\! \left({\mathbf{A}},{\mathbf{B}}, {\mathbf{C}}\right) $ given by \eqref{LABC}
for any $ {\mathbf{A}}\in\mathbb{C}^{N \times N} $, $ {\mathbf{B}} \in\mathbb{C}^{N \times N} $, and $ {\mathbf{C}}\in \mathbb{C}^{N \times N} $. \end{proposition}
\proof The proof of Proposition~\ref{Prop:optimPhase} is given in Appendix~\ref{optimPhase}.\endproof
The RBM beamforming design, based on the gradient ascent, results in an outstanding
performance since the gradient ascent is obtained in a closed-from with low computational complexity based on simple matrix operations. In particular, the complexity of \eqref{der10} is $ \mathcal{O}\left(MN^{2}+M^{2}N+M^{2}K\right) $. Obviously, it depends on all fundamental system parameters, i.e., $ K $, $ M $, and $N $ but with a higher (square) dependence on $ M $ and $ N $.
\begin{remark}
If we do not have a closed-form expression for the SE, we cannot apply the proposed method. Also, although the proposed algorithm, given by \eqref{sol1} and \eqref{sol2}, does not provide a global optimum but a locally optimal solution due to the non-convexity of the initial optimization problem with respect to the phase shifts, it offers a good preliminary tool to study IRS-aided systems under realistic conditions in terms of imperfect CSI and HWIs.
\end{remark}
\begin{remark}
The dependence of the algorithm on the large-scale channel statistics achieves a reduction in the signal exchange overhead between the IRS controller and the BS since it will take place every several coherence intervals defined by the variation of these statistics. On the contrary, on models, relying on the instantaneous CSI, the optimization should take place at every coherence interval, which results in large overhead, especially, when the IRS is large. We highlight that the proposed method can be exploited in both low-speed and fast-speed scenarios. The only difference is that in fast-speed scenarios, the large-scale statistics change faster. For this reason, in such cases, the optimization should take place more frequently. Also, the simple expression of $ \pdv{ \mathcal{R}}{\phi_{n}^{*}} $ results in a significant decrease of the computational complexity. These two reasons make the proposed method quite beneficial.
\end{remark}
\begin{remark}
The proposed methodology for the RB design, i.e., the optimization of the phase matrix after having derived the performance expression in terms of large-scale statistics, could also be used in the case of negligence of HWIs, where another property of an IRS-assisted system would be the main topic of study. In such a case, the optimization in terms of the derivative will be simplified even more since the additional terms, concerning the T-HWIs which depend on the overall channel (the optimization variables $ \phi_{n} $), will be omitted.
\end{remark}
%
%
%
%
%
%
%
%
%
%
\section{Numerical Results}\label{Numerical}
In this section, we depict and discuss the analytical results corresponding to the uplink performance in terms of CE and achievable sum SE of an IRS-aided MU-MISO system with imperfect CSI and HWIs. Monte-Carlo (MC) simulations ($ 10^{3} $ independent channel realizations) represented by "\ding{53}" marks in Figs. \ref{Fig0} and \ref{Fig2} below, corroborate our analysis and the tightness of the UatF bound. For the sake of comparison, we have modified \cite{Nadeem2020} to describe the uplink transmission.
\subsection{Simulation Setup}
We consider a uniform linear array (ULA) of $ M $ antennas ($ M=16 $) at the BS, assisted by an IRS with a uniform planar array (UPA) of $ N $ elements ($ N=60 $) that serve $ K=5 $ UEs. The spatial correlation coefficient for the IRS between elements $ n $ and $ n' $ corresponding to UE $ k $ is given by \cite{Kammoun2020}.
\!\! \begin{align}
[{\mathbf{R}}_{\mathrm{IRS},k}]_{n,n'}\!=\!\mathbb{E}\big[\!\exp\!\big(j \frac{2 \pi}{\lambda}d_{\mathrm{IRS}}\!\left(n-n'\right)\sin \phi_{k}\sin \theta_{k}\big)\!\big],\label{RIS}
\end{align}
where $ d_{\mathrm{IRS}}=0.5\lambda $ while $ \phi_{k} $ and $ \theta_{k} $ express the elevation and azimuth angles for UE $ k $, and are generated by the Laplace and the Von Mises distribution, respectively. In the former case, we assume that the mean angle of departure and the spread are $ 90^{\circ} $ and $ 8^{\circ} $, respectively. The latter distribution is generated with mean angle of departure $ 0 $ and spread $ 0.2 $. Regarding the parameters for the channel matrix $ {\mathbf{H}}_{1} $ between the BS and the IRS, we assume $ d_{\mathrm{BS}} = 0.5\lambda $ and $ \theta_{1,n} $, $ \psi_{1,n} $ are uniformly distributed between $ 0 $ to $ \pi $ and $ 0 $ to $ 2\pi $, respectively. Also, $ \theta_{2,n}= \pi- \theta_{1,n} $, $ \psi_{2,n}=\pi+ \psi_{1,n}$. The correlation matrix $ {\mathbf{R}}_{\mathrm{BS},k} $ is obtained similar to \cite{Hoydis2013}. Moreover, the overall path loss for the IRS-assisted link is given by \cite{Wu2019a,Bjoernson2019b,Kammoun2020}
\begin{align}
\bar{\beta}_{k}=C_{1}C_{2} d_{\mathrm{BS}-\mathrm{IRS}}^{-\alpha_{1}}d_{\mathrm{IRS}-\mathrm{UE}_{k}}^{-\alpha_{2}},
\end{align}
where $ \beta_{1}=\frac{C_{1}}{d_{\mathrm{BS}-\mathrm{IRS}}^{\alpha_{1}}} $ and $ \beta_{2,k}=\frac{C_{2}}{d_{\mathrm{IRS}-\mathrm{UE}_{k}}^{\alpha_{2}}} $ are the channel attenuation coefficients between the BS and the IRS and between the IRS and UE $ k $, respectively. Note that $ \alpha_{1} $ and $ d_{\mathrm{BS}-\mathrm{IRS}} $ are the path-loss exponent and distance concerning the link BS-to-IRS, while $ \alpha_{2} $ and $ d_{\mathrm{IRS}-\mathrm{UE}_{k}} $ are the path-loss exponent and distance concerning the link IRS-to-UE $ k $.
The parameters values are chosen relied on the 3GPP Urban Micro (UMi) scenario from TR36.814 for a carrier frequency of $ 2.5 $ GHz and noise level $ -80 $ dBm, where the path losses for $ {\mathbf{h}}_{2,k} $ and $ {\mathbf{H}}_{1} $ are generated based on the NLOS and LOS versions \cite{3GPP2017}. Hence, we have $ \alpha_{1} =2.2$ and $ \alpha_{2} =3.67$ while $ d_{\mathrm{BS}-\mathrm{IRS}} =8~\mathrm{m}$ and $ d_{\mathrm{IRS}-\mathrm{UE}_{k}} =60~\mathrm{m}$. Also, $ C_{1}=26 $ dB, $ C_{2}=28 $ dB by assuming $ 5 $ dBi antennas at the BS and IRS while each UE includes a single
0dBi antenna \cite{Bjoernson2019b}. For $ \beta_{\mathrm{d},k} $, we assume the same parameters as for $ \beta_{2,k} $, but we also consider an additional penetration loss equal to $ 15~\mathrm{dB} $. In addition, we assume that the coherence bandwidth is $B_{\mathrm{c}}= 200~\mathrm{KHz} $ and the coherence time is $ T_{\mathrm{c}}=1~\mathrm{ms} $, i.e., each coherence block consists of $\tau_{\mathrm{c}}=200$ samples. Also, we assume $ P=6~\mathrm{dB}$ and the same value for $ p_{i}=\rho,~\forall i $ during the uplink data transmission. Note that $ \sigma^2=-174+10\log_{10}B_{\mathrm{c}} $.
For the study of the T-HWIs, we assume that we have an Analog-to-Digital Converter (ADC) at the BS, which quantizes the received signal to a $b$ bit resolution. As a consequence, the receive distortion can be written as $\kappa_{\mathrm{BS}}=2^{-2b}/\left(1-2^{-2b}\right)$, which gives $ \kappa_{\mathrm{BS}}= 0.258^{2},~0.126^{2}$, and $0.062^{2}$ for $ b=2, 3 $, and $ 4 $ bits, respectively ~\cite{Bjornson2015,Papazafeiropoulos2017}. In other words, a smaller resolution results in more severe distortion. The same value is used for $ \kappa_{\mathrm{UE}} $. Note that the trend in 5G networks and beyond is the use of lower precision ADCs. Regarding, the IRS-HWIs, if the Von Mises PDF is assumed to model the phase noise, the concentration parameter is set to $ \kappa_{\tilde{\theta}}=2 $. Unless otherwise stated, this set of parameters is used during the simulations.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.95\linewidth]{Graph0.pdf}
\caption{\footnotesize{NMSE of UE $ k $ versus the SNR of an IRS-assisted MIMO system with imperfect CSI ($ M=16 $, $ N=60 $, $ K=5 $) for varying T-HWIs $ \kappa_{\mathrm{BS}} $, $\kappa_{\mathrm{UE}}$ in the cases of uniform PDF for the phase noise or uncorrelated fading at the IRS (Analytical results and MC simulations). }}
\label{Fig0}
\end{center}
\end{figure}
Fig. \ref{Fig0} illustrates the relative estimation error per channel element, i.e., the normalized mean square error (NMSE) with respect to the uplink SNR for different values of the T-HWIs defining certain noise floors (asymptotic limits as $ p \to \infty $). In addition, we show the result corresponding to perfect hardware. Obviously, this line decreases without bound. Moreover, it is shown that the error floors go higher with increasing the severity of T-HWIs. Even at mild values of T-HWIs, we observe that the NMSE approaches the corresponding floor after $ 20~\mathrm{dB}$, which means that IRS-assisted systems require a high SNR to operate since they are dependent on a conventional MU-MISO architecture. For the sake of exposition, we have assumed uniform phase noise or uncorrelated fading at the IRS, in order to avoid any RB optimization since the NMSE does not depend on $ {\boldsymbol{\Theta}} $ in these cases (Remark~\ref{rem1}). Optimization with respect to $ {\boldsymbol{\Theta}} $ has been performed only in the case of the sum SE. Based on Remark \ref{rem3}, comparisons with other methods cannot be made at this stage. However, a comparison with respect to \cite{Nadeem2020} takes place below (Fig. \ref{Fig1}) in the case of achievable sum SE. Notably, MC simulations verify the analytical results.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.95\linewidth]{Graph1.pdf}
\caption{\footnotesize{Uplink achievable sum SE versus the number of IRS elements $N$ of an IRS-assisted MIMO system with imperfect CSI ($ M=16 $, $ K=5 $) for varying T-HWIs $ \kappa_{\mathrm{BS}} $, $\kappa_{\mathrm{UE}}$ and transmit power $ p $. }}
\label{Fig1}
\end{center}
\end{figure}
In Fig. \ref{Fig1}, we depict the achievable sum SE versus the number of IRS elements $ N $ for two different SNR values, $ \rho=0~\mathrm{dB}$ and $ \rho=20~\mathrm{dB}$. Also, we have considered different values of T-HWIs. First, we observe an increase of $ \mathcal{R} $ with $ N $, which increases unboundedly in the case of ideal hardware, but saturates for imperfect hardware met in practice. As expected, the degradation is higher when T-HWIs are more severe, probably, in the case of cheaper hardware used for a cost-efficient implementation. Furthermore, the convergence speed is faster at the higher SNR group because the T-HWIs are power-dependent. Thus, the largest part of the gain is achieved at lower values of $ N $. However, an increase of the IRS elements still allows for a further increase of $ \mathcal{R} $. For the sake of comparison, we have considered the CE from \cite{Nadeem2020} in terms of simulation ("dashed-star" lines) when $ \kappa_{\mathrm{BS}}=\kappa_{\mathrm{UE}}=0.258^{2} $. The achievable sum SE is much lower than the proposed method because of the high training overhead.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.95\linewidth]{Graph2.pdf}
\caption{\footnotesize{Uplink achievable sum SE versus the number of BS antennas $M$ of an IRS-assisted MIMO system with imperfect CSI ($ N=60 $, $ K=5 $) for varying T-HWIs $ \kappa_{\mathrm{BS}} $, $\kappa_{\mathrm{UE}}$ and transmit power $ p $ (Analytical results and MC simulations). }}
\label{Fig2}
\end{center}
\end{figure}
Fig. \ref{Fig2} illustrates the achievable sum SE with respect to the number of BS antennas $ M $ for varying T-HWIs and SNR. Notably, this figure resembles with the previous figure, i.e., $ \mathcal{R} $ presents a similar dependence on $ M $ and $ N $. Hence, when $ M $ grows large, the sum SE increases without limit, if perfect hardware is assumed while it appears ceilings in practice, where T-HWIs exist. In fact, lower hardware quality results in larger degradation. In addition, by increasing the SNR, the sum SE becomes larger. Also, the sum SE saturates faster in the instance of a larger SNR ($ 20~\mathrm{dB} $). Thus, these two figures indicate that an IRS-assisted system performs better at higher SNR values, and with larger values of IRS elements and BS antennas. Note that the latter is further appealing since it agrees with the massive MIMO technology of which the implementation has already started. Moreover, the analytical results are accompanied by MC simulations showing the tightness and correctness of the lower bounds since they coincide.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.95\linewidth]{Graph3.pdf}
\caption{\footnotesize{Uplink achievable sum SE versus the number of IRS elements $N$ of an IRS-assisted MIMO system ($ M=16 $, $ K=5 $) for varying BS distortion $ \kappa_{\mathrm{BS}} $ and both cases of perfect and imperfect CSI. }}
\label{Fig3}
\end{center}
\end{figure}
Fig. \ref{Fig3} shows the achievable sum SE with respect to the number of IRS elements by varying the impact of the BS distortion $ \kappa_{\mathrm{BS}} $ while the distortion at the UE side is assumed zero. Notably, the lines converge to the same finite limit as $ N $ increases, which means that the impact of $ \kappa_{\mathrm{BS}}\ $ becomes negligible when $ N $ increases. In other words, despite the unavoidable existence of imperfect hardware, the IRS is suggested as $ N $ increases. Actually, it allows the use of low-cost hardware. Moreover, in the same figure, we have provided a comparison between perfect and imperfect CSI. We notice that, in both cases, the sum SE increases with the number of IRS elements while, again, the impact of $ \kappa_{\mathrm{BS}} $ becomes negligible at large $ N $. Obviously, in the realistic case of imperfect CSI, the achievable sum SE is lower, and the gap between the lines of ideal and imperfect hardware increases with larger $ N $ because the impact from the estimation error becomes larger.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.95\linewidth]{Graph9.pdf}
\caption{\footnotesize{ Uplink achievable sum SE versus the number of UEs $K$ of an IRS-assisted MIMO system ($ M=16 $, $ \rho=20~\mathrm{dB}$) for varying T-HWIs $ \kappa_{\mathrm{BS}} $, $\kappa_{\mathrm{UE}}$ and IRS elements $ N $.}}
\label{Fig9}
\end{center}
\end{figure}
Fig. \ref{Fig9} examines the achievable sum SE with respect to the number of UEs by varying the T-HWIs for different numbers of IRS elements. The sum SE increases with $ K $ almost linearly at the beginning but the gradient decreases as $ K $ increases. This result is reasonable since the increase of $ K $ increases the multi-user interference and the received distortion as described by \eqref{eta_rU}. Also, a larger IRS in terms of $ N $ results in a larger sum SE as has been already shown.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.95\linewidth]{Graph4.pdf}
\caption{\footnotesize{Uplink achievable sum SE versus the number of IRS elements $N$ of an IRS-assisted MIMO system with imperfect CSI ($ M=16 $, $ K=5 $) for varying IRS-HWIs and T-HWIs in the cases of correlated/non-correlated Rayleigh fading. }}
\label{Fig4}
\end{center}
\end{figure}
In Fig. \ref{Fig4}, we show the impact of correlated Rayleigh fading on the achievable sum SE in the cases of correlation at both the BS and the IRS while varying the number of IRS elements. Also, we vary the quality of T-HWIs, and we show how for the same T-HWIs at the BS and UE sides, the sum SE decreases in the case of correlated fading ("dot" lines) with comparison to no correlation ("solid" lines). Furthermore, this figure allows shedding light on the impact of phase noise at the IRS. In particular, when uniform phase noise is assumed ("star" symbols), $ \mathcal{R} $ is the lowest because the design cannot take benefit (random fluctuations) from the IRS optimization since $ {\mathbf{R}}_k $ does not depend on the phase matrix (see Rem. \ref{rem5}). However, in the case that the phase noise is distributed according to the Von Mises distribution ("dashed" lines), the achievable sum SE increases because the presence of the IRS becomes advantageous since it can adjust the phase shifts of its passive elements towards better performance. Especially, we have considered variation of the concentration parameter $ \kappa_{\tilde{\theta}} $. As $ \kappa_{\tilde{\theta}} $ decreases, the achievable sum SE decreases. Actually, we show that when $ \kappa_{\tilde{\theta}} =0$, the corresponding line coincides with the line describing the uniform PDF since the Von Mises PDF coincides with the uniform distribution in this case. Note that we have also depicted the performance in the absence of the IRS. From the figure, we can verify our observation in Rem. \ref{rem1} explaining that the IRS contributes to the performance even in the worst-case IRS phase noise scenario since the line describing the case with no IRS is lower, i.e., the performance is worse.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.95\linewidth]{Graph8.pdf}
\caption{\footnotesize{ Uplink achievable sum SE versus the T-HWIs quality $ \sqrt[]{\bar{\kappa}} $ ($ \rho=20~\mathrm{dB}$, $ K=5 $) for varying BS antennas $ M $ and IRS elements $ N $. }}
\label{Fig8}
\end{center}
\end{figure}
Fig.~\ref{Fig8} shows the performance of the achievable sum SE versus $ \sqrt[]{\bar{\kappa}} $, where $ \kappa_{\mathrm{BS}}=\bar{\kappa} $ and $ \kappa_{\mathrm{UE}}=\bar{\kappa} +0.03 $. We have assumed that the distortion at the UE is larger due to its simpler hardware. The impact of IRS phase noise is not considered as we focus on the impact of the T-HWIs. The "solid" and "dashed" lines correspond to the variation of $ M $ BS antennas and $ N $ IRS elements, respectively. The horizontal axis starts from the case of no T-HWIs at the BS, i.e., when $ \bar{\kappa}=0 $ and ends with severe additive HWIs. We observe that as T-HWIs increase, the performance decreases. Moreover, we observe that at severe T-HWIs the variations of the $ M $ and $ N $ do not affect the performance. Also, we notice that the number of IRS elements $ N $ affects more the performance (higher sum SE) than the number of BS antennas $ M $ while the same variation with respect to $ N $ has a larger impact since the gaps between the solid lines are larger.
\begin{figure}%
\centering
\subfigure[]{ \includegraphics[width=0.95\linewidth]{Graph7.pdf}}\qquad
\subfigure[]{ \includegraphics[width=0.95\linewidth]{Graph10.pdf}}\\
\caption{ Uplink achievable sum SE of an IRS-assisted MIMO system with imperfect CSI versus: (a) the number of iterations $N$ ( $ K=5 $, $ p=0\mathrm{dB }$, $ \kappa_{\mathrm{BS}}=\kappa_{\mathrm{UE}}=0.126^2 $) for varying BS antennas $ M $ and IRS elements $ N $; (b) $ 30 $ channel realizations ($ M=16 $, $ N=20 $, $ K=5 $, $ p=20\mathrm{dB }$, $ \kappa_{\mathrm{BS}}=\kappa_{\mathrm{UE}}=0.126^{2} $).}
\label{fig9}
\end{figure}
In Fig. \ref{fig9}.(a), we show the convergence of the proposed algorithm, i.e., Algorithm \ref{Algoa1}. In particular, we have depicted the uplink achievable sum SE versus the number of iterations for various sets of BS antennas and IRS elements. Notably, the algorithm converges fast in all cases. For example, when $ M=20 $ and $ N=60 $, the algorithm converges in $ 7 $ iterations. Moreover, we notice that by increasing the IRS and BS sizes in terms of their elements and antennas, respectively, more iterations are required for convergence because the amount of optimization variables increases and the relevant search space is enlarged. On top of this, an increase in terms of BS antennas or IRS elements results in higher complexity of each iteration of the proposed algorithm as mentioned in Sec. \ref{IRSdesign}.
The non-convexity of the optimization problem suggests that its solution depends on the initial point, i.e., different initial points result in different locally optimal solutions. Fig. \ref{fig9}.(b) investigates this dependence on the initializations by accounting for $ 30 $ channel realizations. The initialization of Alg. 1 assumes that $ {\mathbf{s}}^{0} =\exp\left(j\pi/2\right)\mat{\mathrm{1}}_{N}$ as mentioned in its description. "Alg. 1-Test" in the figure assumes the best initial point out of $ 100 $ random initial points for each channel instance. The figure shows that different initializations result in different solutions and that the sum SE in both cases is almost the same, which means that this phase shifts selection for initialization is a good choice.
\section{Conclusion} \label{Conclusion}
In this paper, not only we studied the impact of both T-HWIs and IRS-HWIs on a general IRS-assisted MU-MISO system with imperfect CSI and correlated Rayleigh fading, but we also proposed a novel optimization methodology regarding the optimization of the RBM with low computational cost, being quite useful in IRS-assisted systems that have a large number of elements. In particular, we obtained the LMMSE estimate of the channel with T-HWIs and IRS-HWIs. Moreover, we derived the uplink achievable sum SE with MRC in closed form, being dependent only on large-scale statistics, and performed high computationally efficient optimization with respect to the IRS RBM. In general, we provided a methodology resulting in analytical and tractable expressions being advantageous over previous works as shown by the simulation results. Furthermore, we evaluated the impact of HWIs at both the transceiver and the IRS on the system SE, and shed insightful light on their interplay with other system parameters towards efficient IRS design. Remarkably, this work opens new research directions for IRS-assisted systems such as the studies of energy efficiency and power scaling laws with imperfect CSI and T-HWIs.
\begin{appendices}
\section{Proof of Proposition~\ref{PropositionDirectChannel}}\label{Proposition1}
According to \cite[Ch. 12]{Kay} The LMMSE estimator of $ {\mathbf{h}}_{k} $ is obtained by $ \hat{{\mathbf{h}}}_{k} ={\mathbf{F}}{\mathbf{r}}_{k}$,
where ${\mathbf{F}} $ is derived my minimizing $ \mathop{\mathrm{tr}}\nolimits\big(\mathbb{E}\big[\left(\hat{{\mathbf{h}}}_{k}-{{\mathbf{h}}}_{k}\right)\left(\hat{{\mathbf{h}}}_{k}-{{\mathbf{h}}}_{k}\right)^{\H}\big]\big) $ as
\begin{align}
{\mathbf{F}}=\mathbb{E}\left[{\mathbf{r}}_{k}{\mathbf{h}}_{k}^{\H}\right]\left(\mathbb{E}\left[{\mathbf{r}}_{k}{\mathbf{r}}_{k}^{\H}\right]\right)^{-1}.\label{Cor6}
\end{align}
Given that the additive distortions and the receiver noise are uncorrelated with overall channel $ {\mathbf{h}}_{k} $, the first term of \eqref{Cor6} becomes
\begin{align}
\mathbb{E}\left[{\mathbf{r}}_{k}{\mathbf{h}}_{k}^{\H}\right]&=\mathbb{E}\big[\big({\mathbf{h}}_{k}+ \sum_{i=1}^{K}\frac{ \tilde{\delta}_{\mathrm{t},i}}{\tau P}{\mathbf{h}}_{i}+ \frac{\tilde{\hbox{\boldmath$\delta$}}_{\mathrm{r}} +{\mathbf{w}}_{k}}{ \tau P}\big){\mathbf{h}}_{k}^{\H}\big]\\
&=\mathbb{E}\left[{\mathbf{h}}_{k}{\mathbf{h}}_{k}^{\H}\right]={\mathbf{R}}_{k}.\label{Cor0}
\end{align}
Regarding the second term, we have
\begin{align}
\mathbb{E}\left[{\mathbf{r}}_{k}{\mathbf{r}}_{k}^{\H}\right]&={\mathbf{R}}_{k}+ \frac{\kappa_{\mathrm{UE}}}{\tau }\sum_{i=1}^{K} {\mathbf{R}}_{i}+ \frac{\kappa_{\mathrm{BS}}}{\tau}\!\sum_{i=1}^{K}\!\mat{\mathrm{I}}_{M}\!\circ\! {\mathbf{R}}_{i} +\frac{\sigma^2}{ \tau P }\mat{\mathrm{I}}_{M},\label{Cor1}
\end{align}
where we have taken into account that the additive distortions and the receiver noise are uncorrelated with each other. Also, we have used that the variance of $ \tilde{\hbox{\boldmath$\delta$}}_{\mathrm{r}} $ is $ \tau P { \bm \Upsilon}$ with ${ \bm \Upsilon}=\kappa_{\mathrm{BS}}P\sum_{i=1}^{K} \mathop{\mathrm{tr}}\nolimits\left({\mathbf{R}}_{i}\right)$.
As a result, the LMMSE estimate is given by inserting \eqref{Cor0} and \eqref{Cor1} into \eqref{Cor6} as
\begin{align}
\hat{{\mathbf{h}}}_{k}={\mathbf{R}}_{k}{\mathbf{Q}}_{k}{\mathbf{r}}_{k}.
\end{align}
Furthermore, the covariance matrix of the estimated channel is obtained as
\begin{align}
\mathbb{E}\left[\hat{{\mathbf{h}}}_{k} \hat{{\mathbf{h}}}_{k}^{\H}\right]={\mathbf{R}}_{k}{\mathbf{Q}}_{k}{\mathbf{R}}_{k}.
\end{align}
\section{Proof of Theorem~\ref{theorem:ULDEMMSE}}\label{theorem1}
For the derivation of $\gamma_{k}$ for finite $M$, we recall each term of~\eqref{sig11} and \eqref{int1}. Generally, we are going to apply a useful property suggesting that ${\mathbf{x}}^{\H}{\mathbf{y}} = \mathop{\mathrm{tr}}\nolimits({\mathbf{y}} {\mathbf{x}}^{\H})$ for any vectors ${\mathbf{x}}$, ${\mathbf{y}}$. First, we obtain the $ S_{k} $ given by~\eqref{sig11}. Specifically, the desired signal part $ \mathrm{DS}_{k} $ (without $ \rho_{k} $) is written as
\begin{align}
\mathbb{E}\big[
\hat{{\mathbf{h}}}_{k}^{\H}{{\mathbf{h}}}_{k}\big]&=\mathop{\mathrm{tr}}\nolimits\big( \mathbb{E}\left[{{\mathbf{h}}}_{k} \hat{{\mathbf{h}}}_{k}^{\H} \right] \big) \\
&=\mathop{\mathrm{tr}}\nolimits\left( \mathbb{E}\left[{{\mathbf{h}}}_{k} {\mathbf{r}}_{k}^{\H}{\mathbf{Q}}_{k} {\mathbf{R}}_{k}\right] \right)\label{term1}\\
&=\mathop{\mathrm{tr}}\nolimits\left({\boldsymbol{\Psi}}_{k}\right)\label{term2},
\end{align}
where, in \eqref{term1}, we have used \eqref{estim1}, while the last step is obtained by applying the expectation between $ {\mathbf{h}}_{k} $ and $ {\mathbf{r}}_{k} $ and by considering that the mean value of $ \tilde{\delta}_{\mathrm{t},k} $ is zero.
Regarding the second-order moment in the denominator, expressing the MU interference part $ \mathrm{MUI}_{ik} $ for $i\ne k$, it is written as
\begin{align}
&\mathbb{E}\big[ \big| \hat{{\mathbf{h}}}_{k}^{\H}{{\mathbf{h}}}_{i}\big|^{2}\big]=\mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k} {\mathbf{R}}_{i}\right),\label{54}
\end{align}
which relies on the independence between the two random vectors.
For the power of the beamforming uncertainty, we have
\begin{align}
\mathbb{E}\left\{|\mathrm{BU}_{k}|^{2}\right\} &=\mathbb{E}\big[ \big| \hat{{\mathbf{h}}}_{k}^{\H}{{\mathbf{h}}}_{k}-\mathbb{E}\big[
\hat{{\mathbf{h}}}_{k}^{\H}{{\mathbf{h}}}_{k}\big]\big|^{2}\big]\label{est0}\\
&=\mathbb{E}\big[ \big| \hat{{\mathbf{h}}}_{k}^{\H}{{\mathbf{h}}}_{k}\big|^{2}\big]-\big|\mathbb{E}\big[
\hat{{\mathbf{h}}}_{k}^{\H}{{\mathbf{h}}}_{k}\big]\big|^{2} \label{est2}\\
&=\mathbb{E}\big[ \big| \hat{{\mathbf{h}}}_{k}^{\H} \hat{{\mathbf{h}}}_{k} +\hat{{\mathbf{h}}}_{k}^{\H}\tilde{{\mathbf{h}}}_{k}\big|^{2}\big]-\big|\mathbb{E}\big[
\hat{{\mathbf{h}}}_{k}^{\H}\hat{{\mathbf{h}}}_{k}\big]\big|^{2}\label{est3} \\
&=\mathbb{E}\big[|\hat{{\mathbf{h}}}_{k}^{\H}\tilde{{\mathbf{h}}}_{k}|^{2}\big] \label{est5}\\
&=\mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k} {\mathbf{R}}_{k}\right)-\mathop{\mathrm{tr}}\nolimits\left( {\boldsymbol{\Psi}}_{k}^{2}\right),\label{est4}
\end{align}
where in~\eqref{est3}, we have substituted \eqref{current}, and in~\eqref{est3}, we have applied the property $\mathbb{E}\left[ |X+Y|^{2}\right] =\mathbb{E}\left[ |X|^{2}\right] +\mathbb{E}\left[ |Y^{2}|\right]$, which holds between two independent random variables when one of them has zero mean value, e.g., $\mathbb{E}\left[ X\right]=0 $. Equation \eqref{est4} follows from the facts that $ \mathbb{E}\big[|\hat{{\mathbf{h}}}_{k}^{\H}\tilde{{\mathbf{h}}}_{k}|^{2}\big]={\boldsymbol{\Psi}}_{k}\left({\mathbf{R}}_{k}-{\boldsymbol{\Psi}}_{k}\right) $ by taking advantage of the independence between the two random vectors.
The derivation of the term $\sigma_{\mathrm{UE}}^{2}$, corresponding to the additive transmit distortion from all UEs, is straightforward. Specifically, we have
\begin{align}
\sigma_{\mathrm{UE}}^{2}&= \sum_{i=1}^{K}\rho_{i}\kappa_{\mathrm{UE}}\mathbb{E}\big\{|{\mathbf{v}}_{k}^{\H}{\mathbf{h}}_{i}|^{2} \big\}\label{simgaue1}\\
&=\kappa_{\mathrm{UE}}\left( \rho_{k}\mathbb{E}\big\{|{\mathbf{v}}_{k}^{\H}{\mathbf{h}}_{k}|^{2} \right\} +\sum_{i\ne k}^{K}\rho_{i}\mathbb{E}\left\{|{\mathbf{v}}_{k}^{\H}{\mathbf{h}}_{i}|^{2} \right\}\!\big).\label{simgaue}
\end{align}
In \eqref{simgaue1}, we have taken the expectation with respect to the transmit distortion for fixed channel realizations. The first part in \eqref{simgaue} is obtained as
\begin{align}
\mathbb{E}\left\{|{\mathbf{v}}_{k}^{\H}{\mathbf{h}}_{k}|^{2} \right\}&=\mathbb{E}\big[ \big| \hat{{\mathbf{h}}}_{k}^{\H} \hat{{\mathbf{h}}}_{k} +\hat{{\mathbf{h}}}_{k}^{\H}\tilde{{\mathbf{h}}}_{k}\big|^{2}\big]\\
&=\mathbb{E}\big[|\hat{{\mathbf{h}}}_{k}^{\H}\hat{{\mathbf{h}}}_{k}|^{2}\big]+\mathbb{E}\big[|\hat{{\mathbf{h}}}_{k}^{\H}\tilde{{\mathbf{h}}}_{k}|^{2}\big]\\
&=|\mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k}\right)|^{2}+\mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k} {\mathbf{R}}_{k}\right)-\mathop{\mathrm{tr}}\nolimits\left( {\boldsymbol{\Psi}}_{k}^{2}\right),\label{esto1}
\end{align}
where we have used similar steps as before. Hence, inserting \eqref{esto1} into \eqref{simgaue}, and noticing that the second part of \eqref{simgaue} is identical to $ \mathrm{MUI}_{ik} $, $ \sigma_{\mathrm{UE}}^{2} $ becomes
\begin{align}
\!\!\sigma_{\mathrm{UE}}^{2}\!=\!\kappa_{\mathrm{UE}}\big(\rho_{k}|\mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k}\right)|^{2}+\sum_{i=1}^{K}\!\rho_{i} \mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k} {\mathbf{R}}_{i}\right)-\rho_{k}\mathop{\mathrm{tr}}\nolimits\left( {\boldsymbol{\Psi}}_{k}^{2}\right)\!\big)\!.\label{sue1}
\end{align}
The term $\sigma_{\mathrm{BS}}^{2}$, concerning the additive receive distortion at the BS, is obtained as
\begin{align}
\sigma_{\mathrm{BS}}^{2}
&={\boldsymbol{\kappa}}_{\mathrm{BS}}\mathbb{E}\big\{\hat{{\mathbf{h}}}_{k}^{\H}\big(\sum_{i=1}^{K}\rho_{i} \mat{\mathrm{I}}_{M}\circ {\mathbf{h}}_{i}{\mathbf{h}}_{i}^{\H}\big)\hat{{\mathbf{h}}}_{k}\big\}\label{bs1} \\
&=\underbrace{{\boldsymbol{\kappa}}_{\mathrm{BS}}\mathbb{E}\big\{\hat{{\mathbf{h}}}_{k}^{\H}\big(\rho_{k} \mat{\mathrm{I}}_{M}\circ {\mathbf{h}}_{k}{\mathbf{h}}_{k}^{\H}\big)\hat{{\mathbf{h}}}_{k}\big\}}_{\mathcal{I}_1}\nonumber\\&+\underbrace{{\boldsymbol{\kappa}}_{\mathrm{BS}}\mathbb{E}\big\{\hat{{\mathbf{h}}}_{k}^{\H}\big(\sum_{i\ne k}^{K}\rho_{i} \mat{\mathrm{I}}_{M}\circ {\mathbf{h}}_{i}{\mathbf{h}}_{i}^{\H}\big)\hat{{\mathbf{h}}}_{k}\big\}}_{\mathcal{I}_2}\label{bs2},
\end{align}
where,
in \eqref{bs1}, we have taken the expectation with respect to the receive distortion for fixed channel realizations, we have accounted for MRC, i.e, $ {\mathbf{v}}_{k}={\mathbf{h}}_{k} $, and we have used the Hadamard product to write the diagonal matrix. In the next equation, we have simply split the sum and denoted the two parts as $ \mathcal{I}_1 $ and $ \mathcal{I}_2 $. In the case of the former part, we have
\begin{align}
\mathcal{I}_1&={\boldsymbol{\kappa}}_{\mathrm{BS}}\mathbb{E}\!\big\{ \!\mathop{\mathrm{tr}}\nolimits\! \big(\!\big(\rho_{k} \mat{\mathrm{I}}_{M}\circ \big(\hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H}+\tilde{{\mathbf{h}}}_{k}\tilde{{\mathbf{h}}}_{k}^{\H}\big)\!\big)\!\hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H}\big)\!\!\big\}\label{key}\\
&=\underbrace{{\boldsymbol{\kappa}}_{\mathrm{BS}} \mathbb{E}\!\big\{ \!\mathop{\mathrm{tr}}\nolimits\! \big(\!\big(\rho_{k} \mat{\mathrm{I}}_{M}\circ \hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H}\!\big)\!\hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H}\big)\!\!\big\}}_{\mathcal{I}_{11}}\nonumber\\&+\underbrace{{\boldsymbol{\kappa}}_{\mathrm{BS}} \mathbb{E}\!\big\{ \!\mathop{\mathrm{tr}}\nolimits\! \big(\!\big(\rho_{k} \mat{\mathrm{I}}_{M}\circ \tilde{{\mathbf{h}}}_{k}\tilde{{\mathbf{h}}}_{k}^{\H}\!\big)\!\hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H}\big)\!\!\big\}}_{\mathcal{I}_{12}},\label{I1}
\end{align}
where, at first, we have used \eqref{current} and exploited that $ \hat{{\mathbf{h}}}_{k} $ and $ \tilde{{\mathbf{h}}}_{k} $ are uncorrelated. Next, we have split \eqref{I1} into $ \mathcal{I}_{11} $ and $ \mathcal{I}_{12} $. For $ \mathcal{I}_{11} $, we exploit that the diagonal matrix can be written as $ \mat{\mathrm{I}}_{M}\circ \hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H}= \sum_{m=1}^{M}|{\mathbf{e}}_{m}^{\H} \hat{{\mathbf{h}}}_{k} |^{2}{\mathbf{e}}_{m}{\mathbf{e}}_{m}^{\H}$, where $ {\mathbf{e}}_{m} $ is the $ m $th column of $ \mat{\mathrm{I}}_{M} $ \cite{Bjornson2015}. Hence, we have
\begin{align}
& \mathcal{I}_{11}={\boldsymbol{\kappa}}_{\mathrm{BS}} \rho_{k}\sum_{m=1}^{M} \mathbb{E}\!\big\{ \!\mathop{\mathrm{tr}}\nolimits\! \big({\mathbf{e}}_{m}{\mathbf{e}}_{m}^{\H} \hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H} {\mathbf{e}}_{m}{\mathbf{e}}_{m}^{\H}\hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H}\big)\!\!\big\}\label{I111}\\
&={\boldsymbol{\kappa}}_{\mathrm{BS}}\rho_{k}\sum_{m=1}^{M} \mathbb{E}\!\big\{\! | \hat{{\mathbf{h}}}_{k}^{\H} {\mathbf{e}}_{m}{\mathbf{e}}_{m}^{\H} \hat{{\mathbf{h}}}_{k}|^{2}\!\big\}\label{I112}\\
&={\boldsymbol{\kappa}}_{\mathrm{BS}}\rho_{k} \sum_{m=1}^{M} \big(| \mathop{\mathrm{tr}}\nolimits \big( {\mathbf{e}}_{m} {\mathbf{e}}_{m}^{\H } {\boldsymbol{\Psi}}_{k}\big)|^{2}\! +\!\mathop{\mathrm{tr}}\nolimits \big({\mathbf{e}}_{m} {\mathbf{e}}_{m}^{\H } {\boldsymbol{\Psi}}_{k} {\mathbf{e}}_{m} {\mathbf{e}}_{m}^{\H } {\boldsymbol{\Psi}}_{k}\big)\big)\label{I113}\\
&={\boldsymbol{\kappa}}_{\mathrm{BS}} \rho_{k}\big(| \mathop{\mathrm{tr}}\nolimits \big( \mat{\mathrm{I}}_{M}\circ {\boldsymbol{\Psi}}_{k}\big)|^{2}\! +\!\mathop{\mathrm{tr}}\nolimits \big(\big(\mat{\mathrm{I}}_{M}\circ {\boldsymbol{\Psi}}_{k}\big) {\boldsymbol{\Psi}}_{k}\big)\big)\label{I114},
\end{align}
where, in \eqref{I113}, we have used \cite[Lemma~2]{Bjornson2015}. In the next equation, we have reverted the matrix expansion. More easily, in the case of $ \mathcal{I}_{12} $, we have
\begin{align}
\mathcal{I}_{12}&={\boldsymbol{\kappa}}_{\mathrm{BS}} \rho_{k}\sum_{m=1}^{M} \mathbb{E}\!\big\{ \!\mathop{\mathrm{tr}}\nolimits\! \big({\mathbf{e}}_{m}{\mathbf{e}}_{m}^{\H} \tilde{{\mathbf{h}}}_{k}\tilde{{\mathbf{h}}}_{k}^{\H} {\mathbf{e}}_{m}{\mathbf{e}}_{m}^{\H}\hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H}\big)\!\!\big\}\label{I121}\\
&={\boldsymbol{\kappa}}_{\mathrm{BS}}\rho_{k} \mathop{\mathrm{tr}}\nolimits \big(\!\big( \mat{\mathrm{I}}_{M}\circ \big({\mathbf{R}}_{k}-{\boldsymbol{\Psi}}_{k}\big)\!\big){\boldsymbol{\Psi}}_{k}\big)
\label{I122},
\end{align}
where, in the first equation, we have used the diagonal matrix expansion, and in the second equation, we have reverted this expansion after taking advantage of the independence between $ \tilde{{\mathbf{h}}}_{k} $ and $ \hat{{\mathbf{h}}}_{k} $. Substitution of \eqref{I114} and \eqref{I122} into \eqref{I1} gives $\mathcal{I}_1 $ after simple algebraic manipulations as
\begin{align}
\mathcal{I}_1={\boldsymbol{\kappa}}_{\mathrm{BS}}\rho_{k} \big(| \mathop{\mathrm{tr}}\nolimits \big( \mat{\mathrm{I}}_{M}\circ {\boldsymbol{\Psi}}_{k}\big)|^{2}\! +\!\mathop{\mathrm{tr}}\nolimits \big(\big(\mat{\mathrm{I}}_{M}\circ {\mathbf{R}}_{k}\big) {\boldsymbol{\Psi}}_{k}\big)\big)\label{i1final}.
\end{align}
Regarding $ \mathcal{I}_2 $, it follows that
\begin{align}
\mathcal{I}_2&={\boldsymbol{\kappa}}_{\mathrm{BS}}\mathbb{E}\big\{\!\!\mathop{\mathrm{tr}}\nolimits\!\big(\!\!\big(\sum_{i\ne k}^{K}\rho_{i} \mat{\mathrm{I}}_{M}\circ {\mathbf{h}}_{i}{\mathbf{h}}_{i}^{\H}\!\big)\!\hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H}\big)\!\!\big\}\label{bs3}\\
&={\boldsymbol{\kappa}}_{\mathrm{BS}}\sum_{i\ne k}^{K}\rho_{i}\mathop{\mathrm{tr}}\nolimits\big(\mathbb{E}\big\{ \left(\mat{\mathrm{I}}_{M}\circ {\mathbf{h}}_{i}{\mathbf{h}}_{i}^{\H}\right)\!\hat{{\mathbf{h}}}_{k}\hat{{\mathbf{h}}}_{k}^{\H}\big\}\big)\label{bs4}\\
&={\boldsymbol{\kappa}}_{\mathrm{BS}}\sum_{i\ne k}^{K}\rho_{i}\mathop{\mathrm{tr}}\nolimits\left(\!\left(\mat{\mathrm{I}}_{M}\circ {\mathbf{R}}_{i} \right)\!{\boldsymbol{\Psi}}_{k}\right).\label{bs5}
\end{align}
In the first equality, $ \mathcal{I}_2 $ has been written in terms of the trace. Next, we have exchanged the orders among summation, trace, and expectations since they are linear operators. In \eqref{bs5}, given the independence between indices $ i $ and $ k $, we have computed the separate expectations.
Having obtained $ \mathcal{I}_1 $ and $ \mathcal{I}_2 $, we replace them in \eqref{bs2} to obtain $ \sigma_{\mathrm{BS}}^{2} $ as
\begin{align}
\sigma_{\mathrm{BS}}^{2}={\boldsymbol{\kappa}}_{\mathrm{BS}} \big(\rho_{k}| \mathop{\mathrm{tr}}\nolimits \left( \mat{\mathrm{I}}_{M}\circ {\boldsymbol{\Psi}}_{k}\right)|^{2}\! +\sum_{i=1}^{K}\!\rho_{i}\mathop{\mathrm{tr}}\nolimits \left(\left(\mat{\mathrm{I}}_{M}\circ {\mathbf{R}}_{i}\right) {\boldsymbol{\Psi}}_{k}\right)\big)\label{i1final1}.
\end{align}
In the case of $ \mathbb{E}\left\{|\mathrm{RN}_{k}|^{2}\right\} $, we easily obtain
\begin{align}
\mathbb{E}\left\{|\mathrm{RN}_{k}|^{2}\right\}
&=\sigma^2 \mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k}\right).\label{th2}
\end{align}
Use of \eqref{term2}, \eqref{est4}, \eqref{sue1}, \eqref{i1final1}, and \eqref{th2} concludes the proof by resulting in $ \bar{S}_{k} $ and $ \bar{I}_{k} $.
\section{Proof of Proposition~\ref{Prop:optimPhase}}\label{optimPhase}
We aim at finding the gradient of $ \mathcal{R} $ with respect to $ \phi_{n}, n=1, \ldots, N$. We use the facts that $ \pdv{ \mathcal{R}}{\phi_{n}^{*}} =\frac{\pdv{ \gamma_{k}}{\phi_{n}^{*}}}{\ln\left(2\right)\left(1+\frac{S_{k}}{I_{k}}\right)} $, which
requires the derivation of $ \pdv{\gamma_{k}}{\phi^{*}_{n}} $. A closer observation of Theorem \ref{theorem:ULDEMMSE}, providing $ \gamma_{k} $, reveals that it is a fraction consisting of terms including functions of traces. Hence, the standard quotient rule derivative gives
\begin{align}
\pdv{\gamma_{k}}{\phi^{*}_{n}}=\frac{\pdv{S_{k}}{\phi^{*}_{n}}I_{k}-S_{k}\pdv{I_{k}}{\phi^{*}_{n}}}{I_{k}^{2}},\label{gam1}
\end{align}
where the partial derivatives follow. Henceforth, for the sake of simplicity, we replace the notation for the partial derivative with respect to $ \phi^{*}_{n} $ by $ \left(\cdot\right)' $. Specifically, in the case of $ S_{k}' $, we obtain
\begin{align}
S_{k}'&=\rho_{k}\left(|\mathop{\mathrm{tr}}\nolimits\left({\boldsymbol{\Psi}}_{k}\right)|^{2}\right)'\\
&=2\rho_{k}\mathop{\mathrm{tr}}\nolimits\left({\boldsymbol{\Psi}}_{k}\right)\mathop{\mathrm{tr}}\nolimits\left({\boldsymbol{\Psi}}_{k}'\right),\label{sk1}
\end{align}
where \eqref{sk1} includes a simple derivative. Since all the terms in $ {\boldsymbol{\Psi}}_{k}$ depend on $ \phi^{*}_{n} $, we have
\begin{align}
&\!\!\!\! \mathop{\mathrm{tr}}\nolimits\left({\boldsymbol{\Psi}}_{k}'\right)=\mathop{\mathrm{tr}}\nolimits\left({\mathbf{R}}_{k}'{\mathbf{Q}}_{k}{\mathbf{R}}_{k}+{\mathbf{R}}_{k}{\mathbf{Q}}_{k}'{\mathbf{R}}_{k}+{\mathbf{R}}_{k}{\mathbf{Q}}_{k}{\mathbf{R}}_{k}'\right)\label{sk2}\\
&\!\!\!\!=\mathop{\mathrm{tr}}\nolimits\!\big({\mathbf{R}}_{k}'{\mathbf{Q}}_{k}{\mathbf{R}}_{k}-{\mathbf{R}}_{k}{\mathbf{Q}}_{k}\left({\mathbf{Q}}_{k}^{-1}\right)^{'}{\mathbf{Q}}_{k}{\mathbf{R}}_{k}+{\mathbf{R}}_{k}{\mathbf{Q}}_{k}{\mathbf{R}}_{k}'\big),\label{sk3}
\end{align}
where the derivative of ${\mathbf{Q}}_{k} $, being an inverse matrix, is obtained by \cite[Eq. 40]{Petersen2012} while $ \left({\mathbf{Q}}_{k}^{-1}\right)^{'} $ is obtained as
\begin{align}
\!\! \left({\mathbf{Q}}_{k}^{-1}\right)^{'}
&\!=\!{\mathbf{R}}_{k}'\!+\! \frac{\kappa_{\mathrm{UE}}}{\tau }\!\sum_{i=1}^{K} {\mathbf{R}}_{i}'\!+\!\frac{ \kappa_{\mathrm{BS}}\!\sum_{i=1}^{K}\!\left(\mat{\mathrm{I}}_{M}\!\circ \!{\mathbf{R}}_{i} \right)'}{ \tau }.\label{sk4}
\end{align}
Substitution of \eqref{sk4} into \eqref{sk3} gives
\begin{align}
& \mathop{\mathrm{tr}}\nolimits\left({\boldsymbol{\Psi}}_{k}'\right)=\mathop{\mathrm{tr}}\nolimits\left({\mathbf{R}}_{k}'{\mathbf{Q}}_{k}{\mathbf{R}}_{k}\right)-\mathop{\mathrm{tr}}\nolimits\big(\big({\mathbf{R}}_{k}{\mathbf{Q}}_{k}\big)^{2}{\mathbf{R}}_{k}'\big)\nonumber\\
&\!-\!\frac{\kappa_{\mathrm{UE}}}{\tau }\!\sum_{i=1}^{K}\mathop{\mathrm{tr}}\nolimits\!\big(\!\big({\mathbf{R}}_{k}{\mathbf{Q}}_{k}\big)^{2}{\mathbf{R}}_{i}'\big)\!-\!\frac{ \kappa_{\mathrm{BS}}}{ \tau }\!\sum_{i=1}^{K}\!\mathop{\mathrm{tr}}\nolimits\!\big(\!\big({\mathbf{R}}_{k}{\mathbf{Q}}_{k}\big)^{\!2}\big(\mat{\mathrm{I}}_{M}\!\circ\! {\mathbf{R}}_{i} \big)'\big)\nonumber\\
&+\mathop{\mathrm{tr}}\nolimits\left({\mathbf{R}}_{k}{\mathbf{Q}}_{k}{\mathbf{R}}_{k}'\right).\label{sk5}
\end{align}
To proceed further, for the sake of exposition, let the matrices $ {\mathbf{A}}, {\mathbf{B}}, {\mathbf{C}} \in\mathbb{C}^{M \times M} $,
we have denoted
\begin{align}
&{\mathbf{L}}\! \left({\mathbf{A}},{\mathbf{B}}, {\mathbf{C}}\right)=\alpha\beta_{2,i}\!\left[{\mathbf{H}}_{1}^{\H}{\mathbf{A}}{\mathbf{Q}}_{k} {\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},i}\right]_{n,n}\nonumber\\
&
-\big(1\!+\!\frac{\kappa_{\mathrm{UE}}}{\tau }\big)\alpha\beta_{2,k}\!\left[{\mathbf{H}}_{1}^{\H}{\mathbf{A}} {\mathbf{Q}}_{k}{\mathbf{C}} {\mathbf{Q}}_{k} {\mathbf{A}}{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}\right]_{n,n}\nonumber\\
&\!\!\!-\!\alpha\frac{ \kappa_{\mathrm{BS}}}{ \tau }\!\sum_{i=1}^{K}\beta_{2,i}\!\left[{\mathbf{H}}_{1}^{\H}{\mathbf{A}} {\mathbf{Q}}_{k}{\mathbf{C}} {\mathbf{Q}}_{k} {\mathbf{A}}{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},i}\right]_{n,n}\!
\nonumber\\
&+\!\alpha\beta_{2,k}\!\left[{\mathbf{H}}_{1}^{\H}{\mathbf{B}} {\mathbf{Q}}_{k} {\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}\right]_{n,n}.\label{LABC}
\end{align}
Also, we are going to use the following useful lemma.
\begin{lemma}\label{traceProd}
Let $ {\mathbf{A}} \in\mathbb{C}^{M\times M} $ be independent of $ {\boldsymbol{\Theta}}$ and ${\mathbf{R}}_{k}= \beta_{2,k}{\mathbf{H}}_{1} {\boldsymbol{\Theta}}$ ${\mathbf{R}}_{\mathrm{IRS},k}{\boldsymbol{\Theta}}^{\H}{\mathbf{H}}_{1}^{\H} $, then
\begin{align}
\mathop{\mathrm{tr}}\nolimits\left( \!\!{\mathbf{A}}\pdv{{\mathbf{R}}_{k}}{\phi^{*}_{n}}\!\right) =\alpha\beta_{2,k}[{\mathbf{H}}_{1}^{\H}{\mathbf{A}}{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}]_{n,n}.
\end{align}\end{lemma}
\proof We have
\begin{align}
\!\mathop{\mathrm{tr}}\nolimits\!\left(\! \!{\mathbf{A}}\pdv{{\mathbf{R}}_{k}}{\phi^{*}_{n}}\!\right) &\!=\!\sum_{i,j}[{\mathbf{A}}]_{ij}\pdv{[{\mathbf{R}}_{k}]_{ji}}{\phi^{*}_{n}}\\
&\!=\!\alpha\beta_{2,k}\sum_{i,j}[{\mathbf{A}}]_{ij}[{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}]_{jn}[{\mathbf{H}}_{1}^{\H} ]_{in}^{{\scriptscriptstyle\mathsf{T}}}\\
&\!=\!\alpha\beta_{2,k}[{\mathbf{H}}_{1}^{\H}{\mathbf{A}}{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}]_{nn},
\end{align}
since $ \pdv{[{\mathbf{R}}_{k}]_{ji}}{\phi^{*}_{n}}=\alpha\beta_{2,k} [{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}]_{jn}[{\mathbf{H}}_{1}^{\H} ]_{in}^{{\scriptscriptstyle\mathsf{T}}} $.
\endproof
By exploiting Lemma \ref{traceProd} for each term of \eqref{sk5} and that the trace of the transpose of a matrix equals the trace of this matrix, after several algebraic manipulations, we obtain
\begin{align}
\!\!\mathop{\mathrm{tr}}\nolimits\left({\boldsymbol{\Psi}}_{k}'\right)&\!=\!\alpha\beta_{2,k}\!\big[{\mathbf{H}}_{1}^{\H}{\mathbf{R}}_{k}{\mathbf{Q}}_{k}\!\big(\!2\mat{\mathrm{I}}_{M}\!-\!\big(1\!+\!\frac{\kappa_{\mathrm{UE}}}{\tau }\big)\!{\mathbf{R}}_{k}{\mathbf{Q}}_{k}\!\big)\!{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}\big]_{n,n}\nonumber\\
&-\!\alpha\frac{ \kappa_{\mathrm{BS}}}{ \tau }\!\sum_{i=1}^{K}\beta_{2,i}\!\big[{\mathbf{H}}_{1}^{\H}\!\left({\mathbf{R}}_{k}{\mathbf{Q}}_{k}\!\right)^{2}\!{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},i}\big]_{n,n}\!\\
&={\mathbf{L}}\! \left({\mathbf{R}}_{k},{\mathbf{R}}_{k},\mat{\mathrm{I}}_{M} \right)
\label{sk6},
\end{align}
which completes the derivation of $ S_{k}' $ after its insertion into \eqref{sk1}.
The computation of $ I_{k}' $ consists of the sum of the derivatives of different terms, requiring separate manipulations. Thus, we start by the computation of the derivative of the first term in \eqref{Den1} as
\begin{align}
&\!\! \left(\mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k} {\mathbf{R}}_{i}\right)\right)'=\mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k}' {\mathbf{R}}_{i}+{\boldsymbol{\Psi}}_{k} {\mathbf{R}}_{i}'\right)\\
&\!\!={\mathbf{L}}\! \left({\mathbf{R}}_{i}{\mathbf{R}}_{k},{\mathbf{R}}_{i}{\mathbf{R}}_{k},{\mathbf{R}}_{i} \right)+\alpha\beta_{2,k}[{\mathbf{H}}_{1}^{\H}{\boldsymbol{\Psi}}_{k}{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}]_{n,n},\label{I6}
\end{align}
where we have applied Lemma \ref{traceProd}. \
Moreover, we have
\begin{align}
\left(\mathop{\mathrm{tr}}\nolimits\left( {\boldsymbol{\Psi}}_{k}^{2}\right)\right)'&=2\mathop{\mathrm{tr}}\nolimits\!\left({\boldsymbol{\Psi}}_{k}' {\boldsymbol{\Psi}}_{k}\right)\\
&=2{\mathbf{L}}\! \left({\boldsymbol{\Psi}}_{k}{\mathbf{R}}_{k},{\boldsymbol{\Psi}}_{k}{\mathbf{R}}_{k},{\boldsymbol{\Psi}}_{k}\right).\label{I8}
\end{align}
In addition, we have
\begin{align}
\left(| \mathop{\mathrm{tr}}\nolimits \left( \mat{\mathrm{I}}_{M}\circ {\boldsymbol{\Psi}}_{k}\right)|^{2}\right)'&=2 \mathop{\mathrm{tr}}\nolimits \left( \mat{\mathrm{I}}_{M}\circ {\boldsymbol{\Psi}}_{k}\right) \mathop{\mathrm{tr}}\nolimits \left( \mat{\mathrm{I}}_{M}\circ {\boldsymbol{\Psi}}_{k}\right)'\\
&=2 \mathop{\mathrm{tr}}\nolimits \left( \mat{\mathrm{I}}_{M}\circ {\boldsymbol{\Psi}}_{k}\right) {\mathbf{L}}\! \left({\mathbf{R}}_{k},{\mathbf{R}}_{k}, \mat{\mathrm{I}}_{M}\right).\label{I9}
\end{align}
Furthermore, we obtain
\begin{align}
&\!\!\left(\mathop{\mathrm{tr}}\nolimits\! \left(\!\left(\mat{\mathrm{I}}_{M}\!\circ\! {\mathbf{R}}_{i}\right) \!{\boldsymbol{\Psi}}_{k}\right)\!\right)'\!=\!\mathop{\mathrm{tr}}\nolimits\!\left(\! \left(\!\left(\mat{\mathrm{I}}_{M}\!\circ\! {\mathbf{R}}_{i}\right)'\! {\boldsymbol{\Psi}}_{k}\!+ \!\left(\mat{\mathrm{I}}_{M}\!\circ \!{\mathbf{R}}_{i}\right)\!{\boldsymbol{\Psi}}_{k}'\right)\!\right)\\
&\!\!=\alpha\beta_{2,k}[{\mathbf{H}}_{1}^{\H}{\boldsymbol{\Psi}}_{k}{\mathbf{H}}_{1} {\boldsymbol{\Theta}}{\mathbf{R}}_{\mathrm{IRS},k}]_{n,n}\nonumber\\
&\!\!+{\mathbf{L}}\! \left(\left(\mat{\mathrm{I}}_{M}\circ {\mathbf{R}}_{i}\right){\mathbf{R}}_{k},\left(\mat{\mathrm{I}}_{M}\circ {\mathbf{R}}_{i}\right){\mathbf{R}}_{k},\mat{\mathrm{I}}_{M}\circ {\mathbf{R}}_{i}\right),\label{I10}
\end{align}
where we have used Lemma \ref{traceProd}. Equations \eqref{I6}-\eqref{I10} give $ I_{k}' $, which together with $ S_{k}' $ provide $ \pdv{\gamma_{k}}{\phi^{*}_{n}} $.
\end{appendices}
\bibliographystyle{IEEEtran}
\section*{Solutions}%
}
|
\section{Introduction}
Consider the problem of signal reconstruction for a large random matrix system (LRMS):
\begin{equation}\label{Eqn:linear_system}
\bf{y}=\bf{Ax}+\bf{n}
\end{equation}
where $\bf{A}$ is an $M\times N$ matrix with independent and identically distributed (IID) entries and $\bf{x}$ a length-$N$ vector with IID entries. Furthermore, we assume that the entries of $\bf{A}$ are Gaussian, but those of $\bf{x}$ are not necessarily Gaussian.
In a special case when $\bf{x}$ is un-coded, if $\bf{x}$ is Gaussian, the optimal solution can be obtained using the standard linear minimum mean square error (MMSE) methods. Otherwise, the problem is in general NP hard \cite{Micciancio2001,verdu1984_1}. Approximate message passing (AMP), derived from belief-propagation (BP) with Gaussian approximation and first order Taylor approximation, has attracted extensive research interest for this problem \cite{Bayati2011, Donoho2009}. A basic assumption of AMP is that $\bf{A}$ is IID Gaussian (IIDG). This assumption will hold throughout this paper.
AMP works by iterating between two local processors: namely, a linear detector (LD) and a non-linear detector (NLD). There is no matrix inversion involved, so its complexity is low. AMP has been studied for various signal processing and communication applications \cite{Ma2014denoising, Rush2017SSC, Liang2020, Barbier2017SSC, Liang2017CL}. Recently, it has been observed that AMP and its variations such as expectation propagation (EP) \cite{Cakmak2018, Minka2001} and orthogonal AMP (OAMP) \cite{Ma2016} can outperform the conventional Turbo linear MMSE (Turbo-LMMSE) in coded linear systems involving FEC coding \cite{Santos2017, MengVTC2015, MaTWC}. Most works on AMP in coded systems are simulation based. There is still a lack of rigorous analysis on the information theoretical limits of AMP in coded systems.
In this paper, we discuss the LRMS in \eqref{Eqn:linear_system} with FEC coding. The receiver is a variation of AMP with NLD formed by an \emph{a posteriori probability} (APP) decoder. For convenience of discussions, we define two classes of optimality for a receiver.
\begin{itemize}
\item A receiver is MMSE-optimal if it can achieve MMSE when $\bf{x}$ is an IID sequence.
\item A receiver is information theoretically optimal if it can achieve error free performance when $\bf{x}$ is coded with a rate which equals to the mutual information $I(\bf{x}; \bf{y})$.
\end{itemize}
The state evolution (SE) technique of AMP was originally derived to track the mean square error (MSE) in AMP during iterative processing. SE involves a scalar recursion of the transfer functions of LD and NLD. It has been shown via SE analysis that AMP can achieve MMSE asymptotically in the un-coded case when the transfer functions of LD and NLD have only one fixed-point \cite{ Tulino2013, Barbier2017arxiv, Reeves_TIT2019}. In this paper, we will show via SE analysis that AMP is information theoretically optimal, while the conventional methods, such as the well-known Turbo-MMSE algorithm \cite{Lei20161b,Yuan2014}, are not.
Our discussions are based on the following background works: (i) the I-MMSE relationship between mutual information and MMSE \cite{Guo2005}, (ii) the area property of iterative decoding systems \cite{Bhattad2007}, and (iii) the MMSE-optimality of AMP \cite{Tulino2013, Barbier2017arxiv, Reeves_TIT2019}. Similarly to \cite{Yuan2014, Lei20161b }, the performance of AMP can be optimized by matching the transfer functions of LD and decoder. The achievable rate can be analyzed using an area property similar to that for low density parity check (LDPC) decoders \cite{Yuan2014, Lei20161b}. We find that perfect matching is impossible for AMP: there is an inherent gap between the two transfer functions. Interestingly, AMP is still information theoretically optimal despite this gap, in the sense that its achievable rate can approach the mutual information $I(\bf{x}; \bf{y})$. The following are the main contributions of this paper.
\begin{itemize}
\item We show that the constrained capacity of a coded LRMS with an arbitrary input distribution (Gaussian or non-Gaussian) can be graphically interpreted as the area determined by the transfer functions of LD and MMSE NLD of an AMP. We establish an area property for AMP and derive its achievable rate under a matching condition. We prove that this achievable rate equals to the constrained capacity of an LRMS derived in \cite{Barbier2017arxiv, Reeves_TIT2019}, thereby showing the potential information theoretic optimality of AMP in coded linear systems.
\item We develop a matching strategy for AMP. We provide numerical results to demonstrate the efficiency of the matching strategy for binary signaling. These findings provide a promising direction to significantly enhance the performance of coded linear systems.
\end{itemize}
\section{Preliminaries}
{In this section, we briefly outline the area property, the capacity of an LRMS, and the AMP algorithm.
\subsection{Area Property for SISO-AWGN Channel}\label{Sec:SISO_area}
A SISO-AWGN channel is defined as \begin{equation}\label{Eqn:SISO_AWGN}
y=\sqrt{\rho}x+z,
\end{equation}
where $z\sim \mathcal{CN}(0,1)$, $\rho$ denotes the signal-to-noise-ratio (SNR), $x\sim P_X(x)$ and $P_X(x)$ is an arbitrary distribution on a constellation $\mathcal{S}$. The MMSE of \eqref{Eqn:SISO_AWGN} is denoted as
\begin{equation}
\omega_{\cal S}(\rho) \equiv \mathrm{mmse}(x|\sqrt{\rho}x+z, x\sim P_X(x)).
\end{equation}
The following theorem, proved in \cite{Guo2005}, establishes the connection between MMSE and the capacity given $P_X(x)$ for a SISO-AWGN channel.
\begin{theorem}[Scalar I-MMSE]\label{Lem:S-I-MMSE}
Let SNR$=\rho^*$. The capacity of a SISO-AWGN channel equals to the area under $\omega_{\cal S}(\rho)$ from $\rho=0$ to $\rho=\rho^*$, i.e.,
\begin{equation}\label{Eqn:C_mmse}
\!\!C_{\rm SISO}(\rho^*) \!=\!\! I({x}; \sqrt{\!\rho^*}x\!+\!z) = \!\!\int_{0}^{\rho^*} \omega_{\cal S}(\rho) d\rho.
\E
\end{theorem}
The following is an instance of Theorem \ref{Lem:S-I-MMSE}.
\emph{Code-Rate-MMSE Lemma \cite{Bhattad2007}:} Let the code length be $N$ and code rate $R_{\cal C} =K/N$. We treat the code-book $\bf{{\mathcal{C}}}=\{\bf{c}_1,\cdots,\bf{c}_{2^K}\}$ as a uniformly distributed $N$-dimension constellation with ${2^K}$ discrete points. When SNR$\to \infty$, the capacity per length-$N$ code block approaches to the entropy of $\bf{{\mathcal{C}}}$, i.e., $\log(2^K)=K$. The entropy per dimension is $K/N$. Hence,
\begin{equation}\label{Eqn:R_mmse}
R_{\cal C} = \int_{0}^{\infty}\omega_{\mathcal{C}}(\rho)d\rho =K/N,
\end{equation}
where $ \omega_\mathcal{C}(\rho) \!\equiv\! \tfrac{1}{N}\mathrm{mmse}(\bf{x}|\sqrt{\rho}\bf{x}\!+\!\bf{z},{\bf{x}\!\in \!{\bf{{\mathcal{C}}}}})$.
\subsection{LRMS Capacity}
Return to the LRMS in \eqref{Eqn:linear_system}: $\bf{y}=\bf{Ax}+\bf{n}$,
where $\bm{y}\!\in\!\mathbb{C}^{M\!\times\!1}$ is a vector of observations, $\bf{A}\!\in\!\mathbb{C}^{M\!\times\! N}$ an IIDG matrix with $A_{ij}\sim \mathcal{CN}({0},1/M)$, $\{x_i\sim P_X(x), \forall i\}$, and $\bm{n}\!\sim\!\mathcal{CN}(\mathbf{0},\sigma^2\bm{I}_M)$ a vector of Gaussian additive noise samples. Fig. \ref{Fig:model}(a) shows
a modulated LRMS. In this paper, we consider a large system with $M,N\to\infty$ and a fixed $\beta=N/M$. The transmit SNR is defined as $snr = \sigma^{-2}$.
We assume that $\bf{A}$ is known at the receiver, but unknown at the transmitter.
\begin{figure}[t]
\centering
\includegraphics[width=6cm]{System_model_a.pdf}\\ \vspace{-0.2cm}
\caption{Un-coded LRMS: transmitter and AMP receiver, where ``Demodulate'' and LD in (b) correspond to ``Modulate'' and ``$\bf{A}\bf{x}+\bf{n}$'' in (a) respectively.}\label{Fig:model}\vspace{-0.2cm}
\end{figure}
The constrained capacity of an LRMS given $P_X(x)$ was proved in \cite{Barbier2017arxiv, Reeves_TIT2019}.
\begin{theorem}[Capacity]\label{Pro:dis_cap}
Assume that the signal distribution $P_X(x)$ satisfies the single-crossing property, i.e., $\zeta \!=\!\beta \cdot snr \cdot \omega\big(1/[\beta(1+\zeta)]\big)$ has exactly one positive fixed point $\zeta^*$. Then, the capacity of the LRMS in \eqref{Eqn:linear_system} is given by \vspace{-0.1cm}
\begin{equation}\label{Eqn:dis_cap}
C \!=\! \beta^{-\!1}\!\big[\! \log({1\!+\!\zeta^*})\!-\!{\zeta^*}\!/\!({1\!+\!\zeta^*}) \!\big] \!+ C_{\rm SISO} \big(snr/(1\!+\!\zeta^*)\!\big),
\end{equation}
where $C_{\rm SISO}(\cdot)$ is defined in \eqref{Eqn:C_mmse}.
\end{theorem}
In \cite{Lei2019arXiv}, we provide a concise derivation of the LRMS capacity, using the properties of AMP.
\subsection{Overview of AMP}\label{Sec:AMP_uncoded}
AMP \cite{Donoho2009} finds an approximate MMSE solution to the problem in \eqref{Eqn:linear_system} using the following iterative process (initialized with $t=0$ and $\bf{s}^0=\bf{r}^0_{\mathrm{Onsager}}=\bf{0}$):
\begin{subequations}\label{Eqn:AMP}\begin{align}
\mathrm{LD:}\;\; & \bf{r}^t\!=\! f(\bf{s}^t) \!\equiv \! \bf{s}^t \!+\! \bf{A}^H(\bf{y}\!-\!\bf{A}\bf{s}^t) \!+ \!\bf{r}^t_{\mathrm{Onsager}},\label{Eqn:LD}\\
\mathrm{NLD_{SBS}}: \;\; &\bf{s}^{t+1} = \eta(\bf{r}^{t})\equiv \mathrm{E}\{\bf{x}|\bf{r}^{t}\},\label{Eqn:NLD}
\end{align}\end{subequations}
where $\eta(\bf{r}^{t})$ is a symbol-by-symbol (SBS) MMSE demodulate function, and $\bf{r}^t_{\mathrm{Onsager}}$ is an ``Onsager term'' defined by $\bf{r}^t_{\mathrm{Onsager}}\!=\!\beta \langle\eta'(\bf{r}^{t-1})\rangle (\bf{r}^{t-1}\!-\!\bf{s}^{t-1})$ \cite{Donoho2009}. Fig. \ref{Fig:model}(b) is a graphical illustration of AMP, where the linear detector (LD) and non-linear detector (NLD) correspond to \eqref{Eqn:LD} and \eqref{Eqn:NLD} respectively.
We define the errors as
\begin{equation}\label{Eqn:errors}
\bf{h}^t \equiv \bf{r}^t -\bf{x} \quad {\rm and} \quad \bf{q}^t \equiv \bf{s}^t -\bf{x}.
\end{equation}
Let $\rho^t$ be the signal-to-interference-plus-noise-ratio (SINR) for $\bf{r}^t$ and $v^t$ the MSE for $\bf{s}^t$:
\begin{equation}\label{Eqn:rho_v}
\rho^t \equiv N\big[{\rm E}\big\{\|\bf{h}^t\|^2\}\big]^{-1},\qquad
v^t \equiv \tfrac{1}{N}{\rm E}\big\{\|\bf{q}^t\|^2\big\}.
\end{equation}
The following theorem summarizes the findings in \cite{Bayati2011}.
\begin{theorem}\label{Pro:SE}
Let $M,N\to\infty$ with a fixed $\beta=N/M$. For AMP, $\bf{h}^t$ defined in \eqref{Eqn:errors} can be modeled by a sequence of IIDG samples independent of $\bf{x}$. The LD and NLD of AMP can be characterized by the following transfer functions \cite{Bayati2011
\begin{subequations} \begin{align}
\mathrm{LD:}& \quad \rho^t =\phi(v^t) = ({\beta v^t + \sigma^2})^{-1},\label{Eqn:LD_form}\\
\mathrm{NLD_{SBS}:}& \quad v^{t+1}= \omega_{\cal S}(\rho^t), \label{Eqn:NLD_MMSE}
\end{align}\end{subequations}
where $\omega_{\cal S}(\cdot)$ is the MMSE function given in \ref{Sec:SISO_area}.
\end{theorem}
\begin{figure}[t]
\centering
\includegraphics[width=6cm]{SE_track_uncoded.pdf}\\%\vspace{-0.3cm}
\caption{Graphical illustration of an AMP, where $\phi^{-1}$ is the inverse of $\phi$ given in \eqref{Eqn:LD_form} and $\omega$ is defined in \eqref{Eqn:NLD_MMSE}. The iterative process of AMP is illustrated by the evolution trajectory, and the fixed point $(\rho^*, v^*)$ gives the MMSE. From \eqref{Eqn:LD_form}, we have $\phi(0)=snr$.}\label{Fig:TF_chart
\end{figure}
{\begin{assumption}\label{Pro:SCP}
There is exactly one fixed point for $\omega_{\cal S}(\rho) = \phi^{-1}(\rho)$ for $\rho>0$, where $\phi^{-1}(\cdot)$ is the inverse of $\phi(\cdot)$.
\end{assumption}}
It is proved in \cite{Lei2019arXiv} that Assumption \ref{Pro:SCP} rigorously holds for Gaussian signaling. Fig. \ref{Fig:TF_chart} provides a graphical illustration of Assumption \ref{Pro:SCP}. The evolution trajectory of AMP converges to a unique fixed point $(\rho^*, v^*)$. The following theorem was first established in \cite{Tulino2013} via replica method, and then was rigorously proved in \cite{Barbier2017arxiv, Reeves_TIT2019}.
\begin{theorem}[MMSE Optimality]\label{Lem:mmse}
Let $\hat{{\bf{x}}}_{\mathrm{MMSE}}=\mathrm{E}\{\bf{x}|\bf{y},x_i\!\sim\! P_X(x), \forall i\}$ be the conditional mean of $\bf{x}$ given $\bf{y}$ and $\{x_i\!\sim\! P_X(x), \forall i\}$ and suppose that {Assumption \ref{Pro:SCP} holds.} Then
\begin{equation}\label{Eqn:snr_rho}
v^*=\tfrac{1}{N}\mathrm{E}\big\{ \|\bf{x}\!-\!\hat{\bf{x}}_{\mathrm{MMSE}}\|^2 \big\}
\end{equation}
i.e., AMP converges to the MMSE of the un-coded LRMS.
\end{theorem}
\section{Capacity Optimality of AMP}\label{Sec:Cap_Opt}
In this section, we investigate the achievable rate of the iterative AMP receiver with FEC decoding.
\subsection{Coded System Model and AMP}
\begin{figure}[b!]
\centering
\includegraphics[width=6.5cm]{System_model.pdf}\\\vspace{-2mm}
\caption{Coded linear system: Transmitter and AMP. ``APP DEC'' (\emph{a-posteriori} probability decoding), ``Demodulate'' and LD in (b) correspond to ``ENC'' (encode), ``Modulate'' and ``$\bf{A}\bf{x}+\bf{n}$'' in (a) respectively.}\label{Fig:model_coded}
\end{figure}
Fig. \ref{Fig:model_coded}(a) shows an LRMS involving FEC coding. We write $\bf{x}\in \bf{{\mathcal{C}}}$ for coded $\bf{x}$. The other conditions are the same as that in Fig. \ref{Fig:model}. We focus on the AMP receiver in Fig. \ref{Fig:model_coded}(b).
\underline{\emph{AMP:}} Initialized with $t\!=\!0$ and $\bf{s}^0\!=\!\bf{r}^0_{\mathrm{Onsager}}\!=\!\bf{0}$,
\begin{subequations}\label{Eqn:Turbo-AMP}\begin{align}
&\mathrm{LD:}\quad\; \bf{r}^t\!=\! f(\bf{s}^t) \!\equiv \! \bf{s}^t + \bf{A}^H(\bf{y}\!-\!\bf{A}\bf{s}^t) + \bf{r}^t_{\mathrm{Onsager}}, \\
&\mathrm{NLD_{DEC}:} \;\; \bf{s}^{t+1} = \eta_{\cal C}(\bf{r}^{t})\equiv \mathrm{E}\{\bf{x}|\bf{r}^{t}, {\bf{x}\in {\bf{{\mathcal{C}}}}}\}.\label{Eqn:Turbo-NLD}
\end{align}\end{subequations}
The symbol-wise NLD in \eqref{Eqn:AMP} of AMP is replaced by an \emph{a-posteriori} probability (APP) decoder in \eqref{Eqn:Turbo-AMP} for coded $\bf{x}$.
Theorem \ref{Pro:SE} gives the IIDG property for AMP for un-coded $\bf{x}$. The discussions in this paper are based on the following assumption for coded $\bf{x}$.
\begin{assumption}\label{Pro:SE_new}
Theorem \ref{Pro:SE} still holds for AMP for coded $\bf{x}$, i.e., $\bf{h}^t$ is IIDG and independent of $\bf{x}$ and LD and NLD of AMP can be characterized by
\begin{subequations}\label{Eqn:Turbo_AMP_SE}\begin{alignat}{2}
\mathrm{LD:} &\quad &&\rho =\phi(v), \\
\mathrm{NLD_{DEC}:} &\quad &&v\!= \!\omega_\mathcal{C}(\rho)\!\equiv\! \tfrac{1}{N}\mathrm{mmse}(\bf{x}|\sqrt{\rho}\bf{x}\!+\!\bf{z},{\bf{x}\!\in\! {\bf{{\mathcal{C}}}}}).\label{Eqn:NLD_MMSE_coded}
\end{alignat}
\end{subequations}
\end{assumption}
The $\phi(v)$ in AMP is the same as that in AMP, and $\omega_\mathcal{C}(\rho)$ depends on the code constraint.
\subsection{Area Property and Capacity Optimality of AMP}\label{Sec:area_AMP}
In the un-coded case in \eqref{Eqn:AMP}, AMP converges to a fixed $(\rho^*,v^*)$ in Fig. \ref{Fig:TF_chart}. Detection is not error free as $v^*>0$. In the coded case, it is possible to achieve error-free detection using a properly designed $\omega_\mathcal{C}(\rho)$. As illustrated in Fig. \ref{Fig:track}, the key is to create a detection tunnel that converges to $v=0$, implying zero error rate. There should be no fixed point between $\omega_\mathcal{C}(\rho)$ and $\phi^{-1}({\rho })$, since otherwise the tunnel will be closed at $v>0$. This requires that
\begin{subequations}\begin{equation}\label{Eqn:error-free}
\omega_{{\mathcal{C}}}(\rho) <\phi^{-1}(\rho), \;\;\; {\rm for} \;\;0\le\rho\le snr.
\end{equation}
Also, by definition, the MMSE $\rm NLD_{DEC}$ in \eqref{Eqn:NLD_MMSE_coded} should achieve an MSE lower than that of a SBS detector, i.e.,
\begin{equation}\label{Eqn:coding_gain}
\omega_{{\mathcal{C}}}(\rho) < \omega_{\cal S}(\rho), \;\;\; {\rm for} \;\; \rho \geq 0.\vspace{-0.1cm}
\end{equation}\end{subequations}
Combining \eqref{Eqn:error-free} and \eqref{Eqn:coding_gain}, we obtain a necessary and sufficient condition for AMP to achieve error-free detection:
\begin{subequations}\label{Eqn:upper_bound}\begin{equation}
\omega_{\mathcal{C}}({\rho})< \omega_{\mathcal{C}}^*({\rho}),\;\;\; {\rm for} \;\; 0 \leq {\rho } \leq snr,
\end{equation}
where
\begin{equation} \label{Eqn:w_star}
\omega_{\mathcal{C}}^*({\rho})= \min \{\omega_{\cal S}(\rho), \phi^{-1}({\rho }) \}.
\end{equation}\end{subequations}
\begin{figure}[t]
\centering
\includegraphics[width=6.5cm]{SE_track_coded.pdf}\\%\vspace{-0.2cm}
\caption{Graphical illustration of AMP, where $\omega_{\mathcal{S}}$ is a demodulation function (un-coded case) and $\omega_{{\mathcal{C}}}$ is a transfer function of a decoder (coded case). The iterative process of AMP is illustrated by the evolution trajectory between $\phi^{-1}$ and $\omega_{{\mathcal{C}}}$. }\label{Fig:track}
\end{figure}
\begin{proposition}\label{The:area_LRMS}
Suppose that Assumption \ref{Pro:SCP} holds. Then the constrained capacity of an LRMS with the given $\mathcal{\bf{S}}$ is
\begin{subequations}\label{Eqn:dis_cap_new2}\begin{equation}
C=A_{\omega_{\mathcal{C}}^*},
\end{equation}
where $A_{\omega_{\mathcal{C}}^*}$ is the area covered by $\omega_{\mathcal{C}}^*$, i.e.,
\begin{align}
A_{\omega_{\mathcal{C}}^*} = \beta^{-1}[\rho^{*}/snr\!-\!\log(\rho^{*}/snr)\!-\!1] \!+\!\! \int_0^{\rho^{*}} \!\!\! \omega_{\cal S}(\rho) d \rho.
\end{align} \end{subequations}
\end{proposition}
\begin{IEEEproof}
See APPENDIX \ref{APP:Consistency}.
\end{IEEEproof}
Combining \eqref{Eqn:R_mmse}, \eqref{Eqn:upper_bound} and \eqref{Eqn:dis_cap_new2}, we obtain the capacity optimality of AMP below.
\begin{theorem}[Capacity Optimality]\label{The:cap_opt}
Assume that Assumptions \ref{Pro:SCP} and \ref{Pro:SE_new} hold and AMP converges to $v=0$. Then,
\begin{equation}
R_{\cal C}\to C,
\end{equation}
if $\omega_{\mathcal{C}}(\rho)\to \omega_{\mathcal{C}}^*(\rho)$ in $[0, snr]$.
\end{theorem}
Fig. \ref{Fig:Area_A} gives a graphical illustration of Theorem \ref{The:cap_opt}. Note that Theorem \ref{The:cap_opt} is based on a matching condition:
\begin{equation}\label{Eqn:C2Copt}
\omega_{\mathcal{C}}(\rho)\to \omega_{\mathcal{C}}^*(\rho).
\end{equation}
A proof for the existence of a code achieving \eqref{Eqn:C2Copt} can be found in \cite{Lei2019arXiv} for Gaussian signaling. For other signaling, the existence of such a code is a conjecture only.
\begin{figure}[t]
\centering
\includegraphics[width=6.5cm]{perfect_match.pdf}\\%\vspace{-0.2cm}
\caption{Graphical illustration of the capacity, the maximum achievable rate of AMP and the optimal transfer function of decoder. The maximum achievable rate of AMP equals to the capacity, which is the area covered by $\omega_{\mathcal{C}}^*$. }\label{Fig:Area_A}
\end{figure}
\subsection{Rate Comparison with Turbo-LMMSE}
It is proved in \cite{Lei20161b,Yuan2014} that Turbo-LMMSE is capacity achieving for Gaussian signaling. In the following, we show that Turbo-LMMSE is sub-optimal for non-Gaussian signaling.
{The main difference between AMP and Turbo-LMMSE is as follows. To avoid the correlation problem in the iterative process, Turbo-LMMSE uses extrinsic local processors (e.g. an extrinsic LD and an extrinsic decoder), while AMP uses an ``Onsager"-term.}
Assume that the transfer functions of the detector and the decoder in Turbo-LMMSE are matched. The achievable rate of Turbo-LMMSE is given in \cite{Yuan2014}
\begin{equation}
R_{\mathrm{LMMSE}}=\log|\mathcal{S}|-\int_0^{+\infty} \omega_\mathcal{S}(\rho+\phi(\omega_\mathcal{S}(\rho))) d\rho.
\end{equation}
Fig. \ref{Fig:Rate_AMP_Turbo} shows the capacity and the achievable rates of AMP and Turbo-LMMSE. The capacity for Gaussian signaling is achieved by both AMP and Turbo-LMMSE. For QPSK, 8PSK and 16QAM modulation, the achievable rate of AMP equals to the capacity when {Assumption \ref{Pro:SCP} holds}, while Turbo-LMMSE always has rate loss. Similar results can be obtained for other non-Gaussian signaling. In addition, the gap between AMP and Turbo-LMMSE increases with $\beta$. This gap $\to0$ when $\beta\to0$. {The reason why Turbo-LMMSE has performance loss is that extrinsic update leads to performance loss for non-Gaussian signal processing, which was first pointed out in \cite{MaTWC}.
\begin{figure}[t
\centering
\includegraphics[width=8.cm]{Rate_AMP_Turbo.pdf}\\%\vspace{-0.2cm}
\caption{Comparison between the capacity and the achievable rates of AMP and Turbo-LMMSE of an LRMS with $\beta=N/M=\{1, 1.5\}$, where $C_{\rm Gau}$ denotes the Gaussian capacity and also the achievable rates of AMP and Turbo-LMMSE with Gaussian signaling, $R_{\mathrm{T-AMP}}$ and $R_{\mathrm{T-LMMSE}}$ respectively denote the achievable rates of AMP and Turbo-LMMSE with QPSK, 16QAM and 8PSK modulations.}\label{Fig:Rate_AMP_Turbo
\end{figure}
\subsection{Rate Comparison with Cascading AMP and Decoding}\label{Sec:AMP_DEC}
We define a cascading AMP and decoding (AMP-DEC) scheme \cite{Guo2005random, Tanaka2002} as follows. We run AMP until it converges. The result is used by decoder. There is no iteration between AMP and the decoder. The achievable rate of AMP-DEC is
\begin{equation}\label{Eqn:dis_ach}
R_{\mathrm{AMP-DEC}}= \int_0^{{{\rho}}^{*}} \!\!\! \omega_\mathcal{S}(\rho) d \rho.
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=8.cm]{AMP_AWGN.pdf}\\%\vspace{-0.2cm}
\caption{Comparison between the achievable rates of AMP, and separate optimal MMSE detection and ideal SISO decoding in \cite{Guo2005random, Tanaka2002} with $\beta=N/M=\{1, 1.5\}$, where $C_{\rm Gau}$ denotes the Gaussian capacity and the achievable rates of AMP with Gaussian signaling, $R_{\mathrm{T-AMP}}$ and $R_{\mathrm{AMP-DEC}}$ respectively denote the achievable rates of AMP and AMP-DEC with QPSK, 16QAM and 8PSK modulations. }\label{Fig:Rate_AMP_AWGN
\end{figure}
Fig. \ref{Fig:Rate_AMP_AWGN} compares AMP and AMP-DEC. For QPSK, 8PSK and 16QAM modulations, the achievable rate of AMP-DEC is lower than that of AMP. This gap increases with $\beta$, but is negligible if $\beta$ is small (e.g. $\beta<0.5$ based on our experimental findings). Furthermore, different from the rate of AMP that always increases with the size of constellation, the rate of AMP-DEC decreases with the increasing of the constellation size for large $\beta$.
\newpage
\section{Simulation Results}
This section provides BER simulations for AMP with optimized irregular LDPC codes. The details of irregular LDPC code optimization for AMP can be found in \cite{Lei2019arXiv}.
\subsection{BER Comparison with Irregular and Regular LDPC Codes}\label{Sec:BER_per}
Fig. \ref{Fig:BER_AMP} provides the BER simulations for an LRMS, in which $\bf{x}$ is generated using optimized irregular LDPC codes \cite{Yuan2008Low, Chung2001} {with code length $=10^5$}. The AMP (see Fig. \ref{Fig:model_coded}) for an optimized LDPC coded LRMS is denoted as ``Opt-Irreg''. The APP decoder is implemented using a standard sum-product decoder. The channel loads are $\beta=\{ 0.5, 1, 2\}$ with $(N, M)=(250, 500), (500, 500)$ and $(500, 250)$, respectively.
To verify the finite-length performance of the irregular LDPC codes with code rate $\approx0.5$, we provide the BER performances of the optimized codes. QPSK modulation is used. The rate of each symbol is $R_{\cal C}\approx1$ bits/symbol, and the sum rate is $R_{sum}\approx N$ bits per channel use. The maximum iteration number is $200\sim700$. Fig.~\ref{Fig:BER_AMP} shows that for all $\beta$, gaps between the BER curves of the codes at $10^{-5}$ and the corresponding Shannon limits are within $0.7 \sim 1$~dB.
To validate the advantage of matching principle, we provide AMP for a standard regular (3, 6) LDPC code (denoted as ``(3, 6)'') \cite{Gallager1962}, and a SISO irregular LDPC code \cite{Chung01} (denoted as ``SISO-Irreg''), corresponding to $R_{\mathrm{AMP-DEC}}$ discussed in Section \ref{Sec:AMP_DEC}. As shown in Fig.~\ref{Fig:BER_AMP}, when the BER curves of three systems are at $10^{-5}$, the optimized irregular LDPC codes have $0.8 \sim 2$~dB performance gains over the un-optimized regular (3, 6) LDPC code for $\beta=\{0.5, 1, 2\}$, and $0.5 \sim 6$~dB performance gains over ``SISO-Irreg'' for $\beta=\{0.5, 1, 2\}$. These results show that code optimization can provide attractive performance improvement, especially for the large $\beta$.
\begin{figure}[htb]
\centering
\includegraphics[width=8.5cm]{BER_AMP.pdf}\\\vspace{-2mm}
\caption{BER performances of AMP, where $C$ deontes the capacity limit, ``Thre'' the BER threshold, ``Opt-Irreg'' the BER of AMP-optimized irregular LDPC codes, ``SISO-Irreg'' the BER with SISO-optimized irregular LDPC codes, ``(3, 6)'' the BER of AMP with regular (3, 6) LDPC code. Code length = $10^5$, code rate $\approx$ 0.5, QPSK modulation, and iterations = $200\sim700$, and $\beta=N/M=\{0.5,1,2\}$. }\label{Fig:BER_AMP
\end{figure}
\subsection{BER Comparison with Optimized Turbo-LMMSE}
We now compare AMP and Turbo-LMMSE \cite{ YC2018TWC}. We consider a $500\times 333$ QPSK LRMS with $\beta=1.5$. As shown in Fig. \ref{Fig:Rate_AMP_Turbo}(b), the SNR limits of AMP and Turbo-LMMSE for the target rate $R_{\cal C}=1.48\approx1.5$ are $5.38$ dB and $7.99$ dB respectively. Fig. \ref{Fig:BER_AMP_Turbo} shows the BER performances of AMP and Turbo-LMMSE (with iterations $=200$) using optimized LDPC codes. The thresholds of AMP and Turbo-LMMSE are $5.62$ dB and $8.50$ dB respectively, 0.24 dB and 0.51 dB away from the corresponding achievable rate limits, and 0.6 dB and 1.2 dB away from their respective thresholds. We can see that, AMP has 3.5 dB improvement in BER over Turbo-LMMSE.
\begin{figure}[t]
\centering
\includegraphics[width=6.5cm]{BER_Turbo_AMP.pdf}\\%\vspace{-0.2cm}
\caption{BER performances (right) of AMP and Turbo-LMMSE \cite{YC2018TWC, Lei20161b} with optimized irregular LDPC codes, where $C_{\rm Gau}$ denotes the Gaussian capacity, ``SE'' the state evolution, ``lim'' the QPSK achievable rate limits of AMP/Turbo-LMMSE, ``sim'' the simulated BERs. Code length = $10^5$, code rate $\approx$ 0.74, QPSK modulation, iterations = $200$, $N=500$ and $M=333$. }\label{Fig:BER_AMP_Turbo
\end{figure}
\section{Conclusion}
This paper is on an AMP based scheme for a coded LRMS with arbitrary input distributions. We show that AMP is information theoretically optimal using a curve matching principle and the IIDG assumption. In addition, a code design principle is provided for AMP, and the irregular LDPC codes are considered for binary signaling as an example. The numerical results show that AMP is capacity-approaching (i.e. within 1dB away from the limit) based on optimized irregular LDPC codes, and significant performance improvements ($0.8$ dB $\sim$ $4$ dB) are observed over the system without code optimization. Apart from that, AMP has lower complexity and better performance that the well-known Turbo-LMMSE.
\appendices
\section{Proof of Proposition \ref{The:area_LRMS}}\label{APP:Consistency}
Let ${\rho}^* \!=\! snr/(1\!+\!\zeta^*)$, i.e. $\zeta^*\!=\!snr/{\rho}^* \!-\!1$. Then the fixed point function in \eqref{Eqn:dis_cap} is rewritten to
\begin{equation}
snr/{\rho}^* -1 = \beta \, snr \,\omega_{\cal S}({\rho^*} ),
\end{equation}
which is equivalent to the fixed point function $\omega_{\cal S}(\rho)=\phi^{-1}(\rho)$. Substituting \eqref{Eqn:C_mmse} and ${\rho}^* = snr/(1+\zeta^*)$ into \eqref{Eqn:dis_cap_new2}, we have\vspace{-2mm}
\begin{subequations} \begin{align}
&\!\!\!A_{\omega_{\cal C}^*} = \beta^{-1}\big[\rho^{*}/snr\!-\!\log(\rho^{*}/snr)\!-\!1\big] \!+\!\! \int_0^{\rho^{*}} \!\!\! \omega_{\cal S}(\rho) d \rho\\
&\!\!\!= \!\beta^{-1}\!\big[\! \log({1\!+\!\zeta^*})\!-\!{\zeta^*}\!/\!({1\!+\!\zeta^*})\! \big] \!\!+\! C_{\rm SISO} \big(snr\!/\!(1\!+\!\zeta^*)\!\big).
\end{align}\end{subequations}
This is the same as the capacity $C$ in \eqref{Eqn:dis_cap}. Hence, we complete the proof of Proposition \ref{The:area_LRMS}.
|
\subsection{\pfref{thm:bandit-unknown}}
\begin{lemma}\label{lem:optimistic-transition-bound}
With probability at least $1-4\delta$, we have
\begin{align*}
\sum_{k=1}^K\inner{u_k-q_k}{c_k} \leq 32\sqrt{S^3A^2\ln^2\rbr{\frac{HKS^2A^2}{\delta}}\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{SAH^3\sqrt{K}}}} + \tilO{\Qumq}.
\end{align*}
\end{lemma}
\begin{proof}
Denote $c^{(s, a)}_k(s', a')=c_k(s', a')\field{I}\{s'=s, a'=a\}$. Then by \pref{lem:transition-bias} (with $\delta$ there set to $\delta/|\Gamma|$) and a union bound over all $(s,a)$, we have with probability $1-4\delta$, for any $(s, a)\in\Gamma$:
\begin{align*}
\sum_{k=1}^K(u_k(s, a) - q_k(s, a))c_k(s, a) &= \sum_{k=1}^K\inner{u_k-q_k}{c_k^{(s, a)}}\\
&\leq 32\sqrt{S^2A\ln^2\rbr{\frac{HKS^2A^2}{\delta}}\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k^{(s, a)}}} + \tilO{H^3\sqrt{K}}}} + \tilO{\Qdtd}.
\end{align*}
Hence,
\begin{align*}
\sum_{k=1}^K\inner{u_k-q_k}{c_k} &\leq 32\sumsa\sqrt{S^2A\ln^2\rbr{\frac{HKS^2A^2}{\delta}}\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k^{(s, a)}}} + \tilO{H^3\sqrt{K}}}} + \tilO{\Qumq}\\
&\leq 32\sqrt{S^3A^2\ln^2\rbr{\frac{HKS^2A^2}{\delta}}\sumsa\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k^{(s, a)}}} + \tilO{H^3\sqrt{K}}}} + \tilO{\Qumq}\\
&= 32\sqrt{S^3A^2\ln^2\rbr{\frac{HKS^2A^2}{\delta}}\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{SAH^3\sqrt{K}}}} + \tilO{\Qumq},
\end{align*}
where in the second line we use Cauchy-Schwarz inequality.
\end{proof}
Below we present the proof of \pref{thm:bandit-unknown}.
It decomposes the regret into several terms, each of which is bounded by a lemma included after the proof.
\begin{proof}[\pfref{thm:bandit-unknown}]
By \pref{lem:loop-free}, with probability at least $1-\delta$,
\[
R_K \leq \sum_{k=1}^K\inner{\widetilde{N}_{k}-q_{{\widetilde{\pi}^\star}}}{c_k} + \bigO{D^{3/2}S^2A(\ln\tfrac{1}{\delta})^2}.
\]
We define a slightly perturbed benchmark $q^\star = (1-\frac{1}{TK})q_{{\widetilde{\pi}^\star}} + \frac{1}{TK}q_0 \in {\widetilde{\Delta}}(T, \{P\})$ (note that ${\widetilde{\Delta}}(T,\{P\})\subseteq{\widetilde{\Delta}}(T, {\mathcal{P}}_{i_k}),\forall k$ under the event of \pref{lem:transition-interval}) for some $q_0 \in {\widetilde{\Delta}}(T, \{P\})$ with $q_0(s,a) \geq \frac{1}{K^3}$ for all $(s,a) \in \widetilde{\SA}$, so that $\qfeat^\star=q^\star+\lambda\h{q^\star} \in \Omega_k,\forall k$.
Also define
$b_k \in \field{R}^{\widetilde{\SA}}$ such that $b_k(s,a) = \frac{\sum_h h u'_k(s, a, h)c_k(s, a)}{u_k(s, a)}$, which clearly satisfies $\field{E}_k[\widehat{b}_k] = b_k$.
We then decompose $\sum_{k=1}^K\inner{\widetilde{N}_{k}-q^\star}{c_k}$ as
\begin{align*}
&\sum_{k=1}^K\inner{\widetilde{N}_{k}-q^\star}{c_k} \\
&= \sum_{k=1}^K\inner{u_{k}}{\widehat{c}_k} - \sum_{k=1}^K\inner{q^\star}{c_k}
\tag{$\inners{\widetilde{N}_{k}}{c_k} = \inner{u_k}{\widehat{c}_k}$} \\
&= \sum_{k=1}^K\inner{u_{k}-\widehat{q}_k}{\widehat{c}_k} + \sum_{k=1}^K\inner{\widehat{q}_k}{\widehat{c}_k} - \inner{q^\star}{c_k}\\
&= \textsc{Err}_1 + \sum_{k=1}^K\inner{\qfeat_{k}-\qfeat^\star}{\widehat{c}_k} + \sum_{k=1}^K\inner{\qfeat^\star}{\widehat{c}_k-c_k} + \lambda\sum_{k=1}^K\inner{\h{q^\star}}{c_k} - \lambda\sum_{k=1}^K\inner{\h{\widehat{q}_k}}{\widehat{c}_k} \tag{define $\textsc{Err}_1=\sum_{k=1}^K\inner{u_k-\widehat{q}_k}{\widehat{c}_k}$} \\
&= \textsc{Err}_1 + \sum_{k=1}^K\inner{\qfeat_{k}-\qfeat^\star}{\widehat{c}_k} + \sum_{k=1}^K\inner{\qfeat^\star}{\widehat{c}_k-c_k} + \tilO{\lambda DTK} - \lambda\sum_{k=1}^K\inner{\h{\widehat{q}_k}}{\widehat{c}_k} \\
&= \sum_{k=1}^K\inner{\qfeat_{k}-\qfeat^\star}{\widehat{c}_k} + \tilO{\lambda DTK} + \textsc{Err}_1 + \textsc{Bias}_1 + \textsc{Bias}_2 - \lambda\sum_{k=1}^K\inner{\h{\widehat{q}_k}}{c_k}
\tag{define $\textsc{Bias}_1 = \sum_{k=1}^K\inner{\qfeat^\star}{\widehat{c}_k-c_k}$ and $\textsc{Bias}_2 = \lambda\sum_{k=1}^K\inner{\h{\widehat{q}_k}}{c_k - \widehat{c}_k}$} \\
&= \textsc{Reg}_\qfeat + \tilO{\lambda DTK} + \textsc{Err}_1 + \textsc{Bias}_1 + \textsc{Bias}_2 + \gamma\sum_{k=1}^K \inner{\qfeat_{k}-\qfeat^\star}{\widehat{b}_k} - \lambda\sum_{k=1}^K\inner{\h{\widehat{q}_k}}{c_k}
\tag{define $\textsc{Reg}_\qfeat = \sum_{k=1}^K\inners{\qfeat_{k}-\qfeat^\star}{\widehat{c}_k-\gamma\widehat{b}_k}$} \\
&= \textsc{Reg}_\qfeat + \tilO{\lambda DTK} + \textsc{Err}_1 + \textsc{Bias}_1 + \textsc{Bias}_2 + \textsc{Bias}_3 + \textsc{Bias}_4 + \gamma\sum_{k=1}^K \inner{\qfeat_{k}-\qfeat^\star}{b_k} - \lambda\sum_{k=1}^K\inner{\h{\widehat{q}_k}}{c_k}
\tag{define $\textsc{Bias}_3 = \gamma\sum_{k=1}^K \inners{\qfeat_{k}}{\widehat{b}_k - b_k}$ and $\textsc{Bias}_4 = \gamma\sum_{k=1}^K \inners{\qfeat^\star}{b_k - \widehat{b}_k }$} \\
&\leq \textsc{Reg}_\qfeat + \tilO{\lambda DTK} + \textsc{Err}_1 + \textsc{Bias}_1 + \textsc{Bias}_2 + \textsc{Bias}_3 + \textsc{Bias}_4 \\
&\hspace{4cm} +2\gamma \sum_{k=1}^K \inner{\widehat{q}_k}{b_k} - \gamma \sum_{k=1}^K\inner{\qfeat^\star}{b_k} - \lambda\sum_{k=1}^K\inner{\h{\widehat{q}_k}}{c_k} \\
&\leq \textsc{Reg}_\qfeat + \tilO{\lambda DTK} + \textsc{Err}_1 + \textsc{Bias}_1 + \textsc{Bias}_2 + \textsc{Bias}_3 + \textsc{Bias}_4 \\
&\hspace{4cm} + (2\gamma-\lambda) \sum_{k=1}^K \inner{\widehat{q}_k}{\h{c_k}} + 2\gamma \sum_{k=1}^K \inner{u'_k-\widehat{q}_k}{\h{c_k}} - \gamma \sum_{k=1}^K\inner{\qfeat^\star}{b_k}
\tag{$\inner{\widehat{q}_k}{b_k}\leq\inner{u_k}{b_k} = \inner{u'_k}{\h{c_k}}$}\\
&= \textsc{Reg}_\qfeat + \tilO{\lambda DTK} + \textsc{Err}_1 + \textsc{Err}_2 + \textsc{Bias}_1 + \textsc{Bias}_2 + \textsc{Bias}_3 + \textsc{Bias}_4 \\
&\hspace{4cm} + (2\gamma-\lambda) \sum_{k=1}^K \inner{\widehat{q}_k}{\h{c_k}} - \gamma \sum_{k=1}^K\inner{\qfeat^\star}{b_k}.
\tag{define $\textsc{Err}_2=2\gamma \sum_{k=1}^K \inner{u'_k-\widehat{q}_k}{\h{c_k}}$}
\end{align*}
The $\textsc{Reg}_\qfeat$ term can be upper bounded by the OMD analysis (see \pref{lem:OMD-increasing-eta}), the four bias terms $\textsc{Bias}_1, \textsc{Bias}_2, \textsc{Bias}_3$, and $\textsc{Bias}_4$ can be bounded using Azuma's or Freedman's inequality (see \pref{lem:BIAS_1_2} and \pref{lem:BIAS_3_4}), and $\textsc{Err}_1+\textsc{Err}_2$ can be bounded by \pref{lem:optimistic-transition-bound} (see \pref{lem:ERR_1_2}).
Combining everything, we obtain
\begin{align*}
R_K &\leq \tilO{\frac{SA}{\eta}} - \frac{\inner{\qfeat^\star}{\rho_K}}{140\eta\ln K} + 40\eta\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{\lambda DTK}\\
&\qquad+ 33\lambda'\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{SAH^3\sqrt{K}} } + \tilO{\frac{S^3A^2}{\lambda'}}\\
&\qquad+ 2C\sqrt{\ln\rbr{\frac{CSA}{\delta}}}\rbr{\frac{\inner{\qfeat^\star}{\rho_K}}{\eta'}+\eta'\inner{\qfeat^\star}{\sum_{k=1}^K b_k}} + 2CH\ln\rbr{\frac{CSA}{\delta}}\inner{\qfeat^\star}{\rho_K}\\
&\qquad+ \rbr{\frac{1}{\eta'}+1}\inner{\qfeat^\star}{\rho_K} + \eta'\inner{\qfeat^\star}{\sum_{k=1}^K b_k} + (2\gamma-\lambda) \sum_{k=1}^K \inner{\widehat{q}_k}{\h{c_k}} - \gamma \sum_{k=1}^K\inner{\qfeat^\star}{b_k} + \tilO{\Qumq}\\
&= \tilO{\frac{SA}{\eta} + \lambda DTK + \lambda'SAH^3\sqrt{K} + \frac{S^3A^2}{\lambda'} + H^3S^3A^2} + \rbr{40\eta + 33\lambda' + 2\gamma - \lambda}\sum_{k=1}^K\inner{\widehat{q}_k}{\h{c_k}}\\
&\qquad+ \rbr{ \frac{2C\sqrt{\ln\rbr{\frac{CSA}{\delta}}} + 1}{\eta'} + 2CH\ln\rbr{\frac{CSA}{\delta}} + 1 - \frac{1}{140\eta\ln K} }\inner{\qfeat^\star}{\rho_K}\\
&\qquad+ \rbr{2C\sqrt{\ln\rbr{\frac{CSA}{\delta}}}\eta' + \eta' - \gamma}\inner{\qfeat^\star}{\sum_{k=1}^K b_k} + (40\eta + 33\lambda')\sum_{k=1}^K\inner{q_k-\widehat{q}_k}{\h{c_k}}.
\end{align*}
Finally, taking $\eta'=\gamma/\rbr{2C\sqrt{\ln\rbr{\frac{CSA}{\delta}}}+1}, \lambda'=\sqrt{\frac{S^3A^2}{DTK}}$, and noticing $\frac{1}{\eta'}\geq 2CH\ln\rbr{\frac{CSA}{\delta}} + 1$, we have the coefficients multiplying $\sum_{k=1}^K\inner{q_k}{\h{c_k}}, \inner{\qfeat^\star}{\rho_K}$, and $\inner{\qfeat^\star}{\sum_{k=1}^K b_k}$ are all non-positive.
Moreover, by \pref{lem:transition-bias}, $\frac{1}{H}(\h{c_k})(s, a)\leq c_k(s, a)$, and $\sum_{k=1}^K\inner{q_k}{\h{c_k}}\leq H^2K$:
\begin{align*}
(40\eta + 33\lambda')\sum_{k=1}^K\inner{q_k-\widehat{q}_k}{\h{c_k}} &= (40\eta + 33\lambda')H\sum_{k=1}^K\inner{q_k-\widehat{q}_k}{\frac{\h{c_k}}{H}} \\
&= \tilO{(40\eta + 33\lambda')H\sqrt{S^2A(H^2K+H^3\sqrt{K})} } = \tilO{H^3S^3A^2}.
\end{align*}
Thus, we arrive at $R_K = \tilO{\sqrt{S^3A^2DTK} + H^3S^3A^2}$.
\end{proof}
\begin{lemma}\label{lem:OMD-increasing-eta}
\pref{alg:bandit-unknown} ensures with probability at least $1-\delta$:
\[
\textsc{Reg}_\qfeat \leq \tilO{\frac{SA}{\eta}} - \frac{\inner{\qfeat^\star}{\rho_K}}{140\eta\ln K} + 40\eta\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{H^2\sqrt{SA}}.
\]
\end{lemma}
\begin{proof}
Denote by $n(s, a)$ the number of times the learning rate for $(s, a)$ increases, such that $\eta_K(s, a)=\eta\beta^{n(s, a)}$,
and by $k_1, \ldots, k_{n(s, a)}$ the episodes where $\eta_k(s, a)$ is increased, such that $\eta_{k_{t}+1}(s, a)=\beta\cdot\eta_{k_{t}}(s, a)$.
Since $\rho_1(s,a) = 2T$ and
\[
\rho_1(s, a)2^{n(s, a)-1}\leq \cdots \leq
\rho_{k_{n(s,a)}}(s, a) \leq \frac{1}{u_{k_{n(s,a)}+1}(s,a)}
\leq \frac{1}{\widehat{q}_{k_{n(s,a)}+1}(s,a)} \leq TK^4,
\]
we have $n(s, a)\leq 1+ \log_2\frac{K^4}{2}\leq 7\log_2K$.
Therefore, $\eta_K(s, a)\leq \eta e^{\frac{7\log_2K}{7\ln K}}\leq 5\eta$.
Now, notice that $\sum_hu'_k(s, a, h)\leq u_k(s, a)$ by definition and $\gamma\leq\frac{1}{H}$ for large enough $K$ ($K \gtrsim S^3A^2H^2$).
Thus,
\[
\gamma\widehat{b}_k(s,a) \leq \frac{\gamma H \sum_h u'_k(s,a,h)\widehat{c}_k(s,a)}{u_k(s,a)}
\leq \gamma H \widehat{c}_k(s,a) \leq \widehat{c}_k(s,a).
\]
This means that the cost $\widehat{c}_k - \gamma\widehat{b}_k$ we feed to OMD is always non-negative, and thus by the same argument of~\citep[Lemma~12]{agarwal2017corralling}, we have
\begin{align*}
&\textsc{Reg}_\qfeat = \sum_{k=1}^K\inner{\qfeat_k-\qfeat^\star}{\widehat{c}_k-\gamma\widehat{b}_k}\\
&\leq \sum_{k=1}^K D_{\psi_k}(\qfeat^\star, \qfeat_k) - D_{\psi_k}(\qfeat^\star, \qfeat_{k+1}) + \sum_{k=1}^K\sumsa\eta_k(s, a)\qfeat^2_k(s, a)(\widehat{c}_k(s, a) - \gamma\widehat{b}_k(s, a))^2\\
&\leq D_{\psi_1}(\qfeat^\star, \qfeat_1) + \sum_{k=1}^{K-1}\rbr{D_{\psi_{k+1}}(\qfeat^\star, \qfeat_{k+1}) - D_{\psi_k}(\qfeat^\star, \qfeat_{k+1})} + 5\eta\sum_{k=1}^K\sumsa \qfeat_k^2(s, a)\widehat{c}_k^2(s, a)\\
&\leq D_{\psi_1}(\qfeat^\star, \qfeat_1) + \sum_{k=1}^{K-1}\rbr{D_{\psi_{k+1}}(\qfeat^\star, \qfeat_{k+1}) - D_{\psi_k}(\qfeat^\star, \qfeat_{k+1})} + 20\eta\sum_{k=1}^K\sumsa \widehat{q}_k^2(s, a)\widehat{c}_k^2(s, a)\\
&\leq D_{\psi_1}(\qfeat^\star, \qfeat_1) + \sum_{k=1}^{K-1}\rbr{D_{\psi_{k+1}}(\qfeat^\star, \qfeat_{k+1}) - D_{\psi_k}(\qfeat^\star, \qfeat_{k+1})} + 20\eta\sum_{k=1}^K\sumsa\widetilde{N}_k^2(s, a)c_k^2(s, a). \tag{$u_k(s, a)\geq \widehat{q}_k(s, a)$}
\end{align*}
For the first term, since $\qfeat_1$ minimizes $\psi_1$ and thus $\inner{\nabla\regz_1(\qfeat_1)}{\qfeat^\star-\qfeat_1}\geq 0$, we have
\begin{align*}
D_{\psi_1}(\qfeat^\star, \qfeat_1) \leq \psi_1(\qfeat^\star) - \psi_1(\qfeat_1) = \frac{1}{\eta}\sumsa\ln\frac{\qfeat_1(s, a)}{\qfeat^\star(s, a)}
\leq \frac{1}{\eta}\sumsa\ln\frac{2H}{q^\star(s,a)} = \tilO{\frac{SA}{\eta}}.
\end{align*}
For the second term, we define $\chi(y) = y-1-\ln y$ and proceed similarly to~\citep{agarwal2017corralling}:
\begin{align*}
&\sum_{k=1}^{K-1} D_{\psi_{k+1}}(\qfeat^\star, \qfeat_{k+1}) - D_{\psi_k}(\qfeat^\star, \qfeat_{k+1})\\
&= \sum_{k=1}^{K-1}\sumsa\rbr{\frac{1}{\eta_{k+1}(s, a)} - \frac{1}{\eta_k(s, a)}}\chi\rbr{\frac{\qfeat^\star(s, a)}{\qfeat_{k+1}(s, a)}}\\
&\leq\sumsa\frac{1-\beta}{\eta\beta^{n(s, a)}} \chi\rbr{\frac{\qfeat^\star(s, a)}{\qfeat_{k_{n(s, a)}+1}(s, a)}}\\
&=\sumsa\frac{1-\beta}{\eta\beta^{n(s, a)}}\rbr{\frac{\qfeat^\star(s, a)}{\qfeat_{k_{n(s, a)}+1}(s, a)} - 1 - \ln\frac{\qfeat^\star(s, a)}{\qfeat_{k_{n(s, a)}+1}(s, a)}}\\
&\leq -\frac{1}{35\eta\ln K}\sumsa\rbr{ \qfeat^\star(s, a)\frac{\rho_K(s, a)}{4} - 1 - \ln\frac{\qfeat^\star(s, a)}{\qfeat_{k_{n(s, a)}+1}(s, a)} }\\
&\leq \frac{SA(1 + 6\ln K)}{35\eta\ln K} - \frac{\inner{\qfeat^\star}{\rho_K}}{140\eta\ln K} = \tilO{\frac{SA}{\eta}}- \frac{\inner{\qfeat^\star}{\rho_K}}{140\eta\ln K},
\end{align*}
where in the last two lines we use the facts $1-\beta \leq -\frac{1}{7\ln K}, \beta^{n(s, a)} \leq 5$, $\rho_K(s, a)=\frac{2}{u_{k_{n(s, a)}+1}(s, a)}$, and $\ln\frac{\qfeat^\star(s, a)}{\qfeat_{k_{n(s, a)}+1}(s, a)} \leq \ln(HTK^4) \leq 6\ln K$.
Finally, for the third term,
since $\sumsa \widetilde{N}_k^2(s, a)c_k^2(s, a) \leq \left(\sumsa \widetilde{N}_k(s, a)\right)^2 \leq H^2$, we apply Azuma's inequality (\pref{lem:azuma}) and obtain, with probability at least $1-\delta$:
\begin{align*}
\eta\sum_{k=1}^K\sumsa \widetilde{N}_k^2(s, a)c_k^2(s, a) &\leq \eta\sum_{k=1}^K\field{E}_k\sbr{\sumsa \widetilde{N}_k^2(s,a )c_k^2(s, a)} + \tilO{\eta H^2\sqrt{K}} \\
&\leq \eta\sum_{k=1}^K\field{E}_k\sbr{\inner{\widetilde{N}_k}{c_k}^2} + \tilO{H^2\sqrt{SA}} \tag{$\sum_i a_i^2\leq (\sum_i a_i)^2$ for $a_i>0$ and $\eta=\sqrt{\frac{SA}{DTK}}$}\\
&\leq 2\eta\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{H^2\sqrt{SA}} \tag{\pref{lem:deviation_loop_free}}.
\end{align*}
Combining everything shows
\begin{align*}
\textsc{Reg}_\qfeat \leq \tilO{\frac{SA}{\eta}} - \frac{\inner{\qfeat^\star}{\rho_K}}{140\eta\ln K} + 40\eta\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{H^2\sqrt{SA}}.
\end{align*}
finishing the proof.
\end{proof}
\begin{lemma}\label{lem:BIAS_1_2}
For any $\eta' > 0$,
with probability at least $1-\delta$,
\[
\textsc{Bias}_1 \leq 2C\sqrt{\ln\rbr{\frac{CSA}{\delta}}}\rbr{\frac{\inner{\qfeat^\star}{\rho_K}}{\eta'}+\eta'\inner{\qfeat^\star}{\sum_{k=1}^K b_k}} + 2CH\ln\rbr{\frac{CSA}{\delta}}\inner{\qfeat^\star}{\rho_K}.
\]
Also, with probability at least $1-5\delta$,
$
\textsc{Bias}_2 = \tilO{ S^3A^2H^2 }.
$
\end{lemma}
\begin{proof}
Define $X_k(s, a) = \widehat{c}_k(s, a) - \field{E}_k[\widehat{c}_k(s, a)]$.
Note that $X_k(s, a)\leq \frac{H}{u_k(s,a)} \leq HTK^4$, and
\begin{align*}
\sum_{k=1}^K\field{E}_k \sbr{X_k^2(s, a)} &\leq \sum_{k=1}^K \frac{\field{E}_k\sbr{\widetilde{N}_k^2(s, a)c^2_k(s, a)}}{u^2_k(s, a)} \\
&\leq \rho_k(s,a) \sum_{k=1}^K \frac{\field{E}_k\sbr{\widetilde{N}_k^2(s, a)c^2_k(s, a)}}{u_k(s, a)} \\
&\leq 2\rho_k(s,a) \sum_{k=1}^K b_k(s,a). \tag{\pref{lem:simpler_form_N_times_c}}
\end{align*}
Therefore, by applying a strengthened Freedman's inequality (\pref{lem:extended-freedman}) with $b = HTK^4$, $B_k=H\rho_k(s, a), \max_kB_k=H\rho_K(s, a)$, and $V = 2\rho_k(s,a) \sum_{k=1}^K b_k(s,a)$, we have with probability at least $1-\delta/(SA)$,
\begin{align*}
&\sum_{k=1}^K \widehat{c}_k(s, a) - \field{E}_k[\widehat{c}_k(s, a)] \\
&\leq C\rbr{ 4\sqrt{\rho_K(s, a)\sum_{k=1}^K b_k(s, a)\ln\rbr{\frac{CSA}{\delta}}} + 2H\rho_K(s, a)\ln\rbr{\frac{CSA}{\delta}} } \\
&\leq 2C\sqrt{\ln\rbr{\frac{CSA}{\delta}}}\rbr{\frac{\rho_K(s,a)}{\eta'}+\eta'\sum_{k=1}^K b_k(s,a)} + 2CH\rho_K(s, a)\ln\rbr{\frac{CSA}{\delta}},
\end{align*}
where the last step is by AM-GM inequality.
Further using a union bound shows that the above holds for all $(s, a)\in\widetilde{\SA}$ with probability at least $1-\delta$.
Thus, by $\field{E}_k[\widehat{c}_k(s, a)]\leq c_k(s, a)$,
\begin{align*}
\textsc{Bias}_1 &= \sum_{k=1}^K\inner{\qfeat^\star}{\widehat{c}_k-c_k} \leq \sum_{k=1}^K\inner{\qfeat^\star}{\widehat{c}_k-\field{E}_k[\widehat{c}_k]} \\
&\leq 2C\sqrt{\ln\rbr{\frac{CSA}{\delta}}}\rbr{\frac{\inner{\qfeat^\star}{\rho_K}}{\eta'}+\eta'\inner{\qfeat^\star}{\sum_{k=1}^K b_k}} + 2CH\ln\rbr{\frac{CSA}{\delta}}\inner{\qfeat^\star}{\rho_K}.
\end{align*}
To bound $\textsc{Bias}_2$, simply note that $\left|\inner{\h{\widehat{q}_k}}{\field{E}_k[\widehat{c}_k]-\widehat{c}_k}\right| \leq 2H^2$ and apply Azuma's inequality (\pref{lem:azuma}): with probability $1-5\delta$,
\begin{align*}
\textsc{Bias}_2 &= \lambda\sum_{k=1}^K\inner{\h{\widehat{q}_k}}{c_k - \field{E}_k[\widehat{c}_k]} + \lambda\sum_{k=1}^K\inner{\h{\widehat{q}_k}}{\field{E}_k[\widehat{c}_k] - \widehat{c}_k} \\
&\leq \lambda H\sum_{k=1}^K\inner{u_k-q_k}{c_k} + \tilO{\lambda H^2\sqrt{K}} \\
&= \tilO{\lambda H\sqrt{S^3A^2\sum_{k=1}^K\inner{q_k}{\h{c_k}}} + H^2\sqrt{S^3A^2} } = \tilO{S^3A^2H^2},\tag{$\lambda=\tilO{\sqrt{\frac{S^3A^2}{DTK}}}, \sum_{k=1}^K\inner{q_k}{\h{c_k}}=\bigO{H^2K}$ and \pref{lem:optimistic-transition-bound}}
\end{align*}
where in the second line we use:
\begin{align*}
\sum_{k=1}^K\inner{\h{\widehat{q}_k}}{c_k-\field{E}_k[\widehat{c}_k]} &= \sum_{k=1}^K\sum_{(s, a), h}h\cdot\widehat{q}_k(s, a, h)c_k(s, a)\rbr{1 - \frac{q_k(s, a)}{u_k(s, a)} }\\
&\leq H\sum_{k=1}^K\sum_{(s, a), h}\widehat{q}_k(s, a, h)c_k(s, a)\frac{u_k(s, a)-q_k(s, a)}{u_k(s, a)}\\
&\leq H\sum_{k=1}^K\inner{u_k-q_k}{c_k}. \tag{$u_k(s, a)\geq \max\{q_k(s, a), \widehat{q}_k(s, a)\}$}
\end{align*}
\end{proof}
\begin{lemma}\label{lem:BIAS_3_4}
With probability at least $1-\delta$, we have $\textsc{Bias}_3 = \tilO{H^2(SA)^{3/2}}$.
Also, for any $\eta' > 0$,
with probability at least $1-\delta$, we have
\[
\textsc{Bias}_4 \leq \rbr{\frac{1}{\eta'}+1}\inner{\qfeat^\star}{\rho_K} + \eta'\inner{\qfeat^\star}{\sum_{k=1}^K b_k} + \tilO{1}.
\]
\end{lemma}
\begin{proof}
To bound $\textsc{Bias}_3$, simply note that $\field{E}_k[\widehat{b}_k(s, a)]\leq b_k(s, a)$,
\begin{align*}
&\left|\inner{\phi_k}{\widehat{b}_k-\field{E}_k[\widehat{b}_k]}\right| \leq \abr{\inner{\phi_k}{\widehat{b}_k}} + \abr{\inner{\phi_k}{\field{E}_k\widehat{b}_k}}\\
&\leq 2\abr{\sumsa\sum_hh\cdot u'_k(s, a, h)\widehat{c}_k(s, a)} + 2\abr{\sumsa\sum_hh\cdot u'_k(s, a, h)\field{E}_k[\widehat{c}_k(s, a)]} \tag{$\phi_k(s, a)\leq 2\widehat{q}_k(s, a)\leq 2u_k(s, a)$} \\
&\leq 2H\abr{\sumsa\widetilde{N}_k(s, a)c_k(s, a)} + 2H\abr{\sumsa\sum_hu'_k(s, a, h)c_k(s, a)} \leq 4SAH^2, \tag{$\sum_hu'_k(s, a, h)\leq u_k(s, a), q_k(s, a)\leq u_k(s, a), \sumsa \widetilde{N}_k(s, a)\leq H$, and $\sum_hu'_k(s, a, h)\leq H$}
\end{align*}
$\gamma=\tilO{\sqrt{\frac{SA}{DTK}}}$, and apply Azuma's inequality (\pref{lem:azuma}): with probability at least $1-\delta$,
\begin{align*}
\textsc{Bias}_3 = \gamma\sum_{k=1}^K \inners{\qfeat_{k}}{\widehat{b}_k - b_k} \leq \gamma\sum_{k=1}^K\inner{\phi_k}{\widehat{b}_k - \field{E}_k[\widehat{b}_k]} = \tilO{\gamma SAH^2\sqrt{K}} = \tilO{H^2(SA)^{3/2}}.
\end{align*}
To bound $\textsc{Bias}_4 = \gamma\sum_{k=1}^K\inner{\qfeat^\star}{b_k - \field{E}_k[\widehat{b}_k]} + \gamma\sum_{k=1}^K\inner{\qfeat^\star}{\field{E}_k[\widehat{b}_k] - \widehat{b}_k}$, first note that
\begin{align*}
\gamma\sum_{k=1}^K\inner{\qfeat^\star}{b_k - \field{E}_k[\widehat{b}_k]} &\leq \gamma \sum_{k=1}^K\sumsa\inner{\qfeat^\star}{b_k} = \gamma\sum_{k=1}^K\sumsa \qfeat^\star(s, a)\frac{\sum_{h=1}^Hh\cdot u'_k(s, a, h)}{u_k(s, a)}c_k(s, a)\\
&\leq \gamma H\sum_{k=1}^K\sumsa\qfeat^\star(s, a) \leq \sum_{k=1}^K\inner{\qfeat^\star}{\rho_K}. \tag{$\gamma H\leq 1$ and $\rho_K(s, a)\geq\rho_1(s, a)\geq 1$}
\end{align*}
Then, we note that $\field{E}_k[\widehat{b}_k(s, a)] - \widehat{b}_k(s,a) \leq \field{E}_k[\widehat{b}_k(s,a)] \leq H$, and
\begin{align*}
\field{E}_k\sbr{\rbr{\field{E}_k[\widehat{b}_k(s, a)] - \widehat{b}_k(s, a)}^2} &\leq \field{E}_k[\widehat{b}_k^2(s, a)] \leq \frac{\field{E}_k\sbr{(\sum_h h\cdot u'_k(s, a, h)\widehat{c}_k(s, a))^2}}{u_k^2(s, a)}\\
&\leq \frac{H^2\field{E}_k\sbr{\widetilde{N}_k^2(s, a)c_k^2(s, a)}}{u^2_k(s, a)} \tag{$h\leq H$ and $\sum_hu'_k(s, a, h)\leq u_k(s, a)$}\\
&\leq H^2\rho_K(s,a)\frac{\field{E}_k\sbr{\widetilde{N}_k^2(s, a)c_k^2(s, a)}}{u_k(s, a)} \\
& \leq 2H^2\rho_K(s, a)b_k(s, a) \tag{\pref{lem:simpler_form_N_times_c}}.
\end{align*}
Hence, applying a strengthened Freedman's inequality (\pref{lem:extended-freedman}) with $b = B_k = H$, $V = 2H^2\rho_K(s, a)\sum_{k=1}^K b_k(s, a)$, and $C' = \lceil \log_2 H \rceil \lceil \log_2 (H^2K) \rceil$, we have
with probability at least $1-\delta/(SA)$,
\begin{align*}
&\sum_{k=1}^K b_k(s, a) - \widehat{b}_k(s, a) \\
&\leq 4C'H\sqrt{\ln\rbr{\frac{C'SA}{\delta}}}\sqrt{\rho_K(s,a) \sum_{k=1}^K b_k(s,a)} + 2C'H\ln\rbr{\frac{C'SA}{\delta}} \\
&= 2C'H\sqrt{\ln\rbr{\frac{C'SA}{\delta}}}\rbr{\frac{\rho_K(s,a)}{\eta'} + \eta' \sum_{k=1}^K b_k(s,a)} + 2C'H\ln\rbr{\frac{C'SA}{\delta}},
\end{align*}
where the last step is by AM-GM inequality.
Finally, applying a union bound shows that the above holds for all $(s,a)\in\widetilde{\SA}$ with probability at least $1-\delta$ and thus
\begin{align*}
\textsc{Bias}_4 \leq \gamma\sum_{k=1}^K\inner{\qfeat^\star}{b_k - \widehat{b}_k} \leq \rbr{\frac{1}{\eta'}+1}\inner{\qfeat^\star}{\rho_K} +\eta'\inner{\qfeat^\star}{\sum_{k=1}^K b_k} + \tilO{1},
\end{align*}
where we bound $2\gamma C'H\sqrt{\ln\rbr{\frac{C'SA}{\delta}}}$ by $\tilO{1}$ since $\gamma H\leq 1$ when $K$ is large enough ($K \gtrsim S^3A^2H^2$).
\end{proof}
\begin{lemma}\label{lem:ERR_1_2}
For any $\lambda' \in (0, \frac{2}{H}]$, with probability at least $1-9\delta$ we have
\begin{align*}
\textsc{Err}_1 \leq 33\lambda'\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{SAH^3\sqrt{K}} } + \frac{34S^3A^2\ln^2\rbr{\frac{HKS^2A^2}{\delta}}}{\lambda'} + \tilO{\Qumq}.
\end{align*}
Also, with probability at least $1-8\delta$, we have $\textsc{Err}_2=\tilO{ H^3S^3A^2 }$,
\end{lemma}
\begin{proof}
We write $\textsc{Err}_1$ as
\begin{align*}
\textsc{Err}_1 &= \sum_{k=1}^K\inner{u_k-\widehat{q}_k}{\widehat{c}_k - \field{E}_k[\widehat{c}_k]} + \sum_{k=1}^K\inner{u_k-\widehat{q}_k}{\field{E}_k[\widehat{c}_k]}.
\end{align*}
For the first term, note that $\inner{u_k-\widehat{q}_k}{\widehat{c}_k}\leq\sumsa\widetilde{N}_k(s, a)\leq H$, and
\begin{align*}
\field{E}_k\sbr{\rbr{\sumsa (u_k(s, a)-\widehat{q}_k(s, a))\frac{\widetilde{N}_k(s, a)}{u_k(s, a)}c_k(s, a) }^2} \leq \field{E}_k\sbr{\rbr{\sumsa \widetilde{N}_k(s, a)c_k(s, a)}^2 } \leq 2\inner{q_k}{\h{c_k}}.
\end{align*}
Hence, by Freedman's inequality (\pref{lem:freedman}), for any $0<\lambda'\leq\frac{2}{H}$, with probability at least $1-\delta$ we have:
\begin{align*}
\sum_{k=1}^K\inner{u_k-\widehat{q}_k}{\widehat{c}_k - \field{E}_k[\widehat{c}_k]} = \lambda'\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \frac{2\ln(1/\delta)}{\lambda'}.
\end{align*}
For the second term, with probability at least $1-8\delta$:
\begin{align*}
\sum_{k=1}^K\inner{u_k-\widehat{q}_k}{\field{E}_k[\widehat{c}_k]} &\leq \sum_{k=1}^K\inner{u_k-\widehat{q}_k}{c_k} = \sum_{k=1}^K\inner{u_k-q_k}{c_k} + \inner{q_k-\widehat{q}_k}{c_k}\\
&\leq 64\sqrt{S^3A^2\ln^2\rbr{\frac{HKS^2A^2}{\delta}}\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{SAH^3\sqrt{K}}}} + \tilO{\Qumq}. \tag{\pref{lem:optimistic-transition-bound} and \pref{lem:transition-bias}}\\
&\leq 32\lambda'\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{SAH^3\sqrt{K}} } + \frac{32S^3A^2\ln^2\rbr{\frac{HKS^2A^2}{\delta}}}{\lambda'} + \tilO{\Qumq}. \tag{AM-GM inequality}
\end{align*}
Putting everything together, we have:
\begin{align*}
\textsc{Err}_1 \leq 33\lambda'\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{SAH^3\sqrt{K}} } + \frac{34S^3A^2\ln^2\rbr{\frac{HKS^2A^2}{\delta}}}{\lambda'} + \tilO{\Qumq}.
\end{align*}
For $\textsc{Err}_2$, use the bound for $\sum_{k=1}^K\inner{u_k-\widehat{q}_k}{c_k}$ from above, we have with probability at least $1-8\delta$:
\begin{align*}
\textsc{Err}_2 &= 2\gamma \sum_{k=1}^K \inner{u'_k-\widehat{q}_k}{\h{c_k}} \leq 2\gamma H\sum_{k=1}^K\inner{u_k-\widehat{q}_k}{c_k}\\
&= \tilO{\gamma H\sqrt{S^3A^2\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{SAH^3\sqrt{K}} }} + \gamma H^4S^3A^2 } \tag{\pref{lem:optimistic-transition-bound} and \pref{lem:transition-bias}}\\
&= \tilO{\gamma H\sqrt{S^3A^2(H^2K + SAH^3\sqrt{K})}} = \tilO{ H^3S^3A^2 }. \tag{$\gamma=\tilO{\sqrt{\frac{SA}{DTK}}}$ and $\gamma H\leq 1$}
\end{align*}
\end{proof}
\begin{lemma}
\label{lem:simpler_form_N_times_c}
For any episode $k$ and $(s, a)\in\widetilde{\SA}$, we have $\field{E}_k \sbr{\widetilde{N}_k(s, a)^2c_k(s, a)^2} \leq 2u_k(s, a)b_k(s, a)$.
\end{lemma}
\begin{proof}
We use the inequality $(\sum_{i=1}^I a_i)^2\leq 2\sum_{i}a_i(\sum_{i'= i}^I a_{i'})$:
\begin{align*}
\field{E}_k\sbr{ \widetilde{N}_k(s, a)^2c_k(s, a)^2} &\leq \field{E}_k\sbr{\rbr{ \sum_{h=1}^H \widetilde{N}_k(s, a, h)}^2c_k(s, a) }\\
&\leq 2\field{E}_k\sbr{\rbr{\sum_{h=1}^H \widetilde{N}_k(s, a, h)}\rbr{\sum_{h'\geq h}^H \widetilde{N}_k(s, a, h')c_k(s, a)}}\\
&\leq 2\field{E}_k\sbr{\sum_{h=1}^H\sum_{h'\geq h}^H \widetilde{N}_k(s, a, h')c_k(s, a)} \tag{$\widetilde{N}_k(s, a, h)\leq 1$}\\
&= 2 \sum_{h=1}^H\sum_{h'\geq h}^H q_k(s, a, h')c_k(s, a) = 2\sum_{h=1}^Hh \cdot q_k(s, a, h)c_k(s, a)\\
&\leq 2\sum_{h=1}^H h\cdot u'_k(s, a, h)c_k(s, a) = 2u_k(s, a)b_k(s, a),
\end{align*}
where the last step is by the definition of $b_k$.
\end{proof}
\subsection{Omitted Details for Transition Estimation}
\label{sec:trans-est}
In this subsection, we introduce a sub-procedure $\text{TransEst}$ (\pref{alg:update_transition}) used in all algorithms proposed in this paper for transition estimation. It takes the learner's trajectory as input and outputs a confidence set ${\mathcal{P}}_{k+1}$ (and additionally the counter $\widetilde{N}_k$ for the bandit setting) at the end of episode $k$.
We first introduce some notations.
Denote by $\tcs$ the set of valid transitions $\widetilde{P}$ for $\widetilde{M}$ based on its layer structure.
Also define the entries of unknown transition in $\widetilde{M}$ as $\unk=\Gamma\times({\mathcal{S}}\cup \{g\})\times[H_1-1]$, such that $\widetilde{P}(s'|s, a, h)=P(s'|s, a),\forall (s, a), s', h\in\unk$.
We implement the confidence sets of transition function by maintaining a separate (empirical) Bernstein confidence bound for each state, action, and next state in $M$ around its empirical estimate:
\begin{align*}
\epsilon_k(s, a, s') &= 4\sqrt{\bar{P}_k(s'| s, a)\A_k(s, a)} + 28\A_k(s, a),
\end{align*}
for any $(s, a)\in\Gamma, s'\in{\mathcal{S}}\cup\{g\}$, where $\bar{P}_k(s'| s, a) = \frac{\M_k(s, a, s')}{\N^+_k(s, a)}$ is the empirical transition estimation, $\N^+_k(s, a)=\max\{1, \N_k(s, a)\}$, $\N_k(s, a)$ is the number of visits to $(s, a)$ before episode $k$, $\M_k(s, a. s')$ is the number of visits to $(s, a), s'$ before episode $k$, and $\A_k(s, a) = \frac{\ln(\frac{HKSA}{\delta})}{\N^+_k(s, a)}$.
We then define the confidence set in episode $k$ as follows:
\begin{align}
\label{eq:conf-set}
{\mathcal{P}}_k &= \Big\{\widehat{P}\in \tcs: \lvert \widehat{P}(s'| s, a, h) - \bar{P}_k(s'| s,a) \rvert \leq \epsilon_k(s, a, s'), \forall(s, a), s', h\in \unk \Big\}.
\end{align}
\begin{algorithm}[t]
\caption{$\text{TransEst}(\N, \mathbf{M}, \delta, H_1, H_2, \tau)$}
\label{alg:update_transition}
\textbf{Input:} counters $\N, \mathbf{M}$, confidence parameter $\delta$, horizon parameters $H_1, H_2$, trajectory $\tau=\{s^1, a^1, s^2, a^2, \ldots, a^{h-1}, s^h\}$.
\textbf{Initialization:} $\widetilde{N}(s, a)=0$, $\N_{k+1}(s, a)=\N_{k}(s, a)$, and $\mathbf{M}_{k+1}(s, a)=\mathbf{M}_{k}(s, a)$ for any $(s, a)\in\widetilde{\SA}$.
\For{$t=1,\ldots, h-1$}{
Update counters: $\N_{k+1}(s^t, a^t)\leftarrow \N_k(s^t, a^t)+1$, $\mathbf{M}_{k+1}(s^t, a^t, s^{t+1})\leftarrow \mathbf{M}_k(s^t, a^t, s^{t+1}) + 1$, $\widetilde{N}(s^t, a^t)\leftarrow\widetilde{N}(s^t, a^t)+1$.
}
Compute confidence set ${\mathcal{P}}_{k+1}$ defined in \pref{eq:conf-set}.
\If{$s^h\neq g$}{
$\widetilde{N}(s_f, a_f)\leftarrow H_2$.
}
\textbf{Return:} ${\mathcal{P}}_{k+1}$ (and additionally $\widetilde{N}_k$ for the bandit setting).
\end{algorithm}
Next, we introduce several lemmas useful for bounding the bias of transition estimation.
\begin{lemma}{(Lemma 4.2 in \citep{cohen2020near})}\label{lem:transition-interval}
With probability at least $1-\delta$, $P\in{\mathcal{P}}_k$ for all $k$.
\end{lemma}
\begin{lemma}
\label{lem:transition-bound}
Assume $\widehat{P}\in{\mathcal{P}}_k$.
Then, under the event of \pref{lem:transition-interval} we have
\begin{align*}
|\widehat{P}(s'|s, a, h) - P(s'|s, a, h)| \leq \field{I}\{(s, a), s', h\in U\}\rbr{ 8\sqrt{P(s'|s, a, h)\A_k(s, a)} + 136\A_k(s, a)} \stackrel{\rm def}{=} \epsilon^\star_k(s, a, s').
\end{align*}
\end{lemma}
\begin{proof}
First note that $\widehat{P}(s'|s, a, h)=P(s'|s, a, h)$ when $(s, a), s', h\notin \unk$.
When $(s, a), s', h\in\unk$, the result follows from $P(s'|s, a, h)=P(s'|s, a)$, definition of $\epsilon_k(s, a, s')$ and \citep[Lemma B.13]{cohen2020near}.
\end{proof}
\begin{lemma}\label{lem:om-diff}
For any occupancy measure $q$ and $\widehat{q}$ induced by the same policy $\pi$ but different transition functions $P$ and $\widehat{P}$ respectively, we have:
\begin{align*}
\widehat{q}(s, a, h) - q(s, a, h) = \sum_{(s', a'), s''}\sum_{m=1}^{h-1}q(s', a', m)(\widehat{P}(s''|s', a', m) - P(s''|s', a', m))\widehat{q}_{(s'', m+1)}(s, a, h),
\end{align*}
where $\widehat{q}_{(s'', m+1)}$ is the occupancy measure with respect to $\pi$, $\widehat{P}$, and initial state $(s'', m+1)$.
As a result, under the event of \pref{lem:transition-interval},
\begin{align*}
\abr{\inner{\widehat{q}-q}{c}} &= \abr{\sum_{(s, a), s', h}q(s, a, h)(\widehat{P}(s'|s, a, h) - P(s'|s, a, h))J^{\widehat{P}, \pi}(s', h+1)}\\
&\leq \inner{|\widehat{q}-q|}{c} \leq H\sum_{(s, a), s', h\in\unk}q(s, a, h)\epsilon^\star_k(s, a, s').
\end{align*}
\end{lemma}
\begin{proof}
We prove the first statement by induction on $h$.
Denote by $q_{(s, h)}$ the occupancy measure of $\pi_q$ with initial state $(s, h)$.
When $h=1$, the statement is true by $q(s, a, h)=\widehat{q}(s, a, h)=\pi(a|s, 1)\field{I}\{s=s_0\}$.
For the induction step with $h>1$:
\begin{align*}
&\widehat{q}(s, a, h) - q(s, a, h) = \pi(a|s, h)(\widehat{q}(s, h) - q(s, h))\\
&= \pi(a|s, h)\sumsa[s', a']\rbr{\widehat{P}(s|s', a', h-1)\widehat{q}(s', a', h-1) - P(s|s', a', h-1)q(s', a', h-1)} \tag{$q(s, h)=\sumsa[s', a']P(s|s', a', h-1)q(s', a', h-1)$}\\
&= \underbrace{\pi(a|s, h)\sumsa[s', a']\widehat{P}(s|s', a', h-1)\rbr{\widehat{q}(s', a', h-1) - q(s', a', h-1)}}_{\chi_1}\\
&\qquad\qquad\qquad + \underbrace{\pi(a|s, h)\sumsa[s', a']q(s', a', h-1)\rbr{\widehat{P}(s|s', a', h-1)-P(s|s', a', h-1)}}_{\chi_2}\\
\end{align*}
For $\chi_1$, by the induction step, we have:
\[
\widehat{q}(s', a', h-1) - q(s', a', h-1) = \sum_{(s'', a''), s'''}\sum_{m=1}^{h-2}q(s'', a'', m)(\widehat{P}-P)(s'''|s'', a'', m)\widehat{q}_{(s''', m+1)}(s', a', h-1).
\]
Thus, by $\sum_{(s', a')}\widehat{q}_{(s''', m+1)}(s', a', h-1)\widehat{P}(s|s', a', h-1)\pi(a|s, h)=\widehat{q}_{(s''', m+1)}(s, a, h)$:
\begin{align*}
\chi_1 &= \pi(a|s, h)\sumsa[s', a']\widehat{P}(s|s', a', h-1)\sum_{(s'', a''), s'''}\sum_{m=1}^{h-2}q(s'', a'', m)(\widehat{P}-P)(s'''|s'', a'', m)\widehat{q}_{(s''', m+1)}(s', a', h-1)\\
&= \sum_{(s'', a''), s'''}\sum_{m=1}^{h-2}q(s'', a'', m)\rbr{\widehat{P}(s'''|s'', a'', m) - P(s'''|s'', a'', m)}\widehat{q}_{(s''', m+1)}(s, a, h).
\end{align*}
For $\chi_2$, note that $\pi(a|s'', h)\field{I}\{s''=s\}=q_{(s'', h)}(s, a, h)$.
Thus,
\begin{align*}
&\pi(a|s, h)\rbr{\widehat{P}(s|s', a', h-1)-P(s|s', a', h-1)}\\
&= \sum_{s''}\pi(a|s'', h)\field{I}\{s''=s\}\rbr{\widehat{P}(s''|s', a', h-1)-P(s''|s', a', h-1)}\\
&= \sum_{s''}q_{(s'', h)}(s, a, h)\rbr{\widehat{P}(s''|s', a', h-1)-P(s''|s', a', h-1)}.\\
\end{align*}
and,
\begin{align*}
\chi_2 = \sum_{(s', a'), s''}q(s', a', h-1)(\widehat{P}-P)(s''|s', a', h-1)q_{(s'', h)}(s, a, h).
\end{align*}
Plugging these back and changing variables $(s'', a''), s'''$ in $\chi_1$ to $(s', a'), s''$, we get:
\begin{align*}
\widehat{q}(s, a, h) - q(s, a, h) &= \chi_1 + \chi_2\\
&= \sum_{(s', a'), s''}\sum_{m=1}^{h-1}q(s', a', m)\rbr{\widehat{P}(s''|s', a', m) - P(s''|s', a', m)}\widehat{q}_{(s'', m+1)}(s, a, h).
\end{align*}
This completes the proof of the first statement.
For the second statement,
\begin{align*}
\inner{\widehat{q}-q}{c} &= \sum_{(s, a), h}\sum_{(s', a'), s''}\sum_{m=1}^{h-1}q(s', a', m)(\widehat{P}(s''|s', a', m) - P(s''|s', a', m))\widehat{q}_{(s'', m+1)}(s, a, h)c(s, a, h)\\
&= \sum_{m=1}^H\sum_{(s', a'), s''}q(s', a', m)(\widehat{P}(s''|s', a', m) - P(s''|s', a', m))\sum_{(s, a), h > m}\widehat{q}_{(s'', m+1)}(s, a, h)c(s, a, h)\\
&= \sum_{m=1}^H\sum_{(s', a'), s''}q(s', a', m)(\widehat{P}(s''|s', a', m) - P(s''|s', a', m))J^{\widehat{P}, \pi}(s'', m+1)\\
&= \sum_{h=1}^H\sum_{(s, a), s'}q(s, a, h)(\widehat{P}(s'|s, a, h) - P(s'|s, a, h))J^{\widehat{P}, \pi}(s', h+1). \tag{change of variables}
\end{align*}
Similarly, $\inner{|\widehat{q}-q|}{c} \leq \sum_{h=1}^H\sum_{(s, a), s'}q(s, a, h)\abr{\widehat{P}(s'|s, a, h) - P(s'|s, a, h)}J^{\widehat{P}, \pi}(s', h+1)$.
Applying the fact $J^{\widehat{P}, \pi}(s', h+1)\leq H$ and \pref{lem:transition-bound} completes the proof.
\end{proof}
\begin{lemma}\label{lem:transition-sum}
With probability at least $1-\delta$ we have: $\sumsa \sum_{k=1}^K\frac{q_k(s, a)}{\N^+_k(s, a)} = \tilO{SAH}$.
\end{lemma}
\begin{proof}
We first prove that for each $(s,a)$,
$
\sum_{k=1}^K\frac{\widetilde{N}_k(s, a)}{\N^+_k(s, a)} = \ensuremath{\mathcal{O}}(\ln (HK) + H)
$.
Indeed, we have
\begin{align*}
\sum_{k=1}^K\frac{\widetilde{N}_k(s, a)}{\N^+_k(s, a)}
&= \sum_{k=1}^K\frac{\widetilde{N}_{k}(s, a)}{\N_{k+1}^+(s, a)} + \sum_{k=1}^K\widetilde{N}_{k}(s, a)\rbr{\frac{1}{\N_{k}^+(s, a)} - \frac{1}{\N_{k+1}^+(s, a)}} \\
&\leq \sum_{k=1}^K\frac{\widetilde{N}_{k}(s, a)}{\N_{k+1}^+(s, a)} + H\sum_{k=1}^K \rbr{\frac{1}{\N_{k}^+(s, a)} - \frac{1}{\N_{k+1}^+(s, a)}} \\
&\leq \sum_{k=1}^K\frac{\widetilde{N}_{k}(s, a)}{\N_{k+1}^+(s, a)} + H \\
&= \ensuremath{\mathcal{O}}(\ln (HK) + H). \tag{$\widetilde{N}_{k}(s, a) = \N_{k+1}(s, a) - \N_k(s, a)$}
\end{align*}
Therefore, applying the fact $\field{E}_k[\widetilde{N}_k(s, a)] = q_k(s,a)$ and \pref{lem:e2r}, we have with probability at least $1-\delta$:
\begin{align*}
\sumsa\sum_{k=1}^K\frac{q_k(s, a)}{\N^+_k(s, a)} = \sum_{k=1}^K\field{E}_k\sbr{\sumsa\frac{n_k(s, a)}{\N^+_k(s, a)}} \leq 2\sumsa\sum_{k=1}^K\frac{n_k(s, a)}{\N^+_k(s, a)} + \bigO{H\ln\frac{2K}{\delta}} = \tilO{SAH},
\end{align*}
competing the proof.
\end{proof}
\begin{lemma}
\label{lem:transition-bias}
Consider interacting with the environment for $K$ episodes, where in episode $k$ the executed policy is ${\widetilde{\pi}}_k$ and the cost function denoted by $c_k: \widetilde{\SA}\rightarrow[0, 1]$ is arbitrary.
Also let $P_k$ be any transition function within the confidence set ${\mathcal{P}}_k$ defined in \pref{eq:conf-set},
and define $q_k=q_{P, {\widetilde{\pi}}_k}$ and $\widehat{q}_k=q_{P_k, {\widetilde{\pi}}_k}$.
Then with probability at least $1-4\delta$, we have
\begin{align*}
\sum_{k=1}^K\abr{\inner{q_k-\widehat{q}_k}{c_k}} &\leq 32\sqrt{S^2A\ln^2\rbr{\frac{HKSA}{\delta}} \rbr{\sum_{k=1}^K\text{\rm Var}_k\sbr{\inner{\widetilde{N}_k}{c_k}} + \tilO{H^3\sqrt{K}}}} + \tilO{\Qdtd}\\
&\leq 32\sqrt{S^2A\ln^2\rbr{\frac{HKSA}{\delta}} \rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{H^3\sqrt{K}}}} + \tilO{\Qdtd} \\
&\leq 16\rbr{\lambda'S^2A\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{H^3\sqrt{K}} } + \frac{\ln^2\rbr{\frac{HKSA}{\delta}}}{\lambda'}} + \tilO{\Qdtd}.
\end{align*}
where $\lambda'>0$ is arbitrary.
\end{lemma}
\begin{proof}
Define $\mu_k(s, a, h)=\sum_{s'}P(s'|s, a, h)J^{{\widetilde{\pi}}_k}_k(s', h+1)$.
Then by \pref{lem:om-diff}:
\begin{align*}
&\sum_{k=1}^K\abr{\inner{q_k-\widehat{q}_k}{c_k}} = \sum_{k=1}^K\abr{\sum_{(s, a), s', h}q_k(s, a, h)(P(s'|s, a, h)-P_k(s'|s, a, h))J^{P_k, {\widetilde{\pi}}_k}_k(s', h+1)}\\
&\leq\sum_{k=1}^K\abr{\sum_{(s, a), s', h}q_k(s, a, h)(P(s'|s, a, h)-P_k(s'|s, a, h))J^{{\widetilde{\pi}}_k}_k(s', h+1)} + \tilO{\Qdtd} \tag{\pref{lem:double transition difference}}\\
&=\sum_{k=1}^K\abr{\sum_{(s, a), s', h}q_k(s, a, h)(P(s'|s, a, h)-P_k(s'|s, a, h))\rbr{J^{{\widetilde{\pi}}_k}_k(s', h+1) - \mu_k(s, a, h)} } + \tilO{\Qdtd} \tag{$\sum_{s'}P(s'|s, a, h)-P_k(s'|s, a, h)=0$ and $\mu_k(s, a, h)$ is independent of $s'$}\\
&\leq\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}q_k(s, a, h)\epsilon_k^{\star}(s, a, s')\left|J^{{\widetilde{\pi}}_k}_k(s', h+1) - \mu_k(s, a, h)\right| \tag{$P(s'|s, a, h)-P_k(s'|s, a, h)\leq \epsilon_k^{\star}(s, a, s')$ by \pref{lem:transition-bound}} + \tilO{\Qdtd}\\
&\leq 8\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}q_k(s, a, h)\sqrt{P(s'|s, a, h)\A_k(s, a)\rbr{ J^{{\widetilde{\pi}}_k}_k(s', h+1) - \mu_k(s, a, h) }^2}\\
&\qquad\qquad + \tilO{HS\sum_{k=1}^K\sumsa\frac{q_k(s, a)}{\N^+_k(s, a)} + \Qdtd} \tag{definition of $\epsilon_k^{\star}$ from \pref{lem:transition-bound}} \\
&\leq 8\sum_{k=1}^K\field{E}_k\sbr{\sum_{(s, a), s', h\in\unk}\widetilde{N}_k(s, a, h)\sqrt{P(s'|s, a, h)\A_k(s, a)\rbr{ J^{{\widetilde{\pi}}_k}_k(s', h+1) - \mu_k(s, a, h) }^2}} + \tilO{\Qdtd} \tag{\pref{lem:transition-sum}}\\
&= 8\sum_{k=1}^K \field{E}_kX_k + \tilO{\Qdtd} \tag{$X_k=\sum_{(s, a), s', h\in\unk}\widetilde{N}_k(s, a, h)\sqrt{P(s'|s, a, h)\A_k(s, a)\rbr{ J^{{\widetilde{\pi}}_k}_k(s', h+1) - \mu_k(s, a, h) }^2}$}.
\end{align*}
Note that $0 \leq X_k = \tilo{ \sum_{(s, a), s', h\in\unk}\widetilde{N}_k(s, a, h)H } = \tilo{H^2S}$.
Hence, by \pref{lem:e2r}, with probability at least $1-\delta$:
\begin{align*}
&\sum_{k=1}^K \field{E}_kX_k \leq 2\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}\widetilde{N}_k(s, a, h)\sqrt{P(s'|s, a, h)\A_k(s, a)\rbr{ J^{{\widetilde{\pi}}_k}_k(s', h+1) - \mu_k(s, a, h) }^2} + \tilO{H^2S}\\
&\leq 2\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}\widetilde{N}_k(s, a, h)\sqrt{P(s'|s, a, h)A_{k+1}(s, a)\rbr{ J^{{\widetilde{\pi}}_k}_k(s', h+1) - \mu_k(s, a, h) }^2}\\
&\qquad\qquad + 2H^2\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}\rbr{\sqrt{\A_k(s, a)} - \sqrt{A_{k+1}(s, a)}} + \tilO{ H^2S }\\
&\leq 2\sqrt{\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}\widetilde{N}_k(s, a, h)P(s'|s, a, h)\rbr{ J^{{\widetilde{\pi}}_k}_k(s', h+1) - \mu_k(s, a, h) }^2}\sqrt{\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}\widetilde{N}_k(s, a, h)A_{k+1}(s, a) } \\
&\qquad\qquad + \tilO{H^3S^2A} \tag{Cauchy-Schwarz inequality}.
\end{align*}
Note that
\begin{align*}
&\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}\widetilde{N}_k(s, a, h)A_{k+1}(s, a) = \ln\rbr{\frac{HKSA}{\delta}}\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}\frac{\widetilde{N}_k(s, a, h)}{\N_{k+1}(s, a)}\\
&= S\ln\rbr{\frac{HKSA}{\delta}}\sum_{k=1}^K\sumsa\frac{\N_{k+1}(s, a)-\N^+_k(s, a)}{\N_{k+1}(s, a)} \leq 2S^2A\ln^2\rbr{\frac{HKSA}{\delta}}.
\end{align*}
Moreover, define $\field{V}_k(s, a, h)=P(s'|s, a, h)\rbr{ J^{{\widetilde{\pi}}_k}_k(s', h+1) - \mu_k(s, a, h) }^2$, with probability at least $1-\delta$,
\begin{align*}
&\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}\widetilde{N}_k(s, a, h)P(s'|s, a, h)\rbr{ J^{{\widetilde{\pi}}_k}_k(s', h+1) - \mu_k(s, a, h) }^2\\
&=\sum_{k=1}^K\sum_{(s, a), s', h\in\unk}\widetilde{N}_k(s, a, h)\field{V}_k(s, a, h) \leq\sum_{k=1}^K\sum_{(s, a), h}\widetilde{N}_k(s, a, h)\field{V}_k(s, a, h)\\
&=\sum_{k=1}^K\sum_{(s, a), h}q_k(s, a, h)\field{V}_k(s, a, h) + \sum_{k=1}^K\sum_{(s, a), h}(\widetilde{N}_k(s, a, h) - q_k(s, a, h))\field{V}_k(s, a, h)\\
&\leq 2\sum_{k=1}^K\text{\rm Var}\sbr{\inner{\widetilde{N}_k}{c_k}} + \tilO{H^3\sqrt{K}},
\end{align*}
where in the last inequality we apply \pref{lem:var} and \pref{lem:azuma} with
\begin{align*}
\sum_{(s, a), h}|\widetilde{N}_k(s, a, h) - q_k(s, a, h)|\field{V}_k(s, a, h) \leq \sum_{(s, a), h}\widetilde{N}_k(s, a, h)H^2 + \sum_{(s, a), h}q_k(s, a, h)H^2 \leq 2H^3.
\end{align*}
Therefore, combining everything we arrive at
\begin{align*}
\abr{\sum_{k=1}^K\inner{q_k-\widehat{q}_k}{c_k}} &\leq 32\sqrt{S^2A\ln^2\rbr{\frac{HKSA}{\delta}}\rbr{\sum_{k=1}^K\text{\rm Var}\sbr{\inner{\widetilde{N}_k}{c_k}} + \tilO{H^3\sqrt{K}}}} + \tilO{\Qdtd}.\\
\end{align*}
When the value of $c_k$ is at most $1$, we have by \pref{lem:deviation_loop_free} and $\text{\rm Var}\sbr{\inner{\widetilde{N}_k}{c_k}}\leq \field{E}_k\sbr{\inner{\widetilde{N}_k}{c_k}^2}$:
\begin{align*}
\abr{\sum_{k=1}^K\inner{q_k-\widehat{q}_k}{c_k}} &\leq 32\sqrt{S^2A\ln^2\rbr{\frac{HKSA}{\delta}}\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{H^3\sqrt{K}}}} + \tilO{\Qdtd}\\
&\leq 16\rbr{\lambda'S^2A\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{H^3\sqrt{K}} } + \frac{\ln^2\rbr{\frac{HKSA}{\delta}}}{\lambda'}} + \tilO{\Qdtd},
\end{align*}
where the last step is by AM-GM inequality.
\end{proof}
\begin{lemma}
\label{lem:double transition difference}
Under the event of \pref{lem:transition-interval},
\begin{align*}
\sum_{k=1}^K\abr{\sum_{(s, a), s', h}q_k(s, a, h)(P(s'|s, a, h)-P_k(s'|s, a, h))\rbr{J^{P_k, {\widetilde{\pi}}_k}_k(s', h+1) - J^{P, {\widetilde{\pi}}_k}_k(s', h+1) }} = \tilO{\Qdtd}.
\end{align*}
\end{lemma}
\begin{proof}
Define $q_{k,(s', h+1)}=q_{P,{\widetilde{\pi}}_k,(s',h+1)}$.
Then,
\begin{align*}
&\sum_{k=1}^K \abr{\sum_{(s, a), s', h}q_k(s, a, h)(P(s'|s, a, h)-P_k(s'|s, a, h))\rbr{J^{P_k, {\widetilde{\pi}}_k}_k(s', h+1) - J^{P, {\widetilde{\pi}}_k}_k(s', h+1) } }\\
&\leq \sum_{k=1}^K\sum_{(s, a), s', h\in \unk}q_k(s, a, h)\epsilon_k^\star(s, a, s')\abr{\inner{q_{P_k, {\widetilde{\pi}}_k, (s', h+1)}-q_{P, {\widetilde{\pi}}_k, (s', h+1)}}{c_k}} \tag{\pref{lem:transition-bound} and $J^{P', {\widetilde{\pi}}_k}_k(s',h+1)=\inner{q_{P', {\widetilde{\pi}}_k, (s', h+1)}}{c_k}$}\\
&\leq \sum_{k=1}^K\sum_{(s, a), s', h\in \unk}q_k(s, a, h)\epsilon_k^\star(s, a, s')\sum_{(\widetilde{s}, \widetilde{a}), \widetilde{s}', h'\in\unk}q_{k,(s',h+1)}(\widetilde{s}, \widetilde{a}, h')\epsilon_k^\star(\widetilde{s}, \widetilde{a}, \widetilde{s}')H \tag{\pref{lem:om-diff}}\\
&=\tilO{H\sum_{k=1}^K\sum_{\substack{(s, a), s', h\in \unk \\ (\widetilde{s}, \widetilde{a}), \widetilde{s}', h'\in\unk } }q_k(s, a, h)\sqrt{\frac{P(s'|s, a, h)}{\N^+_k(s, a)}}q_{k,(s', h+1)}(\widetilde{s}, \widetilde{a}, h')\sqrt{\frac{P(\widetilde{s}'|\widetilde{s}, \widetilde{a}, h')}{\N^+_k(\widetilde{s}, \widetilde{a})}} }\\
&= \tilO{H\sum_{k=1}^K\sum_{\substack{(s, a), s', h\in \unk \\ (\widetilde{s}, \widetilde{a}), \widetilde{s}', h'\in\unk } }\sqrt{\frac{q_k(s, a, h)P(\widetilde{s}'|\widetilde{s}, \widetilde{a}, h')q_{k, (s', h+1)}(\widetilde{s}, \widetilde{a}, h')}{\N^+_k(s, a)}}\sqrt{\frac{q_k(s, a, h)P(s'|s, a, h)q_{k, (s', h+1)}(\widetilde{s}, \widetilde{a}, h')}{\N^+_k(\widetilde{s}, \widetilde{a})}} }\\
&= \tilO{H\sqrt{\sum_{\substack{k, (s, a), s', h\in \unk \\ (\widetilde{s}, \widetilde{a}), \widetilde{s}', h'\in\unk }}\frac{q_k(s, a, h)P(\widetilde{s}'|\widetilde{s}, \widetilde{a}, h')q_{k, (s', h+1)}(\widetilde{s}, \widetilde{a}, h')}{\N^+_k(s, a)}}\sqrt{\sum_{\substack{k, (s, a), s', h\in \unk \\ (\widetilde{s}, \widetilde{a}), \widetilde{s}', h'\in\unk }}\frac{q_k(s, a, h)P(s'|s, a, h)q_{k, (s', h+1)}(\widetilde{s}, \widetilde{a}, h')}{\N^+_k(\widetilde{s}, \widetilde{a})}} } \tag{Cauchy-Schwarz inequality}\\
&= \tilO{ H\sqrt{HS\sum_{k,(s, a)}\frac{q_k(s, a)}{\N^+_k(s, a)} }\sqrt{HS\sum_{k, (\widetilde{s}, \widetilde{a})}\frac{q_k(\widetilde{s}, \widetilde{a})}{\N^+_k(\widetilde{s}, \widetilde{a})} } } = \tilO{\Qdtd}. \tag{\pref{lem:transition-sum}}
\end{align*}
where in the first square root of the last line we simply sum over $s', h, (\widetilde{s}, \widetilde{a}), \widetilde{s}', h'$, and in the second square root of the last line we apply $\sum_{(s, a), s'}q_k(s, a, h)P(s'|s, a, h)q_{k,(s', h+1)}(\widetilde{s}, \widetilde{a}, h')=q_k(\widetilde{s}, \widetilde{a}, h')$ for any $\widetilde{s}, \widetilde{a}, h<h'$.
\end{proof}
\subsection{\pfref{thm:full-unknown}}
\label{sec:thm-full-unknown}
\begin{proof}
First apply \pref{lem:loop-free}: with probability at least $1-\delta$, we have
\begin{align*}
R_K \leq \sum_{k=1}^K\inner{\widetilde{N}_k-q_{{\widetilde{\pi}^\star}}}{c_k} + \tilO{D^{3/2}S^2A(\ln\tfrac{1}{\delta})^2}.
\end{align*}
Define $P_k=P_{\widehat{q}_k}$, and $q_k=q_{P, {\widetilde{\pi}}_k}$.
We decompose the regret in $\widetilde{M}$ into three terms:
\begin{align}
\sum_{k=1}^K\inner{\widetilde{N}_k - q_{{\widetilde{\pi}^\star}}}{c_k} = \sum_{k=1}^K\inner{\widetilde{N}_k-q_k}{c_k} + \sum_{k=1}^K\inner{q_k-\widehat{q}_k}{c_k} + \sum_{k=1}^K\inner{\widehat{q}_k - q_{{\widetilde{\pi}^\star}}}{c_k}\label{eq:full-unknown-reg}
\end{align}
For the first term, by \pref{lem:freedman} with $\frac{\lambda}{4} \leq \frac{1}{H}$ (because $K\geq 16S^2AH^2$) and \pref{lem:deviation_loop_free}, we have with probability at least $1-\delta$:
\begin{align*}
\sum_{k=1}^K\inner{\widetilde{N}_k-q_k}{c_k} \leq \frac{\lambda}{2}\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \frac{4\ln(1/\delta)}{\lambda} .
\end{align*}
For the second term, by \pref{lem:transition-bias}, with probability $1-4\delta$, we have for any $\lambda'>0$:
\begin{align*}
\sum_{k=1}^K\inner{q_k-\widehat{q}_k}{c_k} &\leq 16\rbr{\lambda'S^2A\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}} + \tilO{H^3\sqrt{K}} } + \frac{\ln^2\rbr{\frac{HKSA}{\delta}}}{\lambda'} } + \tilO{H^3S^2A}.
\end{align*}
For the third term, by standard OMD analysis (see for example Eq.~(12) of~\citep{rosenberg2020adversarial}):
\begin{align}
\sum_{k=1}^K\inner{\qfeat_k - \qfeat_{{\widetilde{\pi}^\star}}}{c_k} &\leq D_{\regz}(\qfeat_{{\widetilde{\pi}^\star}}, \qfeat_1) + \sum_{k=1}^K\inner{\qfeat_k - \qfeat'_{k+1}}{c_k},\label{eq:omd}
\end{align}
where $\qfeat'_{k+1} = \argmin_{\phi\in\field{R}^{\widetilde{\SA}\times{\mathcal{S}}\times[H]}}\inner{\phi}{c_k} + D_{\psi}(\phi, \qfeat_k)$.
It can be shown that $\qfeat'_{k+1}(s, a, s', h)=\qfeat_k(s, a, s', h)e^{-\eta c_k(s, a)}$ with the choice of the entropy regularizer.
Applying the inequality $1-e^{-x}\leq x$, we get:
\begin{align*}
\sum_{k=1}^K\inner{\qfeat_k-\qfeat'_{k+1}}{c_k} &\leq \eta\sum_{k=1}^K\sum_{(s, a), s', h}\qfeat_k(s, a, s', h)c_k^2(s, a) \leq 2\eta\sum_{k=1}^K\sumsa\widehat{q}_k(s, a)c_k(s, a) = 2\eta\sum_{k=1}^K\inner{\widehat{q}_k}{c_k},
\end{align*}
where in the last inequality we apply $\qfeat_k(s, a, s', h) = (1+\lambda h)\widehat{q}_k(s, a, s', h)\leq 2\widehat{q}_k(s, a, s', h)$.
To bound $D_{\psi}(\qfeat_{{\widetilde{\pi}^\star}}, \qfeat_1)$, note that $\inner{\nabla\psi(\qfeat_1)}{\qfeat_{{\widetilde{\pi}^\star}}-\qfeat_1}\geq 0$ by $\qfeat_1=\argmin_{\phi\in\Omega_1}\psi(\phi)$.
Thus,
\begin{align*}
D_{\psi}(\qfeat_{{\widetilde{\pi}^\star}}, \qfeat_1) &\leq \psi(\qfeat_{{\widetilde{\pi}^\star}}) - \psi(\qfeat_1) \\
&= \frac{1}{\eta}\sum_{(s, a), s', h}\qfeat_{{\widetilde{\pi}^\star}}(s, a, s', h)\ln\qfeat_{{\widetilde{\pi}^\star}}(s, a, s', h) - \frac{1}{\eta}\sum_{(s, a), s', h}\qfeat_1(s, a, s', h)\ln\qfeat_1(s, a, s', h)\\
&\leq \frac{1}{\eta}\sum_{(s, a), s', h}\qfeat_{{\widetilde{\pi}^\star}}(s, a, s', h)\ln(2T) -\frac{2T}{\eta}\sum_{(s, a), s', h}\frac{\qfeat_1(s, a, s', h)}{2T}\ln\frac{\qfeat_1(s, a, s', h)}{2T}\\
&\leq \frac{2T\ln(2T)}{\eta} + \frac{2T\ln(S^2AH)}{\eta} \leq \frac{2T\ln(2S^2AHT)}{\eta}.
\end{align*}
Substituting these back to \pref{eq:omd} and rearranging terms, we get:
\begin{align}
\sum_{k=1}^K\inner{\widehat{q}_k-q_{{\widetilde{\pi}^\star}}}{c_k} &\leq \frac{1}{1-2\eta}\rbr{ \frac{2T\ln(2S^2AHT)}{\eta} + 2\eta\sum_{k=1}^K\inner{q_{{\widetilde{\pi}^\star}}}{c_k} + \sum_{k=1}^K\lambda\inner{q_{{\widetilde{\pi}^\star}}-\widehat{q}_k}{\h{c_k}} }\notag\\
&\leq \frac{4T\ln(2S^2AHT)}{\eta} + 4\eta DK + 2\lambda DTK - \lambda\sum_{k=1}^K\inner{\widehat{q}_k}{\h{c_k}},\label{eq:full unknown omd}
\end{align}
where we apply $\frac{1}{1-2\eta}\leq 2, \sum_{k=1}^K\inner{q_{{\widetilde{\pi}^\star}}}{c_k}\leq DK$ and $\sum_{k=1}^K\inner{q_{{\widetilde{\pi}^\star}}}{\h{c_k}}=\sum_{k=1}^K\inner{q_{{\widetilde{\pi}^\star}}}{J_k^{{\widetilde{\pi}^\star}}}\leq DTK$ in the last inequality.
Substituting everything back to \pref{eq:full-unknown-reg} and set $\lambda'=\frac{1}{8}\sqrt{\frac{1}{S^2ADTK}}$, we get:
\begin{align*}
\sum_{k=1}^K\inner{\widetilde{N}_k-q_{{\widetilde{\pi}^\star}}}{c_k} &\leq \tilO{\frac{1}{\lambda} + \frac{1}{\lambda'} + \frac{T}{\eta} } + \rbr{16\lambda'S^2A - \frac{\lambda}{2}}\sum_{k=1}^K\inner{q_k}{\h{c_k}}\\
&+ \lambda\sum_{k=1}^K\inner{q_k-\widehat{q}_k}{\h{c_k}} + 4\eta DK + 2\lambda DTK + \tilO{\lambda'S^2AH^3\sqrt{K} + H^3S^2A}\\
&= \tilO{\sqrt{S^2ADTK} + \sqrt{DTK} + \lambda\sqrt{S^2AH^2K} + H^3S^2A } = \tilO{\sqrt{S^2ADTK} + H^3S^2A},
\end{align*}
where in the last line we apply $\eta\leq \sqrt{\frac{T}{DK}}, \lambda=4\sqrt{\frac{S^2A}{DTK}}$, and
\begin{align*}
&\lambda\sum_{k=1}^K\inner{q_k-\widehat{q}_k}{\h{c_k}}=\tilO{ \sqrt{\frac{S^2A}{DTK}} \cdot \sqrt{S^2A\rbr{\sum_{k=1}^K\inner{q_k}{\h{c_k}}+H^3\sqrt{K} } }}\\
&=\tilO{\sqrt{ \frac{S^2A}{K}\cdot S^2A\rbr{ H^2K + H^3\sqrt{K} } }} = \tilO{H^3S^2A}.
\end{align*}
\end{proof}
\subsection{\pfref{lem:loop-free}}
\begin{proof}
Denote by $N'_k(s, a)$ the number of visits to $(s, a)$ during episode $k$ before switching to Bernstein-SSP,
by $N''_k(s, a)$ the number of visits to $(s, a)$ after switching to Bernstein-SSP,
and by $N_f$ the number of episodes where Bernstein-SSP is invoked.
We have: $N_k(s, a)=N'_k(s, a)+N''_k(s, a)$ and $\sum_{k=1}^K\inner{\widetilde{N}_k}{c_k}=\sum_{k=1}^K\inner{N'_k}{c_k}+H_2N_f$.
Recall that the regret of running Bernstein-SSP~\citep{cohen2020near} for $K'$ episodes under uniform cost is of $\bigO{ DS\sqrt{AK'}\ln\frac{K'DSA}{\delta} + \sqrt{D^3S^4A^2}\ln^2\frac{K'DSA}{\delta} }$ with probability at least $1-\delta$.
Conditioned on the event above, we have:
\begin{align*}
\sum_{k=1}^K\inner{N''_k}{c_k} - H_2N_f &\leq \rbr{\sum_{k=1}^K\inner{N''_k}{c_k} - DN_f} - DN_f \tag{$H_2\geq 2D$}\\
&\leq \bigO{DS\sqrt{AN_f}\ln\frac{KDSA}{\delta} + \sqrt{D^3S^4A^2}\ln^2\frac{KDSA}{\delta}} - DN_f\\
&= \bigO{D^{3/2}S^2A\ln^2\frac{KDSA}{\delta}},
\end{align*}
where in the last inequality we solve for the maximum of a quadratic function with variable $N_f$.
Therefore,
\begin{align*}
\sum_{k=1}^K\inner{N_k}{c_k} = \sum_{k=1}^K\inner{N'_k}{c_k} + H_2N_f + \sum_{k=1}^K\inner{N''_k}{c_k} - H_2N_f \leq \sum_{k=1}^K\inner{\widetilde{N}_k}{c_k} + \tilO{D^{3/2}S^2A\ln^2\frac{1}{\delta}}.
\end{align*}
On the other hand, by \pref{lem:hitting}, the probability that the goal state is not reached within $H_1$ steps when executing $\pi^\star$ is at most $2e^{-\frac{H_1}{4\ensuremath{T_{\max}}}} \leq \frac{2}{K^2}$.
Hence, the expected cost of $\pi^\star$ in $M$ and the expected cost of ${\widetilde{\pi}^\star}$ in $\widetilde{M}$ is very similar:
\begin{align*}
J^{{\widetilde{\pi}^\star}}_k(\widetilde{s}_0) \leq J^{\pi^\star}_k(s_0) + \frac{2H_2}{K^2} =
J^{\pi^\star}_k(s_0) + \tilO{\frac{1}{K}}.
\end{align*}
Putting everything together, and by $K\geq D$, we get:
\begin{align*}
R_K = \sum_{k=1}^K\inner{N_k}{c_k} - J^{\pi^\star}_k(s_0) \leq \sum_{k=1}^K\inner{\widetilde{N}_k - q_{{\widetilde{\pi}^\star}}}{c_k} +\tilO{D^{3/2}S^2A\ln^2\frac{1}{\delta}}.
\end{align*}
\end{proof}
\subsection{\pfref{lem:deviation_loop_free}}
\begin{proof}
With the inequality $(\sum_{i=1}^I a_i)^2\leq 2\sum_{i}a_i(\sum_{i'= i}^I a_{i'})$, we proceed as
\begin{align*}
&\field{E}_k\left[\left( \sum_{(s, a)\in\widetilde{\SA}, h} \widetilde{N}_{k}(s, a, h)c_k(s, a, h) \right)^2\right] \\
&\leq 2\field{E}_k\left[\sum_{h=1}^{H}\sumtilsaf \widetilde{N}_{k}(s, a, h)c_{k}(s, a, h)\left( \sum_{h'= h}^{H}\sumtilsaf[s', a'] \widetilde{N}_{k}(s', a', h')c_k(s', a', h') \right)\right] \\
&= 2\field{E}_k\left[\sum_{h=1}^{H}\sumtilsaf \widetilde{N}_{k}(s, a, h)c_{k}(s, a, h)\field{E}\left[\left. \sum_{h'=h}^{H}\sumtilsaf[s', a'] \widetilde{N}_{k}(s', a', h') c_k(s', a', h') \right| \widetilde{s}^h_k=(s, h), a^h_k=a\right]\right]\\
&= 2\field{E}_k\left[\sum_{h=1}^{H}\sumtilsaf \widetilde{N}_{k}(s, a, h)c_k(s, a, h)Q^{{\widetilde{\pi}}}_k(s, a, h)\right] = 2\inner{q_{{\widetilde{\pi}}}}{c_k\odot Q^{{\widetilde{\pi}}}_k}\\
&\leq 2\field{E}_k\left[\sum_{h=1}^{H}\sumtilsaf \widetilde{N}_{k}(s, a, h)Q^{{\widetilde{\pi}}}_k(s, a, h)\right] = 2 \sum_{h=1}^{H}\sumtilsaf q_{{\widetilde{\pi}}}(s, a, h)Q^{{\widetilde{\pi}}}_k(s, a, h)\\
&= 2\sum_{h=1}^H\sum_{s\in{\mathcal{S}}\cup\{s_f\}} q_{{\widetilde{\pi}}}(s, h)\sum_a{\widetilde{\pi}}(a|s)Q^{{\widetilde{\pi}}}_k(s, a, h) = 2\sum_{h=1}^H\sum_{s\in{\mathcal{S}}\cup\{s_f\}}q_{{\widetilde{\pi}}}(s, h)J^{{\widetilde{\pi}}}_k(s, h) = 2\inner{q_{\pi}}{J^{\pi}_k}.
\end{align*}
This proves the first two inequalities.
Denote by $q_{{\widetilde{\pi}}, (s, h)}$ the occupancy measure of policy ${\widetilde{\pi}}$ with initial state $(s, h)$, so that
\[
J^{{\widetilde{\pi}}}_k(s, h) = \sum_{(s', a')\in\widetilde{\SA}}\sum_{h'\geq h}q_{{\widetilde{\pi}},(s, h)}(s', a', h')c_k(s', a', h').
\]
Then, we continue with the following equalities:
\begin{align*}
&\sum_{h=1}^H\sum_{s\in{\mathcal{S}}\cup\{s_f\}}q_{{\widetilde{\pi}}}(s, h)J^{{\widetilde{\pi}}}_k(s, h) \\
&= \sum_{h=1}^{H}\sum_{s\in{\mathcal{S}}\cup\{s_f\}} q_{{\widetilde{\pi}}}(s, h)\sum_{(s', a')\in\widetilde{\SA}}\sum_{h'\geq h}q_{{\widetilde{\pi}},(s, h)}(s', a', h')c_k(s', a', h')\\
&= \sum_{h=1}^{H}\sum_{(s', a')\in\widetilde{\SA}}\sum_{h'\geq h}\left(\sum_{s\in{\mathcal{S}}\cup\{s_f\}} q_{{\widetilde{\pi}}}(s, h)q_{{\widetilde{\pi}},(s, h)}(s', a', h')\right) c_k(s', a', h')\\
&= \sum_{h=1}^{H}\sum_{(s', a')\in\widetilde{\SA}}\sum_{h'\geq h}q_{{\widetilde{\pi}}}(s', a', h')c_k(s', a', h') = \sum_{h'=1}^{H}\sum_{(s', a')\in\widetilde{\SA}}\sum_{h\leq h'}q_{{\widetilde{\pi}}}(s', a', h')c_k(s', a', h') \\
&= \sum_{h'=1}^{H}\sum_{(s', a')\in\widetilde{\SA}} h'\cdot q_{{\widetilde{\pi}}}(s', a', h')c_k(s', a', h')
= \inner{q_{{\widetilde{\pi}}}}{\h{c_k}},
\end{align*}
where in the third line we use the equality $\sum_{s\in{\mathcal{S}}\cup\{s_f\}} q_{{\widetilde{\pi}}}(s, h)q_{{\widetilde{\pi}},(s, h)}(s', a', h') = q_{{\widetilde{\pi}}}(s', a', h')$ by definition (since both sides are the probability of visiting $(s', a', h')$).
This proves the last equality and completes the proof.
\end{proof}
\begin{lemma}{\citep[Lemma E.1]{rosenberg2020adversarial}}\label{lem:hitting}
Let $\pi$ be a policy with expected hitting time at most $\tau$ starting from any state.
Then, the probability that $\pi$ takes more than $m$ steps to reach the goal state is at most $2e^{-\frac{m}{4\tau}}$.
\end{lemma}
\section{Learning without knowing $\ensuremath{T_\star}$ or $\ensuremath{T_{\max}}$}
\label{app:tune T}
In this section, we discuss how to instantiate our proposed algorithms without knowing $\ensuremath{T_\star}$ or $\ensuremath{T_{\max}}$ (or both).
We apply our ideas to \pref{alg:full-unknown} to give concrete examples, and they are applicable to other proposed algorithms similarly.
The modifications described below can be applied separately or jointly depending on the knowledge we have.
Moreover, they are compatible with ideas in \pref{sec:tund-d} for learning without knowing the SSP-diameter $D$.
\paragraph{Learning without knowing $\ensuremath{T_\star}$} We assume knowledge of $\ensuremath{c_{\min}}=\min_{s, a, k}c_k(s, a)$ and $\ensuremath{c_{\min}}>0$, which is the assumption made in \cite{rosenberg2020adversarial}.
Then by the inequality $\ensuremath{T_\star}\leq\frac{T^{\pi^f}(s_0)}{\ensuremath{c_{\min}}}$, it suffices to obtain an upper bound of $T^{\pi^f}(s_0)$ to obtain an upper bound of $\ensuremath{T_\star}$.
To this end, we first run a Bernstein-SSP instance for $L=\tilo{\sqrt{AK}}$ episodes with uniform costs equal to $1$ and obtain $\widetilde{D}_{s_0}$, where both $L$ and $\widetilde{D}_{s_0}$ are defined in \pref{sec:tund-d}. Then, we simply run \pref{alg:full-unknown} with $T=\widetilde{D}_{s_0}/\ensuremath{c_{\min}}$.
Following the arguments in \pref{sec:tund-d}, we know that the extra costs of estimating $\widetilde{D}_{s_0}$ in the regret is $\tilo{DL+D^{3/2}S^2A+D^4S^4AH}=\tilo{D\sqrt{AK}+D^4S^4AH}$.
Thus, we obtain the following result:
\begin{theorem}
If $H_1\geq 8\ensuremath{T_{\max}}\ln K$, and $K\geq 16S^2AH^2$, then with probability at least $1-6\delta$, the algorithm described above ensures $R_K=\tilo{D\sqrt{\frac{S^2AK}{\ensuremath{c_{\min}}}} + H^3S^2A + D^4S^4AH}$.
\end{theorem}
Compared to \cite{rosenberg2020adversarial}, the bound above is $\sqrt{\frac{1}{\ensuremath{c_{\min}}}}$ better in the dominating term.
When $\ensuremath{c_{\min}}=0$, similarly to \cite{rosenberg2020adversarial}, we can solve a modified MDP with perturbed cost functions $c'_k(s, a)=\max\{c_k(s, a),\epsilon\}$, where $\epsilon=K^{-1/3}$.
The bias brought by the perturbation is of order $\tilO{\epsilon\ensuremath{T_\star} K}$, and thus the overall regret is $\tilo{K^{2/3}}$ ignoring other parameters and constant terms.
This is asymptomatically better than the $\tilO{K^{3/4}}$ regret in \cite{rosenberg2020adversarial} for $\ensuremath{c_{\min}}=0$.
We conclude that \pref{alg:full-unknown} improves over previous work even without knowledge of $\ensuremath{T_\star}$.
\paragraph{Learning without knowing $\ensuremath{T_{\max}}$} Similarly to \cite{chen2020minimax}, we run \pref{alg:full-unknown} with $H_1=8(K/S^2A)^{1/6}\ln K$.
Note that when $K\leq \ensuremath{T_{\max}}^6S^2A$, by the regret guarantee of Bernstein-SSP, we have:
\begin{align*}
R_K \leq KH_1 + KD + \tilO{ DS\sqrt{AK} + D^{\frac{3}{2}}S^2A } = \tilO{ \ensuremath{T_{\max}}^7S^2A }.
\end{align*}
Otherwise, $H_1\geq\ensuremath{T_{\max}}$, and by the regret guarantee of \pref{alg:full-unknown}, we have:
$$R_K=\tilO{\sqrt{S^2ADTK} + H^3S^2A} = \tilO{\sqrt{S^2ADTK}}.$$
Combining these two cases, we have the following result:
\begin{theorem}
If $T\geq\ensuremath{T_\star}+1$, and $K\geq 16S^2AH^2$, then with probability at least $1-6\delta$, running \pref{alg:full-unknown} with $H_1=8(K/S^2A)^{1/6}\ln K$ ensures $R_K=\tilo{\sqrt{S^2ADTK} + \ensuremath{T_{\max}}^7S^2A}$.
\end{theorem}
Applying both modifications to \pref{alg:full-unknown}, we obtain: $R_K = \tilO{ D\sqrt{\frac{S^2AK}{\ensuremath{c_{\min}}}} + D^4S^4AK^{1/6} + \ensuremath{T_{\max}}^7S^2A }$, which is still asymptomatically better compared to \cite{rosenberg2020adversarial}.
\section{Extra Notations}
\section{Loop-free reduction}
\label{app:loop-free}
\input{appendix-loop-free}
\section{Omitted details for \pref{sec:full-unknown}}
\label{app:full-unknown}
\input{appendix-full-unknown}
\section{Omitted details for \pref{sec:bandit-unknown}}
\label{app:bandit-unknown}
\input{appendix-bandit-unknown}
\section{Omitted details for \pref{sec:iid}}
\label{app:iid}
\input{appendix-iid}
\section{Learning without knowing SSP-diameter}
\label{app:tune-d}
\input{appendix-tune-d}
\section{Concentration Inequalities}
\label{app:concentration}
\input{appendix-concentration}
\subsection{Extended MDP with fast action}
\subsection{Related work}
\section{Introduction}\label{sec:intro}
\input{introduction}
\section{Preliminaries}\label{sec:prelim}
\input{preliminaries}
\section{Loop-free Reduction with Unknown Transition}\label{sec:reduction}
\input{loop-free}
\section{Adversarial Costs with Full Information}
\label{sec:full-unknown}
\input{full-unknown}
\section{Adversarial Costs with Bandit Feedback}
\label{sec:bandit-unknown}
\input{bandit-unknown}
\section{Stochastically Oblivious Adversary}
\label{sec:iid}
\input{iid}
\section*{Acknowledgements}
We thank Aviv Rosenberg, Chen-Yu Wei, and the anonymous reviewers for many helpful discussions and feedback.
This work is supported by NSF Award IIS-1943607 and a Google Faculty Research Award.
\bibliographystyle{plainnat}
\section{Electronic Submission}
\label{submission}
Submission to ICML 2021 will be entirely electronic, via a web site
(not email). Information about the submission process and \LaTeX\ templates
are available on the conference web site at:
\begin{center}
\textbf{\texttt{http://icml.cc/}}
\end{center}
The guidelines below will be enforced for initial submissions and
camera-ready copies. Here is a brief summary:
\begin{itemize}
\item Submissions must be in PDF\@.
\item Submitted papers can be up to eight pages long, not including references, plus unlimited space for references. Accepted papers can be up to nine pages long, not including references, to allow authors to address reviewer comments. Any paper exceeding this length will automatically be rejected.
\item \textbf{Do not include author information or acknowledgements} in your
initial submission.
\item Your paper should be in \textbf{10 point Times font}.
\item Make sure your PDF file only uses Type-1 fonts.
\item Place figure captions \emph{under} the figure (and omit titles from inside
the graphic file itself). Place table captions \emph{over} the table.
\item References must include page numbers whenever possible and be as complete
as possible. Place multiple citations in chronological order.
\item Do not alter the style template; in particular, do not compress the paper
format by reducing the vertical spaces.
\item Keep your abstract brief and self-contained, one paragraph and roughly
4--6 sentences. Gross violations will require correction at the
camera-ready phase. The title should have content words capitalized.
\end{itemize}
\subsection{Submitting Papers}
\textbf{Paper Deadline:} The deadline for paper submission that is
advertised on the conference website is strict. If your full,
anonymized, submission does not reach us on time, it will not be
considered for publication.
\textbf{Anonymous Submission:} ICML uses double-blind review: no identifying
author information may appear on the title page or in the paper
itself. Section~\ref{author info} gives further details.
\textbf{Simultaneous Submission:} ICML will not accept any paper which,
at the time of submission, is under review for another conference or
has already been published. This policy also applies to papers that
overlap substantially in technical content with conference papers
under review or previously published. ICML submissions must not be
submitted to other conferences and journals during ICML's review
period.
Informal publications, such as technical
reports or papers in workshop proceedings which do not appear in
print, do not fall under these restrictions.
\medskip
Authors must provide their manuscripts in \textbf{PDF} format.
Furthermore, please make sure that files contain only embedded Type-1 fonts
(e.g.,~using the program \texttt{pdffonts} in linux or using
File/DocumentProperties/Fonts in Acrobat). Other fonts (like Type-3)
might come from graphics files imported into the document.
Authors using \textbf{Word} must convert their document to PDF\@. Most
of the latest versions of Word have the facility to do this
automatically. Submissions will not be accepted in Word format or any
format other than PDF\@. Really. We're not joking. Don't send Word.
Those who use \textbf{\LaTeX} should avoid including Type-3 fonts.
Those using \texttt{latex} and \texttt{dvips} may need the following
two commands:
{\footnotesize
\begin{verbatim}
dvips -Ppdf -tletter -G0 -o paper.ps paper.dvi
ps2pdf paper.ps
\end{verbatim}}
It is a zero following the ``-G'', which tells dvips to use
the config.pdf file. Newer \TeX\ distributions don't always need this
option.
Using \texttt{pdflatex} rather than \texttt{latex}, often gives better
results. This program avoids the Type-3 font problem, and supports more
advanced features in the \texttt{microtype} package.
\textbf{Graphics files} should be a reasonable size, and included from
an appropriate format. Use vector formats (.eps/.pdf) for plots,
lossless bitmap formats (.png) for raster graphics with sharp lines, and
jpeg for photo-like images.
The style file uses the \texttt{hyperref} package to make clickable
links in documents. If this causes problems for you, add
\texttt{nohyperref} as one of the options to the \texttt{icml2021}
usepackage statement.
\subsection{Submitting Final Camera-Ready Copy}
The final versions of papers accepted for publication should follow the
same format and naming convention as initial submissions, except that
author information (names and affiliations) should be given. See
Section~\ref{final author} for formatting instructions.
The footnote, ``Preliminary work. Under review by the International
Conference on Machine Learning (ICML). Do not distribute.'' must be
modified to ``\textit{Proceedings of the
$\mathit{38}^{th}$ International Conference on Machine Learning},
Online, PMLR 139, 2021.
Copyright 2021 by the author(s).''
For those using the \textbf{\LaTeX} style file, this change (and others) is
handled automatically by simply changing
$\mathtt{\backslash usepackage\{icml2021\}}$ to
$$\mathtt{\backslash usepackage[accepted]\{icml2021\}}$$
Authors using \textbf{Word} must edit the
footnote on the first page of the document themselves.
Camera-ready copies should have the title of the paper as running head
on each page except the first one. The running title consists of a
single line centered above a horizontal rule which is $1$~point thick.
The running head should be centered, bold and in $9$~point type. The
rule should be $10$~points above the main text. For those using the
\textbf{\LaTeX} style file, the original title is automatically set as running
head using the \texttt{fancyhdr} package which is included in the ICML
2021 style file package. In case that the original title exceeds the
size restrictions, a shorter form can be supplied by using
\verb|\icmltitlerunning{...}|
just before $\mathtt{\backslash begin\{document\}}$.
Authors using \textbf{Word} must edit the header of the document themselves.
\section{Format of the Paper}
All submissions must follow the specified format.
\subsection{Dimensions}
The text of the paper should be formatted in two columns, with an
overall width of 6.75~inches, height of 9.0~inches, and 0.25~inches
between the columns. The left margin should be 0.75~inches and the top
margin 1.0~inch (2.54~cm). The right and bottom margins will depend on
whether you print on US letter or A4 paper, but all final versions
must be produced for US letter size.
The paper body should be set in 10~point type with a vertical spacing
of 11~points. Please use Times typeface throughout the text.
\subsection{Title}
The paper title should be set in 14~point bold type and centered
between two horizontal rules that are 1~point thick, with 1.0~inch
between the top rule and the top edge of the page. Capitalize the
first letter of content words and put the rest of the title in lower
case.
\subsection{Author Information for Submission}
\label{author info}
ICML uses double-blind review, so author information must not appear. If
you are using \LaTeX\/ and the \texttt{icml2021.sty} file, use
\verb+\icmlauthor{...}+ to specify authors and \verb+\icmlaffiliation{...}+ to specify affiliations. (Read the TeX code used to produce this document for an example usage.) The author information
will not be printed unless \texttt{accepted} is passed as an argument to the
style file.
Submissions that include the author information will not
be reviewed.
\subsubsection{Self-Citations}
If you are citing published papers for which you are an author, refer
to yourself in the third person. In particular, do not use phrases
that reveal your identity (e.g., ``in previous work \cite{langley00}, we
have shown \ldots'').
Do not anonymize citations in the reference section. The only exception are manuscripts that are
not yet published (e.g., under submission). If you choose to refer to
such unpublished manuscripts \cite{anonymous}, anonymized copies have
to be submitted
as Supplementary Material via CMT\@. However, keep in mind that an ICML
paper should be self contained and should contain sufficient detail
for the reviewers to evaluate the work. In particular, reviewers are
not required to look at the Supplementary Material when writing their
review.
\subsubsection{Camera-Ready Author Information}
\label{final author}
If a paper is accepted, a final camera-ready copy must be prepared.
For camera-ready papers, author information should start 0.3~inches below the
bottom rule surrounding the title. The authors' names should appear in 10~point
bold type, in a row, separated by white space, and centered. Author names should
not be broken across lines. Unbolded superscripted numbers, starting 1, should
be used to refer to affiliations.
Affiliations should be numbered in the order of appearance. A single footnote
block of text should be used to list all the affiliations. (Academic
affiliations should list Department, University, City, State/Region, Country.
Similarly for industrial affiliations.)
Each distinct affiliations should be listed once. If an author has multiple
affiliations, multiple superscripts should be placed after the name, separated
by thin spaces. If the authors would like to highlight equal contribution by
multiple first authors, those authors should have an asterisk placed after their
name in superscript, and the term ``\textsuperscript{*}Equal contribution"
should be placed in the footnote block ahead of the list of affiliations. A
list of corresponding authors and their emails (in the format Full Name
\textless{}[email protected]\textgreater{}) can follow the list of affiliations.
Ideally only one or two names should be listed.
A sample file with author names is included in the ICML2021 style file
package. Turn on the \texttt{[accepted]} option to the stylefile to
see the names rendered. All of the guidelines above are implemented
by the \LaTeX\ style file.
\subsection{Abstract}
The paper abstract should begin in the left column, 0.4~inches below the final
address. The heading `Abstract' should be centered, bold, and in 11~point type.
The abstract body should use 10~point type, with a vertical spacing of
11~points, and should be indented 0.25~inches more than normal on left-hand and
right-hand margins. Insert 0.4~inches of blank space after the body. Keep your
abstract brief and self-contained, limiting it to one paragraph and roughly 4--6
sentences. Gross violations will require correction at the camera-ready phase.
\subsection{Partitioning the Text}
You should organize your paper into sections and paragraphs to help
readers place a structure on the material and understand its
contributions.
\subsubsection{Sections and Subsections}
Section headings should be numbered, flush left, and set in 11~pt bold
type with the content words capitalized. Leave 0.25~inches of space
before the heading and 0.15~inches after the heading.
Similarly, subsection headings should be numbered, flush left, and set
in 10~pt bold type with the content words capitalized. Leave
0.2~inches of space before the heading and 0.13~inches afterward.
Finally, subsubsection headings should be numbered, flush left, and
set in 10~pt small caps with the content words capitalized. Leave
0.18~inches of space before the heading and 0.1~inches after the
heading.
Please use no more than three levels of headings.
\subsubsection{Paragraphs and Footnotes}
Within each section or subsection, you should further partition the
paper into paragraphs. Do not indent the first line of a given
paragraph, but insert a blank line between succeeding ones.
You can use footnotes\footnote{Footnotes
should be complete sentences.} to provide readers with additional
information about a topic without interrupting the flow of the paper.
Indicate footnotes with a number in the text where the point is most
relevant. Place the footnote in 9~point type at the bottom of the
column in which it appears. Precede the first footnote in a column
with a horizontal rule of 0.8~inches.\footnote{Multiple footnotes can
appear in each column, in the same order as they appear in the text,
but spread them across columns and pages if possible.}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{icml_numpapers}}
\caption{Historical locations and number of accepted papers for International
Machine Learning Conferences (ICML 1993 -- ICML 2008) and International
Workshops on Machine Learning (ML 1988 -- ML 1992). At the time this figure was
produced, the number of accepted papers for ICML 2008 was unknown and instead
estimated.}
\label{icml-historical}
\end{center}
\vskip -0.2in
\end{figure}
\subsection{Figures}
You may want to include figures in the paper to illustrate
your approach and results. Such artwork should be centered,
legible, and separated from the text. Lines should be dark and at
least 0.5~points thick for purposes of reproduction, and text should
not appear on a gray background.
Label all distinct components of each figure. If the figure takes the
form of a graph, then give a name for each axis and include a legend
that briefly describes each curve. Do not include a title inside the
figure; instead, the caption should serve this function.
Number figures sequentially, placing the figure number and caption
\emph{after} the graphics, with at least 0.1~inches of space before
the caption and 0.1~inches after it, as in
Figure~\ref{icml-historical}. The figure caption should be set in
9~point type and centered unless it runs two or more lines, in which
case it should be flush left. You may float figures to the top or
bottom of a column, and you may set wide figures across both columns
(use the environment \texttt{figure*} in \LaTeX). Always place
two-column figures at the top or bottom of the page.
\subsection{Algorithms}
If you are using \LaTeX, please use the ``algorithm'' and ``algorithmic''
environments to format pseudocode. These require
the corresponding stylefiles, algorithm.sty and
algorithmic.sty, which are supplied with this package.
Algorithm~\ref{alg:example} shows an example.
\begin{algorithm}[tb]
\caption{Bubble Sort}
\label{alg:example}
\begin{algorithmic}
\STATE {\bfseries Input:} data $x_i$, size $m$
\REPEAT
\STATE Initialize $noChange = true$.
\FOR{$i=1$ {\bfseries to} $m-1$}
\IF{$x_i > x_{i+1}$}
\STATE Swap $x_i$ and $x_{i+1}$
\STATE $noChange = false$
\ENDIF
\ENDFOR
\UNTIL{$noChange$ is $true$}
\end{algorithmic}
\end{algorithm}
\subsection{Tables}
You may also want to include tables that summarize material. Like
figures, these should be centered, legible, and numbered consecutively.
However, place the title \emph{above} the table with at least
0.1~inches of space before the title and the same after it, as in
Table~\ref{sample-table}. The table title should be set in 9~point
type and centered unless it runs two or more lines, in which case it
should be flush left.
\begin{table}[t]
\caption{Classification accuracies for naive Bayes and flexible
Bayes on various data sets.}
\label{sample-table}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\begin{tabular}{lcccr}
\toprule
Data set & Naive & Flexible & Better? \\
\midrule
Breast & 95.9$\pm$ 0.2& 96.7$\pm$ 0.2& $\surd$ \\
Cleveland & 83.3$\pm$ 0.6& 80.0$\pm$ 0.6& $\times$\\
Glass2 & 61.9$\pm$ 1.4& 83.8$\pm$ 0.7& $\surd$ \\
Credit & 74.8$\pm$ 0.5& 78.3$\pm$ 0.6& \\
Horse & 73.3$\pm$ 0.9& 69.7$\pm$ 1.0& $\times$\\
Meta & 67.1$\pm$ 0.6& 76.5$\pm$ 0.5& $\surd$ \\
Pima & 75.1$\pm$ 0.6& 73.9$\pm$ 0.5& \\
Vehicle & 44.9$\pm$ 0.6& 61.5$\pm$ 0.4& $\surd$ \\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -0.1in
\end{table}
Tables contain textual material, whereas figures contain graphical material.
Specify the contents of each row and column in the table's topmost
row. Again, you may float tables to a column's top or bottom, and set
wide tables across both columns. Place two-column tables at the
top or bottom of the page.
\subsection{Citations and References}
Please use APA reference format regardless of your formatter
or word processor. If you rely on the \LaTeX\/ bibliographic
facility, use \texttt{natbib.sty} and \texttt{icml2021.bst}
included in the style-file package to obtain this format.
Citations within the text should include the authors' last names and
year. If the authors' names are included in the sentence, place only
the year in parentheses, for example when referencing Arthur Samuel's
pioneering work \yrcite{Samuel59}. Otherwise place the entire
reference in parentheses with the authors and year separated by a
comma \cite{Samuel59}. List multiple references separated by
semicolons \cite{kearns89,Samuel59,mitchell80}. Use the `et~al.'
construct only for citations with three or more authors or after
listing all authors to a publication in an earlier reference \cite{MachineLearningI}.
Authors should cite their own work in the third person
in the initial version of their paper submitted for blind review.
Please refer to Section~\ref{author info} for detailed instructions on how to
cite your own papers.
Use an unnumbered first-level section heading for the references, and use a
hanging indent style, with the first line of the reference flush against the
left margin and subsequent lines indented by 10 points. The references at the
end of this document give examples for journal articles \cite{Samuel59},
conference publications \cite{langley00}, book chapters \cite{Newell81}, books
\cite{DudaHart2nd}, edited volumes \cite{MachineLearningI}, technical reports
\cite{mitchell80}, and dissertations \cite{kearns89}.
Alphabetize references by the surnames of the first authors, with
single author entries preceding multiple author entries. Order
references for the same authors by year of publication, with the
earliest first. Make sure that each reference includes all relevant
information (e.g., page numbers).
Please put some effort into making references complete, presentable, and
consistent. If using bibtex, please protect capital letters of names and
abbreviations in titles, for example, use \{B\}ayesian or \{L\}ipschitz
in your .bib file.
\section*{Software and Data}
If a paper is accepted, we strongly encourage the publication of software and data with the
camera-ready version of the paper whenever appropriate. This can be
done by including a URL in the camera-ready copy. However, \textbf{do not}
include URLs that reveal your institution or identity in your
submission for review. Instead, provide an anonymous URL or upload
the material as ``Supplementary Material'' into the CMT reviewing
system. Note that reviewers are not required to look at this material
when writing their review.
\section*{Acknowledgements}
\textbf{Do not} include acknowledgements in the initial version of
the paper submitted for blind review.
If a paper is accepted, the final camera-ready version can (and
probably should) include acknowledgements. In this case, please
place such acknowledgements in an unnumbered section at the
end of the paper. Typically, this will include thanks to reviewers
who gave useful comments, to colleagues who contributed to the ideas,
and to funding agencies and corporate sponsors that provided financial
support.
\nocite{langley00}
\section{Electronic Submission}
\label{submission}
Submission to ICML 2021 will be entirely electronic, via a web site
(not email). Information about the submission process and \LaTeX\ templates
are available on the conference web site at:
\begin{center}
\textbf{\texttt{http://icml.cc/}}
\end{center}
The guidelines below will be enforced for initial submissions and
camera-ready copies. Here is a brief summary:
\begin{itemize}
\item Submissions must be in PDF\@.
\item Submitted papers can be up to eight pages long, not including references, plus unlimited space for references. Accepted papers can be up to nine pages long, not including references, to allow authors to address reviewer comments. Any paper exceeding this length will automatically be rejected.
\item \textbf{Do not include author information or acknowledgements} in your
initial submission.
\item Your paper should be in \textbf{10 point Times font}.
\item Make sure your PDF file only uses Type-1 fonts.
\item Place figure captions \emph{under} the figure (and omit titles from inside
the graphic file itself). Place table captions \emph{over} the table.
\item References must include page numbers whenever possible and be as complete
as possible. Place multiple citations in chronological order.
\item Do not alter the style template; in particular, do not compress the paper
format by reducing the vertical spaces.
\item Keep your abstract brief and self-contained, one paragraph and roughly
4--6 sentences. Gross violations will require correction at the
camera-ready phase. The title should have content words capitalized.
\end{itemize}
\subsection{Submitting Papers}
\textbf{Paper Deadline:} The deadline for paper submission that is
advertised on the conference website is strict. If your full,
anonymized, submission does not reach us on time, it will not be
considered for publication.
\textbf{Anonymous Submission:} ICML uses double-blind review: no identifying
author information may appear on the title page or in the paper
itself. Section~\ref{author info} gives further details.
\textbf{Simultaneous Submission:} ICML will not accept any paper which,
at the time of submission, is under review for another conference or
has already been published. This policy also applies to papers that
overlap substantially in technical content with conference papers
under review or previously published. ICML submissions must not be
submitted to other conferences and journals during ICML's review
period.
Informal publications, such as technical
reports or papers in workshop proceedings which do not appear in
print, do not fall under these restrictions.
\medskip
Authors must provide their manuscripts in \textbf{PDF} format.
Furthermore, please make sure that files contain only embedded Type-1 fonts
(e.g.,~using the program \texttt{pdffonts} in linux or using
File/DocumentProperties/Fonts in Acrobat). Other fonts (like Type-3)
might come from graphics files imported into the document.
Authors using \textbf{Word} must convert their document to PDF\@. Most
of the latest versions of Word have the facility to do this
automatically. Submissions will not be accepted in Word format or any
format other than PDF\@. Really. We're not joking. Don't send Word.
Those who use \textbf{\LaTeX} should avoid including Type-3 fonts.
Those using \texttt{latex} and \texttt{dvips} may need the following
two commands:
{\footnotesize
\begin{verbatim}
dvips -Ppdf -tletter -G0 -o paper.ps paper.dvi
ps2pdf paper.ps
\end{verbatim}}
It is a zero following the ``-G'', which tells dvips to use
the config.pdf file. Newer \TeX\ distributions don't always need this
option.
Using \texttt{pdflatex} rather than \texttt{latex}, often gives better
results. This program avoids the Type-3 font problem, and supports more
advanced features in the \texttt{microtype} package.
\textbf{Graphics files} should be a reasonable size, and included from
an appropriate format. Use vector formats (.eps/.pdf) for plots,
lossless bitmap formats (.png) for raster graphics with sharp lines, and
jpeg for photo-like images.
The style file uses the \texttt{hyperref} package to make clickable
links in documents. If this causes problems for you, add
\texttt{nohyperref} as one of the options to the \texttt{icml2021}
usepackage statement.
\subsection{Submitting Final Camera-Ready Copy}
The final versions of papers accepted for publication should follow the
same format and naming convention as initial submissions, except that
author information (names and affiliations) should be given. See
Section~\ref{final author} for formatting instructions.
The footnote, ``Preliminary work. Under review by the International
Conference on Machine Learning (ICML). Do not distribute.'' must be
modified to ``\textit{Proceedings of the
$\mathit{38}^{th}$ International Conference on Machine Learning},
Online, PMLR 139, 2021.
Copyright 2021 by the author(s).''
For those using the \textbf{\LaTeX} style file, this change (and others) is
handled automatically by simply changing
$\mathtt{\backslash usepackage\{icml2021\}}$ to
$$\mathtt{\backslash usepackage[accepted]\{icml2021\}}$$
Authors using \textbf{Word} must edit the
footnote on the first page of the document themselves.
Camera-ready copies should have the title of the paper as running head
on each page except the first one. The running title consists of a
single line centered above a horizontal rule which is $1$~point thick.
The running head should be centered, bold and in $9$~point type. The
rule should be $10$~points above the main text. For those using the
\textbf{\LaTeX} style file, the original title is automatically set as running
head using the \texttt{fancyhdr} package which is included in the ICML
2021 style file package. In case that the original title exceeds the
size restrictions, a shorter form can be supplied by using
\verb|\icmltitlerunning{...}|
just before $\mathtt{\backslash begin\{document\}}$.
Authors using \textbf{Word} must edit the header of the document themselves.
\section{Format of the Paper}
All submissions must follow the specified format.
\subsection{Dimensions}
The text of the paper should be formatted in two columns, with an
overall width of 6.75~inches, height of 9.0~inches, and 0.25~inches
between the columns. The left margin should be 0.75~inches and the top
margin 1.0~inch (2.54~cm). The right and bottom margins will depend on
whether you print on US letter or A4 paper, but all final versions
must be produced for US letter size.
The paper body should be set in 10~point type with a vertical spacing
of 11~points. Please use Times typeface throughout the text.
\subsection{Title}
The paper title should be set in 14~point bold type and centered
between two horizontal rules that are 1~point thick, with 1.0~inch
between the top rule and the top edge of the page. Capitalize the
first letter of content words and put the rest of the title in lower
case.
\subsection{Author Information for Submission}
\label{author info}
ICML uses double-blind review, so author information must not appear. If
you are using \LaTeX\/ and the \texttt{icml2021.sty} file, use
\verb+\icmlauthor{...}+ to specify authors and \verb+\icmlaffiliation{...}+ to specify affiliations. (Read the TeX code used to produce this document for an example usage.) The author information
will not be printed unless \texttt{accepted} is passed as an argument to the
style file.
Submissions that include the author information will not
be reviewed.
\subsubsection{Self-Citations}
If you are citing published papers for which you are an author, refer
to yourself in the third person. In particular, do not use phrases
that reveal your identity (e.g., ``in previous work \cite{langley00}, we
have shown \ldots'').
Do not anonymize citations in the reference section. The only exception are manuscripts that are
not yet published (e.g., under submission). If you choose to refer to
such unpublished manuscripts \cite{anonymous}, anonymized copies have
to be submitted
as Supplementary Material via CMT\@. However, keep in mind that an ICML
paper should be self contained and should contain sufficient detail
for the reviewers to evaluate the work. In particular, reviewers are
not required to look at the Supplementary Material when writing their
review.
\subsubsection{Camera-Ready Author Information}
\label{final author}
If a paper is accepted, a final camera-ready copy must be prepared.
For camera-ready papers, author information should start 0.3~inches below the
bottom rule surrounding the title. The authors' names should appear in 10~point
bold type, in a row, separated by white space, and centered. Author names should
not be broken across lines. Unbolded superscripted numbers, starting 1, should
be used to refer to affiliations.
Affiliations should be numbered in the order of appearance. A single footnote
block of text should be used to list all the affiliations. (Academic
affiliations should list Department, University, City, State/Region, Country.
Similarly for industrial affiliations.)
Each distinct affiliations should be listed once. If an author has multiple
affiliations, multiple superscripts should be placed after the name, separated
by thin spaces. If the authors would like to highlight equal contribution by
multiple first authors, those authors should have an asterisk placed after their
name in superscript, and the term ``\textsuperscript{*}Equal contribution"
should be placed in the footnote block ahead of the list of affiliations. A
list of corresponding authors and their emails (in the format Full Name
\textless{}[email protected]\textgreater{}) can follow the list of affiliations.
Ideally only one or two names should be listed.
A sample file with author names is included in the ICML2021 style file
package. Turn on the \texttt{[accepted]} option to the stylefile to
see the names rendered. All of the guidelines above are implemented
by the \LaTeX\ style file.
\subsection{Abstract}
The paper abstract should begin in the left column, 0.4~inches below the final
address. The heading `Abstract' should be centered, bold, and in 11~point type.
The abstract body should use 10~point type, with a vertical spacing of
11~points, and should be indented 0.25~inches more than normal on left-hand and
right-hand margins. Insert 0.4~inches of blank space after the body. Keep your
abstract brief and self-contained, limiting it to one paragraph and roughly 4--6
sentences. Gross violations will require correction at the camera-ready phase.
\subsection{Partitioning the Text}
You should organize your paper into sections and paragraphs to help
readers place a structure on the material and understand its
contributions.
\subsubsection{Sections and Subsections}
Section headings should be numbered, flush left, and set in 11~pt bold
type with the content words capitalized. Leave 0.25~inches of space
before the heading and 0.15~inches after the heading.
Similarly, subsection headings should be numbered, flush left, and set
in 10~pt bold type with the content words capitalized. Leave
0.2~inches of space before the heading and 0.13~inches afterward.
Finally, subsubsection headings should be numbered, flush left, and
set in 10~pt small caps with the content words capitalized. Leave
0.18~inches of space before the heading and 0.1~inches after the
heading.
Please use no more than three levels of headings.
\subsubsection{Paragraphs and Footnotes}
Within each section or subsection, you should further partition the
paper into paragraphs. Do not indent the first line of a given
paragraph, but insert a blank line between succeeding ones.
You can use footnotes\footnote{Footnotes
should be complete sentences.} to provide readers with additional
information about a topic without interrupting the flow of the paper.
Indicate footnotes with a number in the text where the point is most
relevant. Place the footnote in 9~point type at the bottom of the
column in which it appears. Precede the first footnote in a column
with a horizontal rule of 0.8~inches.\footnote{Multiple footnotes can
appear in each column, in the same order as they appear in the text,
but spread them across columns and pages if possible.}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{icml_numpapers}}
\caption{Historical locations and number of accepted papers for International
Machine Learning Conferences (ICML 1993 -- ICML 2008) and International
Workshops on Machine Learning (ML 1988 -- ML 1992). At the time this figure was
produced, the number of accepted papers for ICML 2008 was unknown and instead
estimated.}
\label{icml-historical}
\end{center}
\vskip -0.2in
\end{figure}
\subsection{Figures}
You may want to include figures in the paper to illustrate
your approach and results. Such artwork should be centered,
legible, and separated from the text. Lines should be dark and at
least 0.5~points thick for purposes of reproduction, and text should
not appear on a gray background.
Label all distinct components of each figure. If the figure takes the
form of a graph, then give a name for each axis and include a legend
that briefly describes each curve. Do not include a title inside the
figure; instead, the caption should serve this function.
Number figures sequentially, placing the figure number and caption
\emph{after} the graphics, with at least 0.1~inches of space before
the caption and 0.1~inches after it, as in
Figure~\ref{icml-historical}. The figure caption should be set in
9~point type and centered unless it runs two or more lines, in which
case it should be flush left. You may float figures to the top or
bottom of a column, and you may set wide figures across both columns
(use the environment \texttt{figure*} in \LaTeX). Always place
two-column figures at the top or bottom of the page.
\subsection{Algorithms}
If you are using \LaTeX, please use the ``algorithm'' and ``algorithmic''
environments to format pseudocode. These require
the corresponding stylefiles, algorithm.sty and
algorithmic.sty, which are supplied with this package.
Algorithm~\ref{alg:example} shows an example.
\begin{algorithm}[tb]
\caption{Bubble Sort}
\label{alg:example}
\begin{algorithmic}
\STATE {\bfseries Input:} data $x_i$, size $m$
\REPEAT
\STATE Initialize $noChange = true$.
\FOR{$i=1$ {\bfseries to} $m-1$}
\IF{$x_i > x_{i+1}$}
\STATE Swap $x_i$ and $x_{i+1}$
\STATE $noChange = false$
\ENDIF
\ENDFOR
\UNTIL{$noChange$ is $true$}
\end{algorithmic}
\end{algorithm}
\subsection{Tables}
You may also want to include tables that summarize material. Like
figures, these should be centered, legible, and numbered consecutively.
However, place the title \emph{above} the table with at least
0.1~inches of space before the title and the same after it, as in
Table~\ref{sample-table}. The table title should be set in 9~point
type and centered unless it runs two or more lines, in which case it
should be flush left.
\begin{table}[t]
\caption{Classification accuracies for naive Bayes and flexible
Bayes on various data sets.}
\label{sample-table}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\begin{tabular}{lcccr}
\toprule
Data set & Naive & Flexible & Better? \\
\midrule
Breast & 95.9$\pm$ 0.2& 96.7$\pm$ 0.2& $\surd$ \\
Cleveland & 83.3$\pm$ 0.6& 80.0$\pm$ 0.6& $\times$\\
Glass2 & 61.9$\pm$ 1.4& 83.8$\pm$ 0.7& $\surd$ \\
Credit & 74.8$\pm$ 0.5& 78.3$\pm$ 0.6& \\
Horse & 73.3$\pm$ 0.9& 69.7$\pm$ 1.0& $\times$\\
Meta & 67.1$\pm$ 0.6& 76.5$\pm$ 0.5& $\surd$ \\
Pima & 75.1$\pm$ 0.6& 73.9$\pm$ 0.5& \\
Vehicle & 44.9$\pm$ 0.6& 61.5$\pm$ 0.4& $\surd$ \\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -0.1in
\end{table}
Tables contain textual material, whereas figures contain graphical material.
Specify the contents of each row and column in the table's topmost
row. Again, you may float tables to a column's top or bottom, and set
wide tables across both columns. Place two-column tables at the
top or bottom of the page.
\subsection{Citations and References}
Please use APA reference format regardless of your formatter
or word processor. If you rely on the \LaTeX\/ bibliographic
facility, use \texttt{natbib.sty} and \texttt{icml2021.bst}
included in the style-file package to obtain this format.
Citations within the text should include the authors' last names and
year. If the authors' names are included in the sentence, place only
the year in parentheses, for example when referencing Arthur Samuel's
pioneering work \yrcite{Samuel59}. Otherwise place the entire
reference in parentheses with the authors and year separated by a
comma \cite{Samuel59}. List multiple references separated by
semicolons \cite{kearns89,Samuel59,mitchell80}. Use the `et~al.'
construct only for citations with three or more authors or after
listing all authors to a publication in an earlier reference \cite{MachineLearningI}.
Authors should cite their own work in the third person
in the initial version of their paper submitted for blind review.
Please refer to Section~\ref{author info} for detailed instructions on how to
cite your own papers.
Use an unnumbered first-level section heading for the references, and use a
hanging indent style, with the first line of the reference flush against the
left margin and subsequent lines indented by 10 points. The references at the
end of this document give examples for journal articles \cite{Samuel59},
conference publications \cite{langley00}, book chapters \cite{Newell81}, books
\cite{DudaHart2nd}, edited volumes \cite{MachineLearningI}, technical reports
\cite{mitchell80}, and dissertations \cite{kearns89}.
Alphabetize references by the surnames of the first authors, with
single author entries preceding multiple author entries. Order
references for the same authors by year of publication, with the
earliest first. Make sure that each reference includes all relevant
information (e.g., page numbers).
Please put some effort into making references complete, presentable, and
consistent. If using bibtex, please protect capital letters of names and
abbreviations in titles, for example, use \{B\}ayesian or \{L\}ipschitz
in your .bib file.
\section*{Software and Data}
If a paper is accepted, we strongly encourage the publication of software and data with the
camera-ready version of the paper whenever appropriate. This can be
done by including a URL in the camera-ready copy. However, \textbf{do not}
include URLs that reveal your institution or identity in your
submission for review. Instead, provide an anonymous URL or upload
the material as ``Supplementary Material'' into the CMT reviewing
system. Note that reviewers are not required to look at this material
when writing their review.
\section*{Acknowledgements}
\textbf{Do not} include acknowledgements in the initial version of
the paper submitted for blind review.
If a paper is accepted, the final camera-ready version can (and
probably should) include acknowledgements. In this case, please
place such acknowledgements in an unnumbered section at the
end of the paper. Typically, this will include thanks to reviewers
who gave useful comments, to colleagues who contributed to the ideas,
and to funding agencies and corporate sponsors that provided financial
support.
\nocite{langley00}
\section{Introduction}
This is where the content of your paper goes.
\begin{itemize}
\item Limit the main text (not counting references and appendices) to 12 PMLR-formatted pages, using this template. Please add any additional appendix to the same file after references - there is no page limit for the appendix.
\item Include, either in the main text or the appendices, \emph{all} details, proofs and derivations required to substantiate the results.
\item The contribution, novelty and significance of submissions will be judged primarily based on
\textit{the main text of 12 pages}. Thus, include enough details, and overview of key arguments,
to convince the reviewers of the validity of result statements, and to ease parsing of technical material in the appendix.
\item Use the \textbackslash documentclass[anon,12pt]\{colt2021\} option during submission process -- this automatically hides the author names listed under \textbackslash coltauthor. Submissions should NOT include author names or other identifying information in the main text or appendix. To the extent possible, you should avoid including directly identifying information in the text. You should still include all relevant references, discussion, and scientific content, even if this might provide significant hints as to the author identity. But you should generally refer to your own prior work in third person. Do not include acknowledgments in the submission. They can be added in the camera-ready version of accepted papers.
Please note that while submissions must be anonymized, and author names are withheld from reviewers, they are known to the area chair overseeing the paper’s review. The assigned area chair is allowed to reveal author names to a reviewer during the rebuttal period, upon the reviewer’s request, if they deem such information is needed in ensuring a proper review.
\item Use \textbackslash documentclass[final,12pt]\{colt2021\} only during camera-ready submission.
\end{itemize}
\acks{We thank a bunch of people and funding agency.}
\section{Introduction}
This is where the content of your paper goes.
\begin{itemize}
\item Limit the main text (not counting references and appendices) to 12 PMLR-formatted pages, using this template. Please add any additional appendix to the same file after references - there is no page limit for the appendix.
\item Include, either in the main text or the appendices, \emph{all} details, proofs and derivations required to substantiate the results.
\item The contribution, novelty and significance of submissions will be judged primarily based on
\textit{the main text of 12 pages}. Thus, include enough details, and overview of key arguments,
to convince the reviewers of the validity of result statements, and to ease parsing of technical material in the appendix.
\item Use the \textbackslash documentclass[anon,12pt]\{colt2021\} option during submission process -- this automatically hides the author names listed under \textbackslash coltauthor. Submissions should NOT include author names or other identifying information in the main text or appendix. To the extent possible, you should avoid including directly identifying information in the text. You should still include all relevant references, discussion, and scientific content, even if this might provide significant hints as to the author identity. But you should generally refer to your own prior work in third person. Do not include acknowledgments in the submission. They can be added in the camera-ready version of accepted papers.
Please note that while submissions must be anonymized, and author names are withheld from reviewers, they are known to the area chair overseeing the paper’s review. The assigned area chair is allowed to reveal author names to a reviewer during the rebuttal period, upon the reviewer’s request, if they deem such information is needed in ensuring a proper review.
\item Use \textbackslash documentclass[final,12pt]\{colt2021\} only during camera-ready submission.
\end{itemize}
\acks{We thank a bunch of people and funding agency.}
|
\section{Conclusion}
In this work, we presented the results from a survey of 2260 adults about their
experiences using online learning resources to learn about different topics. Our
survey was balanced to include a diverse sample.
We found that over 90\% of respondents reported learning online about subjects
as diverse as math, politics, and the arts. The most popular subjects that
respondents learned about online included DIY, personal health care, history,
physical fitness, and travel \& geography. The proportion of respondents
interested in the various subjects in our survey followed a power law distribution.
The most popular types of online learning resources included YouTube, informational
articles, and Wikipedia. On average, participants used YouTube to learn about three
different subjects, and Wikipedia and informational articles to learn about
more than two different subjects. Respondents reported using every resource in
our survey to learn about every subject. The most popular subject on YouTube
was makeup \& fashion, while the most popular subject on Wikipedia was history,
and personal health care for informational articles.
Further analysis of our results showed that respondents learned about subjects
like math, law, and computer programming from more formal online resources. However,
for subjects like DIY, physical fitness, and travel \& geography, respondents
more often learned from informal resources. Some respondents used formal and
informal online learning resources equally to learn about subjects like history,
the arts, and politics. Finally, a clustering analysis revealed that YouTube is
the resource that has the most overlap among all of our respondents' online
learning experiences.
To conclude, we discussed the importance of video and text as technologies that
are fundamental to the web itself, and serve as especially important
infrastructure for online education. We also explored how the forces of supply
and demand drive the availability of online educational experiences. Finally,
we set our sights on the future by examining factors that impact resources
that learners seek out, and how online education will be shaped by the COVID-19
pandemic.
\section{Discussion}
\label{sec:discussion}
Overall, we find that online learning is a ubiquitous experience: 93\% of
respondents in our survey reported learning online. YouTube, in particular, is by far
the most popular online learning resource. Not only is YouTube used by
the most respondents for learning (88\%), it is the most popular resource
for learning about 11 out of the 19 subjects examined in our survey. The
fact that, on average, respondents used YouTube to learn about more than
three subjects further underscores the importance of this resource. Moreover,
YouTube is not only popular, it is core to the experience of
online learning. Our clustering analyses consistently show that
YouTube is at the core of other online learning experiences. This result is consistent with previous studies that have found that
YouTube spans across formal and informal online learning resources~\cite{Rosenthal-2018}. YouTube recently redesigned an influential page where trending videos are displayed, adding a ``Learning'' category~\cite{yt-edu}, perhaps in response to this trend.
In this section, we discuss
which mechanisms of online learning are most pervasive and what the
prevalence of particular mechanisms implies about the online learning
ecosystem and how we design for it.
Additionally, we point out directions for future work based on our results.
We conclude with a look toward the
future: what implications our findings have for the future of online
learning in a world after COVID-19.
\subsection{Supply, Familiarity, and Ease As Potential Drivers of Leaning Resource Choice}
As noted earlier, YouTube is the most popular resource for online learning. Perhaps
unsurprisingly, informational articles and Wikipedia are the next most
popular types of online learning resources considering that text-based
articles and web pages were the first and most fundamental parts of the
internet ~\cite{tbl-www}. Indeed, the four core resources identified by our clustering -- YouTube,
informational articles, Wikipedia, and how-to guides -- are built around
three legacy internet technologies: video, static text, and wikis. YouTube
was founded over 15 years ago ~\cite{yt-old}, static text websites have
been around since the beginning of the internet, and wikis are a technology
that is over 25 years old ~\cite{ebersbach-2008}.
We hypothesize that the popularity of these legacy resources has four potential causes. First, adult learners may find it easiest -- or prefer -- to learn through informal video media (e.g., YouTube) and informational articles, and thus leverage these resources the most. Second, due in part to having a long period of time over which to develop content, these popular learning resources are able to supply a high volume of content across many topics. For example, YouTube is the leading video platform -- in terms
of both hours of content and revenue~\cite{yt-content,yt-money}. As such, it may be easiest for learners to find learning content on YouTube, thus leading them to learn using video media. Third, over extended use, Internet users' may have developed \textit{familiarity} with learning from video and/or informational-article style content. This familiarity may lead them to continue to turn to learning modes from which they are comfortable receiving content instead of exploring newer learning technologies such as interactive tutorials, MOOCs and question-and-answer forums, which may also have a lower supply of content. Finally, the popularity of these resources may create a cyclical effect: because these resources are more popular, and offer more supply, they are easier to find in search results, and thus learners are more likely to turn to them. For example, prior work has suggested that some of
Wikipedia's popularity could be attributed to how highly it is ranked in search engine results ~\cite{mcmahon-2017}. These findings raise important questions for future work to investigate regarding whether these resources actually best meet learner needs, or are merely used out of the convenience due to supply, familiarity, and/or ease of access. We delve deeper into how our findings support or refute each of these potential drivers behind learners' choice of online learning resources below.
\subsection{Low Subject Diversity Among Formal Resources}
Our work finds that adult online learners
have extremely diverse interests, \hl{with more than half (54\%)} learning about 5 or more different subjects. As mentioned above, one of the advantages of
YouTube and Wikipedia (as well as informational articles) is supply:
these resources host content about a wide array of subjects. However, newer
online learning resources may lack this volume of content, either by choice or due to
their relative novelty. This lack of diversity may inherently influence the mechanism
of online learning. Let us take as a case study law and programming.
Law and computer programming carry wide-ranging societal influence but were
two of the least popular subjects among our respondents, with about 15\% of
respondents indicating they learned about either subject. Despite the equal
popularity of these subjects, programming was one of the most popular
subjects for two types of resources (course materials, interactive tutorials),
while law was 12th most popular of the 19 subjects respondents learned
using Wikipedia, and ranked lower in terms of popularity for every other
type of online learning resource. While it is possible that respondents
use a wide variety of resources for learning about law because of the
nature of the subject, this may also be the case because less online
course content is available for law vs. for programming. Programming
has numerous free, online course style resources such as Codecademy and open source
scholarly projects like Python Tutor ~\cite{sims-2011,guo-2013} that put
interactive programming tutorials online, in addition to more traditional
online courses -- either free or payment-based -- such as those offered by universities or through
MOOCs. Our findings suggest that learners may turn to newer resources tailored to their learning needs for a particular subject if those resources are available. \hl{In the absence of such tailored resources, however,} online learners may turn to the resources that they know will offer sufficient, affordable, supply: YouTube and Wikipedia.
\hl{By highlighting this contrast between law and computer science we do not mean
to suggest that low subject diversity among formal online learning resources is
necessarily a problem to be addressed. We believe that this example illustrates
with data that the way people learn different subjects online is nuanced in ways
that are not intuitive. Understanding these kinds of gradations in the online
learning landscape may also call attention to opportunities for new kinds of
online learning experiences.}
\subsection{Issues of Supply May Relate to Ease of Resource Creation}
These issues of supply may relate to the ease with which resource creators can develop new content.
The relatively low popularity of new learning resources (e.g., MOOCS, interactive tutorials) highlights the absence of the diffusion of
innovations ~\cite{diffusion-1985} we might expect given years of interest
in new types of educational media and interactive online learning
technologies ~\cite{hazen2012proposed}. This absence may be related to the ease with which these resources can be
created. Phones, tablets, and personal computers are now often embedded with more
than one camera, making the creation of (educational) videos easier than
ever before, and the tools for publishing text-based articles are just as
widespread. However, relatively less robust and cost-effective support is available for the creation of newer resources. For example, many universities have had to scale up production studios in which to create MOOCS~\cite{baker2016value}, purchasing expensive equipment out of reach for many of the learning content creators who utilize YouTube. Similarly, many interactive tutorials involve the creation of new software products, requiring significant grant funding and time investment for development~\cite{sims-2011,guo-2013}. Thus, future work on the democratization of online education~\cite{kross2020democratization} may wish to consider how to democritize the \textit{creation} of content, in an effort to improve the learning content available for consumption.
\input{learning-spectrum}
\subsection{The Rapidly Changing Online Learning Landscape}
During the preparation of this manuscript, COVID-19 emerged causing major
disruptions and reorganization to how educational experiences are structured
and delivered. Major universities had to transition all of their courses online
quickly to minimize in-person interactions and to comply with new government
regulations designed to protect public health ~\cite{ed-corona}. We believe
it is reasonable to say that the emergence of COVID-19 is the most significant
event in the nascent history of online education, as the pandemic precipitated
a situation where education in any form became nearly synonymous with online
education.
Although the long-term effects of COVID-19 on education at large have yet to
be realized, several short-term effects related to this study are taking form.
Since the proliferation of the disease and the resulting lockdowns, more
adult learners have been seeking online educational
experiences ~\cite{adult-learners-online}. This may be related to the
financial recession caused by COVID-19, since it is well understood that
recessions drive increased interest in adult educational
programs ~\cite{economy-school}. However, given that most college
campuses are closed, adult learners are more frequently turning to online
learning resources ~\cite{adult-learners-online}. Massively open online
course providers in particular have seen a record-breaking growth
in enrollments ~\cite{remember-moocs}. This is a significant development
in the trajectory of MOOCs as prior work has reflected on the role of MOOCs
in online education as modeled by their path through the Gartner Hype
Cycle ~\cite{hype-cycle,kross-2018}.
The data in this study represent a snapshot of the adult online learning
landscape just months before COVID-19. We believe that our results hold
valuable insights that can inform the future design of online learning
technologies. However, this study also inadvertently and advantageously
provides a baseline that can be used in the future to measure how
this monumental shift in online learning is changing interests in
subjects or the differential use of types of online learning resources as
both are influenced by the global pandemic.
\section{Introduction}
The subject of how people learn has been fascinating researchers since long
before the invention of the
internet ~\cite{bradford-1958,mower-1960,rogers-1957,stengel-1939}. The
innovation of online learning has opened the door for more people to learn,
and thus, the door for more research into what people are learning about and
how they are doing so.
One of the most promising aspects of the internet since its early days has
been the opportunities it offers for widespread participation thanks to the
myriad of resources it makes freely available ~\cite{Benkler-2006}. While
much scholarly work has investigated what this means for the development of
resources like Wikipedia and other wikis (e.g., Shaw \& Hill, 2014),
researchers have paid much less attention to how people use diverse online
resources for educational purposes ~\cite{shaw-2014}. The internet offers a
plethora of formal and informal online resources available for free or at
significantly lower cost compared to traditional, offline educational
opportunities, and has thus widened the availability of educational
resources. Massive open online courses (MOOCs), video tutorials, how-to articles,
online discussion groups, among others, can help people gain new skills with
the potential to improve their job prospects, social mobility,
and personal welfare~\cite{hadavand2018can}.
Prior work has considered how a specific subset of people learn (e.g.,
young adults), how people learn a specific subject, or how people use a
particular resource for
learning ~\cite{shorey-2020,kross2020democratization,Narayan-2017,Torrey-2009}. Yet, little
is known about the \emph{full ecosystem} of what adults learn about online and what
educational resources they use to do so. These can include both formal online
courses (e.g., Coursera, those offered by local universities) and less formal
resources (e.g., YouTube videos, online discussion groups). No prior work, to
our knowledge, has addressed this full ecosystem by looking at more than one
subject or one resource at a time across all online learners. An important
novel contribution for this paper is that it considers the \emph{combination} of
subjects and resources (e.g., an online university course for learning
history, YouTube for learning math) that may generalize to many learners.
In this work, we explore three research questions:
\textbf{RQ1)} What do people learn about online?
\textbf{RQ2)} How do people learn online (i.e., what learning resources do they use)?
\textbf{RQ3)} What are the core online learning experiences - which subject through what resource - that are common across the majority of online learners?
To answer these research questions, we conducted an online survey of 2,260 adults age 18 and over in the U.S. To improve the generalizability of our findings, our survey sample was demographically balanced to match U.S. Census statistics on gender, age, and education. We draw on frameworks and findings of past literature on online learning ~\cite{Piety-2014,Ferguson-2012,kross-2018} to analyze our findings critically and characterize the online learning ecosystem.
We find that the vast majority (93\%) of those we surveyed had learned something online. At a high level, we find that online learners' interests span a very wide set of subjects (with the median subject being learned by 26\% of respondents) and resources (with the median resource being used by 62.5\% of respondents). Further, examining online learning experiences (pairs of subjects and resources), we find 12 core online learning experiences shared by the plurality of online learners. The most common of these core online learning experiences, which was reported by over half of our respondents, was learning how to do something yourself (DIY) using YouTube.
The findings from this work provide insight into the ecosystem and experience of online learning; suggest a combination of factors that may drive learners' choices and affect our progress toward a democratization of online learning; and identify directions for future work on improving the design and targeting of online learning technologies to draw in new learners and onboard new subjects.
\subsection{Future work: Understanding Learners' Motivations}
Our results illustrate the breadth of adult online learning and raise a number of questions regarding why adults choose to learn using particular resources. Above, we hypothesized that learners may choose to learn using certain resources due to the supply of content available from certain resources, learners' familiarity with the mode of content, the ease with which learners can find that content, and with which educators can create it. However, online learners may also choose to learn particular topics with particular resources for reasons related to their own internal motivations, rather than externalities of the online learning ecosystem.
As discussed in prior work, there exists a spectrum of learning methods that may be used by, or which appeal to, different learners. This spectrum ranges from free choice learning -- e.g., visiting a museum, watching a documentary for entertainment~\cite{falk-2007} -- to traditional, academic, structured learning (e.g., taking a course). One of the contributions of our work is to answer the call~\cite{schwier-2012} to compare online learners' use of such informal vs. formal online resources. We find significant differences in the use of such resources (Table~\ref{tab:pair}).
We hypothesize that these differences may relate not only to supply, but to learners' goals. For example, credentials earned online are increasingly popular
across many fields, and demand for these credentials only seems to be growing
~\cite{caudill2017emerging}. We hypothesize that learners who want to earn a credential
are using online courses, practice exams, and course materials more often
compared to other online learning resources. In Figure \ref{fig:avgsub} we
can see that some of the subjects that most commonly use those resources are
math, programming, science, and languages, all subjects in which it can be
valuable to obtain credentials for professional purposes as opposed to the topics that we find are learned more frequently using informal resources such DIY and personal health care. Thus, we encourage future work building on the foundation laid by these findings, to investigate how the professional and credential-related motivations of online learners inform their choice of learning resources.
In addition to professionally-related credentials, our findings also suggest that learners may be motivated to seek more authoritative information about some topics, even if they are not pursuing a professional qualification. For example, respondents indicated that they used Wikipedia most for learning
about history, politics, religion \& ethics, and law (see Figure
\ref{fig:avgsub}), all topics that may be presented with significant bias
elsewhere. The Wikipedia community actively moderates their articles, lending a sense of \textit{community authority} ~\cite{sahut2017wikipedia} to their content, which may appeal to certain learners, or learners of particular subjects. As such, future work may seek not only to investigate learners' goals, but also the criteria through which they evaluate potential sources of learning information -- much like prior work has studied news consumers' evaluation of media and misinformation~\cite{scheufele2019science,lazer2018science}
\begin{comment}
Library science colleague raises the following "I think something that educational resources such as Youtube and Wikipedia have in common is some form of community endorsement or authority. What I mean is, they are created/produced by amateurs/non-professionals so they require community buy-in in order to get 'authority,' if that makes sense? So something descriptive about the 'amateur' educational community, perhaps vs. course materials which are authoritative because they come from a known source...
I think what's really interesting about these platforms with respect to information is precisely the role of authority-granting. Traditionally, education and journalism, and other information sources, established their authority through certifications, vetting processes, licensing, professional organizations, etc. To be a 'school' meant to be 'authorized' by some government, there was a curriculum approval process, an accreditation process, etc. Lots of hiring requirements. And all that was done in order to grant authority that the school was the best to provide an education to the masses.
Now we have these distributed platforms of information sharing and we are developing different ways of granting this authority. There is a tension between the 'established' old guard authorities (academies, universities, professors, teachers, school boards, etc) and these amateur resources. It's a really fraught space and yet, extremely exciting from an information/education POV."
This is an interesting point we could make, I think -- maybe people are willing to rely on community authority for some topics but not for others (e.g., math)
I think maybe writing a new section of discussion that focuses on supply/demand as an explanation for the topic distribution is a good idea, and I like the direction you're going here. E.g., I think we can just explain that MOOCs met a demand for credentialed education at scale, and thus we see people learning more "traditional" topics using them vs. free-choice learning where we see a wider spectrum of topics, and less emphasis on traditional courses.
No Credential <---> Credential
Free Choice <---> Motivated
Non Authority <---> Authority
No Credential
Free Choice
Non Authority
\end{comment}
\section{Methods}
We conducted a national survey of American adults to examine the relationship between
the subjects that people learn about online and the types of online learning
resources they use to do so. Here, we provide details about the data
collection, the statistical procedures we applied to the data, and the
limitations of our analysis.
\subsection{Data Collection}
We contracted with the survey research firm Cint to administer the study to
American adults in June-July, 2019. People were able to fill out the survey
hosted on the Qualtrics platform using a computer or a mobile device. We
included two attention-check verification questions and only those who passed
both are included in the analyses. We received 2,260 valid responses to our
survey.
We quota-sampled on gender, age, and education to obtain a diverse sample.
Specifically, our
respondents are 58.6\% female and the mean age of our respondents is 41 years
(SD=15 years). Additionally, 23.6\% of our sample has a high school diploma or
less education, 39.8\% of our sample has some college education, while the
remaining 36.6\% of our sample has a college degree or more education. The
survey panel provider compensated respondents for their participation based
on their preference of cash, gift cards, or donations to charity. All study
procedures and materials were approved by our Institutional Review Board.
\subsection{Survey Questionnaire}
To understand respondents' online learning experiences, we asked respondents
what \textbf{subjects} they had learned about online and what types of \textbf{resources} they
used to learn about each of those subjects\footnote{Respondents had the option
to select none of the options, thus indicating that they had not learned
online.}. A respondent who indicated that they used a particular type of
resource to learn about a particular subject had what we define as an
\textbf{\emph{online learning experience}}.
To understand what subjects people learned online, we asked ``Which of the
following topics have you tried to learn about through online resources?" This
was followed by nineteen subjects, the option of ``other" and specifying
something else, and ``none of the above." Subjects ranged from traditional
academic subjects like history, math, and science to general welfare and
lifestyle subjects such as: do-it-yourself (DIY)\footnote{This was phrased
as ``how-to (e.g., around the house, cooking/baking, etc.)''}, travel/geography,
personal health/health care, makeup/fashion, and online safety, security,
or privacy. The categories of subjects were developed through iterative
rounds of 10 cognitive interviews with participants of varied demographics (age, gender, socioeconomic status). In these interviews we prompted participants about whether there were any answers they wanted to provide, but which were not available for them to select. These interviews are a commonly-used technique for testing survey
questions~\cite{Presser-2004} and are not intended as research artifacts (see Section~\ref{sec:survey:validation} below for more details).
To understand how people learned online -- that is, what resources they use
to learn -- we included a matrix question that asked about how participants learned
each of the subjects they indicated in the prior question. Specifically,
respondents could select multiple options from a list of ten resources:
watched a video (e.g., YouTube), read Wikipedia, took an online course,
read an informational article, read a how-to-guide, used an interactive
tutorial, used a practice exam website, used materials from a course,
read answers to questions on an online discussion community or forum,
and asked questions in an online discussion community or forum.
Other than the first three of these response options, the others included
at least two examples (see below) in parentheses to help respondents
understand our categories. Informational articles include articles on
non-Wikipedia websites, such as privately-run wikis like Wikia, online
references like dictionaries and thesauri, and research-based resources
like museum websites. How-to guides include websites that help people
complete projects around their home as well as step-by-step guides for
understanding mathematics like BetterExplained.com. Examples of interactive
tutorials on the web we mentioned include platforms like Khan Academy,
Codecademy, and Duolingo. Practice exam sites like Kaplan and Magoosh focus
on preparation for standardized exams, while course materials refer to open
repositories of slides and other artifacts from courses usually taught in
person, like the MIT OpenCourseWare project. Finally, discussion communities
and forums include social-media forums like Facebook Groups and Reddit, in
addition to forums designed for professional communities like StackOverflow.
\hl{Given our interest in comparing informal online learning to formal online learning, which we understood as the online courses, those who indicated taking online courses were then presented with an additional question: ``What kinds of organizations have provided the online course(s) that you have enrolled in? (Check all that apply)'' The answer options were: Online course at a local university; Online course at a university elsewhere; Massive open online course (Coursera, FutureLearn, edX, Udacity, Udemy); Other online course (e.g. The Great Courses, Lynda.com).}
Overall, respondents could report up to 190 possible online learning
experiences (pairs of subjects and resources such as learning math on
YouTube or learning languages using an interactive tutorial), since we
asked about how respondents learned up to 19 subjects using up to 10 online
learning resources.\footnote{The parts of the survey questionnaire that were analyzed in this work can be found at \url{https://doi.org/10.5281/zenodo.4088916
}.}
\subsection{Survey Validation and Pretesting}
\label{sec:survey:validation}
During the in-person cognitive interviews, we asked interviewees to think
aloud as they were filling out the survey, and they were encouraged to share
their thoughts if they believed that they were missing answer choices or if
they were confused about the wording of a question. Based on the feedback
from these interviews we rewrote and rearranged our survey questions to
maximize their clarity and to ensure the completeness of the answer choices
we offered. Once our survey was deployed, less than one percent (0.88\%) of
respondents indicated that their desired answer choice was not available for
any of our questions (they selected ``Other" in response to the question),
confirming the completeness of the choices we provided.
\subsection{Data Analysis}
A respondent is considered to have learned online if they learned about at
least one subject using at least one resource.
To investigate what subjects people learn online and what resources they use
to learn them (RQ1 and RQ2), we use descriptive data analysis including
hypothesis tests and $\chi^{2}$ proportions tests as appropriate. Additionally,
as reviewed in
Section~\ref{sec:related},
prior work has considered formal
and informal learning resources independently, but has not examined
differences in what these resources are used to learn about. To fill this
gap, we make subject-by-subject comparisons between online courses and all other
types of online educational resources combined. We used $\chi^{2}$ proportions
tests to make inferences about whether online courses are used more, less, or
equally as often to learn about certain subjects compared to other types of
resources. To reduce the Type I error rate we applied the Bonferroni-Holm
correction to the resulting p-values to account for multiple hypothesis testing.
To answer RQ3 we examined commonalities in subjects and resources by using k-modes
clustering ~\cite{huang-1998} to organize respondents into clusters based on
the subjects they reported learning about online and, separately, the types
of resources they reported using. K-modes clustering is an extension of
k-means clustering for categorical data. While k-means minimizes the distance
between the center of each cluster and points belonging to that cluster in
euclidean space, k-modes aims to maximize the similarity in categories shared
between observations (respondents in our case) within a cluster. The number of
clusters we selected for clustering respondents according to shared types of
resources and subjects was chosen using the silhouette method
~\cite{Rousseeuw-1987}, where we calculated the average silhouette width
for $k$ equal to 3 through 10, and then selected the value of $k$ that
maximized the average silhouette width. This method for selecting $k$ ensures
that the observations being clustered are the most similar to the cluster
they are in and that they are the most dissimilar from clusters from which they
are excluded.
\subsection{Limitations}
Respondents were asked to self-report their online learning experiences,
therefore this study exhibits many of the same limitations as other
self-report studies. These limitations include over- and under-reporting,
which occur when survey results reflect an over- or under-estimate of the
rate or abundance of an experience. This discrepancy can be caused by
respondents misinterpreting survey questions, or it can be induced by one
of several biases, including desirability bias (when respondents give
socially desirable, instead of honest answers), and recall bias (when
respondents incorrectly remember an experience). To alleviate the potential
for these inconsistencies, we revised the survey questions iteratively
through a series of interviews, followed by pre-testing the survey to ensure
that respondents thoroughly answered all questions and interpreted them the
way we had intended.
Additionally, our survey was conducted online. As our focus is to study online learning, collecting the data online is appropriate. Cint, the firm through which we collected our data, uses a double opt-in procedure to recruit respondents to its panel. Research in the past decade has established that there are ``few or no significant differences between traditional modes [of survey administration] and opt-in online survey approaches'' for research such as that presented here, which is intended to improve our understanding of human behavior and experiences~\cite{ansolabehere2014does,prepared2010research}.
The results of this survey only reflect the usage of certain types of online
learning resources, and the subjects that respondents were interested in
learning about. Therefore we cannot make any claims about how often these
resources were consulted, or the duration of time that respondents spent
pursuing particular subjects. Our findings are accompanied by a number of
theoretical explanations that include suggestions for how online educational
experiences could be improved and how they could be studied in the future.
This is not an experimental study and we have data from one point in time
so it is inappropriate to interpret any of the relationships that we
present as causally linked.
\hl{Finally, it is important to clarify that the conception of ``learning''
presented in this paper is based in the computer and cognitive sciences,
and therefore represents only one perspective among many. For example, much
of the prior work that inspired the design of this study is related
or adjacent to studies of ``online learning.'' This includes both empirical
studies of what kinds of activities benefit specific learning
outcomes}~\cite{koed-15}, \hl{and theoretical work that centers
``instructional events'' and ``learning events'' within greater learning
frameworks}~\cite{koed-12}. \hl{This is in contrast to many other approaches in the
learning sciences that focus on the institutions that create learning
opportunities, instruction styles, the roles of facilitators in learning
environments, and taxonomies of learning outcomes}~\cite{means2014learning}.
\section{Related Work}
\label{sec:related}
\subsection{Frameworks for Quantitative Analyses of Online Education}
This work is modeled after surveys of participants in sociotechnical systems
that are typical throughout the Computer Supported Cooperative Work and Human
Computer Interaction communities. Our work is inspired specifically by three
analysis frames posited by prior work.
First, we draw from the work of Piety et al., who present a framework for the
educational data sciences, which delineates studies by (1) the educational
stage of the learners studied, and (2) the unit of analysis through which the
learning is analyzed~\cite{Piety-2014}. We use this framework to design our
study: we choose to study (1) adult learners and (2) our unit of analysis is the individual with a focus on the subjects and resources they consult online for learning.
Piety and colleagues additionally explain that a challenge of educational
data is that they are often limited to specific types of online learning
resources. In order to address this limitation and maximize the possible set
of resources available for us to analyze, we sample a large and
demographically diverse set of respondents in our study. By collecting
survey responses from people whose demographics are proportional to the
overall US population on multiple demographic axes, we aim to capture a
set of learners who use a broad set of resources.
In addition to the Piety et al. framework for educational data science
studies, we draw from the analysis frame of Ferguson and Shum. Ferguson
and Shum argue for always using networked analysis to understand online
learning~\cite{Ferguson-2012}. Their rationale for this methodological
approach is that online learners often consult multiple types of online
learning resources when trying to understand a subject, and therefore the
relationship between the different types of resources should be understood
better. They point to the proliferation of Open Educational Resources (OERs)
that learners are able to consult online as a basis for the need to
characterize how the uses of these resources intersect. Following their
suggestion, we use clustering analysis to understand our data and draw from
these analyses to identify opportunities for developing new online learning
experiences.
Finally, our analysis has been influenced by
the work of Kross and Guo, who provide a systematic
review of the literature on how different types of online learning resources
connect students to each other to create new interactions that shape
students' learning trajectories~\cite{kross-2018}. They highlight how
software systems can be purpose-built to enhance learning experiences for
certain subjects. Applying this purpose-driven lens to our results,
we will discuss the implications of our empirical analysis on the structure of
the online learning communities described in their study.
\subsection{Surveys of Online Educational Experiences}
Prior work has investigated various aspects of learners' online educational
experiences. Kizilcec and Schneider developed and deployed the Online Learning
Enrollment Intentions scale as an instrument for understanding the motivations
for learners to seek online education ~\cite{Kizilcec-2015}. They found that
online learners often have social goals when taking a course, which contrasted
to the typical course design built upon assignments meant to be completed
alone. They further found that the design of online educational resources
should incorporate an awareness that learners are using multiple resources,
and that online instructors should direct learners toward other resources.
In our study we extend these findings beyond the domain of online courses to
quantify the extent to which ten types of online learning resources are used
alongside other types of resources.
Additionally, Swanson and Walker surveyed young adults between the ages of
18 and 25 to gain insight into their usage of several different digital
technologies for academic and recreational purposes ~\cite{swanson-2015}.
They found that young adults spend a majority of out-of-classroom
academically-focused time using technology. Their results inform how
different types of online learning resources could be deployed depending on
how young adults are using different devices. Our study builds upon their work
by analyzing how different types of online learning resources are used, which
may be influenced by what subjects can be studied effectively given the device
that is being used. Both of these studies aim to understand how technology
can be harnessed to meet learners' expectations in terms of where they can
find online educational resources and how they can be supported to have
fulfilling learning experiences via those resources.
\subsection{Formal and Informal Online Learning Experiences}
The online education literature includes studies of both more formal online
educational experiences like online courses offered for college credit and
massively open online courses that award credentials
\cite{hew-2014,pastore-2009}. The literature also includes studies of more
casual online educational experiences that are related to the field of
``free-choice learning,'' which includes going to a museum or reading a non-fiction
book for pleasure~\cite{falk-2007}.
For our discussion about online learning resources we will differentiate between
\emph{formal} and \emph{informal} resources based on Rosenthal's work on
free-choice learning about YouTube, where both \emph{formal} and \emph{informal}
resources have learning outcomes but only \emph{formal} resources have
prescribed learning objectives and \emph{informal} learning experiences take
place ``outside a formal learning environment''~\cite{Rosenthal-2018}.
Additionally, Wenger characterizes informal learning as requiring community participation, versus formal learning that takes place in a classroom or a structured learning environment~\cite{wenger-2002}.
In line with prior work, we refer to enrollment in an online course as use of a
\emph{formal}
online learning resource, while we refer to
all other types of online learning resources as
\emph{informal}.
We believe that
this terminology is justified considering the time and financial commitment
that many online courses require, and the resemblance that online courses
have to traditional in-person learning experiences in the classroom.
Studies of formal online learning resources greatly outnumber studies of
informal learning resources, perhaps due to the high level of access that
university researchers have to such learners and data generated from online
university courses~\cite{kross-2018,kross2020democratization,Zheng-2015}. These studies
often focus on students' level of engagement with online course materials,
how many students complete the course compared to how many enroll, and how
these dynamics differ from in-person educational experiences.
Here we highlight a few examples of studies about informal online learning
that are especially salient to the goals of our study. Many studies of
informal online learning are centered around one specific aspect of the
educational experience, usually a specific type of online learning resource,
or a study about how a particular subject is pursued online. In an example
of the former, Narayan et al. developed and studied the success of an
interactive tutorial to onboard new members of the Wikipedia
community~\cite{Narayan-2017}. Although the results from their intervention
did not cause participants to be any more likely to contribute to Wikipedia,
it made participants feel more integrated into the Wikipedia community.
Torrey et al. explored teaching and learning crafts online via a series of
interviews focused on how members of craft communities communicate ideas
about space and aesthetics that are not easily communicated with typical
online resources ~\cite{Torrey-2009}. They found that curation of specific
online resources and persistence in communicating complex ideas about their
craft were key to their success in this informal learning community.
Yet other studies on informal online learning focused on how people learn a
specific subject. Exemplar studies include that by Shorey et al., who studied how
participation in social subgroups within an online community for the Scratch
educational programming language ~\cite{resnick-2009} leads to higher levels
of engagement and enriching interactions that may lead to better learning
outcomes ~\cite{shorey-2020}. These results are further supported by
Yang et al. and Gelman et al., who studied how communities of informal
learning can grow, and how learning trajectories can be mapped in informal
learning settings ~\cite{Yang-2015,Gelman-2016}.
Our study differs from past work on both formal and informal learning in that
we (a) study a far broader set of resources and subjects, aiming to
characterize and critically examine the full ecosystem of online learning,
and (b) while prior work has called for the comparison of formal and
informal resources ~\cite{schwier-2012}, our work is the first that we are
aware of to compare these mechanisms of online learning empirically
(specifically, in terms of the subjects that people learn using differing resource types).
\section{Results}
Overall, we find that 93\% of our 2260 survey respondents reported learning
something online. In this section, we detail the results of our three research
questions, examining: (RQ1) what subjects people learn about online; (RQ2) how
people learn online; and (RQ3) which online learning experiences are shared
among the majority of online learners.
\subsection{What do people learn about online?}
\begin{figure}[htb]
\includegraphics[width=1\textwidth]{fig1.png}
\vspace{-3ex}
\caption{(Left) The popularity of different subjects that respondents learned online.
(Right) The proportion of books in circulation according to a study by Littman and Connaway ~\cite{Littman-2004}. Both the internet and library collections are public repositories for learning resources on many different subjects. This comparison shows that the distribution of subject popularity in our survey mirrors the same power law distribution found in the circulation of different subjects of library books.}
\label{fig1}
\vspace{-1ex}
\end{figure}
We asked survey respondents to report what subjects they learned about online.
They reported learning about 19 subjects included in our question\footnote{As discussed in more detail
in the methodology section, we also offered a free-text ``Other'' option for
inputting additional subjects. Given that less than 1\% of respondents reported a subject
not included in the subject list, and there was very little overlap in the
subjects reported by these respondents.}. These subjects ranged from
non-academic topics such as travel and personal physical fitness to
traditional academic subjects such as law and mathematics
(see Figure ~\ref{fig1} [Left] for a summary).
We examine how many of our respondents chose to learn about these subjects. The only subject that more than half of our respondents reported learning about was DIY (62.4\%), which includes around-the-house activities such as cooking, baking, and arts-and-crafts. After DIY, the next most popular subjects, which were statistically less popular (p < 0.001 with Bonferroni-Holm correction for multiple comparisons)~\footnote{Statistical comparisons for each of the subject categories reported in this section can be found in the appendix.}, were those about general welfare and lifestyle -- physical health and healthcare (49\%), personal physical fitness (37\%), and travel and geography (32\%) -- and history (40\%).
Our results show that nearly all general-interest subjects were more popular than academic-focused subjects.
Academic subjects, aside from history, were significantly less popular (p < 0.005 Bonferroni-Holm correction for multiple comparisons) than general interest subjects. Less than 30\% of respondents reported learning about: fine arts (28\%), mathematics (26\%), the natural sciences (26\%), languages (25\%), medicine (23\%), politics (22\%), social sciences (22\%), religion and ethics (19\%), computer programming (15\%), law (15\%), graphic design (14\%), and literature (14\%).
To contextualize our findings, we compare the popularity of subjects learned online with the popularity of subjects learned offline. We find that while the subjects adults learn online vs. offline differ, the popularity of learning various subjects online and offline follows a similar trend: a few subjects are very popular while each of the remaining subjects are learned by a sizable minority. This is illustrated by the comparison in Figure ~\ref{fig1}: on the left, we show the popularity of subjects for online learning in our data set while on the right we show the distribution of offline learning interests as represented by the distribution of library books currently in circulation according to the Library of Congress Classification~\cite{Lancaster-1982}. We observe that this distribution follows a power law relationship like the Zipf distribution, where only a few subjects are very popular and most subjects are all approximately equal in their middling popularity.
\subsection{How people learn online}
\begin{figure}[htb]
\includegraphics[width=0.8\textwidth]{er.png}
\vspace{-3ex}
\caption{Proportion of respondents that reported using each type of educational resource.}
\label{fig:er}
\vspace{-3ex}
\end{figure}
Respondents used at least one and at most 10 (mean: 2.93, median: 2) resources to learn about a particular subject.
Of the 2094 respondents to our survey who indicated using at least one resource, the vast majority (88\%) relied on YouTube to learn online.\footnote{A respondent is considered to have learned online using a resource if they learned about at least one subject using this resource.}Additionally, 77\% used Wikipedia or informational articles, respectively, to learn online.
Over half of respondents reported learning online using how-to guides (70\%) or by using question-asking forums: either asking questions (60\%) or reading answers to others' questions (65\%).
The least used online learning resources, but still used by at least half of respondents, were academic-style resources: interactive tutorials (58\%), course materials (e.g., slides, course notes) (52\%), and practice exam websites (50\%).
Further, 55\% of respondents reported enrolling in formal online courses to learn something. Of these respondents, 40.5\% reported having taken an online course from a local university, while 22.7\% said they had taken an online course at a university that was not local to them. The difference of these two proportions suggests that respondents' awareness of online education is not as global as the potential online educational opportunities that may be open to them. However, it is also possible that the higher percentage of locally-based online course enrollment may be influenced by college degree programs that combine in-person and online coursework, rather than by a lack of awareness about non-local options.
Respondents also took online courses through professional services (20.5\%) such as Lynda, or through massively open online courses such as those offered on Coursera (16.5\%). A tenth (10.2\%) of respondents reported that they had used more than one type of online course.
Figure ~\ref{fig:er} provides a summary of respondents' online learning resources.
\subsubsection{Exploring the relationship between how people learn and what they learn online}
Having identified the popularity of subjects and of resources, the next step is to look at the relationship between the two. Are certain resources used for more subjects than others? Are certain resources more popular for certain subjects than others?
\begin{figure}[htb]
\includegraphics[width=0.8\textwidth]{avgsub.png}
\vspace{-3ex}
\caption{The average number of subjects that an individual respondent learned about for each resource.}
\label{fig:avgsub}
\vspace{0.1em}
\end{figure}
We find that YouTube is not only the resource used by the most respondents, it is also used by respondents to learn about the most subjects: a respondent who uses YouTube for learning uses the platform to learn about an average of 3.56 subjects. YouTube is used to learn about significantly more subjects (Mann-Whitney U Test, Bonferroni-Holm corrected p < 0.001) than Wikipedia (mean = 2.54 subjects) and other informational articles (mean = 2.25 subjects) or reading (mean = 2.54 subjects) and asking questions on Q\&A forums (mean = 2.25 subjects). All of these resources were used by respondents to learn significantly more subjects (Mann-Whitney U Test, Bonferroni-Holm corrected p < 0.002) than how-to guides, academic-style resources, and online courses, all of which were used to learn about less than two subjects, on average. Figure ~\ref{fig:avgsub} summarizes the average number of subjects that respondents learned about using each resource.
\begin{figure}[htb]
\includegraphics[width=1\textwidth]{karyo.png}
\vspace{-3ex}
\caption{Proportion of each type of resource that was used to learn about each subject, in order from most popular to least popular subject.}
\label{fig:karyo}
\vspace{-3ex}
\end{figure}
While YouTube is the most used resource overall (used by 88\% of respondents), it is used particularly heavily to learn about certain subjects. Nearly a quarter of respondents who used YouTube learned about these five subjects: makeup and fashion (used by 36.5\% of learners), physical fitness (30.6\%), DIY (28.4\%), travel/geography (24.0\%) and the arts (23.7\%). Across all subjects, YouTube was at least the third most used type of learning resource.
The next most popular resource types include informational articles and Wikipedia (used by 77\% of learners, respectively). Informational articles were particularly popular for learning about subjects related to health, including being the most popular resource for personal health care (27.2\%), medicine and nursing (24.6\%), and the second most popular resource for physical fitness (15.2\%). Wikipedia on the other hand was the most popular resource for learning social science and humanities subjects such as history (25.7\%), religion and ethics (22.7\%), literature and poetry (19.8\%), politics (19.7\%), and law (18.3\%).
While how-to guides are, unsurprisingly, used by 20.7\% of respondents to learn about DIY, they are not otherwise used heavily to learn about any other particular subject (an average of 8.2\% of participants used how-to guides to learn across all subjects, SD = 4.04\%). Similarly, reading and asking questions on Q\&A forums made up 11.0\% (SD = 2.04\%) and 8.9\% (SD = 1.64\%), respectively, of the resources used by respondents to learn about any given subject.
Together, interactive online tutorials, course materials, and practice exams made up 40.5\% of the resources that respondents used to learn about Math. Similarly, these academic-style resources made up 33.3\% of the resources used to learn languages, 29.0\% of those used to learn programming, and 26.7\% and 29.3\%, respectively, of the resources used to learn social and natural science. The specific type of academic-style resource used most heavily to learn these subjects differed, with practice exams (14.0\% of the resources used) and interactive tutorials (13.6\% of the resources used) being particularly popular for math, interactive tutorials being particularly popular for languages (17.8\% of the resources used), and course materials being particularly popular for the social and natural sciences (~11\% of the resources used, respectively) and programming (10.2\% of the resources used).
\begin{table}[ht]
\centering
\begin{tabular}{lrrll}
\hline
Subject & Proportion Formal & Proportion Informal & Significance & Greater Proportion \\
\hline
History & 0.080 & 0.078 & & \\
Arts & 0.056 & 0.055 & & \\
Medicine/Nursing & 0.049 & 0.045 & & \\
Online Safety & 0.039 & 0.032 & & \\
Religion/Ethics & 0.035 & 0.032 & & \\
Politics & 0.034 & 0.040 & & \\
Law & 0.026 & 0.025 & & \\
Math & 0.101 & 0.063 & *** & Formal \\
Languages & 0.086 & 0.045 & *** & Formal \\
Science & 0.076 & 0.061 & * & Formal \\
Social science & 0.075 & 0.048 & *** & Formal \\
Programming & 0.058 & 0.036 & *** & Formal \\
Graphic Design & 0.042 & 0.026 & *** & Formal \\
Literature/Poetry & 0.037 & 0.024 & ** & Formal \\
DIY & 0.063 & 0.132 & *** & Informal \\
Health Care & 0.047 & 0.091 & *** & Informal \\
Physical Fitness & 0.035 & 0.063 & *** & Informal \\
Travel/Geography & 0.024 & 0.052 & *** & Informal \\
Makeup/Fashion & 0.022 & 0.043 & *** & Informal \\
\hline
\end{tabular}
\caption{Pairwise statistical tests comparing the proportion of respondents who used formal versus informal online learning resources to pursue each subject. Zero * indicates no significant difference, one * indicates p < 0.05, two ** indicates p < 0.01, three *** indicates p < 0.001. `Greater Proportion' indicates whether formal or informal resources were used significantly more often.}
\label{tab:pair}
\end{table}
As mentioned in the introduction of this paper, the vast majority of prior work on online learning focuses on formal online courses rather than informal resources. This paper is the first, to our knowledge, to consider both types of resources. As such, we were especially interested in the differences between how people use online courses vs. informal online learning resources. These differences are summarized in Table ~\ref{tab:pair}.
We find that STEM-related academic subjects as well as social science, literature, languages, and graphic design are more often learned about using online courses than informal resources.
On the other hand, informal resources were significantly more likely to be used to learn about general interest subjects: DIY, health care, physical fitness, travel and geography, and makeup and fashion. The contrast between these two groups of subjects may be related to the need (or lack of need) for a formal environment to become proficient in a subject. Mastering a language requires an experience approximating immersion into a new social environment, and learning graphic design may require hours of orientation with complex user interfaces. In contrast, successfully planning a trip or learning how to repair something specific around the house does not require intense prolonged study. Another difference between these two groups is the extent to which these activities are learned through physical interactions. Mathematics and computer programming are subjects concerned with manipulating abstract symbols, while training one's physical fitness or learning a new sewing or makeup technique have to be physically practiced to be understood.
Finally, respondents in our survey were equally as likely to report having used an online course or an informal resource to learn about: history, arts, religion and ethics, politics, law, medicine and nursing, and online safety. We hypothesize that this is because these subjects straddle the line between academic and general interest: for example, an online learner could be studying art history as part of a degree program or job preparation or an online learner could be passively curious about the art of Michelangelo after a recent trip to Rome.
\subsection{Identifying core online learning experiences}
\begin{figure
\includegraphics[width=0.8\textwidth]{cr.png}
\vspace{-3ex}
\caption{Euler diagram representing k-modes clustering of types of online learning resources. Our clustering analysis revealed three overlapping core sets of online learning resources. Resources that are contained within a greater number of circles are more central overall to online learning experiences. For example, YouTube is contained within three circles, compared to Wikipedia which is only contained within two circles. Therefore YouTube is a more central resource compared to Wikipedia.}
\label{fig:cr}
\vspace{-3ex}
\end{figure}
Finally, going beyond the popularity of what and \emph{how} people learn online, we consider common patterns in people's practice of online learning. We seek to identify shared experiences in online learners' interests (what they learn), learning tools (how they learn), and overall online learning experiences (common pairs of subjects and resources).
First, to identify commonalities in online learners' interests, we used k-modes clustering to group respondents who reported learning more than one thing online based on the subjects they reported learning.
Of those who learned at least two things online, 76.1\% of respondents had in common only that they learned about DIY subjects online. These respondents did not necessarily indicate that they learned exclusively about DIY subjects, however their choice to learn about DIY subjects is what they had most in common.
An additional 18.8\% of respondents had in common that they learned both about DIY \emph{and} about at least one of: languages, health care, math, history, natural science, social science, and physical fitness. For the remaining 5.1\% of respondents, our clustering was unable to identify a common pattern among these respondents' online learning interests.
Second, we consider commonalities in how people learn online. We again used k-modes clustering to group those respondents who used more than one online learning resource, this time clustering the respondents based on the resources they used rather than the subjects they learned.
Of respondents who used at least two online learning resources, 17.1\% had in common only that they used YouTube (the blue cluster in Figure ~\ref{fig:cr}), while 12.9\% of respondents had in common that they used both YouTube and at least one of the following: how-to guides, Wikipedia and informational articles. An additional 18.8\% of respondents had in common that they (1) used YouTube, (2) used at least one of the aforementioned three resources, and (3) asked questions in Q\&A forums. We could not find any pattern in resource use for the remaining 51.2\% of respondents.
Of the 190 online learning experiences respondents reported, these twelve were experienced most often by our respondents: using YouTube to learn about DIY subjects (54.9\% of respondents), health care (24.0\%), makeup/fashion (23.2\%), history (21.2\%), arts (19.9\%), or travel/geography (18.3\%); using how-to guides to learn about DIY subjects (40.0\%), using informational articles to learn about health care (36.5\%), using Wikipedia to learn about history (30.6\%), or reading Q\&A forums or asking questions on those forums to learn about DIY subjects (25.4\% and 20.3\%, respectively).
|
\section{Introduction}
Let $(\Omega,{\cal F},P)$ be a probability space equipped with
a filtration ${\mathbb F}=({\cal F}_t)_{t\in[0,T]}$.
We consider a one-dimensional semimartingle $X=(X_t)_{t\in[0,T]}$
having a decomposition
\begin{eqnarray}\label{0301191155}
X_t &=& X_0+\int_0^t b_sds+\int_0^t{\color{black}}% {\color[rgb]{0,0,0.8} \sigma_s}dw_s+J_t\quad(t\in[0,T])
\end{eqnarray}
where
$X_0$ is an ${\cal F}_0$-measurable random variable,
$b=(b_t)_{t\in[0,T]}$ and $\sigma=(\sigma_t)_{t\in[0,T]}$ are
c\`adl\`ag\ ${\mathbb F}$-adapted processes, and
$w=(w_t)_{t\in[0,T]}$ is an ${\mathbb F}$-standard Wiener process.
$J=(J_t)_{t\in[0,T]}$ is the jump part of $X$.
We will assumed that $J$ is finitely active, that is,
$J_t=\sum_{s\in(0,t]}{\color{black}}% {\color[rgb]{0.8,0,0}\Delta J_s}$ for ${\color{black}}% {\color[rgb]{0.8,0,0}\Delta J_s=J_s-J_{s-}}$
and $\sum_{t\in[0,T]}1_{\{\Delta J_t\not=0\}}<\infty$ a.s.
In this paper, we are interested in the estimation of the integrated volatility
\begin{eqnarray}\label{0211281514}
\Theta
&=&
\int_0^T\sigma_t^2d
\end{eqnarray}
based on the data $(X_{t_j})_{j=0,1,...n}$, where
${t_j}={t_j}^n=jT/n$.
The jump part $J$ can be endogenous or exogenous, as well as $b$ and $\sigma$,
however, $J$ is a nuisance in any case.
The simple realized volatility is heavily damaged when jumps exist.
To avoid the effects of the jumps, various methods have been proposed so far.
For example,
the bipower variation (Barndorff-Nielsen and Shephard. \cite{Barndorff-NielsenShephard2004b}, Barndorff-Nielsen et al. \cite{Barndorff-NielsenShephardWinkel2006}) and the minimum
realized volatility (Andersen et al. \cite{Andersen2012}) are shown to
be consistent estimators of the integrated volatility even in the presence of jumps.
The idea of these methods is that, to mitigate the effect of jumps,
they employ adjacent increments in constructing the estimator.
Another direction to handle jumps is to introduce a threshold to detect jumps.
{\color{black}}% {\color[rgb]{0.8,0,0}
Parametric inference for sampled
{\color{black}}% {\color[rgb]{1,0.851,0} diffusion type processes} was studied by
Dohnal \cite{Dohnal1987},
Prakasa Rao \cite{PrakasaRao1983,prakasa1988statistical},
Yoshida \cite{Yoshida1992b, yoshida2011},
Kessler \cite{Kessler1997},
Genon-Catalot and Jacod \cite{Genon-CatalotJacod1993},
Uchida and Yoshida \cite{UchidaYoshida2013, UchidaYoshida2012Adaptive,uchida2014adaptive},
Ogihara and Yoshida \cite{ogihara2014quasi},
Kamatani and Uchida \cite{KamataniUchida2014}
and others.
{\color{black}}% {\color[rgb]{0,0.5,0} Limit theorems used to analyse the realized volatility appeared
in the studies of parametric inference.}
If no jump part exists, then the distribution of the increment $\Delta_jX=X_{t_j}-X_{t_{j-1}}$
admits Gaussian approximation in a short time interval, and
a quasi-likelihood function can be constructed with the conditional Gaussian density.
When the jump part exists, the local Gaussian approximation is no longer valid.
Then it is necessary to detect jumps and classify the increments to apply the local Gaussian quasi-likelihood function
for estimation of the parameters in the continuous part.}
Threshold method was investigated {\color{black}}% {\color[rgb]{0.8,0,0} by} Shimizu and Yoshida \cite{ShimizuYoshida2006}
{\color{black}}% {\color[rgb]{0.8,0,0} and Ogihara and Yoshida \cite{OgiharaYoshida2011}}
in the context of the parametric inference for a stochastic differential equation with jumps.
The idea of thresholding {\color{black}}% {\color[rgb]{0,0.5,0} is rather old, going}
back to the studies of limit theorems for L\'evy processes
as latest. Mancini {\color{black}}% {\color[rgb]{0,0,0.8} \cite{Mancini2001}} {\color{black}}% {\color[rgb]{0.8,0,0} used this idea in a nonparametric situation.
{\color{black}}% {\color[rgb]{0,0.5,0} Koike \cite{koike2014estimator} applied the threshold method to covariance estimation for asynchronously observed semimartingales with jumps.}
The classical jump filters compare
the size of increment with a threshold determined
by a {\color{black}}% {\color[rgb]{0,0.5,0} (conditionally/unconditionally deterministic)} function
of the length of {\color{black}}% {\color[rgb]{0,0.5,0} the} time interval.
If an increment
is so large that exceeds the threshold, it is regarded as having jumps.
Otherwise, the increment is regarded as having no jump.
Once classified, the increments are used to estimate the parameters
in continuous and jump parts, respectively.}
{\color{black}}% {\color[rgb]{0.8,0,0}
Though the efficiency of the traditional thresholding parametric estimators
has been established theoretically,
it is known that their real performance strongly depends on a choice of
tuning parameters; see, e.g.,
Iacus and Yoshida \cite{iacus2017simulation}.}
{\color{black}}% {\color[rgb]{1,0.851,0} Examining each individual increment without other data
is not always effective in
finding jumps.} It sometimes overlooks relatively small jumps
{\color{black}}% {\color[rgb]{0.8,0,0}
due to a conservative level of threshold to try to incorporate all Brownian increments.
To resolve this problem,}
Inatsugu and Yoshida {\color{black}}% {\color[rgb]{0,0,0.8}\cite{InatsuguYoshida2020Accepted}} introduced
{\color{black}}% {\color[rgb]{0.8,0,0} the so-called} global filters that
examine all increments simultaneously and regard an increment of high rank in order of absolute size as a jump. Using the information about the size of other increments helps us detect jumps more accurately than {\color{black}}% {\color[rgb]{0.8,0,0} the previous methods that ignore
such information.
Moreover, Inatsugu and Yoshida {\color{black}}% {\color[rgb]{0,0,0.8}\cite{InatsuguYoshida2020Accepted}} also removed
the assumption of low intensity of small jumps, {\color{black}}% {\color[rgb]{0,0.5,0} that} was used in
Shimizu and Yoshida \cite{ShimizuYoshida2006}
and Ogihara and Yoshida \cite{OgiharaYoshida2011}.
This is a theoretical advantage of the global jump filters,
in addition to their outperformance
in practice.
}
In this paper, we {\color{black}}% {\color[rgb]{0.8,0,0} will} apply the global filtering method to nonparametric volatility estimation.
Specifically, we {\color{black}}% {\color[rgb]{0.8,0,0} will} construct the ``global realized volatility (GRV) estimator" of
the integrated volatility
for the semimartingale $X$ having the decomposition (\ref{0301191155}).
{\color{black}}% {\color[rgb]{0.8,0,0}
Though $J$ and the jump part of $\sigma$ are assumed to be finitely active for each $n$,
we permit the number of jumps to diverge as $n$ tends to infinity.}
We will investigate the theoretical properties of GRV and then
conduct numerical simulations to study their performance compared with
{\color{black}}% {\color[rgb]{0.8,0,0} traditional} methods, that is,
{\color{black}}% {\color[rgb]{0.8,0,0} the deterministic threshold estimator,}
the bipower variation and the minimum
realized volatility.
The organization of this paper is as follows.
{\color{black}}% {\color[rgb]{0.8,0,0} Section \ref{Sec2} introduces} the GRV and its variant, the winsorized GRV (WGRV).
In Section \ref{Sec3},
we introduce the local-global realized volatility (LGRV) and
prove its convergence to {\color{black}}% {\color[rgb]{0.8,0,0} the spot volatility}.
{\color{black}}% {\color[rgb]{0.8,0,0} The LGRV will be used
for normalizing the increments to compute the global filter.
Section \ref{0211181814} gives}
the rate of convergence of the GRV and WGRV in the
situation where the intensity of jumps is high. In this case,
we need a high and
fixed cut-off rate $\alpha$ {\color{black}}% {\color[rgb]{0.8,0,0} to eliminate harmful jumps}.
In Section \ref{0211181815},
we allow the cut-off rate to vary according to the sample size.
This ``moving threshold" method is for the situation where the intensity of jumps is
moderate and small cut-off rate is applicable.
{\color{black}}% {\color[rgb]{0.8,0,0} {\color{black}}% {\color[rgb]{0,0.5,0} Section \ref{0211281401}}
briefly discusses the situation where true volatility is constant.
In this case, normalizing increments is not necessry,} so the estimator gets
a little simpler.
{\color{black}}% {\color[rgb]{0.8,0,0} Section \ref{Sec7} presents some}
simulation results to compare the {\color{black}}% {\color[rgb]{0.8,0,0} real} performance
of the GRV, WGRV, bipower varition, and the mininum realized volatility.
{\color{black}}% {\color[rgb]{0.8,0,0}
Concluding, let us mention some technical aspects.
The global jump filter causes theoretical difficulty.
By nature, it uses all the data to classify {\color{black}}% {\color[rgb]{1,0.851,0} each increment} $\Delta_jX$.
This completely destroys the martingale structure in the model, which makes it difficult
to use orthogonality between the selected increments to validate
the law of large numbers and the central limit theorem.
However, it is possible to asymptotically recover the orthogonality
by {\color{black}}% {\color[rgb]{0,0.5,0} the} glocal and global filtering lemmas presented in
Sections \ref{Sec3} and \ref{0211181814}.
Technically, the argument here is closed within the semimartingale theory,
although the global filter breaks adaptivity of {\color{black}}% {\color[rgb]{0,0.5,0} the functionals}, in other words,
a quadratic variation with anticipative weights is treated.
On the other hand, Yoshida \cite{yoshida2020asymptotic}
suggests a use of the Malliavin calculus to {\color{black}}% {\color[rgb]{0,0.5,0} analyse}
robustified volatility estimators with anticipative weights.
}
\section{Realized volatilities with a global jump filter}\label{Sec2}
The global jump filter introduced by Inatsugu and Yoshida {\color{black}}% {\color[rgb]{0,0,0.8} \cite{InatsuguYoshida2020Accepted}} uses
the order statistics of the transformed increments of the observations.
Suppose that an estimator $S_{n,j-1}$ of the spot volatility $\sigma(X_{t_{j-1}})^2$
(up to a common scaling factor)
is given for each $j\in I_n=\{1,...,n\}$.
Denote $\Delta_jU=U_{t_j}-U_{t_{j-1}}$ for a process $U=(U_t)_{t\in[0,T]}$.
Then the distribution of
the scaled increment $S_{n,j-1}^{-1/2}\Delta_jX$
is expected to be well approximated by the standard normal distribution $N(0,1)$.
Therefore, if the value
\begin{eqnarray}\label{0211281531}
V_j &=& \big|(S_{n,j-1})^{-1/2}\Delta_jX\big|
\end{eqnarray}
is relatively very large among ${\cal V}_n=\{V_k\}_{k\in I_n}$, then
plausibly we can infer that the $V_j$ involves jumps with high probability.
The idea of the global jump filter is to eliminate the increment $\Delta_jX$ from the data
if the corresponding $V_j$ is ranked {\color{black}}% {\color[rgb]{0,0,0.8} within} the top $100\alpha$\% in ${\cal V}_n$.
More precisely, let
\begin{eqnarray*}
{\cal J}_n(\alpha)
&=& \big\{j\in I_n;\>V_j<V_{(s_n(\alpha))}\big\}
\end{eqnarray*}
where
\begin{eqnarray*}
s_n(\alpha) &=& \lfloor n(1-\alpha) \rfloor
\end{eqnarray*}
for $\alpha\in[0,1)$,
and we denote by $r_n(U_j)$ the rank of $U_j$ among the variables $\{U_i\}_{i\in I_n}$.
Let
\begin{eqnarray}\label{0302110914}
q(\alpha)
&=&
\int_{\{|z|\leq c(\alpha)^{1/2}\}}z^2\phi(z;0,1)dz
\end{eqnarray}
where $\phi(z;0,1)$ is the density function of $N(0,1)$ and
$c(\alpha)$ defined by
\begin{eqnarray*}
P\big[\zeta^2\leq c(\alpha)\big]&=&1-\alpha
\end{eqnarray*}
for $\zeta\sim N(0,1)$ and $\alpha\in[0,1)$.
Then the {\bf global realized volatility} (globally truncated realized volatility, GRV)
with cut-off ratio $\alpha$ is defined by
\begin{eqnarray}\label{0211260834}
{\mathbb V}_n(\alpha)
&=&
\sum_{j\in{\cal J}_n(\alpha)}q(\alpha)^{-1}|\Delta_jX|^2K_{n,j}
\end{eqnarray}
where
$K_{n,j}=1_{\{|\Delta_jX|\leq n^{-1/4}\}}$.
As remarked in Inatsugu and Yoshida {\color{black}}% {\color[rgb]{0,0,0.8} \cite{InatsuguYoshida2020Accepted}}, the indicator function $K_{n,j}$ is set
just for relaxing the conditions for validation.
Generalization by using like $1_{\{|\Delta_jX|\leq B_1n^{-\delta_1}\}}$
with constants $B_1>0$ and $\delta_1\in(0,1/4]$ is straightforward, but
we prefer simplicity in presentation of this article.
In practice, the probability {\color{black}}% {\color[rgb]{0,0,0.8} that} $K_{n,j}$ executes the task is exponentially small
by the large deviation principle.
However, the moments of $\Delta J_t$ are not controllable without assumption,
and we can simply avoid it by the cut-off function $K_{n,j}$.
{\color{black}}% {\color[rgb]{0,0,0.8} Winsorization} is a popular technique in robust statistics.
In the present context,
the {\bf Winsorized global realized volatility} (WGRV) is given by
\begin{eqnarray*}
{\mathbb W}_n(\alpha)
&=&
\sum_{j=1}^n{\sf w}(\alpha)^{-1}
\big\{|\Delta_jX|\wedge\big(S_{n,j-1}^{1/2}V_{(s_n(\alpha))}\big)\big\}^2
K_{n,j}
\end{eqnarray*}
where
\begin{eqnarray*}
{\sf w}(\alpha)
&=&
\int_{\mathbb R}\big(z^2\wedge c(\alpha)\big)\phi(z;0,1)dz.
\end{eqnarray*}
The cut-off ratio $\alpha\in[0,1)$ is a tuning parameter in estimation procedures.
The bigger $\alpha$ provides the more stable estimates even in high intensity of jumps.
On the other hand, the smaller $\alpha$ gives the more precise estimates
if the intensity of jumps is low.
Making trade-off between stability and precision is necessary in practice.
As a matter of fact, these cases require different theoretical treatments.
We will consider fixed $\alpha$ in Section \ref{0211181814},
and shrinking $\alpha$ in Section \ref{0211181815}.
\section{Local-global filter}\label{Sec3}
{\color{black}}% {\color[rgb]{0.8,0,0}
An estimator $S_{n,j-1}$ for the spot volatility (up to a constant scaling) is necessary to construct a global realized volatility.
Naturally, we use the data around time $t$ to estimate $\sigma_t$.
Since these data are also contaminated with jumps, we need a jump filter to construct
a temporally-local estimator $S_{n,j-1}$. The idea of the global jump filter with the order statistics
of the data around $t$ serves to eliminate the effects of jumps,
not only theoretically but also practically as demonstrated by the simulation studies
of Section \ref{Sec7}.
In this section, we propose a local-global realized volatility and validate it
by establishing in Section \ref{0301241722} the rate of convergence of the estimator.
Since the local-global filter involves the order statistics, that destroy the martingale structure,
we try to recover it by somewhat sophisticated lemmas given in Section \ref{0301241725}.
The minimum realized volatility (minRV) made of the temporally-local data is also a candidate of
an estimator for the spot volatility.
A rate of convergence of the local minRV is mentioned in Section \ref{0301241726}.
}
\subsection{Glocal filtering lemmas}\label{0301241725}
For each $j\in I_n$, let
\begin{eqnarray*}
\underline{j}_n
&=&
\left\{\begin{array}{cl}
1&(j\leq \kappa_n)\vspace*{3mm}\\
j-\kappa_n&(\kappa_n+1\leq j\leq n-\kappa_n)\vspace*{3mm}\\
n-2\kappa_n&(j\geq n-\kappa_n+1)
\end{array}\right.
\end{eqnarray*}
for $\kappa_n\in{\mathbb Z}_+$ satisfying $2\kappa_n+1\leq n$.
Let
$I_{n,j}=\{\underline{j}_n,\underline{j}_n+1,...,\underline{j}_n+2\kappa_n\}$.
Let
\begin{eqnarray*}
\widehat{U}_{j,k}\>=\> h^{-1/2}
\sigma_{t_{\underline{j}_n-1}}^{-1}
\Delta_kX\quad \text{and}\quad
W_j\>=\> h^{-1/2}\Delta_jw
\end{eqnarray*}
for $j,k\in I_n$.
Both variables $\widehat{U}_{j,k}$ and $W_j$ depend on $n$.
Let
\begin{eqnarray*}
\widehat{R}_{j,k}
&=&
\widehat{U}_{j,k}-W_k-h^{-1/2}\sigma_{t_{\underline{j}_n-1}}^{-1}\Delta_kJ
\end{eqnarray*}
for $j,k\in I_n$.
Denote $L^\infm=\cap_{p>1}L^p$.
{\color{black}}% {\color[rgb]{0.8,0,0}
Let $N=\sum_{s\in(0,\cdot]}1_{\{\Delta J_s\not=0\}}$.
Let $\widetilde{\sigma}=\sigma-J^\sigma$ for $J^\sigma=\sum_{s\in(0,\cdot]}\Delta\sigma_s$,
and let $N^\sigma=\sum_{s\in(0,\cdot]}1_{\{\Delta J^\sigma_s\not=0\}}$.
{\color{black}}% {\color[rgb]{0,0,0.8} We assume that
$N^\sigma_T<\infty$ a.s.
Moreover, let $\overline{N}=N+N^\sigma$.
Let $\widetilde{X}=X-J$.
A counting process will be identified with a random measure.
Let ${\tt I}_{n,j}=\big(t_{\underline{j}_n-1},t_{\underline{j}_n+2\kappa_n}\big]$.
{\color{black}}% {\color[rgb]{0.8,0,0}
\begin{description}
\item[{\bf [G1]}]
{\bf (i)} For every $p>1$, $\sup_{t\in[0,T]}\|\sigma_t\|_p<\infty$ and
\begin{eqnarray*}
\big\|\widetilde{\sigma}_t-\widetilde{\sigma}_s\big\|_p &\leq& C(p)|t-s|^{1/2}
\quad(t,s\in[0,T])
\end{eqnarray*}
for some constant $C(p)$ for every $p>1$.
\item[{\bf (ii)}] $\sup_{t\in[0,T]}\|b_t\|_p<\infty$ for every $p>1$.
\item[{\bf (iii)}] $\sigma_t\not=0$ a.s. for every $t\in[0,T]$, an
$\sup_{t\in[0,T]}\big\|\sigma_t^{-1}\big\|_p<\infty$ for every $p>1$.
\end{description}
\begin{en-text}
\begin{lemma}\label{0211190303}
Under $[G1]$,
\begin{eqnarray*}
\sup_{j\in I_n}\sup_{k\in I_{n,j}}
\bigg\|\widehat{R}_{j,k}1_{\big\{\Delta_kN=0,\>\sum_{k'\in I_{n,j}}\Delta_{k'}N^\sigma=0\big\}}
\bigg\|_p
&=&
O\left(\left(\frac{\kappa_n}{n}\right)^{1/2}\right)
\end{eqnarray*}
as $n\to\infty$ for every $p>1$.
\end{lemma}
\proof
For $j\in I_n$,
on the event $\big\{\>\sum_{k'\in I_{n,j}}\Delta_{k'}N^\sigma=0\big\}$,
\begin{eqnarray*}
\widehat{R}_{j,k}1_{\{\Delta_kN=0\}}
&=&
\big(h^{-1/2}\sigma_{t_{j-1}}^{-1}\Delta_k\widetilde{X}-h^{-1/2}\Delta_jw \big)1_{\{\Delta_kN=0\}}
\nonumber\\&=&
\big(h^{-1/2}\sigma_{t_{j-1}}^{-1}\Delta_k\widetilde{X}-h^{-1/2}\Delta_jw \big)
-\big(h^{-1/2}\sigma_{t_{j-1}}^{-1}\Delta_k\widetilde{X}-h^{-1/2}\Delta_jw \big)1_{\{\Delta_kN>0\}}
\end{eqnarray*}
Use orthogonality in
$\sum_{k\in I_{n,j}}\int_{t_{k-1}}^{t_k}(\widetilde{\sigma}_s-\widetilde{\sigma}_{t_{k-1}})dw_s$.
\qed
\begin{lemma}\label{0211190303}
Under $[G1]$,
\begin{eqnarray}\label{0211191046}
\sup_{j\in I_n}\sup_{k\in I_{n,j}}
\bigg\|\widehat{R}_{j,k}1_{\big\{\sum_{k'\in I_{n,j}}\Delta_{k'}N^\sigma=0\big\}}
\bigg\|_p
&=&
O\left(\left(\frac{\kappa_n}{n}\right)^{1/2}\right)
\end{eqnarray}
as $n\to\infty$ for every $p>1$.
\end{lemma}
\proof
For $j\in I_n$, let
$E(j)=\big\{\>\sum_{k'\in I_{n,j}}\Delta_{k'}N^\sigma=0\big\}$.
Then, for $k\in I_{n,j}$,
\begin{eqnarray*}
\widehat{R}_{j,k}1_{E(j)}
&=&
\big(h^{-1/2}\sigma_{t_{\underline{j}_n-1}}^{-1}\Delta_k\widetilde{X}-h^{-1/2}\Delta_kw \big)1_{E(j)}
\nonumber\\&=&
R^{(\ref{0211191011})}+R^{(\ref{0211191012})}
\end{eqnarray*}
where
\begin{eqnarray}\label{0211191011}
R^{(\ref{0211191011})}
&=&
h^{-1/2}\int_{t_{k-1}}^{t_k}
\sigma_{t_{\underline{j}_n-1}}^{-1}(\widetilde{\sigma}_t-\widetilde{\sigma}_{t_{\underline{j}_n-1}})dw_t
+
h^{-1/2}\sigma_{t_{\underline{j}_n-1}}^{-1}\int_{t_{k-1}}^{t_k} b_tdt
\end{eqnarray}
and
\begin{eqnarray}\label{0211191012}
R^{(\ref{0211191012})}
&=&
-\big(h^{-1/2}\sigma_{t_{\underline{j}_n-1}}^{-1}\Delta_k\widetilde{X}-h^{-1/2}\Delta_jw \big)1_{E(j)^c}.
\end{eqnarray}
Apply the Burkholder-Davis-Gundy inequality to the martingale part of (\ref{0211191011}),
and the H\"older inequality gives
\begin{eqnarray*}
\big\|R^{(\ref{0211191012})}\big\|_p
&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
\end{eqnarray*}
Then we obtain the estimate (\ref{0211191046}).
\qed
\end{en-text}
\begin{lemma}\label{0211190303}
Under $[G1]$,
\begin{eqnarray}\label{0211191046}
\sup_{j\in I_n}\sup_{k\in I_{n,j}}
\big\|\widehat{R}_{j,k}1_{\{N^\sigma({\tt I}_{n,j})=0\}}\big\|_p
&=&
O\left(\left(\frac{\kappa_n}{n}\right)^{1/2}\right)
\end{eqnarray}
as $n\to\infty$ for every $p>1$.
\end{lemma}
\proof
For $j\in I_n$, let
$E(j)=\{N^\sigma({\tt I}_{n,j})=0\}$.
Then, for $k\in I_{n,j}$,
\begin{eqnarray}\label{0211191011}
\widehat{R}_{j,k}1_{E(j)}
&=&
\big(h^{-1/2}\sigma_{t_{\underline{j}_n-1}}^{-1}\Delta_k\widetilde{X}-h^{-1/2}\Delta_kw \big)1_{E(j)}
\nonumber\\&=&
h^{-1/2}\int_{t_{k-1}}^{t_k}
\sigma_{t_{\underline{j}_n-1}}^{-1}(\widetilde{\sigma}_t-\widetilde{\sigma}_{t_{\underline{j}_n-1}})dw_t1_{E(j)}
\nonumber\\&&
+
h^{-1/2}\sigma_{t_{\underline{j}_n-1}}^{-1}\int_{t_{k-1}}^{t_k} b_tdt1_{E(j)}
\end{eqnarray}
We obtain (\ref{0211191046})
by applying the Burkholder-Davis-Gundy inequality to the martingale part of (\ref{0211191011})
after the trivial estimate $1_{E(j)}\leq1$.
\qed\vspace*{3mm}
For $j\in I_n$, denote by $r_{n,j}(U_k)$ the rank of the element $U_k$
among a collection of random variables $\{U_\ell\}_{k\in I_{n,{\color{black}}% {\color[rgb]{0.8,0,0} j}}}$.
Let
\begin{eqnarray*} &&
0<\eta_2<\eta_1,\qquad
\overline{\kappa}_n \>=\> 2\kappa_n+1,\quad
\nonumber\\&&
{\sf a}_n\>=\>\lfloor (1-\alpha_0)\overline{\kappa}_n-\overline{\kappa}_n^{\>1-\eta_2}\rfloor,\qquad
\widehat{\sf a}_n\>=\>\lfloor{\sf a}_n-\overline{\kappa}_n^{\>1-\eta_2}\rfloor
\end{eqnarray*}
for $\alpha_0\in[0,1)$.
Let
\begin{eqnarray}\label{0211201930}
L_{n,j,k}
&=&
\big\{r_{n,j}(|W_k|)\leq {\sf a}_n-\overline{\kappa}_n^{\>1-\eta_2}\big\}
\cap
\big\{|W|_{(j,{\sf a}_n)}-|W_k|<\overline{\kappa}_n^{\>-\eta_1}\big\}
\end{eqnarray}
where $\big(|W|_{(j,k)}\big)_{k\in I_{n,j}}$ are the ordered statistics made from $\{|W_k|\}_{k\in I_{n,j}}$.
In the same way as Lemma 1 of Inatsugu and Yoshida {\color{black}}% {\color[rgb]{0,0,0.8}\cite{InatsuguYoshida2020Accepted}},
we obtain the following result.
\begin{lemma}\label{0211191227}
Let $\alpha_0\in(0,1)$.
Suppose that ${\color{black}}% {\color[rgb]{0.8,0,0} \eta_1}<1/2$
and that
$n^{-\epsilon}\kappa_n\to\infty$ as $n\to\infty$
for some $\epsilon\in(0,1)$.
Then
\begin{eqnarray*}
\sup_{j\in I_n}P\bigg[\bigcup_{k\in I_{n,j}}L_{n,j,k}\bigg]\>=\> O(n^{-L})
\end{eqnarray*}
as $n\to\infty$ for every $L>0$.
\end{lemma}
\vspace*{3mm}
Define ${\cal K}_{n,j}(\alpha_0)$ by
\begin{eqnarray*}
{\cal K}_{n,j}(\alpha_0)
&=&
\big\{k\in I_{n,j};\>r_{n,j}
({\color{black}}% {\color[rgb]{0,0,0.8} |\Delta_kX|})
\leq (1-\alpha_0)\overline{\kappa}_n\big\},
\end{eqnarray*}
where
{\color{black}}% {\color[rgb]{0,0,0.8}
$r_{n,j}(|\Delta_kX|)$
}
is the rank of ${\color{black}}% {\color[rgb]{0,0,0.8} |\Delta_kX|}$
among
$\{ {\color{black}}% {\color[rgb]{0,0,0.8} |\Delta_{k'}X| }\}_{k'\in I_{n,j}}$.
Let
\begin{eqnarray*}
\widehat{{\cal K}}_{n,j}(\alpha_0)
&=&
\big\{k\in I_{n,j};\>r_{n,j}(|W_k|)\>\leq\>\hat{\sf a}_n\big\}{\color{black}}% {\color[rgb]{0,0,0.8} .}
\end{eqnarray*}
Let
\begin{eqnarray*}
\Omega_{n,j}
&=&
\bigcap_{k\in I_{n,j}}\bigg[
\bigg\{|\widehat{R}_{j,k}|1_{\{N^\sigma({\tt I}_{n,j})=0\}}<2^{-1}\overline{\kappa}_n^{\>-\eta_1}\bigg\}
\cap L_{n,j,k}^c\bigg]{\color{black}}% {\color[rgb]{0,0,0.8} .}
\end{eqnarray*}
Let {\color{black}}% {\color[rgb]{0.8,0,0}
\begin{eqnarray}\label{0301021550}
{\cal L}_n=\big\{j\in I_n;\>\overline{N}({\tt I}_{n,j})\not=0\big\}.
\end{eqnarray}
}
\begin{lemma}\label{0211200634}
\begin{description}
\item[{\bf (a)}]
$\widehat{{\cal K}}_{n,j}(\alpha_0) \subset {\cal K}_{n,j}(\alpha_0)$
on $\Omega_{n,j}$ if $j\in{\cal L}_n^c$.
\item[{\bf (b)}]
$\d
1_{\Omega_{n,j}}1_{\{j\in{\cal L}_n^c\}}\>
\#\big({\cal K}_{n,j}(\alpha_0)\setminus\widehat{{\cal K}}_{n,j}(\alpha_0)\big)
\leq
{\sred4
\>\overline{\kappa}_n^{\>1-\eta_2}
\quad(j\in I_{n.j},\>n\in{\mathbb N}).
\end{description}
\end{lemma}
\proof
Let $n\in{\mathbb N}$ and suppose that ${\color{black}}% {{\color[rgb]{0.8,0,0} j\in{\cal L}_n^c}$.
We will work on $\Omega_{n,j}$.
For a pair
{\color{black}}% {{\color[rgb]{0.8,0,0}$(k_1,k_2)\in I_{n,j}^2$},
suppose that
\begin{eqnarray}\label{0211201926}
r_{n,j}(|W_{k_1}|)\leq\widehat{\sf a}_n\quad\text{and}\quad
r_{n,j}(|W_{k_2}|)\geq{\sf a}_n.
\end{eqnarray}
Then
$
|\widehat{U}_{j,k_1}|
<
|W_{k_1}|+2^{-1}\overline{\kappa}_n^{\>-\eta_1}
$, since
$\Delta_{k_1}N=0$ and
{\color{black}}% {{\color[rgb]{0.8,0,0}$N^\sigma({\tt I}_{n,j})=0$} when
{\color{black}}% {{\color[rgb]{0.8,0,0} $j\in{\cal L}_n^c$},
and then
$|\widehat{R}_{j,{k_1}}|<2^{-1}\overline{\kappa}_n^{\>-\eta_1}$
on $\Omega_{n,j}$.
By the first inequality of (\ref{0211201926}),
$r_{n,j}(|W_{k_1}|)\leq{\sf a}_n-\overline{\kappa}_n^{\>1-\eta_2}$,
and hence on $\Omega_{n,j}\subset L_{n,j,{\color{black}}% {\color[rgb]{0.8,0,0} k_1}}^c$, we have
$|W|_{(j,{\sf a}_n)}-|W_{k_1}|\geq\overline{\kappa}_n^{\>-\eta_1}$
by the definition (\ref{0211201930}) of $L_{n,j,k}$.
Therefore
\begin{eqnarray}\label{0211201012}
|\widehat{U}_{j,k_1}| &<& |W|_{(j,{\sf a}_n)}-2^{-1}\overline{\kappa}_n^{\>{\color{black}}% {\color[rgb]{0.8,0,0} -\eta_1}}.
\end{eqnarray}
The assumption
{\color{black}}% {{\color[rgb]{0.8,0,0}$j\in{\cal L}_n^c$}
entails
$|\widehat{R}_{j,{k_2}}|<2^{-1}\overline{\kappa}_n^{\>-\eta_1}$
on $\Omega_{n,j}$, and hence
$|W_{k_2}|-2^{-1}\overline{\kappa}_n^{\>-\eta_1}<|\widehat{U}_{j,k_2}|$
due to $\Delta_{k_2}J=0$.
From (\ref{0211201012}), we have got
\begin{eqnarray}\label{0211202027}
|\widehat{U}_{j,k_1}|<|\widehat{U}_{j,k_2}|
\end{eqnarray}
on $\Omega_{n,j}$
if
$j\in{\cal L}_n^c$ and if a pair $(k_1,k_2)\in I_{n,j}^2$ satisfies (\ref{0211201926}).
We are working on $\Omega_{n,j}$ yet.
Suppose that $j\in{\cal L}_n^c$ and $k_1\in\widehat{{\cal K}}_{n,j}(\alpha_0)$.
Then the inequality (\ref{0211202027}) holds
for any $k_2\in I_{n,j}$ satisfying $r_{n,j}(|W_{k_2}|)\geq{\sf a}_n$.
So, there are at least
$\lfloor\alpha_0\overline{\kappa}_n+1\rfloor\big(\leq
\alpha_0\overline{\kappa}_n{\color{black}}% {\color[rgb]{0,0.5,0}+\overline{\kappa}_n^{\>1-\eta_2}}+1\leq
\overline{\kappa}_n-{\sf a}_n+1\big)$ variables $\widehat{U}_{j,k_2}$ that satisfy (\ref{0211202027}).
Then $r_{n,j}(|\widehat{U}_{j,k_1}|)\leq (1-{\color{black}}% {\color[rgb]{0.8,0,0}\alpha_0})\overline{\kappa}_n$, and hence
$k_1\in{\cal K}_{n,j}(\alpha_0)$.
Thus, we found
\begin{eqnarray*}
\widehat{{\cal K}}_{n,j}(\alpha_0) &\subset& {\cal K}_{n,j}(\alpha_0)
\end{eqnarray*}
on $\Omega_{n,j}$ if $j\in{\cal L}_n^c$, that is, (a).
We still work on $\Omega_{n,j}$.
Suppose that $j\in{\cal L}_n^c$ and
$k_2\in{\cal K}_{n,j}(\alpha_0)\setminus\widehat{{\cal K}}_{n,j}(\alpha_0)$.
When $r_{n,j}(|W_{k_2}|)<{\sf a}_n$,
since $r_{n,j}(|W_{k_2}|)>\widehat{\sf a}_n$ due to
$k_2\in{\color{black}}% {\color[rgb]{0.8,0,0}\widehat{{\cal K}}}_{n,j}(\alpha_0)^c$,
we see
\begin{eqnarray}\label{0211211229}
1_{\{j\in{\cal L}_n^c\}}\>
\#\big\{k_2\in {\cal K}_{n,j}(\alpha_0)\setminus\widehat{{\cal K}}_{n,j}(\alpha_0);\>
r_{n,j}(|W_{k_2}|)<{\sf a}_n\big\}
&\leq&
\overline{\kappa}_n^{\>1-\eta_2}
\end{eqnarray}
on $\Omega_{n,j}$.
When $r_{n,j}(|W_{k_2}|)\geq{\sf a}_n$, for
any $k_1$ satisfying $r_{n,j}(|W_{k_1}|)\leq\widehat{\sf a}_n$, we have (\ref{0211202027}).
Therefore
\begin{eqnarray*}
\#\big\{k_1\in I_{n,j};\>|\widehat{U}_{j,k_1}|<|\widehat{U}_{j,k_2}|\big\}
&\geq&
1_{\{j\in{\cal L}_n^c\}}1_{\{r_{n,j}(|W_{k_2}|)\geq{\sf a}_n\}}\>{\color{black}}% {\color[rgb]{0.8,0,0}\widehat{\sf a}}_n,
\end{eqnarray*}
in other words,
\begin{eqnarray}\label{0211211225}
r_{n,j}(|\widehat{U}_{j,k_2}|) &>& {\color{black}}% {\color[rgb]{0.8,0,0}\widehat{\sf a}}_n
\end{eqnarray}
on $\Omega_{n,j}$ if $j\in{\cal L}_n^c$ and $r_{n,j}(|W_{k_2}|)\geq{\sf a}_n$.
Moreover,
$r_{n,j}(|\widehat{U}_{j,k_2}|)\leq\lfloor(1-\alpha_0)\overline{\kappa}_n\rfloor$
since $k_2\in{\cal K}_{n,j}(\alpha_0)$.
Combining this estimate with (\ref{0211211225}), we obtain
\begin{eqnarray}\label{0211211232}
1_{\{j\in{\cal L}_n^c\}}\>
\#\big\{
k_2\in{\cal K}_{n,j}(\alpha_0)\setminus\widehat{{\cal K}}_{n,j}(\alpha_0);\>
r_{n,j}(|W_{k_2}|)\geq{\sf a}_n\big\}
&\leq&
(1-\alpha_0)\overline{\kappa}_n-{\color{black}}% {\color[rgb]{0.8,0,0}\widehat{\sf a}}_n
\nonumber\\&\leq&
{\sred2\>}\overline{\kappa}_n^{\>1-\eta_2}+1
\end{eqnarray}
From (\ref{0211211229}) and (\ref{0211211232}), we obtain (b).
\qed\vspace*{3mm}
For $\eta_3\in{\mathbb R}$, $j\in I_n$ and a sequence of random variables $(V_j)_{j\in I_n}$, let
\begin{eqnarray*}
{\cal D}_{n,j}
&=&
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg|\frac{1}{{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}}\sum_{k\in{\cal K}_{n,j}(\alpha_0)}V_k
-
\frac{1}{{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}}\sum_{k\in\widehat{{\cal K}}_{n,j}(\alpha_0)}V_k
\bigg|
\end{eqnarray*}
The following lemma follows from Lemma \ref{0211200634} immediately.
\begin{lemma}\label{0211231513}
{\bf (i)} Let $p\geq1$. Then
\begin{eqnarray*}
\big\|{\cal D}_{n,j}\big\|_p
&\leq&
{\sred4\>
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3-\eta_2}}\bigg\|\max_{k\in I_{n,j}}|V_k|
{\color{black}}% {{\color[rgb]{0.8,0,0} 1_{\Omega_{n,j}\cap\{j\in{\cal L}_n^c\}}}
\bigg\|_
+{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|\max_{k\in I_{n,j}}|V_k|1_{\Omega_{n,j}^c}\bigg\|_p
+{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|\max_{k\in I_{n,j}}|V_k|1_{\{j\in{\cal L}_n\}}\bigg\|_p
\end{eqnarray*}
for $j\in I_n$, $n\in{\mathbb N}$.
\begin{description}\item[{\bf (ii)}]
Let $p\geq1$ and $\eta_4>0$. Then
\begin{eqnarray*}
\big\|{\cal D}_{n,j}\big\|_p
&\leq&
{\sred4\>
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3-\eta_2}}
\bigg(\kappa_n^{\eta_4}
+{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}\max_{k\in I_{n,j}}\bigg\||V_k|1_{\{|V_k|>\kappa_n^{\eta_4}\}}
{\color{black}}% {{\color[rgb]{0.8,0,0} 1_{\Omega_{n,j}\cap\{j\in{\cal L}_n^c\}}}
\bigg\|_
\bigg)
\nonumber\\&&
+{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|\max_{k\in I_{n,j}}|V_k|1_{\Omega_{n,j}^c}\bigg\|_p
+{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|\max_{k\in I_{n,j}}|V_k|1_{\{j\in{\cal L}_n\}}\bigg\|_p
\end{eqnarray*}
for $j\in I_n$, $n\in{\mathbb N}$.
\end{description}
\end{lemma}
\vspace*{3mm}
Let
\begin{eqnarray*}
\widetilde{{\cal K}}_{n,j}(\alpha_0)
&=&
\big\{k\in I_{n,j};\>|W_k|\leq c(\alpha_0)^{1/2}\big\}
\end{eqnarray*}
For $\eta_3>0$, $j\in I_n$ and a sequence of random variables $(V_j)_{j\in I_n}$, let
\begin{eqnarray*}
\widetilde{{\cal D}}_{n,j}
&=&
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg|\frac{1}{{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}}\sum_{k\in\widehat{{\cal K}}_{n,j}{\color{black}}% {\color[rgb]{0.8,0,0}(\alpha_0)}}V_k
-
\frac{1}{{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}}\sum_{k\in\widetilde{{\cal K}}_{n,j}{\color{black}}% {\color[rgb]{0.8,0,0}(\alpha_0)}}V_k
\bigg|
\end{eqnarray*}
Let
\begin{eqnarray}\label{0211241445}
\widetilde{\Omega}_{n,j}=\big\{\big||W|_{(j,\widehat{\sf a}_n)}-c(\alpha_0)^{1/2}\big|
<\check{C}\kappa_n^{-\eta_2}\big\}
\end{eqnarray}
for $j\in I_n$, where $\check{C}$ is a positive constant.
\begin{lemma}\label{0211231405}
Let $\eta_3\in{\mathbb R}$.
Then
\begin{description}
\item[(i)] For $p\geq1$ and $j\in I_n$,
\begin{eqnarray*}
\big\|\widetilde{{\cal D}}_{n,j}\big\|_p
&\leq&
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|\max_{k'\in I_{n,j}}|V_{k'}|\>
\frac{1}{{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}}\sum_{k\in I_{n,j}}
1_{\big\{\big||{\color{black}}% {\color[rgb]{0.8,0,0} W_k}|-c(\alpha_0)^{1/2}\big|{\color{black}}% {\color[rgb]{0.8,0,0}\><\>}\check{C}\kappa_n^{-\eta_2}\big\}}\bigg\|_p
+
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|1_{\widetilde{\Omega}_{n,j}^c}\max_{k'\in I_{n,j}}|V_{k'}|\bigg\|_p
\end{eqnarray*}
\item[(ii)] For $p_1>p\geq1$ and $j\in I_n$,
\begin{eqnarray*}
\big\|\widetilde{{\cal D}}_{n,j}\big\|_p
&\leq&
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|\max_{k\in I_{n,j}}|V_k|\bigg\|_p
P\bigg[\big||{\color{black}}% {\color[rgb]{0.8,0,0} W_1}|-c(\alpha_0)^{1/2}\big|{\color{black}}% {\color[rgb]{0.8,0,0}\><\>}\check{C}\kappa_n^{-\eta_2}\bigg]
\nonumber\\&&
+
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|\max_{k\in I_{n,j}}|V_k|\bigg\|_{pp_1(p_1-p)^{-1}}
\bigg\|\frac{1}{{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}}\sum_{k\in I_{n,j}}\bigg(
1_{\big\{\big|{\color{black}}% {\color[rgb]{0,0,0.8} |W_k|}-c(\alpha_0)^{1/2}\big|{\color{black}}% {\color[rgb]{0.8,0,0}\><\>} \check{C}\kappa_n^{-\eta_2}\big\}}
\nonumber\\&&\hspace{180pt}
-P\bigg[\big||W_k|-c(\alpha_0)^{1/2}\big|{\color{black}}% {\color[rgb]{0.8,0,0}\><\>}\check{C}\kappa_n^{-\eta_2}\bigg]
\bigg)\bigg\|_{p_1}
\nonumber\\&&
+
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}P\big[\widetilde{\Omega}_{n,j}^c\big]^{1/p_1}
\bigg\|\max_{k\in I_{n,j}}|V_k|\bigg\|_{pp_1(p_1-p)^{-1}}
\end{eqnarray*}
\end{description}
\end{lemma}
\proof
For $k\in I_{n,j}$,
\begin{eqnarray*}&&
\widetilde{\Omega}_{n,j}\cap\big\{\>r_{n,j}(|W_k|)\>\leq\>\widehat{\sf a}_n\big\}^c\cap\big\{|W_k|\leq c(\alpha_0)^{1/2}\big\}
\nonumber\\&=&
\big\{\big||W|_{(j,\widehat{\sf a}_n)}-c(\alpha_0)^{1/2}\big|
<\check{C}\kappa_n^{-\eta_2}\big\}\cap\big\{|W_k|>|W|_{(j,\widehat{\sf a}_n)}\big\}\cap\big\{|W_k|\leq c(\alpha_0)^{1/2}\big\}
\nonumber\\&\subset&
\big\{\big||W_k|-c(\alpha_0)^{1/2}\big|<\check{C}\kappa_n^{-\eta_2}\big\}
\end{eqnarray*}
and
\begin{eqnarray*}&&
\widetilde{\Omega}_{n,j}\cap\big\{\>r_{n,j}(|W_k|)\>\leq\>\widehat{\sf a}_n\big\}\cap\big\{|W_k|\leq c(\alpha_0)^{1/2}\big\}^c
\nonumber\\&=&
\big\{\big||W|_{(j,\widehat{\sf a}_n)}-c(\alpha_0)^{1/2}\big|
<\check{C}\kappa_n^{-\eta_2}\big\}\cap\big\{|W_k|\leq|W|_{(j,\widehat{\sf a}_n)}\big\}\cap\big\{|W_k|> c(\alpha_0)^{1/2}\big\}
\nonumber\\&\subset&
\big\{\big||W_k|-c(\alpha_0)^{1/2}\big|<\check{C}\kappa_n^{-\eta_2}\big\}.
\end{eqnarray*}
Thus we obtain (i).
Property (ii) follows from (i).
\qed\vspace*{3mm}
\begin{lemma}\label{0211241300}
If the constant $\check{C}$ in (\ref{0211280310}) is sufficiently large, then
\begin{eqnarray*}
\sup_{j\in I_n}P\big[\widetilde{\Omega}_{n,j}^c\big]&=&O(n^{-L})
\end{eqnarray*}
as $n\to\infty$ for any $L>0$.
\end{lemma}
\proof
We have
\begin{eqnarray}\label{0211241430}
&&
P\big[|W|_{(j,\widehat{\sf a}_n)}-c(\alpha_0)^{1/2}<-\check{C}\kappa_n^{-\eta_2}\big]
\nonumber\\&\leq&
P\bigg[|W|_{\big(j,\lfloor{\sf a}_n-\overline{\kappa}_n^{\>1-\eta_2}-1\rfloor\big)
}<c(\alpha_0)^{1/2}-\check{C}\kappa_n^{-\eta_2}\bigg]
\nonumber\\&\leq&
P\bigg[\sum_{k\in I_{n,j}}1_{A_{{\color{black}}% {\color[rgb]{0.8,0,0} n,k}}}\geq \lfloor{\sf a}_n-\overline{\kappa}_n^{\>1-\eta_2}-1\rfloor\bigg]
\nonumber\\&=&
P\bigg[\overline{\kappa}_n^{-1/2}\sum_{k\in I_{n,j}}\big\{1_{A_{n,k}}-P[A_{n,k}]\big\}
\geq {\color{black}}% {\color[rgb]{0,0,0.8} C_{n}}
\bigg]
\end{eqnarray}
where
\begin{eqnarray*}
A_{n,k}
&=&
\big\{|W_k|< c(\alpha_0)^{1/2}-\check{C}\kappa_n^{-\eta_2}\big\},
\nonumber\\
{\color{black}}% {\color[rgb]{0,0,0.8} C_{n}}
&=&
\overline{\kappa}_n^{\>-1/2}\big({\sf a}_n-\overline{\kappa}_n^{\>1-\eta_2}-2-\overline{\kappa}_n
P[A_{n,1}]\big).
\end{eqnarray*}
{\color{black}}% {\color[rgb]{0.8,0,0} By using the mean-value theorem, we obtain}
\begin{eqnarray*}
{\color{black}}% {\color[rgb]{0,0,0.8} C_{n}}
&\sim&
\overline{\kappa}_n^{\>-1/2}\bigg[
(1-\alpha_0)\overline{\kappa}_n-2\overline{\kappa}_n^{\>1-\eta_2}
-\overline{\kappa}_n\big\{1-\alpha_0-
{\color{black}}% {\color[rgb]{0,0,0.8} 2 \phi\big(c(\alpha_0)^{1/2};0,1\big)\>}
\check{C}\kappa_n^{-\eta_2}\big\}\bigg]
\nonumber\\&\gtrsim&
\overline{\kappa}_n^{\>{\color{black}}% {\color[rgb]{0,0,0.8} \frac{1}{2}}-\eta_2}
\end{eqnarray*}
as $n\to\infty$ if we choose a sufficiently large $\check{C}$.
Therefore, the $L^p$-boundedness of the random variables in (\ref{0211241430}) gives
\begin{eqnarray}\label{0211241424}
\sup_{j\in I_n}
P\big[|W|_{(j,\widehat{\sf a}_n)}-c(\alpha_0)^{1/2}<-\check{C}\kappa_n^{-\eta_2}\big]
&=&
O(n^{-L})
\end{eqnarray}
as $n\to\infty$ for any $L>0$.
In a similar way, we know
\begin{eqnarray}\label{0211241441}
P\big[|W|_{(j,\widehat{\sf a}_n)}-c(\alpha_0)^{1/2}>\check{C}\kappa_n^{-\eta_2}\big]
&=&
O(n^{-L})
\end{eqnarray}
as $n\to\infty$ for any $L>0$.
Then we obtain the result from (\ref{0211241424}) and (\ref{0211241441}).
\qed\vspace*{3mm}
\begin{en-text}
\begin{eqnarray*}
\widetilde{\Omega}_{n,j}^c
&=&
\big\{|W|_{(j,\widehat{\sf a}_n)}-c(\alpha_0)^{1/2}>\check{C}\kappa_n^{-\eta_2}\big\}
\cup
\big\{|W|_{(j,\widehat{\sf a}_n)}-c(\alpha_0)^{1/2}<-\check{C}\kappa_n^{-\eta_2}\big\}
\end{eqnarray*}
\end{en-text}
\subsection{{\color{black}}% {\color[rgb]{0.8,0,0} Local-global realized volatility}}\label{0301241722}
We introduce the {local-global realized volatility} (LGRV)
\begin{eqnarray}\label{0211260215}
{\mathbb L}_{n,j}(\alpha_0)
&=&
\frac{n}{\overline{\kappa}_nT}
\sum_{k\in{\cal K}_{n,j}(\alpha_0)}q(\alpha_0)^{-1}|\Delta_kX|^2K_{n,k}{\color{black}}% {\color[rgb]{0,0,0.8} .}
\end{eqnarray}
\begin{theorem}\label{0301021509}
Suppose that $[G1]$ is fulfilled.
\begin{en-text}
and that
\begin{eqnarray*}
&c_0\in(0,1), \quad \eta_1\in\bigg(0,\frac{1}{2}\big(\frac{1}{c_0}-1\big)\bigg), \quad
\eta_2\in(0,1/2),&
\nonumber\\
&\eta_3>0,\quad\eta_4>0,\quad \eta_3+\eta_4<\eta_2,
\quad B>0.&
\end{eqnarray*}
\end{en-text}
For $c_0\in(0,1)$ and $B>0$,
suppose that $\kappa_n\sim Bn^{c_0}$ as $n\to\infty$.
Then
\begin{eqnarray}\label{0211221359}
\sup_{n\in{\mathbb N}}
\sup_{j\in I_n}\sup_{k\in I_{n,j}}n^{\>\gamma_*}
\big\|1_{\{j\in{\cal L}_n^c\}}\>\big({\mathbb L}_{n,j}(\alpha_0)-\sigma_{t_k}^2\big)\big\|_p
&<&
\infty
\end{eqnarray}
as $n\to\infty$ for any constant $\gamma_*$ satisfying
\begin{eqnarray*}
\gamma_* &<& \min\bigg\{\frac{1}{2}(1-c_0),\frac{1}{2} c_0\bigg\}.
\end{eqnarray*}
\end{theorem}
\proof
(I)
We have $\kappa_n\sim n^{c_0}\sim h^{-c_0}$ and
$n/\overline{\kappa}_n\sim n^{1-c_0}\sim h^{c_0-1}$.
{\color{black}}% {\color[rgb]{0.8,0,0} Let
\begin{eqnarray*}
{\cal D}_{n,j}^*
&=&
\overline{\kappa}_n^{\eta_3}
\bigg\{
\frac{n}{\overline{\kappa}_n}
\sum_{k\in{\cal K}_{n,j}(\alpha_0)}|\Delta_kX|^2K_{n,k}
-
\frac{n}{\overline{\kappa}_n}
\sum_{k\in\widehat{{\cal K}}_{n,j}(\alpha_0)}|\Delta_kX|^2K_{n,k}\bigg\}
\end{eqnarray*}
{\color{black}}% {\color[rgb]{0.8,0,0}
Applied to $V_k=n|\Delta_kX|^2K_{n,k}1_{\{j\in{\cal L}_n^c\}}$,
Lemma \ref{0211231513} (ii)
}
gives
\begin{eqnarray}\label{0211241150}
\big\|{\cal D}_{n,j}^*1_{\{j\in{\cal L}_n^c\}}\big\|_p
&\leq&
\Phi^{(\ref{0211240511})}_{n,j}+\Phi^{(\ref{0211240512})}_{n,j}
\end{eqnarray}
{\color{black}}% {\color[rgb]{0.8,0,0} for every $p>1$,}
where
\begin{eqnarray}\label{0211240511}
\Phi^{(\ref{0211240511})}_{n,j}
&=&
{\sred4}{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3-\eta_2}}
\bigg(\kappa_n^{\eta_4}
+{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}\max_{k\in I_{n,j}}\bigg\|n|\Delta_kX|^
1_{\{n|\Delta_kX|^
>\kappa_n^{\eta_4}\}}
{\color{black}}% {{\color[rgb]{0.8,0,0} 1_{\{j\in{\cal L}_n^c\}}}
\bigg\|_
\bigg)
\end{eqnarray}
and
\begin{eqnarray}\label{0211240512}
\Phi^{(\ref{0211240512})}_{n,j}
&=&
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|\max_{k\in I_{n,j}}n|\Delta_kX|^2K_{n,k}1_{\Omega_{n,j}^c}\bigg\|_p{\color{black}}% {\color[rgb]{0.8,0,0}.}
\end{eqnarray}
\begin{en-text}
\begin{eqnarray}\label{0211240513}
\Phi^{(\ref{0211240513})}_{n,j}
&=&
\kappa_n^{\eta_3}\bigg\|\max_{k\in I_{n,j}}n|\Delta_kX|^2K_{n,k}|1_{\{j\in{\cal L}_n\}}\bigg\|_p
\end{eqnarray}
for $j\in I_n$, $n\in{\mathbb N}$.
\end{en-text}
Since there is no jump of $J$ on $\{j\in{\cal L}_n^c\}$, we see
\begin{eqnarray}\label{0211240530}
\sup_{j\in I_n}\sup_{k\in I_{n,j}}\big\|n|\Delta_kX|^21_{\{j\in{\cal L}_n^c\}}\big\|_p=O(1)
\end{eqnarray}
for every $p>1$,
as a result, the $L^p$-norm on the right-hand side of (\ref{0211240511})
is of $O(n^{-L})$ for arbitrary $L>0$, and hence
\begin{eqnarray}\label{0211240541}
\Phi^{(\ref{0211240511})}_{n,j}
&=&
O\big(\kappa_n^{\eta_3-\eta_2+\eta_4}\big)
\end{eqnarray}
as $n\to\infty$.
Similarly to (\ref{0211240530}), we obtain
\begin{eqnarray}\label{0211240531}
\sup_{j\in I_n}P\big[\Omega_{n,j}^{{\color{black}}% {\color[rgb]{0.8,0,0} c}}\big] &=& O(n^{-L})
\end{eqnarray}
as $n\to\infty$ for every $L>0$, from
{\color{black}}% {\color[rgb]{0.8,0,0} Lemma \ref{0211191227} as well as}
Lemma \ref{0211190303}
because $(n/\kappa_n)^{1/2}\kappa_n^{-\eta_1}\gg1$ when
$2^{-1}(c_0^{-1}-1)>\eta_1$.
{\color{black}}% {\color[rgb]{0.8,0,0} Then
\begin{eqnarray}\label{0211240533}
\Phi^{(\ref{0211240512})}_{n,j}
&\leq&
\kappa_n^{\eta_3}n^{1/2}
P\big[\Omega_{n,j}^c\big]^{1/p}
\>=\> O(n^{-L})
\end{eqnarray}
for every $L>0$ and $p>1$.
\begin{en-text}
\begin{eqnarray}\label{0211240536}{\coloroy{koko}}
\Phi^{(\ref{0211240513})}_{n,j}
&=&
\kappa_n^{\eta_3}\bigg\|\max_{k\in I_{n,j}}n|\Delta_kX|^2K_{n,k}|1_{\{j\in{\cal L}_n\}}\bigg\|_p
\end{eqnarray}
\end{en-text}
\begin{en-text}
On the event $\{j\in{\cal L}_n^c\}\cap\Omega_{n,j}$, we have
\begin{eqnarray}\label{0211230448}
\sup_{j\in I_n}\big\|1_{\{j\in{\cal L}_n^c\}}1_{\Omega_{n,j}}
{\cal D}_{n,j}^*\big\|_p
&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
\kappa_n^{\eta_3}\times
\frac{n}{\kappa_n}\times\kappa_N^{1-\eta_2}\times h
\>\sim\>
h^{c_0(\eta_2-\eta_3)}
\end{eqnarray}
by Lemma \ref{0211200634}.
\end{en-text}
From (\ref{0211241150}), (\ref{0211240541}) and (\ref{0211240533}),
\begin{eqnarray}\label{0211241151}
\big\|{\cal D}_{n,j}^*1_{\{j\in{\cal L}_n^c\}}\big\|_p
&=&
O\big(\kappa_n^{\eta_3-\eta_2+\eta_4}\big)
\>=\>
O\big(n^{-c_0(\eta_2-\eta_3-\eta_4)}\big)
\end{eqnarray}
as $n\to\infty$ for every $p>1$.
We recall that the parameters should satisfy
\begin{eqnarray*}\label{0211241203}&&
0<\eta_2<\eta_1<\min\left\{\frac{1}{2},\frac{1}{2}\left(\frac{1}{c_0}-1\right)\right\},
\quad
\eta_3+\eta_4<\eta_2{\color{black}}% {\color[rgb]{0,0,0.8} .}
\end{eqnarray*}
{\color{black}}% {\color[rgb]{0.8,0,0}[
In particular, if $c_0=1/2$, then $0<\eta_2<\eta_1<1/2$.
The positive parameters $\eta_3$ and $\eta_4$ can be sufficiently small
at this stage.
Remark that $c_0\eta_2<1/4$ when $c_0\leq1/2$. ]
\vspace*{3mm}
\noindent
(II)
Let
{\color{black}}% {\color[rgb]{0.8,0,0}
\begin{eqnarray*}
\widetilde{{\cal D}}_{n,j}^*
&=&
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\{
\frac{n}{{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}}
\sum_{k\in\widehat{{\cal K}}_{n,j}(\alpha_0)}|\Delta_kX|^2K_{n,k}
-
\frac{n}{{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}}
\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}|\Delta_kX|^2K_{n,k}
\bigg\}{\color{black}}% {\color[rgb]{0,0,0.8} .}
\end{eqnarray*}
{\color{black}}% {\color[rgb]{0.8,0,0} Applying
Lemma \ref{0211231405} (ii) to
$V_k=n|\Delta_kX|^2K_{n,k}1_{\{j\in{\cal L}_n^c\}}$},
we have
\begin{eqnarray}\label{0211241220}
\big\|\widetilde{{\cal D}}_{n,j}^*1_{\{j\in{\cal L}_n^c\}}\big\|_p
&\leq&
\Phi^{(\ref{0211241221})}_{n,j}
+{\color{black}}% {\color[rgb]{0,0,0.8} \Phi^{(\ref{0211241222})}_{n,j}}
+\Phi^{(\ref{0211241223})}_{n,j},
\end{eqnarray}
where
\begin{eqnarray}\label{0211241221}
\Phi^{(\ref{0211241221})}_{n,j}
&=&
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|\max_{k\in I_{n,j}}n|\Delta_kX|^2K_{n,k}1_{\{j\in{\cal L}_n^c\}}\bigg\|_p
P\bigg[\big||W_{{\sred1}}|-c(\alpha_0)^{1/2}\big|{\color{black}}% {\color[rgb]{0,0,0.8} <
\check{C}\kappa_n^{-\eta_2}\bigg],
\end{eqnarray}
\begin{eqnarray}\label{0211241222}
\Phi^{(\ref{0211241222})}_{n,j}
&=&
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}\bigg\|\max_{k\in I_{n,j}}n|\Delta_kX|^2K_{n,k}1_{\{j\in{\cal L}_n^c\}}\bigg\|_{pp_1(p_1-p)^{-1}}
\nonumber\\&&\times
\bigg\|\frac{1}{{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n}}\sum_{k\in I_{n,j}}\bigg(
1_{\big\{\big|{\color{black}}% {\color[rgb]{0,0,0.8} |W_k|}-c(\alpha_0)^{1/2}\big|{\color{black}}% {\color[rgb]{0,0,0.8} <
\check{C}\kappa_n^{-\eta_2}\big\}}
-P\bigg[\big||W_k|-c(\alpha_0)^{1/2}\big|{\color{black}}% {\color[rgb]{0,0,0.8} <
\check{C}\kappa_n^{-\eta_2}\bigg]
\bigg)\bigg\|_{p_1}
\nonumber\\&&
\end{eqnarray}
and
\begin{eqnarray}\label{0211241223}
\Phi^{(\ref{0211241223})}_{n,j}
&=&
{\color{black}}% {\color[rgb]{0.8,0,0}\overline{\kappa}_n^{\eta_3}}P\big[\widetilde{\Omega}_{n,j}^c\big]^{1/p_1}
\bigg\|\max_{k\in I_{n,j}}n|\Delta_kX|^2K_{n,k}1_{\{j\in{\cal L}_n^c\}}\bigg\|_{pp_1(p_1-p)^{-1}}
\end{eqnarray}
for $j\in I_n$, $n\in{\mathbb N}$.
Then,
{\color{black}}% {\color[rgb]{0.8,0,0} paying $\kappa_n^{\eta_4}$ for the maximum,}
we have the following estimates for any $p_1>p\geq1$:
\begin{eqnarray}\label{0211241241}
\sup_{j\in I_n}\Phi^{(\ref{0211241221})}_{n,j}
&=&
O\big(\kappa_n^{\eta_3{\color{black}}% {\color[rgb]{0.8,0,0}+\eta_4}-\eta_2}\big)
=O\big(n^{-c_0(\eta_2-\eta_3{\color{black}}% {\color[rgb]{0.8,0,0}-\eta_4})}\big),
\end{eqnarray}
\begin{eqnarray}\label{0211241242}
\sup_{j\in I_n}\Phi^{(\ref{0211241222})}_{n,j}
&=&
O\big(\kappa_n^{\eta_3}{\color{black}}% {\color[rgb]{0.8,0,0} \times\kappa_n^{\eta_4}}\times\kappa_n^{-(1+\eta_2)/2}\big)
\>=\>
O\big(n^{-c_0\big(\frac{1+\eta_2}{2}-\eta_3{-\color{black}}% {\color[rgb]{0.8,0,0} \eta_4}\big)}\big),
\end{eqnarray}
and
\begin{eqnarray}\label{0211241243}
\sup_{j\in I_n}\Phi^{(\ref{0211241223})}_{n,j}
&=&
O(n^{-L})
\end{eqnarray}
as $n\to\infty$ for any $L>0$
for a sufficiently large $\check{C}$;
the estimate (\ref{0211241243}) follows from Lemma \ref{0211241300}.
In this way,
\begin{eqnarray}\label{0211241453}
\big\|\widetilde{{\cal D}}_{n,j}^*1_{\{j\in{\cal L}_n^c\}}\big\|_p
&=&
O\big(n^{-c_0(\eta_2-\eta_3{\color{black}}% {\color[rgb]{0.8,0,0}-{\eta_4}})}\big)
+O\big(n^{-c_0\big(\frac{1+\eta_2}{2}-\eta_3{\color{black}}% {\color[rgb]{0.8,0,0}-{\eta_4}}\big)}\big)
\end{eqnarray}
as $n\to\infty$ for every $p\geq1$.
\vspace*{3mm}
\noindent
(III)
On the event $\{j\in{\cal L}_n^c\}$, we have
\begin{eqnarray}\label{0301021407}
\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}|\Delta_kX|^2K_{n,k}
&=&
\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\left(\int_{t_{k-1}}^{t_k} \sigma_tdw_t+\int_{t_{k-1}}^{t_k} b_tdt\right)^2K_{n,k}
\nonumber\\&=&
\Phi^{(\ref{0211221131})}_{n,j}+\Phi^{(\ref{0211221132})}_{n,j}+\Phi^{(\ref{0211221133})}_{n,j}+
\Phi^{(\ref{0211221134})}_{n,j}+\Phi^{(\ref{0211221135})}_{n,j}
\end{eqnarray}
where
\begin{eqnarray}\label{0211221131}
\Phi^{(\ref{0211221131})}_{n,j}
&=&
\sum_{k\in I_{n,j}}\big(\sigma_{t_{\underline{j}_n}}\big)^2
hW_k^21_{\{|W_k|\leq c(\alpha_0)^{1/2}\}},
\end{eqnarray}
\begin{eqnarray}\label{0211221132}
\Phi^{(\ref{0211221132})}_{n,j}
&=&
\sum_{k\in I_{n,j}}\big(\sigma_{t_{\underline{j}_n}}\big)^2
hW_k^21_{\{|W_k|\leq c(\alpha_0)^{1/2}\}}\big(K_{n,k}-1\big)
\nonumber\\&&
+2\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\int_{t_{k-1}}^{t_k} \int_{t_{k-1}}^t\big(\widetilde{\sigma}_s-\widetilde{\sigma}_{t_{\underline{j}_n}}\big)dw_s\sigma_tdw_tK_{n,k}
\nonumber\\&&
+2\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\int_{t_{k-1}}^{t_k} \int_{t_{k-1}}^t\sigma_{t_{\underline{j}_n}} dw_s\big(\widetilde{\sigma}_t-\widetilde{\sigma}_{t_{\underline{j}_n}}\big) dw_tK_{n,k}
\nonumber\\&&
{\color{black}}% {{\color[rgb]{0.8,0,0}
+2\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\int_{t_{k-1}}^{t_k} \widetilde{\sigma}_{t_{\underline{j}_n}}
\big(\widetilde{\sigma}_t-\widetilde{\sigma}_{t_{\underline{j}_n}}\big)dtK_{n,k}
}
\nonumber\\&&
{\color{black}}% {{\color[rgb]{0.8,0,0}
+\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\bigg(\int_{t_{k-1}}^{t_k}
\big(\widetilde{\sigma}_t-\widetilde{\sigma}_{t_{\underline{j}_n}}\big)dw_t\bigg)^2K_{n,k}
},
\end{eqnarray}
\begin{eqnarray}\label{0211221133}
\Phi^{(\ref{0211221133})}_{n,j}
&=&
2\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\int_{t_{k-1}}^{t_k} \int_{t_{k-1}}^tb_sds\sigma_tdw_tK_{n,k}{\color{black}}% {\color[rgb]{0.8,0,0},}
\end{eqnarray}
\begin{eqnarray}\label{0211221134}
\Phi^{(\ref{0211221134})}_{n,j}
&=&
2\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\int_{t_{k-1}}^{t_k} \int_{t_{k-1}}^t\sigma_sdw_sb_tdtK_{n,k},
\end{eqnarray}
and
\begin{eqnarray}\label{0211221135}
\Phi^{(\ref{0211221135})}_{n,j}
&=&
2\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\int_{t_{k-1}}^{t_k} \int_{t_{k-1}}^tb_sdsb_tdtK_{n,k}.
\end{eqnarray}
\begin{en-text}
{\color{black}}% {{\color[rgb]{0.8,0,0}
On $\{j\in{\cal L}_n^c\}$,
\begin{eqnarray*}
\Phi^{(\ref{0211221132})}_{n,j}
&=&
\sum_{k\in I_{n,j}}\big(\sigma_{t_{\underline{j}_n}}\big)^2
W_k^21_{\{|W_k|\leq c(\alpha_0)^{1/2}\}}\big(K_{n,k}-1\big)
\nonumber\\&&
+2\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\int_{t_{k-1}}^{t_k} \int_{t_{k-1}}^t\big(\widetilde{\sigma}_s-\widetilde{\sigma}_{t_{\underline{j}_n}}\big)dw_s\sigma_tdw_tK_{n,k}
\nonumber\\&&
+2\int_{t_{k-1}}^{t_k} \int_{t_{k-1}}^t\sigma_{t_{\underline{j}_n}} dw_s\big(\widetilde{\sigma}_t-\widetilde{\sigma}_{t_{\underline{j}_n}}\big) dw_tK_{n,k}
\end{eqnarray*}
}
\end{en-text}
By assumption,
\begin{eqnarray}\label{0211260123}&&
\sup_{j\in I_n}{{\color{black}}% {\color[rgb]{0.8,0,0} \sup_{s\in[t_{\underline{j}_n-1},t_{\underline{j}_n+\overline{\kappa}_n}]}}}
\big\|1_{\{j\in{\cal L}_n^c\}}\big(\sigma_{{\color{black}}% {\color[rgb]{0.8,0,0} s}}^2-\sigma_{t_{\underline{j}_n-1}}^2\big)\big\|_p
\nonumber\\&\>\leq\> &
\sup_{j\in I_n}{{\color{black}}% {\color[rgb]{0.8,0,0} \sup_{s\in[t_{\underline{j}_n-1},t_{\underline{j}_n+\overline{\kappa}_n}]}}}
\big\|\widetilde{\sigma}_{{\color{black}}% {\color[rgb]{0.8,0,0} s}}^2-\widetilde{\sigma}_{t_{\underline{j}_n-1}}^2\big\|_p
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
(\kappa_nh)^{1/2}
\>\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ \>
h^{\frac{1}{2}(1-c_0)}
\end{eqnarray}
for every $p>1$.
First, {\color{black}}% {\color[rgb]{0.8,0,0} a primitive estimate gives}
\begin{en-text}
\begin{eqnarray}\label{0211221143}
\frac{n}{\kappa_n}
\big\|\Phi^{(\ref{0211221132})}_{n,j}1_{\{j\in{\cal L}_n^c\}}\big\|_p
&\leq&
\frac{n}{\kappa_n}
\bigg\|\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\int_{t_{k-1}}^{t_k} \int_{t_{k-1}}^t\sigma_sdw_s\sigma_tdw_t\bigg\|_p
\nonumber\\&&
+
\frac{n}{\kappa_n}
\bigg\|\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}
\int_{t_{k-1}}^{t_k} \int_{t_{k-1}}^t\sigma_sdw_s\sigma_tdw_t(1-K_{n,k})1_{\{j\in{\cal L}_n^c\}}\bigg\|_p
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
\frac{n}{\kappa_n}\big(\kappa_nh^2)^{1/2}
+
P\bigg[\max_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}(1-K_{n,k})1_{\{j\in{\cal L}_n^c\}}\bigg]^{1/(2p)}
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
h^{c_0/2}+\kappa_n h^L
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
h^{c_0/2}
\end{eqnarray}
as $n\to\infty$, where $L$ is any number bigger than $3c_0/2$.
\end{en-text}
\begin{eqnarray}\label{0211221143}
\sup_{j\in I_n}
\frac{n}{\kappa_n}
\big\|\Phi^{(\ref{0211221132})}_{n,j}1_{\{j\in{\cal L}_n^c\}}\big\|_p
&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
\frac{n}{\kappa_n}\times\frac{\kappa_n^{3/2}}{n^{3/2}}
\>\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ \>
h^{\frac{1}{2} (1-c_0)}
\end{eqnarray}
as $n\to\infty$;
we note that the orthogonality cannot apply due to $\widetilde{{\cal K}}_{n,j}(\alpha_0)$
even after $K_{n,k}$ is decoupled.
We also have
\begin{eqnarray}\label{0211221212}
\sup_{j\in I_n}
\frac{n}{\kappa_n}
\big\|\Phi^{(\ref{0211221133})}_{n,j}1_{\{j\in{\cal L}_n^c\}}\big\|_p
&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
{\color{black}}% {{\color[rgb]{0.8,0,0} h^{1/2}}{\color{black}}% {\color[rgb]{0,0,0.8} .
\end{eqnarray}
For {\color{black}}% {\color[rgb]{0.8,0,0}$\Phi^{(\ref{0211221134})}_{n,j}$ and $\Phi^{(\ref{0211221135})}_{n,j}$},
by the same way, we can get
\begin{eqnarray}\label{0211221214}
\sup_{j\in I_n}
\frac{n}{\kappa_n}
\big\|\Phi^{(\ref{0211221134})}_{n,j}1_{\{j\in{\cal L}_n^c\}}\big\|_p
&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
h^{1/2},
\end{eqnarray}
and
\begin{eqnarray}\label{0211221215}
\sup_{j\in I_n}
\frac{n}{\kappa_n}
\big\|\Phi^{(\ref{0211221135})}_{n,j}1_{\{j\in{\cal L}_n^c\}}\big\|_p
&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
{\color{black}}% {{\color[rgb]{0.8,0,0} h}
\end{eqnarray}
as $n\to\infty$.
Furthermore, we hav
\begin{eqnarray}\label{0211260110}&&
\sup_{j\in I_n}
\bigg\|
\bigg\{
\frac{1}{\overline{\kappa}_nh}
\Phi^{(\ref{0211221131})}_{n,j}
-\big(\sigma_{t_{\underline{j}_n}}\big)^2q(\alpha_0)\bigg\}
1_{\{j\in{\cal L}_n^c\}}
\bigg\|_p
\nonumber\\&\leq&
\sup_{j\in I_n}
\bigg\|
\frac{1}{\overline{\kappa}_nh}\bigg\{
\Phi^{(\ref{0211221131})}_{n,j}
-\sum_{k\in I_{n,j}}\big(\sigma_{t_{\underline{j}_n}}\big)^2q(\alpha_0)h\bigg\}
1_{\{j\in{\cal L}_n^c\}}
\bigg\|_p
\nonumber\\&\leq&
\sup_{j\in I_n}
\bigg\|
\frac{1}{\overline{\kappa}_n}
\sum_{k\in I_{n,j}}\big(\sigma_{t_{\underline{j}_n}}\big)^2
\big(W_k^21_{\{|W_k|\leq c(\alpha_0)^{1/2}\}}-q(\alpha_0)\big)
\bigg\|_p
\nonumber\\&=&
O(\kappa_n^{-1/2})
\>=\>
O(h^{c_0/2})
\end{eqnarray}
for every $p>1${\color{black}}% {\color[rgb]{0.8,0,0}.}
\begin{en-text}
\begin{eqnarray}\label{0211260119}&&
\sup_{j\in I_n}
\bigg\|
\frac{1}{\overline{\kappa}_nh}
\sum_{k\in I_{n,j}}\bigg\{
\big(\sigma_{t_{\underline{j}_n}}\big)^2q(\alpha_0)h
-\int_{t_{k-1}}^{t_k}\sigma_t^2dt
\bigg\}
1_{\{j\in{\cal L}_n^c\}}
\bigg\|_p
\nonumber\\&=&
O(h^{\frac{1}{2}(1-c_0)})
\end{eqnarray}
by using (\ref{0211260123}).
We have {\coloroy{koko}}
\begin{eqnarray*} &&
\bigg\|1_{\{j\in{\cal L}_n^c\}}\bigg(
\frac{n}{\overline{\kappa}_nT}\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}\int_{t_{k-1}}^{t_k} \sigma_t^2dtK_{n,k}
-q(\alpha_0)\sigma_{t_{\underline{j}_n-1}}^2\bigg)\bigg\|_p
\nonumber\\&\leq&
\bigg\|1_{\{j\in{\cal L}_n^c\}}
\frac{n}{\overline{\kappa}_nT}\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}\int_{t_{k-1}}^{t_k}
\big(\widetilde{\sigma}_t^2-\widetilde{\sigma}_{t_{\underline{j}_n-1}}^2\big)dtK_{n,k}\bigg\|_p
\nonumber\\&&
+
\bigg\
\frac{1}{\overline{\kappa}_n}\sum_{k\in I_{n,j}}
\sigma_{t_{\underline{j}_n-1}}^2
\big(1_{\{|W_k|<c(\alpha_0)^{1/2}\}}-q(\alpha_0)\big)\bigg\|_p
\nonumber\\&&
+
\bigg\|1_{\{j\in{\cal L}_n^c\}}
\frac{1}{\overline{\kappa}_n}\sum_{k\in I_{n,j}}
\sigma_{t_{\underline{j}_n-1}}^2
1_{\{|W_k|<c(\alpha_0)^{1/2}\}}
(K_{n,k}-1)\bigg\|_p
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
(\kappa_nh)^{1/2}+\kappa_n^{-1/2}+O(n^{-L})
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
(\kappa_nh)^{1/2}+\kappa_n^{-1/2}
\end{eqnarray*}
as well as
\begin{eqnarray*}
\big\|1_{\{j\in{\cal L}_n^c\}}\big(\sigma_{t_k}^2-\sigma_{t_{\underline{j}_n-1}}^2\big)\big\|_p
\>\leq\>
\big\|\widetilde{\sigma}_{t_k}^2-\widetilde{\sigma}_{t_{\underline{j}_n-1}}^2\big\|_p
\>\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ \>
(\kappa_nh)^{1/2}
\end{eqnarray*}
for every $p>1$.
\end{en-text}
Combining {\color{black}}% {\color[rgb]{0.8,0,0} (\ref{0301021407}) and}
(\ref{0211260123})-(\ref{0211260110}), we obtain
\begin{eqnarray}\label{0211221357}&&
\sup_{j\in I_n}
\sup_{k'\in I_{n,j}}
\bigg\|1_{\{j\in{\cal L}_n^c\}}\bigg(
\frac{n}{\overline{\kappa}_nT}\sum_{k\in\widetilde{{\cal K}}_{n,j}(\alpha_0)}|\Delta_kX|^2K_{n,k}
-\sigma_{t_{k'}}^2q(\alpha_0)\bigg)\bigg\|_p
\nonumber\\&=&
O\big(n^{-(1-c_0)/2}\big)+O(n^{-c_0/2})
\end{eqnarray}
as $n\to\infty$
for every $p>1$.
\begin{en-text}
The estimates
(\ref{0211221143}), (\ref{0211221212}),
(\ref{0211221214}), (\ref{0211221215}),
and (\ref{0211221357}) hold uniformly in $j\in I_n$.
\end{en-text}
\noindent
(IV)
From (\ref{0211241151}), (\ref{0211241453}) and (\ref{0211221357}),
we obtain the estimate
\begin{eqnarray}\label{0301021434}&&
\sup_{j\in I_n}
\sup_{k'\in I_{n,j}}{\color{black}}% {\color[rgb]{0.8,0,0}\kappa_n^{\eta_3}}
\bigg\|1_{\{j\in{\cal L}_n^c\}}\bigg(
\frac{n}{\overline{\kappa}_nT}\sum_{k\in{{\color{black}}% {\color[rgb]{0.8,0,0}{\cal K}}}_{n,j}(\alpha_0)}|\Delta_kX|^2K_{n,k}
-\sigma_{t_{k'}}^2q(\alpha_0)\bigg)\bigg\|_p
\nonumber\\&=&
O\big(n^{-c_0(\eta_2-\eta_3-\eta_4)}\big)
+\bigg\{
O\big(n^{-c_0(\eta_2-\eta_3{\color{black}}% {\color[rgb]{0.8,0,0}-{\eta_4}})}\big)
+O\big(n^{-c_0\big(\frac{1+\eta_2}{2}-\eta_3{\color{black}}% {\color[rgb]{0.8,0,0}-{\eta_4}}\big)}\big)\bigg\}
\nonumber\\&&
+
{\color{black}}% {\color[rgb]{0.8,0,0}\kappa_n^{\eta_3}}\bigg\{O\big(n^{-(1-c_0)/2}\big)+O(n^{-c_0/2})\bigg\}
\nonumber\\&=&
O(n^{-c_0(\eta_2-\eta_3-\eta_4)})+O\big(n^{c_0{\color{black}}% {\color[rgb]{0.8,0,0}(\eta_3+\eta_4)}-(1-c_0)/2}\big)
\>=:\> {\mathbb O}_n
\end{eqnarray}
as $n\to\infty$ for every $p>1$.
Here we are assuming the parameters satisfy
\begin{eqnarray}\label{0211260224}
&c_0\in(0,1), \quad B>0, \quad \eta_1\in\bigg(0,\min\bigg\{\frac{1}{2}\big(\frac{1}{c_0}-1\big),\frac{1}{2}\bigg\}\bigg), &
\nonumber\\
&\eta_2\in(0,\eta_1),\quad\eta_3>0,\quad\eta_4>0,\quad \eta_3+\eta_4<\eta_2.&
\end{eqnarray}
{\color{black}}% {\color[rgb]{0.8,0,0}
To obtain the last error bound in (\ref{0301021434}), we used the inequalities
\begin{eqnarray*}
-c_0\bigg(\frac{1+\eta_2}{2}-\eta_3-\eta_4\bigg)
<
-c_0\big(\eta_2-\eta_3-\eta_4\big)
\end{eqnarray*}
and
\begin{eqnarray*}
c_0\eta_3-\frac{c_0}{2}
<
c_0\eta_3-c_0\eta_2
<
-c_0(\eta_2-\eta_3-\eta_4).
\end{eqnarray*}
The LGRV ${\mathbb L}_{n,j}(\alpha_0)$ of (\ref{0211260215}) does not depend on
$\eta_i$ ($i=1,2,3,4$) within the ranges (\ref{0211260224}).
When ${\color{black}}% {\color[rgb]{0.8,0,0} c_0}>1/2$, we make
\begin{eqnarray*}
\frac{1}{2}>\frac{1}{2}\left(\frac{1}{c_0}-1\right)>\eta_1>\eta_2>\eta_3
\uparrow\frac{1}{2}\left(\frac{1}{c_0}-1\right),\quad \eta_4\down0
\end{eqnarray*}
to obtain {\color{black}}% {\color[rgb]{0.8,0,0}${\mathbb O}_n=O(1)$}.
When ${\color{black}}% {\color[rgb]{0.8,0,0} c_0}\leq1/2$, we make
\begin{eqnarray*}
\frac{1}{2}>\eta_1>\eta_2>\eta_3\uparrow\frac{1}{2},\quad\eta_4\down0
\end{eqnarray*}
to obtain {\color{black}}% {\color[rgb]{0.8,0,0}${\mathbb O}_n=O(1)$}.
Thus, the proof {\color{black}}% {\color[rgb]{0.8,0,0} of Theorem \ref{0301021509}} is concluded.
\qed\vspace*{3mm}
According to the error bound (\ref{0211221359}),
we should in general take
$c_0=1/2$, i.e.,
$\kappa_n\sim B n^{1/2}$ to obtain an optimal error estimate.
However, this is not always true. If the process $\sigma$ is (unknown) constant for example, then
we do not need any spot volatility estimator to construct
a global jump filter, and the convergence of the resulting estimator for $\Theta$
becomes much faster than that in the non-constant $\sigma$ case.
\subsection{Local minimum RV}\label{0301241726}
Estimation of spot volatilities can be done by the minimum realized volatility
{\color{black}}% {\color[rgb]{0.8,0,0} (minRV)} method
of Andersen et al. \cite{Andersen2012}.
{\color{black}}% {\color[rgb]{0.8,0,0} This method is localized to define
the local minRV by }
\begin{eqnarray} \label{loc_minRV}
{\mathbb M}_{n, j} = \frac{\pi}{\pi-2} \frac{n}{\bar{\kappa}_n T}
\sum_{k \in I_{n,j}} \big\{ |\Delta_k X| \wedge |\Delta_{k+1} X| \}^2.
\end{eqnarray}
\begin{theorem}
Suppose that $[G1]$ is fulfilled.
For $c_0\in(0,1)$ and $B>0$,
suppose that $\kappa_n\sim Bn^{c_0}$ as $n\to\infty$.
Then
\begin{eqnarray*}
\sup_{n\in{\mathbb N}}
\sup_{j\in I_n}\sup_{k\in I_{n,j}}n^{{\vred \gamma_{**}}
\big\
{\fred1_{\{j\in{\cal L}_n^c\}}\>\big(
{\mathbb M}_{n,j} - \sigma_{t_k}^2 \big)}
\big\|_p
&<&
\infty
\end{eqnarray*}
as $n\to\infty$
{\color{black}}% {\color[rgb]{0.8,0,0} for any $p>1$ and any constant ${\vred \gamma_{**}}$ satisfying
\begin{eqnarray*}
{\vred \gamma_{**}} &{\vred=}& \min\bigg\{\frac{1}{2}(1-c_0),\frac{1}{2} c_0\bigg\}.
\end{eqnarray*}
}
\end{theorem}
{\fblue
\begin{proof}
Consider $k \in {\vred{ I}}_{n,j}$ for $ j \in {\vred{\cal L}}_n^c$. Then we can decompose
$\Delta_k X$ as
\begin{eqnarray*}
\Delta_k X = \sigma_{t_{\underline{j}_n}} \Delta_k w
+ \int_{t_{k-1}}^{t_k} (\sigma_t - \sigma_{t_{\underline{j}_n}}) dw_t
+ \int_{t_{k-1}}^{t_k} b_t dt.
\end{eqnarray*}
By (\ref{0211260123}),
\begin{eqnarray*}&&
\sup_{j \in I_n} \sup_{k \in{\vred { I}}_{n,j}} \Bigg\|
\int_{t_{k-1}}^{t_k} (\sigma_t - \sigma_{t_{\underline{j}_n}}) dw_t \> 1_{\{ j \in {\vred{\cal L}}_n^c \}} \Bigg\|_{{\vred p}}
\nonumber\\&{\vred\leq} &
{\vred
\sup_{j \in I_n} \sup_{k \in { I}_{n,j}} \Bigg\|
\int_{t_{k-1}}^{t_k} (\widetilde{\sigma}_t - \widetilde{\sigma}_{t_{\underline{j}_n}}) dw_t \Bigg\|_{2p}
\> \big\|1_{\{ j \in {\cal J}_n^c \}} \big\|_{2p}
}\\
&{\vred\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ }&
\sup_{j \in I_n} \sup_{k \in {\vred I}_{n,j}}
\sqrt{
\int_{t_{k-1}}^{t_k} \Big\|
{\vred
\big(\widetilde{\sigma}_t - \widetilde{\sigma}_{t_{\underline{j}_n}} \big)^2
}
\Big\|_{p} dt } \\
&{\vred=}&
\sqrt{O\big( h \times \kappa_n h \big)} = O\big(h^{1-\frac{1}{2}c_0} \big)
\end{eqnarray*}
{\vred for every $p>1$.}
Hence, we obtain
\begin{eqnarray*}
|\Delta_k X|^2 = \sigma_{t_{\underline{j}_n}}^2 h W_k^2 + \mathcal{X}_k,
\end{eqnarray*}
where $\mathcal{X}_k$ is a random variable satisfying
$\sup_{j \in I_n} \sup_{k \in {\vred I}_{n,j}} \| \mathcal{X}_k\|_p = O(h^{\frac{1}{2}(3-c_0)})$. By using this approximation (and the equality $a \wedge b = \frac{1}{2}(a+b - |a-b|)$ for $a,b > 0$), we have
\begin{align*}
|\Delta_k X|^2 \wedge |\Delta_{k+1} X|^2 =
\sigma_{t_{\underline{j}_n}}^2 h \>
{\vred\big(}W_k^2 \wedge W_{k+1}^2{\vred\big)} + \mathcal{X}_k',
\end{align*}
where $\mathcal{X}_k'$ is a random variable satisfying
$\sup_{j \in I_n} \sup_{k \in {\vred I}_{n,j}} \| \mathcal{X}_k'\|_p = O(h^{\frac{1}{2}(3-c_0)})$.
Hence, we obtain
\begin{eqnarray} \label{locminRV-1}
{\mathbb M}_{n, j} - \sigma_{t_k}^2
&=&
\frac{\pi}{\pi-2} \frac{n}{\overline{\kappa}_n T}
\sum_{k \in {\vred I}_{n,j}}
{\vred\big\{}|\Delta_k X|^2 \wedge |\Delta_{k+1} X|^2{\vred\big\}}
- \sigma_{t_k}^2 \nonumber \\
&=&
\frac{1}{\overline{\kappa}_n}
\sum_{k \in {\vred I}_{n,j}} \sigma_{t_{\underline{j}_n}}^2
\Bigg( \frac{\pi}{\pi-2} {\vred\big\{}W_k^2 \wedge W_{k+1}^2{\vred\big\}} - 1 \Bigg)
+ \big(\sigma_{t_{\underline{j}_n}}^2 - \sigma_{t_k}^2 \big)
\nonumber\\&&
+ \frac{\pi}{\pi-2} \frac{n}{\overline{\kappa}_n T} \sum_{k \in {\vred I}_{n,j}}\mathcal{X}_k'
\end{eqnarray}
{\vred The} first term on the right-hand side of (\ref{locminRV-1}) is $O(\kappa_n^{-1/2}) = O(h^{c_0/2})$.
As for the second term, (\ref{0211260123}) gives
$\| \sigma_{t_{\underline{j}_n}}^2 - \sigma_{t_k}^2 \|_p = O(h^{\frac{1}{2}(1-c_0)})$.
Finally, as for the third term, we can estimate as
\begin{eqnarray*}
\Bigg\| \frac{n}{\overline{\kappa}_n T}
\sum_{k \in {\vred I}_{n,j}}\mathcal{X}_k' \Bigg\|_p
\lesssim
n \times O\Big(h^{\frac{1}{2}(3-c_0)}\Big) =
O\Big(h^{\frac{1}{2}(1-c_0)}\Big).
\end{eqnarray*}
With these estimates, we obtain the desired result.
\end{proof}
}
\section{Rate of convergence of the global realized volatilities
in high intensity of jumps}\label{0211181814}
{\color{black}}% {\color[rgb]{0.8,0,0}
In this section, we present a rate of convergence of the GRV and WGRV,
both defined in Section \ref{Sec2}.}
When the frequency of the jumps is high,
it is recommend that one should choose a value of $\alpha$ that is not extremely small
in order to cover the jumps by the index set ${\cal J}_n(\alpha)^c$.
{\color{black}}% {\color[rgb]{0.8,0,0}
We will assume the properties of $S_{n,j-1}$ below,
that we already proved in Section \ref{Sec3} for the LGRV and the local minRV.
Thus, GRV and WGRV with a LGRV or
the local minRV are global realized volatilities.
}
\begin{description}
\item[{\bf [G2]}]
{\bf (i)} $S_{n,j-1}$ is positive a.s. and
\begin{eqnarray*}
{\color{black}}% {\color[rgb]{0.8,0,0}\sup_{n\in{\mathbb N}}\sup_{j\in I_n}}\
\big\|S_{n,j-1}^{-1}\big\|_p&<&\infty
\end{eqnarray*}
for every $p>1$.
\begin{description}
\item[(ii)]
There exist positive constants $\gamma_0$ and $c$ such that
\begin{eqnarray*}
\sup_{n\in{\mathbb N}}\sup_{j\in I_n}
n^{\gamma_0}\big\|
1_{\{j\in{\cal L}_n^c\}}\big(
\sigma_{t_{j-1}}^2-c\>S_{n,j-1}\big)\big\|_p
&<& \infty
\end{eqnarray*}
for every $p>1$.
\end{description}
\end{description}
In $[G2]$, we do not assume that the value of constant $c$ is known.
We note that
\begin{eqnarray*}
{\color{black}}% {\color[rgb]{0.8,0,0}\sup_{n\in{\mathbb N}}\sup_{j\in I_n}}\
\big\|1_{\{j\in{\cal L}_n^c\}}S_{n,j-1}\big\|_p
&<& \infty
\end{eqnarray*}
for every $p>1$ under $[G1]$ and $[G2]$.
{\color{black}}% {\color[rgb]{0.8,0,0} As shown in Theorem \ref{0301021509},
the LGRV in (\ref{0211260215}) can serve as $S_{n,j-1}$.
}
If $\sigma_t$ is equal to a (possibly unknown) constant, then
$\gamma_0$ can be arbitrarily large since we can let $S_{n,j-1}=1$.
In other words, we do not need any pre-estimate of $\sigma_{t_{j-1}}^2$.
So, the constant volatility case is very special and it will
be discussed briefly in Section \ref{0211281401} separately.
This section logically includes the constant volatility case
(hence a less efficient way for it)
but we will consider a general non-constant volatility
and assume a given local estimator attains a limited rate of convergence.
\begin{remark}\rm
When ${\sf v}=2^{-1}\inf_{\omega\in\Omega,t\in[0,T]}\sigma_t^2>0$
for a priori known constant ${\sf v}$,
given a local estimator
${\mathbb L}^{loc}_{n,j-1}$ of $\sigma_{t_{j-1}}^2$,
we can use $S_{n,j-1}({\sf v})={\mathbb L}^{loc}_{n,j}\vee{\sf v}$ for $S_{n,j-1}$.
For example, it is the case when $X$ satisfies a stochastic differential equation
with jumps and its diffusion coefficient is uniformly elliptic.
When ${\sf v}=0$, an appropriate modification of ${\mathbb L}^{loc}_{n,j}$ is necessary
and possible. We only give an idea without going into details here.
Preset a positive constant ${\sf v}$.
Using $S_{n,j-1}({\sf v})$ for $S_{n,j-1}$, we obtain an estimator $\widetilde{{\mathbb V}}_n[{\sf v}]$ of
$\Theta({\sf v})=\int_0^T\sigma_t^21_{\{\sigma_t^2\geq{\sf v}\}}dt$, and indeed,
the rate of convergence $\widetilde{{\mathbb V}}_n[{\sf v}]$ is established in this paper.
Then it is natural to use $\widetilde{{\mathbb V}}_n[{\sf v}_n]$ to estimate
$\Theta=\int_0^T\sigma_t^2dt$ with a sequence of numbers ${\sf v}_n$
tending to $0$ as $n\to\infty$.
Consistency does not matter because the mappting ${\sf v}\mapsto \Theta({\sf v})$ is
continuous and the operation ${\sf v}_n\down0$ is stable.
Some work is necessary to give an explicit rate of convergence since
the constant of the error bound for each ${\sf v}_n$ depends on ${\sf v}_n$.
However, the cause of the error by the truncation at level ${\sf v}_n$ is
the difference
$\int_0^T\sigma_t^21_{\{\sigma_t^2<{\sf v}_n\}}dt$, and it is rather easy to control
for small ${\sf v}_n$.
\end{remark}
\subsection{Rate of convergence of the GRV with a fixed $\alpha$}
We consider the GRV given by (\ref{0211260834}):
\begin{eqnarray*}
{\mathbb V}_n(\alpha)
&=&
\sum_{j\in{\cal J}_n(\alpha)}q(\alpha)^{-1}|\Delta_jX|^2K_{n,j}{\color{black}}% {\color[rgb]{0,0,0.8} .}
\end{eqnarray*}
Denote by $r_n({\tt U}_j)$ the rank of ${\tt U}_j$ among the variables $\{{\tt U}_i\}_{i\in I_n}$
as before, and
$|{\tt U}|_{(r)}$ denotes the $r$-th ordered statistic of $\{|{\tt U}_i|\}_{i\in I_n}$.
Let $0<\gamma_2<\gamma_1<\gamma_0$, and
define numbers $a_n$ and $\widehat{a}_n$ by
\begin{eqnarray*}
a_n\>=\>\lfloor (1-\alpha)n-n^{1-\gamma_2}\rfloor\quad\text{and}\quad
\widehat{a}_n\>=\>\lfloor a_n-n^{1-\gamma_2}\rfloor,
\end{eqnarray*}
respectively.
Define the event $N_{n,j}$ by
\begin{eqnarray*}
N_{n,j}
&=&
\big\{r_n(|W_j|)\leq a_n-n^{1-\gamma_2}\big\}\cap
\big\{|W|_{(a_n)}-|W_j|< n^{-\gamma_1}\big\}
\end{eqnarray*}
The following lemma is
Lemma 2.6 of {\color{black}}% {\color[rgb]{0.8,0,0} Inatsugu} and Yoshida {\color{black}}% {\color[rgb]{0,0,0.8}\cite{InatsuguYoshida2020Accepted}}{\color{black}}% {\color[rgb]{0,0,0.8} .}
\begin{lemma}\label{0211260910}
\begin{eqnarray*}
P {\color{black}}% {\color[rgb]{0,0,0.8} \Bigg[} \bigcup_{j=1,..,,n}N_{n,j} {\color{black}}% {\color[rgb]{0,0,0.8} \Bigg]} \>=\> O(n^{-L})
\end{eqnarray*}
as $n\to\infty$ for every $L>0$.
\end{lemma}
We need some notation:
{\color{black}}% {\color[rgb]{0.8,0,0}
\begin{eqnarray*}
\widehat{{\cal J}}_n(\alpha)
&=&
\big\{j\in I_n;\> r_n(|W_j|)\leq \widehat{a}_n\big\},\quad
\nonumber\\
U_j &=& {\color{black}}% {\color[rgb]{0.8,0,0} c}^{-1/2}h^{-1/2}(S_{n,j-1})^{-1/2}\Delta_jX
\nonumber\\
R_j &=& U_j-W_j- {\color{black}}% {\color[rgb]{0.8,0,0} c}^{-1/2}h^{-1/2}(S_{n,j-1})^{-1/2}\Delta_jJ,
\end{eqnarray*}
as well
\begin{en-text}
\begin{eqnarray*}
\Omega_n &=& \bigg\{\overline{N}_T< n^{1-\gamma_2}\bigg\} {\color{black}}% {\color[rgb]{0.8,0,0} \bigcap}
\bigg(\bigcap_{j=1,...,n}\bigg[\big\{|R_j|1_{\{\Delta_jN^\sigma=0\}}<2^{-1}n^{-\gamma_1}\big\}\cap(N_{n,j})^c\bigg]
\bigg).
\end{eqnarray*}
\end{en-text}
\begin{eqnarray*}
\Omega_n &=& \bigg\{
{\xred\#\>{\cal L}_n}
< n^{1-\gamma_2}\bigg\} {\color{black}}% {\color[rgb]{0.8,0,0} \bigcap}
\bigg(\bigcap_{j=1,...,n}\bigg[\big\{|R_j|1_{\{{\xred j\in{\cal L}_n^c}\}}<2^{-1}n^{-\gamma_1}\big\}\cap(N_{n,j})^c\bigg]
\bigg).
\end{eqnarray*}
{\xred
Let
\begin{eqnarray}\label{0211280352}
{\mathfrak L}_n &=& \big\{j\in I_n;\> \Delta_j\overline{N}\not=0\big\}.
\end{eqnarray}
The definition of ${\cal L}_n^{(k)}$
in Inatsugu and Yoshida \cite{InatsuguYoshida2020Accepted}
of the extended version arXiv:1806.10706v3 is essentially
the same as ${\mathfrak L}_n$, and different from
${\cal L}_n$ defined by (\ref{0301021550}).
The random set ${\sf L}_n^{(k)}$ therein corresponds to ${\cal L}_n$.
}
{\xred
We assume that the distribution of the variable $\overline{N}_T$ depends on $n$,
and consider the case where $\overline{N}_T$ may diverge as $n\to\infty$.
More precisely, we will assume the following situation.
\footnote{We slightly relaxed Condition $[G3]$ of arXiv:2102.05307v1.}
}
{\color{black}}% {\color[rgb]{0.8,0,0}
\begin{description}
\item[{\bf [G3]}]
There exists a constant $\xi\geq0$
such that
$\|{\xred\#\>{\mathfrak L}_n}\|_p=O(n^{\xi})$
as $n\to\infty$
for every $p>1$.
\end{description}
\begin{lemma}\label{0211261336}
Suppose that $[G1]$ and $[G2]$ are satisfied.
Suppose that $0<\gamma_1<\gamma_0<1/2$. Then
\begin{eqnarray}\label{02112q61347}
\sup_{j\in I_n} P\big[
{\color{black}}% {\color[rgb]{0.8,0,0} |R_j|}1_{\{{\xred j\in{\cal L}_n^c}\}}\geq2^{-1}n^{-\gamma_1}\big]
&=&
O(n^{-L})
\end{eqnarray}
as $n\to\infty$ for every $L>0$.
In particular,
{\color{black}}% {\color[rgb]{0.8,0,0} if
{\xred the conditions $\kappa_n=O(n^{1/2})$,
$\gamma_2<\frac{1}{2}-\xi$ and $[G3]$ are additionally satisfied}, then}
\begin{eqnarray}\label{0211261348}
P[\Omega_n^c] &=& O(n^{-L})
\end{eqnarray}
as $n\to\infty$ for every $L>0$.
\end{lemma}
\proof
We have
\begin{eqnarray*}
\sup_{{\color{black}}% {\color[rgb]{0,0,0.8} j} \in I_n}\big\|R_j1_{\{{\xred j\in{\cal L}_n^c}\}}\big\|_p
&=&
O(n^{-\gamma_0})
\end{eqnarray*}
for every $p>1$.
The Markov inequality implies (\ref{02112q61347}).
This estimate and Lemma \ref{0211260910} give (\ref{0211261348})
{\xred
if the Markov inequality is used with the estimate $\|\#\>{\cal L}_n\|_p=O(n^{\xi+1/2})$ from $[G3]$.
}
\qed\vspace*{3mm}
{\xred
Lemma 2.7 of Inatsugu and Yoshida {\color{black}}% {\color[rgb]{0,0,0.8}\cite{InatsuguYoshida2020Accepted}}
(or see an extended version arXiv:1806.10706v3)
is rephrased as follows.
Recall that ${\cal L}_n$ is given by (\ref{0301021550})}.
\begin{lemma}
\begin{eqnarray}\label{300217-1}
\widehat{{\cal J}}_n(\alpha)\cap{\xred {\cal L}_n^c}
&\subset&
{\cal J}_n(\alpha)
\end{eqnarray}
on $\Omega_n$.
In particular
\begin{eqnarray}\label{300217-2}
\# \big[{\cal J}_n(\alpha)\ominus\widehat{{\cal J}}_n(\alpha)\big] &\leq&
c_*n^{1-\gamma_2}
+{\xred \#\>{\cal L}_n
}
\end{eqnarray}
on $\Omega_n$, where $c_*$ is a positive constant.
{\color{black}}% {\color[rgb]{0.8,0,0} Here $\ominus$ denotes the symmetric difference operator of sets. }
\end{lemma}
\vspace*{3mm}
For $\gamma_3>0$ and random variables $({\tt U}_j)_{j=1,...,n}$, let
\begin{eqnarray*}
{\cal D}_n &=& n^{\gamma_3}
\Bigg| \frac{1}{n}\sum_{j\in{\cal J}_n(\alpha)}{\tt U}_j-\frac{1}{n}\sum_{j\in\widehat{{\cal J}}_n(\alpha)}{\tt U}_j \Bigg|.
\end{eqnarray*}
{\xred
We refer the reader to Lemmas 2.8 and 2.9 of Inatsugu and Yoshida \cite{InatsuguYoshida2020Accepted} (or see arXiv:1806.10706v3)
for proof of the following two lemmas.
}
\begin{lemma}\label{300217-5}
{\bf (i)} Let $p_1>1$. Then
\begin{eqnarray*}
\|{\cal D}_n\|_p&\leq&
\big(c_*n^{\gamma_3-\gamma_2}+n^{-1+\gamma_3}\|{\xred \#\>{\cal L}_n}\|_{p_1}\big)
\bigg\|\max_{j=1,...,n}\big|{\tt U}_j\big|\bigg\|_{pp_1(p_1-p)^{-1}}
\\&&
+n^{\gamma_3}\bigg\|\max_{j=1,...,n}\big|{\tt U}_j\big|1_{\Omega_n^c}\bigg\|_{p}
\end{eqnarray*}
for $p\in(1,p_1)$.
\begin{description}
\item[(ii)] Let $\gamma_4>0$ and $p_1>1$. Then
\begin{eqnarray*}
\|{\cal D}_n\|_p&\leq&
\big(c_*n^{\gamma_3-\gamma_2}+n^{-1+\gamma_3}\|{\xred \#\>{\cal L}_n}\|_{p_1}\big)
\nonumber\\&&\times
\bigg(n^{\gamma_4}
+n\max_{j=1,...,n}\bigg\|\big|{\tt U}_j\big|1_{\{|{\tt U}_j|>n^{\gamma_4}\}}\bigg\|_{pp_1(p_1-p)^{-1}}
\bigg)
\\&&
+n^{\gamma_3}\bigg\|\max_{j=1,...,n}\big|{\tt U}_j\big|1_{\Omega_n^c}\bigg\|_{p}
\end{eqnarray*}
for $p\in(1,p_1)$.
\end{description}
\end{lemma}
\vspace*{3mm}
Let
\begin{eqnarray*}
\widetilde{{\cal D}}_n &=& n^{\gamma_3}
\bigg|\frac{1}{n}\sum_{j\in\widehat{{\cal J}}_n(\alpha)}{\tt U}_j-\frac{1}{n}\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}{\tt U}_j\bigg|.
\end{eqnarray*}
for a collection of random variables $\{{\tt U}_j\}_{j\in I_n}$ and
\begin{eqnarray}\label{0211280343}
\widetilde{{\cal J}}_n(\alpha)
&=&
\big\{j\in I_n;\> |W_j|\leq c(\alpha)^{1/2}\big\}.
\end{eqnarray}
Let
\begin{eqnarray}\label{0211280310}
\widetilde{\Omega}_n
&=&
\big\{\big| |W|_{({\color{black}}% {\color[rgb]{0.8,0,0}\widehat{a}}_n)}-{\color{black}}% {\color[rgb]{0.8,0,0} c(\alpha)^{1/2}}\big|
< \check{C}\>n^{-\gamma_2}\big\},
\end{eqnarray}
where $\check{C}$ is a positive constant.
See Lemma 4 of Inatsugu and Yoshida {\color{black}}% {\color[rgb]{0,0,0.8}\cite{InatsuguYoshida2020Accepted}} for a proof of the following lemma.
\begin{lemma}\label{0211261213}
Let $\check{C}>0$ and $\gamma_3>0$. Then
\begin{description}
\item[(i)] For $p\geq1$,
\begin{eqnarray*}
\|\widetilde{{\cal D}}_n\|_p
&\leq&
n^{\gamma_3}\bigg\|\max_{j'=1,...,n}|{\tt U}_{j'}|\>
\frac{1}{n}\sum_{j=1}^n
1_{\big\{ \big||W_j|-{\color{black}}% {\color[rgb]{0.8,0,0} c(\alpha)^{1/2}}\big|{\color{black}}% {\color[rgb]{0,0,0.8} <
\check{C}\>n^{-\gamma_2}\big\}}
\bigg\|_p
+n^{\gamma_3}\bigg\|
1_{\widetilde{\Omega}_n^c}
\max_{j'=1,...,n}|{\tt U}_{j'}|\bigg\|_p.
\end{eqnarray*}
\item[(ii)] For $p_1>p\geq1$,
\begin{eqnarray*}
\|\tilde{{\cal D}}_n^{(k)}\|_p
&\leq&
n^{\gamma_3}\bigg\|\max_{j=1,...,n}|{\tt U}_{j}|\>
\bigg\|_p
P\bigg[\big||W_1|-{\color{black}}% {\color[rgb]{0.8,0,0} c(\alpha)^{1/2}}\big|{\color{black}}% {\color[rgb]{0,0,0.8} <
\check{C}\>n^{-\gamma_2}\bigg]
\\&&
+n^{\gamma_3}\bigg\|\max_{j=1,...,n}|{\tt U}_{j}|
\bigg\|_{pp_1(p_1-p)^{-1}}
\nonumber\\&&\hspace{30pt}\times
\bigg\|
\frac{1}{n}\sum_{j=1}^n\bigg(
1_{\big\{ \big|{\color{black}}% {\color[rgb]{0.8,0,0}|W_j|-}{\color{black}}% {\color[rgb]{0.8,0,0} c(\alpha)^{1/2}}\big| {\color{black}}% {\color[rgb]{0,0,0.8} <
\check{C}\>n^{-\gamma_2}\big\}}
-
P\bigg[\big||W_1|-{\color{black}}% {\color[rgb]{0.8,0,0} c(\alpha)^{1/2}}\big|{\color{black}}% {\color[rgb]{0,0,0.8} <
\check{C}\>n^{-\gamma_2}\bigg]
\bigg)
\bigg\|_{p_1}
\\&&
+n^{\gamma_3}
P[\widetilde{\Omega}_n^c]^{1/p_1}
\bigg\|\max_{j=1,...,n}|{\tt U}_{j}|\bigg\|_{pp_1(p_1-p)^{-1}}.
\end{eqnarray*}
\end{description}
\end{lemma}
\vspace*{3mm}
\begin{lemma}\label{021127f1234}
If the constant $\check{C}$ in (\ref{0211241445}) is sufficiently large, then
\begin{eqnarray*}
P\big[\widetilde{\Omega}_n^c\big]
&=&
O(n^{-L})
\end{eqnarray*}
as $n\to\infty$ for any $L>0$.
\end{lemma}
{\color{black}}% {\color[rgb]{0.8,0,0}
Now we shall investigate the rate of convergence of ${\mathbb V}_n(\alpha)$
for a constant $\alpha\in(0,1)$.
We note that, under $[G1]$ and $[G3]$,
\begin{eqnarray}\label{0211261352}
\bigg\|
\sum_{j\i
{\mathfrak L}_n}|\Delta_jX|^2K_{n,j}\bigg\|_p
&\leq&
n^{-1/2}\big\|{\xred\#\>{\mathfrak L}_n}\big\|_p
\>=\>
O(n^{-1/2+\xi}){\color{black}}% {\color[rgb]{0.8,0,0}.}
\end{eqnarray}
Let
\begin{eqnarray*}
\widehat{{\mathbb V}}_n(\alpha)
&=&
\sum_{j\in\widehat{{\cal J}}_n(\alpha)}q(\alpha)^{-1}|\Delta_jX|^2K_{n,j}.
\end{eqnarray*}
\begin{lemma}\label{0211271153}
Suppose that $[G1]$ $[G2]$ and $[G3]$ are fulfilled.
Suppose that $\xi<\frac{1}{2}$.
Let $\gamma_5<\min\big\{\gamma_0,\frac{1}{2}-\xi\big\}$
{\xred and $\kappa_n=O(n^{1/2})$}. Then
\begin{eqnarray*}
\sup_{n\in{\mathbb N}}n^{\gamma_5}\big\|{\mathbb V}_n(\alpha)-\widehat{{\mathbb V}}_n(\alpha)\big\|_p
&<&
\infty.
\end{eqnarray*}
\end{lemma}
\proof
By (\ref{0211261352}), we obtain
\begin{eqnarray*}
\big\|{\mathbb V}_n(\alpha)-\widehat{{\mathbb V}}_n(\alpha)\big\|_p
&=&
\bigg\|\sum_{j\in{\cal J}_n(\alpha)}
q(\alpha)^{-1}|\Delta_jX|^21_{\{\Delta_j\overline{N}=0\}}K_{n,j}
\nonumber\\&&\hspace{30pt}
-\sum_{j\in\widehat{{\cal J}}_n(\alpha)}
q(\alpha)^{-1}|\Delta_jX|^21_{\{\Delta_j\overline{N}=0\}}K_{n,j}
\bigg\|_p
+O(n^{-1/2+\xi}){\color{black}}% {\color[rgb]{0.8,0,0}.}
\end{eqnarray*}
By Lemmas \ref{300217-5} and \ref{0211261336},
\begin{eqnarray*} &&
n^{\gamma_3}\big\|{\mathbb V}_n(\alpha)-\widehat{{\mathbb V}}_n(\alpha)\big\|_p
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
\big(c_*n^{\gamma_3-\gamma_2}+n^{-1+\gamma_3}\|{\xred \#\>{\cal L}_n}\|_{p_1}\big)
\nonumber\\&&\times
\bigg(n^{\gamma_4}
+n\max_{j=1,...,n}\bigg\|n|\Delta_jX|^21_{\{\Delta_j\overline{N}=0\}}K_{n,j}1_{\{n|\Delta_jX|^21_{\{\Delta_j\overline{N}=0\}}K_{n,j}>n^{\gamma_4}\}}\bigg\|_{pp_1(p_1-p)^{-1}}
\bigg)
\\&&
+n^{\gamma_3}\bigg\|\max_{j=1,...,n}\big(n|\Delta_jX|^21_{\{\Delta_j\overline{N}=0\}}K_{n,j}\big)1_{\Omega_n^c}\bigg\|_{p} {\color{black}}% {\color[rgb]{0,0,0.8} +O(n^{-1/2+\gamma_3+\xi})}
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
c_*n^{\gamma_3+\gamma_4-\gamma_2}+n^{{\xred-1/2}+\gamma_3+\gamma_4+\xi}
+n^{-1/2+\gamma_3+\xi}
,
\end{eqnarray*}
where $1<p<p_1$.
The parameters should satisfy
\begin{eqnarray*}
0<\gamma_3<\gamma_2<\gamma_1<\gamma_0<\frac{1}{2},
\quad
\gamma_2<\frac{1}{2}-\xi,
\quad
\gamma_4>0.
\end{eqnarray*}
We make
\begin{eqnarray*}
\gamma_4\down0,\qquad
{\color{black}}% {\color[rgb]{0,0,0.8} \gamma_5 < }\gamma_3<\uparrow\gamma_2<\uparrow\gamma_1<\uparrow\min\bigg\{\gamma_0,\frac{1}{2}-\xi\bigg\}
\end{eqnarray*}
to obtain the desired exponent.
\qed\vspace*{3mm}
For $\widetilde{{\cal J}}_n(\alpha)$ defined in (\ref{0211280343}), let
\begin{eqnarray*}
\widetilde{{\mathbb V}}_n(\alpha)
&=&
\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}q(\alpha)^{-1}|\Delta_jX|^2K_{n,j}.
\end{eqnarray*}
\begin{lemma}\label{0211271219}
Suppose that $[G1]$ and $[G3]$ are fulfilled.
Suppose that $\xi<\frac{1}{2}$.
Let
$\gamma_6<\frac{1}{2}-\xi$.
Then
\begin{eqnarray*}
\sup_{n\in{\mathbb N}}
n^{\gamma_6}\big\|\widehat{{\mathbb V}}_n(\alpha)-\widetilde{{\mathbb V}}_n(\alpha)\big\|_p
&<&
\infty{\color{black}}% {\color[rgb]{0,0,0.8} .}
\end{eqnarray*}
\end{lemma}
\proof
By (\ref{0211261352}), we obtain
\begin{eqnarray*}
\big\|\widehat{{\mathbb V}}_n(\alpha)-\widetilde{{\mathbb V}}_n(\alpha)\big\|_p
&=&
\bigg\|\sum_{j\in\widehat{{\cal J}}_n(\alpha)}
q(\alpha)^{-1}|\Delta_jX|^21_{\{\Delta_j\overline{N}=0\}}K_{n,j}
\nonumber\\&&\hspace{30pt}
-\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}
q(\alpha)^{-1}|\Delta_jX|^21_{\{\Delta_j\overline{N}=0\}}K_{n,j}
\bigg\|_p
+O(n^{-1/2+\xi}){\color{black}}% {\color[rgb]{0,0,0.8} .}
\end{eqnarray*}
By Lemma \ref{0211261213}, we obtain
\begin{eqnarray*} &&
n^{\gamma_3}\big\|\widehat{{\mathbb V}}_n(\alpha)-\widetilde{{\mathbb V}}_n(\alpha)\big\|_p
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{\gamma_3}\bigg\|\max_{j=1,...,n}\big(n|\Delta_jX|^21_{\{\Delta_j\overline{N}=0\}}K_{n,j}\big)\>
\bigg\|_p
P\bigg[\big||W_1|-{\color{black}}% {\color[rgb]{0.8,0,0} c(\alpha)^{1/2}}\big|{\color{black}}% {\color[rgb]{0,0,0.8} <
\check{C}\>n^{-\gamma_2}\bigg]
\\&&
+n^{\gamma_3}\bigg\|\max_{j=1,...,n}\big(n|\Delta_jX|^21_{\{\Delta_j\overline{N}=0\}}K_{n,j}\big)
\bigg\|_{pp_1(p_1-p)^{-1}}
\nonumber\\&&\hspace{30pt}\times
\bigg\|
\frac{1}{n}\sum_{j=1}^n\bigg(
1_{\big\{ \big|{\color{black}}% {\color[rgb]{0.8,0,0} |W_j|-c(\alpha)^{1/2}}\big|{\color{black}}% {\color[rgb]{0,0,0.8} <
\check{C}\>n^{-\gamma_2}\big\}}
-
P\bigg[\big||W_1|-{\color{black}}% {\color[rgb]{0.8,0,0} c(\alpha)^{1/2}}\big|{\color{black}}% {\color[rgb]{0,0,0.8} <
\check{C}\>n^{-\gamma_2}\bigg]
\bigg)
\bigg\|_{p_1}
\\&&
+n^{\gamma_3}
P[\widetilde{\Omega}_n^c]^{1/p_1}
\bigg\|\max_{j=1,...,n}\big(n|\Delta_jX|^21_{\{\Delta_j\overline{N}=0\}}K_{n,j}\big)\bigg\|_{pp_1(p_1-p)^{-1}}
+O(n^{\gamma_3-1/2+\xi})
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{\gamma_3-\gamma_2}+n^{\gamma_3-\frac{1}{2}-\frac{\gamma_2}{2}}
{\color{black}}% {\color[rgb]{0.8,0,0}+n^{\gamma_3-1/2+\xi}}
\>\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ \>
n^{\gamma_3{\color{black}}% {\color[rgb]{0.8,0,0}+\gamma_4}-\gamma_2}+n^{\gamma_3{\color{black}}% {\color[rgb]{0.8,0,0}+\gamma_4}-1/2+\xi},
\end{eqnarray*}
where $1\leq p<p_1$ and {\color{black}}% {\color[rgb]{0.8,0,0}$\gamma_4$ is {\color{black}}% {\color[rgb]{0,0,0.8} an} arbitrary positive number}.
{\color{black}}% {\color[rgb]{0.8,0,0} Lemma \ref{021127f1234} was used in the above derivation.}
Making
\begin{eqnarray*}
{\color{black}}% {\color[rgb]{0.8,0,0}
\gamma_4\down0\quad\text{and}\quad}
\gamma_6<
\gamma_3<\uparrow\gamma_2<\u
\frac{1}{2}-\xi,
\end{eqnarray*}
we conclude the proof.
\qed\vspace*{3mm}
\begin{lemma}\label{0211280809}
Suppose that ${\color{black}}% {\color[rgb]{0.8,0,0}[G1]}$ and $[G3]$ are satisfied. Suppose that $\xi<1/2$.
Then
\begin{eqnarray*}
\bigg\|\widetilde{{\mathbb V}}_n(\alpha)
-\int_0^T \sigma_t^2dt\bigg\|_p
&=&
O\big(n^{-\frac{1}{2}+\xi}\big)
\end{eqnarray*}
as $n\to\infty$ for every $p>1$.
\end{lemma}
\proof
Recall that ${\color{black}}% {\color[rgb]{0.8,0,0}\mathfrak L_n}$ is defined by (\ref{0211280352}).
We have
\begin{eqnarray}\label{0211280728}
\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}|\Delta_jX|^2K_{n,j}
1_{\{j\in{\mathfrak L}_n^c\}}
&=&
\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}
\left(\int_{t_{j-1}}^{t_j} \sigma_tdw_t+\int_{t_{j-1}}^{t_j} b_tdt\right)^2K_{n,j}1_{\{j\in{\mathfrak L}_n^c\}}
\nonumber\\&=&
\Phi^{(\ref{0211280401})}_n+\Phi^{(\ref{0211280402})}_n+
\Phi^{(\ref{0211280403})}_n
\end{eqnarray}
where
\begin{eqnarray}\label{0211280401}
\Phi^{(\ref{0211280401})}_n
&=&
\sum_{j\in I_n }\sigma_{t_{j-1}}^2
hW_j^21_{\{|W_j|\leq c(\alpha)^{1/2}\}},
\end{eqnarray}
\begin{eqnarray}\label{0211280402}
\Phi^{(\ref{0211280402})}_n
&=&
\sum_{j\in I_n}\sigma_{t_{j-1}}^2
hW_j^21_{\{|W_j|\leq c(\alpha_0)^{1/2}\}}
\big(K_{n,j}1_{\{j\in{\mathfrak L}_n^c\}}
-1\big)
\nonumber\\&&
+2\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}
\int_{t_{j-1}}^{t_j} \int_{t_{j-1}}^t\big(\widetilde{\sigma}_s-\widetilde{\sigma}_{t_{j-1}}\big)dw_s\sigma_tdw_tK_{n,j}
1_{\{j\in{\mathfrak L}_n^c\}}
\nonumber\\&&
+2\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}
\int_{t_{j-1}}^{t_j} \int_{t_{j-1}}^t\sigma_{t_{j-1}} dw_s\big(\widetilde{\sigma}_t-\widetilde{\sigma}_{t_{j-1}}\big) dw_tK_{n,j}
1_{\{j\in{\mathfrak L}_n^c\}}
\nonumber\\&&
{\color{black}}% {{\color[rgb]{0.8,0,0}
+2\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}
\int_{t_{j-1}}^{t_j} \widetilde{\sigma}_{t_{j-1}}
\big(\widetilde{\sigma}_t-\widetilde{\sigma}_{t_{j-1}}\big)dtK_{n,j}1_{\{j\in{\mathfrak L}_n^c\}}
}
\nonumber\\&&
{\color{black}}% {{\color[rgb]{0.8,0,0}
+\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}
\bigg(\int_{t_{j-1}}^{t_j}
\big(\widetilde{\sigma}_t-\widetilde{\sigma}_{t_{j-1}}\big)dw_t\bigg)^2K_{n,j}1_{\{j\in{\mathfrak L}_n^c\}}
}
\end{eqnarray}
and
\begin{eqnarray}\label{0211280403}
\Phi^{(\ref{0211280403})}_n
&=&
2\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}
\int_{t_{j-1}}^{t_j} \int_{t_{j-1}}^tb_sds\sigma_tdw_tK_{n,j}1_{\{j\in{\mathfrak L}_n^c\}}
\nonumber\\&&+
2\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}
\int_{t_{j-1}}^{t_j} \int_{t_{j-1}}^t\sigma_sdw_sb_tdtK_{n,j}1_{\{j\in{\mathfrak L}_n^c\}}
\nonumber\\&&+
2\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}
\int_{t_{j-1}}^{t_j} \int_{t_{j-1}}^tb_sdsb_tdtK_{n,j}1_{\{j\in{\mathfrak L}_n^c\}}.
\end{eqnarray}
Let $\epsilon>\xi$.
For $p>1$ and $\epsilon'>0$,
\begin{eqnarray}\label{0211280645} &&
\bigg\|
\sum_{j\in I_n}\sigma_{t_{j-1}}^2
hW_j^21_{\{|W_j|\leq c({\color{black}}% {\color[rgb]{0.8,0,0}\alpha})^{1/2}\}}
\big(K_{n,j}1_{\{j\in{\mathfrak L}_n^c\}}
-1\big)\bigg\|_p
\nonumber\\&\leq&
\bigg\|
\sum_{j\in I_n}\sigma_{t_{j-1}}^2
hW_j^21_{\{|W_j|\leq c({\color{black}}% {\color[rgb]{0.8,0,0}\alpha})^{1/2}\}
{\xred1_{\{j\in{\mathfrak L}_n\}}}
\bigg\|_p
\\&&
+\bigg\|
\sum_{j\in I_n}\sigma_{t_{j-1}}^2
hW_j^21_{\{|W_j|\leq c({\color{black}}% {\color[rgb]{0.8,0,0}\alpha})^{1/2}\}}
\big(K_{n,j
-1\big){\xred1_{\{j\in{\mathfrak L}_n^c\}}}\bigg\|_p
\nonumber\\&\leq&
\bigg\|
\max_{j\in I_n}\bigg(\sigma_{t_{j-1}}^2hW_j^21_{\{|W_j|\leq c({\color{black}}% {\color[rgb]{0.8,0,0}\alpha})^{1/2}\}}K_{n,j}\bigg)
{\>\xred\#{\mathfrak L}_n
\bigg\|_p
+O(n^{-L})
\nonumber\\&\leq&
\big\|{\xred\#{\mathfrak L}_n}1_{\{{\xred\#{\mathfrak L}_n}>n^{\epsilon}\}}\big\|_{2p}
\bigg\|
\max_{j\in I_n}\bigg(\sigma_{t_{j-1}}^2hW_j^21_{\{|W_j|\leq c({\color{black}}% {\color[rgb]{0.8,0,0}\alpha})^{1/2}\}}K_{n,j}\bigg)
\bigg\|_{2p}
\nonumber\\&&
+
n^{\epsilon}
\bigg\|
\max_{j\in I_n}\bigg(\sigma_{t_{j-1}}^2hW_j^21_{\{|W_j|\leq c({\color{black}}% {\color[rgb]{0.8,0,0}\alpha})^{1/2}\}}K_{n,j}\bigg)
\bigg\|_{p}
+O(n^{-L})
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{-\frac{L\epsilon}{2p}}\big\|{\xred\#{\mathfrak L}_n}\big\|_{2p+L}^{\frac{2p+L}{2p}}\times n^{-1+\epsilon'}
+n^{-1+\epsilon+\epsilon'}
+O(n^{-L})
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{-\frac{L(\epsilon-\xi)}{2p}+\xi-1+\epsilon'}+n^{-1+\epsilon+\epsilon'}+O(n^{-L})
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{-1+\epsilon+\epsilon'}
\nonumber
\end{eqnarray}
since $\epsilon>\xi$, where $L$ is a sufficiently large number
chosen suitably depending on $(\epsilon,\xi,p,\epsilon')$.
From the estimate (\ref{0211280645}), we have
\begin{eqnarray}\label{0211280407}
\big\|\Phi^{(\ref{0211280402})}_n
\big\|_p
&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
h^{1/2}+h^{1-\epsilon-\epsilon'}
\>\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ \>
h^{1/2}
\end{eqnarray}
if letting $\epsilon\downarrow\xi<1/2$ and $\epsilon'\down0$.
{\color{black}}% {\color[rgb]{0,0,0.8}
By the Burkholder-Davis-Gundy inequality, we have
\begin{align*}
\Bigg\| \sum_{j \in \widetilde{{\cal J}}_n(\alpha)}
\int_{{t_{j-1}}}^{t_j} \int _{{t_{j-1}}}^{t} b_s ds \sigma_t dw_t K_{n, j} 1_{\{j \in \mathfrak{L}_n^c \}} \Bigg\|_p
&\leq
\sum_{j \in I_n} \Bigg\|
\int_{{t_{j-1}}}^{t_j} \int _{{t_{j-1}}}^{t} b_s ds \sigma_t dw_t \Bigg\|_p \\
&\lesssim
\sum_{j \in I_n} \sqrt{
\Bigg\|
\int_{{t_{j-1}}}^{t_j} \Bigg( \int _{{t_{j-1}}}^{t} b_s ds \sigma_t \Bigg)^2 dt \Bigg\|_{p/2}} \\
&\leq
\sum_{j \in I_n} \sqrt{
\int_{{t_{j-1}}}^{t_j} \Bigg\|
\int _{{t_{j-1}}}^{t} b_s ds \sigma_t \Bigg\|_p^2 dt }\\
&\lesssim
h^{1/2}.
\end{align*}
From this and similar estimates, we have
}
\begin{eqnarray}\label{0211280408}
\big\|\Phi^{(\ref{0211280403})}_
\big\|_p
&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
h^{1/2}
\end{eqnarray}
as $n\to\infty$ for every $p>1$.
Moreover,
\begin{eqnarray}\label{0211280411}
\bigg\|
\Phi^{(\ref{0211280401})}_j
-\sum_{j\in I_n}\sigma_{t_{j-1}}^2 q({\color{black}}% {\color[rgb]{0.8,0,0}\alpha})h
\bigg\|_p
&\leq&
\bigg\|h
\sum_{j\in I_n}\sigma_{t_{j-1}}^2
\big(W_j^21_{\{|W_j|\leq c({\color{black}}% {\color[rgb]{0.8,0,0}\alpha})^{1/2}\}}-q({\color{black}}% {\color[rgb]{0.8,0,0}\alpha})\big)
\bigg\|_p
\nonumber\\&=&
O(h^{1/2})
\end{eqnarray}
for every $p>1$.
Obviously,
\begin{eqnarray}\label{0211280406}
\sup_{j\in I_n}
\big\|1_{\{j\in{\mathfrak L}_n^c\}}\big(\sigma_{t_k}^2-\sigma_{t_{j-1}}^2\big)
\big\|_p
&\>\leq\> &\sup_{j\in I_n}\big\|\widetilde{\sigma}_{t_k}^2-\widetilde{\sigma}_{t_{j-1}}^2\big\|_p
\>\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ \>
h^{1/2}
\end{eqnarray}
for every $p>1$.
In view of (\ref{0211280406}), we deduce that
\begin{eqnarray}\label{0211280745}&&
\bigg\|
\sum_{j\in I_n}\sigma_{t_{j-1}}^2 h
-\int_0^T \sigma_t^2dt
\bigg\|_p
\nonumber\\&\leq&
\bigg\|
\sum_{j\in I_n}\int_{t_{j-1}}^{t_j} \big|\widetilde{\sigma}_t^2
-\widetilde{\sigma}_{t_{j-1}}^2 \big| d
\bigg\|_p
+
\bigg\|
\sum_{j\in I_n}\int_{t_{j-1}}^{t_j} \big(\sigma_t^2
-\sigma_{t_{j-1}}^2 \big) dt1_{\{j\in{\mathfrak L}_n\}}
\bigg\|_p
\nonumber\\&\leq&
O(h^{1/2})
+
\bigg\|
\max_{j\in I_n}\bigg\{\int_{t_{j-1}}^{t_j} \big(|\sigma_t^2|
+|\sigma_{t_{j-1}}^2| \big) dt\bigg\}\>
{\xred\#{\mathfrak L}_n
\bigg\|_p
\nonumber\\&=&
O(h^{1/2}),
\end{eqnarray}
following the passage from (\ref{0211280645}) to (\ref{0211280407}).
Easily,
\begin{eqnarray}\label{0211280722}
\bigg\|
\sum_{j\in\widetilde{{\cal J}}_n(\alpha)}|\Delta_jX|^2K_{n,j}
1_{\{j\in{\mathfrak L}_n\}}
\bigg\|_p
&\leq&
\big\|n^{-1/2}{\xred\#{\mathfrak L}_n
\big\|_p
\>\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ \>
n^{-\frac{1}{2}+\xi}.
\end{eqnarray}
Combining (\ref{0211280722}), (\ref{0211280728}),
(\ref{0211280407}), (\ref{0211280408})
(\ref{0211280411}) and (\ref{0211280745}),
we obtain
\begin{eqnarray*}
\bigg\|\widetilde{{\mathbb V}}_n(\alpha)
-\int_0^T \sigma_t^2dt\bigg\|_p
&=&
O\big(n^{-\frac{1}{2}+\xi}\big)
\end{eqnarray*}
as $n\to\infty$ for every $p>1$.
\qed\vspace*{3mm}
\begin{theorem}\label{0211280814}
Suppose that $[G1]$ $[G2]$ and $[G3]$ are fulfilled.
Suppose that $\xi<\frac{1}{2}$ {\xred and $\kappa_n=O(n^{1/2})$}.
Let $\alpha\in(0,1)$ and $\beta_0<\min\big\{\gamma_0,\frac{1}{2}-\xi\big\}$. Then
\begin{eqnarray*}
\big\| {\mathbb V}_n(\alpha)-\Theta \big\|_p
&=&
O(n^{-\beta_0})
\end{eqnarray*}
as $n\to\infty$ for every $p>1$.
\end{theorem}
\proof
Use Lemmas \ref{0211271153}, \ref{0211271219} and \ref{0211280809}.
\qed\vspace*{3mm}
\subsection{Rate of convergence of the WGRV with a fixed $\alpha$}
Next, we discuss the convergence of the WGRV with a fixed $\alpha$.
Recall that the WGRV is defined as
$$
{\mathbb W}_n(\alpha) = \sum_{j \in I_n} {\sf w}(\alpha)^{-1}
\big\{ |\Delta_j X| \wedge (S_{n,j-1}^{1/2} V_{(s_n(\alpha))}) \big\}^2 K_{n,j}.
$$
The WGRV has entirely the same rate of convergence as the GRV.
\begin{theorem}
Suppose that {\color{black}}% {\color[rgb]{0,0,0.8} $[G1]$}, {\color{black}}% {\color[rgb]{0,0,0.8} $[G2]$}, and {\color{black}}% {\color[rgb]{0,0,0.8} $[G3]$} are fulfilled. Suppose that
$\xi < \frac{1}{2}$. Let $\alpha \in (0, 1)$ and
$\beta_0 {\xred<} \min \big\{ \gamma_0, \frac{1}{2} - \xi \big\}$. Moreover, assume that $\kappa_n{\xred =O( n^{1/2})}$. Then
$$
\left\| {\mathbb W}_n(\alpha) - \Theta \right\|_p = O(n^{-\beta_0})
$$
as $n \to \infty$ for every $p > 1$.
\end{theorem}
\proof Decompose {\color{black}}% {\color[rgb]{0,0,0.8} ${\mathbb W}_n(\alpha)$} as
\begin{align*}
{\color{black}}% {\color[rgb]{0,0,0.8} {\mathbb W}_n(\alpha)}
&=
\sum_{j \in {\cal J}_n(\alpha)} {\sf w}(\alpha)^{-1} |\Delta_j X|^2 K_{n,j}
+ \sum_{j \in {\cal J}_n(\alpha)^c} {\sf w}(\alpha)^{-1} S_{n,j-1} V_{(s_n(\alpha))}^2 K_{n,j}\\
&=
\frac{q(\alpha)}{{\sf w}(\alpha)} {{\mathbb V}}_n(\alpha)
+ \sum_{j \in {\cal J}_n(\alpha)^c} {\sf w}(\alpha)^{-1} S_{n,j-1} V_{(s_n(\alpha))}^2 K_{n,j}.
\end{align*}
Note that ${\sf w}(\alpha) = q(\alpha) + \alpha c(\alpha)$.
Hence, it suffices to show that
$$
\left \| \sum_{j \in {\cal J}_n(\alpha)^c} S_{n,j-1} V_{(s_n(\alpha))}^2 K_{n,j}
- \alpha c(\alpha) \Theta \right\|_p = O(n^{-\beta_0})
$$
{\color{black}}% {\color[rgb]{0,0,0.8} as $n \to \infty$ for every $p>1$.}
Decompose the left-hand side as
\begin{align*}
\sum_{j \in {\cal J}_n(\alpha)^c} S_{n,j-1} V_{(s_n(\alpha))}^2 K_{n,j}
- \alpha c(\alpha) \Theta
&=
\sum_{j \in {\cal J}_n(\alpha)^c} S_{n,j-1} V_{(s_n(\alpha))}^2 K_{n,j}
1_{\{j \in {\cal L}_n^c \}} - \alpha c(\alpha) \Theta \\
&\qquad + \sum_{j \in {\cal J}_n(\alpha)^c} S_{n,j-1} V_{(s_n(\alpha))}^2 K_{n,j} 1_{\{j \in \mathfrak{L}_n \cap {\cal L}_n\}} \\
&\qquad + \sum_{j \in {\cal J}_n(\alpha)^c} S_{n,j-1} V_{(s_n(\alpha))}^2 K_{n,j} 1_{\{j \in \mathfrak{L}_n^c \cap {\cal L}_n\}} \\
&{\color{black}}% {\color[rgb]{0,0,0.8}=:} A_1 + A_2 + A_3.
\end{align*}
Since $S_{n,j-1} V_{(s_n(\alpha))}^2 K_{n,j} \leq |\Delta_j X|^2 K_{n,j} \leq
n^{-1/2}$ for $j \in {\cal J}_n(\alpha)^c$, we have
$\big\| A_2 \big\|_p \lesssim n^{-1/2+\xi}$.
As for $A_3$, note that $\# {\cal L}_n \lesssim n^{\xi} \times \bar{\kappa}_n
{\xred=O( n^{\xi + 1/2})}$
and that $\Delta_j X = \Delta_j \tilde{X}$ for $j \in \mathfrak{L}_n^c$.
Hence we have
$$
\| A_3 \|_p \leq
\left \| \max_{j \in I_n} | \Delta_j \tilde{X}|^2 \# (\mathfrak{L}_n^c \cap {\cal L}_n) \right \|_p
\lesssim n^{-1/2 + \xi + \epsilon}
$$
where $\epsilon$ is an arbitrarily small positive number.
As for $A_1$, we can set $c = 1$ in the condition {\color{black}}% {\color[rgb]{0,0,0.8} $[G2]$}(ii) without loss of generality.
\begin{align*}
\left\| A_1 - \alpha c(\alpha) \Theta \right\|_p
&\leq
\left\| \Big( {\xred h^{-1}} V_{(s_n(\alpha))}^2 - c(\alpha) \Big)
{\xred h} \sum_{j \in {\cal J}_n(\alpha)^c} S_{n,j-1} K_{n,j} 1_{\{ j \in {\cal L}_n^c \}} \right\|_p \\
&\qquad +
c(\alpha) \left\| {\xred h} \sum_{j \in {\cal J}_n(\alpha)^c} \big( S_{n,j-1} - \sigma_{t_{j-1}}^2 \big) 1_{\{ j \in {\cal L}_n^c \}} \right \|_p\\
&\qquad +
c(\alpha) \left\| {\xred h} \sum_{j \in {\cal J}_n(\alpha)^c} S_{n,j-1} \big( 1 - K_{n,j} \big) 1_{\{ j \in {\cal L}_n^c \}} \right \|_p\\
&\qquad +
c(\alpha) \left\| {\xred h}\sum_{j \in {\cal J}_n(\alpha)^c} \sigma_{t_{j-1}}^2 1_{\{ j \in {\cal L}_n^c \}}
- \alpha \Theta \right\|_p \\
&{\color{black}}% {\color[rgb]{0,0,0.8}=:}
B_1 + B_2 + B_3 + B_4.
\end{align*}
By condition {\color{black}}% {\color[rgb]{0,0,0.8} $[G2]$}, $B_2 = O(n^{-\gamma_0})$.
As for $B_3$, with the estimate $\big\| 1_{\{j \in {\cal L}_n^c\}}(1 - K_{n,j}) \big\|_p \leq P[|\Delta_j \tilde{X}| > n^{-1/4}]^{1/p} = O(n^{-L})$ (for all $p > 1$ and $L >0$) and the Cauchy-Schwarz inequality, we have
$$
\| B_3 \|_p \leq
{\xred h} \sum_{j \in I_n} \big\| 1_{\{j \in {\cal L}_n^c\}} S_{n,j-1} \big\|_{2p}
\big\| 1_{\{j \in {\cal L}_n^c\}} (1 - K_{n,j}) \big\|_{2p} = O(n^{-L}).
$$
For $B_4$, we use the following decomposition:
\begin{eqnarray*}&&
{\xred h}\sum_{j \in {\cal J}_n(\alpha)^c} \sigma_{{t_{j-1}}}^2 {\xred 1_{\{ j \in {\cal L}_n^c \}}}
- \alpha \Theta
\\&= &
\left( {\xred h} \sum_{j \in I_n} \sigma_{{t_{j-1}}}^2 - \Theta \right)
- {\xred h} \sum_{j \in I_n} \sigma_{{t_{j-1}}}^2 1_{\{ j \in {\cal L}_n\}}
+ {\xred h} \sum_{j \in {\cal J}_n(\alpha)} \sigma_{{t_{j-1}}}^2 1_{\{ j \in {\cal L}_n\}} \\
\\&&\qquad +
\left( (1-\alpha) \Theta - {\xred h} \sum_{j \in \widetilde{{\cal J}}_n(\alpha)} \sigma_{{t_{j-1}}}^2 \right) +
\left( {\xred h} \sum_{j \in \widetilde{{\cal J}}_n(\alpha)} \sigma_{{t_{j-1}}}^2
- {\xred h} \sum_{j \in {\cal J}_n(\alpha)} \sigma_{{t_{j-1}}}^2 \right).
\end{eqnarray*}
Hence, with the aid of {\color{black}}% {\color[rgb]{0,0,0.8} Lemmas} \ref{300217-5}, \ref{0211261213}
and the estimate
{\xred $\|{\cal L}_n\|_p\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ n^{\xi+1/2}$,}
we have
\begin{eqnarray}\label{Ineq001}&&
\Bigg\| {\xred h} \sum_{j \in {\cal J}_n(\alpha)^c} \sigma_{{t_{j-1}}}^2
{\xred 1_{\{ j \in {\cal L}_n^c \}}}
- \alpha \Theta \Bigg\|_p
\nonumber\\&\lesssim&
\Bigg\| {\xred h} \sum_{j \in I_n} \sigma_{{t_{j-1}}}^2 - \Theta \Bigg\|_p
+
\Bigg\| (1-\alpha) \Theta - {\xred h} \sum_{j \in \widetilde{{\cal J}}_n(\alpha)} \sigma_{{t_{j-1}}}^2 \Bigg\|_p + {\xred O(n^{-\beta_0})
\end{eqnarray}
{\xred since $\beta_0<\frac{1}{2}-\xi$.}
{\color{black}}% {\color[rgb]{0,0,0.8}
The first term of the right-hand side of the above inequality is $O(n^{-1/2})$
by (\ref{0211280745}). }
As for the second term {\xred on the right-hand side} of (\ref{Ineq001}),
\begin{align*}
\Bigg\| (1-\alpha) \Theta - {\xred h}\sum_{j \in \widetilde{{\cal J}}_n(\alpha)} \sigma_{{t_{j-1}}}^2 \Bigg\|_p
&=
\Bigg\| (1-\alpha) \Theta - {\xred h} \sum_{j \in I_n} \sigma_{{t_{j-1}}}^2
1_{\{ |W_j| \leq c(\alpha)^{1/2} \}} \Bigg\|_p \\
&\leq
\Bigg\| {\xred h} \sum_{j \in I_n} {\xred\sigma_{{t_{j-1}}}^2 }
\Big( 1_{\{ |W_j| \leq c(\alpha)^{1/2}\}} -
P\Big[ |W_j| \leq c(\alpha)^{1/2} \Big] \Big) \Bigg\|_p \\
&\qquad
+ (1-\alpha) \Bigg\| {\xred h} \sum_{j \in I_n} \sigma_{t_{j-1}}^2 - \Theta \Bigg\|_p \\
&= {\color{black}}% {\color[rgb]{0,0,0.8} O(n^{-1/2}) }.
\end{align*}
since
Hence we have ${\color{black}}% {\color[rgb]{0,0,0.8} B_4} = O(n^{-\frac{1}{2} + \xi})$.
{\xred
Finally, for $B_1$, it suffices to show that
\begin{eqnarray}\label{0302141014}
P \left[ \big| {\xred h^{-1}} V_{(s_n(\alpha))} - c(\alpha)^{1/2} \big| > n^{-\beta_0} \right]
&=&
O(n^{-L})
\end{eqnarray}
as $n\to\infty$ for every $L>0$ and
for every $\beta_0 < \min \{ \gamma_0, \frac{1}{2} - \xi \}$.
Let
\begin{eqnarray*}
A_{n,j} &=& \big\{ h^{-1/2} V_j < c(\alpha)^{1/2} - n^{-\beta_0} \big\}
\\
{\cal V}_{n,j}
&=&
1_{\big\{|W_j|\leq c(\alpha)^{1/2}-n^{-\beta_0}
+2^{-1}n^{-\gamma_1}
\big\}}
\end{eqnarray*}
and
\begin{eqnarray*}
\mu_n
&=&
(1-\alpha)n - 1-n^{\frac{1}{2}+\xi+\epsilon} -nE[{\cal V}_{n,j}]
\end{eqnarray*}
for $\epsilon>0$.
Then
\begin{eqnarray*}&&
P \left[ h^{-1/2} V_{(s_n(\alpha))} - c(\alpha)^{1/2} < - n^{-\beta_0} \right]
\\&\leq &
P \left[ \sum_{j \in I_n} 1_{A_{n,j}} \geq (1-\alpha)n - 1 \right]
\\&=&
P \left[ \sum_{j \in I_n} 1_{A_{n,j}\cap\{j\in{\cal L}_n^c\}}
+\sum_{j \in I_n} 1_{A_{n,j}\cap\{j\in{\cal L}_n\}} \geq (1-\alpha)n - 1 \right]
\\&\leq&
P \left[ \sum_{j \in I_n} 1_{A_{n,j}\cap\{j\in{\cal L}_n^c\}}
+\#{\cal L}_n \geq (1-\alpha)n - 1 \right]
\\&\leq&
P \left[ \sum_{j \in I_n} {\cal V}_{n,j}\geq (1-\alpha)n - 1-n^{\frac{1}{2}+\xi+\epsilon} \right]
+P[\#{\cal L}_n >n^{\frac{1}{2}+\xi+\epsilon}]
+P[\Omega_n^c]
\\&\leq&
P \left[ \sum_{j \in I_n} \big({\cal V}_{n,j}-E[{\cal V}_{n,j}]\big)
\geq \mu_n \right]
+P[\#{\cal L}_n >n^{\frac{1}{2}+\xi+\epsilon}]
+P[\Omega_n^c].
\end{eqnarray*}
We see
\begin{eqnarray*}
\mu_n
&\sim&
(1-\alpha)n - 1-n^{\frac{1}{2}+\xi+\epsilon}
-n\big\{(1-\alpha)-c^*(n^{-\beta_0}-2^{-1}n^{-\gamma_1}\big)\big\}
\>\geq\
\frac{1}{2} c^*n^{1-\beta_0}
\end{eqnarray*}
for large $n$, where $c^*$ is some positive constant,
if we take a sufficiently small $\epsilon$ and $\gamma_1\in(\beta_0,\gamma_0)$ thanks to
$\beta_0 {\xred<} \min \big\{ \gamma_0, \frac{1}{2} - \xi \big\}$.
Since $n^{-1/2}\mu_n\geq 2^{-1}c^*n^{\frac{1}{2}-\beta_0}$, from $\beta_0<1/2$, we obtain
\begin{eqnarray*}
P \left[ n^{-1/2}\sum_{j \in I_n} \big({\cal V}_{n,j}-E[{\cal V}_{n,j}]\big)
\geq n^{-1/2}\mu_n \right]
&=&
O(n^{-L})
\end{eqnarray*}
for every $L>0$. Therefore,
\begin{eqnarray*}
P \left[ h^{-1/2} V_{(s_n(\alpha))} - c(\alpha)^{1/2} < - n^{-\beta_0} \right]
&=&
O(n^{-L})
\end{eqnarray*}
as $n\to\infty$ for every $L>0$.
Similarly, we can obtain the estimate
$P \left[ h^{-1/2} V_{(s_n(\alpha))} - c(\alpha)^{1/2} > n^{-\beta_0} \right] =O(n^{-L})$
to show (\ref{0302141014}),
which concludes the proof.
}
\qed\vspace*{3mm}
\begin{en-text}
\noindent
Let $A_j = \Big\{ {\color{black}}% {\color[rgb]{0,0,0.8} n^{1/2}} V_j < c(\alpha)^{1/2} - n^{-\beta_0} \Big\}$.
Then we have
\begin{align*}
P \left[ n^{1/2} V_{(s_n(\alpha))} - c(\alpha)^{1/2} < {\color{black}}% {\color[rgb]{0,0,0.8} -} n^{-\beta_0} \right]
&\leq
P \left[ \sum_{j \in I_n} 1_{A_j} \geq (1-\alpha)n - 1 \right] \\
&=
P \left[ n^{-1/2} \sum_{j \in I_n} \Big( 1_{A_j} - P[A_j] \Big) \geq C_n \right],
\end{align*}
where
$$
C_n = n^{-1/2} \left( (1-\alpha)n - 1 - \sum_{j \in I_n} P[A_j] \right).
$$
\noindent
We can estimate $\sum_{j \in I_n} P[A_j] $ as follows:
\begin{align*}
\sum_{j \in I_n} P[A_j]
&\leq
\sum_{j \in I_n} P \Big[ n^{1/2} \tilde{V}_j < c(\alpha)^{1/2} - n^{-\beta_0}, \ j \in {\cal L}_n^c \Big] +E[ \# {\cal L}_n]\\
&\lesssim
\sum_{j \in I_n} P \Big[ n^{1/2} \tilde{V}_j < c(\alpha)^{1/2} - n^{-\beta_0}, \ j \in {\cal L}_n^c \Big] + n^{\xi + \frac{1}{2}},
\end{align*}
where $\tilde{V}_j = S_{n,j-1}^{-1/2} |\Delta_j \tilde{X}|$.
The probability on the right-hand side can be estimated as follows:
\begin{align*}
&P \Big[ n^{1/2} \tilde{V}_j < c(\alpha)^{1/2} - n^{-\beta_0}, \ j \in {\cal L}_n^c \Big] \\
&\leq
P \Bigg[ {\color{black}}% {\color[rgb]{0,0,0.8} |W_j|} < c(\alpha)^{1/2} - n^{-\beta_0}
+ n^{1/2} S_{n, j-1}^{-1/2} \bigg| \int_{t_{j-1}}^{t_j} b_t dt \bigg|
+ n^{1/2} S_{n, j-1}^{-1/2} \bigg| \int_{t_{j-1}}^{t_j} (\tilde{\sigma}_t - \tilde{\sigma}_{t_{j-1}}) dw_t \bigg| \\
&\qquad \qquad \qquad
+ {\color{black}}% {\color[rgb]{0,0,0.8} |W_j|} {\color{black}}% {\color[rgb]{0,0,0.8} S_{n, j-1}^{-1/2}} \Big| {\color{black}}% {\color[rgb]{0,0,0.8} S_{n, j-1}^{1/2}} - \sigma_{t_{j-1}} \Big|,
\ j \in {\cal L}_n^c \Bigg] \\
&\leq
P \Bigg[ {\color{black}}% {\color[rgb]{0,0,0.8} |W_j|}< c(\alpha)^{1/2} - \frac{n^{-\beta_0}}{2} \Bigg] +
P \Bigg[ n^{1/2} S_{n, j-1}^{-1/2} \bigg| \int_{t_{j-1}}^{t_j} b_t dt \bigg|
> \frac{n^{-\beta_0}}{6}, \ j \in {\cal L}_n^c \Bigg] \\
&\qquad \qquad
+ P \Bigg[ n^{1/2} S_{n, j-1}^{-1/2} \bigg| \int_{t_{j-1}}^{t_j} (\tilde{\sigma}_t - \tilde{\sigma}_{t_{j-1}}) dw_t \bigg|
> \frac{n^{-\beta_0}}{6}, \ j \in {\cal L}_n^c \Bigg] \\
&\qquad \qquad
+ P \Bigg[ {\color{black}}% {\color[rgb]{0,0,0.8} |W_j|}{\color{black}}% {\color[rgb]{0,0,0.8} S_{n, j-1}^{-1/2}} \Big| {\color{black}}% {\color[rgb]{0,0,0.8} S_{n, j-1}^{1/2}}- \sigma_{t_{j-1}} \Big|
> \frac{n^{-\beta_0}}{6}, \ j \in {\cal L}_n^c \Bigg]\\
&\lesssim
P \Bigg[ {\color{black}}% {\color[rgb]{0,0,0.8} |W_j|} < c(\alpha)^{1/2} - \frac{1}{2} n^{-\beta_0} \Bigg]
+ O(n^{-p(\gamma_0 - \beta_0)})\\
&\sim (1-\alpha) - \phi(c(\alpha)^{1/2}; 0, 1) n^{-\beta_0}
+ O(n^{-p(\gamma_0 - \beta_0)}), \text{ for all } p > 1.
\end{align*}
Here we used the following inequality{\color{black}}% {\color[rgb]{0,0,0.8} :} $C>0$,
\begin{align*}
&P \Bigg[ {\color{black}}% {\color[rgb]{0,0,0.8} |W_j|}{\color{black}}% {\color[rgb]{0,0,0.8} S_{n, j-1}^{-1/2}} \Big| {\color{black}}% {\color[rgb]{0,0,0.8} S_{n, j-1}^{1/2}} - \sigma_{t_{j-1}} \Big|
> Cn^{-\beta_0}, \ j \in {\cal L}_n^c \Bigg] \\
&\!\!\leq
P \Bigg[ {\color{black}}% {\color[rgb]{0,0,0.8} |W_j|}{\color{black}}% {\color[rgb]{0,0,0.8} S_{n, j-1}^{-1/2}} n^{\gamma_0} \Big| {\color{black}}% {\color[rgb]{0,0,0.8} S_{n, j-1}^{1/2}} - \sigma_{t_{j-1}} \Big| 1_{\{ j \in {\cal L}_n^c \}}
> Cn^{\gamma_0 - \beta_0} \Bigg] \\
&{\color{black}}% {\color[rgb]{0,0,0.8} \lesssim
n^{-p(\gamma_0 - \beta_0)}
\Bigg\| W_j S_{n,j-1}^{-1/2} \Big( S_{n,j-1}^{1/2} + \sigma_{t_{j-1}} \Big)^{-1} \times n^{\gamma_0} \Big( S_{n, j-1}- \sigma_{t_{j-1}}^2 \Big) 1_{\{ j \in {\cal L}_n^c \}} \Bigg\|_p^p} \\
&{\color{black}}% {\color[rgb]{0,0,0.8} \leq
n^{-p(\gamma_0 - \beta_0)}
\Bigg\| W_j S_{n,j-1}^{-1/2} \Big( S_{n,j-1}^{1/2} + \sigma_{t_{j-1}} \Big)^{-1} \Bigg\|_{2p}^p
\Bigg\| n^{\gamma_0} \Big( S_{n, j-1}- \sigma_{t_{j-1}}^2 \Big) 1_{\{ j \in {\cal L}_n^c \}} \Bigg\|_{2p}^p} \\
&{\color{black}}% {\color[rgb]{0,0,0.8} \lesssim
n^{-p(\gamma_0 - \beta_0)}
\Bigg\| W_j S_{n,j-1}^{-1/2} \Big( S_{n,j-1}^{-1/2} + \sigma_{t_{j-1}}^{-1} \Big) \Bigg\|_{2p}^p
\Bigg\| n^{\gamma_0} \Big( S_{n, j-1}- \sigma_{t_{j-1}}^2 \Big) 1_{\{ j \in {\cal L}_n^c \}} \Bigg\|_{2p}^p} \\
&{\color{black}}% {\color[rgb]{0,0,0.8} \lesssim
n^{-p(\gamma_0 - \beta_0)}
\Big\| n^{\gamma_0} \Big( S_{n, j-1}- \sigma_{t_{j-1}}^2 \Big) 1_{\{ j \in {\cal L}_n^c \}} \Big\|_{2p}^p = O(n^{-p(\gamma_0 - \beta_0)}) },
\end{align*}
{\color{black}}% {\color[rgb]{0,0,0.8} with the aid of the assumptions $[G1]$(iii), $[G2]$, and the inequality $(x + y)^{-1} \leq \frac{1}{4}(x^{-1} + y^{-1})$ for
$x, y > 0$ by the convexity of $f(x)=x^{-1}$. }
\noindent
For $\beta_0 < \min \Big\{ \gamma_0, \ \frac{1}{2} - \xi \Big\}$,
the order of the lower bound $C_n$ is given by
{\color{black}}% {\color[rgb]{0,0,0.8}
\begin{align*}
C_n &\gtrsim
n^{-1/2} \Bigg\{ (1-\alpha)n - 1 -
\sum_{j \in I_n} \Bigg( (1-\alpha) - \phi(c(\alpha)^{1/2};0, 1) n^{-\beta_0} + O(n^{-p(\gamma_0 - \beta_0)}) \Bigg) + O(n^{\xi + \frac{1}{2} }) \Bigg\} \\
&= n^{-1/2} \Bigg\{ - 1
+ \phi(c(\alpha)^{1/2};0, 1) n^{1-\beta_0} + O(n^{1-p(\gamma_0 - \beta_0)}) +
O(n^{\xi + \frac{1}{2} }) \Bigg\} \\
&\gtrsim
n^{\frac{1}{2} - \beta_0} + n^{\xi} \gtrsim n^{\frac{1}{2} - \beta_0}.
\end{align*}
}
Since the random variable $n^{-1/2} \sum_{j \in I_n} (1_{A_j} - P[A_j])$ is
bounded in $L^p$ for all $p > 1$, we have
\begin{align*}
P \left[ n^{1/2} V_{(s_n(\alpha))} - c(\alpha)^{1/2} < {\color{black}}% {\color[rgb]{0,0,0.8} -} n^{-\beta_0} \right]
&\leq
P \left[ n^{-1/2} \sum_{j \in I_n} \Big( 1_{A_j} - P[A_j] \Big) \gtrsim
n^{\frac{1}{2}-\beta_0} \right] = O(n^{-L}).
\end{align*}
\noindent
As the other inequality $P \left[ n^{1/2} V_{(s_n(\alpha))} - c(\alpha)^{1/2} > n^{-\beta_0} \right] = O(n^{-L})$ can be shown similarly,
the proof is completed.
\qed\vspace*{3mm}
\end{en-text}
\section{Asymptotic mixed normality of the global realized volatilities
with a moving threshold}\label{0211181815}
\subsection{{\color{black}}% {\color[rgb]{0.8,0,0} The GRV with a moving threshold}}
In this section, we will consider a situation where the intensity of jumps
is moderate. Then it is possible to keep the cut-off ratio of the data small, and
to get a precise estimate for the integrated volatility.
\begin{en-text}
{\xred
The ordinary realized volatility is useless when
the observations $(X_{t_j})_{j\in I_n}$, $I_n=\{1,...,n\}$, are contaminated with jumps.
To remove the influence of jumps, we will consider a realized volatility estimator with
a global jump filter.
}
\end{en-text}
Let
\begin{eqnarray}\label{0301030324}
\delta_0\in\big(0,1/4\big)\quad\text{and}\quad\delta_1\in\big(0,1/2\big).
\end{eqnarray}
In the context of the global jump filtering,
given a collection $({\mathfrak S}_{n,j-1})_{j\in I_n}$
of {\xred nonnegative}
random variables, we {\xred consider} the index set {\xred${\cal J}_n$} given by
\begin{eqnarray}\label{0301030241}
{\xred{\cal J}}_n &=&
\big\{j\in I_n;\> V_j<V_{(s_n)}\big\}
\end{eqnarray}
where
\begin{eqnarray}\label{0301030242}
V_j &=& \big|{\xred({\mathfrak S}_{n,j-1}^-)^{1/2}
\>\Delta_jX\big|
\end{eqnarray}
and
\begin{eqnarray}\label{0302110546}
s_n \>=\> n-\lfloor Bn^{\delta_1}\rfloor
\end{eqnarray}
for a positive constant $B$.
{\xred
Here $x^-=1_{\{x\not=0\}}x^{-1}$ for $x\in{\mathbb R}$.
\begin{remark}\rm
It is natural to set a spot volatility estimator of $\sigma_{t_{j-1}}^2$ in ${\mathfrak S}_{n,j-1}$
though not definitively necessary (Remark \ref{0302110521}).
In Section \ref{Sec3}, we discussed some constructions of ${\mathfrak S}_{n,j-1}$.
In the terminology of Section \ref{Sec2}, the cut-off rate by ${\cal J}_n$ is
$\alpha_n=\lfloor Bn^{\delta_1}\rfloor/n$,
${\cal J}_n={\cal J}_n(\alpha_n)$
and $\alpha_n$ goes to $0$ as $n$ tends to $\infty$.
We note that the definition of $V_j$ is different from that in (\ref{0211281531}).
\end{remark}
}
For estimation of $\Theta$ of (\ref{0211281514}), we consider
the global realized volatility {\color{black}}% {\color[rgb]{0.8,0,0}(GRV)} with a moving threshold
\begin{eqnarray}\label{0302101836
{\xred{\bf G}}_n
&=&
\sum_{j\in{\xred{\cal J}}_n}q_n^{-1}\big|\Delta_jX\big|^2H_{n,j}
\end{eqnarray}
where $(q_n)_{n\in{\mathbb N}}$ is a sequence of positive numbers,
and
\begin{eqnarray}\label{0301030351}
H_{n,j}&=&1_{\{|\Delta_jX|<B_0n^{-\frac{1}{4}-\delta_0}\}}
\end{eqnarray}
for a positive constant $B_0$.
{\xred
Here, $\sigma=(\sigma_t)_{t\in[0,T]}$ and
$b=(b_t)_{t\in[0,T]}$ are c\`adl\`ag\ adapted processes.
We will assume (\ref{0301030324}) and the following conditions.
\begin{en-text}
{\color{black}}% {\color[rgb]{0.8,0,0}
\begin{description}
\item[{\bf [G1]}]
{\bf (i)} For every $p>1$, $\sup_{t\in[0,T]}\|\sigma_t\|_p<\infty$ and
\begin{eqnarray*}
\big\|\widetilde{\sigma}_t-\widetilde{\sigma}_s\big\|_p &\leq& C(p)|t-s|^{1/2}
\quad(t,s\in[0,T])
\end{eqnarray*}
for some constant $C(p)$ for every $p>1$.
\item[{\bf (ii)}] $\sup_{t\in[0,T]}\|b_t\|_p<\infty$ for every $p>1$.
\item[{\bf (iii)}] $\sigma_t\not=0$ a.s. for every $t\in[0,T]$, an
$\sup_{t\in[0,T]}\big\|\sigma_t^{-1}\big\|_p<\infty$ for every $p>1$.
\end{description}
\end{en-text}
{\xred
\begin{description}
\item[{\bf [G1$^o$]}]
For every $p>1$, $\sup_{t\in[0,T]}\|\sigma_t\|_p+\|b_t\|_p<\infty$.
\end{description}
{\xred
\begin{description}
\item[{\bf [G2$^o$]}]
$q_n>0$ $(n\in{\mathbb N})$ and
$q_n-1=o(n^{-1/2})$ as $n\to\infty$.
\end{description}
\vspace*{3mm}
{\xred
\begin{remark}\label{0302110521}\rm
Theoretically, we may set ${\mathfrak S}_{n,j-1}=1$.
Condition ${\xred[G2^o]}$ is satisfied with $q_n=1$.
Asymptotically the choice $({\mathfrak S}_{n,j-1},q_n)=(1,1)$ is sufficient and valid.
However, in practice, a natural choice is a pair
$S_{n,j-1}$ {\xred satisfying $[G2]$ in Section \ref{Sec2}} and $q_n=q(\alpha_n)$,
{\xred where the function $q$ is defined by (\ref{0302110914}) in Section \ref{Sec2}.}
\end{remark}
{\xred
\begin{en-text}
The jump part $J$ of $X$ is assumed to be finitely active, i.e.,
$N_T<\infty$ a.s., $N_t=\sum_{s\leq t}1_{\{\Delta J_s\not=0\}}$.
We assume that the distribution of the variable $N_T$
depends on $n$
and $N_T$
may diverge as $n\to\infty$.
More precisely, the following condition is assumed
for $\Lambda_n=\#\{j\in I_n;\>${\coloroy{koko}}
\end{en-text}
For the jump part $J$ of the semimartingale $X$, we only assume
\begin{eqnarray*}
\Lambda_n &:=&\#\{j\in I_n;\>\Delta_jJ\not=0\}<\infty\quad a.s.
\end{eqnarray*}
for every $n\in{\mathbb N}$,
and the following estimate:
\begin{en-text}
{\xred The following condition models a fairly high jump-intensity of $\overline{N}$,
in particular, exogenous heavy contamination with jumps. }
\begin{description}
\item[{\bf [G3]}]
There exists a constant $\xi\geq0$
such that
$\|\overline{N}_T\|_p=O(n^{\xi})$ as $n\to\infty$
for every $p>1$.
\end{description}
\end{en-text}
{\xred
\begin{description}
\item[{\bf [G3$^o$]}]
There exists a constant $\xi\geq0$
such that
$\|{\xred \Lambda_n}\|_p=O(n^{\xi})$
as $n\to\infty$
for every $p>1$.
\end{description}
\begin{remark}\rm
The diverging $\Lambda_n$ models high intensity of the jump part for a fixed $n$
in practice.
Mathematically, we are assuming that the process $\sigma$ is independent of $n$.
This makes sense naturally in particular when the jumps are exogenous.
It is sufficient for the limit theorem by using the c\`adl\`ag\ property of $\sigma$.
On the other hand, though details are omitted, we can treat $\sigma$ depending on $n$
if
uniform $L^\infm$-continuity of $\sigma$
and uniformity in $[G1^o]$ are satisfied.
\end{remark}
{\xred
Define $\Gamma$ by
\begin{eqnarray*}
\Gamma &=& 2T\int_0^T \sigma_t^4dt.
\end{eqnarray*}
Extend $(\Omega,{\cal F},P)$ so that there is
a standard normal random variable $\zeta$ independent of ${\cal F}$
on the extension.
The ${\cal F}$-stable convergence is denoted by $\to^{d_s}$.
We obtain asymptotic mixed normality of the global realized volatility ${\bf G}_n$ with
a moving threshold.
\begin{theorem}\label{0302110536
Suppose that $[G1^o]$, $[G2^o]$ and $[G3^o]$ are satisfied.
Suppose that $\xi<2\delta_0$. Then
\begin{eqnarray*}
n^{1/2}\big(
{\xred {\bf G}_n}
-\Theta\big)
&\to^{d_s}&
\Gamma^{1/2}\zeta
\end{eqnarray*}
as $n\to\infty$.
\end{theorem}
{\xred
Theorem \ref{0302110536} follows from Theorem \ref{0211291415},
that is presented in a slightly more general setting.
}
\begin{en-text}
{\color{gray}
Just combine Lemmas
\ref{0211281630}, \ref{0211281845} and \ref{0211291358}.
}
\end{en-text}
\subsection{The WGRV with a moving threshold}
Suppose that a collection $({\mathfrak S}_{n,j-1})_{j\in I_n}$ ($n\in{\mathbb N}$)
of positive random variables is given.
Consider constants $\delta_0$ and $\delta_1$ satisfying (\ref{0301030324}),
and
$V_j$ and $s_n$ given by
(\ref{0301030242})
and (\ref{0302110546}), respectively.
We define
{\color{black}}% {\color[rgb]{0.8,0,0}
the Winsorized global realized volatility (WGRV) with a moving threshold by
}
\begin{eqnarray}\label{0302111147}
{\bf W}_n
&=&
\sum_{j \in I_n} q_n^{-1}
\big\{ |\Delta_j X| \wedge \mathfrak{S}_{n,j-1}^{1/2} V_{(s_n)} \big) \big\}^2 H_{n,j},
\end{eqnarray}
where $(q_n)_{n \in \mathbb{N}}$ is a sequence of positive numbers.
The error of the WGRV has the same limit as GRV ${\bf G}_n$.
}}
\begin{theorem}
Suppose that {\xred$[G1^o]$}, {\xred$[G2^o]$} and {\xred$[G3^o]$} are satisfied.
Suppose that $\xi < 2\delta_0$.
Then
\begin{eqnarray*}
n^{1/2}\big(
{\bf W}_n-\Theta\big)
&\to^{d_s}&
\Gamma^{1/2}\zeta
\end{eqnarray*}
as $n\to\infty$
where $\zeta$ is a standard Gaussian random variable independent of ${\cal F}$.
\end{theorem}
\proof
{\xred Let $\widetilde{X}=X-J$.}
It suffices to show that
$n^{1/2} \| {\bf W}_n - {\xred{\bf G}}_n \|_p \to 0$
as $n\to\infty$ for every $p > 1$.
{\xred From (\ref{0302111147}), }
\begin{eqnarray*}
{\bf W}_n - {\xred{\bf G}}_n
&=&
\sum_{j \in {\xred{\cal J}}_n^c} q_n^{-1} \mathfrak{S}_{n,j-1} V_{(s_n)}^2 H_{n,j}.
\end{eqnarray*}
We have
\begin{align*}
\left\| \sum_{j \in {\xred{\cal J}}_n^c} \mathfrak{S}_{n,j-1} V_{(s_n)}^2 H_{n,j} \right\|_p
&\leq
\left\| \sum_{j \in {\xred{\cal J}}_n^c} |\Delta_j X|^2 1_{\{ j \in \mathfrak{L}_n\}} H_{n,j} \right\|_p
+ \left\| \sum_{j \in {\xred{\cal J}}_n^c} |\Delta_j \tilde{X}|^2 1_{\{ j \in \mathfrak{L}_n^c\}} H_{n,j} \right\|_p \\
&\lesssim n^{-1/2-2\delta_0+\xi} + n^{-1+\epsilon+\delta_1}
\end{align*}
{\xred as $n\to\infty$ for any $\epsilon>0$.}
Since $\delta_1 < 1/2$, $\xi < 2\delta_0$
{\xred and $\epsilon$ is arbitrary}, we obtain the desired
{\xred convergence from Theorem \ref{0302110536}.}
\qed\vspace*{3mm}
\subsection{Stability of the realized volatility under missing}
{\color{black}}% {\color[rgb]{0.8,0,0}
We are about establishing asymptotic mixed normality of
the integrated volatility estimator having a moving threshold.
We will solve this problem by showing a stability of estimation under elimination of a certain portion of the data.
{\xred In other words, this is a question of stability under missing data.
In what follows, we will consider the variable ${\bf V}_n$ defined by
\begin{eqnarray}\label{0301030318}
{\bf V}_n
&=&
\sum_{j\in{\cal M}_n}q_n^{-1}\big|\Delta_jX\big|^2H_{n,j}
\end{eqnarray}
where $(q_n)_{n\in{\mathbb N}}$ is a sequence of positive numbers,
$H_{n,j}$ is given in (\ref{0301030351}), and
${\cal M}_n$ is an abstract random index set in $I_n$.}
It is not necessary to specify {\xred${\cal M}_n$ like ${\cal J}_n$}
by (\ref{0301030241}) and (\ref{0301030242}).
\begin{en-text}
\begin{description}
\item[{\bf [G2$'$]}]
\begin{description}
\item[{\bf (i)}]
The positive random variables ${\mathfrak S}_{n,j-1}$ $(j\in I_n, n\in{\mathbb N})$ satisfy
\begin{eqnarray*}
\sup_{n\in{\mathbb N}}\sup_{j\in I_n}\big(\|{\mathfrak S}_{n,j-1}\|_p+\|{\mathfrak S}_{n,j-1}^{-1}\|_p\big)
&<&
\infty
\end{eqnarray*}
for every $p>1$.
\item[{\bf (ii)}]
$|q_n-1|=o(n^{-1/2})$ as $n\to\infty$.
\end{description}
\end{description}
\end{en-text}
\begin{en-text}
{\color{black}}% {\color[rgb]{0,0.5,0} The following green part is not necessary. Delete it later.
Define two events ${\mathfrak A}_n$ and ${\mathfrak B}_n$ by
\begin{eqnarray*}
{\mathfrak A}_n
&=&
\bigcup_{j\in I_n}
\bigg[\big\{j\in{\cal M}_n^c\big\}\cap\big\{\Delta_jN=0\big\}\bigg]
\end{eqnarray*}
and
\begin{eqnarray*}
{\mathfrak B}_n
&=&
\bigcap_{j\in I_n}\bigg[\big\{V_j\geq V_{(s_n)}\big\}\cup\big\{|\Delta_jJ|\leq n^{-\frac{1}{4}-\delta_0}\big\}\bigg]
\end{eqnarray*}
\begin{lemma}\label{0211281550}
For any $L>0$, the following estimates hold as $n\to\infty$:
\begin{description}
\item[{\bf (i)}]
$P\big[{\mathfrak A}_n\cap{\mathfrak B}_n^{\>c}\big]\>=\> O(n^{-L})$
\item[{\bf (ii)}]
$P\big[{\mathfrak A}_n^{\>c}\big]\>=\> O(n^{-L})$
\item[{\bf (iii)}]
$P\big[{\mathfrak B}_n\big]\>=\> 1-O(n^{-L})$.
\end{description}
\end{lemma}
\proof {\coloroy{koko}}
}
\end{en-text}
Let
\begin{eqnarray*}
{\bf V}_n^\dagger
&=&
\sum_{j\in{\cal M}_n}q_n^{-1}\big|\Delta_j\widetilde{X}\big|^2H_{n,j}.
\end{eqnarray*}
Recall $\widetilde{X}=X-J$.
\begin{description}
\item[{\bf [G2$'$]}]
\begin{description}
\item[{\bf (i)}]
For every $n\in{\mathbb N}$, ${\cal M}_n$ is a random set in $I_n$ such that
$\#\big(I_n\setminus{\cal M}_n\big)\leq B_1n^{\delta_1}$ $(n\in{\mathbb N})$ for some positive constant $B_1$.
\item[{\bf (ii)}] $q_n>0$ $(n\in{\mathbb N})$ and
$q_n-1=o(n^{-1/2})$ as $n\to\infty$.
\end{description}
\end{description}
\vspace*{3mm}
\begin{lemma}\label{0211281630}
Suppose that {\xred$[G1^o]$}, $[G2']$ and {\xred$[G3^o]$} are satisfied.
Suppose that $\xi<2\delta_0$. Then
\begin{eqnarray*}
n^{1/2}\big\|{\bf V}_n-{\bf V}_n^\dagger\big\|_p
&\to&
0
\end{eqnarray*}
as $n\to\infty$ for every $p>1$.
\end{lemma}
\proof We have the estimate
\begin{eqnarray}\label{0211281840}
n^{1/2}\big\|{\bf V}_n-{\bf V}_n^\dagger\big\|_p
&\leq&
2q_n^{-1}\Phi_n^{(\ref{0211281652})}+q_n^{-1}\Phi_n^{(\ref{0211281653})},
\end{eqnarray}
where
\begin{eqnarray}\label{0211281652}
\Phi_n^{(\ref{0211281652})}
&=&
n^{1/2}\bigg\|
\sum_{j\in{\cal M}_n}\big|\Delta_j\widetilde{X}\Delta_jJ\big|
H_{n,j}
\bigg\|_p,
\end{eqnarray}
and
\begin{eqnarray}\label{0211281653}
\Phi_n^{(\ref{0211281653})}
&=&
n^{1/2}\bigg\|
\sum_{j\in{\cal M}_n}\big|\Delta_jJ\big|^2
H_{n,j}
\bigg\|_p
\end{eqnarray}
for $p>1$.
By using the inequality
\begin{eqnarray*}
|\Delta_jJ|H_{n,j}&\leq&\big(|\Delta_j\widetilde{X}|
+{\color{black}}% {\color[rgb]{0.8,0,0} B_0\>}
n^{-\frac{1}{4}-\delta_0}\big)1_{\{\Delta_jJ\not=0\}},
\end{eqnarray*}
we obtain
\begin{en-text}
\begin{eqnarray}\label{0211281656}
\Phi_n^{(\ref{0211281652})}
&\leq&
n^{1/2}\bigg\|
\sum_{j\in{\cal M}_n}\big|\Delta_j\widetilde{X}\Delta_jJ\big|
H_{n,j}1_{{\mathfrak B}_n}
\bigg\|_p
+
n^{1/2}\bigg\|
\sum_{j\in{\cal M}_n}\big|\Delta_j\widetilde{X}\Delta_jJ\big|
H_{n,j}1_{{\mathfrak B}_n^c}
\bigg\|_p
\nonumber\\&\leq
n^{1/2}\bigg\|
\max_{j\in I_n}\big\{\big|\Delta_j\widetilde{X}\big|(|\Delta_j\widetilde{X}|
+{\color{black}}% {\color[rgb]{0.8,0,0} B_0\>}n^{-\frac{1}{4}-\delta_0}\big)\big\}
N_T
\bigg\|_p
+
n^{1/2}\bigg\|
\max_{j\in{\cal M}_n}\big|\Delta_j\widetilde{X}\big|
\sum_{j\in{\cal M}_n}\big|\Delta_jJ\big|
H_{n,j}1_{{\mathfrak B}_n^c}
\bigg\|_p
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{-2\delta_0}\big\|N_T\big\|_p
+
\end{eqnarray}
\end{en-text}
\begin{eqnarray*}
\Phi_n^{(\ref{0211281652})}
&\leq&
n^{1/2}\bigg\|
\max_{j\in I_n}\big\{\big|\Delta_j\widetilde{X}\big|(|\Delta_j\widetilde{X}|
+{\color{black}}% {\color[rgb]{0.8,0,0} B_0\>}n^{-\frac{1}{4}-\delta_0}\big)\big\}
\bigg\|_{2p}
\big\|
{\xred\Lambda_n
\big\|_{2p}
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{-\frac{1}{4}-\delta_0+\xi+\epsilon}
\end{eqnarray*}
as $n\to\infty$ for any $\epsilon>0$ and $p>1$.
Therefore,
\begin{eqnarray}\label{0211281656}
\Phi_n^{(\ref{0211281652})}
&\to&
0
\end{eqnarray}
for every $p>1$
since $\xi<2\delta_0<\frac{1}{4}+\delta_0$.
Similarly,
\begin{eqnarray*}
\Phi_n^{(\ref{0211281653})}
&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{1/2}\bigg\|
\max_{j\in I_n}(|\Delta_j\widetilde{X}|^2+n^{-\frac{1}{2}-2\delta_0}\big)
\bigg\|_{2p}
\big\|
{\xred\Lambda_n
\big\|_{2p}
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{-2\delta_0+\xi+\epsilon}
\end{eqnarray*}
as $n\to\infty$ for any $\epsilon>0$ and $p>1$ since
{\xred$\delta_0<1/4$}.
In particular,
\begin{eqnarray}\label{0211281834}
\Phi_n^{(\ref{0211281653})}
&\to&
0
\end{eqnarray}
as $n\to\infty$ since $\xi<2\delta_0$.
Now the proof is completed with
(\ref{0211281840}), (\ref{0211281656}) and (\ref{0211281834}).
\qed\vspace*{3mm}
Define $\widetilde{\bf V}_n$ by
\begin{eqnarray*}
\widetilde{\bf V}_n
&=&
\sum_{j\in I_n}\big|\Delta_j\widetilde{X}\big|^2{\color{black}}% {\color[rgb]{0.8,0,0}.}
\end{eqnarray*}
\begin{lemma}\label{0211281845}
Suppose that $\xi<1/2$. Then
\begin{eqnarray*}
n^{1/2}\big\|{\bf V}_n^\dagger-\widetilde{\bf V}_n\big\|_p
&\to&
0
\end{eqnarray*}
as $n\to\infty$ for every $p>1$.
\end{lemma}
\proof
{\color{black}}% {\color[rgb]{0.8,0,0}
Recall that $\delta_0<1/4$ and $\delta_1<1/2$.
Define ${\bf V}_n^\ddagger$ by
\begin{eqnarray*}
{\bf V}_n^\ddagger
&=&
\sum_{j\in{\cal M}_n}q_n^{-1}\big|\Delta_j\widetilde{X}\big|^2
\end{eqnarray*}
Then
\begin{eqnarray*}
n^{1/2}\big\|{\bf V}_n^\dagger-{\bf V}_n^\ddagger\big\|_p
&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{1/2}\bigg\|\sum_{j\in{\cal M}_n}\big|\Delta_j\widetilde{X}\big|^2|H_{n,j}-1|\bigg\|_p
\nonumber\\&\leq&
n^{1/2}\bigg\|\sum_{j\in{\cal M}_n}\big|\Delta_j\widetilde{X}\big|^2|H_{n,j}-1|
1_{\{{\xred\Delta_jJ\not=0}\}}\bigg\|_p
\nonumber\\&&
+n^{1/2}\bigg\|\sum_{j\in{\cal M}_n}\big|\Delta_j\widetilde{X}\big|^2
1_{\{|\Delta_j\widetilde{X}|>n^{-\frac{1}{4}-\delta_0}\}}1_{\{{\xred\Delta_jJ=0}\}}\bigg\|_p
\nonumber\\&\leq&
n^{1/2}\bigg\|\max_{j\in I_n}\big|\Delta_j\widetilde{X}\big|^2{\xred\Lambda_n}\bigg\|_p
+O(n^{-L})
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{-\frac{1}{2}+\epsilon+\xi}
\end{eqnarray*}
for any positive number $\epsilon>0$.
Here $L$ is an arbitrary positive number greater than $1/2$, and
we used the inequality $\delta_0<1/4$ to get $O(n^{-L})$.
Since $\xi<1/2$, we obtain
\begin{eqnarray}\label{0301021857}
n^{1/2}\big\|{\bf V}_n^\dagger-{\bf V}_n^\ddagger\big\|_p
&=&
o(1)
\end{eqnarray}
as $n\to\infty$ for every $p>1$.
{\color{black}}% {\color[rgb]{0.8,0,0}
From the condition $q_n-1=o(n^{-1/2})$ of $[G2']$ (ii),} obviously,
\begin{eqnarray}\label{0211281849}
n^{1/2}\big\|{\color{black}}% {\color[rgb]{0.8,0,0}{\bf V}_n^\ddagger}-\widetilde{\bf V}_n\big\|_p
&\leq&
n^{1/2}\bigg\|\sum_{j\in I_n\setminus{\cal M}_n}\big|\Delta_j\widetilde{X}\big|^2\bigg\|_p
+o(1)
\nonumber\\&\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ &
n^{-\frac{1}{2}{\color{black}}% {\color[rgb]{0.8,0,0} + \epsilon}+\delta_1}+o(1)
\>=\>
o(1)
\end{eqnarray}
as $n\to\infty$ for every $p>1$ since
$\#(I_n\setminus{\cal M}_n)\ \raisebox{-.7ex}{$\stackrel{{\textstyle <}}{\sim}$}\ n^{\delta_1}$ with $\delta_1<1/2$
{\color{black}}% {\color[rgb]{0.8,0,0}
thanks to $[G2']$ (i) and (\ref{0301030324}),}
and
{\color{black}}% {\color[rgb]{0.8,0,0}
\begin{eqnarray*}
\bigg\|\max_{j\in I_n}\big|\Delta_j\widetilde{X}\big|^2\bigg\|_p
&=&
O(n^{-1+\epsilon})
\end{eqnarray*}
for any $p>1$ and any positive number $\epsilon$.
Proof ends with (\ref{0301021857}) and (\ref{0211281849}).
\qed\vspace*{3mm}
\begin{lemma}\label{0211291358}
Suppose that {\xred$[G1^o]$} is satisfied. Then
\begin{eqnarray*}
n^{1/2}\big(\widetilde{\bf V}_n-\Theta\big)
&\to^{d_s}&
\Gamma^{1/2}\zeta
\end{eqnarray*}
as $n\to\infty$.
\end{lemma}
\proof
We have
\begin{eqnarray}\label{0211291352}
\widetilde{\bf V}_n
&=&
\sum_{j\in I_n}
\left(\int_{t_{k-1}}^{t_k} \sigma_tdw_t+\int_{t_{k-1}}^{t_k} b_tdt\right)^2
\nonumber\\&=&
\Phi^{(\ref{0211280431})}_n+\Phi^{(\ref{0211280432})}_n+
2\Phi^{(\ref{0211280433})}_n
+\Phi^{(\ref{0211280435})}_n
\end{eqnarray}
where
\begin{eqnarray}\label{0211280431}
\Phi^{(\ref{0211280431})}_n
&=&
\sum_{j\in I_n}
2\int_{t_{j-1}}^{t_j}\int_{t_{j-1}}^t\sigma_sdw_s\sigma_tdw_t{\color{black}}% {\color[rgb]{0,0,0.8} , }
\end{eqnarray}
\begin{eqnarray}\label{0211280432}
\Phi^{(\ref{0211280432})}_n
&=&
\sum_{j\in I_n}
\int_{t_{j-1}}^{t_j} \sigma_t^2dt
{\xred\>\>=\>\Theta},
\end{eqnarray}
\begin{eqnarray}\label{0211280433}
\Phi^{(\ref{0211280433})}_n
&=&
\sum_{j\in I_n}
\int_{t_{j-1}}^{t_j} \sigma_tdw_t\int_{t_{j-1}}^{t_j} b_tdt,
\end{eqnarray}
and
\begin{eqnarray}\label{0211280435}
\Phi^{(\ref{0211280435})}_n
&=&
\sum_{j\in I_n}
\bigg(\int_{t_{j-1}}^{t_j} b_tdt\bigg)^2.
\end{eqnarray}
Since $b$ is a c\`adl\`ag\ process,
for any $\epsilon>0$, there exists a number $\delta>0$ such that
$P\big[w'(b,\delta)\geq\epsilon\big]<\epsilon$. Here $w'(b,\delta)$ is
a modulus of continuity defined by
\begin{eqnarray*}
w'(b,\delta)
&=&
\inf_{(s_i)\in{\cal S}_\delta}\max_i\sup_{r_1,r_2\in[s_{i-1},s_i)}\big|b_{r_1}-b_{r_2}\big|,
\end{eqnarray*}
where ${\cal S}_\delta$ is the set of sequences $(s_i)$ such that
$0=s_0<s_1<\cdots<s_v=T$ and $\min_{i=1,...,v-1}(s_i-s_{i-1})>\delta$.
Let
\begin{eqnarray}\label{0211290604}
\dot{\Phi}^{(\ref{0211280433})}_n
&=&
\sum_{j\in I_n}
\int_{t_{k-1}}^{t_k} \sigma_tdw_t\int_{t_{j-1}}^{t_j} (b_t-b_{t_{j-1}})dt{\color{black}}% {\color[rgb]{0,0,0.8} .}
\end{eqnarray}
Write
\begin{eqnarray*}
{\sf E}_j &=& \int_{t_{j-1}}^{t_j} \sigma_tdw_t,
\nonumber\\
{\sf V}_j &=& n^{1/2}\bigg|\int_{t_{j-1}}^{t_j} \sigma_tdw_t \bigg| \int_{t_{j-1}}^{t_j}\big(|b_t|+|b_{t_{j-1}}|\big)dt.
\end{eqnarray*}
For $\omega\in\Omega$ such that
$w'(b(\omega),\delta)<\epsilon$,
there exists a $(s_i)$ (depending on $\omega$) such that
\begin{eqnarray*} &&
\max_i\sup_{r_1,r_2\in[s_{i-1},s_i)}\big|b_{r_1}(\omega)-b_{r_2}(\omega)\big|<\epsilon,
\nonumber\\&&
\min_{i=1,...,v-1}(s_i-s_{i-1})>\delta.
\end{eqnarray*}
For $n>T/\delta$,
all intervals $[{t_{j-1}},{t_j})$ ($j\in I_n$) includes at most one point among $(s_i)$,
therefore the number of intervals $[{t_{j-1}},{t_j})$ that include some one $s_i$
is at most $T/\delta$.
The increment of $b(\omega)$ in $[{t_{j-1}},{t_j})$ is less than $\epsilon$
if $[{t_{j-1}},{t_j})\cap\{s_i\}{\xred=\emptyset}$.
Thus, we have the inequality
\begin{eqnarray*}
\big\|n^{1/2}\dot{\Phi}^{(\ref{0211280433})}_n\big\|_p
&\leq&
\bigg\|\sum_{j\in I_n}n^{1/2}|{\sf E}_j|\bigg\|_p\epsilon h
+\bigg\|\max_{j\in I_n}{\sf V}_j\bigg\|_p\frac{T}{\delta}
+\bigg\|\sum_{j\in I_n}{\sf V}_j\bigg\|_{2p}P\big[w'(b,\delta)
{\xred\geq}\epsilon\big]^{\frac{1}{2p}}
\end{eqnarray*}
for every $p>1$.
Therefore,
\begin{eqnarray*}
\big\|n^{1/2}\dot{\Phi}^{(\ref{0211280433})}_n\big\|_p
&\leq&
C\bigg[
\epsilon
+\bigg(n^{-1/2}+\sum_{j\in I_n}\big\|{\sf V}_j1_{\{{\sf V}_j>n^{-1/2}\}}\big\|_p\bigg)\frac{T}{\delta}
+\epsilon^{\frac{1}{2p}}\bigg]
\nonumber\\&\leq&
C'\big(\epsilon+n^{-1/2}+\epsilon^{\frac{1}{2p}}\big)
\end{eqnarray*}
for all $n>T/\delta$, where $C$ and $C'$ are some constants independent of $n$.
Consequently,
\begin{eqnarray}\label{0211291342}
\lim_{n\to\infty}\big\|n^{1/2}\dot{\Phi}^{(\ref{0211280433})}_n\big\|_p
&=&
0
\end{eqnarray}
for every $p>1$.
Moreover, for
\begin{eqnarray}
\ddot{\Phi}^{(\ref{0211280433})}_n
&=&
\sum_{j\in I_n}
\int_{t_{j-1}}^{t_j} \sigma_tdw_t\int_{t_{k-1}}^{t_k} b_{t_{j-1}} dt
\>=\>
\sum_{j\in I_n}hb_{t_{j-1}}
\int_{t_{j-1}}^{t_j} \sigma_tdw_t,
\end{eqnarray*}
we have
\begin{eqnarray}\label{0211291346}
\lim_{n\to\infty}\big\|n^{1/2}\ddot{\Phi}^{(\ref{0211280433})}_n\big\|_p
&=&
0
\end{eqnarray}
for every $p>1$, by orthogonality.
From (\ref{0211291342}) and (\ref{0211291346}),
\begin{eqnarray}\label{0211291347}
\lim_{n\to\infty}\big\|n^{1/2}\Phi^{(\ref{0211280433})}_n\big\|_p
&=&
0
\end{eqnarray}
for every $p>1$.
Obviously,
\begin{eqnarray}\label{0211291355}
\big\|n^{1/2}\Phi^{(\ref{0211280435})}_n\big\|_p
&=&
0
\end{eqnarray}
for every $p>1$.
Now, we can show the claim of the lemma by using
(\ref{0211291352}), (\ref{0211291347}) and (\ref{0211291355})
together with the mixture type of martingale central limit theorem
applied to ${\xred n^{1/2}}\Phi^{(\ref{0211280431})}_n$,
{\xred with the aid of the c\`adl\`ag\ property of $\sigma$.}
\qed\vspace*{3mm}
\begin{theorem}\label{0211291415}
Suppose that {\xred$[G1^o]$}, $[G2']$ and {\xred$[G3^o]$} are satisfied.
Suppose that $\xi<2\delta_0$. Then
\begin{eqnarray*}
n^{1/2}\big(
{\color{black}}% {\color[rgb]{0,0,0.8} {\bf V}_n}
-\Theta\big)
&\to^{d_s}&
\Gamma^{1/2}\zeta
\end{eqnarray*}
as $n\to\infty$.
\end{theorem}
\proof
Just combine Lemmas
\ref{0211281630}, \ref{0211281845} and \ref{0211291358}.
\qed\vspace*{3mm}
\section{Constant volatility}\label{0211281401}
The case of constant $\sigma$ is specific and theoretical treatments can be slightly different from those of the previous sections.
In this situation, we do not need to pre-estimate the local spot volatility,
and hence, we can take $S_{n,j}=1$ constantly and
no approximation error is caused.
$\sigma_t=\theta \tilde{\sigma}_t$ is also the case if $\tilde{\sigma}_{t_{j-1}}$ are observable.
For example, the GRV with a fixed cut-off rate $\alpha$ is redefined as
$$
{\mathbb V}_n^0 (\alpha) = \sum_{j \in {\cal J}_n^0(\alpha)} q(\alpha)^{-1} |\Delta_j X|^2 K_{n,j},
$$
where
$$
{\cal J}_n^0(\alpha) = \big\{ j\in I_n;\> |\Delta_j X| < |\Delta X|_{(s_n(\alpha))} \big\}.
$$
Then we have the following theorem.
Note that we do not need the condition $[G2]$,
and $\gamma_0$ in $[G2]$(ii) can be arbitrarily close to $1/2$.
\begin{theorem}
Suppose that $[G1]$ and $[G3]$ are fulfilled.
Suppose that $\xi < \frac{1}{2}$. Let $\alpha \in (0, 1)$ and
$\beta_0 < \frac{1}{2} - \xi$.
Then
$$
\| {\mathbb V}_n^0(\alpha) - \Theta\|_p = O(n^{-\beta_0})
$$
as $n \to \infty$ for every $p>1$.
\end{theorem}
The other global-threshold estimators are discussed similarly.
\section{Simulation studies}\label{Sec7}
In this section, we conduct several numerical simulations to see that
our global realized volatility estimators outperform those proposed in previous studies.
\subsection{The case of compound Poisson jumps}\label{SubSec7-1}
Here we consider {\color{black}}% {\color[rgb]{0.8,0,0} a process $X=(X_t)_{t\in[0,1]}$ satisfying
the stochastic differential equation}
\begin{equation}
dX_t = \theta X_t dt +
(\sigma + \eta X_t^2 )^{\frac{1}{4}} dw_t + dJ_t, \quad t \in [0, 1]{\color{black}}% {\color[rgb]{0,0,0.8} ,}
\end{equation}
{\color{black}}% {\color[rgb]{0.8,0,0} with $X_0=1$},
where $J_t$ is the jump part of $X$.
{\color{black}}% {\color[rgb]{0.8,0,0} In this section,}
we assume that $J$ is a compound Poisson process of the form
$
J_t = \sum_{i =1}^{N_t} \xi_i,
$
where $(N_t)_t$ is a Poisson process with intensity $\lambda > 0$ and
$(\xi_i)_i$ are independently and normally distributed random variable with
mean $\mu$ and variance $\nu^2$.
For the intensity parameter, we consider both cases where
$\lambda$ is high and low.
Our aim is to estimate the integrated volatility $\Theta = \int_0^1 (\sigma + \eta {\color{black}}% {\color[rgb]{0.8,0,0} X}_t^2)^{\frac{1}{2}} dt$.
{\color{black}}% {\color[rgb]{0.8,0,0} By simulation, we will}
compare {\color{black}}% {\color[rgb]{0.8,0,0} the performance of}
the {\color{black}}% {\color[rgb]{0,0,0.8} threshold realized volatility (TRV),} bipower variation (BV), minimum realized volatility (minRV), the GRV, and the WGRV,
{\color{black}}% {\color[rgb]{0.8,0,0} where TRV, BV and minRV are given by
\begin{align*}
\mathrm{TRV}_n &= \sum_{j=1}^n |\Delta_j X|^2 1_{\{ |\Delta_j X| \leq n^{-\rho} \}}, \quad \rho \in (0, 1/2), \\
\mathrm{BV}_n &= \frac{\pi}{2} \sum_{j=1}^{n-1} |\Delta_j X| |\Delta_{j+1} X|, \\
\mathrm{minRV}_n &= \frac{\pi}{\pi-2} \sum_{j=1}^{n-1} |\Delta_j X|^2 \wedge |\Delta_{j+1} X|^2,
\end{align*}
respectively.
The package YUIMA (cf. \cite{Yuima2014}, \cite{iacus2017simulation})
was used for the simulation studies below.
}
Note that, although TRV is based on threshold method, it is completely
different from our GRV, since TRV employs a deterministic threshold
and never uses information of other increments. In this sense,
TRV is based on a ``local" approach.
The set-up of simulation is as follows. The number of samples is $n = 2000$.
We repeat calculating the estimators 500 times to obtain their average and quantile.
The true parameters are
$\theta = 0.2, \ \sigma = 1, \ \eta = 3, \ \mu = 0.3, \ \nu = 0.2$.
Throughout this subsection, we set the cut-off ratio $\alpha = 0.2$ for
{\color{black}}% {\color[rgb]{0,0,0.8} GRV and WRGV} with {\color{black}}% {\color[rgb]{0.8,0,0} a local volatility estimator $S_{n,j-1}$}.
That is, we trim the upper 20\% of absolute increments.
While it may seem that we eliminate too many observations and
the estimator suffers from downside bias, GRV and WGRV estimate the integrated volatility
well thanks to the adjustment coefficient by $q(\alpha)$ and ${\sf w}(\alpha)$.
{\color{black}}% {\color[rgb]{0,0,0.8}
In calculating the TRV, we set $\rho = 0.45, 0.2, 0.1$ to see
the effect of the choice of this parameter on the accuracy of estimation. }
Note that {\color{black}}% {\color[rgb]{0.8,0,0} $\sigma_s$ in (\ref{0301191155})}
is not directly observable and depends on {\color{black}}% {\color[rgb]{0.8,0,0} $X_t$}.
Hence, {we need \color{black}}% {\color[rgb]{0.8,0,0} $S_{n,j-1}$}
to normalize the increment
$\Delta_i {\color{black}}% {\color[rgb]{0.8,0,0} X}$ when constructing the GRV. In this simulation,
we use the LGRV {\color{black}}% {\color[rgb]{0.8,0,0}${\mathbb L}_{n,j}(\alpha_0)$ of} (\ref{0211260215})
{\color{black}}% {\color[rgb]{0.8,0,0} with $\alpha_0=0.2$,}
and {\color{black}}% {\color[rgb]{0.8,0,0} the local minRV ${\mathbb M}_{n, j}$ of} (\ref{loc_minRV})
{\color{black}}% {\color[rgb]{0.8,0,0} for $S_{n,j-1}$.
We adopt $\kappa_n = \lfloor 10 \times n^{0.45} \rfloor$
for the length of a subinterval to calculate these local volatilities.}
Moreover, we calculate GRV without normalization (defined in Section \ref{0211281401})
for comparison.
{\color{black}}% {\color[rgb]{0,0,0.8}
Note that $\kappa_n$ depends on two tuning parameters, the choice of
which can affect the precision of estimation.
We argue this point in the final Section \ref{SubSec7-3}.
}
We use the following labels as in Table \ref{Table01}
to describe the estimators.
\renewcommand{\arraystretch}{1.5}
\begin{table}[H
\label{Table01}
\centering
\caption{Definitions of estimators}
\begin{tabular}[t]{lllll}
\hline
Label & Method & Spot volatility & Cut-off ratio $\alpha$ & Exponent $\rho$ for truncation \\
\hline \hline
{\tt trv[$\rho$]} & TRV & -- & -- & 0.45, 0.2, 0.1 \\
{\tt bv} & BV & -- & -- & -- \\
{\tt mrv} & minRV & -- & -- & -- \\
{\tt grv.lgrv[$\alpha$]} & GRV & GRV & 0.2 & --\\
{\tt grv.mrv[$\alpha$]} & GRV & minRV & 0.2 & -- \\
{\tt wgrv.lgrv[$\alpha$]} & WGRV & GRV & 0.2 & -- \\
{\tt wgrv.mrv[$\alpha$]} & WGRV & minRV & 0.2 & -- \\
{\tt grv[$\alpha$]} & GRV & -- & 0.2, 0.1, 0.05 & -- \\
{\tt grv.lgrv.mov} &GRV & GRV & depends on $n$& -- \\
{\tt wgrv.lgrv.mov} & WGRV & GRV & depends on $n$ & -- \\
\hline
\end{tabular}
\end{table}
\renewcommand{\arraystretch}{1.0}
\subsubsection{The case of high intensity: GRV with fixed cut-off ratio}
First, we deal with the case of high intensity.
Here we set $\lambda = 30$ so that the data includes many jumps.
The example of a sample path and its increments are shown in Figure \ref{fig1}.
Obviously, there are many large spikes in the data, suggesting the
existence of jumps.
Note that the volatility is non-constant here.
In fact, in Panel (b) of Figure \ref{fig1}, the size of increments tend to increase
as time passes.
Hence, to estimate the volatility, we have to use estimated spot volatilities
to normalize the increments.
{\color{black}}% {\color[rgb]{0.8,0,0}
In this example, we show the error ratios of GRV and WGRV with shrinking cut-off ratio (tuning parameters that determine the cut-off ratio $\alpha_n = \lfloor B n^{\delta_1} \rfloor$ for these estimators are $B=10$ and $\delta_1=0.45$,
the same as those used in the next subsection).
Theoretically, they are available in the case of moderate intensity of jumps.
We show their results just for reference.
We will discuss the case of moderate intensity in more detail in the next subsection.
}
\begin{figure}[H
\begin{center}
\begin{tabular}{c}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[keepaspectratio, scale=0.55]{Fig/Sim03_SamplePath10.eps}
{(a) Sample path of $X$}
\end{center}
\end{minipage}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[keepaspectratio, scale=0.55]{Fig/Sim03_SamplePath10_diff.eps}
{(b) Increment of $X$}
\end{center}
\end{minipage}
\end{tabular}
\caption{Sample path of $X$ and its increments ($\lambda = 30$)}
\label{fig1}
\end{center}
\end{figure}
{\color{black}}% {\color[rgb]{0,0.5,0} Table \ref{Table05} shows the summary of error ratios (percentage deviation of estimated values from the true value
for each estimator)}, and
Figure \ref{fig2}
{\color{black}}% {\color[rgb]{0.8,0,0} gives their box plots.}
In this case, both BV ({\tt bv}) and minRV ({\tt mrv}) seem to suffer from upward bias due to jumps. In particular, the BV deviates from the true value considerably.
On the other hand, GRV with normalization perform well with errors concentrating around zero ({\tt grv.lgrv, grv.mrv}). Note that, although WGRV performs relatively well, it seems to have a small upward bias ({\tt wgrv.lgrv, wgrv.mrv}).
This suggests that, if there are many large jumps,
using an upper quantile ($V_{(s_n(\alpha))}$) may sometimes lead to biases
rather than obtaining a robust estimate.
The three right {\color{black}}% {\color[rgb]{0.8,0,0} box plots}
in this figure ({\tt grv[0.20], grv[0.10], grv[0.05]}) are the results of GRV without
normalizing increments by local-global filters, with the cut-off ratio
$\alpha = 0.2, \ 0.1, \ 0.05$, respectively.
We see that they seem to be less precise (especially when $\alpha$ is large {\color{black}}% {\color[rgb]{0.8,0,0} extremely small}) than
GRV or WGRV with local volatility.
This result suggests that, if we do not normalize increments by spot volatilities in the case of non-constant volatility, we end up obtaining
inappropriate estimates.
Intuitively, when we ignore normalization, we tend to eliminate increments
where volatility is high (because they are typically large), even if they come from the Brownian motion, while keeping relatively small jumps which we should actually remove.
In addition, theoretically, the adjusting constant $q(\alpha)$ in the definition of GRV
(\ref{0211260834}) comes from the standard normal distribution.
Therefore, when the volatility is non-constant, we should normalize the increments
$|\Delta_i {\color{black}}% {\color[rgb]{0.8,0,0} X}|$ by local volatility to make them
approximately {\color{black}}% {\color[rgb]{0.8,0,0} standard} normally distributed.
\renewcommand{\arraystretch}{1.5}
\begin{table}[H] \centering
\caption{Summary of error ratios: $\lambda = 30$}
\vspace{2mm}
\begin{tabular}{@{\extracolsep{5pt}} lccccc}
\hline
& Min. & 1st Qu. & Median & 3rd Qu. & Max. \\
\hline \hline
{\tt trv[0.45]} & -29.88 & -28.16 & -27.42 & -26.45 & -21.64 \\
{\tt trv[0.20]} & -9.40 & -3.57 & -1.95 & -0.67 & 3.54 \\
{\tt trv[0.10]} & 0.32 & 3.38 & 4.55 & 6.11 & 21.61 \\
{\tt bv} & -0.25 & 2.66 & 3.87 & 5.11 & 8.98 \\
{\tt mrv} & -2.61 & 0.36 & 1.43 & 2.48 & 5.75 \\
{\tt grv.lgrv[0.20]} & -6.65 & -1.03 & -0.10 & 0.88 & 3.64 \\
{\tt grv.mrv[0.20]} & -6.65 & -1.33 & -0.40 & 0.64 & 3.65 \\
{\tt wgrv.lgrv[0.20]} & -3.50 & -0.39 & 0.60 & 1.46 & 3.68 \\
{\tt wgrv.mrv[0.20]} & -3.53 & -0.50 & 0.44 & 1.32 & 3.72 \\
{\tt grv[0.20]} & -9.29 & -4.39 & -3.21 & -1.91 & 2.15 \\
{\tt grv[0.10]} & -10.26 & -2.84 & -1.76 & -0.76 & 2.71 \\
{\tt grv[0.05]} & -13.85 & -4.41 & -1.77 & -0.42 & 3.38 \\
{\tt grv.lgrv.mov} & -9.13 & -1.18 & -0.00 & 0.95 & 3.55 \\
{\tt wgrv.lgrv.mov} & -3.19 & -0.14 & 0.80 & 1.62 & 4.61 \\
\hline \\[-1.8ex]
\end{tabular}
\label{Table05}
\end{table}
\renewcommand{\arraystretch}{1.0}
\begin{figure}[H
\begin{center}
\begin{tabular}{c}
\begin{minipage}{0.4\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim01_2021-01-23_lambda30_relA.eps}
\end{minipage}
\begin{minipage}{0.6\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim01_2021-01-23_lambda30_relB.eps}
\end{minipage}
%
\end{tabular}
\caption{Error ratios [\%] for the case of high intensity: $\lambda = 30$}
\label{fig2}
\end{center}
\end{figure}
{\color{black}}% {\color[rgb]{0,0.5,0}
The good news is that they also perform well even in the case of
extremely high intensity.
We consider $\lambda = 50$ here to see their accuracy.
Figure \ref{fig2EX} shows a sample path and its increments.
It is obvious there are numerous jumps and one can easily imagine
that the standard realized volatility estimator can never
estimate the true volatility.
{\color{black}}% {\color[rgb]{0,0.5,0}
Table \ref{Table08} and Figure \ref{fig9} show error ratios of each estimator for $\lambda=50$.
It shows that GRV and WGRV with cut-off ratio $\alpha=0.2$ perform well even
in the case of high intensity of jumps.
}
\begin{figure}[H
\begin{center}
\begin{tabular}{c}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[keepaspectratio, scale=0.55]{Fig/Sim01EX_2021-01-21_Path12_lambda50.eps}
{(a) Sample path of $X$}
\end{center}
\end{minipage}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[keepaspectratio, scale=0.55]{Fig/Sim01EX_2021-01-21_Path12_lambda50_diff.eps}
{(b) Increment of $X$}
\end{center}
\end{minipage}
\end{tabular}
\caption{Sample path of $X$ and its increments ($\lambda = 50$)}
\label{fig2EX}
\end{center}
\end{figure}
}
\renewcommand{\arraystretch}{1.5}
\begin{table}[H] \centering
\caption{Summary of error ratios: $\lambda = 50, \ \alpha = 0.2$}
\vspace{2mm}
\begin{tabular}{lccccc}
\hline
& Min. & 1st Qu. & Median & 3rd Qu. & Max. \\
\hline \hline
{\tt trv[0.45]} & -99.52 & -98.90 & -98.52 & -98.09 & -94.42 \\
{\tt trv[0.20]} & -52.18 & -27.61 & -19.86 & -12.98 & 2.35 \\
{\tt trv[0.10]} & 0.78 & 11.76 & 15.51 & 20.33 & 37.69 \\
{\tt bv} & 2.71 & 13.60 & 17.47 & 20.89 & 37.19 \\
{\tt mrv} & -6.36 & 3.40 & 7.72 & 11.58 & 27.20 \\
{\tt grv.lgrv[0.20]} & -38.58 & -7.51 & -2.38 & 1.82 & 18.41 \\
{\tt grv.mrv[0.20]} & -39.65 & -8.54 & -3.51 & 0.81 & 16.58 \\
{\tt wgrv.lgrv[0.20]} & -10.64 & 0.87 & 3.99 & 7.44 & 25.22 \\
{\tt wgrv.mrv[0.20]} & -12.60 & -0.08 & 3.18 & 6.74 & 23.23 \\
{\tt grv[0.20]} & -33.92 & -15.51 & -11.40 & -7.01 & 5.42 \\
{\tt grv[0.10]} & -56.46 & -23.57 & -11.63 & -4.84 & 9.21 \\
{\tt grv[0.05]} & -66.14 & -40.55 & -29.94 & -17.97 & 5.88 \\
{\tt grv.lgrv.mov} & -48.17 & -16.03 & -6.55 & -0.14 & 16.16 \\
{\tt wgrv.lgrv.mov} & -10.31 & 2.16 & 5.85 & 9.61 & 31.21 \\
\hline
\end{tabular}
\label{Table08}
\end{table}
\renewcommand{\arraystretch}{1.0}
\begin{figure}[H
\begin{center}
\begin{tabular}{c}
\begin{minipage}{0.4\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim01EX2_2021-01-23_lambda50_relA.eps}
\end{minipage}
\begin{minipage}{0.6\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim01EX2_2021-01-23_lambda50_relB.eps}
\end{minipage}
%
\end{tabular}
\caption{Error ratios [\%] for the case of high intensity: $\lambda = 50$}
\label{fig9}
\end{center}
\end{figure}
{\color{black}}% {\color[rgb]{0,0.5,0}
By taking $\alpha$ larger, accuracy improves.
Table \ref{Table04} shows the error ratios of each estimator in the case of $\lambda = 50$,
with $\alpha = \alpha_0$ ranging from 0.1 to 0.5.
We can see that, for large cut-off ratio ($\alpha=\alpha_0 = 0.4, 0.5$),
GRV and WGRV with spot volatilities
({\tt grv.lgrv}, {\tt grv.mrv}, {\tt wgrv.lgrv}, {\tt wgrv.mrv}) still
perform well.
Looking in more detail, we see that the local GRV outperforms
local minRV for both GRV and WGRV.
This example imply that we should take a cut-off ratio
quite large in order to obtain a precise estimate.
}
\vspace{5mm}
\renewcommand{\arraystretch}{1.5}
\begin{table}[H]
\begin{center}
\caption{Error ratios [\%] for the case of extremely high intensity: $\lambda = 50$}
\vspace{2mm}
\begin{tabular}{lccccc}
\hline
& \multicolumn{5}{c}{Cut-off ratio ($\alpha = \alpha_0$)} \\ \cline{2-6}
& 0.1 & 0.2 & 0.3 & 0.4 & 0.5 \\ \hline
{\tt trv[0.45]} & -98.40 & -98.40 & -98.40 & -98.40 & -98.40 \\
{\tt trv[0.20]} & -20.35 & -20.35 & -20.35 & -20.35 & -20.35 \\
{\tt trv[0.10]} & 16.19 & 16.19 & 16.19 & 16.19 & 16.19 \\
{\tt bv} & 17.60 & 17.60 & 17.60 & 17.60 & 17.60 \\
{\tt mrv} & 7.71 & 7.71 & 7.71 & 7.71 & 7.71 \\
{\tt grv.lgrv}[$\alpha$] & -18.04 & -3.49 & -0.41 & -0.96 & -1.47 \\
{\tt grv.mrv}[$\alpha$] & -18.05 & -4.35 & -2.62 & -3.90 & -4.89 \\
{\tt wgrv.lgrv}[$\alpha$] & 10.28 & 4.05 & 2.27 & 1.35 & 0.59 \\
{\tt wgrv.mrv} [$\alpha$] & 10.87 & 3.21 & 1.06 & -0.30 & -1.40 \\
{\tt grv[0.20]} & -11.50 & -11.50 & -11.50 & -11.50 & -11.50 \\
{\tt grv[0.10]} & -14.60 & -14.60 & -14.60 & -14.60 & -14.60 \\
{\tt grv[0.05]} & -28.85 & -28.85 & -28.85 & -28.85 & -28.85 \\
{\tt grv.lgrv.mov} & -8.31 & -8.76 & -8.78 & -8.80 & -8.83 \\
{\tt wgrv.lgrv.mov} & 4.99 & 6.26 & 6.48 & 6.63 & 6.73 \\
\hline
\end{tabular}
\label{Table04}
\end{center}
\end{table}
\renewcommand{\arraystretch}{1.0}
\subsubsection{The case of moderate intensity: GRV with a shrinking cut-off ratio}
Next, we consider the case of low intensity.
In this case, we can use shrinking cut-off rate.
Recall that the shrinking cut-off rate is defined by
$\alpha_n = \lfloor B n^{\delta_1} \rfloor / n$.
In this simulation, we set $B = 10$ and $\delta_1 = 0.45$, so the
cut-off rate is then $\alpha_n = 0.1525$.
The {\color{black}}% {\color[rgb]{0,0.5,0} error ratios} are shown in Table \ref{Table07} Figure \ref{fig3}.
All global-filtering estimators perform well (for GRVs with fixed cut-off ratio,
we set $\alpha = 0.2$ as before).
These results suggest that if there are not so many jumps in the data,
it would be advisable to use as many data as possible by making
the cut-off ratio small.
{\color{black}}% {\color[rgb]{0,0,0.8}
Note that TRV still has bias, especially for $\rho=0.45$.
This implies that the accuracy of estimation is still highly
vulnerable to the choice of $\rho$ for TRV,
even in the case of moderate intensity of jumps.
}
\renewcommand{\arraystretch}{1.5}
\begin{table}[H] \centering
\caption{Summary of error ratios: $\lambda = 5$}
\vspace{2mm}
\begin{tabular}{@{\extracolsep{5pt}} lccccc}
\hline
& Min. & 1st Qu. & Median & 3rd Qu. & Max. \\
\hline \hline
{\tt trv[0.45]} & -29.61 & -27.40 & -26.09 & -23.89 & -18.27 \\
{\tt trv[0.20]} & -2.79 & -0.58 & 0.14 & 0.85 & 3.86 \\
{\tt trv[0.10]} & -2.39 & 1.16 & 2.52 & 4.25 & 13.51 \\
{\tt bv} & -2.26 & 0.31 & 1.29 & 2.24 & 6.34 \\
{\tt mrv} & -3.63 & -0.69 & 0.29 & 1.24 & 4.27 \\
{\tt grv.lgrv[0.20]} & -3.62 & -1.44 & -0.55 & 0.21 & 4.79 \\
{\tt grv.mrv[0.20]} & -3.59 & -1.40 & -0.55 & 0.30 & 5.00 \\
{\tt wgrv.lgrv[0.20]} & -3.52 & -1.11 & -0.39 & 0.44 & 4.55 \\
{\tt wgrv.mrv[0.20]} & -3.53 & -1.06 & -0.40 & 0.46 & 4.62 \\
{\tt grv[0.20]} & -7.14 & -2.82 & -1.75 & -0.75 & 2.90 \\
{\tt grv[0.10]} & -5.46 & -2.15 & -1.32 & -0.46 & 3.17 \\
{\tt grv[0.05]} & -4.15 & -1.63 & -0.86 & -0.10 & 3.23 \\
{\tt grv.lgrv.mov} & -3.54 & -1.32 & -0.50 & 0.27 & 4.87 \\
{\tt wgrv.lgrv.mov} & -3.15 & -0.96 & -0.34 & 0.47 & 4.17 \\
\hline
\end{tabular}
\label{Table07}
\end{table}
\renewcommand{\arraystretch}{1.0}
\begin{figure}[H
\begin{tabular}{c}
\begin{minipage}{0.4\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim01_2021-01-21_lambda5_relA.eps}
\end{minipage}
\begin{minipage}{0.6\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim01_2021-01-21_lambda5_relB.eps}
\end{minipage}
%
\end{tabular}
\caption{Error ratios [\%] for the case of low intensity: $\lambda = 5$}
\label{fig3}
\end{figure}
{\color{black}}% {\color[rgb]{0,0.5,0}
For GRV and WGRV with shrinking cut-off ratio,
we proved the asymptotic mixed normality.
Hence, the distribution of the Studentized errors
$\Gamma^{-1/2} \sqrt{n} ({\bf V}_n - \Theta)$ and
$\Gamma^{-1/2} \sqrt{n} ({\bf W}_n - \Theta)$ are expected to follow the standard normally distribution.
Figure \ref{fig7} shows QQ plots comparing theoretical quantiles of the standard normal distribution and the Studentized errors of GRV and WGRV estimators with shrinking threshold, BV and minRV.
In this example, {\tt wgrv.lgrv.mov} outperforms the others.
It is close to the standard normal distribution.
On the other hand, {\tt grv.lgrv.mov} seems to deviate {\color{black}}% {\color[rgb]{0.8,0,0} from} $N(0,1)$.
We can also see that {\tt bv} are far from $N(0,1)$, implying
that it is not appropriate even in the case of low intensity.
\begin{figure}[H
\begin{center}
\begin{tabular}{c}
\begin{minipage}{0.5\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim01_2021-01-21_lambda5_QQ_id13.eps}
\end{minipage}
\begin{minipage}{0.5\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim01_2021-01-21_lambda5_QQ_id14.eps}
\end{minipage} \\
\begin{minipage}{0.5\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim01_2021-01-21_lambda5_QQ_id4.eps}
\end{minipage}
\begin{minipage}{0.5\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim01_2021-01-21_lambda5_QQ_id5.eps}
\end{minipage}
%
\end{tabular}
\caption{QQ plot for Studentized errors: $\lambda = 5$}
\label{fig7}
\end{center}
\end{figure}
}
{\color{black}}% {\color[rgb]{0,0,0.8}
{\color{black}}% {\color[rgb]{0,0.5,0} The important tuning parameter for the shrinking threshold GRV} is the exponent $\delta_1$,
an appropriate choice of which may strongly depend on the intensity of jumps. Recall that small $\delta_1$ means that we keep almost all the samples untrimmed.
Table \ref{Table02} shows average {\color{black}}% {\color[rgb]{0,0.5,0} error ratios}
of GRV and WGRV with shrinking
cut-off ratio for several values of intensity $\lambda$ and the parameter $\delta_1$.
For moderate intensity ($\lambda = 5, 10$),
{\color{black}}% {\color[rgb]{0,0.5,0} the average ratios are not so large for small $\delta_1$. }
On the other hand, for high intensity ($\lambda = 30, 50$), this is not the case.
Indeed, as for GRV, estimation errors are quite large downward for small $\delta_1$.
This can be interpreted that its multiplication of $q(\alpha_n)^{-1}$ for GRV is insufficient to compensate its
elimination of jumps (small $\delta_1$ implies small $\alpha_n$, making $q(\alpha_n)^{-1}$ close to 1).
Moreover, as for WGRV, there occur large upward biases for small $\delta_1$,
since it keeps almost large increments and uses an extremely large increment
for winsorization.
It is worth noting that large $\delta_1$ makes both GRV and WGRV accurate
{\color{black}}% {\color[rgb]{0,0.5,0} to a certain extent},
even in the case of high intensity of jumps.
Thus, in practice, one may use shrinking cut-off GRV and WGRV
by setting the tuning parameter $\delta_1$ sufficiently close to $1/2$.
{\color{black}}% {\color[rgb]{0,0.5,0}
However, as Figure \ref{fig8} implies, the errors are not normally distributed as
theory predicts when the intensity of jumps is extremely high.
We should be aware that GRV and WGRV with shrinking cut-off may suffer from
some biases in the case of extremely intensive jumps.
We may consider using a large fixed cut-off ratio (as discussed in
the previous subsection) in such a situation.
}
\begin{figure}[H
\begin{center}
\begin{tabular}{c}
\begin{minipage}{0.5\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim04_2021-01-23_lambda50_QQ_id13.eps}
\end{minipage}
\begin{minipage}{0.5\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim04_2021-01-23_lambda50_QQ_id14.eps}
\end{minipage}
%
\end{tabular}
\caption{QQ plot for Studentized errors: $\lambda = 50, \ \delta_1 = 0.49$}
\label{fig8}
\end{center}
\end{figure}
}
\renewcommand{\arraystretch}{1.5}
\begin{table}[H]
\caption{Average error ratios [$\%$] of GRV and WGRV with shrinking cut-off ratio}
\begin{minipage}{0.5\hsize}
\begin{center}
\vspace{3mm}
{(a) GRV}\\
\vspace{3mm}
\begin{tabular}{ccccc}\hline
\multicolumn{1}{c}{} & \multicolumn{4}{c}{Intensity of jumps ($\lambda$) }\\
\cline{2-5}
$\delta_1$ & 5 & 10 & 30 & 50 \\ \hline
0.10 & -0.41 & -3.07 & -23.89 & -43.57 \\
0.20 & -0.28 & -1.56 & -18.95 & -39.12 \\
0.30 & -0.58 & -0.63 & -11.40 & -31.14 \\
0.40 & -1.22 & -0.65 & -2.91 & -17.39 \\
0.45 & -1.57 & -1.09 & -0.57 & -8.76 \\
0.49 & -1.82 & -1.50 & -0.43 & -3.01 \\ \hline
\end{tabular}
\end{center}
\end{minipage}
\hfill
\begin{minipage}{0.5\hsize}
\begin{center}
\vspace{3mm}
{(b) WGRV}\\
\vspace{3mm}
\begin{tabular}{ccccc}\hline
\multicolumn{1}{c}{} & \multicolumn{4}{c}{Intensity of jumps ($\lambda$) }\\
\cline{2-5}
$\delta_1$ & 5 & 10 & 30 & 50 \\ \hline
0.10 & 2.03 & 5.88 & 60.71 & 208.18 \\
0.20 & 1.18 & 3.43 & 22.96 & 66.24 \\
0.30 & 0.43 & 1.76 & 10.27 & 29.11 \\
0.40 & -0.32 & 0.60 & 4.30 & 11.13 \\
0.45 & -0.80 & 0.00 & 2.74 & 6.26 \\
0.49 & -1.11 & -0.46 & 1.85 & 3.83\\ \hline
\end{tabular}
\end{center}
\end{minipage}
\label{Table02}
\end{table}
\renewcommand{\arraystretch}{1.0}
\subsubsection{The case of constant volatility}
Since we assumed that the volatility is location-dependent in the previous
sections, the normalization by estimated spot volatilities is needed to
obtain an accurate estimator.
However, if the true volatility of data is constant,
we may ignore normalization.
Here we set $\eta = 0$ so that the data is driven by a constant-volatility diffusion process. The intensity is $\lambda = 30$.
{
\color{black}}% {\color[rgb]{0,0.5,0}
The summary table of estimated values are shown in Table \ref{Table06}.
Obviously, all types of GRV and WGRV outperform other estimators.
}
Figure \ref{fig4} shows the {\color{black}}% {\color[rgb]{0,0.5,0} error ratios} of this case.
The GRVs without normalization ({\tt grv[0.20]}, {\tt grv[0.10]} and {\tt grv[0.05]}) perform as well as those with normalization.
This suggests that, if the true process can be thought as constant-volatility,
we may skip normalization (calculation of spot volatilities) procedure.
However, it would be more typical that the volatility is non-constant.
Thus, basically, it would be advisable to use normalization.
\renewcommand{\arraystretch}{1.5}
\begin{table}[H] \centering
\caption{Summary table of estimated values: $\lambda = 30$}
\vspace{2mm}
\begin{tabular}{@{\extracolsep{5pt}} lccccc}
\hline
& Min. & 1st Qu. & Median & 3rd Qu. & Max. \\
\hline \hline
{\tt trv[0.45]} & 0.41 & 0.44 & 0.45 & 0.46 & 0.49 \\
{\tt trv[0.20]} & 0.97 & 1.11 & 1.16 & 1.22 & 1.45 \\
{\tt trv[0.10]} & 1.67 & 2.49 & 2.76 & 3.08 & 4.81 \\
{\tt bv} & 1.16 & 1.43 & 1.53 & 1.63 & 2.52 \\
{\tt mrv} & 0.92 & 1.03 & 1.07 & 1.13 & 2.52 \\
{\tt grv.lgrv[0.20]} & 0.94 & 1.01 & 1.05 & 1.07 & 1.17 \\
{\tt grv.mrv[0.20]} & 0.94 & 1.02 & 1.05 & 1.09 & 1.32 \\
{\tt wgrv.lgrv[0.20]} & 0.95 & 1.02 & 1.05 & 1.08 & 1.16 \\
{\tt wgrv.mrv[0.20]} & 0.95 & 1.02 & 1.05 & 1.08 & 1.17 \\
{\tt grv[0.20]} & 0.94 & 1.01 & 1.05 & 1.07 & 1.17 \\
{\tt grv[0.10]} & 0.95 & 1.02 & 1.06 & 1.08 & 1.16 \\
{\tt grv[0.05]} & 0.96 & 1.04 & 1.07 & 1.09 & 1.17 \\
{\tt grv.lgrv.mov} & 0.95 & 1.02 & 1.05 & 1.08 & 1.16 \\
{\tt wgrv.lgrv.mov} & 0.95 & 1.02 & 1.05 & 1.08 & 1.15 \\
\hline \hline
{\tt True Value} & 1.00 & 1.00 & 1.00 & 1.00 & 1.00 \\
\hline
\end{tabular}
\label{Table06}
\end{table}
\renewcommand{\arraystretch}{1.0}
\begin{figure}[H
\begin{tabular}{c}
\begin{minipage}{0.4\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim03_2021-01-21_lambda30_relA.eps}
\end{minipage}
\begin{minipage}{0.6\hsize}
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim03_2021-01-21_lambda30_relB.eps}
\end{minipage}
%
\end{tabular}
\caption{Error ratios [\%] results for the constant volatility: $\lambda = 30$}
\label{fig4}
\end{figure}
\subsection{The case of Neyman-Scott type clustering jumps}
As the previous examples show, the minRV performs relatively well in the case of
compound Poisson type jumps.
However, even if the intensity of jumps is small, the minRV
may suffer from an upward bias depending on the structure of jumps.
In particular, if there are consecutive jumps (which is quite rare
for compound Poisson processes), the minRV loses it advantage.
Here we show an example of such a situation.
We consider the case that the data-generating process is given by
{\color{black}}% {\color[rgb]{0.8,0,0} $X=U+J$,}
where {\color{black}}% {\color[rgb]{0.8,0,0} $U$} is the continuous part and $J$ is the jump part.
Here we assume that $J$ {\color{black}}% {\color[rgb]{0.8,0,0} is a {\it marked Neyman-Scott clustering process}}
({\color{black}}% {\color[rgb]{0.8,0,0} simply denoted by} NS hereafter), instead of a compound Poisson process.
{\color{black}}% {\color[rgb]{0.8,0,0} The} NS process is a typical point process representing consecutive jumps.
That is, there may be jumps within some consecutive intervals.
This leads to upward bias of BV and minRV because the both of two adjacent increments can consist of large jumps.
The NS process is constructed as follows.
{\color{black}}% {\color[rgb]{0,0,0.8}
\begin{description}
\item[(1)]
Set ``centers" on the time interval $[0, {\color{black}}% {\color[rgb]{0.8,0,0} 1}]$ by a Poisson process $(N_t^0)$ with intensity $\lambda_0$. A center is defined as the point $t \in [0, {\color{black}}% {\color[rgb]{0.8,0,0} 1}]$ which
satisfies $\Delta N_t = 1$.
\item[(2)]
For each center $c \in [0, {\color{black}}% {\color[rgb]{0.8,0,0} 1}]$, choose the number $N_c$ of ``children,"
assuming $N_c$ is Poisson-distributed with mean $\lambda_c$.
\item[(3)]
For each center $c \in [0, {\color{black}}% {\color[rgb]{0.8,0,0} 1}]$, generate independently and exponentially distributed random variables $\left( v_i^{(c)} \right)_{1 \leq i \leq N_c}$
with mean $h$.
Then the location of child $i$ derived from center $c$ is defined as $c - v_i^{(c)}$.
This defines the location of a jump.
\item[(4)]
For each child $i$, generate an independently and normally distributed random variable $\xi_i \sim N(0, \nu_J^2)$. This determines the size and direction of a jump $\Delta J_s$.
\item[(5)]
The NS process is defined as $J_t = \sum_{s \in [0, t]} \Delta J_s$.
\end{description}
}
We
{\color{black}}% {\color[rgb]{0.8,0,0} generate $X=U+J$,}
where ${\color{black}}% {\color[rgb]{0.8,0,0} U}$ is the Brownian semimartingale {\color{black}}% {\color[rgb]{0.8,0,0} independent of $J$},
satisfying the {\color{black}}% {\color[rgb]{0.8,0,0} stochastic differential equation
\begin{equation}
dU_t = \theta\> U_t dt +
(\sigma + \eta\> U_t^2 )^{\frac{1}{4}} dw_t
\end{equation}
with $U_0=1$.
}
We set $\lambda_0 = \lambda_c = 5$ and $\nu_J = 0.5$.
For the continuous part {\color{black}}% {\color[rgb]{0,0.5,0} $U$}, we use $\theta = 0.2, \ \sigma = 1, \ \eta = 3$.
{\color{black}}% {\color[rgb]{0.8,0,0} As before, the number $n$ of samples is $n = 2000$, and the number
of trials is 500. }
\renewcommand{\arraystretch}{1.5}
\begin{table}[H] \centering
\caption{Summary table of error ratios: Neyman-Scott clustering jumps}
\vspace{2mm}
\begin{tabular}{@{\extracolsep{5pt}} lccccc}
\hline
& Min. & 1st Qu. & Median & 3rd Qu. & Max. \\
\hline \hline
{\tt trv[0.45]} & -97.04 & -88.67 & -82.29 & -75.74 & -59.83 \\
{\tt trv[0.20]} & -74.35 & -29.07 & -11.09 & 6.46 & 138.67 \\
{\tt trv[0.10]} & -68.02 & -14.24 & 3.17 & 24.43 & 157.66 \\
{\tt bv} & -54.60 & 5.52 & 27.19 & 69.17 & 369.71 \\
{\tt mrv} & -67.31 & -1.40 & 19.59 & 61.35 & 300.83 \\
{\tt grv.lgrv[0.20]} & -74.39 & -31.45 & -14.02 & 3.64 & 136.14 \\
{\tt grv.mrv[0.20]} & -70.53 & -26.44 & -9.19 & 8.39 & 139.11 \\
{\tt wgrv.lgrv[0.20]} & -74.49 & -31.31 & -13.35 & 4.32 & 137.83 \\
{\tt wgrv.mrv[0.20]} & -74.48 & -30.70 & -12.73 & 4.23 & 136.96 \\
{\tt grv[0.20]} & -74.89 & -33.38 & -16.67 & 0.73 & 134.64 \\
{\tt grv[0.10]} & -74.70 & -32.63 & -15.03 & 1.88 & 136.44 \\
{\tt grv[0.05]} & -74.38 & -31.68 & -14.04 & 3.40 & 136.67 \\
{\tt grv.lgrv.mov} & -74.50 & -31.63 & -13.79 & 3.84 & 138.23 \\
{\tt wgrv.lgrv.mov} & -74.32 & -30.82 & -12.89 & 4.34 & 139.78 \\
\hline
\end{tabular}
\label{Table10}
\end{table}
\renewcommand{\arraystretch}{1.0}
{\color{black}}% {\color[rgb]{0,0.5,0} Table \ref{Table10} and Figure \ref{fig5} show the error ratios} in the case of NS jumps.
Because of the possible consecutive jumps, both bipower variation and
minRV have upward bias, whereas {\color{black}}% {\color[rgb]{0,0,0.8} GRV and WRGV} are all robust to such clustering jumps.
This suggests that the {\color{black}}% {\color[rgb]{0,0,0.8} GRV and WRGV perform} very well for various
structures of jumps.
\begin{figure}[H
\includegraphics[keepaspectratio, scale=0.7]{Fig/Sim02_2021-01-21_rel.eps}
\caption{Error ratios [\%] for the case of Neyman-Scott clustering jumps}
\label{fig5}
\end{figure}
{\color{black}}% {\color[rgb]{0,0,0.8}
\subsection{A remark on estimation of spot volatilities} \label{SubSec7-3}
Finally, we argue how estimation of spot volatilities affect the
accuracy of GRV and WGRV.
We have used $\kappa_n = \lfloor B n^c \rfloor = \lfloor 10n^{0.45} \rfloor = 305$ for local GRV and local minRV and seen that GRV and WGRV with these spot volatilities perform highly well.
However, the choice of $\kappa_n$ may affect the accuracy of
GRV and WGRV.
In fact, if the true volatility varies greatly, a wide subinterval (a large $\kappa_n$)
leads to imprecise estimation of spot volatilities and causes
misdetection of jumps by using such information.
Therefore, it ends up obtaining biases of GRV and WGRV.
To see this, consider the following SDE:
\begin{align*}
dX_t = \theta X_t dt + (\sigma + \eta \sin^2 X_t) dw_t + dJ_t,
\end{align*}
where $J_t = \sum_{j=1}^{N_t} \xi_j$ is the same compound Poisson process with intensity $\lambda$ as in Section \ref{SubSec7-1}.
We set $\sigma = 1, \eta = 5, \lambda = 10, \mu = 0.3, \nu = 0.2$.
{\color{black}}% {\color[rgb]{0.8,0,0} Again, the number $n$ of samples is $n = 2000$, and the number
of trials is 500. }
{\color{black}}% {\color[rgb]{0.8,0,0} In this example}, the volatility $(\sigma + \eta \sin^2 X_t)^2$ swings in the range $[1, 36]$.
A sample path of this model is shown in Figure \ref{fig6}.
The volatility alternates between low and high in short time intervals, so the
estimation of spot volatility requires an appropriate choice of $\kappa_n$.
\begin{figure}[H
\begin{center}
\begin{tabular}{c}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[keepaspectratio, scale=0.55]{Fig/Sim08_SamplePath10_lambda10.eps}
{(a) Sample path of $X$}
\end{center}
\end{minipage}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[keepaspectratio, scale=0.55]{Fig/Sim08_SamplePath10_diff_lambda10.eps}
{(b) Increment of $X$}
\end{center}
\end{minipage}
\end{tabular}
\caption{Sample path of $X$ and its increments}
\label{fig6}
\end{center}
\end{figure}
{\color{black}}% {\color[rgb]{0,0.5,0} Table \ref{Table03} shows the summary and average error ratios of GRV and WGRV, respectively.}
for several values of $c$ and $B$ that determine the width $\overline{\kappa}_n = 2\kappa_n + 1$ of subintervals for
spot volatility estimation.
This indicates that large $B$ and $c$ (wide subinterval) tend to give imprecise estimates.
Since the volatility varies in a wide subinterval as Figure \ref{fig6} shows,
the estimated spot volatility is prone to deviate the true value.
This leads to misdetection of jumps, and thus distorts the estimate of GRV and WGRV. For instance, an underestimated
spot volatility makes normalized increments too large, so the increments
are likely to be regarded as jumps and eliminated from calculation of the estimates. As a result, GRV and WGRV are underestimated.
In this example, it seems that small values such as $c=0.1, 0.2$ and $B = 1, 5$ are
preferable.
This example suggests that we should choose the tuning parameters $B$ and $c$ carefully,
especially when volatility switches between high and low states frequently.
After all, the proper choice of tuning parameters, such as $B$ and $c$, while observing the data in detail, is needed to obtain precise estimates by GRV and WGRV.
\renewcommand{\arraystretch}{1.5}
\begin{table}[H]
\caption{Average error ratios of GRV and WGRV for different $c$ and $B$ determining the width $\overline{\kappa}_n$}
\vspace{3mm}
\begin{minipage}{0.5\hsize}
\begin{center}
\vspace{3mm}
{(a) GRV with local GRV ({\tt grv.lgrv})}\\
\vspace{3mm}
\begin{tabular}{ccccc}\hline
\multicolumn{1}{c}{} & \multicolumn{4}{c}{$B$}\\
\cline{2-5}
$c$ & 1 & 5 & 10 & 20 \\ \hline
0.10 & 3.41 & -5.80 & -9.31 & -15.17 \\
0.20 & -1.19 & -9.31 & -15.17 & -24.28 \\
0.30 & -5.26 & -16.48 & -26.09 & -36.26 \\
0.40 & -9.31 & -27.63 & -37.64 & -44.13 \\
0.45 & -12.36 & -33.71 & -42.02 & -46.39 \\
0.49 & -15.46 & -37.92 & -44.30 & -47.37 \\ \hline
\end{tabular}
\end{center}
\end{minipage}
\hfill
\begin{minipage}{0.5\hsize}
\begin{center}
\vspace{3mm}
{(b) GRV with local minRV ({\tt grv.mrv})}\\
\vspace{3mm}
\begin{tabular}{ccccc}\hline
\multicolumn{1}{c}{} & \multicolumn{4}{c}{$B$}\\
\cline{2-5}
$c$ & 1 & 5 & 10 & 20 \\ \hline
0.10 & 6.21 & -4.21 & -6.92 & -11.85 \\
0.20 & 0.69 & -6.92 & -11.85 & -20.91 \\
0.30 & -3.85 & -13.08 & -22.75 & -33.82 \\
0.40 & -6.92 & -24.44 & -35.37 & -43.14 \\
0.45 & -9.33 & -31.01 & -40.38 & -46.01 \\
0.49 & -12.11 & -35.72 & -43.37 & -47.22 \\ \hline
\end{tabular}
\end{center}
\end{minipage} \\
\vspace{3mm}
\begin{minipage}{0.5\hsize}
\begin{center}
\vspace{3mm}
{(c) WGRV with local GRV ({\tt wgrv.lgrv})}\\
\vspace{3mm}
\begin{tabular}{ccccc}\hline
\multicolumn{1}{c}{} & \multicolumn{4}{c}{$B$}\\
\cline{2-5}
$c$ & 1 & 5 & 10 & 20 \\ \hline
0.10 & 15.27 & 1.33 & -3.69 & -9.64 \\
0.20 & 8.81 & -3.69 & -9.64 & -17.78 \\
0.30 & 1.88 & -10.83 & -19.33 & -28.89 \\
0.40 & -3.69 & -20.74 & -30.20 & -37.17 \\
0.45 & -6.97 & -26.38 & -34.80 & -39.68 \\
0.49 & -9.94 & -30.50 & -37.36 & -40.95 \\ \hline
\end{tabular}
\end{center}
\end{minipage}
\hfill
\begin{minipage}{0.5\hsize}
\begin{center}
\vspace{3mm}
{(d) WGRV with local minRV ({\tt wgrv.mrv})}\\
\vspace{3mm}
\begin{tabular}{ccccc}\hline
\multicolumn{1}{c}{} & \multicolumn{4}{c}{$B$}\\
\cline{2-5}
$c$ & 1 & 5 & 10 & 20 \\ \hline
0.10 & 22.42 & 3.91 & -1.64 & -7.54 \\
0.20 & 13.26 & -1.64 & -7.54 & -15.38 \\
0.30 & 4.77 & -8.71 & -16.88 & -26.58 \\
0.40 & -1.64 & -18.29 & -28.05 & -35.98 \\
0.45 & -4.96 & -23.98 & -33.10 & -39.12 \\
0.49 & -7.78 & -28.37 & -36.22 & -40.71 \\ \hline
\end{tabular}
\end{center}
\end{minipage}
\label{Table03}
\end{table}
\renewcommand{\arraystretch}{1.0}
}
\section{Concluding remarks}
In this paper, we construct the global realized volatility estimator
in the nonparametric context.
We proved the consistency and the asymptotic normality
of
GRV and WGRV, and, by numerical simulations, we show that these new approaches outperform
previous studies which use increments within a single or two intervals.
Our new approach for eliminating jumps is highly versatile.
For example, by normalization, it works well when the volatility of data
is driven by a nonconstant-volatility process.
Moreover, both GRV and WGRV
are accurate enough in the case of not only compound-Poisson sporadic jumps but also Neyman-Scott consecutive jumps.
The global-filtering method could be extended to the covariance estimation
even under the nonsynchronous sampling scheme.
Furthermore, this approach could also be applied to construct
a test statistic for jump.
Also, it is valuable to apply our approach to empirical research of
high-frequency time series data.
These are important topics for future research.
\bibliographystyle{spmpsci}
|
\section{Introduction}
The average motion of particles in space and time in samples in which at least the first two moments of the distribution of spacing are finite follows a diffusion equation in Brownian models. Coherent-transport phenomena of classical and
quantum waves have been also studied within Brownian approaches \cite{Beenakker-review,Mello_book}. Diffusion is suppressed as a result coherent backscattering in which waves returning to points in the medium along time-reversed paths interfere constructively. Anderson localization occurs as diffusion ceases in sufficiently large systems of dimensions $d \le 2$ and in higher dimensions above a critical value of disorder \cite{Anderson,Abrahams1979,Abrahams2009,Kramer}. In 1D, all waves are localized. The average transmission falls exponentially asymptotically, $\langle T \rangle \sim \exp{(-L/\ell)}$, while $\langle \ln T \rangle =- L/\ell$. Indeed, the full statistics of transmission in standard light-tailed distributions of separation between scattering elements is determined in accord to the single parameter scaling theory of localization in terms of the dimensionless parameter $L/\ell$~\cite{Anderson_1,Mello_book}.
Most of studies of coherent transport in random media consider standard light-tailed
distributions of disorder that lead to Anderson localization. Such systems include mesoscopic electronic systems in micron scale devices at low temperatures and classical waves in stationary media. However, heavy tailed distributions are common in biology and geology~\cite{Zaburdaev} and may lead to advantageous mesoscopic devices.
Heavy-tailed L\'evy $\alpha$-stable distribution are characterized by power-law tails. Thus, for a random variable $z$ following a L\'evy $\alpha$-stable distribution $\rho(z)$~\cite{Levy,Kolmogorov,Uchaikin}:
\begin{equation}
\label{qalpha}
\rho(z) \sim 1/z^{1+\alpha}
\end{equation}
for $z \gg 1$
and $0 < \alpha < 2$. For $\alpha < 1$, both the first and second moments diverge.
In L\'evy-type disorder, waves can travel long distances without being scattered and thus have a profound impact on the transport properties.
Measurements and analytic calculations of wave transmission in L\'evy
$\alpha$-stable media give different scaling than in standard 1D random media
~\cite{Lambert1998,Barthelemy2008, Mercadier2009, Falceto2010, Ilias2012,Fernandez2014,AntonioF2012,Asatryan2018,Barbosa2019, Razo2020,Boose, Beenakker2009,Burioni2010,Sibatov,Bouchad1990,Zaburdaev}.
In this work, we treat the energy inside heavy-tailed L\'evy disordered media.
The distinctive impacts of L\'evy $\alpha$-stable disorder and Anderson localization upon wave propagation are manifest. Potential application of novel states in L\'evy disordered media for low threshold lasing are discussed.
In particular, we investigate the statistics of waves in-
side a 1D L\'evy disordered structures via the statistics of
intensity $I(x)$ at the observation point $x$ (See Fig. 1).
\begin{figure}
\includegraphics[width=0.95\columnwidth]{fig_1_sub.jpg}
\caption{Schematic of a random waveguide with scatterers (slabs) randomly separated according to a L\'evy distribution. The scattering processes at the left and right of the observation point $x$ are described by the transfer matrices $M_l$ and $M_r$, respectively.}
\label{fig1}
\end{figure}
As we show below, for L\'evy-type disorder, the average of the intensity and its logarithm follows a power-law dependence with the observation point. We will contrast these results for disordered systems with standard disorder, which have been studied
experimentally and theoretically with random-matrix theory~\cite{Yamilov2014, Mello2015, Cheng}. Calculations of intensity
inside L\'evy disordered samples \cite{Xujun_arxiv} using the concept of leap-over to compute the density of scatterers gives different results from those presented in Fig.~\ref{fig4} (solid lines).
The intensity inside L\'evy disordered samples has been analyzed in \cite{Xujun_arxiv}
using the concept of leap-over to calculate the density of scatterers,
but discrepancies exist between the calculated average intensity and numerical simulations.
The present manuscript is organized as follows. In Section II, we present general expressions for the intensity and transmission in a single sample in terms of transfer matrices. We then introduce some known results for the statistics of the transmission of standard disordered systems that will be contrasted with L\'evy disorder in the subsequent section. In Section III, L\'evy type disorder is introduced and compared to transmission in standard systems. The results for the transmission are useful in the study of the statistics of the intensity inside the medium. The averages of the intensity and of the logarithm of intensity are given as functions of depth into the sample. Examples of the complete distribution of the logarithmic intensity are shown in Section III. A summary of the results and discussion are given in Section IV.
\section{Preliminaries}
\subsection{Transfer matrix: transmission and intensity}
Let us assume that we measure the intensity $I(x)$ at a point $x$. If $A$ and $B$ are the amplitudes of the forward and backward waves going at this point (see Fig.~ \ref{fig1}), $I(x)$ is given
by
\begin{equation}
\label{I}
I(x)=|A \exp{(ikx)}+ B \exp{(-ikx)}|^2 ,
\end{equation}
where $k$ is the wavenumber. We now introduce the transfer matrices
$M_l$ and $M_r$ associated with the segments of the sample at the left and right-hand side of the observation point, respectively:
\begin{equation}
\label{mmatrices}
M_{l(r)}= \begin{bmatrix}
\gamma_{l(r)} & \beta_{l(r)} \\
\beta_{l(r)}^* & \gamma_{l(r)}^*
\end{bmatrix} ,
\end{equation}
where $\gamma_{l(r)}$ and $\beta_{l(r)}$ are complex numbers satisfying $|\gamma_{l(r)}|^2-|\beta_{l(r)}|^2=1$. The amplitudes $A$ and $B$
can be written in terms of the transfer matrices, $M_{l(r)}$ and from Eq. (\ref{I}), the intensity is given by
\begin{eqnarray}
\label{Iofx}
I(x)&=&\frac{1}{|\gamma_r|^2}\left|\gamma_l \gamma^{*}_r+\beta_l \beta^*_r \right|^2
\left|1-\frac{\beta^*_r}{\gamma^*_r} \exp{-2ikx} \right|^2 \nonumber \\
&=& \frac{T}{T_r}\left|1-\frac{\beta^*_r}{\gamma^*_r} \exp{(-2ikx)} \right|^2 ,
\end{eqnarray}
where $T=\left|\gamma_l \gamma^{*}_r+\beta_l \beta^*_r \right|^2$ is the transmission coefficient of the entire sample and $T_r=1/|\gamma_r|^2$ is the transmission coefficient of the right segment.
The transfer matrices $M_{l(r)}$ are conveniently written in the polar representation as \cite{Mello_book}
\begin{eqnarray}
\label{mpolar}
M_{l (r)} = \begin{bmatrix}
\sqrt{1+\lambda_{l(r)}}e^{i \theta_{l(r)}} & \sqrt{\lambda_{l(r)}}e^{i(2\mu_{l(r)}-\theta_{l(r)})} \\
\sqrt{\lambda_{l(r)}}e^{-i(2\mu_{l(r)}-\theta_{l(r)})} & \sqrt{1+\lambda_{l(r)}}e^{-i \theta_{l(r)}} \nonumber \\
\end{bmatrix} ,
\end{eqnarray}
with phases $\theta_{l (r)}, \mu_{l (r)}$$\in$$[0,2\pi]$ and $\lambda_{l(r)} \ge 0$. An advantage of using the polar representation is that the radial variables $\lambda_{l(r)}$ are directly related to the transmission coefficients: $\lambda_{l(r)}=\left(1-T_{l(r)}\right)/T_{l(r)}$. Therefore Eq. (\ref{Iofx}) can be written as
\begin{equation}
\label{Iofx_polar}
I(x)= \frac{T}{T_r}\left|1-\sqrt{1-T_r} \exp{(-2i(\mu_r-\theta_r+kx))} \right|^2 ,
\end{equation}
while the total transmission $T$ is given by
\begin{eqnarray}
\frac{1}{T}&=&\frac{1}{T_r T_l}\Big( 2+T_r T_l -T_r-T_l \nonumber \\
&&+\left. 2\sqrt{(T_r-1)(T_l-1)\cos{2(\mu_l-\mu_r +\theta_r)}}\right) ,
\end{eqnarray}
\subsection{Statistics of the transmission in standard disordered 1D media}
Now that we have obtained analytical expressions for the intensity and transmission of a single sample in the previous section, we consider an ensemble
of random samples. In particular,
we assume that the disordered structures composed of
randomly separated weak scatterers or slabs. Thus, the intensity $I(x)$ is a random quantity. From Eq.~(\ref{Iofx_polar}), the statistics of $I(x)$ depend on the statistical properties of
the transmission and the angular variables $\theta_{l (r)}$ and $\mu_{l (r)}$.
Before considering the case of disordered samples with L\'evy disorder, we introduce the distribution of the transmission for standard disorder. The statistics of transmission through standard disordered systems with light-tailed distributions have been
extensively studied using random matrix theory\cite{Mello_book,Beenakker-review}. The
distribution of the transmission $p_s (T)$ is given by \cite{Molina,Kleftogiannis2013}
\begin{equation}
\label{pofT}
p_s(T)=C \frac{\left[\mathrm{acosh}(1/\sqrt{T})\right]^{1/2}}{T^{3/2}(1-T)^{1/4}}e^{-s^{-1}\mathrm{acosh}^2(1/\sqrt{T})} ,
\end{equation}
where $C$ is a normalization constant, $s=L/\ell$ with $L$ the length of the system and $\ell$ the mean free path. The complete distribution of transmission is determined by the parameter $s$, which is proportional to the number of scatterers $n$ in the sample with proportionality constant $a$: $s=an$ \cite{Mello_groups}.
For later comparisons with systems with L\'evy disorder, we point out the asymptotic exponential decay with $L$ of the average transmission in standard disordered systems \cite{Beenakker-review}:
\begin{equation}
\label{averageT}
\langle T \rangle \propto \exp{\left(-L/2\ell \right)}
\end{equation}
and the linear behavior of the average of the logarithmic transmission
\begin{equation}
\label{averagelnT}
\langle - \ln T \rangle \propto L
\end{equation}
To illustrate some statistical properties of the transmission of standard disordered systems, we show in Fig.~\ref{fig_2}(a) the distribution of the logarithmic transmission $p_s(\ln T)$, which is obtained from Eq. (\ref{pofT}) and the linear behavior of $\langle - \ln T \rangle$ with $L$, given by Eq. (\ref{averagelnT}). The histogram and symbols in Fig.~\ref{fig_2} are obtained from numerical simulations as explained next.
The numerical simulations performed in this work are based on the transfer matrix approach \cite{Markos2008,AntonioF2012}. The numerical model consists
of layers of thickness 2.5 mm with refraction index $n_2=1.1$, and reflection coefficient 0.007, randomly placed in a background of index of refraction $n_1=1$ with separations
following a Gaussian distribution for standard disorder and a L\'evy $\alpha$-stable distribution for L\'evy disorder.
We have fixed the frequency at 1 THz in all the calculations.
The statistics is collected for $10^6$ realizations of the disorder.
\section{Statistics of the intensity inside 1D media with L\'evy disorder}
We utilize the model of L\'evy disordered media introduced in ~\cite{Falceto2010}
with the asymptotic decay given in Eq. (\ref{qalpha}).
We briefly summarize the main results of Ref.~\cite{Falceto2010} for the transmission that will be useful for obtaining the statistical properties of the intensity. We will consider the case $\alpha<1$, where the effects of L\'evy disorder on transport are strong.
\begin{figure}
\includegraphics[width=\columnwidth]{fig_2_sub.jpg}
\caption{(a) The distribution of the logarithmic transmission for standard disorder systems. The distribution is determined by the parameter $s=\langle -\ln T \rangle=8$. Inset: linear behavior of $\langle \ln T \rangle$ with the system length. (b) The distribution
of the logarithmic transmission for L\'evy disordered systems. The distribution is determined by the parameters $\langle -\ln T \rangle=8$ and $\alpha=1/2$. Inset: the power-law behavior of $\langle \ln T \rangle$ with the system length, Eq.~ (\ref{lngofl_a}).}
\label{fig_2}
\end{figure}
In L\'evy disordered samples of fixed length $L$, the number of scattering units
$n$ is a random variable with strong sample-to-sample fluctuations; thus, it is crucial to know the complete distribution of $n$. The probability density
$\Pi_L(n;\alpha)$ of these fluctuations is given by \cite{Falceto2010}
\begin{equation}
\label{pi_alpha}
\Pi_{L}(n;\alpha)=\frac2\alpha\frac{L}{(2n)^{\frac{1+\alpha}{\alpha}}}
q_{\alpha,c}\left({L}/{(2n)^{1/\alpha}}\right) ,
\end{equation}
for $0 < \alpha <1 $, in the limit $L\gg c^{1/\alpha}$ with $c$ a scaling parameter. The probability density $q_{\alpha,c}(x)$ has a power-law tail: $q_{\alpha,c}(x) \sim c/x^{1+\alpha}$ for $x \gg 1$.
Using the distribution of the transmission for a fixed number of scatterers $p_{s}(T)$ given in Eq. (\ref{pofT}) and the distribution $\Pi_L(n,\alpha)$, Eq. (\ref{pi_alpha}), we write the distribution of the transmission for
L\'evy disordered systems $p_{\alpha,\xi}(T)$ as
\begin{eqnarray}
\label{pofT_alphaxi}
P_{\alpha,\xi}(T)=\int_0^\infty p_{s(\alpha,\xi,z)}(T) q_{\alpha,1}(z){\rm d}z ,
\end{eqnarray}
where $p_{s(\alpha,\xi,z)}(T)$ is given by Eq. (\ref{pofT}) with $s$ replaced by
$s(\alpha,\xi,z)={\xi}/(2{z^\alpha \mathcal{I}_\alpha)}$ and $\mathcal{I_{\alpha}}$ is half of the mean value
$\langle z^{-\alpha} \rangle$: $\mathcal{I}_\alpha=(1/2)\int z^{-\alpha}q_{\alpha,1}dz=\cos(\pi\alpha/2)/2\Gamma(1+\alpha)$, where $\Gamma$ denotes the Gamma function \cite{Koren}. The parameter $\xi$ introduced in Eq. (\ref{pofT_alphaxi}) is the average of the logarithmic transmission for a fixed length $L$: $\xi = \langle -\ln T \rangle_L = \int_0^{\infty} an \Pi_L(n)dn$, which is given by
\begin{equation}
\label{lngofl_a}
\langle \ln T\rangle_L = -\left( \frac{a}{c} \mathcal{I}_\alpha \right) L^\alpha,
\end{equation}
Since the factor in parentheses in Eq. (\ref{lngofl_a}) is a constant,
the average of the logarithmic transmission is a power-law function of $L$, in contrast to the linear dependence for standard disorder in Eq.~(\ref{averagelnT}). Similarly, a power-law is found for the average transmission \cite{AntonioF2012} $\langle T \rangle \propto L^{-\alpha}$ in contrast to the exponentially decay for standard disordered systems in Eq.~(\ref{averageT}).
In Fig.~\ref{fig_2}(b), we show an example of the distribution of the logarithm of transmission for L\'evy disordered structures characterized by $\alpha=1/2$. The theoretical results as given in Eq.~(\ref{pofT_alphaxi}) are compared to numerical simulations show as the histogram. The power-law behavior
of $\langle \ln T \rangle$ in L\'evy structures is shown in the inset
in Fig. \ref{fig_2}(b). Thus, by comparing Fig.~ \ref{fig_2}(a) and \ref{fig_2}(b), the
strong influence of L\'evy disorder on the statistical properties of the transmission are clearly seen.
We also note that the only parameters that enter into Eq.~(\ref{pofT_alphaxi}) are $\alpha$ and
$\xi=\langle \ln T \rangle$. Thus, the complete statistics of the transmission is determined by these parameters.
With the above results for the statistics of transmission, we now study the statistical properties of the intensity.
As we show next, the presence of L\'evy disorder is revealed in basic statistical quantities such as the ensemble averages $\langle \ln I(x) \rangle_L$ and $\langle I(x) \rangle_L$.
Let us consider first the average of the logarithmic intensity, $\langle \ln I(x) \rangle$. The calculations are lenghty but a simple analytical result can be provided. This quantity is of particular importance since it is directly related to
the average of the logarithmic transmission, which along $\alpha$ determines all the statistical properties of the transport in L\'evy disordered systems.
We perform the average over the uniformly distributed phases in Eq.~(\ref{Iofx_polar}) to obtain
$\langle \ln I(x) \rangle = \langle \ln T \rangle_L - \langle \ln T_r \rangle_{L-x}$. Since $\ln T$ is an additive quantity, we obtain $\langle \ln T \rangle_{L-x}=\langle \ln T \rangle_L - \langle \ln T \rangle_{x}$, and
from Eq.~(\ref{pi_alpha}), we have
\begin{eqnarray}
\label{logofIx}
\langle \ln I(x) \rangle &=& - \left( \frac{a}{c} \mathcal{I}_\alpha \right) x^\alpha \nonumber \\
&=& \langle \ln T \rangle_L \left( \frac{x}{L}\right)^\alpha .
\end{eqnarray}
Thus, $ \langle \ln I(x) \rangle \propto x^\alpha$ has a power-law behavior in L\'evy disordered media, in contrast
to the linear dependence in standard disordered media~\cite{Cheng}.
We verify numerically the result given in Eq. (\ref{logofIx}). The results (symbols) are shown in Fig.~\ref{fig3}(a) together with the theoretical results (solid lines) from Eq.~(\ref{logofIx}) for $\alpha=1/2$ and 3/4. The linear behavior of $ \langle \ln I(x) \rangle$ for standard disorder is also shown in Fig.~\ref{fig3}(a) (green squares).
The $\langle \ln I(x) \rangle$ does not fall linearly in L\'evy structures as it does in standard (Fig.~\ref{fig3}(a)), however,
the role of coherent backscattering in inhibiting propagation is unchanged. To gain
an insight into this nonlinear behavior, we note that $\langle \ln I(x) \rangle$ is given by the difference between average of the logarithmic transmission of the complete sample ($L$) and the right segment ($L-x$), as we have shown above.
The power-law behavior of $\langle \ln I(x) \rangle$ finds its origin in the power-law of the variation of the number of scatterers up to the depth $x$. This is illustrated in Fig. \ref{fig3}(b), where the average number of scatterers $\langle n \rangle=\int n \Pi_x(n)dn=\mathcal{I_\alpha}x^\alpha/c$ is plotted at the position $x/L$ for L\'evy ($\alpha=1/2, 3/4$) and standard disordered structures. For L\'evy disordered samples, the average number of scatterers follows a power law with exponent $\alpha$. In contrast, in standard disorder, the average number of scatterers is a linear function with the system size. Thus, for both L\'evy and standard disorder, $\langle \ln I \rangle$ is a linear function of $\langle n \rangle$, as shown in Fig.~\ref{fig3}(c) and $\langle \ln I(x) \rangle$ is additive, as in standard disordered structures \cite{Cheng,Yiming}.
\begin{figure}
\includegraphics[width=\columnwidth]{fig_3_sub.jpg}
\caption{(a) Average of the logarithmic intensity for L\'evy disordered systems with $\alpha=1/2$ and 3/4 (red and blue symbols). The solid lines are given by Eq. (\ref{logofIx}). For comparison, the linear dependence with $x$ in the case of standard disorder is also shown (green). In all cases, $\langle -\ln T \rangle_L=5$. (b) The average number of slabs $\langle n\rangle$ in the L\'evy ($\alpha=1/2, 3/4$) and standard disordered structures at the position of observation $x$. (c) $\langle \ln I \rangle$ as a function of $\langle n \rangle$ for $\alpha=1/2, 3/4$, and standard disorder.}
\label{fig3}
\end{figure}
We now study the ensemble average intensity $\langle I(x) \rangle$, after averaging $I(x)$, Eq. ~(\ref{Iofx}), over the uniformly distributed random phases, we write $\langle I(x) \rangle$ as~\cite{Mello2015}
\begin{eqnarray}
\label{averageIx}
&& \langle I(x) \rangle = \nonumber \\ && \int_0^1 \int_0^1 \frac{T_l\left(2-T_r \right)}{T_l+T_r-T_lT_r} P_{\alpha,\xi_l}(T_l) P_{\alpha,\xi_r}(T_r) dT_l
dT_r , \nonumber \\
\end{eqnarray}
where the distributions $P_{\alpha,\xi_l}(T_l)$ and $P_{\alpha,\xi_r}(T_r)$
are the distributions for the left and right segments, respectively, with
$\xi_l =\langle \ln T_l \rangle= (x/L)^\alpha \langle \ln T \rangle_L$. For the right segment, $P_{\alpha,\xi_r}(T_r)$
$\xi_r =\langle \ln T_r \rangle= (1-(x/L)^\alpha) \langle \ln T \rangle_L$, according to Eq. (\ref{lngofl_a}). We can verify the particular cases at $x=0$ and $x=L$: for $x=0$, $T_l=1$ and $\langle I(0) \rangle=\langle (2-T_r) \rangle=2-\langle T \rangle$, while at $x=L$, $T_r=1$ and therefore
$\langle I(L) \rangle= \langle T_r \rangle =\langle T \rangle$.
We perform numerical simulations to support Eq.~(\ref{averageIx}), where the double integral is performed numerically. The distribution for the left segment $P_{\alpha,\xi_l}(T_l)$ in Eq.~(\ref{averageIx}) is obtained from Eq.~(\ref{pofT_alphaxi}), while for the right segment,
$P_{\alpha,\xi_r}(T_r)$ is obtained by considering the corresponding
probability density of scatterers which is generated numerically.
The results are shown in Fig. \ref{fig4}(a) for $\alpha=1/2$ and 3/4. The numerical simulations and Eq.~(\ref{averageIx}) are seen in good agreement. The average intensity for standard disorder is also shown (green dots and solid line) in Fig \ref{fig4}(a) to provided a contrast with the power law dependence found in media with L\'evy disorder.
\begin{figure}
\includegraphics[width=\columnwidth]{fig_4_sub.jpg}
\caption{(a) Ensemble average intensity $\langle I \rangle$ for L\'evy $\alpha=1/2$ and 3/4 and standard disorder. The solid lines are obtained from Eq.~(\ref{averageIx}). In all these 3 cases $\langle -\ln T \rangle=1$. (b) The average
$\langle I \rangle$ for $\alpha=1/2$, as in (a), but for waves incident from the left and right incidences. The black squares show that the sum of both incidences is constant.}
\label{fig4}
\end{figure}
The profile of $\langle I(x) \rangle$ is not symmetric about the center, as it is in standard homogeneously disordered systems. See Fig. \ref{fig4}(a).
In L\'evy disordered structures, the disorder is inhomogeneous; the density of scatterers is greatest near the left side of the sample, as can be seen in Fig. \ref{fig3}(b), where waves launched causing the intensity to fall more rapidly there.
However, the sum of intensities for waves incident from the left and right is constant and equal to twice the intensity of the incident beam from one side, as shown in Fig. \ref{fig4}(b). This can be understood by noting that the integrand
of Eq.~(\ref{averageIx}) for the wave incident from the right, $T_l$ and $\xi_l$ are replaced by
$T_r$ and $\xi_r$, respectively, and similarly, $T_r$ and $\xi_r$ are replaced by $T_l$ and $\xi_l$ for waves incident from the right. Adding the contributions from the wave incident from the left and right gives $T_l(2-T_r)/(T_l+T_r-T_l T_r)+T_r(2-T_l)/(T_l+T_r-T_l T_r)=2$. Since $P_{\alpha,\xi_{l,(r)}}(T_{l,(r)})$ are normalized, the average of the sum of intensities excited by waves incident on the left and right is equal to 2. This result is illustrated in Fig. \ref{fig4} (b) for the case of $\alpha=1/2$, where the black squares are the sum of the averages intensity for incident waves from the left and right ends of the samples.
The previous discussion is general and gives an average intensity profile for standard homogeneously disordered media which is symmetric about the center of the sample ~\cite{Yiming}. This symmetry is summarized by the expression
$\langle I(x) \rangle +
\langle I(L-x) \rangle =2$ and reflects the fact that the sum of intensity from the right and left is equal to the local density of states (LDOS) relative to LDOS outside the medium, which is unchanged by disorder.
The complete distribution of the logarithmic is obtained numerically and shown in Fig.~\ref{fig5} for $\langle \ln T \rangle = 10$. There is a higher probability of large fluctuations of intensity in L\'evy disordered samples (blue and red histograms) as compared to standard disordered systems (green histogram).
\begin{figure}
\includegraphics[width=\columnwidth]{fig_5_sub.jpg}
\caption{Distribution of the logarithmic intensity for L\'evy and standard disordered structures. $\langle -\ln T \rangle=10$.}
\label{fig5}
\end{figure}
\section{Summary and discussion}
We have studied the wave intensity statistics inside random 1D media with disorder described by L\'evy type distributions characterized by an asymptotic
power-law decay. For both $\langle I(x) \rangle$ and $\langle \ln I(x) \rangle$,
we find a power-law decay with position.
In contrast, $\langle I(x) \rangle$ falls linearly in standard disordered systems.
The slower decay with $x$ than for the standard disorder indicates that
wave localization in space is weaker in L\'evy disorder than in standard disorder.
The equivalence of the statistics of intensity in $\alpha$-stable L\'evy disordered systems and standard random media at the corresponding layer number $n$, suggests opportunities for engineering structures for analyzing and controlling waves. The forward and background amplitudes within a layer of the medium are constant so that they interfere and create an oscillatory pattern with high peak intensity within the layer. When a thick layer is near the spatial and spectral peak of a mode, the lifetime of quasi-normal mode increases and the line narrows
as the thickness of the layer increases. The material could therefore serve as a filter. If gain is introduced into this system, the correspondingly long lifetime of the mode would enhance the opportunity for emitted photons to stimulate emission before escaping the sample~\cite{Milner2005}. In addition, the large spatial extent of the mode allows the system to be efficiently pumped without saturating the gain medium.
The prospects for an $\alpha$-stable laser will be considered in future work.
\begin{acknowledgments}
We acknowledge discussions with Xujun Ma.
This work is supported by the National Science Foundation under EAGER Award No. 2022629 and by PSC-CUNY under Award No. 63822-00 51, and by MCIU (Spain) under the Project number PGC2018-094684-B-C22. L.A.R.-L. thanks UCA$^{\rm{JEDI}}$ for the award of pre-doctoral studies fellowship. .
\end{acknowledgments}
$^\dagger$ Email: <EMAIL>
\bibliographystyle{unsrt}
|
\section{Introduction}
How to promote a spectral triple on an algebra to a spectral triple on a crossed product $C^*$-algebra has been the subject of various papers \cite{Skalski,Paterson,BMR,GaGr,IoMa}; the same has been recently done for the structure of compact quantum metric spaces \cite{KaKy}.
The aim of this paper is to tackle the following question: is it possibile to extend the construction of a spectral triple on a crossed product $C^*$-algebra based on a spectral triple on the base algebra to the case of crossed products with a single endomorphism?
Even though we do not have yet a general answer to this problem, we are able to propose a procedure - some steps of which can be completely described, while for others we can give several examples - which explains what we expect to be the general case.
Before describing our plan, we draw attention to a feature of our construction, namely we more or less explicitly assume that our endomorphism is in a sense expansive, a notion which has been often considered both in the commutative and in the noncommutative case, see e.g. \cite{DGMW}. Such property has important consequences: the compact resolvent property for the Dirac operator forces the spectral triple on the crossed product to be semifinite, and the bounded commutator property requires a reduction of the crossed product $C^*$-algebra, namely a new definition of crossed product by an endomorphism.
Indeed, even though there are now various notions of crossed product of a $C^*$-algebra with an endomorphism, see e.g. \cite{Murphy,Exel,KwLe}, we essentially follow a path outlined by Cuntz \cite{Cuntz} and then further developed by Stacey \cite{StaceyCrossed}, but we are forced to adapt it to the case of expansive endomorphisms.
According to Cuntz, given a $C^*$-algebra ${\mathcal A}$ together with a unital injective endomorphism $\alpha$, one constructs a direct system of $C^*$-algebras ${\mathcal A}_n$ with endomorphisms $\alpha_n$, whence the direct limit $C^*$-algebra ${\mathcal A}_\infty$ is obtained. The key point is that the endomorphism $\alpha$ of ${\mathcal A}$ becomes an automorphism $\alpha_\infty$ on ${\mathcal A}_\infty$, so that one may define the crossed product ${\mathcal A}\rtimes_\alpha{\mathbb N}$ as the crossed product ${\mathcal A}_\infty\rtimes_{\alpha_\infty}{\mathbb Z}$. Let us note that this definition gives back the original algebra when $\alpha$ is an automorphism.
The first and second step of our construction have been studied in \cite{AiGuIs01,AGI3}, where one assumes that a spectral triple ${\mathcal T}$ on ${\mathcal A}$ is given.
Let us observe that unital injective endomorphisms of a $C^*$-algebra ${\mathcal A}$ can be seen as noncommutative self-coverings of the underlying noncommutative space; the first step is then to endow any of the $C^*$-algebras ${\mathcal A}_n$ described above with a spectral triple ${\mathcal T}_n$ which makes the self-covering locally isometric or, equivalently, such that the Lip-norms induced by the Dirac operators are compatible with the connecting maps (this property can and will be weakened in some cases, cf. Section \ref{UHF}). This means that the sequence of covering spaces consists of dilated copies of the original space. This request is the reason of the expansivity mentioned above. Even if we do not give a general procedure for this step, this is not a difficult task in all the examples considered in \cite{AiGuIs01,AGI3}.
The second step consists of constructing a spectral triple ${\mathcal T}_\infty$ on the direct limit ${\mathcal A}_\infty$ which is in some sense naturally associated with the original spectral triple on ${\mathcal A}$. We note here that the algebra ${\mathcal A}_\infty$ can be naturally seen as the solenoid algebra associated with the pair (${\mathcal A},\alpha$), see \cite{AiGuIs01,AGI3,DGMW,LP2} for related constructions. In the abelian case, an intrinsic notion of solenoid, called compact universal cover, has been studied in \cite{Plaut} in great generality.
Coming back to ${\mathcal T}_\infty$, we wish to define it as a suitable limit of the triples ${\mathcal T}_n$ on ${\mathcal A}_n$. This step is far from being obvious, firstly because there is no general procedure to define a limit of a sequence of spectral triples (however in some circumstances one may follow \cite{FloGho}), secondly because the situations we consider are quite different, ranging from regular coverings associated with an action of an abelian group to (possibly ramified) coverings with trivial group of deck transformations. Examples illustrating this step are contained in \cite{AiGuIs01,AGI3} and briefly described below. In all cases, the coverings becoming wider and wider, the spectra of the Dirac operators turn more and more closely packed, so that the limit has no longer compact resolvent. However, a corresponding rescaling of the traces gives rise to a (semicontinuous semifinite) trace on a suitable $C^*$-algebra ${\mathcal B}$ of geometric operators, which contains ${\mathcal A}_\infty$ and the resolvents of the limiting Dirac operator, finally producing a semifinite spectral triple on ${\mathcal A}_\infty$. This means in particular that the semifiniteness property is true already at the level of ${\mathcal A}_\infty$, therefore determines the analogous semifiniteness property for the spectral triple on the crossed product.
The third and final step, which is the main object of this paper, consists in defining a new kind of crossed product of a $C^*$-algebra w.r.t. an endomorphism, which can be seen as a variant of the crossed product considered by Cuntz in \cite{Cuntz} and Stacey in \cite{StaceyCrossed}, and which turns out to be tailored to accomodate a spectral triple in the case of expansive endomorphisms.
The notion of this new crossed product with an endomorphism is given in Definition \ref{1.1}. On the one hand it is a universal object, therefore defines a unique object up to isomorphisms, on the other hand, as shown in Theorem \ref{prop:crossedProd}, it coincides with a reduction by a projection of the $C^*$-algebra crossed product defined in \cite{StaceyCrossed}, Proposition 1.13. While the latter is nothing else than the crossed product of ${\mathcal A}_\infty$ with ${\mathbb Z}$ w.r.t. $\alpha_\infty$, our notion can be considered as the crossed product of ${\mathcal A}_\infty$ with ${\mathbb N}$ w.r.t. $\alpha_\infty$.
Indeed, while Stacey crossed product with an endomorphism reduces to the usual crossed product for an automorphism $\alpha$, ours produces a ``corner'' of it, in such a way that only positive powers of $\alpha$ are implemented.
The advantage of such a choice is to allow the weakening of the request of metric equicontinuity (Lip-boundedness in our paper) of \cite{Skalski}, which, for an action $\alpha$ of ${\mathbb Z}$ and a Lipschitz element $a$ reads $\displaystyle \sup_{n\in{\mathbb Z}} L(\alpha^{-n}(a)) < \infty$ and makes sense for automorphisms, to a condition on $\alpha$ that we call Lip-semiboundedness, namely $\displaystyle \sup_{n\in{\mathbb N}} L(\alpha^{-n}(a)) < \infty$.
More precisely, in Section \ref{subsec:CrossedProd}, we first generalize the construction of a spectral triple on a crossed product described in \cite{Skalski} to the case of a semifinite spectral triple, maintaining the Lip-boundedness assumption, and then modify it by replacing the crossed product of Cuntz-Stacey with our crossed product, and noting that in this case the request of the endomorphism being Lip-semibounded
is sufficient to guarantee the bounded commutator property of the spectral triple, cf. Theorem \ref{triple-cross-prod-N}. Moreover, such theorem shows that the metric dimension of the crossed product spectral triple equals the metric dimension of the base triple increased by 1.
On the same grounds, a theory for the action of suitable semigroups (e.g. ${\mathbb N}^k$) can be established, but this will not be discussed here.
In the last section of this paper we show that the self-coverings considered in \cite{AiGuIs01,AGI3} satisfy the Lip-semiboundedness condition, hence give rise to a semifinite spectral triple on the crossed product.
The first example deals with the self-covering of a $p$-torus, which is a regular covering. Given a purely expanding integer valued matrix $B$, the covering projection goes from ${\mathbb R}^p/B{\mathbb Z}^p$ to ${\mathbb R}^p/{\mathbb Z}^p$ and the canonical Dirac operator on the covering makes the covering projection locally isometric. A natural embedding of the $C^*$-algebra ${\mathcal A}_n$ in ${\mathcal B}({\mathcal H}_0)\otimes M_{{\molt}^n}({\mathbb C})$ gives rise to the embedding of the direct limit $C^*$-algebra ${\mathcal A}_\infty$ in ${\mathcal B}({\mathcal H}_0)\otimes \mathrm{UHF}_{\molt}$, which is the algebra ${\mathcal B}$ mentioned above, where $r=|\det(B)|$
and $\mathrm{UHF}_r$ denotes the infinite tensor product of $M_r({\mathbb C})$.
Moreover, the Dirac operators $D_n$ converge in the norm resolvent sense to a Dirac operator affiliated with ${\mathcal B}({\mathcal H}_0)\otimes \mathrm{UHF}_{\molt}$. This structure produces a semifinite spectral triple on ${\mathcal A}_\infty$, as shown in \cite{AiGuIs01}. Theorem \ref{teo-p-toro} shows that the condition of Lip-semiboundedness is satisfied, hence we get a semifinite spectral triple on our crossed product with ${\mathbb N}$.
The second example treats the case of regular noncommutative coverings of the rational rotation algebra with abelian group of deck transformations as defined in \cite{AiGuIs01}. The procedure and the results are essentially the same as the previous example, but the condition $r \equiv_q \pm 1$ has to be further assumed in order to get a self-covering.
The third example concerns the UHF-algebra with the covering map given by the shift endomorphism and the spectral triple described in \cite{Chris}. In this case the Lip-norms given by the spectral triples are not compatible, namely $\|[D_n,\alpha^n(a)]\|\ne\|[D_0,a]\|$ for $a$ Lipschitz in ${\mathcal A}_0$, however $\|[D_{n+p},\alpha^p(a)]\|$ is bounded in $p$ (indeed converges) for any Lipschitz element in ${\mathcal A}_n$. Again we show that the condition of Lip-semiboundedness is satisfied, cf. Theorem \ref{UHFcrossedprod}.
The fourth and last example describes the crossed product associated with a ramified covering of the fractal called Sierpi\'nski gasket. Such covering is not given by an action of a group of deck transformations. Here the spectral triple on ${\mathcal A}$ is the one described in \cite{GuIs16}, and the spectral triples on ${\mathcal A}_n$ make the covering maps locally isometric. The $C^*$-algebra ${\mathcal B}$ containing both ${\mathcal A}_\infty$ and the resolvents of $D_\infty$ is an algebra of geometric operators acting on the $\ell^2$ space on the edges of the infinite Sierpi\'nski gasket with one boundary point \cite{Tep}. The proof of the condition of Lip-semiboundedness is contained in Theorem \ref{teo-gasket}.
In all cases, by Theorem \ref{triple-cross-prod-N}, the spectral triples are finitely summable and their metric dimension is equal to the metric dimension of ${\mathcal T}$ plus 1, namely it is the sum of the metric dimension of ${\mathcal T}$ and the growth of ${\mathbb N}$.
Finally, we mention that even though in all of our examples the functional given by the norm of the commutator with the Dirac operator is a Lip-norm in the sense of Rieffel \cite{Rieffel99} on ${\mathcal A}$, such property does not hold for the spectral triple on the crossed product. In fact any distance on the state space of a unital $C^*$-algebra inducing the weak$^*$-topology should necessarily be bounded, and this is not the case for our construction. The reason is that the expansiveness of the endomorphism $\alpha$ produces larger and larger (quantum) covering spaces and eventually an unbounded solenoid space. This property leads to an analogous unboundedness for the distance on the state space of the crossed product $C^*$-algebra.
\section{Crossed products for $C^*$-algebras}
\subsection{Preliminaries}
\textbf{Inductive limit}. We begin by recalling the construction of the inductive limit $C^*$-algebra, due to Takeda \cite{Takeda}, for the particular case of interest in this paper, to fix some notation. Let ${\mathcal A}$ be a unital $C^*$-algebra, $\alpha\in \End(A)$ an injective, unital $*$-endomorphism. Consider the following inductive system
\begin{equation} \label{eq:CstarIndLim1}
\begin{CD}
A_0 @ > \varphi_0 >> A_1 @ > \varphi_1 >> \cdots
\end{CD}
\end{equation}
where, for all $n\in{\mathbb N}=\{0,1,2,\ldots\}$, $A_n={\mathcal A}$, $\varphi_n=\alpha$, and define, for $m< n$, $\varphi_{nm} : A_m \to A_n$ by $\varphi_{nm} := \varphi_{n-1}\circ \cdots \circ \varphi_m \equiv \alpha^{n-m}$,
and $\varphi_{mm}:=\id$. Consider the direct product $\prod_{n=0}^\infty A_n$, with pointwise operations, and set
$$
A_\infty := \left\{ (a_n) \in \prod_{n=0}^\infty A_n : \exists m\in{\mathbb N} \textrm{ such that } a_n=\varphi_{nm}(a_m)= \alpha^{n-m}(a_m), n\geq m \right\}/\!\sim \; ,$$
where $(a_n)\sim(b_n) \iff a_n=b_n$ for all large enough $n$. Then, $A_\infty$ is a $^*$-algebra. For all $n\in{\mathbb N}$, define $\varphi_{\infty n}: a\in A_n\mapsto \varphi_{\infty n}(a)\in A_\infty$, where $\varphi_{\infty n}(a) \equiv (a_k)$, and
$$
a_k :=
\begin{cases}
0, & k<n,\\
\varphi_{kn}(a)=\alpha^{k-n}(a), & k\geq n.
\end{cases}
$$
We can introduce a norm $p$ on $A_\infty$ given by
$$
p(a):= \limsup\limits_{n\to\infty} \Arrowvert \varphi_{nm}(a_m) \Arrowvert = \norm{a_m}\; ,
$$
if $a=\varphi_{\infty m}(a_m)$, which is independent of the representative, and is a $C^*$-norm. Upon completion, we get the desired inductive limit $C^*$-algebra, which is denoted ${\mathcal A}_\infty \equiv \varinjlim A_n$.
\smallskip
\textbf{Crossed product}. Let us recall the definition of the crossed product by an automorphism, in the case of unital $C^*$-algebras, to fix some notation.
Let ${\mathcal A}$ be a unital $C^*$-algebra, $\alpha \in \Aut({\mathcal A})$ an automorphism. Denote by $C_c({\mathcal A},{\mathbb Z},\alpha)$ the $^*$-algebra of functions $f: {\mathbb Z} \to {\mathcal A}$ with finite support, pointwise addition and scalar multiplication, with product $(fg)(n):= \sum_{k\in{\mathbb Z}} f(k) \alpha^k(g(n-k))$, and involution $f^*(n):= \alpha^n(f(-n)^*)$, $f,g\in C_c({\mathcal A},{\mathbb Z},\alpha)$, $n\in{\mathbb Z}$. Define a norm on $C_c({\mathcal A},{\mathbb Z},\alpha)$ by $\Arrowvert f \Arrowvert_1 := \sum_{n\in{\mathbb Z}} \Arrowvert f(n)\Arrowvert$, and denote by $\ell^1({\mathcal A},{\mathbb Z},\alpha)$ the Banach $^*$-algebra obtained by completing $C_c({\mathcal A},{\mathbb Z},\alpha)$ with respect to this norm. A different description of $\ell^1({\mathcal A},{\mathbb Z},\alpha)$
is obtained by introducing the functions $\dd_n(k):= \delta} \def\D{\Delta_{k, n}$. Then, $\ell^1({\mathcal A},{\mathbb Z},\alpha)$ is the set of all sums $\sum_{n\in{\mathbb Z}} a_n \dd_n$, with $a_n\in{\mathcal A}$, for all $n\in{\mathbb Z}$, and $\sum_{n\in{\mathbb Z}} \Arrowvert a_n\Arrowvert <+\infty$. Let now $\pi$ be a representation of ${\mathcal A}$ on ${\mathcal H}$, $V$ a unitary operator on ${\mathcal H}$, such that $\pi(\alpha(a)) = V \pi(a) V^*$, $a\in{\mathcal A}$. The triple $({\mathcal H}, \pi,V)$ is called a covariant representation of $({\mathcal A},\alpha)$. Then, the integrated form of $({\mathcal H}, \pi,V)$ is the representation $\pi\rtimes V$ of $C_c({\mathcal A},{\mathbb Z},\alpha)$ on ${\mathcal H}$ given by
\begin{eqnarray} \label{IntegratedForm}
\pi\rtimes V(\sum_{n\in{\mathbb Z}} a_n\dd_n) &:=& \sum_{n\in{\mathbb Z}} \pi(a_n) V^n.
\end{eqnarray}
It can be proved (\cite{Ped} Proposition 7.6.4) that there is a bijection between the set of non-degenerate covariant representations $({\mathcal H}, \pi,V)$ of $({\mathcal A},\alpha)$ on a Hilbert space ${\mathcal H}$, and the set of non-degenerate continuous representations of $\ell^1({\mathcal A},{\mathbb Z},\alpha)$ on ${\mathcal H}$. Define the universal representation $\ppuniv$ of $\ell^1({\mathcal A},{\mathbb Z},\alpha)$ to be the direct sum of all non-degenerate continuous representations of $\ell^1({\mathcal A},{\mathbb Z},\alpha)$ on Hilbert spaces. The crossed product of ${\mathcal A}$ by the action $\alpha$ of ${\mathbb Z}$ is the $C^*$-algebra ${\mathcal A}\rtimes_\alpha{\mathbb Z}$ obtained as the norm closure of $\ppuniv(\ell^1({\mathcal A},{\mathbb Z},\alpha))$.
\smallskip
\textbf{Reduced crossed product}. Since ${\mathbb Z}$ is an amenable group, a different description (\cite{Ped}, 7.7.7) of the crossed product (called the reduced crossed product, in the case of non amenable groups) can be given. Let $\pi$ be
a faithful, non-degenerate representation
of ${\mathcal A}$ on ${\mathcal H}$, set $\widetilde{\mathcal H} := \ell^2({\mathbb Z},{\mathcal H}) \equiv \{ \xi : {\mathbb Z}\to{\mathcal H} | \sum_{n\in{\mathbb Z}} \Arrowvert \xi(n) \Arrowvert^2 <+\infty \}$, and, for $n\in{\mathbb Z}$, $a\in{\mathcal A}$, $\xi\in\widetilde{\mathcal H}$,
\begin{align*}
(\uu \xi)(n) := \xi(n-1), \qquad (\widetilde{\pi}(a)\xi )(n) := \pi( \alpha^{-n}(a) )(\xi(n)).
\end{align*}
Observe that, $\widetilde{\pi}(\alpha(a))= \uu\widetilde{\pi}(a) \uu^* $, $a\in {\mathcal A}$.
Therefore, $(\widetilde{\mathcal H}, \widetilde{\pi},\uu)$ is a covariant representation of $({\mathcal A},{\mathbb Z},\alpha)$, and the representation $\widetilde{\pi} \rtimes \uu$ is called a regular representation of $\ell^1({\mathcal A},{\mathbb Z},\alpha)$. In particular, if $a = \sum_{n\in{\mathbb Z}} a_n \dd_n \in C_c({\mathcal A},{\mathbb Z},\alpha)$, then $(\widetilde{\pi} \rtimes \uu (a) \xi )(n) = \sum_{k\in{\mathbb Z}} \pi( \alpha^{-n}(a_k)) (\xi(n-k))$, $n\in{\mathbb Z}$.
Define the universal regular representation $\lambda} \def\La{\Lambda_u$ of $\ell^1({\mathcal A},{\mathbb Z},\alpha)$ to be the direct sum of all regular representations of $\ell^1({\mathcal A},{\mathbb Z},\alpha)$ on Hilbert spaces. The (reduced) crossed product of ${\mathcal A}$ by the action $\alpha$ of ${\mathbb Z}$ is the $C^*$-algebra obtained as the norm closure of $\lambda} \def\La{\Lambda_u(\ell^1({\mathcal A},{\mathbb Z},\alpha))$. Observe that (\cite{Ped}, 7.7.4), if $\pi_u$ is the universal representation of ${\mathcal A}$, then ${\mathcal A}\rtimes_\alpha{\mathbb Z}$ coincides with the norm closure of $\widetilde{\pi_u}\rtimes \uu (\ell^1({\mathcal A},{\mathbb Z},\alpha))$.
Therefore, we get ${\mathcal A} \rtimes_{\alpha} {\mathbb Z} = \langle\, \widetilde{\pi_u}({\mathcal A}),\uu \,\rangle$, where $\langle\, \widetilde{\pi_u}({\mathcal A}),\uu \,\rangle$ stands for the $C^*$-algebra generated by $\widetilde{\pi_u}({\mathcal A})$ and $\uu$.
\smallskip
\textbf{Lift of a spectral triple to a crossed product}.
First of all we recall the definition of spectral triples.
\begin{definition}\label{deftripla}
An odd spectral triple $({\mathcal L},{\mathcal H},D)$ consists of a Hilbert space ${\mathcal H}$,
an algebra ${\mathcal L}$ acting (faithfully) on it,
a self adjoint operator $D$ on the same Hilbert space such that $a\dom(D) \subset \dom(D)$ and $[D, a]$ is bounded for any $a\in{\mathcal L}$, and with $D$ having compact resolvent.
A spectral triple is said to be even if there exists a self-adjoint unitary operator $\G$ such that $\pi(a)\G = \G\pi(a)$, $\forall a\in{\mathcal A}$, and $D\G=-\G D$.
\end{definition}
In \cite{BMR}, Bellissard, Marcolli and Reihani show how to lift a spectral triple from a unital C*-algebra ${\mathcal A}$, endowed with an automorphism $\alpha$, to the crossed product ${\mathcal A}\rtimes_\alpha {\mathbb Z}$. Their setting is generalised in (\cite{Skalski}, Theorem 2.8) to the case of the action of a discrete group. In the particular case of an automorphism, one obtains
\begin{definition}\label{equi}
Let ${\mathcal A}$ be a unital C*-algebra, $\alpha\in \Aut({\mathcal A})$ a unital automorphism, $({\mathcal L},{\mathcal H},D)$ a spectral triple on ${\mathcal A}$ such that $\alpha({\mathcal L})\subset{\mathcal L}$.
The automorphism is said to be Lip-bounded if
$$
\sup_{n\in{\mathbb Z}} \| [D,\alpha^{-n}(a)] \| < \infty, \qquad \forall a\in{\mathcal L}.
$$
\end{definition}
The previous notion was introduced in \cite{Skalski} where it is called the metric equicontinuity of the action.
\begin{theorem}
Let ${\mathcal A}$ be a unital C*-algebra, $({\mathcal L},{\mathcal H},D)$ an odd spectral triple on ${\mathcal A}$, and
$\alpha\in \Aut({\mathcal A})$ a unital Lip-bounded automorphism. Set
\begin{flalign*}
& {\mathcal L}_\rtimes := {}^*\mathrm{alg}(\widetilde{\pi_u}({\mathcal L}),U), & \qquad & {\mathcal H}_\rtimes := {\mathcal H} \otimes \ell^2({\mathbb Z}) \otimes {\mathbb C}^2, & \\
& D_\rtimes := D \otimes I \otimes \eps_1 + I \otimes D_{\mathbb Z} \otimes \eps_2, & \qquad & \G_\rtimes := I \otimes I \otimes \eps_3, &
\end{flalign*}
where ${}^*\mathrm{alg}(\widetilde{\pi_u}({\mathcal L}),U)$ is the $^*$-algebra generated by $\widetilde{\pi_u}({\mathcal L})$ and $U$, $(D_{\mathbb Z} \xi)(n) := n\xi(n)$, $\forall \xi\in\ell^2({\mathbb Z})$,
and
\begin{align}\label{pauli}
\eps_1 := \begin{pmatrix}
0 & 1 \\
1 & 0
\end{pmatrix}, \; \eps_2 := \begin{pmatrix}
0 & -i \\
i & 0
\end{pmatrix}, \; \eps_3 := \begin{pmatrix}
1 & 0\\
0 & -1
\end{pmatrix}
\end{align}
are the Pauli matrices. %
\smallskip
Then $({\mathcal L}_\rtimes,{\mathcal H}_\rtimes,D_\rtimes,\Gamma_\rtimes)$ is an even spectral triple on ${\mathcal A}\rtimes_\alpha {\mathbb Z}$.
\end{theorem}
\begin{remark}
In \cite{KaKy} a more general notion than Lip-boundeness of an automorphism is introduced, there called quasi-isometricity (see their Definition 4.4), and it is shown (in Example 4.5) that quasi-isometric automorphisms arise naturally in differential geometry.
\end{remark}
\subsection{A new definition of crossed product by an endomorphism}
There are many different definitions of the crossed product with an endomorphism, see e.g. \cite{Murphy}, \cite{Exel}, and the very general one given in \cite{KwLe}. We will work with a modification of the one introduced in \cite{Cuntz,StaceyCrossed}. Indeed, Cuntz (\cite{Cuntz}, pag. 101) considers the inductive sequence \eqref{eq:CstarIndLim1}, and its inductive limit C*-algebra ${\mathcal A}_\infty$, which is endowed with an automorphism $\alpha_\infty$, uniquely defined by the diagram \eqref{eq:CstarIndLim2}
\begin{equation} \label{eq:CstarIndLim2}
\xymatrix{
& {\mathcal A} \ar[rr]^{ \alpha } \ar[dd]^{\alpha} && {\mathcal A} \ar[rr]^{ \alpha } \ar[dd]^{\alpha} && {\mathcal A} \ar[rr]^{ \alpha } \ar[dd]^{\alpha} && \cdots \ar[r] & {\mathcal A}_\infty \ar[dd]^{\alpha_\infty} \\
&&&&&&&&\\
& {\mathcal A} \ar[rr]^{ \alpha } \ar[uurr]^{id} && {\mathcal A} \ar[rr]^{ \alpha } \ar[uurr]^{id} && {\mathcal A} \ar[rr]^{ \alpha } \ar[uurr]^{id} && \cdots \ar[r] & {\mathcal A}_\infty
}
\end{equation}
where the diagonal maps define the inverse $\alpha_\infty^{-1}$. Then Cuntz defined ${\mathcal A}\rtimes_\alpha {\mathbb N} := q({\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z})q$, where $q\in{\mathcal A}_\infty$ is the image of $1\in{\mathcal A}$, and turns out to be $q=1$ in our case, since $\alpha$ is unital. Subsequently, Stacey \cite{StaceyCrossed} characterised ${\mathcal A}\rtimes_\alpha {\mathbb N}$ as the solution of a universal problem.
In this paper, our interest is in lifting suitable spectral triples from $({\mathcal A},\alpha)$, where $\alpha\in \End({\mathcal A})$, to ${\mathcal A} \rtimes_\alpha {\mathbb N}$. Since we already know how to lift a spectral triple from $({\mathcal A},\alpha)$ to $({\mathcal A}_\infty,\alpha_\infty)$, at least in some examples \cite{AiGuIs01,AGI3}, and the lift from $({\mathcal A}_\infty,\alpha_\infty)$ to ${\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z}$ is well known \cite{BMR}, we found only natural to use Cuntz' definition of the crossed product ${\mathcal A}\rtimes_\alpha {\mathbb N}$. Unfortunately, the spectral triples $({\mathcal L}_\infty,{\mathcal H}_\infty,D_\infty)$ on $({\mathcal A}_\infty,\alpha_\infty)$ we constructed in \cite{AiGuIs01,AGI3} satisfy, besides $\alpha_\infty({\mathcal L}_\infty)\subset {\mathcal L}_\infty$, only $\sup_{n\in{\mathbb N}} \| [D_\infty,\alpha_\infty^{-n}(a)] \| < \infty$, $\forall a\in{\mathcal L}_\infty$. This fact forces us to introduce a modification in Cuntz' procedure, namely to consider ${\mathcal A} \rtimes_\alpha{\mathbb N} :=p({\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z})p$, where $p \in {\mathcal B}(\ell^2({\mathbb Z},{\mathcal H}_u))$ is the projection on the non-negative ``frequencies''
\begin{equation} \label{eq:pProjection}
(p \xi)(n) =
\begin{cases}
\xi(n), & n\geq 0,\\
0, & n<0.
\end{cases}
\end{equation}
Actually, we prefer to define our version of the crossed product by an endomorphism, in the same spirit of Stacey, as the solution to a universal problem, see Definition \ref{1.1}, and then prove in Theorem \ref{prop:crossedProd} that it coincides with $p({\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z})p$.
\begin{definition}\label{Def:CovRep}
Let ${\mathcal A}$ be a unital $C^*$-algebra, $\alpha\in \End(A)$ a $^*$-endomorphism. Let $\piA:{\mathcal A}\to {\mathcal B}({\mathcal H})$ be a representation, $W\in {\mathcal B}({\mathcal H})$ an isometry. We say that $({\mathcal H},\piA,W)$ is a covariant representation of $({\mathcal A},\alpha)$ on ${\mathcal H}$, if
\begin{align*}
\piA(\alpha(a))W & = W\piA(a), \quad a\in {\mathcal A},\\
W^kW^{*k} & \in \piA({\mathcal A})', \quad k\in{\mathbb N}.
\end{align*}
\end{definition}
\begin{definition}
\label{1.1}
Let ${\mathcal A}$ be a unital $C^*$-algebra, $\alpha\in \End(A)$ an injective, unital $*$-endomorphism. The crossed product of ${\mathcal A}$ with ${\mathbb N}$ by $\alpha$ is a unital $C^*$-algebra ${\mathcal B}$, together with a unital $^*$-monomorphism $\zci_{\mathcal A}: {\mathcal A}\to{\mathcal B}$, and an isometry $t\in{\mathcal B}$, such that
\begin{enumerate}
\item[$(1)$] ${\mathcal B}$ is the $C^*$-algebra generated by $\zci_{\mathcal A}({\mathcal A})$ and $t$,
\item[$(2)$] $\zci_{\mathcal A}(\alpha(a))t = t\zci_{\mathcal A}(a)$, $a\in{\mathcal A}$,
\item[$(3)$] $t^k(t^*)^k$ commutes with $\zci_{\mathcal A}({\mathcal A})$, $k\in{\mathbb N}$,
\item[$(4)$] for every covariant representation $({\mathcal H},\piA,W)$ of $({\mathcal A},\alpha)$, there exists a non-degenerate representation $\widehat{\piA}$ of ${\mathcal B}$ on ${\mathcal H}$, such that $\widehat{\piA}\circ \zci_{\mathcal A} = \piA$, and $\widehat{\piA}(t)=W$.
\end{enumerate}
\end{definition}
We denote by ${\mathcal A} \rtimes_\alpha {\mathbb N}$ the above algebra ${\mathcal B}$. We have defined our crossed product as a universal object, which guarantees its uniqueness. For its existence, we will prove in Proposition \ref{prop:crossedProd} that it is a reduction by a projection of the $C^*$-algebra crossed product defined by Cuntz in \cite{Cuntz}.
\subsection{Existence of the universal object}
Let us now consider the commutative diagram \eqref{eq:CstarIndLim2}.
It follows from (\cite{WO}, Theorem L.2.1) that the vertical maps determine a $^*$-homomorphism $\alpha_\infty:{\mathcal A}_\infty\to{\mathcal A}_\infty$, and the diagonal maps define the inverse of $\alpha_\infty$.
\begin{proposition} \label{prop:exists}
Let ${\mathcal A}$ be a unital $C^*$-algebra, $\alpha$ a unital, injective $^*$-endomorphism of ${\mathcal A}$. Then, there exists a covariant representation $({\mathcal H},\piA,W)$ of $({\mathcal A},\alpha)$.
\end{proposition}
\begin{proof}
Let $\repr$ be a faithful representation of ${\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z}$ on a Hilbert space $H$. If $\pi_u$ is the universal representation of ${\mathcal A}_\infty$, let $\ppcov : {\mathcal A}_\infty \to {\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z}$, $\uu \in {\mathcal U}( {\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z})$ be such that ${\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z} = \langle \ppcov({\mathcal A}_\infty), \uu \rangle$, $\piA := \repr \circ \ppcov \circ \varphi_{\infty 0} : {\mathcal A} \to {\mathcal B}(H)$, which is a representation of ${\mathcal A}$ on $H$, and $W:= \repr(\uu) \in {\mathcal B}(H)$, which is a unitary operator acting on $H$. Moreover, for all $a\in{\mathcal A}$, $k\in{\mathbb N}$, by using that $\varphi_{\infty 0}\circ\alpha=\alpha_\infty \circ \varphi_{\infty 0}$ and $\ppcov(\alpha_\infty(x))=\uu\ppcov(x)\uu^*$, we get
\begin{align*}
\piA(\alpha(a)) W & = (\repr\circ \ppcov \circ \varphi_{\infty 0}(\alpha(a)) \cdot \repr(\uu) ) = \repr( \ppcov \circ \alpha_\infty \circ \varphi_{\infty 0}(a) \cdot \uu ) \\
& = \repr( \uu \cdot \ppcov \circ \varphi_{\infty 0}(a)) = (\repr(\uu)) (\repr\circ \ppcov \circ \varphi_{\infty 0}(a)) \\
& = W\piA(a), \\
W^kW^{*k} & = \repr(\uu^k \uu^{*k}) = 1 \in \piA({\mathcal A})'.
\end{align*}
\end{proof}
We now prove that any covariant representation of $({\mathcal A},\alpha)$ lifts to a covariant representation of $({\mathcal A}_\infty,\alpha_\infty)$.
\begin{proposition} \label{prop:RepIndLim}
Let ${\mathcal A}$ be a unital $C^*$-algebra, $\alpha$ a unital, injective $^*$-endomorphism of ${\mathcal A}$, and denote by ${\mathcal A}_\infty$ the $C^*$-algebra inductive limit of the inductive system \eqref{eq:CstarIndLim1},
and denote by $\alpha_\infty$ the automorphism of ${\mathcal A}_\infty$ induced by $\alpha$.
Let $({\mathcal H},\piA,W)$ be a covariant representation of $(A,\alpha)$, and denote by ${\mathcal H}_\infty \equiv \varinjlim H_n$ the Hilbert space inductive limit of the inductive system
\begin{equation} \label{HilbertIndLim}
\begin{CD}
H_0 @ > S_0 >> H_1 @ > S_1 >> \cdots
\end{CD}
\end{equation}
where, for all $n\in{\mathbb N}$, $H_n:={\mathcal H}$, $S_n:= W$. Then, there exist $W_\infty\in{\mathcal U}({\mathcal H}_\infty)$, and a covariant representation $({\mathcal H}_\infty, \ppinf, W_\infty)$ of $({\mathcal A}_\infty,\alpha_\infty)$, such that
\begin{align*}
\ppinf\circ \varphi_{\infty n}(a) S_{\infty n} & = S_{\infty n} \piA(a), \quad n\in\bn\cup\{0\}, a\in{\mathcal A}, \\
W_\infty S_{\infty 0} & = S_{\infty 0} W,
\end{align*}
where $S_{\infty n}: \xi\in H_n \mapsto (\xi_k)\in {\mathcal H}_\infty$, $\xi_k:=
\begin{cases}
0, & k<n, \\
W^{k-n}\xi, & k\geq n.
\end{cases}$
\end{proposition}
\begin{proof}
Denote by $W_\infty$ the unitary operator on the inductive limit ${\mathcal H}_\infty \equiv \varinjlim H_n$ defined by the following diagram
\begin{equation*}
\xymatrix{
& {\mathcal H} \ar[rr]^{ W } \ar[dd]^{ W } && {\mathcal H} \ar[rr]^{ W } \ar[dd]^{ W } && {\mathcal H} \ar[rr]^{ W } \ar[dd]^{ W } && \cdots \ar[r] & {\mathcal H}_\infty \ar[dd]^{ W_\infty } \\
&&&&&&&&\\
& {\mathcal H} \ar[rr]^{ W } \ar[uurr]^{id} && {\mathcal H} \ar[rr]^{ W } \ar[uurr]^{id} && {\mathcal H} \ar[rr]^{ W } \ar[uurr]^{id} && \cdots \ar[r] & {\mathcal H}_\infty
}
\end{equation*}
so that $W_\infty S_{\infty n} = S_{\infty,n-1}$, for all $n\in{\mathbb N}$, $n\geq 1$, and $W_\infty S_{\infty 0} = S_{\infty 0}W$.
Introduce a map $\piAinf_0:{\mathcal A} \to {\mathcal B}({\mathcal H}_\infty)$ by
\begin{align*}
\piAinf_0(a) S_{\infty m}\xi := S_{\infty m} \piA(\alpha^m(a)) \xi, \quad a\in{\mathcal A}, m\in{\mathbb N}, \xi\in H_m \equiv {\mathcal H},
\end{align*}
which is well defined, because, if $S_{\infty m}\xi = S_{\infty,m-1}\eta = S_{\infty m}W\eta$, then $\xi=W\eta$, and
\begin{align*}
S_{\infty m} \piA(\alpha^m(a)) \xi & = S_{\infty m} \piA(\alpha^m(a)) W\eta = S_{\infty m} W\piA(\alpha^{m-1}(a))\eta = S_{\infty,m-1} \piA(\alpha^{m-1}(a))\eta.
\end{align*}
Let us prove that $\piAinf_0$ is a representation of ${\mathcal A}$. Indeed, for $a,b\in{\mathcal A}$, we get, for all $m\in{\mathbb N}$, $\xi\in H_m$,
\begin{align*}
\piAinf_0(ab) S_{\infty m}\xi & = S_{\infty m} \piA(\alpha^m(ab)) \xi = S_{\infty m} \piA(\alpha^m(a)) \piA(\alpha^m(b)) \xi \\
& = \piAinf_0(a) S_{\infty m} \piA(\alpha^m(b)) \xi = \piAinf_0(a) \piAinf_0(b) S_{\infty m} \xi.
\end{align*}
Moreover, for $a\in{\mathcal A}_\infty$, $\xi,\eta\in{\mathcal H}$, $m,n\in{\mathbb Z}$, we get, if $n<m$,
\begin{align*}
(S_{\infty m}\xi, \piAinf_0(a)^*S_{\infty n}\eta) & = (\piAinf_0(a)S_{\infty m}\xi, S_{\infty n}\eta) = (S_{\infty m}\piA(\alpha^m(a))\xi, S_{\infty n}\eta) \\
& = (S_{\infty m}\piA(\alpha^m(a))\xi, S_{\infty m}S_{mn}\eta) = (\piA(\alpha^m(a))\xi, S_{mn}\eta) \\
& = (\xi, \piA(\alpha^m(a^*))W^{m-n}\eta) = (\xi, W^{m-n}\piA(\alpha^n(a^*))\eta) \\
& = (S_{\infty m}\xi, S_{\infty n}\piA(\alpha^n(a^*))\eta) = (S_{\infty m}\xi, \piAinf_0(a^*)S_{\infty n}\eta).
\end{align*}
Setting, for all $n\in{\mathbb N}$, $\piAinf_n := \textrm{Ad}(W_\infty^*)^n\circ\piAinf_0$, we get, for $m\geq n+1$,
\begin{align*}
\piAinf_{n+1}(\alpha(a)) S_{\infty m} & = (W_\infty^*)^{n+1}\piAinf_0(\alpha(a))W_\infty^{n+1} S_{\infty m} = (W_\infty^*)^{n+1} \piAinf_0(\alpha(a)) S_{\infty,m-n-1} \\
& = (W_\infty^*)^{n+1} S_{\infty,m-n-1} \piA(\alpha^{m-n}(a)) = (W_\infty^*)^{n} S_{\infty,m-n} \piA(\alpha^{m-n}(a)) \\
& = (W_\infty^*)^{n} \piAinf_0(a) S_{\infty,m-n} = (W_\infty^*)^{n} \piAinf_0(a) W_\infty^n S_{\infty m} = \piAinf_n(a) S_{\infty m},
\end{align*}
so that $\piAinf_{n+1}(\alpha(a)) = \piAinf_n(a)$. Therefore, the following diagram commutes
\begin{equation*}
\begin{CD}
A_0 @ > \varphi_0 >> A_1 @ > \varphi_1 >> A_2 @ > \varphi_2 >> \cdots @ >>> {\mathcal A}_\infty \\
@V \piAinf_0 VV @V \piAinf_1 VV @V \piAinf_2 VV @. @V \ppinf VV \\
{\mathcal B}({\mathcal H}_\infty) @ > \id >> {\mathcal B}({\mathcal H}_\infty) @> \id >> {\mathcal B}({\mathcal H}_\infty) @> \id >> \cdots @>>> {\mathcal B}({\mathcal H}_\infty)
\end{CD}
\end{equation*}
so that there is a unique $^*$-homomorphism $\ppinf:{\mathcal A}_\infty\to{\mathcal B}({\mathcal H}_\infty)$ such that $\ppinf\circ \varphi_{\infty n} = \piAinf_n$, for all $n\in{\mathbb N}$. Therefore, for all $n\in{\mathbb N}$, $a\in{\mathcal A}$, we have
\begin{align}
\ppinf\circ \varphi_{\infty n}(a) S_{\infty n} & = \piAinf_n(a)S_{\infty n} = W_\infty^{*n}\piAinf_0(a) W_\infty^n S_{\infty n} = W_\infty^{*n}\piAinf_0(a) S_{\infty 0} \notag \\
& = W_\infty^{*n} S_{\infty 0} \piA(a) = S_{\infty n} \piA(a). \label{allaccia}
\end{align}
Finally, for all $n\in{\mathbb N}$, $n\geq 1$, $a\in A_n = {\mathcal A}$, we have
\begin{align*}
\ppinf\circ\alpha_\infty\circ\varphi_{\infty n}(a) & = \ppinf\circ\varphi_{\infty n}\circ\alpha(a) = \piAinf_n\circ\alpha(a) = \piAinf_{n-1}(a) \\
& = \textrm{Ad}(W_\infty) \circ \piAinf_n(a) = \textrm{Ad}(W_\infty) \circ \ppinf\circ \varphi_{\infty n} (a),
\end{align*}
so that $\ppinf\circ \alpha_\infty = \textrm{Ad}(W_\infty) \circ \ppinf$, that is $({\mathcal H}_\infty,\ppinf,W_\infty)$ is a covariant representation of $({\mathcal A}_\infty,\alpha_\infty)$.
\end{proof}
\bigskip
We recall that in the construction of ${\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z}$ we
denoted by $\pi_u$ the universal representation of ${\mathcal A}_\infty$ on ${\mathcal H}_u$, so that ${\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z} = \langle\, \ppcov({\mathcal A}_\infty),\uu \,\rangle$.
\noindent Define the projection $p\in {\mathcal B}( \ell^2({\mathbb Z},{\mathcal H}_u) )$ as in \eqref{eq:pProjection}, so that $p\ppcov(a)=\ppcov(a)p$, $a\in{\mathcal A}_\infty$, and set $t:= p\uu p \equiv \uu p$, so that $t^*t = p$, and $t\ppcov(a) = \ppcov(\alpha_\infty(a))t$, $a\in {\mathcal A}_\infty$. Set $\zci_{\mathcal A}(a) := \ppcov\circ \varphi_{\infty 0}(a)p$, which is a representation of ${\mathcal A}$ on $p\ell^2({\mathbb Z},{\mathcal H}_u)$, and denote by $C^*({\mathcal A},\alpha,{\mathbb N})$ the $C^*$-algebra generated by $\zci_{\mathcal A}({\mathcal A})$ and $t$ on $p\ell^2({\mathbb Z},{\mathcal H}_u)$.
\begin{proposition} \label{prop:generate}
For any $a\in{\mathcal A}$, $k\in{\mathbb N}$, we have that
\begin{itemize}
\item[$(1)$] $\zci_{\mathcal A}(\alpha(a))t = t\zci_{\mathcal A}(a)$,
\item[$(2)$] $t^k(t^*)^k \zci_{\mathcal A}(a) = \zci_{\mathcal A}(a) t^k(t^*)^k$,
\item[$(3)$] $C^*({\mathcal A},\alpha,{\mathbb N}) \equiv \langle\, \zci_{\mathcal A}({\mathcal A}), t \,\rangle = \langle\, p\ppcov({\mathcal A}_\infty)p, p\uu p \,\rangle = \langle\, t^{*m}\zci_{\mathcal A}(a)t^n: a\in{\mathcal A}, m,n\in{\mathbb N} \,\rangle = \langle\, \zci_{\mathcal A}(a)t^mt^{*n}: a\in{\mathcal A}, m,n\in{\mathbb N} \,\rangle$.
\end{itemize}
\end{proposition}
\begin{proof}
$(1)$ Indeed, for all $a\in{\mathcal A}$,
\begin{align*}
\zci_{\mathcal A}(\alpha(a))t & = \ppcov\circ\varphi_{\infty 0}\circ\alpha(a)t = \ppcov\circ\alpha_\infty\circ\varphi_{\infty 0}(a)t = t\ppcov\circ\varphi_{\infty 0}(a)p = t\zci_{\mathcal A}(a).
\end{align*}
\noindent $(2)$ Indeed, since $t^k = \uu^kp = p\uu^kp$, we get
\begin{align*}
t^k(t^*)^k \zci_{\mathcal A}(a) & = U^kp(U^*)^k \ppcov \circ\varphi_{\infty 0}(a)p = U^kp \ppcov \circ\alpha_\infty^{-k} \circ \varphi_{\infty 0}(a) (U^*)^kp \\
& = U^k \ppcov \circ\alpha_\infty^{-k} \circ \varphi_{\infty 0}(a) p(U^*)^kp = \ppcov \circ \varphi_{\infty 0}(a) U^k p(U^*)^kp \\
& = \ppcov \circ \varphi_{\infty 0}(a) pU^k p(U^*)^kp = \zci_{\mathcal A}(a) t^k(t^*)^k.
\end{align*}
\noindent $(3)$ Indeed, ${\mathcal A}_\infty = \overline{\ssv} \{ \varphi_{\infty m}(a) : a\in{\mathcal A}, m\in{\mathbb N} \}$, and
\begin{align*}
p\ppcov\circ \varphi_{\infty m}(a)p & = p\ppcov\circ \varphi_{\infty m}(a) \uu^{-m}\uu ^{m}p = p \uu^{-m} \ppcov \circ \alpha_\infty^m \circ\varphi_{\infty m}(a) \uu ^{m} p \\
& = p \uu^{-m} \ppcov \circ \varphi_{\infty m} \circ \alpha^m(a) \uu ^{m }p = p \uu^{-m} \ppcov \circ \varphi_{\infty 0}(a) \uu ^{m }p = t^{*m} \zci_{\mathcal A}(a) t^{m},
\end{align*}
so that $\langle\, p\ppcov({\mathcal A}_\infty)p, p\uu p \,\rangle = \langle\, t^{*m}\zci_{\mathcal A}(a)t^n: a\in{\mathcal A}, m,n\in{\mathbb N} \,\rangle = \langle\, \zci_{\mathcal A}(a)t^mt^{*n}: a\in{\mathcal A}, m,n\in{\mathbb N} \,\rangle = \langle\, \zci_{\mathcal A}({\mathcal A}), t \,\rangle$.
\end{proof}
We want to prove that $C^*({\mathcal A},\alpha,{\mathbb N})$ is isomorphic to the crossed product of ${\mathcal A}$ with $\alpha$ by ${\mathbb N}$. Actually, property $(1)$ in Definition \ref{1.1} follows by definition, while properties $(2)$ and $(3)$ have been proved in Proposition \ref{prop:generate}. Unfortunately, the proof of property $(4)$ in Definition \ref{1.1} will force us to a long detour. First of all, we need a $C^*$-algebra which contains ${\mathcal A}_\infty\rtimes_{\alpha_\infty}{\mathbb Z}$ and a projection on the ``positive frequencies'' of ${\mathbb Z}$, and to which we can lift, in a canonical way, any representation of ${\mathcal A}_\infty$. We start with some preliminary results.
Denote by ${\mathbb Z}_\infty:={\mathbb Z}\cup\{+\infty\}$ the spectrum of the $C^*$-algebra of functions on ${\mathbb Z}$, vanishing at $-\infty$, and having finite limit for $n\to+\infty$, and let $\betaC$ be the automorphism of $C_0({\mathbb Z}_\infty)$ given by $\betaC(f)(n):= f(n-1)$, $n\in{\mathbb Z}$.
It follows from \cite{Sakai}, Proposition 1.22.3, that ${\mathcal A}_\infty \otimes C_0({\mathbb Z}_\infty) \cong C_0({\mathbb Z}_\infty,{\mathcal A}_\infty)$, that is two-sided sequences of elements in ${\mathcal A}_\infty$, vanishing at $-\infty$, and having norm-limit for $n\to+\infty$. It follows from \cite{Take}, Proposition IV.4.22 that there is a unique automorphism $\gammaCA\in \Aut( C_0({\mathbb Z}_\infty;{\mathcal A}_\infty) )$ such that $\gammaCA(a\otimes f) = \alpha_\infty(a)\otimes \betaC(f)$, $a\in{\mathcal A}_\infty$, $f\in C_0({\mathbb Z}_\infty)$.
In Proposition \ref{prop:reprOfCrossed}, we construct a representation of $C_0({\mathbb Z}_\infty;{\mathcal A}_\infty) \rtimes_{\gammaCA} {\mathbb Z}$ on $\ell^2({\mathbb Z},{\mathcal H}_u)$. Let $\rruniv$ be the representation of $C_0({\mathbb Z}_\infty)$ on ${\mathcal H}_u$ given by $\rruniv(f)\xi = f(0)\xi$, $f\in C_0({\mathbb Z}_\infty)$, $\xi\in{\mathcal H}_u$. It follows from \cite{Take}, Proposition IV.4.7, that there is a unique representation $\ssuniv$ of $C_0({\mathbb Z}_\infty;{\mathcal A}_\infty)$ on ${\mathcal H}_u$, such that $\ssuniv(a\otimes f) = \ppuniv(a)\rruniv(f)$, $a\in{\mathcal A}_\infty$, $f\in C_0({\mathbb Z}_\infty)$.
Introduce the representations $\rrcov$ of $C_0({\mathbb Z}_\infty)$ and $\sscov$ of $C_0({\mathbb Z}_\infty;{\mathcal A}_\infty)$ on $\ell^2({\mathbb Z},{\mathcal H}_u)$ given by, for $a\in{\mathcal A}_\infty$, $f\in C_0({\mathbb Z}_\infty)$, $\xi \in \ell^2({\mathbb Z},{\mathcal H}_u)$, $n\in{\mathbb Z}$,
\begin{align*}
(\rrcov(f) \xi)(n) & := \rruniv( \betaC^{-n}(f) ) \xi(n) = f(n)\xi(n), \\
(\sscov(a\otimes f) \xi)(n) & := \ssuniv( \gammaCA^{-n}(a\otimes f) ) \xi(n).
\end{align*}
\begin{proposition} \label{prop:reprOfCrossed}
\begin{itemize}
\item[$(1)$] $\uu \rrcov(f)\uu^* = \rrcov(\betaC(f))$, $f\in C_0({\mathbb Z}_\infty)$.
\item[$(2)$] The representation $\sscov$ of $C_0({\mathbb Z}_\infty;{\mathcal A}_\infty)$ on $\ell^2({\mathbb Z},{\mathcal H}_u)$ is faithful, and
\begin{align*}
\sscov( C_0({\mathbb Z}_\infty;{\mathcal A}_\infty) ) & = \langle\, \ppcov({\mathcal A}_\infty), \rrcov(C_0({\mathbb Z}_\infty)) \,\rangle, \\
\uu \sscov(a\otimes f)\uu^* & = \sscov(\gammaCA(a\otimes f)), \quad a\in{\mathcal A}_\infty, f\in C_0({\mathbb Z}_\infty).
\end{align*}
\item[$(3)$] The regular representation $\chi:=\ssU$ of $C_0({\mathbb Z}_\infty,{\mathcal A}_\infty) \rtimes_{\gammaCA} {\mathbb Z}$, induced from $\ssuniv$ on $\ell^2({\mathbb Z},{\mathcal H}_u)$, is faithful.
\end{itemize}
\end{proposition}
\begin{proof}
$(1)$ is a computation.
\noindent $(2)$ It is easy to see that $(\sscov(g)\xi)(k) = \pi_u(\alpha_\infty^{-k}(g(k)))\xi(k)$, $k\in{\mathbb Z}$, $\xi\in\ell^2({\mathbb Z},{\mathcal H}_u)$, $g\in C_0({\mathbb Z}_\infty,{\mathcal A}_\infty)$, from which it follows that $\sscov$ is faithful. Moreover, for $a\in{\mathcal A}_\infty$, $f\in C_0({\mathbb Z}_\infty)$, one has
\begin{align*}
\sscov(\gammaCA( a\otimes f)) & = \sscov(\alpha_\infty(a)\otimes \betaC(f)) = \ppcov(\alpha_\infty(a)) \rrcov(\betaC(f)) \\
& = \uu \ppcov(a) \uu^* \uu \rrcov(f) \uu^* = \uu \sscov(a\otimes f) \uu^*.
\end{align*}
\noindent $(3)$ This follows from \cite{Will}, Theorem 7.13.
\end{proof}
It follows from the previous Proposition that ${\mathcal C}:= \langle\, \ppcov({\mathcal A}_\infty), \rrcov(C_0({\mathbb Z}_\infty)), \uu \,\rangle \subset {\mathcal B}(\ell^2({\mathbb Z},{\mathcal H}_u))$ is isomorphic, via $\chi^{-1}$, to $({\mathcal A}_\infty \otimes C_0({\mathbb Z}_\infty)) \rtimes_{\gammaCA} {\mathbb Z}$, and contains $C^*({\mathcal A},\alpha,{\mathbb N})$. It follows from its construction that we can lift canonically to ${\mathcal C}$ any representation of ${\mathcal A}_\infty$, as we prove in Proposition \ref{prop:Sigma}.
We now begin the proof of property $(4)$ in Definition \ref{1.1}. In rough terms, starting from a covariant representation $\pi$ of $({\mathcal A},\alpha)$ on a Hilbert space ${\mathcal H}$, we construct a covariant representation $\pi_\infty$ of $({\mathcal A}_\infty,\alpha_\infty)$ on ${\mathcal H}_\infty$. Then we construct a suitable representation $\rho_\infty$ of $C_0({\mathbb Z}_\infty)$ on ${\mathcal H}_\infty$, which allows us to construct a representation $\sigma} \def\S{\Sigma_\infty$ of $C_0({\mathbb Z}_\infty;{\mathcal A}_\infty)$ on ${\mathcal H}_\infty$, and then a representation $\sigma} \def\S{\Sigma_\infty \rtimes W_\infty$ of $C_0({\mathbb Z}_\infty;{\mathcal A}_\infty) \rtimes_\gamma} \def\G{\Gamma {\mathbb Z}$, viz. a representation $\pi_{\mathcal C}$ of ${\mathcal C}$, on ${\mathcal H}_\infty$, that we can restrict to $C^*({\mathcal A},\alpha,{\mathbb N})$, and compress to a representation $\widehat{\pi}$ on ${\mathcal H}$ that satisfies property $(4)$ in Definition \ref{1.1}.
In order to the help the reader with the understanding of the following statements and proofs, we exhibit two tables with
the $C^*$-algebras considered, and their representations on the various Hilbert spaces
\begin{center}
\begin{tabular}{ | c | c | c | c | c | }
\hline
& $\Aut(\cdot)$ & ${\mathcal H}_\infty$ & ${\mathcal H}_u$ & $\ell^2({\mathbb Z};{\mathcal H}_u)$ \\ \hline
${\mathcal A}_\infty$ & $\alpha_\infty$ & $\pi_\infty$ & $\pi_u$ & $\widetilde{\pi_u}$ \\ \hline
$C_0({\mathbb Z}_\infty)$ & $\beta$ & $\rho_\infty$ & $\rho_u$ & $\widetilde{\rho_u}$ \\ \hline
$C_0({\mathbb Z}_\infty;{\mathcal A}_\infty)$ & $\gamma} \def\G{\Gamma\equiv \alpha_\infty\otimes\beta$ & $\sigma} \def\S{\Sigma_\infty$ & $\sigma} \def\S{\Sigma_u$ & $\widetilde{\sigma} \def\S{\Sigma_u}$ \\ \hline
${\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z}$ & - & $\pi_\infty \rtimes W_\infty$ & - & $\widetilde{\pi_u} \rtimes U$ \\ \hline
$C_0({\mathbb Z}_\infty;{\mathcal A}_\infty) \rtimes_{\gamma} \def\G{\Gamma} {\mathbb Z}$ & - & $\sigma} \def\S{\Sigma_\infty \rtimes W_\infty$ & - & $\chi\equiv \widetilde{\sigma} \def\S{\Sigma_u} \rtimes U$ \\ \hline
${\mathcal C} \equiv \chi(C_0({\mathbb Z}_\infty;{\mathcal A}_\infty) \rtimes_{\gamma} \def\G{\Gamma} {\mathbb Z})$ & - & $\pi_{\mathcal C} \equiv (\sigma} \def\S{\Sigma_\infty \rtimes W_\infty)\circ \chi^{-1}$ & - & id \\ \hline
\end{tabular}
\end{center}
and
\begin{center}
\begin{tabular}{ | c | c | c | c | c | }
\hline
& $\End(\cdot)$ & ${\mathcal H}$ & ${\mathcal H}_\infty$ & $p\ell^2({\mathbb Z};{\mathcal H}_u)$ \\\hline
${\mathcal A}$ & $\alpha$ & $\pi$ & $\psi_0$ & $\zci_{\mathcal A} \equiv \widetilde{\pi_u}\circ\varphi_{\infty0}(\cdot)p$ \\\hline
$C^*({\mathcal A},\alpha,{\mathbb N})$ & - & $\widehat{\pi}\equiv S_{\infty0}^*\pi_{\mathcal C}(\cdot)S_{\infty0}$ & $\pi_{\mathcal C}|_{C^*({\mathcal A},\alpha,{\mathbb N})}$ & id\\\hline
\end{tabular}
\end{center}
Let $({\mathcal H},\piA,W)$ be a covariant representation of $(A,\alpha)$, and recall from Proposition \ref{prop:RepIndLim} that there exist $W_\infty\in{\mathcal U}({\mathcal H}_\infty)$, and a covariant representation $({\mathcal H}_\infty, \ppinf, W_\infty)$ of $({\mathcal A}_\infty,\alpha_\infty)$, on ${\mathcal H}_\infty \equiv \varinjlim H_n$, the Hilbert space inductive limit of the inductive system \eqref{HilbertIndLim},
such that $\ppinf\circ \varphi_{\infty n}(a) S_{\infty n} = S_{\infty n} \piA(a)$, for all $n\in{\mathbb N}$, $a\in{\mathcal A}$, and $W_\infty S_{\infty 0} = S_{\infty 0} W$.
We now construct a representation $\rrinf$ of $C_0({\mathbb Z}_\infty)$ on ${\mathcal H}_\infty$ such that $[\ppinf(a),\rrinf(f)]=0$, for all $a\in{\mathcal A}_\infty$, $f\in C_0({\mathbb Z}_\infty)$.
\begin{proposition} \label{prop:spectralFamily}
Set $P_0:= S_{\infty 0}S_{\infty 0}^*$, $P_n:= \textrm{Ad}(W_\infty^n)(P_0)$, $n\in{\mathbb Z}$. Then
\begin{itemize}
\item[$(1)$] $\{ P_n:n\in{\mathbb Z}\}$ is a decreasing family of projections in ${\mathcal B}({\mathcal H}_\infty)$,
\item[$(2)$] there exists $P_{+\infty}:= \lim_{n\to+\infty} P_n$, in the strong operator topology of ${\mathcal B}({\mathcal H}_\infty)$,
\item[$(3)$] $\lim_{n\to-\infty} P_n = 1$, in the strong operator topology of ${\mathcal B}({\mathcal H}_\infty)$,
\item[$(4)$] $\{ P_n : n\in{\mathbb Z}_\infty \} \subset \ppinf({\mathcal A}_\infty)'$.
\end{itemize}
\end{proposition}
\begin{proof}
$(1)$ Let $n\in{\mathbb Z}$. If $n\geq0$, then
$$
P_n = W_\infty^nS_{\infty 0}S_{\infty 0}^*W_\infty^{n*} = S_{\infty 0}W^nW^{n*}S_{\infty 0}^* \geq S_{\infty 0}W^{n+1}(W^{*})^{n+1}S_{\infty 0}^* = P_{n+1}.
$$
If $n=-k\leq 0$, then
$$
P_n = W_\infty^{*k}S_{\infty 0}S_{\infty 0}^*W_\infty^{k} = S_{\infty k}S_{\infty k}^* = S_{\infty,k+1}WW^*S_{\infty,k+1}^* \leq S_{\infty,k+1}S_{\infty,k+1}^* = P_{n-1}.
$$
\noindent $(2)$ follows from $(1)$.
\noindent $(3)$ We have to prove that $\lim_{k\to+\infty} S_{\infty k}S_{\infty k}^* = 1$, in the strong operator topology, and it suffices to prove it on the dense subset of ${\mathcal H}_\infty$ spanned by $\{ S_{\infty n}\xi : n\in{\mathbb N}, \xi\in{\mathcal H}\}$. Let us fix $n\in{\mathbb N}$, $\xi\in{\mathcal H}$, and compute, for $k>n$, $S_{\infty k}S_{\infty k}^*S_{\infty n}\xi = S_{\infty k}S_{\infty k}^*S_{\infty k}S_{kn}\xi = S_{\infty k}S_{kn}\xi = S_{\infty n}\xi$, and the thesis follows.
\noindent $(4)$ Let us first prove that $\ppinf(x) P_0 = P_0 \ppinf(x)$ for $x \in {\mathcal A}_\infty$. It suffices to show the equality for $x\in \{ \varphi_{\infty n}(a) : n\in{\mathbb N},a\in{\mathcal A}\}$. We have, from equation \eqref{allaccia},
\begin{align*}
\ppinf \circ \varphi_{\infty n}(a)P_0 & = \ppinf\circ \varphi_{\infty n}(a)S_{\infty 0}S_{\infty 0}^* = \ppinf\circ \varphi_{\infty n}(a)S_{\infty n}W^nS_{\infty 0}^* \\
& = S_{\infty n} \piA(a) W^n W^{*n} S_{\infty n}^* = S_{\infty n} W^n W^{*n} \piA(a) S_{\infty n}^* \\
& = S_{\infty 0} W^{*n} S_{\infty n}^* \ppinf \circ \varphi_{\infty n}(a) = P_0 \ppinf \circ \varphi_{\infty n}(a).
\end{align*}
Then, for any $x \in {\mathcal A}_\infty$, $k\in{\mathbb Z}$,
\begin{align*}
\ppinf(x) P_k & = \ppinf(x)W_\infty^kP_0W_\infty^{*k} = W_\infty^k\ppinf(\alpha_\infty^{-k}(x)) P_0W_\infty^{*k} \\
& = W_\infty^kP_0\ppinf(\alpha_\infty^{-k}(x)) W_\infty^{*k} = W_\infty^kP_0W_\infty^{*k} \ppinf(x) = P_k \ppinf(x).
\end{align*}
Finally, $P_{+\infty} \in \ppinf({\mathcal A}_\infty)'$, because of $(2)$.
\end{proof}
\begin{proposition} \label{prop:repC0}
There exists a representation $\rrinf$ of $C_0({\mathbb Z}_\infty)$ on ${\mathcal H}_\infty$, such that, for any $f\in C_0({\mathbb Z}_\infty)$,
\begin{align*}
\rrinf(f) & \in\ppinf({\mathcal A}_\infty)', \\
\rrinf(\betaC(f)) & = W_\infty \rrinf(f)W_\infty^*.
\end{align*}
\end{proposition}
\begin{proof}
Set $E_n:= P_n-P_{n+1}$, $n\in{\mathbb Z}$, $E_{+\infty}:=P_{+\infty}$. Then, $\{ E_n: n\in{\mathbb Z}_\infty\}$ is a spectral family on ${\mathcal H}_\infty$, and $E_{n+1} = W_\infty E_n W_\infty^*$, $n\in{\mathbb N}$, $E_{+\infty} = W_\infty E_{+\infty} W_\infty^*$. Define, for $f\in C_0({\mathbb Z}_\infty)$, $\rrinf(f):= \sum_{n\in{\mathbb Z}_\infty} f(n)E_n$, where the series converges in the strong operator topology of ${\mathcal B}({\mathcal H}_\infty)$. Then, $\rrinf$ is a representation of $C_0({\mathbb Z}_\infty)$ on ${\mathcal H}_\infty$, such that $\rrinf(f)\in\ppinf({\mathcal A}_\infty)'$, for any $f\in C_0({\mathbb Z}_\infty)$, and $\rrinf(\betaC(f)) = W_\infty \rrinf(f)W_\infty^*$, $f\in C_0({\mathbb Z}_\infty)$.
\end{proof}
\begin{proposition} \label{prop:Sigma}
\begin{itemize}
\item[$(1)$] There is a unique representation $\ssinf$ of $C_0({\mathbb Z}_\infty;{\mathcal A}_\infty)$ on ${\mathcal H}_\infty$, such that $\ssinf(a\otimes f) = \ppinf(a)\rrinf(f)$, $a\in{\mathcal A}_\infty$, $f\in C_0({\mathbb Z}_\infty)$. Moreover, $\ssinf(\gammaCA(g)) = W_\infty \ssinf(g)W_\infty^*$, $g\in C_0({\mathbb Z}_\infty;{\mathcal A}_\infty)$.
\item[$(2)$] There is a unique representation $\ssW$ of $C_0({\mathbb Z}_\infty;{\mathcal A}_\infty)\rtimes_{\gammaCA} {\mathbb Z}$ on ${\mathcal H}_\infty$ such that $\ssW(g\delta} \def\D{\Delta_n) = \ssinf(g)W_\infty^n$, $g\in C_0({\mathbb Z}_\infty;{\mathcal A}_\infty)$, $n\in{\mathbb Z}$.
\end{itemize}
\end{proposition}
\begin{proof}
$(1)$ This follows from \cite{Take}, Proposition IV.4.7.
\noindent $(2)$ This follows from \cite{Ped}, Proposition 7.6.4.
\end{proof}
Let us set $\pi_{\mathcal C} := \ssW \circ \chi^{-1}$, which is a representation of ${\mathcal C}$ on ${\mathcal H}_\infty$.
\begin{proposition} \label{prop:repr}
\begin{itemize}
\item[$(1)$] $\pi_{\mathcal C}( x ) = \ppW( x )$, for all $x\in{\mathcal A}_\infty \rtimes_{\alpha_\infty} {\mathbb Z} \equiv \langle\, \ppcov({\mathcal A}_\infty), \uu \,\rangle$,
\item[$(2)$] $\pi_{\mathcal C}(p) = P_0 = S_{\infty 0}S_{\infty 0}^*$.
\end{itemize}
\end{proposition}
\begin{proof}
It follows from Proposition \ref{prop:reprOfCrossed} that, for $\sum_{n\in{\mathbb Z}} (a_n\otimes f_n) \dd_n \in C_c( C_0({\mathbb Z}_\infty;{\mathcal A}_\infty), {\mathbb Z}, \gammaCA)$, we have $\chi( \sum_{n\in{\mathbb Z}} (a_n\otimes f_n) \dd_n ) = \sum_{n\in{\mathbb Z}} \ppcov(a_n)\rrcov(f_n)u^n$, so that
\begin{align*}
\pi_{\mathcal C} \bigg( \sum_{n\in{\mathbb Z}} \ppcov(a_n)\rrcov(f_n) \uu^n \bigg) & = \ssW \bigg( \sum_{n\in{\mathbb Z}} (a_n\otimes f_n) \dd_n \bigg) \\
& = \sum_{k\in{\mathbb Z}} \ssinf( a_n \otimes f_n) W_\infty^n = \sum_{k\in{\mathbb Z}} \ppinf( a_n) \rrinf( f_n) W_\infty^n.
\end{align*}
\noindent $(1)$ Indeed, with $\set{e_n:n\in{\mathbb N}}$ an approximate unit of $C_0({\mathbb Z}_\infty)$, we get, for all $a\in{\mathcal A}_\infty$, $k\in{\mathbb Z}$,
\begin{align*}
\pi_{\mathcal C}( \ppcov(a) \uu^k ) & = \lim_{n\to\infty} \pi_{\mathcal C}( \ppcov(a)\rrcov(e_n) \uu^k) = \lim_{n\to\infty} \ppinf(a) \rrinf(e_n)W_\infty^k \\
& = \ppinf(a)W_\infty^k = \ppW( \ppcov(a) \uu^k ),
\end{align*}
and the thesis follows.
\noindent $(2)$ If $f(n) =
\begin{cases}
0, & n<0,\\
1, & n\geq 0,
\end{cases}\ $ then $\widehat{\chi}(p) = \widehat{\chi}(\rrcov(f)) = \rrinf(f) = P_0$.
\end{proof}
Let us still denote by $\pi_{\mathcal C}$ the restriction of $\pi_{\mathcal C}$ to the subalgebra $C^*({\mathcal A},\alpha,{\mathbb N}) \equiv \langle\, p\ppcov({\mathcal A}_\infty)p, p\uu p \,\rangle$ of ${\mathcal C} \equiv \langle\, \ppcov({\mathcal A}_\infty), \rrcov(C_0({\mathbb Z}_\infty)), \uu \,\rangle$.
\begin{theorem} \label{prop:crossedProd}
$C^*({\mathcal A},\alpha,{\mathbb N})$ satisfies all the properties in Definition \ref{1.1}, namely is the crossed product of ${\mathcal A}$ with ${\mathbb N}$ by $\alpha$.
\end{theorem}
\begin{proof}
As it was already noticed, property $(1)$ in Definition \ref{1.1} follows by definition, while properties $(2)$ and $(3)$ have been proved in Proposition \ref{prop:generate}.
It remains to prove property $(4)$. Let $({\mathcal H},\piA,W)$ be a covariant representation of $(A,\alpha)$, and recall from Proposition \ref{prop:RepIndLim} that there exist $W_\infty\in{\mathcal U}({\mathcal H}_\infty)$, and a covariant representation $({\mathcal H}_\infty, \ppinf, W_\infty)$ of $({\mathcal A}_\infty,\alpha_\infty)$, on ${\mathcal H}_\infty \equiv \varinjlim H_n$, the Hilbert space inductive limit of the inductive system \eqref{HilbertIndLim},
such that $\ppinf\circ \varphi_{\infty n}(a) S_{\infty n} = S_{\infty n} \piA(a)$, for all $n\in{\mathbb N}$, $a\in{\mathcal A}$, and $W_\infty S_{\infty 0} = S_{\infty 0} W$.
Let $\pi_{\mathcal C}$ be the representation of $C^*({\mathcal A},\alpha,{\mathbb N})$ on ${\mathcal H}_\infty$ constructed in Proposition \ref{prop:repr}.
Let us now prove that $P_0\in \pi_{\mathcal C}(C^*({\mathcal A},\alpha,{\mathbb N}))'$, that is $\pi_{\mathcal C}(C^*({\mathcal A},\alpha,{\mathbb N})) S_{\infty 0}{\mathcal H} \subset S_{\infty 0}{\mathcal H}$. Because of Proposition \ref{prop:generate} it is enough to prove that $\pi_{\mathcal C}(t)S_{\infty 0}{\mathcal H} \subset S_{\infty 0}{\mathcal H}$, $\pi_{\mathcal C}(t^*)S_{\infty 0}{\mathcal H} \subset S_{\infty 0}{\mathcal H}$, and $\pi_{\mathcal C}(\zci_{\mathcal A}(a))S_{\infty 0}{\mathcal H} \subset S_{\infty 0}{\mathcal H}$, for all $a\in{\mathcal A}$. Indeed, for all $a\in{\mathcal A}$, $\xi\in{\mathcal H}$, we have
\begin{align*}
\pi_{\mathcal C}(t)S_{\infty 0}\xi & = \pi_{\mathcal C}(p\uu p)S_{\infty 0}\xi = P_0W_\infty P_0S_{\infty 0}\xi \in S_{\infty 0}{\mathcal H}, \\
\pi_{\mathcal C}(t^*)S_{\infty 0}\xi & = \pi_{\mathcal C}(p\uu^* p)S_{\infty 0}\xi = P_0W_\infty^* P_0S_{\infty 0}\xi \in S_{\infty 0}{\mathcal H}, \\
\pi_{\mathcal C}( \zci_{\mathcal A}(a) )S_{\infty 0}\xi & = \pi_{\mathcal C}\circ \ppcov \circ \varphi_{\infty 0} (a) P_0S_{\infty 0}\xi = \ppinf \circ \varphi_{\infty 0} (a) S_{\infty 0}\xi \\
& = S_{\infty 0} \piA(a)\xi \in S_{\infty 0}{\mathcal H}.
\end{align*}
Recall from the proof of Proposition \ref{prop:RepIndLim} that there is a representation $\piAinf_0$ of ${\mathcal A}$ on ${\mathcal H}_\infty$ such that $\piAinf_0(a)S_{\infty 0} = S_{\infty 0}\piA(a)$, $a\in{\mathcal A}$, and $\ppinf\circ\varphi_{\infty 0} = \piAinf_0$. Finally, define
\begin{align*}
\widehat{\piA}(x):= S_{\infty 0}^* \pi_{\mathcal C}(x)S_{\infty 0},\quad x\in C^*({\mathcal A},\alpha,{\mathbb N}),
\end{align*}
which is a representation of $C^*({\mathcal A},\alpha,{\mathbb N})$ on ${\mathcal H}$, because $P_0\in \pi_{\mathcal C}(C^*({\mathcal A},\alpha,{\mathbb N}))'$. Then,
\begin{align*}
\widehat{\piA}(t) & = S_{\infty 0}^* \pi_{\mathcal C}(t)S_{\infty 0} = S_{\infty 0}^* P_0W_\infty P_0S_{\infty 0} = S_{\infty 0}^* W_\infty S_{\infty 0} \\
& = S_{\infty 0}^* S_{\infty 0}W = W,
\end{align*}
and, for all $a\in{\mathcal A}$,
\begin{align*}
\widehat{\piA}(\zci_{\mathcal A}(a)) & = S_{\infty 0}^* \pi_{\mathcal C}( \ppcov \circ \varphi_{\infty 0} (a) p) S_{\infty 0} = S_{\infty 0}^* \ppinf \circ \varphi_{\infty 0} (a) S_{\infty 0}S_{\infty 0}^* S_{\infty 0} \\
& = S_{\infty 0}^* \piAinf_0 (a) S_{\infty 0} = S_{\infty 0}^* S_{\infty 0}\piA(a) = \piA(a).
\end{align*}
\end{proof}
\subsection{An example: the noncommutative torus}
As mentioned in the introduction, the crossed product ${\mathcal A} \rtimes_\alpha {\mathbb N}$ given in Definition \ref{1.1} coincides with a reduction by a projection of the ordinary crossed product when $\alpha$ is an automorphism.
We now give two equivalent descriptions of ${\mathcal A} \rtimes_\alpha {\mathbb N}$, when ${\mathcal A}=C({\mathbb R}/{\mathbb Z})$ and $\alpha$ is a rotation by $2\pi \theta$, where $\theta$ is irrational.
The first description is the following.
As it is known, the noncommutative torus $A_\theta$ can be described as the crossed product $C({\mathbb R}/{\mathbb Z})\rtimes_{\alpha_\theta}{\mathbb Z}$, where $(\alpha_\theta(f))(t)=f(t-\theta)$. Given the Hilbert space $H=\ell^2({\mathbb Z},L^2({\mathbb R}/{\mathbb Z}))$, the representation $\pi:C({\mathbb R}/{\mathbb Z})\to {\mathcal B}(H)$, $(\pi(f)\xi)(n)=\alpha_\theta^{-n}(f)\xi(n)$ and the unitary $V$ acting on $H$ as $(V\xi)(n)=\xi(n-1)$, $A_\theta$ can be identified with the $C^*$-algebra generated by $V$ and $\pi(C({\mathbb R}/{\mathbb Z}))$ on the Hilbert space $H$. Since $C({\mathbb R}/{\mathbb Z})$ is generated as a $C^*$-algebra by the unitary $U_0=\exp(2\pi i t)$, $A_\theta$ is generated by the unitary $V$ and the unitary $U$ given by $(U\xi)(n)=\exp(2\pi i n\theta)U_0\xi(n)$.
It is easy to check that $UV=\exp(2\pi i\theta)VU$.
Since $\alpha_\theta$ is an automorphism, Theorem \ref{prop:crossedProd} implies that $C({\mathbb R}/{\mathbb Z})\rtimes_{\alpha_\theta}{\mathbb N}$ is the reduction of $C({\mathbb R}/{\mathbb Z})\rtimes_{\alpha_\theta}{\mathbb Z}$ by the projection $p$ on the Hilbert space $H_+=\ell^2({\mathbb N}_0,L^2({\mathbb R}/{\mathbb Z}))$.
We have proved the following theorem.
\begin{theorem}
The C$^*$-algebra $C({\mathbb R}/{\mathbb Z})\rtimes_{\alpha_\theta}{\mathbb N}$ can be identified with the $C^*$-algebra generated by the unitary $U$ and the isometry $pVp$ acting on $H_+$.
\end{theorem}
We now provide a description of ${\mathcal A} \rtimes_\alpha {\mathbb N}$ as a universal object.
\begin{theorem}
The $C^*$-algebra $C({\mathbb R}/{\mathbb Z})\rtimes_{\alpha_\theta}{\mathbb N}$ coincides with the universal $C^*$-algebra generated by a unitary $U$ and an isometry $V$ satisfying the conditions $UV=\exp(2\pi i\theta)VU$.
\end{theorem}
\begin{proof}
By definition, the universal $C^*$-algebra generated by a unitary $U$ and an isometry $V$ satisfying the conditions $UV=\exp(2\pi i\theta)VU$
is the unique $C^*$-algebra $B$ satisfyng the following universal property: for any triple $({\mathcal H},u,v)$, where ${\mathcal H}$ is a Hilbert space, $u$ is a unitary and $v$ is an isometry acting on ${\mathcal H}$ satisfying $uv=\exp(2\pi i\theta)vu$, there exists a representation $\pi:B\to {\mathcal B}({\mathcal H})$ such that $\pi(U)=u$ and $\pi(V)=v$.
By definition, also $C({\mathbb R}/{\mathbb Z})\rtimes_{\alpha_\theta}{\mathbb N}$ satisfies a universal property, given by properties $(1) - (4)$ of Definition 2.4. Therefore, given a triple $({\mathcal H},u,v)$ as above, we get indeed a covariant representation $({\mathcal H},\rho, v)$ of $(C({\mathbb R}/{\mathbb Z}),\alpha_\theta)$, where we set $\rho(f)=f(u)$, in fact the commutation relations imply that $v^kv^{*k}u=u v^kv^{*k}$. The properties of $C({\mathbb R}/{\mathbb Z})\rtimes_{\alpha_\theta}{\mathbb N}$ imply the thesis.
\end{proof}
\begin{remark}
If $\theta$ is rational, the projection $p$ in the first description is the identity and, therefore, $C({\mathbb R}/{\mathbb Z})\rtimes_{\alpha_\theta}{\mathbb N}$ coincides with
$A_\theta$.
\end{remark}
\section{Some results on semifinite spectral triples}
In this section we discuss some generalizations of results well-known for type I spectral triples. Some of these results have already been proved in \cite{Jordans} and some are new.
First of all we recall the following definitions:
\begin{definition}
Let $({\mathcal M},\tau)$ be a von Neumann algebra with a normal semifinite faithful (n.s.f.) trace, $T\, \widehat{\in}\, {\mathcal M}$ a self-adjoint operator\footnote{By $T\, \widehat{\in}\, {\mathcal M}$ we mean that the operator $T$ is affiliated with ${\mathcal M}$. Another common notation is $T\, \eta\, {\mathcal M}$.}. We use the notation $e_T(\Omega)$ for the spectral projection of $T$ relative to the measurable set $\Omega\subset {\mathbb R}$, $\lambda} \def\La{\Lambda_t(T):=\tau(e_{|T|}(t,+\infty))$, $\Lambda_t(T):=\tau(e_{|T|}[0,t))$, $\mu_t(T):=\inf \{s>0: \lambda_T(s)\leq t\}$, $t>0$. The operator $T$ is said to be $\tau$-measurable if $\lambda} \def\La{\Lambda_t(T) \to 0$, $t\to+\infty$, and $\tau$-compact if $\mu_t(T) \to 0$, $t\to+\infty$, or equivalently, $\lambda} \def\La{\Lambda_t(T)<+\infty$, $\forall\; t>0$.
\end{definition}
\begin{definition}
Let ${\mathcal A}$ be a unital $C^*$-algebra. An odd semifinite spectral triple $({\mathcal L},{\mathcal H},D; {\mathcal M},\tau)$ on ${\mathcal A}$, with respect to a semifinite von Neumann algebra ${\mathcal M}\subset{\mathcal B}({\mathcal H})$ endowed with a n.s.f. trace $\tau$, is given by a unital, norm-dense, $^*$-subalgebra ${\mathcal L}\subset{\mathcal A}$, a (separable) Hilbert space ${\mathcal H}$, a faithful representation $\pi:{\mathcal A}\to{\mathcal B}({\mathcal H})$ such that $\pi({\mathcal A})\subset{\mathcal M}$, and an unbounded self-adjoint operator $D\, \widehat{\in}\, {\mathcal M}$ such that
\begin{itemize}
\item[$(1)$] $(1+D^2)^{-1}$ is a $\tau$-compact operator, \textit{i.e.} $\lambda} \def\La{\Lambda_t((1+D^2)^{-1})<+\infty$, $\forall\; t>0$ or, equivalently, $\Lambda_t(D)<+\infty$, $\forall \; t>0$,
\item[$(2)$] $\pi(a)(\dom D) \subset \dom D$, and $[D,\pi(a)] \in{\mathcal M}$, for all $a\in{\mathcal L}$.
\end{itemize}
\par\noindent
The spectral triple $({\mathcal L},{\mathcal H},D; {\mathcal M},\tau)$ is even if, in addition,
\begin{itemize}
\item[$(3)$] there is a self-adjoint unitary operator (\textit{i.e.} a ${\mathbb Z}_2$-grading) $\G\in{\mathcal M}$ such that $\pi(a)\G = \G\pi(a)$, $\forall a\in{\mathcal A}$, and $D\G=-\G D$.
\par
The spectral triple $({\mathcal L},{\mathcal H},D; {\mathcal M},\tau)$ is finitely summable if, in addition,
\item[$(4)$] there exists a $\delta} \def\D{\Delta>0$ such that $\tau((1+D^2)^{-\delta} \def\D{\Delta/2})<+\infty$.
\end{itemize}
\end{definition}
\begin{definition}
Given a finitely summable semifinite spectral triple $({\mathcal L},{\mathcal H},D; {\mathcal M},\tau)$, the number $d=\inf\{\alpha>0:\tau((1+D^2)^{-\alpha/2})<+\infty\}$ is called the metric or Hausdorff dimension of the triple, since it is the unique exponent, if any, such that the logarithmic Dixmier trace is finite non-zero on $(1+D^2)^{-\alpha/2}$ (cf. \cite{GuIs09}, Theorem 2.7).
\end{definition}
We note that the usual definition of spectral triple, which was recalled in Definition \ref{deftripla}, can be recovered by taking ${\mathcal M}={\mathcal B}({\mathcal H})$.
\begin{proposition} \label{dimension-Lambda}
Let $({\mathcal L},{\mathcal H},D; {\mathcal M},\tau)$ be a finitely summable semifinite spectral triple. Then $d= \limsup_{t\to\infty}\frac{\log\Lambda_t(D)}{\log t}$.
\end{proposition}
\begin{proof}
We first observe that, by \cite{FaKo}, Proposition 2.7,
$$
\tau((1+D^2)^{-\alpha/2})=\int_0^{+\infty}\mu_t((1+D^2)^{-\alpha/2})dt
= \int_0^{+\infty}\mu^\alpha_t((1+D^2)^{-1/2})dt.
$$
Therefore,
\begin{align*}
d&=\left(\liminf_{t\to\infty}\frac{\log\mu_t((1+D^2)^{-1/2})}{\log(1/t)}\right)^{-1}
=\limsup_{s\to0}\frac{\log\lambda_s((1+D^2)^{-1/2})}{\log(1/s)}\\
&=\limsup_{t\to\infty}\frac{\log\Lambda_t((1+D^2)^{1/2})}{\log t}
=\limsup_{t\to\infty}\frac{\log\Lambda_t(D)}{\log t},
\end{align*}
where the first equality follows by \cite{GuIs09} Theorem 1.4, the second by \cite {GuIs05} Proposition 1.13, the third by definition of $\Lambda$, the last by simple estimates.
\end{proof}
\subsection{The case of the tensor product}
Let us recall the definition of tensor product of semifinite spectral triples.
\begin{definition}
Let ${\mathcal A}_1,{\mathcal A}_2$ be unital $C^*$-algebras, with respective semifinite spectral triples ${\mathcal T}_1:=({\mathcal L}_1,{\mathcal H}_1,D_1,\G_1;{\mathcal M}_1,\tau_1)$, ${\mathcal T}_2:=({\mathcal L}_2,{\mathcal H}_2,D_2,\G_2;{\mathcal M}_2,\tau_2)$, and define ${\mathcal T}_1\times {\mathcal T}_2 \equiv ({\mathcal L},{\mathcal H},D,\G;{\mathcal M},\tau)$ as follows:
\\
if ${\mathcal T}_1$ and ${\mathcal T}_2$ are both even
$$\begin{array}{ccc}
{\mathcal L}:={\mathcal L}_1 \odot {\mathcal L}_2, &
{\mathcal H}:= {\mathcal H}_1 \otimes {\mathcal H}_2, &
D := D_1 \otimes I_2 + \G_1 \otimes D_2, \\
\G:=\G_1\otimes \G_2, &
{\mathcal M} := {\mathcal M}_1 \otimes {\mathcal M}_2, &
\tau := \tau_1 \otimes \tau_2,
\end{array}$$
\\
if ${\mathcal T}_1$ is even, and ${\mathcal T}_2$ is odd,
$$\begin{array}{ccc}
{\mathcal L}:={\mathcal L}_1 \odot {\mathcal L}_2, &
{\mathcal H}:= {\mathcal H}_1 \otimes {\mathcal H}_2, &
D := D_1 \otimes I_2 + \G_1 \otimes D_2, \\
\G:=I_1\otimes I_2, &
{\mathcal M} := {\mathcal M}_1 \otimes {\mathcal M}_2, &
\tau := \tau_1 \otimes \tau_2,
\end{array}$$
\\
if ${\mathcal T}_1$ is odd, and ${\mathcal T}_2$ is even,
$$\begin{array}{ccc}
{\mathcal L}:={\mathcal L}_1 \odot {\mathcal L}_2, &
{\mathcal H}:= {\mathcal H}_1 \otimes {\mathcal H}_2, &
D := D_1 \otimes \G_2 + I_1 \otimes D_2, \\
\G:=I_1\otimes I_2, &
{\mathcal M} := {\mathcal M}_1 \otimes {\mathcal M}_2, &
\tau := \tau_1 \otimes \tau_2,
\end{array}$$
\\
if ${\mathcal T}_1$ and ${\mathcal T}_2$ are both odd,
$$\begin{array}{ccc}
{\mathcal L}:={\mathcal L}_1 \odot {\mathcal L}_2, &
{\mathcal H}:= {\mathcal H}_1 \otimes {\mathcal H}_2 \otimes{\mathbb C}^2, &
D := D_1 \otimes I_2 \otimes \eps_1 + I_1 \otimes D_2 \otimes \eps_2, \\
\G:=I_1\otimes I_2\otimes \eps_3, &
{\mathcal M} := {\mathcal M}_1 \otimes {\mathcal M}_2\otimes M_2({\mathbb C}), &
\tau := \tau_1 \otimes \tau_2\otimes Tr,
\end{array}$$
where $\eps_1$, $\eps_2$, $\eps_3$
are the Pauli matrices, see \eqref{pauli}.
\end{definition}
\begin{proposition} \label{tensorProductTriple}
Let ${\mathcal A}_1,{\mathcal A}_2$ be unital $C^*$-algebras, with respective semifinite spectral triples ${\mathcal T}_1:=({\mathcal L}_1,{\mathcal H}_1,D_1,\G_1;{\mathcal M}_1,\tau_1)$, ${\mathcal T}_2:=({\mathcal L}_2,{\mathcal H}_2,D_2,\G_2;{\mathcal M}_2,\tau_2)$. Then ${\mathcal T}_1 \times {\mathcal T}_2$ is a semifinite spectral triple on the spatial tensor product ${\mathcal A}_1 \otimes {\mathcal A}_2$. Moreover, the Hausdorff dimension $d$ of ${\mathcal T}_1 \times {\mathcal T}_2$ satisfies $d\leq d_1 + d_2$, where $d_1, d_2$ are the Hausdorff dimensions of the factor spectral triples. Finally, if $\displaystyle\lim_{t\to\infty}\frac{\log\Lambda_t(D_1)}{\log t}$ exists, the equality $d= d_1+d_2$ holds.
\end{proposition}
\begin{proof}
In case ${\mathcal T}_1$ and ${\mathcal T}_2$ are not both odd, the result is proved in \cite{Jordans}, Theorem 2.13, and Lemma 2.19. In the remaining case, one can proceed analogously.
We now give an alternative proof of the formula for the Hausdorff dimension, valid in all cases.
Since $D^2=D_1^2\otimes I+I\otimes D_2^2$, in all cases, if $d$ denotes the dimension of $({\mathcal L},{\mathcal H},D; {\mathcal M},\tau)$, we have that
\begin{align*}
d & = \limsup_{t\to\infty}\frac{\log\Lambda_t(D)}{\log t}
= \limsup_{t\to\infty}\frac{\log\tau(e_{D}(-t,t))}{\log t} \\
& = \limsup_{t\to\infty}\frac{\log\tau(\chi_{[0,t^2)}(D_1^2\otimes I+I\otimes D_2^2))}{\log t}.
\end{align*}
If $\sigma} \def\S{\Sigma_i$ denotes the spectrum of $D_i$, $i=1,2$, the representations of $C_0(\sigma} \def\S{\Sigma_i)$ on ${\mathcal H}_i$ with image in ${\mathcal M}_i$ given by functional calculus, $i=1,2$, together with the Radon measures $\nu_i$ on $\sigma} \def\S{\Sigma_i$ induced by the traces $\tau_i$, $i=1,2$, give rise to a representation $j$ of $C_0(\sigma} \def\S{\Sigma_1\times\sigma} \def\S{\Sigma_2)$ on ${\mathcal H}_1\otimes{\mathcal H}_2$ with image in ${\mathcal M}_1\otimes{\mathcal M}_2$ together with the Radon measure $\nu:=\nu_1\otimes\nu_2$ on $\sigma} \def\S{\Sigma_1\times\sigma} \def\S{\Sigma_2$ induced by the trace $\tau:=\tau_1\otimes\tau_2$ such that $j(f_1\otimes f_2)=f_1(D_1)\otimes f_2(D_2)$ and $\int f_1\otimes f_2 d\nu=\tau_1(f_1(D_1)) \tau_2(f_2(D_2))$. Then, denoting by $B_r$ the disk of radius $r$ centered in the origin of the plane, and by $Q_r$ the square $[-r,r]\times[-r,r]$ in the plane,
$$
\chi_{[0,t^2)}(D_1^2\otimes I+I\otimes D_2^2) = j(\chi_{B_t}).
$$
Then the inclusions $Q_{t/\sqrt2}\subset B_t\subset Q_t$ give the inequalities
$$
\tau_1(\Lambda_{t/\sqrt2}(D_1))\cdot\tau_2(\Lambda_{t/\sqrt2}(D_2)) \leq \nu(Q_{t/\sqrt2}) \leq
\nu(B_t)\leq\nu(Q_t)\leq\tau_1(\Lambda_t(D_1))\cdot\tau_2(\Lambda_t(D_2)),
$$
from which we get
\begin{align*}
\liminf_{t\to\infty}\frac{\log\Lambda_t(D_1)}{\log t}+\limsup_{t\to\infty}\frac{\log\Lambda_t(D_2)}{\log t}
& \leq \limsup_{t\to\infty}\frac{\log\Lambda_t(D)}{\log t} \\
& \leq \limsup_{t\to\infty}\frac{\log\Lambda_t(D_1)}{\log t}+\limsup_{t\to\infty}\frac{\log\Lambda_t(D_2)}{\log t}.
\end{align*}
\end{proof}
\subsection{The cases of the crossed products}\label{subsec:CrossedProd}
Let ${\mathcal A}$ be a unital $C^*$-algebra, $\alpha\in \Aut({\mathcal A})$ a unital automorphism, and $({\mathcal L},{\mathcal H},D;{\mathcal M},\tau)$ a semifinite spectral triple on ${\mathcal A}$. Assume that $\alpha$ is Lip-bounded, that is $\alpha({\mathcal L})\subset{\mathcal L}$, and, for any $a\in{\mathcal L}$, $\sup_{n\in{\mathbb Z}} \| [D,\alpha^{-n}(a)] \| < \infty$. Then, following \cite{BMR}, we can construct a semifinite spectral triple $({\mathcal L}_\rtimes,{\mathcal H}_\rtimes,D_\rtimes;{\mathcal M}_\rtimes,\tau_\rtimes)$ on the crossed product $C^*$-algebra ${\mathcal A}\rtimes_\alpha {\mathbb Z} = \langle \widetilde{\pi_u}({\mathcal A}),U \rangle$, which is defined as follows:
\begin{itemize}
\item[$(1)$] if $({\mathcal L},{\mathcal H},D,\G;{\mathcal M},\tau)$ is even,
\begin{flalign*}
& {\mathcal L}_\rtimes := {}^*\mathrm{alg}(\widetilde{\pi_u}({\mathcal L}),U), & \qquad & {\mathcal H}_\rtimes := {\mathcal H} \otimes \ell^2({\mathbb Z}), & \\
& D_\rtimes := D \otimes I + \G \otimes D_{\mathbb Z}, & \qquad & \G_\rtimes := I\otimes I, & \\
& {\mathcal M}_\rtimes := {\mathcal M} \otimes {\mathcal B}(\ell^2({\mathbb Z})), & \qquad & \tau_\rtimes := \tau \otimes Tr, &
\end{flalign*}
where ${}^*\mathrm{alg}(\widetilde{\pi_u}({\mathcal L}),U)$ is the $^*$-algebra generated by $\widetilde{\pi_u}({\mathcal L})$ and $U$, $(D_{\mathbb Z} \xi)(n) := n\xi(n)$, $\forall \xi\in\ell^2({\mathbb Z})$, and $Tr$ is the usual trace on ${\mathcal B}(\ell^2({\mathbb Z}))$,
\item[$(2)$] if $({\mathcal L},{\mathcal H},\G;{\mathcal M},\tau)$ is odd,
\begin{flalign*}
& {\mathcal L}_\rtimes := {}^*\mathrm{alg}(\widetilde{\pi_u}({\mathcal L}),U), & \qquad & {\mathcal H}_\rtimes := {\mathcal H} \otimes \ell^2({\mathbb Z}) \otimes {\mathbb C}^2, & \\
& D_\rtimes := D \otimes I \otimes \eps_1 + I \otimes D_{\mathbb Z} \otimes \eps_2, & \qquad & \G_\rtimes := I \otimes I \otimes \eps_3, & \\
& {\mathcal M}_\rtimes := {\mathcal M} \otimes {\mathcal B}(\ell^2({\mathbb Z})) \otimes M_2({\mathbb C}), & \qquad & \tau_\rtimes := \tau \otimes Tr \otimes tr, &
\end{flalign*}
where $tr$ is the normalized trace on $M_2({\mathbb C})$.
\end{itemize}
In case $\alpha$ satisfies a weaker condition, we have the following result.
\begin{definition}\label{semiequi}
Let ${\mathcal A}$ be a unital C*-algebra, $\alpha\in \Aut({\mathcal A})$ a unital automorphism, $({\mathcal L},{\mathcal H},D)$ a spectral triple on ${\mathcal A}$ such that $\alpha({\mathcal L})\subset{\mathcal L}$.
The automorphism is said to be Lip-semibounded if
$$
\sup_{n\in{\mathbb N}} \| [D,\alpha^{-n}(a)] \| < \infty, \qquad \forall a\in{\mathcal L}.
$$
\end{definition}
\begin{proposition}
Let ${\mathcal A}$ be a unital $C^*$-algebra, $\alpha\in \Aut({\mathcal A})$ a unital automorphism, $({\mathcal L},{\mathcal H},D;{\mathcal M},\tau)$ a semifinite spectral triple on ${\mathcal A}$, and assume $\alpha$ is Lip-semibounded. Then we can construct a semifinite spectral triple $({\mathcal L}_\rtimes,{\mathcal H}_\rtimes,D_\rtimes;{\mathcal M}_\rtimes,\tau_\rtimes)$ on the crossed product $C^*$-algebra ${\mathcal A}\rtimes_\alpha {\mathbb N} = \langle \zci_{\mathcal A}({\mathcal A}),t \rangle$, which is defined as follows:
\begin{itemize}
\item[$(1)$] if $({\mathcal L},{\mathcal H},D,\G;{\mathcal M},\tau)$ is even,
\begin{flalign*}
& {\mathcal L}_\rtimes := {}^*\mathrm{alg}(\zci_{\mathcal A}({\mathcal L}),t), & \qquad & {\mathcal H}_\rtimes := {\mathcal H} \otimes \ell^2({\mathbb N}), & \\
& D_\rtimes := D \otimes I + \G \otimes D_{\mathbb N}, & \qquad & \G_\rtimes := I\otimes I, & \\
& {\mathcal M}_\rtimes := {\mathcal M} \otimes {\mathcal B}(\ell^2({\mathbb N})), & \qquad & \tau_\rtimes := \tau \otimes Tr, &
\end{flalign*}
where ${}^*\mathrm{alg}(\zci_{\mathcal A}({\mathcal L}),t)$ is the $^*$-algebra generated by $\zci_{\mathcal A}({\mathcal L})$ and $t$, $(D_{\mathbb N} \xi)(n) := n\xi(n)$, $\forall \xi\in\ell^2({\mathbb N})$, and $Tr$ is the usual trace on ${\mathcal B}(\ell^2({\mathbb N}))$,
\item[$(2)$] if $({\mathcal L},{\mathcal H},\G;{\mathcal M},\tau)$ is odd,
\begin{flalign*}
& {\mathcal L}_\rtimes := {}^*\mathrm{alg}(\zci_{\mathcal A}({\mathcal L}),t), & \qquad & {\mathcal H}_\rtimes := {\mathcal H} \otimes \ell^2({\mathbb N}) \otimes {\mathbb C}^2, & \\
& D_\rtimes := D \otimes I \otimes \eps_1 + I \otimes D_{\mathbb N} \otimes \eps_2, & \qquad & \G_\rtimes := I \otimes I \otimes \eps_3, & \\
& {\mathcal M}_\rtimes := {\mathcal M} \otimes {\mathcal B}(\ell^2({\mathbb N})) \otimes M_2({\mathbb C}), & \qquad & \tau_\rtimes := \tau \otimes Tr \otimes tr, &
\end{flalign*}
where $tr$ is the normalized trace on $M_2({\mathbb C})$.
\noindent Moreover, in both cases, if $\som$ is the dimension of the original spectral triple, then the dimension of the new spectral triple is $\som+1$.
\end{itemize}
\end{proposition}
\begin{proof}
We only prove the even case, the odd case being similar. Let us first observe that, since $\alpha$ is an automorphism, ${\mathcal A}_\infty = {\mathcal A}$, $\alpha_\infty=\alpha$, and $\zci_{\mathcal A}(a) = \ppcov(a)p$, $\forall a\in{\mathcal A}$. Let $\pi:{\mathcal A}\to{\mathcal B}({\mathcal H})$ be the representation implied by the spectral triple $({\mathcal L},{\mathcal H},D,\G;{\mathcal M},\tau)$, and consider $(\widetilde{\pi}(a)\xi)(n) := \pi(\alpha^{-n}(a))\xi(n)$, $\forall a\in{\mathcal A}$, $\xi\in{\mathcal H}\otimes\ell^2({\mathbb N})$, $n\in{\mathbb N}$, which is a representation of ${\mathcal A}$ on ${\mathcal H}\otimes\ell^2({\mathbb N})$, and the shift operator
$$
(W\xi)(n) :=
\begin{cases}
0, & n=0,\\
\xi(n-1), & n\geq 1.
\end{cases}\
$$
Then, it is easy to see that $({\mathcal H}\otimes\ell^2({\mathbb N}),\widetilde{\pi},W)$ is a covariant representation of $({\mathcal A},\alpha,{\mathbb N})$ on ${\mathcal H}\otimes\ell^2({\mathbb N})$, in the sense of Definition \ref{Def:CovRep}. Therefore it induces a non-degenerate representation $\widehat{\pi}$ of ${\mathcal A}\rtimes_\alpha {\mathbb N} = \langle \zci({\mathcal A}), t \rangle$ on ${\mathcal H}\otimes\ell^2({\mathbb N})$, such that $\widehat{\pi}\circ \zci_{\mathcal A} = \widetilde{\pi}$, and $\widehat{\pi}(t)=W$. Hence $\widehat{\pi}({\mathcal A}\rtimes_\alpha {\mathbb N}) \subset {\mathcal M}_\rtimes$, while the facts that $D_\rtimes\widehat{\in} {\mathcal M}_\rtimes$, and $(1+D_\rtimes^2)^{-1}$ is $\tau_\rtimes$-compact follow from Proposition \ref{tensorProductTriple}. It remains to prove that $\| [ D_\rtimes, \widehat{\pi}(a) ] \| < \infty$, $\forall a\in{\mathcal L}_\rtimes$. Since the commutators $[\G\otimes D_{\mathbb N},\widehat{\pi}(a)]$ and $[D\otimes I,W]$ vanish, while $\| [\G\otimes D_{\mathbb N}, W] \| \leq 1$, it is enough to estimate the commutators $\| [D\otimes I,\widehat{\pi}(a) ] \| = \| \diag \{ [D,\pi(\alpha^{-n}(a)) ] : n\in{\mathbb N} \} \| = \sup_{n\in{\mathbb N}} \| [D,\pi(\alpha^{-n}(a)) ] \| < \infty$, and the claim follows.
\noindent We now prove the statement about the dimension, which in turn implies (again) the $\tau$-compactness of the resolvent.
By Proposition \ref{dimension-Lambda}, the Hausdorff dimension of $D_\rtimes$ is given by
\begin{displaymath}
\limsup_{t\to +\infty}\frac{\log(\Lambda_t(D_\rtimes))}{\log t}.
\end{displaymath}
We observe that $\Lambda_{t}(D_{\mathbb N})=[t]$ and thus
$$
\limsup_{t\to \infty} \frac{\log \Lambda_t (D_{\mathbb N})}{\log t}=\lim_{t\to \infty} \frac{\log ([t])}{\log t }=1.
$$
Now by applying Proposition \ref{tensorProductTriple} we are done.
\end{proof}
The next result has to do with the case of crossed products with respect to endomorphisms.
\begin{theorem} \label{triple-cross-prod-N}
Let ${\mathcal A}$ be a unital $C^*$-algebra, $\alpha\in \End(A)$ an injective, unital $*$-endomorphism, ${\mathcal A}_\infty=\varinjlim {\mathcal A}$ the inductive limit described in \eqref{eq:CstarIndLim1}, and $({\mathcal L}_\infty,{\mathcal H}_\infty,D_{\infty};{\mathcal M}_\infty,\tau_\infty)$ a semifinite spectral triple of dimension $p$ on ${\mathcal A}_\infty$. If the morphism $\alpha_\infty\in {\rm Aut}({\mathcal A}_\infty)$ is Lip-semibounded, then
there exists a semifinite spectral triple $({\mathcal L}_\rtimes,{\mathcal H}_\rtimes,D_\rtimes;{\mathcal M}_\rtimes,\tau_\rtimes)$ of dimension $p+1$ on the crossed product $C^*$-algebra ${\mathcal A}\rtimes_\alpha {\mathbb N}$.
\end{theorem}
\begin{proof}
Note that ${\mathcal A}\rtimes_\alpha {\mathbb N}={\mathcal A}_\infty\rtimes_{\alpha_\infty} {\mathbb N}$. Now the claim follows by applying the previous proposition.
\end{proof}
\section[Spectral triples for crossed products]{Spectral triples for crossed products generated by self-coverings}
In this section we exhibit some examples of semifinite spectral triples for crossed products with respect to an endomorphism: the self-covering of a $p$-torus, the self-covering of the rational rotation algebra, the endomorphism UHF algebra given by the shift, and the self-covering of the Sierpi\'nski gasket.
In this paper we consider two pictures of the inductive limits. One is what we call the Cuntz picture. The other one deals with an increasing sequence of algebras ${\mathcal A}_i$ with the morphisms $\varphi_i: {\mathcal A}_i\to {\mathcal A}_{i+1}$ being the inclusions, which entails that the morphisms $\alpha_i: {\mathcal A}_i\to{\mathcal A}_i$ are injective.
The following result gives a more detailed description of the second picture.
\begin{proposition}\label{prop41}
Given a family of algebras $\{{\mathcal A}_i\}_{i\geq 1}$, a morphism $\alpha_1: {\mathcal A}_1\to{\mathcal A}_1$, a collection of isomorphisms $\beta_i: {\mathcal A}_i\to{\mathcal A}_{i+1}$ for all $i\in {\mathbb N}$, one can obtain the following commuting diagram
\begin{equation*}
\xymatrix{
& {\mathcal A}_1 \ar[rr]^{ \varphi_1 } \ar[dd]^{\alpha_1} && {\mathcal A}_2 \ar[rr]^{ \varphi_2 } \ar[dd]^{\alpha_2} && {\mathcal A}_3 \ar[rr]^{ \varphi_3 } \ar[dd]^{\alpha_3} && \cdots \ar[r] & {\mathcal A}_\infty \ar[dd]^{\alpha_\infty} \\
&&&&&&&&\\
& {\mathcal A}_1 \ar[rr]^{ \varphi_1 } \ar[uurr]^{\beta_1} && {\mathcal A}_2 \ar[rr]^{ \varphi_2 } \ar[uurr]^{\beta_2} && {\mathcal A}_3 \ar[rr]^{ \varphi_3 } \ar[uurr]^{\beta_3} && \cdots \ar[r] & {\mathcal A}_\infty
}
\end{equation*}
where the morphisms $\alpha_i: {\mathcal A}_i\to{\mathcal A}_i$ are defined by the formula $\alpha_i :=\beta_{i-1}\circ \alpha_{i-1}\circ \beta_{i-1}^{-1}$ for $i\geq 2$,
$\varphi_1 := \beta_1\circ \alpha_1$, $\varphi_i := \alpha_{i+1}\circ \beta_i =\beta_i\circ \alpha_i$ for $i\geq 2$.
Moreover, the morphisms $\{\varphi_i\}_{i\geq 1}$ give rise to an inductive limit that we denote by ${\mathcal A}_\infty$ and
the former morphisms $\{\alpha_i\}_{i\geq 1}$ and $\{\beta_i\}_{i\geq 1}$ induce morphisms $\alpha_\infty, \beta_\infty: {\mathcal A}_\infty\to{\mathcal A}_\infty$ that are inverses of each other.
\end{proposition}
\begin{proof}
The first part of the statement, namely the one concerning the commuting diagram, follows by direct computations.
Now we take care of the second part concerning the morphisms $\alpha_\infty$ and $\beta_\infty$. We observe that
\begin{align*}
\alpha_\infty(f_1, f_2, \ldots) & = (\alpha_1(f_1), \alpha_{2}(f_2),\ldots)\\
\beta_\infty(f_1, f_2, \ldots) & = (0,\beta_{1}(f_1), \beta_{2} (f_2),\ldots)
\end{align*}
for all $(f_1, f_2,\ldots)\in{\mathcal A}_\infty$.
On the one hand, we have that
\begin{align*}
\alpha_\infty\circ \beta_\infty(f_1, f_2, \ldots) & = \alpha_\infty(0,\beta_{1}(f_1), \beta_{2} (f_2),\ldots)\\
& = (0,\alpha_2\circ\beta_{1}(f_1), \alpha_3\circ\beta_{2} (f_2),\ldots)\; .
\end{align*}
On the other hand, we have that
\begin{align*}
\beta_\infty\circ\alpha_\infty (f_1, f_2, \ldots) & = \beta_\infty(\alpha_1(f_1), \alpha_{2}(f_2),\ldots)\\
& = (0,\beta_1\circ \alpha_1(f_1), \beta_2\circ\alpha_{2}(f_2),\ldots)\; .
\end{align*}
Since $ \alpha_{i+1}\circ \beta_i =\beta_i\circ \alpha_i$ we are done.
\end{proof}
\begin{notation}\label{notazioneBs}
Before the discussion of the examples, we introduce some notation. We will consider an invertible matrix $B\in M_p({\mathbb Z})$ and we will set $A:=(B^T)^{-1}$. The following exact sequence will play a role in the definition of some of the Dirac operators
$$
0\to {\mathbb Z}^p\to A{\mathbb Z}^p\to \widehat{{\mathbb Z}_B}:= A{\mathbb Z}^p/{\mathbb Z}^p\to 0\; .
$$
Moreover, we will consider a section $s: \widehat{{\mathbb Z}_B}\to A{\mathbb Z}^p$ such that $s(\cdot)\in [0,1)^p$. We set $s_h(x):=A^{h-1}s(x)$ as in \cite{AiGuIs01}, p. 1387-1388. Note that $|\widehat{{\mathbb Z}_B}|=|\det(B)|=:r$.
\end{notation}
\medskip
\subsection{The crossed product for the self-coverings of the $p$-torus}
We begin with the case of tori.
The $p$-torus ${\mathbb T}^p :=\mathbb{R}^p/\mathbb{Z}^p$ can be endowed with a Dirac operator acting on the Hilbert space ${\mathcal H}_0 := {\mathbb C}^{2^{[p/2]}} \otimes L^2({\mathbb T}^p,dm)$
$$
D_0 :=-i\sum_{a=1}^p \eps_a \otimes \partial^a,
$$
where the matrices $\eps_a= (\eps_a)^*\in M_{2^{[p/2]}}({\mathbb C})$, $\eps_a \eps_b + \eps_b \eps_a=2 \delta_{a,b}$, furnish a representation of the Clifford algebra for the $p$-torus (see \cite{Spin} for more information on Dirac operators).
Then, we may consider the following spectral triple
$$
({\mathcal L}_0 :=C^1({\mathbb T}^p), {\mathcal H}_0, D_0).
$$
We recall that the spectral triple considered for the torus is even precisely when $p$ is even.
With the above notation and $B\in M_p({\mathbb Z})$, let $\pi:t\in{\mathbb T}^p\mapsto Bt\in{\mathbb T}^p$ be the self-covering, $\alpha(f)(t)=f(Bt)$ the associated endomorphism of ${\mathcal A} =C({\mathbb T}^p)$.
Then we consider the inductive system \eqref{eq:CstarIndLim1} and construct the inductive limit ${\mathcal A}_\infty=\displaystyle\varinjlim{\mathcal A}_n$.
An alternative description is given by the following isomorphic inductive family: ${\mathcal A}_n$ consists of continuous $B^n{\mathbb Z}^p$-periodic functions on ${\mathbb R}^p$, and the embedding is the inclusion.
In the following we denote by ${\mathbb T}_n$ the $p$-torus ${\mathbb R}^p/B^n{\mathbb Z}^p$.
Assume now that $B$ is purely expanding, namely $\|B^n v\|$ goes to infinity
for all vectors $v\neq 0$, hence $\|A\|<1$, where
$A=(B^T)^{-1}$.
In \cite{AiGuIs01}, we produced a semifinite spectral triple on ${\mathcal A}_\infty=\varinjlim C({\mathbb T}_n)$. More precisely, we constructed a Dirac operator\footnote{The symbol $s_h(\cdot)$ denotes the section defined in Notation \ref{notazioneBs}.} $D_\infty$ acting on ${\mathcal H}_\infty:= {\mathbb C}^{2^{[p/2]}} \otimes L^2({\mathbb T}^p,dm)\otimes L^2({\mathcal R},\tau)$
$$
D_\infty := D_0 \otimes I - 2\pi \sum_{a=1}^p \eps_a \otimes I \otimes \bigg( \sum_{h=1}^\infty I^{\otimes h-1} \otimes \diag(s_{h}(\cdot)^a) \bigg),
$$
the algebra ${\mathcal L}_\infty := \cup_{n\in{\mathbb N}} C^1({\mathbb T}_n)\subset {\mathcal A}_\infty$ embeds into the injective limit
$$
\varinjlim {\mathcal B}({\mathcal H}_0)\otimes M_{{\molt}^n}({\mathbb C}) = {\mathcal B}({\mathbb C}^{2^{[p/2]}} \otimes L^2({\mathbb T}^p,dm)) \otimes \mathrm{UHF}_r\; ,
$$
where $\mathrm{UHF}_r$ denotes the infinite tensor product of $M_r({\mathbb C})$, see Section 4.3 for more details. The C$^*$-algebra ${\mathcal B}({\mathbb C}^{2^{[p/2]}} \otimes L^2({\mathbb T}^p,dm)) \otimes \mathrm{UHF}_r$ in turn embeds into ${\mathcal M}_\infty := {\mathcal B}({\mathbb C}^{2^{[p/2]}} \otimes L^2({\mathbb T}^p,dm)) \otimes {\mathcal R}$, where ${\mathcal R}$ denotes the unique injective type II$_1$ factor obtained as the weak closure of the UHF algebra in the GNS representation of the unital trace, and we denote by $\tau_\infty := Tr\otimes\tau_{\mathcal R}$ the trace on ${\mathcal M}_\infty$. Then
$( {\mathcal L}_\infty, {\mathcal H}_\infty,D_\infty; {\mathcal M}_\infty, \tau_\infty)$ is a finitely summable, semifinite, spectral triple on $\varinjlim {\mathcal A}_n$, with Hausdorff dimension $p$.
\begin{theorem}\label{teo-p-toro}
Under the above hypotheses and with the notation of the former section, $C({\mathbb T}^p)\rtimes_\alpha{\mathbb N}$ can be endowed with the finitely summable semifinite spectral triple $({\mathcal L}_\rtimes, {\mathcal H}_\rtimes,D_\rtimes; {\mathcal M}_\rtimes, \tau_\rtimes)$ of Theorem \ref{triple-cross-prod-N}, with Hausdorff dimension $p+1$.
\end{theorem}
\begin{proof}
In order to construct a spectral triple on $C({\mathbb T}^p)\rtimes_\alpha \mathbb{N}$, according to Theorem \ref{triple-cross-prod-N}, we only need to check that $\alpha_\infty$ is Lip-semibounded, that is
\begin{displaymath}
\sup \{\Arrowvert [D_\infty ,\alpha_\infty^{-n}(f)]\Arrowvert, n\in{\mathbb N}\}<\infty, \quad \forall f\in {\mathcal L}_\infty = \cup_{n\in{\mathbb N}} C^1({\mathbb T}_n).
\end{displaymath}
Let $f\in C^1({\mathbb T}_k)$.
As observed in \cite{AiGuIs01}, the seminorms $L_{D_\infty}$, $L_{D_1}$, $L_{D_2}$, \ldots \; are compatible and we have that
$$
\| [D_\infty ,\alpha_\infty^{-n}(f)] \| = \| [D_0,f\circ B^{-n}] \|
$$
Moreover, by using the relation $\eps_a \eps_b + \eps_b \eps_a=2 \delta_{a,b}$ we obtain the following equalities
\begin{align*}
\| [D_0,f] \|^2 & = \left\| \sum_{a=1}^p \eps_a \otimes \partial^a(f) \right\|^2\\
& = \left\| \left(\sum_{a=1}^p \eps_a \otimes \partial^a(f)\right)^* \left(\sum_{a=1}^p \eps_a \otimes \partial^a(f)\right)\right\| \\
& = \left\| \sum_{a=1}^p (\eps_a)^2 \otimes |\partial^a f|^2\right\| = \left\| \sum_{a=1}^p 1 \otimes |\partial^a f|^2\right\|
\end{align*}
Now we compute $\| [D_0,f\circ B^{-n}] \|$. Setting $X=B^{-n}$ for simplicity, we have that
\begin{align*}
\| [D_0,f\circ X] \|^2 & = \left\| \sum_{a=1}^p \eps_a \otimes \partial^a(f\circ X) \right\|^2 \\
& = \left\| \left(\sum_{a=1}^p \eps_a \otimes \partial^a(f\circ X)\right)^* \left(\sum_{a=1}^p \eps_a \otimes \partial^a(f\circ X)\right)\right\| \\
& = \left\| \left(\sum_{a=1}^p \eps_a \otimes \left( \sum_{i=1}^p X_{a,i} (\partial^i \overline{f})\circ X \right) \right) \left(\sum_{b=1}^p \eps_b \otimes \left( \sum_{j=1}^p X_{b,j}(\partial^j f)\circ X \right) \right)\right\| \\
& = \left\| \sum_{a=1}^p (\eps_a)^2 \otimes \sum_{i,j=1}^p X_{a,i} X_{a,j} (\partial^i \overline{f})\circ X \cdot (\partial^j f)\circ X \right.\\
& \quad + \sum_{a<b} \eps_a \eps_b \otimes \sum_{i,j=1}^p X_{a,i} X_{b,j} (\partial^i \overline{f})\circ X \cdot (\partial^j f)\circ X \\
& \quad \left. + \sum_{a>b} \eps_a \eps_b \otimes \sum_{i,j=1}^p X_{a,i} X_{b,j} (\partial^i \overline{f})\circ X \cdot (\partial^j f)\circ X \right\| \\
& = \left\| \sum_{a=1}^p 1 \otimes \left( \sum_{i,j=1}^p X_{a,i} X_{a,j} (\partial^i \overline{f})\circ X \cdot (\partial^j f)\circ X \right) \right\| \\
& = \| \big( (\nabla f)\circ X, X^* X (\nabla f)\circ X \big)\| \\
& \leq \| X^* X \| \left\| \sum_{a=1}^p 1 \otimes (\partial^a f)^2\right\| = \| X\|^2 \| [D,f] \|^2 \; .
\end{align*}
These computations and the hypothesis on $B$ being purely expanding (cf. Proposition 2.6 in \cite{AiGuIs01}) imply that
\begin{displaymath}
\sup\{\Arrowvert [D_\infty ,\alpha_\infty^{-n}(f)]\Arrowvert, n\in{\mathbb N}\}\leq\sup\{ \Arrowvert B^{-n}\Arrowvert \Arrowvert [D_\infty,f]\Arrowvert, n\in{\mathbb N}\}<\infty\; .
\end{displaymath}
\end{proof}
\medskip
\subsection{The crossed product for the self-coverings of the rational rotation algebra}
The present example is associated with a regular noncommutative self-covering with finite abelian group of deck transformations \cite{AiGuIs01}.
\begin{definition} \label{def-reg-cov}
A finite (noncommutative) covering with abelian group is an inclusion of (unital) $C^*$-algebras ${\mathcal A}\subset {\mathcal B}$ together with an action of a finite abelian group $\Gamma$ on ${\mathcal B}$ such that ${\mathcal A}={\mathcal B}^\Gamma$. We will say that ${\mathcal B}$ is a covering of ${\mathcal A}$ with deck transformations given by the group $\Gamma$.
\end{definition}
We are now going to give a description of the rational rotation algebra making small modifications to the description of $A_\theta$, $\theta=p/q\in\mathbb{Q}$, seen in \cite{BEEK}. We observe that $A_\theta$ reduces to $C({\mathbb T}^2)$ in the case $\theta\in{\mathbb Z}$.
Consider the following matrices
\begin{eqnarray*}
(U_0)_{hk} = \delta_{h,k}e^{2\pi i(k-1) \theta }, \quad
(V_0)_{hk} = \delta_{h+1,k} +\delta_{h,q}\delta_{k,1} \in M_q(\mathbb{C})
\end{eqnarray*}
and $W_0(n) := U_0^{n_1}V_0^{n_2}$, for all $n=(n_1,n_2)\in{\mathbb Z}^2$. Let $p',p''\in{\mathbb N}$, $p',p''<q$, be such that $pp'+1=n'q$, $pp''-1=n''q$, for some $n',n''\in{\mathbb N}$, and introduce
$P := \begin{pmatrix}
0 & p' \\
p'' & 0
\end{pmatrix}
$, and
$$
\widetilde{\gamma} \def\G{\Gamma}_n(f)(t):=\textrm{ad}(W_0(P n))[f(t+n)]=V_0^{-p''n_1}U_0^{-p'n_2}f(t+n)U_0^{p'n_2}V_0^{p''n_1},
$$
for all $t\in{\mathbb R}^2$, $n\in{\mathbb Z}^2$. We have the following description of $A_\theta$ (cf. \cite{BEEK})
$$
A_\theta=\{f\in C(\mathbb{R}^2, M_q(\mathbb{C})) \, : \, f = \widetilde{\gamma} \def\G{\Gamma}_{n}(f), n\in{\mathbb Z}^2 \}.
$$
This algebra comes with a natural trace
$$
\tau(f):= \frac{1}{q}\int_{{\mathbb T}_0} \tr(f(t))dt,
$$
where we are considering the Haar measure on ${\mathbb T}_0:={\mathbb R}^2/B{\mathbb Z}^2$ and $\tr(A)=\sum_i a_{ii}$. We observe that the function $\tr(f(t))$ is ${\mathbb Z}^2$-periodic.
Define
\begin{align*}
U(t_1,t_2)&:=e^{-2\pi i t_1/q} U_0\\
V(t_1,t_2)&:=e^{-2\pi i t_2/q} V_0
\end{align*}
and
\begin{displaymath}
{\mathcal L}_\theta :=\left\{\sum_{r,s}a_{rs}U^rV^s : (a_{rs})\in S(\mathbb{Z}^2) \right\},
\end{displaymath}
where $S(\mathbb{Z}^2)$ is the set of rapidly decreasing sequences. It is clear that the derivations $\partial_1$ and $\partial_2$, defined as follows on the generators, extend to ${\mathcal L}_\theta$
\begin{eqnarray*}
\partial_1(U^hV^k)&=&2\pi ihU^hV^k\\
\partial_2(U^hV^k)&=&2\pi ikU^hV^k.
\end{eqnarray*}
Moreover, the above derivations extend to densely defined derivations both on $A_\theta$ and $L^2(A_\theta,\tau)$.
We still denote these extensions with the same symbols.
We may consider the following spectral triple (see \cite{GBFV})
\begin{eqnarray*}
({\mathcal L}_0:={\mathcal L}_\theta , {\mathcal H}_0:={\mathbb C}^{2} \otimes L^2(A_\theta,\tau), D_0 :=-i(\eps_1 \otimes \partial_1+\eps_2 \otimes \partial_2)),
\end{eqnarray*}
where $\eps_1, \eps_2$ denote the Pauli matrices.
Given the integer-valued matrix $B\in M_2({\mathbb Z})$ such that $\det(B)\equiv_q 1$, there is an associated endomorphism $\alpha: A_\theta\to A_\theta$ defined by $\alpha(f)(t)=f(Bt)$, \cite{Stacey}. Then, we consider the inductive limit ${\mathcal A}_\infty=\displaystyle\varinjlim{\mathcal A}_n$ as in \eqref{eq:CstarIndLim1}.
As in the case of the torus one can consider the following isomorphic inductive family: ${\mathcal A}_n$ consists of continuous $B^n{\mathbb Z}^2$-invariant matrix-valued functions on ${\mathbb R}^2$, i.e
$$
{\mathcal A}_n:=\{f\in C(\mathbb{R}^2, M_q(\mathbb{C})) \, : \, f =\widetilde{\gamma} \def\G{\Gamma}_{B^n k}(f), k\in{\mathbb Z}^2 \},
$$
with trace
$$
\tau_n(f)=\frac{1}{q |\!\det B^n|}\int_{{\mathbb T}_n}\tr(f(t))dt,
$$
and the embedding is unital inclusion $\alpha_{n+1,n}: {\mathcal A}_n\hookrightarrow {\mathcal A}_{n+1}$. In particular, ${\mathcal A}_0={\mathcal A}$, and ${\mathcal A}_1={\mathcal B}$. This means that ${\mathcal A}_\infty$ may be considered as a solenoid $C^*$-algebra (cf. \cite{McCord}, \cite{LP2}).
On the $n$-th noncommutative covering ${\mathcal A}_n$, the formula of the Dirac operator doesn't change and we can consider the following spectral triple
\begin{eqnarray*}
({\mathcal L}_\theta^{(n)} , {\mathbb C}^{2} \otimes L^2({\mathcal A}_n,\tau), D=-i(\eps_1 \otimes \partial_1 + \eps_2 \otimes \partial_2)).
\end{eqnarray*}
In \cite{AiGuIs01}, we produced a semifinite spectral triple on ${\mathcal A}_\infty=\varinjlim {\mathcal A}_n$. More precisely, we constructed a Dirac operator $D_\infty$ acting on ${\mathcal H}_\infty:= {\mathbb C}^{2} \otimes L^2({\mathcal A}_0,\tau_0)\otimes L^2({\mathcal R},\tau)$
\[
D_\infty := D_0 \otimes I - 2\pi \sum_{a=1}^2 \eps_a \otimes I \otimes \bigg( \sum_{h=1}^\infty I^{\otimes h-1} \otimes \diag(s_{h}(\cdot)^a) \bigg),
\]
the algebra ${\mathcal A}_\infty$ embeds into the injective limit
$$
\varinjlim {\mathcal B}({\mathbb C}^{2} \otimes L^2({\mathcal A}_0,\tau_0))\otimes M_{{\molt}^n}({\mathbb C}) = {\mathcal B}({\mathbb C}^{2} \otimes L^2({\mathcal A}_0,\tau_0)) \otimes \mathrm{UHF}_{\molt}
$$
which in turn embeds into ${\mathcal M}_\infty := {\mathcal B}({\mathbb C}^{2} \otimes L^2({\mathcal A}_0,\tau_0)) \otimes {\mathcal R}$, which is endowed with the trace $\tau_\infty := Tr\otimes\tau_{\mathcal R}$. Then
$({\mathcal L}_\infty, {\mathcal H}_\infty,D_\infty; {\mathcal M}_\infty, \tau_\infty)$ is a finitely summable, semifinite, spectral triple on $\varinjlim {\mathcal A}_n$, with Hausdorff dimension $2$ (\cite{AiGuIs01}, Theorem 3.7).
\begin{theorem}
Under the above hypotheses and with the notation of the former section, $A_\theta\rtimes_\alpha{\mathbb N}$ can be endowed with the finitely summable semifinite spectral triple $({\mathcal L}_\rtimes, {\mathcal H}_\rtimes,D_\rtimes; {\mathcal M}_\rtimes, \tau_\rtimes)$ of Theorem \ref{triple-cross-prod-N}, with Hausdorff dimension $3$.
\end{theorem}
\begin{proof}
According to Theorem \ref{triple-cross-prod-N} we only need to check that $\alpha_\infty$ is Lip-semibounded, that is
$$
\sup\{\Arrowvert [D_\infty,\alpha_\infty^{-n}(f)]\Arrowvert, n\in{\mathbb N}\} < \infty, \quad \forall f\in {\mathcal L}_\infty \; .
$$
This is true because similar computations to those in the proof of Theorem \ref{teo-p-toro} yield
\begin{align*}
\sup\{\Arrowvert [D_\infty,\alpha_\infty^{-n}(f)]\Arrowvert, n\in{\mathbb N}\}
\leq \sup\{ \| B^{-n}\| \Arrowvert [D_\infty,f]\|, n\in{\mathbb N}\}
\end{align*}
The hypothesis of $B$ being purely expanding ensures that $\sup\{\Arrowvert [D_\infty,\alpha_\infty^{-n}(f)]\Arrowvert, n\in{\mathbb N}\}$ is finite.
\end{proof}
\medskip
\subsection{The crossed product for the shift-endomorphism of the UHF-algebra}\label{UHF}
Consider now the case of the UHF-algebra.
This algebra is defined as the inductive limit of the following sequence of finite dimensional matrix algebras:
\begin{eqnarray*}
M_0 & = & M_\molt(\mathbb{C})\\
M_n & = & M_{n-1}\otimes M_\molt(\mathbb{C}) \quad n\geq 1,
\end{eqnarray*}
with maps $\phi_{ij}: M_j\to M_i$ given by $\phi_{ij}(a_j)=a_j\otimes 1$.
We denote by ${\mathcal A}$ the $C^*$-algebra UHF$_r$ and set $M_{-1}=\mathbb{C}1_{\mathcal A}$ in the inductive limit defining the above algebra. The $C^*$-algebra ${\mathcal A}$ has a unique normalized trace that we denote by $\tau$.
Consider the projection $P_n:L^2({\mathcal A},\tau)\to L^2(M_n,{\rm Tr})$, where ${\rm Tr}: M_n \to \mathbb{C}$ is the normalized trace, and define
\begin{eqnarray*}
Q_n&:=& P_n-P_{n-1}, \quad n\geq 0,\\
E(x)&:=&\tau(x)1_{\mathcal A}\,.
\end{eqnarray*}
For any $s>1$, Christensen and Ivan \cite{Chris} defined the following spectral triple for the algebra UHF$_\molt$
\begin{eqnarray*}
({\mathcal L}_0, L^2({\mathcal A},\tau),D_0=\sum_{n\geq 0} \molt^{ns}Q_n )
\end{eqnarray*}
where ${\mathcal L}_0$ is the algebra consisting of the elements of ${\mathcal A}$ with bounded commutator with $D_0$. It was proved that for any such value of the parameter $s$, this spectral triple induces a metric which defines a topology equivalent to the weak$^*$-topology on the state space (\cite[Theorem 3.1]{Chris}).
We consider the endomorphism of ${\mathcal A}$ given by the right shift, $\alpha(x)=1\otimes x$. Then as in \eqref{eq:CstarIndLim1} we may consider the inductive limit ${\mathcal A}_\infty=\displaystyle\varinjlim{\mathcal A}_n$. As in the previous sections, we have the following isomorphic inductive family: ${\mathcal A}_i$ is defined as
\begin{eqnarray*}
{\mathcal A}_0&=& {\mathcal A};\\
{\mathcal A}_n &=& M_\molt(\mathbb{C})^{\otimes n} \otimes {\mathcal A}_0;\\
{\mathcal A}_\infty &=& \varinjlim {\mathcal A}_i
\end{eqnarray*}
and the embedding is the inclusion.
It is easy to see that ${\mathcal A}_\infty$ is again the UHF-algebra of the same type, since the corresponding supernatural number is the same.
In \cite{AiGuIs01}, we produced a semifinite spectral triple on $\varinjlim {\mathcal A}_n$. More precisely, we defined the following Dirac operator acting on ${\mathcal H}_\infty := L^2({\mathcal R}, \tau)\otimes L^2({\mathcal A}_0, \tau)$
\begin{equation}
D_{\infty}=I_{-\infty,-1} \otimes D_0+\sum_{k=1}^\infty \molt^{-sk} I_{-\infty,-k-1}\otimes F\otimes E,
\end{equation}
where $I_{-\infty,k}$ is the identity on the factors with indices in $[-\infty, k]$, $F: M_r({\mathbb C})\to M_r({\mathbb C})$ is defined as $F(x):=x-\tr(x)1$ for $x\in M_r({\mathbb C})$, and the algebra ${\mathcal A}_\infty$ embeds in the injective limit
$$
\varinjlim {\mathcal B}(L^2({\mathcal A}_0,\tau)) \otimes M_{{\molt}^n}({\mathbb C})={\mathcal B}(L^2({\mathcal A}_0,\tau))\otimes \mathrm{UHF}_r
$$
Set ${\mathcal L}_\infty = \cup_n{\mathcal L}_n$, ${\mathcal M}_\infty = {\mathcal R} \otimes {\mathcal B}(L^2({\mathcal A}_{0},\tau))$, $\tau_\infty:=\tau_{\mathcal R}\otimes Tr$.
Then $({\mathcal L}_\infty,{\mathcal H}_\infty,D_{\infty}; {\mathcal M}_\infty,\tau_\infty)$ is a finitely summable, semifinite, spectral triple, with Hausdorff dimension $2/s$ (\cite{AiGuIs01}, Theorem 5.6).
\begin{theorem}\label{UHFcrossedprod}
Under the above hypotheses and the notation of the former section, ${\rm UHF}_r\rtimes_\alpha {\mathbb N}$ can be endowed with the finitely summable semifinite spectral triple $({\mathcal L}_\rtimes, {\mathcal H}_\rtimes,D_\rtimes; {\mathcal M}_\rtimes,\tau_\rtimes)$ of Theorem \ref{triple-cross-prod-N}, with Hausdorff dimension $1+2/s$.
\end{theorem}
\begin{proof}
According to Theorem \ref{triple-cross-prod-N}, in order to construct a spectral triple on ${\mathcal A} \rtimes_\alpha{\mathbb N}$ we only need to check that $\alpha_\infty$ is Lip-semibounded, that is
\begin{displaymath}
\sup\{\Arrowvert [D_\infty,\alpha_\infty^{-k}(f)]\Arrowvert, k\in{\mathbb N}\}< \infty, \quad \forall f\in {\mathcal L}_\infty.
\end{displaymath}
This is true because
\begin{align*}
\Arrowvert [D_\infty,\alpha_\infty^{-k}(f)]\Arrowvert = \molt^{-ks}\Arrowvert [D_\infty,f]\Arrowvert.
\end{align*}
In fact, let $f=(\bigotimes_{k=-\infty}^{-n-1}I)\otimes a\in {\mathcal A}_n$,
$\alpha_\infty^{k}(f)=(\bigotimes_{j=-\infty}^{-n+k-1}I)\otimes a\in {\mathcal A}_{n-k}$ for $k\in{\mathbb Z}$.
The Hilbert space on which $D_\infty$ acts is the completion of ${\mathcal A}_\infty$.
On this Hilbert space, we consider the right shift on the factors and we denote it by $U_\alpha$. We set $\Phi := \textrm{ad} (U_\alpha)$. Then we have that
\begin{eqnarray*}
[D_\infty,\alpha_\infty^{-k}(f)] &=& \sum_{h\in{\mathbb Z}} \molt^{hs}[Q_h,\left(\bigotimes_{j=-\infty}^{-n-k-1}I\right)\otimes a]\\
&=& \Phi^{-k} \left(\sum_{h\in{\mathbb Z}} \molt^{hs}[Q_{k+h},\left(\bigotimes_{j=-\infty}^{n-1}I\right)\otimes a]\right)\\
&=& \molt^{-ks} \Phi^{-k}([D_\infty,f])
\end{eqnarray*}
where we used that $\Phi(Q_h)=Q_{h+1}$ and $\Phi\upharpoonright_{{\mathcal A}_\infty} = \alpha_\infty$.
\end{proof}
In the theorem above, we considered the $C^*$-algebra ${\rm UHF}_r\rtimes_\alpha {\mathbb N}$. We note that the crossed product of the UHF of type $2^\infty$ under the action of the bilateral shift, namely the $C^*$-algebra ${\rm UHF}_2 \rtimes_\alpha {\mathbb Z}$,
was studied in \cite{BKRS}.
\medskip
\subsection{The crossed product for the self-coverings of the Sierpi\'nski gasket}
We conclude this paper with the case of a self-covering of the Sierpi\'nski gasket that was studied by the authors in \cite{AGI3}.
The Sierpi\'nski gasket is the self-similar fractal determined by $3$ similarities with scaling parameter 1/2 centered in the vertices
$v_0=(0,0)$, $v_1=(1/2,\sqrt{3}/2)$, $v_2=(1,0)$,
namely
the non-empty, compact set $K$, such that
$$
K=\bigcup_{j=0,1,2}w_j(K),
$$
where $w_j$ is the dilation around $v_j$ with contraction parameter $1/2$ (see Figure \ref{fig:covering}).
Denote by $V_0(K)$ the set $\{v_0, v_1, v_2\}$, and let $E_0(K):=\{ (p,q) : p,q\in V_0, p\neq q\}$.
We call an element of the family $\{w_{i_1} \circ \dots \circ w_{i_k}(K):k\geq0\}$ a {\it cell}, and call its diameter the size of the cell.
We call an element of the family $E(K)=\{w_{i_1} \circ \dots \circ w_{i_k}(e):k\geq0, e\in E_{0}(K)\}$ an {\it (oriented) edge} of $K$ and we denote by $e^-$ (resp. $e^+$) the source (resp. the target) of the oriented edge $e$.
Note that a cell $C:=w_{i_1} \circ \dots \circ w_{i_k}(K)$ has $\mathrm{size}(C)=2^{-k}$ and, if $e_0\in E_{0}(K)$, then $e=w_{i_1} \circ \dots \circ w_{i_k}(e_0)$ has length $2^{-k}$.
In the following we shall consider $K_0:=K$, $E_0:=E(K)$, $K_n:=w_0^{-n}K_0$.
Let us now consider the middle point $x_{i,i+1}$ of the segment $(w_0^{-1}v_i,w_0^{-1}v_{i+1})$, $i=0,1,2$, the map $R_{i+1,i}:w_0^{-1}w_{i}K\to w_0^{-1}w_{i+1}K$ consisting of the rotation of $\frac43\pi$ around the point $x_{i,i+1}$, $i=0,1,2$.
We then construct the coverings $p:K_1\to K$ and $\phi: K\to K$ given by
$$
p(x)=
\begin{cases}
x,&x\in K,\\
R_{0,1}(x),&x\in w_0^{-1}w_1 K,\\
R_{0,2}(x),&x\in w_0^{-1}w_2 K,
\end{cases}
$$
and
\begin{displaymath}
\phi(x)=\left\{
\begin{array}{ll}
w_0^{-1}x & \textrm{if $x\in C_0$} \\
R_{0,1}(w_0^{-1}(x)) & \textrm{if $x\in C_1$}\\
R_{0,2}(w_0^{-1}(x)) & \textrm{if $x\in C_2$}\\
\end{array}
\right.\\
\end{displaymath}
Note that $p(x)=\phi(w_0(x))$ for all $x\in K_1$ (see Figure \ref{fig:covering}).
\begin{figure}
\scalebox{1.75}{\def\trianglewidth{2cm}%
\pgfdeclarelindenmayersystem{Sierpinski triangle}{
\symbol{X}{\pgflsystemdrawforward}
\symbol{Y}{\pgflsystemdrawforward}
\rule{X -> X-Y+X+Y-X}
\rule{Y -> YY}
}
\foreach \mathrm{level} in {6}{
\tikzset{
l-system={step=\trianglewidth/(2^\mathrm{level}), order=\mathrm{level}, angle=-120}
}%
\begin{tikzpicture}
\fill [black] (0,0) -- ++(0:\trianglewidth) -- ++(120:\trianglewidth) -- cycle;
\draw [draw=none] (0,0) l-system
[l-system={Sierpinski triangle, axiom=X},fill=white];
\node (bbb) at (-.05,-0.075) {$\scalebox{.5}{$v_0$}$};
\node (bbb) at (1,1.85) {$\scalebox{.5}{$v_2$}$};
\node (bbb) at (2.09,-.075) {$\scalebox{.5}{$v_1$}$};
\node (bbb) at (0,-.87) {$\scalebox{.5}{$\;$}$};
\end{tikzpicture}
}%
}
\scalebox{3}{
\def\trianglewidth{2cm}%
\foreach \mathrm{level} in {6}{
\tikzset{
l-system={step=\trianglewidth/(2^\mathrm{level}), order=\mathrm{level}, angle=-120}
}%
\begin{tikzpicture}
\fill [black] (0,0) -- ++(0:\trianglewidth) -- ++(120:\trianglewidth) -- cycle;
\draw [draw=none] (0,0) l-system
[l-system={Sierpinski triangle, axiom=X},fill=white];
\draw[line width=0.05mm, <-] (.75,-.1) to[out=-90,in=-90] (1.25,-.1);
\draw[line width=0.05mm, <-] (0.25,.6) to[out=135,in=120] (.5,1);
\node (bbb) at (-.05,-0.075) {$\scalebox{.25}{$v_0$}$};
\node (bbb) at (.395,.87) {$\scalebox{.25}{$x_{2,0}$}$};
\node (bbb) at (1.65,.87) {$\scalebox{.25}{$x_{1,2}$}$};
\node (bbb) at (1,1.85) {$\scalebox{.25}{$w_0^{-1}v_2$}$};
\node (bbb) at (1,-.075) {$\scalebox{.25}{$x_{0,1}=v_1$}$};
\node (bbb) at (2.07,-.075) {$\scalebox{.25}{$w_0^{-1}v_1$}$};
\node (bbb) at (0.1,.89) {$\scalebox{.25}{$R_{0,2}$}$};
\node (bbb) at (1.1,-.35) {$\scalebox{.25}{$R_{0,1}$}$};
\end{tikzpicture}
}
}
\label{fig:covering}
\caption{The Sierpi\'nski gasket $K=K_0$ and the covering map $p_1=p: K_1\to K$.}
\end{figure}
Similarly, for every $n\geq 0$, we define a family of coverings $p_n: K_{n+1}\to K_{n}$ and $\phi_n: K_n\to K_n$ by $p_{n+1} := w_0^{-n} \circ p \circ w_0^{n}$ and $\phi_n := w_0^{-n} \circ \phi \circ w_0^{n}$.
\begin{proposition}
The following diagrams are commutative
\begin{equation*}
\xymatrix{
& K_0 && K_1 \ar[ll]^{ p_1 } && K_2 \ar[ll]^{ p_2 } && \cdots \ar[ll]^{p_3} \\
&&&&&&&&\\
& K_0 \ar[uu]^{ \phi_0 } && K_1 \ar[ll]^{ p_1 } \ar[uu]^{ \phi_1 } && K_2 \ar[ll]^{ p_2 } \ar[uu]^{ \phi_2 } && \cdots \ar[ll]^{ p_3 }
}
\end{equation*}
\end{proposition}
\begin{proof}
Indeed, first note that $\phi_0 \circ p_1=\phi \circ p=p\circ \phi_1$ and $w_0\circ \phi_1=p_1$, which implies that $p_1\circ\phi_1=\phi_0\circ p_1$.
Then, for any $n\geq 1$ we have
\begin{align*}
p_n \circ \phi_n & = w_0^{-n+1} \circ p \circ w_0^{n-1} \circ w_0^{-n} \circ \phi \circ w_0^n= w_0^{-n+1} \circ p \circ w_0^{-1} \circ \phi \circ w_0 \circ w_0^{n-1} \\
& = w_0^{-n+1} \circ p_1 \circ \phi_1 \circ w_0^{n-1} = w_0^{-n+1} \circ \phi_0 \circ p_1 \circ w_0^{n-1} = \phi_{n-1} \circ p_n.
\end{align*}
\end{proof}
It follows that the maps $\{\phi_n\}_{n\geq 0}$ induce a map in the projective limit and by functoriality a map on $\varinjlim C(K_i)$ which we denote by $\alpha_\infty$.
An element $f\in C(K_n)$ can be seen in $\varinjlim C(K_i)$ as the sequence $[f]=(0_{n}, f, f\circ p_{n+1}, f\circ p_{n, n+2}, \ldots )$, where $p_{n, n+k}:= p_{n+1}\circ \cdots\circ p_{n+k}$.
Accordingly the map $\alpha_\infty$ reads as
$$
\alpha_\infty[f] := (0_{n}, f\circ \phi_n, f\circ p_{n+1} \circ \phi_{n+1}, f\circ p_{n, n+2} \circ \phi_{n+2}, \ldots ).
$$
By functoriality each $(\phi_n)^*: C(K_n)\to C(K_n)$ is a proper endomorphism, that is, it is injective, but not surjective.
With the notation of Proposition \ref{prop41}, we set $\beta_i$ equal to $w_0^*$ for all $i\geq 0$.
Thanks to Proposition 3.1, the map $\alpha_\infty$ is invertible and its inverse is given by
$$
\alpha_\infty^{-1}[f] := (0_{n+1}, f\circ w_0, f\circ p_{n+1} \circ w_0, f\circ p_{n, n+2} \circ w_0, \ldots ).
$$
Denote by $E_n:=\{w_0^{-n}e, e\in E(K)\}$, $E_\infty:=\cup_{n\geq 0} E_n$, $E^n:=\{e\in E_\infty, \mathrm{length}(e)=2^n\}$, $P^n$ the projection of $\ell_2(E_\infty)$ onto $\ell_2(E^n)$.
It was shown in \cite[Sec. 6]{AGI3} that ${\mathcal A}_\infty:=\varinjlim C(K_n)$ supports a semifinite spectral triple $({\mathcal L}_\infty,{\mathcal H}_\infty, D_\infty; {\mathcal M}_\infty, \tau_\infty)$,
where ${\mathcal M}_\infty:=\pi_\tau(B_\infty)''$ is a suitable closure of the geometric operators (see \cite[Sec. 5]{AGI3} for a precise definition),
$D_\infty:=F|D|: \ell^2(E_\infty)\to \ell^2(E_\infty)$,
$F$ is the orientation reversing operator on edges and
$$
|D_\infty|:=\sum_{n\in{\mathbb Z}}2^{-n}P^n.
$$
\begin{theorem}\label{teo-gasket}
Under the above hypotheses and with the notation of the former section, $C(K)\rtimes_\alpha{\mathbb N}$ can be endowed with the finitely summable semifinite spectral triple $({\mathcal L}_\rtimes, {\mathcal H}_\rtimes,D_\rtimes; {\mathcal M}_\rtimes, \tau_\rtimes)$ of Theorem \ref{triple-cross-prod-N}, with Hausdorff dimension $\log_23+1$.
\end{theorem}
\begin{proof}
According to Theorem \ref{triple-cross-prod-N}, in order to construct a spectral triple on $C(K)\rtimes_\alpha \mathbb{N}$ we only need to check that $\alpha_\infty$ is Lip-semibounded, that is
$$
\sup_{k\geq 0} \Arrowvert[D_\infty, \alpha_\infty^{-k}(f)]\Arrowvert <\infty, \quad \forall f\in {\mathcal L}_\infty := \cup_{n\geq 0}\ \mathrm{Lip}(K_n).
$$
We are going to show that for any $f\in C(K_n)$ it holds that
$$
\Arrowvert [D_\infty, \alpha_\infty^{-k}(f)] \Arrowvert =\frac{\Arrowvert [D_\infty, f]\Arrowvert}{2^k} \quad k\in \mathbb{N}.
$$
Indeed, since both $p_n$ and $\phi_n$ are isometries, we have that
\begin{eqnarray*}
\Arrowvert [D_\infty, \alpha_\infty^{-k}(f)] \Arrowvert &=& \left\Arrowvert \oplus_{e\in E_\infty} \frac{\alpha_\infty^{-k}(f)(e^+)-\alpha_\infty^{-k}(f)(e^-)}{l(e)} F \right\Arrowvert \\
&=& \left\Arrowvert \oplus_{e\in E_\infty} \frac{f(w_0^k(e^+))-f(w_0^k(e^-))}{l(e)} F \right\Arrowvert \\
&=& \left\Arrowvert \oplus_{e\in E_\infty} \frac{f(w_0^k(e^+))-f(w_0^k(e^-))}{2^k l(w_0^k(e))} F \right\Arrowvert \\
&=& \left\Arrowvert \oplus_{e'\in E_\infty} \frac{f(e'^+)-f(e'^-)}{2^k l(e')} F\right\Arrowvert \\
&=& \frac{\Arrowvert [D_\infty, f]\Arrowvert}{2^k}.
\end{eqnarray*}
\end{proof}
\section*{Acknowledgement}
We thank the referee for the attentive reading of this article and for useful suggestions.
This work was supported by the following institutions:
the ERC Advanced Grant 669240 QUEST "Quantum Algebraic Structures and Models",
the MIUR PRIN ``Operator Algebras, Noncommutative Geometry and Applications'',
the INdAM-CNRS GREFI GENCO, and the INdAM GNAMPA.
V. A. acknowledges the support by the Swiss National Science foundation through the SNF project no. 178756 (Fibred links, L-space covers and algorithmic knot theory).
D. G. and T. I. acknowledge the MIUR Excellence Department Project awarded to the Department of Mathematics, University of Rome Tor Vergata, CUP E83C18000100006.
|
\section{Background}
\subsection{Long range dependence}
LRD refers to a process where correlations decay slower over time such
that the future is non-trivially dependent on the past no matter how
far forward we proceed. A sample path of an LRD fractional Gaussian noise
process, with $H=0.8$, is shown in Figure~\ref{fig: fgn_sample_path}.
As is typical for LRD processes, there is the
appearance of long periods of upwards and downwards ``trends", even
though the process is stationary.
LRD can be defined in two equivalent ways, via the autocorrelation or
spectral density. These definitions are related through the Fourier
Transform~\cite[pg. 117]{brockwell_davis_1986}, and are equivalent via
the Kolmogorov Isomorphism Theorem (\cite{bingham2012}). The following
statement defines the concept of long range dependence in terms of its
autocorrelation function. For reference, the autocovariance function,
$\gamma(k) = \mathds{E}[\left(X_{n+k} - \mu\right)\left(X_n - \mu\right)]$ for a process with mean $\mu$, and autocorrelation coefficient for $k$ lags, $\rho(k)$, are related by
$\rho(k) = \gamma(k)/ \gamma(0) = \gamma(k)/\sigma^2$, where
$\sigma^2$ is the variance of the stochastic process. Note we are using the autocorrelation coefficient and the the autocorrelation function, $\mathds{E}[X_{n+k}X_n]$, this is consistent with many works in the analysis of LRD processes.
\begin{definition}
Let $\{X_n\}_{n \in \mathbb{N}}$ be a stationary process. If
there exists $\alpha \in (0,1)$, and $c_\gamma >
0$, such that the auto-covariance $\gamma(k)$ satisfies
\begin{align*}
\lim\limits_{k \rightarrow \infty} \frac{\gamma(k)}{c_\gamma k ^{-\alpha}} = 1,
\end{align*}
then we say that the process is long range dependent.
\label{def:lrd_alpha}
\end{definition}
\noindent The equivalent definition in the frequency domain considers
the limit of the spectral density near the origin.
\begin{definition}
Let $\{X_n\}_{n \in \mathbb{N}}$ be a stationary process. If
there exists $\beta \in (0,1)$, and $c_f > 0$,
such that the spectral density $f(\lambda)$ satisfies
\begin{align*}
\lim\limits_{\lambda \rightarrow 0} \frac{f(\lambda)}{c_f |\lambda| ^{-\beta}} = 1,
\end{align*}
then we say that the process is long range dependent.
\end{definition}
The concept of LRD is often characterised by the Hurst parameter, $H$, which measures the strength of the correlations between the past and present of a stochastic process. It is
perhaps unfortunate that $H$ is used as standard notation both for the
Hurst parameter, and for Shannon entropy, but we shall side-step that
issue here as we are mainly concerned with entropy rates, which we
will designate with lower-case $h$.
\cite{hurst1951} developed the parameter when he was measuring flows in
the Nile River. There is a
relationship between $H$ and the $\alpha$ in Definition~\ref{def:lrd_alpha},
namely
\[ H = 1 - \alpha/2. \]
LRD processes have $\alpha \in (0,1)$ and hence have $H \in (0.5,
1)$. The parameter
takes on values between 0 and 1, with $H > \nicefrac{1}{2}$ representing a
positively correlated process and $H < \nicefrac{1}{2}$ representing a
negatively correlated process and $H = \nicefrac{1}{2}$ being a short range
correlated process, such as white Gaussian noise.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{Plots/fgn_example_realisation_08.pdf}
\caption{Sample path of Fractional Gaussian Noise with Hurst parameter, $H=0.8$. We can see that the high correlations lead to the appearance of longer trends than we would expect for an independent and identically distributed process.}
\label{fig: fgn_sample_path}
\end{figure}
Another property of note, that has been used as the definition of LRD
processes, is the (un)summability of the autocorrelation function. For
a LRD processes the sum of the autocovariances diverges, {\em i.e.,}
$\sum_{k=1}^{\infty} \gamma(k) \rightarrow \infty$, whereas for SRD
processes this is finite (\cite{beran1994statistics}). Then we can
interpret LRD as having such strong correlations that the
autocovariance values decay such that the distant past still
influences the future. The intuition for non-LRD processes is
that the autocovariance function decays exponentially, or quicker, and
in many cases this means, for analysis, we can ignore correlations beyond some short
lag.
The negatively correlated processes, $H < \nicefrac{1}{2}$, have not received as
much consideration at the SRD and LRD cases but we include them in the
analysis here. These are similar to short range dependent processes,
such as white Gaussian noise, as they still have a summable
autocorrelation function (see~\cite{beran1994statistics,
gefferth2003nature}). In fact their structure enforces that
$\sum_{k=1}^{\infty} \gamma(k) = 0$. This is quite a strict and
surprising property, and hence~\cite{gefferth2003nature} called these processes
constrained short range dependent (CSRD).
We will be working with Gaussian processes in this paper.
\begin{definition}
A stochastic process is a Gaussian process if and only if each
finite collection of random variables from the process has a
multivariate Gaussian distribution.
\end{definition}
This definition applies to both discrete and continuous time
processes, though here we are principally interested in the
former. Gaussian processes are completely characterised by their
second order statistics (the mean and autocovariance
function) (\cite{box2015time}), which makes them the primary type of
stochastic model used in this context.
\subsection{Entropy rate}
As we are considering continuous random variables in this paper, we
will be considering differential entropy, which is a continuous
extension of Shannon entropy for discrete random variables. In this
paper, we will be using the natural logarithm in all of the
definitions, and hence the units of entropy that we will be working
with are nats. We include standard definitions in order that all
notation be precisely defined.
\begin{definition}
The differential entropy, $h(X)$, of a continuous random variable with probability density function, $f(x)$, is defined as,
\begin{align*}
h(X) = - \int_{\Omega} f(x) \log f(x) \, dx,
\end{align*}
where $\Omega$ is the support of the random variable.
\end{definition}
Differential entropy has some important properties which are different
from Shannon entropy. For example, differential entropy can be
negative, or even diverge to $-\infty$, which we can see by
considering the Dirac delta function, $\delta(x)$. The delta, {\em i.e.,} the
unit impulse, is defined by the properties $\delta(x) =0$ for $x \neq
0$ and $\int_{-\infty}^{\infty} \delta(x) \, dx = 1$. The Dirac delta
can be thought of in terms of probability as a completely determined
point in time, that is, a function possessing no uncertainty. It can
be constructed as the limit of rectangular pulses of constant area 1
as their width decreases, and hence we can calculate the entropy of
the delta as
\begin{align*}
h(X) &= -\int_{-a}^{a} \frac{1}{2a} \log\left(\frac{1}{2a}\right) dx,\\
&= \log(2a),
\end{align*}
which tends to $-\infty$ as $a \rightarrow 0$.
The intuition for $h(X) = -\infty$ from~\cite{cover_thomas_2006} is that the number of bits (note we are working in nats) on
average required to fix a random variable, $X$ to n-bit accuracy
is $h(X) + n$. Meaning $h(X) = -\infty$, can be read as requiring
$n-\infty$ bits, or that we can describe the random variable
arbitrarily accurately without using any bits.
We will see the same type of asymptotic behaviour for LRD processes as
$H \rightarrow 1$. Effectively, in the limit the correlations in the
process straight-jacket it, such that the future is completely
determined by the past, and so the incremental uncertainty in the
process is the same as that of the delta.
The differential entropy can be extended into the multivariate case
and hence to stochastic processes using the joint entropy for a
collection of random variables.
\begin{definition}
The joint differential entropy of a collection of random variables $X_1, X_2, ..., X_n$, with density $f(x_1, x_2, ... , x_n) = f(\mathbf{x})$ is defined as,
\begin{align*}
h(X_1, X_2, ... , X_n) = - \int_{\Omega}f(\mathbf{x}) \log f(\mathbf{x}) \, d\mathbf{x},
\end{align*}
where $\Omega = \Omega_1 \times \Omega_2 \times ... \times \Omega_n$ is the support of the random variables.
\end{definition}
Similarly, the conditional differential entropy can be defined for a random variable, given knowledge of other variables.
\begin{definition}
The conditional differential entropy of a random variable, $X_n$, given a collection of random variables $X_1, X_2, ... , X_{n-1}$, with a joint density $f(x_1, x_2, ... , x_n) = f(\mathbf{x})$, is defined as,
\begin{align*}
h(X_n | X_{n-1}, ... , X_1) = - \int_{\Omega} f(\mathbf{x}) \log f(x_n | x_{n-1}, ... ,x_1) d\mathbf{x},
\end{align*}
where $\Omega = \Omega_1 \times \Omega_2 \times ... \times \Omega_n$ is the support of the random variables.
\end{definition}
Finally, we define the concept of differential entropy rate, which can
be thought of as the average amount of new information from each
sample of a random variable in a discrete time process.
\begin{definition}
Where the limit exists, the differential entropy rate of a
stochastic process $\raisebox{\depth}{\(\chi\)} = \{X_i\}_{i \in \mathbb{N}}$
is defined to be,
\begin{align*}
h(\raisebox{\depth}{\(\chi\)}) &= \lim\limits_{n \rightarrow \infty} \frac{h(X_1, ... , X_n)}{n}.
\end{align*}
\end{definition}
An example of a process which is non-stationary but has a differential
entropy rate is the Gaussian walk, $S_n$. This is defined as the
process of sums of i.i.d. normally distributed random variables, {\em i.e.,}
$S_n = \sum_{i=1}^{n} X_i$, where $X_i \sim \mathcal{N}(0, \sigma^2)$.
The process has mean 0 for all $n$, however it is non-stationary as
the variance depends on $n$ as,
\begin{align*}
\mbox{Var}(S_n) = \mbox{Var}\left(\sum_{i=1}^{n} X_i\right) = n\sigma^2.
\end{align*}
However, the entropy rate converges and is equal to
\begin{align*}
\lim\limits_{n \rightarrow \infty} \frac{h(S_1, ... , S_n)}{n} &= \lim\limits_{n \rightarrow \infty} \frac{\sum_{i=1}^{n} h(S_i | S_{i-1}, \ldots, S_1)}{n},\\
&= \lim\limits_{n \rightarrow \infty}\frac{nh(X_i)}{n}, \\
&= h(X_i),
\end{align*}
as each random variable $X_i$ is independent.
An alternative characterisation of the differential entropy rate is given by the following theorem for stationary processes. This was developed for the Shannon entropy of discrete processes~\cite[Theorem 4.2.1]{cover_thomas_2006}, however this has been extended to differential entropy~\cite[pg. 416]{cover_thomas_2006}.
\begin{theorem}[Theorem 4.2.1~{\cite{cover_thomas_2006}}]\label{conditional_entropy_rate}
For a stationary stochastic process, $\raisebox{\depth}{\(\chi\)} = \{X_i\}_{i \in \mathbb{N}}$, the entropy rate is equal to,
\begin{align*}
h(\raisebox{\depth}{\(\chi\)}) &= \lim\limits_{n \rightarrow \infty} h(X_n | X_{n-1}, ... , X_1).
\end{align*}
\end{theorem}
\noindent The second equivalent definition is useful because it will
allow us to analyse the convergence rates of conditional entropy to
the differential entropy rate, which is important in estimation of
entropy rates.
\section{Conclusion}
In this paper, we are concerned with the behaviour of the differential entropy rate to understand and characterise the behaviour of LRD and SRD processes. Analysing two common LRD processes, FGN and ARFIMA(0,d,0), we have shown that the maximum occurs in the absence of correlations, \emph{i.e.}, $H=0.5$, and the differential entropy rate tends to the minimum, $-\infty$ as the strength of positive correlations increase, \emph{i.e.}, as we receive more information from correlations, the entropy of the process decreases. However, there is very different behaviour for negatively correlated processes, where ARFIMA(0,d,0) processes do not tend to $-\infty$ as the strength of the negative correlations increases. Further research is required to understand this behaviour for these processes.
In addition, we have made a link, similar to Shannon entropy, between the mutual information between past and future and excess entropy, meaning that the amount of shared information between the complete past of future of a process is the same as the additional information that accrues when converging to the entropy rate, based on past observations. This leads to a characterisation of LRD processes, as those having infinite mutual information between past and future. Using this and Szeg{\"o}'s limit theorems we then can classify LRD and SRD processes by their convergence rates, and show that LRD processes have slower convergence of the conditional entropy, conditioned on the past, to the entropy rate.
\section{Entropy rate function for ARFIMA(\MakeLowercase{p},\textbf{d,q)}}
We consider the differential entropy rate function of a related process to Fractional Gaussian Noise, which is ARFIMA(p,d,q), the fractional extension of the ARIMA (Autoregressive Integrated Moving Average) processes, by extending to non-integer differencing parameters, $d$ (see~\cite{hosking1981, granger1980}). FGN and ARFIMA(0,d,0) are commonly used stationary LRD processes for modelling real phenomena, and in particular FGN and ARFIMA(0,d,0) have very similar properties in the time and frequency domains. Additionally, these processes have been linked by by limit of their autocorrelation coefficients, $\rho(k) := \gamma(k)/\sigma^2$, under aggregation and rescaling (see~\cite{gefferth2003nature}). However, ARFIMA processes do differ from FGN in that you could change the fixed point, \emph{i.e.}, alter the eventual limit under aggregation and rescaling, with the addition of additive noise (\cite{veitch2013farima}), \emph{i.e.}, this class is less robust to the addition of noise. Hence, there may be some differences in behaviour when looking through an entropic lens.
Before we define an ARFIMA(p,d,q) process, we define two polynomials that are required for the ARFIMA definition. These are polynomials of the lag operator, $L$, where $LX_n = X_{n-1}$, defined by
\begin{align*}
\phi(x) &= 1 - \sum_{j=1}^{p} \phi_j x^j, \text{for coefficients }\phi_j \text{ and } p \in \mathbb{Z^+},\\
\psi(x) &= 1 + \sum_{j=1}^{q} \psi_j x^j, \text{for coefficients }\psi_j \text{ and } q \in \mathbb{Z^+}.
\end{align*}
\noindent Now we can define an ARFIMA(p,d,q) process,
\begin{definition}
For a stationary stochastic process $\{X_i\}_{i \in \mathbb{Z}}$, such that
\begin{align*}
\phi(L)(1-L)^dX_i &= \psi(L)\epsilon_i ,
\end{align*}
for some $-\nicefrac{1}{2} < d < \nicefrac{1}{2}$ and $\epsilon_i$, a zero mean normally distributed random variable with variance $\sigma_\epsilon^2$. $\{X_i\}_{i \in \mathbb{Z}}$ is called a ARFIMA$(p,d,q)$ process.
\end{definition}
There is a connection between the differencing parameter, $d$, and the Hurst parameter, $H$, for these processes
An ARFIMA(0,d,0) process is a special case of an ARFIMA(p,d,q), where $\phi(x) = \psi(x) = 1$, that is there is no lag on the noise, $\epsilon$, and all the auto-regressive lags on the previous values come from the differencing operator.
The spectral density of an ARFIMA(p,d,q) is given by~\cite{beran1994statistics} as
\begin{align}
f(\lambda) &= \frac{\sigma_\epsilon^2|\psi(e^{i\lambda})|^2}{2\pi|\phi(e^{i\lambda})|^2}|1 - e^{i\lambda}|^{-2d}.\label{eqn:arfima_spec_dens}
\end{align}
\noindent The following theorem from \cite{hosking1981} (and in~\cite{beran1994statistics}) gives infinite autoregressive and moving average representations for ARFIMA(0,d,0) processes.
\begin{theorem}[Proposition 2.2~{\cite{beran1994statistics}}]
Let $X_n$ be a fractional ARIMA(0,d,0) process with -$\frac{1}{2} < d < \frac{1}{2}$ . Then\\
(i) the following infinite autoregressive representation holds:\\
\begin{align*}
\sum_{k=0}^{\infty} \pi_k X_{n-k} = \epsilon_n,
\end{align*}
where $\epsilon_n (n = 1, 2, ...)$ are independent identically distributed random variables and\\
\begin{align*}
\pi_k = \frac{\Gamma(k-d)}{\Gamma(k+1)\Gamma(-d)}.
\end{align*}
For $k \rightarrow \infty$ we have,
\begin{align*}
\pi_k \sim \frac{1}{\Gamma(-d)}k^{-d-1}.
\end{align*}
(ii) The following infinite moving average representation holds:\\
\begin{align*}
X_n = \sum_{k=0}^{\infty} a_k \epsilon_{n-k}
\end{align*}
where $\epsilon_n (n = 1, 2, ...)$ are independent identically distributed random variables and\\
\begin{align*}
a_k = \frac{\Gamma(k+d)}{\Gamma(k+1)\Gamma(d)}.
\end{align*}
For $k \rightarrow \infty$ we have
\begin{align*}
a_k \sim \frac{1}{\Gamma(d)}k^{d-1}.
\end{align*}
\end{theorem}
We will express an entropy rate characterisation for ARMA processes in terms of its innovation process variance, from \cite{ihara1993information}, and show that this can be extended to ARFIMA(0,d,0) and ARFIMA(p,d,q) processes. Then we will use the result to characterise the entropy rate of an ARFIMA(0,d,0) process in terms of its process variance.
\begin{theorem}[Proposition 2.2~{\cite{ihara1993information}}]\label{thm:ihara}
The entropy rate of an ARMA(p,q) process is given by,
$h(\raisebox{\depth}{\(\chi\)}) = \frac{1}{2}\log(2\pi e \sigma_\epsilon^2)$.
\end{theorem}
\noindent Now, we state our extension to ARFIMA(0,d,0) and present a proof based on Ihara's proof of Theorem~\ref{thm:ihara}
\begin{theorem}\label{arfima(0,d,0)}
The entropy rate of an ARFIMA(0,d,0) process is given by,
$h(\raisebox{\depth}{\(\chi\)}) = \frac{1}{2}\log(2\pi e \sigma_\epsilon^2)$.
\end{theorem}
\begin{proof}
First we calculate $\int_{-\pi}^{\pi} \log(2\pi f(\lambda)) d\lambda$, using the spectral density of an ARFIMA process given in~(\ref{eqn:arfima_spec_dens})
\begin{align*}
\int_{-\pi}^{\pi} \log (2\pi f(\lambda)) d\lambda &= \int_{-\pi}^{\pi} \log(\sigma_\epsilon^2|1 - e^{i \lambda}|^{1-2H}) d\lambda,\\
&= \int_{-\pi}^{\pi} \log(\sigma_\epsilon^2) d\lambda + (1-2H)\int_{-\pi}^{\pi} \log|1 - e^{i \lambda}| d\lambda.
\end{align*}
Now we transform the elements in the last term using their trigonometric representation,
\begin{align*}
|1 - e^{i\lambda}| &= |1 - \cos(\lambda) - i\sin(\lambda)|,\\
&= \sqrt{(1 - \cos(\lambda))^2 + \sin^2(\lambda)},\\
&= \sqrt{2 - 2\cos(\lambda)},\\
&= \sqrt{4 \sin^2\left(\frac{\lambda}{2}\right)},\\
&= 2|\sin\left(\frac{\lambda}{2}\right)|.
\end{align*}
This makes the integral of the log spectral density,
\begin{align*}
\int_{-\pi}^{\pi} \log |1 - e^{i\lambda}| d\lambda &= 2 \int_{0}^{\pi} \log\left(2\sin\left(\frac{\lambda}{2}\right)\right) d\lambda,\\
&= 2 \int_{0}^{\pi} \log(2) d\lambda + 2 \int_{0}^{\pi} \log\left(\sin\left(\frac{\lambda}{2}\right)\right) d\lambda,
\end{align*}
We substitute $y = \lambda/2$,
\begin{align*}
\int_{-\pi}^{\pi} \log |1 - e^{i\lambda}| d\lambda &= 2 \pi \log(2) + 2 \int_{0}^{\frac{\pi}{2}} \log(\sin y) 2 dy,\\
&= 2 \pi \log(2) + 4\left(-\frac{\pi}{2} \log(2)\right), \\
&= 0.
\end{align*}
Where the equality $\int_{0}^{\frac{\lambda}{2}} \log(\sin y) dy = -\frac{\pi}{2} \log(2)$ is given by~\cite{koyama2005}.
\noindent So the last term of the spectral density vanishes, and
\begin{align*}
\int_{-\pi}^{\pi} \log (2\pi f(\lambda)) d\lambda &= \int_{-\pi}^{\pi} \log(\sigma_\epsilon^2) d\lambda + (1-2H)\int_{-\pi}^{\pi} \log|1 - e^{i \lambda}| d\lambda,\\
&= 2\pi \log(\sigma_\epsilon^2).
\end{align*}
Using Kolmogorov's entropy rate expression, the entropy rate is therefore,
\begin{align*}
h(\raisebox{\depth}{\(\chi\)}) &= \frac{1}{2} \log (2 \pi e) + \frac{1}{4\pi}\left(2\pi \log(\sigma_\epsilon^2)\right),\\
&= \frac{1}{2} \log(2 \pi e \sigma_\epsilon^2).
\end{align*}
\end{proof}
\begin{remark}
This can be shown also using the infinite autoregressive expression above, $X_n = \epsilon_n - \sum_{k=1}^{\infty} \pi_kX_{n-k}$, and substituting into the conditional entropy rate for stationary processes, $h(\raisebox{\depth}{\(\chi\)}) = \lim\limits_{n \rightarrow \infty} h(X_n | X_n-1, ... , X_0)$. Then we can remove the conditioning from the entropy rate calculation, \begin{align*}
h(\raisebox{\depth}{\(\chi\)}) = \lim\limits_{n \rightarrow \infty} h\left(\epsilon_n - \sum_{k=1}^{\infty} \pi_kX_{n-k} \bigg| X_n-1, ... , X_0\right) = \lim\limits_{n \rightarrow \infty} h(\epsilon_n).
\end{align*} Which then implies that $h(\raisebox{\depth}{\(\chi\)}) = \frac{1}{2}\log(2\pi e \sigma_\epsilon^2)$, \emph{i.e.}, the entropy rate of the process depends only on the entropy introduced at each step by the innovations. Therefore, we conclude that the entropy rate of an ARFIMA(0,d,0) process depends on the innovation variance and not $H$.
\end{remark}
We can generalise to ARFIMA(p,d,q) processes by adding an additional condition, the invertibility of the moving average polynomial. This is an extremely common condition applied in the theory of autoregressive-moving average, \emph{i.e.}, ARMA(p,q) processes. The condition implies that all roots of the moving average polynomial lie outside of the unit circle, and similarly the stationarity condition of the process ensures that all roots of the autoregressive polynomial lie outside of the unit circle (\cite{box2015time}). We will use these conditions on the ARFIMA processes to analyse their properties.
\begin{theorem}\label{arfima(p,d,q)}
The entropy rate of a stationary ARFIMA(p,d,q) process with invertible moving average polynomial is given by,
$h(\raisebox{\depth}{\(\chi\)}) = \frac{1}{2}\log(2\pi e \sigma_\epsilon^2)$.
\end{theorem}
\begin{proof}
Since ARFIMA(p,d,q) processes are stationary and invertible, this implies that the polynomials $\phi(x)$ and $\psi(x)$, have roots outside of the unit circle, i.e. each root $z \in \mathbb{C}$ is such that $|z| > 1$. By the Fundamental Theorem of Algebra, both the autoregressive and moving average polynomials can be factored into linear factors. As the constant terms are 1, this implies the polynomials can be factored as $\phi(x) = \prod_{i=1}^p (1 - a_i e^{i \lambda})$ and $\psi(x) = \prod_{i=i}^{q} (1 - b_i e^{i \lambda})$, where $|a_i|, |b_i| < 1, \forall i$.
Recall from equation~(\ref{eqn:arfima_spec_dens})that the spectral density is given by,
\begin{align*}
f(\lambda) &= \frac{\sigma_\epsilon^2}{2\pi} |1 - e^{i\lambda}|^{-2d} \frac{|\psi(e^{i\lambda})|^2}{|\phi(e^{i\lambda})|^2}.
\end{align*}
\noindent Hence,
\begin{align*}
\log((2\pi f(\lambda)))
&= \log(\sigma_\epsilon^2) -2d\log|1-e^{i\lambda}| \\
&\qquad + 2\log\left|\prod_{j=1}^{q}(1 - a_j e^{i\lambda})\right| -2\log\left|\prod_{j=1}^{p}(1 - b_j e^{i\lambda})\right|,\\
&= \log(\sigma_\epsilon^2) -2d\log|1-e^{i\lambda}| \\
&\qquad + \sum_{j=1}^{q} 2\log|1 - a_j e^{i\lambda}| - \sum_{j=1}^{p} 2\log|1 - b_j e^{i\lambda}|.
\end{align*}
Now we calculate the integral of the log spectral density,
\begin{align*}
\int_{-\pi}^{\pi} \log(2\pi f(\lambda)) d \lambda
&= \int_{-\pi}^{\pi} \log(\sigma_\epsilon^2) d \lambda - \int_{-\pi}^{\pi} 2d\log|1-e^{i\lambda}| d \lambda\\
&\qquad + \sum_{j=1}^{q} 2 \int_{-\pi}^{\pi} \log|1 - a_j e^{i\lambda}| d \lambda - \sum_{j=1}^{p} 2 \int_{-\pi}^{\pi} \log|1 - b_j e^{i\lambda}| d \lambda,\\
&= 2 \pi \log(\sigma_\epsilon^2).
\end{align*}
Where the third equality is given as all the integrals of $\log|1 - ae^{i\lambda}|$ over $[-\pi, \pi]$ vanish for $|a| \le 1$ (\cite{shiu_2004}).
We substitute this expression into Kolmogorov's entropy rate expression for Gaussian processes.
\begin{align*}
h(\raisebox{\depth}{\(\chi\)})
&= \frac{1}{2}\log(2 \pi e) + \frac{1}{4 \pi} (2 \pi \log(\sigma_\epsilon^2)),\\
&= \frac{1}{2}\log(2 \pi e \sigma_\epsilon^2).
\end{align*}
\end{proof}
This result leads to the following corollary, which can finalise the discussion of the differential entropy rate in terms of innovation variance for the classes of AR, MA, ARMA processes. This is relevant as the definition in terms of the innovation variance is the perspective that is commonly used in the time series literature, when modelling real world processes.
\begin{corollary}
The differential entropy rate of stationary AR(p), invertible MA(q) and, stationary and invertible ARMA(p,q) processes is $h(\raisebox{\depth}{\(\chi\)}) = \frac{1}{2} \log(2 \pi e \sigma_\epsilon^2)$.
\end{corollary}
\noindent Hence, for these models the entropy rate can be calculated in terms of the variance of its innovations. However we want to compare the entropy rates, as a function of their Hurst parameter, between ARFIMA(0,d,0) and FGN, so we want to fix the variance of process itself, $\sigma^2$. We will use the autocovariance function of ARFIMA(0,d,0), from~\cite{beran1994statistics},
\begin{align*}
\gamma(k) &= \sigma_\epsilon^2 \frac{(-1)^k \Gamma(1-2d)}{\Gamma(k-d+1)\Gamma(1-k-d)}.\\
\end{align*}
Note that $\gamma(0) = \sigma^2$,
\begin{align*}
\sigma^2 = \gamma(0) &= \sigma_\epsilon^2 \frac{ \Gamma(1-2d)}{\Gamma(1-d)^2},
\end{align*}
and hence,
\begin{align*}
\sigma_\epsilon^2 &= \sigma^2\frac{\Gamma(1-d)^2}{\Gamma(1-2d)}.
\end{align*}
This leads to the following characterisation of ARFIMA(0,d,0) processes in terms of the Hurst parameter, $H$, noting that $d = H - \nicefrac{1}{2}$.
\begin{corollary}
The entropy rate of an ARFIMA(0,d,0) process for a fixed process variance, $\sigma^2$, is given by,
\begin{align}
h(\raisebox{\depth}{\(\chi\)}) = \frac{1}{2} \log(2 \pi e \sigma^2) + \log\left(\Gamma\left(\frac{3}{2} - \mathcal{H}\right)\right) - \frac{1}{2} \log \bigg(\Gamma(2 - 2\mathcal{H})\bigg). \label{eq:arfima_entropy_rate}
\end{align}
\end{corollary}
\begin{proof}
By Theorem~\ref{arfima(0,d,0)} and from the characterisation of $\sigma_\epsilon^2$ above,
\begin{align*}
h(\raisebox{\depth}{\(\chi\)}) &= \frac{1}{2} \log \left(2 \pi e \sigma^2\frac{\Gamma(1-d)^2}{\Gamma(1-2d)}\right),\\
&= \frac{1}{2} \log(2 \pi e \sigma^2) + \log(\Gamma(1-d)) - \frac{1}{2} \log(\Gamma(1-2d)),\\
&= \frac{1}{2} \log(2 \pi e \sigma^2) + \log\left(\Gamma\left(\frac{3}{2} - \mathcal{H}\right)\right) - \frac{1}{2} \log \bigg(\Gamma(2 - 2\mathcal{H})\bigg).
\end{align*}
\end{proof}
\begin{remark}
The same approach can be used for more general ARFIMA(p,d,q) processes. However in this case, there is no general closed form for the autocovariance function, so the variance must be calculated for each process and then substituted for the innovation variance. Interestingly, this result indicates that the effect of the changing the process variance is balanced by the effect of the change in the Hurst parameter, with respect to the innovation variance. This results in the constant differential entropy rate when considered in terms of its innovation variance.
\end{remark}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Plots/arfima_entropy_rate_comparison_variance.pdf}
\caption{The entropy rate of ARFIMA(0,d,0) as a function of the Hurst parameter, $H$, for variance, $\sigma^2 = 1, 2, 3, 4$. On the positively correlated side, $H > 0.5$, we see a similar asymptotic behaviour to FGN. However, for negatively correlated processes, the amount of entropy in the process, stays quite high. We see the maximum of the function at $H=0.5$, which intuitively shows that the highest uncertainty occurs for the white Gaussian noise process.}
\label{fig: arfima_entropy_rate_function}
\end{figure}
We show the plot of the ARFIMA(0,d,0) entropy rate as a function of the Hurst parameter, $H$, with process variance, $\sigma^2 = 1,2,3,4$, in Figure~\ref{fig: arfima_entropy_rate_function}. The plot shows some interesting behaviour, particularly when compared to the FGN entropy rate function in Figure~\ref{fig: arfima_fgn_entropy_rate_function_comparison}. Some of these observed properties are:
\begin{itemize}
\item The entropy rate is not symmetric, much less so than FGN. The positively correlated side has a dramatic drop, however the negatively correlated side stays relatively high. In order words, there is a demonstrable difference between FGN and ARFIMA(0,d,0) in the behaviour as CSRD processes.
\item The entropy rate asymptotically tends to $-\infty$ as $H \rightarrow$ 1 only.
\item The maximum entropy rate occurs at the same point as FGN, $\mathcal{H} =0.5$, indicating that the maximum entropy occurs for white Gaussian noise.
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Plots/arfima_fgn_entropy_rate_comparison_paper.pdf}
\caption{The comparison of the entropy rate as function of the Hurst parameter, for both ARFIMA(0,d,0) and FGN processes, with variance 1. It appears that the ARFIMA(0,d,0) process has an entropy rate which is greater than or equal to FGN for all values of $H$. The negatively correlated portion falls away quickly as $H \rightarrow 0$ for FGN but stays relatively high for the ARFIMA(0,d,0) process. The maximum of the functions coincide at $H=0.5$.}
\label{fig: arfima_fgn_entropy_rate_function_comparison}
\end{figure}
\noindent Similar to the previous section, we will prove the asymptotics of the entropy rate function, and show that the maximum occurs at $H=0.5$.
\begin{corollary}
The differential entropy rate of ARFIMA(0,d,0), $h(\raisebox{\depth}{\(\chi\)})$, tends to negative infinity as $H \rightarrow 1$, for a fixed variance $\sigma^2 \in \mathbb{R}$.
\end{corollary}
\begin{proof}
As $H \rightarrow 1$, the term $\Gamma(\frac{3}{2} - H)$ is finite, as well as the $\frac{1}{2} \log(2 \pi e \sigma^2)$, for a fixed variance $0 < \sigma^2 < \infty$.
Now, as $H \rightarrow 1$, the term $\Gamma(2 - 2H) \rightarrow \Gamma(0)$. There exists a singularity for the gamma function at 0, which diverges to infinity. Which implies that the term $-\frac{1}{2} \log(\Gamma(2 - 2H)) \rightarrow -\infty$, since $\Gamma(x) \rightarrow \infty$, as $x \rightarrow 0$. This implies that $h(\raisebox{\depth}{\(\chi\)}) \rightarrow -\infty$, as $H \rightarrow 1$.
\end{proof}
\begin{remark}
Note that the value of the entropy rate function for an ARFIMA(0,d,0) process as $H \rightarrow 0$, is
\begin{align*}
h(\raisebox{\depth}{\(\chi\)}) = \frac{1}{2} \log(2 \pi e \sigma^2) + \log(\Gamma(\frac{3}{2})) - \frac{1}{2} \log(\Gamma(2)) \approx 1.298.
\end{align*}
\end{remark}
To complete this section of the analysis, we will consider the maximum of the entropy rate function of ARFIMA(0,d,0), and conclude which Hurst parameter has the highest uncertainty, in the sense of maximum differential entropy rate.
\begin{theorem}
The differential entropy rate of ARFIMA(0,d,0) as a function of $H$ attains the maximum at $H = \nicefrac{1}{2}$.
\end{theorem}
\begin{proof}
We differentiate the entropy rate function with respect to $H$, and we get
\begin{align*}
\frac{dh(\raisebox{\depth}{\(\chi\)})}{dH} &= \frac{d}{dH}\left(\frac{1}{2} \log(2 \pi e \sigma^2) + \log\left(\Gamma(\frac{3}{2} - H)\right) - \frac{1}{2} \log\left(\Gamma(2 - 2H)\right)\right),\\
&= \frac{d}{dH}\left(\log(\Gamma(\frac{3}{2} - H))\right) - \frac{d}{dH}\left(\log(\Gamma(2 - 2H))\right),\\
&= \frac{\Gamma(\frac{3}{2} - H)\psi\left(\frac{3}{2} - H\right)}{\Gamma(\frac{3}{2} - H)} - \frac{\Gamma(2 - 2H)\psi(2 - 2H)}{\Gamma(2 - 2H)},\\
&= \psi\left(\frac{3}{2} - H\right) - \psi(2 - 2H),
\end{align*}
where $\psi(x)$ is the digamma function.
Then we set $\frac{dh(\raisebox{\depth}{\(\chi\)})}{dH} = 0$, and solve for $H$. Since $\psi(x)$ is a monotonically increasing function on $\mathbb{R^+}$, this implies that $\frac{dh(\raisebox{\depth}{\(\chi\)})}{dH}$ has one solution. Since $\frac{3}{2} - H = 2 - 2H$ only when $H = \nicefrac{1}{2}$, this implies that $h(\raisebox{\depth}{\(\chi\)})$ achieves a unique maximum at this point.
\end{proof}
This aligns with our intuition, that the highest uncertainty occurs for this model when it is uncorrelated and equal to white Gaussian noise, as it simplifies to $X_n = \epsilon_n$, identical to FGN processes. This explains why the maxima coincide for the two processes, given the same process variance, although ARFIMA(0,d,0) appears to higher differential entropy across the entire parameter range, when not at $H = 0.5$. This provides support that the maximum entropy process for LRD covariance constraints is the ARFIMA class, which echos previous results in this area such as Burg's Theorem (\cite{choi1984information}), that the AR and ARMA class of processes are the maximum entropy models given appropriate constraints on the covariances and impulse responses (see~\cite{ihara1984maximum, franke1985arma}).
We have shown in this section that the behaviour for the ARFIMA(0,d,0) model differs from that of FGN in the behaviour of their CSRD processes. This is a surprising discovery and warrants further investigation. Both models, however, have much less uncertainty as the strength of the positive correlations increases, as well as a maximum uncertainty occurring for uncorrelated processes. Hence, we may be able to characterise the behaviour of LRD processes on the entropy rate as tending to $-\infty$ as the strength of correlations increases.
In remainder of the paper we look at other information theoretic measures as way to characterise the behaviour of SRD and LRD processes.
\section{Entropy rate convergence for LRD processes}
From the previous sections we have gained an understanding of some of the properties of the entropy rate function for common LRD models. In this section, we classify the convergence rate of the conditional entropy to the entropy rate for SRD and LRD processes.
In Section~\ref{excess_entropy} we used Szeg{\"o}'s theorem to show,
\begin{align*}
\lim\limits_{n \to \infty} \frac{|K^{(n)}|}{|K^{(n-1)}|} = \exp\left(\frac{1}{2\pi} \int_{-\pi}^{\pi} \log f(\lambda) d\lambda\right),
\end{align*}
if $\log f(\lambda) \in L_1[-\pi, \pi]$. This provides another perspective to explain why the conditional entropy converges to the entropy rate. The Strong Szeg{\"o} Theorem, provides an additional term to this limit with the same regularity conditions. The Szeg{\"o} theorem showed why the convergence occurred for the conditional entropy, and the strong Szeg{\"o} theorem we explain the convergence rate. We will state the version as given in~\cite{bingham2012}, which is more suited to probabilistic analysis.
For ease of notation we define the limit of the Szeg{\"o} theorem,
\begin{align*}
G(\mu) := \exp\left( \frac{1}{2} \int_{-\pi}^{\pi} \log f(\lambda) d\lambda\right),
\end{align*}
and the partial autocorrelation coefficients, $\alpha_n$, as,
\begin{align*}
\alpha_n = corr(X_n - P_{[1, n-1]}X_n, X_0 - P_{[1, n-1]}X_0),
\end{align*}
where $P_{[1, n-1]}$ is the projection onto the linear space spanned by $\{X_{-n}, ... , X_{-1}\}$ and the correlation function is defined as,
\begin{align*}
corr(X,Y) = \frac{\mathds{E}[X\bar{Y}]}{\sqrt{\mathds{E}[|X|^2]\mathds{E}[|Y|^2]}},
\end{align*}
for $X$ and $Y$ zero mean random variables. Note that $P_{[1, n-1]}X_0$ is the best linear predictor of $X_0$ given the finite past of length $n$.
The partial autocorrelation function is related to the autocorrelation function by the removal of the linear dependence on the variables within n lags. For example, in the case of finite lag processes, such as AR(p), the partial correlation function is 0 for a lag greater than p. In the case of the ARFIMA(p,d,q) process, the decay is slower for LRD parametrisations (\cite{inoue2002asymptotic}).
We define the Hardy space $\mathcal{H}^{\frac{1}{2}}$, which is the subspace of $\ell_2$ of sequences $a = (a_n)$ such that $||a||^2 := \sum_{n} (1 + |n|)|a_n|^2 < \infty$, which is a well defined norm on $\ell_2$. We use this to describe the conditions on the strong Szeg{\"o} theorem. Also, the cepstral coefficients that were defined in Theorem~\ref{Li_thm_1} will be used in the definition of the additional term of the Szeg{\"o} limit theorem, reinforcing their connection with the information theoretic perspective of stochastic processes.
\begin{theorem}[Strong Szeg{\"o} Theorem (\cite{bingham2012})]
For a sequence of Toeplitz matrices, $\Gamma^{(n)} = [\gamma(|i-j|)]_{i,j}$, of increasing size $n$, for covariance function $\gamma(n)$, with associated spectral density $f(\lambda)$, such that $\log f(\lambda) \in L_1[-\pi, \pi]$, then
\begin{align*}
\lim\limits_{n \to \infty} \frac{|\Gamma^{(n)}|}{G(\mu)^n} \to E(\mu),
\end{align*}
where,
\begin{align*}
E(\mu) := \prod_{j=1}^{\infty} (1 - |\alpha_j|^2)^{-j} = \exp\left(\sum_{k=1}^\infty k b_k^2\right),
\end{align*}
noting that these expressions may be infinite. The infinite product converges if and only if
\begin{itemize}
\item The strong Szeg{\"o} condition, $\alpha \in \mathcal{H}^{\frac{1}{2}}$ holds; or,
\item The sum of cepstral coefficients $\sum_{k=1}^\infty k b_k^2 \in \mathcal{H}^{\frac{1}{2}}$.
\end{itemize}
\end{theorem}
\noindent Using this theorem we will be able to characterise the convergence rate of the short range dependent processes that we have been considering in this paper.
The SRD and CSRD versions of ARFIMA and Fractional Gaussian Noise meet the conditions of the convergence of the infinite product and sum of Strong Szeg{\"o} Theorem. This is shown by Theorem~\ref{Li_thm_1}. For a positive, continuous spectral density $\sum_{k=1}^{\infty} k \gamma(k)^2 < \infty$ if and only if $I_{p-f} < \infty$, which in turn holds if and only if $\sum_{k=-\infty}^{\infty} |k||b_k|^2 < \infty$. In the proof of Theorem~\ref{mutual_infinite}, we showed that the boundary of the finiteness and infiniteness of $\sum_{k=1}^{\infty} kb_k^2$, coincided with the boundary between SRD and LRD processes. For example, for ARFIMA and Fractional Gaussian Noise, $\sum_{k=1}^{\infty} k \gamma(k)^2 < \infty$ when $H \le \frac{1}{2}$ and therefore by Theorem~\ref{mutual_infinite}, $\sum_{k=1}^{\infty} |k||b_k|^2 < \infty$.
One additional note about the conditions of the infinite product and sum in this theorem, the conditions for $\ell_1$ and $H < \frac{1}{2}$, coincide for sequences that decay as power laws, {\em i.e.,} $n^\alpha$, which is common when considering the convergence or divergence of sequences of LRD processes. However, it may be the case that a process may be in $\ell_1$ and not in $\mathcal{H}^\frac{1}{2}$, \emph{e.g.}, $a_n = \frac{1}{n\log n}$ (\cite{bingham2012}).
Before we continue, we require a lemma to prove a theorem about the convergence rate of the differential entropy rate of SRD and LRD processes, which gives two different determinant limits for the results of the Szeg{\"o} limit theorems.
\begin{lemma}\label{determinant_ratio}
For all discrete-time stationary Gaussian processes,
\begin{align*}
\lim\limits_{n \to \infty} \left(|K^{(n)}| \right)^{\frac{1}{n}} &= \lim\limits_{n \to \infty} \frac{|K^{(n)}|}{|K^{(n-1)}|},\\
\end{align*}
where $K^{(n)}$ is the $n \times n$ autocovariance matrix of the process.
\end{lemma}
\begin{proof}
For discrete-time stationary Gaussian processes,
\begin{align*}
\lim\limits_{n \to \infty} \frac{1}{n} h(X_1, .. , X_n) &= \lim\limits_{n \to \infty} h(X_n|X_{n-1}, ... , X_1),
\end{align*}
by Theorem~\ref{conditional_entropy_rate} and the fact that both of the limits exist for stationary processes. The form of the joint entropy of a multivariate Gaussian random vector of length $n$~\cite[pg. 249]{cover_thomas_2006} is
\begin{align*}
h(X_1, .. , X_n) = \frac{1}{2} \log\left(2 \pi e |K^{(n)}|\right).
\end{align*}
Therefore as,
\begin{align*}
\lim\limits_{n \to \infty} \frac{1}{n} h(X_1, .. , X_n) &= \lim\limits_{n \to \infty} \frac{1}{n} \left(\frac{1}{2} \log\left((2 \pi e)^n |K^{(n)}|\right)\right),\\
&= \lim\limits_{n \to \infty} \frac{1}{2} \log\left(2 \pi e |K^{(n)}|^{\frac{1}{n}}\right).
\end{align*}
Equating this for the entropy rate characterised using conditional entropy we get
\begin{align*}
\lim\limits_{n \to \infty} \frac{1}{2} \log\left(2 \pi e |K^{(n)}|^{\frac{1}{n}}\right)
&= \lim\limits_{n \to \infty} h(X_n, ... , X_1) - h(X_{n-1}, ... , X_1),\\
&= \lim\limits_{n \to \infty} \frac{1}{2} \log\left((2 \pi e)^n |K^{(n)}|\right) - \frac{1}{2} \log\left((2 \pi e)^{n-1} |K^{(n-1)}|\right),\\
&= \lim\limits_{n \to \infty} \frac{1}{2} \log\left((2 \pi e) \frac{|K^{(n)}|}{|K^{(n-1)}|}\right),
\end{align*}
Hence, we conclude that,
\begin{align*}
\lim\limits_{n \to \infty} (|K^{(n)}|)^{\frac{1}{n}} &= \lim\limits_{n \to \infty} \frac{|K^{(n)}|}{|K^{(n-1)}|}.
\end{align*}
\end{proof}
\noindent Now we can characterise the convergence rate of the conditional entropy of SRD ARFIMA and Fractional Gaussian Noise processes to the differential entropy rate. Then we will show that a slower decay exists in the case of long range dependence.
\begin{theorem}\label{srd_convergence}
For all discrete-time Gaussian SRD processes, such that the autocovariance function $\gamma \in \ell_1$, the convergence rate of the conditional entropy, $h(X_n|X_{n-1}, ..., X_1)$, to the differential entropy rate of the process is $O(n^{-1})$.
\end{theorem}
\begin{proof}
We rearrange the asymptotic expression of Szeg{\"o}'s strong theorem to get
\begin{align*}
\frac{|K^{(n)}|}{G(\mu)^n} &\sim E(\mu),\\
\implies |K^{(n)}|^{\frac{1}{n}} &\sim G(\mu)E(\mu)^{\frac{1}{n}}.
\end{align*}
Then we use the asymptotic form of the determinant limit from Lemma~\ref{determinant_ratio},
\begin{align*}
\frac{|K^{(n)}|}{|K^{(n-1)}|} &\sim G(\mu) \exp\left(\sum_{k=1}^\infty k b_k^2\right)^{\frac{1}{n}}, \\
\implies \frac{|K^{(n)}|}{|K^{(n-1)}|} &\sim G(\mu) \exp\left(\frac{1}{n} \sum_{k=1}^\infty k b_k^2\right).
\end{align*}
\noindent From the asymptotic form of the Szeg{\"o} theorem, Theorem~\ref{thm:szego_theorem}, the LHS will converge to $G(\mu)$ as $n$ increases. This implies that the convergence rate of the conditional entropy is controlled by the term, $\exp(\frac{1}{n} \sum_{k=1}^\infty k b_k^2)$, and its convergence rate to 1. That is, $n \to \infty$
\begin{align*}
\left|G(\mu) \exp\left(\frac{1}{n} \sum_{k=1}^\infty k b_k^2\right) - G(\mu)\right| &= G(\mu)\left|\exp\left(\frac{1}{n} \sum_{k=1}^\infty k b_k^2\right) - \exp(0)\right|,\\
&= G(\mu).
\end{align*}
Since $\frac{1}{n} \sum_{k=1}^\infty k b_k^2 \to 0$, as $n \rightarrow \infty$.
Since $\sum_{k=1}^\infty k b_k^2$ is finite for $H \le \frac{1}{2}$, as $\gamma \in \ell_1$, this implies that the convergence is at the rate of $O(n^{-1})$ as
\begin{align*}
\left|\exp\left(\frac{1}{n} \sum_{k=1}^\infty k b_k^2\right) - \exp(0)\right| &\to 0,\\
\implies \left|\frac{1}{n} \sum_{k=1}^\infty k b_k^2 - 0\right| &\to 0,\\
\text{and, } \left|\frac{1}{n} \sum_{k=1}^\infty k b_k^2\right| &\le \frac{C}{n}, C \in \mathbb{R^+}, \forall n \in \mathbb{Z^+}.
\end{align*}
\noindent For $C < \infty$, such that $C \ge \sum_{k=1}^\infty k b_k^2$.
\end{proof}
The key part of the proof that allows us to characterise the convergence rate of the conditional entropy is that the series, $\sum_{k=1}^\infty k b_k^2$, is finite. By Theorem~\ref{Li_thm_1}, we know that this is the boundary between LRD and SRD processes. This behaviour intuitively indicates that a limit would take a longer time to converge in the LRD case, similar to many estimators for LRD processes. Hence, we expect that the convergence to the entropy rate from the conditional entropy rate will be slower for LRD processes. This leads to the following theorem and characterisation of LRD processes through convergence of conditional entropy.
\begin{theorem}
For all Gaussian LRD processes the convergence rate of the conditional entropy, $h(X_n|X_{n-1}, ..., X_1)$, to the differential entropy rate of the process is
\begin{align*}
O(\frac{\log(n^{(1-2H)^2})}{n}).
\end{align*}
\end{theorem}
\begin{proof}
Similar to the theorem above, we consider the convergence of the term, $\exp(\frac{1}{n} \sum_{k=1}^\infty k b_k^2)$, to 0, and use the following expansion of the term,
\begin{align*}
\lim\limits_{n \rightarrow \infty} \exp\left(\frac{1}{n} \sum_{k=1}^\infty k b_k^2\right) &= \lim\limits_{n \rightarrow \infty} \exp\left(\frac{1}{n} \lim\limits_{m \rightarrow \infty}\sum_{k=1}^m k b_k^2\right),\\
&= \lim\limits_{n \rightarrow \infty} \lim\limits_{m \rightarrow \infty} \exp\left(\frac{1}{n} \sum_{k=1}^m k b_k^2\right).
\end{align*}
Where we use the continuity in the exponential function to exchange the limit in the function.
From Theorem~\ref{mutual_infinite}, we showed that the rate of divergence of the partial sum of $kb_k^2$ is equal to,
\begin{align*}
\sum_{k=1}^{m} k b_k^2 &\sim \frac{\left(1-2H\right)^2\ Si\left(\pi\right)^2}{\pi^2} \log m,\\
&= \frac{Si\left(\pi\right)^2}{\pi^2} \log \left(m^{\left(1-2H\right)^2}\right).
\end{align*}
Hence we can consider the limits as $n$ and $m$ tend to infinity,
\begin{align*}
\lim\limits_{n \rightarrow \infty} \lim\limits_{m \rightarrow \infty} \exp\left(\frac{1}{n} \sum_{k=1}^m k b_k^2\right) &\sim \lim\limits_{n \rightarrow \infty} \lim\limits_{m \rightarrow \infty} \exp\left(\frac{1}{n} \frac{Si\left(\pi\right)^2}{\pi^2} \log \left(m^{\left(1-2H\right)^2}\right) \right).
\end{align*}
If we take the $n \rightarrow \infty$ and $m \rightarrow \infty$ such that $m=n$, then we have
\begin{align*}
\lim\limits_{n \rightarrow \infty} \exp\left(\frac{1}{n} \frac{Si\left(\pi\right)^2}{\pi^2} \log \left(n^{\left(1-2H\right)^2}\right) \right) \rightarrow 1.
\end{align*}
Hence,
\begin{align*}
\lim\limits_{n \rightarrow \infty} \frac{\log\left(n^{\left(1-2H\right)^2}\right)}{n} \rightarrow 0.
\end{align*}
Which gives the rate of convergence of the conditional entropy to entropy rate.
\end{proof}
\noindent Therefore, the convergence of the conditional entropy to entropy rate is slower for LRD processes than SRD processes. This provides an information theoretic characterisation of LRD by convergence properties, similar to the covariance function, the sample mean and the parameters of linear predictors on the infinite past. The rate of convergence to the entropy rate decreases rapidly as $H \rightarrow 1$, because in $\left(1-2H\right)^2$ the influence of the Hurst parameter is squared. The convergence becomes much quicker as the Hurst parameter approaches 0.5, and eventually reaches the point where it converges immediately at $H=0.5$, as $h(X_i) = h(\raisebox{\depth}{\(\chi\)}), \forall i \in \mathbb{Z^+}$ in the absence of any correlations.
\section{Entropy rate function for Fractional Gaussian Noise}\label{entropy_fgn}
We want to understand the effect of memory on the entropic properties
of a stochastic process. We start with the entropy rate
characterisation for Gaussian processes originally derived by
Kolmogorov (\cite[pg. 76]{ihara1993information})
\begin{equation}
\label{eqn:ihara}
h(\raisebox{\depth}{\(\chi\)})
= \frac{1}{2} \log(2\pi e) + \frac{1}{4 \pi} \int_{-\pi}^{\pi} \log (2\pi f(\lambda)) \, d\lambda,
\end{equation}
where $f(\lambda)$ is the spectral density, {\em i.e.,} the Fourier transform
of the autocovariance function for a mean zero process.
We'll begin by investigating the spectral density of Fractional
Gaussian Noise (FGN), which is given by~\cite{beran1994statistics} as
\begin{equation}
\label{eqn:fgn_spectral_density}
f(\lambda) = 2 c_f (1 - \cos\lambda) \sum_{j=-\infty}^{\infty} |2 \pi j + \lambda |^{-2H -1},
\end{equation}
where, $c_f = \frac{\sigma^2}{2 \pi} \sin(\pi H) \Gamma(2H + 1)$, $H$
is the Hurst parameter, and $\sigma^2$ is the variance of the process.
This spectral density is difficult to analyse as it has an infinite sum of absolute values. In particular, when we apply the entropy rate characterisation~\ref{eqn:ihara}, as it involves taking a logarithm of a sum, making analytical calculation prohibitively difficult. However, we can still use this expression to derive some properties of the entropy rate of FGN processes.
\subsection{Comparison of approximate and analytical spectral density for entropy rate calculation}
\noindent Substituting the spectral density of FGN
(\ref{eqn:fgn_spectral_density}) into the second term in the entropy
rate expression (\ref{eqn:ihara}) we get
\begin{align*}
\int_{-\pi}^{\pi} &\log\big( 2 \pi f(\lambda) \big) \, d \lambda \\
&= 2 \pi \log(4 \pi c_f) + \int_{-\pi}^{\pi} \log( 1-\cos \lambda ) d \lambda + \int_{-\pi}^{\pi} \log\left(\sum_{j=-\infty}^{\infty} |2 \pi j + \lambda|^{-2H-1}\right) d\lambda,\\
&=2 \pi \log(4 \pi c_f) - 2 \pi \log 2 + \int_{-\pi}^{\pi} \log\left(\sum_{j=-\infty}^{\infty} |2 \pi j + \lambda|^{-2H-1}\right) d\lambda.
\end{align*}
\noindent The last term is finite for all $H \in (0,1)$, since the singularity that exists when $\lambda = j = 0$ in the absolute value term is integrable. This is
important as we can then see that this does not affect the asymptotic
behaviour of FGN processes. The resulting entropy rate is
\begin{align*}
h(\raisebox{\depth}{\(\chi\)}) = \frac{1}{2} &\log(2\pi e) + \frac{1}{2}\log\left(2\sigma^2\sin(\pi H) \Gamma(2 H + 1)\right)\\
&+ \frac{1}{4\pi} \int_{-\pi}^{\pi} \log\left(\sum_{j=-\infty}^{\infty} |2 \pi j + \lambda|^{-2H - 1}\right) d \lambda.
\end{align*}
We calculate $h(\raisebox{\depth}{\(\chi\)})$ using numerical integration via Python's Scipy library (\cite{SciPy-NMeth2020}). We plot the differential entropy rate of Fractional Gaussian Noise as a function of the Hurst parameter, H, in Figure~\ref{fig: fgn_entropy_rate}. The plot shows the impact of the variance on entropy rate calculation, and hence that the entropy rate of Fractional Gaussian Noise has a large dependence on the variance, \emph{i.e.} the second order properties. Each unit increase in variance has a smaller effect on the value of the differential entropy, due to the $\log(\sigma^2)$ term.
The spectral density expression is quite cumbersome to work with and an approximation is often used, which is accurate at low frequencies~\cite[pg. 53]{beran1994statistics}. It is derived from a Taylor series expansion of the spectral density and is given by,
\begin{align*}
f(\lambda) \approx c_f |\lambda|^{1-2H}.
\end{align*}
\noindent We can obtain a closed form for the entropy rate if we substitute this
approximation into the integral in the entropy
rate expression (\ref{eqn:ihara}) to get
\begin{align*}
\int_{-\pi}^{\pi} \log (2\pi f(\lambda)) \, d\lambda
&=\int_{-\pi}^{\pi} \log\left(2\pi c_f\right) d\lambda + \int_{-\pi}^{\pi} \log\left(|\lambda|^{1-2H}\right) \, d\lambda, \\
&= 2\pi \log\left(2\pi c_f\right) + 2\left(1 -2H\right)\int_{0}^{\pi} \log\left(\lambda\right) \, d\lambda,\\
&= 2\pi \log\left(2\pi c_f\right) + 2\left(1 -2H\right)\left(\pi \log \pi - \pi\right).
\end{align*}
\noindent Note that there is a singularity at the origin of the
spectral density of LRD processes. However, the integral is still well
defined and finite in this case. Therefore the entropy rate
approximation is,
\begin{align*}
\tilde{h}(\raisebox{\depth}{\(\chi\)}) = \frac{1}{2}\log(2\pi e)
+ \frac{1}{2}\log 2 \pi c_f
+ (1-2H)(\log\pi - 1),
\end{align*}
which differs from the exact formulation only in the last term.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Plots/entropy_rate_fgn_var.pdf}
\caption{Entropy rate of Fractional Gaussian Noise as a function of the Hurst Parameter. The maximum is at $H=0.5$, where the process is white Gaussian noise. As $H \rightarrow 0$ or $1$, the function tends towards $-\infty$, as the strength of the negative or positive correlations increase. The impact of changing variance decreases as the variance increase, due to the $\log(\sigma^2)$ term.}
\label{fig: fgn_entropy_rate}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Plots/entropy_rate_comparison_actual_approx_grid.pdf}
\caption{Comparison of the numerically integrated spectral
density and the spectral density approximation. The
approximation is relatively good for $H \geq \nicefrac{1}{2}$ but an
underestimate for $H \leq \nicefrac{1}{2}$.}
\label{fig: fgn_entropy_rate_comparison}
\end{figure}
Figure~\ref{fig: fgn_entropy_rate_comparison} shows the
entropy rate and its approximation. We can see that the entropy rate
approximation is very good for the positively correlated cases $H \geq
0.5$ and at the limits around $H=0$ or $1$. However for moderately,
negatively-correlated processes the approximation is a noticeable
underestimate of the entropy rate.
\subsection{Properties of Entropy rate for Fractional Gaussian Noise}
Figure~\ref{fig: fgn_entropy_rate_comparison} shows some interesting
properties
\begin{itemize}
\item The entropy rate as a function of $H$ is not symmetric. Negatively
correlated processes seem to have higher uncertainty the
same distance from $H=0.5$.
\item The entropy rate asymptotically tends to $-\infty$ as $H
\rightarrow$ 0 or 1.
\item The maximum entropy rate occurs at 0.5. Indicating that
the maximum entropy occurs for white Gaussian noise.
\end{itemize}
We explain how these properties emerge below.
\subsubsection{Asymptotic behaviour}
\begin{theorem}
The approximate differential entropy rate of Fractional
Gaussian Noise, $\tilde{h}(\raisebox{\depth}{\(\chi\)}) \rightarrow -\infty$ as $H
\rightarrow 0$ or $1$.
\end{theorem}
\begin{proof}
When $H \rightarrow 0$ or $1$, the term $c_f \rightarrow 0$,
as the gamma function terms are finite, however the
trigonometric terms tend to 0 as $H$ tends to an integer
value. Hence, asymptotically the approximate entropy rate
expression is dominated by $\log c_f \rightarrow -\infty$, as
$c_f \rightarrow 0$.
\end{proof}
\begin{remark}
Note that the approximation works well in the limits $\mathcal{H} \rightarrow
0$ or 1, and so the theorem describes the asymptotic behaviour of
entropy rate well. Moreover, the theorem lines up with the
intuition for an LRD process. As we move closer to either perfectly
positively or negatively correlated, the process becomes ``less
uncertain", \emph{i.e.}, we have less entropy on average. When the
uncertainty disappears, by viewing the entire past we can
accurately infer the current value. It's important to reiterate
that the differential entropy can be $-\infty$, which can be
interpreted as least uncertainty for a process.
\end{remark}
\subsubsection{Maximum}
We want to understand the maximum of differential entropy rate, as a
function of the Hurst parameter. This will provide an understanding of
which parameter choices represent the highest uncertainty. We
differentiate the entropy rate, with respect to $H$ and then solve for
$H$ when the derivative equals zero. Here we need to apply this to the
exact formula because the approximation distorts the location of the
maximum. Therefore, dropping constant terms, we get
\begin{align*}
\frac{dh}{dH}
&= \frac{1}{2} \frac{d}{dH} \log\left(\sigma^2\sin(\pi H) \Gamma(2 H + 1)\right)\\
& \qquad+ \frac{1}{4 \pi} \frac{d}{dH} \int_{-\pi}^{\pi} \log\left(\sum_{j=-\infty}^{\infty} |2 \pi j + \lambda|^{-2H - 1}\right) d \lambda \\
&= \frac{1}{2} \frac{d}{dH} \log\big(\sin(\pi H)\big) + \frac{1}{2} \frac{d}{dH} \log\big(\Gamma(2 H + 1)\big)\\
&\qquad - \frac{1}{2 \pi} \int_{-\pi}^{\pi} \frac{\sum_{j} \log(|2 \pi j + \lambda|) |2 \pi j + \lambda|^{-2H -1}}{\sum_{j} |2 \pi j + \lambda|^{-2H -1}} d\lambda \\
&= \frac{\pi}{2} \cot(\pi H) + \psi(2 H + 1) - \frac{1}{2 \pi} \int_{-\pi}^{\pi} \frac{\sum_{j} \log(|2 \pi j + \lambda|) |2 \pi j + \lambda|^{-2H -1}}{\sum_{j} |2 \pi j + \lambda|^{-2H -1}} d\lambda.
\end{align*}
\noindent where $\psi(z) = \Gamma^\prime(z)/\Gamma(z)$ is the digamma
function.
Then we set this expression to zero, and solve for $\mathcal{H}$. This is a
transcendental equation with no closed form. We solve it numerically
using Python's SciPy package (\cite{SciPy-NMeth2020}), which yields $\mathcal{H}
\approx 0.500$. Therefore we conclude that the maximum entropy rate, using the exact spectral density, is at $\mathcal{H}=0.5$, which aligns with the idea that a SRD process has more uncertainty than any equivalent LRD
process.
Note that from the solution of the spectral density approximation is
$H \approx 0.516$. So although using the spectral density
approximation is acceptable for many purposes, it can lead to false
conclusions about the properties of the differential entropy rate.
\section{Excess Entropy for Stationary Gaussian Processes}\label{excess_entropy}
In this section we investigate the behaviour of the excess entropy, $E$, for Gaussian processes which have an autocorrelation function which decays as a power law.
For stationary processes, the terms in the excess entropy will tend to zero as $n$ increases because
\begin{align*}
\lim\limits_{n \rightarrow \infty} h_e(n) = \lim\limits_{n \rightarrow \infty} h(X_n | X_{n-1}, ... , X_1) = h(\raisebox{\depth}{\(\chi\)}).
\end{align*}
However, we will investigate the nature of the convergence using the conditional entropy to gain some additional insight.
We begin by looking at the behaviour of the individual terms of the excess entropy series to understand why these terms decay to zero.
Since $h(X_n ,... , X_1) = \frac{1}{2} \log\left((2 \pi e)^n |K^{(n)}| \right)$ for a finite collection of random variables of a Gaussian process~\cite[pg. 416]{cover_thomas_2006}, we have
\begin{align}
h_e(n) &= h(X_n ,... , X_1) - h(X_{n-1}, ... , X_1),\nonumber\\
&= \frac{1}{2} \log\left((2 \pi e)^n |K^{(n)}| \right) - \frac{1}{2} \log\left((2 \pi e)^{n-1} |K^{(n-1)}| \right),\nonumber\\
&= \frac{1}{2} \log\left(2 \pi e \frac{|K^{(n)}|}{|K^{(n-1)}|}\right),\label{eq:excess_entropy_term}
\end{align}
where $K^{(n)}$ is the autocovariance matrix of the process $X_n$. Note that $K^{(n)}$ is a Toeplitz matrix of size $n \times n$.
We analyse each element of the infinite series of differential excess entropy by substituting the characterisation of $h_e(n)$ in~(\ref{eq:excess_entropy_term}) and utilising the entropy rate characterisation for Gaussian processes given in~(\ref{eqn:ihara}). This gives
\begin{align*}
h_e(n) - h(\raisebox{\depth}{\(\chi\)}) &= \frac{1}{2} \log\left(2 \pi e \frac{|K^{(n)}|}{|K^{(n-1)}|}\right) - \frac{1}{2} \log (2 \pi e) - \frac{1}{4\pi} \int_{-\pi}^{\pi} \log f(\lambda) d\lambda,\\
&= \frac{1}{2} \log\frac{|K^{(n)}|}{|K^{(n-1)}|} - \frac{1}{4\pi} \int_{-\pi}^{\pi} \log f(\lambda) d\lambda. \numberthis \label{eqn:excess_ent_convergence}
\end{align*}
\noindent The expression, (\ref{eqn:excess_ent_convergence}), informs us about the rate of convergence of the conditional entropy to the entropy rate, and about the general convergence of the excess entropy. If $\log f(\lambda)$ is integrable, there is a well known limit theorem by Szeg{\"o} (see~\cite{szego1915, bingham2012}), given below and used to evaluate the limit of $\frac{|K^{(n)}|}{|K^{(n-1)}|}$ as $n \rightarrow \infty$.
First we show that $\log f(\lambda) \in L_1[-\pi, \pi]$, for LRD processes and hence that Szeg{\"o}'s limit theorem can be applied. We use the asymptotic form to analyse this case, as the issue with integrability is the singularity that exists at the origin of the spectral density functions. Therefore, the integrability of the asymptotic form implies the integrability for LRD processes.
\begin{theorem}
For a spectral density, $f(\lambda)$ of a LRD stochastic process, $\log f(\lambda) \in L_1[-\pi, \pi]$.
\end{theorem}
\begin{proof}
Using the asymptotic expression for spectral density for LRD processes,
$f(\lambda) \sim c_f|\lambda|^{-\beta}, \text{ } \beta \in (0,1), \text{ } 0 < c_f < \infty$.\\
Hence,
\begin{align*}
\int_{-\pi}^{\pi} \log(c_f|\lambda|^{-\beta}) d \lambda &= \int_{-\pi}^{\pi} \log c_f d \lambda - \beta \int_{-\pi}^{\pi} \log|\lambda| d \lambda,\\
&= 2 \pi \log c_f - 2 \beta \int_{0}^{\pi} \log \lambda\ d\lambda,\\
&= 2 \pi \log c_f - 2 \beta \left[\lambda \log \lambda - \lambda \right]_0^\pi,\\
&= 2 \pi \log c_f - 2 \beta \left[\pi \log \pi - \pi \right],
\end{align*}
Since we have that $0\log 0 = 0$. By the finiteness of all of the terms, this implies that,
\begin{align*}
\int_{-\pi}^{\pi} \log f(\lambda) d \lambda < \infty.
\end{align*}
\end{proof}
The following theorem was originally formulated by~\cite{szego1915}, and then extended to include another term in~\cite{szego1952certain}. This theorem has a history of having the conditions generalised, and applications being found in many areas in functional analysis, statistics and probability~\cite[pg. 145-228]{grenander1958toeplitz}, to calculate functions of eigenvalues of Toeplitz matrices. The statement of the theorem that we will be using is by~\cite{bingham2012}, which is from the probabilistic perspective and includes the most recent generalisations in the conditions.
\begin{theorem}[Szeg{\"o}'s Theorem (\cite{bingham2012})]\label{thm:szego_theorem}
For a sequence of Toeplitz matrices, $\Gamma^{(n)} = [\gamma(|i-j|)]_{i,j}$, of increasing size $n$, where $\gamma(n) = \int_{-\pi}^{\pi} f(\lambda) e^{in\lambda} d \lambda$, and $f(\lambda)$ is a spectral density such that $\log f(\lambda) \in L_1[-\pi, \pi]$, then there exists a limit
\begin{align*}
\lim\limits_{n \to \infty} | \Gamma^{(n)} |^{\frac{1}{n}} = \exp \left\lbrace\frac{1}{2\pi}\int_{-\pi}^{\pi} \log f(\lambda) d\lambda \right\rbrace.
\end{align*}
\end{theorem}
\noindent We apply this theorem to the ratios of the autocovariance function, $K^{(n)}$, of the process, $\{X_n\}_{n \in \mathbb{N}}$. Due to the finiteness of this limit we can take the $n$th power, for a covariance matrix $K^{(n)}$. Which has the following form,
\begin{align*}
\lim\limits_{n \to \infty} | K^{(n)} | &= \exp \left\lbrace\frac{1}{2\pi}\int_{-\pi}^{\pi} \log f(\lambda) d\lambda \right\rbrace^n,
\end{align*}
taking the $n$th power. Applying this result to the limit of the $n$th term of the differential excess entropy (\ref{eqn:excess_ent_convergence}), gives
\begin{align*}
\lim\limits_{n \to \infty} h_e(n) - h(\raisebox{\depth}{\(\chi\)})
&= \frac{1}{2} \log \left( \exp \left\lbrace\frac{1}{2\pi}\int_{-\pi}^{\pi} \log f(\lambda) d\lambda \right\rbrace\right) - \frac{1}{4\pi} \int_{-\pi}^{\pi} \log f(\lambda) d\lambda,\\
&= 0.
\end{align*}
\noindent As $n \to \infty, h_e(n) - h(\raisebox{\depth}{\(\chi\)}) \to 0$, which implies the convergence of the conditional entropy, conditioned on the infinite past, to the entropy rate, equivalent to the result of Theorem~\ref{conditional_entropy_rate}. We have gained some additional insight that the summands converge to zero in the limit as $n \rightarrow \infty$ due to the convegence of the ratio of subsequent covariance matrix determinants. However, this does not tell us anything about the rate of convergence of the terms or whether $E$ converges at all, but this approach can be extended to garner such information.
In following section we will use a stronger version of the Szeg{\"o} theorem, with an additional term in the limit. This gives an approach to analyse the convergence properties to the entropy rate that arise from observing the conditional entropy, conditioned on the past observations.
\section*{Acknowledgment}
This research was funded by CSIRO's Data61, the Australian Research Council's Centre of Excellence in Mathematical and Statistical Frontiers and Defence Science and Technology Group. Thanks to Adam Hamilton for his assistance in providing feedback on this manuscript.
\def\href#1#2{#2}
\section{Introduction}
The entropy rate of discrete time stochastic
processes has been studied as a measure of the average
uncertainty. Most investigations of this type have focussed on
processes whose correlations decay quickly, and hence the dependence
on past observations disappears rapidly. However, many real processes
from a variety of contexts, {\em i.e.,} data
networks ({\em e.g.,} \cite{Leland:1993:SNE:166237.166255,
willinger_self_similar_high_variability_1997, willinger1995}),
climate ({\em e.g.,} \cite{varotsos2006}), hydrology ({\em e.g.,} \cite{beran1994statistics,
lawrence1977}), and economics ({\em e.g.,} \cite{willinger_stock_1999, cont2005}), have
been shown to exhibit long range dependence, meaning correlations
exist between past and future observations that cannot be ignored at
any time lag.
Information and coding theory have had profoundly important uses in
signal processing and communication. Noise processes are an important
part of this story. However, in most works, for instance on designing
optimal codes on noisy channels, the noise processes are presumed to
be short range dependent. However, as far back as 1965, ~\cite{1089090}
showed that some noise processes are also long-range
dependent (though the terminology was still developing).
Recent work has investigated an information theoretic characterisation
of long range and short range processes (e.g. \cite{Li_2004, Chavez_2016,
ding2016entropic}), using the finiteness of mutual information
between past and future. We aim to clarify this characterisation and
investigate its implications.
This paper calculates the differential entropy rate for the two most
common stationary Gaussian Long Range Dependent (LRD) processes:
Fractional Gaussian Noise (FGN) and the Auto-Regressive
Fractionally-Integrated Moving Average (ARFIMA) process. We start by
deriving the entropy rate for these processes, and show that they both
have negative poles as the processes tend towards strong long-range
correlations, but that their behaviour when anti-correlated is
surprisingly different: FGN has a pole similar to that for positive
correlations, but ARFIMA does not. This contradicts common intuition
based on their similar spectral densities that FGN and ARFIMA(0,d,0) are close to equivalent.
We also investigate the links between the two information measures:
excess entropy and the mutual information between past and future
processes, and compare these to the differential entropy rate. We show
that the differential entropy rate definition for excess entropy is
equivalent to the mutual information between past and future for
continuous valued discrete time Gaussian processes, and hence that
excess entropy is infinite for all long range dependent Gaussian
processes.
Finally, estimators, such as the sample mean, applied to LRD processes
have been shown to have slow convergence rates, which can lead to a
larger than expected uncertainty when investigating these
processes. We ask, \textit{``Does this behaviour apply to estimators
of entropy?"} and \textit{``What is the impact of the degree of
positive or negative correlations on the entropy rate?"} We show
that while the convergence rate of the conditional entropy of short
range dependent Gaussian processes is in the order of $n^{-1}$, the
rate of convergence for LRD Gaussian processes is slower. Although
this parallels many of the other results for LRD processes, the actual
rate of convergence is different at $O\Big((1-2H)^2 \log(n)/n \Big)$, where $H$ is
the Hurst parameter. As $H \rightarrow 1$ we can see that this
convergence is at its worst, though $\mathcal{H}$ only appears as a factor not in the exponent.
\section{Mutual Information and Excess Entropy for Long Range Dependent Processes}\label{mutual_information}
In this section we continue analysing of the differential entropy rate for stochastic processes with power-law decaying covariance function. We investigate the links between the amount of entropy that is accumulated during the convergence of the conditional entropy to the entropy rate and the amount of information that is shared between the past and future of a stochastic process.
We extend the standard notion of mutual information to the special case of mutual information between past and future, $I_\pf$, which will measure the amount of information about the infinite future, given knowledge of the infinite past of stochastic processes.
\begin{definition}
The mutual information for continuous random variables is defined as
\begin{align*}
I(X;Y) = \int_{\mathcal{Y}} \int_{\mathcal{X}} f(x, y) \log\left(\frac{f(x,y)}{f(x)f(y)}\right)dx dy,
\end{align*}
and in particular the mutual information between past and future with $n$ lags, $I^{(n)}$, for a stochastic process $\{X_i\}_{i \in \mathbb{Z}}$ is defined as $I(\{X_s, s < 0\},\{X_s, s \ge n\})$. The case with $n=0$ is called the mutual information between past and future, $I_\pf$, and is of special interest.
\end{definition}
We present a theorem from~\cite{Li_2004}, that links the value of $I_\pf$, and autocovariance function and the Fourier coefficients of the logarithm of the spectral density function.
\begin{theorem}[From~\cite{Li_2004}]\label{Li_thm_1}
Let $\{X_i\}_{i \in \mathbb{Z}}$ be a stationary Gaussian stochastic process:
\begin{itemize}
\item if the spectral density $f(\lambda)$ is continuous and $f(\lambda) > 0$, then $I_\pf$ is finite if and only if the autocovariance function satisfies the condition $\sum_{k=-\infty}^{\infty} k \gamma(k)^2 < \infty$.
\item $I_\pf$ is finite if and only if the cepstrum coefficients, $b_k = \frac{1}{2 \pi} \int_{-\pi}^{\pi} \log f(\lambda)e^{-ik\lambda} d\lambda$, satisfy the condition $\sum_{k=1}^{\infty} kb_k^2 < \infty$. In this case, $I_{p-f} = \frac{1}{2} \sum_{k=1}^{\infty} kb_k^2$.
\end{itemize}
\end{theorem}
\begin{remark}
The convergence of the sum $\sum_{k=-\infty}^{\infty} k b_k^2$ requires that $\sum_{k=1}^{\infty} k b_k^2 < \infty$ and $\sum_{k=1}^{\infty} -k b_{-k}^2 < \infty$ separately.
\end{remark}
\noindent This theorem gives us a way to classify whether processes have infinite mutual information between past and future, in this paper we will use this quantity to analyse convergence towards the entropic rate, in particular for LRD stochastic processes.
In the next result we make an explicit link between LRD processes and the finiteness of the mutual information between past and future. This perspective provides us with a characterisation of LRD processes, these are processes that ``share infinite information from the past to the future".
\begin{theorem}\label{mutual_infinite}
The Mutual Information between past and future, $I_\pf$, for stationary LRD Gaussian processes is infinite
\end{theorem}
\begin{proof}
To analyse LRD processes we use the spectral density asymptotic representation around the origin, $f(\lambda) \sim c_f|\lambda|^{1-2H}$, as we can see the divergence by considering the asymptotic behaviour around the singularity at the origin. Hence, as $b_k = \frac{1}{2 \pi} \int_{-\pi}^{\pi} \log f(\lambda) e^{-ik\lambda} d \lambda$, then
\begin{align*}
b_k &\sim \frac{1}{2 \pi} \int_{-\pi}^{\pi} \log\left(c_f|\lambda|^{1-2H}\right) e^{-ik\lambda} d \lambda,\\
&= \frac{\log c_f}{2 \pi} \int_{-\pi}^{\pi} e^{-ik\lambda} d \lambda + \frac{1-2H}{2 \pi} \int_{-\pi}^{\pi} \log |\lambda| e^{-ik\lambda} d \lambda.\\
\end{align*}
We split the integral into positive and negative components,
\begin{align*}
b_k &= \frac{2 \log c_f \sin\left(\pi k\right)}{2 \pi k} + \frac{1-2H}{2 \pi} \int_{0}^{\pi} \log \lambda \left(e^{-ik\lambda} + e^{ik\lambda}\right) d \lambda.
\end{align*}
Since $\sin(\pi k) = 0, \forall k \in \mathbb{Z}$, the first term vanishes and we only need to consider the integral. This can be decomposed into a trigonometric representation since $e^{-ik\lambda} + e^{ik\lambda} = 2\cos\left(k \lambda\right)$, so
\begin{align*}
b_k &\sim \frac{1-2H}{2 \pi} \int_{0}^{\pi} 2 \log \lambda \cos\left(k \lambda\right) d \lambda.
\end{align*}
Integrating this expression by parts, we get
\begin{align*}
b_k &\sim \frac{1-2H}{2 \pi} \left[\frac{\log \lambda \sin\left(k\lambda\right)}{k}\right]_0^\pi - \int_{0}^{\pi} \frac{\sin\left(k\lambda\right)}{k\lambda} d \lambda.
\end{align*}
To analyse the integral in the second term we use the substitution, $u = k\lambda$, and therefore
\begin{align*}
\int_{0}^{\pi} \frac{\sin\left(k\lambda\right)}{k\lambda} d \lambda &= \int_{0}^{k\pi} \frac{\sin\left(u\right)}{u} du.
\end{align*}
Then undoing the u-substitution and noting that Si$(x) = \int_{0}^{x} \frac{\sin t}{t} dt$, is the Sine integral,
\begin{align*}
b_k &\sim \frac{1-2H}{\pi}\left[\frac{\log \lambda \sin\left(k \lambda\right) - \text{Si}(\lambda)}{k}\right]_0^\pi,\\
&= \frac{\left(1-2H\right)\left(-\text{Si}\left(\pi\right)\right)}{\pi k}.
\end{align*}
The partial sum has the asymptotic form,
\begin{align*}
\sum_{k=1}^{m} k b_k^2 &\sim \sum_{k=1}^{m} \frac{\left(1-2H\right)^2\ Si\left(\pi\right)^2}{\pi^2 k},\\
&= \frac{\left(1-2H\right)^2\ Si\left(\pi\right)^2}{\pi^2} \sum_{k=1}^{m} \frac{1}{k},\\
&\sim \frac{\left(1-2H\right)^2\ Si\left(\pi\right)^2}{\pi^2} \log m,
\end{align*}
because the harmonic series, $\sum_{k=1}^{m} \frac{1}{k} \sim \log(m)$ as $m \rightarrow \infty$. Hence, as $m \rightarrow \infty$, the sum, $\sum_{k=1}^{m} k b_k^2$, diverges. This implies that the sum, $\sum_{k=-\infty}^{\infty} k b_k^2$, diverges and therefore by Theorem~\ref{Li_thm_1}, $I_\pf$ is infinite.
\end{proof}
\begin{remark}
We can quite easily show this result with the additional assumptions that the spectral density, $f(\lambda)$ is positive and continuous. The asymptotic expression of the autocovariance function, $\gamma(k) \sim \sigma^2 c_{\rho} |k|^{-\alpha}$. Hence, considering the sum, $\sum_{k=1}^{\infty} k \gamma(k)^2$, from the first part of Theorem~\ref{Li_thm_1},
\begin{align*}
\sum_{k=1}^{\infty} k \gamma(k)^2 &\sim \sum_{k=1}^{\infty} k (\sigma^2 c_{\rho} |k|^{-\alpha})^2 \\
&= \sigma^4 c_{\rho}^2 \sum_{k=1}^{\infty} k^{1-2\alpha}.
\end{align*}
This sum diverges in the parameter range, $\alpha \in (0,1)$, and hence by Theorem~\ref{Li_thm_1}, $I_\pf$ is infinite for LRD processes.
There exist many processes that have infinite excess entropy but are not long range dependent. Some examples are given, including deterministic processes, in~\cite{crutchfield_feldman_2003}.
\end{remark}
\cite{crutchfield_feldman_2003} analysed a quantity named excess entropy, $\sum_{n=1}^{\infty} \left(H(X_n|X_{n-1}, \ldots, X_0) - H(\raisebox{\depth}{\(\chi\)})\right)$, for the Shannon entropy $H$ and corresponding entropy rate $H(\raisebox{\depth}{\(\chi\)})$, which has been shown to be equivalent to the mutual information between past and future. This has been named, with implicit interpretation, as stored information ({\em e.g.,} \cite{shaw1984dripping}), effective measure complexity ({\em e.g.,} \cite{grassberger1986toward, lindgren1988complexity}), predictive information ({\em e.g.,} \cite{nemenman2000information}). Importantly, it has been used to measure the convergence rate of the conditional entropy, based on past observations, to the entropy rate. We aim to extend this result to differential entropy, and then the question of classification of LRD processes via the amount of shared information can be made by the convergence rate to the entropy rate. We extend the definition of the excess entropy to the case of differential entropy.
\begin{definition}\label{def:differential_excess_entropy}
The differential \emph{excess} entropy, $E$, of a stochastic process, $\{X_i\}_{i \in \mathbb{N}}$, is defined as,
\begin{align*}
E &= \sum_{n=1}^{\infty} \left(h_e(n) - h(\raisebox{\depth}{\(\chi\)})\right),\\
&= \lim\limits_{n \rightarrow \infty} \left[ h(X_n, \ldots, X_0) - n h(\raisebox{\depth}{\(\chi\)})\right].
\end{align*}
where
\begin{align*}
h_e(n) &= h(X_1, .. , X_n) - h(X_1, ... , X_{n-1}),\\
&= h(X_n | X_{n-1}, ... , X_1).
\end{align*}
\end{definition}
We have the tools available to make an explicit link between the mutual information between past and future and the excess entropy of a continuous-valued, discrete-time stochastic process. This is an exact analogue of Proposition 8 from~\cite{crutchfield_feldman_2003} and has been stated utilising a different approach by~\cite{ding2016entropic}.
\begin{theorem}\label{mutual_excess}
For a stationary, continuous-valued stochastic process, the mutual information between past and future, $I_\pf$, is equal to the differential excess entropy, $E$.
\end{theorem}
\begin{proof}
The mutual information for a process $X$, with a past and future of $n$ observations,
\begin{align*}
I[\{X_s, -n \le s < 0\}; \{X_s, 0 \le s \le n\}] &= h(X_0, ..., X_{n-1}) - h(X_0, ... , X_{n-1} | X_{-n}, ... , X_{-1}),\\
&= \sum_{i=0}^{n-1} h(X_i | X_{i-1}, ... , X_0) - \sum_{i=0}^{n-1} h(X_i | X_{i-1}, ... , X_{-n}),\\
&= \sum_{i=0}^{n-1} \bigg(h(X_i | X_{i-1}, ... , X_0) - h(X_i | X_{i-1}, ... , X_{-n})\bigg),
\end{align*}
by the chain rule of differential entropy~\cite[pg. 253]{cover_thomas_2006}.
Then we consider the mutual information between past and future, by taking the limit of the above expression as $n \rightarrow \infty$, which leads to
\begin{align*}
I_\pf &= \lim\limits_{n \rightarrow \infty} \left[ \sum_{i=0}^{n-1} \bigg(h(X_i | X_{i-1}, ... , X_0) - h(X_i | X_{i-1}, ... , X_{-n})\bigg)\right],\\
&= \lim\limits_{n \rightarrow \infty} \left[ \sum_{i=0}^{\infty} \bigg( h(X_i | X_{i-1}, ... , X_0) - h(X_i | X_{i-1}, ... , X_{-n}) \bigg)\mathbbm{1}_{\{i \le n\}}\right].
\end{align*}
We define the sequence of measurable functions, $f_n(i)$ as
\begin{align*}
f_n(i) &= \bigg( h(X_i | X_{i-1}, ... , X_0) - h(X_i | X_{i-1}, ... , X_{-n}) \bigg)\mathbbm{1}_{\{i \le n\}},
\end{align*}
and we define the function, $g(i)$ as
\begin{align*}
g(i) &= h(X_i | X_{i-1}, ... , X_0) - h(\raisebox{\depth}{\(\chi\)}).
\end{align*}
We want to show that $|f_n(i)| \le g(i)$ for all $n$ and for all $i \in \mathbb{N}$. In this case it is equivalent to showing that $f_n(i) \le g(i)$, since $f_n(i) \ge 0$ for all $n,i \in \mathbb{N}$, as the second term of $f_n(i)$ conditions on more random variables, and since conditioning cannot increase entropy this implies that $h(X_i | X_{i-1}, ... , X_0) \ge h(X_i | X_{i-1}, ... , X_{-n})$. We consider two cases, $i \le n$ and $i > n$, separately. In the case, $i > n$, we have that $f_n(i) = 0$, and since $g(i) \ge 0$ for all $i$, this implies that $f_n(i) \le g(i)$. Considering the second case, $i \le n$, we have that
\begin{align*}
f_n(i) &= \bigg(h(X_i | X_{i-1}, ... , X_0) - h(X_i | X_{i-1}, ... , X_{-n})\bigg),
\end{align*}
and therefore,
\begin{align*}
g(i) - f_n(i) &= h(X_i | X_{i-1}, ... , X_{-n}) - h(\raisebox{\depth}{\(\chi\)}).
\end{align*}
Again, since conditioning does not increase entropy and the characterisation of entropy rate for stationary processes from Theorem~\ref{conditional_entropy_rate} this implies that $g(i) - f_n(i) \ge 0$ and therefore $g(i) \ge f_n(i)$ for all $n,i$ such that $i \le n$. Then we can apply the dominated convergence theorem~\cite[pg. 26]{Durrett2010}, since $f_n(i) \rightarrow (X_i | X_{i-1}, ... , X_0) - h(\raisebox{\depth}{\(\chi\)})$ pointwise, this implies that
\begin{align*}
I_\pf &= \lim\limits_{n \rightarrow \infty} \sum_{i=0}^{\infty} \bigg( h(X_i | X_{i-1}, ... , X_0) - h(X_i | X_{i-1}, ... , X_{-n}) \bigg)\mathbbm{1}_{\{i \le n\}},\\
&= \sum_{i=0}^{\infty} \lim\limits_{n \rightarrow \infty} \bigg( h(X_i | X_{i-1}, ... , X_0) - h(X_i | X_{i-1}, ... , X_{-n}) \bigg)\mathbbm{1}_{\{i \le n\}},\\
&= \sum_{i=0}^{\infty} \bigg( h(X_i | X_{i-1}, ... , X_0) - h(\raisebox{\depth}{\(\chi\)}) \bigg).
\end{align*}
\end{proof}
\begin{remark}
This proof is similar to that of Proposition 8 from~\cite{crutchfield_feldman_2003}. However, it is more rigorous since the limit is kept out the front of the sum while simultaneously applied to the second term in the sum. This approach using dominated convergence can resolve the issue in their proof.
\end{remark}
\noindent In~\cite{crutchfield_feldman_2003}, they analyse the excess entropy of discrete random variables to understand the convergence rate of the conditional entropy to the entropy rate. From this link we can utilise our knowledge of the mutual information between the past and future to classify the rate of convergence of the conditional entropy to the entropy rate.
This result and Theorem~\ref{mutual_infinite}, lead to the following corollary which gives us an approach to understand the entropy rate convergence by conditional entropy, of Gaussian LRD processes and is a generalisation of a theorem stated for LRD FGN processes (\cite{ding2016entropic}).
\begin{corollary}
The excess entropy of an LRD Gaussian process is infinite.
\end{corollary}
\begin{proof}
This is shown by combining Theorem~\ref{mutual_infinite} and Theorem~\ref{mutual_excess}.
\end{proof}
\noindent We will use this idea in the subsequent sections to analyse the excess entropy, given its relationship to convergence to entropy rate which is noted for Shannon entropy rate by~\cite{crutchfield_feldman_2003}.
|
\section{Introduction}
\label{sec:introduction}
Quantum computing promises to achieve an exponential speedup to tackle certain computational tasks compared with the classical computers~\cite{doi:10.1137/S0097539795293172,cao2019quantum,egger2020quantum,farhi2014quantum,lanyon2010towards,tang2020quantum,kerenidis2020quantum}. Although quantum technologies are continuously improving, current quantum devices are still qualified as Noisy Intermediate-Scale Quantum (NISQ) hardware~\cite{Preskill2018quantumcomputingin}, with several physical constraints. For example, for superconducting devices which we target in this paper, connections are only allowed between two neighbouring qubits. Besides, the gate operations of NISQ devices are noisy and have unavoidable error rates. As we do not have enough number of qubits to realize Quantum Error Correction~\cite{calderbank1998quantum,calderbank1996good,fowler2012surface}, only small circuits with limited depth can obtain reliable results when executed on quantum hardware, which leads to the waste of hardware resource. Moreover, with the growing demand to access to quantum hardware, its under-utilization issue increases the waiting time for users, which indicates the need to improve the hardware throughput.
As the qubit number of the hardware increases and the error rates improve, it becomes possible to execute multiple circuits on a quantum chip simultaneously. The multi-programming mapping problem was firstly introduced by~\cite{das2019case}, which demonstrated that the throughput and utilization of NISQ hardware can be enhanced by executing several circuits at the same time. Ref~\cite{dou2020new} further improved it in terms of fidelity and gate number by proposing a Community Detection Assisted Partition algorithm along with the X-SWAP scheme (we refer to this algorithm as CDAP for brevity). However, their results showed that when executing multiple quantum circuits simultaneously, the activity of one circuit can negatively impact the fidelity of others, due to the difficulty of allocating reliable regions to each circuit, higher chance of crosstalk error~\cite{sheldon2016procedure}, and the qubit movement limitation (only inside of the partition). Previous works~\cite{das2019case,dou2020new} have left these issues largely unexplored and have not addressed the problem holistically: (1) Hardware topology and calibration data are not fully analyzed where allocation is done on unreliable or sparse-connected partitions to circuits ignoring the robust qubits and links. (2) These works use only \texttt{SWAP} gate for mapping transition process and the modified circuits always have a large number of additional gates. (3) Crosstalk error is not considered when allocating partitions for circuits. For example, the X-SWAP scheme~\cite{dou2020new} for reducing the inserted \texttt{SWAP} number can only be performed when the two circuits are allocated to neighbouring partitions, which can introduce crosstalk effect and decrease the circuit output fidelity. Detrimental crosstalk impact when executing multiple parallel instructions has been reported in ~\cite{murali2020software,ash2020analysis,ash2020experimental} by using Simultaneous Randomized Benchmarking (SRB)~\cite{gambetta2012characterization}. In presence of crosstalk, gate error can be increased by an order of magnitude. Ref~\cite{ash2020analysis} even proposed a fault-attack model using crosstalk in a multi-programming environment.
It is important to investigate the multi-programming approach in the NISQ era especially for Variational Quantum Algorithms (VQAs)~\cite{cerezo2020variational}. For example, the multi-programming mechanism can enable to execute several ansatz states in parallel in one quantum processor, such as in Variational Quantum Eigensolver (VQE)~\cite{peruzzo2014variational,kandala2017hardware}, Variational Quantum Linear Solver (VQLS)~\cite{bravo2020variational,huang2019near}, or Variational Quantum Classifier (VQC)~\cite{havlivcek2019supervised,romero2019variational} with reliability. It is also general enough to be applied to other quantum circuits regardless of applications or algorithms.
In this work, we address the problem of multi-programming while considering the impact of hardware topology, calibration data, and crosstalk without losing the circuit fidelity. First, we introduce a parallelism manager that can optimally select the number of circuits being executed on the quantum hardware simultaneously. Second, we present two different qubit partition algorithms to allocate reliable partitions to different circuits. One is a greedy partition algorithm which provides optimal choices. The other one is based on a heuristic which can give nearly optimal results and significantly reduce the time complexity. Third, we consider crosstalk error during the partition process to lower the crosstalk effect during simultaneous executions. Then, we improve the mapping transition step of the qubit mapping problem to make quantum circuits executable on quantum hardware with a reduced number of additional gates. Finally, we evaluate our algorithm on real quantum hardware by first executing circuits of different sizes at the same time and then applying it to VQE algorithm to estimate the ground state energy of deuteron. To the best of our knowledge, this is the first attempt to propose a complete multi-programming process flow for executing an optimal number of workloads in parallel ensuring the output fidelity by analyzing the hardware limitations.
\section{Results}
\subsection{Multi-programming workflow}
\begin{figure*}[h]
\centering
\includegraphics[scale=0.5]{overflow3.pdf}
\caption{Overview of the proposed multi-programming framework. {\normalfont The input layer includes the quantum hardware information and multiple quantum circuit workloads. The parallelism manager helps to decide whether executing circuits simultaneously or independently. For simultaneous executions, it works with the hardware-aware multi-programming compiler to select an optimal number of shared workloads to be executed at the same time. Then, the scheduler makes all the circuits executable on the quantum hardware and we can obtain the results of the output circuits.}
}
\label{fig:workflow}
\end{figure*}
The multi-programming workflow is schematically shown in Fig.~\ref{fig:workflow}, which includes the following steps:
\begin{itemize}
\item Input layer. It contains a list of small quantum circuits written in OpenQASM language~\cite{cross2017open}, and the quantum hardware information, including the hardware topology, calibration data, and crosstalk effect.
\item Parallelism manager. It can determine whether executing circuits concurrently or separately. If the simultaneous execution is allowed, it can further decide the number of circuits to be executed on the hardware at the same time without losing fidelity based on the fidelity metric included in the hardware-aware multi-programming compiler.
\item Hardware-aware multi-programming Compiler. Qubits are partitioned to several reliable regions and are allocated to different quantum circuits using qubit partition algorithms. Then, the partition fidelity is evaluated by the post qubit partition process. We introduce a fidelity metric here which helps to decide whether this number of circuits can be executed simultaneously or the number needs to be reduced based on their properties.
\item Scheduler. The mapping transition algorithm is applied and circuits are transpiled to be executable on real quantum hardware.
\item Output layer. Output circuits are executed on the quantum hardware simultaneously or independently according to the previous steps and the experimental results are obtained.
\end{itemize}
\subsection{Parallelism manager}
\begin{figure*}[t]
\centering
\includegraphics[scale=0.6]{overflow4.pdf}
\caption{Process flow of each block that constitutes our multi-programming approach. (a) {\normalfont The parallelism manager selects $K$ circuits according to their densities and passes them to the hardware-aware multi-programming compiler.} (b) {\normalfont The qubit partition algorithms allocate reliable regions to multiple circuits. $\Delta S$ is the difference between partition scores when partitioning independently and simultaneously, which is the fidelity metric. $\delta$ is the threshold set by the user. The fidelity metric helps to select the optimal number of simultaneous circuits to be executed.} (c) {\normalfont The scheduler performs mapping transition algorithm and makes quantum circuits executable on real quantum hardware.} }
\label{fig:workflow2}
\end{figure*}
In order to determine the optimal number of circuits that can be executed on the hardware in parallel without losing fidelity, here, we introduce the parallelism manager, shown in Fig.~\ref{fig:workflow2}a.
Suppose we have a list of $n$ circuit workloads with $n_i$ qubits for each of them, that are expected to be executed on $N$-qubit hardware. Firstly, the circuits are sorted according to their densities. The density of a circuit is defined as the number of \texttt{CNOTs} divided by the qubit number of the circuit, $\#CNOTs/n_i$,~\cite{dou2020new}. Then, we pick $K$ circuits which is the maximum number of circuits that are able to be executed on the hardware at the same time, $\sum_{n=1}^{K} n_i \leq N$. If $K$ is equal to one, then all the circuits should be executed independently. Otherwise, these circuits are passed to the hardware-aware multi-programming compiler. They work together to decide an optimal number of simultaneous circuits to be executed.
\subsection{Hardware-aware multi-programming compiler}
\subsubsection{Qubit partition}
\
\newline
Here, we present the key features of the qubit partition algorithms. A motivational example can be found in Supplementary Note 2.
\vspace{2mm}
\noindent
\textit{Crosstalk effect characterization.}
\vspace{2mm}
\
\newline
Crosstalk is one of the major noise sources in NISQ devices, which can corrupt a quantum state due to quantum operations on other qubits~\cite{sarovar2020detecting}. There are two types of crosstalk. The first one is quantum crosstalk, which is caused by the always-on-ZZ interaction~\cite{mundada2019suppression,zhao2020high}. The second one is classical crosstalk caused by the incorrect control of the qubits. The calibration data provided by IBM do not include the crosstalk error. To consider the crosstalk effect in partition algorithms, we must first characterize it in the hardware. There are several protocols presented in~\cite{gambetta2012characterization,bialczak2010quantum,proctor2019direct,erhard2019characterizing,huang2020alibaba} to benchmark the crosstalk effect in quantum devices. In this paper, we choose the mostly used protocol -- Simultaneous Randomized Benchmarking (SRB)~\cite{gambetta2012characterization} to detect and quantify the crosstalk between \texttt{CNOT} pairs when executing them in parallel.
We characterize the crosstalk effect followed by the optimization methods presented in~\cite{murali2020software}. On IBM quantum devices, the crosstalk effect is significant only at one hop distance between \texttt{CNOT} pairs~\cite{murali2020software}, such as ($CX_{0,1} | CX_{2,3}$) shown in Fig.~\ref{fig:one_hop}, when the control pulse of one qubit propagates an unwanted drive to the nearby qubits that have similar resonate frequencies. Therefore, we perform SRB only on \texttt{CNOT} pairs that are separated by one-hop distance. For those pairs whose distance is greater than one hop, the crosstalk effects are very weak and we ignore them. It allows us to parallelize SRB experiments of multiple \texttt{CNOT} pairs when they are separated by two or more hops. For example, in IBM Q 27 Toronto (ibmq\_toronto)~\cite{IBMQ27Toronto}, the pairs ($CX_{0,1} | CX_{4,7}$), ($CX_{12,15} | CX_{17,18}$), ($CX_{5,8} | CX_{11,14}$) can be characterized in parallel.
We perform the crosstalk characterization on IBM Q 27 Toronto twice. The results show that, although the absolute gate errors vary every day, the pairs that have strong crosstalk effect remain the same across days. SRB experiment on \texttt{CNOT} pairs ($g_i|g_j$) gives error rate $E(g_i|g_j)$ and $E(g_j|g_i)$. Here, $E(g_i|g_j)$ represents the \texttt{CNOT} error rate of $g_i$ when $g_i$ and $g_j$ are executed in parallel. If there is a crosstalk effect between the two pairs, it will lead to $E(g_i|g_j) > E(g_i)$ or $E(g_j|g_i) > E(g_j)$. The crosstalk effect characterization is expensive and time costly. Some of the pairs do not have crosstalk effect whereas the \texttt{CNOT} error of the pair affected the most by crosstalk effect is increased by more than five times. Therefore, we extract the pairs with significant crosstalk effect, i.e., $E(g_i|g_j) > 3 \times E(g_i)$ and only characterize these pairs when crosstalk properties are needed. We choose the same factor 3 to quantify the pairs with strong crosstalk error like~\cite{murali2020software}. The result of crosstalk effect characterization on IBM Q 27 Toronto is shown in Fig.~\ref{fig:crosstalk}.
\begin{figure}
\begin{subfigure}{\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.7]{one-hop.pdf}
\label{fig:one_hop}
\end{subfigure}
\begin{subfigure}{\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.7]{srb_toronto.pdf}
\label{fig:crosstalk}
\end{subfigure}
\caption{Characterization of crosstalk effect. (a) {\normalfont Crosstalk pairs separated by one-hop distance. The crosstalk pairs should be able to be executed at the same time. Therefore, they cannot share the same qubit. One-hop is the minimum distance between crosstalk pairs. }(b) {\normalfont Crosstalk effect results of IBM Q 27 Toronto using SRB. The arrow of the red dash line points to the \texttt{CNOT} pair that is affected significantly by crosstalk effect, e.g., $CX_{2,3}$ and $CX_{5,8}$ affect each other when they are executed simultaneously. In our experiments, $E(CX_{10,12}|CX_{15,18}) > 3 \times E(CX_{10,12})$, whereas $E(CX_{15,18}|CX_{10,12}) \approx 2.2 \times E(CX_{15,18})$. As we choose 3 as the factor to pick up pairs with strong crosstalk effect, there is no arrow at pair $CX_{15,18}$.}}
\label{fig:characterization}
\end{figure}
\vspace{2mm}
\noindent
\textit{Greedy sub-graph partition algorithm.}
\vspace{2mm}
\
\newline
We develop a Greedy Sub-graph Partition algorithm (GSP) for qubit partition process which is able to provide theoretically the optimal partitions for different quantum circuits (see Supplementary Note 3 for pseudo-code of GSP).
The first step of the GSP algorithm is to traverse the overall hardware to find all the possible partitions for a given circuit. For example, suppose we have a five-qubit circuit, we find all the subgraphs of the hardware topology (also called coupling graph) containing five qubits as the partition candidates. Each candidate has a score to represent its fidelity depending on the topology and calibration data. The partition with the best fidelity is selected and all the qubits inside of the partition are marked as used qubits so they cannot be assigned to other circuits. For the next circuit, a subgraph with the required number of qubits is assigned and we check if there is an overlap on this partition to partitions of previous circuits. If not, the subgraph is a partition candidate for the given circuit and the same process is applied to each subsequent circuit. To account for crosstalk, we check if any pairs in a subgraph have strong crosstalk effect caused by the allocated partitions of other circuits. If so, the score of the subgraph is adjusted to take crosstalk error into account.
In order to evaluate the reliability of a partition, there are two factors that need to be considered: partition topology and error rates of two-qubit links and readout error of each qubit. One-qubit gates are ignored for simplicity and because of their relatively low error rates compared to the other quantum operations. If there is a qubit pair in a partition that has strong crosstalk affected by other partitions, then \texttt{CNOT} error of this pair is added to the crosstalk effect. Note that the most recent calibration data should be retrieved through the IBM Quantum Experience before each usage to ensure that the algorithm has access to the most accurate and up-to-date information.
To evaluate the partition topology, we determine the longest shortest path (also called graph diameter) of the partition, denoted $L$. The smaller the longest shortest path is, the better the partition is connected and eventually fewer \texttt{SWAP} gates would be needed to make a connection between two qubits in a well-connected partition.
We devise a reliability score metric for a partition that is the sum of the graph diameter $L$, average \texttt{CNOT} error rate of the links times the number of \texttt{CNOTs} of the circuit, and the sum of the readout error rate of each qubit in a partition (Eq.~\ref{eq:1}). Note that the \texttt{CNOT} error rate includes the crosstalk effect if it exists.
\begin{equation}
Score_g = L + Avg_{CNOT} \times \#CNOTs + \sum_{Q_i \in P}R_{Q_i}
\label{eq:1}
\end{equation}
The graph diameter $L$ is always prioritized in this equation, since it is more than one order of magnitude larger than the other two factors. The partition with the smallest reliability score is selected. It is supposed to have the best connectivity and the lowest error rate. Moreover, the partition algorithm prioritizes the quantum circuit with a large density because the input circuits are ordered by their densities during the parallelism manager process. The partition algorithm is then called for each circuit in order. However, GSP algorithm is expensive and time costly. For small circuits, GSP algorithm gives the best choice of partition. It is also useful to use it as a baseline to compare with other partition algorithms. For beyond NISQ, a better approach should be explored to overcome the complexity overhead.
\vspace{2mm}
\noindent
\textit{Qubit fidelity degree-based heuristic sub-graph partition algorithm.}
\vspace{2mm}
\
\newline
The Qubit fidelity degree-based Heuristic Sub-graph Partition algorithm (QHSP) should perform as well as GSP but without the large runtime overhead.
In QHSP, when allocating partitions, we favor qubits with high fidelity. We define the fidelity degree of qubit based on the \texttt{CNOT} and readout fidelities of this qubit as in Eq.~\ref{eq:2}.
\begin{equation}
F\_Degree_{Q_i} = \sum_{Q_j \in N(Q_i)} \lambda \times (1 - E[Q_i][Q_j]) + (1 - R_{Q_i})
\label{eq:2}
\end{equation}
$Q_j$ are the neighbour qubits connected to $Q_i$, $E$ is the \texttt{CNOT} error matrix, and $R$ is the readout error rate. $\lambda$ is a user defined parameter to weight between the \texttt{CNOT} error rate and readout error rate. Such parameter is useful for two reasons: (1) Typically, in a quantum circuit, the number of \texttt{CNOT} operations is different from the number of measurement operations. Hence, the user can decide on $\lambda$ based on the relative number of operations. (2) For some qubits, the readout error rate is one or more orders of magnitude larger than the \texttt{CNOT} error rate. Thus, it is reasonable to add a weight parameter.
The fidelity degree metric reveals two aspects of the qubit. The first one is the connectivity of the qubit. The more neighbours a qubit has, the larger its fidelity degree is. The second one is the reliability of the qubit accounting \texttt{CNOT} and readout error rates. Thus, the metric allows us to select a reliable qubit with good connectivity. Instead of trying all the possible subgraph combinations (as in GSP algorithm), we propose a QHSP algorithm to build partitions that contain qubits with high fidelity degree while significantly reducing runtime.
To further improve the algorithm, we construct a list of qubits with good connectivity as starting points. We sort all physical qubits (qubits used in hardware) by their physical node degree, which is defined as the number of links in a physical qubit. Note that, the physical node degree is different from the fidelity degree. Similarly, we also obtain the largest logical node degree of the logical qubit (qubits used in the quantum circuit) by checking the number of different qubits that are connected to a qubit through \texttt{CNOT} operations. Next, we compare these two metrics.
If the largest physical node degree is less than the largest logical node degree, it means we cannot find a suitable physical qubit to map the logical qubit with the largest logical node degree that satisfies all the connections. In this case, we only collect the physical qubits with the largest physical node degree. Otherwise, the physical qubits whose physical node degree is greater than or equal to the largest logical node degree are collected as starting points. By limiting the starting points, this heuristic partition algorithm becomes even faster.
For each qubit in the starting points list, it explores its neighbours and finds the neighbour qubit with the highest fidelity degree calculated in Eq.~\ref{eq:2}, and merges it into the sub-partition. Then, the qubit inside of the sub-partition with the highest fidelity degree explores its neighbour qubits and merges the best one. The process is repeated until the number of qubits inside of the sub-partition is equal to the number of qubits needed. This sub-partition is considered as a subgraph and is added to the partition candidates (see Supplementary Note 3 for pseudo-code of QHSP).
After obtaining all the partition candidates, we compute the fidelity score for each of them. As we start from a qubit with high physical node degree and merge to neighbour qubits with high fidelity degree, the constructed partition is supposed to be well-connected, hence, we do not need to check the connectivity of the partition using the longest shortest path $L$ as in Eq.~\ref{eq:1}, GSP algorithm. We can only compare the error rates. The fidelity score metric is simplified by only calculating the \texttt{CNOT} and readout error rates as in Eq.~\ref{eq:3}. It is calculated for each partition candidate and the best one is selected. See supplementary note 3 for an example of explaining QHSP in detail.
\begin{equation}
Score_h = Avg_{CNOT} \times \#CNOTs + \sum_{Q_i \in P}R_{Q_i}
\label{eq:3}
\end{equation}
\vspace{2mm}
\noindent
\textit{Runtime analysis}
\vspace{2mm}
\
\newline
Let $n$ be the number of hardware qubits, $k$ the number of circuit qubits to be allocated in a partition, $g$ the number of gates that the circuit has.
For GSP algorithm, in most cases, the number of circuit qubits is less than the number of hardware qubits, thus the time cost is $O(k^3n^k)$. It increases exponentially as the number of circuit qubits augments. QHSP algorithm starts by collecting a list of $m$ starting points where $m \leq n$. It takes $O(mk^2 + nlog(n) + g)$, which is polynomial. For the detailed explanation of runtime analysis, see Supplementary Note 3.
\subsubsection{Post qubit partition}
\
\newline
By default multi-programming mechanism reduces circuit fidelity compared to standalone circuit execution mode. If the fidelity reduction is significant, circuits should be executed independently or the number of simultaneous circuits should be reduced even though the hardware throughput can be decreased as well. Therefore, we consistently check the circuit fidelity difference between independent versus concurrent execution.
We start with qubit partition process for each circuit independently and obtain the fidelity score of the partition. Next, this qubit partition process is applied to these circuits to compute the fidelity score when executing them simultaneously. The difference between the fidelity scores is denoted $\Delta S$, which is the fidelity metric. If $\Delta S$ is less than a specific threshold $\delta$, it means simultaneous circuit execution does not detriment significantly the fidelity score, thus circuits can be executed concurrently, otherwise, independently or reduce the number of simultaneous circuits. The fidelity metric along with the parallelism manager help to define the optimal number of simultaneous circuits to be executed.
\subsection{Scheduler}
\subsubsection{Mapping transition algorithm}
\
\newline
The circuits need to be transformed to be executable on real quantum hardware, which includes two steps: initial mapping and mapping transition. The initial mapping of each circuit is created while taking into account swap error rate and swap distance to perform qubit movement operations \cite{niu2020hardware}. The initial mapping of the simultaneous mapping transition process is obtained by merging the initial mapping of each circuit according to its partition. We further improve the mapping transition algorithm \cite{niu2020hardware} by modifying the heuristic cost function to better select the inserted gate. We also introduce the \texttt{Bridge} gate to the simultaneous mapping transition process for multi-programming.
First, each quantum circuit is transformed into a more convenient format -- Directed Acyclic Graph (DAG) circuit which represents the operation dependencies of the circuit without considering the connectivity constraints. Then, the compiler traverses the DAG circuit and goes through each quantum gate sequentially. The gate that does not depend on other gates (i.e., all the gates before it have been executed) is allocated to the first layer, denoted $F$. The compiler checks if the gates on the first layer are hardware-compliant. The hardware-compliant gates can be executed on the hardware directly without modification. They are added to the scheduler, removed from the first layer and marked as executed. If the first layer is not empty, which means some gates are non-executable on hardware, a \texttt{SWAP} or \texttt{Bridge} gate is needed. We collect all the possible \texttt{SWAPs} and \texttt{Bridges}, and use the cost function $H$ (see Eq.~\ref{eq:5}) to find the best candidate. The process is repeated until all the gates are marked as executed (see Supplementary Note 4 for pseudo-code of simultaneous mapping transition algorithm).
A \texttt{SWAP} gate requires three \texttt{CNOTs} and inserting a \texttt{SWAP} gate can change the current mapping. A \texttt{Bridge} gate requires four \texttt{CNOTs} and inserting a \texttt{Bridge} gate does not change the current mapping and it can only be used to execute a \texttt{CNOT} when the distance between the control qubit and the target qubit is exactly two. Both gates need three supplementary \texttt{CNOTs}. The \texttt{SWAP} gate is preferred when it has a positive impact on the following gates, allocated in the extended layer $E$, hence it makes these gates executable or reduces the distance between control and target quits. Otherwise, a \texttt{Bridge} gate is preferred.
A cost function $H$ is introduced to evaluate the cost of inserting a \texttt{SWAP} or \texttt{Bridge}. We use the following distance matrix (Eq.~\ref{eq:4}) as in~\cite{niu2020hardware} to quantify the impact of the \texttt{SWAP} or \texttt{Bridge} gate,
\begin{equation}
D = \alpha_1 \times S + \alpha_2 \times \mathcal{E}
\label{eq:4}
\end{equation}
where $S$ is the swap distance matrix and $\mathcal{E}$ is the swap error matrix. We set $\alpha_1$ and $\alpha_2$ to 0.5 to equally consider the swap distance and swap error rate. In~\cite{niu2020hardware}, only the impact of a \texttt{SWAP} and \texttt{Bridge} on other gates (first and extended layer) was considered without considering their impact on the gate itself. As each of them is composed of either three or four \texttt{CNOTs}, their impact cannot be ignored. Hence, in our multi-programming mapping transition algorithm, we take self impact into account and create a list of both \texttt{SWAP} and \texttt{Bridge} candidates, labeled as "tentative gates" and the heuristic cost function is as:
\begin{multline}
H = \frac{1}{\vert F + N_{Tent} \vert}(\sum_{g \in F} D[\pi(g.q_1)][\pi(g.q_2)] \\ + \sum_{g \in Tent}D[\pi(g.q_1)][\pi(g.q_2)])
+ W \times \frac{1}{\vert E \vert}\sum_{g \in E}D[\pi(g.q_1)][\pi(g.q_2)]
\label{eq:5}
\end{multline}
where $W$ is the parameter that weights the impact of the extended layer, $N_{Tent}$ is the number of gates of the tentative gate, $Tent$ represents a \texttt{SWAP} or \texttt{Bridge} gate, and $\pi$ represents the mapping. \texttt{SWAP} gate has three \texttt{CNOTs}, thus $N_{Tent}$ is three and we consider the impact of three \texttt{CNOTs} on the first layer. The mapping is the new mapping after inserting a \texttt{SWAP}. For \texttt{Bridge} gate, $N_{Tent}$ is four and we consider four \texttt{CNOTs} on the first layer, and the mapping is the current mapping as \texttt{Bridge} gate does not change the current mapping. We weight the impact on the extended layer to prioritize the first layer. This cost function can help the compiler select the best gate to insert between a \texttt{SWAP} and \texttt{Bridge} gate.
\subsection{Application: simultaneous executions of multiple circuits of different size}
\subsubsection{Experimental results}
\
\newline
\begin{figure}[h]
\begin{subfigure}{0.45\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.4]{result_toronto_fidelity_line_chart.pdf}
\label{fig:toronto_fidelity}
\end{subfigure}
\hspace{5mm}
\begin{subfigure}{0.45\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.4]{result_toronto_gate_line_chart.pdf}
\label{fig:toronto_gate}
\end{subfigure}
\caption{Comparison of fidelity and number of additional gates on IBM Q 27 Toronto when executing two circuits simultaneously. (a) {\normalfont Fidelity}. (b) {\normalfont Number of additional gates.}}
\label{fig:result_toronto_figure}
\end{figure}
\begin{figure}[h]
\begin{subfigure}{0.45\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.4]{result_mahattan_fidelity3_line_chart.pdf}
\label{fig:manhattan_fidelity_3}
\end{subfigure}
\hspace{5mm}
\begin{subfigure}{0.45\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.4]{result_mahattan_gate3_line_chart.pdf}
\label{fig:manhattan_gate_3}
\end{subfigure}
\caption{Comparison of fidelity and number of additional gates on IBM Q 65 Manhattan when executing three circuits simultaneously. (a) {\normalfont Fidelity}. (b) {\normalfont Number of additional gates.}}
\label{fig:manhattan_result_3}
\end{figure}
\begin{figure}[h]
\begin{subfigure}{0.45\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.4]{result_mahattan_fidelity4_line_chart.pdf}
\label{fig:manhattan_fidelity_4}
\end{subfigure}
\hspace{5mm}
\begin{subfigure}{0.45\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.4]{result_mahattan_gate4_line_chart.pdf}
\label{fig:manhattan_gate_4}
\end{subfigure}
\caption{Comparison of fidelity and number of additional gates on IBM Q 65 Manhattan when executing four circuits simultaneously. (a) {\normalfont Fidelity}. (b) {\normalfont Number of additional gates.}}
\label{fig:manhattan_result_4}
\end{figure}
We first evaluated our multi-programming approach by executing a list of different-size benchmarks at the same time on two quantum devices, IBM Q 27 Toronto and IBM Q 65 Manhattan (ibmq\_manhattan)~\cite{IBMQ65Manhattan} (see Supplementary Note 1 for further information about the selected quantum hardware). All the benchmarks are collected from the previous work~\cite{zulehner2018efficient}, including several functions taken from RevLib~\cite{WGT+:2008} as well as some quantum algorithms written in Quipper~\cite{green2013quipper} or Scaffold~\cite{abhari2012scaffold}. These benchmarks are widely used in the quantum community and their details are shown in Table~\ref{fig:benchmarks}. We chose small quantum circuits with shallow-depth since only small circuits can obtain reliable results when executed on real quantum hardware. The metrics we used to evaluate our algorithm include Probability of a Successful Trial (PST), number of additional CNOT gates, and Trial Reduction Factor (TRF), see Method for detailed explanation.
\begin{table}[b]
\centering
\includegraphics[scale=0.8]{Benchmarks.pdf}
\caption{Information of benchmarks}
\label{fig:benchmarks}
\end{table}
Several published qubit mapping algorithms~\cite{li2019tackling,wille2019mapping,murali2019noise,zhu2020dynamic,niu2020hardware,guerreschi2018two,itoko2020optimization} and multi-programming mapping algorithms are available as discussed in section~\ref{sec:introduction}. HA~\cite{niu2020hardware} seems to be the best qubit mapping algorithm in terms of the number of additional gates and circuit fidelity. We use HA as the baseline for independent executions of multiple circuits. CDAP algorithm proposed in~\cite{dou2020new} seems to be the best multi-programming mapping algorithm and is considered as the baseline for concurrent executions of multiple circuits.
To summarize, we compare our multi-programming algorithms, 1) GSP + improved mapping transition (labeled as GSP) and 2) QHSP + improved mapping transition (labeled as QHSP), with the baseline CDAP. The loss of fidelity due to simultaneous executions of multiple circuits is reported by comparing concurrent versus independent executions. Moreover, we compare the partition + improved mapping transition algorithm based on HA (labeled as PHA) versus HA on independent executions to show the impact of partition in large quantum hardware for a small circuit. The details of the configuration of algorithms are presented in Methods.
We first ran two quantum circuits on IBM Q 27 Toronto simultaneously. Results on output state fidelity and the number of additional gates are shown in Fig.~\ref{fig:result_toronto_figure}.
For independent executions, the fidelity is improved by 46.8\% and the number of additional gates is reduced by 8.7\% comparing PHA to HA. For simultaneous executions, QHSP and GSP allocate the same partitions except for the first experiment -- (ID1, ID1). In this experiment, GSP improves the fidelity by 6\% compared to QHSP. Partition results might be different due to the various calibration data and the choice of $\lambda$, but the difference of the partition fidelity score between the two algorithms is small. The results show that QHSP is able to allocate nearly optimal partitions while reducing runtime significantly. Therefore, for the rest experiments, we only evaluate QHSP algorithm. QHSP can improve the fidelity by 28.9\% and reduce the additional gate number by 52.3\% compared to CDAP. Comparing simultaneous (QHSP) versus independent (PHA) executions for two circuits, fidelity decreases by 5.8\% and the number of additional gates is almost the same. During the post-partition process, $\Delta S$ does not pass the threshold and TRF is two.
Next, we executed on IBM 65 Manhattan three and four simultaneous quantum circuits. Fig.~\ref{fig:manhattan_result_3} and Fig.~\ref{fig:manhattan_result_4} show the comparison of fidelity and the number of additional gates. PHA always outperforms HA for independent executions. QSHP significantly outperforms CDAP with the number of simultaneous circuits increasing. The output fidelity is increased by 74.8\% and 55.3\% on average for the two cases. The reduction of inserted gate number is always more than 50\%. The threshold is still not passed and TRF becomes three and four. Moreover, fidelities decrease by 1.5\% and 6.7\% when comparing simultaneous (QHSP) versus independent (PHA) executions.
Finally, to evaluate the hardware limitations of executing multiple circuits in parallel, we set the threshold $\delta$ to 0.2. All the five benchmarks are able to be executed simultaneously on IBM Q 65 Manhattan. Partition fidelity difference is 0.18. Results show that fidelity of simultaneous executions (QHSP) is decreased by 9.5\% compared to independent executions (PHA). Both fidelity and additional gate number improvement of QHSP are more than 50\% compared to CDAP. The complete experimental results can be found in Supplementary Note 5.
\subsubsection{Result analysis}
\
\newline
For independent executions, algorithm PHA is always better than HA due to two reasons: (1) The initial mapping of the two algorithms is based on a random process. During the experiment, we perform the initial mapping generation process ten times and select the best one. However, for PHA, we first limit the random process into a reliable and well-connected small partition space rather than the overall hardware space used by HA. Therefore, with only ten trials, PHA finds a better initial mapping. (2) We improve the mapping transition process of PHA, which can make a better selection between \texttt{SWAP} and \texttt{Bridge} gate. HA is shown to be sufficient for hardware with a small number of qubits for example a 5-qubit quantum chip. If we want to map a circuit on large hardware, it is better to first limit the search space into a reliable small partition and then find the initial mapping. This qubit partition approach can be applied to general qubit mapping problem for search space limitation when large hardware is selected to map.
For simultaneous executions, QHSP performs better than CDAP because of the following reasons: (1) CDAP constructs a hierarchy tree according to the modularity-based FN community detection algorithm~\cite{newman2004fast}. The tree is constructed by calculating the modularity of the overall hardware coupling graph. However, when allocating a partition to a circuit, we focus on the topology and calibration data inside of the partition, rather than the whole hardware. As the number of partitions to allocate increases, the performance of CDAP becomes worse. (2) CDAP only considers the \texttt{SWAP} gate to realize the connection ignoring the \texttt{Bridge} gate, which can significantly reduce the number of additional gates. (3) CDAP does not consider the crosstalk effect. Although the X-SWAP scheme used in CDAP can slightly reduce the number of additional gates, it only works when the allocated partitions are close to each other, which will increase the crosstalk effect. However, QHSP takes the partition topology, error rate, and crosstalk effect into consideration and can provide better partitions. QHSP uses almost the same number of additional gates whereas fidelity is decreased less than 10\% compared to PHA if the threshold is set to 0.1.
\subsection{Application: Estimate the ground state energy of deuteron}
In order to demonstrate the potential interest to apply the multi-programming mechanism to existing quantum algorithms, we investigated it on VQE algorithm.
To do this, we performed the same experiment as~\cite{gokhale2020optimization,dumitrescu2018cloud} on IBM Q 65 Manhattan, estimating the ground state energy of deuteron, which is the nucleus of a deuterium atom, an isotope of hydrogen.
Deuteron can be modeled using a 2-qubit Hamiltonian spanning four Pauli strings: $ZI, IZ, XX,$ and $YY$,~\cite{gokhale2020optimization,dumitrescu2018cloud}.
If we use the naive measurement to calculate the state energy, one ansatz corresponds to four different measurements. Pauli operator grouping has been proposed to reduce this overhead by utilizing simultaneous measurement~\cite{gokhale2020optimization,kandala2017hardware,crawford2019efficient}. For example, the Pauli strings can be partitioned into two commuting families: \{$ZI, IZ$\} and \{$XX,YY$\} using the approach proposed in~\cite{gokhale2020optimization}. It allows one parameterized ansatz to be measured twice instead of four measurements in naive method.
We used a simplified Unitary Coupled Cluster ansatz with a single parameter and three gates, as described in~\cite{gokhale2020optimization,dumitrescu2018cloud}. The algorithm configuration of this experiment is explained in Methods. We applied our multi-programming method on the top of the Pauli operator grouping approach (labeled as PG)~\cite{gokhale2020optimization}. We performed this experiment twice across different days. For the first experiment, the parallelism manager worked with the hardware-aware multi-programming compiler to finally select ten circuits for simultaneous execution without passing the fidelity threshold. It corresponds to perform five optimisations (five different parameterized circuits) at the same time (one parameterized circuit needs two measurements). The selected ten circuits were passed to the scheduler to be executed in parallel. The required circuit number is reduced by ten times compared to PG. Note that, if we use the naive measurement, the number of circuits needed will be reduced by a factor of 20. The result is shown in Fig.~\ref{fig:deuteron1}. The error rate is quite high for the two executions, 29.7\% for PG and 64.4\% for multi-programming + PG. The result of the second experiment is shown in Fig.~\ref{fig:deuteron2}. In this case, four optimisations (eight circuits) were selected to be executed at the same time with respect to the fidelity threshold. The error rate is 9.3\% and 7\% for the two methods. Applying multi-programming can even improve the output fidelity. The huge fidelity difference is due to the different calibration data of the device which are the input of our multi-programming approach. The complete result of the two experiments including hardware throughput is shown in Fig.~\ref{fig:deuteron_result}.
\begin{figure}[h]
\begin{subfigure}{0.45\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.23]{Deuteron_results_100.pdf}
\label{fig:deuteron1}
\end{subfigure}
\hspace{5mm}
\begin{subfigure}{0.45\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.23]{Deuteron_results_100_2.pdf}
\label{fig:deuteron2}
\end{subfigure}
\begin{subfigure}{\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.6]{deuteron.pdf}
\label{fig:deuteron_result}
\end{subfigure}
\caption{The estimation of the ground state energy of deuteron under PG and muti-programming + PG. (a) {\normalfont Five optimisations with ten measurements.} (b) {\normalfont Four optimisations with eight measurements.} (c) {\normalfont The complete result of the two experiments. $n_c$ is the number of simultaneous circuit number.}}
\label{fig:deuteron}
\end{figure}
\section{Discussion}
In this article, we presented a multi-programming approach that allows to execute multiple circuits on a quantum chip simultaneously without losing fidelity. We introduced the parallelism manager and fidelity metric to select optimally the number of circuits to be executed at the same time. Moreover, we proposed a hardware-aware multi-programming compiler which contains two qubit partition algorithms taking hardware topology, calibration data, and crosstalk effect into account to allocate reliable partitions to different quantum circuits. We also demonstrated an improved simultaneous mapping transition algorithm which helps to transpile the circuits on quantum hardware with a reduced number of inserted gates.
We first executed a list of circuits of different sizes simultaneously and compared our algorithm with the state-of-the-art multi-programming approach. Experimental results showed that our approach can outperform the state of the art in terms of both output fidelity and the number of additional gates. Then, we investigated our multi-programming approach on VQE algorithm to estimate the ground state energy of deuteron, showing the added value of applying our approach to existing quantum algorithms. The multi-programming approach is evaluated on IBM hardware, but it is general enough to be adapted to other quantum hardware.
Based on the experimental result, we found that the main concern with multi-programming mechanism is a trade-off between output fidelity and the hardware throughput. For example, how one can decide which programs to execute simultaneously and how many of them to execute without losing fidelity. Here, we list several guidelines to help the user to utilize our multi-programming approach.
\begin{itemize}
\item Check the target hardware topology and calibration data. The multi-programming mechanism is more suitable for a relatively large quantum chip compared to the quantum circuit and with low error rate.
\item Choose appropriate fidelity threshold for post qubit partition process. A high threshold can improve the hardware throughput but lead to the reduction of output fidelity. It should be set carefully depending on the size of the benchmark. For benchmarks of small size that we used in experiments, it is reasonable to set the threshold to 0.1.
\item The number of circuits that can be executed simultaneously will mainly depend on the fidelity threshold and the calibration data of the hardware.
\item QHSP algorithm is suggested for the partition process due to efficiency and GSP is recommended to evaluate the quality of the partition algorithm. Using both algorithms, one can explore which circuits can be executed simultaneously and how many of them within the given fidelity threshold.
\end{itemize}
Quantum hardware development with more and more qubits will enable execution of multiple quantum programs simultaneously and possibly a linchpin for quantum algorithms requiring parallel sub-problem executions. Variational Quantum Algorithm is becoming a leading strategy to demonstrate quantum advantages for practical applications. In such algorithms, the preparation of parameterized quantum state and the measurement of expectation value are realized on shallow circuits~\cite{zhang2020shallow}. Taking VQE as an example, the Hamiltonian can be decomposed into several Pauli operators and simultaneous measurement by grouping Pauli operators have been proposed in~\cite{gokhale2020optimization,kandala2017hardware,crawford2019efficient} to reduce the overhead of the algorithm. Based on our experiment, we have shown that the overhead of VQE can be further improved by executing several sets of Pauli operators at the same time using multi-programming mechanism.
For future work, we would like to apply our multi-programming algorithm to other variational quantum algorithms such as VQLS or VQC to enable the preparation of states in parallel and to reduce the overhead of these algorithms. Moreover, in our qubit partition algorithms, we take the crosstalk effects into consideration by characterizing them and adding them to the fidelity score of the partition, which is able to avoid the crosstalk error in a high level. There are some other approaches of eliminating the crosstalk error in a cheaper way instead of performing SRB protocol, for example using commutativity rules to reorder the simultaneous gate operations~\cite{murali2020software,itoko2020optimization}. However, these methods have some challenges such as trading off between crosstalk and decoherence. More interesting tricks for crosstalk mitigation need to be targeted for simultaneous executions. In addition, not all the benchmarks have the same circuit depth.
Taking the time-dependency into consideration, choosing the optimal combination of circuits of different depth to run simultaneously can also be the focus of future work.
\section{Methods}
\subsection{Metrics}
Here are the detailed explanations of the metrics that we use to evaluate our algorithm.
\begin{enumerate}
\item Probability of a Successful Trial (PST)~\cite{tannu2019not}. This metric is defined by the number of trials that give the expected result divided by the total number of trials. The expected result is obtained by executing the quantum circuit on the simulator. To have a precise estimation of the PST, we execute each quantum circuit on the quantum hardware for a large number of trials (8192).
\item Number of additional \texttt{CNOT} gates. This metric is related to the number of \texttt{SWAP} or \texttt{Bridge} gates inserted. This metric can show the ability of the algorithm to reduce the number of additional gates.
\item Trial Reduction Factor (TRF). This metric is introduced in~\cite{das2019case} to evaluate the improvement of the throughput thanks to the multi-programming mechanism. It is defined as the ratio of trials needed when quantum circuits are executed independently to the trials that when they are executed simultaneously.
\end{enumerate}
\subsection{Algorithm configurations}
Here, we consider the algorithm configurations of different multi-programming and standalone mapping approaches. We select the best initial mapping out of ten attempts for HA, PHA, GSP, and QHSP. Weight parameter $W$ in the cost function (Eq.~\ref{eq:5}) is set to 0.5 and the size of the extended layer is set to 20. Parameters $\alpha_1$ and $\alpha_2$ are set to 0.5 respectively to consider equally the swap distance and swap error rate. For the experiments of multiple different-size circuits, the weight parameter $\lambda$ of QHSP (Eq.~\ref{eq:2}) is set to $2$ because of the relatively large number of \texttt{CNOT} gates in benchmarks, whereas for deuteron experiment, $\lambda$ is set to $1$ because of the small number of \texttt{CNOTs} of the parameterized circuit. The threshold $\delta$ for post qubit partition is set to 0.1 to ensure the multi-programming fidelity. Due to the expensive cost of SRB, we perform SRB only on IBM Q 27 Toronto and collect the pairs with significant crosstalk effect. Only the collected pairs are characterized and their crosstalk properties are provided to the partition process. The experimental results on IBM Q 65 Manhattan do not consider the crosstalk effect. For each algorithm, we only evaluate the mapping transition process, which means no optimisation methods like gate commutation or cancellation are applied.
The algorithm is implemented in Python and evaluated on a PC with 1 Intel i5-5300U CPU and 8 GB memory. Operating System is Ubuntu 18.04. All the experiments were performed on the IBM quantum information science kit (qiskit)~\cite{Qiskit} and the version used is 0.21.0.
\section{Data Availability}
The source code of the algorithms used in this paper is available on the Github repository~\cite{Github}.
\section{Supplementary Information}
\subsection{Supplementary note - Hardware information}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{ibmq_toronto.pdf}
Single qubit error rate: $2.168\mathrm{e}{-4}$ to $1.455\mathrm{e}{-3}$\\
CNOT error rate: $5.697\mathrm{e}{-3}$ to $2.735\mathrm{e}{-2}$\\
Readout error rate: $7.3\mathrm{e}{-3}$ to $2.317\mathrm{e}{-1}$\\
\caption{IBM Q 27 Toronto topology and error rates.}
\label{fig:ibmq_toronto}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{ibmq_manhattan.pdf}
Single qubit error rate: $2.561\mathrm{e}{-4}$ to $5.407\mathrm{e}{-3}$\\
CNOT error rate: $7.619\mathrm{e}{-3}$ to $4.638\mathrm{e}{-2}$\\
Readout error rate: $9.2\mathrm{e}{-3}$ to $1.118\mathrm{e}{-1}$\\
\caption{IBM Q 65 Manhattan topology and calibration data.}
\label{fig:ibmq_manhattan}
\end{figure}
Noise can cause several errors during the execution process such as (1) coherence errors due to the fragile nature of qubits. The qubit can only maintain information for a limited amount of time. (2) Operational errors including gate errors and measurement errors (readout errors). (3) Crosstalk errors that violate the isolated qubit state due to operations on other qubits.
Supplementary Fig.~\ref{fig:ibmq_toronto} shows the hardware topology and the calibration data of IBM Q 27 Toronto. We list the calibration data of single-qubit error rate, \texttt{CNOT} error rate, and readout error rate. Note that these errors are not constant and change at each re-calibration of the chip, and IBM does not provide the statistics of crosstalk error. The other device that we choose to evaluate our algorithm is IBM Q 65 Manhattan. Its topology and calibration data are shown in Supplementary Fig.~\ref{fig:ibmq_manhattan}. \texttt{CNOT} error rate is one order of magnitude higher than their one-qubit counterparts. Moreover, the readout error rate is of the same order of magnitude or higher than \texttt{CNOT} error rate.
In this paper, we only focus on \texttt{CNOT} error rate and readout error rate because of the relatively low error rates of one-qubit gates.
It is important to note that all the interconnects between qubits as well as the reliability of qubit are not equal with respect to \texttt{CNOT} error rate and readout error rate. Taking IBM Q 27 Toronto as an example, the best \texttt{CNOT} gate has an error rate of 4.8 times lower than the worst \texttt{CNOT}, and the most reliable qubit has a readout error rate of 31.7 times lower than the worst qubit. Therefore, each qubit cannot be treated equally, and we need to consider the error difference between the links and qubits.
In this article, we mainly focus on IBM architectures. But the proposed methods are general enough to be applied to any other quantum chips that use the quantum-gate model of computation, such as Google's Sycamore~\cite{google-quantum-supremacy} or Rigetti's Aspen-8.
\subsection{Supplementary note - Motivational Example}
\begin{figure}
\begin{subfigure}{\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.8]{motivition1.pdf}
\label{fig:multi_1}
\end{subfigure}\\
\begin{subfigure}{\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.8]{motivition2.pdf}
\label{fig:multi_2}
\end{subfigure}\\
\begin{subfigure}{\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.8]{motivition3.pdf}
\label{fig:multi_3}
\end{subfigure}
\caption{A motivational example of qubit partition problem (error rate in \%). (a) {\normalfont Partition without considering operational error.} (b) {\normalfont Partition considering operational error without considering crosstalk effect.} (c) {\normalfont Partition considering both operational error and crosstalk effect.}}
\label{fig:motivation}
\end{figure}
To motivate the qubit partition problem, we execute two small circuits QC1 and QC2 simultaneously on IBM Q 27 Toronto with different partitions (Supplementary Fig.~\ref{fig:motivation}). \texttt{CNOT} error rate of each link is shown in the figure and the unreliable links and qubits with high readout error rates are highlighted in red. Both circuits have five qubits with a different number of gates as listed in Supplementary Fig.~\ref{fig:motivation_result}.
There are two constraints to be considered when executing multiple circuits concurrently. First, each circuit should be allocated to a partition containing reliable physical qubits. Allocated physical qubits can not be shared among quantum circuits. Second, qubits can be moved only inside of their circuit partition, in other words, qubits can be swapped within the same partition only. Thus, finding reliable partitions for multiple circuits is an important step in the multi-programming mapping problem.
We compare three partitions with the same topology to show the impact of different error sources on the output fidelity: (1) Partition P1 without considering the operational error (Supplementary Fig.~\ref{fig:multi_1}). (2) Partition P2 only considering operational error without the crosstalk effect (Supplementary Fig.~\ref{fig:multi_2}). (3) Partition P3 considering both operational error and crosstalk effect (Supplementary Fig.~\ref{fig:multi_3}). Note that the operational error includes \texttt{CNOT} error and readout error. For illustration, we fix the partition of QC2 to $\{5, 8, 9, 11, 14\}$ and only change the partition of QC1. It is important to note that if we have different topologies, the fidelity of the circuit will be different as well because the number of additional gates is strongly related to the hardware topology.
Results in Supplementary Fig.~\ref{fig:motivation_result}b show that Partition P1 has the lowest fidelity. Partition P2 considers operational error and selects $\{0, 1, 2, 3, 4\}$ with reliable qubits and links. However, it does not consider the crosstalk effect. Since $Q_3$ is the neighour of $Q_5$, when $CX_{2,3}$ and $CX_{5,8}$ are executed at the same time, they can affect each other and violate the qubit state. Partition P3 includes $\{1, 4, 6, 7, 10\}$ and considers both operational error and crosstalk effect. P3 does not have the crosstalk effect and is slightly better than P2 in terms of the operational error, however, the output fidelity of QC1 is increased by $61.6\%$.
\begin{figure}[t]
\begin{subfigure}{\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.8]{motivation_table.pdf}
\label{fig:motivation_table}
\end{subfigure}\\
\begin{subfigure}{\columnwidth}
\centering
\caption{}
\includegraphics[scale=0.6]{motivation_result.pdf}
\label{fig:motivation_fidelity}
\end{subfigure}
\caption{Results of the motivational example. (a) {\normalfont Circuit information and output fidelity results of different partitions. n: qubit number. g: gate number of the circuit.} (b) {\normalfont Output fidelity results of different partitions.}}
\label{fig:motivation_result}
\end{figure}
\subsection{Supplementary note - Qubit Partition}
In this note, we first demonstrate the pseudo-code of GSP algorithm. Then, we show an example of QHSP algorithm and its pseudo-code. Finally, we explain the runtime analysis of the two algorithms in detail.
\subsubsection{Greedy sub-graph partition algorithm}
\
\newline
The pseudo-code of GSP is shown in Algorithm~\ref{algo1}.
\begin{algorithm}[!htbp]%
\SetAlgoLined
\caption{GSP algorithm} \label{algo1}
\SetKwInOut{Input}{input}
\SetKwInOut{Output}{output}
\SetKwFunction{combinations}{combinations}
\SetKwFunction{subgraph}{subgraph}
\SetKwFunction{partitionScore}{Set\_Partition\_Score}
\SetKwFunction{Append}{append}
\SetKwFunction{findCrosstalk}{Find\_Crosstalk\_pairs}
\SetKwData{CouplingGraph}{$G$}
\SetKwData{CalibrationData}{$C$}
\SetKwData{QubitsUsed}{$q_{\text{used}}$}
\SetKwData{QuantumCircuit}{$QC$}
\SetKwData{SubGraphList}{sub\_graph\_list}
\SetKwData{crosstalkProperties}{crosstalk\_props}
\SetKwData{SubGraph}{sub\_graph}
\SetKwData{QubitNum}{qubit\_num}
\SetKwData{GSub}{G\_sub}
\SetKwData{crosstalkPairs}{crosstalk\_pairs}
\Input{Quantum circuit \QuantumCircuit
, Coupling graph \CouplingGraph, Calibration data \CalibrationData,
Crosstalk properties \crosstalkProperties,
Used\_qubits \QubitsUsed}
\Output{A list of candidate partitions \SubGraphList}
\BlankLine
\Begin{
\QubitNum $\leftarrow$ \QuantumCircuit.qubit\_num\;
Set \SubGraphList to empty list\;
\For {\SubGraph $\in$ \combinations(\CouplingGraph, \QubitNum)}{
\If{\SubGraph is connected}{
\If {\QubitsUsed is empty}{
\SubGraph.\partitionScore(\CouplingGraph, \CalibrationData, \QuantumCircuit)\;
\SubGraphList.\Append(\SubGraph)\;
}
{
\If {no qubit in \SubGraph is in \QubitsUsed}{
\crosstalkPairs $\leftarrow$ \findCrosstalk(\SubGraph, \crosstalkProperties, \QubitsUsed)\; \SubGraph.\partitionScore(\CouplingGraph, \CalibrationData, \QuantumCircuit, \crosstalkPairs)\;
\SubGraphList.\Append(\SubGraph)\;
}
}
}
}
\Return \SubGraphList \;
}
\end{algorithm}
\subsubsection{Qubit fidelity degree-based heuristic sub-graph partition algorithm}
\
\newline
Supplementary Fig.~\ref{fig:partition_example} shows an example of applying QHSP on IBM Q 5 Valencia (5-qubit ibmq\_valencia)~\cite{IBMQ5Valencia} for a four-qubit circuit. The calibration data of IBM Q 5 Valencia, including readout error rate and \texttt{CNOT} error rate is shown in Supplementary Fig.~\ref{fig:ibmq_valencia_info}. The fidelity degree of qubit calculated by Eq.~\ref{eq:2} is shown in Supplementary Fig.~\ref{fig:qubit degree}. Here, we consider a circuit of medium size and set $\lambda$ to two. Suppose the largest logical degree is three. Therefore, $Q_1$ is selected as the starting point since it is the only physical qubit that has the same physical node degree as the largest logical degree. It has three neighbour qubits: $Q_0$, $Q_2$, and $Q_3$. $Q_3$ is merged into the sub-partition because it has the highest fidelity degree among neighbour qubits. The sub-partition becomes $\{Q_1, Q_3\}$. As the fidelity degree of $Q_1$ is larger than $Q_3$, the algorithm will select again the left neighbour qubit with the largest fidelity degree of $Q_1$, which is $Q_0$. The sub-partition becomes $\{Q_1, Q_3, Q_0\}$. $Q_1$ is still the qubit with the largest fidelity degree in the current sub-partition, its neighbour qubit -- $Q_2$ is merged. The final sub-partition is $\{Q_1,Q_3,Q_0,Q_2\}$ and it can be considered as a partition candidate. The merging process is shown in Supplementary Fig.~\ref{fig:Qubit partition}.
The pseudo-code of QHSP is shown in Algorithm~\ref{algo2}.
\begin{algorithm}%
\SetAlgoLined
\caption{QHSP algorithm} \label{algo2}
\SetKwInOut{Input}{input}
\SetKwInOut{Output}{output}
\SetKwFunction{findBestQubit}{find\_best\_qubit}
\SetKwFunction{partitionScore}{Set\_Partition\_Error}
\SetKwFunction{Append}{append}
\SetKwFunction{Range}{range}
\SetKwFunction{findCrosstalk}{Find\_Crosstalk\_pairs}
\SetKwFunction{Len}{len}
\SetKwData{CouplingGraph}{$G$}
\SetKwData{CalibrationData}{$C$}
\SetKwData{QubitsUsed}{$q_{\text{used}}$}
\SetKwData{QuantumCircuit}{$QC$}
\SetKwData{SubGraphList}{sub\_graph\_list}
\SetKwData{crosstalkProperties}{crosstalk\_props}
\SetKwData{SubGraph}{sub\_graph}
\SetKwData{CircQubitNum}{circ\_qubit\_num}
\SetKwData{QubitNum}{qubit\_num}
\SetKwData{GSub}{G\_sub}
\SetKwData{BestQubit}{best\_qubit}
\SetKwData{crosstalkPairs}{crosstalk\_pairs}
\SetKwData{Starting}{starting\_points}
\Input{Quantum circuit \QuantumCircuit
, Coupling graph \CouplingGraph, Calibration data \CalibrationData,
Crosstalk properties \crosstalkProperties,
Used\_qubits \QubitsUsed,
Starting points \Starting}
\Output{A list of candidate partitions \SubGraphList}
\BlankLine
\Begin{
\CircQubitNum $\leftarrow$ \QuantumCircuit.qubit\_num\;
Set \SubGraphList to empty list\;
\For{i $\in$ \Starting}{
Set \SubGraph to empty list\;
\QubitNum $\leftarrow$ $0$\;
\While{\QubitNum $<$ \CircQubitNum}{
\If{\SubGraph is empty}{
\SubGraph.\Append(i)\;
\QubitNum $\leftarrow$ \QubitNum + 1 \;
continue\;
}
\BestQubit $\leftarrow$ \findBestQubit(\SubGraph, \CouplingGraph, \CalibrationData)\;
\If{\BestQubit $\neq$ None}{
\SubGraph.\Append(\BestQubit)\;
\QubitNum $\leftarrow$ \QubitNum + 1 \;
continue\;
}
}
\If{\Len(\SubGraph) = \CircQubitNum}{
\If{\QubitsUsed is empty}{
\SubGraph.\partitionScore(\CouplingGraph, \CalibrationData,
\QuantumCircuit,)\;
\SubGraphList.\Append(\SubGraph)\;
}
{\If{no qubit in \SubGraph is in \QubitsUsed}{
\crosstalkPairs $\leftarrow$ \findCrosstalk(\SubGraph, \crosstalkProperties, \QubitsUsed)\; \SubGraph.\partitionScore(\CouplingGraph, \CalibrationData, \QuantumCircuit, \crosstalkPairs)\;
\SubGraphList.\Append(\SubGraph)\;
}
}
}
}
\Return \SubGraphList \;
}
\end{algorithm}
\begin{figure}[t]
\begin{subfigure}{0.4\columnwidth}
\centering
\caption{
}
\includegraphics[scale=0.7,height=2.5cm]{ibmq_valencia_info.pdf}
\label{fig:ibmq_valencia_info}
\end{subfigure}
\hspace{5mm}
\begin{subfigure}{0.4\columnwidth}
\centering
\caption{
}
\includegraphics[scale=0.7,height=2.5cm]{partition_heuristic.pdf}
\label{fig:Qubit partition}
\end{subfigure}
\begin{subfigure}{\columnwidth}
\centering
\caption{
}
\includegraphics[scale=0.7]{qubit_degree.pdf}
\label{fig:qubit degree}
\end{subfigure}
\caption{Example of qubit partition on IBM Q 5 Valencia for a four-qubit circuit using QHSP. {\normalfont Suppose the largest logical degree of the target circuit is three.} (a) {\normalfont Calibration data of IBM Q 5 Valencia. The value inside of the node represents the readout error rate (in\%), and the value above the link represents the \texttt{CNOT} error rate (in\%).} (b) {\normalfont Process of constructing a partition candidate using QHSP. (c) The physical node degree and the fidelity degree of each qubit calculated by Eq.~\ref{eq:2}.}}
\label{fig:partition_example}
\end{figure}
\subsubsection{Runtime analysis}
\
\newline
Let $n$ be the number of hardware qubits and $k$ the number of qubits in the circuit to be allocated in a partition. GSP algorithm selects all the combinations of $k$ subgraphs from $n$-qubit hardware and takes $O(C(n,k))$ time, which is $O(n \, choose \, k)$. For each subgraph, it computes its fidelity score including calculating the longest shortest path, which scales at $O(k^3)$. It ends up being equivalent to $O(k^3min(n^k, n^{n-k}))$. In most cases, the number of circuit qubits is less than the number of hardware qubits, thus the time complexity becomes $O(k^3n^k)$. It increases exponentially as the number of qubits of the circuit augments.
QHSP algorithm starts by collecting a list of $m$ starting points where $m \leq n$. To get the starting points, we sort the $n$ physical qubits by their physical node degree, which takes $O(nlog(n))$. Then, we iterate over all the gates of the circuit (e.g. circuit has $g$ gates) and sort the $k$ logical qubits according to the logical node degree, which takes $O(g+klog(k))$. Next, for each starting point, it iteratively merges the best neighbour qubit until each sub-partition contains $k$ qubits. To find the best neighbour qubit, the algorithm finds the best qubit in a sub-partition and traverses all its neighbours to select the one with the highest fidelity degree. Finding the best qubit in the sub-partition is $O(p)$ where $p$ is the number of qubits in a sub-partition. The average number of qubits $p$ is $k/2$, so this process takes $O(k)$ time on average. Finding the best neighbour qubit is $O(1)$ because of the nearest-neighbor connectivity of superconducting devices. Overall, the QHSP takes $O(mk^2 + nlog(n) + g + klog(k))$ time, and it can be truncated to $O(mk^2 + nlog(n) + g)$, which is polynomial.
\subsection{Supplementary note - Mapping Transition Algorithm}
In this note, we present the pseudo-code of our simultaneous mapping transition algorithm (see Algorithm~\ref{algo3}).
\begin{algorithm}
\SetAlgoLined
\caption{Simultaneous mapping transition algorithm} \label{algo3}
\SetKwInOut{Input}{input}
\SetKwInOut{Output}{output}
\SetKwFunction{FindSwapBridgePairs}{FindSwapBridgePairs}
\SetKwFunction{MapUpdate}{Map\_Update}
\SetKwFunction{Append}{append}
\SetKwFunction{UpdateLayer}{Update\_Layer}
\SetKwFunction{cost}{$cost$}
\SetKwData{DAGs}{$DAGs$}
\SetKwData{DAG}{$DAG_i$}
\SetKwData{Fs}{$Fs$}
\SetKwData{F}{$F_i$}
\SetKwData{DistanceMatrices}{$Ds$}
\SetKwData{DistanceMatrix}{$D_i$}
\SetKwData{CouplingGraph}{$G$}
\SetKwData{Schedule}{schedule}
\SetKwData{gate}{gate}
\SetKwData{InitialMapping}{$\pi_i$}
\SetKwData{CurrentMapping}{$\pi_c$}
\SetKwData{swapBridgeCandidateList}{swap\_bridge\_candidate\_list}
\SetKwData{TentativeGate}{$g_{\text{tmp}}$}
\SetKwData{swapBridgeLists}{swap\_bridge\_lists}
\SetKwData{TempMapping}{$\pi_{\text{tmp}}$}
\SetKwData{HBasic}{$H_{\text{basic}}$}
\SetKwData{HTentative}{$H_{\text{tentative}}$}
\SetKwData{HTentativeNumber}{$N_{\text{tent}}$}
\SetKwData{HExtended}{$H_{\text{extend}}$}
\SetKwData{ExtendedLayer}{$E$}
\SetKwData{Weight}{$W$}
\SetKwData{H}{$H$}
\SetKwData{gFinal}{$g_n$}
\Input{Circuits \DAGs
, Coupling graph \CouplingGraph, Distance matrices \DistanceMatrices, Initial mapping \InitialMapping, First layers \Fs
}
\Output{Final schedule \Schedule}
\BlankLine
\Begin{
\CurrentMapping $\leftarrow$ \InitialMapping \;
\While{ not all gates are executed}{
Set \swapBridgeLists to empty list\;
\For{\F in \Fs}{
\For{\gate in \F}{
\If{\gate is hardware-compliant}{
\Schedule.\Append(\gate)\;
Remove \gate from \F\;
}
}
\If{\F is not empty}{
\swapBridgeCandidateList $\leftarrow$ \FindSwapBridgePairs(\F, \CouplingGraph)\;
\swapBridgeLists.\Append(\swapBridgeCandidateList)\;
}
}
\For{\swapBridgeCandidateList $\in$ \swapBridgeLists}{
\For{\TentativeGate $\in$ \swapBridgeCandidateList}{
\TempMapping $\leftarrow$ \MapUpdate(\TentativeGate, \CurrentMapping)\;
\HBasic $\leftarrow$ $0$\;
\For{\gate $\in$ \F}{
\HBasic $\leftarrow$ \HBasic + \DistanceMatrix(\gate, \TempMapping)
}
\HTentative $\leftarrow$ \TentativeGate.\cost(\CouplingGraph, \DistanceMatrix, \TempMapping)\;
Update the extended layer \ExtendedLayer\;
\HExtended $\leftarrow$ $0$\;
\For{\gate $\in$ \ExtendedLayer}{
\HExtended $\leftarrow$ \HExtended + \DistanceMatrix(\gate, \TempMapping)\;
}
\H $\leftarrow \frac{1}{\vert F + \HTentativeNumber \vert} \text{(\HBasic + \HTentative)} + \frac{\text{\Weight}}{\vert E \vert} \text{\HExtended}$
}
Choose the best gate \gFinal\;
\CurrentMapping $\leftarrow$ \MapUpdate(\gFinal, \CurrentMapping)\;
}
Update the First layers\;
}
\Return schedule
}
\end{algorithm}
\subsection{Supplementary note - Experimental results}
In this note, we demonstrate the exact experimental results when executing a different number of circuits on the two devices, IBM Q 27 Toronto and IBM Q 65 Manhattan, at the same time.
\begin {table*}[t]
\caption {Comparison of fidelity when executing two circuits simultaneously on IBM Q 27 Toronto.} \label{tab:result_toronto_fidelity}
\begin{center}
\resizebox{\textwidth}{!}{%
\begin{threeparttable}
\centering
\resizebox{\textwidth}{!}{%
\includegraphics{result_toronto.pdf}
}
\begin{tablenotes}
\footnotesize
\centering
\item $\mathbf{Avg}$: average of PSTs. $\mathbf{t}$: runtime in seconds of the partition process. $\mathbf{\Delta_{PST}}$: comparison of average fidelity.
\end{tablenotes}
\end{threeparttable}
}
\end{center}
\end{table*}
\begin {table*}
\caption {Comparison of number of additional gates when executing two circuits simultaneously on IBM Q 27 Toronto.} \label{tab:result_toronto_gate}
\begin{center}
\begin{threeparttable}
\centering
\includegraphics[scale=0.8]{result_gate_toronto.pdf}
\begin{tablenotes}
\footnotesize
\centering
\item $\mathbf{g}$: number of additional gates. $\mathbf{Sum}$: sum of number of additional gates. $\mathbf{\Delta_g}$: comparison of sum of number of additional gates.
\end{tablenotes}
\end{threeparttable}
\end{center}
\end{table*}
\begin {table*}
\caption {Comparison of fidelity when executing three circuits simultaneously on IBM Q 65 Manhattan.} \label{tab:result_manhattan_3_fidelity}
\begin{center}
\resizebox{\textwidth}{!}{%
\begin{threeparttable}
\centering
\resizebox{\textwidth}{!}{%
\includegraphics{result_mahattan_fidelity3.pdf}
}
\begin{tablenotes}
\footnotesize
\centering
\item $\mathbf{Avg}$: average of PSTs. $\mathbf{t}$: runtime in seconds of the partition process. $\mathbf{\Delta_{PST}}$: comparison of average fidelity.
\end{tablenotes}
\end{threeparttable}
}
\end{center}
\end{table*}
\begin {table*}
\caption {Comparison of number of additional gates when executing three circuits simultaneously on IBM Q 65 Manhattan.} \label{tab:result_manhattan_3_gate}
\begin{center}
\begin{threeparttable}
\centering
\includegraphics[scale=0.8]{result_mahattan_gate3.pdf}
\begin{tablenotes}
\footnotesize
\centering
\item$\mathbf{g}$: number of additional gates. $\mathbf{Sum}$: sum of number of additional gates. $\mathbf{\Delta_g}$: comparison of sum of number of additional gates.
\end{tablenotes}
\end{threeparttable}
\end{center}
\end{table*}
\begin {table*}
\caption {Comparison of fidelity when executing four circuits simultaneously on IBM Q 65 Manhattan.} \label{tab:result_manhattan_4}
\begin{center}
\resizebox{\textwidth}{!}{%
\begin{threeparttable}
\centering
\resizebox{\textwidth}{!}{%
\includegraphics{result_mahattan_fidelity4.pdf}
}
\begin{tablenotes}
\footnotesize
\centering
\item $\mathbf{Avg}$: average of PSTs. $\mathbf{t}$: runtime in seconds of the partition process. $\mathbf{\Delta_{PST}}$: comparison of average fidelity.
\end{tablenotes}
\end{threeparttable}
}
\end{center}
\end{table*}
\begin {table*}
\caption {Comparison of number of additional gates when executing three circuits simultaneously on IBM Q 65 Manhattan.} \label{tab:result_manhattan_gate_4}
\begin{center}
\begin{threeparttable}
\centering
\includegraphics[scale=0.8]{result_mahattan_gate4.pdf}
\begin{tablenotes}
\footnotesize
\centering
\item $\mathbf{g}$: number of additional gates. $\mathbf{Sum}$: sum of number of additional gates. $\mathbf{\Delta_g}$: comparison of sum of number of additional gates.
\end{tablenotes}
\end{threeparttable}
\end{center}
\end{table*}
|
\section{Random sections}
\section{Introduction}
We present a generalization of Ripley's $K$-function for shape-valued point processes, in particular, for point processes where each observation is a curve in $\mathbb{R}^3$, a fiber. Fiber structures appear naturally in the human body, for example in tracts in the central nervous system and in skeletal muscles. The introduced $K$-function captures both spatial and shape clustering or repulsion, thus providing a powerful descriptive statistic for analysis of medical image of sets of fiber or more general shape data.
As an example, Fig. 1 displays myelin sheaths in four configurations from different debts in a mouse brain. We develop the methodology to quantify the visually apparent differences in both spatial and shape distribution of the fibers.
\subsection{Background}
Ripley's $K$-function \cite{ripley_second-order_1976} is a well-known tool for analyzing second order moment structure of point processes \cite{baddeley_spatial_2015} providing a measure of deviance from complete spatial randomness in point sets. For a stationary point process, $K(t)$ gives the expected number of points within distance $t$ from a typical point. An estimator of Ripley's $K$-function for a point set $\{p_i\}_{i=1}^n$ inside an observation window $W$ is,
\begin{align}\label{RipleysKfunction}
\hat{K}(t) = \frac{1}{n\hat{\lambda}} \sum_{i\neq j }1[\text{dist}(p_i,p_j)<t]
\end{align}
where $\hat{\lambda} = \frac{n}{|W|}$ is the sample intensity, $|W|$ is the volume of the observation window, and $1$ is the indicator function. By comparing $\hat{K}(t)$ with the
$K$-function corresponding to complete spatial randomness, we can measure the deviation from spatial homogeneity. Smaller values of $\hat{K}(t)$ indicate clustering whereas the points tend to repel each other for greater values.\par
\begin{figure}
\centering
\begin{tabular}{cccc}
ST01 & ST06 & ST17 & ST20
\\\includegraphics[width=0.23\textwidth,trim={3.5cm 1.8cm 1.5cm 3.5cm}, clip]{plots/ST01_3d.pdf}
&\includegraphics[width=0.23\textwidth,trim={3.5cm 1.8cm 1.5cm 3.5cm}, clip]{plots/ST06_3d.pdf}
&\includegraphics[width=0.23\textwidth,trim={3.5cm 1.8cm 1.5cm 3.5cm}, clip]{plots/ST17_3d.pdf}
&\includegraphics[width=0.23\textwidth,trim={3.5cm 1.8cm 1.5cm 3.5cm}, clip]{plots/ST20_3d.pdf}
\\{\includegraphics[width=0.23\textwidth, trim={ 1.2cm 0cm 10.45cm 1.2cm}, clip]{plots/K_ST01.pdf}}
&{\includegraphics[width=0.23\textwidth, trim={ 1.2cm 0cm 10.45cm 1.2cm}, clip]{plots/K_ST06.pdf}}
&{\includegraphics[width=0.23\textwidth, trim={ 1.2cm 0cm 10.45cm 1.2cm}, clip]{plots/K_ST17.pdf}}
&{\includegraphics[width=0.23\textwidth, trim={ 1.2cm 0cm 10.45cm 1.2cm}, clip]{plots/K_ST20.pdf}}
\\\includegraphics[width=0.23\textwidth, trim={ 9.85cm 0cm 1.8cm 1.2cm}, clip]{plots/K_ST01.pdf}
&\includegraphics[width=0.23\textwidth, trim={ 9.85cm 0cm 1.8cm 1.2cm}, clip]{plots/K_ST06.pdf}
&\includegraphics[width=0.23\textwidth, trim={ 9.85cm 0cm 1.8cm 1.2cm}, clip]{plots/K_ST17.pdf}
&\includegraphics[width=0.23\textwidth, trim={ 9.85cm 0cm 1.8cm 1.2cm}, clip]{plots/K_ST20.pdf}
\end{tabular}
\caption{$K$-functions for samples of myelin sheaths. Each column corresponds to measured on a data set. (Top row): the centerlines of the myelin sheathed axons. (Middle row): The $K$-function for fixed values of $t$. (Bottom row): The $K$-function for fixed values of $s$.}
\label{myelin3d}
\end{figure}
Generalizations of Ripley's $K$-function have previously been considered for curve pieces in \cite{chiu_stochastic_2013} and several approaches were presented in \cite{sporring_generalizations_2019} for space curves. In this paper, we present a $K$-function inspired by the currents approach from \cite{sporring_generalizations_2019} and provide the theoretical account for the statistical foundation.
\par
The challenges when generalizing the $K$-function to shape-valued point processes arise in defining a distance measure on the shape space and determining a meaningful descriptive quantity that is well-defined and that we are able to estimate. In this paper, we provide a well-defined $K$-function for point processes in general metric spaces and use the embedding of shapes as currents to obtain a distance measure of shapes.
\subsection{Contributions and outline}
We construct the following two-parameter $K$-function for a curve-valued point process $X$
\begin{equation} \hat{K}(t,s) = \frac{1}{|W|\lambda}
\sum_{\gamma \in X: c(\gamma) \in W} \sum_{\gamma' \in X \setminus \{\gamma\} }
1[\|c(\gamma)-c(\gamma')\| \le t, d_m(\gamma,\gamma') \le s]
\end{equation}
for $t,s>0$, where $c(\gamma)$ is the center point of the curve $\gamma$, $d_m$ the minimal currents distance with respect to translation, and $\lambda$ is the spatial intensity of the center points. By introducing a second distance parameter, we are able to separate the spatial distance of the curves from the difference in shape, allowing us to measure both spatial and shape homogeneity.
\par
The paper thereby presents the following contributions:
\begin{enumerate}
\item A $K$-function for shape-valued point processes along with a theoretical account for the statistical foundation.
\item We suggest a certain fiber process which we argue corresponds to complete randomness of points, an analogue of the Poisson process.
\item An application of the $K$-function to several generated data set and a real data set of myelin sheaths.
\end{enumerate}
\section{Shapes as currents}
\subsection{Shape-valued point processes}
We model a random collection of shapes as a point process on the space of shapes. Shape spaces are usually defined as the space of embeddings $B_e(\mathcal{M},\mathbb{R}^d)$ of a manifold $\mathcal{M}$ into $\mathbb{R}^d$ \cite{bauer_overview_2014}. For example, the space of closed curves in $\mathbb{R}^3$ is $B_e(S^1,\mathbb{R}^3)$, where $S^1$ denotes the 1-sphere, and the space of fibers is $B_e(I,\mathbb{R}^3)$ for some real interval $I$.
\par
Formally, a point process $X$ on a metric space $S$ is a measurable map from some probability space $(\Omega,\mathcal{F},P)$ into the space of locally finite subsets of $S$. Thus, for each $\omega\in \Omega$, $X(\omega) \subseteq S$, and for every compact Borel set $B\subseteq S$, $X(\omega) \cap B$ is a finite set. Measurability of $X$ means that all sets of the form $\{ \omega \in \Omega | \#(X(\omega)\cap B)=m \}$, where $m\in \mathbb{N}_0$ and $B\subseteq S$ is a Borel set, must be measurable.
\par
There are different ways to endow $B_e(\mathcal{M},\mathbb{R}^d)$ with a metric \cite{michor_metric_2008}. In this paper, we consider the representation of shapes as currents embedded in the dual space of a reproducing kernel Hilbert space (RKHS). Thus, the RKHS metric induces a metric for our shape space. This can be combined with the Euclidean metric on $\mathbb{R}^d$ to obtain a suitable metric on $B_e(\mathcal{M},\mathbb{R}^d)$. This approach is very useful due to its generality and computability, as it requires very little information about the shape.
\subsection{Shapes as currents}
Shapes are usually more difficult to work with than points, as they usually cannot be captured in any finite dimensional vector space. An approach already considered for anatomical structures \cite{durrleman_statistical_2009} \cite{vaillant_surface_2005} is embedding shapes as currents.
We will give a brief introduction to this setup and refer to \cite{durrleman_statistical_2009} for a detailed description.
We can characterize a piece-wise smooth curve $\gamma\in B_e(I,\mathbb{R}^d)$ by computing its path-integral of all vector fields $w$
\begin{align}\label{Vgamme}
V_\gamma(w) = \int_\gamma w(x)^t\tau(x)d\lambda(x),
\end{align}
where $\tau(x)$ is the unit tangent of $\gamma$ at $x$ and $\lambda$ is the length measure on the curve. Likewise, an oriented hypersurface $S$ embedded in $\mathbb{R}^d$ can be characterized by its flux integral of vector fields $w$
\begin{align}
V_S(w) = \int_S w(x)^tn(x)d\lambda(x),
\end{align}
where $n(x)$ is the unit normal at $x$ and $\lambda$ is the surface area measure on $S$. These are both examples of representing shapes as currents, i.e.\ as elements in the dual space of the space of vector fields on $\mathbb{R}^d$. Formally, the space of $m$-currents $C_m$ is the dual space of the space $C^0(\mathbb{R}^d, (\Lambda^m\mathbb{R}^d)^*)$ of differential $m$-forms.
It is not only curves and hypersurfaces that can be represented as currents. Let $\mathcal{M}$ be an oriented rectifiable sub-manifold of dimension $m$ in $\mathbb{R}^d$ with positively oriented basis of the tangent space $u_1(x),...,u_m(x)$ for all $x\in \mathcal{M}$. The sub-manifold $\mathcal{M}$ can be embedded into the space of $m$-currents as the current
\begin{equation}\label{generalembedding}
T_\mathcal{M}(w) = \int_\mathcal{M} I(x) w(x) \Big( \frac{u_1(x) \wedge ... \wedge u_m }{|u_1(x) \wedge ... \wedge u_m|} \Big) d\lambda(x)
\end{equation}
where $w\in C^0(\mathbb{R}^d, (\Lambda^m\mathbb{R}^d)^*)$ is an $m$-differential form and $I:T\to \mathbb{R}$ is a scalar function satisfying $\int_T |I(x)| d\lambda(x)<\infty$ \cite{durrleman_statistical_2009}. Since shapes are embedded sub-manifolds, this means that shapes can be embedded into $C_m$.
\subsection{Reproducing kernel Hilbert space metric on shapes}
The space of $m$-currents $C_m$ is continuously embedded into the dual space of a reproducing kernel Hilbert space (RKHS) $H$ with arbitrary kernel $K_H:\mathbb{R}^d\times\mathbb{R}^d \to \mathbb{R}^{d\times d}$ \cite{durrleman_statistical_2009}.
It follows from Riesz representation theorem that $v\in H$ can be embedded in the dual space $H^*$ as the functional $\mathcal{L}_H(v)\in H^*$ defined by $\mathcal{L}_H(v)(w)= \langle v,w\rangle_H$ for $w\in H$.
\par
Elements $v(y) = K_H(x,y)\alpha$ form a basis for $H$ where $x,\alpha\in \mathbb{R}^d$, and basis elements in $H$ are lifted to basis elements in $H^*$ as $\delta_x^\alpha := \mathcal{L}_H(v)$ which are called the Dirac delta currents.
The element $V_\gamma$ from \eqref{Vgamme} can be written in terms of the basis elements $\delta^{\tau_{\gamma}(x_i)}_{x_i}$ where $\tau_{\gamma}(x_i)$ are the unit tangent vectors of $\gamma$ at $x_i$. This means that the curve $\gamma$ is embedded into $H^*$ as the $1$-current \begin{equation}
V_\gamma(w) = \int_\gamma w(x)^t\tau_{\gamma}(x)d\lambda(x) = \int_\gamma \delta_{x}^{\tau_{\gamma}(x)}(w)d\lambda(x)
\label{oneCurrent}
\end{equation}
where $\lambda$ is the length measure on the curve. Furthermore, it is approximated by the Riemann sum of Dirac delta currents $V_\gamma(w) \approx \tilde{V}_\gamma(w) =\sum_i \delta_{x_i}^{\tau(x_i)\Delta x_i}(w)$ where $x_i$ are sampled points along the curve according to $\lambda$.
The dual space $H^*$ inherits the inner product from the inner product on the RKHS via the inverse mapping $\mathcal{L}_H^{-1}$, so that the inner product for two curves $\gamma_1$ and $\gamma_2$ in $H^*$ is
\begin{align}
\langle V_{\gamma_1}, V_{\gamma_2}\rangle_{H^*} = \int_{\gamma_1}\int_{\gamma_2} \tau_{\gamma_1}^t(x)K_H(x,y)\tau_{\gamma_2}(y)d\lambda_{\gamma_2}(x)d\lambda_{\gamma_1}(y).
\end{align}
Writing $||V_{\gamma}||^2_{H^*}=\langle V_{\gamma}, V_{\gamma}\rangle_{H^*}$, we finally arrive at the currents distance of two curves $\gamma_1$ and $\gamma_2$
\begin{align}
d_c(V_{\gamma_1}, V_{\gamma_2}) =||V_{\gamma_1}- V_{\gamma_2}||_{H^*} = \Big( ||V_{\gamma_1}||_{H^*}^2 + ||V_{\gamma_2}||_{H^*}^2 -2\langle V_{\gamma_1}, V_{\gamma_2}\rangle_{H^*} \Big)^{1/2}.
\end{align}
In practice, we usually don't know the orientation of the curves, thus we choose to consider the minimal distance between them,
\begin{equation}
d(V_{\gamma_1}, V_{\gamma_2}) = \min\{ d_c(V_{\gamma_1}, V_{\gamma_2}), d_c(V_{\gamma_1}, V_{-\gamma_2}) \}
\label{mindistance}
\end{equation}
where $-\gamma_2$ denotes the curve with opposite orientation of $\gamma_2$. If the orientation of the data is important, this step may be omitted. From \eqref{oneCurrent} we see that $K_H$ serves as a weight of the inner product between $\tau_1(x_i)$ and $\tau_2(y_j)$ depending on the positions $x_i$ and $y_j$.
\subsection{A note on short lines and generalized Gaussian kernels}
To illustrate the distance metric, consider the generalized Gaussian kernel
\begin{equation}\label{gauskernel}
K^p_\sigma(x,y) = \exp\Big({\frac{-|x-y|^p}{2\sigma^p}}\Big)\text{Id},
\end{equation}
where $\sigma , p\in (0,\infty] $, and consider two lines of equal length parametrized by $l_u(t) = x_u + u t,\, l_v(t)=x_v+v t$ where $x_u,x_v,u,v \in \Re^d$ and $0<t\leq T\in\Re$. For very short lines far from each other, i.e., $T/|x_u-x_v|\rightarrow 0$, we have,
\begin{equation}
\frac{d(\tilde{V}_{l_u},\tilde{V}_{l_v})^2}{{T^2}} \rightarrow
d_0 -2\exp\big({\frac{-|x_u-x_v|^p}{2\sigma^p}}\big)d_1,
\end{equation}
where $d_0 = u^tu + v^tv$ and $d_1=\max\left(u^tv,-u^tv\right)$. Since the $d_0$ and $d_1$ are constants, and since the exponential and the square root functions are both monotonic, then in the limit, $d(\tilde{V}_{l_u},\tilde{V}_{l_v})/T$ is one-to-one with $|x_u-x_v|^p$ which is one-to-one with $|x_u-x_v|$. Thus, for very short lines, $d/T$ is one-to-one with the euclidean distance between the points $x_u$, and $x_v$. Further, in the limit $p\rightarrow\infty$ we have,
\begin{equation}
\frac{d(\tilde{V}_{l_u},\tilde{V}_{l_v})^2}{T^2}\rightarrow
\begin{cases}
d_0 -2\,d_1, &\text{ when } |x_u-x_v| < \sigma,\\
d_0 -2\exp\left(-\frac12\right)d_1, &\text{ when } |x_u-x_v| = \sigma,\\
d_0, &\text{ otherwise}.
\end{cases}
\end{equation}
Thus, for very short lines and very large exponents, $(d_0 - d^2/T^2)/(2d_1)$ converges to a unit step function in $|x_u-x_v|$ where the step is at $\sigma$.
\section{The $K$-function}
\subsection{Statistical Setup}
Let $S$ be the image of the embedding of $B_e(I,\mathbb{R}^d)$ into $C_1$. For brevity, $\gamma$ is identified with its representation in $C_1$. We model a random collection of curves as a point process $X$ in $S$.
\par
Let $c:S\to \mathbb{R}^d$ be a center function on the space of fibers in $\mathbb{R}^d$ that associates a center point to each fiber. A center function should be translation covariant in the sense that $c(\gamma + x) = c(\gamma) + x$ for all $x\in \mathbb{R}^d$. It could be the center of mass or the midpoint of the curve with respect to curve length.
Let $S_c$ denote the space of centered fibers wrt. $c$, i.e., those $\gamma\in S$ for which $c(\gamma)=0$. We define $\gamma_c := \gamma -c(\gamma)\in S_c$ to be the centering of $\gamma$.
\par
For Borel sets $B_1,A_1 \subset \mathbb{R}^d$ and $B_2,A_2 \subset S_c$, define the first moment measure
\[ \mu(B_1 \times B_2) ={\mathbb E} \sum_{\gamma \in X}1[ c(\gamma) \in B_1, \gamma_c \in B_2] \]
and the second moment measure
\[ \alpha((A_1 \times A_2) \times (B_1 \times B_2)) = {\mathbb E} \sum_{\gamma ,\gamma' \in X}^{\neq} 1[ c(\gamma) \in A_1, \gamma_c \in A_2]1[ c(\gamma') \in B_1, \gamma'_c \in B_2]. \]
We assume that $\mu$ is translation invariant in its first argument, i.e.\
\[ \mu(B_1 \times B_2)= \mu((B_1+h)\times B_2) \]
for any $h \in \mathbb{R}^d$. This is for instance the case if the distribution of $X$ is invariant under translations. This implies that $\mu(\cdot \times B_2)$ is proportional to the Lebesgue measure for all $B_2$. Thus we can write
\[ \mu(B_1 \times B_2) = |B_1| \nu(B_2) \]
for some measure $\nu(\cdot)$ on $S_c$. Note that the total measure $\nu(S_c)$ is the spatial intensity of the center points, i.e.\ the expected number of center points in a unit volume window. In applications, this will typically be finite. In this case, we may normalize $\nu$ to obtain a probability measure which could be interpreted as the distribution of a single centered fiber.
\par
We define the reduced Campbell measure
\[ C^! (A_1\times A_2 \times F)= {\mathbb E} \sum_{\gamma \in X}1[ c(\gamma) \in A_1, \gamma_c \in A_2 , X \setminus \{\gamma \} \in F] \]
where the "$!$" represents the removal of the point $\gamma$ from $X$. By disintegration,
\[ C^! (A_1\times A_2 \times F)= \int_{A_1 \times A_2} P^!_{c,\gamma_c}(F) \mu(\mathrm{d} (c,\gamma_c)) .\]
By the standard proof, we get for any measurable function $h:\mathbb{R}^d \times S_c \times \mathcal{N} \to [0,\infty)$
\begin{equation}\label{eq:cm}
{\mathbb E} \sum_{\gamma \in X}h(c(\gamma),\gamma_c,X \setminus \{\gamma \}) = \int_{\mathbb{R}^d \times S_c} {\mathbb E}^!_{c,\gamma_c}h(c,\gamma_c,X) \mu(\mathrm{d} (c,\gamma_c)).
\end{equation}
In particular,
\begin{equation}
\alpha((A_1 \times A_2) \times (B_1 \times B_2)) = \int_{A_1 \times A_2} {\mathbb E}^!_{c,\gamma_c} \sum_{\gamma' \in X} 1[ c(\gamma') \in B_1, \gamma'_c \in B_2] \mu(\mathrm{d} (c,\gamma_c)).
\end{equation}
Assume also that $\alpha$ is invariant under joint translation of the arguments $A_1,B_1$. Then
\begin{align}{\cal K}_{c,\gamma_c}(B_1 \times B_2) &:= {\mathbb E}^!_{c,\gamma_c} \sum_{\gamma' \in X} 1[ c(\gamma') \in B_1, \gamma_c' \in B_2] \\
&\phantom{:}= {\mathbb E}^!_{0,\gamma_0} \sum_{\gamma' \in X} 1[ c(\gamma') \in B_1-c, \gamma_c' \in B_2] \\
&\phantom{:}= {\cal K}_{0,\gamma_0}((B_1-c) \times B_2)
.\end{align}
Assume that also ${\mathbb E}^!_{c,\gamma_c}h(c,\gamma_c,X)$ does not depend on $c$, which is true if the distribution of $X$ is invariant over translations. Then, using \eqref{eq:cm} and the factorization of $\mu$,
\begin{align} \nonumber {\mathbb E} \sum_{\gamma \in X } 1[c(\gamma) \in W]h(c(\gamma),\gamma_c,X \setminus \{\gamma \}) &= \int_{W\times S_c} {\mathbb E}^!_{c,\gamma_c}h(c,\gamma_c,X) \mu(\mathrm{d} (c,\gamma_c))\\ = \int_{W\times S_c} {\mathbb E}^!_{0,\gamma_0}h(0,\gamma_0,X) \mu(\mathrm{d} (c,\gamma_0)) &= |W| \int_{S_c} {\mathbb E}^!_{0,\gamma_0}h(0,\gamma_0,X) \nu(\mathrm{d} \gamma_0)
\end{align}
From this it follows that
\[ E_h = \frac{1}{|W|} \sum_{\gamma \in X} 1[c(\gamma) \in W] h(c(\gamma),\gamma_c,X \setminus \{\gamma \}) \]
is an unbiased estimator of
$ \int_{S_c} {\mathbb E}^!_{0,\gamma_0}h(0,\gamma_0,X) \nu(\mathrm{d} \gamma_0).$
Furthermore, if $\nu(S_c)$ is finite,
\[ \int_{S_c} {\mathbb E}^!_{0,\gamma_0}h(0,\gamma_0,X) \nu(\mathrm{d} \gamma_0) = \nu(S_c) {\mathbb E}_{\tilde{\nu}} {\mathbb E}^!_{0,\Gamma_0}h(0,\Gamma_0,X)\]
where $\Gamma_0$ is a random centered fiber with distribution $\tilde{\nu}(\cdot)=\nu(\cdot)/\nu(S_c)$ and ${\mathbb E}_{\tilde{\nu}}$ is expectation with respect to this distribution of $\Gamma_0$.
\subsection{$K$-function for fibers}
In order to define a $K$-function, we must make an appropriate choice of $h$. A seemingly natural choice for $h$ that coincides with \cite{sporring_generalizations_2019}, is
\[ h(c,\gamma_c,X) = \sum_{\gamma' \in X} 1[d(\gamma_c+c,\gamma') \le t] = \sum_{\gamma' \in X} 1[d(\gamma,\gamma') \le t]. \]
However this choice allows the $K$-function to be a.s.\ infinite, due to the fact that $d(\gamma,\gamma')\leq \sqrt{2(||\gamma||_{H^*}^2+||\gamma'||_{H^*}^2)}$. If every curve in $X$ has $||\gamma||_{H^*}\leq M$, e.g.\ if the length of fibers is bounded, then choosing $t\geq 2M$ results in any fiber in $X$ having infinitely many neighbors within distance $t$.
\par
A solution is to separate the spatial distance of the curves from the difference in shape by introducing another radius parameter for the distance between center points. Accounting for spatial distance with this parameter, we choose to minimize the influence of spatial distance by measuring the currents distance between the centered curves.
Thus, we choose $h$ as
\begin{equation}
h(c,\gamma_c,X) = \sum_{\gamma' \in X\setminus \{\gamma\}}1[\|c(\gamma)-c(\gamma')\| \le t, d(\gamma_c,\gamma_c') \le s]
\end{equation}
for $s,t>0$, where $||c(\gamma)-c(\gamma')||$ is the usual distance in $\mathbb{R}^d$ between center points.
Thus we define the empirical $K$-function for $t,s>0$ as
\begin{equation}\label{eq:Kfix} \hat{K}(t,s) = \frac{1}{|W|\nu(S_c)}
\sum_{\substack{\gamma \in X: c(\gamma) \in W, \\ \gamma' \in X \setminus \{\gamma\} }}
1[\|c(\gamma)-c(\gamma')\| \le t, d(\gamma,\gamma') \le s] .\end{equation}
Since $\nu(S_c) $ is the intensity of fiber centers, it is estimated by $N/|W|$ where $N$ is the observed number of centers $c(\gamma)$ inside $W$. The $K$-function is the expectation of the empirical $K$-function
\[K(t,s)=\mathbb{E} \hat{K}(t,s) ={\mathbb E}_{\tilde{\nu}} {\mathbb E}^!_{0,\Gamma_0}h(0,\Gamma_0,X) = {\mathbb E}_{\tilde{\nu}} {\cal K}_{0,\Gamma_0}(B(0,t)\times B_c(\Gamma_0,s)) \]
where $B(x,r) = \{y: ||x-y||\le r \}$ and $B_c(\gamma,s) = \{\gamma' : d(\gamma,\gamma')\le s\}$.
\subsection{$K$-function for general shapes}
The currents metric and the $K$-function easily extends to shape-valued point processes with values in $B_e(\mathcal M, \Omega)$ for more general manifolds $\mathcal M$ and $\Omega\subset \mathbb{R}^d$. Shapes $\mathcal A, \mathcal B\in B_e(\mathcal M,\mathbb{R}^d)$ are embedded as $m$-currents $V_{\mathcal A}$ and $V_{\mathcal B}$ as in \eqref{generalembedding}. Since $C_m$ is continuously embedded into the dual RKHS $H^*$, we get the distance measure $d_c$ between shapes.
\par
If $c: B_e(\mathcal M,\Omega)\to \mathbb{R}^d$ is a center function, then we can generalize the $K$-function to a point process $X$ with values in $B_e(\mathcal M,\Omega)$. Identifying elements of $B_e(\mathcal M, \Omega)$ with their embedding in $H^*$, we can write the same $K$-function
\begin{equation}
\hat{K}(t,s) = \frac{1}{|W|\lambda}
\sum_{\substack{\mathcal U \in X: c(\mathcal U) \in W, \\ \mathcal U' \in X \setminus \{\mathcal U\}} }
1[\|c(\mathcal U)-c(\mathcal U')\| \le t, d_m(\mathcal U,\mathcal U') \le s]
\end{equation}
for $t,s>0$, where $d_m$ is constructed as in Section $3.2$ and $\lambda$ is the spatial intensity of the center points.
\section{Experiments}
To obtain a measure of spatial homogeneity, Ripley's $K$-function for points is usually compared with the $K$-function for a Poisson process, $K_P(t) = \text{vol}(B_d(t))$, corresponding to complete spatial randomness.
We are now in a more complicated situation where the $K$-function has two parameters and we do not have a notion of complete randomness of fibers.
The aim of the experiments on generated data sets is to analyze the behavior of the $K$-function on different types of distributions and suggest
a fiber process that corresponds to complete randomness. This will serve as a way to compare the results in 4.2.
\subsection{Generated data sets}
The four generated data sets $X_1, X_2, X_3$ and $X_4$ each contain $500$ fibers with curve length $l = 40$ and center points in $[0,100]^3$ and is visualized in the first row of Fig. \ref{GeneratedDataset}. Each data set is created by sampling center points from a distribution on $\mathbb{R}^3$ and fibers from a distribution on $S_0$, that is then translated by the center points. For the first three data sets, the center points are generated by a Poisson process and the fibers are uniformly rotated lines in $X_1$, uniformly rotated spirals in $X_2$ and Brownian motions in $X_3$. The data set $X_4$ has clustered center points and within each cluster the fibers are slightly perturbed lines.
\par
To avoid most edge effects, we choose the window $W \approx [13,87]^3\subset \mathbb{R}^3$ for the calculation of the $K$-function. Furthermore, we choose a Gaussian kernel $K_\sigma$ as in \eqref{gauskernel} with $p=2$ and $\sigma = \frac{100}{3}$. Finally, $c$ is defined to be the mass center of the curve.
\begin{figure}
\centering
\begin{tabular}{cccc}
Uniform Lines & Uniform Spirals & Uniform Brownian & Clustered Lines
\\{\includegraphics[width=0.23\textwidth,trim={3.5cm 2.8cm 2cm 3.5cm}, clip]{plots/uni_lin_randrot_3d.png}}
&{\includegraphics[width=0.23\textwidth,trim={3.5cm 2.8cm 2cm 3.5cm}, clip]{plots/uni_spiral_randrot_3d.png}}
&{\includegraphics[width=0.23\textwidth,trim={3.5cm 2.8cm 2cm 3.5cm}, clip]{plots/uni_brownian_3d.png}}
&{\includegraphics[width=0.23\textwidth,trim={3.5cm 2.8cm 2cm 3.5cm}, clip]{plots/MoG_linear_each_randrot_3d.pdf}}
\\{\includegraphics[width=0.23\textwidth, trim={ 1.2cm .2cm 10.45cm 1.2cm}, clip]{plots/K_uni_lin_randrot_v2_0.pdf}}
&{\includegraphics[width=0.23\textwidth, trim={ 1.2cm .2cm 10.45cm 1.2cm}, clip]{plots/K_uni_spiral_randrot_v2_0.pdf}}
&{\includegraphics[width=0.23\textwidth, trim={ 1.2cm .2cm 10.45cm 1.2cm}, clip]{plots/K_uni_brownian_v2_0.pdf}}
&{\includegraphics[width=0.23\textwidth, trim={ 1.2cm .2cm 10.45cm 1.2cm}, clip]{plots/K_Mog_linear_each_randrot_v2_0.pdf}}
\\\includegraphics[width=0.23\textwidth, trim={ 9.85cm 0.2cm 1.8cm 1.2cm}, clip]{plots/K_uni_lin_randrot_v2_0.pdf}
&\includegraphics[width=0.23\textwidth, trim={ 9.85cm 0.2cm 1.8cm 1.2cm}, clip]{plots/K_uni_spiral_randrot_v2_0.pdf}
&\includegraphics[width=0.23\textwidth, trim={ 9.85cm 0.2cm 1.8cm 1.2cm}, clip]{plots/K_uni_brownian_v2_0.pdf}
&\includegraphics[width=0.23\textwidth, trim={ 9.85cm 0.2cm 1.8cm 1.2cm}, clip]{plots/K_Mog_linear_each_randrot_v2_0.pdf}
\end{tabular}
\caption{$K$-function on the generated data, where each column corresponds to one data set. (Top row): The data sets $X_1, X_2, X_3$ and $X_4$ described in 4.1. (Middle row): The $K$-function of the data set above for fixed values of $t$. (Bottom row): The $K$-function of the data set above for fixed values of $s$. }
\label{GeneratedDataset}
\end{figure}
The first row of Fig. \ref{GeneratedDataset} shows the generated data sets and the respective $K$-functions are visualized the second and third row. In the second row, $s\mapsto K(t,s)$ is plotted for fixed values of $t$. For example, the graphs with $t=50$ show the expected numbers of fibers within currents distance $s$, where the distance of center points are $50$ at most. Lastly in the third row, $t\mapsto K(t,s)$ is plotted for fixed values of $s$. Similarly, the graphs with $s=70$ show the expected numbers of fibers with center point distance $t$ when the currents distances are $70$ at most. Thus, the graphs in the second row capture the fiber shape difference of each data set whereas the graphs in the third row capture spatial difference.
\par
It is distribution $X_3$ that we consider to a natural suggestion for a uniform randomness distribution of fibers. This is because Brownian motions are well-known for modelling randomness, thus representing shape randomness. And by translating these Brownian motion with a Poisson process, we argue that this distribution is a good choice.
\par
Considering only the data sets with uniformly distributed center points, i.e., the first three columns of Fig. 2, we see a big difference in the second row of plots. This indicates that the $K$-function is sensitive to the change in shape. The $K$-function for the Brownian motions captures much more mass for smaller radii compared to the lines, with the spirals being somewhere in-between. The plots in the third row are very much as expected, since we generated the center points from a Poisson process. Finally, the second row plot for $X_4$ indicate a slight shape clustering when compared to the uniformly rotated lines. This makes sense, since each cluster is directed differently.
\subsection{Application to myelin sheaths}
Myelin surrounds the nerve cell axons and is an example of a fiber structure in the brain. Based on 3D reconstructions from the region motor cortex of the mouse brain, centre lines were generated in the myelin sheaths. The data sets ST01, ST06, ST17 and ST20 displayed in the first row of Fig. 1 represent the myelin sheaths from four samples at different debts.
\par
For real shape-valued data sets, it very common that only parts of the shapes are observed. This is the case for many fiber data sets as well. This fact is important to have in mind when choosing $c$, since we should have a clear idea of when $c(\gamma)$ is observed, in order to get an unbiased estimate.
\par
Since myelin sheaths tend to be quite long, we chose to divide the fibers of length greater than $40$ into several fibers segments of length $40$. This has the benefits, that the mass center is an appropriate choice for $c$ and that the results are comparable with the results of Fig. 2, since the curves are of similar length.
\par
The results of the estimated K-function on the four data sets ST01, ST06, ST17 and ST20 are visualized in Fig. 1, where $s\mapsto K(t,s)$ is plotted in the second row for fixed values of $t$ and $t\mapsto K(t,s)$ is plotted in the third row for fixed values of $s$. The plots in the second row showing the fiber shape are very similar, resembling the fiber distribution of $X_2$. We notice a slight difference in ST20, where the graphs have a more pronounced cut off.
When noticing the scale of the $y$-axis, we see that the expected number of neighbor fibers vary significantly between the data sets.
\par
The third row plots indicate that the center point distributions of each data set is similar to the center point distribution of $X_1$, $X_2$ and $X_3$, which we generated from a Poisson process. For ST17, we notice a slight clustering of center points for $t\in [10,15]$. The biggest difference is for the graphs for $t=30$, indicating that the neighbors for fibers i ST20 are of more similar shape than the others.
\par
\bibliographystyle{splncs04}
|
\section{Introduction}
\label{section introduction}
The study of near-rationality properties of algebraic varieties is a central problem in algebraic geometry. This paper is motivated by the following question - do the standard norm varieties over a field constructed by Rost-Voevodsky \cite{Rost-ICM2002} (see also \cite{Suslin-Joukhovitski}) admit any near-rationality properties, after passing to the algebraic closure of the base field? The standard norm varieties are a crucial input in the proof of the Bloch-Kato conjecture \cite{Voevodsky-Bloch-Kato}. It has been shown in \cite[Proposition 2.6 and Remark 2.9]{Asok-unramified} that standard norm varieties over a field $k$ (corresponding to a symbol in the mod-$\ell$ Milnor $K$-group of $k$) are rationally connected after base change to $\-k$. However, it is not known if they are retract rational. Over a field $k$ of characteristic $0$, various near-rationality properties of varieties are related as follows:
\[
\begin{split}
\text{rational}~ \Rightarrow ~ \text{stably rational} ~ \Rightarrow ~ & \text{retract rational} ~ \Rightarrow ~ \text{unirational} ~ \Rightarrow ~ \text{rationally connected} \\
& \quad \quad \quad \Downarrow\\
& ~ \text{${\mathbb A}^1$-connected}
\end{split}
\]
where a variety $X$ is ${\mathbb A}^1$-connected if the sheaf of its ${\mathbb A}^1$-connected components $\pi_0^{{\mathbb A}^1}(X)$ (in the sense of \cite{Morel-Voevodsky}) is the trivial (single point) sheaf. See \cite[Section 1]{CT-Sansuc-rationality}, for example, for the horizontal implications (over an arbitrary field) and \cite[Theorem 2.3.6]{Asok-Morel} for the vertical implication. It is an open question, whether ${\mathbb A}^1$-connectedness is equivalent to retract rationality \cite[Remark 2.3.11]{Asok-Morel}.
Pfister quadrics corrresponding to symbols in mod-$2$ Milnor $K$-theory are norm varieties and are obviously ${\mathbb A}^1$-connected over an algebraically closed field, being rational. By the main results of \cite{Nguyen}, Pfister quadrics are birational to standard norm varieties. Consequently, standard norm varieties in the case $\ell = 2$ are rational and hence, ${\mathbb A}^1$-connected.
Let $\ell$ be an arbitrary prime number. Given an ordered sequence $\alpha = \{a_1, \ldots, a_n\}$ in $k^\times$, the standard norm variety $X_\alpha$ is constructed by an inductive procedure (see Section \ref{section norm varieties} for an outline of the construction following \cite{Rost-ICM2002} and \cite{Suslin-Joukhovitski}). The standard norm variety $X_\alpha$ has nice splitting properties with respect to the symbol $\{a_1, \ldots, a_n\} \in \KM_n(k)/\ell$ (see \cite{Suslin-Joukhovitski}). Note that it is not clear how standard norm varieties associated with given two ordered sequences of units $a_1, \ldots, a_n$ and $b_1, \ldots, b_n$ in $k$ such that equality of the associated symbols $\{a_1, \ldots, a_n\} = \{b_1, \ldots, b_n\} \in \KM_n(k)/\ell$ holds are related. It has been conjectured \cite[Conjecture 1.1]{Nguyen} that such standard norm varieties are birational. For a pair of units $\{a_1, a_2\}$, the standard norm variety is the Severi-Brauer variety associated with the cyclic algebra corresponding to the symbol $\{a_1, a_2\} \in \KM_2(k)/\ell$, which is ${\mathbb A}^1$-connected precisely when it admits a $k$-rational point (since it is isomorphic to a projective space). The main result of this paper shows that a standard norm variety is ${\mathbb A}^1$-connected, after passing to the algebraic closure $\-k$ of the base field $k$ (see Theorem \ref{theorem:norm varieties}).
\begin{introtheorem}
\label{intro thm: norm varieties}
Let $k$ be a field of characteristic $0$ and let $\ell$ be any prime number. Let $n\geq 2$ be an integer and $a_1, \ldots, a_n \in k^{\times}$. Let $X_\alpha$ denote the standard norm variety associated with the ordered sequence $\alpha = \{a_1, \ldots, a_n\}$. Then $(X_\alpha)_{\-k}$ is ${\mathbb A}^1$-connected.
\end{introtheorem}
Since ${\mathbb A}^1$-connectedness is equivalent to universal $R$-triviality for smooth proper varieties, Theorem \ref{intro thm: norm varieties} says that the standard norm varieties over a field of characteristic $0$ are universally $R$-trivial after base change to the algebraic closure of the base field. A result of Karpenko and Merkurjev \cite{Karpenko-Merkurjev} states that the standard norm variety $X_\alpha$ is universally ${\rm CH}_0$-trivial. It is known that universal ${\rm CH}_0$-triviality is a strictly stronger condition than universal $R$-triviality. Thus, Theorem \ref{intro thm: norm varieties} gives additional information about $X_\alpha$ on base change to $\-k$.
The proof of Theorem \ref{intro thm: norm varieties} contains two key ingredients, both of which are of independent interest. The first ingredient is the following criterion for ${\mathbb A}^1$-connectedness of a variety $X$ over a field (see Theorem \ref{theorem:criterion2}).
\begin{introtheorem}
\label{intro thm: criterion}
Let $X$ be a variety over a perfect field $k$. Suppose that there exists a point $x_0 \in X(k)$ such that for any $x \in X$, one of the following conditions hold:
\begin{enumerate}[label=$(\alph*)$]
\item $\overline{\{x\}}$ contains $x_0$, and $x$ and $x_0$ have the same image in $\pi_0^{\#A^1}(\overline{\{x\}})(k(x))$;
\item there exists a smooth, irreducible curve in $X$ passing through $x$ and not contained in $\overline{\{x\}}$.
\end{enumerate}
Then $X$ is $\#A^1$-connected.
\end{introtheorem}
A straightforward consequence of Theorem \ref{intro thm: criterion} is the characterization of ${\mathbb A}^1$-connectedness of a smooth proper variety $X$ over a field $k$ in terms of the condition that the generic point of $X$ can be connected to a $k$-rational point using naive ${\mathbb A}^1$-homotopies or equivalently, a chain of $\P^1$'s (see Corollary \ref{cor criterion}). The hypotheses of Theorem \ref{intro thm: criterion} imply that the generic point of $X$ can be connected to $x_0$ by an \emph{$n$-ghost homotopy} in the sense of \cite[Definition 2.7]{Balwe-Sawant-ruled}, thanks to the results of \cite{Balwe-Hogadi-Sawant}. The proof of Theorem \ref{intro thm: criterion} involves showing that the given ${\mathbb A}^1$-ghost homotopy between the generic point of $X$ and $x_0$ can be systematically enlarged to show that any two points of $X$ over a given essentially smooth field extension $F$ of $k$ map to the same element of $\pi_0^{{\mathbb A}^1}(X)(F)$, which allows one to conclude using \cite[Lemma 6.1.3]{Morel-connectivity}. Theorem \ref{intro thm: criterion} and several useful consequences of it are proved in Section \ref{section criterion}.
The second key ingredient in the proof is ${\mathbb A}^1$-connectedness of smooth compactifications of the symmetric powers of an ${\mathbb A}^1$-connected variety over a field, studied in Section \ref{section symmetric}. We show that the symmetric power of an ${\mathbb A}^1$-connected smooth projective variety is ${\mathbb A}^1$-connected (see Theorem \ref{theorem:symmetric}) as a consequence of Theorem \ref{intro thm: criterion} in the appendix. However, this is not sufficient to conclude ${\mathbb A}^1$-connectedness of smooth compactifications of the symmetric powers as they are not smooth in general. Since ${\mathbb A}^1$-connectedness is a birationally invariant property of smooth proper schemes, we only need to find one smooth birational proper model of the symmetric power of an ${\mathbb A}^1$-connected smooth projective variety. This is done with the help of some tricky geometric arguments involving approximating points on rational varieties by rational curves and the geometric structure of symmetric powers. The main result of Section \ref{section symmetric} is as follows (see Theorem \ref{theorem:symmetric_model}).
\begin{introtheorem}
\label{intro thm: symmetric}
Let $X$ be an ${\mathbb A}^1$-connected smooth projective variety over an algebraically closed field $k$ of characteristic $0$. Then any smooth proper variety birational to ${\rm Sym}^m X$ is ${\mathbb A}^1$-connected, where $m \geq 1$ and ${\rm Sym}^m X$ denotes the quotient of $X^m$ by the action of the symmetric group $S_m$ by permutation of factors.
\end{introtheorem}
Given the above two ingredients, the proof of Theorem \ref{intro thm: norm varieties} follows from the inductive construction of the standard norm varieties and the observation that compactifications of norm hypersurfaces are ${\mathbb A}^1$-connected (see Proposition \ref{prop:norm-hypersurface}). The details can be found in Section \ref{section norm varieties}.
\subsection*{Acknowledgement}
We thank Aravind Asok, Jean-Louis Colliot-Th\'el\`ene, Bruno Kahn, Burt Totaro, Suraj Yadav and the anonymous referee for their comments.
\subsection*{Conventions and notation}
By a variety over a field $k$, we mean a reduced and irreducible scheme of finite type over $k$. We will always work with the big Nisnevich site $Sm/k$ of smooth schemes over a field $k$. Given a simplicial sheaf of sets $\@X$ on $Sm/k$ and an affine scheme ${\rm Spec \,} A$ over $k$, we will write $\@X(A)$ for $\@X({\rm Spec \,} A)$ for the sake of brevity.
For a scheme $X$ (over ${\rm Spec \,} {\mathbb Z}$), the residue field of a scheme-theoretic point $x$ will be denoted by $\kappa(x)$. In case $X$ is a scheme over over a field $k$, we will sometimes denote the residue field of $x$ by $k(x)$.
\section{Geometric criteria for \texorpdfstring{$\#A^1$}{A1}-connectedness}
\label{section criterion}
We will freely use the notation and terminology from \cite{Morel-Voevodsky}, \cite{Balwe-Hogadi-Sawant} and \cite{Balwe-Sawant-ruled}, especially regarding the sheaves of naive and genuine ${\mathbb A}^1$-connected components.
For any scheme $U$ over $k$, we let $\sigma_0$ and $\sigma_1$ denote the morphisms $U \to U \times \#A^1$ given by $u \mapsto (u,0)$ and $u \mapsto (u,1)$, respectively. An \emph{$\#A^1$-homotopy} of $U$ in a scheme $X$ over $k$ is a morphism $h: U \times {\mathbb A}^1 \to X$ and we say that $h$ connects $h(0)$ and $h(1)$. An \emph{$\#A^1$-chain homotopy} of $U$ in $X$ is a finite sequence $h=(h_1, \ldots, h_r)$ where each $h_i$ is an $\#A^1$-homotopy of $U$ in $X$ such that $h_i(1) = h_{i+1}(0)$ for $1 \leq i \leq r-1$. In this case, we say that $h_1(0)$ and $h_r(1)$ are \emph{$\#A^1$-chain homotopic}.
\begin{definition}
\label{defn movable}
A point $x$ of a scheme $X$ is said to be \emph{movable}, if there exists a smooth, irreducible curve $C$ over $\kappa(x)$, a point $c_0 \in C(\kappa(x))$ and a morphism $\gamma: C \to X$ such that $\gamma(c_0) = x$ and the image of $\gamma$ is not contained in the closure $\overline{\{x\}}$ of $x$ in $X$.
\end{definition}
It is easy to see that every non-generic smooth point of a variety is movable. In fact, a non-movable point on a variety has to be a singular point of every closed subvariety that properly contains it. The main result of this section is the following geometric criterion for ${\mathbb A}^1$-connectedness.
\begin{theorem}
\label{theorem:criterion2}
Let $X$ be a variety over a perfect field $k$. Suppose that there exists a point $x_0 \in X(k)$ such that for any $x \in X$, one of the following conditions hold:
\begin{enumerate}[label=$(\alph*)$]
\item $\overline{\{x\}}$ contains $x_0$, and the $x$ and $x_0$ have the same image in $\pi_0^{\#A^1}(\overline{\{x\}})(k(x))$.
\item $x$ is a movable point of $X$.
\end{enumerate}
Then $X$ is $\#A^1$-connected.
\end{theorem}
\begin{proof}
Let $L_{{\mathbb A}^1}$ denote the fibrant approximation functor \cite[\textsection 2, Lemma 2.6, page 107]{Morel-Voevodsky} for the $\#A^1$-model structure. We define $\@X := L_{\#A^1}(X)$, so that $\pi_0^{\#A^1}(X) = \pi_0(\@X)$.
We will prove that for any point $x$ of $X$, the images of $x$ and $x_0$ in $\pi_0^{\#A^1} (X)(k(x))$ are the same. This will prove that $\pi_0^{\#A^1}(X)(F)$ is a singleton for every finitely generated extension $F/k$. We can then deduce by \cite[Lemma 6.1.3]{Morel-connectivity} that $X$ is $\#A^1$-connected. The strategy of the proof is to inductively construct a sequence of open subsets
\[
\phi=:U_{-1} \subset U_0 \subset U_1 \subset \cdots \subset X
\]
such that the following conditions hold:
\begin{itemize}
\item[(A)] For every $i \geq 0$, $Z_i:= U_i\backslash U_{i-1}$ is irreducible and $Z_i$ is non-empty unless $U_{i-1} = X$.
\item[(B)] For every $i \geq 0$, let $j_i, g_i \in \@X(Z_i)$ be defined as the compositions $Z_i \hookrightarrow X \to \@X$ and $ Z_i \to {\rm Spec \,} k \stackrel{x_0}{\to} X \to \@X$ respectively. Then, there exists a morphism $H_i: Z_i \times \Delta^1 \to \@X$ such that $H_i|_{Z_i \times \{0\}} = j_i$ and $H_i|_{Z_i \times \{1\}} = g_i$.
\end{itemize}
As $X$ is noetherian, this sequence of open sets is of finite length and must stabilize with $U_N = X$ for some integer $N$. Note that
$$\pi_0(\@X)(k(X)) = \mathop{\lim}_{U \subset X \text{ open}} \pi_0(\@X)(U).$$ Thus, there exists an open subset $U_0 \subset X$ such that the elements $j_0, g_0$ of $\@X(U_0)$, defined as in condition (B) above, are simplicially homotopic. Thus, we have a morphism of simplicial sheaves $H_0: U_0 \times \Delta^1 \to \@X$ connecting $j_0$ and $g_0$.
We now suppose that $U_{-1} \subset U_0 \subset \cdots \subset U_i$ have been chosen. Let $Z$ be a component of the closed subset $X \backslash U_i$, let $z$ be its generic point and let $L = k(z)$. We claim that the images of $z$ and $x_0$ in $\@X(L)$ are simplicially homotopic. One of the two conditions in the statement of the theorem must hold for $z$. If condition (a) holds, our claim is trivially true. Thus, we now assume that (a) does not hold for $z$ and hence, $z$ is movable.
Thus, there exists a smooth, irreducible curve $C$ over $L$, a point $c_0 \in C(L)$ and a non-constant morphism $\gamma: C \to X$ such that $\gamma(c_0) = z$ and the image of $\gamma$ is not contained in $Z$. Clearly, $\gamma$ maps the generic point of the curve $C$ into some $Z_{i_0}$ where $i_0 \leq i$.
Let $V_1 := \gamma^{-1}(U_{i_0}) \cup \{c_0\}$, which is an open subset of $C$. Choose an \'etale morphism $f_1: V_1 \to \#A^1_L$ such that $f(c_0) = 0$. By replacing $V_1$ by a smaller open neighbourhood of $c_0$ if necessary, we may also assume that $f_1^{-1}(0) = \{c_0\}$. Let $V_2 = \#A^1_L \backslash \{0\}$ and let $f_2: V_2 \to \#A^1_L$ be the inclusion map. Then $\{f_i:V_i \to \#A^1_L|i = 1,2\}$ is an elementary Nisnevich cover of $\#A^1_L$. Let $W:= V_1 \times_{\#A^1_L} V_2$. We observe that the morphism $W \to V_1$ is an open immersion.
Let $h_1: V_1 \to \@X$ be the morphism
\[
V_1 \hookrightarrow C \stackrel{\gamma}{\to} X \to \@X
\]
and let $h_2: V_2 \to \@X$ be the morphism
\[
V_2 \to {\rm Spec \,} k \stackrel{x_0}{\to} X \to \@X.
\]
The morphism $h_1|_W$ is the same as $H_{i_0}|_{U \times \{0\}} \circ \gamma|_W$ and the morphism $h_2|_W$ is the same as $H_{i_0}|_{U \times \{1\}} \circ \gamma|_W$. Thus, $H_{i_0} \circ \gamma|_W$ gives us a simplicial homotopy connecting $h_1|_W$ and $h_2|_W$.
Thus, we have the diagram
\[
\xymatrix{
\ast \ar[rr] \ar[d] && \@X(V_1) \ar[d] \\
\Delta^1 \ar[rr]_{H_{i_0} \circ \gamma|_W} \ar@{-->}[urr] && \@X(W)
}
\]
of simplicial sets, where the morphism on the right is a fibration since $W \hookrightarrow V_1$ is a cofibration and $\@X$ is simplicially fibrant. Thus, there exists a morphism from $\Delta^1$ to $\@X(V_1)$, indicated by the dotted arrow in the above diagram, making the diagram commutative. We denote this morphism by $H'$. Let $h'_1:= H'|_{V_1 \times \{1\}}$. Then $h'_1$ and $h_2$ can be glued together to define morphism from $\#A^1_L$ to $X$. This shows that the images of $z$ and $x_0$ in $\@X(L)$ are simplicially homotopic, concluding the proof of our claim.
This simplicial homotopy can be extended to a suitable open subset of $Z$. In other words, there exists an open subset $Z_{i+1}$ of $Z$ such that the following conditions hold:
\begin{itemize}
\item[(1)] The morphisms $j_{i+1}, g_{j+1}: Z_{i+1} \to \@X$, defined as $Z_{i+1} \hookrightarrow X \to \@X$ and $Z_{i+1} \to {\rm Spec \,} k \stackrel{x_0}{\to} X \to \@X$ respectively, are connected by a simplicial homotopy $H_{i+1}: Z_{i+1} \times \Delta^1 \to \@X$.
\item[(2)] $Z_{i+1}$ does not meet any component of $X \backslash U_i$ other than $Z$.
\end{itemize}
Condition (2) ensures that $Z_{i+1} \cup U_i$ is an open subset of $X$, which we denote by $U_{i+1}$. This completes the induction step.
\end{proof}
\begin{remark}
Let $X$ be a variety over $k$ and let $x, y \in X(k)$ be distinct points such that $x$ and $y$ map to the same element of $\pi_0^{\#A^1}(X)(k)$. Then $x$ must be a movable point of $X$. Indeed, let $\-X$ be a compactification of $X$. Then $\pi_0^{\#A^1}(\-X)(k) = \@S(\-X)(k)$, by \cite[Theorem 2.4.3]{Asok-Morel} or \cite[Theorem 2]{Balwe-Hogadi-Sawant}. Thus, $x$ and $y$ are $\#A^1$-chain connected on $\-X$. It follows from this that $x$ is a movable point of $X$.
This observation implies that if $X/k$ is $\#A^1$-connected and has more than one $k$-rational point, then every $k$-rational point must be movable. This allows us to see that Theorem \ref{theorem:criterion2} may not hold for a variety $X$ if some points of $X$ are not movable. A simple example is given by the curve $C$ in $\#P^2_{\#R}$ defined by the homogeneous polynomial $Y^2 - X^2(X-1)$. This is a rational curve, and so its generic point can certainly be connected to an $\#R$-valued point. However, it is easy to see that the point $(0:0:1)$ is not movable.
\end{remark}
A useful consequence of Theorem \ref{theorem:criterion2} for smooth proper varieties is that ${\mathbb A}^1$-connectedness can be characterized as the condition that the generic point of the variety can be connected to a $k$-rational point of it by a chain of $\P^1$'s (also see \cite[Theorem 8.5.1]{Kahn-Sujatha} for an equivalent treatment in terms of $R$-equivalence classes).
\begin{corollary}
\label{cor criterion}
A smooth proper variety $X$ over a field $k$ is ${\mathbb A}^1$-connected if and only if $X$ has a $k$-rational point and the generic point of $X$ can be connected to it (and hence, to any $k$-rational point of $X$) by an ${\mathbb A}^1$-chain homotopy.
\end{corollary}
\begin{proof}
By Theorem \ref{theorem:criterion2}, it follows that any smooth variety $X$ is ${\mathbb A}^1$-connected if and only if the generic point of $X$ has the same image in $\pi_0^{{\mathbb A}^1}(X)(k(X))$ as a $k$-rational point of $X$. If $X$ is proper in addition, the latter statement is equivalent to the statement that the generic point of $X$ can be connected to a $k$-rational point of $X$ by an ${\mathbb A}^1$-chain homotopy, by \cite[Theorem 2.4.3]{Asok-Morel} or \cite[Theorem 2]{Balwe-Hogadi-Sawant}.
\end{proof}
We end this section with two direct consequences of Corollary \ref{cor criterion}, which will be used to show ${\mathbb A}^1$-connectedness of norm varieties in Section \ref{section norm varieties}.
\begin{corollary}
\label{cor:fiberbase}
Let $k$ be a field of characteristic $0$ and $Y\xrightarrow{f} X$ be a generically smooth morphism of smooth proper varieties over $k$. If $X$ and the generic fiber of $f$ are ${\mathbb A}^1$-connected, then $Y$ is ${\mathbb A}^1$-connected.
\end{corollary}
\begin{proof}
Let $\eta$ be the generic point of $X$ and $Y_{\eta}$ be the generic fiber of $f$. Let $\xi$ be the generic point of $Y_{\eta}$. Note that this is also the generic point of $Y$.
Since $Y_{\eta}$ is ${\mathbb A}^1$-connected, it has a $k(\eta)$-rational point. In other words, the map $f$ has a rational section $X\stackrel{s}{\dashrightarrow}Y$. Since $k$ has characteristic $0$, we may resolve the singularities of the section and find a smooth proper variety $\tilde{X}$ over $k$ and a birational proper morphism $\tilde{X}\xrightarrow{\pi} X$ such that the rational section $s$ defines a morphism $\tilde{X}\xrightarrow{\~s} Y$.
\[
\xymatrix{
\tilde{X} \ar[rd]_\pi \ar[r]^{\~s} & Y \ar[d]^-f \\
& X\ar@{-->}@/_2pc/[u]_s
}
\]
Since $\tilde{X}$ is birational to $X$ and both $X, \~X$ are smooth, we see that ${\mathbb A}^1$-connectedness of $X$ implies that of $\tilde{X}$ by \cite[Theorem 3]{Asok-crelle}. Thus, its generic point $\eta$ is ${\mathbb A}^1$-chain homotopic to a $k$-rational point $x_0\in X(k)$, which can be chosen to lie in the open set over which $\pi$ is an isomorphism. Note that $\~s(\eta)$ is a $k(\eta)$-rational point of $Y_{\eta}$. Thus by applying $\~s$, we see that $\~s(\eta)$ is ${\mathbb A}^1$-chain homotopic to the $k$-rational point $\~s(x_0)$ in $Y$. However since $Y_\eta$ is an ${\mathbb A}^1$-connected variety over $k(\eta)$, we see that $\~s(\eta)$ is ${\mathbb A}^1$-chain homotopic to its generic point $\xi$, which is also the generic point of $Y$. Thus, the generic point of $Y$ is ${\mathbb A}^1$-chain homotopic to the $k$-rational point $\~s(x_0)$.
\end{proof}
\begin{remark}
Since $\mathbb A^1$-connectedness is equivalent to universal $R$-triviality, Corollary \ref{cor:fiberbase} is, in a sense, well-known to experts. Indeed, birational invariance of $R$-equivalence classes in a smooth proper variety has been already observed in \cite[Proposition 10]{Colliot-Thelene-Sansuc-1979} (in characteristic $0$) and in \cite[Corollary 6.6.6]{Kahn-Sujatha} (in arbitrary characteristic).
\end{remark}
\begin{corollary}\label{cor:open}
Let $k$ be a field and $U$ be a variety over $k$, which has a $k$-rational point $x_0\in U(k)$. Assume that for any field extension $F/k$ and any point $x \in U(F)$, there exists a rational map $h:\#P^1_F \dashrightarrow U$, defined at $0,1\in \#P^1_k$ such that
$h(0)=x$ and $h(1)=x_0$. Then any smooth proper model of $U$ is ${\mathbb A}^1$-connected.
\end{corollary}
\begin{proof}
Let $X$ be a smooth proper model of $U$ and $F=k(X)$ be its function field. Let $\eta \in U(F)$ be the generic point. By hypothesis, we get a morphism $\#P^1\dashrightarrow X$ connecting $x$ to $x_0$. Since $X$ is proper, this map extends to give a morphism, which in turn gives an ${\mathbb A}^1$-homotopy connecting $x$ to $x_0$. Theorem \ref{cor criterion} now implies that $X$ is ${\mathbb A}^1$-connected.
\end{proof}
\section{Compactifications of symmetric products of \texorpdfstring{${\mathbb A}^1$}{A1}-connected varieties}
\label{section symmetric}
The aim of this section is to show that any smooth proper compactification of the symmetric power of an ${\mathbb A}^1$-connected smooth projective variety is ${\mathbb A}^1$-connected. The proof of this result is quite tricky and relies on two geometric inputs. The first one is that any point on a rational smooth proper variety can be moved by a rational curve into an open subset that is isomorphic to the open subset of an affine space. The proof of this result, which uses jet schemes, is discussed in Section \ref{subsection approximation}. The second input, discussed in Section \ref{subsection symmetric powers}, is the construction of a particular ${\mathbb A}^1$-connected smooth proper compactification of the symmetric power of an ${\mathbb A}^1$-connected smooth projective variety. The ${\mathbb A}^1$-connectedness of any smooth proper compactification of the symmetric power of an ${\mathbb A}^1$-connected smooth projective variety then follows from birational invariance of ${\mathbb A}^1$-connectedness for smooth proper varieties.
Throughout this section, we will assume that $k$ is an algebraically closed field of characteristic $0$.
\subsection{Rational approximation of curves}
\label{subsection approximation}
Let $t$ be a variable and for any $n \geq 0$, set $\#D_n:= {\rm Spec \,} k[[t]]/\<t^{n+1}\>$. Let $\#D_{\infty}$ denote the scheme ${\rm Spec \,} k[[t]]$. For any variety $X$ over $k$ and any non-negative integer $n$, let $J_n(X)$ denote the sheaf $U \mapsto X(U \times \#D_n)$. It is well-known that (see \cite{Greenberg-schemata1}) for any finite $n$, the sheaf $J_n(X)$ is a scheme and is called the \emph{$m$-jet scheme} of $X$.
\begin{definition}
\label{defn parameterized pointed curve}
A \emph{parametrized pointed curve} is defined to be a triple $(C, c, \pi)$ where $C$ is a smooth curve, $c$ is a point of $C$ and $\pi$ is a uniformizing parameter of the ring $\@O_{C,c}$. There exists a unique isomorphism $\widehat{\@O}_{C,c} \xrightarrow{\simeq} \kappa(c)[[t]]$ such that $\pi \mapsto t$. Thus, if $(X,x)$ is a variety, a morphism $f: (C,c) \to (X,x)$ determines a $\kappa(c)$-valued point of $J_{\infty}(X)$ (which depends on $\pi$), which we denote by $f_{\infty}$. The image of $f_{\infty}$ in $J_n(X)$ will be denoted by $f_n$ for any non-negative integer $n$.
\end{definition}
\begin{definition}
\label{defn approximation}
We say that a variety $X$ over $k$ \emph{admits rational approximation of curves} if it has the property that for any point $x$ of $X$, non-negative integer $n$, parametrized pointed curve $(C, c, \pi)$ and morphism $f: (C,c) \to (X,x)$, there exists a morphism $g: (V,0) \to (X,x)$ for some open subset $U$ of ${\mathbb A}^1$ containing $0$ such that $f_n = g_n$ (here we use the parameter $T$ on $V \subset \#A^1 = {\rm Spec \,} k[T]$). We say that $g$ approximates $f$ to order $n$. We will abuse the notation and call such a $g$ a \emph{pointed rational curve} $(\#A^1, 0) \to (X,x)$.
\end{definition}
It is easy to see that the property of admitting rational approximation of curves is respected by blow-downs. Lemma \ref{lemma rational approximation} below shows that this property is also respected under blowups at smooth centers.
\begin{lemma}
\label{lemma rational approximation}
Let $X$ be a smooth variety which admits rational approximation of curves. If $\~X$ is a blowup of $X$ at a smooth center $Z$, then $\~X$ too admits rational approximation of curves.
\end{lemma}
\begin{proof}
We fix a point $x$ of $X$. Let $n$ be a non-negative integer. We prove that for any parametrized pointed curve $(C,c, \pi)$ and morphism $f: (C,c) \to (X,x)$, there exists an integer $m$ such that if $g:(\#A^1, 0) \to (X,x)$ approximates $f$ to order $n + m$, then the lift of $g$ to $\~X$ approximates the lift of $f$ to order $n$. Clearly, this will imply the result. Without any loss of generality, we will assume that $\kappa(c) = \kappa(x) = k$.
If $x \notin Z$, the result is trivial. Thus, we now assume that $x$ is in $Z$. Choose a system of parameters $\pi_1, \ldots, \pi_d$ at $x$ (where $d = \dim(X)$). By replacing $X$ with an open neighbourhood if necessary, we may assume that the $\pi_i$ are regular on $X$ and that $Z$ is defined by the ideal $\<\pi_1, \ldots, \pi_p\>$ for some $p \leq d$. Thus, $\~X$ is the closed subscheme of $X \times \#P^{p-1}$ defined by the equations $T_i \pi_j = T_j \pi_i$ for $1 \leq i,j \leq p$, where $T_1, \ldots, T_p$ denote homogeneous coordinates on $\#P^{p-1}$.
For $1 \leq i \leq d$, let $\alpha_i = f_{\infty}^*(\pi_i)$. Let $i_0$ be such that ${\rm ord}_t(\alpha_{i_0}) \leq {\rm ord}_t (\alpha_i)$ for $1 \leq i \leq p$. Without loss of generality, we may take $i_0 = 1$. We set $m:= {\rm ord}_{t}(\alpha_1)$. Since $X$ admits rational approximation of curves, there exists a pointed rational curve $g: (\#A^1,0) \to (X,x)$, which approximates $f$ to order $n + m$. For $1 \leq i \leq d$, let $\beta_i = g_{\infty}^*(\pi_i)$. Since $g$ approximates $f$ to order $n + m$, we have $\alpha_i \equiv \beta_i \mod(t^{m + n + 1})$ for all $i$. Thus, ${\rm ord}_t(\beta_1) = m \leq {\rm ord}_t(\beta_i)$ for $1 \leq i \leq p$.
Let $\~f$ and $\~g$ denote the lifts of $f$ and $g$ to $\~X$. Then $\~f$ maps the closed point of $\#D$ to a point $\~x$ in the open subset of $\~X$ defined by $T_1 \neq 0$. Let us denote the generic point of $\#D$ by $\eta_0$ and the closed point by $\eta_1$. Then
\[
\~f(\eta_0) = (f(\eta_0), (\alpha_1: \cdots : \alpha_p)) = (f(\eta_0), (\gamma_1: \cdots : \gamma_p))\in (X \times \#P^{p-1})(k((t))) \text{,}
\]
where $\gamma_i = \alpha_i/\alpha_1 \in k[[t]] \subset k((t))$. We also have
\[
\~f(\eta_1) = (f(\eta_1), (\gamma_1(0): \cdots : \gamma_p(0)) \in (X \times \#P^{p-1})(k),
\]
where $\gamma_i(0)$ is the image of $\gamma_i$ under the quotient homomorphism $k[[t]] \to k[[t]]/\<t\> = k$. As $\gamma_1 = 1$, the point $\~f(\eta_1)$ lies in the open subscheme of $\~X$ defined by the condition $T_1 \neq 0$. Similarly, we have
\[
\~g(\eta_0) = (g(\eta_0), (\beta_1: \cdots : \beta_p)) (g(\eta_0), (\delta_1: \cdots : \delta_p))\in (X \times \#P^{p-1})(k((t))),
\]
where $\delta_i = \beta_i/\beta_1 \in k[[t]]$. Also,
\[
\~g(\eta_1) = (g(\eta_1), (\delta_1(0): \cdots : \delta_p(0)) \in (X \times \#P^{p-1})(k).
\]
Again, we observe that $\delta_{1} = 1$ and so the point $\~g(\eta_1)$ lies in the open subcheme $T_1 \neq 0$.
In fact, since $\gamma_i \equiv \delta_i \mod(t^{n+1})$ for all $i$, we have $\~f(\eta_1) = \~g(\eta_1)$. At the point $\~f(\eta_1)$, the functions
\[
\pi_1, T_2/T_1, \ldots, T_p/T_1, \pi_{p+1}, \ldots , \pi_d
\]
form a system of parameters. Since $\alpha_1 \equiv \beta_1 \mod(t^{m+n+1})$, $\gamma_i \equiv \delta_i \mod(t^{n+1})$ for $2 \leq i \leq p$ and $\alpha_j \equiv \delta_j \mod(t^{m+n+1})$ for $p \leq j \leq d$, we conclude that $\~g$ approximates $\~f$ to order $n$.
\end{proof}
\begin{proposition}
\label{prop rational approximation}
Any smooth rational variety admits rational approximation of curves.
\end{proposition}
\begin{proof}
It is clear that the projective space admits rational approximation of curves. Now, the result follows from Lemma \ref{lemma rational approximation} and the weak factorization theorem \cite[Theorem 0.1.1]{AKMW}.
\end{proof}
\begin{corollary}
Let $X$ be a smooth projective variety and $U \subset X$ be an open subset. Then, for any point $x \in X \backslash U$, there exists a rational curve through $x$ which intersects $U$.
\end{corollary}
\begin{proof}
Let $Z = X \backslash U$. Choose a pointed pointed parametrized curve $(C,c,\pi)$ with $\kappa(c) = \kappa(x)$, and a morphism $f: (C,c) \to (X,x)$ such that $f(C)$ intersects $U$. If we have $f_n \in J_n(Z)(\kappa(x))$ for all $n \geq 0$, then we must have $f_{\infty} \in J_{\infty}(Z)$. Thus, $f$ must map the generic point of $C$ into $Z$, and hence $f(C) \subset Z$, which is a contradiction. Thus, there exists some positive integer $n$ such that $f_n \notin J_n(Z)(\kappa(x))$. Since $X$ admits rational approximation of curves, there exists a pointed rational curve $g: (\#A^1,0) \to (X,x)$ which approximates $f$ to order $n$. Then, $g(\#A^1)$ must also intersect $U$.
\end{proof}
\subsection{\texorpdfstring{${\mathbb A}^1$}{A1}-connectedness of symmetric powers}
\label{subsection symmetric powers}
\begin{notation}\label{notation}
For an integer $m\geq 2$, we let ${\rm Sym}^m X$ denote the $m$-th symmetric power of $X$, that is, the quotient of $X^m$ by the action of the symmetric group $S_m$. Note that ${\rm Sym}^m X$ is singular in general. Let $\Delta$ denote the union of all the partial diagonals of $X^m$. The geometric quotient $U:= (X^m \setminus \Delta)/S_m$ is a smooth scheme. We let $\eta$, $\xi$ and $\theta$ denote the generic points of $X$, $X^m$ and ${\rm Sym}^m X$, respectively. Let $K:=k(\eta)$ and $E:=k(\xi)$ and $L:=k(\theta)$. Since $X$ is geometrically integral, it follows from the Grothendieck-Sharp theorem (see \cite[Remarque (4.2.1.4)]{EGA4-IV}, \cite[Theorem]{Sharp}) that ${\rm Spec \,} K^{\otimes m}$ is an integral domain and its function field is isomorphic to $E$.
Let $F$ be any field extension of $k$ in which $k$ is algebraically closed. Let $F'$ denote the function field of ${\rm Spec \,} F^{\otimes m}$. For any $F$-rational point ${\rm Spec \,} F \xrightarrow{p} X$, we denote by $p^m$ the corresponding ${\rm Spec \,} F'$-valued point of $X^m$.
\end{notation}
Given an ${\mathbb A}^1$-connected smooth projective variety $X$ over $k$, we first construct an explicit smooth proper compactification of ${\rm Sym}^m X$, which is ${\mathbb A}^1$-connected. This compactification will be constructed by first considering a particular blowup $\~X$ of $X$ and then taking a suitable compactification of ${\rm Sym}^m \~X$. Lemma \ref{lemma pre-arrange-x} below gives the required $\~X$.
\begin{lemma}
\label{lemma pre-arrange-x}
Let $X$ be a smooth projective ${\mathbb A}^1$-connected variety over an algebraically closed field $k$ of characteristic $0$. Let $x_0\in X(k)$. Let $M$ be a finitely generated field extension of $k$ and let $C$ be a smooth curve over $M$ with a point $0\in C(M)$. Let $f: C \to X$ be a morphism with $f(0)=x_0$. Further, assume that $k$ is algebraically closed in $M$ and that the image of generic point of $C$ in $X$ is a point of dimension $\geq 2$. Then there exists a proper birational morphism $X'\xrightarrow{\alpha} X$ such that
\begin{enumerate}[label=$(\arabic*)$]
\item $\alpha$ is obtained by successively blowing up $k$-rational points lying over $x_0$;
\item $x'_0:=f'(0)$ is a $k$-rational point, where $f'$ is the lift of $f$ to $X'$; and
\item if $\tilde{X}\xrightarrow{\pi} X'$ is the blowup of $X'$ at $x'_0$, then $\tilde{f}(0)$ is a positive-dimensional point of $\tilde{X}$, where $\tilde{f}$ is the lift of $f'$ to $\tilde{X}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $f_M:C\to X_M$ be the induced map obtained by the base change of $X$ to $M$. By replacing $C$ with the normalization of the image of $f_M$, we may assume without loss of generality that $C$ is the normalization of its image.
We first claim that the image of $f_M$ cannot be normal at $f_M(0)$. To see this, we first \emph{spread} the morphism $f$, that is, consider the diagram
$$\xymatrix{
C \ar[r]\ar[d]\ar@/^2pc/[rr]^f & \@C \ar[r]_-{\hat{f}}\ar[d] & X \\
{\rm Spec \,} M \ar@/^1.5pc/[u]^0 \ar[r]^-\eta & B\ar@/^1.5pc/[u]^s &
}$$
in which
\begin{itemize}
\item $B$ is an integral $k$-scheme and $\eta:{\rm Spec \,} M \to B$ is the generic point of $B$;
\item $\@C \to B$ is smooth curve with a section $s$; and
\item the square is cartesian, so the curve $(C,0)$ is the generic fiber of $\@C\to B$.
\end{itemize}
Since $f(0)$ is a $k$-rational point, $\hat{f}\circ s$ is a constant map. If the image $f_M(C)$ is normal at $f_M(0)$, then since $C$ also happens to be the normalization of its image, we deduce that $f_M$ is an immersion in a neighbourhood of $0$. By a standard limiting argument, we can conclude that there exists an open set $U\subset B$ and an open neighbourhood $\@C^0 \subset \@C_U$ of the image of the section $s(U)$, such that $(\hat{f}|_{\@C^0})_M: \@C^0 \to X_{M}$ is an immersion. However, this is impossible if $\hat{f}\circ s$ is a constant map, since $\dim(s(U)) = \dim(B)\geq 2$ by assumption. This shows that if $f_M(C)$ is normal at the image of $0$, then $f(0)$ cannot be a $k$-rational point.
Let $\tilde{X}\to X$ be the blowup of $X$ at $f(0)$. The map $f$ lifts uniquely to give a map $\tilde{f}:C \to \tilde{X}$. If $\tilde{f}(0)$ is a positive dimensional point, then we are done. Assume the contrary; then $\tilde{f}(0)$ has to be a $k$-rational point, since $k$ is algebraically closed in $M$. Note that the map $\tilde{X}_M \to X_M$ is also the blowup of $X_M$ at $f_M(0)$. We may replace $X$ with $\tilde{X}$ without loss of generality. However, by the embedded desingularization of a curve, this process can only be repeated finitely many times until $\tilde{f}_M(C)$ becomes regular at $\tilde{f}_M(0)$ in which case, $\tilde{f}(0)$ cannot be a $k$-rational point as observed in the above paragraph. This completes the proof of the lemma.
\end{proof}
\begin{lemma}\label{be_positive_be_free}
We follow the notation described in Notation \ref{notation}. Let $K$ be the function field of a variety $X$ over $k$ and let $p\in X(K)$ be a $K$-rational point, which is not in the image of $X(k)$. Then the image of $p^m$ is not contained in $\Delta$.
\end{lemma}
\begin{proof}
Let $Z$ be the closure of $p$ (that is, the image of $p$). Since $K$ is the function field of a $X$, which is a geometrically irreducible variety, we see that $Z$ is geometrically irreducible. In particular, it is zero dimensional if and only if it is a $k$-rational point. However, by hypothesis, $p$ is not $k$-rational. Hence $\dim(Z)>0$. The closure of the image of $p^m$, that is, the closure of the image of
$${\rm Spec \,}(E) \xrightarrow{\simeq} {\rm Spec \,}(K)^m \xrightarrow{p^m} X^m$$
is precisely $Z^m$. Since $\dim(Z)>0$, it follows that $Z^m$ is not contained in any partial diagonal in $X^m$.
\end{proof}
\begin{lemma}\label{lemma singlep1}
Let $X, K, E, L$ be as in Notation \ref{notation}. Let $p,q \in X(E)$. Let $\phi: \P^1_K \to X$ be a morphism with $\phi(0) = p$ and $\phi(1)=q$. Then there exists a morphism $$\psi:\P^1_L \to S^mX$$ connecting the images of $p^m$ and $q^m$ in $S^m(X)$.
\end{lemma}
\begin{proof}
Note that $E$ is the function field of the integral scheme ${\rm Spec \,} K^{\otimes m}$. Note that $S_m$ acts on ${\rm Spec \,} K^{\otimes m}$ by permutation of factors and hence we have an induced $S_m$-action on $E$. We have an $S_m$-equivariant isomorphism
\[
(\P^1_K)^m \xrightarrow{\simeq} (\P^1_k)^m \times_k E,
\]
where the $S_m$-action on the right is the diagonal action. Taking $m$-fold product of the morphism $\phi$, we get an $S_m$-equivariant morphism
\[
\phi^m: (\P^1_k)^m \times_k E \cong (\P^1_K)^m \to X^m.
\]
Now precomposing the above map with the diagonal $\P^1_k \times_k E \xrightarrow {\Delta \times {\rm id}_{E}} (\P^1_k)^m \times_k E$, we get an $S_m$-equivariant morphism
\[
\P^1_k \times_k E \to X^m.
\]
We obtain the desired morphism $\psi$ after taking quotient by $S_m$.
\end{proof}
We are now set to prove the main result of this section.
\begin{theorem}
\label{theorem:symmetric_model}
Let $X$ be a smooth projective ${\mathbb A}^1$-connected variety over an algebraically closed field $k$ of characteristic $0$ and fix a positive integer $m$. Then any smooth proper variety birational to ${\rm Sym}^m X$ is ${\mathbb A}^1$-connected.
\end{theorem}
\begin{proof}
Since ${\mathbb A}^1$-connectedness is a birationally invariant property of smooth proper varieties \cite[Theorem 3.9]{Asok-crelle}, it suffices to obtain one smooth proper ${\mathbb A}^1$-connected variety birational to ${\rm Sym}^m X$.
Since $X$ is ${\mathbb A}^1$-connected, there exists an ${\mathbb A}^1$-chain homotopy connecting a $k$-rational point $x_0\in X(k)$ to the generic point $\eta$ of $X$ by Corollary \ref{cor criterion}. Since $X$ is proper, this ${\mathbb A}^1$-chain homotopy gives rise to a morphism
\[
f :C = \bigcup^r_{i=1} \P^1_L \xrightarrow{\cup f_i} X
\]
satisfying $f_1(0)=x_0$, $f_i(1) = f_{i+1}(0)$ for $1 \leq i \leq r-1$, and $f_r(1)=\eta$. Applying Lemma \ref{lemma pre-arrange-x}, we can find a blowup $\pi:\tilde{X}\to X$ (obtained by successively blowing up points over $x_0$) and $\tilde{f_i}$. If we denote the lift of $f_i$ to $\tilde{X}$ by $\tilde{f}_i$ for every $i$, we see by Lemma \ref{lemma pre-arrange-x} that $\tilde{x}_0:=\tilde{f_1}(0)$ is a positive dimensional point of $\~X$. Since $\pi:\tilde{X}\to X$ is an isomorphism outside $x_0$, the lifts $\tilde{f_i}$ glue to give a well-defined map $\tilde{f}: C \to \tilde{X}$ lifting $f$. Thus, $\tilde{f}$ is a chain of rational curves in $\~X$ connecting $\tilde{x}_0$ to $\eta$. Note that $\tilde{x}_0$ lies in the exceptional divisor $E$ of $\pi$.
We now apply Lemma \ref{lemma singlep1} to each of the $\tilde{f_i}$ and get a chain of rational curves defined over the generic point $\theta$ of ${\rm Sym}^m\tilde{X}$. This chain of curves joins $\tilde{x}_0^m$ to $\theta$. All nodes of this chain are in the free locus of ${\rm Sym}^m X$, thanks to Lemma \ref{be_positive_be_free}. The point $\tilde{x}_0^m$ is also in the free locus by the same argument. In addition, $\tilde{x}_0^m \in {\rm Sym}^m E$ (considered as a subscheme of ${\rm Sym}^m \tilde{X}$).
Note that $E$ is a projective space and consequently, ${\rm Sym}^m E$ is rational by \cite{Mattuck} (see also \cite[Chapter 4, Theorem 2.8]{GKZ}). Let $U$ be an open subscheme of ${\rm Sym}^m E$ that is isomorphic to an open subscheme of an affine space. Therefore, any point of $U(F)$ can be connected by an ${\mathbb A}^1$-homotopy to a point of $U(F)$ lying in the image of the natural map $U(k) \to U(F)$. If $\tilde{x}_0^m \in U$, then we are done. If not, we apply Proposition \ref{prop rational approximation} to obtain an ${\mathbb A}^1$-homotopy $h: {\mathbb A}^1 \to {\rm Sym}^m E$ connecting $\tilde{x}_0^m$ to a point in $U$. We thus get a chain of rational curves in ${\rm Sym}^m \tilde{X}$, which joins a $k$-rational point in the free locus of ${\rm Sym}^m \tilde{X}$ to the generic point of ${\rm Sym}^m \tilde{X}$, and all whose nodes are in the free locus.
Let $Y\to {\rm Sym}^m \tilde{X}$ be a desingularization which is an isomorphism over the smooth locus of ${\rm Sym}^m \tilde{X}$. Such a desingularization exists since $k$ is assumed to be of characteristic $0$. Then the chain of rational curves found in the above paragraph lifts to give a chain of rational curves in $Y$. This gives an ${\mathbb A}^1$-chain homotopy connecting a $k$-rational point of $Y$ to its generic point. Hence, $Y$ is ${\mathbb A}^1$-connected by Corollary \ref{cor criterion}. Since $Y$ is also a desingularization of ${\rm Sym}^m X$, this completes the proof of the theorem.
\end{proof}
\section{\texorpdfstring{${\mathbb A}^1$}{A1}-connectedness of norm varieties}
\label{section norm varieties}
In this section, we use the results of Sections \ref{section criterion} and \ref{section symmetric} to prove Theorem \ref{intro thm: norm varieties} stated in the introduction. We begin with some preliminaries on norm hypersurfaces.
\begin{definition}
\label{definition norm hypersurface}
Let $F$ be a field and let $K/F$ be a finite field extension of degree $n$. If we pick a basis for the $F$-vector space $K$, the norm function $N_{K/F}: K \to F$ is seen to be given by a polynomial in $n$ variables. Thus, there exists a homomorphism of algebraic groups $N: R_{K/F}(\#G_{m,K}) \to \#G_{m,F}$ such that $N_{K/F}$ describes the behaviour of $N$ on $F$-valued points. For any $a \in F^{\times} = \#G_m(F)$, we denote the scheme $N^{-1}(a)$ by $H_{K,a}$ and call it the \emph{norm hypersurface} defined by the extension $K/F$ and the element $a \in F^{\times}$.
\end{definition}
Clearly, $H_{K,1}(F) \neq \emptyset$. For any $x \in K^{\times}$, the morphism $m_x: \#G_{m,K} \to \#G_{m,K}$, $y \mapsto xy$ induces a $k$-automorphism of $R_{K/k}(m_x)$ of $R_{K/F}(\#G_{m,K})$. This automorphism maps $H_{K,1}$ isomorphically onto $H_{K,N_{K/F}(x)}$. Note that an element $x$ of $K^{\times}$ corresponds to a $F$-rational point of $R_{K/F}(\#G_{K,m})$. Thus, we see that if $H_{K,a}(F) \neq \emptyset$, then $H_{K,a}$ is isomorphic to $H_{K,1}$.
\begin{proposition}
\label{prop:norm-hypersurface}
Let $\ell$ be a prime number. Let $K/F$ be a separable extension of a field $F$ and let $L/F$ be its Galois closure. Suppose that ${\rm Gal}(L/F) = S_{\ell}$, the symmetric group on $\ell$ letters, and that ${\rm Gal}(L/K)$ is the stabilizer of one of the letters (so that ${\rm Gal}(L/K) \cong S_{\ell-1}$). If $a\in F^{\times}$ is such that $H_{K,a}(F) \neq \emptyset$, then any birational smooth proper model $X$ of $H_{K,a}$ is ${\mathbb A}^1$-connected.
\end{proposition}
\begin{proof}
As we saw above, the condition $H_{K,a}(F) \neq \emptyset$ implies that $H_{K,a} \cong H_{K,1}$. Thus, we may assume that $a = 1$. By \cite[Theorem 4.1]{Endo}, $H_{K,1}$ is retract rational. Now, it follows from \cite[Theorem 2.3.6]{Asok-Morel} that any smooth proper model of $H_{K,1}$ is $\#A^1$-connected.
\end{proof}
We are now set to put all the ingredients together and give a proof of Theorem \ref{intro thm: norm varieties} stated in the introduction. We start with a brief recollection of the construction of norm varieties given in \cite[\textsection 2]{Suslin-Joukhovitski}.
Let $X$ be a smooth projective variety over a field $k$ of characteristic $0$ and let $\ell$ be a prime number. Let ${\rm Sym}^{\ell} X = X^\ell/S_\ell$ be the $\ell$th symmetric power of $X$, which need not be smooth in general. Let $\Delta$ denote the union of all the partial diagonals of $X^{\ell}$. The geometric quotient $U:= (X^{\ell} \setminus \Delta)/S_{\ell}$ is a smooth scheme. There is a finite, surjective morphism $\pi: X \times {\rm Sym}^{\ell-1} X \to {\rm Sym}^{\ell} X$ of degree $\ell$ and its restriction $\pi^{-1}(U) \to U$ is a finite \'etale morphism of degree $\ell$. The sheaf $$\@E:=\pi_*(\@O_{X \times {\rm Sym}^{\ell-1}X})|_U$$ is a locally free sheaf of $\@O_U$-algebras of rank $\ell$. Let $\#V(\@E)$ denote the associated vector bundle on $U$. Since $\@E$ is a locally free sheaf of $\@O_U$-algebras, there is a well-defined norm morphism
\[
N: \@E \to \@O_U,
\]
which can be viewed as a section of the degree $\ell$ component ${\rm Sym}^{\ell} \@E^{\vee}$ of the symmetric algebra on the dual of $\@E$.
Any element $a \in k^{\times}$ can be viewed as a section of $\@O_U$. Let $Y$ denote the closed subscheme of $\#V(\@E)$ defined by the equation $N = a$. Then, $Y$ is smooth over $U$ and geometrically irreducible, by \cite[Lemma 2.1]{Suslin-Joukhovitski}. Define $N(X, a, \ell)$ to be a smooth compactification of $Y$, which exists since $k$ has characteristic $0$. We may further arrange that $N(X, a, \ell)$ is proper and surjective over a smooth compactification of $U$ with smooth generic fiber, by resolution of indeterminacy and resolution of singularities. In other words, $N(X, a, \ell)$ admits a surjective proper morphism to a smooth proper variety birational to ${\rm Sym}^{\ell} X$. This observation will be crucial to our proof of Theorem \ref{intro thm: norm varieties} (see Theorem \ref{theorem:norm varieties} below).
\begin{definition}
\label{def:norm variety}
Let $n \geq 2$ be an integer. Let $a_1, \ldots, a_n \in k^{\times}$ and consider the symbol $\alpha := \{a_1, \ldots, a_n\} \in \KM_n(k)/\ell$. The \emph{norm variety} $X_\alpha$ associated with the symbol $\alpha$ is inductively defined as follows. Define $X_{\{a_1, a_2\}}$ to be the Severi-Brauer variety associated with the cyclic algebra corresponding to $\{a_1, a_2\}$ (see \cite[Chapter 5]{Gille-Szamuely}). Inductively proceeding, we define
\[
X_\alpha = X_{\{a_1, \ldots, a_n\}} := N(X_{\{a_1, \ldots, a_{n-1}\}}, a_n, \ell),
\]
for $n>2$.
\end{definition}
\begin{theorem}
\label{theorem:norm varieties}
Let $k$ be an algebraically closed field of characteristic $0$. Let $n\geq 2$ be an integer and $a_1, \ldots, a_n \in k^{\times}$. Then the norm variety $X_\alpha$ associated with the ordered sequence $\alpha = \{a_1, \ldots, a_n\}$ is ${\mathbb A}^1$-connected.
\end{theorem}
\begin{proof}
We proceed by induction on $n$. As noted in the introduction, in the case $n=2$, the norm variety $X_\alpha$ is the Severi-Brauer variety associated with the cyclic algebra corresponding to the symbol $\{a_1, a_2\} \in \KM_2(k)/\ell$. In this case, $X_\alpha$ is clearly ${\mathbb A}^1$-connected as it is isomorphic to a projective space.
Now, let $n>2$ and assume by induction that $X = X_{\{a_1, \ldots, a_{n-1}\}}$ is ${\mathbb A}^1$-connected. Recall from the construction of $X_\alpha$ that it admits a surjective proper morphism to a smooth proper variety birational to ${\rm Sym}^{\ell} X$. The generic fiber of this morphism is a smooth compactification of the norm hypersurface $N=a_n$ over $k({\rm Sym}^\ell X)$. Since $k$ is algebraically closed and $a_n \in k^\times$, this generic fiber admits a $k({\rm Sym}^\ell X)$-rational point. We can apply the criterion given by Corollary \ref{cor:fiberbase} to this morphism to conclude that $X_\alpha$ is ${\mathbb A}^1$-connected. Note that the hypotheses of Corollary \ref{cor:fiberbase} are satisfied by Proposition \ref{prop:norm-hypersurface} and Theorem \ref{theorem:symmetric_model}.
\end{proof}
|
\section{Introduction}
\label{sec:intro}
Short video has been one of the most important social media and entertaining fashion for people worldwide recently. As millions of short videos are uploaded to the platforms by users each day, requirement of generic object re-identification from a large amount of short videos increases rapidly.
However, building a generic object re-identification system for short videos poses a number of challenges due to characteristics of short videos as follows:
\begin{itemize}
\item \textbf{Great variety in pattern and object category:} Most short videos in short video platforms are uploaded by users. Topics of video vary a lot according to interest of users, and numerous categories of objects may occur in the same video simultaneously.
\item \textbf{Various visual effects (VFX):} Visual effects are common when different video shots switching or emphasizing objects in short videos. These visual effects may cause color change of video frame, object distortion, etc.
\item \textbf{Complicated motions and postures:} Different from scenarios such as traffic monitoring, motion and posture of people in short videos are much more unpredictable. Moreover, frequent strenuous motion may also bring great burden to object detection and tracking.
\item \textbf{Rapid changes in object appearance:} Different from traffic scenarios in most person re-identification tasks, people dress differently in different short videos. Even more, some short videos are consist of multiple video clips (or shots), and appearance of the same object changes rapidly when the video shot switches, such as in dress switching videos.
\end{itemize}
As a result, short video has become one of the most challenging scenarios in recent years.
To extract the trajectories of objects in videos, there exist a series of studies on video object detection and video object tracking. However, most of the public benchmark datasets only focus on relatively few categories. For example, ImageNet VID \cite{ILSVRC15}, which is most frequently used for evaluation in video object detection, only contains 30 categories, most of which are animals and vehicles. By contrast, there are hundreds of various categories, complicated visual effects (VFX) and multiple video shots in short video scenario, which is opposite to the scenario that most recent approaches assumed.
In order to satisfy the mentioned challenge of generic object re-identification in short video scenarios, a novel system composed of a detection module, a tracking module, and a generic object re-identification module is proposed in this paper. In this framework, the trajectories of the major objects in the input query videos and their features are extracted by the three modules in the proposed system. As follow-up applications, the processed features can be applied in short video identical object retrieval tasks. The flowchart of the proposed system is illustrated in Figure \ref{fig:system}.
In particular, in order to handle the video object detection task with hundreds of categories, we build our detection module based on an image object detector, and propose a Temporal Information Fusion Network (TIFN) to take full advantage of temporal information among continuous frames. Further, a Cross-Layer Pointwise Siamese Network (CPSN) is proposed in the tracking module to mitigate the fragmented issue of tracklets, enhancing the robustness to match the same target despite the variety in appearance of objects in short videos.
To evaluate the proposed system, two challenge datasets are built for short video object trajectory extraction and generic object re-identification respectively. Extensive experiments for proposed methods and the whole proposed system are conducted with both the proposed short video datasets and public benchmark datasets, and the results demonstrate the high effectiveness and efficiency of the proposed system.
The contributions of this paper are summarized as follows:
\begin{itemize}
\item We design a generic object re-identification system for short videos. To the best of our knowledge, it is the first work to tackle the generic object re-identification problem under short video scenario, which is much more complicated and unpredictable in variety than regular benchmarks.
\item We propose TIFN, which takes full advantage of temporal information in videos to improve image object detector, and shows comparable accuracy and improved time efficiency to the state-of-the-art video object detector.
\item We propose CPSN, which makes full use of the local features in different scales, and excludes the disturb of the background and rapidly changing regions of foreground. This module can thus effectively reduce the fragmented issue of tracklets.
\item Two challenging datasets for short video object trajectory extraction and generic object re-identification are built, which have many potential applications in follow-up study of short video.
\end{itemize}
The rest of the paper is organized as follows. Section 2 discusses related work. The proposed framework and methods are detailedly described in Section 3. Proposed datasets introduction and experimental evaluations are shown in Section 4. Finally, Section 5 concludes this paper.
\section{Related Work}
Building a system to handle generic object re-identification task among different short videos includes three primary sections: a video object detection module to locate objects, a video object tracking module to extract object tracklets in each shot, and a generic object re-identification module to retrieve tracklets of the same object in different shots to a complete trajectory. Therefore, the related work of the proposed can be concluded as follows.
{\bf Video Object Detection:} Video object detection (VOD) focuses on temporal information between adjacent frames. Combining detection and tracking, \cite{kang2017t} propose a multi-phase framework, including image object detection, bounding box tracking and temporal convolutional network to re-scoring tubelet, to obtain enhanced performance. \cite{luo2019detect} propose a scheduler network, which determines to detect or track at a certain frame. Without aid of tracking module, \cite{zhu2017flow} improve the per-frame features by aggregation of nearby features along the motion paths obtained by optical flow. Innovatively, \cite{wang2018fully} jointly calibrate the features of objects on both pixel-level and instance-level by optical flow, capturing detailed motion and global motion features. Instead of the time-consuming optical flow, \cite{Lu_2017_ICCV} use LSTM to extract temporal motion information. \cite{xiao2018video} propose a novel RNN architecture called the Spatial-Temporal Memory Network (STMM) to model both changing appearance and motion of objects over time. Considering the full-sequence level features, \cite{wu2019sequence} devise Sequence Level Semantics Aggregation (SELSA) module to obtain more discriminative and robust features for video object detection. \cite{chen2020memory} take full consideration of both global and local information and enlarge range of accessible content of the key frame by introducing a Long Range Memory (LRM) module. However, all the methods mentioned above either associate temporal information in frames by optical flow and RNN, or formed as a complicated and time-consuming framework. Moreover, all of them need training dataset where bounding boxes in every frame are well-annotated, which is difficult for labelling when categories increase. In contrast, the Temporal Information Fusion Network (TIFN) we propose is based on fast one-stage detector without extra supervision, so it can be trained with dataset for image object detection.
{\bf Video Object Tracking:} Most of the recent multiple object tracking methods only focus on tracking single category, such as pedestrians or vehicles.
A standard tracking system can be divided into four stages: detection stage, feature extraction stage, affinity stage and association stage, while recent works pay more attention to the second and third stages. DeepSort \cite{wojke2017simple} integrates the deep neural network to learn appearance information, and obtains improved performance compared with previously published approach SORT \cite{bewley2016simple}. DAN \cite{sun2019deep} combines the feature extraction stage and affinity stage together to learn compact, yet comprehensive features in an end-to-end manner. Unlike most methods which use the off-the-shelf detector to detect region of interest, Tracktor \cite{bergmann2019tracking} creates a new tracking paradigm, in which a single detector solve most of tracking problems.
Unfortunately, most of these methods can only handle the pedestrian tracking problems under the traffic surveillance scenarios, which are not applicable to generic objects. In this paper, we propose a tracking system to adapt to generic object tracking for short videos.
{\bf Re-identification:} Recently most re-identification (ReID) methods focus on person retrieval. A standard person re-identification system contains three main components: feature representation learning, deep metric learning and ranking optimization. Feature representation in ReID can be furtherly divided into global feature and local feature. Due to the complicated scenario in person ReID, local features such as partial region feature, human pose and landmark feature, are becoming principal components, rather than global features. \cite{zhang2017alignedreid} extract global feature which is jointly learned with local partial region features to enhance global feature learning. \cite{Zhao_2017_CVPR} employ human pose estimation and landmark detection in person re-identification. \cite{sun2018beyond} target in learning discriminative part-informed features, and propose Part-based Convolutional Baseline (PCB) and refined part pooling (RPP), resulting in refined parts with enhanced within-part consistency.
However, most methods mentioned above are not designed for short video scenarios. Due to the challenging characteristics of short videos described in Section \ref{sec:intro}, existing methods are difficult to be applied to generic object re-identification for short videos directly. Considering numerous categories, various VFX, multiple video shots and rapid appearance changes, we propose a novel system to handle generic object re-identification for short videos in this paper. To our knowledge, it is the first system to tackle generic object re-identification problem for short video, which is one of the most challenge task in recent years.
\section{Proposed Method}
\noindent As is shown in Figure \ref{fig:system}, a generic object re-identification system for short videos can be formulated as follows. Given a sequence of video frames $F_q = \{f_t\}_{t=1}^T$ which contains $N$ objects and $K$ shots, the goal is to obtain the trajectories of the major objects $\{{\rm traj}_n\}_{n \in {\rm major}(N)}$ and use their features $\{{\rm feat}_n\}_{n \in {\rm major}(N)}$ to retrieve other videos $\{F_g\}$ with identical objects in gallery, where ${\rm major}(N)$ denotes the major objects of the query video. Shots in the query video are denoted as $S = \{s_k\}_{k=1}^K$, where $s_k = \{f_t\}_{t=t^*}^{t^*+{\rm length}(s_k)-1}$ denotes the $k$th shot in video which started at $t=t^*$.
Considering the complex scenario in short video and robustness of algorithm, we decompose the generic object re-identification for short videos problem into several sub-problems, and formulate them as an end-to-end system, which consists of three major modules: detection module, tracking module and re-identification module (colored in Figure \ref{fig:system}). An algorithm summary follows the introduction of the three modules.
\subsection{Detection Module}
\noindent In order to extract feature embedding of each object in video accurately, all the candidate objects need to be localized at first. The aim of detection module is to detect bounding boxes $B^t=\{b_n^t\}_{n=1}^N$ of the $N$ objects in each frame $f_t$.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{network_tifn.pdf}
\vskip -0.1in
\caption{\small Visualization of the architecture of the proposed TIFN. The details of the $3$ structure striding over the time are shown in Figure \ref{fig:net_tifn_sub}.}
\vskip -0.1in
\label{fig:net_tifn}
\end{figure}
\begin{figure}[t]
\centering
\subfigure[]{
\centering
\includegraphics[width=0.9\linewidth]{detect_ls.pdf}
\label{subfig:tifn_ls}
}
\subfigure[]{
\centering
\includegraphics[width=0.9\linewidth]{detect_ll.pdf}
\label{subfig:tifn_ll}
}
\subfigure[]{
\centering
\includegraphics[width=0.9\linewidth]{detect_gs.pdf}
\label{subfig:tifn_gs}
}
\caption{\small \subref{subfig:tifn_ls} is the local semantic information fusion module, \subref{subfig:tifn_ll} is the local location information fusion module and \subref{subfig:tifn_gs} is the global semantic information fusion module.}
\label{fig:net_tifn_sub}
\end{figure}
To detect objects in video, most works utilize optical flow motion or on the basis of two-stage detector like Faster-RCNN \cite{ren2015faster}, which are time consuming. Moreover, to train these models, dataset that objects are well-annotated in every frame is required. When category of object increases, the time cost of labelling all frames in video increases proportionally as well. In view of the great variety of categories in short video and the efficiency requirements in practical applications, we start from a fast one-stage detector YOLOv3-SPP \cite{redmon2018yolov3, he2015spatial} for single image detection, and take full advantage of temporal information in video to improve the accuracy by introducing the proposed Temporal Information Fusion Network (TIFN) without extra supervision.
Inspired by \cite{chen2020memory}, the temporal information in video can be summarized as three parts: local semantic information, local localization information and global semantic information. Local semantic information is the appearance similarity of the same object between adjacent frames, and local localization information represents the continuity of spatial information (i.e., height, width and center location) of the identical object in continuous frames. Global semantic information, which is similar to the local one, means the appearance similarity of the same object in the entire video, but due to practical implementation and efficiency, we only consider frames before the frame at current $t^*$. Figure \ref{fig:net_tifn} is an illustration of the detection module, and Figure \ref{fig:net_tifn_sub} shows the three time information fusion modules.
According to the candidate boxes predicted in each frame, the corresponding regions in the three feature maps before each detection head of three scales are cropped and saved with their spatial information, i.e., center position, width and height, denoted as $\{{\rm crop}_n^t\}_{t=1, n=1}^{T, N}$ and $\{{\rm spat}_n^t\}_{t=1, n=1}^{T, N}$.
For local semantic information, as illustrated in Figure \ref{fig:net_tifn_sub}\subref{subfig:tifn_ls}, when detecting objects at $t=t^*$, cross-correlation is calculated between each cropped feature map $\{{\rm crop}_n^t\}_{t=t^*-\tau, n=1}^{t^*-1, N}$ in the past $\tau$ frames and feature map of $f_{t^*}$
{\small\begin{equation}
\label{con:attn_ls}
{\rm Attn_{ls}}(f_{t^*}) = \frac{1}{\tau} \sum_{t=t^*-\tau}^{t^*-1}{\rm clip} \Big( {\sum_{n=1}^N{{\rm xcorr} \big( {\rm crop}_n^t, \Phi(f_{t^*}) \big) }} \Big)
\end{equation}
}where ${\rm Attn_{ls}}$ denotes the attention map of $f_{t^*}$ obtained by local semantic information, i.e., the mean cross-correlation heatmap obtained, ${\rm clip}(\cdot)$ denotes clipping value to $[0, 1]$, ${\rm xcorr}(\cdot)$ denotes calculating cross-correlation, and $\Phi(\cdot)$ denotes the neural network for feature extraction in detector.
For local localization information, as illustrated in Figure \ref{fig:net_tifn_sub}\subref{subfig:tifn_ll}, attention map is obtained simply by applying truncated 2-Dimension hanning window at the corresponding coordinates to the saved center position
{\small\begin{equation}
\label{con:attn_ll}
{\rm Attn_{ll}}(f_{t^*}) = \frac{1}{\tau}\sum_{t=t^*-\tau}^{t^*-1} \sum_{n=1}^N {\rm hanning}({\rm spat}_n^t)
\end{equation}
}where ${\rm Attn_{ll}}$ denotes the attention map of $f_{t^*}$ obtained by local localization information and ${\rm hanning}(\cdot)$ denotes applying truncated 2D hanning window with threshold $\lambda$.
For global semantic information, as illustrated in Figure \ref{fig:net_tifn_sub}\subref{subfig:tifn_gs}, the key is to utilize high-quality frames in global time domain of the input video to improve detection of local low-quality frames. Since the appearance and position of objects in short video may change rapidly and with low-quality due to the frequent video shot switching or rapid motion, local semantic information and local location information are not sufficient enough. However, high-quality frames in global with similar semantic information can be used to lead the attention to focus on objects in these low-quality frames.
To capture the global feature of objects, we build a global set $G$ with fixed size $\alpha_1$ which stores features of objects with high frequency. Similar to local semantic information, cropped features in each frame are obtained, and firstly added to a candidate pool $C$ with fixes size $\alpha_2$ ($\alpha_2 > \alpha_1$). Features in the candidate pool are matched with cropped feature of each bounding box in the next frame by similarity and updated according to the strategy as follows:
\begin{itemize}
\item if not matched and $C$ is not full, add the new feature to $C$ directly.
\item if not matched, $C$ is full and confidence of the new feature is higher than the lowest confidence in pool, replace the feature with lowest confidence.
\item if matched, recorded frequency of matched feature adds 1, and if confidence of the new feature is higher than which matched, replace it.
\end{itemize}
When the recorded frequency of features in $C$ exceed the threshold $\gamma$, features are added to the global set $G$ if $G$ is not full. If $G$ is full, the feature with lowest confidence in $G$ is replaced only if the feature in $C$ is with higher confidence and frequency. Then cross-correlation map is obtained between each cropped feature map ${\rm crop}_n^t \in G$ and feature map of $f_{t^*}$, which is similar to the local semantic information.
{\small\begin{equation}
\label{con:attn_gs}
{\rm Attn_{gs}}(f_{t^*}) = {\rm clip} \Big( \sum_{{\rm crop}_n \in G}{{\rm xcorr} \big( {\rm crop}_n, \mathcal{N}(f_{t^*}) \big) } \Big)
\end{equation}
}where ${\rm Attn_{gs}}(f_{t^*})$ denotes the attention map of $f_{t^*}$ obtained by global semantic information. All the formulations mentioned above are with no extra parameters to be trained and can be done by processing each video frame only once.
At last, $3$ groups of candidate boxes are predicted by the feature maps applied with Equation (\ref{con:attn_ls})-(\ref{con:attn_gs}) respectively, and proceeded to post-processing phase such as Non-Maximum Suppression (NMS) together to obtain the final predicted bounding boxes $B^t$.
\subsection{Tracking Module}
\noindent After the detection stage, we could obtain a series of bounding boxes $B^t=\{b_n^t\}_{n=1}^N$. The aim of the tracking module is to link each bounding box of the same object. So we have to extract the features of those bounding boxes, compute the affinity matrix between detection responses and tracklets, and assign the matched bounding boxes to the existing tracklet frame-by-frame. The paper in this part focus on the feature extraction stage. Most existing methods focus on the global features, i.e., image-level features. \cite{li2019revisiting} argues that a measure in such a level may not be effective enough in light of the scarcity of examples under our scenario. Inspired by the paper, we could obtain a more fine-grained feature based on the local features.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{network_cpsn.pdf}
\caption{\small Visualization of the architecture of the proposed CPSN.}
\label{fig:net_cpsn}
\end{figure}
In order to be consistent with the feature extractor in the following re-identification module, we first employ ResNet50 \cite{he2016deep} to learn the feature map with spatial feature. Given the cropped feature map ${\rm crop}_n \in C \times H_1 \times W_1$ derived by the detected bounding box $b_n$ ($t$ is ignored for convenience of description) and a single ${\rm crop}_d \in C \times H_2 \times W_2$ from existing tracklet, we try to obtain a pointwise response map.
$x_{ij}$ denotes the cell in feature map, i.e., the local feature of the target. For convenience of calculation, cosine similarity metric is used to measure the similarity of the local region between detection response and tracklet. Naturally, the pointwise similarity could be formulated as follows:
{\small\begin{equation}
\label{con:simi}
{\rm sim}(x^d_{ij}, x^t_{i'j'})=\cos(x^d_{ij}, x^t_{i'j'})
\end{equation}
}
{\small\begin{equation}
\label{con:simi2}
A^s=x^{d^{\mathsf{T}}}x^t
\end{equation}
}where ${\rm sim}(\cdot)$ denotes the similarity between different cells and $A^s$ denotes the pointwise response map. We can just use matrix multiplication to calculate the response map, and the amplitude of each cell of it represents the similarity of the part from the pair. So we could use the mean value of the top$k$ maximum response to approximate the original similarity of the pair. This score could exclude the disturb of the background and rapidly changing parts of foreground.
The operation above just happens on one stage. Under short-video scenarios, there always be scale changes, so the same target may occupy in different scales in different two frames. As a result, we use multi-scale feature of region in two frames to find the regions matched with highest response. Briefly, $x_1^d, x_2^d, \cdots, x_n^d$ denote the feature map in different scales of the detection response, and similarly, $x_1^t, x_2^t, \cdots, x_n^t$ denote the feature map in different scales of the tracklet. So we could integrate multiple layers in different scale and averaging the score mentioned above, and get the final similarity score.
{\small\begin{equation}
{\rm SIM}(x^d,x^t)=\frac 1 {N^2} \sum\limits_m\sum\limits_n\{{\rm top} k\{A^s_{mn}\}\}_{m=1, n=1}^{N, N}
\end{equation}
}where ${\rm top} k(\cdot)$ denotes the top$k$ values of cells in the matrix $A^s$, and ${\rm SIM}(\cdot)$ denotes the similarity between matrices, which is different from the similarity between cells in Equation (\ref{con:simi}).
For other parts such as motion prediction branch, bipartite matching, trajectory management, we just keep the same as DeepSort, due to its high efficiency.
\subsection{Re-Identification Module}
\noindent In re-identification module, tracklets of the $n$th object from $K$ shots are grouped into a complete trajectory $traj_n = \{{\rm tr}_{nk}\}_{k=1}^K$, and features of objects per frame in trajectory ${\rm traj}_n$ are sampled and averaged as the features of trajectory ${\rm traj}_n$, denoted as ${\rm feat}_n$. Then major trajectories $\{{\rm traj}_n\}_{n \in {\rm major}(N)}$ are selected by duration time and mean size, and features of major objects $\{{\rm feat}_n\}_{n \in {\rm major}(N)}$ are selected relatively. A similarity score is computed between each pair of major object features in query video and gallery video to retrieve videos with identical objects.
In retrieving objects of general categories for short videos, for person, we need to identify each instance since they are different in appearance. But for non-person object, such as animal or commodity, instances of same breed or same pattern are unrecognizable in appearance mostly. As a result, the re-identification module is divided to two branches: person branch and non-person object branch, dealing with each kind of object respectively.
In the person branch, we apply the person re-identification process with face detection and recognition, since in different short videos, the same person may dress in different clothes. As a result, a face detection and recognition model is introduced to support person re-identification. We use the PCB network \cite{sun2018beyond} as the person ReID feature extractor and ArcFace \cite{deng2019arcface} as the face feature extractor. The person similarity is decided by cosine similarity computed between each pair of features obtained by the two models together as follows
{\small\begin{equation}
S_{\rm person} = \left\{
\begin{array}{ll}
\lambda_1 \times S_{\rm ReID} + \lambda_2 \times S_{\rm Face} &,\ {\rm face}\ {\rm detected} \\
S_{\rm ReID} &,\ {\rm not}\ {\rm detected}
\end{array} \right.
\label{con:simi_person}
\end{equation}}
In the non-person object branch, since we regard unrecognizable objects of same breed or same pattern as same instance (unless they appear simultaneously), we formulate the non-person object re-identification as a fine-grained classification problem. The first four blocks of ResNet50 \cite{he2016deep} trained by ImageNet dataset \cite{deng2009imagenet} with a average pooling layer followed are applied in non-person object branch as feature extractor. ImageNet dataset consists of $1000$ fine-grained categories, as a result of which, discriminative feature can be obtained.
Furthermore, since general re-identification architectures retain a fixed query set, tracklets appeared in the first frame in most cases, to retrieve identical objects in the gallery set, which is hard to satisfy the complicated variety of appearance and pose in short videos and is liable to cause mismatching, an updating mechanism is introduced to query set in both person branch and non-person object branch in re-identification module. Setting tracklets appeared in the first frame as query set similarly, when each tracklet remained in gallery set is accessed individually, similarities between query tracklets and it are first computed. Contrast to general formulation, the accessed tracklet is updated to the query set if the difference between mean similarity among matched tracklets and mean similarity among unmatched tracklets is larger than a threshold $\Delta$. Therefore, the proposed updating mechanism can mitigate mismatching caused by complicated variety of appearance and pose in short videos.
In identical object retrieval among different videos, first the major objects are selected according to the length and the average size of predicted trajectories. Then the average features of each major object are calculated by re-identification module and used to match other features of major objects in the short video gallery according to cosine similarity metrics.
\section{Experiments}
\noindent To demonstrate the performance of the proposed generic object re-identification system for short videos, two novel datasets for evaluation are proposed in this paper, which comprised of various short videos collected from short video platform. Besides, the proposed generic object re-identification system is compared with other approaches under public benchmark, such as ImageNet VID \cite{ILSVRC15} and MOTChallenge \cite{PoseTrack}, to indicate the generalization ability.
\begin{figure}[t]
\begin{tikzpicture}[scale=0.40]
\begin{axis}[
height = 7cm,
width = 20cm,
grid=major,
x tick label style = {rotate=45,anchor=east,font=\LARGE},
y tick label style = {font=\LARGE},
ylabel style = {font=\Huge,yshift=0pt},
ylabel=Number of videos,
enlarge x limits =0.04,
enlarge y limits =0.04,
legend style={area legend,at={(0.7,0.9),font=\Huge},anchor=north, legend columns=-1},
ybar=0.45pt
bar width=25pt,
point meta=y
xticklabels={,0-5,5-10,10-15,15-20,20-25,25-30,30-35,35-40,40-45,45-50,50-55,55-60,$>=$60},
]
\addplot[draw=blue,fill=blue]
coordinates {(0,1)(1,58)(2,162)(3,80)(4,25)(5,15)(6,6)(7,4)(8,9)(9,11)(10,5)(11,6)(12,16)};
\legend{Number of videos}
\end{axis}
\end{tikzpicture}
\caption{\small Video length distribution of the proposed SVD-IOR dataset.}
\vskip -0.15in
\label{fig:video_len}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.9\linewidth]{diff_dataset.pdf}
\vskip -0.05in
\caption{Visualization of the comparison among ImageNet VID, MOTChallenge, SVD-IOR and SVD-ReID.}
\vskip -0.15in
\label{fig:diff_dataset}
\end{figure}
\subsection{Proposed Datasets}
To compare with existing methods and determine state-of-the-art approach under short-video scenarios, we propose two novel short video dataset for evaluation in identical object retrieval among different shots and different videos. Videos in the proposed datasets are both collected from real-world short video platform.
The short video dataset for identical object retrieval among different shots, abbreviated as SVD-IOR, includes $400$ videos in amount with a variety of characteristics, e.g., multiple categories, complex scenarios and frequent shot changes. Foreground objects are annotated every $10$ frames in the form of bounding boxes with corresponding track ID in each video. As is shown in Figure \ref{fig:video_len}, the lengths of most videos are distributed between $10$ and $20$ seconds. In total, it contains $232964$ frames at various resolutions, with $1221$ different identities and $66180$ bounding boxes annotated. Comparisons among ImageNet VID, MOTChallenge and the proposed SVD-IOR, SVD-ReID are shown as Figure \ref{fig:diff_dataset}.
The short video dataset for generic object re-identification among different videos, abbreviated as SVD-ReID, contains $400$ short videos. Objects of four categories, i.e., person, cat, dog, landmark, are annotated as same as which in SVD-IOR. Videos of each class contain $10$ different object instances and each instance contains $10$ different videos collected from corresponding user. Specially, track ID of each identical object among different videos is annotated as same value. The $10$ videos of each object instance are split into $2$ query videos and $8$ gallery videos, and aggregated to build a query set and a gallery set in order to evaluate generic object re-identification among different videos.
To demonstrate the generalization ability of our method on public benchmarks, the proposed system is evaluated on benchmark dataset ImageNet VID and MOTChallenge respectively as well.
\subsection{Implementation Details}
\noindent For detection module during training, we follow the settings in \cite{redmon2018yolov3} to train YOLOv3-SPP model with training set of OpenImage \cite{OpenImages2} and ImageNet VID, respectively evaluating on the proposed short video dataset and ImageNet VID validation set. During evaluation, the input videos are first split by shots with shot boundary detection approach based on color histogram \cite{mas2003video} to ensure tracklets of objects are continuous in each shot. The input video frames are resized and padded to $608 \times 608$.
For tracking module, We use the training set of MOTChallenge to train, while using the validation set to evaluate. For the proposed CPSN, we rescale the image pair to $256\times 256$ as inputs during training. With ResNet50 as the backbone, outputs of the second and forth stage are used to build the response map $A^s$ in Equation \ref{con:simi2}. It is worth noting that the calculation of the response map don't introduce any extra parameters. Circle loss and Adam optimizer are employed to optimize the proposed CPSN.
For re-identification module during training, we follow the settings in \cite{sun2018beyond} and train the PCB network as the person re-identification model with Market-1501 \cite{zheng2015scalable}, while using ResNet50 pretrained by ImageNet \cite{deng2009imagenet} as the non-person object re-identification model to extract fine-grained feature.
\subsection{Evaluation of Detection Module}
For video object detection under short video dataset, ablation experiments with SVD-IOR are applied to demonstrate the effectiveness of the proposed TIFN. Moreover, performance comparison between the proposed TIFN and the state-of-the-art video object detection models on ImageNet VID validation set are reported. Quantitative comparison results are summarized in Table \ref{tab:det_ablation} and Table \ref{tab:det_compare}.
Table \ref{tab:det_ablation} describes the ablation experiment results of aggregation of local semantic information, local localization and global semantic information under short video dataset. As is shown in the table, the proposed TIFN obtains improvement of approximately $5\%$ mAP rather than the base model YOLOv3-SPP, without extra supervision. The introduction of local semantic information contributes most, since the temporal information between adjacent frames is most significant in videos. The improvement brought by local localization and global semantic information indicate their effectiveness in dealing with the rapid change of appearance in short videos. Besides, comparison results of execution efficiency show that the proposed TIFN can still keep high efficiency after aggregating multiple temporal information.
\newsavebox{\taba}
\begin{lrbox}{\taba}
\begin{tabular}{c|ccc|c|c}
\toprule[2pt]
\multirow{2}{*}{Methods}& local & local & global & \multirow{2}{*}{mAP(\%)} & \multirow{2}{*}{FPS} \\
& semantic & localization & semantic & & \\
\midrule[1pt]
base model \cite{redmon2018yolov3, he2015spatial} & & & & 60.91 & 15.27 \\
+ ls &\checkmark & & & 64.73 & 15.01 \\
+ ls + ll &\checkmark & \checkmark & & 65.06 & 12.51 \\
+ ls + gs &\checkmark & &\checkmark & 65.47 & 10.79 \\
+ ls + ll + gs &\checkmark & \checkmark &\checkmark & \textbf{65.69} & 10.65 \\
\bottomrule[2pt]
\end{tabular}
\end{lrbox}
\begin{table}[t]
\caption{\small Ablation study on aggregation of local semantic information, local localization and global semantic information for short videos. }
\vskip -0.1in
\centering
\scalebox{0.8}{\usebox{\taba}}
\label{tab:det_ablation}
\end{table}
As is shown in Table \ref{tab:det_compare}, TIFN performs comparable accuracy and improved time efficiency to the state-of-the-art approaches. Since most methods shown in Table \ref{tab:det_compare} are formulated in much more complex structure with complicated post-processing, the proposed TIFN with a simple backbone is more suitable for real-world applications.
\newsavebox{\tabb}
\begin{lrbox}{\tabb}
\begin{tabular}{c|c|c|c}
\toprule[2pt]
Methods & Backbone & mAP(\%) & FPS \\
\midrule[1pt]
T-CNN \cite{kang2017t} & DeepID-Net+CRAFT \cite{ouyang2015deepid, baek2019character} & 73.8 & --- \\
FGFA \cite{zhu2017flow} & ResNet-101 \cite{he2016deep} & 78.4 & 1.14 \\
DoT \cite{luo2019detect} & ResNet-101 & 79.8 & --- \\
MANet \cite{wang2018fully} & ResNet-101 & 80.3 & 4.96 \\
SELSA \cite{wu2019sequence} & ResNet-101 & 80.5 & --- \\
MEGA \cite{chen2020memory} & ResNet-101 & \textbf{82.9} & 8.73 \\
\midrule[1pt]
FGFA \cite{zhu2017flow} & Inception-ResNet \cite{szegedy2017inception} & 80.1 & 1.05 \\
DoT \cite{luo2019detect} & Inception-v4 \cite{szegedy2017inception} & 82.1 & --- \\
SELSA \cite{wu2019sequence} & ResNeXt-101 \cite{xie2017aggregated} & 84.3 & --- \\
MEGA \cite{chen2020memory} & ResNeXt-101 & \textbf{85.4} & 1.10 \\
\midrule[1pt]
base model \cite{redmon2018yolov3, he2015spatial} & DarkNet53 & 74.3 & \textbf{17.40} \\
TIFN (ours) & DarkNet53 & \textbf{83.2} & \textbf{10.14} \\
\bottomrule[2pt]
\end{tabular}
\end{lrbox}
\begin{table}[H]
\caption{\small Comparison results among the proposed TIFN and other video object detection methods under ImageNet VID dataset. }
\vskip -0.1in
\centering
\scalebox{0.85}{\usebox{\tabb}}
\label{tab:det_compare}
\vskip -0.1in
\end{table}
\subsection{Evaluation of Tracking Module}
For evalution of tracking module, the proposed CPSN is compared with numerous classic and recent object tracking methods both benchmark dataset MOTChallenge and the proposed short video dataset SVD-IOR.
As is shown in Table \ref{tab:track_compare_motchallenge}, the proposed multi-object tracking method in this paper performs poor results on the public datasets, since it is not designed for the surveillance scenarios in MOTChallenge dataset. The better performance than jCC and FAMNet demonstrate the great generalization of the proposed CPSN to some extent.
\newsavebox{\tabc}
\begin{lrbox}{\tabc}
\begin{tabular}{c|c|c|c|c}
\toprule[2pt]
\diagbox{Methods}{Metrics} & MOTA & FP & FN & IDS \\
\midrule[1pt]
Tracktor++v2 \cite{bergmann2019tracking} & 56.3 & \textbf{8866} & 235449 & 1987 \\
GSM\_Tracktor \cite{liugsm} & 56.4 & 14379 & 231074 & 1485 \\
Lif\_TsimInt \cite{hornakova2020lifted} & 58.2 & 16850 & 217944 & \textbf{1022} \\
MPNTrack \cite{braso2020learning} & 58.8 & 17413 & 213594 & 1185 \\
Lif\_T \cite{hornakova2020lifted} & 60.5 & 14966 & 206619 & 1189 \\
CTTrackPub \cite{zhou2020tracking} & \textbf{61.5} & 14076 & \textbf{200672} & 2583 \\
jCC \cite{keuper2018motion} & 51.2 & 25937 & 247822 & 1802 \\
FAMNet \cite{chu2019famnet} & 52.0 & 14138 & 253616 & 3072 \\
\midrule[1pt]
CPSN(Ours) & 52.1 & 16253 & 243672 & 3012 \\
\bottomrule[2pt]
\end{tabular}
\end{lrbox}
\begin{table}[H]
\caption{\small Comparison results among the proposed system and other multiple object tracking methods under MOTChallenge dataset. }
\vskip -0.1in
\centering
\scalebox{0.85}{\usebox{\tabc}}
\label{tab:track_compare_motchallenge}
\vskip -0.1in
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{result_kdd1.pdf}
\caption{\small Visualization of the detection results (odd rows) and re-identification results (even rows).}
\label{fig:result_traj}
\end{figure}
\begin{lrbox}{\tabc}
\begin{tabular}{c|c|c|c|c|c}
\toprule[2pt]
\diagbox{Methods}{Metrics} & MOTA & FP & FN & IDS & FPS \\
\midrule[1pt]
Tracktor \cite{bergmann2019tracking} & 36.98 & 22.06 & 65.17 & 4.21 & 1.5 \\
Tracktor++v2 \cite{bergmann2019tracking} & 41.56 & 20.01 & 64.16 & 4.01 & 1.5 \\
GSM\_Tracktor \cite{liugsm} & 42.96 & 19.56 & \textbf{61.07} & 3.90 & 8.7 \\
MPNTrack \cite{braso2020learning} & \textbf{43.10} & \textbf{18.76} & 61.52 & \textbf{3.54} & 1.8 \\
\midrule[1pt]
CPSN(Ours) & 41.85 & 20.16 & 63.06 & 3.94 & \textbf{12.94}\\
\bottomrule[2pt]
\end{tabular}
\end{lrbox}
\begin{table}[t]
\caption{\small Comparison results among the proposed system and other object tracking methods under the proposed SVD-IOR dataset. }
\vskip -0.1in
\centering
\scalebox{0.85}{\usebox{\tabc}}
\label{tab:track_compare_svd}
\vskip -0.1in
\end{table}
On the contrary, Table \ref{tab:track_compare_svd} shows that our system achieves higher MOTA than Tracktor and Tracktor++v2 in short video dataset, which are the state-of-the-art video object tracking methods available in 2019. It indicates that our system has promising performance in video object tracking. Moreover, the proposed CPSN keep a higher efficiency than most of those methods.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{result_kdd2.pdf}
\caption{\small Visualization of the re-identification results among different short videos. The first column denotes the query short videos, and the following columns denote query results in order of retrieve confidence.}
\label{fig:result_reid}
\end{figure}
\subsection{Evaluation of Re-Identification Module}
For identical object trajectory extraction among different video shot in single short videos, the detection and re-identification results under the proposed SVD-IOR dataset are visualized in Figure \ref{fig:result_traj}. As illustrated in the odd rows, objects with rare categories and complicated motion can be still well detected. The corresponding re-identification results in the even rows shows that the identical object can keep same track-ID among different short video frames and shots, due to the robust detection results and the proposed system.
\newsavebox{\tabd}
\begin{lrbox}{\tabd}
\begin{tabular}{c|c|c|c|c|c|c}
\toprule[2pt]
\multirow{2}{*}{Methods} & \multirow{2}{*}{Rank-1 (\%)} & \multirow{2}{*}{Rank-5 (\%)} & \multicolumn{4}{c}{Rank-1 Accuracy of specific category (\%)} \\
\cline{4-7}
& & & person & cat & dog & landmark \\
\midrule[1pt]
baseline & 58.5 & 68.8 & 75.0 & 36.5 & 52.5 & 70.0 \\
Ours & \textbf{78.8} & \textbf{87.5} & \textbf{90.0} & \textbf{75.0} & \textbf{70.0} & \textbf{80.0} \\
\bottomrule[2pt]
\end{tabular}
\end{lrbox}
\begin{table}[t]
\caption{\small Evaluation results of the proposed system on SOT-ReID dataset. }
\vskip -0.1in
\centering
\scalebox{0.8}{\usebox{\tabd}}
\label{tab:reid}
\vskip -0.1in
\end{table}
For identical object retrieval among different videos, Table \ref{tab:reid} indicates that the proposed system achieves promising performance evaluated on SVD-ReID dataset. Compared with the baseline model, which uses YOLOv3-SPP in detection module and disables the query set update mechanism in re-identification module, the proposed system performs better in all $4$ categories. Due to the introduction of face detection and recognition, our system obtains high rank-1 accuracy in person re-identification among short videos. Moreover, landmark buildings are also well retrieved since each of them generally share a relatively consistent appearance in different videos. Cats and dogs in same breed but different ID are difficult to distinguish, as a result of which, these categories achieve a lower rank-1 accuracy. Re-identification results among different short videos are shown in Figure \ref{fig:result_reid}.
\section{Conclusion}
This paper makes two contributions to solving generic object re-identification problem for short videos. First, we propose a system composed of a detection module, a tracking module and a re-identification module, which formulate the challenging problem into three main sub-problems. In order to satisfy the high efficiency requested in practical application and get over the complicated variety of appearance of objects in short videos, we propose Temporal Information Fusion Network (TIFN) and Cross-layer Pointwise Siamese Network (CPSN) in detection module and tracking module respectively. Moreover, we propose two novel real-world short video datasets collected from short video platform for evaluating object trajectory extraction and generic object re-identification among different short videos. Quantitative experiments demonstrate the high effectiveness and efficiency of our system.
\input{KDD_paper.bbl}
\bibliographystyle{ACM-Reference-Format}
\end{document}
\endinput
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.